linux-stable/kernel/workqueue.c

6906 lines
192 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* kernel/workqueue.c - generic async execution with shared worker pool
*
* Copyright (C) 2002 Ingo Molnar
*
* Derived from the taskqueue/keventd code by:
* David Woodhouse <dwmw2@infradead.org>
* Andrew Morton
* Kai Petzke <wpp@marie.physik.tu-berlin.de>
* Theodore Ts'o <tytso@mit.edu>
*
* Made to use alloc_percpu by Christoph Lameter.
*
* Copyright (C) 2010 SUSE Linux Products GmbH
* Copyright (C) 2010 Tejun Heo <tj@kernel.org>
*
* This is the generic async execution mechanism. Work items as are
* executed in process context. The worker pool is shared and
* automatically managed. There are two worker pools for each CPU (one for
* normal work items and the other for high priority ones) and some extra
* pools for workqueues which are not bound to any specific CPU - the
* number of these backing pools is dynamic.
*
* Please read Documentation/core-api/workqueue.rst for details.
*/
#include <linux/export.h>
#include <linux/kernel.h>
#include <linux/sched.h>
#include <linux/init.h>
#include <linux/signal.h>
#include <linux/completion.h>
#include <linux/workqueue.h>
#include <linux/slab.h>
#include <linux/cpu.h>
#include <linux/notifier.h>
#include <linux/kthread.h>
#include <linux/hardirq.h>
#include <linux/mempolicy.h>
#include <linux/freezer.h>
#include <linux/debug_locks.h>
#include <linux/lockdep.h>
#include <linux/idr.h>
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
#include <linux/jhash.h>
#include <linux/hashtable.h>
#include <linux/rculist.h>
#include <linux/nodemask.h>
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
#include <linux/moduleparam.h>
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
#include <linux/uaccess.h>
#include <linux/sched/isolation.h>
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
#include <linux/sched/debug.h>
#include <linux/nmi.h>
#include <linux/kvm_para.h>
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
#include <linux/delay.h>
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
#include "workqueue_internal.h"
enum {
/*
* worker_pool flags
2012-07-17 19:39:27 +00:00
*
* A bound pool is either associated or disassociated with its CPU.
2012-07-17 19:39:27 +00:00
* While associated (!DISASSOCIATED), all workers are bound to the
* CPU and none has %WORKER_UNBOUND set and concurrency management
* is in effect.
*
* While DISASSOCIATED, the cpu may be offline and all workers have
* %WORKER_UNBOUND set and concurrency management disabled, and may
* be executing on any CPU. The pool behaves as an unbound one.
2012-07-17 19:39:27 +00:00
*
* Note that DISASSOCIATED should be flipped only while holding
* wq_pool_attach_mutex to avoid changing binding state while
* worker_attach_to_pool() is in progress.
2012-07-17 19:39:27 +00:00
*/
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
/* worker flags */
WORKER_DIE = 1 << 1, /* die die die */
WORKER_IDLE = 1 << 2, /* is idle */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
WORKER_PREP = 1 << 3, /* preparing to run works */
WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
WORKER_UNBOUND = 1 << 7, /* worker is unbound */
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
WORKER_REBOUND = 1 << 8, /* worker was rebound */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
WORKER_UNBOUND | WORKER_REBOUND,
NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
/* call for help after 10ms
(min two ticks) */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
CREATE_COOLDOWN = HZ, /* time to breath after fail */
/*
* Rescue workers are used only on emergencies and shared by
* all cpus. Give MIN_NICE.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
RESCUER_NICE_LEVEL = MIN_NICE,
HIGHPRI_NICE_LEVEL = MIN_NICE,
WQ_NAME_LEN = 24,
};
/*
* Structure fields follow one of the following exclusion rules.
*
* I: Modifiable by initialization/destruction paths and read-only for
* everyone else.
*
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* P: Preemption protected. Disabling preemption is enough and should
* only be modified and accessed from the local cpu.
*
* L: pool->lock protected. Access with pool->lock held.
*
* K: Only modified by worker while holding pool->lock. Can be safely read by
* self, while holding pool->lock or from IRQ context if %current is the
* kworker.
*
* S: Only modified by worker self.
*
* A: wq_pool_attach_mutex protected.
*
* PL: wq_pool_mutex protected.
*
* PR: wq_pool_mutex protected for writes. RCU protected for reads.
*
* PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
*
* PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
* RCU for reads.
*
* WQ: wq->mutex protected.
*
* WR: wq->mutex protected for writes. RCU protected for reads.
*
* MD: wq_mayday_lock protected.
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
*
* WD: Used internally by the watchdog.
*/
/* struct worker is defined in workqueue_internal.h */
struct worker_pool {
raw_spinlock_t lock; /* the pool lock */
int cpu; /* I: the associated cpu */
int node; /* I: the associated node ID */
int id; /* I: pool ID */
unsigned int flags; /* L: flags */
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
unsigned long watchdog_ts; /* L: watchdog timestamp */
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
bool cpu_stall; /* WD: stalled cpu bound pool */
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
/*
* The counter is incremented in a process context on the associated CPU
* w/ preemption disabled, and decremented or reset in the same context
* but w/ pool->lock held. The readers grab pool->lock and are
* guaranteed to see if the counter reached zero.
*/
int nr_running;
struct list_head worklist; /* L: list of pending works */
workqueue: reimplement idle worker rebinding Currently rebind_workers() uses rebinds idle workers synchronously before proceeding to requesting busy workers to rebind. This is necessary because all workers on @worker_pool->idle_list must be bound before concurrency management local wake-ups from the busy workers take place. Unfortunately, the synchronous idle rebinding is quite complicated. This patch reimplements idle rebinding to simplify the code path. Rather than trying to make all idle workers bound before rebinding busy workers, we simply remove all to-be-bound idle workers from the idle list and let them add themselves back after completing rebinding (successful or not). As only workers which finished rebinding can on on the idle worker list, the idle worker list is guaranteed to have only bound workers unless CPU went down again and local wake-ups are safe. After the change, @worker_pool->nr_idle may deviate than the actual number of idle workers on @worker_pool->idle_list. More specifically, nr_idle may be non-zero while ->idle_list is empty. All users of ->nr_idle and ->idle_list are audited. The only affected one is too_many_workers() which is updated to check %false if ->idle_list is empty regardless of ->nr_idle. After this patch, rebind_workers() no longer performs the nasty idle-rebind retries which require temporary release of gcwq->lock, and both unbinding and rebinding are atomic w.r.t. global_cwq->lock. worker->idle_rebind and global_cwq->rebind_hold are now unnecessary and removed along with the definition of struct idle_rebind. Changed from V1: 1) remove unlikely from too_many_workers(), ->idle_list can be empty anytime, even before this patch, no reason to use unlikely. 2) fix a small rebasing mistake. (which is from rebasing the orignal fixing patch to for-next) 3) add a lot of comments. 4) clear WORKER_REBIND unconditionaly in idle_worker_rebind() tj: Updated comments and description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2012-09-18 16:59:22 +00:00
int nr_workers; /* L: total number of workers */
int nr_idle; /* L: currently idle workers */
struct list_head idle_list; /* L: list of idle workers */
struct timer_list idle_timer; /* L: worker idle timeout */
struct work_struct idle_cull_work; /* L: worker idle cleanup */
struct timer_list mayday_timer; /* L: SOS timer for workers */
/* a workers is either on busy_hash or idle_list, or the manager */
DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
/* L: hash of busy workers */
struct worker *manager; /* L: purely informational */
struct list_head workers; /* A: attached workers */
struct list_head dying_workers; /* A: workers about to die */
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
struct completion *detach_completion; /* all workers detached */
struct ida worker_ida; /* worker IDs for task name */
struct workqueue_attrs *attrs; /* I: worker attributes */
struct hlist_node hash_node; /* PL: unbound_pool_hash node */
int refcnt; /* PL: refcnt for unbound pools */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/*
* Destruction of pool is RCU protected to allow dereferences
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
* from get_work_pool().
*/
struct rcu_head rcu;
};
/*
* Per-pool_workqueue statistics. These can be monitored using
* tools/workqueue/wq_monitor.py.
*/
enum pool_workqueue_stats {
PWQ_STAT_STARTED, /* work items started execution */
PWQ_STAT_COMPLETED, /* work items completed execution */
PWQ_STAT_CPU_TIME, /* total CPU time consumed */
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
PWQ_STAT_MAYDAY, /* maydays to rescuer */
PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
PWQ_NR_STATS,
};
/*
* The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
* of work_struct->data are used for flags and the remaining high bits
* point to the pwq; thus, pwqs need to be aligned at two's power of the
* number of flag bits.
*/
struct pool_workqueue {
struct worker_pool *pool; /* I: the associated pool */
struct workqueue_struct *wq; /* I: the owning workqueue */
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
int work_color; /* L: current color */
int flush_color; /* L: flushing color */
int refcnt; /* L: reference count */
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
int nr_in_flight[WORK_NR_COLORS];
/* L: nr of in_flight works */
workqueue: Mark barrier work with WORK_STRUCT_INACTIVE Currently, WORK_NO_COLOR has two meanings: Not participate in flushing Not participate in nr_active And only non-barrier work items are marked with WORK_STRUCT_INACTIVE when they are in inactive_works list. The barrier work items are not marked INACTIVE even linked in inactive_works list since these tail items are always moved together with the head work item. These definitions are simple, clean and practical. (Except a small blemish that only the first meaning of WORK_NO_COLOR is documented in include/linux/workqueue.h while both meanings are in workqueue.c) But dual-purpose WORK_NO_COLOR used for barrier work items has proven to be problematical[1]. Only the second purpose is obligatory. So we plan to make barrier work items participate in flushing but keep them still not participating in nr_active. So the plan is to mark barrier work items inactive without using WORK_NO_COLOR in this patch so that we can assign a flushing color to them in next patch. The reasonable way is to add or reuse a bit in work data of the work item. But adding a bit will double the size of pool_workqueue. Currently, WORK_STRUCT_INACTIVE is only used in try_to_grab_pending() for user-queued work items and try_to_grab_pending() can't work for barrier work items. So we extend WORK_STRUCT_INACTIVE to also mark barrier work items no matter which list they are in because we don't need to determind which list a barrier work item is in. So the meaning of WORK_STRUCT_INACTIVE becomes just "the work items don't participate in nr_active" (no matter whether it is a barrier work item or a user-queued work item). And WORK_STRUCT_INACTIVE for user-queued work items means they are in inactive_works list. This patch does it by setting WORK_STRUCT_INACTIVE for barrier work items in insert_wq_barrier() and checking WORK_STRUCT_INACTIVE first in pwq_dec_nr_in_flight(). And the meaning of WORK_NO_COLOR is reduced to only "not participating in flushing". There is no functionality change intended in this patch. Because WORK_NO_COLOR+WORK_STRUCT_INACTIVE represents the previous WORK_NO_COLOR in meaning and try_to_grab_pending() doesn't use for barrier work items and avoids being confused by this extended WORK_STRUCT_INACTIVE. A bunch of comment for nr_active & WORK_STRUCT_INACTIVE is also added for documenting how WORK_STRUCT_INACTIVE works in nr_active management. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:37 +00:00
/*
* nr_active management and WORK_STRUCT_INACTIVE:
*
* When pwq->nr_active >= max_active, new work item is queued to
* pwq->inactive_works instead of pool->worklist and marked with
* WORK_STRUCT_INACTIVE.
*
* All work items marked with WORK_STRUCT_INACTIVE do not participate
* in pwq->nr_active and all work items in pwq->inactive_works are
* marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
* work items are in pwq->inactive_works. Some of them are ready to
* run in pool->worklist or worker->scheduled. Those work itmes are
* only struct wq_barrier which is used for flush_work() and should
* not participate in pwq->nr_active. For non-barrier work item, it
* is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
*/
int nr_active; /* L: nr of active works */
int max_active; /* L: max active works */
struct list_head inactive_works; /* L: inactive works */
struct list_head pwqs_node; /* WR: node on wq->pwqs */
struct list_head mayday_node; /* MD: node on wq->maydays */
u64 stats[PWQ_NR_STATS];
/*
* Release of unbound pwq is punted to a kthread_worker. See put_pwq()
* and pwq_release_workfn() for details. pool_workqueue itself is also
* RCU protected so that the first pwq can be determined without
* grabbing wq->mutex.
*/
struct kthread_work release_work;
struct rcu_head rcu;
} __aligned(1 << WORK_STRUCT_FLAG_BITS);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/*
* Structure used to wait for workqueue flush.
*/
struct wq_flusher {
struct list_head list; /* WQ: list of flushers */
int flush_color; /* WQ: flush color waiting for */
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
struct completion done; /* flush completion */
};
struct wq_device;
/*
* The externally visible workqueue. It relays the issued work items to
* the appropriate worker_pool through its pool_workqueues.
*/
struct workqueue_struct {
struct list_head pwqs; /* WR: all pwqs of this wq */
struct list_head list; /* PR: list of all workqueues */
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
struct mutex mutex; /* protects this wq */
int work_color; /* WQ: current work color */
int flush_color; /* WQ: current flush color */
atomic_t nr_pwqs_to_flush; /* flush in progress */
struct wq_flusher *first_flusher; /* WQ: first flusher */
struct list_head flusher_queue; /* WQ: flush waiters */
struct list_head flusher_overflow; /* WQ: flush overflow list */
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
struct list_head maydays; /* MD: pwqs requesting rescue */
struct worker *rescuer; /* MD: rescue worker */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
int nr_drainers; /* WQ: drain in progress */
int saved_max_active; /* WQ: saved pwq max_active */
struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
#ifdef CONFIG_SYSFS
struct wq_device *wq_dev; /* I: for sysfs interface */
#endif
#ifdef CONFIG_LOCKDEP
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
char *lock_name;
struct lock_class_key key;
struct lockdep_map lockdep_map;
#endif
char name[WQ_NAME_LEN]; /* I: workqueue name */
/*
* Destruction of workqueue_struct is RCU protected to allow walking
* the workqueues list without grabbing wq_pool_mutex.
* This is used to dump all workqueues from sysrq.
*/
struct rcu_head rcu;
/* hot fields used during command issue, aligned to cacheline */
unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
};
static struct kmem_cache *pwq_cache;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
/*
* Each pod type describes how CPUs should be grouped for unbound workqueues.
* See the comment above workqueue_attrs->affn_scope.
*/
struct wq_pod_type {
int nr_pods; /* number of pods */
cpumask_var_t *pod_cpus; /* pod -> cpus */
int *pod_node; /* pod -> node */
int *cpu_pod; /* cpu -> pod */
};
static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
[WQ_AFFN_DFL] = "default",
[WQ_AFFN_CPU] = "cpu",
[WQ_AFFN_SMT] = "smt",
[WQ_AFFN_CACHE] = "cache",
[WQ_AFFN_NUMA] = "numa",
[WQ_AFFN_SYSTEM] = "system",
};
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
/*
* Per-cpu work items which run for longer than the following threshold are
* automatically considered CPU intensive and excluded from concurrency
* management to prevent them from noticeably delaying other per-cpu work items.
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
* ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
* The actual value is initialized in wq_cpu_intensive_thresh_init().
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
*/
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
/* see the comment above the definition of WQ_POWER_EFFICIENT */
static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
module_param_named(power_efficient, wq_power_efficient, bool, 0444);
static bool wq_online; /* can kworkers be created yet? */
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
static struct workqueue_attrs *wq_update_pod_attrs_buf;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
/* wait for manager to go away */
static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
static LIST_HEAD(workqueues); /* PR: list of all workqueues */
static bool workqueue_freezing; /* PL: have wqs started freezing? */
/* PL&A: allowable cpus for unbound wqs and work items */
static cpumask_var_t wq_unbound_cpumask;
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
/* PL: user requested unbound cpumask via sysfs */
static cpumask_var_t wq_requested_unbound_cpumask;
/* PL: isolated cpumask to be excluded from unbound cpumask */
static cpumask_var_t wq_isolated_cpumask;
/* for further constrain wq_unbound_cpumask by cmdline parameter*/
static struct cpumask wq_cmdline_cpumask __initdata;
/* CPU where unbound work was last round robin scheduled from this CPU */
static DEFINE_PER_CPU(int, wq_rr_cpu_last);
/*
* Local execution of unbound work items is no longer guaranteed. The
* following always forces round-robin CPU selection on unbound work items
* to uncover usages which depend on it.
*/
#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
static bool wq_debug_force_rr_cpu = true;
#else
static bool wq_debug_force_rr_cpu = false;
#endif
module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
/* the per-cpu worker pools */
tags: Fix DEFINE_PER_CPU expansions $ make tags GEN tags ctags: Warning: drivers/acpi/processor_idle.c:64: null expansion of name pattern "\1" ctags: Warning: drivers/xen/events/events_2l.c:41: null expansion of name pattern "\1" ctags: Warning: kernel/locking/lockdep.c:151: null expansion of name pattern "\1" ctags: Warning: kernel/rcu/rcutorture.c:133: null expansion of name pattern "\1" ctags: Warning: kernel/rcu/rcutorture.c:135: null expansion of name pattern "\1" ctags: Warning: kernel/workqueue.c:323: null expansion of name pattern "\1" ctags: Warning: net/ipv4/syncookies.c:53: null expansion of name pattern "\1" ctags: Warning: net/ipv6/syncookies.c:44: null expansion of name pattern "\1" ctags: Warning: net/rds/page.c:45: null expansion of name pattern "\1" Which are all the result of the DEFINE_PER_CPU pattern: scripts/tags.sh:200: '/\<DEFINE_PER_CPU([^,]*, *\([[:alnum:]_]*\)/\1/v/' scripts/tags.sh:201: '/\<DEFINE_PER_CPU_SHARED_ALIGNED([^,]*, *\([[:alnum:]_]*\)/\1/v/' The below cures them. All except the workqueue one are within reasonable distance of the 80 char limit. TJ do you have any preference on how to fix the wq one, or shall we just not care its too long? Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: David S. Miller <davem@davemloft.net> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Cc: Tejun Heo <tj@kernel.org> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-15 21:52:49 +00:00
static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
/* PL: hash of all unbound pools keyed by pool->attrs */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
/* I: attributes used when instantiating standard unbound pools on demand */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
/* I: attributes used when instantiating ordered pools on demand */
static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
/*
* I: kthread_worker to release pwq's. pwq release needs to be bounced to a
* process context while holding a pool lock. Bounce to a dedicated kthread
* worker to avoid A-A deadlocks.
*/
static struct kthread_worker *pwq_release_worker __ro_after_init;
struct workqueue_struct *system_wq __ro_after_init;
EXPORT_SYMBOL(system_wq);
struct workqueue_struct *system_highpri_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_highpri_wq);
struct workqueue_struct *system_long_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_long_wq);
struct workqueue_struct *system_unbound_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_unbound_wq);
struct workqueue_struct *system_freezable_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_freezable_wq);
struct workqueue_struct *system_power_efficient_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_power_efficient_wq);
struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
static int worker_thread(void *__worker);
static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
static void show_pwq(struct pool_workqueue *pwq);
static void show_one_worker_pool(struct worker_pool *pool);
#define CREATE_TRACE_POINTS
#include <trace/events/workqueue.h>
#define assert_rcu_or_pool_mutex() \
RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
!lockdep_is_held(&wq_pool_mutex), \
"RCU or wq_pool_mutex should be held")
#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
!lockdep_is_held(&wq->mutex) && \
!lockdep_is_held(&wq_pool_mutex), \
"RCU, wq->mutex or wq_pool_mutex should be held")
#define for_each_cpu_worker_pool(pool, cpu) \
for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
(pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
(pool)++)
/**
* for_each_pool - iterate through all worker_pools in the system
* @pool: iteration cursor
* @pi: integer used for iteration
*
* This must be called either with wq_pool_mutex held or RCU read
* locked. If the pool needs to be used beyond the locking in effect, the
* caller is responsible for guaranteeing that the pool stays online.
*
* The if/else clause exists only for the lockdep assertion and can be
* ignored.
*/
#define for_each_pool(pool, pi) \
idr_for_each_entry(&worker_pool_idr, pool, pi) \
if (({ assert_rcu_or_pool_mutex(); false; })) { } \
else
/**
* for_each_pool_worker - iterate through all workers of a worker_pool
* @worker: iteration cursor
* @pool: worker_pool to iterate workers of
*
* This must be called with wq_pool_attach_mutex.
*
* The if/else clause exists only for the lockdep assertion and can be
* ignored.
*/
#define for_each_pool_worker(worker, pool) \
list_for_each_entry((worker), &(pool)->workers, node) \
if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
else
/**
* for_each_pwq - iterate through all pool_workqueues of the specified workqueue
* @pwq: iteration cursor
* @wq: the target workqueue
*
* This must be called either with wq->mutex held or RCU read locked.
* If the pwq needs to be used beyond the locking in effect, the caller is
* responsible for guaranteeing that the pwq stays online.
*
* The if/else clause exists only for the lockdep assertion and can be
* ignored.
*/
#define for_each_pwq(pwq, wq) \
list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
lockdep_is_held(&(wq->mutex)))
#ifdef CONFIG_DEBUG_OBJECTS_WORK
static const struct debug_obj_descr work_debug_descr;
static void *work_debug_hint(void *addr)
{
return ((struct work_struct *) addr)->func;
}
debugobjects: insulate non-fixup logic related to static obj from fixup callbacks When activating a static object we need make sure that the object is tracked in the object tracker. If it is a non-static object then the activation is illegal. In previous implementation, each subsystem need take care of this in their fixup callbacks. Actually we can put it into debugobjects core. Thus we can save duplicated code, and have *pure* fixup callbacks. To achieve this, a new callback "is_static_object" is introduced to let the type specific code decide whether a object is static or not. If yes, we take it into object tracker, otherwise give warning and invoke fixup callback. This change has paassed debugobjects selftest, and I also do some test with all debugobjects supports enabled. At last, I have a concern about the fixups that can it change the object which is in incorrect state on fixup? Because the 'addr' may not point to any valid object if a non-static object is not tracked. Then Change such object can overwrite someone's memory and cause unexpected behaviour. For example, the timer_fixup_activate bind timer to function stub_timer. Link: http://lkml.kernel.org/r/1462576157-14539-1-git-send-email-changbin.du@intel.com [changbin.du@intel.com: improve code comments where invoke the new is_static_object callback] Link: http://lkml.kernel.org/r/1462777431-8171-1-git-send-email-changbin.du@intel.com Signed-off-by: Du, Changbin <changbin.du@intel.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Josh Triplett <josh@kernel.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tejun Heo <tj@kernel.org> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-05-20 00:09:41 +00:00
static bool work_is_static_object(void *addr)
{
struct work_struct *work = addr;
return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
}
/*
* fixup_init is called when:
* - an active object is initialized
*/
static bool work_fixup_init(void *addr, enum debug_obj_state state)
{
struct work_struct *work = addr;
switch (state) {
case ODEBUG_STATE_ACTIVE:
cancel_work_sync(work);
debug_object_init(work, &work_debug_descr);
return true;
default:
return false;
}
}
/*
* fixup_free is called when:
* - an active object is freed
*/
static bool work_fixup_free(void *addr, enum debug_obj_state state)
{
struct work_struct *work = addr;
switch (state) {
case ODEBUG_STATE_ACTIVE:
cancel_work_sync(work);
debug_object_free(work, &work_debug_descr);
return true;
default:
return false;
}
}
static const struct debug_obj_descr work_debug_descr = {
.name = "work_struct",
.debug_hint = work_debug_hint,
debugobjects: insulate non-fixup logic related to static obj from fixup callbacks When activating a static object we need make sure that the object is tracked in the object tracker. If it is a non-static object then the activation is illegal. In previous implementation, each subsystem need take care of this in their fixup callbacks. Actually we can put it into debugobjects core. Thus we can save duplicated code, and have *pure* fixup callbacks. To achieve this, a new callback "is_static_object" is introduced to let the type specific code decide whether a object is static or not. If yes, we take it into object tracker, otherwise give warning and invoke fixup callback. This change has paassed debugobjects selftest, and I also do some test with all debugobjects supports enabled. At last, I have a concern about the fixups that can it change the object which is in incorrect state on fixup? Because the 'addr' may not point to any valid object if a non-static object is not tracked. Then Change such object can overwrite someone's memory and cause unexpected behaviour. For example, the timer_fixup_activate bind timer to function stub_timer. Link: http://lkml.kernel.org/r/1462576157-14539-1-git-send-email-changbin.du@intel.com [changbin.du@intel.com: improve code comments where invoke the new is_static_object callback] Link: http://lkml.kernel.org/r/1462777431-8171-1-git-send-email-changbin.du@intel.com Signed-off-by: Du, Changbin <changbin.du@intel.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Josh Triplett <josh@kernel.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tejun Heo <tj@kernel.org> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-05-20 00:09:41 +00:00
.is_static_object = work_is_static_object,
.fixup_init = work_fixup_init,
.fixup_free = work_fixup_free,
};
static inline void debug_work_activate(struct work_struct *work)
{
debug_object_activate(work, &work_debug_descr);
}
static inline void debug_work_deactivate(struct work_struct *work)
{
debug_object_deactivate(work, &work_debug_descr);
}
void __init_work(struct work_struct *work, int onstack)
{
if (onstack)
debug_object_init_on_stack(work, &work_debug_descr);
else
debug_object_init(work, &work_debug_descr);
}
EXPORT_SYMBOL_GPL(__init_work);
void destroy_work_on_stack(struct work_struct *work)
{
debug_object_free(work, &work_debug_descr);
}
EXPORT_SYMBOL_GPL(destroy_work_on_stack);
void destroy_delayed_work_on_stack(struct delayed_work *work)
{
destroy_timer_on_stack(&work->timer);
debug_object_free(&work->work, &work_debug_descr);
}
EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
#else
static inline void debug_work_activate(struct work_struct *work) { }
static inline void debug_work_deactivate(struct work_struct *work) { }
#endif
/**
* worker_pool_assign_id - allocate ID and assign it to @pool
* @pool: the pool pointer of interest
*
* Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
* successfully, -errno on failure.
*/
static int worker_pool_assign_id(struct worker_pool *pool)
{
int ret;
lockdep_assert_held(&wq_pool_mutex);
ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
GFP_KERNEL);
Linux 3.9-rc5 -----BEGIN PGP SIGNATURE----- Version: GnuPG v2.0.19 (GNU/Linux) iQEcBAABAgAGBQJRWLTrAAoJEHm+PkMAQRiGe8oH/iMy48mecVWvxVZn74Tx3Cef xmW/PnAIj28EhSPqK49N/Ow6AfQToFKf7AP0ge20KAf5teTq95AY+tH74DAANt8F BjKXXTZiR5xwBvRkq7CR5wDcCvEcBAAz8fgTEd6SEDB2d2VXFf5eKdKUqt1avTCh Z6Hup5kuwX+ddtwY2DCBXtp2n6fL0Rm5yLzY1A3OOBye1E7VyLTF7M5BR603Q44P 4kRLxn8+R7jy3hTuZIhAeoS8TKUoBwVk7DmKxEzrhTHZVOmvwE9lEHybRnIyOpd/ k1JnbRbiPsLsCVFOn10SQkGDAIk00lro3tuWP2C1ljERiD/OOh5Ui9nXYAhMkbI= =q15K -----END PGP SIGNATURE----- Merge tag 'v3.9-rc5' into wq/for-3.10 Writeback conversion to workqueue will be based on top of wq/for-3.10 branch to take advantage of custom attrs and NUMA support for unbound workqueues. Mainline currently contains two commits which result in non-trivial merge conflicts with wq/for-3.10 and because block/for-3.10/core is based on v3.9-rc3 which contains one of the conflicting commits, we need a pre-merge-window merge anyway. Let's pull v3.9-rc5 into wq/for-3.10 so that the block tree doesn't suffer from workqueue merge conflicts. The two conflicts and their resolutions: * e68035fb65 ("workqueue: convert to idr_alloc()") in mainline changes worker_pool_assign_id() to use idr_alloc() instead of the old idr interface. worker_pool_assign_id() goes through multiple locking changes in wq/for-3.10 causing the following conflict. static int worker_pool_assign_id(struct worker_pool *pool) { int ret; <<<<<<< HEAD lockdep_assert_held(&wq_pool_mutex); do { if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) return -ENOMEM; ret = idr_get_new(&worker_pool_idr, pool, &pool->id); } while (ret == -EAGAIN); ======= mutex_lock(&worker_pool_idr_mutex); ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); if (ret >= 0) pool->id = ret; mutex_unlock(&worker_pool_idr_mutex); >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 return ret < 0 ? ret : 0; } We want locking from the former and idr_alloc() usage from the latter, which can be combined to the following. static int worker_pool_assign_id(struct worker_pool *pool) { int ret; lockdep_assert_held(&wq_pool_mutex); ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); if (ret >= 0) { pool->id = ret; return 0; } return ret; } * eb2834285c ("workqueue: fix possible pool stall bug in wq_unbind_fn()") updated wq_unbind_fn() such that it has single larger for_each_std_worker_pool() loop instead of two separate loops with a schedule() call inbetween. wq/for-3.10 renamed pool->assoc_mutex to pool->manager_mutex causing the following conflict (earlier function body and comments omitted for brevity). static void wq_unbind_fn(struct work_struct *work) { ... spin_unlock_irq(&pool->lock); <<<<<<< HEAD mutex_unlock(&pool->manager_mutex); } ======= mutex_unlock(&pool->assoc_mutex); >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 schedule(); <<<<<<< HEAD for_each_cpu_worker_pool(pool, cpu) ======= >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 atomic_set(&pool->nr_running, 0); spin_lock_irq(&pool->lock); wake_up_worker(pool); spin_unlock_irq(&pool->lock); } } The resolution is mostly trivial. We want the control flow of the latter with the rename of the former. static void wq_unbind_fn(struct work_struct *work) { ... spin_unlock_irq(&pool->lock); mutex_unlock(&pool->manager_mutex); schedule(); atomic_set(&pool->nr_running, 0); spin_lock_irq(&pool->lock); wake_up_worker(pool); spin_unlock_irq(&pool->lock); } } Signed-off-by: Tejun Heo <tj@kernel.org>
2013-04-02 00:08:13 +00:00
if (ret >= 0) {
pool->id = ret;
Linux 3.9-rc5 -----BEGIN PGP SIGNATURE----- Version: GnuPG v2.0.19 (GNU/Linux) iQEcBAABAgAGBQJRWLTrAAoJEHm+PkMAQRiGe8oH/iMy48mecVWvxVZn74Tx3Cef xmW/PnAIj28EhSPqK49N/Ow6AfQToFKf7AP0ge20KAf5teTq95AY+tH74DAANt8F BjKXXTZiR5xwBvRkq7CR5wDcCvEcBAAz8fgTEd6SEDB2d2VXFf5eKdKUqt1avTCh Z6Hup5kuwX+ddtwY2DCBXtp2n6fL0Rm5yLzY1A3OOBye1E7VyLTF7M5BR603Q44P 4kRLxn8+R7jy3hTuZIhAeoS8TKUoBwVk7DmKxEzrhTHZVOmvwE9lEHybRnIyOpd/ k1JnbRbiPsLsCVFOn10SQkGDAIk00lro3tuWP2C1ljERiD/OOh5Ui9nXYAhMkbI= =q15K -----END PGP SIGNATURE----- Merge tag 'v3.9-rc5' into wq/for-3.10 Writeback conversion to workqueue will be based on top of wq/for-3.10 branch to take advantage of custom attrs and NUMA support for unbound workqueues. Mainline currently contains two commits which result in non-trivial merge conflicts with wq/for-3.10 and because block/for-3.10/core is based on v3.9-rc3 which contains one of the conflicting commits, we need a pre-merge-window merge anyway. Let's pull v3.9-rc5 into wq/for-3.10 so that the block tree doesn't suffer from workqueue merge conflicts. The two conflicts and their resolutions: * e68035fb65 ("workqueue: convert to idr_alloc()") in mainline changes worker_pool_assign_id() to use idr_alloc() instead of the old idr interface. worker_pool_assign_id() goes through multiple locking changes in wq/for-3.10 causing the following conflict. static int worker_pool_assign_id(struct worker_pool *pool) { int ret; <<<<<<< HEAD lockdep_assert_held(&wq_pool_mutex); do { if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) return -ENOMEM; ret = idr_get_new(&worker_pool_idr, pool, &pool->id); } while (ret == -EAGAIN); ======= mutex_lock(&worker_pool_idr_mutex); ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); if (ret >= 0) pool->id = ret; mutex_unlock(&worker_pool_idr_mutex); >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 return ret < 0 ? ret : 0; } We want locking from the former and idr_alloc() usage from the latter, which can be combined to the following. static int worker_pool_assign_id(struct worker_pool *pool) { int ret; lockdep_assert_held(&wq_pool_mutex); ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); if (ret >= 0) { pool->id = ret; return 0; } return ret; } * eb2834285c ("workqueue: fix possible pool stall bug in wq_unbind_fn()") updated wq_unbind_fn() such that it has single larger for_each_std_worker_pool() loop instead of two separate loops with a schedule() call inbetween. wq/for-3.10 renamed pool->assoc_mutex to pool->manager_mutex causing the following conflict (earlier function body and comments omitted for brevity). static void wq_unbind_fn(struct work_struct *work) { ... spin_unlock_irq(&pool->lock); <<<<<<< HEAD mutex_unlock(&pool->manager_mutex); } ======= mutex_unlock(&pool->assoc_mutex); >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 schedule(); <<<<<<< HEAD for_each_cpu_worker_pool(pool, cpu) ======= >>>>>>> c67bf5361e7e66a0ff1f4caf95f89347d55dfb89 atomic_set(&pool->nr_running, 0); spin_lock_irq(&pool->lock); wake_up_worker(pool); spin_unlock_irq(&pool->lock); } } The resolution is mostly trivial. We want the control flow of the latter with the rename of the former. static void wq_unbind_fn(struct work_struct *work) { ... spin_unlock_irq(&pool->lock); mutex_unlock(&pool->manager_mutex); schedule(); atomic_set(&pool->nr_running, 0); spin_lock_irq(&pool->lock); wake_up_worker(pool); spin_unlock_irq(&pool->lock); } } Signed-off-by: Tejun Heo <tj@kernel.org>
2013-04-02 00:08:13 +00:00
return 0;
}
return ret;
}
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
static unsigned int work_color_to_flags(int color)
{
return color << WORK_STRUCT_COLOR_SHIFT;
}
static int get_work_color(unsigned long work_data)
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
{
return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
((1 << WORK_STRUCT_COLOR_BITS) - 1);
}
static int work_next_color(int color)
{
return (color + 1) % WORK_NR_COLORS;
}
/*
* While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
* contain the pointer to the queued pwq. Once execution starts, the flag
* is cleared and the high bits contain OFFQ flags and pool ID.
*
* set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
* and clear_work_data() can be used to set the pwq, pool or clear
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* work->data. These functions should only be called while the work is
* owned - ie. while the PENDING bit is set.
*
* get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
* corresponding to a work. Pool is available once the work has been
* queued anywhere after initialization until it is sync canceled. pwq is
* available only while the work item is queued.
*
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* %WORK_OFFQ_CANCELING is used to mark a work item which is being
* canceled. While being canceled, a work item may have its PENDING set
* but stay off timer and worklist for arbitrarily long and nobody should
* try to steal the PENDING bit.
*/
static inline void set_work_data(struct work_struct *work, unsigned long data,
unsigned long flags)
{
WARN_ON_ONCE(!work_pending(work));
atomic_long_set(&work->data, data | flags | work_static(work));
}
static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
unsigned long extra_flags)
{
set_work_data(work, (unsigned long)pwq,
WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
}
static void set_work_pool_and_keep_pending(struct work_struct *work,
int pool_id)
{
set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
WORK_STRUCT_PENDING);
}
static void set_work_pool_and_clear_pending(struct work_struct *work,
int pool_id)
{
/*
* The following wmb is paired with the implied mb in
* test_and_set_bit(PENDING) and ensures all updates to @work made
* here are visible to and precede any updates by the next PENDING
* owner.
*/
smp_wmb();
set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
workqueue: fix ghost PENDING flag while doing MQ IO The bug in a workqueue leads to a stalled IO request in MQ ctx->rq_list with the following backtrace: [ 601.347452] INFO: task kworker/u129:5:1636 blocked for more than 120 seconds. [ 601.347574] Tainted: G O 4.4.5-1-storage+ #6 [ 601.347651] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 601.348142] kworker/u129:5 D ffff880803077988 0 1636 2 0x00000000 [ 601.348519] Workqueue: ibnbd_server_fileio_wq ibnbd_dev_file_submit_io_worker [ibnbd_server] [ 601.348999] ffff880803077988 ffff88080466b900 ffff8808033f9c80 ffff880803078000 [ 601.349662] ffff880807c95000 7fffffffffffffff ffffffff815b0920 ffff880803077ad0 [ 601.350333] ffff8808030779a0 ffffffff815b01d5 0000000000000000 ffff880803077a38 [ 601.350965] Call Trace: [ 601.351203] [<ffffffff815b0920>] ? bit_wait+0x60/0x60 [ 601.351444] [<ffffffff815b01d5>] schedule+0x35/0x80 [ 601.351709] [<ffffffff815b2dd2>] schedule_timeout+0x192/0x230 [ 601.351958] [<ffffffff812d43f7>] ? blk_flush_plug_list+0xc7/0x220 [ 601.352208] [<ffffffff810bd737>] ? ktime_get+0x37/0xa0 [ 601.352446] [<ffffffff815b0920>] ? bit_wait+0x60/0x60 [ 601.352688] [<ffffffff815af784>] io_schedule_timeout+0xa4/0x110 [ 601.352951] [<ffffffff815b3a4e>] ? _raw_spin_unlock_irqrestore+0xe/0x10 [ 601.353196] [<ffffffff815b093b>] bit_wait_io+0x1b/0x70 [ 601.353440] [<ffffffff815b056d>] __wait_on_bit+0x5d/0x90 [ 601.353689] [<ffffffff81127bd0>] wait_on_page_bit+0xc0/0xd0 [ 601.353958] [<ffffffff81096db0>] ? autoremove_wake_function+0x40/0x40 [ 601.354200] [<ffffffff81127cc4>] __filemap_fdatawait_range+0xe4/0x140 [ 601.354441] [<ffffffff81127d34>] filemap_fdatawait_range+0x14/0x30 [ 601.354688] [<ffffffff81129a9f>] filemap_write_and_wait_range+0x3f/0x70 [ 601.354932] [<ffffffff811ced3b>] blkdev_fsync+0x1b/0x50 [ 601.355193] [<ffffffff811c82d9>] vfs_fsync_range+0x49/0xa0 [ 601.355432] [<ffffffff811cf45a>] blkdev_write_iter+0xca/0x100 [ 601.355679] [<ffffffff81197b1a>] __vfs_write+0xaa/0xe0 [ 601.355925] [<ffffffff81198379>] vfs_write+0xa9/0x1a0 [ 601.356164] [<ffffffff811c59d8>] kernel_write+0x38/0x50 The underlying device is a null_blk, with default parameters: queue_mode = MQ submit_queues = 1 Verification that nullb0 has something inflight: root@pserver8:~# cat /sys/block/nullb0/inflight 0 1 root@pserver8:~# find /sys/block/nullb0/mq/0/cpu* -name rq_list -print -exec cat {} \; ... /sys/block/nullb0/mq/0/cpu2/rq_list CTX pending: ffff8838038e2400 ... During debug it became clear that stalled request is always inserted in the rq_list from the following path: save_stack_trace_tsk + 34 blk_mq_insert_requests + 231 blk_mq_flush_plug_list + 281 blk_flush_plug_list + 199 wait_on_page_bit + 192 __filemap_fdatawait_range + 228 filemap_fdatawait_range + 20 filemap_write_and_wait_range + 63 blkdev_fsync + 27 vfs_fsync_range + 73 blkdev_write_iter + 202 __vfs_write + 170 vfs_write + 169 kernel_write + 56 So blk_flush_plug_list() was called with from_schedule == true. If from_schedule is true, that means that finally blk_mq_insert_requests() offloads execution of __blk_mq_run_hw_queue() and uses kblockd workqueue, i.e. it calls kblockd_schedule_delayed_work_on(). That means, that we race with another CPU, which is about to execute __blk_mq_run_hw_queue() work. Further debugging shows the following traces from different CPUs: CPU#0 CPU#1 ---------------------------------- ------------------------------- reqeust A inserted STORE hctx->ctx_map[0] bit marked kblockd_schedule...() returns 1 <schedule to kblockd workqueue> request B inserted STORE hctx->ctx_map[1] bit marked kblockd_schedule...() returns 0 *** WORK PENDING bit is cleared *** flush_busy_ctxs() is executed, but bit 1, set by CPU#1, is not observed As a result request B pended forever. This behaviour can be explained by speculative LOAD of hctx->ctx_map on CPU#0, which is reordered with clear of PENDING bit and executed _before_ actual STORE of bit 1 on CPU#1. The proper fix is an explicit full barrier <mfence>, which guarantees that clear of PENDING bit is to be executed before all possible speculative LOADS or STORES inside actual work function. Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Cc: Gioh Kim <gi-oh.kim@profitbricks.com> Cc: Michael Wang <yun.wang@profitbricks.com> Cc: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: stable@vger.kernel.org Signed-off-by: Tejun Heo <tj@kernel.org>
2016-04-26 11:15:35 +00:00
/*
* The following mb guarantees that previous clear of a PENDING bit
* will not be reordered with any speculative LOADS or STORES from
* work->current_func, which is executed afterwards. This possible
* reordering can lead to a missed execution on attempt to queue
workqueue: fix ghost PENDING flag while doing MQ IO The bug in a workqueue leads to a stalled IO request in MQ ctx->rq_list with the following backtrace: [ 601.347452] INFO: task kworker/u129:5:1636 blocked for more than 120 seconds. [ 601.347574] Tainted: G O 4.4.5-1-storage+ #6 [ 601.347651] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 601.348142] kworker/u129:5 D ffff880803077988 0 1636 2 0x00000000 [ 601.348519] Workqueue: ibnbd_server_fileio_wq ibnbd_dev_file_submit_io_worker [ibnbd_server] [ 601.348999] ffff880803077988 ffff88080466b900 ffff8808033f9c80 ffff880803078000 [ 601.349662] ffff880807c95000 7fffffffffffffff ffffffff815b0920 ffff880803077ad0 [ 601.350333] ffff8808030779a0 ffffffff815b01d5 0000000000000000 ffff880803077a38 [ 601.350965] Call Trace: [ 601.351203] [<ffffffff815b0920>] ? bit_wait+0x60/0x60 [ 601.351444] [<ffffffff815b01d5>] schedule+0x35/0x80 [ 601.351709] [<ffffffff815b2dd2>] schedule_timeout+0x192/0x230 [ 601.351958] [<ffffffff812d43f7>] ? blk_flush_plug_list+0xc7/0x220 [ 601.352208] [<ffffffff810bd737>] ? ktime_get+0x37/0xa0 [ 601.352446] [<ffffffff815b0920>] ? bit_wait+0x60/0x60 [ 601.352688] [<ffffffff815af784>] io_schedule_timeout+0xa4/0x110 [ 601.352951] [<ffffffff815b3a4e>] ? _raw_spin_unlock_irqrestore+0xe/0x10 [ 601.353196] [<ffffffff815b093b>] bit_wait_io+0x1b/0x70 [ 601.353440] [<ffffffff815b056d>] __wait_on_bit+0x5d/0x90 [ 601.353689] [<ffffffff81127bd0>] wait_on_page_bit+0xc0/0xd0 [ 601.353958] [<ffffffff81096db0>] ? autoremove_wake_function+0x40/0x40 [ 601.354200] [<ffffffff81127cc4>] __filemap_fdatawait_range+0xe4/0x140 [ 601.354441] [<ffffffff81127d34>] filemap_fdatawait_range+0x14/0x30 [ 601.354688] [<ffffffff81129a9f>] filemap_write_and_wait_range+0x3f/0x70 [ 601.354932] [<ffffffff811ced3b>] blkdev_fsync+0x1b/0x50 [ 601.355193] [<ffffffff811c82d9>] vfs_fsync_range+0x49/0xa0 [ 601.355432] [<ffffffff811cf45a>] blkdev_write_iter+0xca/0x100 [ 601.355679] [<ffffffff81197b1a>] __vfs_write+0xaa/0xe0 [ 601.355925] [<ffffffff81198379>] vfs_write+0xa9/0x1a0 [ 601.356164] [<ffffffff811c59d8>] kernel_write+0x38/0x50 The underlying device is a null_blk, with default parameters: queue_mode = MQ submit_queues = 1 Verification that nullb0 has something inflight: root@pserver8:~# cat /sys/block/nullb0/inflight 0 1 root@pserver8:~# find /sys/block/nullb0/mq/0/cpu* -name rq_list -print -exec cat {} \; ... /sys/block/nullb0/mq/0/cpu2/rq_list CTX pending: ffff8838038e2400 ... During debug it became clear that stalled request is always inserted in the rq_list from the following path: save_stack_trace_tsk + 34 blk_mq_insert_requests + 231 blk_mq_flush_plug_list + 281 blk_flush_plug_list + 199 wait_on_page_bit + 192 __filemap_fdatawait_range + 228 filemap_fdatawait_range + 20 filemap_write_and_wait_range + 63 blkdev_fsync + 27 vfs_fsync_range + 73 blkdev_write_iter + 202 __vfs_write + 170 vfs_write + 169 kernel_write + 56 So blk_flush_plug_list() was called with from_schedule == true. If from_schedule is true, that means that finally blk_mq_insert_requests() offloads execution of __blk_mq_run_hw_queue() and uses kblockd workqueue, i.e. it calls kblockd_schedule_delayed_work_on(). That means, that we race with another CPU, which is about to execute __blk_mq_run_hw_queue() work. Further debugging shows the following traces from different CPUs: CPU#0 CPU#1 ---------------------------------- ------------------------------- reqeust A inserted STORE hctx->ctx_map[0] bit marked kblockd_schedule...() returns 1 <schedule to kblockd workqueue> request B inserted STORE hctx->ctx_map[1] bit marked kblockd_schedule...() returns 0 *** WORK PENDING bit is cleared *** flush_busy_ctxs() is executed, but bit 1, set by CPU#1, is not observed As a result request B pended forever. This behaviour can be explained by speculative LOAD of hctx->ctx_map on CPU#0, which is reordered with clear of PENDING bit and executed _before_ actual STORE of bit 1 on CPU#1. The proper fix is an explicit full barrier <mfence>, which guarantees that clear of PENDING bit is to be executed before all possible speculative LOADS or STORES inside actual work function. Signed-off-by: Roman Pen <roman.penyaev@profitbricks.com> Cc: Gioh Kim <gi-oh.kim@profitbricks.com> Cc: Michael Wang <yun.wang@profitbricks.com> Cc: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: linux-block@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: stable@vger.kernel.org Signed-off-by: Tejun Heo <tj@kernel.org>
2016-04-26 11:15:35 +00:00
* the same @work. E.g. consider this case:
*
* CPU#0 CPU#1
* ---------------------------- --------------------------------
*
* 1 STORE event_indicated
* 2 queue_work_on() {
* 3 test_and_set_bit(PENDING)
* 4 } set_..._and_clear_pending() {
* 5 set_work_data() # clear bit
* 6 smp_mb()
* 7 work->current_func() {
* 8 LOAD event_indicated
* }
*
* Without an explicit full barrier speculative LOAD on line 8 can
* be executed before CPU#0 does STORE on line 1. If that happens,
* CPU#0 observes the PENDING bit is still set and new execution of
* a @work is not queued in a hope, that CPU#1 will eventually
* finish the queued @work. Meanwhile CPU#1 does not see
* event_indicated is set, because speculative LOAD was executed
* before actual STORE.
*/
smp_mb();
}
static void clear_work_data(struct work_struct *work)
{
smp_wmb(); /* see set_work_pool_and_clear_pending() */
set_work_data(work, WORK_STRUCT_NO_POOL, 0);
}
workqueue: clean up WORK_* constant types, clarify masking Dave Airlie reports that gcc-13.1.1 has started complaining about some of the workqueue code in 32-bit arm builds: kernel/workqueue.c: In function ‘get_work_pwq’: kernel/workqueue.c:713:24: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast] 713 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); | ^ [ ... a couple of other cases ... ] and while it's not immediately clear exactly why gcc started complaining about it now, I suspect it's some C23-induced enum type handlign fixup in gcc-13 is the cause. Whatever the reason for starting to complain, the code and data types are indeed disgusting enough that the complaint is warranted. The wq code ends up creating various "helper constants" (like that WORK_STRUCT_WQ_DATA_MASK) using an enum type, which is all kinds of confused. The mask needs to be 'unsigned long', not some unspecified enum type. To make matters worse, the actual "mask and cast to a pointer" is repeated a couple of times, and the cast isn't even always done to the right pointer, but - as the error case above - to a 'void *' with then the compiler finishing the job. That's now how we roll in the kernel. So create the masks using the proper types rather than some ambiguous enumeration, and use a nice helper that actually does the type conversion in one well-defined place. Incidentally, this magically makes clang generate better code. That, admittedly, is really just a sign of clang having been seriously confused before, and cleaning up the typing unconfuses the compiler too. Reported-by: Dave Airlie <airlied@gmail.com> Link: https://lore.kernel.org/lkml/CAPM=9twNnV4zMCvrPkw3H-ajZOH-01JVh_kDrxdPYQErz8ZTdA@mail.gmail.com/ Cc: Arnd Bergmann <arnd@arndb.de> Cc: Tejun Heo <tj@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Nathan Chancellor <nathan@kernel.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-06-23 19:08:14 +00:00
static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
{
return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
}
static struct pool_workqueue *get_work_pwq(struct work_struct *work)
{
unsigned long data = atomic_long_read(&work->data);
if (data & WORK_STRUCT_PWQ)
workqueue: clean up WORK_* constant types, clarify masking Dave Airlie reports that gcc-13.1.1 has started complaining about some of the workqueue code in 32-bit arm builds: kernel/workqueue.c: In function ‘get_work_pwq’: kernel/workqueue.c:713:24: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast] 713 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); | ^ [ ... a couple of other cases ... ] and while it's not immediately clear exactly why gcc started complaining about it now, I suspect it's some C23-induced enum type handlign fixup in gcc-13 is the cause. Whatever the reason for starting to complain, the code and data types are indeed disgusting enough that the complaint is warranted. The wq code ends up creating various "helper constants" (like that WORK_STRUCT_WQ_DATA_MASK) using an enum type, which is all kinds of confused. The mask needs to be 'unsigned long', not some unspecified enum type. To make matters worse, the actual "mask and cast to a pointer" is repeated a couple of times, and the cast isn't even always done to the right pointer, but - as the error case above - to a 'void *' with then the compiler finishing the job. That's now how we roll in the kernel. So create the masks using the proper types rather than some ambiguous enumeration, and use a nice helper that actually does the type conversion in one well-defined place. Incidentally, this magically makes clang generate better code. That, admittedly, is really just a sign of clang having been seriously confused before, and cleaning up the typing unconfuses the compiler too. Reported-by: Dave Airlie <airlied@gmail.com> Link: https://lore.kernel.org/lkml/CAPM=9twNnV4zMCvrPkw3H-ajZOH-01JVh_kDrxdPYQErz8ZTdA@mail.gmail.com/ Cc: Arnd Bergmann <arnd@arndb.de> Cc: Tejun Heo <tj@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Nathan Chancellor <nathan@kernel.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-06-23 19:08:14 +00:00
return work_struct_pwq(data);
else
return NULL;
}
/**
* get_work_pool - return the worker_pool a given work was associated with
* @work: the work item of interest
*
* Pools are created and destroyed under wq_pool_mutex, and allows read
* access under RCU read lock. As such, this function should be
* called under wq_pool_mutex or inside of a rcu_read_lock() region.
*
* All fields of the returned pool are accessible as long as the above
* mentioned locking is in effect. If the returned pool needs to be used
* beyond the critical section, the caller is responsible for ensuring the
* returned pool is and stays online.
*
* Return: The worker_pool @work was last associated with. %NULL if none.
*/
static struct worker_pool *get_work_pool(struct work_struct *work)
{
unsigned long data = atomic_long_read(&work->data);
int pool_id;
assert_rcu_or_pool_mutex();
if (data & WORK_STRUCT_PWQ)
workqueue: clean up WORK_* constant types, clarify masking Dave Airlie reports that gcc-13.1.1 has started complaining about some of the workqueue code in 32-bit arm builds: kernel/workqueue.c: In function ‘get_work_pwq’: kernel/workqueue.c:713:24: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast] 713 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); | ^ [ ... a couple of other cases ... ] and while it's not immediately clear exactly why gcc started complaining about it now, I suspect it's some C23-induced enum type handlign fixup in gcc-13 is the cause. Whatever the reason for starting to complain, the code and data types are indeed disgusting enough that the complaint is warranted. The wq code ends up creating various "helper constants" (like that WORK_STRUCT_WQ_DATA_MASK) using an enum type, which is all kinds of confused. The mask needs to be 'unsigned long', not some unspecified enum type. To make matters worse, the actual "mask and cast to a pointer" is repeated a couple of times, and the cast isn't even always done to the right pointer, but - as the error case above - to a 'void *' with then the compiler finishing the job. That's now how we roll in the kernel. So create the masks using the proper types rather than some ambiguous enumeration, and use a nice helper that actually does the type conversion in one well-defined place. Incidentally, this magically makes clang generate better code. That, admittedly, is really just a sign of clang having been seriously confused before, and cleaning up the typing unconfuses the compiler too. Reported-by: Dave Airlie <airlied@gmail.com> Link: https://lore.kernel.org/lkml/CAPM=9twNnV4zMCvrPkw3H-ajZOH-01JVh_kDrxdPYQErz8ZTdA@mail.gmail.com/ Cc: Arnd Bergmann <arnd@arndb.de> Cc: Tejun Heo <tj@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Nathan Chancellor <nathan@kernel.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-06-23 19:08:14 +00:00
return work_struct_pwq(data)->pool;
pool_id = data >> WORK_OFFQ_POOL_SHIFT;
if (pool_id == WORK_OFFQ_POOL_NONE)
return NULL;
return idr_find(&worker_pool_idr, pool_id);
}
/**
* get_work_pool_id - return the worker pool ID a given work is associated with
* @work: the work item of interest
*
* Return: The worker_pool ID @work was last associated with.
* %WORK_OFFQ_POOL_NONE if none.
