Impact: cleanup
BALANCE_FOR_MC_POWER and similar macros defined in sched.h are
not constants and have various condition checks and significant
amount of code that is not suitable to be contain in a macro.
Also there could be side effects on the expressions passed to
some of them like test_sd_parent().
This patch converts all complex macros related to power savings
balance to inline functions.
Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Acked-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Provide a way to pause the function graph tracer
As suggested by Steven Rostedt, the previous patch that prevented from
spinlock function tracing shouldn't use the raw_spinlock to fix it.
It's much better to follow lockdep with normal spinlock, so this patch
adds a new flag for each task to make the function graph tracer able
to be paused. We also can send an ftrace_printk whithout worrying of
the irrelevant traced spinlock during insertion.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds the file:
/debugfs/tracing/set_graph_function
which can be used along with the function graph tracer.
When this file is empty, the function graph tracer will act as
usual. When the file has a function in it, the function graph
tracer will only trace that function.
For example:
# echo blk_unplug > /debugfs/tracing/set_graph_function
# cat /debugfs/tracing/trace
[...]
------------------------------------------
| 2) make-19003 => kjournald-2219
------------------------------------------
2) | blk_unplug() {
2) | dm_unplug_all() {
2) | dm_get_table() {
2) 1.381 us | _read_lock();
2) 0.911 us | dm_table_get();
2) 1. 76 us | _read_unlock();
2) + 12.912 us | }
2) | dm_table_unplug_all() {
2) | blk_unplug() {
2) 0.778 us | generic_unplug_device();
2) 2.409 us | }
2) 5.992 us | }
2) 0.813 us | dm_table_put();
2) + 29. 90 us | }
2) + 34.532 us | }
You can add up to 32 functions into this file. Currently we limit it
to 32, but this may change with later improvements.
To add another function, use the append '>>':
# echo sys_read >> /debugfs/tracing/set_graph_function
# cat /debugfs/tracing/set_graph_function
blk_unplug
sys_read
Using the '>' will clear out the function and write anew:
# echo sys_write > /debug/tracing/set_graph_function
# cat /debug/tracing/set_graph_function
sys_write
Note, if you have function graph running while doing this, the small
time between clearing it and updating it will cause the graph to
record all functions. This should not be an issue because after
it sets the filter, only those functions will be recorded from then on.
If you need to only record a particular function then set this
file first before starting the function graph tracer. In the future
this side effect may be corrected.
The set_graph_function file is similar to the set_ftrace_filter but
it does not take wild cards nor does it allow for more than one
function to be set with a single write. There is no technical reason why
this is the case, I just do not have the time yet to implement that.
Note, dynamic ftrace must be enabled for this to appear because it
uses the dynamic ftrace records to match the name to the mcount
call sites.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Impact: extend information in /proc/sched_debug
This patch adds uid information in sched_debug for CONFIG_USER_SCHED
Signed-off-by: Arun R Bharadwaj <arun@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
This patch changes the name of the "return function tracer" into
function-graph-tracer which is a more suitable name for a tracing
which makes one able to retrieve the ordered call stack during
the code flow.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: restructure DS memory allocation to be done by the usage site of DS
Require pre-allocated buffers in ds.h.
Move the BTS buffer allocation for ptrace into ptrace.c.
The pointer to the allocated buffer is stored in the traced task's
task_struct together with the handle returned by ds_request_bts().
Removes memory accounting code.
Signed-off-by: Markus Metzger <markus.t.metzger@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: generalize the DS code to shared buffers
Change the in-kernel ds.h interface to identify the tracer via a
handle returned on ds_request_~().
Tracers used to be identified via their task_struct.
The changes are required to allow DS to be shared between different
tasks, which is needed for perfmon2 and for ftrace.
For ptrace, the handle is stored in the traced task's task_struct.
This should probably go into a (arch-specific) ptrace context some
time.
