2018-10-31 18:21:09 +00:00
|
|
|
// SPDX-License-Identifier: GPL-2.0
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2013-04-30 22:27:37 +00:00
|
|
|
* Kernel internal timers
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* Copyright (C) 1991, 1992 Linus Torvalds
|
|
|
|
*
|
|
|
|
* 1997-01-28 Modified by Finn Arne Gangstad to make timers scale better.
|
|
|
|
*
|
|
|
|
* 1997-09-10 Updated NTP code according to technical memorandum Jan '96
|
|
|
|
* "A Kernel Model for Precision Timekeeping" by Dave Mills
|
|
|
|
* 1998-12-24 Fixed a xtime SMP race (we need the xtime_lock rw spinlock to
|
|
|
|
* serialize accesses to xtime/lost_ticks).
|
|
|
|
* Copyright (C) 1998 Andrea Arcangeli
|
|
|
|
* 1999-03-10 Improved NTP compatibility by Ulrich Windl
|
|
|
|
* 2002-05-31 Move sys_sysinfo here and make its locking sane, Robert Love
|
|
|
|
* 2000-10-05 Implemented scalable SMP per-CPU timer handling.
|
|
|
|
* Copyright (C) 2000, 2001, 2002 Ingo Molnar
|
|
|
|
* Designed by David S. Miller, Alexey Kuznetsov and Ingo Molnar
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/kernel_stat.h>
|
2011-05-23 18:51:41 +00:00
|
|
|
#include <linux/export.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/interrupt.h>
|
|
|
|
#include <linux/percpu.h>
|
|
|
|
#include <linux/init.h>
|
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/swap.h>
|
2007-10-19 06:40:14 +00:00
|
|
|
#include <linux/pid_namespace.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <linux/notifier.h>
|
|
|
|
#include <linux/thread_info.h>
|
|
|
|
#include <linux/time.h>
|
|
|
|
#include <linux/jiffies.h>
|
|
|
|
#include <linux/posix-timers.h>
|
|
|
|
#include <linux/cpu.h>
|
|
|
|
#include <linux/syscalls.h>
|
2006-01-08 09:02:17 +00:00
|
|
|
#include <linux/delay.h>
|
2007-02-16 09:28:03 +00:00
|
|
|
#include <linux/tick.h>
|
[PATCH] Add debugging feature /proc/timer_stat
Add /proc/timer_stats support: debugging feature to profile timer expiration.
Both the starting site, process/PID and the expiration function is captured.
This allows the quick identification of timer event sources in a system.
Sample output:
# echo 1 > /proc/timer_stats
# cat /proc/timer_stats
Timer Stats Version: v0.1
Sample period: 4.010 s
24, 0 swapper hrtimer_stop_sched_tick (hrtimer_sched_tick)
11, 0 swapper sk_reset_timer (tcp_delack_timer)
6, 0 swapper hrtimer_stop_sched_tick (hrtimer_sched_tick)
2, 1 swapper queue_delayed_work_on (delayed_work_timer_fn)
17, 0 swapper hrtimer_restart_sched_tick (hrtimer_sched_tick)
2, 1 swapper queue_delayed_work_on (delayed_work_timer_fn)
4, 2050 pcscd do_nanosleep (hrtimer_wakeup)
5, 4179 sshd sk_reset_timer (tcp_write_timer)
4, 2248 yum-updatesd schedule_timeout (process_timeout)
18, 0 swapper hrtimer_restart_sched_tick (hrtimer_sched_tick)
3, 0 swapper sk_reset_timer (tcp_delack_timer)
1, 1 swapper neigh_table_init_no_netlink (neigh_periodic_timer)
2, 1 swapper e1000_up (e1000_watchdog)
1, 1 init schedule_timeout (process_timeout)
100 total events, 25.24 events/sec
[ cleanups and hrtimers support from Thomas Gleixner <tglx@linutronix.de> ]
[bunk@stusta.de: nr_entries can become static]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: john stultz <johnstul@us.ibm.com>
Cc: Roman Zippel <zippel@linux-m68k.org>
Cc: Andi Kleen <ak@suse.de>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-16 09:28:13 +00:00
|
|
|
#include <linux/kallsyms.h>
|