*/
static int get_work_pool_id(struct work_struct *work)
{
unsigned long data = atomic_long_read(&work->data);
if (data & WORK_STRUCT_PWQ)
workqueue: clean up WORK_* constant types, clarify masking Dave Airlie reports that gcc-13.1.1 has started complaining about some of the workqueue code in 32-bit arm builds: kernel/workqueue.c: In function ‘get_work_pwq’: kernel/workqueue.c:713:24: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast] 713 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); | ^ [ ... a couple of other cases ... ] and while it's not immediately clear exactly why gcc started complaining about it now, I suspect it's some C23-induced enum type handlign fixup in gcc-13 is the cause. Whatever the reason for starting to complain, the code and data types are indeed disgusting enough that the complaint is warranted. The wq code ends up creating various "helper constants" (like that WORK_STRUCT_WQ_DATA_MASK) using an enum type, which is all kinds of confused. The mask needs to be 'unsigned long', not some unspecified enum type. To make matters worse, the actual "mask and cast to a pointer" is repeated a couple of times, and the cast isn't even always done to the right pointer, but - as the error case above - to a 'void *' with then the compiler finishing the job. That's now how we roll in the kernel. So create the masks using the proper types rather than some ambiguous enumeration, and use a nice helper that actually does the type conversion in one well-defined place. Incidentally, this magically makes clang generate better code. That, admittedly, is really just a sign of clang having been seriously confused before, and cleaning up the typing unconfuses the compiler too. Reported-by: Dave Airlie <airlied@gmail.com> Link: https://lore.kernel.org/lkml/CAPM=9twNnV4zMCvrPkw3H-ajZOH-01JVh_kDrxdPYQErz8ZTdA@mail.gmail.com/ Cc: Arnd Bergmann <arnd@arndb.de> Cc: Tejun Heo <tj@kernel.org> Cc: Nick Desaulniers <ndesaulniers@google.com> Cc: Nathan Chancellor <nathan@kernel.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-06-23 19:08:14 +00:00
return work_struct_pwq(data)->pool->id;
return data >> WORK_OFFQ_POOL_SHIFT;
}
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
static void mark_work_canceling(struct work_struct *work)
{
unsigned long pool_id = get_work_pool_id(work);
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
pool_id <<= WORK_OFFQ_POOL_SHIFT;
set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
}
static bool work_is_canceling(struct work_struct *work)
{
unsigned long data = atomic_long_read(&work->data);
return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
* Policy functions. These define the policies on how the global worker
* pools are managed. Unless noted otherwise, these functions assume that
* they're being called with pool->lock held.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
[PATCH] WorkStruct: Use direct assignment rather than cmpxchg() Use direct assignment rather than cmpxchg() as the latter is unavailable and unimplementable on some platforms and is actually unnecessary. The use of cmpxchg() was to guard against two possibilities, neither of which can actually occur: (1) The pending flag may have been unset or may be cleared. However, given where it's called, the pending flag is _always_ set. I don't think it can be unset whilst we're in set_wq_data(). Once the work is enqueued to be actually run, the only way off the queue is for it to be actually run. If it's a delayed work item, then the bit can't be cleared by the timer because we haven't started the timer yet. Also, the pending bit can't be cleared by cancelling the delayed work _until_ the work item has had its timer started. (2) The workqueue pointer might change. This can only happen in two cases: (a) The work item has just been queued to actually run, and so we're protected by the appropriate workqueue spinlock. (b) A delayed work item is being queued, and so the timer hasn't been started yet, and so no one else knows about the work item or can access it (the pending bit protects us). Besides, set_wq_data() _sets_ the workqueue pointer unconditionally, so it can be assigned instead. So, replacing the set_wq_data() with a straight assignment would be okay in most cases. The problem is where we end up tangling with test_and_set_bit() emulated using spinlocks, and even then it's not a problem _provided_ test_and_set_bit() doesn't attempt to modify the word if the bit was set. If that's a problem, then a bitops-proofed assignment will be required - equivalent to atomic_set() vs other atomic_xxx() ops. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 11:33:26 +00:00
/*
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* Need to wake up a worker? Called from anything but currently
* running workers.
*
* Note that, because unbound workers never contribute to nr_running, this
* function will always return %true for unbound pools as long as the
* worklist isn't empty.
[PATCH] WorkStruct: Use direct assignment rather than cmpxchg() Use direct assignment rather than cmpxchg() as the latter is unavailable and unimplementable on some platforms and is actually unnecessary. The use of cmpxchg() was to guard against two possibilities, neither of which can actually occur: (1) The pending flag may have been unset or may be cleared. However, given where it's called, the pending flag is _always_ set. I don't think it can be unset whilst we're in set_wq_data(). Once the work is enqueued to be actually run, the only way off the queue is for it to be actually run. If it's a delayed work item, then the bit can't be cleared by the timer because we haven't started the timer yet. Also, the pending bit can't be cleared by cancelling the delayed work _until_ the work item has had its timer started. (2) The workqueue pointer might change. This can only happen in two cases: (a) The work item has just been queued to actually run, and so we're protected by the appropriate workqueue spinlock. (b) A delayed work item is being queued, and so the timer hasn't been started yet, and so no one else knows about the work item or can access it (the pending bit protects us). Besides, set_wq_data() _sets_ the workqueue pointer unconditionally, so it can be assigned instead. So, replacing the set_wq_data() with a straight assignment would be okay in most cases. The problem is where we end up tangling with test_and_set_bit() emulated using spinlocks, and even then it's not a problem _provided_ test_and_set_bit() doesn't attempt to modify the word if the bit was set. If that's a problem, then a bitops-proofed assignment will be required - equivalent to atomic_set() vs other atomic_xxx() ops. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 11:33:26 +00:00
*/
static bool need_more_worker(struct worker_pool *pool)
{
return !list_empty(&pool->worklist) && !pool->nr_running;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
[PATCH] WorkStruct: Use direct assignment rather than cmpxchg() Use direct assignment rather than cmpxchg() as the latter is unavailable and unimplementable on some platforms and is actually unnecessary. The use of cmpxchg() was to guard against two possibilities, neither of which can actually occur: (1) The pending flag may have been unset or may be cleared. However, given where it's called, the pending flag is _always_ set. I don't think it can be unset whilst we're in set_wq_data(). Once the work is enqueued to be actually run, the only way off the queue is for it to be actually run. If it's a delayed work item, then the bit can't be cleared by the timer because we haven't started the timer yet. Also, the pending bit can't be cleared by cancelling the delayed work _until_ the work item has had its timer started. (2) The workqueue pointer might change. This can only happen in two cases: (a) The work item has just been queued to actually run, and so we're protected by the appropriate workqueue spinlock. (b) A delayed work item is being queued, and so the timer hasn't been started yet, and so no one else knows about the work item or can access it (the pending bit protects us). Besides, set_wq_data() _sets_ the workqueue pointer unconditionally, so it can be assigned instead. So, replacing the set_wq_data() with a straight assignment would be okay in most cases. The problem is where we end up tangling with test_and_set_bit() emulated using spinlocks, and even then it's not a problem _provided_ test_and_set_bit() doesn't attempt to modify the word if the bit was set. If that's a problem, then a bitops-proofed assignment will be required - equivalent to atomic_set() vs other atomic_xxx() ops. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-07 11:33:26 +00:00
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* Can I start working? Called from busy but !running workers. */
static bool may_start_working(struct worker_pool *pool)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
return pool->nr_idle;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
/* Do I need to keep working? Called from currently running workers. */
static bool keep_working(struct worker_pool *pool)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
/* Do we need a new worker? Called from manager. */
static bool need_to_create_worker(struct worker_pool *pool)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
return need_more_worker(pool) && !may_start_working(pool);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* Do we have too many workers and should some go away? */
static bool too_many_workers(struct worker_pool *pool)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
bool managing = pool->flags & POOL_MANAGER_ACTIVE;
int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
int nr_busy = pool->nr_workers - nr_idle;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
}
/**
* worker_set_flags - set worker flags and adjust nr_running accordingly
* @worker: self
* @flags: flags to set
*
* Set @flags in @worker->flags and adjust nr_running accordingly.
*/
static inline void worker_set_flags(struct worker *worker, unsigned int flags)
{
struct worker_pool *pool = worker->pool;
lockdep_assert_held(&pool->lock);
/* If transitioning into NOT_RUNNING, adjust nr_running. */
if ((flags & WORKER_NOT_RUNNING) &&
!(worker->flags & WORKER_NOT_RUNNING)) {
pool->nr_running--;
}
worker->flags |= flags;
}
/**
* worker_clr_flags - clear worker flags and adjust nr_running accordingly
* @worker: self
* @flags: flags to clear
*
* Clear @flags in @worker->flags and adjust nr_running accordingly.
*/
static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
{
struct worker_pool *pool = worker->pool;
unsigned int oflags = worker->flags;
lockdep_assert_held(&pool->lock);
worker->flags &= ~flags;
/*
* If transitioning out of NOT_RUNNING, increment nr_running. Note
* that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
* of multiple flags, not a single flag.
*/
if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
if (!(worker->flags & WORKER_NOT_RUNNING))
pool->nr_running++;
}
/* Return the first idle worker. Called with pool->lock held. */
static struct worker *first_idle_worker(struct worker_pool *pool)
{
if (unlikely(list_empty(&pool->idle_list)))
return NULL;
return list_first_entry(&pool->idle_list, struct worker, entry);
}
/**
* worker_enter_idle - enter idle state
* @worker: worker which is entering idle state
*
* @worker is entering idle state. Update stats and idle timer if
* necessary.
*
* LOCKING:
* raw_spin_lock_irq(pool->lock).
*/
static void worker_enter_idle(struct worker *worker)
{
struct worker_pool *pool = worker->pool;
if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
WARN_ON_ONCE(!list_empty(&worker->entry) &&
(worker->hentry.next || worker->hentry.pprev)))
return;
/* can't use worker_set_flags(), also called from create_worker() */
worker->flags |= WORKER_IDLE;
pool->nr_idle++;
worker->last_active = jiffies;
/* idle_list is LIFO */
list_add(&worker->entry, &pool->idle_list);
if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
/* Sanity check nr_running. */
WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
}
/**
* worker_leave_idle - leave idle state
* @worker: worker which is leaving idle state
*
* @worker is leaving idle state. Update stats.
*
* LOCKING:
* raw_spin_lock_irq(pool->lock).
*/
static void worker_leave_idle(struct worker *worker)
{
struct worker_pool *pool = worker->pool;
if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
return;
worker_clr_flags(worker, WORKER_IDLE);
pool->nr_idle--;
list_del_init(&worker->entry);
}
/**
* find_worker_executing_work - find worker which is executing a work
* @pool: pool of interest
* @work: work to find worker for
*
* Find a worker which is executing @work on @pool by searching
* @pool->busy_hash which is keyed by the address of @work. For a worker
* to match, its current execution should match the address of @work and
* its work function. This is to avoid unwanted dependency between
* unrelated work executions through a work item being recycled while still
* being executed.
*
* This is a bit tricky. A work item may be freed once its execution
* starts and nothing prevents the freed area from being recycled for
* another work item. If the same work item address ends up being reused
* before the original execution finishes, workqueue will identify the
* recycled work item as currently executing and make it wait until the
* current execution finishes, introducing an unwanted dependency.
*
* This function checks the work item address and work function to avoid
* false positives. Note that this isn't complete as one may construct a
* work function which can introduce dependency onto itself through a
* recycled work item. Well, if somebody wants to shoot oneself in the
* foot that badly, there's only so much we can do, and if such deadlock
* actually occurs, it should be easy to locate the culprit work function.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*
* Return:
* Pointer to worker which is executing @work if found, %NULL
* otherwise.
*/
static struct worker *find_worker_executing_work(struct worker_pool *pool,
struct work_struct *work)
{
struct worker *worker;
hash_for_each_possible(pool->busy_hash, worker, hentry,
(unsigned long)work)
if (worker->current_work == work &&
worker->current_func == work->func)
return worker;
return NULL;
}
/**
* move_linked_works - move linked works to a list
* @work: start of series of works to be scheduled
* @head: target list to append @work to
* @nextp: out parameter for nested worklist walking
*
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
* Schedule linked works starting from @work to @head. Work series to be
* scheduled starts at @work and includes any consecutive work with
* WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
* @nextp.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*/
static void move_linked_works(struct work_struct *work, struct list_head *head,
struct work_struct **nextp)
{
struct work_struct *n;
/*
* Linked worklist will always end before the end of the list,
* use NULL for list head.
*/
list_for_each_entry_safe_from(work, n, NULL, entry) {
list_move_tail(&work->entry, head);
if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
break;
}
/*
* If we're already inside safe list traversal and have moved
* multiple works to the scheduled queue, the next position
* needs to be updated.
*/
if (nextp)
*nextp = n;
}
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
/**
* assign_work - assign a work item and its linked work items to a worker
* @work: work to assign
* @worker: worker to assign to
* @nextp: out parameter for nested worklist walking
*
* Assign @work and its linked work items to @worker. If @work is already being
* executed by another worker in the same pool, it'll be punted there.
*
* If @nextp is not NULL, it's updated to point to the next work of the last
* scheduled work. This allows assign_work() to be nested inside
* list_for_each_entry_safe().
*
* Returns %true if @work was successfully assigned to @worker. %false if @work
* was punted to another worker already executing it.
*/
static bool assign_work(struct work_struct *work, struct worker *worker,
struct work_struct **nextp)
{
struct worker_pool *pool = worker->pool;
struct worker *collision;
lockdep_assert_held(&pool->lock);
/*
* A single work shouldn't be executed concurrently by multiple workers.
* __queue_work() ensures that @work doesn't jump to a different pool
* while still running in the previous pool. Here, we should ensure that
* @work is not executed concurrently by multiple workers from the same
* pool. Check whether anyone is already processing the work. If so,
* defer the work to the currently executing one.
*/
collision = find_worker_executing_work(pool, work);
if (unlikely(collision)) {
move_linked_works(work, &collision->scheduled, nextp);
return false;
}
move_linked_works(work, &worker->scheduled, nextp);
return true;
}
/**
* kick_pool - wake up an idle worker if necessary
* @pool: pool to kick
*
* @pool may have pending work items. Wake up worker if necessary. Returns
* whether a worker was woken up.
*/
static bool kick_pool(struct worker_pool *pool)
{
struct worker *worker = first_idle_worker(pool);
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
struct task_struct *p;
lockdep_assert_held(&pool->lock);
if (!need_more_worker(pool) || !worker)
return false;
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
p = worker->task;
#ifdef CONFIG_SMP
/*
* Idle @worker is about to execute @work and waking up provides an
* opportunity to migrate @worker at a lower cost by setting the task's
* wake_cpu field. Let's see if we want to move @worker to improve
* execution locality.
*
* We're waking the worker that went idle the latest and there's some
* chance that @worker is marked idle but hasn't gone off CPU yet. If
* so, setting the wake_cpu won't do anything. As this is a best-effort
* optimization and the race window is narrow, let's leave as-is for
* now. If this becomes pronounced, we can skip over workers which are
* still on cpu when picking an idle worker.
*
* If @pool has non-strict affinity, @worker might have ended up outside
* its affinity scope. Repatriate.
*/
if (!pool->attrs->affn_strict &&
!cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
struct work_struct *work = list_first_entry(&pool->worklist,
struct work_struct, entry);
p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
}
#endif
wake_up_process(p);
return true;
}
#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
/*
* Concurrency-managed per-cpu work items that hog CPU for longer than
* wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
* which prevents them from stalling other concurrency-managed work items. If a
* work function keeps triggering this mechanism, it's likely that the work item
* should be using an unbound workqueue instead.
*
* wq_cpu_intensive_report() tracks work functions which trigger such conditions
* and report them so that they can be examined and converted to use unbound
* workqueues as appropriate. To avoid flooding the console, each violating work
* function is tracked and reported with exponential backoff.
*/
#define WCI_MAX_ENTS 128
struct wci_ent {
work_func_t func;
atomic64_t cnt;
struct hlist_node hash_node;
};
static struct wci_ent wci_ents[WCI_MAX_ENTS];
static int wci_nr_ents;
static DEFINE_RAW_SPINLOCK(wci_lock);
static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
static struct wci_ent *wci_find_ent(work_func_t func)
{
struct wci_ent *ent;
hash_for_each_possible_rcu(wci_hash, ent, hash_node,
(unsigned long)func) {
if (ent->func == func)
return ent;
}
return NULL;
}
static void wq_cpu_intensive_report(work_func_t func)
{
struct wci_ent *ent;
restart:
ent = wci_find_ent(func);
if (ent) {
u64 cnt;
/*
* Start reporting from the fourth time and back off
* exponentially.
*/
cnt = atomic64_inc_return_relaxed(&ent->cnt);
if (cnt >= 4 && is_power_of_2(cnt))
printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
ent->func, wq_cpu_intensive_thresh_us,
atomic64_read(&ent->cnt));
return;
}
/*
* @func is a new violation. Allocate a new entry for it. If wcn_ents[]
* is exhausted, something went really wrong and we probably made enough
* noise already.
*/
if (wci_nr_ents >= WCI_MAX_ENTS)
return;
raw_spin_lock(&wci_lock);
if (wci_nr_ents >= WCI_MAX_ENTS) {
raw_spin_unlock(&wci_lock);
return;
}
if (wci_find_ent(func)) {
raw_spin_unlock(&wci_lock);
goto restart;
}
ent = &wci_ents[wci_nr_ents++];
ent->func = func;
atomic64_set(&ent->cnt, 1);
hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
raw_spin_unlock(&wci_lock);
}
#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
static void wq_cpu_intensive_report(work_func_t func) {}
#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
/**
* wq_worker_running - a worker is running again
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* @task: task waking up
*
* This function is called when a worker returns from schedule()
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
void wq_worker_running(struct task_struct *task)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct worker *worker = kthread_data(task);
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
if (!READ_ONCE(worker->sleeping))
return;
workqueue: Fix unbind_workers() VS wq_worker_running() race At CPU-hotplug time, unbind_worker() may preempt a worker while it is waking up. In that case the following scenario can happen: unbind_workers() wq_worker_running() -------------- ------------------- if (!(worker->flags & WORKER_NOT_RUNNING)) //PREEMPTED by unbind_workers worker->flags |= WORKER_UNBOUND; [...] atomic_set(&pool->nr_running, 0); //resume to worker atomic_inc(&worker->pool->nr_running); After unbind_worker() resets pool->nr_running, the value is expected to remain 0 until the pool ever gets rebound in case cpu_up() is called on the target CPU in the future. But here the race leaves pool->nr_running with a value of 1, triggering the following warning when the worker goes idle: WARNING: CPU: 3 PID: 34 at kernel/workqueue.c:1823 worker_enter_idle+0x95/0xc0 Modules linked in: CPU: 3 PID: 34 Comm: kworker/3:0 Not tainted 5.16.0-rc1+ #34 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba527-rebuilt.opensuse.org 04/01/2014 Workqueue: 0x0 (rcu_par_gp) RIP: 0010:worker_enter_idle+0x95/0xc0 Code: 04 85 f8 ff ff ff 39 c1 7f 09 48 8b 43 50 48 85 c0 74 1b 83 e2 04 75 99 8b 43 34 39 43 30 75 91 8b 83 00 03 00 00 85 c0 74 87 <0f> 0b 5b c3 48 8b 35 70 f1 37 01 48 8d 7b 48 48 81 c6 e0 93 0 RSP: 0000:ffff9b7680277ed0 EFLAGS: 00010086 RAX: 00000000ffffffff RBX: ffff93465eae9c00 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffff9346418a0000 RDI: ffff934641057140 RBP: ffff934641057170 R08: 0000000000000001 R09: ffff9346418a0080 R10: ffff9b768027fdf0 R11: 0000000000002400 R12: ffff93465eae9c20 R13: ffff93465eae9c20 R14: ffff93465eae9c70 R15: ffff934641057140 FS: 0000000000000000(0000) GS:ffff93465eac0000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000000 CR3: 000000001cc0c000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> worker_thread+0x89/0x3d0 ? process_one_work+0x400/0x400 kthread+0x162/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x22/0x30 </TASK> Also due to this incorrect "nr_running == 1", further queued work may end up not being served, because no worker is awaken at work insert time. This raises rcutorture writer stalls for example. Fix this with disabling preemption in the right place in wq_worker_running(). It's worth noting that if the worker migrates and runs concurrently with unbind_workers(), it is guaranteed to see the WORKER_UNBOUND flag update due to set_cpus_allowed_ptr() acquiring/releasing rq->lock. Fixes: 6d25be5782e4 ("sched/core, workqueues: Distangle worker accounting from rq lock") Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Paul E. McKenney <paulmck@kernel.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Daniel Bristot de Oliveira <bristot@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-12-01 15:19:44 +00:00
/*
* If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
* and the nr_running increment below, we may ruin the nr_running reset
* and leave with an unexpected pool->nr_running == 1 on the newly unbound
* pool. Protect against such race.
*/
preempt_disable();
if (!(worker->flags & WORKER_NOT_RUNNING))
worker->pool->nr_running++;
workqueue: Fix unbind_workers() VS wq_worker_running() race At CPU-hotplug time, unbind_worker() may preempt a worker while it is waking up. In that case the following scenario can happen: unbind_workers() wq_worker_running() -------------- ------------------- if (!(worker->flags & WORKER_NOT_RUNNING)) //PREEMPTED by unbind_workers worker->flags |= WORKER_UNBOUND; [...] atomic_set(&pool->nr_running, 0); //resume to worker atomic_inc(&worker->pool->nr_running); After unbind_worker() resets pool->nr_running, the value is expected to remain 0 until the pool ever gets rebound in case cpu_up() is called on the target CPU in the future. But here the race leaves pool->nr_running with a value of 1, triggering the following warning when the worker goes idle: WARNING: CPU: 3 PID: 34 at kernel/workqueue.c:1823 worker_enter_idle+0x95/0xc0 Modules linked in: CPU: 3 PID: 34 Comm: kworker/3:0 Not tainted 5.16.0-rc1+ #34 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba527-rebuilt.opensuse.org 04/01/2014 Workqueue: 0x0 (rcu_par_gp) RIP: 0010:worker_enter_idle+0x95/0xc0 Code: 04 85 f8 ff ff ff 39 c1 7f 09 48 8b 43 50 48 85 c0 74 1b 83 e2 04 75 99 8b 43 34 39 43 30 75 91 8b 83 00 03 00 00 85 c0 74 87 <0f> 0b 5b c3 48 8b 35 70 f1 37 01 48 8d 7b 48 48 81 c6 e0 93 0 RSP: 0000:ffff9b7680277ed0 EFLAGS: 00010086 RAX: 00000000ffffffff RBX: ffff93465eae9c00 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffff9346418a0000 RDI: ffff934641057140 RBP: ffff934641057170 R08: 0000000000000001 R09: ffff9346418a0080 R10: ffff9b768027fdf0 R11: 0000000000002400 R12: ffff93465eae9c20 R13: ffff93465eae9c20 R14: ffff93465eae9c70 R15: ffff934641057140 FS: 0000000000000000(0000) GS:ffff93465eac0000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000000 CR3: 000000001cc0c000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> worker_thread+0x89/0x3d0 ? process_one_work+0x400/0x400 kthread+0x162/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x22/0x30 </TASK> Also due to this incorrect "nr_running == 1", further queued work may end up not being served, because no worker is awaken at work insert time. This raises rcutorture writer stalls for example. Fix this with disabling preemption in the right place in wq_worker_running(). It's worth noting that if the worker migrates and runs concurrently with unbind_workers(), it is guaranteed to see the WORKER_UNBOUND flag update due to set_cpus_allowed_ptr() acquiring/releasing rq->lock. Fixes: 6d25be5782e4 ("sched/core, workqueues: Distangle worker accounting from rq lock") Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Paul E. McKenney <paulmck@kernel.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Daniel Bristot de Oliveira <bristot@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-12-01 15:19:44 +00:00
preempt_enable();
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
/*
* CPU intensive auto-detection cares about how long a work item hogged
* CPU without sleeping. Reset the starting timestamp on wakeup.
*/
worker->current_at = worker->task->se.sum_exec_runtime;
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
WRITE_ONCE(worker->sleeping, 0);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
/**
* wq_worker_sleeping - a worker is going to sleep
* @task: task going to sleep
*
* This function is called from schedule() when a busy worker is
* going to sleep.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
void wq_worker_sleeping(struct task_struct *task)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct worker *worker = kthread_data(task);
struct worker_pool *pool;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
* Rescuers, which may not have all the fields set up like normal
* workers, also reach here, let's not access anything before
* checking NOT_RUNNING.
*/
if (worker->flags & WORKER_NOT_RUNNING)
return;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
pool = worker->pool;
workqueue: Remove the warning in wq_worker_sleeping() The kernel test robot triggered a warning with the following race: task-ctx A interrupt-ctx B worker -> process_one_work() -> work_item() -> schedule(); -> sched_submit_work() -> wq_worker_sleeping() -> ->sleeping = 1 atomic_dec_and_test(nr_running) __schedule(); *interrupt* async_page_fault() -> local_irq_enable(); -> schedule(); -> sched_submit_work() -> wq_worker_sleeping() -> if (WARN_ON(->sleeping)) return -> __schedule() -> sched_update_worker() -> wq_worker_running() -> atomic_inc(nr_running); -> ->sleeping = 0; -> sched_update_worker() -> wq_worker_running() if (!->sleeping) return In this context the warning is pointless everything is fine. An interrupt before wq_worker_sleeping() will perform the ->sleeping assignment (0 -> 1 > 0) twice. An interrupt after wq_worker_sleeping() will trigger the warning and nr_running will be decremented (by A) and incremented once (only by B, A will skip it). This is the case until the ->sleeping is zeroed again in wq_worker_running(). Remove the WARN statement because this condition may happen. Document that preemption around wq_worker_sleeping() needs to be disabled to protect ->sleeping and not just as an optimisation. Fixes: 6d25be5782e48 ("sched/core, workqueues: Distangle worker accounting from rq lock") Reported-by: kernel test robot <lkp@intel.com> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Cc: Tejun Heo <tj@kernel.org> Link: https://lkml.kernel.org/r/20200327074308.GY11705@shao2-debian
2020-03-27 23:29:59 +00:00
/* Return if preempted before wq_worker_running() was reached */
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
if (READ_ONCE(worker->sleeping))
return;
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
WRITE_ONCE(worker->sleeping, 1);
raw_spin_lock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
workqueue: Fix unbind_workers() VS wq_worker_sleeping() race At CPU-hotplug time, unbind_workers() may preempt a worker while it is going to sleep. In that case the following scenario can happen: unbind_workers() wq_worker_sleeping() -------------- ------------------- if (worker->flags & WORKER_NOT_RUNNING) return; //PREEMPTED by unbind_workers worker->flags |= WORKER_UNBOUND; [...] atomic_set(&pool->nr_running, 0); //resume to worker atomic_dec_and_test(&pool->nr_running); After unbind_worker() resets pool->nr_running, the value is expected to remain 0 until the pool ever gets rebound in case cpu_up() is called on the target CPU in the future. But here the race leaves pool->nr_running with a value of -1, triggering the following warning when the worker goes idle: WARNING: CPU: 3 PID: 34 at kernel/workqueue.c:1823 worker_enter_idle+0x95/0xc0 Modules linked in: CPU: 3 PID: 34 Comm: kworker/3:0 Not tainted 5.16.0-rc1+ #34 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba527-rebuilt.opensuse.org 04/01/2014 Workqueue: 0x0 (rcu_par_gp) RIP: 0010:worker_enter_idle+0x95/0xc0 Code: 04 85 f8 ff ff ff 39 c1 7f 09 48 8b 43 50 48 85 c0 74 1b 83 e2 04 75 99 8b 43 34 39 43 30 75 91 8b 83 00 03 00 00 85 c0 74 87 <0f> 0b 5b c3 48 8b 35 70 f1 37 01 48 8d 7b 48 48 81 c6 e0 93 0 RSP: 0000:ffff9b7680277ed0 EFLAGS: 00010086 RAX: 00000000ffffffff RBX: ffff93465eae9c00 RCX: 0000000000000000 RDX: 0000000000000000 RSI: ffff9346418a0000 RDI: ffff934641057140 RBP: ffff934641057170 R08: 0000000000000001 R09: ffff9346418a0080 R10: ffff9b768027fdf0 R11: 0000000000002400 R12: ffff93465eae9c20 R13: ffff93465eae9c20 R14: ffff93465eae9c70 R15: ffff934641057140 FS: 0000000000000000(0000) GS:ffff93465eac0000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000000000 CR3: 000000001cc0c000 CR4: 00000000000006e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> worker_thread+0x89/0x3d0 ? process_one_work+0x400/0x400 kthread+0x162/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x22/0x30 </TASK> Also due to this incorrect "nr_running == -1", all sorts of hazards can happen, starting with queued works being ignored because no workers are awaken at insert_work() time. Fix this with checking again the worker flags while pool->lock is locked. Fixes: b945efcdd07d ("sched: Remove pointless preemption disable in sched_submit_work()") Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Paul E. McKenney <paulmck@kernel.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Paul E. McKenney <paulmck@kernel.org> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Daniel Bristot de Oliveira <bristot@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-12-01 15:19:45 +00:00
/*
* Recheck in case unbind_workers() preempted us. We don't
* want to decrement nr_running after the worker is unbound
* and nr_running has been reset.
*/
if (worker->flags & WORKER_NOT_RUNNING) {
raw_spin_unlock_irq(&pool->lock);
return;
}
pool->nr_running--;
if (kick_pool(pool))
worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
raw_spin_unlock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
/**
* wq_worker_tick - a scheduler tick occurred while a kworker is running
* @task: task currently running
*
* Called from scheduler_tick(). We're in the IRQ context and the current
* worker's fields which follow the 'K' locking rule can be accessed safely.
*/
void wq_worker_tick(struct task_struct *task)
{
struct worker *worker = kthread_data(task);
struct pool_workqueue *pwq = worker->current_pwq;
struct worker_pool *pool = worker->pool;
if (!pwq)
return;
pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
if (!wq_cpu_intensive_thresh_us)
return;
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
/*
* If the current worker is concurrency managed and hogged the CPU for
* longer than wq_cpu_intensive_thresh_us, it's automatically marked
* CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
*
* Set @worker->sleeping means that @worker is in the process of
* switching out voluntarily and won't be contributing to
* @pool->nr_running until it wakes up. As wq_worker_sleeping() also
* decrements ->nr_running, setting CPU_INTENSIVE here can lead to
* double decrements. The task is releasing the CPU anyway. Let's skip.
* We probably want to make this prettier in the future.
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
*/
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle() Currently, pool->nr_running can be modified from timer tick, that means the timer tick can run nested inside a not-irq-protected section that's in the process of modifying nr_running. Consider the following scenario: CPU0 kworker/0:2 (events) worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); ->pool->nr_running++; (1) process_one_work() ->worker->current_func(work); ->schedule() ->wq_worker_sleeping() ->worker->sleeping = 1; ->pool->nr_running--; (0) .... ->wq_worker_running() .... CPU0 by interrupt: wq_worker_tick() ->worker_set_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running--; (-1) ->worker->flags |= WORKER_CPU_INTENSIVE; .... ->if (!(worker->flags & WORKER_NOT_RUNNING)) ->pool->nr_running++; (will not execute) ->worker->sleeping = 0; .... ->worker_clr_flags(worker, WORKER_CPU_INTENSIVE); ->pool->nr_running++; (0) .... worker_set_flags(worker, WORKER_PREP); ->pool->nr_running--; (-1) .... worker_enter_idle() ->WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); if the nr_workers is equal to nr_idle, due to the nr_running is not zero, will trigger WARN_ON_ONCE(). [ 2.460602] WARNING: CPU: 0 PID: 63 at kernel/workqueue.c:1999 worker_enter_idle+0xb2/0xc0 [ 2.462163] Modules linked in: [ 2.463401] CPU: 0 PID: 63 Comm: kworker/0:2 Not tainted 6.4.0-rc2-next-20230519 #1 [ 2.463771] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014 [ 2.465127] Workqueue: 0x0 (events) [ 2.465678] RIP: 0010:worker_enter_idle+0xb2/0xc0 ... [ 2.472614] Call Trace: [ 2.473152] <TASK> [ 2.474182] worker_thread+0x71/0x430 [ 2.474992] ? _raw_spin_unlock_irqrestore+0x28/0x50 [ 2.475263] kthread+0x103/0x120 [ 2.475493] ? __pfx_worker_thread+0x10/0x10 [ 2.476355] ? __pfx_kthread+0x10/0x10 [ 2.476635] ret_from_fork+0x2c/0x50 [ 2.477051] </TASK> This commit therefore add the check of worker->sleeping in wq_worker_tick(), if the worker->sleeping is not zero, directly return. tj: Updated comment and description. Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Anders Roxell <anders.roxell@linaro.org> Closes: https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230519/testrun/17078554/suite/boot/test/clang-nightly-lkftconfig/log Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-05-24 03:53:39 +00:00
if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
worker->task->se.sum_exec_runtime - worker->current_at <
wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
return;
raw_spin_lock(&pool->lock);
worker_set_flags(worker, WORKER_CPU_INTENSIVE);
wq_cpu_intensive_report(worker->current_func);
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
if (kick_pool(pool))
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
pwq->stats[PWQ_STAT_CM_WAKEUP]++;
raw_spin_unlock(&pool->lock);
}
psi: fix aggregation idle shut-off psi has provisions to shut off the periodic aggregation worker when there is a period of no task activity - and thus no data that needs aggregating. However, while developing psi monitoring, Suren noticed that the aggregation clock currently won't stay shut off for good. Debugging this revealed a flaw in the idle design: an aggregation run will see no task activity and decide to go to sleep; shortly thereafter, the kworker thread that executed the aggregation will go idle and cause a scheduling change, during which the psi callback will kick the !pending worker again. This will ping-pong forever, and is equivalent to having no shut-off logic at all (but with more code!) Fix this by exempting aggregation workers from psi's clock waking logic when the state change is them going to sleep. To do this, tag workers with the last work function they executed, and if in psi we see a worker going to sleep after aggregating psi data, we will not reschedule the aggregation work item. What if the worker is also executing other items before or after? Any psi state times that were incurred by work items preceding the aggregation work will have been collected from the per-cpu buckets during the aggregation itself. If there are work items following the aggregation work, the worker's last_func tag will be overwritten and the aggregator will be kept alive to process this genuine new activity. If the aggregation work is the last thing the worker does, and we decide to go idle, the brief period of non-idle time incurred between the aggregation run and the kworker's dequeue will be stranded in the per-cpu buckets until the clock is woken by later activity. But that should not be a problem. The buckets can hold 4s worth of time, and future activity will wake the clock with a 2s delay, giving us 2s worth of data we can leave behind when disabling aggregation. If it takes a worker more than two seconds to go idle after it finishes its last work item, we likely have bigger problems in the system, and won't notice one sample that was averaged with a bogus per-CPU weight. Link: http://lkml.kernel.org/r/20190116193501.1910-1-hannes@cmpxchg.org Fixes: eb414681d5a0 ("psi: pressure stall information for CPU, memory, and IO") Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reported-by: Suren Baghdasaryan <surenb@google.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-01 22:20:42 +00:00
/**
* wq_worker_last_func - retrieve worker's last work function
* @task: Task to retrieve last work function of.
psi: fix aggregation idle shut-off psi has provisions to shut off the periodic aggregation worker when there is a period of no task activity - and thus no data that needs aggregating. However, while developing psi monitoring, Suren noticed that the aggregation clock currently won't stay shut off for good. Debugging this revealed a flaw in the idle design: an aggregation run will see no task activity and decide to go to sleep; shortly thereafter, the kworker thread that executed the aggregation will go idle and cause a scheduling change, during which the psi callback will kick the !pending worker again. This will ping-pong forever, and is equivalent to having no shut-off logic at all (but with more code!) Fix this by exempting aggregation workers from psi's clock waking logic when the state change is them going to sleep. To do this, tag workers with the last work function they executed, and if in psi we see a worker going to sleep after aggregating psi data, we will not reschedule the aggregation work item. What if the worker is also executing other items before or after? Any psi state times that were incurred by work items preceding the aggregation work will have been collected from the per-cpu buckets during the aggregation itself. If there are work items following the aggregation work, the worker's last_func tag will be overwritten and the aggregator will be kept alive to process this genuine new activity. If the aggregation work is the last thing the worker does, and we decide to go idle, the brief period of non-idle time incurred between the aggregation run and the kworker's dequeue will be stranded in the per-cpu buckets until the clock is woken by later activity. But that should not be a problem. The buckets can hold 4s worth of time, and future activity will wake the clock with a 2s delay, giving us 2s worth of data we can leave behind when disabling aggregation. If it takes a worker more than two seconds to go idle after it finishes its last work item, we likely have bigger problems in the system, and won't notice one sample that was averaged with a bogus per-CPU weight. Link: http://lkml.kernel.org/r/20190116193501.1910-1-hannes@cmpxchg.org Fixes: eb414681d5a0 ("psi: pressure stall information for CPU, memory, and IO") Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reported-by: Suren Baghdasaryan <surenb@google.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-01 22:20:42 +00:00
*
* Determine the last function a worker executed. This is called from
* the scheduler to get a worker's last known identity.
*
* CONTEXT:
* raw_spin_lock_irq(rq->lock)
psi: fix aggregation idle shut-off psi has provisions to shut off the periodic aggregation worker when there is a period of no task activity - and thus no data that needs aggregating. However, while developing psi monitoring, Suren noticed that the aggregation clock currently won't stay shut off for good. Debugging this revealed a flaw in the idle design: an aggregation run will see no task activity and decide to go to sleep; shortly thereafter, the kworker thread that executed the aggregation will go idle and cause a scheduling change, during which the psi callback will kick the !pending worker again. This will ping-pong forever, and is equivalent to having no shut-off logic at all (but with more code!) Fix this by exempting aggregation workers from psi's clock waking logic when the state change is them going to sleep. To do this, tag workers with the last work function they executed, and if in psi we see a worker going to sleep after aggregating psi data, we will not reschedule the aggregation work item. What if the worker is also executing other items before or after? Any psi state times that were incurred by work items preceding the aggregation work will have been collected from the per-cpu buckets during the aggregation itself. If there are work items following the aggregation work, the worker's last_func tag will be overwritten and the aggregator will be kept alive to process this genuine new activity. If the aggregation work is the last thing the worker does, and we decide to go idle, the brief period of non-idle time incurred between the aggregation run and the kworker's dequeue will be stranded in the per-cpu buckets until the clock is woken by later activity. But that should not be a problem. The buckets can hold 4s worth of time, and future activity will wake the clock with a 2s delay, giving us 2s worth of data we can leave behind when disabling aggregation. If it takes a worker more than two seconds to go idle after it finishes its last work item, we likely have bigger problems in the system, and won't notice one sample that was averaged with a bogus per-CPU weight. Link: http://lkml.kernel.org/r/20190116193501.1910-1-hannes@cmpxchg.org Fixes: eb414681d5a0 ("psi: pressure stall information for CPU, memory, and IO") Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reported-by: Suren Baghdasaryan <surenb@google.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-01 22:20:42 +00:00
*
* This function is called during schedule() when a kworker is going
* to sleep. It's used by psi to identify aggregation workers during
* dequeuing, to allow periodic aggregation to shut-off when that
* worker is the last task in the system or cgroup to go to sleep.
*
* As this function doesn't involve any workqueue-related locking, it
* only returns stable values when called from inside the scheduler's
* queuing and dequeuing paths, when @task, which must be a kworker,
* is guaranteed to not be processing any works.
*
psi: fix aggregation idle shut-off psi has provisions to shut off the periodic aggregation worker when there is a period of no task activity - and thus no data that needs aggregating. However, while developing psi monitoring, Suren noticed that the aggregation clock currently won't stay shut off for good. Debugging this revealed a flaw in the idle design: an aggregation run will see no task activity and decide to go to sleep; shortly thereafter, the kworker thread that executed the aggregation will go idle and cause a scheduling change, during which the psi callback will kick the !pending worker again. This will ping-pong forever, and is equivalent to having no shut-off logic at all (but with more code!) Fix this by exempting aggregation workers from psi's clock waking logic when the state change is them going to sleep. To do this, tag workers with the last work function they executed, and if in psi we see a worker going to sleep after aggregating psi data, we will not reschedule the aggregation work item. What if the worker is also executing other items before or after? Any psi state times that were incurred by work items preceding the aggregation work will have been collected from the per-cpu buckets during the aggregation itself. If there are work items following the aggregation work, the worker's last_func tag will be overwritten and the aggregator will be kept alive to process this genuine new activity. If the aggregation work is the last thing the worker does, and we decide to go idle, the brief period of non-idle time incurred between the aggregation run and the kworker's dequeue will be stranded in the per-cpu buckets until the clock is woken by later activity. But that should not be a problem. The buckets can hold 4s worth of time, and future activity will wake the clock with a 2s delay, giving us 2s worth of data we can leave behind when disabling aggregation. If it takes a worker more than two seconds to go idle after it finishes its last work item, we likely have bigger problems in the system, and won't notice one sample that was averaged with a bogus per-CPU weight. Link: http://lkml.kernel.org/r/20190116193501.1910-1-hannes@cmpxchg.org Fixes: eb414681d5a0 ("psi: pressure stall information for CPU, memory, and IO") Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reported-by: Suren Baghdasaryan <surenb@google.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-01 22:20:42 +00:00
* Return:
* The last work function %current executed as a worker, NULL if it
* hasn't executed any work yet.
*/
work_func_t wq_worker_last_func(struct task_struct *task)
{
struct worker *worker = kthread_data(task);
return worker->last_func;
}
/**
* get_pwq - get an extra reference on the specified pool_workqueue
* @pwq: pool_workqueue to get
*
* Obtain an extra reference on @pwq. The caller should guarantee that
* @pwq has positive refcnt and be holding the matching pool->lock.
*/
static void get_pwq(struct pool_workqueue *pwq)
{
lockdep_assert_held(&pwq->pool->lock);
WARN_ON_ONCE(pwq->refcnt <= 0);
pwq->refcnt++;
}
/**
* put_pwq - put a pool_workqueue reference
* @pwq: pool_workqueue to put
*
* Drop a reference of @pwq. If its refcnt reaches zero, schedule its
* destruction. The caller should be holding the matching pool->lock.
*/
static void put_pwq(struct pool_workqueue *pwq)
{
lockdep_assert_held(&pwq->pool->lock);
if (likely(--pwq->refcnt))
return;
/*
* @pwq can't be released under pool->lock, bounce to a dedicated
* kthread_worker to avoid A-A deadlocks.
*/
kthread_queue_work(pwq_release_worker, &pwq->release_work);
}
/**
* put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
* @pwq: pool_workqueue to put (can be %NULL)
*
* put_pwq() with locking. This function also allows %NULL @pwq.
*/
static void put_pwq_unlocked(struct pool_workqueue *pwq)
{
if (pwq) {
/*
* As both pwqs and pools are RCU protected, the
* following lock operations are safe.
*/
raw_spin_lock_irq(&pwq->pool->lock);
put_pwq(pwq);
raw_spin_unlock_irq(&pwq->pool->lock);
}
}
static void pwq_activate_inactive_work(struct work_struct *work)
{
struct pool_workqueue *pwq = get_work_pwq(work);
trace_workqueue_activate_work(work);
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
if (list_empty(&pwq->pool->worklist))
pwq->pool->watchdog_ts = jiffies;
move_linked_works(work, &pwq->pool->worklist, NULL);
__clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
pwq->nr_active++;
}
static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
{
struct work_struct *work = list_first_entry(&pwq->inactive_works,
struct work_struct, entry);
pwq_activate_inactive_work(work);
}
/**
* pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
* @pwq: pwq of interest
* @work_data: work_data of work which left the queue
*
* A work either has completed or is removed from pending queue,
* decrement nr_in_flight of its pwq and handle workqueue flushing.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*/
static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
{
int color = get_work_color(work_data);
workqueue: Mark barrier work with WORK_STRUCT_INACTIVE Currently, WORK_NO_COLOR has two meanings: Not participate in flushing Not participate in nr_active And only non-barrier work items are marked with WORK_STRUCT_INACTIVE when they are in inactive_works list. The barrier work items are not marked INACTIVE even linked in inactive_works list since these tail items are always moved together with the head work item. These definitions are simple, clean and practical. (Except a small blemish that only the first meaning of WORK_NO_COLOR is documented in include/linux/workqueue.h while both meanings are in workqueue.c) But dual-purpose WORK_NO_COLOR used for barrier work items has proven to be problematical[1]. Only the second purpose is obligatory. So we plan to make barrier work items participate in flushing but keep them still not participating in nr_active. So the plan is to mark barrier work items inactive without using WORK_NO_COLOR in this patch so that we can assign a flushing color to them in next patch. The reasonable way is to add or reuse a bit in work data of the work item. But adding a bit will double the size of pool_workqueue. Currently, WORK_STRUCT_INACTIVE is only used in try_to_grab_pending() for user-queued work items and try_to_grab_pending() can't work for barrier work items. So we extend WORK_STRUCT_INACTIVE to also mark barrier work items no matter which list they are in because we don't need to determind which list a barrier work item is in. So the meaning of WORK_STRUCT_INACTIVE becomes just "the work items don't participate in nr_active" (no matter whether it is a barrier work item or a user-queued work item). And WORK_STRUCT_INACTIVE for user-queued work items means they are in inactive_works list. This patch does it by setting WORK_STRUCT_INACTIVE for barrier work items in insert_wq_barrier() and checking WORK_STRUCT_INACTIVE first in pwq_dec_nr_in_flight(). And the meaning of WORK_NO_COLOR is reduced to only "not participating in flushing". There is no functionality change intended in this patch. Because WORK_NO_COLOR+WORK_STRUCT_INACTIVE represents the previous WORK_NO_COLOR in meaning and try_to_grab_pending() doesn't use for barrier work items and avoids being confused by this extended WORK_STRUCT_INACTIVE. A bunch of comment for nr_active & WORK_STRUCT_INACTIVE is also added for documenting how WORK_STRUCT_INACTIVE works in nr_active management. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:37 +00:00
if (!(work_data & WORK_STRUCT_INACTIVE)) {
pwq->nr_active--;
if (!list_empty(&pwq->inactive_works)) {
/* one down, submit an inactive one */
if (pwq->nr_active < pwq->max_active)
pwq_activate_first_inactive(pwq);
}
}
pwq->nr_in_flight[color]--;
/* is flush in progress and are we at the flushing tip? */
if (likely(pwq->flush_color != color))
goto out_put;
/* are there still in-flight works? */
if (pwq->nr_in_flight[color])
goto out_put;
/* this pwq is done, clear flush_color */
pwq->flush_color = -1;
/*
* If this was the last pwq, wake up the first flusher. It
* will handle the rest.
*/
if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
complete(&pwq->wq->first_flusher->done);
out_put:
put_pwq(pwq);
}
/**
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* try_to_grab_pending - steal work item from worklist and disable irq
* @work: work item to steal
* @is_dwork: @work is a delayed_work
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* @flags: place to store irq state
*
* Try to grab PENDING bit of @work. This function can handle @work in any
* stable state - idle, on timer or on worklist.
*
* Return:
*
* ======== ================================================================
* 1 if @work was pending and we successfully stole PENDING
* 0 if @work was idle and we claimed PENDING
* -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* -ENOENT if someone else is canceling @work, this state may persist
* for arbitrarily long
* ======== ================================================================
*
* Note:
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
* On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
* interrupted while holding PENDING and @work off queue, irq must be
* disabled on entry. This, combined with delayed_work->timer being
* irqsafe, ensures that we return -EAGAIN for finite short period of time.
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
*
* On successful return, >= 0, irq is disabled and the caller is
* responsible for releasing it using local_irq_restore(*@flags).
*
* This function is safe to call from any context including IRQ handler.
*/
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
unsigned long *flags)
{
struct worker_pool *pool;
struct pool_workqueue *pwq;
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
local_irq_save(*flags);
/* try to steal the timer if it exists */
if (is_dwork) {
struct delayed_work *dwork = to_delayed_work(work);
/*
* dwork->timer is irqsafe. If del_timer() fails, it's
* guaranteed that the timer is not queued anywhere and not
* running on the local CPU.
*/
if (likely(del_timer(&dwork->timer)))
return 1;
}
/* try to claim PENDING the normal way */
if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
return 0;
rcu_read_lock();
/*
* The queueing is in progress, or it is already queued. Try to
* steal it from ->worklist without clearing WORK_STRUCT_PENDING.
*/
pool = get_work_pool(work);
if (!pool)
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
goto fail;
raw_spin_lock(&pool->lock);
workqueue: simplify is-work-item-queued-here test Currently, determining whether a work item is queued on a locked pool involves somewhat convoluted memory barrier dancing. It goes like the following. * When a work item is queued on a pool, work->data is updated before work->entry is linked to the pending list with a wmb() inbetween. * When trying to determine whether a work item is currently queued on a pool pointed to by work->data, it locks the pool and looks at work->entry. If work->entry is linked, we then do rmb() and then check whether work->data points to the current pool. This works because, work->data can only point to a pool if it currently is or were on the pool and, * If it currently is on the pool, the tests would obviously succeed. * It it left the pool, its work->entry was cleared under pool->lock, so if we're seeing non-empty work->entry, it has to be from the work item being linked on another pool. Because work->data is updated before work->entry is linked with wmb() inbetween, work->data update from another pool is guaranteed to be visible if we do rmb() after seeing non-empty work->entry. So, we either see empty work->entry or we see updated work->data pointin to another pool. While this works, it's convoluted, to put it mildly. With recent updates, it's now guaranteed that work->data points to cwq only while the work item is queued and that updating work->data to point to cwq or back to pool is done under pool->lock, so we can simply test whether work->data points to cwq which is associated with the currently locked pool instead of the convoluted memory barrier dancing. This patch replaces the memory barrier based "are you still here, really?" test with much simpler "does work->data points to me?" test - if work->data points to a cwq which is associated with the currently locked pool, the work item is guaranteed to be queued on the pool as work->data can start and stop pointing to such cwq only under pool->lock and the start and stop coincide with queue and dequeue. tj: Rewrote the comments and description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
/*
* work->data is guaranteed to point to pwq only while the work
* item is queued on pwq->wq, and both updating work->data to point
* to pwq on queueing and to pool on dequeueing are done under
* pwq->pool->lock. This in turn guarantees that, if work->data
* points to pwq which is associated with a locked pool, the work
workqueue: simplify is-work-item-queued-here test Currently, determining whether a work item is queued on a locked pool involves somewhat convoluted memory barrier dancing. It goes like the following. * When a work item is queued on a pool, work->data is updated before work->entry is linked to the pending list with a wmb() inbetween. * When trying to determine whether a work item is currently queued on a pool pointed to by work->data, it locks the pool and looks at work->entry. If work->entry is linked, we then do rmb() and then check whether work->data points to the current pool. This works because, work->data can only point to a pool if it currently is or were on the pool and, * If it currently is on the pool, the tests would obviously succeed. * It it left the pool, its work->entry was cleared under pool->lock, so if we're seeing non-empty work->entry, it has to be from the work item being linked on another pool. Because work->data is updated before work->entry is linked with wmb() inbetween, work->data update from another pool is guaranteed to be visible if we do rmb() after seeing non-empty work->entry. So, we either see empty work->entry or we see updated work->data pointin to another pool. While this works, it's convoluted, to put it mildly. With recent updates, it's now guaranteed that work->data points to cwq only while the work item is queued and that updating work->data to point to cwq or back to pool is done under pool->lock, so we can simply test whether work->data points to cwq which is associated with the currently locked pool instead of the convoluted memory barrier dancing. This patch replaces the memory barrier based "are you still here, really?" test with much simpler "does work->data points to me?" test - if work->data points to a cwq which is associated with the currently locked pool, the work item is guaranteed to be queued on the pool as work->data can start and stop pointing to such cwq only under pool->lock and the start and stop coincide with queue and dequeue. tj: Rewrote the comments and description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
* item is currently queued on that pool.
*/
pwq = get_work_pwq(work);
if (pwq && pwq->pool == pool) {
debug_work_deactivate(work);
/*
workqueue: Mark barrier work with WORK_STRUCT_INACTIVE Currently, WORK_NO_COLOR has two meanings: Not participate in flushing Not participate in nr_active And only non-barrier work items are marked with WORK_STRUCT_INACTIVE when they are in inactive_works list. The barrier work items are not marked INACTIVE even linked in inactive_works list since these tail items are always moved together with the head work item. These definitions are simple, clean and practical. (Except a small blemish that only the first meaning of WORK_NO_COLOR is documented in include/linux/workqueue.h while both meanings are in workqueue.c) But dual-purpose WORK_NO_COLOR used for barrier work items has proven to be problematical[1]. Only the second purpose is obligatory. So we plan to make barrier work items participate in flushing but keep them still not participating in nr_active. So the plan is to mark barrier work items inactive without using WORK_NO_COLOR in this patch so that we can assign a flushing color to them in next patch. The reasonable way is to add or reuse a bit in work data of the work item. But adding a bit will double the size of pool_workqueue. Currently, WORK_STRUCT_INACTIVE is only used in try_to_grab_pending() for user-queued work items and try_to_grab_pending() can't work for barrier work items. So we extend WORK_STRUCT_INACTIVE to also mark barrier work items no matter which list they are in because we don't need to determind which list a barrier work item is in. So the meaning of WORK_STRUCT_INACTIVE becomes just "the work items don't participate in nr_active" (no matter whether it is a barrier work item or a user-queued work item). And WORK_STRUCT_INACTIVE for user-queued work items means they are in inactive_works list. This patch does it by setting WORK_STRUCT_INACTIVE for barrier work items in insert_wq_barrier() and checking WORK_STRUCT_INACTIVE first in pwq_dec_nr_in_flight(). And the meaning of WORK_NO_COLOR is reduced to only "not participating in flushing". There is no functionality change intended in this patch. Because WORK_NO_COLOR+WORK_STRUCT_INACTIVE represents the previous WORK_NO_COLOR in meaning and try_to_grab_pending() doesn't use for barrier work items and avoids being confused by this extended WORK_STRUCT_INACTIVE. A bunch of comment for nr_active & WORK_STRUCT_INACTIVE is also added for documenting how WORK_STRUCT_INACTIVE works in nr_active management. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:37 +00:00
* A cancelable inactive work item must be in the
* pwq->inactive_works since a queued barrier can't be
* canceled (see the comments in insert_wq_barrier()).
*
* An inactive work item cannot be grabbed directly because
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
* it might have linked barrier work items which, if left
* on the inactive_works list, will confuse pwq->nr_active
* management later on and cause stall. Make sure the work
* item is activated before grabbing.