Signed-off-by: Markus Metzger <markus.t.metzger@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Trivial API conversion
NR_CPUS -> nr_cpu_ids
cpumask_t -> struct cpumask
sizeof(cpumask_t) -> cpumask_size()
cpumask_a = cpumask_b -> cpumask_copy(&cpumask_a, &cpumask_b)
cpu_set() -> cpumask_set_cpu()
first_cpu() -> cpumask_first()
cpumask_of_cpu() -> cpumask_of()
cpus_* -> cpumask_*
There are some FIXMEs where we all archs to complete infrastructure
(patches have been sent):
cpu_coregroup_map -> cpu_coregroup_mask
node_to_cpumask* -> cpumask_of_node
There is also one FIXME where we pass an array of cpumasks to
partition_sched_domains(): this implies knowing the definition of
'struct cpumask' and the size of a cpumask. This will be fixed in a
future patch.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: (future) size reduction for large NR_CPUS.
Dynamically allocating cpumasks (when CONFIG_CPUMASK_OFFSTACK) saves
space for small nr_cpu_ids but big CONFIG_NR_CPUS. cpumask_var_t
is just a struct cpumask for !CONFIG_CPUMASK_OFFSTACK.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: (future) size reduction for large NR_CPUS.
We move the 'cpumask' member of sched_group to the end, so when we
kmalloc it we can do a minimal allocation: saves space for small
nr_cpu_ids but big CONFIG_NR_CPUS. Similar trick for 'span' in
sched_domain.
This isn't quite as good as converting to a cpumask_var_t, as some
sched_groups are actually static, but it's safer: we don't have to
figure out where to call alloc_cpumask_var/free_cpumask_var.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: trivial wrap of member accesses
This eases the transition in the next patch.
We also get rid of a temporary cpumask in find_idlest_cpu() thanks to
for_each_cpu_and, and sched_balance_self() due to getting weight before
setting sd to NULL.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: use deeper function tracing depth safely
Some tests showed that function return tracing needed a more deeper depth
of function calls. But it could be unsafe to store these return addresses
to the stack.
So these arrays will now be allocated dynamically into task_struct of current
only when the tracer is activated.
Typical scheme when tracer is activated:
- allocate a return stack for each task in global list.
- fork: allocate the return stack for the newly created task
- exit: free return stack of current
- idle init: same as fork
I chose a default depth of 50. I don't have overruns anymore.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: help to find the better depth of trace
We decided to arbitrary define the depth of function return trace as
"20". Perhaps this is not enough. To help finding an optimal depth, we
measure now the overrun: the number of functions that have been missed
for the current thread. By default this is not displayed, we have to
do set a particular flag on the return tracer: echo overrun >
/debug/tracing/trace_options And the overrun will be printed on the
right.
As the trace shows below, the current 20 depth is not enough.
update_wall_time+0x37f/0x8c0 -> update_xtime_cache (345 ns) (Overruns: 2838)
update_wall_time+0x384/0x8c0 -> clocksource_get_next (1141 ns) (Overruns: 2838)
do_timer+0x23/0x100 -> update_wall_time (3882 ns) (Overruns: 2838)
tick_do_update_jiffies64+0xbf/0x160 -> do_timer (5339 ns) (Overruns: 2838)
tick_sched_timer+0x6a/0xf0 -> tick_do_update_jiffies64 (7209 ns) (Overruns: 2838)
vgacon_set_cursor_size+0x98/0x120 -> native_io_delay (2613 ns) (Overruns: 274)
vgacon_cursor+0x16e/0x1d0 -> vgacon_set_cursor_size (33151 ns) (Overruns: 274)
set_cursor+0x5f/0x80 -> vgacon_cursor (36432 ns) (Overruns: 274)
con_flush_chars+0x34/0x40 -> set_cursor (38790 ns) (Overruns: 274)
release_console_sem+0x1ec/0x230 -> up (721 ns) (Overruns: 274)
release_console_sem+0x225/0x230 -> wake_up_klogd (316 ns) (Overruns: 274)
con_flush_chars+0x39/0x40 -> release_console_sem (2996 ns) (Overruns: 274)
con_write+0x22/0x30 -> con_flush_chars (46067 ns) (Overruns: 274)
n_tty_write+0x1cc/0x360 -> con_write (292670 ns) (Overruns: 274)
smp_apic_timer_interrupt+0x2a/0x90 -> native_apic_mem_write (330 ns) (Overruns: 274)
irq_enter+0x17/0x70 -> idle_cpu (413 ns) (Overruns: 274)
smp_apic_timer_interrupt+0x2f/0x90 -> irq_enter (1525 ns) (Overruns: 274)
ktime_get_ts+0x40/0x70 -> getnstimeofday (465 ns) (Overruns: 274)
ktime_get_ts+0x60/0x70 -> set_normalized_timespec (436 ns) (Overruns: 274)
ktime_get+0x16/0x30 -> ktime_get_ts (2501 ns) (Overruns: 274)
hrtimer_interrupt+0x77/0x1a0 -> ktime_get (3439 ns) (Overruns: 274)
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
This function no longer exists, so remove the defintion.