*/
if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
pwq_activate_inactive_work(work);
list_del_init(&work->entry);
pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
/* work->data points to pwq iff queued, point to pool */
set_work_pool_and_keep_pending(work, pool->id);
raw_spin_unlock(&pool->lock);
rcu_read_unlock();
return 1;
}
raw_spin_unlock(&pool->lock);
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
fail:
rcu_read_unlock();
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
local_irq_restore(*flags);
if (work_is_canceling(work))
return -ENOENT;
cpu_relax();
return -EAGAIN;
}
/**
* insert_work - insert a work into a pool
* @pwq: pwq @work belongs to
* @work: work to insert
* @head: insertion point
* @extra_flags: extra WORK_STRUCT_* flags to set
*
* Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
* work_struct flags.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*/
static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
struct list_head *head, unsigned int extra_flags)
implement flush_work() A basic problem with flush_scheduled_work() is that it blocks behind _all_ presently-queued works, rather than just the work whcih the caller wants to flush. If the caller holds some lock, and if one of the queued work happens to want that lock as well then accidental deadlocks can occur. One example of this is the phy layer: it wants to flush work while holding rtnl_lock(). But if a linkwatch event happens to be queued, the phy code will deadlock because the linkwatch callback function takes rtnl_lock. So we implement a new function which will flush a *single* work - just the one which the caller wants to free up. Thus we avoid the accidental deadlocks which can arise from unrelated subsystems' callbacks taking shared locks. flush_work() non-blockingly dequeues the work_struct which we want to kill, then it waits for its handler to complete on all CPUs. Add ->current_work to the "struct cpu_workqueue_struct", it points to currently running "struct work_struct". When flush_work(work) detects ->current_work == work, it inserts a barrier at the _head_ of ->worklist (and thus right _after_ that work) and waits for completition. This means that the next work fired on that CPU will be this barrier, or another barrier queued by concurrent flush_work(), so the caller of flush_work() will be woken before any "regular" work has a chance to run. When wait_on_work() unlocks workqueue_mutex (or whatever we choose to protect against CPU hotplug), CPU may go away. But in that case take_over_work() will move a barrier we queued to another CPU, it will be fired sometime, and wait_on_work() will be woken. Actually, we are doing cleanup_workqueue_thread()->kthread_stop() before take_over_work(), so cwq->thread should complete its ->worklist (and thus the barrier), because currently we don't check kthread_should_stop() in run_workqueue(). But even if we did, everything should be ok. [akpm@osdl.org: cleanup] [akpm@osdl.org: add flush_work_keventd() wrapper] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:33:52 +00:00
{
debug_work_activate(work);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
workqueue: kasan: record workqueue stack Patch series "kasan: add workqueue stack for generic KASAN", v5. Syzbot reports many UAF issues for workqueue, see [1]. In some of these access/allocation happened in process_one_work(), we see the free stack is useless in KASAN report, it doesn't help programmers to solve UAF for workqueue issue. This patchset improves KASAN reports by making them to have workqueue queueing stack. It is useful for programmers to solve use-after-free or double-free memory issue. Generic KASAN also records the last two workqueue stacks and prints them in KASAN report. It is only suitable for generic KASAN. [1] https://groups.google.com/g/syzkaller-bugs/search?q=%22use-after-free%22+process_one_work [2] https://bugzilla.kernel.org/show_bug.cgi?id=198437 This patch (of 4): When analyzing use-after-free or double-free issue, recording the enqueuing work stacks is helpful to preserve usage history which potentially gives a hint about the affected code. For workqueue it has turned out to be useful to record the enqueuing work call stacks. Because user can see KASAN report to determine whether it is root cause. They don't need to enable debugobjects, but they have a chance to find out the root cause. Link: https://lkml.kernel.org/r/20201203022148.29754-1-walter-zh.wu@mediatek.com Link: https://lkml.kernel.org/r/20201203022442.30006-1-walter-zh.wu@mediatek.com Signed-off-by: Walter Wu <walter-zh.wu@mediatek.com> Suggested-by: Marco Elver <elver@google.com> Acked-by: Marco Elver <elver@google.com> Acked-by: Tejun Heo <tj@kernel.org> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Marco Elver <elver@google.com> Cc: Matthias Brugger <matthias.bgg@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 03:09:09 +00:00
/* record the work call stack in order to print it in KASAN reports */
workqueue, kasan: avoid alloc_pages() when recording stack Shuah Khan reported: | When CONFIG_PROVE_RAW_LOCK_NESTING=y and CONFIG_KASAN are enabled, | kasan_record_aux_stack() runs into "BUG: Invalid wait context" when | it tries to allocate memory attempting to acquire spinlock in page | allocation code while holding workqueue pool raw_spinlock. | | There are several instances of this problem when block layer tries | to __queue_work(). Call trace from one of these instances is below: | | kblockd_mod_delayed_work_on() | mod_delayed_work_on() | __queue_delayed_work() | __queue_work() (rcu_read_lock, raw_spin_lock pool->lock held) | insert_work() | kasan_record_aux_stack() | kasan_save_stack() | stack_depot_save() | alloc_pages() | __alloc_pages() | get_page_from_freelist() | rm_queue() | rm_queue_pcplist() | local_lock_irqsave(&pagesets.lock, flags); | [ BUG: Invalid wait context triggered ] The default kasan_record_aux_stack() calls stack_depot_save() with GFP_NOWAIT, which in turn can then call alloc_pages(GFP_NOWAIT, ...). In general, however, it is not even possible to use either GFP_ATOMIC nor GFP_NOWAIT in certain non-preemptive contexts, including raw_spin_locks (see gfp.h and commmit ab00db216c9c7). Fix it by instructing stackdepot to not expand stack storage via alloc_pages() in case it runs out by using kasan_record_aux_stack_noalloc(). While there is an increased risk of failing to insert the stack trace, this is typically unlikely, especially if the same insertion had already succeeded previously (stack depot hit). For frequent calls from the same location, it therefore becomes extremely unlikely that kasan_record_aux_stack_noalloc() fails. Link: https://lkml.kernel.org/r/20210902200134.25603-1-skhan@linuxfoundation.org Link: https://lkml.kernel.org/r/20210913112609.2651084-7-elver@google.com Signed-off-by: Marco Elver <elver@google.com> Reported-by: Shuah Khan <skhan@linuxfoundation.org> Tested-by: Shuah Khan <skhan@linuxfoundation.org> Acked-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Acked-by: Tejun Heo <tj@kernel.org> Reviewed-by: Andrey Konovalov <andreyknvl@gmail.com> Cc: Alexander Potapenko <glider@google.com> Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: "Gustavo A. R. Silva" <gustavoars@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Taras Madan <tarasmadan@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vijayanand Jitta <vjitta@codeaurora.org> Cc: Vinayak Menon <vinmenon@codeaurora.org> Cc: Walter Wu <walter-zh.wu@mediatek.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-05 20:35:50 +00:00
kasan_record_aux_stack_noalloc(work);
workqueue: kasan: record workqueue stack Patch series "kasan: add workqueue stack for generic KASAN", v5. Syzbot reports many UAF issues for workqueue, see [1]. In some of these access/allocation happened in process_one_work(), we see the free stack is useless in KASAN report, it doesn't help programmers to solve UAF for workqueue issue. This patchset improves KASAN reports by making them to have workqueue queueing stack. It is useful for programmers to solve use-after-free or double-free memory issue. Generic KASAN also records the last two workqueue stacks and prints them in KASAN report. It is only suitable for generic KASAN. [1] https://groups.google.com/g/syzkaller-bugs/search?q=%22use-after-free%22+process_one_work [2] https://bugzilla.kernel.org/show_bug.cgi?id=198437 This patch (of 4): When analyzing use-after-free or double-free issue, recording the enqueuing work stacks is helpful to preserve usage history which potentially gives a hint about the affected code. For workqueue it has turned out to be useful to record the enqueuing work call stacks. Because user can see KASAN report to determine whether it is root cause. They don't need to enable debugobjects, but they have a chance to find out the root cause. Link: https://lkml.kernel.org/r/20201203022148.29754-1-walter-zh.wu@mediatek.com Link: https://lkml.kernel.org/r/20201203022442.30006-1-walter-zh.wu@mediatek.com Signed-off-by: Walter Wu <walter-zh.wu@mediatek.com> Suggested-by: Marco Elver <elver@google.com> Acked-by: Marco Elver <elver@google.com> Acked-by: Tejun Heo <tj@kernel.org> Reviewed-by: Dmitry Vyukov <dvyukov@google.com> Reviewed-by: Andrey Konovalov <andreyknvl@google.com> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Alexander Potapenko <glider@google.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Marco Elver <elver@google.com> Cc: Matthias Brugger <matthias.bgg@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-15 03:09:09 +00:00
/* we own @work, set data and link */
set_work_pwq(work, pwq, extra_flags);
list_add_tail(&work->entry, head);
get_pwq(pwq);
implement flush_work() A basic problem with flush_scheduled_work() is that it blocks behind _all_ presently-queued works, rather than just the work whcih the caller wants to flush. If the caller holds some lock, and if one of the queued work happens to want that lock as well then accidental deadlocks can occur. One example of this is the phy layer: it wants to flush work while holding rtnl_lock(). But if a linkwatch event happens to be queued, the phy code will deadlock because the linkwatch callback function takes rtnl_lock. So we implement a new function which will flush a *single* work - just the one which the caller wants to free up. Thus we avoid the accidental deadlocks which can arise from unrelated subsystems' callbacks taking shared locks. flush_work() non-blockingly dequeues the work_struct which we want to kill, then it waits for its handler to complete on all CPUs. Add ->current_work to the "struct cpu_workqueue_struct", it points to currently running "struct work_struct". When flush_work(work) detects ->current_work == work, it inserts a barrier at the _head_ of ->worklist (and thus right _after_ that work) and waits for completition. This means that the next work fired on that CPU will be this barrier, or another barrier queued by concurrent flush_work(), so the caller of flush_work() will be woken before any "regular" work has a chance to run. When wait_on_work() unlocks workqueue_mutex (or whatever we choose to protect against CPU hotplug), CPU may go away. But in that case take_over_work() will move a barrier we queued to another CPU, it will be fired sometime, and wait_on_work() will be woken. Actually, we are doing cleanup_workqueue_thread()->kthread_stop() before take_over_work(), so cwq->thread should complete its ->worklist (and thus the barrier), because currently we don't check kthread_should_stop() in run_workqueue(). But even if we did, everything should be ok. [akpm@osdl.org: cleanup] [akpm@osdl.org: add flush_work_keventd() wrapper] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:33:52 +00:00
}
/*
* Test whether @work is being queued from another work executing on the
* same workqueue.
*/
static bool is_chained_work(struct workqueue_struct *wq)
{
struct worker *worker;
worker = current_wq_worker();
/*
* Return %true iff I'm a worker executing a work item on @wq. If
* I'm @worker, it's safe to dereference it without locking.
*/
return worker && worker->current_pwq->wq == wq;
}
/*
* When queueing an unbound work item to a wq, prefer local CPU if allowed
* by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
* avoid perturbing sensitive tasks.
*/
static int wq_select_unbound_cpu(int cpu)
{
int new_cpu;
if (likely(!wq_debug_force_rr_cpu)) {
if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
return cpu;
} else {
pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
}
new_cpu = __this_cpu_read(wq_rr_cpu_last);
new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
if (unlikely(new_cpu >= nr_cpu_ids)) {
new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
if (unlikely(new_cpu >= nr_cpu_ids))
return cpu;
}
__this_cpu_write(wq_rr_cpu_last, new_cpu);
return new_cpu;
}
static void __queue_work(int cpu, struct workqueue_struct *wq,
struct work_struct *work)
{
struct pool_workqueue *pwq;
struct worker_pool *last_pool, *pool;
unsigned int work_flags;
unsigned int req_cpu = cpu;
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
/*
* While a work item is PENDING && off queue, a task trying to
* steal the PENDING will busy-loop waiting for it to either get
* queued or lose PENDING. Grabbing PENDING and queueing should
* happen with IRQ disabled.
*/
lockdep_assert_irqs_disabled();
/*
* For a draining wq, only works from the same workqueue are
* allowed. The __WQ_DESTROYING helps to spot the issue that
* queues a new work item to a wq after destroy_workqueue(wq).
*/
if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
WARN_ON_ONCE(!is_chained_work(wq))))
return;
rcu_read_lock();
retry:
/* pwq which will be used unless @work is executing elsewhere */
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
if (req_cpu == WORK_CPU_UNBOUND) {
if (wq->flags & WQ_UNBOUND)
cpu = wq_select_unbound_cpu(raw_smp_processor_id());
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
else
cpu = raw_smp_processor_id();
}
workqueue: make all workqueues non-reentrant By default, each per-cpu part of a bound workqueue operates separately and a work item may be executing concurrently on different CPUs. The behavior avoids some cross-cpu traffic but leads to subtle weirdities and not-so-subtle contortions in the API. * There's no sane usefulness in allowing a single work item to be executed concurrently on multiple CPUs. People just get the behavior unintentionally and get surprised after learning about it. Most either explicitly synchronize or use non-reentrant/ordered workqueue but this is error-prone. * flush_work() can't wait for multiple instances of the same work item on different CPUs. If a work item is executing on cpu0 and then queued on cpu1, flush_work() can only wait for the one on cpu1. Unfortunately, work items can easily cross CPU boundaries unintentionally when the queueing thread gets migrated. This means that if multiple queuers compete, flush_work() can't even guarantee that the instance queued right before it is finished before returning. * flush_work_sync() was added to work around some of the deficiencies of flush_work(). In addition to the usual flushing, it ensures that all currently executing instances are finished before returning. This operation is expensive as it has to walk all CPUs and at the same time fails to address competing queuer case. Incorrectly using flush_work() when flush_work_sync() is necessary is an easy error to make and can lead to bugs which are difficult to reproduce. * Similar problems exist for flush_delayed_work[_sync](). Other than the cross-cpu access concern, there's no benefit in allowing parallel execution and it's plain silly to have this level of contortion for workqueue which is widely used from core code to extremely obscure drivers. This patch makes all workqueues non-reentrant. If a work item is executing on a different CPU when queueing is requested, it is always queued to that CPU. This guarantees that any given work item can be executing on one CPU at maximum and if a work item is queued and executing, both are on the same CPU. The only behavior change which may affect workqueue users negatively is that non-reentrancy overrides the affinity specified by queue_work_on(). On a reentrant workqueue, the affinity specified by queue_work_on() is always followed. Now, if the work item is executing on one of the CPUs, the work item will be queued there regardless of the requested affinity. I've reviewed all workqueue users which request explicit affinity, and, fortunately, none seems to be crazy enough to exploit parallel execution of the same work item. This adds an additional busy_hash lookup if the work item was previously queued on a different CPU. This shouldn't be noticeable under any sane workload. Work item queueing isn't a very high-frequency operation and they don't jump across CPUs all the time. In a micro benchmark to exaggerate this difference - measuring the time it takes for two work items to repeatedly jump between two CPUs a number (10M) of times with busy_hash table densely populated, the difference was around 3%. While the overhead is measureable, it is only visible in pathological cases and the difference isn't huge. This change brings much needed sanity to workqueue and makes its behavior consistent with timer. I think this is the right tradeoff to make. This enables significant simplification of workqueue API. Simplification patches will follow. Signed-off-by: Tejun Heo <tj@kernel.org>
2012-08-20 21:51:23 +00:00
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
pool = pwq->pool;
/*
* If @work was previously on a different pool, it might still be
* running there, in which case the work needs to be queued on that
* pool to guarantee non-reentrancy.
*/
last_pool = get_work_pool(work);
if (last_pool && last_pool != pool) {
struct worker *worker;
raw_spin_lock(&last_pool->lock);
worker = find_worker_executing_work(last_pool, work);
if (worker && worker->current_pwq->wq == wq) {
pwq = worker->current_pwq;
pool = pwq->pool;
WARN_ON_ONCE(pool != last_pool);
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
} else {
/* meh... not running there, queue here */
raw_spin_unlock(&last_pool->lock);
raw_spin_lock(&pool->lock);
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
}
} else {
raw_spin_lock(&pool->lock);
}
/*
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
* pwq is determined and locked. For unbound pools, we could have raced
* with pwq release and it could already be dead. If its refcnt is zero,
* repeat pwq selection. Note that unbound pwqs never die without
* another pwq replacing it in cpu_pwq or while work items are executing
* on it, so the retrying is guaranteed to make forward-progress.
*/
if (unlikely(!pwq->refcnt)) {
if (wq->flags & WQ_UNBOUND) {
raw_spin_unlock(&pool->lock);
cpu_relax();
goto retry;
}
/* oops */
WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
wq->name, cpu);
}
/* pwq determined, queue */
trace_workqueue_queue_work(req_cpu, pwq, work);
if (WARN_ON(!list_empty(&work->entry)))
goto out;
pwq->nr_in_flight[pwq->work_color]++;
work_flags = work_color_to_flags(pwq->work_color);
if (likely(pwq->nr_active < pwq->max_active)) {
if (list_empty(&pool->worklist))
pool->watchdog_ts = jiffies;
trace_workqueue_activate_work(work);
pwq->nr_active++;
insert_work(pwq, work, &pool->worklist, work_flags);
kick_pool(pool);
} else {
work_flags |= WORK_STRUCT_INACTIVE;
insert_work(pwq, work, &pwq->inactive_works, work_flags);
}
out:
raw_spin_unlock(&pool->lock);
rcu_read_unlock();
}
/**
* queue_work_on - queue work on specific cpu
* @cpu: CPU number to execute work on
* @wq: workqueue to use
* @work: work to queue
*
* We queue the work to a specific CPU, the caller must ensure it
* can't go away. Callers that fail to ensure that the specified
* CPU cannot go away will execute on a randomly chosen CPU.
* But note well that callers specifying a CPU that never has been
* online will get a splat.
*
* Return: %false if @work was already on a queue, %true otherwise.
*/
bool queue_work_on(int cpu, struct workqueue_struct *wq,
struct work_struct *work)
{
bool ret = false;
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
unsigned long flags;
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
local_irq_save(flags);
if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
__queue_work(cpu, wq, work);
ret = true;
}
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
local_irq_restore(flags);
return ret;
}
EXPORT_SYMBOL(queue_work_on);
/**
* select_numa_node_cpu - Select a CPU based on NUMA node
* @node: NUMA node ID that we want to select a CPU from
*
* This function will attempt to find a "random" cpu available on a given
* node. If there are no CPUs available on the given node it will return
* WORK_CPU_UNBOUND indicating that we should just schedule to any
* available CPU if we need to schedule this work.
*/
static int select_numa_node_cpu(int node)
{
int cpu;
/* Delay binding to CPU if node is not valid or online */
if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
return WORK_CPU_UNBOUND;
/* Use local node/cpu if we are already there */
cpu = raw_smp_processor_id();
if (node == cpu_to_node(cpu))
return cpu;
/* Use "random" otherwise know as "first" online CPU of node */
cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
/* If CPU is valid return that, otherwise just defer */
return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
}
/**
* queue_work_node - queue work on a "random" cpu for a given NUMA node
* @node: NUMA node that we are targeting the work for
* @wq: workqueue to use
* @work: work to queue
*
* We queue the work to a "random" CPU within a given NUMA node. The basic
* idea here is to provide a way to somehow associate work with a given
* NUMA node.
*
* This function will only make a best effort attempt at getting this onto
* the right NUMA node. If no node is requested or the requested node is
* offline then we just fall back to standard queue_work behavior.
*
* Currently the "random" CPU ends up being the first available CPU in the
* intersection of cpu_online_mask and the cpumask of the node, unless we
* are running on the node. In that case we just use the current CPU.
*
* Return: %false if @work was already on a queue, %true otherwise.
*/
bool queue_work_node(int node, struct workqueue_struct *wq,
struct work_struct *work)
{
unsigned long flags;
bool ret = false;
/*
* This current implementation is specific to unbound workqueues.
* Specifically we only return the first available CPU for a given
* node instead of cycling through individual CPUs within the node.
*
* If this is used with a per-cpu workqueue then the logic in
* workqueue_select_cpu_near would need to be updated to allow for
* some round robin type logic.
*/
WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
local_irq_save(flags);
if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
int cpu = select_numa_node_cpu(node);
__queue_work(cpu, wq, work);
ret = true;
}
local_irq_restore(flags);
return ret;
}
EXPORT_SYMBOL_GPL(queue_work_node);
workqueue: Convert callback to use from_timer() In preparation for unconditionally passing the struct timer_list pointer to all timer callbacks, switch workqueue to use from_timer() and pass the timer pointer explicitly. Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mips@linux-mips.org Cc: Petr Mladek <pmladek@suse.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Sebastian Reichel <sre@kernel.org> Cc: Kalle Valo <kvalo@qca.qualcomm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Pavel Machek <pavel@ucw.cz> Cc: linux1394-devel@lists.sourceforge.net Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: linux-s390@vger.kernel.org Cc: linux-wireless@vger.kernel.org Cc: "James E.J. Bottomley" <jejb@linux.vnet.ibm.com> Cc: Wim Van Sebroeck <wim@iguana.be> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Ursula Braun <ubraun@linux.vnet.ibm.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Harish Patil <harish.patil@cavium.com> Cc: Stephen Boyd <sboyd@codeaurora.org> Cc: Guenter Roeck <linux@roeck-us.net> Cc: Manish Chopra <manish.chopra@cavium.com> Cc: Len Brown <len.brown@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: linux-pm@vger.kernel.org Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Julian Wiedmann <jwi@linux.vnet.ibm.com> Cc: John Stultz <john.stultz@linaro.org> Cc: Mark Gross <mark.gross@intel.com> Cc: linux-watchdog@vger.kernel.org Cc: linux-scsi@vger.kernel.org Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Stefan Richter <stefanr@s5r6.in-berlin.de> Cc: Michael Reed <mdr@sgi.com> Cc: netdev@vger.kernel.org Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linuxppc-dev@lists.ozlabs.org Cc: Sudip Mukherjee <sudipm.mukherjee@gmail.com> Link: https://lkml.kernel.org/r/1507159627-127660-14-git-send-email-keescook@chromium.org
2017-10-04 23:27:07 +00:00
void delayed_work_timer_fn(struct timer_list *t)
{
workqueue: Convert callback to use from_timer() In preparation for unconditionally passing the struct timer_list pointer to all timer callbacks, switch workqueue to use from_timer() and pass the timer pointer explicitly. Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mips@linux-mips.org Cc: Petr Mladek <pmladek@suse.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Sebastian Reichel <sre@kernel.org> Cc: Kalle Valo <kvalo@qca.qualcomm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Pavel Machek <pavel@ucw.cz> Cc: linux1394-devel@lists.sourceforge.net Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: linux-s390@vger.kernel.org Cc: linux-wireless@vger.kernel.org Cc: "James E.J. Bottomley" <jejb@linux.vnet.ibm.com> Cc: Wim Van Sebroeck <wim@iguana.be> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Ursula Braun <ubraun@linux.vnet.ibm.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Harish Patil <harish.patil@cavium.com> Cc: Stephen Boyd <sboyd@codeaurora.org> Cc: Guenter Roeck <linux@roeck-us.net> Cc: Manish Chopra <manish.chopra@cavium.com> Cc: Len Brown <len.brown@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: linux-pm@vger.kernel.org Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Julian Wiedmann <jwi@linux.vnet.ibm.com> Cc: John Stultz <john.stultz@linaro.org> Cc: Mark Gross <mark.gross@intel.com> Cc: linux-watchdog@vger.kernel.org Cc: linux-scsi@vger.kernel.org Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Stefan Richter <stefanr@s5r6.in-berlin.de> Cc: Michael Reed <mdr@sgi.com> Cc: netdev@vger.kernel.org Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linuxppc-dev@lists.ozlabs.org Cc: Sudip Mukherjee <sudipm.mukherjee@gmail.com> Link: https://lkml.kernel.org/r/1507159627-127660-14-git-send-email-keescook@chromium.org
2017-10-04 23:27:07 +00:00
struct delayed_work *dwork = from_timer(dwork, t, timer);
/* should have been called from irqsafe timer with irq already off */
workqueue: add delayed_work->wq to simplify reentrancy handling To avoid executing the same work item from multiple CPUs concurrently, a work_struct records the last pool it was on in its ->data so that, on the next queueing, the pool can be queried to determine whether the work item is still executing or not. A delayed_work goes through timer before actually being queued on the target workqueue and the timer needs to know the target workqueue and CPU. This is currently achieved by modifying delayed_work->work.data such that it points to the cwq which points to the target workqueue and the last CPU the work item was on. __queue_delayed_work() extracts the last CPU from delayed_work->work.data and then combines it with the target workqueue to create new work.data. The only thing this rather ugly hack achieves is encoding the target workqueue into delayed_work->work.data without using a separate field, which could be a trade off one can make; unfortunately, this entangles work->data management between regular workqueue and delayed_work code by setting cwq pointer before the work item is actually queued and becomes a hindrance for further improvements of work->data handling. This can be easily made sane by adding a target workqueue field to delayed_work. While delayed_work is used widely in the kernel and this does make it a bit larger (<5%), I think this is the right trade-off especially given the prospect of much saner handling of work->data which currently involves quite tricky memory barrier dancing, and don't expect to see any measureable effect. Add delayed_work->wq and drop the delayed_work->work.data overloading. tj: Rewrote the description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
__queue_work(dwork->cpu, dwork->wq, &dwork->work);
}
EXPORT_SYMBOL(delayed_work_timer_fn);
static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
struct delayed_work *dwork, unsigned long delay)
{
struct timer_list *timer = &dwork->timer;
struct work_struct *work = &dwork->work;
WARN_ON_ONCE(!wq);
WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
WARN_ON_ONCE(timer_pending(timer));
WARN_ON_ONCE(!list_empty(&work->entry));
workqueue: mod_delayed_work_on() shouldn't queue timer on 0 delay 8376fe22c7 ("workqueue: implement mod_delayed_work[_on]()") implemented mod_delayed_work[_on]() using the improved try_to_grab_pending(). The function is later used, among others, to replace [__]candel_delayed_work() + queue_delayed_work() combinations. Unfortunately, a delayed_work item w/ zero @delay is handled slightly differently by mod_delayed_work_on() compared to queue_delayed_work_on(). The latter skips timer altogether and directly queues it using queue_work_on() while the former schedules timer which will expire on the closest tick. This means, when @delay is zero, that [__]cancel_delayed_work() + queue_delayed_work_on() makes the target item immediately executable while mod_delayed_work_on() may induce delay of upto a full tick. This somewhat subtle difference breaks some of the converted users. e.g. block queue plugging uses delayed_work for deferred processing and uses mod_delayed_work_on() when the queue needs to be immediately unplugged. The above problem manifested as noticeably higher number of context switches under certain circumstances. The difference in behavior was caused by missing special case handling for 0 delay in mod_delayed_work_on() compared to queue_delayed_work_on(). Joonsoo Kim posted a patch to add it - ("workqueue: optimize mod_delayed_work_on() when @delay == 0")[1]. The patch was queued for 3.8 but it was described as optimization and I missed that it was a correctness issue. As both queue_delayed_work_on() and mod_delayed_work_on() use __queue_delayed_work() for queueing, it seems that the better approach is to move the 0 delay special handling to the function instead of duplicating it in mod_delayed_work_on(). Fix the problem by moving 0 delay special case handling from queue_delayed_work_on() to __queue_delayed_work(). This replaces Joonsoo's patch. [1] http://thread.gmane.org/gmane.linux.kernel/1379011/focus=1379012 Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-tested-by: Anders Kaseorg <andersk@MIT.EDU> Reported-and-tested-by: Zlatko Calusic <zlatko.calusic@iskon.hr> LKML-Reference: <alpine.DEB.2.00.1211280953350.26602@dr-wily.mit.edu> LKML-Reference: <50A78AA9.5040904@iskon.hr> Cc: Joonsoo Kim <js1304@gmail.com>
2012-12-02 00:23:42 +00:00
/*
* If @delay is 0, queue @dwork->work immediately. This is for
* both optimization and correctness. The earliest @timer can
* expire is on the closest next tick and delayed_work users depend
* on that there's no such delay when @delay is 0.
*/
if (!delay) {
__queue_work(cpu, wq, &dwork->work);
return;
}
workqueue: add delayed_work->wq to simplify reentrancy handling To avoid executing the same work item from multiple CPUs concurrently, a work_struct records the last pool it was on in its ->data so that, on the next queueing, the pool can be queried to determine whether the work item is still executing or not. A delayed_work goes through timer before actually being queued on the target workqueue and the timer needs to know the target workqueue and CPU. This is currently achieved by modifying delayed_work->work.data such that it points to the cwq which points to the target workqueue and the last CPU the work item was on. __queue_delayed_work() extracts the last CPU from delayed_work->work.data and then combines it with the target workqueue to create new work.data. The only thing this rather ugly hack achieves is encoding the target workqueue into delayed_work->work.data without using a separate field, which could be a trade off one can make; unfortunately, this entangles work->data management between regular workqueue and delayed_work code by setting cwq pointer before the work item is actually queued and becomes a hindrance for further improvements of work->data handling. This can be easily made sane by adding a target workqueue field to delayed_work. While delayed_work is used widely in the kernel and this does make it a bit larger (<5%), I think this is the right trade-off especially given the prospect of much saner handling of work->data which currently involves quite tricky memory barrier dancing, and don't expect to see any measureable effect. Add delayed_work->wq and drop the delayed_work->work.data overloading. tj: Rewrote the description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
dwork->wq = wq;
dwork->cpu = cpu;
timer->expires = jiffies + delay;
Revert "workqueue: make sure delayed work run in local cpu" This reverts commit 874bbfe600a660cba9c776b3957b1ce393151b76. Workqueue used to implicity guarantee that work items queued without explicit CPU specified are put on the local CPU. Recent changes in timer broke the guarantee and led to vmstat breakage which was fixed by 176bed1de5bf ("vmstat: explicitly schedule per-cpu work on the CPU we need it to run on"). vmstat is the most likely to expose the issue and it's quite possible that there are other similar problems which are a lot more difficult to trigger. As a preventive measure, 874bbfe600a6 ("workqueue: make sure delayed work run in local cpu") was applied to restore the local CPU guarnatee. Unfortunately, the change exposed a bug in timer code which got fixed by 22b886dd1018 ("timers: Use proper base migration in add_timer_on()"). Due to code restructuring, the commit couldn't be backported beyond certain point and stable kernels which only had 874bbfe600a6 started crashing. The local CPU guarantee was accidental more than anything else and we want to get rid of it anyway. As, with the vmstat case fixed, 874bbfe600a6 is causing more problems than it's fixing, it has been decided to take the chance and officially break the guarantee by reverting the commit. A debug feature will be added to force foreign CPU assignment to expose cases relying on the guarantee and fixes for the individual cases will be backported to stable as necessary. Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: 874bbfe600a6 ("workqueue: make sure delayed work run in local cpu") Link: http://lkml.kernel.org/g/20160120211926.GJ10810@quack.suse.cz Cc: stable@vger.kernel.org Cc: Mike Galbraith <umgwanakikbuti@gmail.com> Cc: Henrique de Moraes Holschuh <hmh@hmh.eng.br> Cc: Daniel Bilik <daniel.bilik@neosystem.cz> Cc: Jan Kara <jack@suse.cz> Cc: Shaohua Li <shli@fb.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Ben Hutchings <ben@decadent.org.uk> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Daniel Bilik <daniel.bilik@neosystem.cz> Cc: Jiri Slaby <jslaby@suse.cz> Cc: Michal Hocko <mhocko@kernel.org>
2016-02-09 21:11:26 +00:00
if (unlikely(cpu != WORK_CPU_UNBOUND))
add_timer_on(timer, cpu);
else
add_timer(timer);
}
/**
* queue_delayed_work_on - queue work on specific CPU after delay
* @cpu: CPU number to execute work on
* @wq: workqueue to use
* @dwork: work to queue
* @delay: number of jiffies to wait before queueing
*
* Return: %false if @work was already on a queue, %true otherwise. If
* @delay is zero and @dwork is idle, it will be scheduled for immediate
* execution.
*/
bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
struct delayed_work *dwork, unsigned long delay)
{
struct work_struct *work = &dwork->work;
bool ret = false;
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
unsigned long flags;
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
/* read the comment in __queue_work() */
local_irq_save(flags);
if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
__queue_delayed_work(cpu, wq, dwork, delay);
ret = true;
}
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
local_irq_restore(flags);
return ret;
}
EXPORT_SYMBOL(queue_delayed_work_on);
/**
* mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
* @cpu: CPU number to execute work on
* @wq: workqueue to use
* @dwork: work to queue
* @delay: number of jiffies to wait before queueing
*
* If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
* modify @dwork's timer so that it expires after @delay. If @delay is
* zero, @work is guaranteed to be scheduled immediately regardless of its
* current state.
*
* Return: %false if @dwork was idle and queued, %true if @dwork was
* pending and its timer was modified.
*
* This function is safe to call from any context including IRQ handler.
* See try_to_grab_pending() for details.
*/
bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
struct delayed_work *dwork, unsigned long delay)
{
unsigned long flags;
int ret;
do {
ret = try_to_grab_pending(&dwork->work, true, &flags);
} while (unlikely(ret == -EAGAIN));
if (likely(ret >= 0)) {
__queue_delayed_work(cpu, wq, dwork, delay);
local_irq_restore(flags);
}
/* -ENOENT from try_to_grab_pending() becomes %true */
return ret;
}
EXPORT_SYMBOL_GPL(mod_delayed_work_on);
static void rcu_work_rcufn(struct rcu_head *rcu)
{
struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
/* read the comment in __queue_work() */
local_irq_disable();
__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
local_irq_enable();
}
/**
* queue_rcu_work - queue work after a RCU grace period
* @wq: workqueue to use
* @rwork: work to queue
*
* Return: %false if @rwork was already pending, %true otherwise. Note
* that a full RCU grace period is guaranteed only after a %true return.
* While @rwork is guaranteed to be executed after a %false return, the
* execution may happen before a full RCU grace period has passed.
*/
bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
{
struct work_struct *work = &rwork->work;
if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
rwork->wq = wq;
workqueue: Make queue_rcu_work() use call_rcu_hurry() Earlier commits in this series allow battery-powered systems to build their kernels with the default-disabled CONFIG_RCU_LAZY=y Kconfig option. This Kconfig option causes call_rcu() to delay its callbacks in order to batch them. This means that a given RCU grace period covers more callbacks, thus reducing the number of grace periods, in turn reducing the amount of energy consumed, which increases battery lifetime which can be a very good thing. This is not a subtle effect: In some important use cases, the battery lifetime is increased by more than 10%. This CONFIG_RCU_LAZY=y option is available only for CPUs that offload callbacks, for example, CPUs mentioned in the rcu_nocbs kernel boot parameter passed to kernels built with CONFIG_RCU_NOCB_CPU=y. Delaying callbacks is normally not a problem because most callbacks do nothing but free memory. If the system is short on memory, a shrinker will kick all currently queued lazy callbacks out of their laziness, thus freeing their memory in short order. Similarly, the rcu_barrier() function, which blocks until all currently queued callbacks are invoked, will also kick lazy callbacks, thus enabling rcu_barrier() to complete in a timely manner. However, there are some cases where laziness is not a good option. For example, synchronize_rcu() invokes call_rcu(), and blocks until the newly queued callback is invoked. It would not be a good for synchronize_rcu() to block for ten seconds, even on an idle system. Therefore, synchronize_rcu() invokes call_rcu_hurry() instead of call_rcu(). The arrival of a non-lazy call_rcu_hurry() callback on a given CPU kicks any lazy callbacks that might be already queued on that CPU. After all, if there is going to be a grace period, all callbacks might as well get full benefit from it. Yes, this could be done the other way around by creating a call_rcu_lazy(), but earlier experience with this approach and feedback at the 2022 Linux Plumbers Conference shifted the approach to call_rcu() being lazy with call_rcu_hurry() for the few places where laziness is inappropriate. And another call_rcu() instance that cannot be lazy is the one in queue_rcu_work(), given that callers to queue_rcu_work() are not necessarily OK with long delays. Therefore, make queue_rcu_work() use call_rcu_hurry() in order to revert to the old behavior. [ paulmck: Apply s/call_rcu_flush/call_rcu_hurry/ feedback from Tejun Heo. ] Signed-off-by: Uladzislau Rezki <urezki@gmail.com> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org> Acked-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
2022-10-16 16:23:03 +00:00
call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
return true;
}
return false;
}
EXPORT_SYMBOL(queue_rcu_work);
static struct worker *alloc_worker(int node)
{
struct worker *worker;
worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
if (worker) {
INIT_LIST_HEAD(&worker->entry);
INIT_LIST_HEAD(&worker->scheduled);
INIT_LIST_HEAD(&worker->node);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* on creation a worker is in !idle && prep state */
worker->flags = WORKER_PREP;
}
return worker;
}
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
{
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
if (pool->cpu < 0 && pool->attrs->affn_strict)
return pool->attrs->__pod_cpumask;
else
return pool->attrs->cpumask;
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
}
/**
* worker_attach_to_pool() - attach a worker to a pool
* @worker: worker to be attached
* @pool: the target pool
*
* Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
* cpu-binding of @worker are kept coordinated with the pool across
* cpu-[un]hotplugs.
*/
static void worker_attach_to_pool(struct worker *worker,
struct worker_pool *pool)
{
mutex_lock(&wq_pool_attach_mutex);
/*
* The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
* stable across this function. See the comments above the flag
* definition for details.
*/
if (pool->flags & POOL_DISASSOCIATED)
worker->flags |= WORKER_UNBOUND;
else
kthread_set_per_cpu(worker->task, pool->cpu);
if (worker->rescue_wq)
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
list_add_tail(&worker->node, &pool->workers);
worker->pool = pool;
mutex_unlock(&wq_pool_attach_mutex);
}
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
/**
* worker_detach_from_pool() - detach a worker from its pool
* @worker: worker which is attached to its pool
*
* Undo the attaching which had been done in worker_attach_to_pool(). The
* caller worker shouldn't access to the pool after detached except it has
* other reference to the pool.
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
*/
static void worker_detach_from_pool(struct worker *worker)
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
{
struct worker_pool *pool = worker->pool;
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
struct completion *detach_completion = NULL;
mutex_lock(&wq_pool_attach_mutex);
kthread_set_per_cpu(worker->task, -1);
list_del(&worker->node);
worker->pool = NULL;
if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
detach_completion = pool->detach_completion;
mutex_unlock(&wq_pool_attach_mutex);
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
/* clear leftover flags without pool->lock after it is detached */
worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
if (detach_completion)
complete(detach_completion);
}
/**
* create_worker - create a new workqueue worker
* @pool: pool the new worker will belong to
*
* Create and start a new worker which is attached to @pool.
*
* CONTEXT:
* Might sleep. Does GFP_KERNEL allocations.
*
* Return:
* Pointer to the newly created worker.
*/
2012-07-17 19:39:27 +00:00
static struct worker *create_worker(struct worker_pool *pool)
{
struct worker *worker;
int id;
char id_buf[23];
/* ID is needed to determine kthread name */
id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
if (id < 0) {
pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
ERR_PTR(id));
return NULL;
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
}
worker = alloc_worker(pool->node);
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
if (!worker) {
pr_err_once("workqueue: Failed to allocate a worker\n");
goto fail;
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
}
worker->id = id;
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
if (pool->cpu >= 0)
snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
pool->attrs->nice < 0 ? "H" : "");
else
snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
"kworker/%s", id_buf);
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
if (IS_ERR(worker->task)) {
if (PTR_ERR(worker->task) == -EINTR) {
pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
id_buf);
} else {
pr_err_once("workqueue: Failed to create a worker thread: %pe",
worker->task);
}
goto fail;
workqueue: Warn when a new worker could not be created The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. The progress is guaranteed by using idle workers or creating new workers for pending work items. There are several reasons why a new worker could not be created: + there is not enough memory + there is no free pool ID (IDR API) + the system reached PID limit + the process creating the new worker was interrupted + the last idle worker (manager) has not been scheduled for a long time. It was not able to even start creating the kthread. None of these failures is reported at the moment. The only clue is that show_one_worker_pool() prints that there is a manager. It is the last idle worker that is responsible for creating a new one. But it is not clear if create_worker() is failing and why. Make the debugging easier by printing errors in create_worker(). The error code is important, especially from kthread_create_on_node(). It helps to distinguish the various reasons. For example, reaching memory limit (-ENOMEM), other system limits (-EAGAIN), or process interrupted (-EINTR). Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN for each stuck worker pool. Ratelimited printk() might be better. It would help to know if the problem remains. It would be more clear if the create_worker() errors and workqueue stalls are related. Also old messages might get lost when the internal log buffer is full. The problem is that printk() might touch the watchdog. For example, see touch_nmi_watchdog() in serial8250_console_write(). It would require synchronization of the begin and length of the ratelimit interval with the workqueue watchdog. Otherwise, the error messages might break the watchdog. This does not look worth the complexity. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:32 +00:00
}
set_user_nice(worker->task, pool->attrs->nice);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
/* successful, attach the worker to the pool */
worker_attach_to_pool(worker, pool);
/* start the newly created worker */
raw_spin_lock_irq(&pool->lock);
worker->pool->nr_workers++;
worker_enter_idle(worker);
kick_pool(pool);
/*
* @worker is waiting on a completion in kthread() and will trigger hung
* check if not woken up soon. As kick_pool() might not have waken it
* up, wake it up explicitly once more.
*/
wake_up_process(worker->task);
raw_spin_unlock_irq(&pool->lock);
return worker;
fail:
ida_free(&pool->worker_ida, id);
kfree(worker);
return NULL;
}
static void unbind_worker(struct worker *worker)
{
lockdep_assert_held(&wq_pool_attach_mutex);
kthread_set_per_cpu(worker->task, -1);
if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
else
WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
}
static void wake_dying_workers(struct list_head *cull_list)
{
struct worker *worker, *tmp;
list_for_each_entry_safe(worker, tmp, cull_list, entry) {
list_del_init(&worker->entry);
unbind_worker(worker);
/*
* If the worker was somehow already running, then it had to be
* in pool->idle_list when set_worker_dying() happened or we
* wouldn't have gotten here.
*
* Thus, the worker must either have observed the WORKER_DIE
* flag, or have set its state to TASK_IDLE. Either way, the
* below will be observed by the worker and is safe to do
* outside of pool->lock.
*/
wake_up_process(worker->task);
}
}
/**
* set_worker_dying - Tag a worker for destruction
* @worker: worker to be destroyed
* @list: transfer worker away from its pool->idle_list and into list
*
* Tag @worker for destruction and adjust @pool stats accordingly. The worker
* should be idle.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*/
static void set_worker_dying(struct worker *worker, struct list_head *list)
{
struct worker_pool *pool = worker->pool;
lockdep_assert_held(&pool->lock);
lockdep_assert_held(&wq_pool_attach_mutex);
/* sanity check frenzy */
if (WARN_ON(worker->current_work) ||
workqueue: destroy_worker() should destroy idle workers only We used to have the CPU online failure path where a worker is created and then destroyed without being started. A worker was created for the CPU coming online and if the online operation failed the created worker was shut down without being started. But this behavior was changed. The first worker is created and started at the same time for the CPU coming online. It means that we had already ensured in the code that destroy_worker() destroys only idle workers and we don't want to allow it to destroy any non-idle worker in the future. Otherwise, it may be buggy and it may be extremely hard to check. We should force destroy_worker() to destroy only idle workers explicitly. Since destroy_worker() destroys only idle workers, this patch does not change any functionality. We just need to update the comments and the sanity check code. In the sanity check code, we will refuse to destroy the worker if !(worker->flags & WORKER_IDLE). If the worker entered idle which means it is already started, so we remove the check of "worker->flags & WORKER_STARTED", after this removal, WORKER_STARTED is totally unneeded, so we remove WORKER_STARTED too. In the comments for create_worker(), "Create a new worker which is bound..." is changed to "... which is attached..." due to we change the name of this behavior to attaching. tj: Minor description / comment updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:28 +00:00
WARN_ON(!list_empty(&worker->scheduled)) ||
WARN_ON(!(worker->flags & WORKER_IDLE)))
return;
workqueue: destroy_worker() should destroy idle workers only We used to have the CPU online failure path where a worker is created and then destroyed without being started. A worker was created for the CPU coming online and if the online operation failed the created worker was shut down without being started. But this behavior was changed. The first worker is created and started at the same time for the CPU coming online. It means that we had already ensured in the code that destroy_worker() destroys only idle workers and we don't want to allow it to destroy any non-idle worker in the future. Otherwise, it may be buggy and it may be extremely hard to check. We should force destroy_worker() to destroy only idle workers explicitly. Since destroy_worker() destroys only idle workers, this patch does not change any functionality. We just need to update the comments and the sanity check code. In the sanity check code, we will refuse to destroy the worker if !(worker->flags & WORKER_IDLE). If the worker entered idle which means it is already started, so we remove the check of "worker->flags & WORKER_STARTED", after this removal, WORKER_STARTED is totally unneeded, so we remove WORKER_STARTED too. In the comments for create_worker(), "Create a new worker which is bound..." is changed to "... which is attached..." due to we change the name of this behavior to attaching. tj: Minor description / comment updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:28 +00:00
pool->nr_workers--;
pool->nr_idle--;
worker->flags |= WORKER_DIE;
list_move(&worker->entry, list);
list_move(&worker->node, &pool->dying_workers);
}
/**
* idle_worker_timeout - check if some idle workers can now be deleted.
* @t: The pool's idle_timer that just expired
*
* The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
* worker_leave_idle(), as a worker flicking between idle and active while its
* pool is at the too_many_workers() tipping point would cause too much timer
* housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
* it expire and re-evaluate things from there.
*/
static void idle_worker_timeout(struct timer_list *t)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct worker_pool *pool = from_timer(pool, t, idle_timer);
bool do_cull = false;
if (work_pending(&pool->idle_cull_work))
return;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
raw_spin_lock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
if (too_many_workers(pool)) {
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
struct worker *worker;
unsigned long expires;
/* idle_list is kept in LIFO order, check the last one */
worker = list_entry(pool->idle_list.prev, struct worker, entry);
expires = worker->last_active + IDLE_WORKER_TIMEOUT;
do_cull = !time_before(jiffies, expires);
if (!do_cull)
mod_timer(&pool->idle_timer, expires);
}
raw_spin_unlock_irq(&pool->lock);
if (do_cull)
queue_work(system_unbound_wq, &pool->idle_cull_work);
}
/**
* idle_cull_fn - cull workers that have been idle for too long.
* @work: the pool's work for handling these idle workers
*
* This goes through a pool's idle workers and gets rid of those that have been
* idle for at least IDLE_WORKER_TIMEOUT seconds.
*
* We don't want to disturb isolated CPUs because of a pcpu kworker being
* culled, so this also resets worker affinity. This requires a sleepable
* context, hence the split between timer callback and work item.
*/
static void idle_cull_fn(struct work_struct *work)
{
struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
LIST_HEAD(cull_list);
/*
* Grabbing wq_pool_attach_mutex here ensures an already-running worker
* cannot proceed beyong worker_detach_from_pool() in its self-destruct
* path. This is required as a previously-preempted worker could run after
* set_worker_dying() has happened but before wake_dying_workers() did.
*/
mutex_lock(&wq_pool_attach_mutex);
raw_spin_lock_irq(&pool->lock);
while (too_many_workers(pool)) {
struct worker *worker;
unsigned long expires;
worker = list_entry(pool->idle_list.prev, struct worker, entry);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
expires = worker->last_active + IDLE_WORKER_TIMEOUT;
if (time_before(jiffies, expires)) {
mod_timer(&pool->idle_timer, expires);
break;
}
set_worker_dying(worker, &cull_list);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
raw_spin_unlock_irq(&pool->lock);
wake_dying_workers(&cull_list);
mutex_unlock(&wq_pool_attach_mutex);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
static void send_mayday(struct work_struct *work)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct pool_workqueue *pwq = get_work_pwq(work);
struct workqueue_struct *wq = pwq->wq;
lockdep_assert_held(&wq_mayday_lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
if (!wq->rescuer)
return;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* mayday mayday mayday */
if (list_empty(&pwq->mayday_node)) {
/*
* If @pwq is for an unbound wq, its base ref may be put at
* any time due to an attribute change. Pin @pwq until the
* rescuer is done with it.
*/
get_pwq(pwq);
list_add_tail(&pwq->mayday_node, &wq->maydays);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
wake_up_process(wq->rescuer->task);
pwq->stats[PWQ_STAT_MAYDAY]++;
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
static void pool_mayday_timeout(struct timer_list *t)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct worker_pool *pool = from_timer(pool, t, mayday_timer);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
struct work_struct *work;
raw_spin_lock_irq(&pool->lock);
raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
if (need_to_create_worker(pool)) {
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
* We've been trying to create a new worker but
* haven't been successful. We might be hitting an
* allocation deadlock. Send distress signals to
* rescuers.
*/
list_for_each_entry(work, &pool->worklist, entry)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
send_mayday(work);
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
raw_spin_unlock(&wq_mayday_lock);
raw_spin_unlock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/**
* maybe_create_worker - create a new worker if necessary
* @pool: pool to create a new worker for
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*
* Create a new worker for @pool if necessary. @pool is guaranteed to
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* have at least one idle worker on return from this function. If
* creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
* sent to all rescuers with works scheduled on @pool to resolve
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* possible allocation deadlock.
*
* On return, need_to_create_worker() is guaranteed to be %false and
* may_start_working() %true.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*
* LOCKING:
* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* multiple times. Does GFP_KERNEL allocations. Called only from
* manager.
*/
workqueue: fix subtle pool management issue which can stall whole worker_pool A worker_pool's forward progress is guaranteed by the fact that the last idle worker assumes the manager role to create more workers and summon the rescuers if creating workers doesn't succeed in timely manner before proceeding to execute work items. This manager role is implemented in manage_workers(), which indicates whether the worker may proceed to work item execution with its return value. This is necessary because multiple workers may contend for the manager role, and, if there already is a manager, others should proceed to work item execution. Unfortunately, the function also indicates that the worker may proceed to work item execution if need_to_create_worker() is false at the head of the function. need_to_create_worker() tests the following conditions. pending work items && !nr_running && !nr_idle The first and third conditions are protected by pool->lock and thus won't change while holding pool->lock; however, nr_running can change asynchronously as other workers block and resume and while it's likely to be zero, as someone woke this worker up in the first place, some other workers could have become runnable inbetween making it non-zero. If this happens, manage_worker() could return false even with zero nr_idle making the worker, the last idle one, proceed to execute work items. If then all workers of the pool end up blocking on a resource which can only be released by a work item which is pending on that pool, the whole pool can deadlock as there's no one to create more workers or summon the rescuers. This patch fixes the problem by removing the early exit condition from maybe_create_worker() and making manage_workers() return false iff there's already another manager, which ensures that the last worker doesn't start executing work items. We can leave the early exit condition alone and just ignore the return value but the only reason it was put there is because the manage_workers() used to perform both creations and destructions of workers and thus the function may be invoked while the pool is trying to reduce the number of workers. Now that manage_workers() is called only when more workers are needed, the only case this early exit condition is triggered is rare race conditions rendering it pointless. Tested with simulated workload and modified workqueue code which trigger the pool deadlock reliably without this patch. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Eric Sandeen <sandeen@sandeen.net> Link: http://lkml.kernel.org/g/54B019F4.8030009@sandeen.net Cc: Dave Chinner <david@fromorbit.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: stable@vger.kernel.org
2015-01-16 19:21:16 +00:00
static void maybe_create_worker(struct worker_pool *pool)
__releases(&pool->lock)
__acquires(&pool->lock)
{
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
restart:
raw_spin_unlock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
while (true) {
if (create_worker(pool) || !need_to_create_worker(pool))
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
break;
schedule_timeout_interruptible(CREATE_COOLDOWN);
if (!need_to_create_worker(pool))
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
break;
}
del_timer_sync(&pool->mayday_timer);
raw_spin_lock_irq(&pool->lock);
/*
* This is necessary even after a new worker was just successfully
* created as @pool->lock was dropped and the new worker might have
* already become busy.
*/
if (need_to_create_worker(pool))
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
goto restart;
}
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/**
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* manage_workers - manage worker pool
* @worker: self
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*
* Assume the manager role and manage the worker pool @worker belongs
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* to. At any given time, there can be only zero or one manager per
* pool. The exclusion is handled automatically by this function.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*
* The caller can safely start processing works on false return. On
* true return, it's guaranteed that need_to_create_worker() is false
* and may_start_working() is true.
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* multiple times. Does GFP_KERNEL allocations.
*
* Return:
workqueue: fix subtle pool management issue which can stall whole worker_pool A worker_pool's forward progress is guaranteed by the fact that the last idle worker assumes the manager role to create more workers and summon the rescuers if creating workers doesn't succeed in timely manner before proceeding to execute work items. This manager role is implemented in manage_workers(), which indicates whether the worker may proceed to work item execution with its return value. This is necessary because multiple workers may contend for the manager role, and, if there already is a manager, others should proceed to work item execution. Unfortunately, the function also indicates that the worker may proceed to work item execution if need_to_create_worker() is false at the head of the function. need_to_create_worker() tests the following conditions. pending work items && !nr_running && !nr_idle The first and third conditions are protected by pool->lock and thus won't change while holding pool->lock; however, nr_running can change asynchronously as other workers block and resume and while it's likely to be zero, as someone woke this worker up in the first place, some other workers could have become runnable inbetween making it non-zero. If this happens, manage_worker() could return false even with zero nr_idle making the worker, the last idle one, proceed to execute work items. If then all workers of the pool end up blocking on a resource which can only be released by a work item which is pending on that pool, the whole pool can deadlock as there's no one to create more workers or summon the rescuers. This patch fixes the problem by removing the early exit condition from maybe_create_worker() and making manage_workers() return false iff there's already another manager, which ensures that the last worker doesn't start executing work items. We can leave the early exit condition alone and just ignore the return value but the only reason it was put there is because the manage_workers() used to perform both creations and destructions of workers and thus the function may be invoked while the pool is trying to reduce the number of workers. Now that manage_workers() is called only when more workers are needed, the only case this early exit condition is triggered is rare race conditions rendering it pointless. Tested with simulated workload and modified workqueue code which trigger the pool deadlock reliably without this patch. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Eric Sandeen <sandeen@sandeen.net> Link: http://lkml.kernel.org/g/54B019F4.8030009@sandeen.net Cc: Dave Chinner <david@fromorbit.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: stable@vger.kernel.org
2015-01-16 19:21:16 +00:00
* %false if the pool doesn't need management and the caller can safely
* start processing works, %true if management function was performed and
* the conditions that the caller verified before calling the function may
* no longer be true.
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*/
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
static bool manage_workers(struct worker *worker)
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
{
struct worker_pool *pool = worker->pool;
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
if (pool->flags & POOL_MANAGER_ACTIVE)
workqueue: fix subtle pool management issue which can stall whole worker_pool A worker_pool's forward progress is guaranteed by the fact that the last idle worker assumes the manager role to create more workers and summon the rescuers if creating workers doesn't succeed in timely manner before proceeding to execute work items. This manager role is implemented in manage_workers(), which indicates whether the worker may proceed to work item execution with its return value. This is necessary because multiple workers may contend for the manager role, and, if there already is a manager, others should proceed to work item execution. Unfortunately, the function also indicates that the worker may proceed to work item execution if need_to_create_worker() is false at the head of the function. need_to_create_worker() tests the following conditions. pending work items && !nr_running && !nr_idle The first and third conditions are protected by pool->lock and thus won't change while holding pool->lock; however, nr_running can change asynchronously as other workers block and resume and while it's likely to be zero, as someone woke this worker up in the first place, some other workers could have become runnable inbetween making it non-zero. If this happens, manage_worker() could return false even with zero nr_idle making the worker, the last idle one, proceed to execute work items. If then all workers of the pool end up blocking on a resource which can only be released by a work item which is pending on that pool, the whole pool can deadlock as there's no one to create more workers or summon the rescuers. This patch fixes the problem by removing the early exit condition from maybe_create_worker() and making manage_workers() return false iff there's already another manager, which ensures that the last worker doesn't start executing work items. We can leave the early exit condition alone and just ignore the return value but the only reason it was put there is because the manage_workers() used to perform both creations and destructions of workers and thus the function may be invoked while the pool is trying to reduce the number of workers. Now that manage_workers() is called only when more workers are needed, the only case this early exit condition is triggered is rare race conditions rendering it pointless. Tested with simulated workload and modified workqueue code which trigger the pool deadlock reliably without this patch. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Eric Sandeen <sandeen@sandeen.net> Link: http://lkml.kernel.org/g/54B019F4.8030009@sandeen.net Cc: Dave Chinner <david@fromorbit.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: stable@vger.kernel.org
2015-01-16 19:21:16 +00:00
return false;
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
pool->flags |= POOL_MANAGER_ACTIVE;
pool->manager = worker;
workqueue: fix subtle pool management issue which can stall whole worker_pool A worker_pool's forward progress is guaranteed by the fact that the last idle worker assumes the manager role to create more workers and summon the rescuers if creating workers doesn't succeed in timely manner before proceeding to execute work items. This manager role is implemented in manage_workers(), which indicates whether the worker may proceed to work item execution with its return value. This is necessary because multiple workers may contend for the manager role, and, if there already is a manager, others should proceed to work item execution. Unfortunately, the function also indicates that the worker may proceed to work item execution if need_to_create_worker() is false at the head of the function. need_to_create_worker() tests the following conditions. pending work items && !nr_running && !nr_idle The first and third conditions are protected by pool->lock and thus won't change while holding pool->lock; however, nr_running can change asynchronously as other workers block and resume and while it's likely to be zero, as someone woke this worker up in the first place, some other workers could have become runnable inbetween making it non-zero. If this happens, manage_worker() could return false even with zero nr_idle making the worker, the last idle one, proceed to execute work items. If then all workers of the pool end up blocking on a resource which can only be released by a work item which is pending on that pool, the whole pool can deadlock as there's no one to create more workers or summon the rescuers. This patch fixes the problem by removing the early exit condition from maybe_create_worker() and making manage_workers() return false iff there's already another manager, which ensures that the last worker doesn't start executing work items. We can leave the early exit condition alone and just ignore the return value but the only reason it was put there is because the manage_workers() used to perform both creations and destructions of workers and thus the function may be invoked while the pool is trying to reduce the number of workers. Now that manage_workers() is called only when more workers are needed, the only case this early exit condition is triggered is rare race conditions rendering it pointless. Tested with simulated workload and modified workqueue code which trigger the pool deadlock reliably without this patch. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Eric Sandeen <sandeen@sandeen.net> Link: http://lkml.kernel.org/g/54B019F4.8030009@sandeen.net Cc: Dave Chinner <david@fromorbit.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: stable@vger.kernel.org
2015-01-16 19:21:16 +00:00
maybe_create_worker(pool);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
pool->manager = NULL;
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
pool->flags &= ~POOL_MANAGER_ACTIVE;
rcuwait_wake_up(&manager_wait);
workqueue: fix subtle pool management issue which can stall whole worker_pool A worker_pool's forward progress is guaranteed by the fact that the last idle worker assumes the manager role to create more workers and summon the rescuers if creating workers doesn't succeed in timely manner before proceeding to execute work items. This manager role is implemented in manage_workers(), which indicates whether the worker may proceed to work item execution with its return value. This is necessary because multiple workers may contend for the manager role, and, if there already is a manager, others should proceed to work item execution. Unfortunately, the function also indicates that the worker may proceed to work item execution if need_to_create_worker() is false at the head of the function. need_to_create_worker() tests the following conditions. pending work items && !nr_running && !nr_idle The first and third conditions are protected by pool->lock and thus won't change while holding pool->lock; however, nr_running can change asynchronously as other workers block and resume and while it's likely to be zero, as someone woke this worker up in the first place, some other workers could have become runnable inbetween making it non-zero. If this happens, manage_worker() could return false even with zero nr_idle making the worker, the last idle one, proceed to execute work items. If then all workers of the pool end up blocking on a resource which can only be released by a work item which is pending on that pool, the whole pool can deadlock as there's no one to create more workers or summon the rescuers. This patch fixes the problem by removing the early exit condition from maybe_create_worker() and making manage_workers() return false iff there's already another manager, which ensures that the last worker doesn't start executing work items. We can leave the early exit condition alone and just ignore the return value but the only reason it was put there is because the manage_workers() used to perform both creations and destructions of workers and thus the function may be invoked while the pool is trying to reduce the number of workers. Now that manage_workers() is called only when more workers are needed, the only case this early exit condition is triggered is rare race conditions rendering it pointless. Tested with simulated workload and modified workqueue code which trigger the pool deadlock reliably without this patch. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Eric Sandeen <sandeen@sandeen.net> Link: http://lkml.kernel.org/g/54B019F4.8030009@sandeen.net Cc: Dave Chinner <david@fromorbit.com> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: stable@vger.kernel.org
2015-01-16 19:21:16 +00:00
return true;
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
}
/**
* process_one_work - process single work
* @worker: self
* @work: work to process
*
* Process @work. This function contains all the logics necessary to
* process a single work including synchronization against and
* interaction with other workers on the same cpu, queueing and
* flushing. As long as context requirement is met, any worker can
* call this function to process a work.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock) which is released and regrabbed.