Signed-off-by: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: add infrastructure for function-return tracing
Add low level support for ftrace return tracing.
This plug-in stores return addresses on the thread_info structure of
the current task.
The index of the current return address is initialized when the task
is the first one (init) and when a process forks (the child). It is
not needed when a task does a sys_execve because after this syscall,
it still needs to return on the kernel functions it called.
Note that the code of return_to_handler has been suggested by Steven
Rostedt as almost all of the ideas of improvements in this V3.
For purpose of security, arch/x86/kernel/process_32.c is not traced
because __switch_to() changes the current task during its execution.
That could cause inconsistency in the stored return address of this
function even if I didn't have any crash after testing with tracing on
this function enabled.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: fix hang/crash on ia64 under high load
This is ugly, but the simplest patch by far.
Unlike other similar routines, account_group_exec_runtime() could be
called "implicitly" from within scheduler after exit_notify(). This
means we can race with the parent doing release_task(), we can't just
check ->signal != NULL.
Change __exit_signal() to do spin_unlock_wait(&task_rq(tsk)->lock)
before __cleanup_signal() to make sure ->signal can't be freed under
task_rq(tsk)->lock. Note that task_rq_unlock_wait() doesn't care
about the case when tsk changes cpu/rq under us, this should be OK.
Thanks to Ingo who nacked my previous buggy patch.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reported-by: Doug Chapman <doug.chapman@hp.com>
__scm_destroy() walks the list of file descriptors in the scm_fp_list
pointed to by the scm_cookie argument.
Those, in turn, can close sockets and invoke __scm_destroy() again.
There is nothing which limits how deeply this can occur.
The idea for how to fix this is from Linus. Basically, we do all of
the fput()s at the top level by collecting all of the scm_fp_list
objects hit by an fput(). Inside of the initial __scm_destroy() we
keep running the list until it is empty.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'v28-range-hrtimers-for-linus-v2' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (37 commits)
hrtimers: add missing docbook comments to struct hrtimer
hrtimers: simplify hrtimer_peek_ahead_timers()
hrtimers: fix docbook comments
DECLARE_PER_CPU needs linux/percpu.h
hrtimers: fix typo
rangetimers: fix the bug reported by Ingo for real
rangetimer: fix BUG_ON reported by Ingo
rangetimer: fix x86 build failure for the !HRTIMERS case
select: fix alpha OSF wrapper
select: fix alpha OSF wrapper
hrtimer: peek at the timer queue just before going idle
hrtimer: make the futex() system call use the per process slack value
hrtimer: make the nanosleep() syscall use the per process slack
hrtimer: fix signed/unsigned bug in slack estimator
hrtimer: show the timer ranges in /proc/timer_list
hrtimer: incorporate feedback from Peter Zijlstra
hrtimer: add a hrtimer_start_range() function
hrtimer: another build fix
hrtimer: fix build bug found by Ingo
hrtimer: make select() and poll() use the hrtimer range feature
...
* 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
sched: disable the hrtick for now
sched: revert back to per-rq vruntime
sched: fair scheduler should not resched rt tasks
sched: optimize group load balancer
sched: minor fast-path overhead reduction
sched: fix the wrong mask_len, cleanup
sched: kill unused scheduler decl.
sched: fix the wrong mask_len
sched: only update rq->clock while holding rq->lock
a patch from Henrik Austad did this:
>> Do not declare select_task_rq as part of sched_class when CONFIG_SMP is
>> not set.