*/
static void process_one_work(struct worker *worker, struct work_struct *work)
__releases(&pool->lock)
__acquires(&pool->lock)
{
struct pool_workqueue *pwq = get_work_pwq(work);
struct worker_pool *pool = worker->pool;
unsigned long work_data;
#ifdef CONFIG_LOCKDEP
/*
* It is permissible to free the struct work_struct from
* inside the function that is called from it, this we need to
* take into account for lockdep too. To avoid bogus "held
* lock freed" warnings as well as problems when looking into
* work->lockdep_map, make a copy and use that here.
*/
lockdep: fix oops in processing workqueue Under memory load, on x86_64, with lockdep enabled, the workqueue's process_one_work() has been seen to oops in __lock_acquire(), barfing on a 0xffffffff00000000 pointer in the lockdep_map's class_cache[]. Because it's permissible to free a work_struct from its callout function, the map used is an onstack copy of the map given in the work_struct: and that copy is made without any locking. Surprisingly, gcc (4.5.1 in Hugh's case) uses "rep movsl" rather than "rep movsq" for that structure copy: which might race with a workqueue user's wait_on_work() doing lock_map_acquire() on the source of the copy, putting a pointer into the class_cache[], but only in time for the top half of that pointer to be copied to the destination map. Boom when process_one_work() subsequently does lock_map_acquire() on its onstack copy of the lockdep_map. Fix this, and a similar instance in call_timer_fn(), with a lockdep_copy_map() function which additionally NULLs the class_cache[]. Note: this oops was actually seen on 3.4-next, where flush_work() newly does the racing lock_map_acquire(); but Tejun points out that 3.4 and earlier are already vulnerable to the same through wait_on_work(). * Patch orginally from Peter. Hugh modified it a bit and wrote the description. Signed-off-by: Peter Zijlstra <peterz@infradead.org> Reported-by: Hugh Dickins <hughd@google.com> LKML-Reference: <alpine.LSU.2.00.1205070951170.1544@eggly.anvils> Signed-off-by: Tejun Heo <tj@kernel.org>
2012-05-15 15:06:19 +00:00
struct lockdep_map lockdep_map;
lockdep_copy_map(&lockdep_map, &work->lockdep_map);
#endif
/* ensure we're on the correct CPU */
WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
raw_smp_processor_id() != pool->cpu);
workqueue: reimplement CPU online rebinding to handle idle workers Currently, if there are left workers when a CPU is being brough back online, the trustee kills all idle workers and scheduled rebind_work so that they re-bind to the CPU after the currently executing work is finished. This works for busy workers because concurrency management doesn't try to wake up them from scheduler callbacks, which require the target task to be on the local run queue. The busy worker bumps concurrency counter appropriately as it clears WORKER_UNBOUND from the rebind work item and it's bound to the CPU before returning to the idle state. To reduce CPU on/offlining overhead (as many embedded systems use it for powersaving) and simplify the code path, workqueue is planned to be modified to retain idle workers across CPU on/offlining. This patch reimplements CPU online rebinding such that it can also handle idle workers. As noted earlier, due to the local wakeup requirement, rebinding idle workers is tricky. All idle workers must be re-bound before scheduler callbacks are enabled. This is achieved by interlocking idle re-binding. Idle workers are requested to re-bind and then hold until all idle re-binding is complete so that no bound worker starts executing work item. Only after all idle workers are re-bound and parked, CPU_ONLINE proceeds to release them and queue rebind work item to busy workers thus guaranteeing scheduler callbacks aren't invoked until all idle workers are ready. worker_rebind_fn() is renamed to busy_worker_rebind_fn() and idle_worker_rebind() for idle workers is added. Rebinding logic is moved to rebind_workers() and now called from CPU_ONLINE after flushing trustee. While at it, add CPU sanity check in worker_thread(). Note that now a worker may become idle or the manager between trustee release and rebinding during CPU_ONLINE. As the previous patch updated create_worker() so that it can be used by regular manager while unbound and this patch implements idle re-binding, this is safe. This prepares for removal of trustee and keeping idle workers across CPU hotplugs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: "Rafael J. Wysocki" <rjw@sisk.pl>
2012-07-17 19:39:27 +00:00
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
/* claim and dequeue */
debug_work_deactivate(work);
hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
worker->current_work = work;
workqueue: consider work function when searching for busy work items To avoid executing the same work item concurrenlty, workqueue hashes currently busy workers according to their current work items and looks up the the table when it wants to execute a new work item. If there already is a worker which is executing the new work item, the new item is queued to the found worker so that it gets executed only after the current execution finishes. Unfortunately, a work item may be freed while being executed and thus recycled for different purposes. If it gets recycled for a different work item and queued while the previous execution is still in progress, workqueue may make the new work item wait for the old one although the two aren't really related in any way. In extreme cases, this false dependency may lead to deadlock although it's extremely unlikely given that there aren't too many self-freeing work item users and they usually don't wait for other work items. To alleviate the problem, record the current work function in each busy worker and match it together with the work item address in find_worker_executing_work(). While this isn't complete, it ensures that unrelated work items don't interact with each other and in the very unlikely case where a twisted wq user triggers it, it's always onto itself making the culprit easy to spot. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Andrey Isakov <andy51@gmx.ru> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=51701 Cc: stable@vger.kernel.org
2012-12-18 18:35:02 +00:00
worker->current_func = work->func;
worker->current_pwq = pwq;
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
worker->current_at = worker->task->se.sum_exec_runtime;
work_data = *work_data_bits(work);
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
worker->current_color = get_work_color(work_data);
/*
* Record wq name for cmdline and debug reporting, may get
* overridden through set_worker_desc().
*/
strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
list_del_init(&work->entry);
/*
* CPU intensive works don't participate in concurrency management.
* They're the scheduler's responsibility. This takes @worker out
* of concurrency management and the next code block will chain
* execution of the pending work items.
*/
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
worker_set_flags(worker, WORKER_CPU_INTENSIVE);
/*
* Kick @pool if necessary. It's always noop for per-cpu worker pools
* since nr_running would always be >= 1 at this point. This is used to
* chain execution of the pending work items for WORKER_NOT_RUNNING
* workers such as the UNBOUND and CPU_INTENSIVE ones.
*/
kick_pool(pool);
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
/*
* Record the last pool and clear PENDING which should be the last
* update to @work. Also, do this inside @pool->lock so that
* PENDING and queued state changes happen together while IRQ is
* disabled.
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
*/
set_work_pool_and_clear_pending(work, pool->id);
workqueue: fix data race with the pwq->stats[] increment KCSAN has discovered a data race in kernel/workqueue.c:2598: [ 1863.554079] ================================================================== [ 1863.554118] BUG: KCSAN: data-race in process_one_work / process_one_work [ 1863.554142] write to 0xffff963d99d79998 of 8 bytes by task 5394 on cpu 27: [ 1863.554154] process_one_work (kernel/workqueue.c:2598) [ 1863.554166] worker_thread (./include/linux/list.h:292 kernel/workqueue.c:2752) [ 1863.554177] kthread (kernel/kthread.c:389) [ 1863.554186] ret_from_fork (arch/x86/kernel/process.c:145) [ 1863.554197] ret_from_fork_asm (arch/x86/entry/entry_64.S:312) [ 1863.554213] read to 0xffff963d99d79998 of 8 bytes by task 5450 on cpu 12: [ 1863.554224] process_one_work (kernel/workqueue.c:2598) [ 1863.554235] worker_thread (./include/linux/list.h:292 kernel/workqueue.c:2752) [ 1863.554247] kthread (kernel/kthread.c:389) [ 1863.554255] ret_from_fork (arch/x86/kernel/process.c:145) [ 1863.554266] ret_from_fork_asm (arch/x86/entry/entry_64.S:312) [ 1863.554280] value changed: 0x0000000000001766 -> 0x000000000000176a [ 1863.554295] Reported by Kernel Concurrency Sanitizer on: [ 1863.554303] CPU: 12 PID: 5450 Comm: kworker/u64:1 Tainted: G L 6.5.0-rc6+ #44 [ 1863.554314] Hardware name: ASRock X670E PG Lightning/X670E PG Lightning, BIOS 1.21 04/26/2023 [ 1863.554322] Workqueue: btrfs-endio btrfs_end_bio_work [btrfs] [ 1863.554941] ================================================================== lockdep_invariant_state(true); → pwq->stats[PWQ_STAT_STARTED]++; trace_workqueue_execute_start(work); worker->current_func(work); Moving pwq->stats[PWQ_STAT_STARTED]++; before the line raw_spin_unlock_irq(&pool->lock); resolves the data race without performance penalty. KCSAN detected at least one additional data race: [ 157.834751] ================================================================== [ 157.834770] BUG: KCSAN: data-race in process_one_work / process_one_work [ 157.834793] write to 0xffff9934453f77a0 of 8 bytes by task 468 on cpu 29: [ 157.834804] process_one_work (/home/marvin/linux/kernel/linux_torvalds/kernel/workqueue.c:2606) [ 157.834815] worker_thread (/home/marvin/linux/kernel/linux_torvalds/./include/linux/list.h:292 /home/marvin/linux/kernel/linux_torvalds/kernel/workqueue.c:2752) [ 157.834826] kthread (/home/marvin/linux/kernel/linux_torvalds/kernel/kthread.c:389) [ 157.834834] ret_from_fork (/home/marvin/linux/kernel/linux_torvalds/arch/x86/kernel/process.c:145) [ 157.834845] ret_from_fork_asm (/home/marvin/linux/kernel/linux_torvalds/arch/x86/entry/entry_64.S:312) [ 157.834859] read to 0xffff9934453f77a0 of 8 bytes by task 214 on cpu 7: [ 157.834868] process_one_work (/home/marvin/linux/kernel/linux_torvalds/kernel/workqueue.c:2606) [ 157.834879] worker_thread (/home/marvin/linux/kernel/linux_torvalds/./include/linux/list.h:292 /home/marvin/linux/kernel/linux_torvalds/kernel/workqueue.c:2752) [ 157.834890] kthread (/home/marvin/linux/kernel/linux_torvalds/kernel/kthread.c:389) [ 157.834897] ret_from_fork (/home/marvin/linux/kernel/linux_torvalds/arch/x86/kernel/process.c:145) [ 157.834907] ret_from_fork_asm (/home/marvin/linux/kernel/linux_torvalds/arch/x86/entry/entry_64.S:312) [ 157.834920] value changed: 0x000000000000052a -> 0x0000000000000532 [ 157.834933] Reported by Kernel Concurrency Sanitizer on: [ 157.834941] CPU: 7 PID: 214 Comm: kworker/u64:2 Tainted: G L 6.5.0-rc7-kcsan-00169-g81eaf55a60fc #4 [ 157.834951] Hardware name: ASRock X670E PG Lightning/X670E PG Lightning, BIOS 1.21 04/26/2023 [ 157.834958] Workqueue: btrfs-endio btrfs_end_bio_work [btrfs] [ 157.835567] ================================================================== in code: trace_workqueue_execute_end(work, worker->current_func); → pwq->stats[PWQ_STAT_COMPLETED]++; lock_map_release(&lockdep_map); lock_map_release(&pwq->wq->lockdep_map); which needs to be resolved separately. Fixes: 725e8ec59c56c ("workqueue: Add pwq->stats[] and a monitoring script") Cc: Tejun Heo <tj@kernel.org> Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Link: https://lore.kernel.org/lkml/20230818194448.29672-1-mirsad.todorovac@alu.unizg.hr/ Signed-off-by: Mirsad Goran Todorovac <mirsad.todorovac@alu.unizg.hr> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-26 14:51:03 +00:00
pwq->stats[PWQ_STAT_STARTED]++;
raw_spin_unlock_irq(&pool->lock);
lock_map_acquire(&pwq->wq->lockdep_map);
lock_map_acquire(&lockdep_map);
/*
* Strictly speaking we should mark the invariant state without holding
* any locks, that is, before these two lock_map_acquire()'s.
*
* However, that would result in:
*
* A(W1)
* WFC(C)
* A(W1)
* C(C)
*
* Which would create W1->C->W1 dependencies, even though there is no
* actual deadlock possible. There are two solutions, using a
* read-recursive acquire on the work(queue) 'locks', but this will then
* hit the lockdep limitation on recursive locks, or simply discard
* these locks.
*
* AFAICT there is no possible deadlock scenario between the
* flush_work() and complete() primitives (except for single-threaded
* workqueues), so hiding them isn't a problem.
*/
lockdep_invariant_state(true);
trace_workqueue_execute_start(work);
workqueue: consider work function when searching for busy work items To avoid executing the same work item concurrenlty, workqueue hashes currently busy workers according to their current work items and looks up the the table when it wants to execute a new work item. If there already is a worker which is executing the new work item, the new item is queued to the found worker so that it gets executed only after the current execution finishes. Unfortunately, a work item may be freed while being executed and thus recycled for different purposes. If it gets recycled for a different work item and queued while the previous execution is still in progress, workqueue may make the new work item wait for the old one although the two aren't really related in any way. In extreme cases, this false dependency may lead to deadlock although it's extremely unlikely given that there aren't too many self-freeing work item users and they usually don't wait for other work items. To alleviate the problem, record the current work function in each busy worker and match it together with the work item address in find_worker_executing_work(). While this isn't complete, it ensures that unrelated work items don't interact with each other and in the very unlikely case where a twisted wq user triggers it, it's always onto itself making the culprit easy to spot. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Andrey Isakov <andy51@gmx.ru> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=51701 Cc: stable@vger.kernel.org
2012-12-18 18:35:02 +00:00
worker->current_func(work);
/*
* While we must be careful to not use "work" after this, the trace
* point will only record its address.
*/
trace_workqueue_execute_end(work, worker->current_func);
pwq->stats[PWQ_STAT_COMPLETED]++;
lock_map_release(&lockdep_map);
lock_map_release(&pwq->wq->lockdep_map);
if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2019-03-25 19:32:28 +00:00
" last function: %ps\n",
workqueue: consider work function when searching for busy work items To avoid executing the same work item concurrenlty, workqueue hashes currently busy workers according to their current work items and looks up the the table when it wants to execute a new work item. If there already is a worker which is executing the new work item, the new item is queued to the found worker so that it gets executed only after the current execution finishes. Unfortunately, a work item may be freed while being executed and thus recycled for different purposes. If it gets recycled for a different work item and queued while the previous execution is still in progress, workqueue may make the new work item wait for the old one although the two aren't really related in any way. In extreme cases, this false dependency may lead to deadlock although it's extremely unlikely given that there aren't too many self-freeing work item users and they usually don't wait for other work items. To alleviate the problem, record the current work function in each busy worker and match it together with the work item address in find_worker_executing_work(). While this isn't complete, it ensures that unrelated work items don't interact with each other and in the very unlikely case where a twisted wq user triggers it, it's always onto itself making the culprit easy to spot. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Andrey Isakov <andy51@gmx.ru> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=51701 Cc: stable@vger.kernel.org
2012-12-18 18:35:02 +00:00
current->comm, preempt_count(), task_pid_nr(current),
worker->current_func);
debug_show_held_locks(current);
dump_stack();
}
/*
* The following prevents a kworker from hogging CPU on !PREEMPTION
* kernels, where a requeueing work item waiting for something to
* happen could deadlock with stop_machine as such work item could
* indefinitely requeue itself while all other CPUs are trapped in
* stop_machine. At the same time, report a quiescent RCU state so
* the same condition doesn't freeze RCU.
*/
cond_resched();
raw_spin_lock_irq(&pool->lock);
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE If a per-cpu work item hogs the CPU, it can prevent other work items from starting through concurrency management. A per-cpu workqueue which intends to host such CPU-hogging work items can choose to not participate in concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be error-prone and difficult to debug when missed. This patch adds an automatic CPU usage based detection. If a concurrency-managed work item consumes more CPU time than the threshold (10ms by default) continuously without intervening sleeps, wq_worker_tick() which is called from scheduler_tick() will detect the condition and automatically mark it CPU_INTENSIVE. The mechanism isn't foolproof: * Detection depends on tick hitting the work item. Getting preempted at the right timings may allow a violating work item to evade detection at least temporarily. * nohz_full CPUs may not be running ticks and thus can fail detection. * Even when detection is working, the 10ms detection delays can add up if many CPU-hogging work items are queued at the same time. However, in vast majority of cases, this should be able to detect violations reliably and provide reasonable protection with a small increase in code complexity. If some work items trigger this condition repeatedly, the bigger problem likely is the CPU being saturated with such per-cpu work items and the solution would be making them UNBOUND. The next patch will add a debug mechanism to help spot such cases. v4: Documentation for workqueue.cpu_intensive_thresh_us added to kernel-parameters.txt. v3: Switch to use wq_worker_tick() instead of hooking into preemptions as suggested by Peter. v2: Lai pointed out that wq_worker_stopping() also needs to be called from preemption and rtlock paths and an earlier patch was updated accordingly. This patch adds a comment describing the risk of infinte recursions and how they're avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com>
2023-05-18 03:02:08 +00:00
/*
* In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
* CPU intensive by wq_worker_tick() if @work hogged CPU longer than
* wq_cpu_intensive_thresh_us. Clear it.
*/
worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
psi: fix aggregation idle shut-off psi has provisions to shut off the periodic aggregation worker when there is a period of no task activity - and thus no data that needs aggregating. However, while developing psi monitoring, Suren noticed that the aggregation clock currently won't stay shut off for good. Debugging this revealed a flaw in the idle design: an aggregation run will see no task activity and decide to go to sleep; shortly thereafter, the kworker thread that executed the aggregation will go idle and cause a scheduling change, during which the psi callback will kick the !pending worker again. This will ping-pong forever, and is equivalent to having no shut-off logic at all (but with more code!) Fix this by exempting aggregation workers from psi's clock waking logic when the state change is them going to sleep. To do this, tag workers with the last work function they executed, and if in psi we see a worker going to sleep after aggregating psi data, we will not reschedule the aggregation work item. What if the worker is also executing other items before or after? Any psi state times that were incurred by work items preceding the aggregation work will have been collected from the per-cpu buckets during the aggregation itself. If there are work items following the aggregation work, the worker's last_func tag will be overwritten and the aggregator will be kept alive to process this genuine new activity. If the aggregation work is the last thing the worker does, and we decide to go idle, the brief period of non-idle time incurred between the aggregation run and the kworker's dequeue will be stranded in the per-cpu buckets until the clock is woken by later activity. But that should not be a problem. The buckets can hold 4s worth of time, and future activity will wake the clock with a 2s delay, giving us 2s worth of data we can leave behind when disabling aggregation. If it takes a worker more than two seconds to go idle after it finishes its last work item, we likely have bigger problems in the system, and won't notice one sample that was averaged with a bogus per-CPU weight. Link: http://lkml.kernel.org/r/20190116193501.1910-1-hannes@cmpxchg.org Fixes: eb414681d5a0 ("psi: pressure stall information for CPU, memory, and IO") Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Reported-by: Suren Baghdasaryan <surenb@google.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-02-01 22:20:42 +00:00
/* tag the worker for identification in schedule() */
worker->last_func = worker->current_func;
/* we're done with it, release */
hash_del(&worker->hentry);
worker->current_work = NULL;
workqueue: consider work function when searching for busy work items To avoid executing the same work item concurrenlty, workqueue hashes currently busy workers according to their current work items and looks up the the table when it wants to execute a new work item. If there already is a worker which is executing the new work item, the new item is queued to the found worker so that it gets executed only after the current execution finishes. Unfortunately, a work item may be freed while being executed and thus recycled for different purposes. If it gets recycled for a different work item and queued while the previous execution is still in progress, workqueue may make the new work item wait for the old one although the two aren't really related in any way. In extreme cases, this false dependency may lead to deadlock although it's extremely unlikely given that there aren't too many self-freeing work item users and they usually don't wait for other work items. To alleviate the problem, record the current work function in each busy worker and match it together with the work item address in find_worker_executing_work(). While this isn't complete, it ensures that unrelated work items don't interact with each other and in the very unlikely case where a twisted wq user triggers it, it's always onto itself making the culprit easy to spot. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Andrey Isakov <andy51@gmx.ru> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=51701 Cc: stable@vger.kernel.org
2012-12-18 18:35:02 +00:00
worker->current_func = NULL;
worker->current_pwq = NULL;
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
worker->current_color = INT_MAX;
pwq_dec_nr_in_flight(pwq, work_data);
}
/**
* process_scheduled_works - process scheduled works
* @worker: self
*
* Process all scheduled works. Please note that the scheduled list
* may change while processing a work, so this function repeatedly
* fetches a work from the top and executes it.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock) which may be released and regrabbed
* multiple times.
*/
static void process_scheduled_works(struct worker *worker)
{
struct work_struct *work;
bool first = true;
while ((work = list_first_entry_or_null(&worker->scheduled,
struct work_struct, entry))) {
if (first) {
worker->pool->watchdog_ts = jiffies;
first = false;
}
process_one_work(worker, work);
}
}
static void set_pf_worker(bool val)
{
mutex_lock(&wq_pool_attach_mutex);
if (val)
current->flags |= PF_WQ_WORKER;
else
current->flags &= ~PF_WQ_WORKER;
mutex_unlock(&wq_pool_attach_mutex);
}
/**
* worker_thread - the worker thread function
* @__worker: self
*
* The worker thread function. All workers belong to a worker_pool -
* either a per-cpu one or dynamic unbound one. These workers process all
* work items regardless of their specific target workqueue. The only
* exception is work items which belong to workqueues with a rescuer which
* will be explained in rescuer_thread().
*
* Return: 0
*/
static int worker_thread(void *__worker)
{
struct worker *worker = __worker;
struct worker_pool *pool = worker->pool;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* tell the scheduler that this is a workqueue worker */
set_pf_worker(true);
woke_up:
raw_spin_lock_irq(&pool->lock);
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
/* am I supposed to die? */
if (unlikely(worker->flags & WORKER_DIE)) {
raw_spin_unlock_irq(&pool->lock);
set_pf_worker(false);
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
set_task_comm(worker->task, "kworker/dying");
ida_free(&pool->worker_ida, worker->id);
worker_detach_from_pool(worker);
WARN_ON_ONCE(!list_empty(&worker->entry));
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
kfree(worker);
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
return 0;
}
worker_leave_idle(worker);
recheck:
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* no more worker necessary? */
if (!need_more_worker(pool))
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
goto sleep;
/* do we need to manage? */
if (unlikely(!may_start_working(pool)) && manage_workers(worker))
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
goto recheck;
/*
* ->scheduled list can only be filled while a worker is
* preparing to process a work or actually processing it.
* Make sure nobody diddled with it while I was sleeping.
*/
WARN_ON_ONCE(!list_empty(&worker->scheduled));
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
* Finish PREP stage. We're guaranteed to have at least one idle
* worker or that someone else has already assumed the manager
* role. This is where @worker starts participating in concurrency
* management if applicable and concurrency management is restored
* after being rebound. See rebind_workers() for details.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
do {
struct work_struct *work =
list_first_entry(&pool->worklist,
struct work_struct, entry);
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (assign_work(work, worker, NULL))
process_scheduled_works(worker);
} while (keep_working(pool));
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
worker_set_flags(worker, WORKER_PREP);
sleep:
/*
* pool->lock is held and there's no work to process and no need to
* manage, sleep. Workers are woken up only while holding
* pool->lock or from local cpu, so setting the current state
* before releasing pool->lock is enough to prevent losing any
* event.
*/
worker_enter_idle(worker);
__set_current_state(TASK_IDLE);
raw_spin_unlock_irq(&pool->lock);
schedule();
goto woke_up;
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/**
* rescuer_thread - the rescuer thread function
* @__rescuer: self
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*
* Workqueue rescuer thread function. There's one rescuer for each
* workqueue which has WQ_MEM_RECLAIM set.
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*
* Regular work processing on a pool may block trying to create a new
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* worker which uses GFP_KERNEL allocation which has slight chance of
* developing into deadlock if some works currently on the same queue
* need to be processed to satisfy the GFP_KERNEL allocation. This is
* the problem rescuer solves.
*
* When such condition is possible, the pool summons rescuers of all
* workqueues which have works queued on the pool and let them process
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* those works so that forward progress can be guaranteed.
*
* This should happen rarely.
*
* Return: 0
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
*/
static int rescuer_thread(void *__rescuer)
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
{
struct worker *rescuer = __rescuer;
struct workqueue_struct *wq = rescuer->rescue_wq;
bool should_stop;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
set_user_nice(current, RESCUER_NICE_LEVEL);
/*
* Mark rescuer as worker too. As WORKER_PREP is never cleared, it
* doesn't participate in concurrency management.
*/
set_pf_worker(true);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
repeat:
set_current_state(TASK_IDLE);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
* By the time the rescuer is requested to stop, the workqueue
* shouldn't have any work pending, but @wq->maydays may still have
* pwq(s) queued. This can happen by non-rescuer workers consuming
* all the work items before the rescuer got to them. Go through
* @wq->maydays processing before acting on should_stop so that the
* list is always empty on exit.
*/
should_stop = kthread_should_stop();
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/* see whether any pwq is asking for help */
raw_spin_lock_irq(&wq_mayday_lock);
while (!list_empty(&wq->maydays)) {
struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
struct pool_workqueue, mayday_node);
struct worker_pool *pool = pwq->pool;
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
struct work_struct *work, *n;
__set_current_state(TASK_RUNNING);
list_del_init(&pwq->mayday_node);
raw_spin_unlock_irq(&wq_mayday_lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
worker_attach_to_pool(rescuer, pool);
raw_spin_lock_irq(&pool->lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
/*
* Slurp in all works issued via this workqueue and
* process'em.
*/
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
list_for_each_entry_safe(work, n, &pool->worklist, entry) {
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (get_work_pwq(work) == pwq &&
assign_work(work, rescuer, &n))
pwq->stats[PWQ_STAT_RESCUED]++;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
}
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
workqueue: Factor out work to worker assignment and collision handling The two work execution paths in worker_thread() and rescuer_thread() use move_linked_works() to claim work items from @pool->worklist. Once claimed, process_schedule_works() is called which invokes process_one_work() on each work item. process_one_work() then uses find_worker_executing_work() to detect and handle collisions - situations where the work item to be executed is still running on another worker. This works fine, but, to improve work execution locality, we want to establish work to worker association earlier and know for sure that the worker is going to excute the work once asssigned, which requires performing collision handling earlier while trying to assign the work item to the worker. This patch introduces assign_work() which assigns a work item to a worker using move_linked_works() and then performs collision handling. As collision handling is handled earlier, process_one_work() no longer needs to worry about them. After the this patch, collision checks for linked work items are skipped, which should be fine as they can't be queued multiple times concurrently. For work items running from rescuers, the timing of collision handling may change but the invariant that the work items go through collision handling before starting execution does not. This patch shouldn't cause noticeable behavior changes, especially given that worker_thread() behavior remains the same. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (!list_empty(&rescuer->scheduled)) {
workqueue: allow rescuer thread to do more work. When there is serious memory pressure, all workers in a pool could be blocked, and a new thread cannot be created because it requires memory allocation. In this situation a WQ_MEM_RECLAIM workqueue will wake up the rescuer thread to do some work. The rescuer will only handle requests that are already on ->worklist. If max_requests is 1, that means it will handle a single request. The rescuer will be woken again in 100ms to handle another max_requests requests. I've seen a machine (running a 3.0 based "enterprise" kernel) with thousands of requests queued for xfslogd, which has a max_requests of 1, and is needed for retiring all 'xfs' write requests. When one of the worker pools gets into this state, it progresses extremely slowly and possibly never recovers (only waited an hour or two). With this patch we leave a pool_workqueue on mayday list until it is clearly no longer in need of assistance. This allows all requests to be handled in a timely fashion. We keep each pool_workqueue on the mayday list until need_to_create_worker() is false, and no work for this workqueue is found in the pool. I have tested this in combination with a (hackish) patch which forces all work items to be handled by the rescuer thread. In that context it significantly improves performance. A similar patch for a 3.0 kernel significantly improved performance on a heavy work load. Thanks to Jan Kara for some design ideas, and to Dongsu Park for some comments and testing. tj: Inverted the lock order between wq_mayday_lock and pool->lock with a preceding patch and simplified this patch. Added comment and updated changelog accordingly. Dongsu spotted missing get_pwq() in the simplified code. Cc: Dongsu Park <dongsu.park@profitbricks.com> Cc: Jan Kara <jack@suse.cz> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: NeilBrown <neilb@suse.de> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-12-08 17:39:16 +00:00
process_scheduled_works(rescuer);
/*
* The above execution of rescued work items could
* have created more to rescue through
* pwq_activate_first_inactive() or chained
workqueue: allow rescuer thread to do more work. When there is serious memory pressure, all workers in a pool could be blocked, and a new thread cannot be created because it requires memory allocation. In this situation a WQ_MEM_RECLAIM workqueue will wake up the rescuer thread to do some work. The rescuer will only handle requests that are already on ->worklist. If max_requests is 1, that means it will handle a single request. The rescuer will be woken again in 100ms to handle another max_requests requests. I've seen a machine (running a 3.0 based "enterprise" kernel) with thousands of requests queued for xfslogd, which has a max_requests of 1, and is needed for retiring all 'xfs' write requests. When one of the worker pools gets into this state, it progresses extremely slowly and possibly never recovers (only waited an hour or two). With this patch we leave a pool_workqueue on mayday list until it is clearly no longer in need of assistance. This allows all requests to be handled in a timely fashion. We keep each pool_workqueue on the mayday list until need_to_create_worker() is false, and no work for this workqueue is found in the pool. I have tested this in combination with a (hackish) patch which forces all work items to be handled by the rescuer thread. In that context it significantly improves performance. A similar patch for a 3.0 kernel significantly improved performance on a heavy work load. Thanks to Jan Kara for some design ideas, and to Dongsu Park for some comments and testing. tj: Inverted the lock order between wq_mayday_lock and pool->lock with a preceding patch and simplified this patch. Added comment and updated changelog accordingly. Dongsu spotted missing get_pwq() in the simplified code. Cc: Dongsu Park <dongsu.park@profitbricks.com> Cc: Jan Kara <jack@suse.cz> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: NeilBrown <neilb@suse.de> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-12-08 17:39:16 +00:00
* queueing. Let's put @pwq back on mayday list so
* that such back-to-back work items, which may be
* being used to relieve memory pressure, don't
* incur MAYDAY_INTERVAL delay inbetween.
*/
if (pwq->nr_active && need_to_create_worker(pool)) {
raw_spin_lock(&wq_mayday_lock);
/*
* Queue iff we aren't racing destruction
* and somebody else hasn't queued it already.
*/
if (wq->rescuer && list_empty(&pwq->mayday_node)) {
get_pwq(pwq);
list_add_tail(&pwq->mayday_node, &wq->maydays);
}
raw_spin_unlock(&wq_mayday_lock);
workqueue: allow rescuer thread to do more work. When there is serious memory pressure, all workers in a pool could be blocked, and a new thread cannot be created because it requires memory allocation. In this situation a WQ_MEM_RECLAIM workqueue will wake up the rescuer thread to do some work. The rescuer will only handle requests that are already on ->worklist. If max_requests is 1, that means it will handle a single request. The rescuer will be woken again in 100ms to handle another max_requests requests. I've seen a machine (running a 3.0 based "enterprise" kernel) with thousands of requests queued for xfslogd, which has a max_requests of 1, and is needed for retiring all 'xfs' write requests. When one of the worker pools gets into this state, it progresses extremely slowly and possibly never recovers (only waited an hour or two). With this patch we leave a pool_workqueue on mayday list until it is clearly no longer in need of assistance. This allows all requests to be handled in a timely fashion. We keep each pool_workqueue on the mayday list until need_to_create_worker() is false, and no work for this workqueue is found in the pool. I have tested this in combination with a (hackish) patch which forces all work items to be handled by the rescuer thread. In that context it significantly improves performance. A similar patch for a 3.0 kernel significantly improved performance on a heavy work load. Thanks to Jan Kara for some design ideas, and to Dongsu Park for some comments and testing. tj: Inverted the lock order between wq_mayday_lock and pool->lock with a preceding patch and simplified this patch. Added comment and updated changelog accordingly. Dongsu spotted missing get_pwq() in the simplified code. Cc: Dongsu Park <dongsu.park@profitbricks.com> Cc: Jan Kara <jack@suse.cz> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: NeilBrown <neilb@suse.de> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-12-08 17:39:16 +00:00
}
}
/*
* Put the reference grabbed by send_mayday(). @pool won't
* go away while we're still attached to it.
*/
put_pwq(pwq);
/*
* Leave this pool. Notify regular workers; otherwise, we end up
* with 0 concurrency and stalling the execution.
*/
kick_pool(pool);
raw_spin_unlock_irq(&pool->lock);
worker_detach_from_pool(rescuer);
raw_spin_lock_irq(&wq_mayday_lock);
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
}
raw_spin_unlock_irq(&wq_mayday_lock);
if (should_stop) {
__set_current_state(TASK_RUNNING);
set_pf_worker(false);
return 0;
}
/* rescuers should never participate in concurrency management */
WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
schedule();
goto repeat;
}
/**
* check_flush_dependency - check for flush dependency sanity
* @target_wq: workqueue being flushed
* @target_work: work item being flushed (NULL for workqueue flushes)
*
* %current is trying to flush the whole @target_wq or @target_work on it.
* If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
* reclaiming memory or running on a workqueue which doesn't have
* %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
* a deadlock.
*/
static void check_flush_dependency(struct workqueue_struct *target_wq,
struct work_struct *target_work)
{
work_func_t target_func = target_work ? target_work->func : NULL;
struct worker *worker;
if (target_wq->flags & WQ_MEM_RECLAIM)
return;
worker = current_wq_worker();
WARN_ONCE(current->flags & PF_MEMALLOC,
2019-03-25 19:32:28 +00:00
"workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
current->pid, current->comm, target_wq->name, target_func);
workqueue: skip flush dependency checks for legacy workqueues fca839c00a12 ("workqueue: warn if memory reclaim tries to flush !WQ_MEM_RECLAIM workqueue") implemented flush dependency warning which triggers if a PF_MEMALLOC task or WQ_MEM_RECLAIM workqueue tries to flush a !WQ_MEM_RECLAIM workquee. This assumes that workqueues marked with WQ_MEM_RECLAIM sit in memory reclaim path and making it depend on something which may need more memory to make forward progress can lead to deadlocks. Unfortunately, workqueues created with the legacy create*_workqueue() interface always have WQ_MEM_RECLAIM regardless of whether they are depended upon memory reclaim or not. These spurious WQ_MEM_RECLAIM markings cause spurious triggering of the flush dependency checks. WARNING: CPU: 0 PID: 6 at kernel/workqueue.c:2361 check_flush_dependency+0x138/0x144() workqueue: WQ_MEM_RECLAIM deferwq:deferred_probe_work_func is flushing !WQ_MEM_RECLAIM events:lru_add_drain_per_cpu ... Workqueue: deferwq deferred_probe_work_func [<c0017acc>] (unwind_backtrace) from [<c0013134>] (show_stack+0x10/0x14) [<c0013134>] (show_stack) from [<c0245f18>] (dump_stack+0x94/0xd4) [<c0245f18>] (dump_stack) from [<c0026f9c>] (warn_slowpath_common+0x80/0xb0) [<c0026f9c>] (warn_slowpath_common) from [<c0026ffc>] (warn_slowpath_fmt+0x30/0x40) [<c0026ffc>] (warn_slowpath_fmt) from [<c00390b8>] (check_flush_dependency+0x138/0x144) [<c00390b8>] (check_flush_dependency) from [<c0039ca0>] (flush_work+0x50/0x15c) [<c0039ca0>] (flush_work) from [<c00c51b0>] (lru_add_drain_all+0x130/0x180) [<c00c51b0>] (lru_add_drain_all) from [<c00f728c>] (migrate_prep+0x8/0x10) [<c00f728c>] (migrate_prep) from [<c00bfbc4>] (alloc_contig_range+0xd8/0x338) [<c00bfbc4>] (alloc_contig_range) from [<c00f8f18>] (cma_alloc+0xe0/0x1ac) [<c00f8f18>] (cma_alloc) from [<c001cac4>] (__alloc_from_contiguous+0x38/0xd8) [<c001cac4>] (__alloc_from_contiguous) from [<c001ceb4>] (__dma_alloc+0x240/0x278) [<c001ceb4>] (__dma_alloc) from [<c001cf78>] (arm_dma_alloc+0x54/0x5c) [<c001cf78>] (arm_dma_alloc) from [<c0355ea4>] (dmam_alloc_coherent+0xc0/0xec) [<c0355ea4>] (dmam_alloc_coherent) from [<c039cc4c>] (ahci_port_start+0x150/0x1dc) [<c039cc4c>] (ahci_port_start) from [<c0384734>] (ata_host_start.part.3+0xc8/0x1c8) [<c0384734>] (ata_host_start.part.3) from [<c03898dc>] (ata_host_activate+0x50/0x148) [<c03898dc>] (ata_host_activate) from [<c039d558>] (ahci_host_activate+0x44/0x114) [<c039d558>] (ahci_host_activate) from [<c039f05c>] (ahci_platform_init_host+0x1d8/0x3c8) [<c039f05c>] (ahci_platform_init_host) from [<c039e6bc>] (tegra_ahci_probe+0x448/0x4e8) [<c039e6bc>] (tegra_ahci_probe) from [<c0347058>] (platform_drv_probe+0x50/0xac) [<c0347058>] (platform_drv_probe) from [<c03458cc>] (driver_probe_device+0x214/0x2c0) [<c03458cc>] (driver_probe_device) from [<c0343cc0>] (bus_for_each_drv+0x60/0x94) [<c0343cc0>] (bus_for_each_drv) from [<c03455d8>] (__device_attach+0xb0/0x114) [<c03455d8>] (__device_attach) from [<c0344ab8>] (bus_probe_device+0x84/0x8c) [<c0344ab8>] (bus_probe_device) from [<c0344f48>] (deferred_probe_work_func+0x68/0x98) [<c0344f48>] (deferred_probe_work_func) from [<c003b738>] (process_one_work+0x120/0x3f8) [<c003b738>] (process_one_work) from [<c003ba48>] (worker_thread+0x38/0x55c) [<c003ba48>] (worker_thread) from [<c0040f14>] (kthread+0xdc/0xf4) [<c0040f14>] (kthread) from [<c000f778>] (ret_from_fork+0x14/0x3c) Fix it by marking workqueues created via create*_workqueue() with __WQ_LEGACY and disabling flush dependency checks on them. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-tested-by: Thierry Reding <thierry.reding@gmail.com> Link: http://lkml.kernel.org/g/20160126173843.GA11115@ulmo.nvidia.com Fixes: fca839c00a12 ("workqueue: warn if memory reclaim tries to flush !WQ_MEM_RECLAIM workqueue")
2016-01-29 10:59:46 +00:00
WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
(WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2019-03-25 19:32:28 +00:00
"workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
worker->current_pwq->wq->name, worker->current_func,
target_wq->name, target_func);
}
struct wq_barrier {
struct work_struct work;
struct completion done;
struct task_struct *task; /* purely informational */
};
static void wq_barrier_func(struct work_struct *work)
{
struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
complete(&barr->done);
}
/**
* insert_wq_barrier - insert a barrier work
* @pwq: pwq to insert barrier into
* @barr: wq_barrier to insert
* @target: target work to attach @barr to
* @worker: worker currently executing @target, NULL if @target is not executing
*
* @barr is linked to @target such that @barr is completed only after
* @target finishes execution. Please note that the ordering
* guarantee is observed only with respect to @target and on the local
* cpu.
*
* Currently, a queued barrier can't be canceled. This is because
* try_to_grab_pending() can't determine whether the work to be
* grabbed is at the head of the queue and thus can't clear LINKED
* flag of the previous work while there must be a valid next work
* after a work with LINKED flag set.
*
* Note that when @worker is non-NULL, @target may be modified
* underneath us, so we can't reliably determine pwq from @target.
*
* CONTEXT:
* raw_spin_lock_irq(pool->lock).
*/
static void insert_wq_barrier(struct pool_workqueue *pwq,
struct wq_barrier *barr,
struct work_struct *target, struct worker *worker)
{
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
unsigned int work_flags = 0;
unsigned int work_color;
struct list_head *head;
/*
* debugobject calls are safe here even with pool->lock locked
* as we know for sure that this will not trigger any of the
* checks and call back into the fixup functions where we
* might deadlock.
*/
INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
locking/lockdep: Explicitly initialize wq_barrier::done::map With the new lockdep crossrelease feature, which checks completions usage, a false positive is reported in the workqueue code: > Worker A : acquired of wfc.work -> wait for cpu_hotplug_lock to be released > Task B : acquired of cpu_hotplug_lock -> wait for lock#3 to be released > Task C : acquired of lock#3 -> wait for completion of barr->done > (Task C is in lru_add_drain_all_cpuslocked()) > Worker D : wait for wfc.work to be released -> will complete barr->done Such a dead lock can not happen because Task C's barr->done and Worker D's barr->done can not be the same instance. The reason of this false positive is we initialize all wq_barrier::done at insert_wq_barrier() via init_completion(), which makes them belong to the same lock class, therefore, impossible circles are reported. To fix this, explicitly initialize the lockdep map for wq_barrier::done in insert_wq_barrier(), so that the lock class key of wq_barrier::done is a subkey of the corresponding work_struct, as a result we won't build a dependency between a wq_barrier with a unrelated work, and we can differ wq barriers based on the related works, so the false positive above is avoided. Also define the empty lockdep_init_map_crosslock() for !CROSSRELEASE to make the code simple and away from unnecessary #ifdefs. Reported-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Boqun Feng <boqun.feng@gmail.com> Cc: Byungchul Park <byungchul.park@lge.com> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20170817094622.12915-1-boqun.feng@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-17 09:46:12 +00:00
init_completion_map(&barr->done, &target->lockdep_map);
barr->task = current;
workqueue: Mark barrier work with WORK_STRUCT_INACTIVE Currently, WORK_NO_COLOR has two meanings: Not participate in flushing Not participate in nr_active And only non-barrier work items are marked with WORK_STRUCT_INACTIVE when they are in inactive_works list. The barrier work items are not marked INACTIVE even linked in inactive_works list since these tail items are always moved together with the head work item. These definitions are simple, clean and practical. (Except a small blemish that only the first meaning of WORK_NO_COLOR is documented in include/linux/workqueue.h while both meanings are in workqueue.c) But dual-purpose WORK_NO_COLOR used for barrier work items has proven to be problematical[1]. Only the second purpose is obligatory. So we plan to make barrier work items participate in flushing but keep them still not participating in nr_active. So the plan is to mark barrier work items inactive without using WORK_NO_COLOR in this patch so that we can assign a flushing color to them in next patch. The reasonable way is to add or reuse a bit in work data of the work item. But adding a bit will double the size of pool_workqueue. Currently, WORK_STRUCT_INACTIVE is only used in try_to_grab_pending() for user-queued work items and try_to_grab_pending() can't work for barrier work items. So we extend WORK_STRUCT_INACTIVE to also mark barrier work items no matter which list they are in because we don't need to determind which list a barrier work item is in. So the meaning of WORK_STRUCT_INACTIVE becomes just "the work items don't participate in nr_active" (no matter whether it is a barrier work item or a user-queued work item). And WORK_STRUCT_INACTIVE for user-queued work items means they are in inactive_works list. This patch does it by setting WORK_STRUCT_INACTIVE for barrier work items in insert_wq_barrier() and checking WORK_STRUCT_INACTIVE first in pwq_dec_nr_in_flight(). And the meaning of WORK_NO_COLOR is reduced to only "not participating in flushing". There is no functionality change intended in this patch. Because WORK_NO_COLOR+WORK_STRUCT_INACTIVE represents the previous WORK_NO_COLOR in meaning and try_to_grab_pending() doesn't use for barrier work items and avoids being confused by this extended WORK_STRUCT_INACTIVE. A bunch of comment for nr_active & WORK_STRUCT_INACTIVE is also added for documenting how WORK_STRUCT_INACTIVE works in nr_active management. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:37 +00:00
/* The barrier work item does not participate in pwq->nr_active. */
work_flags |= WORK_STRUCT_INACTIVE;
/*
* If @target is currently being executed, schedule the
* barrier to the worker; otherwise, put it after @target.
*/
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
if (worker) {
head = worker->scheduled.next;
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
work_color = worker->current_color;
} else {
unsigned long *bits = work_data_bits(target);
head = target->entry.next;
/* there can already be other linked works, inherit and set */
work_flags |= *bits & WORK_STRUCT_LINKED;
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
work_color = get_work_color(*bits);
__set_bit(WORK_STRUCT_LINKED_BIT, bits);
}
workqueue: Assign a color to barrier work items There was no strong reason to or not to flush barrier work items in flush_workqueue(). And we have to make barrier work items not participate in nr_active so we had been using WORK_NO_COLOR for them which also makes them can't be flushed by flush_workqueue(). And the users of flush_workqueue() often do not intend to wait barrier work items issued by flush_work(). That made the choice sound perfect. But barrier work items have reference to internal structure (pool_workqueue) and the worker thread[s] is/are still busy for the workqueue user when the barrrier work items are not done. So it is reasonable to make flush_workqueue() also watch for flush_work() to make it more robust. And a problem[1] reported by Li Zhe shows that we need such robustness. The warning logs are listed below: WARNING: CPU: 0 PID: 19336 at kernel/workqueue.c:4430 destroy_workqueue+0x11a/0x2f0 ***** destroy_workqueue: test_workqueue9 has the following busy pwq pwq 4: cpus=2 node=0 flags=0x0 nice=0 active=0/1 refcnt=2 in-flight: 5658:wq_barrier_func Showing busy workqueues and worker pools: ***** It shows that even after drain_workqueue() returns, the barrier work item is still in flight and the pwq (and a worker) is still busy on it. The problem is caused by flush_workqueue() not watching flush_work(): Thread A Worker /* normal work item with linked */ process_scheduled_works() destroy_workqueue() process_one_work() drain_workqueue() /* run normal work item */ /-- pwq_dec_nr_in_flight() flush_workqueue() <---/ /* the last normal work item is done */ sanity_check process_one_work() /-- raw_spin_unlock_irq(&pool->lock) raw_spin_lock_irq(&pool->lock) <-/ /* maybe preempt */ *WARNING* wq_barrier_func() /* maybe preempt by cond_resched() */ Thread A can get the pool lock after the Worker unlocks the pool lock before running wq_barrier_func(). And if there is any preemption happen around wq_barrier_func(), destroy_workqueue()'s sanity check is more likely to get the lock and catch it. (Note: preemption is not necessary to cause the bug, the unlocking is enough to possibly trigger the WARNING.) A simple solution might be just executing all linked barrier work items once without releasing pool lock after the head work item's pwq_dec_nr_in_flight(). But this solution has two problems: 1) the head work item might also be barrier work item when the user-queued work item is cancelled. For example: thread 1: thread 2: queue_work(wq, &my_work) flush_work(&my_work) cancel_work_sync(&my_work); /* Neiter my_work nor the barrier work is scheduled. */ destroy_workqueue(wq); /* This is an easier way to catch the WARNING. */ 2) there might be too much linked barrier work items and running them all once without releasing pool lock just causes trouble. The only solution is to make flush_workqueue() aslo watch barrier work items. So we have to assign a color to these barrier work items which is the color of the head (user-queued) work item. Assigning a color doesn't cause any problem in ative management, because the prvious patch made barrier work items not participate in nr_active via WORK_STRUCT_INACTIVE rather than reliance on the (old) WORK_NO_COLOR. [1]: https://lore.kernel.org/lkml/20210812083814.32453-1-lizhe.67@bytedance.com/ Reported-by: Li Zhe <lizhe.67@bytedance.com> Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-08-17 01:32:38 +00:00
pwq->nr_in_flight[work_color]++;
work_flags |= work_color_to_flags(work_color);
insert_work(pwq, &barr->work, head, work_flags);
}
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/**
* flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
* @wq: workqueue being flushed
* @flush_color: new flush color, < 0 for no-op
* @work_color: new work color, < 0 for no-op
*
* Prepare pwqs for workqueue flushing.
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*
* If @flush_color is non-negative, flush_color on all pwqs should be
* -1. If no pwq has in-flight commands at the specified color, all
* pwq->flush_color's stay at -1 and %false is returned. If any pwq
* has in flight commands, its pwq->flush_color is set to
* @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
* wakeup logic is armed and %true is returned.
*
* The caller should have initialized @wq->first_flusher prior to
* calling this function with non-negative @flush_color. If
* @flush_color is negative, no flush color update is done and %false
* is returned.
*
* If @work_color is non-negative, all pwqs should have the same
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
* work_color which is previous to @work_color and all will be
* advanced to @work_color.
*
* CONTEXT:
* mutex_lock(wq->mutex).
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*
* Return:
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
* %true if @flush_color >= 0 and there's something to flush. %false
* otherwise.
*/
static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
int flush_color, int work_color)
{
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
bool wait = false;
struct pool_workqueue *pwq;
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
if (flush_color >= 0) {
WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
atomic_set(&wq->nr_pwqs_to_flush, 1);
}
for_each_pwq(pwq, wq) {
struct worker_pool *pool = pwq->pool;
raw_spin_lock_irq(&pool->lock);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
if (flush_color >= 0) {
WARN_ON_ONCE(pwq->flush_color != -1);
if (pwq->nr_in_flight[flush_color]) {
pwq->flush_color = flush_color;
atomic_inc(&wq->nr_pwqs_to_flush);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
wait = true;
}
}
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
if (work_color >= 0) {
WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
pwq->work_color = work_color;
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
}
raw_spin_unlock_irq(&pool->lock);
}
if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
complete(&wq->first_flusher->done);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
return wait;
}
/**
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
* __flush_workqueue - ensure that any scheduled work has run to completion.
* @wq: workqueue to flush
*
* This function sleeps until all work items which were queued on entry
* have finished execution, but it is not livelocked by new incoming ones.
*/
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
void __flush_workqueue(struct workqueue_struct *wq)
{
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
struct wq_flusher this_flusher = {
.list = LIST_HEAD_INIT(this_flusher.list),
.flush_color = -1,
.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
};
int next_color;
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
if (WARN_ON(!wq_online))
return;
lock_map_acquire(&wq->lockdep_map);
lock_map_release(&wq->lockdep_map);
mutex_lock(&wq->mutex);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/*
* Start-to-wait phase
*/
next_color = work_next_color(wq->work_color);
if (next_color != wq->flush_color) {
/*
* Color space is not full. The current work_color
* becomes our flush_color and work_color is advanced
* by one.
*/
WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
this_flusher.flush_color = wq->work_color;
wq->work_color = next_color;
if (!wq->first_flusher) {
/* no flush in progress, become the first flusher */
WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
wq->first_flusher = &this_flusher;
if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
wq->work_color)) {
/* nothing to flush, done */
wq->flush_color = next_color;
wq->first_flusher = NULL;
goto out_unlock;
}
} else {
/* wait in queue */
WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
list_add_tail(&this_flusher.list, &wq->flusher_queue);
flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
}
} else {
/*
* Oops, color space is full, wait on overflow queue.
* The next flush completion will assign us
* flush_color and transfer to flusher_queue.
*/
list_add_tail(&this_flusher.list, &wq->flusher_overflow);
}
check_flush_dependency(wq, NULL);
mutex_unlock(&wq->mutex);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
wait_for_completion(&this_flusher.done);
/*
* Wake-up-and-cascade phase
*
* First flushers are responsible for cascading flushes and
* handling overflow. Non-first flushers can simply return.
*/
if (READ_ONCE(wq->first_flusher) != &this_flusher)
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
return;
mutex_lock(&wq->mutex);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/* we might have raced, check again with mutex held */
if (wq->first_flusher != &this_flusher)
goto out_unlock;
WRITE_ONCE(wq->first_flusher, NULL);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
WARN_ON_ONCE(!list_empty(&this_flusher.list));
WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
while (true) {
struct wq_flusher *next, *tmp;
/* complete all the flushers sharing the current flush color */
list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
if (next->flush_color != wq->flush_color)
break;
list_del_init(&next->list);
complete(&next->done);
}
WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
wq->flush_color != work_next_color(wq->work_color));
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
/* this flush_color is finished, advance by one */
wq->flush_color = work_next_color(wq->flush_color);
/* one color has been freed, handle overflow queue */
if (!list_empty(&wq->flusher_overflow)) {
/*
* Assign the same color to all overflowed
* flushers, advance work_color and append to
* flusher_queue. This is the start-to-wait
* phase for these overflowed flushers.
*/
list_for_each_entry(tmp, &wq->flusher_overflow, list)
tmp->flush_color = wq->work_color;
wq->work_color = work_next_color(wq->work_color);
list_splice_tail_init(&wq->flusher_overflow,
&wq->flusher_queue);
flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
}
if (list_empty(&wq->flusher_queue)) {
WARN_ON_ONCE(wq->flush_color != wq->work_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
break;
}
/*
* Need to flush more colors. Make the next flusher
* the new first flusher and arm pwqs.
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
*/
WARN_ON_ONCE(wq->flush_color == wq->work_color);
WARN_ON_ONCE(wq->flush_color != next->flush_color);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
list_del_init(&next->list);
wq->first_flusher = next;
if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
break;
/*
* Meh... this color is already done, clear first
* flusher and repeat cascading.
*/
wq->first_flusher = NULL;
}
out_unlock:
mutex_unlock(&wq->mutex);
}
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
EXPORT_SYMBOL(__flush_workqueue);
/**
* drain_workqueue - drain a workqueue
* @wq: workqueue to drain
*
* Wait until the workqueue becomes empty. While draining is in progress,
* only chain queueing is allowed. IOW, only currently pending or running
* work items on @wq can queue further work items on it. @wq is flushed
* repeatedly until it becomes empty. The number of flushing is determined
* by the depth of chaining and should be relatively short. Whine if it
* takes too long.
*/
void drain_workqueue(struct workqueue_struct *wq)
{
unsigned int flush_cnt = 0;
struct pool_workqueue *pwq;
/*
* __queue_work() needs to test whether there are drainers, is much
* hotter than drain_workqueue() and already looks at @wq->flags.
* Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
*/
mutex_lock(&wq->mutex);
if (!wq->nr_drainers++)
wq->flags |= __WQ_DRAINING;
mutex_unlock(&wq->mutex);
reflush:
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
__flush_workqueue(wq);
mutex_lock(&wq->mutex);
for_each_pwq(pwq, wq) {
bool drained;
raw_spin_lock_irq(&pwq->pool->lock);
drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
raw_spin_unlock_irq(&pwq->pool->lock);
if (drained)
continue;
if (++flush_cnt == 10 ||
(flush_cnt % 100 == 0 && flush_cnt <= 1000))
pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
wq->name, __func__, flush_cnt);
mutex_unlock(&wq->mutex);
goto reflush;
}
if (!--wq->nr_drainers)
wq->flags &= ~__WQ_DRAINING;
mutex_unlock(&wq->mutex);
}
EXPORT_SYMBOL_GPL(drain_workqueue);
static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
bool from_cancel)
{
struct worker *worker = NULL;
struct worker_pool *pool;
struct pool_workqueue *pwq;
might_sleep();
rcu_read_lock();
pool = get_work_pool(work);
if (!pool) {
rcu_read_unlock();
return false;
}
raw_spin_lock_irq(&pool->lock);
workqueue: simplify is-work-item-queued-here test Currently, determining whether a work item is queued on a locked pool involves somewhat convoluted memory barrier dancing. It goes like the following. * When a work item is queued on a pool, work->data is updated before work->entry is linked to the pending list with a wmb() inbetween. * When trying to determine whether a work item is currently queued on a pool pointed to by work->data, it locks the pool and looks at work->entry. If work->entry is linked, we then do rmb() and then check whether work->data points to the current pool. This works because, work->data can only point to a pool if it currently is or were on the pool and, * If it currently is on the pool, the tests would obviously succeed. * It it left the pool, its work->entry was cleared under pool->lock, so if we're seeing non-empty work->entry, it has to be from the work item being linked on another pool. Because work->data is updated before work->entry is linked with wmb() inbetween, work->data update from another pool is guaranteed to be visible if we do rmb() after seeing non-empty work->entry. So, we either see empty work->entry or we see updated work->data pointin to another pool. While this works, it's convoluted, to put it mildly. With recent updates, it's now guaranteed that work->data points to cwq only while the work item is queued and that updating work->data to point to cwq or back to pool is done under pool->lock, so we can simply test whether work->data points to cwq which is associated with the currently locked pool instead of the convoluted memory barrier dancing. This patch replaces the memory barrier based "are you still here, really?" test with much simpler "does work->data points to me?" test - if work->data points to a cwq which is associated with the currently locked pool, the work item is guaranteed to be queued on the pool as work->data can start and stop pointing to such cwq only under pool->lock and the start and stop coincide with queue and dequeue. tj: Rewrote the comments and description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
/* see the comment in try_to_grab_pending() with the same code */
pwq = get_work_pwq(work);
if (pwq) {
if (unlikely(pwq->pool != pool))
goto already_gone;
} else {
worker = find_worker_executing_work(pool, work);
if (!worker)
goto already_gone;
pwq = worker->current_pwq;
}
check_flush_dependency(pwq->wq, work);
insert_wq_barrier(pwq, barr, work, worker);
raw_spin_unlock_irq(&pool->lock);
/*
* Force a lock recursion deadlock when using flush_work() inside a
* single-threaded or rescuer equipped workqueue.
*
* For single threaded workqueues the deadlock happens when the work
* is after the work issuing the flush_work(). For rescuer equipped
* workqueues the deadlock happens when the rescuer stalls, blocking
* forward progress.
*/
if (!from_cancel &&
(pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
lock_map_acquire(&pwq->wq->lockdep_map);
lock_map_release(&pwq->wq->lockdep_map);
}
rcu_read_unlock();
return true;
already_gone:
raw_spin_unlock_irq(&pool->lock);
rcu_read_unlock();
return false;
}
static bool __flush_work(struct work_struct *work, bool from_cancel)
{
struct wq_barrier barr;
if (WARN_ON(!wq_online))
return false;
if (WARN_ON(!work->func))
return false;
workqueue: don't skip lockdep work dependency in cancel_work_sync() Like Hillf Danton mentioned syzbot should have been able to catch cancel_work_sync() in work context by checking lockdep_map in __flush_work() for both flush and cancel. in [1], being unable to report an obvious deadlock scenario shown below is broken. From locking dependency perspective, sync version of cancel request should behave as if flush request, for it waits for completion of work if that work has already started execution. ---------- #include <linux/module.h> #include <linux/sched.h> static DEFINE_MUTEX(mutex); static void work_fn(struct work_struct *work) { schedule_timeout_uninterruptible(HZ / 5); mutex_lock(&mutex); mutex_unlock(&mutex); } static DECLARE_WORK(work, work_fn); static int __init test_init(void) { schedule_work(&work); schedule_timeout_uninterruptible(HZ / 10); mutex_lock(&mutex); cancel_work_sync(&work); mutex_unlock(&mutex); return -EINVAL; } module_init(test_init); MODULE_LICENSE("GPL"); ---------- The check this patch restores was added by commit 0976dfc1d0cd80a4 ("workqueue: Catch more locking problems with flush_work()"). Then, lockdep's crossrelease feature was added by commit b09be676e0ff25bd ("locking/lockdep: Implement the 'crossrelease' feature"). As a result, this check was once removed by commit fd1a5b04dfb899f8 ("workqueue: Remove now redundant lock acquisitions wrt. workqueue flushes"). But lockdep's crossrelease feature was removed by commit e966eaeeb623f099 ("locking/lockdep: Remove the cross-release locking checks"). At this point, this check should have been restored. Then, commit d6e89786bed977f3 ("workqueue: skip lockdep wq dependency in cancel_work_sync()") introduced a boolean flag in order to distinguish flush_work() and cancel_work_sync(), for checking "struct workqueue_struct" dependency when called from cancel_work_sync() was causing false positives. Then, commit 87915adc3f0acdf0 ("workqueue: re-add lockdep dependencies for flushing") tried to restore "struct work_struct" dependency check, but by error checked this boolean flag. Like an example shown above indicates, "struct work_struct" dependency needs to be checked for both flush_work() and cancel_work_sync(). Link: https://lkml.kernel.org/r/20220504044800.4966-1-hdanton@sina.com [1] Reported-by: Hillf Danton <hdanton@sina.com> Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Fixes: 87915adc3f0acdf0 ("workqueue: re-add lockdep dependencies for flushing") Cc: Johannes Berg <johannes.berg@intel.com> Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-07-29 04:30:23 +00:00
lock_map_acquire(&work->lockdep_map);
lock_map_release(&work->lockdep_map);
if (start_flush_work(work, &barr, from_cancel)) {
wait_for_completion(&barr.done);
destroy_work_on_stack(&barr.work);
return true;
} else {
return false;
}
}
/**
* flush_work - wait for a work to finish executing the last queueing instance
* @work: the work to flush
*
* Wait until @work has finished execution. @work is guaranteed to be idle
* on return if it hasn't been requeued since flush started.
*
* Return:
* %true if flush_work() waited for the work to finish execution,
* %false if it was already idle.
*/
bool flush_work(struct work_struct *work)
{
return __flush_work(work, false);
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
}
EXPORT_SYMBOL_GPL(flush_work);
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
struct cwt_wait {
wait_queue_entry_t wait;
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
struct work_struct *work;
};
static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
{
struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
if (cwait->work != key)
return 0;
return autoremove_wake_function(wait, mode, sync, key);
}
static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
{
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
unsigned long flags;
int ret;
do {
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
ret = try_to_grab_pending(work, is_dwork, &flags);
/*
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
* If someone else is already canceling, wait for it to
* finish. flush_work() doesn't work for PREEMPT_NONE
* because we may get scheduled between @work's completion
* and the other canceling task resuming and clearing
* CANCELING - flush_work() will return false immediately
* as @work is no longer busy, try_to_grab_pending() will
* return -ENOENT as @work is still being canceled and the
* other canceling task won't be able to clear CANCELING as
* we're hogging the CPU.
*
* Let's wait for completion using a waitqueue. As this
* may lead to the thundering herd problem, use a custom
* wake function which matches @work along with exclusive
* wait and wakeup.
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
*/
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
if (unlikely(ret == -ENOENT)) {
struct cwt_wait cwait;
init_wait(&cwait.wait);
cwait.wait.func = cwt_wakefn;
cwait.work = work;
prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
TASK_UNINTERRUPTIBLE);
if (work_is_canceling(work))
schedule();
finish_wait(&cancel_waitq, &cwait.wait);
}
} while (unlikely(ret < 0));
workqueue: mark a work item being canceled as such There can be two reasons try_to_grab_pending() can fail with -EAGAIN. One is when someone else is queueing or deqeueing the work item. With the previous patches, it is guaranteed that PENDING and queued state will soon agree making it safe to busy-retry in this case. The other is if multiple __cancel_work_timer() invocations are racing one another. __cancel_work_timer() grabs PENDING and then waits for running instances of the target work item on all CPUs while holding PENDING and !queued. try_to_grab_pending() invoked from another task will keep returning -EAGAIN while the current owner is waiting. Not distinguishing the two cases is okay because __cancel_work_timer() is the only user of try_to_grab_pending() and it invokes wait_on_work() whenever grabbing fails. For the first case, busy looping should be fine but wait_on_work() doesn't cause any critical problem. For the latter case, the new contender usually waits for the same condition as the current owner, so no unnecessarily extended busy-looping happens. Combined, these make __cancel_work_timer() technically correct even without irq protection while grabbing PENDING or distinguishing the two different cases. While the current code is technically correct, not distinguishing the two cases makes it difficult to use try_to_grab_pending() for other purposes than canceling because it's impossible to tell whether it's safe to busy-retry grabbing. This patch adds a mechanism to mark a work item being canceled. try_to_grab_pending() now disables irq on success and returns -EAGAIN to indicate that grabbing failed but PENDING and queued states are gonna agree soon and it's safe to busy-loop. It returns -ENOENT if the work item is being canceled and it may stay PENDING && !queued for arbitrary amount of time. __cancel_work_timer() is modified to mark the work canceling with WORK_OFFQ_CANCELING after grabbing PENDING, thus making try_to_grab_pending() fail with -ENOENT instead of -EAGAIN. Also, it invokes wait_on_work() iff grabbing failed with -ENOENT. This isn't necessary for correctness but makes it consistent with other future users of try_to_grab_pending(). v2: try_to_grab_pending() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Updated so that try_to_grab_pending() disables irq on success rather than requiring preemption disabled by the caller. This makes busy-looping easier and will allow try_to_grap_pending() to be used from bh/irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:46 +00:00
/* tell other tasks trying to grab @work to back off */
mark_work_canceling(work);
local_irq_restore(flags);
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/*
* This allows canceling during early boot. We know that @work
* isn't executing.
*/
if (wq_online)
__flush_work(work, true);
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
clear_work_data(work);
workqueue: fix hang involving racing cancel[_delayed]_work_sync()'s for PREEMPT_NONE cancel[_delayed]_work_sync() are implemented using __cancel_work_timer() which grabs the PENDING bit using try_to_grab_pending() and then flushes the work item with PENDING set to prevent the on-going execution of the work item from requeueing itself. try_to_grab_pending() can always grab PENDING bit without blocking except when someone else is doing the above flushing during cancelation. In that case, try_to_grab_pending() returns -ENOENT. In this case, __cancel_work_timer() currently invokes flush_work(). The assumption is that the completion of the work item is what the other canceling task would be waiting for too and thus waiting for the same condition and retrying should allow forward progress without excessive busy looping Unfortunately, this doesn't work if preemption is disabled or the latter task has real time priority. Let's say task A just got woken up from flush_work() by the completion of the target work item. If, before task A starts executing, task B gets scheduled and invokes __cancel_work_timer() on the same work item, its try_to_grab_pending() will return -ENOENT as the work item is still being canceled by task A and flush_work() will also immediately return false as the work item is no longer executing. This puts task B in a busy loop possibly preventing task A from executing and clearing the canceling state on the work item leading to a hang. task A task B worker executing work __cancel_work_timer() try_to_grab_pending() set work CANCELING flush_work() block for work completion completion, wakes up A __cancel_work_timer() while (forever) { try_to_grab_pending() -ENOENT as work is being canceled flush_work() false as work is no longer executing } This patch removes the possible hang by updating __cancel_work_timer() to explicitly wait for clearing of CANCELING rather than invoking flush_work() after try_to_grab_pending() fails with -ENOENT. Link: http://lkml.kernel.org/g/20150206171156.GA8942@axis.com v3: bit_waitqueue() can't be used for work items defined in vmalloc area. Switched to custom wake function which matches the target work item and exclusive wait and wakeup. v2: v1 used wake_up() on bit_waitqueue() which leads to NULL deref if the target bit waitqueue has wait_bit_queue's on it. Use DEFINE_WAIT_BIT() and __wake_up_bit() instead. Reported by Tomeu Vizoso. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Rabin Vincent <rabin.vincent@axis.com> Cc: Tomeu Vizoso <tomeu.vizoso@gmail.com> Cc: stable@vger.kernel.org Tested-by: Jesper Nilsson <jesper.nilsson@axis.com> Tested-by: Rabin Vincent <rabin.vincent@axis.com>
2015-03-05 13:04:13 +00:00
/*
* Paired with prepare_to_wait() above so that either
* waitqueue_active() is visible here or !work_is_canceling() is
* visible there.
*/
smp_mb();
if (waitqueue_active(&cancel_waitq))
__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
return ret;
}
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
/**
* cancel_work_sync - cancel a work and wait for it to finish
* @work: the work to cancel
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
*
* Cancel @work and wait for its execution to finish. This function
* can be used even if the work re-queues itself or migrates to
* another workqueue. On return from this function, @work is
* guaranteed to be not pending or executing on any CPU.
*
* cancel_work_sync(&delayed_work->work) must not be used for
* delayed_work's. Use cancel_delayed_work_sync() instead.
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
*
* The caller must ensure that the workqueue on which @work was last
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
* queued can't be destroyed before this function returns.
*
* Return:
* %true if @work was pending, %false otherwise.
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
*/
bool cancel_work_sync(struct work_struct *work)
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
{
return __cancel_work_timer(work, false);
implement flush_work() A basic problem with flush_scheduled_work() is that it blocks behind _all_ presently-queued works, rather than just the work whcih the caller wants to flush. If the caller holds some lock, and if one of the queued work happens to want that lock as well then accidental deadlocks can occur. One example of this is the phy layer: it wants to flush work while holding rtnl_lock(). But if a linkwatch event happens to be queued, the phy code will deadlock because the linkwatch callback function takes rtnl_lock. So we implement a new function which will flush a *single* work - just the one which the caller wants to free up. Thus we avoid the accidental deadlocks which can arise from unrelated subsystems' callbacks taking shared locks. flush_work() non-blockingly dequeues the work_struct which we want to kill, then it waits for its handler to complete on all CPUs. Add ->current_work to the "struct cpu_workqueue_struct", it points to currently running "struct work_struct". When flush_work(work) detects ->current_work == work, it inserts a barrier at the _head_ of ->worklist (and thus right _after_ that work) and waits for completition. This means that the next work fired on that CPU will be this barrier, or another barrier queued by concurrent flush_work(), so the caller of flush_work() will be woken before any "regular" work has a chance to run. When wait_on_work() unlocks workqueue_mutex (or whatever we choose to protect against CPU hotplug), CPU may go away. But in that case take_over_work() will move a barrier we queued to another CPU, it will be fired sometime, and wait_on_work() will be woken. Actually, we are doing cleanup_workqueue_thread()->kthread_stop() before take_over_work(), so cwq->thread should complete its ->worklist (and thus the barrier), because currently we don't check kthread_should_stop() in run_workqueue(). But even if we did, everything should be ok. [akpm@osdl.org: cleanup] [akpm@osdl.org: add flush_work_keventd() wrapper] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:33:52 +00:00
}
EXPORT_SYMBOL_GPL(cancel_work_sync);
implement flush_work() A basic problem with flush_scheduled_work() is that it blocks behind _all_ presently-queued works, rather than just the work whcih the caller wants to flush. If the caller holds some lock, and if one of the queued work happens to want that lock as well then accidental deadlocks can occur. One example of this is the phy layer: it wants to flush work while holding rtnl_lock(). But if a linkwatch event happens to be queued, the phy code will deadlock because the linkwatch callback function takes rtnl_lock. So we implement a new function which will flush a *single* work - just the one which the caller wants to free up. Thus we avoid the accidental deadlocks which can arise from unrelated subsystems' callbacks taking shared locks. flush_work() non-blockingly dequeues the work_struct which we want to kill, then it waits for its handler to complete on all CPUs. Add ->current_work to the "struct cpu_workqueue_struct", it points to currently running "struct work_struct". When flush_work(work) detects ->current_work == work, it inserts a barrier at the _head_ of ->worklist (and thus right _after_ that work) and waits for completition. This means that the next work fired on that CPU will be this barrier, or another barrier queued by concurrent flush_work(), so the caller of flush_work() will be woken before any "regular" work has a chance to run. When wait_on_work() unlocks workqueue_mutex (or whatever we choose to protect against CPU hotplug), CPU may go away. But in that case take_over_work() will move a barrier we queued to another CPU, it will be fired sometime, and wait_on_work() will be woken. Actually, we are doing cleanup_workqueue_thread()->kthread_stop() before take_over_work(), so cwq->thread should complete its ->worklist (and thus the barrier), because currently we don't check kthread_should_stop() in run_workqueue(). But even if we did, everything should be ok. [akpm@osdl.org: cleanup] [akpm@osdl.org: add flush_work_keventd() wrapper] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:33:52 +00:00
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
/**
* flush_delayed_work - wait for a dwork to finish executing the last queueing
* @dwork: the delayed work to flush
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
*
* Delayed timer is cancelled and the pending work is queued for
* immediate execution. Like flush_work(), this function only
* considers the last queueing instance of @dwork.
*
* Return:
* %true if flush_work() waited for the work to finish execution,
* %false if it was already idle.
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
*/
bool flush_delayed_work(struct delayed_work *dwork)
{
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
local_irq_disable();
if (del_timer_sync(&dwork->timer))
workqueue: add delayed_work->wq to simplify reentrancy handling To avoid executing the same work item from multiple CPUs concurrently, a work_struct records the last pool it was on in its ->data so that, on the next queueing, the pool can be queried to determine whether the work item is still executing or not. A delayed_work goes through timer before actually being queued on the target workqueue and the timer needs to know the target workqueue and CPU. This is currently achieved by modifying delayed_work->work.data such that it points to the cwq which points to the target workqueue and the last CPU the work item was on. __queue_delayed_work() extracts the last CPU from delayed_work->work.data and then combines it with the target workqueue to create new work.data. The only thing this rather ugly hack achieves is encoding the target workqueue into delayed_work->work.data without using a separate field, which could be a trade off one can make; unfortunately, this entangles work->data management between regular workqueue and delayed_work code by setting cwq pointer before the work item is actually queued and becomes a hindrance for further improvements of work->data handling. This can be easily made sane by adding a target workqueue field to delayed_work. While delayed_work is used widely in the kernel and this does make it a bit larger (<5%), I think this is the right trade-off especially given the prospect of much saner handling of work->data which currently involves quite tricky memory barrier dancing, and don't expect to see any measureable effect. Add delayed_work->wq and drop the delayed_work->work.data overloading. tj: Rewrote the description. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2013-02-07 02:04:53 +00:00
__queue_work(dwork->cpu, dwork->wq, &dwork->work);
workqueue: disable irq while manipulating PENDING Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access to the target work item. They first try to claim the bit and proceed with queueing only after that succeeds and there's a window between PENDING being set and the actual queueing where the task can be interrupted or preempted. There's also a similar window in process_one_work() when clearing PENDING. A work item is dequeued, gcwq->lock is released and then PENDING is cleared and the worker might get interrupted or preempted between releasing gcwq->lock and clearing PENDING. cancel[_delayed]_work_sync() tries to claim or steal PENDING. The function assumes that a work item with PENDING is either queued or in the process of being [de]queued. In the latter case, it busy-loops until either the work item loses PENDING or is queued. If canceling coincides with the above described interrupts or preemptions, the canceling task will busy-loop while the queueing or executing task is preempted. This patch keeps irq disabled across claiming PENDING and actual queueing and moves PENDING clearing in process_one_work() inside gcwq->lock so that busy looping from PENDING && !queued doesn't wait for interrupted/preempted tasks. Note that, in process_one_work(), setting last CPU and clearing PENDING got merged into single operation. This removes possible long busy-loops and will allow using try_to_grab_pending() from bh and irq contexts. v2: __queue_work() was testing preempt_count() to ensure that the caller has disabled preemption. This triggers spuriously if !CONFIG_PREEMPT_COUNT. Use preemptible() instead. Reported by Fengguang Wu. v3: Disable irq instead of preemption. IRQ will be disabled while grabbing gcwq->lock later anyway and this allows using try_to_grab_pending() from bh and irq contexts. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Fengguang Wu <fengguang.wu@intel.com>
2012-08-03 17:30:45 +00:00
local_irq_enable();
return flush_work(&dwork->work);
}
EXPORT_SYMBOL(flush_delayed_work);
/**
* flush_rcu_work - wait for a rwork to finish executing the last queueing
* @rwork: the rcu work to flush
*
* Return:
* %true if flush_rcu_work() waited for the work to finish execution,
* %false if it was already idle.
*/
bool flush_rcu_work(struct rcu_work *rwork)
{
if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
rcu_barrier();
flush_work(&rwork->work);
return true;
} else {
return flush_work(&rwork->work);
}
}
EXPORT_SYMBOL(flush_rcu_work);
static bool __cancel_work(struct work_struct *work, bool is_dwork)
{
unsigned long flags;
int ret;
do {
ret = try_to_grab_pending(work, is_dwork, &flags);
} while (unlikely(ret == -EAGAIN));
if (unlikely(ret < 0))
return false;
set_work_pool_and_clear_pending(work, get_work_pool_id(work));
local_irq_restore(flags);
return ret;
}
/*
* See cancel_delayed_work()
*/
bool cancel_work(struct work_struct *work)
{
return __cancel_work(work, false);
}
EXPORT_SYMBOL(cancel_work);
/**
* cancel_delayed_work - cancel a delayed work
* @dwork: delayed_work to cancel
*
* Kill off a pending delayed_work.
*
* Return: %true if @dwork was pending and canceled; %false if it wasn't
* pending.
*
* Note:
* The work callback function may still be running on return, unless
* it returns %true and the work doesn't re-arm itself. Explicitly flush or
* use cancel_delayed_work_sync() to wait on it.
*
* This function is safe to call from any context including IRQ handler.
*/
bool cancel_delayed_work(struct delayed_work *dwork)
{
return __cancel_work(&dwork->work, true);
}
EXPORT_SYMBOL(cancel_delayed_work);
/**
* cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
* @dwork: the delayed work cancel
*
* This is cancel_work_sync() for delayed works.
*
* Return:
* %true if @dwork was pending, %false otherwise.
*/
bool cancel_delayed_work_sync(struct delayed_work *dwork)
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
{
return __cancel_work_timer(&dwork->work, true);
make cancel_rearming_delayed_work() reliable Thanks to Jarek Poplawski for the ideas and for spotting the bug in the initial draft patch. cancel_rearming_delayed_work() currently has many limitations, because it requires that dwork always re-arms itself via queue_delayed_work(). So it hangs forever if dwork doesn't do this, or cancel_rearming_delayed_work/ cancel_delayed_work was already called. It uses flush_workqueue() in a loop, so it can't be used if workqueue was freezed, and it is potentially live- lockable on busy system if delay is small. With this patch cancel_rearming_delayed_work() doesn't make any assumptions about dwork, it can re-arm itself via queue_delayed_work(), or queue_work(), or do nothing. As a "side effect", cancel_work_sync() was changed to handle re-arming works as well. Disadvantages: - this patch adds wmb() to insert_work(). - slowdowns the fast path (when del_timer() succeeds on entry) of cancel_rearming_delayed_work(), because wait_on_work() is called unconditionally. In that case, compared to the old version, we are doing "unneeded" lock/unlock for each online CPU. On the other hand, this means we don't need to use cancel_work_sync() after cancel_rearming_delayed_work(). - complicates the code (.text grows by 130 bytes). [akpm@linux-foundation.org: fix speling] Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru> Cc: David Chinner <dgc@sgi.com> Cc: David Howells <dhowells@redhat.com> Cc: Gautham Shenoy <ego@in.ibm.com> Acked-by: Jarek Poplawski <jarkao2@o2.pl> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-09 09:34:46 +00:00
}
EXPORT_SYMBOL(cancel_delayed_work_sync);
/**
* schedule_on_each_cpu - execute a function synchronously on each online CPU
* @func: the function to call
*
* schedule_on_each_cpu() executes @func on each online CPU using the
* system workqueue and blocks until all CPUs have completed.
* schedule_on_each_cpu() is very slow.
*
* Return:
* 0 on success, -errno on failure.
*/
2006-11-22 14:55:48 +00:00
int schedule_on_each_cpu(work_func_t func)
{
int cpu;
struct work_struct __percpu *works;
works = alloc_percpu(struct work_struct);
if (!works)
return -ENOMEM;
cpus_read_lock();
for_each_online_cpu(cpu) {
struct work_struct *work = per_cpu_ptr(works, cpu);
INIT_WORK(work, func);
schedule_work_on(cpu, work);
}
for_each_online_cpu(cpu)
flush_work(per_cpu_ptr(works, cpu));
cpus_read_unlock();
free_percpu(works);
return 0;
}
/**
* execute_in_process_context - reliably execute the routine with user context
* @fn: the function to execute
* @ew: guaranteed storage for the execute work structure (must
* be available when the work executes)
*
* Executes the function immediately if process context is available,
* otherwise schedules the function for delayed execution.
*
* Return: 0 - function was executed
* 1 - function was scheduled for execution
*/
2006-11-22 14:55:48 +00:00
int execute_in_process_context(work_func_t fn, struct execute_work *ew)
{
if (!in_interrupt()) {
2006-11-22 14:55:48 +00:00
fn(&ew->work);
return 0;
}
2006-11-22 14:55:48 +00:00
INIT_WORK(&ew->work, fn);
schedule_work(&ew->work);
return 1;
}
EXPORT_SYMBOL_GPL(execute_in_process_context);
/**
* free_workqueue_attrs - free a workqueue_attrs
* @attrs: workqueue_attrs to free
*
* Undo alloc_workqueue_attrs().
*/
void free_workqueue_attrs(struct workqueue_attrs *attrs)
{
if (attrs) {
free_cpumask_var(attrs->cpumask);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
free_cpumask_var(attrs->__pod_cpumask);
kfree(attrs);
}
}
/**
* alloc_workqueue_attrs - allocate a workqueue_attrs
*
* Allocate a new workqueue_attrs, initialize with default settings and
* return it.
*
* Return: The allocated new workqueue_attr on success. %NULL on failure.
*/
struct workqueue_attrs *alloc_workqueue_attrs(void)
{
struct workqueue_attrs *attrs;
attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
if (!attrs)
goto fail;
if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
goto fail;
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
goto fail;
cpumask_copy(attrs->cpumask, cpu_possible_mask);
attrs->affn_scope = WQ_AFFN_DFL;
return attrs;
fail:
free_workqueue_attrs(attrs);
return NULL;
}
static void copy_workqueue_attrs(struct workqueue_attrs *to,
const struct workqueue_attrs *from)
{
to->nice = from->nice;
cpumask_copy(to->cpumask, from->cpumask);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
to->affn_strict = from->affn_strict;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
/*
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
* Unlike hash and equality test, copying shouldn't ignore wq-only
* fields as copying is used for both pool and wq attrs. Instead,
* get_unbound_pool() explicitly clears the fields.
*/
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
to->affn_scope = from->affn_scope;
to->ordered = from->ordered;
}
/*
* Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
* comments in 'struct workqueue_attrs' definition.
*/
static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
{
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
attrs->affn_scope = WQ_AFFN_NR_TYPES;
attrs->ordered = false;
}
/* hash value of the content of @attr */
static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
{
u32 hash = 0;
hash = jhash_1word(attrs->nice, hash);
hash = jhash(cpumask_bits(attrs->cpumask),
BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
hash = jhash(cpumask_bits(attrs->__pod_cpumask),
BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
hash = jhash_1word(attrs->affn_strict, hash);
return hash;
}
/* content equality test */
static bool wqattrs_equal(const struct workqueue_attrs *a,
const struct workqueue_attrs *b)
{
if (a->nice != b->nice)
return false;
if (!cpumask_equal(a->cpumask, b->cpumask))
return false;
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
return false;
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
if (a->affn_strict != b->affn_strict)
return false;
return true;
}
workqueue: Factor out actual cpumask calculation to reduce subtlety in wq_update_pod() For an unbound pool, multiple cpumasks are involved. U: The user-specified cpumask (may be filtered with cpu_possible_mask). A: The actual cpumask filtered by wq_unbound_cpumask. If the filtering leaves no CPU, wq_unbound_cpumask is used. P: Per-pod subsets of #A. wq->attrs stores #U, wq->dfl_pwq->pool->attrs->cpumask #A, and wq->cpu_pwq[CPU]->pool->attrs->cpumask #P. wq_update_pod() is called to update per-pod pwq's during CPU hotplug. To calculate the new #P for each workqueue, it needs to call wq_calc_pod_cpumask() with @attrs that contains #A. Currently, wq_update_pod() achieves this by calling wq_calc_pod_cpumask() with wq->dfl_pwq->pool->attrs. This is rather fragile because we're calling wq_calc_pod_cpumask() with @attrs of a worker_pool rather than the workqueue's actual attrs when what we want to calculate is the workqueue's cpumask on the pod. While this works fine currently, future changes will add fields which are used differently between workqueues and worker_pools and this subtlety will bite us. This patch factors out #U -> #A calculation from apply_wqattrs_prepare() into wqattrs_actualize_cpumask and updates wq_update_pod() to copy wq->unbound_attrs and use the new helper to obtain #A freshly instead of abusing wq->dfl_pwq->pool_attrs. This shouldn't cause any behavior changes in the current code. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: K Prateek Nayak <kprateek.nayak@amd.com> Reference: http://lkml.kernel.org/r/30625cdd-4d61-594b-8db9-6816b017dde3@amd.com
2023-08-08 01:57:24 +00:00
/* Update @attrs with actually available CPUs */
static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
const cpumask_t *unbound_cpumask)
{
/*
* Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
* @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
* @unbound_cpumask.
*/
cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
if (unlikely(cpumask_empty(attrs->cpumask)))
cpumask_copy(attrs->cpumask, unbound_cpumask);
}
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
/* find wq_pod_type to use for @attrs */
static const struct wq_pod_type *
wqattrs_pod_type(const struct workqueue_attrs *attrs)
{
enum wq_affn_scope scope;
struct wq_pod_type *pt;
/* to synchronize access to wq_affn_dfl */
lockdep_assert_held(&wq_pool_mutex);
if (attrs->affn_scope == WQ_AFFN_DFL)
scope = wq_affn_dfl;
else
scope = attrs->affn_scope;
pt = &wq_pod_types[scope];
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
likely(pt->nr_pods))
return pt;
/*
* Before workqueue_init_topology(), only SYSTEM is available which is
* initialized in workqueue_init_early().
*/
pt = &wq_pod_types[WQ_AFFN_SYSTEM];
BUG_ON(!pt->nr_pods);
return pt;
}
/**
* init_worker_pool - initialize a newly zalloc'd worker_pool
* @pool: worker_pool to initialize
*
* Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
*
* Return: 0 on success, -errno on failure. Even on failure, all fields
* inside @pool proper are initialized and put_unbound_pool() can be called
* on @pool safely to release it.
*/
static int init_worker_pool(struct worker_pool *pool)
{
raw_spin_lock_init(&pool->lock);
pool->id = -1;
pool->cpu = -1;
pool->node = NUMA_NO_NODE;
pool->flags |= POOL_DISASSOCIATED;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
pool->watchdog_ts = jiffies;
INIT_LIST_HEAD(&pool->worklist);
INIT_LIST_HEAD(&pool->idle_list);
hash_init(pool->busy_hash);
timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
INIT_LIST_HEAD(&pool->workers);
INIT_LIST_HEAD(&pool->dying_workers);
ida_init(&pool->worker_ida);
INIT_HLIST_NODE(&pool->hash_node);
pool->refcnt = 1;
/* shouldn't fail above this point */
pool->attrs = alloc_workqueue_attrs();
if (!pool->attrs)
return -ENOMEM;
wqattrs_clear_for_pool(pool->attrs);
return 0;
}
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
#ifdef CONFIG_LOCKDEP
static void wq_init_lockdep(struct workqueue_struct *wq)
{
char *lock_name;
lockdep_register_key(&wq->key);
lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
if (!lock_name)
lock_name = wq->name;
workqueue, lockdep: Fix a memory leak in wq->lock_name The following commit: 669de8bda87b ("kernel/workqueue: Use dynamic lockdep keys for workqueues") introduced a memory leak as wq_free_lockdep() calls kfree(wq->lock_name), but wq_init_lockdep() does not point wq->lock_name to the newly allocated slab object. This can be reproduced by running LTP fallocate04 followed by oom01 tests: unreferenced object 0xc0000005876384d8 (size 64): comm "fallocate04", pid 26972, jiffies 4297139141 (age 40370.480s) hex dump (first 32 bytes): 28 77 71 5f 63 6f 6d 70 6c 65 74 69 6f 6e 29 65 (wq_completion)e 78 74 34 2d 72 73 76 2d 63 6f 6e 76 65 72 73 69 xt4-rsv-conversi backtrace: [<00000000cb452883>] kvasprintf+0x6c/0xe0 [<000000004654ddac>] kasprintf+0x34/0x60 [<000000001c68f311>] alloc_workqueue+0x1f8/0x6ac [<0000000003c2ad83>] ext4_fill_super+0x23d4/0x3c80 [ext4] [<0000000006610538>] mount_bdev+0x25c/0x290 [<00000000bcf955ec>] ext4_mount+0x28/0x50 [ext4] [<0000000016e08fd3>] legacy_get_tree+0x4c/0xb0 [<0000000042b6a5fc>] vfs_get_tree+0x6c/0x190 [<00000000268ab022>] do_mount+0xb9c/0x1100 [<00000000698e6898>] ksys_mount+0x158/0x180 [<0000000064e391fd>] sys_mount+0x20/0x30 [<00000000ba378f12>] system_call+0x5c/0x70 Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Bart Van Assche <bvanassche@acm.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Will Deacon <will.deacon@arm.com> Cc: catalin.marinas@arm.com Cc: jiangshanlai@gmail.com Cc: tj@kernel.org Fixes: 669de8bda87b ("kernel/workqueue: Use dynamic lockdep keys for workqueues") Link: https://lkml.kernel.org/r/20190307002731.47371-1-cai@lca.pw Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-07 00:27:31 +00:00
wq->lock_name = lock_name;
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
}
static void wq_unregister_lockdep(struct workqueue_struct *wq)
{
lockdep_unregister_key(&wq->key);
}
static void wq_free_lockdep(struct workqueue_struct *wq)
{
if (wq->lock_name != wq->name)
kfree(wq->lock_name);
}
#else
static void wq_init_lockdep(struct workqueue_struct *wq)
{
}
static void wq_unregister_lockdep(struct workqueue_struct *wq)
{
}
static void wq_free_lockdep(struct workqueue_struct *wq)
{
}
#endif
static void rcu_free_wq(struct rcu_head *rcu)
{
struct workqueue_struct *wq =
container_of(rcu, struct workqueue_struct, rcu);
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
wq_free_lockdep(wq);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
free_percpu(wq->cpu_pwq);
free_workqueue_attrs(wq->unbound_attrs);
kfree(wq);
}
static void rcu_free_pool(struct rcu_head *rcu)
{
struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
ida_destroy(&pool->worker_ida);
free_workqueue_attrs(pool->attrs);
kfree(pool);
}
/**
* put_unbound_pool - put a worker_pool
* @pool: worker_pool to put
*
* Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
* safe manner. get_unbound_pool() calls this function on its failure path
* and this function should be able to release pools which went through,
* successfully or not, init_worker_pool().
*
* Should be called with wq_pool_mutex held.
*/
static void put_unbound_pool(struct worker_pool *pool)
{
DECLARE_COMPLETION_ONSTACK(detach_completion);
struct worker *worker;
LIST_HEAD(cull_list);
lockdep_assert_held(&wq_pool_mutex);
if (--pool->refcnt)
return;
/* sanity checks */
if (WARN_ON(!(pool->cpu < 0)) ||
WARN_ON(!list_empty(&pool->worklist)))
return;
/* release id and unhash */
if (pool->id >= 0)
idr_remove(&worker_pool_idr, pool->id);
hash_del(&pool->hash_node);
/*
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
* Become the manager and destroy all workers. This prevents
* @pool's workers from blocking on attach_mutex. We're the last
* manager and @pool gets freed with the flag set.
*
* Having a concurrent manager is quite unlikely to happen as we can
* only get here with
* pwq->refcnt == pool->refcnt == 0
* which implies no work queued to the pool, which implies no worker can
* become the manager. However a worker could have taken the role of
* manager before the refcnts dropped to 0, since maybe_create_worker()
* drops pool->lock
*/
while (true) {
rcuwait_wait_event(&manager_wait,
!(pool->flags & POOL_MANAGER_ACTIVE),
TASK_UNINTERRUPTIBLE);
mutex_lock(&wq_pool_attach_mutex);
raw_spin_lock_irq(&pool->lock);
if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
pool->flags |= POOL_MANAGER_ACTIVE;
break;
}
raw_spin_unlock_irq(&pool->lock);
mutex_unlock(&wq_pool_attach_mutex);
}
workqueue: replace pool->manager_arb mutex with a flag Josef reported a HARDIRQ-safe -> HARDIRQ-unsafe lock order detected by lockdep: [ 1270.472259] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 1270.472783] 4.14.0-rc1-xfstests-12888-g76833e8 #110 Not tainted [ 1270.473240] ----------------------------------------------------- [ 1270.473710] kworker/u5:2/5157 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 1270.474239] (&(&lock->wait_lock)->rlock){+.+.}, at: [<ffffffff8da253d2>] __mutex_unlock_slowpath+0xa2/0x280 [ 1270.474994] [ 1270.474994] and this task is already holding: [ 1270.475440] (&pool->lock/1){-.-.}, at: [<ffffffff8d2992f6>] worker_thread+0x366/0x3c0 [ 1270.476046] which would create a new lock dependency: [ 1270.476436] (&pool->lock/1){-.-.} -> (&(&lock->wait_lock)->rlock){+.+.} [ 1270.476949] [ 1270.476949] but this new dependency connects a HARDIRQ-irq-safe lock: [ 1270.477553] (&pool->lock/1){-.-.} ... [ 1270.488900] to a HARDIRQ-irq-unsafe lock: [ 1270.489327] (&(&lock->wait_lock)->rlock){+.+.} ... [ 1270.494735] Possible interrupt unsafe locking scenario: [ 1270.494735] [ 1270.495250] CPU0 CPU1 [ 1270.495600] ---- ---- [ 1270.495947] lock(&(&lock->wait_lock)->rlock); [ 1270.496295] local_irq_disable(); [ 1270.496753] lock(&pool->lock/1); [ 1270.497205] lock(&(&lock->wait_lock)->rlock); [ 1270.497744] <Interrupt> [ 1270.497948] lock(&pool->lock/1); , which will cause a irq inversion deadlock if the above lock scenario happens. The root cause of this safe -> unsafe lock order is the mutex_unlock(pool->manager_arb) in manage_workers() with pool->lock held. Unlocking mutex while holding an irq spinlock was never safe and this problem has been around forever but it never got noticed because the only time the mutex is usually trylocked while holding irqlock making actual failures very unlikely and lockdep annotation missed the condition until the recent b9c16a0e1f73 ("locking/mutex: Fix lockdep_assert_held() fail"). Using mutex for pool->manager_arb has always been a bit of stretch. It primarily is an mechanism to arbitrate managership between workers which can easily be done with a pool flag. The only reason it became a mutex is that pool destruction path wants to exclude parallel managing operations. This patch replaces the mutex with a new pool flag POOL_MANAGER_ACTIVE and make the destruction path wait for the current manager on a wait queue. v2: Drop unnecessary flag clearing before pool destruction as suggested by Boqun. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Josef Bacik <josef@toxicpanda.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: stable@vger.kernel.org
2017-10-09 15:04:13 +00:00
while ((worker = first_idle_worker(pool)))
set_worker_dying(worker, &cull_list);
WARN_ON(pool->nr_workers || pool->nr_idle);
raw_spin_unlock_irq(&pool->lock);
wake_dying_workers(&cull_list);
if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
pool->detach_completion = &detach_completion;
mutex_unlock(&wq_pool_attach_mutex);
if (pool->detach_completion)
wait_for_completion(pool->detach_completion);
/* shut down the timers */
del_timer_sync(&pool->idle_timer);
cancel_work_sync(&pool->idle_cull_work);
del_timer_sync(&pool->mayday_timer);
/* RCU protected to allow dereferences from get_work_pool() */
call_rcu(&pool->rcu, rcu_free_pool);
}
/**
* get_unbound_pool - get a worker_pool with the specified attributes
* @attrs: the attributes of the worker_pool to get
*
* Obtain a worker_pool which has the same attributes as @attrs, bump the
* reference count and return it. If there already is a matching
* worker_pool, it will be used; otherwise, this function attempts to
* create a new one.
*
* Should be called with wq_pool_mutex held.
*
* Return: On success, a worker_pool with the same attributes as @attrs.
* On failure, %NULL.
*/
static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
{
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
u32 hash = wqattrs_hash(attrs);
struct worker_pool *pool;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
int pod, node = NUMA_NO_NODE;
lockdep_assert_held(&wq_pool_mutex);
/* do we already have a matching pool? */
hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
if (wqattrs_equal(pool->attrs, attrs)) {
pool->refcnt++;
return pool;
}
}
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
/* If __pod_cpumask is contained inside a NUMA pod, that's our node */
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
for (pod = 0; pod < pt->nr_pods; pod++) {
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
node = pt->pod_node[pod];
break;
}
}
/* nope, create a new one */
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
if (!pool || init_worker_pool(pool) < 0)
goto fail;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
pool->node = node;
copy_workqueue_attrs(pool->attrs, attrs);
wqattrs_clear_for_pool(pool->attrs);
if (worker_pool_assign_id(pool) < 0)
goto fail;
/* create and start the initial worker */
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
if (wq_online && !create_worker(pool))
goto fail;
/* install */
hash_add(unbound_pool_hash, &pool->hash_node, hash);
return pool;
fail:
if (pool)
put_unbound_pool(pool);
return NULL;
}
static void rcu_free_pwq(struct rcu_head *rcu)
{
kmem_cache_free(pwq_cache,
container_of(rcu, struct pool_workqueue, rcu));
}
/*
* Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
* refcnt and needs to be destroyed.
*/
static void pwq_release_workfn(struct kthread_work *work)
{
struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
release_work);
struct workqueue_struct *wq = pwq->wq;
struct worker_pool *pool = pwq->pool;
workqueue: fix UAF in pwq_unbound_release_workfn() I got a UAF report when doing fuzz test: [ 152.880091][ T8030] ================================================================== [ 152.881240][ T8030] BUG: KASAN: use-after-free in pwq_unbound_release_workfn+0x50/0x190 [ 152.882442][ T8030] Read of size 4 at addr ffff88810d31bd00 by task kworker/3:2/8030 [ 152.883578][ T8030] [ 152.883932][ T8030] CPU: 3 PID: 8030 Comm: kworker/3:2 Not tainted 5.13.0+ #249 [ 152.885014][ T8030] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014 [ 152.886442][ T8030] Workqueue: events pwq_unbound_release_workfn [ 152.887358][ T8030] Call Trace: [ 152.887837][ T8030] dump_stack_lvl+0x75/0x9b [ 152.888525][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.889371][ T8030] print_address_description.constprop.10+0x48/0x70 [ 152.890326][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891163][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891999][ T8030] kasan_report.cold.15+0x82/0xdb [ 152.892740][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.893594][ T8030] __asan_load4+0x69/0x90 [ 152.894243][ T8030] pwq_unbound_release_workfn+0x50/0x190 [ 152.895057][ T8030] process_one_work+0x47b/0x890 [ 152.895778][ T8030] worker_thread+0x5c/0x790 [ 152.896439][ T8030] ? process_one_work+0x890/0x890 [ 152.897163][ T8030] kthread+0x223/0x250 [ 152.897747][ T8030] ? set_kthread_struct+0xb0/0xb0 [ 152.898471][ T8030] ret_from_fork+0x1f/0x30 [ 152.899114][ T8030] [ 152.899446][ T8030] Allocated by task 8884: [ 152.900084][ T8030] kasan_save_stack+0x21/0x50 [ 152.900769][ T8030] __kasan_kmalloc+0x88/0xb0 [ 152.901416][ T8030] __kmalloc+0x29c/0x460 [ 152.902014][ T8030] alloc_workqueue+0x111/0x8e0 [ 152.902690][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.903459][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.904198][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.904929][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.905599][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.906247][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.906916][ T8030] do_syscall_64+0x34/0xb0 [ 152.907535][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.908365][ T8030] [ 152.908688][ T8030] Freed by task 8884: [ 152.909243][ T8030] kasan_save_stack+0x21/0x50 [ 152.909893][ T8030] kasan_set_track+0x20/0x30 [ 152.910541][ T8030] kasan_set_free_info+0x24/0x40 [ 152.911265][ T8030] __kasan_slab_free+0xf7/0x140 [ 152.911964][ T8030] kfree+0x9e/0x3d0 [ 152.912501][ T8030] alloc_workqueue+0x7d7/0x8e0 [ 152.913182][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.913949][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.914703][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.915402][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.916077][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.916729][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.917414][ T8030] do_syscall_64+0x34/0xb0 [ 152.918034][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.918872][ T8030] [ 152.919203][ T8030] The buggy address belongs to the object at ffff88810d31bc00 [ 152.919203][ T8030] which belongs to the cache kmalloc-512 of size 512 [ 152.921155][ T8030] The buggy address is located 256 bytes inside of [ 152.921155][ T8030] 512-byte region [ffff88810d31bc00, ffff88810d31be00) [ 152.922993][ T8030] The buggy address belongs to the page: [ 152.923800][ T8030] page:ffffea000434c600 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x10d318 [ 152.925249][ T8030] head:ffffea000434c600 order:2 compound_mapcount:0 compound_pincount:0 [ 152.926399][ T8030] flags: 0x57ff00000010200(slab|head|node=1|zone=2|lastcpupid=0x7ff) [ 152.927515][ T8030] raw: 057ff00000010200 dead000000000100 dead000000000122 ffff888009c42c80 [ 152.928716][ T8030] raw: 0000000000000000 0000000080100010 00000001ffffffff 0000000000000000 [ 152.929890][ T8030] page dumped because: kasan: bad access detected [ 152.930759][ T8030] [ 152.931076][ T8030] Memory state around the buggy address: [ 152.931851][ T8030] ffff88810d31bc00: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.932967][ T8030] ffff88810d31bc80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.934068][ T8030] >ffff88810d31bd00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.935189][ T8030] ^ [ 152.935763][ T8030] ffff88810d31bd80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.936847][ T8030] ffff88810d31be00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 152.937940][ T8030] ================================================================== If apply_wqattrs_prepare() fails in alloc_workqueue(), it will call put_pwq() which invoke a work queue to call pwq_unbound_release_workfn() and use the 'wq'. The 'wq' allocated in alloc_workqueue() will be freed in error path when apply_wqattrs_prepare() fails. So it will lead a UAF. CPU0 CPU1 alloc_workqueue() alloc_and_link_pwqs() apply_wqattrs_prepare() fails apply_wqattrs_cleanup() schedule_work(&pwq->unbound_release_work) kfree(wq) worker_thread() pwq_unbound_release_workfn() <- trigger uaf here If apply_wqattrs_prepare() fails, the new pwq are not linked, it doesn't hold any reference to the 'wq', 'wq' is invalid to access in the worker, so add check pwq if linked to fix this. Fixes: 2d5f0764b526 ("workqueue: split apply_workqueue_attrs() into 3 stages") Cc: stable@vger.kernel.org # v4.2+ Reported-by: Hulk Robot <hulkci@huawei.com> Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Pavel Skripkin <paskripkin@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-07-14 09:19:33 +00:00
bool is_last = false;
workqueue: fix UAF in pwq_unbound_release_workfn() I got a UAF report when doing fuzz test: [ 152.880091][ T8030] ================================================================== [ 152.881240][ T8030] BUG: KASAN: use-after-free in pwq_unbound_release_workfn+0x50/0x190 [ 152.882442][ T8030] Read of size 4 at addr ffff88810d31bd00 by task kworker/3:2/8030 [ 152.883578][ T8030] [ 152.883932][ T8030] CPU: 3 PID: 8030 Comm: kworker/3:2 Not tainted 5.13.0+ #249 [ 152.885014][ T8030] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014 [ 152.886442][ T8030] Workqueue: events pwq_unbound_release_workfn [ 152.887358][ T8030] Call Trace: [ 152.887837][ T8030] dump_stack_lvl+0x75/0x9b [ 152.888525][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.889371][ T8030] print_address_description.constprop.10+0x48/0x70 [ 152.890326][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891163][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891999][ T8030] kasan_report.cold.15+0x82/0xdb [ 152.892740][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.893594][ T8030] __asan_load4+0x69/0x90 [ 152.894243][ T8030] pwq_unbound_release_workfn+0x50/0x190 [ 152.895057][ T8030] process_one_work+0x47b/0x890 [ 152.895778][ T8030] worker_thread+0x5c/0x790 [ 152.896439][ T8030] ? process_one_work+0x890/0x890 [ 152.897163][ T8030] kthread+0x223/0x250 [ 152.897747][ T8030] ? set_kthread_struct+0xb0/0xb0 [ 152.898471][ T8030] ret_from_fork+0x1f/0x30 [ 152.899114][ T8030] [ 152.899446][ T8030] Allocated by task 8884: [ 152.900084][ T8030] kasan_save_stack+0x21/0x50 [ 152.900769][ T8030] __kasan_kmalloc+0x88/0xb0 [ 152.901416][ T8030] __kmalloc+0x29c/0x460 [ 152.902014][ T8030] alloc_workqueue+0x111/0x8e0 [ 152.902690][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.903459][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.904198][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.904929][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.905599][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.906247][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.906916][ T8030] do_syscall_64+0x34/0xb0 [ 152.907535][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.908365][ T8030] [ 152.908688][ T8030] Freed by task 8884: [ 152.909243][ T8030] kasan_save_stack+0x21/0x50 [ 152.909893][ T8030] kasan_set_track+0x20/0x30 [ 152.910541][ T8030] kasan_set_free_info+0x24/0x40 [ 152.911265][ T8030] __kasan_slab_free+0xf7/0x140 [ 152.911964][ T8030] kfree+0x9e/0x3d0 [ 152.912501][ T8030] alloc_workqueue+0x7d7/0x8e0 [ 152.913182][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.913949][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.914703][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.915402][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.916077][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.916729][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.917414][ T8030] do_syscall_64+0x34/0xb0 [ 152.918034][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.918872][ T8030] [ 152.919203][ T8030] The buggy address belongs to the object at ffff88810d31bc00 [ 152.919203][ T8030] which belongs to the cache kmalloc-512 of size 512 [ 152.921155][ T8030] The buggy address is located 256 bytes inside of [ 152.921155][ T8030] 512-byte region [ffff88810d31bc00, ffff88810d31be00) [ 152.922993][ T8030] The buggy address belongs to the page: [ 152.923800][ T8030] page:ffffea000434c600 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x10d318 [ 152.925249][ T8030] head:ffffea000434c600 order:2 compound_mapcount:0 compound_pincount:0 [ 152.926399][ T8030] flags: 0x57ff00000010200(slab|head|node=1|zone=2|lastcpupid=0x7ff) [ 152.927515][ T8030] raw: 057ff00000010200 dead000000000100 dead000000000122 ffff888009c42c80 [ 152.928716][ T8030] raw: 0000000000000000 0000000080100010 00000001ffffffff 0000000000000000 [ 152.929890][ T8030] page dumped because: kasan: bad access detected [ 152.930759][ T8030] [ 152.931076][ T8030] Memory state around the buggy address: [ 152.931851][ T8030] ffff88810d31bc00: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.932967][ T8030] ffff88810d31bc80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.934068][ T8030] >ffff88810d31bd00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.935189][ T8030] ^ [ 152.935763][ T8030] ffff88810d31bd80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.936847][ T8030] ffff88810d31be00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 152.937940][ T8030] ================================================================== If apply_wqattrs_prepare() fails in alloc_workqueue(), it will call put_pwq() which invoke a work queue to call pwq_unbound_release_workfn() and use the 'wq'. The 'wq' allocated in alloc_workqueue() will be freed in error path when apply_wqattrs_prepare() fails. So it will lead a UAF. CPU0 CPU1 alloc_workqueue() alloc_and_link_pwqs() apply_wqattrs_prepare() fails apply_wqattrs_cleanup() schedule_work(&pwq->unbound_release_work) kfree(wq) worker_thread() pwq_unbound_release_workfn() <- trigger uaf here If apply_wqattrs_prepare() fails, the new pwq are not linked, it doesn't hold any reference to the 'wq', 'wq' is invalid to access in the worker, so add check pwq if linked to fix this. Fixes: 2d5f0764b526 ("workqueue: split apply_workqueue_attrs() into 3 stages") Cc: stable@vger.kernel.org # v4.2+ Reported-by: Hulk Robot <hulkci@huawei.com> Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Pavel Skripkin <paskripkin@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-07-14 09:19:33 +00:00
/*
* When @pwq is not linked, it doesn't hold any reference to the
workqueue: fix UAF in pwq_unbound_release_workfn() I got a UAF report when doing fuzz test: [ 152.880091][ T8030] ================================================================== [ 152.881240][ T8030] BUG: KASAN: use-after-free in pwq_unbound_release_workfn+0x50/0x190 [ 152.882442][ T8030] Read of size 4 at addr ffff88810d31bd00 by task kworker/3:2/8030 [ 152.883578][ T8030] [ 152.883932][ T8030] CPU: 3 PID: 8030 Comm: kworker/3:2 Not tainted 5.13.0+ #249 [ 152.885014][ T8030] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1.1 04/01/2014 [ 152.886442][ T8030] Workqueue: events pwq_unbound_release_workfn [ 152.887358][ T8030] Call Trace: [ 152.887837][ T8030] dump_stack_lvl+0x75/0x9b [ 152.888525][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.889371][ T8030] print_address_description.constprop.10+0x48/0x70 [ 152.890326][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891163][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.891999][ T8030] kasan_report.cold.15+0x82/0xdb [ 152.892740][ T8030] ? pwq_unbound_release_workfn+0x50/0x190 [ 152.893594][ T8030] __asan_load4+0x69/0x90 [ 152.894243][ T8030] pwq_unbound_release_workfn+0x50/0x190 [ 152.895057][ T8030] process_one_work+0x47b/0x890 [ 152.895778][ T8030] worker_thread+0x5c/0x790 [ 152.896439][ T8030] ? process_one_work+0x890/0x890 [ 152.897163][ T8030] kthread+0x223/0x250 [ 152.897747][ T8030] ? set_kthread_struct+0xb0/0xb0 [ 152.898471][ T8030] ret_from_fork+0x1f/0x30 [ 152.899114][ T8030] [ 152.899446][ T8030] Allocated by task 8884: [ 152.900084][ T8030] kasan_save_stack+0x21/0x50 [ 152.900769][ T8030] __kasan_kmalloc+0x88/0xb0 [ 152.901416][ T8030] __kmalloc+0x29c/0x460 [ 152.902014][ T8030] alloc_workqueue+0x111/0x8e0 [ 152.902690][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.903459][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.904198][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.904929][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.905599][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.906247][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.906916][ T8030] do_syscall_64+0x34/0xb0 [ 152.907535][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.908365][ T8030] [ 152.908688][ T8030] Freed by task 8884: [ 152.909243][ T8030] kasan_save_stack+0x21/0x50 [ 152.909893][ T8030] kasan_set_track+0x20/0x30 [ 152.910541][ T8030] kasan_set_free_info+0x24/0x40 [ 152.911265][ T8030] __kasan_slab_free+0xf7/0x140 [ 152.911964][ T8030] kfree+0x9e/0x3d0 [ 152.912501][ T8030] alloc_workqueue+0x7d7/0x8e0 [ 152.913182][ T8030] __btrfs_alloc_workqueue+0x11e/0x2a0 [ 152.913949][ T8030] btrfs_alloc_workqueue+0x6d/0x1d0 [ 152.914703][ T8030] scrub_workers_get+0x1e8/0x490 [ 152.915402][ T8030] btrfs_scrub_dev+0x1b9/0x9c0 [ 152.916077][ T8030] btrfs_ioctl+0x122c/0x4e50 [ 152.916729][ T8030] __x64_sys_ioctl+0x137/0x190 [ 152.917414][ T8030] do_syscall_64+0x34/0xb0 [ 152.918034][ T8030] entry_SYSCALL_64_after_hwframe+0x44/0xae [ 152.918872][ T8030] [ 152.919203][ T8030] The buggy address belongs to the object at ffff88810d31bc00 [ 152.919203][ T8030] which belongs to the cache kmalloc-512 of size 512 [ 152.921155][ T8030] The buggy address is located 256 bytes inside of [ 152.921155][ T8030] 512-byte region [ffff88810d31bc00, ffff88810d31be00) [ 152.922993][ T8030] The buggy address belongs to the page: [ 152.923800][ T8030] page:ffffea000434c600 refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x10d318 [ 152.925249][ T8030] head:ffffea000434c600 order:2 compound_mapcount:0 compound_pincount:0 [ 152.926399][ T8030] flags: 0x57ff00000010200(slab|head|node=1|zone=2|lastcpupid=0x7ff) [ 152.927515][ T8030] raw: 057ff00000010200 dead000000000100 dead000000000122 ffff888009c42c80 [ 152.928716][ T8030] raw: 0000000000000000 0000000080100010 00000001ffffffff 0000000000000000 [ 152.929890][ T8030] page dumped because: kasan: bad access detected [ 152.930759][ T8030] [ 152.931076][ T8030] Memory state around the buggy address: [ 152.931851][ T8030] ffff88810d31bc00: fa fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.932967][ T8030] ffff88810d31bc80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.934068][ T8030] >ffff88810d31bd00: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.935189][ T8030] ^ [ 152.935763][ T8030] ffff88810d31bd80: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb [ 152.936847][ T8030] ffff88810d31be00: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [ 152.937940][ T8030] ================================================================== If apply_wqattrs_prepare() fails in alloc_workqueue(), it will call put_pwq() which invoke a work queue to call pwq_unbound_release_workfn() and use the 'wq'. The 'wq' allocated in alloc_workqueue() will be freed in error path when apply_wqattrs_prepare() fails. So it will lead a UAF. CPU0 CPU1 alloc_workqueue() alloc_and_link_pwqs() apply_wqattrs_prepare() fails apply_wqattrs_cleanup() schedule_work(&pwq->unbound_release_work) kfree(wq) worker_thread() pwq_unbound_release_workfn() <- trigger uaf here If apply_wqattrs_prepare() fails, the new pwq are not linked, it doesn't hold any reference to the 'wq', 'wq' is invalid to access in the worker, so add check pwq if linked to fix this. Fixes: 2d5f0764b526 ("workqueue: split apply_workqueue_attrs() into 3 stages") Cc: stable@vger.kernel.org # v4.2+ Reported-by: Hulk Robot <hulkci@huawei.com> Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Yang Yingliang <yangyingliang@huawei.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Tested-by: Pavel Skripkin <paskripkin@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-07-14 09:19:33 +00:00
* @wq, and @wq is invalid to access.
*/
if (!list_empty(&pwq->pwqs_node)) {
mutex_lock(&wq->mutex);
list_del_rcu(&pwq->pwqs_node);
is_last = list_empty(&wq->pwqs);
mutex_unlock(&wq->mutex);
}
if (wq->flags & WQ_UNBOUND) {
mutex_lock(&wq_pool_mutex);
put_unbound_pool(pool);
mutex_unlock(&wq_pool_mutex);
}
call_rcu(&pwq->rcu, rcu_free_pwq);
/*
* If we're the last pwq going away, @wq is already dead and no one
* is gonna access it anymore. Schedule RCU free.
*/
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
if (is_last) {
wq_unregister_lockdep(wq);
call_rcu(&wq->rcu, rcu_free_wq);
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
}
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
}
/**
* pwq_adjust_max_active - update a pwq's max_active to the current setting
* @pwq: target pool_workqueue
*
* If @pwq isn't freezing, set @pwq->max_active to the associated
* workqueue's saved_max_active and activate inactive work items
* accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
*/
static void pwq_adjust_max_active(struct pool_workqueue *pwq)
{
struct workqueue_struct *wq = pwq->wq;
bool freezable = wq->flags & WQ_FREEZABLE;
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
unsigned long flags;
/* for @wq->saved_max_active */
lockdep_assert_held(&wq->mutex);
/* fast exit for non-freezable wqs */
if (!freezable && pwq->max_active == wq->saved_max_active)
return;
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/* this function can be called during early boot w/ irq disabled */
raw_spin_lock_irqsave(&pwq->pool->lock, flags);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/*
* During [un]freezing, the caller is responsible for ensuring that
* this function is called at least once after @workqueue_freezing
* is updated and visible.
*/
if (!freezable || !workqueue_freezing) {
pwq->max_active = wq->saved_max_active;
while (!list_empty(&pwq->inactive_works) &&
pwq->nr_active < pwq->max_active)
pwq_activate_first_inactive(pwq);
kick_pool(pwq->pool);
} else {
pwq->max_active = 0;
}
raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
}
/* initialize newly allocated @pwq which is associated with @wq and @pool */
static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
struct worker_pool *pool)
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
{
BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
memset(pwq, 0, sizeof(*pwq));
pwq->pool = pool;
pwq->wq = wq;
pwq->flush_color = -1;
pwq->refcnt = 1;
INIT_LIST_HEAD(&pwq->inactive_works);
INIT_LIST_HEAD(&pwq->pwqs_node);
INIT_LIST_HEAD(&pwq->mayday_node);
kthread_init_work(&pwq->release_work, pwq_release_workfn);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
}
/* sync @pwq with the current state of its associated wq and link it */
static void link_pwq(struct pool_workqueue *pwq)
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
{
struct workqueue_struct *wq = pwq->wq;
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
lockdep_assert_held(&wq->mutex);
/* may be called multiple times, ignore if already linked */
if (!list_empty(&pwq->pwqs_node))
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
return;
/* set the matching work_color */
pwq->work_color = wq->work_color;
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/* sync max_active to the current setting */
pwq_adjust_max_active(pwq);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/* link in @pwq */
list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
}
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
const struct workqueue_attrs *attrs)
{
struct worker_pool *pool;
struct pool_workqueue *pwq;
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
lockdep_assert_held(&wq_pool_mutex);
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
pool = get_unbound_pool(attrs);
if (!pool)
return NULL;
workqueue: async worker destruction worker destruction includes these parts of code: adjust pool's stats remove the worker from idle list detach the worker from the pool kthread_stop() to wait for the worker's task exit free the worker struct We can find out that there is no essential work to do after kthread_stop(), which means destroy_worker() doesn't need to wait for the worker's task exit, so we can remove kthread_stop() and free the worker struct in the worker exiting path. However, put_unbound_pool() still needs to sync the all the workers' destruction before destroying the pool; otherwise, the workers may access to the invalid pool when they are exiting. So we also move the code of "detach the worker" to the exiting path and let put_unbound_pool() to sync with this code via detach_completion. The code of "detach the worker" is wrapped in a new function "worker_detach_from_pool()" although worker_detach_from_pool() is only called once (in worker_thread()) after this patch, but we need to wrap it for these reasons: 1) The code of "detach the worker" is not short enough to unfold them in worker_thread(). 2) the name of "worker_detach_from_pool()" is self-comment, and we add some comments above the function. 3) it will be shared by rescuer in later patch which allows rescuer and normal thread use the same attach/detach frameworks. The worker id is freed when detaching which happens before the worker is fully dead, but this id of the dying worker may be re-used for a new worker, so the dying worker's task name is changed to "worker/dying" to avoid two or several workers having the same name. Since "detach the worker" is moved out from destroy_worker(), destroy_worker() doesn't require manager_mutex, so the "lockdep_assert_held(&pool->manager_mutex)" in destroy_worker() is removed, and destroy_worker() is not protected by manager_mutex in put_unbound_pool(). tj: Minor description updates. Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2014-05-20 09:46:29 +00:00
pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
if (!pwq) {
put_unbound_pool(pool);
return NULL;
}
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
init_pwq(pwq, wq, pool);
return pwq;
}
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/**
* wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
* @attrs: the wq_attrs of the default pwq of the target workqueue
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
* @cpu: the target CPU
* @cpu_going_down: if >= 0, the CPU to consider as offline
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
*
* Calculate the cpumask a workqueue with @attrs should use on @pod. If
* @cpu_going_down is >= 0, that cpu is considered offline during calculation.
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
* The result is stored in @attrs->__pod_cpumask.
*
* If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
* and @pod has online CPUs requested by @attrs, the returned cpumask is the
* intersection of the possible CPUs of @pod and @attrs->cpumask.
*
* The caller is responsible for ensuring that the cpumask of @pod stays stable.
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
*/
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
int cpu_going_down)
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
{
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
int pod = pt->cpu_pod[cpu];
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
/* does @pod have any online CPUs @attrs wants? */
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
if (cpu_going_down >= 0)
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (cpumask_empty(attrs->__pod_cpumask)) {
cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
return;
}
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/* yeap, return possible CPUs in @pod that @attrs wants */
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (cpumask_empty(attrs->__pod_cpumask))
pr_warn_once("WARNING: workqueue cpumask: online intersect > "
"possible intersect\n");
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
}
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
/* install @pwq into @wq's cpu_pwq and return the old pwq */
static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
int cpu, struct pool_workqueue *pwq)
{
struct pool_workqueue *old_pwq;
lockdep_assert_held(&wq_pool_mutex);
lockdep_assert_held(&wq->mutex);
/* link_pwq() can handle duplicate calls */
link_pwq(pwq);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
return old_pwq;
}
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
/* context to store the prepared attrs & pwqs before applying */
struct apply_wqattrs_ctx {
struct workqueue_struct *wq; /* target workqueue */
struct workqueue_attrs *attrs; /* attrs to apply */
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
struct list_head list; /* queued for batching commit */
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
struct pool_workqueue *dfl_pwq;
struct pool_workqueue *pwq_tbl[];
};
/* free the resources after success or abort */
static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
{
if (ctx) {
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
int cpu;
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
for_each_possible_cpu(cpu)
put_pwq_unlocked(ctx->pwq_tbl[cpu]);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
put_pwq_unlocked(ctx->dfl_pwq);
free_workqueue_attrs(ctx->attrs);
kfree(ctx);
}
}
/* allocate the attrs and pwqs for later installation */
static struct apply_wqattrs_ctx *
apply_wqattrs_prepare(struct workqueue_struct *wq,
const struct workqueue_attrs *attrs,
const cpumask_var_t unbound_cpumask)
{
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
struct apply_wqattrs_ctx *ctx;
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
struct workqueue_attrs *new_attrs;
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
int cpu;
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
lockdep_assert_held(&wq_pool_mutex);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
if (WARN_ON(attrs->affn_scope < 0 ||
attrs->affn_scope >= WQ_AFFN_NR_TYPES))
return ERR_PTR(-EINVAL);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
new_attrs = alloc_workqueue_attrs();
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (!ctx || !new_attrs)
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
goto out_free;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/*
* If something goes wrong during CPU up/down, we'll fall back to
* the default pwq covering whole @attrs->cpumask. Always create
* it even if we don't use it immediately.
*/
workqueue: Factor out actual cpumask calculation to reduce subtlety in wq_update_pod() For an unbound pool, multiple cpumasks are involved. U: The user-specified cpumask (may be filtered with cpu_possible_mask). A: The actual cpumask filtered by wq_unbound_cpumask. If the filtering leaves no CPU, wq_unbound_cpumask is used. P: Per-pod subsets of #A. wq->attrs stores #U, wq->dfl_pwq->pool->attrs->cpumask #A, and wq->cpu_pwq[CPU]->pool->attrs->cpumask #P. wq_update_pod() is called to update per-pod pwq's during CPU hotplug. To calculate the new #P for each workqueue, it needs to call wq_calc_pod_cpumask() with @attrs that contains #A. Currently, wq_update_pod() achieves this by calling wq_calc_pod_cpumask() with wq->dfl_pwq->pool->attrs. This is rather fragile because we're calling wq_calc_pod_cpumask() with @attrs of a worker_pool rather than the workqueue's actual attrs when what we want to calculate is the workqueue's cpumask on the pod. While this works fine currently, future changes will add fields which are used differently between workqueues and worker_pools and this subtlety will bite us. This patch factors out #U -> #A calculation from apply_wqattrs_prepare() into wqattrs_actualize_cpumask and updates wq_update_pod() to copy wq->unbound_attrs and use the new helper to obtain #A freshly instead of abusing wq->dfl_pwq->pool_attrs. This shouldn't cause any behavior changes in the current code. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: K Prateek Nayak <kprateek.nayak@amd.com> Reference: http://lkml.kernel.org/r/30625cdd-4d61-594b-8db9-6816b017dde3@amd.com
2023-08-08 01:57:24 +00:00
copy_workqueue_attrs(new_attrs, attrs);
wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
if (!ctx->dfl_pwq)
goto out_free;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
for_each_possible_cpu(cpu) {
if (new_attrs->ordered) {
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
ctx->dfl_pwq->refcnt++;
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
} else {
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
wq_calc_pod_cpumask(new_attrs, cpu, -1);
ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
if (!ctx->pwq_tbl[cpu])
goto out_free;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
}
}
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
/* save the user configured attrs and sanitize it. */
copy_workqueue_attrs(new_attrs, attrs);
cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
ctx->attrs = new_attrs;
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
ctx->wq = wq;
return ctx;
out_free:
free_workqueue_attrs(new_attrs);
apply_wqattrs_cleanup(ctx);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
return ERR_PTR(-ENOMEM);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
}
/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
{
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
int cpu;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/* all pwqs have been created successfully, let's install'em */
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
mutex_lock(&ctx->wq->mutex);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/* save the previous pwq and install the new one */
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
for_each_possible_cpu(cpu)
ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
ctx->pwq_tbl[cpu]);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/* @dfl_pwq might not have been used, ensure it's linked */
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
link_pwq(ctx->dfl_pwq);
swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
mutex_unlock(&ctx->wq->mutex);
}
static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
const struct workqueue_attrs *attrs)
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
{
struct apply_wqattrs_ctx *ctx;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
/* only unbound workqueues can change attributes */
if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
return -EINVAL;
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
/* creating multiple pwqs breaks ordering guarantee */
if (!list_empty(&wq->pwqs)) {
if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
return -EINVAL;
wq->flags &= ~__WQ_ORDERED;
}
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
if (IS_ERR(ctx))
return PTR_ERR(ctx);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
/* the ctx has been prepared successfully, let's commit it */
apply_wqattrs_commit(ctx);
workqueue: split apply_workqueue_attrs() into 3 stages Current apply_workqueue_attrs() includes pwqs-allocation and pwqs-installation, so when we batch multiple apply_workqueue_attrs()s as a transaction, we can't ensure the transaction must succeed or fail as a complete unit. To solve this, we split apply_workqueue_attrs() into three stages. The first stage does the preparation: allocation memory, pwqs. The second stage does the attrs-installaion and pwqs-installation. The third stage frees the allocated memory and (old or unused) pwqs. As the result, batching multiple apply_workqueue_attrs()s can succeed or fail as a complete unit: 1) batch do all the first stage for all the workqueues 2) only commit all when all the above succeed. This patch is a preparation for the next patch ("Allow modifying low level unbound workqueue cpumask") which will do a multiple apply_workqueue_attrs(). The patch doesn't have functionality changed except two minor adjustment: 1) free_unbound_pwq() for the error path is removed, we use the heavier version put_pwq_unlocked() instead since the error path is rare. this adjustment simplifies the code. 2) the memory-allocation is also moved into wq_pool_mutex. this is needed to avoid to do the further splitting. tj: minor updates to comments. Suggested-by: Tejun Heo <tj@kernel.org> Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-27 09:58:38 +00:00
apply_wqattrs_cleanup(ctx);
return 0;
}
/**
* apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
* @wq: the target workqueue
* @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
*
* Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
* a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
* work items are affine to the pod it was issued on. Older pwqs are released as
* in-flight work items finish. Note that a work item which repeatedly requeues
* itself back-to-back will stay on its current pwq.
*
* Performs GFP_KERNEL allocations.
*
* Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
*
* Return: 0 on success and -errno on failure.
*/
int apply_workqueue_attrs(struct workqueue_struct *wq,
const struct workqueue_attrs *attrs)
{
int ret;
lockdep_assert_cpus_held();
mutex_lock(&wq_pool_mutex);
ret = apply_workqueue_attrs_locked(wq, attrs);
mutex_unlock(&wq_pool_mutex);
return ret;
}
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/**
* wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
* @wq: the target workqueue
* @cpu: the CPU to update pool association for
* @hotplug_cpu: the CPU coming up or going down
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
* @online: whether @cpu is coming up or going down
*
* This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
* %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
* @wq accordingly.
*
*
* If pod affinity can't be adjusted due to memory allocation failure, it falls
* back to @wq->dfl_pwq which may not be optimal but is always correct.
*
* Note that when the last allowed CPU of a pod goes offline for a workqueue
* with a cpumask spanning multiple pods, the workers which were already
* executing the work items for the workqueue will lose their CPU affinity and
* may execute on any CPU. This is similar to how per-cpu workqueues behave on
* CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
* responsibility to flush the work item from CPU_DOWN_PREPARE.
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
*/
static void wq_update_pod(struct workqueue_struct *wq, int cpu,
int hotplug_cpu, bool online)
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
{
int off_cpu = online ? -1 : hotplug_cpu;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
struct pool_workqueue *old_pwq = NULL, *pwq;
struct workqueue_attrs *target_attrs;
lockdep_assert_held(&wq_pool_mutex);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
return;
/*
* We don't wanna alloc/free wq_attrs for each wq for each CPU.
* Let's use a preallocated one. The following buf is protected by
* CPU hotplug exclusion.
*/
target_attrs = wq_update_pod_attrs_buf;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
workqueue: Factor out actual cpumask calculation to reduce subtlety in wq_update_pod() For an unbound pool, multiple cpumasks are involved. U: The user-specified cpumask (may be filtered with cpu_possible_mask). A: The actual cpumask filtered by wq_unbound_cpumask. If the filtering leaves no CPU, wq_unbound_cpumask is used. P: Per-pod subsets of #A. wq->attrs stores #U, wq->dfl_pwq->pool->attrs->cpumask #A, and wq->cpu_pwq[CPU]->pool->attrs->cpumask #P. wq_update_pod() is called to update per-pod pwq's during CPU hotplug. To calculate the new #P for each workqueue, it needs to call wq_calc_pod_cpumask() with @attrs that contains #A. Currently, wq_update_pod() achieves this by calling wq_calc_pod_cpumask() with wq->dfl_pwq->pool->attrs. This is rather fragile because we're calling wq_calc_pod_cpumask() with @attrs of a worker_pool rather than the workqueue's actual attrs when what we want to calculate is the workqueue's cpumask on the pod. While this works fine currently, future changes will add fields which are used differently between workqueues and worker_pools and this subtlety will bite us. This patch factors out #U -> #A calculation from apply_wqattrs_prepare() into wqattrs_actualize_cpumask and updates wq_update_pod() to copy wq->unbound_attrs and use the new helper to obtain #A freshly instead of abusing wq->dfl_pwq->pool_attrs. This shouldn't cause any behavior changes in the current code. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: K Prateek Nayak <kprateek.nayak@amd.com> Reference: http://lkml.kernel.org/r/30625cdd-4d61-594b-8db9-6816b017dde3@amd.com
2023-08-08 01:57:24 +00:00
wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
/* nothing to do if the target cpumask matches the current pwq */
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
lockdep_is_held(&wq_pool_mutex));
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
if (wqattrs_equal(target_attrs, pwq->pool->attrs))
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
return;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
/* create a new pwq */
pwq = alloc_unbound_pwq(wq, target_attrs);
if (!pwq) {
pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
wq->name);
goto use_dfl_pwq;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
}
/* Install the new pwq. */
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
mutex_lock(&wq->mutex);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
old_pwq = install_unbound_pwq(wq, cpu, pwq);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
goto out_unlock;
use_dfl_pwq:
mutex_lock(&wq->mutex);
raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
get_pwq(wq->dfl_pwq);
raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
out_unlock:
mutex_unlock(&wq->mutex);
put_pwq_unlocked(old_pwq);
}
static int alloc_and_link_pwqs(struct workqueue_struct *wq)
{
bool highpri = wq->flags & WQ_HIGHPRI;
int cpu, ret;
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
if (!wq->cpu_pwq)
goto enomem;
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
if (!(wq->flags & WQ_UNBOUND)) {
for_each_possible_cpu(cpu) {
struct pool_workqueue **pwq_p =
per_cpu_ptr(wq->cpu_pwq, cpu);
struct worker_pool *pool =
&(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
*pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
pool->node);
if (!*pwq_p)
goto enomem;
init_pwq(*pwq_p, wq, pool);
mutex_lock(&wq->mutex);
link_pwq(*pwq_p);
mutex_unlock(&wq->mutex);
}
return 0;
}
cpus_read_lock();
if (wq->flags & __WQ_ORDERED) {
ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
/* there should only be single pwq for ordering guarantee */
WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
"ordering guarantee broken for workqueue %s\n", wq->name);
} else {
ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
}
cpus_read_unlock();
workqueue: Fix UAF report by KASAN in pwq_release_workfn() Currently, for UNBOUND wq, if the apply_wqattrs_prepare() return error, the apply_wqattr_cleanup() will be called and use the pwq_release_worker kthread to release resources asynchronously. however, the kfree(wq) is invoked directly in failure path of alloc_workqueue(), if the kfree(wq) has been executed and when the pwq_release_workfn() accesses wq, this leads to the following scenario: BUG: KASAN: slab-use-after-free in pwq_release_workfn+0x339/0x380 kernel/workqueue.c:4124 Read of size 4 at addr ffff888027b831c0 by task pool_workqueue_/3 CPU: 0 PID: 3 Comm: pool_workqueue_ Not tainted 6.5.0-rc7-next-20230825-syzkaller #0 Hardware name: Google Compute Engine/Google Compute Engine, BIOS Google 07/26/2023 Call Trace: <TASK> __dump_stack lib/dump_stack.c:88 [inline] dump_stack_lvl+0xd9/0x1b0 lib/dump_stack.c:106 print_address_description mm/kasan/report.c:364 [inline] print_report+0xc4/0x620 mm/kasan/report.c:475 kasan_report+0xda/0x110 mm/kasan/report.c:588 pwq_release_workfn+0x339/0x380 kernel/workqueue.c:4124 kthread_worker_fn+0x2fc/0xa80 kernel/kthread.c:823 kthread+0x33a/0x430 kernel/kthread.c:388 ret_from_fork+0x45/0x80 arch/x86/kernel/process.c:147 ret_from_fork_asm+0x11/0x20 arch/x86/entry/entry_64.S:304 </TASK> Allocated by task 5054: kasan_save_stack+0x33/0x50 mm/kasan/common.c:45 kasan_set_track+0x25/0x30 mm/kasan/common.c:52 ____kasan_kmalloc mm/kasan/common.c:374 [inline] __kasan_kmalloc+0xa2/0xb0 mm/kasan/common.c:383 kmalloc include/linux/slab.h:599 [inline] kzalloc include/linux/slab.h:720 [inline] alloc_workqueue+0x16f/0x1490 kernel/workqueue.c:4684 kvm_mmu_init_tdp_mmu+0x23/0x100 arch/x86/kvm/mmu/tdp_mmu.c:19 kvm_mmu_init_vm+0x248/0x2e0 arch/x86/kvm/mmu/mmu.c:6180 kvm_arch_init_vm+0x39/0x720 arch/x86/kvm/x86.c:12311 kvm_create_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:1222 [inline] kvm_dev_ioctl_create_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:5089 [inline] kvm_dev_ioctl+0xa31/0x1c20 arch/x86/kvm/../../../virt/kvm/kvm_main.c:5131 vfs_ioctl fs/ioctl.c:51 [inline] __do_sys_ioctl fs/ioctl.c:871 [inline] __se_sys_ioctl fs/ioctl.c:857 [inline] __x64_sys_ioctl+0x18f/0x210 fs/ioctl.c:857 do_syscall_x64 arch/x86/entry/common.c:50 [inline] do_syscall_64+0x38/0xb0 arch/x86/entry/common.c:80 entry_SYSCALL_64_after_hwframe+0x63/0xcd Freed by task 5054: kasan_save_stack+0x33/0x50 mm/kasan/common.c:45 kasan_set_track+0x25/0x30 mm/kasan/common.c:52 kasan_save_free_info+0x2b/0x40 mm/kasan/generic.c:522 ____kasan_slab_free mm/kasan/common.c:236 [inline] ____kasan_slab_free+0x15b/0x1b0 mm/kasan/common.c:200 kasan_slab_free include/linux/kasan.h:164 [inline] slab_free_hook mm/slub.c:1800 [inline] slab_free_freelist_hook+0x114/0x1e0 mm/slub.c:1826 slab_free mm/slub.c:3809 [inline] __kmem_cache_free+0xb8/0x2f0 mm/slub.c:3822 alloc_workqueue+0xe76/0x1490 kernel/workqueue.c:4746 kvm_mmu_init_tdp_mmu+0x23/0x100 arch/x86/kvm/mmu/tdp_mmu.c:19 kvm_mmu_init_vm+0x248/0x2e0 arch/x86/kvm/mmu/mmu.c:6180 kvm_arch_init_vm+0x39/0x720 arch/x86/kvm/x86.c:12311 kvm_create_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:1222 [inline] kvm_dev_ioctl_create_vm arch/x86/kvm/../../../virt/kvm/kvm_main.c:5089 [inline] kvm_dev_ioctl+0xa31/0x1c20 arch/x86/kvm/../../../virt/kvm/kvm_main.c:5131 vfs_ioctl fs/ioctl.c:51 [inline] __do_sys_ioctl fs/ioctl.c:871 [inline] __se_sys_ioctl fs/ioctl.c:857 [inline] __x64_sys_ioctl+0x18f/0x210 fs/ioctl.c:857 do_syscall_x64 arch/x86/entry/common.c:50 [inline] do_syscall_64+0x38/0xb0 arch/x86/entry/common.c:80 entry_SYSCALL_64_after_hwframe+0x63/0xcd This commit therefore flush pwq_release_worker in the alloc_and_link_pwqs() before invoke kfree(wq). Reported-by: syzbot+60db9f652c92d5bacba4@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=60db9f652c92d5bacba4 Signed-off-by: Zqiang <qiang.zhang1211@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-20 06:07:04 +00:00
/* for unbound pwq, flush the pwq_release_worker ensures that the
* pwq_release_workfn() completes before calling kfree(wq).
*/
if (ret)
kthread_flush_worker(pwq_release_worker);
return ret;
enomem:
if (wq->cpu_pwq) {
for_each_possible_cpu(cpu) {
struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
if (pwq)
kmem_cache_free(pwq_cache, pwq);
}
free_percpu(wq->cpu_pwq);
wq->cpu_pwq = NULL;
}
return -ENOMEM;
}
static int wq_clamp_max_active(int max_active, unsigned int flags,
const char *name)
{
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
max_active, name, 1, WQ_MAX_ACTIVE);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
}
/*
* Workqueues which may be used during memory reclaim should have a rescuer
* to guarantee forward progress.
*/
static int init_rescuer(struct workqueue_struct *wq)
{
struct worker *rescuer;
int ret;
if (!(wq->flags & WQ_MEM_RECLAIM))
return 0;
rescuer = alloc_worker(NUMA_NO_NODE);
if (!rescuer) {
pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
wq->name);
return -ENOMEM;
}
rescuer->rescue_wq = wq;
rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
if (IS_ERR(rescuer->task)) {
ret = PTR_ERR(rescuer->task);
pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
wq->name, ERR_PTR(ret));
kfree(rescuer);
return ret;
}
wq->rescuer = rescuer;
kthread_bind_mask(rescuer->task, cpu_possible_mask);
wake_up_process(rescuer->task);
return 0;
}
__printf(1, 4)
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
struct workqueue_struct *alloc_workqueue(const char *fmt,
unsigned int flags,
int max_active, ...)
{
va_list args;
struct workqueue_struct *wq;
struct pool_workqueue *pwq;
/*
* Unbound && max_active == 1 used to imply ordered, which is no longer
* the case on many machines due to per-pod pools. While
* alloc_ordered_workqueue() is the right way to create an ordered
* workqueue, keep the previous behavior to avoid subtle breakages.
*/
if ((flags & WQ_UNBOUND) && max_active == 1)
flags |= __WQ_ORDERED;
/* see the comment above the definition of WQ_POWER_EFFICIENT */
if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
flags |= WQ_UNBOUND;
/* allocate wq and format name */
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
wq = kzalloc(sizeof(*wq), GFP_KERNEL);
if (!wq)
return NULL;
if (flags & WQ_UNBOUND) {
wq->unbound_attrs = alloc_workqueue_attrs();
if (!wq->unbound_attrs)
goto err_free_wq;
}
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
va_start(args, max_active);
vsnprintf(wq->name, sizeof(wq->name), fmt, args);
va_end(args);
max_active = max_active ?: WQ_DFL_ACTIVE;
max_active = wq_clamp_max_active(max_active, flags, wq->name);
/* init wq */
wq->flags = flags;
wq->saved_max_active = max_active;
mutex_init(&wq->mutex);
atomic_set(&wq->nr_pwqs_to_flush, 0);
INIT_LIST_HEAD(&wq->pwqs);
workqueue: reimplement workqueue flushing using color coded works Reimplement workqueue flushing using color coded works. wq has the current work color which is painted on the works being issued via cwqs. Flushing a workqueue is achieved by advancing the current work colors of cwqs and waiting for all the works which have any of the previous colors to drain. Currently there are 16 possible colors, one is reserved for no color and 15 colors are useable allowing 14 concurrent flushes. When color space gets full, flush attempts are batched up and processed together when color frees up, so even with many concurrent flushers, the new implementation won't build up huge queue of flushers which has to be processed one after another. Only works which are queued via __queue_work() are colored. Works which are directly put on queue using insert_work() use NO_COLOR and don't participate in workqueue flushing. Currently only works used for work-specific flush fall in this category. This new implementation leaves only cleanup_workqueue_thread() as the user of flush_cpu_workqueue(). Just make its users use flush_workqueue() and kthread_stop() directly and kill cleanup_workqueue_thread(). As workqueue flushing doesn't use barrier request anymore, the comment describing the complex synchronization around it in cleanup_workqueue_thread() is removed together with the function. This new implementation is to allow having and sharing multiple workers per cpu. Please note that one more bit is reserved for a future work flag by this patch. This is to avoid shifting bits and updating comments later. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:11 +00:00
INIT_LIST_HEAD(&wq->flusher_queue);
INIT_LIST_HEAD(&wq->flusher_overflow);
INIT_LIST_HEAD(&wq->maydays);
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
wq_init_lockdep(wq);
INIT_LIST_HEAD(&wq->list);
if (alloc_and_link_pwqs(wq) < 0)
goto err_unreg_lockdep;
if (wq_online && init_rescuer(wq) < 0)
goto err_destroy;
if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
goto err_destroy;
/*
* wq_pool_mutex protects global freeze state and workqueues list.
* Grab it, adjust max_active and add the new @wq to workqueues
* list.
*/
mutex_lock(&wq_pool_mutex);
mutex_lock(&wq->mutex);
for_each_pwq(pwq, wq)
pwq_adjust_max_active(pwq);
mutex_unlock(&wq->mutex);
list_add_tail_rcu(&wq->list, &workqueues);
mutex_unlock(&wq_pool_mutex);
return wq;
err_unreg_lockdep:
workqueue, lockdep: Fix an alloc_workqueue() error path This patch fixes a use-after-free and a memory leak in an alloc_workqueue() error path. Repoted by syzkaller and KASAN: BUG: KASAN: use-after-free in __read_once_size include/linux/compiler.h:197 [inline] BUG: KASAN: use-after-free in lockdep_register_key+0x3b9/0x490 kernel/locking/lockdep.c:1023 Read of size 8 at addr ffff888090fc2698 by task syz-executor134/7858 CPU: 1 PID: 7858 Comm: syz-executor134 Not tainted 5.0.0-rc8-next-20190301 #1 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 Call Trace: __dump_stack lib/dump_stack.c:77 [inline] dump_stack+0x172/0x1f0 lib/dump_stack.c:113 print_address_description.cold+0x7c/0x20d mm/kasan/report.c:187 kasan_report.cold+0x1b/0x40 mm/kasan/report.c:317 __asan_report_load8_noabort+0x14/0x20 mm/kasan/generic_report.c:132 __read_once_size include/linux/compiler.h:197 [inline] lockdep_register_key+0x3b9/0x490 kernel/locking/lockdep.c:1023 wq_init_lockdep kernel/workqueue.c:3444 [inline] alloc_workqueue+0x427/0xe70 kernel/workqueue.c:4263 ucma_open+0x76/0x290 drivers/infiniband/core/ucma.c:1732 misc_open+0x398/0x4c0 drivers/char/misc.c:141 chrdev_open+0x247/0x6b0 fs/char_dev.c:417 do_dentry_open+0x488/0x1160 fs/open.c:771 vfs_open+0xa0/0xd0 fs/open.c:880 do_last fs/namei.c:3416 [inline] path_openat+0x10e9/0x46e0 fs/namei.c:3533 do_filp_open+0x1a1/0x280 fs/namei.c:3563 do_sys_open+0x3fe/0x5d0 fs/open.c:1063 __do_sys_openat fs/open.c:1090 [inline] __se_sys_openat fs/open.c:1084 [inline] __x64_sys_openat+0x9d/0x100 fs/open.c:1084 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Allocated by task 7789: save_stack+0x45/0xd0 mm/kasan/common.c:75 set_track mm/kasan/common.c:87 [inline] __kasan_kmalloc mm/kasan/common.c:497 [inline] __kasan_kmalloc.constprop.0+0xcf/0xe0 mm/kasan/common.c:470 kasan_kmalloc+0x9/0x10 mm/kasan/common.c:511 __do_kmalloc mm/slab.c:3726 [inline] __kmalloc+0x15c/0x740 mm/slab.c:3735 kmalloc include/linux/slab.h:553 [inline] kzalloc include/linux/slab.h:743 [inline] alloc_workqueue+0x13c/0xe70 kernel/workqueue.c:4236 ucma_open+0x76/0x290 drivers/infiniband/core/ucma.c:1732 misc_open+0x398/0x4c0 drivers/char/misc.c:141 chrdev_open+0x247/0x6b0 fs/char_dev.c:417 do_dentry_open+0x488/0x1160 fs/open.c:771 vfs_open+0xa0/0xd0 fs/open.c:880 do_last fs/namei.c:3416 [inline] path_openat+0x10e9/0x46e0 fs/namei.c:3533 do_filp_open+0x1a1/0x280 fs/namei.c:3563 do_sys_open+0x3fe/0x5d0 fs/open.c:1063 __do_sys_openat fs/open.c:1090 [inline] __se_sys_openat fs/open.c:1084 [inline] __x64_sys_openat+0x9d/0x100 fs/open.c:1084 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Freed by task 7789: save_stack+0x45/0xd0 mm/kasan/common.c:75 set_track mm/kasan/common.c:87 [inline] __kasan_slab_free+0x102/0x150 mm/kasan/common.c:459 kasan_slab_free+0xe/0x10 mm/kasan/common.c:467 __cache_free mm/slab.c:3498 [inline] kfree+0xcf/0x230 mm/slab.c:3821 alloc_workqueue+0xc3e/0xe70 kernel/workqueue.c:4295 ucma_open+0x76/0x290 drivers/infiniband/core/ucma.c:1732 misc_open+0x398/0x4c0 drivers/char/misc.c:141 chrdev_open+0x247/0x6b0 fs/char_dev.c:417 do_dentry_open+0x488/0x1160 fs/open.c:771 vfs_open+0xa0/0xd0 fs/open.c:880 do_last fs/namei.c:3416 [inline] path_openat+0x10e9/0x46e0 fs/namei.c:3533 do_filp_open+0x1a1/0x280 fs/namei.c:3563 do_sys_open+0x3fe/0x5d0 fs/open.c:1063 __do_sys_openat fs/open.c:1090 [inline] __se_sys_openat fs/open.c:1084 [inline] __x64_sys_openat+0x9d/0x100 fs/open.c:1084 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe The buggy address belongs to the object at ffff888090fc2580 which belongs to the cache kmalloc-512 of size 512 The buggy address is located 280 bytes inside of 512-byte region [ffff888090fc2580, ffff888090fc2780) Reported-by: syzbot+17335689e239ce135d8b@syzkaller.appspotmail.com Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@surriel.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Will Deacon <will.deacon@arm.com> Fixes: 669de8bda87b ("kernel/workqueue: Use dynamic lockdep keys for workqueues") Link: https://lkml.kernel.org/r/20190303220046.29448-1-bvanassche@acm.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-03-03 22:00:46 +00:00
wq_unregister_lockdep(wq);
wq_free_lockdep(wq);
err_free_wq:
free_workqueue_attrs(wq->unbound_attrs);
kfree(wq);
return NULL;
err_destroy:
destroy_workqueue(wq);
return NULL;
}
kernel/workqueue: Use dynamic lockdep keys for workqueues The following commit: 87915adc3f0a ("workqueue: re-add lockdep dependencies for flushing") improved deadlock checking in the workqueue implementation. Unfortunately that patch also introduced a few false positive lockdep complaints. This patch suppresses these false positives by allocating the workqueue mutex lockdep key dynamically. An example of a false positive lockdep complaint suppressed by this patch can be found below. The root cause of the lockdep complaint shown below is that the direct I/O code can call alloc_workqueue() from inside a work item created by another alloc_workqueue() call and that both workqueues share the same lockdep key. This patch avoids that that lockdep complaint is triggered by allocating the work queue lockdep keys dynamically. In other words, this patch guarantees that a unique lockdep key is associated with each work queue mutex. ====================================================== WARNING: possible circular locking dependency detected 4.19.0-dbg+ #1 Not tainted fio/4129 is trying to acquire lock: 00000000a01cfe1a ((wq_completion)"dio/%s"sb->s_id){+.+.}, at: flush_workqueue+0xd0/0x970 but task is already holding lock: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 (&sb->s_type->i_mutex_key#14){+.+.}: down_write+0x3d/0x80 __generic_file_fsync+0x77/0xf0 ext4_sync_file+0x3c9/0x780 vfs_fsync_range+0x66/0x100 dio_complete+0x2f5/0x360 dio_aio_complete_work+0x1c/0x20 process_one_work+0x481/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #1 ((work_completion)(&dio->complete_work)){+.+.}: process_one_work+0x447/0x9f0 worker_thread+0x63/0x5a0 kthread+0x1cf/0x1f0 ret_from_fork+0x24/0x30 -> #0 ((wq_completion)"dio/%s"sb->s_id){+.+.}: lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe other info that might help us debug this: Chain exists of: (wq_completion)"dio/%s"sb->s_id --> (work_completion)(&dio->complete_work) --> &sb->s_type->i_mutex_key#14 Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&sb->s_type->i_mutex_key#14); lock((work_completion)(&dio->complete_work)); lock(&sb->s_type->i_mutex_key#14); lock((wq_completion)"dio/%s"sb->s_id); *** DEADLOCK *** 1 lock held by fio/4129: #0: 00000000a0acecf9 (&sb->s_type->i_mutex_key#14){+.+.}, at: ext4_file_write_iter+0x154/0x710 stack backtrace: CPU: 3 PID: 4129 Comm: fio Not tainted 4.19.0-dbg+ #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014 Call Trace: dump_stack+0x86/0xc5 print_circular_bug.isra.32+0x20a/0x218 __lock_acquire+0x1c68/0x1cf0 lock_acquire+0xc5/0x200 flush_workqueue+0xf3/0x970 drain_workqueue+0xec/0x220 destroy_workqueue+0x23/0x350 sb_init_dio_done_wq+0x6a/0x80 do_blockdev_direct_IO+0x1f33/0x4be0 __blockdev_direct_IO+0x79/0x86 ext4_direct_IO+0x5df/0xbb0 generic_file_direct_write+0x119/0x220 __generic_file_write_iter+0x131/0x2d0 ext4_file_write_iter+0x3fa/0x710 aio_write+0x235/0x330 io_submit_one+0x510/0xeb0 __x64_sys_io_submit+0x122/0x340 do_syscall_64+0x71/0x220 entry_SYSCALL_64_after_hwframe+0x49/0xbe Signed-off-by: Bart Van Assche <bvanassche@acm.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Johannes Berg <johannes.berg@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Waiman Long <longman@redhat.com> Cc: Will Deacon <will.deacon@arm.com> Link: https://lkml.kernel.org/r/20190214230058.196511-20-bvanassche@acm.org [ Reworked the changelog a bit. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-02-14 23:00:54 +00:00
EXPORT_SYMBOL_GPL(alloc_workqueue);
static bool pwq_busy(struct pool_workqueue *pwq)
{
int i;
for (i = 0; i < WORK_NR_COLORS; i++)
if (pwq->nr_in_flight[i])
return true;
if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
return true;
if (pwq->nr_active || !list_empty(&pwq->inactive_works))
return true;
return false;
}
/**
* destroy_workqueue - safely terminate a workqueue
* @wq: target workqueue
*
* Safely destroy a workqueue. All work currently pending will be done first.
*/
void destroy_workqueue(struct workqueue_struct *wq)
{
struct pool_workqueue *pwq;
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
int cpu;
/*
* Remove it from sysfs first so that sanity check failure doesn't
* lead to sysfs name conflicts.
*/
workqueue_sysfs_unregister(wq);
/* mark the workqueue destruction is in progress */
mutex_lock(&wq->mutex);
wq->flags |= __WQ_DESTROYING;
mutex_unlock(&wq->mutex);
/* drain it before proceeding with destruction */
drain_workqueue(wq);
/* kill rescuer, if sanity checks fail, leave it w/o rescuer */
if (wq->rescuer) {
struct worker *rescuer = wq->rescuer;
/* this prevents new queueing */
raw_spin_lock_irq(&wq_mayday_lock);
wq->rescuer = NULL;
raw_spin_unlock_irq(&wq_mayday_lock);
/* rescuer will empty maydays list before exiting */
kthread_stop(rescuer->task);
kfree(rescuer);
}
/*
* Sanity checks - grab all the locks so that we wait for all
* in-flight operations which may do put_pwq().
*/
mutex_lock(&wq_pool_mutex);
mutex_lock(&wq->mutex);
for_each_pwq(pwq, wq) {
raw_spin_lock_irq(&pwq->pool->lock);
if (WARN_ON(pwq_busy(pwq))) {
pr_warn("%s: %s has the following busy pwq\n",
__func__, wq->name);
show_pwq(pwq);
raw_spin_unlock_irq(&pwq->pool->lock);
mutex_unlock(&wq->mutex);
mutex_unlock(&wq_pool_mutex);
show_one_workqueue(wq);
return;
}
raw_spin_unlock_irq(&pwq->pool->lock);
}
mutex_unlock(&wq->mutex);
/*
* wq list is used to freeze wq, remove from list after
* flushing is complete in case freeze races us.
*/
list_del_rcu(&wq->list);
mutex_unlock(&wq_pool_mutex);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
/*
* We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
* to put the base refs. @wq will be auto-destroyed from the last
* pwq_put. RCU read lock prevents @wq from going away from under us.
*/
rcu_read_lock();
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
for_each_possible_cpu(cpu) {
pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
put_pwq_unlocked(pwq);
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
}
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
put_pwq_unlocked(wq->dfl_pwq);
wq->dfl_pwq = NULL;
rcu_read_unlock();
}
EXPORT_SYMBOL_GPL(destroy_workqueue);
/**
* workqueue_set_max_active - adjust max_active of a workqueue
* @wq: target workqueue
* @max_active: new max_active value.
*
* Set max_active of @wq to @max_active.
*
* CONTEXT:
* Don't call from IRQ context.
*/
void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
{
struct pool_workqueue *pwq;
/* disallow meddling with max_active for ordered workqueues */
if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
return;
max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
mutex_lock(&wq->mutex);
wq->flags &= ~__WQ_ORDERED;
wq->saved_max_active = max_active;
for_each_pwq(pwq, wq)
pwq_adjust_max_active(pwq);
mutex_unlock(&wq->mutex);
}
EXPORT_SYMBOL_GPL(workqueue_set_max_active);
/**
* current_work - retrieve %current task's work struct
*
* Determine if %current task is a workqueue worker and what it's working on.
* Useful to find out the context that the %current task is running in.
*
* Return: work struct if %current task is a workqueue worker, %NULL otherwise.
*/
struct work_struct *current_work(void)
{
struct worker *worker = current_wq_worker();
return worker ? worker->current_work : NULL;
}
EXPORT_SYMBOL(current_work);
/**
* current_is_workqueue_rescuer - is %current workqueue rescuer?
*
* Determine whether %current is a workqueue rescuer. Can be used from
* work functions to determine whether it's being run off the rescuer task.
*
* Return: %true if %current is a workqueue rescuer. %false otherwise.
*/
bool current_is_workqueue_rescuer(void)
{
struct worker *worker = current_wq_worker();
return worker && worker->rescue_wq;
}
/**
* workqueue_congested - test whether a workqueue is congested
* @cpu: CPU in question
* @wq: target workqueue
*
* Test whether @wq's cpu workqueue for @cpu is congested. There is
* no synchronization around this function and the test result is
* unreliable and only useful as advisory hints or for debugging.
*
workqueue: workqueue_congested() shouldn't translate WORK_CPU_UNBOUND into node number df2d5ae499 ("workqueue: map an unbound workqueues to multiple per-node pool_workqueues") made unbound workqueues to map to multiple per-node pool_workqueues and accordingly updated workqueue_contested() so that, for unbound workqueues, it maps the specified @cpu to the NUMA node number to obtain the matching pool_workqueue to query the congested state. Before this change, workqueue_congested() ignored @cpu for unbound workqueues as there was only one pool_workqueue and some users (fscache) called it with WORK_CPU_UNBOUND. After the commit, this causes the following oops as WORK_CPU_UNBOUND gets translated to garbage by cpu_to_node(). BUG: unable to handle kernel paging request at ffff8803598d98b8 IP: [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa PGD 2421067 PUD 0 Oops: 0000 [#1] SMP CPU: 1 PID: 2689 Comm: cat Tainted: GF 3.9.0-fsdevel+ #4 task: ffff88003d801040 ti: ffff880025806000 task.ti: ffff880025806000 RIP: 0010:[<ffffffff81043b7e>] [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa RSP: 0018:ffff880025807ad8 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffff8800388a2400 RCX: 0000000000000003 RDX: ffff880025807fd8 RSI: ffffffff81a31420 RDI: ffff88003d8016e0 RBP: ffff880025807ae8 R08: ffff88003d801730 R09: ffffffffa00b4898 R10: ffffffff81044217 R11: ffff88003d801040 R12: 0000000064206e97 R13: ffff880036059d98 R14: ffff880038cc8080 R15: ffff880038cc82d0 FS: 00007f21afd9c740(0000) GS:ffff88003d100000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: ffff8803598d98b8 CR3: 000000003df49000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Stack: ffff8800388a2400 0000000000000002 ffff880025807b18 ffffffff810442ce ffffffff81044217 ffff880000000002 ffff8800371b4080 ffff88003d112ec0 ffff880025807b38 ffffffffa00810b0 ffff880036059d88 ffff880036059be8 Call Trace: [<ffffffff810442ce>] workqueue_congested+0xb7/0x12c [<ffffffffa00810b0>] fscache_enqueue_object+0xb2/0xe8 [fscache] [<ffffffffa007facd>] __fscache_acquire_cookie+0x3b9/0x56c [fscache] [<ffffffffa00ad8fe>] nfs_fscache_set_inode_cookie+0xee/0x132 [nfs] [<ffffffffa009e112>] do_open+0x9/0xd [nfs] [<ffffffff810e804a>] do_dentry_open+0x175/0x24b [<ffffffff810e8298>] finish_open+0x41/0x51 Fix it by using smp_processor_id() if @cpu is WORK_CPU_UNBOUND. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: David Howells <dhowells@redhat.com> Tested-and-Acked-by: David Howells <dhowells@redhat.com>
2013-05-10 18:10:17 +00:00
* If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
*
* With the exception of ordered workqueues, all workqueues have per-cpu
* pool_workqueues, each with its own congested state. A workqueue being
* congested on one CPU doesn't mean that the workqueue is contested on any
* other CPUs.
workqueue: workqueue_congested() shouldn't translate WORK_CPU_UNBOUND into node number df2d5ae499 ("workqueue: map an unbound workqueues to multiple per-node pool_workqueues") made unbound workqueues to map to multiple per-node pool_workqueues and accordingly updated workqueue_contested() so that, for unbound workqueues, it maps the specified @cpu to the NUMA node number to obtain the matching pool_workqueue to query the congested state. Before this change, workqueue_congested() ignored @cpu for unbound workqueues as there was only one pool_workqueue and some users (fscache) called it with WORK_CPU_UNBOUND. After the commit, this causes the following oops as WORK_CPU_UNBOUND gets translated to garbage by cpu_to_node(). BUG: unable to handle kernel paging request at ffff8803598d98b8 IP: [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa PGD 2421067 PUD 0 Oops: 0000 [#1] SMP CPU: 1 PID: 2689 Comm: cat Tainted: GF 3.9.0-fsdevel+ #4 task: ffff88003d801040 ti: ffff880025806000 task.ti: ffff880025806000 RIP: 0010:[<ffffffff81043b7e>] [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa RSP: 0018:ffff880025807ad8 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffff8800388a2400 RCX: 0000000000000003 RDX: ffff880025807fd8 RSI: ffffffff81a31420 RDI: ffff88003d8016e0 RBP: ffff880025807ae8 R08: ffff88003d801730 R09: ffffffffa00b4898 R10: ffffffff81044217 R11: ffff88003d801040 R12: 0000000064206e97 R13: ffff880036059d98 R14: ffff880038cc8080 R15: ffff880038cc82d0 FS: 00007f21afd9c740(0000) GS:ffff88003d100000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: ffff8803598d98b8 CR3: 000000003df49000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Stack: ffff8800388a2400 0000000000000002 ffff880025807b18 ffffffff810442ce ffffffff81044217 ffff880000000002 ffff8800371b4080 ffff88003d112ec0 ffff880025807b38 ffffffffa00810b0 ffff880036059d88 ffff880036059be8 Call Trace: [<ffffffff810442ce>] workqueue_congested+0xb7/0x12c [<ffffffffa00810b0>] fscache_enqueue_object+0xb2/0xe8 [fscache] [<ffffffffa007facd>] __fscache_acquire_cookie+0x3b9/0x56c [fscache] [<ffffffffa00ad8fe>] nfs_fscache_set_inode_cookie+0xee/0x132 [nfs] [<ffffffffa009e112>] do_open+0x9/0xd [nfs] [<ffffffff810e804a>] do_dentry_open+0x175/0x24b [<ffffffff810e8298>] finish_open+0x41/0x51 Fix it by using smp_processor_id() if @cpu is WORK_CPU_UNBOUND. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: David Howells <dhowells@redhat.com> Tested-and-Acked-by: David Howells <dhowells@redhat.com>
2013-05-10 18:10:17 +00:00
*
* Return:
* %true if congested, %false otherwise.
*/
bool workqueue_congested(int cpu, struct workqueue_struct *wq)
{
struct pool_workqueue *pwq;
bool ret;
rcu_read_lock();
preempt_disable();
workqueue: workqueue_congested() shouldn't translate WORK_CPU_UNBOUND into node number df2d5ae499 ("workqueue: map an unbound workqueues to multiple per-node pool_workqueues") made unbound workqueues to map to multiple per-node pool_workqueues and accordingly updated workqueue_contested() so that, for unbound workqueues, it maps the specified @cpu to the NUMA node number to obtain the matching pool_workqueue to query the congested state. Before this change, workqueue_congested() ignored @cpu for unbound workqueues as there was only one pool_workqueue and some users (fscache) called it with WORK_CPU_UNBOUND. After the commit, this causes the following oops as WORK_CPU_UNBOUND gets translated to garbage by cpu_to_node(). BUG: unable to handle kernel paging request at ffff8803598d98b8 IP: [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa PGD 2421067 PUD 0 Oops: 0000 [#1] SMP CPU: 1 PID: 2689 Comm: cat Tainted: GF 3.9.0-fsdevel+ #4 task: ffff88003d801040 ti: ffff880025806000 task.ti: ffff880025806000 RIP: 0010:[<ffffffff81043b7e>] [<ffffffff81043b7e>] unbound_pwq_by_node+0xa1/0xfa RSP: 0018:ffff880025807ad8 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffff8800388a2400 RCX: 0000000000000003 RDX: ffff880025807fd8 RSI: ffffffff81a31420 RDI: ffff88003d8016e0 RBP: ffff880025807ae8 R08: ffff88003d801730 R09: ffffffffa00b4898 R10: ffffffff81044217 R11: ffff88003d801040 R12: 0000000064206e97 R13: ffff880036059d98 R14: ffff880038cc8080 R15: ffff880038cc82d0 FS: 00007f21afd9c740(0000) GS:ffff88003d100000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: ffff8803598d98b8 CR3: 000000003df49000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Stack: ffff8800388a2400 0000000000000002 ffff880025807b18 ffffffff810442ce ffffffff81044217 ffff880000000002 ffff8800371b4080 ffff88003d112ec0 ffff880025807b38 ffffffffa00810b0 ffff880036059d88 ffff880036059be8 Call Trace: [<ffffffff810442ce>] workqueue_congested+0xb7/0x12c [<ffffffffa00810b0>] fscache_enqueue_object+0xb2/0xe8 [fscache] [<ffffffffa007facd>] __fscache_acquire_cookie+0x3b9/0x56c [fscache] [<ffffffffa00ad8fe>] nfs_fscache_set_inode_cookie+0xee/0x132 [nfs] [<ffffffffa009e112>] do_open+0x9/0xd [nfs] [<ffffffff810e804a>] do_dentry_open+0x175/0x24b [<ffffffff810e8298>] finish_open+0x41/0x51 Fix it by using smp_processor_id() if @cpu is WORK_CPU_UNBOUND. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: David Howells <dhowells@redhat.com> Tested-and-Acked-by: David Howells <dhowells@redhat.com>
2013-05-10 18:10:17 +00:00
if (cpu == WORK_CPU_UNBOUND)
cpu = smp_processor_id();
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
ret = !list_empty(&pwq->inactive_works);
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
preempt_enable();
rcu_read_unlock();
return ret;
}
EXPORT_SYMBOL_GPL(workqueue_congested);
/**
* work_busy - test whether a work is currently pending or running
* @work: the work to be tested
*
* Test whether @work is currently pending or running. There is no
* synchronization around this function and the test result is
* unreliable and only useful as advisory hints or for debugging.
*
* Return:
* OR'd bitmask of WORK_BUSY_* bits.
*/
unsigned int work_busy(struct work_struct *work)
{
struct worker_pool *pool;
unsigned long flags;
unsigned int ret = 0;
if (work_pending(work))
ret |= WORK_BUSY_PENDING;
rcu_read_lock();
pool = get_work_pool(work);
if (pool) {
raw_spin_lock_irqsave(&pool->lock, flags);
if (find_worker_executing_work(pool, work))
ret |= WORK_BUSY_RUNNING;
raw_spin_unlock_irqrestore(&pool->lock, flags);
}
rcu_read_unlock();
return ret;
}
EXPORT_SYMBOL_GPL(work_busy);
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
/**
* set_worker_desc - set description for the current work item
* @fmt: printf-style format string
* @...: arguments for the format string
*
* This function can be called by a running work function to describe what
* the work item is about. If the worker task gets dumped, this
* information will be printed out together to help debugging. The
* description can be at most WORKER_DESC_LEN including the trailing '\0'.
*/
void set_worker_desc(const char *fmt, ...)