Peter observed:
> While a proper cleanup, could you do it by re-arranging the methods so
> as to not create an additional ifdef?
Do not declare select_task_rq and some other methods as part of sched_class
when CONFIG_SMP is not set.
Also gather those methods to avoid CONFIG_SMP mess.
Idea-by: Henrik Austad <henrik.austad@gmail.com>
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Henrik Austad <henrik@austad.us>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'v28-timers-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (36 commits)
fix documentation of sysrq-q really
Fix documentation of sysrq-q
timer_list: add base address to clock base
timer_list: print cpu number of clockevents device
timer_list: print real timer address
NOHZ: restart tick device from irq_enter()
NOHZ: split tick_nohz_restart_sched_tick()
NOHZ: unify the nohz function calls in irq_enter()
timers: fix itimer/many thread hang, fix
timers: fix itimer/many thread hang, v3
ntp: improve adjtimex frequency rounding
timekeeping: fix rounding problem during clock update
ntp: let update_persistent_clock() sleep
hrtimer: reorder struct hrtimer to save 8 bytes on 64bit builds
posix-timers: lock_timer: make it readable
posix-timers: lock_timer: kill the bogus ->it_id check
posix-timers: kill ->it_sigev_signo and ->it_sigev_value
posix-timers: sys_timer_create: cleanup the error handling
posix-timers: move the initialization of timer->sigq from send to create path
posix-timers: sys_timer_create: simplify and s/tasklist/rcu/
...
Fix trivial conflicts due to sysrq-q description clahes in
Documentation/sysrq.txt and drivers/char/sysrq.c
This adds a kconfig option to change the /proc/PID/coredump_filter default.
Fedora has been carrying a trivial patch to change the hard-wired value for
this default, since Fedora 8. The default default can't change safely
because there are old GDB versions out there (all before 6.7) that are
confused by the core dump files created by the MMF_DUMP_ELF_HEADERS setting.
Signed-off-by: Roland McGrath <roland@redhat.com>
Cc: Michael Kerrisk <mtk.manpages@googlemail.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Kawai Hidehiro <hidehiro.kawai.ez@hitachi.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Jones <davej@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Presently hugepage's vma has a VM_RESERVED flag in order not to be
swapped. But a VM_RESERVED vma isn't core dumped because this flag is
often used for some kernel vmas (e.g. vmalloc, sound related).
Thus hugepages are never dumped and it can't be debugged easily. Many
developers want hugepages to be included into core-dump.
However, We can't read generic VM_RESERVED area because this area is often
IO mapping area. then these area reading may change device state. it is
definitly undesiable side-effect.
So adding a hugepage specific bit to the coredump filter is better. It
will be able to hugepage core dumping and doesn't cause any side-effect to
any i/o devices.
In additional, libhugetlb use hugetlb private mapping pages as anonymous
page. Then, hugepage private mapping pages should be core dumped by
default.
Then, /proc/[pid]/core_dump_filter has two new bits.
- bit 5 mean hugetlb private mapping pages are dumped or not. (default: yes)
- bit 6 mean hugetlb shared mapping pages are dumped or not. (default: no)
I tested by following method.
% ulimit -c unlimited
% ./crash_hugepage 50
% ./crash_hugepage 50 -p
% ls -lh
% gdb ./crash_hugepage core
%
% echo 0x43 > /proc/self/coredump_filter
% ./crash_hugepage 50
% ./crash_hugepage 50 -p
% ls -lh
% gdb ./crash_hugepage core
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/mman.h>
#include <string.h>
#include "hugetlbfs.h"
int main(int argc, char** argv){
char* p;
int ch;
int mmap_flags = MAP_SHARED;
int fd;
int nr_pages;
while((ch = getopt(argc, argv, "p")) != -1) {
switch (ch) {
case 'p':
mmap_flags &= ~MAP_SHARED;
mmap_flags |= MAP_PRIVATE;
break;
default:
/* nothing*/
break;
}
}
argc -= optind;
argv += optind;
if (argc == 0){
printf("need # of pages\n");
exit(1);
}
nr_pages = atoi(argv[0]);
if (nr_pages < 2) {
printf("nr_pages must >2\n");
exit(1);
}
fd = hugetlbfs_unlinked_fd();
p = mmap(NULL, nr_pages * gethugepagesize(),
PROT_READ|PROT_WRITE, mmap_flags, fd, 0);
sleep(2);
*(p + gethugepagesize()) = 1; /* COW */
sleep(2);
/* crash! */
*(int*)0 = 1;
return 0;
}
Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Reviewed-by: Kawai Hidehiro <hidehiro.kawai.ez@hitachi.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: William Irwin <wli@holomorphy.com>
Cc: Adam Litke <agl@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I noticed that tg_shares_up() unconditionally takes rq-locks for all cpus
in the sched_domain. This hurts.