{
struct worker *worker = current_wq_worker();
va_list args;
if (worker) {
va_start(args, fmt);
vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
va_end(args);
}
}
scsi: zfcp: workqueue: set description for port work items with their WWPN as context As a prerequisite, complement commit 3d1cb2059d93 ("workqueue: include workqueue info when printing debug dump of a worker task") to be usable with kernel modules by exporting the symbol set_worker_desc(). Current built-in user was introduced with commit ef3b101925f2 ("writeback: set worker desc to identify writeback workers in task dumps"). Can help distinguishing work items which do not have adapter scope. Description is printed out with task dump for debugging on WARN, BUG, panic, or magic-sysrq [show-task-states(t)]. Example: $ echo 0 >| /sys/bus/ccw/drivers/zfcp/0.0.1880/0x50050763031bd327/failed & $ echo 't' >| /proc/sysrq-trigger $ dmesg sysrq: SysRq : Show State task PC stack pid father ... zfcp_q_0.0.1880 S14640 2165 2 0x02000000 Call Trace: ([<00000000009df464>] __schedule+0xbf4/0xc78) [<00000000009df57c>] schedule+0x94/0xc0 [<0000000000168654>] rescuer_thread+0x33c/0x3a0 [<000000000016f8be>] kthread+0x166/0x178 [<00000000009e71f2>] kernel_thread_starter+0x6/0xc [<00000000009e71ec>] kernel_thread_starter+0x0/0xc no locks held by zfcp_q_0.0.1880/2165. ... kworker/u512:2 D11280 2193 2 0x02000000 Workqueue: zfcp_q_0.0.1880 zfcp_scsi_rport_work [zfcp] (zrpd-50050763031bd327) ^^^^^^^^^^^^^^^^^^^^^ Call Trace: ([<00000000009df464>] __schedule+0xbf4/0xc78) [<00000000009df57c>] schedule+0x94/0xc0 [<00000000009e50c0>] schedule_timeout+0x488/0x4d0 [<00000000001e425c>] msleep+0x5c/0x78 >>test code only<< [<000003ff8008a21e>] zfcp_scsi_rport_work+0xbe/0x100 [zfcp] [<0000000000167154>] process_one_work+0x3b4/0x718 [<000000000016771c>] worker_thread+0x264/0x408 [<000000000016f8be>] kthread+0x166/0x178 [<00000000009e71f2>] kernel_thread_starter+0x6/0xc [<00000000009e71ec>] kernel_thread_starter+0x0/0xc 2 locks held by kworker/u512:2/2193: #0: (name){++++.+}, at: [<0000000000166f4e>] process_one_work+0x1ae/0x718 #1: ((&(&port->rport_work)->work)){+.+.+.}, at: [<0000000000166f4e>] process_one_work+0x1ae/0x718 ... ============================================= Showing busy workqueues and worker pools: workqueue zfcp_q_0.0.1880: flags=0x2000a pwq 512: cpus=0-255 flags=0x4 nice=0 active=1/1 in-flight: 2193:zfcp_scsi_rport_work [zfcp] pool 512: cpus=0-255 flags=0x4 nice=0 hung=0s workers=4 idle: 5 2354 2311 Work items with adapter scope are already identified by the workqueue name "zfcp_q_<devbusid>" and the work item function name. Signed-off-by: Steffen Maier <maier@linux.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Reviewed-by: Benjamin Block <bblock@linux.ibm.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
2018-05-17 17:14:57 +00:00
EXPORT_SYMBOL_GPL(set_worker_desc);
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
/**
* print_worker_info - print out worker information and description
* @log_lvl: the log level to use when printing
* @task: target task
*
* If @task is a worker and currently executing a work item, print out the
* name of the workqueue being serviced and worker description set with
* set_worker_desc() by the currently executing work item.
*
* This function can be safely called on any task as long as the
* task_struct itself is accessible. While safe, this function isn't
* synchronized and may print out mixups or garbages of limited length.
*/
void print_worker_info(const char *log_lvl, struct task_struct *task)
{
work_func_t *fn = NULL;
char name[WQ_NAME_LEN] = { };
char desc[WORKER_DESC_LEN] = { };
struct pool_workqueue *pwq = NULL;
struct workqueue_struct *wq = NULL;
struct worker *worker;
if (!(task->flags & PF_WQ_WORKER))
return;
/*
* This function is called without any synchronization and @task
* could be in any state. Be careful with dereferences.
*/
worker = kthread_probe_data(task);
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
/*
* Carefully copy the associated workqueue's workfn, name and desc.
* Keep the original last '\0' in case the original is garbage.
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
*/
copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
if (fn || name[0] || desc[0]) {
2019-03-25 19:32:28 +00:00
printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
if (strcmp(name, desc))
workqueue: include workqueue info when printing debug dump of a worker task One of the problems that arise when converting dedicated custom threadpool to workqueue is that the shared worker pool used by workqueue anonimizes each worker making it more difficult to identify what the worker was doing on which target from the output of sysrq-t or debug dump from oops, BUG() and friends. This patch implements set_worker_desc() which can be called from any workqueue work function to set its description. When the worker task is dumped for whatever reason - sysrq-t, WARN, BUG, oops, lockdep assertion and so on - the description will be printed out together with the workqueue name and the worker function pointer. The printing side is implemented by print_worker_info() which is called from functions in task dump paths - sched_show_task() and dump_stack_print_info(). print_worker_info() can be safely called on any task in any state as long as the task struct itself is accessible. It uses probe_*() functions to access worker fields. It may print garbage if something went very wrong, but it wouldn't cause (another) oops. The description is currently limited to 24bytes including the terminating \0. worker->desc_valid and workder->desc[] are added and the 64 bytes marker which was already incorrect before adding the new fields is moved to the correct position. Here's an example dump with writeback updated to set the bdi name as worker desc. Hardware name: Bochs Modules linked in: Pid: 7, comm: kworker/u9:0 Not tainted 3.9.0-rc1-work+ #1 Workqueue: writeback bdi_writeback_workfn (flush-8:0) ffffffff820a3ab0 ffff88000f6e9cb8 ffffffff81c61845 ffff88000f6e9cf8 ffffffff8108f50f 0000000000000000 0000000000000000 ffff88000cde16b0 ffff88000cde1aa8 ffff88001ee19240 ffff88000f6e9fd8 ffff88000f6e9d08 Call Trace: [<ffffffff81c61845>] dump_stack+0x19/0x1b [<ffffffff8108f50f>] warn_slowpath_common+0x7f/0xc0 [<ffffffff8108f56a>] warn_slowpath_null+0x1a/0x20 [<ffffffff81200150>] bdi_writeback_workfn+0x2a0/0x3b0 ... Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ingo Molnar <mingo@redhat.com> Acked-by: Jan Kara <jack@suse.cz> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 22:27:22 +00:00
pr_cont(" (%s)", desc);
pr_cont("\n");
}
}
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
static void pr_cont_pool_info(struct worker_pool *pool)
{
pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
if (pool->node != NUMA_NO_NODE)
pr_cont(" node=%d", pool->node);
pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
}
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
struct pr_cont_work_struct {
bool comma;
work_func_t func;
long ctr;
};
static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
{
if (!pcwsp->ctr)
goto out_record;
if (func == pcwsp->func) {
pcwsp->ctr++;
return;
}
if (pcwsp->ctr == 1)
pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
else
pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
pcwsp->ctr = 0;
out_record:
if ((long)func == -1L)
return;
pcwsp->comma = comma;
pcwsp->func = func;
pcwsp->ctr = 1;
}
static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
{
if (work->func == wq_barrier_func) {
struct wq_barrier *barr;
barr = container_of(work, struct wq_barrier, work);
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
pr_cont("%s BAR(%d)", comma ? "," : "",
task_pid_nr(barr->task));
} else {
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
if (!comma)
pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
pr_cont_work_flush(comma, work->func, pcwsp);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
}
}
static void show_pwq(struct pool_workqueue *pwq)
{
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
struct pr_cont_work_struct pcws = { .ctr = 0, };
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
struct worker_pool *pool = pwq->pool;
struct work_struct *work;
struct worker *worker;
bool has_in_flight = false, has_pending = false;
int bkt;
pr_info(" pwq %d:", pool->id);
pr_cont_pool_info(pool);
pr_cont(" active=%d/%d refcnt=%d%s\n",
pwq->nr_active, pwq->max_active, pwq->refcnt,
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
hash_for_each(pool->busy_hash, bkt, worker, hentry) {
if (worker->current_pwq == pwq) {
has_in_flight = true;
break;
}
}
if (has_in_flight) {
bool comma = false;
pr_info(" in-flight:");
hash_for_each(pool->busy_hash, bkt, worker, hentry) {
if (worker->current_pwq != pwq)
continue;
2019-03-25 19:32:28 +00:00
pr_cont("%s %d%s:%ps", comma ? "," : "",
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
task_pid_nr(worker->task),
worker->rescue_wq ? "(RESCUER)" : "",
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
worker->current_func);
list_for_each_entry(work, &worker->scheduled, entry)
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work(false, work, &pcws);
pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
comma = true;
}
pr_cont("\n");
}
list_for_each_entry(work, &pool->worklist, entry) {
if (get_work_pwq(work) == pwq) {
has_pending = true;
break;
}
}
if (has_pending) {
bool comma = false;
pr_info(" pending:");
list_for_each_entry(work, &pool->worklist, entry) {
if (get_work_pwq(work) != pwq)
continue;
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work(comma, work, &pcws);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
}
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
pr_cont("\n");
}
if (!list_empty(&pwq->inactive_works)) {
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
bool comma = false;
pr_info(" inactive:");
list_for_each_entry(work, &pwq->inactive_works, entry) {
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work(comma, work, &pcws);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
}
workqueue: Make show_pwq() use run-length encoding The show_pwq() function dumps out a pool_workqueue structure's activity, including the pending work-queue handlers: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: test_work_func, test_work_func, test_work_func1, test_work_func1, test_work_func1, test_work_func1, test_work_func1 When large systems are facing certain types of hang conditions, it is not unusual for this "pending" list to contain runs of hundreds of identical function names. This "wall of text" is difficult to read, and worse yet, it can be interleaved with other output such as stack traces. Therefore, make show_pwq() use run-length encoding so that the above printout instead looks like this: Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x1 nice=0 active=10/256 refcnt=11 in-flight: 7:test_work_func, 64:test_work_func, 249:test_work_func pending: 2*test_work_func, 5*test_work_func1 When no comma would be printed, including the WORK_STRUCT_LINKED case, a new run is started unconditionally. This output is more readable, places less stress on the hardware, firmware, and software on the console-log path, and reduces interference with other output. Signed-off-by: Paul E. McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Dave Jones <davej@codemonkey.org.uk> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-01-07 00:10:24 +00:00
pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
pr_cont("\n");
}
}
/**
* show_one_workqueue - dump state of specified workqueue
* @wq: workqueue whose state will be printed
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
*/
void show_one_workqueue(struct workqueue_struct *wq)
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
{
struct pool_workqueue *pwq;
bool idle = true;
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
unsigned long flags;
for_each_pwq(pwq, wq) {
if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
idle = false;
break;
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
}
}
if (idle) /* Nothing to print for idle workqueue */
return;
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
for_each_pwq(pwq, wq) {
raw_spin_lock_irqsave(&pwq->pool->lock, flags);
if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
/*
* Defer printing to avoid deadlocks in console
* drivers that queue work while holding locks
* also taken in their write paths.
*/
printk_deferred_enter();
show_pwq(pwq);
printk_deferred_exit();
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
}
raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
/*
* We could be printing a lot from atomic context, e.g.
* sysrq-t -> show_all_workqueues(). Avoid triggering
* hard lockup.
*/
touch_nmi_watchdog();
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
}
}
/**
* show_one_worker_pool - dump state of specified worker pool
* @pool: worker pool whose state will be printed
*/
static void show_one_worker_pool(struct worker_pool *pool)
{
struct worker *worker;
bool first = true;
unsigned long flags;
unsigned long hung = 0;
raw_spin_lock_irqsave(&pool->lock, flags);
if (pool->nr_workers == pool->nr_idle)
goto next_pool;
/* How long the first pending work is waiting for a worker. */
if (!list_empty(&pool->worklist))
hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
/*
* Defer printing to avoid deadlocks in console drivers that
* queue work while holding locks also taken in their write
* paths.
*/
printk_deferred_enter();
pr_info("pool %d:", pool->id);
pr_cont_pool_info(pool);
pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
if (pool->manager)
pr_cont(" manager: %d",
task_pid_nr(pool->manager->task));
list_for_each_entry(worker, &pool->idle_list, entry) {
pr_cont(" %s%d", first ? "idle: " : "",
task_pid_nr(worker->task));
first = false;
}
pr_cont("\n");
printk_deferred_exit();
next_pool:
raw_spin_unlock_irqrestore(&pool->lock, flags);
/*
* We could be printing a lot from atomic context, e.g.
* sysrq-t -> show_all_workqueues(). Avoid triggering
* hard lockup.
*/
touch_nmi_watchdog();
}
/**
* show_all_workqueues - dump workqueue state
*
* Called from a sysrq handler and prints out all busy workqueues and pools.
*/
void show_all_workqueues(void)
{
struct workqueue_struct *wq;
struct worker_pool *pool;
int pi;
rcu_read_lock();
pr_info("Showing busy workqueues and worker pools:\n");
list_for_each_entry_rcu(wq, &workqueues, list)
show_one_workqueue(wq);
for_each_pool(pool, pi)
show_one_worker_pool(pool);
rcu_read_unlock();
workqueue: dump workqueues on sysrq-t Workqueues are used extensively throughout the kernel but sometimes it's difficult to debug stalls involving work items because visibility into its inner workings is fairly limited. Although sysrq-t task dump annotates each active worker task with the information on the work item being executed, it is challenging to find out which work items are pending or delayed on which queues and how pools are being managed. This patch implements show_workqueue_state() which dumps all busy workqueues and pools and is called from the sysrq-t handler. At the end of sysrq-t dump, something like the following is printed. Showing busy workqueues and worker pools: ... workqueue filler_wq: flags=0x0 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 in-flight: 491:filler_workfn, 507:filler_workfn pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 in-flight: 501:filler_workfn pending: filler_workfn ... workqueue test_wq: flags=0x8 pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 in-flight: 510(RESCUER):test_workfn BAR(69) BAR(500) delayed: test_workfn1 BAR(492), test_workfn2 ... pool 0: cpus=0 node=0 flags=0x0 nice=0 workers=2 manager: 137 pool 2: cpus=1 node=0 flags=0x0 nice=0 workers=3 manager: 469 pool 3: cpus=1 node=0 flags=0x0 nice=-20 workers=2 idle: 16 pool 8: cpus=0-3 flags=0x4 nice=0 workers=2 manager: 62 The above shows that test_wq is executing test_workfn() on pid 510 which is the rescuer and also that there are two tasks 69 and 500 waiting for the work item to finish in flush_work(). As test_wq has max_active of 1, there are two work items for test_workfn1() and test_workfn2() which are delayed till the current work item is finished. In addition, pid 492 is flushing test_workfn1(). The work item for test_workfn() is being executed on pwq of pool 2 which is the normal priority per-cpu pool for CPU 1. The pool has three workers, two of which are executing filler_workfn() for filler_wq and the last one is assuming the manager role trying to create more workers. This extra workqueue state dump will hopefully help chasing down hangs involving workqueues. v3: cpulist_pr_cont() replaced with "%*pbl" printf formatting. v2: As suggested by Andrew, minor formatting change in pr_cont_work(), printk()'s replaced with pr_info()'s, and cpumask printing now uses cpulist_pr_cont(). Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com>
2015-03-09 13:22:28 +00:00
}
/**
* show_freezable_workqueues - dump freezable workqueue state
*
* Called from try_to_freeze_tasks() and prints out all freezable workqueues
* still busy.
*/
void show_freezable_workqueues(void)
{
struct workqueue_struct *wq;
rcu_read_lock();
pr_info("Showing freezable workqueues that are still busy:\n");
list_for_each_entry_rcu(wq, &workqueues, list) {
if (!(wq->flags & WQ_FREEZABLE))
continue;
show_one_workqueue(wq);
}
rcu_read_unlock();
}
/* used to show worker information through /proc/PID/{comm,stat,status} */
void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
{
int off;
/* always show the actual comm */
off = strscpy(buf, task->comm, size);
if (off < 0)
return;
/* stabilize PF_WQ_WORKER and worker pool association */
mutex_lock(&wq_pool_attach_mutex);
if (task->flags & PF_WQ_WORKER) {
struct worker *worker = kthread_data(task);
struct worker_pool *pool = worker->pool;
if (pool) {
raw_spin_lock_irq(&pool->lock);
/*
* ->desc tracks information (wq name or
* set_worker_desc()) for the latest execution. If
* current, prepend '+', otherwise '-'.
*/
if (worker->desc[0] != '\0') {
if (worker->current_work)
scnprintf(buf + off, size - off, "+%s",
worker->desc);
else
scnprintf(buf + off, size - off, "-%s",
worker->desc);
}
raw_spin_unlock_irq(&pool->lock);
}
}
mutex_unlock(&wq_pool_attach_mutex);
}
#ifdef CONFIG_SMP
/*
* CPU hotplug.
*
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* There are two challenges in supporting CPU hotplug. Firstly, there
* are a lot of assumptions on strong associations among work, pwq and
* pool which make migrating pending and scheduled works very
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* difficult to implement without impacting hot paths. Secondly,
* worker pools serve mix of short, long and very long running works making
workqueue: implement concurrency managed dynamic worker pool Instead of creating a worker for each cwq and putting it into the shared pool, manage per-cpu workers dynamically. Works aren't supposed to be cpu cycle hogs and maintaining just enough concurrency to prevent work processing from stalling due to lack of processing context is optimal. gcwq keeps the number of concurrent active workers to minimum but no less. As long as there's one or more running workers on the cpu, no new worker is scheduled so that works can be processed in batch as much as possible but when the last running worker blocks, gcwq immediately schedules new worker so that the cpu doesn't sit idle while there are works to be processed. gcwq always keeps at least single idle worker around. When a new worker is necessary and the worker is the last idle one, the worker assumes the role of "manager" and manages the worker pool - ie. creates another worker. Forward-progress is guaranteed by having dedicated rescue workers for workqueues which may be necessary while creating a new worker. When the manager is having problem creating a new worker, mayday timer activates and rescue workers are summoned to the cpu and execute works which might be necessary to create new workers. Trustee is expanded to serve the role of manager while a CPU is being taken down and stays down. As no new works are supposed to be queued on a dead cpu, it just needs to drain all the existing ones. Trustee continues to try to create new workers and summon rescuers as long as there are pending works. If the CPU is brought back up while the trustee is still trying to drain the gcwq from the previous offlining, the trustee will kill all idles ones and tell workers which are still busy to rebind to the cpu, and pass control over to gcwq which assumes the manager role as necessary. Concurrency managed worker pool reduces the number of workers drastically. Only workers which are necessary to keep the processing going are created and kept. Also, it reduces cache footprint by avoiding unnecessarily switching contexts between different workers. Please note that this patch does not increase max_active of any workqueue. All workqueues can still only process one work per cpu. Signed-off-by: Tejun Heo <tj@kernel.org>
2010-06-29 08:07:14 +00:00
* blocked draining impractical.
*
* This is solved by allowing the pools to be disassociated from the CPU
* running as an unbound one and allowing it to be reattached later if the
* cpu comes back online.
*/
static void unbind_workers(int cpu)
{
struct worker_pool *pool;
struct worker *worker;
for_each_cpu_worker_pool(pool, cpu) {
mutex_lock(&wq_pool_attach_mutex);
raw_spin_lock_irq(&pool->lock);
/*
* We've blocked all attach/detach operations. Make all workers
* unbound and set DISASSOCIATED. Before this, all workers
* must be on the cpu. After this, they may become diasporas.
* And the preemption disabled section in their sched callbacks
* are guaranteed to see WORKER_UNBOUND since the code here
* is on the same cpu.
*/
for_each_pool_worker(worker, pool)
worker->flags |= WORKER_UNBOUND;
pool->flags |= POOL_DISASSOCIATED;
/*
* The handling of nr_running in sched callbacks are disabled
* now. Zap nr_running. After this, nr_running stays zero and
* need_more_worker() and keep_working() are always true as
* long as the worklist is not empty. This pool now behaves as
* an unbound (in terms of concurrency management) pool which
* are served by workers tied to the pool.
*/
pool->nr_running = 0;
/*
* With concurrency management just turned off, a busy
* worker blocking could lead to lengthy stalls. Kick off
* unbound chain execution of currently pending work items.
*/
kick_pool(pool);
raw_spin_unlock_irq(&pool->lock);
for_each_pool_worker(worker, pool)
unbind_worker(worker);
mutex_unlock(&wq_pool_attach_mutex);
}
}
/**
* rebind_workers - rebind all workers of a pool to the associated CPU
* @pool: pool of interest
*
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
* @pool->cpu is coming online. Rebind all workers to the CPU.
*/
static void rebind_workers(struct worker_pool *pool)
{
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
struct worker *worker;
lockdep_assert_held(&wq_pool_attach_mutex);
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
/*
* Restore CPU affinity of all workers. As all idle workers should
* be on the run-queue of the associated CPU before any local
* wake-ups for concurrency management happen, restore CPU affinity
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
* of all workers first and then clear UNBOUND. As we're called
* from CPU_ONLINE, the following shouldn't fail.
*/
for_each_pool_worker(worker, pool) {
kthread_set_per_cpu(worker->task, pool->cpu);
WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
pool_allowed_cpus(pool)) < 0);
}
raw_spin_lock_irq(&pool->lock);
workqueue: fix rebind bound workers warning ------------[ cut here ]------------ WARNING: CPU: 0 PID: 16 at kernel/workqueue.c:4559 rebind_workers+0x1c0/0x1d0 Modules linked in: CPU: 0 PID: 16 Comm: cpuhp/0 Not tainted 4.6.0-rc4+ #31 Hardware name: IBM IBM System x3550 M4 Server -[7914IUW]-/00Y8603, BIOS -[D7E128FUS-1.40]- 07/23/2013 0000000000000000 ffff881037babb58 ffffffff8139d885 0000000000000010 0000000000000000 0000000000000000 0000000000000000 ffff881037babba8 ffffffff8108505d ffff881037ba0000 000011cf3e7d6e60 0000000000000046 Call Trace: dump_stack+0x89/0xd4 __warn+0xfd/0x120 warn_slowpath_null+0x1d/0x20 rebind_workers+0x1c0/0x1d0 workqueue_cpu_up_callback+0xf5/0x1d0 notifier_call_chain+0x64/0x90 ? trace_hardirqs_on_caller+0xf2/0x220 ? notify_prepare+0x80/0x80 __raw_notifier_call_chain+0xe/0x10 __cpu_notify+0x35/0x50 notify_down_prepare+0x5e/0x80 ? notify_prepare+0x80/0x80 cpuhp_invoke_callback+0x73/0x330 ? __schedule+0x33e/0x8a0 cpuhp_down_callbacks+0x51/0xc0 cpuhp_thread_fun+0xc1/0xf0 smpboot_thread_fn+0x159/0x2a0 ? smpboot_create_threads+0x80/0x80 kthread+0xef/0x110 ? wait_for_completion+0xf0/0x120 ? schedule_tail+0x35/0xf0 ret_from_fork+0x22/0x50 ? __init_kthread_worker+0x70/0x70 ---[ end trace eb12ae47d2382d8f ]--- notify_down_prepare: attempt to take down CPU 0 failed This bug can be reproduced by below config w/ nohz_full= all cpus: CONFIG_BOOTPARAM_HOTPLUG_CPU0=y CONFIG_DEBUG_HOTPLUG_CPU0=y CONFIG_NO_HZ_FULL=y As Thomas pointed out: | If a down prepare callback fails, then DOWN_FAILED is invoked for all | callbacks which have successfully executed DOWN_PREPARE. | | But, workqueue has actually two notifiers. One which handles | UP/DOWN_FAILED/ONLINE and one which handles DOWN_PREPARE. | | Now look at the priorities of those callbacks: | | CPU_PRI_WORKQUEUE_UP = 5 | CPU_PRI_WORKQUEUE_DOWN = -5 | | So the call order on DOWN_PREPARE is: | | CB 1 | CB ... | CB workqueue_up() -> Ignores DOWN_PREPARE | CB ... | CB X ---> Fails | | So we call up to CB X with DOWN_FAILED | | CB 1 | CB ... | CB workqueue_up() -> Handles DOWN_FAILED | CB ... | CB X-1 | | So the problem is that the workqueue stuff handles DOWN_FAILED in the up | callback, while it should do it in the down callback. Which is not a good idea | either because it wants to be called early on rollback... | | Brilliant stuff, isn't it? The hotplug rework will solve this problem because | the callbacks become symetric, but for the existing mess, we need some | workaround in the workqueue code. The boot CPU handles housekeeping duty(unbound timers, workqueues, timekeeping, ...) on behalf of full dynticks CPUs. It must remain online when nohz full is enabled. There is a priority set to every notifier_blocks: workqueue_cpu_up > tick_nohz_cpu_down > workqueue_cpu_down So tick_nohz_cpu_down callback failed when down prepare cpu 0, and notifier_blocks behind tick_nohz_cpu_down will not be called any more, which leads to workers are actually not unbound. Then hotplug state machine will fallback to undo and online cpu 0 again. Workers will be rebound unconditionally even if they are not unbound and trigger the warning in this progress. This patch fix it by catching !DISASSOCIATED to avoid rebind bound workers. Cc: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Frédéric Weisbecker <fweisbec@gmail.com> Cc: stable@vger.kernel.org Suggested-by: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
2016-05-11 09:55:18 +00:00
pool->flags &= ~POOL_DISASSOCIATED;
for_each_pool_worker(worker, pool) {
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
unsigned int worker_flags = worker->flags;
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
/*
* We want to clear UNBOUND but can't directly call
* worker_clr_flags() or adjust nr_running. Atomically
* replace UNBOUND with another NOT_RUNNING flag REBOUND.
* @worker will clear REBOUND using worker_clr_flags() when
* it initiates the next execution cycle thus restoring
* concurrency management. Note that when or whether
* @worker clears REBOUND doesn't affect correctness.
*
locking/atomics, workqueue: Convert ACCESS_ONCE() to READ_ONCE()/WRITE_ONCE() For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't currently harmful. However, for some features it is necessary to instrument reads and writes separately, which is not possible with ACCESS_ONCE(). This distinction is critical to correct operation. It's possible to transform the bulk of kernel code using the Coccinelle script below. However, this doesn't handle comments, leaving references to ACCESS_ONCE() instances which have been removed. As a preparatory step, this patch converts the workqueue code and comments to use {READ,WRITE}_ONCE() consistently. ---- virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-12-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 21:07:22 +00:00
* WRITE_ONCE() is necessary because @worker->flags may be
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
* tested without holding any lock in
* wq_worker_running(). Without it, NOT_RUNNING test may
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
* fail incorrectly leading to premature concurrency
* management operations.
*/
WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
worker_flags |= WORKER_REBOUND;
worker_flags &= ~WORKER_UNBOUND;
locking/atomics, workqueue: Convert ACCESS_ONCE() to READ_ONCE()/WRITE_ONCE() For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't currently harmful. However, for some features it is necessary to instrument reads and writes separately, which is not possible with ACCESS_ONCE(). This distinction is critical to correct operation. It's possible to transform the bulk of kernel code using the Coccinelle script below. However, this doesn't handle comments, leaving references to ACCESS_ONCE() instances which have been removed. As a preparatory step, this patch converts the workqueue code and comments to use {READ,WRITE}_ONCE() consistently. ---- virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Acked-by: Tejun Heo <tj@kernel.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-12-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 21:07:22 +00:00
WRITE_ONCE(worker->flags, worker_flags);
}
workqueue: directly restore CPU affinity of workers from CPU_ONLINE Rebinding workers of a per-cpu pool after a CPU comes online involves a lot of back-and-forth mostly because only the task itself could adjust CPU affinity if PF_THREAD_BOUND was set. As CPU_ONLINE itself couldn't adjust affinity, it had to somehow coerce the workers themselves to perform set_cpus_allowed_ptr(). Due to the various states a worker can be in, this led to three different paths a worker may be rebound. worker->rebind_work is queued to busy workers. Idle ones are signaled by unlinking worker->entry and call idle_worker_rebind(). The manager isn't covered by either and implements its own mechanism. PF_THREAD_BOUND has been relaced with PF_NO_SETAFFINITY and CPU_ONLINE itself now can manipulate CPU affinity of workers. This patch replaces the existing rebind mechanism with direct one where CPU_ONLINE iterates over all workers using for_each_pool_worker(), restores CPU affinity, and clears WORKER_UNBOUND. There are a couple subtleties. All bound idle workers should have their runqueues set to that of the bound CPU; however, if the target task isn't running, set_cpus_allowed_ptr() just updates the cpus_allowed mask deferring the actual migration to when the task wakes up. This is worked around by waking up idle workers after restoring CPU affinity before any workers can become bound. Another subtlety is stems from matching @pool->nr_running with the number of running unbound workers. While DISASSOCIATED, all workers are unbound and nr_running is zero. As workers become bound again, nr_running needs to be adjusted accordingly; however, there is no good way to tell whether a given worker is running without poking into scheduler internals. Instead of clearing UNBOUND directly, rebind_workers() replaces UNBOUND with another new NOT_RUNNING flag - REBOUND, which will later be cleared by the workers themselves while preparing for the next round of work item execution. The only change needed for the workers is clearing REBOUND along with PREP. * This patch leaves for_each_busy_worker() without any user. Removed. * idle_worker_rebind(), busy_worker_rebind_fn(), worker->rebind_work and rebind logic in manager_workers() removed. * worker_thread() now looks at WORKER_DIE instead of testing whether @worker->entry is empty to determine whether it needs to do something special as dying is the only special thing now. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-19 20:45:21 +00:00
raw_spin_unlock_irq(&pool->lock);
}
/**
* restore_unbound_workers_cpumask - restore cpumask of unbound workers
* @pool: unbound pool of interest
* @cpu: the CPU which is coming up
*
* An unbound pool may end up with a cpumask which doesn't have any online
* CPUs. When a worker of such pool get scheduled, the scheduler resets
* its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
* online CPU before, cpus_allowed of all its workers should be restored.
*/
static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
{
static cpumask_t cpumask;
struct worker *worker;
lockdep_assert_held(&wq_pool_attach_mutex);
/* is @cpu allowed for @pool? */
if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
return;
cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
/* as we're called from CPU_ONLINE, the following shouldn't fail */
for_each_pool_worker(worker, pool)
workqueue: Fix setting affinity of unbound worker threads With commit e9d867a67fd03ccc ("sched: Allow per-cpu kernel threads to run on online && !active"), __set_cpus_allowed_ptr() expects that only strict per-cpu kernel threads can have affinity to an online CPU which is not yet active. This assumption is currently broken in the CPU_ONLINE notification handler for the workqueues where restore_unbound_workers_cpumask() calls set_cpus_allowed_ptr() when the first cpu in the unbound worker's pool->attr->cpumask comes online. Since set_cpus_allowed_ptr() is called with pool->attr->cpumask in which only one CPU is online which is not yet active, we get the following WARN_ON during an CPU online operation. ------------[ cut here ]------------ WARNING: CPU: 40 PID: 248 at kernel/sched/core.c:1166 __set_cpus_allowed_ptr+0x228/0x2e0 Modules linked in: CPU: 40 PID: 248 Comm: cpuhp/40 Not tainted 4.6.0-autotest+ #4 <..snip..> Call Trace: [c000000f273ff920] [c00000000010493c] __set_cpus_allowed_ptr+0x2cc/0x2e0 (unreliable) [c000000f273ffac0] [c0000000000ed4b0] workqueue_cpu_up_callback+0x2c0/0x470 [c000000f273ffb70] [c0000000000f5c58] notifier_call_chain+0x98/0x100 [c000000f273ffbc0] [c0000000000c5ed0] __cpu_notify+0x70/0xe0 [c000000f273ffc00] [c0000000000c6028] notify_online+0x38/0x50 [c000000f273ffc30] [c0000000000c5214] cpuhp_invoke_callback+0x84/0x250 [c000000f273ffc90] [c0000000000c562c] cpuhp_up_callbacks+0x5c/0x120 [c000000f273ffce0] [c0000000000c64d4] cpuhp_thread_fun+0x184/0x1c0 [c000000f273ffd20] [c0000000000fa050] smpboot_thread_fn+0x290/0x2a0 [c000000f273ffd80] [c0000000000f45b0] kthread+0x110/0x130 [c000000f273ffe30] [c000000000009570] ret_from_kernel_thread+0x5c/0x6c ---[ end trace 00f1456578b2a3b2 ]--- This patch fixes this by limiting the mask to the intersection of the pool affinity and online CPUs. Changelog-cribbed-from: Gautham R. Shenoy <ego@linux.vnet.ibm.com> Reported-by: Abdul Haleem <abdhalee@linux.vnet.ibm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2016-06-16 12:38:42 +00:00
WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
}
int workqueue_prepare_cpu(unsigned int cpu)
{
struct worker_pool *pool;
for_each_cpu_worker_pool(pool, cpu) {
if (pool->nr_workers)
continue;
if (!create_worker(pool))
return -ENOMEM;
}
return 0;
}
int workqueue_online_cpu(unsigned int cpu)
{
struct worker_pool *pool;
workqueue: implement NUMA affinity for unbound workqueues Currently, an unbound workqueue has single current, or first, pwq (pool_workqueue) to which all new work items are queued. This often isn't optimal on NUMA machines as workers may jump around across node boundaries and work items get assigned to workers without any regard to NUMA affinity. This patch implements NUMA affinity for unbound workqueues. Instead of mapping all entries of numa_pwq_tbl[] to the same pwq, apply_workqueue_attrs() now creates a separate pwq covering the intersecting CPUs for each NUMA node which has online CPUs in @attrs->cpumask. Nodes which don't have intersecting possible CPUs are mapped to pwqs covering whole @attrs->cpumask. As CPUs come up and go down, the pool association is changed accordingly. Changing pool association may involve allocating new pools which may fail. To avoid failing CPU_DOWN, each workqueue always keeps a default pwq which covers whole attrs->cpumask which is used as fallback if pool creation fails during a CPU hotplug operation. This ensures that all work items issued on a NUMA node is executed on the same node as long as the workqueue allows execution on the CPUs of the node. As this maps a workqueue to multiple pwqs and max_active is per-pwq, this change the behavior of max_active. The limit is now per NUMA node instead of global. While this is an actual change, max_active is already per-cpu for per-cpu workqueues and primarily used as safety mechanism rather than for active concurrency control. Concurrency is usually limited from workqueue users by the number of concurrently active work items and this change shouldn't matter much. v2: Fixed pwq freeing in apply_workqueue_attrs() error path. Spotted by Lai. v3: The previous version incorrectly made a workqueue spanning multiple nodes spread work items over all online CPUs when some of its nodes don't have any desired cpus. Reimplemented so that NUMA affinity is properly updated as CPUs go up and down. This problem was spotted by Lai Jiangshan. v4: destroy_workqueue() was putting wq->dfl_pwq and then clearing it; however, wq may be freed at any time after dfl_pwq is put making the clearing use-after-free. Clear wq->dfl_pwq before putting it. v5: apply_workqueue_attrs() was leaking @tmp_attrs, @new_attrs and @pwq_tbl after success. Fixed. Retry loop in wq_update_unbound_numa_attrs() isn't necessary as application of new attrs is excluded via CPU hotplug. Removed. Documentation on CPU affinity guarantee on CPU_DOWN added. All changes are suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-04-01 18:23:36 +00:00
struct workqueue_struct *wq;
int pi;
mutex_lock(&wq_pool_mutex);
for_each_pool(pool, pi) {
mutex_lock(&wq_pool_attach_mutex);
if (pool->cpu == cpu)
rebind_workers(pool);
else if (pool->cpu < 0)
restore_unbound_workers_cpumask(pool, cpu);
mutex_unlock(&wq_pool_attach_mutex);
}
/* update pod affinity of unbound workqueues */
list_for_each_entry(wq, &workqueues, list) {
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
struct workqueue_attrs *attrs = wq->unbound_attrs;
if (attrs) {
const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
int tcpu;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
wq_update_pod(wq, tcpu, cpu, true);
}
}
mutex_unlock(&wq_pool_mutex);
return 0;
}
int workqueue_offline_cpu(unsigned int cpu)
{
struct workqueue_struct *wq;
/* unbinding per-cpu workers should happen on the local CPU */
if (WARN_ON(cpu != smp_processor_id()))
return -1;
unbind_workers(cpu);
/* update pod affinity of unbound workqueues */
mutex_lock(&wq_pool_mutex);
list_for_each_entry(wq, &workqueues, list) {
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
struct workqueue_attrs *attrs = wq->unbound_attrs;
if (attrs) {
const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
int tcpu;
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
wq_update_pod(wq, tcpu, cpu, false);
}
}
mutex_unlock(&wq_pool_mutex);
return 0;
}
struct work_for_cpu {
struct work_struct work;
long (*fn)(void *);
void *arg;
long ret;
};
static void work_for_cpu_fn(struct work_struct *work)
{
struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
wfc->ret = wfc->fn(wfc->arg);
}
/**
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
* work_on_cpu_key - run a function in thread context on a particular cpu
* @cpu: the cpu to run on
* @fn: the function to run
* @arg: the function arg
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
* @key: The lock class key for lock debugging purposes
*
* It is up to the caller to ensure that the cpu doesn't go offline.
* The caller must not hold any locks which would prevent @fn from completing.
*
* Return: The value @fn returns.
*/
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
long work_on_cpu_key(int cpu, long (*fn)(void *),
void *arg, struct lock_class_key *key)
{
struct work_for_cpu wfc = { .fn = fn, .arg = arg };
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
schedule_work_on(cpu, &wfc.work);
flush_work(&wfc.work);
destroy_work_on_stack(&wfc.work);
return wfc.ret;
}
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
EXPORT_SYMBOL_GPL(work_on_cpu_key);
/**
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
* work_on_cpu_safe_key - run a function in thread context on a particular cpu
* @cpu: the cpu to run on
* @fn: the function to run
* @arg: the function argument
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
* @key: The lock class key for lock debugging purposes
*
* Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
* any locks which would prevent @fn from completing.
*
* Return: The value @fn returns.
*/
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
void *arg, struct lock_class_key *key)
{
long ret = -ENODEV;
cpus_read_lock();
if (cpu_online(cpu))
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
ret = work_on_cpu_key(cpu, fn, arg, key);
cpus_read_unlock();
return ret;
}
workqueue: Provide one lock class key per work_on_cpu() callsite All callers of work_on_cpu() share the same lock class key for all the functions queued. As a result the workqueue related locking scenario for a function A may be spuriously accounted as an inversion against the locking scenario of function B such as in the following model: long A(void *arg) { mutex_lock(&mutex); mutex_unlock(&mutex); } long B(void *arg) { } void launchA(void) { work_on_cpu(0, A, NULL); } void launchB(void) { mutex_lock(&mutex); work_on_cpu(1, B, NULL); mutex_unlock(&mutex); } launchA and launchB running concurrently have no chance to deadlock. However the above can be reported by lockdep as a possible locking inversion because the works containing A() and B() are treated as belonging to the same locking class. The following shows an existing example of such a spurious lockdep splat: ====================================================== WARNING: possible circular locking dependency detected 6.6.0-rc1-00065-g934ebd6e5359 #35409 Not tainted ------------------------------------------------------ kworker/0:1/9 is trying to acquire lock: ffffffff9bc72f30 (cpu_hotplug_lock){++++}-{0:0}, at: _cpu_down+0x57/0x2b0 but task is already holding lock: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #2 ((work_completion)(&wfc.work)){+.+.}-{0:0}: __flush_work+0x83/0x4e0 work_on_cpu+0x97/0xc0 rcu_nocb_cpu_offload+0x62/0xb0 rcu_nocb_toggle+0xd0/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #1 (rcu_state.barrier_mutex){+.+.}-{3:3}: __mutex_lock+0x81/0xc80 rcu_nocb_cpu_deoffload+0x38/0xb0 rcu_nocb_toggle+0x144/0x1d0 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 -> #0 (cpu_hotplug_lock){++++}-{0:0}: __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 percpu_down_write+0x31/0x200 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 kthread+0xe6/0x120 ret_from_fork+0x2f/0x40 ret_from_fork_asm+0x1b/0x30 other info that might help us debug this: Chain exists of: cpu_hotplug_lock --> rcu_state.barrier_mutex --> (work_completion)(&wfc.work) Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock((work_completion)(&wfc.work)); lock(rcu_state.barrier_mutex); lock((work_completion)(&wfc.work)); lock(cpu_hotplug_lock); *** DEADLOCK *** 2 locks held by kworker/0:1/9: #0: ffff900481068b38 ((wq_completion)events){+.+.}-{0:0}, at: process_scheduled_works+0x212/0x500 #1: ffff9e3bc0057e60 ((work_completion)(&wfc.work)){+.+.}-{0:0}, at: process_scheduled_works+0x216/0x500 stack backtrace: CPU: 0 PID: 9 Comm: kworker/0:1 Not tainted 6.6.0-rc1-00065-g934ebd6e5359 #35409 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014 Workqueue: events work_for_cpu_fn Call Trace: rcu-torture: rcu_torture_read_exit: Start of episode <TASK> dump_stack_lvl+0x4a/0x80 check_noncircular+0x132/0x150 __lock_acquire+0x1538/0x2500 lock_acquire+0xbf/0x2a0 ? _cpu_down+0x57/0x2b0 percpu_down_write+0x31/0x200 ? _cpu_down+0x57/0x2b0 _cpu_down+0x57/0x2b0 __cpu_down_maps_locked+0x10/0x20 work_for_cpu_fn+0x15/0x20 process_scheduled_works+0x2a7/0x500 worker_thread+0x173/0x330 ? __pfx_worker_thread+0x10/0x10 kthread+0xe6/0x120 ? __pfx_kthread+0x10/0x10 ret_from_fork+0x2f/0x40 ? __pfx_kthread+0x10/0x10 ret_from_fork_asm+0x1b/0x30 </TASK Fix this with providing one lock class key per work_on_cpu() caller. Reported-and-tested-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-09-24 15:07:02 +00:00
EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
#endif /* CONFIG_SMP */
#ifdef CONFIG_FREEZER
/**
* freeze_workqueues_begin - begin freezing workqueues
*
* Start freezing workqueues. After this function returns, all freezable
* workqueues will queue new works to their inactive_works list instead of
* pool->worklist.
*
* CONTEXT:
* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
*/
void freeze_workqueues_begin(void)
{
struct workqueue_struct *wq;
struct pool_workqueue *pwq;
mutex_lock(&wq_pool_mutex);
WARN_ON_ONCE(workqueue_freezing);
workqueue_freezing = true;
list_for_each_entry(wq, &workqueues, list) {
mutex_lock(&wq->mutex);
for_each_pwq(pwq, wq)
pwq_adjust_max_active(pwq);
mutex_unlock(&wq->mutex);
}
mutex_unlock(&wq_pool_mutex);
}
/**
* freeze_workqueues_busy - are freezable workqueues still busy?
*
* Check whether freezing is complete. This function must be called
* between freeze_workqueues_begin() and thaw_workqueues().
*
* CONTEXT:
* Grabs and releases wq_pool_mutex.
*
* Return:
* %true if some freezable workqueues are still busy. %false if freezing
* is complete.
*/
bool freeze_workqueues_busy(void)
{
bool busy = false;
struct workqueue_struct *wq;
struct pool_workqueue *pwq;
mutex_lock(&wq_pool_mutex);
WARN_ON_ONCE(!workqueue_freezing);
list_for_each_entry(wq, &workqueues, list) {
if (!(wq->flags & WQ_FREEZABLE))
continue;
/*
* nr_active is monotonically decreasing. It's safe
* to peek without lock.
*/
rcu_read_lock();
for_each_pwq(pwq, wq) {
WARN_ON_ONCE(pwq->nr_active < 0);
if (pwq->nr_active) {
busy = true;
rcu_read_unlock();
goto out_unlock;
}
}
rcu_read_unlock();
}
out_unlock:
mutex_unlock(&wq_pool_mutex);
return busy;
}
/**
* thaw_workqueues - thaw workqueues
*
* Thaw workqueues. Normal queueing is restored and all collected
* frozen works are transferred to their respective pool worklists.
*
* CONTEXT:
* Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
*/
void thaw_workqueues(void)
{
struct workqueue_struct *wq;
struct pool_workqueue *pwq;
mutex_lock(&wq_pool_mutex);
if (!workqueue_freezing)
goto out_unlock;
workqueue_freezing = false;
/* restore max_active and repopulate worklist */
list_for_each_entry(wq, &workqueues, list) {
mutex_lock(&wq->mutex);
for_each_pwq(pwq, wq)
pwq_adjust_max_active(pwq);
mutex_unlock(&wq->mutex);
}
out_unlock:
mutex_unlock(&wq_pool_mutex);
}
#endif /* CONFIG_FREEZER */
static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
{
LIST_HEAD(ctxs);
int ret = 0;
struct workqueue_struct *wq;
struct apply_wqattrs_ctx *ctx, *n;
lockdep_assert_held(&wq_pool_mutex);
list_for_each_entry(wq, &workqueues, list) {
if (!(wq->flags & WQ_UNBOUND))
continue;
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
/* creating multiple pwqs breaks ordering guarantee */
if (!list_empty(&wq->pwqs)) {
if (wq->flags & __WQ_ORDERED_EXPLICIT)
continue;
wq->flags &= ~__WQ_ORDERED;
}
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
if (IS_ERR(ctx)) {
ret = PTR_ERR(ctx);
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
break;
}
list_add_tail(&ctx->list, &ctxs);
}
list_for_each_entry_safe(ctx, n, &ctxs, list) {
if (!ret)
apply_wqattrs_commit(ctx);
apply_wqattrs_cleanup(ctx);
}
if (!ret) {
mutex_lock(&wq_pool_attach_mutex);
cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
mutex_unlock(&wq_pool_attach_mutex);
}
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
return ret;
}
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
/**
* workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
* @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
*
* This function can be called from cpuset code to provide a set of isolated
* CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
* either cpus_read_lock or cpus_write_lock.
*/
int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
{
cpumask_var_t cpumask;
int ret = 0;
if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
return -ENOMEM;
lockdep_assert_cpus_held();
mutex_lock(&wq_pool_mutex);
/* Save the current isolated cpumask & export it via sysfs */
cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
/*
* If the operation fails, it will fall back to
* wq_requested_unbound_cpumask which is initially set to
* (HK_TYPE_WQ HK_TYPE_DOMAIN) house keeping mask and rewritten
* by any subsequent write to workqueue/cpumask sysfs file.
*/
if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
cpumask_copy(cpumask, wq_requested_unbound_cpumask);
if (!cpumask_equal(cpumask, wq_unbound_cpumask))
ret = workqueue_apply_unbound_cpumask(cpumask);
mutex_unlock(&wq_pool_mutex);
free_cpumask_var(cpumask);
return ret;
}
static int parse_affn_scope(const char *val)
{
int i;
for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
return i;
}
return -EINVAL;
}
static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
{
struct workqueue_struct *wq;
int affn, cpu;
affn = parse_affn_scope(val);
if (affn < 0)
return affn;
if (affn == WQ_AFFN_DFL)
return -EINVAL;
cpus_read_lock();
mutex_lock(&wq_pool_mutex);
wq_affn_dfl = affn;
list_for_each_entry(wq, &workqueues, list) {
for_each_online_cpu(cpu) {
wq_update_pod(wq, cpu, cpu, true);
}
}
mutex_unlock(&wq_pool_mutex);
cpus_read_unlock();
return 0;
}
static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
{
return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
}
static const struct kernel_param_ops wq_affn_dfl_ops = {
.set = wq_affn_dfl_set,
.get = wq_affn_dfl_get,
};
module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
#ifdef CONFIG_SYSFS
/*
* Workqueues with WQ_SYSFS flag set is visible to userland via
* /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
* following attributes.
*
* per_cpu RO bool : whether the workqueue is per-cpu or unbound
* max_active RW int : maximum number of in-flight work items
*
* Unbound workqueues have the following extra attributes.
*
* nice RW int : nice value of the workers
* cpumask RW mask : bitmask of allowed CPUs for the workers
* affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
* affinity_strict RW bool : worker CPU affinity is strict
*/
struct wq_device {
struct workqueue_struct *wq;
struct device dev;
};
static struct workqueue_struct *dev_to_wq(struct device *dev)
{
struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
return wq_dev->wq;
}
static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
}
static DEVICE_ATTR_RO(per_cpu);
static ssize_t max_active_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
}
static ssize_t max_active_store(struct device *dev,
struct device_attribute *attr, const char *buf,
size_t count)
{
struct workqueue_struct *wq = dev_to_wq(dev);
int val;
if (sscanf(buf, "%d", &val) != 1 || val <= 0)
return -EINVAL;
workqueue_set_max_active(wq, val);
return count;
}
static DEVICE_ATTR_RW(max_active);
static struct attribute *wq_sysfs_attrs[] = {
&dev_attr_per_cpu.attr,
&dev_attr_max_active.attr,
NULL,
};
ATTRIBUTE_GROUPS(wq_sysfs);
static void apply_wqattrs_lock(void)
{
/* CPUs should stay stable across pwq creations and installations */
cpus_read_lock();
mutex_lock(&wq_pool_mutex);
}
static void apply_wqattrs_unlock(void)
{
mutex_unlock(&wq_pool_mutex);
cpus_read_unlock();
}
static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
int written;
mutex_lock(&wq->mutex);
written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
mutex_unlock(&wq->mutex);
return written;
}
/* prepare workqueue_attrs for sysfs store operations */
static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
{
struct workqueue_attrs *attrs;
lockdep_assert_held(&wq_pool_mutex);
attrs = alloc_workqueue_attrs();
if (!attrs)
return NULL;
copy_workqueue_attrs(attrs, wq->unbound_attrs);
return attrs;
}
static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
const char *buf, size_t count)
{
struct workqueue_struct *wq = dev_to_wq(dev);
struct workqueue_attrs *attrs;
int ret = -ENOMEM;
apply_wqattrs_lock();
attrs = wq_sysfs_prep_attrs(wq);
if (!attrs)
goto out_unlock;
if (sscanf(buf, "%d", &attrs->nice) == 1 &&
attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
ret = apply_workqueue_attrs_locked(wq, attrs);
else
ret = -EINVAL;
out_unlock:
apply_wqattrs_unlock();
free_workqueue_attrs(attrs);
return ret ?: count;
}
static ssize_t wq_cpumask_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
int written;
mutex_lock(&wq->mutex);
written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
cpumask_pr_args(wq->unbound_attrs->cpumask));
mutex_unlock(&wq->mutex);
return written;
}
static ssize_t wq_cpumask_store(struct device *dev,
struct device_attribute *attr,
const char *buf, size_t count)
{
struct workqueue_struct *wq = dev_to_wq(dev);
struct workqueue_attrs *attrs;
int ret = -ENOMEM;
apply_wqattrs_lock();
attrs = wq_sysfs_prep_attrs(wq);
if (!attrs)
goto out_unlock;
ret = cpumask_parse(buf, attrs->cpumask);
if (!ret)
ret = apply_workqueue_attrs_locked(wq, attrs);
out_unlock:
apply_wqattrs_unlock();
free_workqueue_attrs(attrs);
return ret ?: count;
}
static ssize_t wq_affn_scope_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
int written;
mutex_lock(&wq->mutex);
if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
wq_affn_names[WQ_AFFN_DFL],
wq_affn_names[wq_affn_dfl]);
else
written = scnprintf(buf, PAGE_SIZE, "%s\n",
wq_affn_names[wq->unbound_attrs->affn_scope]);
mutex_unlock(&wq->mutex);
return written;
}
static ssize_t wq_affn_scope_store(struct device *dev,
struct device_attribute *attr,
const char *buf, size_t count)
{
struct workqueue_struct *wq = dev_to_wq(dev);
struct workqueue_attrs *attrs;
int affn, ret = -ENOMEM;
affn = parse_affn_scope(buf);
if (affn < 0)
return affn;
apply_wqattrs_lock();
attrs = wq_sysfs_prep_attrs(wq);
if (attrs) {
attrs->affn_scope = affn;
ret = apply_workqueue_attrs_locked(wq, attrs);
}
apply_wqattrs_unlock();
free_workqueue_attrs(attrs);
return ret ?: count;
}
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
static ssize_t wq_affinity_strict_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
struct workqueue_struct *wq = dev_to_wq(dev);
return scnprintf(buf, PAGE_SIZE, "%d\n",
wq->unbound_attrs->affn_strict);
}
static ssize_t wq_affinity_strict_store(struct device *dev,
struct device_attribute *attr,
const char *buf, size_t count)
{
struct workqueue_struct *wq = dev_to_wq(dev);
struct workqueue_attrs *attrs;
int v, ret = -ENOMEM;
if (sscanf(buf, "%d", &v) != 1)
return -EINVAL;
apply_wqattrs_lock();
attrs = wq_sysfs_prep_attrs(wq);
if (attrs) {
attrs->affn_strict = (bool)v;
ret = apply_workqueue_attrs_locked(wq, attrs);
}
apply_wqattrs_unlock();
free_workqueue_attrs(attrs);
return ret ?: count;
}
static struct device_attribute wq_sysfs_unbound_attrs[] = {
__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
__ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
__ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
__ATTR_NULL,
};
static struct bus_type wq_subsys = {
.name = "workqueue",
.dev_groups = wq_sysfs_groups,
};
/**
* workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
* @cpumask: the cpumask to set
*
* The low-level workqueues cpumask is a global cpumask that limits
* the affinity of all unbound workqueues. This function check the @cpumask
* and apply it to all unbound workqueues and updates all pwqs of them.
*
* Return: 0 - Success
* -EINVAL - Invalid @cpumask
* -ENOMEM - Failed to allocate memory for attrs or pwqs.
*/
static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
{
int ret = -EINVAL;
/*
* Not excluding isolated cpus on purpose.
* If the user wishes to include them, we allow that.
*/
cpumask_and(cpumask, cpumask, cpu_possible_mask);
if (!cpumask_empty(cpumask)) {
apply_wqattrs_lock();
cpumask_copy(wq_requested_unbound_cpumask, cpumask);
if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
ret = 0;
goto out_unlock;
}
ret = workqueue_apply_unbound_cpumask(cpumask);
out_unlock:
apply_wqattrs_unlock();
}
return ret;
}
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
static ssize_t __wq_cpumask_show(struct device *dev,
struct device_attribute *attr, char *buf, cpumask_var_t mask)
{
int written;
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
mutex_lock(&wq_pool_mutex);
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
mutex_unlock(&wq_pool_mutex);
return written;
}
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
static ssize_t wq_unbound_cpumask_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
}
static ssize_t wq_requested_cpumask_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
}
static ssize_t wq_isolated_cpumask_show(struct device *dev,
struct device_attribute *attr, char *buf)
{
return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
}
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
static ssize_t wq_unbound_cpumask_store(struct device *dev,
struct device_attribute *attr, const char *buf, size_t count)
{
cpumask_var_t cpumask;
int ret;
if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
return -ENOMEM;
ret = cpumask_parse(buf, cpumask);
if (!ret)
ret = workqueue_set_unbound_cpumask(cpumask);
free_cpumask_var(cpumask);
return ret ? ret : count;
}
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
static struct device_attribute wq_sysfs_cpumask_attrs[] = {
workqueue: Allow modifying low level unbound workqueue cpumask Allow to modify the low-level unbound workqueues cpumask through sysfs. This is performed by traversing the entire workqueue list and calling apply_wqattrs_prepare() on the unbound workqueues with the new low level mask. Only after all the preparation are done, we commit them all together. Ordered workqueues are ignored from the low level unbound workqueue cpumask, it will be handled in near future. All the (default & per-node) pwqs are mandatorily controlled by the low level cpumask. If the user configured cpumask doesn't overlap with the low level cpumask, the low level cpumask will be used for the wq instead. The comment of wq_calc_node_cpumask() is updated and explicitly requires that its first argument should be the attrs of the default pwq. The default wq_unbound_cpumask is cpu_possible_mask. The workqueue subsystem doesn't know its best default value, let the system manager or the other subsystem set it when needed. Changed from V8: merge the calculating code for the attrs of the default pwq together. minor change the code&comments for saving the user configured attrs. remove unnecessary list_del(). minor update the comment of wq_calc_node_cpumask(). update the comment of workqueue_set_unbound_cpumask(); Cc: Christoph Lameter <cl@linux.com> Cc: Kevin Hilman <khilman@linaro.org> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Tejun Heo <tj@kernel.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Original-patch-by: Frederic Weisbecker <fweisbec@gmail.com> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2015-04-30 09:16:12 +00:00
__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
wq_unbound_cpumask_store),
__ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
__ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
__ATTR_NULL,
};
static int __init wq_sysfs_init(void)
{
struct device *dev_root;
int err;
err = subsys_virtual_register(&wq_subsys, NULL);
if (err)
return err;
dev_root = bus_get_dev_root(&wq_subsys);
if (dev_root) {
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
struct device_attribute *attr;
for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
err = device_create_file(dev_root, attr);
if (err)
break;
}
put_device(dev_root);
}
return err;
}
core_initcall(wq_sysfs_init);
static void wq_device_release(struct device *dev)
{
struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
kfree(wq_dev);
}
/**
* workqueue_sysfs_register - make a workqueue visible in sysfs
* @wq: the workqueue to register
*
* Expose @wq in sysfs under /sys/bus/workqueue/devices.
* alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
* which is the preferred method.
*
* Workqueue user should use this function directly iff it wants to apply
* workqueue_attrs before making the workqueue visible in sysfs; otherwise,
* apply_workqueue_attrs() may race against userland updating the
* attributes.
*
* Return: 0 on success, -errno on failure.
*/
int workqueue_sysfs_register(struct workqueue_struct *wq)
{
struct wq_device *wq_dev;
int ret;
/*
* Adjusting max_active or creating new pwqs by applying
* attributes breaks ordering guarantee. Disallow exposing ordered
* workqueues.
*/
if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
return -EINVAL;
wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
if (!wq_dev)
return -ENOMEM;
wq_dev->wq = wq;
wq_dev->dev.bus = &wq_subsys;
wq_dev->dev.release = wq_device_release;
dev_set_name(&wq_dev->dev, "%s", wq->name);
/*
* unbound_attrs are created separately. Suppress uevent until
* everything is ready.
*/
dev_set_uevent_suppress(&wq_dev->dev, true);
ret = device_register(&wq_dev->dev);
if (ret) {
put_device(&wq_dev->dev);
wq->wq_dev = NULL;
return ret;
}
if (wq->flags & WQ_UNBOUND) {
struct device_attribute *attr;
for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
ret = device_create_file(&wq_dev->dev, attr);
if (ret) {
device_unregister(&wq_dev->dev);
wq->wq_dev = NULL;
return ret;
}
}
}
dev_set_uevent_suppress(&wq_dev->dev, false);
kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
return 0;
}
/**
* workqueue_sysfs_unregister - undo workqueue_sysfs_register()
* @wq: the workqueue to unregister
*
* If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
*/
static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
{
struct wq_device *wq_dev = wq->wq_dev;
if (!wq->wq_dev)
return;
wq->wq_dev = NULL;
device_unregister(&wq_dev->dev);
}
#else /* CONFIG_SYSFS */
static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
#endif /* CONFIG_SYSFS */
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
/*
* Workqueue watchdog.
*
* Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
* flush dependency, a concurrency managed work item which stays RUNNING
* indefinitely. Workqueue stalls can be very difficult to debug as the
* usual warning mechanisms don't trigger and internal workqueue state is
* largely opaque.
*
* Workqueue watchdog monitors all worker pools periodically and dumps
* state if some pools failed to make forward progress for a while where
* forward progress is defined as the first item on ->worklist changing.
*
* This mechanism is controlled through the kernel parameter
* "workqueue.watchdog_thresh" which can be updated at runtime through the
* corresponding sysfs parameter file.
*/
#ifdef CONFIG_WQ_WATCHDOG
static unsigned long wq_watchdog_thresh = 30;
timer: Remove users of TIMER_DEFERRED_INITIALIZER This removes uses of TIMER_DEFERRED_INITIALIZER and chooses a location to call timer_setup() from before add_timer() or mod_timer() is called. Adjusts callbacks to use from_timer() as needed. Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mips@linux-mips.org Cc: Petr Mladek <pmladek@suse.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Sebastian Reichel <sre@kernel.org> Cc: Kalle Valo <kvalo@qca.qualcomm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Pavel Machek <pavel@ucw.cz> Cc: linux1394-devel@lists.sourceforge.net Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: linux-s390@vger.kernel.org Cc: linux-wireless@vger.kernel.org Cc: "James E.J. Bottomley" <jejb@linux.vnet.ibm.com> Cc: Wim Van Sebroeck <wim@iguana.be> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Ursula Braun <ubraun@linux.vnet.ibm.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Harish Patil <harish.patil@cavium.com> Cc: Stephen Boyd <sboyd@codeaurora.org> Cc: Guenter Roeck <linux@roeck-us.net> Cc: Manish Chopra <manish.chopra@cavium.com> Cc: Len Brown <len.brown@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: linux-pm@vger.kernel.org Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Julian Wiedmann <jwi@linux.vnet.ibm.com> Cc: John Stultz <john.stultz@linaro.org> Cc: Mark Gross <mark.gross@intel.com> Cc: linux-watchdog@vger.kernel.org Cc: linux-scsi@vger.kernel.org Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Stefan Richter <stefanr@s5r6.in-berlin.de> Cc: Michael Reed <mdr@sgi.com> Cc: netdev@vger.kernel.org Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linuxppc-dev@lists.ozlabs.org Cc: Sudip Mukherjee <sudipm.mukherjee@gmail.com> Link: https://lkml.kernel.org/r/1507159627-127660-7-git-send-email-keescook@chromium.org
2017-10-04 23:27:00 +00:00
static struct timer_list wq_watchdog_timer;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
/*
* Show workers that might prevent the processing of pending work items.
* The only candidates are CPU-bound workers in the running state.
* Pending work items should be handled by another idle worker
* in all other situations.
*/
static void show_cpu_pool_hog(struct worker_pool *pool)
{
struct worker *worker;
unsigned long flags;
int bkt;
raw_spin_lock_irqsave(&pool->lock, flags);
hash_for_each(pool->busy_hash, bkt, worker, hentry) {
if (task_is_running(worker->task)) {
/*
* Defer printing to avoid deadlocks in console
* drivers that queue work while holding locks
* also taken in their write paths.
*/
printk_deferred_enter();
pr_info("pool %d:\n", pool->id);
sched_show_task(worker->task);
printk_deferred_exit();
}
}
raw_spin_unlock_irqrestore(&pool->lock, flags);
}
static void show_cpu_pools_hogs(void)
{
struct worker_pool *pool;
int pi;
pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
rcu_read_lock();
for_each_pool(pool, pi) {
if (pool->cpu_stall)
show_cpu_pool_hog(pool);
}
rcu_read_unlock();
}
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
static void wq_watchdog_reset_touched(void)
{
int cpu;
wq_watchdog_touched = jiffies;
for_each_possible_cpu(cpu)
per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
}
timer: Remove users of TIMER_DEFERRED_INITIALIZER This removes uses of TIMER_DEFERRED_INITIALIZER and chooses a location to call timer_setup() from before add_timer() or mod_timer() is called. Adjusts callbacks to use from_timer() as needed. Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mips@linux-mips.org Cc: Petr Mladek <pmladek@suse.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Sebastian Reichel <sre@kernel.org> Cc: Kalle Valo <kvalo@qca.qualcomm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Pavel Machek <pavel@ucw.cz> Cc: linux1394-devel@lists.sourceforge.net Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: linux-s390@vger.kernel.org Cc: linux-wireless@vger.kernel.org Cc: "James E.J. Bottomley" <jejb@linux.vnet.ibm.com> Cc: Wim Van Sebroeck <wim@iguana.be> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Ursula Braun <ubraun@linux.vnet.ibm.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Harish Patil <harish.patil@cavium.com> Cc: Stephen Boyd <sboyd@codeaurora.org> Cc: Guenter Roeck <linux@roeck-us.net> Cc: Manish Chopra <manish.chopra@cavium.com> Cc: Len Brown <len.brown@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: linux-pm@vger.kernel.org Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Julian Wiedmann <jwi@linux.vnet.ibm.com> Cc: John Stultz <john.stultz@linaro.org> Cc: Mark Gross <mark.gross@intel.com> Cc: linux-watchdog@vger.kernel.org Cc: linux-scsi@vger.kernel.org Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Stefan Richter <stefanr@s5r6.in-berlin.de> Cc: Michael Reed <mdr@sgi.com> Cc: netdev@vger.kernel.org Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linuxppc-dev@lists.ozlabs.org Cc: Sudip Mukherjee <sudipm.mukherjee@gmail.com> Link: https://lkml.kernel.org/r/1507159627-127660-7-git-send-email-keescook@chromium.org
2017-10-04 23:27:00 +00:00
static void wq_watchdog_timer_fn(struct timer_list *unused)
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
{
unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
bool lockup_detected = false;
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
bool cpu_pool_stall = false;
unsigned long now = jiffies;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
struct worker_pool *pool;
int pi;
if (!thresh)
return;
rcu_read_lock();
for_each_pool(pool, pi) {
unsigned long pool_ts, touched, ts;
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
pool->cpu_stall = false;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
if (list_empty(&pool->worklist))
continue;
/*
* If a virtual machine is stopped by the host it can look to
* the watchdog like a stall.
*/
kvm_check_and_clear_guest_paused();
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
/* get the latest of pool and touched timestamps */
workqueue/watchdog: Make unbound workqueues aware of touch_softlockup_watchdog() 84;0;0c84;0;0c There are two workqueue-specific watchdog timestamps: + @wq_watchdog_touched_cpu (per-CPU) updated by touch_softlockup_watchdog() + @wq_watchdog_touched (global) updated by touch_all_softlockup_watchdogs() watchdog_timer_fn() checks only the global @wq_watchdog_touched for unbound workqueues. As a result, unbound workqueues are not aware of touch_softlockup_watchdog(). The watchdog might report a stall even when the unbound workqueues are blocked by a known slow code. Solution: touch_softlockup_watchdog() must touch also the global @wq_watchdog_touched timestamp. The global timestamp can no longer be used for bound workqueues because it is now updated from all CPUs. Instead, bound workqueues have to check only @wq_watchdog_touched_cpu and these timestamps have to be updated for all CPUs in touch_all_softlockup_watchdogs(). Beware: The change might cause the opposite problem. An unbound workqueue might get blocked on CPU A because of a real softlockup. The workqueue watchdog would miss it when the timestamp got touched on CPU B. It is acceptable because softlockups are detected by softlockup watchdog. The workqueue watchdog is there to detect stalls where a work never finishes, for example, because of dependencies of works queued into the same workqueue. V3: - Modify the commit message clearly according to Petr's suggestion. Signed-off-by: Wang Qing <wangqing@vivo.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-03-24 11:40:29 +00:00
if (pool->cpu >= 0)
touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
else
touched = READ_ONCE(wq_watchdog_touched);
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
pool_ts = READ_ONCE(pool->watchdog_ts);
if (time_after(pool_ts, touched))
ts = pool_ts;
else
ts = touched;
/* did we stall? */
if (time_after(now, ts + thresh)) {
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
lockup_detected = true;
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
if (pool->cpu >= 0) {
pool->cpu_stall = true;
cpu_pool_stall = true;
}
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
pr_emerg("BUG: workqueue lockup - pool");
pr_cont_pool_info(pool);
pr_cont(" stuck for %us!\n",
jiffies_to_msecs(now - pool_ts) / 1000);
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
}
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
}
rcu_read_unlock();
if (lockup_detected)
show_all_workqueues();
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
workqueue: Print backtraces from CPUs with hung CPU bound workqueues The workqueue watchdog reports a lockup when there was not any progress in the worker pool for a long time. The progress means that a pending work item starts being proceed. Worker pools for unbound workqueues always wake up an idle worker and try to process the work immediately. The last idle worker has to create new worker first. The stall might happen only when a new worker could not be created in which case an error should get printed. Another problem might be too high load. In this case, workers are victims of a global system problem. Worker pools for CPU bound workqueues are designed for lightweight work items that do not need much CPU time. They are proceed one by one on a single worker. New worker is used only when a work is sleeping. It creates one additional scenario. The stall might happen when the CPU-bound workqueue is used for CPU-intensive work. More precisely, the stall is detected when a CPU-bound worker is in the TASK_RUNNING state for too long. In this case, it might be useful to see the backtrace from the problematic worker. The information how long a worker is in the running state is not available. But the CPU-bound worker pools do not have many workers in the running state by definition. And only few pools are typically blocked. It should be acceptable to print backtraces from all workers in TASK_RUNNING state in the stalled worker pools. The number of false positives should be very low. Signed-off-by: Petr Mladek <pmladek@suse.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-03-07 12:53:35 +00:00
if (cpu_pool_stall)
show_cpu_pools_hogs();
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
wq_watchdog_reset_touched();
mod_timer(&wq_watchdog_timer, jiffies + thresh);
}
notrace void wq_watchdog_touch(int cpu)
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
{
if (cpu >= 0)
per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
workqueue/watchdog: Make unbound workqueues aware of touch_softlockup_watchdog() 84;0;0c84;0;0c There are two workqueue-specific watchdog timestamps: + @wq_watchdog_touched_cpu (per-CPU) updated by touch_softlockup_watchdog() + @wq_watchdog_touched (global) updated by touch_all_softlockup_watchdogs() watchdog_timer_fn() checks only the global @wq_watchdog_touched for unbound workqueues. As a result, unbound workqueues are not aware of touch_softlockup_watchdog(). The watchdog might report a stall even when the unbound workqueues are blocked by a known slow code. Solution: touch_softlockup_watchdog() must touch also the global @wq_watchdog_touched timestamp. The global timestamp can no longer be used for bound workqueues because it is now updated from all CPUs. Instead, bound workqueues have to check only @wq_watchdog_touched_cpu and these timestamps have to be updated for all CPUs in touch_all_softlockup_watchdogs(). Beware: The change might cause the opposite problem. An unbound workqueue might get blocked on CPU A because of a real softlockup. The workqueue watchdog would miss it when the timestamp got touched on CPU B. It is acceptable because softlockups are detected by softlockup watchdog. The workqueue watchdog is there to detect stalls where a work never finishes, for example, because of dependencies of works queued into the same workqueue. V3: - Modify the commit message clearly according to Petr's suggestion. Signed-off-by: Wang Qing <wangqing@vivo.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2021-03-24 11:40:29 +00:00
wq_watchdog_touched = jiffies;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
}
static void wq_watchdog_set_thresh(unsigned long thresh)
{
wq_watchdog_thresh = 0;
del_timer_sync(&wq_watchdog_timer);
if (thresh) {
wq_watchdog_thresh = thresh;
wq_watchdog_reset_touched();
mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
}
}
static int wq_watchdog_param_set_thresh(const char *val,
const struct kernel_param *kp)
{
unsigned long thresh;
int ret;
ret = kstrtoul(val, 0, &thresh);
if (ret)
return ret;
if (system_wq)
wq_watchdog_set_thresh(thresh);
else
wq_watchdog_thresh = thresh;
return 0;
}
static const struct kernel_param_ops wq_watchdog_thresh_ops = {
.set = wq_watchdog_param_set_thresh,
.get = param_get_ulong,
};
module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
0644);
static void wq_watchdog_init(void)
{
timer: Remove users of TIMER_DEFERRED_INITIALIZER This removes uses of TIMER_DEFERRED_INITIALIZER and chooses a location to call timer_setup() from before add_timer() or mod_timer() is called. Adjusts callbacks to use from_timer() as needed. Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mips@linux-mips.org Cc: Petr Mladek <pmladek@suse.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Sebastian Reichel <sre@kernel.org> Cc: Kalle Valo <kvalo@qca.qualcomm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Pavel Machek <pavel@ucw.cz> Cc: linux1394-devel@lists.sourceforge.net Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: linux-s390@vger.kernel.org Cc: linux-wireless@vger.kernel.org Cc: "James E.J. Bottomley" <jejb@linux.vnet.ibm.com> Cc: Wim Van Sebroeck <wim@iguana.be> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Ursula Braun <ubraun@linux.vnet.ibm.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Cc: Harish Patil <harish.patil@cavium.com> Cc: Stephen Boyd <sboyd@codeaurora.org> Cc: Guenter Roeck <linux@roeck-us.net> Cc: Manish Chopra <manish.chopra@cavium.com> Cc: Len Brown <len.brown@intel.com> Cc: Arnd Bergmann <arnd@arndb.de> Cc: linux-pm@vger.kernel.org Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Julian Wiedmann <jwi@linux.vnet.ibm.com> Cc: John Stultz <john.stultz@linaro.org> Cc: Mark Gross <mark.gross@intel.com> Cc: linux-watchdog@vger.kernel.org Cc: linux-scsi@vger.kernel.org Cc: "Martin K. Petersen" <martin.petersen@oracle.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: "Rafael J. Wysocki" <rjw@rjwysocki.net> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Stefan Richter <stefanr@s5r6.in-berlin.de> Cc: Michael Reed <mdr@sgi.com> Cc: netdev@vger.kernel.org Cc: Tejun Heo <tj@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: linuxppc-dev@lists.ozlabs.org Cc: Sudip Mukherjee <sudipm.mukherjee@gmail.com> Link: https://lkml.kernel.org/r/1507159627-127660-7-git-send-email-keescook@chromium.org
2017-10-04 23:27:00 +00:00
timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
wq_watchdog_set_thresh(wq_watchdog_thresh);
}
#else /* CONFIG_WQ_WATCHDOG */
static inline void wq_watchdog_init(void) { }
#endif /* CONFIG_WQ_WATCHDOG */
static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
{
if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
return;
}
cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
}
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/**
* workqueue_init_early - early init for workqueue subsystem
*
* This is the first step of three-staged workqueue subsystem initialization and
* invoked as soon as the bare basics - memory allocation, cpumasks and idr are
* up. It sets up all the data structures and system workqueues and allows early
* boot code to create workqueues and queue/cancel work items. Actual work item
* execution starts only after kthreads can be created and scheduled right
* before early initcalls.
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
*/
void __init workqueue_init_early(void)
{
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
int i, cpu;
BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
if (!cpumask_empty(&wq_cmdline_cpumask))
restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
workqueue: Add workqueue_unbound_exclude_cpumask() to exclude CPUs from wq_unbound_cpumask When the "isolcpus" boot command line option is used to add a set of isolated CPUs, those CPUs will be excluded automatically from wq_unbound_cpumask to avoid running work functions from unbound workqueues. Recently cpuset has been extended to allow the creation of partitions of isolated CPUs dynamically. To make it closer to the "isolcpus" in functionality, the CPUs in those isolated cpuset partitions should be excluded from wq_unbound_cpumask as well. This can be done currently by explicitly writing to the workqueue's cpumask sysfs file after creating the isolated partitions. However, this process can be error prone. Ideally, the cpuset code should be allowed to request the workqueue code to exclude those isolated CPUs from wq_unbound_cpumask so that this operation can be done automatically and the isolated CPUs will be returned back to wq_unbound_cpumask after the destructions of the isolated cpuset partitions. This patch adds a new workqueue_unbound_exclude_cpumask() function to enable that. This new function will exclude the specified isolated CPUs from wq_unbound_cpumask. To be able to restore those isolated CPUs back after the destruction of isolated cpuset partitions, a new wq_requested_unbound_cpumask is added to store the user provided unbound cpumask either from the boot command line options or from writing to the cpumask sysfs file. This new cpumask provides the basis for CPU exclusion. To enable users to understand how the wq_unbound_cpumask is being modified internally, this patch also exposes the newly introduced wq_requested_unbound_cpumask as well as a wq_isolated_cpumask to store the cpumask to be excluded from wq_unbound_cpumask as read-only sysfs files. Signed-off-by: Waiman Long <longman@redhat.com> Signed-off-by: Tejun Heo <tj@kernel.org>
2023-10-25 18:25:52 +00:00
cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
wq_update_pod_attrs_buf = alloc_workqueue_attrs();
BUG_ON(!wq_update_pod_attrs_buf);
workqueue: Generalize unbound CPU pods While renamed to pod, the code still assumes that the pods are defined by NUMA boundaries. Let's generalize it: * workqueue_attrs->affn_scope is added. Each enum represents the type of boundaries that define the pods. There are currently two scopes - WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before - one pod per NUMA node. The latter defines one global pod across the whole system. * struct wq_pod_type is added which describes how pods are configured for each affnity scope. For each pod, it lists the member CPUs and the preferred NUMA node for memory allocations. The reverse mapping from CPU to pod is also available. * wq_pod_enabled is dropped. Pod is now always enabled. The previously disabled behavior is now implemented through WQ_AFFN_SYSTEM. * get_unbound_pool() wants to determine the NUMA node to allocate memory from for the new pool. The variables are renamed from node to pod but the logic still assumes they're one and the same. Clearly distinguish them - walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's NUMA node. * wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA node. Take @cpu instead and determine the cpumask to use from the pod_type matching @attrs. * apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of NULL so that it can indicate -EINVAL on invalid affinity scopes. This patch allows CPUs to be grouped into pods however desired per type. While this patch causes some internal behavior changes, nothing material should change for workqueue users. v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is WQ_AFFN_NR_TYPES which indicates that the function is called with a worker_pool's attrs instead of a workqueue's. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:24 +00:00
/* initialize WQ_AFFN_SYSTEM pods */
pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
pt->nr_pods = 1;
cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
pt->pod_node[0] = NUMA_NO_NODE;
pt->cpu_pod[0] = 0;
/* initialize CPU pools */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
for_each_possible_cpu(cpu) {
struct worker_pool *pool;
i = 0;
for_each_cpu_worker_pool(pool, cpu) {
BUG_ON(init_worker_pool(pool));
pool->cpu = cpu;
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
workqueue: Add workqueue_attrs->__pod_cpumask workqueue_attrs has two uses: * to specify the required unouned workqueue properties by users * to match worker_pool's properties to workqueues by core code For example, if the user wants to restrict a workqueue to run only CPUs 0 and 2, and the two CPUs are on different affinity scopes, the workqueue's attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be associated with two worker_pools, one with attrs->cpumask containing just CPU 0 and the other CPU 2. Workqueue wants to support non-strict affinity scopes where work items are started in their matching affinity scopes but the scheduler is free to migrate them outside the starting scopes, which can enable utilizing the whole machine while maintaining most of the locality benefits from affinity scopes. To enable that, worker_pools need to distinguish the strict affinity that it has to follow (because that's the restriction coming from the user) and the soft affinity that it wants to apply when dispatching work items. Note that two worker_pools with different soft dispatching requirements have to be separate; otherwise, for example, we'd be ping-ponging worker threads across NUMA boundaries constantly. This patch adds workqueue_attrs->__pod_cpumask. The new field is double underscored as it's only used internally to distinguish worker_pools. A worker_pool's ->cpumask is now always the same as the online subset of allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's subset of that ->cpumask. Going back to the example above, both worker_pools would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask would contain 0 while the other's 2. * pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask that the pool's workers must stay within. This is currently always ->__pod_cpumask as all boundaries are still strict. * As a workqueue_attrs can now track both the associated workqueues' cpumask and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external out-argument. Drop @cpumask and instead store the result in ->__pod_cpumask. * The above also simplifies apply_wqattrs_prepare() as the same workqueue_attrs can be used to create all pods associated with a workqueue. tmp_attrs is dropped. * wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq update is needed instead of only comparing ->cpumask so that ->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks but the code is easier to understand and more robust this way. The only user-visible behavior change is that two workqueues with different cpumasks no longer can share worker_pools even when their pod subsets coincide. Going back to the example, let's say there's another workqueue with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped to two worker_pools - one with CPU 0, the other with 2 and 3. The former has the same cpumask as the first pod of the earlier example and would have shared the same worker_pool but that's no longer the case after this patch. The worker_pools would have the same ->__pod_cpumask but their ->cpumask's wouldn't match. While this is necessary to support non-strict affinity scopes, there can be further optimizations to maintain sharing among strict affinity scopes. However, non-strict affinity scopes are going to be preferable for most use cases and we don't see very diverse mixture of unbound workqueue cpumasks anyway, so the additional overhead doesn't seem to justify the extra complexity. v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by using wqattrs_equal() for comparison instead. - Per-cpu worker pools weren't initializing ->__pod_cpumask which caused a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask. Signed-off-by: Tejun Heo <tj@kernel.org>
2023-08-08 01:57:25 +00:00
cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
pool->attrs->nice = std_nice[i++];
workqueue: Implement non-strict affinity scope for unbound workqueues An unbound workqueue can be served by multiple worker_pools to improve locality. The segmentation is achieved by grouping CPUs into pods. By default, the cache boundaries according to cpus_share_cache() define the CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the system has two L3 caches. The workqueue would be mapped to two worker_pools each serving one L3 cache domains. While this improves locality, because the pod boundaries are strict, it limits the total bandwidth a given issuer can consume. For example, let's say there is a thread pinned to a CPU issuing enough work items to saturate the whole machine. With the machine segmented into two pods, no matter how many work items it issues, it can only use half of the CPUs on the system. While this limitation has existed for a very long time, it wasn't very pronounced because the affinity grouping used to be always by NUMA nodes. With cache boundaries as the default and support for even finer grained scopes (smt and cpu), it is now an a lot more pressing problem. This patch implements non-strict affinity scope where the pod boundaries aren't enforced strictly. Going back to the previous example, the workqueue would still be mapped to two worker_pools; however, the affinity enforcement would be soft. The workers in both pools would have their cpus_allowed set to the whole machine thus allowing the scheduler to migrate them anywhere on the machine. However, whenever an idle worker is woken up, the workqueue code asks the scheduler to bring back the task within the pod if the worker is outside. ie. work items start executing within its affinity scope but can be migrated outside as the scheduler sees fit. This removes the hard cap on utilization while maintaining the benefits of affinity scopes. After the earlier ->__pod_cpumask changes, the implementation is pretty simple. When non-strict which is the new default: * pool_allowed_cpus() returns @pool->attrs->cpumask instead of ->__pod_cpumask so that the workers are allowed to run on any CPU that the associated workqueues allow. * If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets the field to a CPU within the pod. This would be the first use of task_struct->wake_cpu outside scheduler proper, so it isn't clear whether this would be acceptable. However, other methods of migrating tasks are significantly more expensive and are likely prohibitively so if we want to do this on every work item. This needs discussion with scheduler folks. There is also a race window where setting ->wake_cpu wouldn't be effective as the target task is still on CPU. However, the window is pretty small and this being a best-effort optimization, it doesn't seem to warrant more complexity at the moment. While the non-strict cache affinity scopes seem to be the best option, the performance picture interacts with the affinity scope and is a bit complicated to fully discuss in this patch, so the behavior is made easily selectable through wqattrs and sysfs and the next patch will add documentation to discuss performance implications. v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-08-08 01:57:25 +00:00
pool->attrs->affn_strict = true;
pool->node = cpu_to_node(cpu);
/* alloc pool ID */
mutex_lock(&wq_pool_mutex);
BUG_ON(worker_pool_assign_id(pool));
mutex_unlock(&wq_pool_mutex);
}
}
/* create default unbound and ordered wq attrs */
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
struct workqueue_attrs *attrs;
BUG_ON(!(attrs = alloc_workqueue_attrs()));
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
attrs->nice = std_nice[i];
unbound_std_wq_attrs[i] = attrs;
/*
* An ordered wq should have only one pwq as ordering is
* guaranteed by max_active which is enforced by pwqs.
*/
BUG_ON(!(attrs = alloc_workqueue_attrs()));
attrs->nice = std_nice[i];
attrs->ordered = true;
ordered_wq_attrs[i] = attrs;
workqueue: implement attribute-based unbound worker_pool management This patch makes unbound worker_pools reference counted and dynamically created and destroyed as workqueues needing them come and go. All unbound worker_pools are hashed on unbound_pool_hash which is keyed by the content of worker_pool->attrs. When an unbound workqueue is allocated, get_unbound_pool() is called with the attributes of the workqueue. If there already is a matching worker_pool, the reference count is bumped and the pool is returned. If not, a new worker_pool with matching attributes is created and returned. When an unbound workqueue is destroyed, put_unbound_pool() is called which decrements the reference count of the associated worker_pool. If the refcnt reaches zero, the worker_pool is destroyed in sched-RCU safe way. Note that the standard unbound worker_pools - normal and highpri ones with no specific cpumask affinity - are no longer created explicitly during init_workqueues(). init_workqueues() only initializes workqueue_attrs to be used for standard unbound pools - unbound_std_wq_attrs[]. The pools are spawned on demand as workqueues are created. v2: - Comment added to init_worker_pool() explaining that @pool should be in a condition which can be passed to put_unbound_pool() even on failure. - pool->refcnt reaching zero and the pool being removed from unbound_pool_hash should be dynamic. pool->refcnt is converted to int from atomic_t and now manipulated inside workqueue_lock. - Removed an incorrect sanity check on nr_idle in put_unbound_pool() which may trigger spuriously. All changes were suggested by Lai Jiangshan. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
2013-03-12 18:30:03 +00:00
}
system_wq = alloc_workqueue("events", 0, 0);
system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
system_long_wq = alloc_workqueue("events_long", 0, 0);
system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
workqueue: Make unbound workqueues to use per-cpu pool_workqueues A pwq (pool_workqueue) represents an association between a workqueue and a worker_pool. When a work item is queued, the workqueue selects the pwq to use, which in turn determines the pool, and queues the work item to the pool through the pwq. pwq is also what implements the maximum concurrency limit - @max_active. As a per-cpu workqueue should be assocaited with a different worker_pool on each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq. However, unbound workqueues were sharing a pwq within each NUMA node by default. The sharing has several downsides: * Because @max_active is per-pwq, the meaning of @max_active changes depending on the machine configuration and whether workqueue NUMA locality support is enabled. * Makes per-cpu and unbound code deviate. * Gets in the way of making workqueue CPU locality awareness more flexible. This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu workqueues do by making the following changes: * wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound workqueues. * numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs the specified pwq to the target CPU's wq->cpu_pwq. * apply_wqattrs_prepare() now always allocates a separate pwq for each CPU unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq. This makes the return value of wq_calc_node_cpumask() unnecessary. It now returns void. * @max_active now means the same thing for both per-cpu and unbound workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer used in workqueue implementation and will be removed later. * All unbound pwq operations which used to be per-numa-node are now per-cpu. For most unbound workqueue users, this shouldn't cause noticeable changes. Work item issue and completion will be a small bit faster, flush_workqueue() would become a bit more expensive, and the total concurrency limit would likely become higher. All @max_active==1 use cases are currently being audited for conversion into alloc_ordered_workqueue() and they shouldn't be affected once the audit and conversion is complete. One area where the behavior change may be more noticeable is workqueue_congested() as the reported congestion state is now per CPU instead of NUMA node. There are only two users of this interface - drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are cc'd. Inputs on the behavior change would be very much appreciated. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Leon Romanovsky <leon@kernel.org> Cc: Karsten Graul <kgraul@linux.ibm.com> Cc: Wenjia Zhang <wenjia@linux.ibm.com> Cc: Jan Karcher <jaka@linux.ibm.com>
2023-08-08 01:57:23 +00:00
WQ_MAX_ACTIVE);
system_freezable_wq = alloc_workqueue("events_freezable",
WQ_FREEZABLE, 0);
system_power_efficient_wq = alloc_workqueue("events_power_efficient",
WQ_POWER_EFFICIENT, 0);
system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
WQ_FREEZABLE | WQ_POWER_EFFICIENT,
0);
BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
!system_unbound_wq || !system_freezable_wq ||
!system_power_efficient_wq ||
!system_freezable_power_efficient_wq);
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
}
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
static void __init wq_cpu_intensive_thresh_init(void)
{
unsigned long thresh;
unsigned long bogo;
pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
BUG_ON(IS_ERR(pwq_release_worker));
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
/* if the user set it to a specific value, keep it */
if (wq_cpu_intensive_thresh_us != ULONG_MAX)
return;
/*
* The default of 10ms is derived from the fact that most modern (as of
* 2023) processors can do a lot in 10ms and that it's just below what
* most consider human-perceivable. However, the kernel also runs on a
* lot slower CPUs including microcontrollers where the threshold is way
* too low.
*
* Let's scale up the threshold upto 1 second if BogoMips is below 4000.
* This is by no means accurate but it doesn't have to be. The mechanism
* is still useful even when the threshold is fully scaled up. Also, as
* the reports would usually be applicable to everyone, some machines
* operating on longer thresholds won't significantly diminish their
* usefulness.
*/
thresh = 10 * USEC_PER_MSEC;
/* see init/calibrate.c for lpj -> BogoMIPS calculation */
bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
if (bogo < 4000)
thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
loops_per_jiffy, bogo, thresh);
wq_cpu_intensive_thresh_us = thresh;
}
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/**
* workqueue_init - bring workqueue subsystem fully online
*
* This is the second step of three-staged workqueue subsystem initialization
* and invoked as soon as kthreads can be created and scheduled. Workqueues have
* been created and work items queued on them, but there are no kworkers
* executing the work items yet. Populate the worker pools with the initial
* workers and enable future kworker creations.
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
*/
void __init workqueue_init(void)
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
{
workqueue: move wq_numa_init() to workqueue_init() While splitting up workqueue initialization into two parts, ac8f73400782 ("workqueue: make workqueue available early during boot") put wq_numa_init() into workqueue_init_early(). Unfortunately, on some archs including power and arm64, cpu to node mapping isn't yet established by the time the early init is called leading to incorrect NUMA initialization and subsequently the following oops due to zero cpumask on node-specific unbound pools. Unable to handle kernel paging request for data at address 0x00000038 Faulting instruction address: 0xc0000000000fc0cc Oops: Kernel access of bad area, sig: 11 [#1] SMP NR_CPUS=2048 NUMA PowerNV Modules linked in: CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.8.0-compiler_gcc-6.2.0-next-20161005 #94 task: c0000007f5400000 task.stack: c000001ffc084000 NIP: c0000000000fc0cc LR: c0000000000ed928 CTR: c0000000000fbfd0 REGS: c000001ffc087780 TRAP: 0300 Not tainted (4.8.0-compiler_gcc-6.2.0-next-20161005) MSR: 9000000002009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE> CR: 48000424 XER: 00000000 CFAR: c0000000000089dc DAR: 0000000000000038 DSISR: 40000000 SOFTE: 0 GPR00: c0000000000ed928 c000001ffc087a00 c000000000e63200 c000000010d6d600 GPR04: c0000007f5409200 0000000000000021 000000000748e08c 000000000000001f GPR08: 0000000000000000 0000000000000021 000000000748f1f8 0000000000000000 GPR12: 0000000028000422 c00000000fb80000 c00000000000e0c8 0000000000000000 GPR16: 0000000000000000 0000000000000000 0000000000000021 0000000000000001 GPR20: ffffffffafb50401 0000000000000000 c000000010d6d600 000000000000ba7e GPR24: 000000000000ba7e c000000000d8bc58 afb504000afb5041 0000000000000001 GPR28: 0000000000000000 0000000000000004 c0000007f5409280 0000000000000000 NIP [c0000000000fc0cc] enqueue_task_fair+0xfc/0x18b0 LR [c0000000000ed928] activate_task+0x78/0xe0 Call Trace: [c000001ffc087a00] [c0000007f5409200] 0xc0000007f5409200 (unreliable) [c000001ffc087b10] [c0000000000ed928] activate_task+0x78/0xe0 [c000001ffc087b50] [c0000000000ede58] ttwu_do_activate+0x68/0xc0 [c000001ffc087b90] [c0000000000ef1b8] try_to_wake_up+0x208/0x4f0 [c000001ffc087c10] [c0000000000d3484] create_worker+0x144/0x250 [c000001ffc087cb0] [c000000000cd72d0] workqueue_init+0x124/0x150 [c000001ffc087d00] [c000000000cc0e74] kernel_init_freeable+0x158/0x360 [c000001ffc087dc0] [c00000000000e0e4] kernel_init+0x24/0x160 [c000001ffc087e30] [c00000000000bfa0] ret_from_kernel_thread+0x5c/0xbc Instruction dump: 62940401 3b800000 3aa00000 7f17c378 3a600001 3b600001 60000000 60000000 60420000 72490021 ebfe0150 2f890001 <ebbf0038> 419e0de0 7fbee840 419e0e58 ---[ end trace 0000000000000000 ]--- Fix it by moving wq_numa_init() to workqueue_init(). As this means that the early intialization may not have full NUMA info for per-cpu pools and ignores NUMA affinity for unbound pools, fix them up from workqueue_init() after wq_numa_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Michael Ellerman <mpe@ellerman.id.au> Link: http://lkml.kernel.org/r/87twck5wqo.fsf@concordia.ellerman.id.au Fixes: ac8f73400782 ("workqueue: make workqueue available early during boot") Signed-off-by: Tejun Heo <tj@kernel.org>
2016-10-19 16:01:27 +00:00
struct workqueue_struct *wq;
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
struct worker_pool *pool;
int cpu, bkt;
workqueue: Scale up wq_cpu_intensive_thresh_us if BogoMIPS is below 4000 wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items. Once detected, they're excluded from concurrency management to prevent them from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is enabled, repeat offenders are also reported so that the code can be updated. The default threshold is 10ms which is long enough to do fair bit of work on modern CPUs while short enough to be usually not noticeable. This unfortunately leads to a lot of, arguable spurious, detections on very slow CPUs. Using the same threshold across CPUs whose performance levels may be apart by multiple levels of magnitude doesn't make whole lot of sense. This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS is below 4000. This is obviously very inaccurate but it doesn't have to be accurate to be useful. The mechanism is still useful when the threshold is fully scaled up and the benefits of reports are usually shared with everyone regardless of who's reporting, so as long as there are sufficient number of fast machines reporting, we don't lose much. Some (or is it all?) ARM CPUs systemtically report significantly lower BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs are, it's at least a missed opportunity and it probably would be a good idea to teach workqueue about it. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
2023-07-17 22:50:02 +00:00
wq_cpu_intensive_thresh_init();
workqueue: move wq_numa_init() to workqueue_init() While splitting up workqueue initialization into two parts, ac8f73400782 ("workqueue: make workqueue available early during boot") put wq_numa_init() into workqueue_init_early(). Unfortunately, on some archs including power and arm64, cpu to node mapping isn't yet established by the time the early init is called leading to incorrect NUMA initialization and subsequently the following oops due to zero cpumask on node-specific unbound pools. Unable to handle kernel paging request for data at address 0x00000038 Faulting instruction address: 0xc0000000000fc0cc Oops: Kernel access of bad area, sig: 11 [#1] SMP NR_CPUS=2048 NUMA PowerNV Modules linked in: CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.8.0-compiler_gcc-6.2.0-next-20161005 #94 task: c0000007f5400000 task.stack: c000001ffc084000 NIP: c0000000000fc0cc LR: c0000000000ed928 CTR: c0000000000fbfd0 REGS: c000001ffc087780 TRAP: 0300 Not tainted (4.8.0-compiler_gcc-6.2.0-next-20161005) MSR: 9000000002009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE> CR: 48000424 XER: 00000000 CFAR: c0000000000089dc DAR: 0000000000000038 DSISR: 40000000 SOFTE: 0 GPR00: c0000000000ed928 c000001ffc087a00 c000000000e63200 c000000010d6d600 GPR04: c0000007f5409200 0000000000000021 000000000748e08c 000000000000001f GPR08: 0000000000000000 0000000000000021 000000000748f1f8 0000000000000000 GPR12: 0000000028000422 c00000000fb80000 c00000000000e0c8 0000000000000000 GPR16: 0000000000000000 0000000000000000 0000000000000021 0000000000000001 GPR20: ffffffffafb50401 0000000000000000 c000000010d6d600 000000000000ba7e GPR24: 000000000000ba7e c000000000d8bc58 afb504000afb5041 0000000000000001 GPR28: 0000000000000000 0000000000000004 c0000007f5409280 0000000000000000 NIP [c0000000000fc0cc] enqueue_task_fair+0xfc/0x18b0 LR [c0000000000ed928] activate_task+0x78/0xe0 Call Trace: [c000001ffc087a00] [c0000007f5409200] 0xc0000007f5409200 (unreliable) [c000001ffc087b10] [c0000000000ed928] activate_task+0x78/0xe0 [c000001ffc087b50] [c0000000000ede58] ttwu_do_activate+0x68/0xc0 [c000001ffc087b90] [c0000000000ef1b8] try_to_wake_up+0x208/0x4f0 [c000001ffc087c10] [c0000000000d3484] create_worker+0x144/0x250 [c000001ffc087cb0] [c000000000cd72d0] workqueue_init+0x124/0x150 [c000001ffc087d00] [c000000000cc0e74] kernel_init_freeable+0x158/0x360 [c000001ffc087dc0] [c00000000000e0e4] kernel_init+0x24/0x160 [c000001ffc087e30] [c00000000000bfa0] ret_from_kernel_thread+0x5c/0xbc Instruction dump: 62940401 3b800000 3aa00000 7f17c378 3a600001 3b600001 60000000 60000000 60420000 72490021 ebfe0150 2f890001 <ebbf0038> 419e0de0 7fbee840 419e0e58 ---[ end trace 0000000000000000 ]--- Fix it by moving wq_numa_init() to workqueue_init(). As this means that the early intialization may not have full NUMA info for per-cpu pools and ignores NUMA affinity for unbound pools, fix them up from workqueue_init() after wq_numa_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Michael Ellerman <mpe@ellerman.id.au> Link: http://lkml.kernel.org/r/87twck5wqo.fsf@concordia.ellerman.id.au Fixes: ac8f73400782 ("workqueue: make workqueue available early during boot") Signed-off-by: Tejun Heo <tj@kernel.org>
2016-10-19 16:01:27 +00:00
mutex_lock(&wq_pool_mutex);
/*
* Per-cpu pools created earlier could be missing node hint. Fix them
* up. Also, create a rescuer for workqueues that requested it.
*/
workqueue: move wq_numa_init() to workqueue_init() While splitting up workqueue initialization into two parts, ac8f73400782 ("workqueue: make workqueue available early during boot") put wq_numa_init() into workqueue_init_early(). Unfortunately, on some archs including power and arm64, cpu to node mapping isn't yet established by the time the early init is called leading to incorrect NUMA initialization and subsequently the following oops due to zero cpumask on node-specific unbound pools. Unable to handle kernel paging request for data at address 0x00000038 Faulting instruction address: 0xc0000000000fc0cc Oops: Kernel access of bad area, sig: 11 [#1] SMP NR_CPUS=2048 NUMA PowerNV Modules linked in: CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.8.0-compiler_gcc-6.2.0-next-20161005 #94 task: c0000007f5400000 task.stack: c000001ffc084000 NIP: c0000000000fc0cc LR: c0000000000ed928 CTR: c0000000000fbfd0 REGS: c000001ffc087780 TRAP: 0300 Not tainted (4.8.0-compiler_gcc-6.2.0-next-20161005) MSR: 9000000002009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE> CR: 48000424 XER: 00000000 CFAR: c0000000000089dc DAR: 0000000000000038 DSISR: 40000000 SOFTE: 0 GPR00: c0000000000ed928 c000001ffc087a00 c000000000e63200 c000000010d6d600 GPR04: c0000007f5409200 0000000000000021 000000000748e08c 000000000000001f GPR08: 0000000000000000 0000000000000021 000000000748f1f8 0000000000000000 GPR12: 0000000028000422 c00000000fb80000 c00000000000e0c8 0000000000000000 GPR16: 0000000000000000 0000000000000000 0000000000000021 0000000000000001 GPR20: ffffffffafb50401 0000000000000000 c000000010d6d600 000000000000ba7e GPR24: 000000000000ba7e c000000000d8bc58 afb504000afb5041 0000000000000001 GPR28: 0000000000000000 0000000000000004 c0000007f5409280 0000000000000000 NIP [c0000000000fc0cc] enqueue_task_fair+0xfc/0x18b0 LR [c0000000000ed928] activate_task+0x78/0xe0 Call Trace: [c000001ffc087a00] [c0000007f5409200] 0xc0000007f5409200 (unreliable) [c000001ffc087b10] [c0000000000ed928] activate_task+0x78/0xe0 [c000001ffc087b50] [c0000000000ede58] ttwu_do_activate+0x68/0xc0 [c000001ffc087b90] [c0000000000ef1b8] try_to_wake_up+0x208/0x4f0 [c000001ffc087c10] [c0000000000d3484] create_worker+0x144/0x250 [c000001ffc087cb0] [c000000000cd72d0] workqueue_init+0x124/0x150 [c000001ffc087d00] [c000000000cc0e74] kernel_init_freeable+0x158/0x360 [c000001ffc087dc0] [c00000000000e0e4] kernel_init+0x24/0x160 [c000001ffc087e30] [c00000000000bfa0] ret_from_kernel_thread+0x5c/0xbc Instruction dump: 62940401 3b800000 3aa00000 7f17c378 3a600001 3b600001 60000000 60000000 60420000 72490021 ebfe0150 2f890001 <ebbf0038> 419e0de0 7fbee840 419e0e58 ---[ end trace 0000000000000000 ]--- Fix it by moving wq_numa_init() to workqueue_init(). As this means that the early intialization may not have full NUMA info for per-cpu pools and ignores NUMA affinity for unbound pools, fix them up from workqueue_init() after wq_numa_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Michael Ellerman <mpe@ellerman.id.au> Link: http://lkml.kernel.org/r/87twck5wqo.fsf@concordia.ellerman.id.au Fixes: ac8f73400782 ("workqueue: make workqueue available early during boot") Signed-off-by: Tejun Heo <tj@kernel.org>
2016-10-19 16:01:27 +00:00
for_each_possible_cpu(cpu) {
for_each_cpu_worker_pool(pool, cpu) {
pool->node = cpu_to_node(cpu);
}
}
list_for_each_entry(wq, &workqueues, list) {
WARN(init_rescuer(wq),
"workqueue: failed to create early rescuer for %s",
wq->name);
}
workqueue: move wq_numa_init() to workqueue_init() While splitting up workqueue initialization into two parts, ac8f73400782 ("workqueue: make workqueue available early during boot") put wq_numa_init() into workqueue_init_early(). Unfortunately, on some archs including power and arm64, cpu to node mapping isn't yet established by the time the early init is called leading to incorrect NUMA initialization and subsequently the following oops due to zero cpumask on node-specific unbound pools. Unable to handle kernel paging request for data at address 0x00000038 Faulting instruction address: 0xc0000000000fc0cc Oops: Kernel access of bad area, sig: 11 [#1] SMP NR_CPUS=2048 NUMA PowerNV Modules linked in: CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.8.0-compiler_gcc-6.2.0-next-20161005 #94 task: c0000007f5400000 task.stack: c000001ffc084000 NIP: c0000000000fc0cc LR: c0000000000ed928 CTR: c0000000000fbfd0 REGS: c000001ffc087780 TRAP: 0300 Not tainted (4.8.0-compiler_gcc-6.2.0-next-20161005) MSR: 9000000002009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE> CR: 48000424 XER: 00000000 CFAR: c0000000000089dc DAR: 0000000000000038 DSISR: 40000000 SOFTE: 0 GPR00: c0000000000ed928 c000001ffc087a00 c000000000e63200 c000000010d6d600 GPR04: c0000007f5409200 0000000000000021 000000000748e08c 000000000000001f GPR08: 0000000000000000 0000000000000021 000000000748f1f8 0000000000000000 GPR12: 0000000028000422 c00000000fb80000 c00000000000e0c8 0000000000000000 GPR16: 0000000000000000 0000000000000000 0000000000000021 0000000000000001 GPR20: ffffffffafb50401 0000000000000000 c000000010d6d600 000000000000ba7e GPR24: 000000000000ba7e c000000000d8bc58 afb504000afb5041 0000000000000001 GPR28: 0000000000000000 0000000000000004 c0000007f5409280 0000000000000000 NIP [c0000000000fc0cc] enqueue_task_fair+0xfc/0x18b0 LR [c0000000000ed928] activate_task+0x78/0xe0 Call Trace: [c000001ffc087a00] [c0000007f5409200] 0xc0000007f5409200 (unreliable) [c000001ffc087b10] [c0000000000ed928] activate_task+0x78/0xe0 [c000001ffc087b50] [c0000000000ede58] ttwu_do_activate+0x68/0xc0 [c000001ffc087b90] [c0000000000ef1b8] try_to_wake_up+0x208/0x4f0 [c000001ffc087c10] [c0000000000d3484] create_worker+0x144/0x250 [c000001ffc087cb0] [c000000000cd72d0] workqueue_init+0x124/0x150 [c000001ffc087d00] [c000000000cc0e74] kernel_init_freeable+0x158/0x360 [c000001ffc087dc0] [c00000000000e0e4] kernel_init+0x24/0x160 [c000001ffc087e30] [c00000000000bfa0] ret_from_kernel_thread+0x5c/0xbc Instruction dump: 62940401 3b800000 3aa00000 7f17c378 3a600001 3b600001 60000000 60000000 60420000 72490021 ebfe0150 2f890001 <ebbf0038> 419e0de0 7fbee840 419e0e58 ---[ end trace 0000000000000000 ]--- Fix it by moving wq_numa_init() to workqueue_init(). As this means that the early intialization may not have full NUMA info for per-cpu pools and ignores NUMA affinity for unbound pools, fix them up from workqueue_init() after wq_numa_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Michael Ellerman <mpe@ellerman.id.au> Link: http://lkml.kernel.org/r/87twck5wqo.fsf@concordia.ellerman.id.au Fixes: ac8f73400782 ("workqueue: make workqueue available early during boot") Signed-off-by: Tejun Heo <tj@kernel.org>
2016-10-19 16:01:27 +00:00
mutex_unlock(&wq_pool_mutex);
workqueue: make workqueue available early during boot Workqueue is currently initialized in an early init call; however, there are cases where early boot code has to be split and reordered to come after workqueue initialization or the same code path which makes use of workqueues is used both before workqueue initailization and after. The latter cases have to gate workqueue usages with keventd_up() tests, which is nasty and easy to get wrong. Workqueue usages have become widespread and it'd be a lot more convenient if it can be used very early from boot. This patch splits workqueue initialization into two steps. workqueue_init_early() which sets up the basic data structures so that workqueues can be created and work items queued, and workqueue_init() which actually brings up workqueues online and starts executing queued work items. The former step can be done very early during boot once memory allocation, cpumasks and idr are initialized. The latter right after kthreads become available. This allows work item queueing and canceling from very early boot which is what most of these use cases want. * As systemd_wq being initialized doesn't indicate that workqueue is fully online anymore, update keventd_up() to test wq_online instead. The follow-up patches will get rid of all its usages and the function itself. * Flushing doesn't make sense before workqueue is fully initialized. The flush functions trigger WARN and return immediately before fully online. * Work items are never in-flight before fully online. Canceling can always succeed by skipping the flush step. * Some code paths can no longer assume to be called with irq enabled as irq is disabled during early boot. Use irqsave/restore operations instead. v2: Watchdog init, which requires timer to be running, moved from workqueue_init_early() to workqueue_init(). Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/CA+55aFx0vPuMuxn00rBSM192n-Du5uxy+4AvKa0SBSOVJeuCGg@mail.gmail.com
2016-09-16 19:49:32 +00:00
/* create the initial workers */
for_each_online_cpu(cpu) {
for_each_cpu_worker_pool(pool, cpu) {
pool->flags &= ~POOL_DISASSOCIATED;
BUG_ON(!create_worker(pool));
}
}
hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
BUG_ON(!create_worker(pool));
wq_online = true;
workqueue: implement lockup detector Workqueue stalls can happen from a variety of usage bugs such as missing WQ_MEM_RECLAIM flag or concurrency managed work item indefinitely staying RUNNING. These stalls can be extremely difficult to hunt down because the usual warning mechanisms can't detect workqueue stalls and the internal state is pretty opaque. To alleviate the situation, this patch implements workqueue lockup detector. It periodically monitors all worker_pools periodically and, if any pool failed to make forward progress longer than the threshold duration, triggers warning and dumps workqueue state as follows. BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 31s! Showing busy workqueues and worker pools: workqueue events: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=17/256 pending: monkey_wrench_fn, e1000_watchdog, cache_reap, vmstat_shepherd, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, release_one_tty, cgroup_release_agent workqueue events_power_efficient: flags=0x80 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=2/256 pending: check_lifetime, neigh_periodic_work workqueue cgroup_pidlist_destroy: flags=0x0 pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/1 pending: cgroup_pidlist_destroy_work_fn ... The detection mechanism is controller through kernel parameter workqueue.watchdog_thresh and can be updated at runtime through the sysfs module parameter file. v2: Decoupled from softlockup control knobs. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Don Zickus <dzickus@redhat.com> Cc: Ulrich Obergfell <uobergfe@redhat.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Chris Mason <clm@fb.com> Cc: Andrew Morton <akpm@linux-foundation.org>
2015-12-08 16:28:04 +00:00
wq_watchdog_init();
}
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
/*
* Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
* @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
* and consecutive pod ID. The rest of @pt is initialized accordingly.
*/
static void __init init_pod_type(struct wq_pod_type *pt,
bool (*cpus_share_pod)(int, int))
{
int cur, pre, cpu, pod;
pt->nr_pods = 0;
/* init @pt->cpu_pod[] according to @cpus_share_pod() */
pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
BUG_ON(!pt->cpu_pod);
for_each_possible_cpu(cur) {
for_each_possible_cpu(pre) {
if (pre >= cur) {
pt->cpu_pod[cur] = pt->nr_pods++;
break;
}
if (cpus_share_pod(cur, pre)) {
pt->cpu_pod[cur] = pt->cpu_pod[pre];
break;
}
}
}
/* init the rest to match @pt->cpu_pod[] */
pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
BUG_ON(!pt->pod_cpus || !pt->pod_node);
for (pod = 0; pod < pt->nr_pods; pod++)
BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
for_each_possible_cpu(cpu) {
cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
}
}
static bool __init cpus_dont_share(int cpu0, int cpu1)
{
return false;
}
static bool __init cpus_share_smt(int cpu0, int cpu1)
{
#ifdef CONFIG_SCHED_SMT
return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
#else
return false;
#endif
}
static bool __init cpus_share_numa(int cpu0, int cpu1)
{
return cpu_to_node(cpu0) == cpu_to_node(cpu1);
}
/**
* workqueue_init_topology - initialize CPU pods for unbound workqueues
*
* This is the third step of there-staged workqueue subsystem initialization and
* invoked after SMP and topology information are fully initialized. It
* initializes the unbound CPU pods accordingly.
*/
void __init workqueue_init_topology(void)
{
struct workqueue_struct *wq;
int cpu;
init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
mutex_lock(&wq_pool_mutex);
/*
* Workqueues allocated earlier would have all CPUs sharing the default
* worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
* combinations to apply per-pod sharing.
*/
list_for_each_entry(wq, &workqueues, list) {
for_each_online_cpu(cpu) {
wq_update_pod(wq, cpu, cpu, true);
}
}
mutex_unlock(&wq_pool_mutex);
}
void __warn_flushing_systemwide_wq(void)
{
pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
dump_stack();
}
workqueue: Wrap flush_workqueue() using a macro Since flush operation synchronously waits for completion, flushing system-wide WQs (e.g. system_wq) might introduce possibility of deadlock due to unexpected locking dependency. Tejun Heo commented at [1] that it makes no sense at all to call flush_workqueue() on the shared WQs as the caller has no idea what it's gonna end up waiting for. Although there is flush_scheduled_work() which flushes system_wq WQ with "Think twice before calling this function! It's very easy to get into trouble if you don't take great care." warning message, syzbot found a circular locking dependency caused by flushing system_wq WQ [2]. Therefore, let's change the direction to that developers had better use their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is inevitable. Steps for converting system-wide WQs into local WQs are explained at [3], and a conversion to stop flushing system-wide WQs is in progress. Now we want some mechanism for preventing developers who are not aware of this conversion from again start flushing system-wide WQs. Since I found that WARN_ON() is complete but awkward approach for teaching developers about this problem, let's use __compiletime_warning() for incomplete but handy approach. For completeness, we will also insert WARN_ON() into __flush_workqueue() after all in-tree users stopped calling flush_scheduled_work(). Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1] Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2] Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3] Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Tejun Heo <tj@kernel.org>
2022-06-01 07:32:47 +00:00
EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
static int __init workqueue_unbound_cpus_setup(char *str)
{
if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
cpumask_clear(&wq_cmdline_cpumask);
pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
}
return 1;
}
__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);