We need the rq-locks whenever we change the weight of the per-cpu group sched
entities. To allevate this a little, only change the weight when the new
weight is at least shares_thresh away from the old value.
This avoids the rq-lock for the top level entries, since those will never
be re-weighted, and fuzzes the lower level entries a little to gain performance
in semi-stable situations.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
I noticed this while making investigations into the tbench
regressions. Please apply.
sched: Remove hrtick_resched() extern decl.
This function was removed by 31656519e1
("sched, x86: clean up hrtick implementation").
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
add /proc/sys/kernel/sched_domain/cpu0/domain0/name, to make
it easier to see which specific scheduler domain remained at
that entry.
Since we process the scheduler domain tree and
simplify it, it's not always immediately clear during debugging
which domain came from where.
depends on CONFIG_SCHED_DEBUG=y.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- fix UP lockup
- another set of UP/SMP cleanups and simplifications
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This is the second resubmission of the posix timer rework patch, posted
a few days ago.
This includes the changes from the previous resubmittion, which addressed
Oleg Nesterov's comments, removing the RCU stuff from the patch and
un-inlining the thread_group_cputime() function for SMP.
In addition, per Ingo Molnar it simplifies the UP code, consolidating much
of it with the SMP version and depending on lower-level SMP/UP handling to
take care of the differences.
It also cleans up some UP compile errors, moves the scheduler stats-related
macros into kernel/sched_stats.h, cleans up a merge error in
kernel/fork.c and has a few other minor fixes and cleanups as suggested
by Oleg and Ingo. Thanks for the review, guys.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Lin Ming reported a 10% OLTP regression against 2.6.27-rc4.
The difference seems to come from different preemption agressiveness,
which affects the cache footprint of the workload and its effective
cache trashing.
Aggresively preempt a task if its avg overlap is very small, this should
avoid the task going to sleep and find it still running when we schedule
back to it - saving a wakeup.
Reported-by: Lin Ming <ming.m.lin@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
fix the UP build:
In file included from arch/x86/kernel/asm-offsets_32.c:9,
from arch/x86/kernel/asm-offsets.c:3:
include/linux/sched.h: In function ‘thread_group_cputime_clone_thread’:
include/linux/sched.h:2272: warning: no return statement in function returning non-void
include/linux/sched.h: In function ‘thread_group_cputime_account_user’:
include/linux/sched.h:2284: error: invalid type argument of ‘->’ (have ‘struct task_cputime’)
include/linux/sched.h:2284: error: invalid type argument of ‘->’ (have ‘struct task_cputime’)
include/linux/sched.h: In function ‘thread_group_cputime_account_system’:
include/linux/sched.h:2291: error: invalid type argument of ‘->’ (have ‘struct task_cputime’)
include/linux/sched.h:2291: error: invalid type argument of ‘->’ (have ‘struct task_cputime’)
include/linux/sched.h: In function ‘thread_group_cputime_account_exec_runtime’:
include/linux/sched.h:2298: error: invalid type argument of ‘->’ (have ‘struct task_cputime’)
distcc[14501] ERROR: compile arch/x86/kernel/asm-offsets.c on a/30 failed
make[1]: *** [arch/x86/kernel/asm-offsets.s] Error 1
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Define USE_SPLIT_PTLOCKS as a constant expression rather than repeating
"NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS" all over the place.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Acked-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