2010-10-14 06:01:34 +00:00
|
|
|
#include <linux/irq_work.h>
|
2017-02-02 18:15:33 +00:00
|
|
|
#include <linux/sched/signal.h>
|
2013-02-07 15:46:59 +00:00
|
|
|
#include <linux/sched/sysctl.h>
|
2017-02-08 17:51:35 +00:00
|
|
|
#include <linux/sched/nohz.h>
|
2017-02-08 17:51:35 +00:00
|
|
|
#include <linux/sched/debug.h>
|
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
|
|
|
#include <linux/slab.h>
|
2013-04-30 22:27:34 +00:00
|
|
|
#include <linux/compat.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2016-12-24 19:46:01 +00:00
|
|
|
#include <linux/uaccess.h>
|
2005-04-16 22:20:36 +00:00
|
|
|
#include <asm/unistd.h>
|
|
|
|
#include <asm/div64.h>
|
|
|
|
#include <asm/timex.h>
|
|
|
|
#include <asm/io.h>
|
|
|
|
|
2015-04-14 21:08:58 +00:00
|
|
|
#include "tick-internal.h"
|
|
|
|
|
2009-08-10 02:48:59 +00:00
|
|
|
#define CREATE_TRACE_POINTS
|
|
|
|
#include <trace/events/timer.h>
|
|
|
|
|
2014-02-08 07:51:59 +00:00
|
|
|
__visible u64 jiffies_64 __cacheline_aligned_in_smp = INITIAL_JIFFIES;
|
2005-10-30 23:03:00 +00:00
|
|
|
|
|
|
|
EXPORT_SYMBOL(jiffies_64);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* The timer wheel has LVL_DEPTH array levels. Each level provides an array of
|
|
|
|
* LVL_SIZE buckets. Each level is driven by its own clock and therefor each
|
|
|
|
* level has a different granularity.
|
|
|
|
*
|
|
|
|
* The level granularity is: LVL_CLK_DIV ^ lvl
|
|
|
|
* The level clock frequency is: HZ / (LVL_CLK_DIV ^ level)
|
|
|
|
*
|
|
|
|
* The array level of a newly armed timer depends on the relative expiry
|
|
|
|
* time. The farther the expiry time is away the higher the array level and
|
|
|
|
* therefor the granularity becomes.
|
|
|
|
*
|
|
|
|
* Contrary to the original timer wheel implementation, which aims for 'exact'
|
|
|
|
* expiry of the timers, this implementation removes the need for recascading
|
|
|
|
* the timers into the lower array levels. The previous 'classic' timer wheel
|
|
|
|
* implementation of the kernel already violated the 'exact' expiry by adding
|
|
|
|
* slack to the expiry time to provide batched expiration. The granularity
|
|
|
|
* levels provide implicit batching.
|
|
|
|
*
|
|
|
|
* This is an optimization of the original timer wheel implementation for the
|
|
|
|
* majority of the timer wheel use cases: timeouts. The vast majority of
|
|
|
|
* timeout timers (networking, disk I/O ...) are canceled before expiry. If
|
|
|
|
* the timeout expires it indicates that normal operation is disturbed, so it
|
|
|
|
* does not matter much whether the timeout comes with a slight delay.
|
|
|
|
*
|
|
|
|
* The only exception to this are networking timers with a small expiry
|
|
|
|
* time. They rely on the granularity. Those fit into the first wheel level,
|
|
|
|
* which has HZ granularity.
|
|
|
|
*
|
|
|
|
* We don't have cascading anymore. timers with a expiry time above the
|
|
|
|
* capacity of the last wheel level are force expired at the maximum timeout
|
|
|
|
* value of the last wheel level. From data sampling we know that the maximum
|
|
|
|
* value observed is 5 days (network connection tracking), so this should not
|
|
|
|
* be an issue.
|
|
|
|
*
|
|
|
|
* The currently chosen array constants values are a good compromise between
|
|
|
|
* array size and granularity.
|
|
|
|
*
|
|
|
|
* This results in the following granularity and range levels:
|
|
|
|
*
|
|
|
|
* HZ 1000 steps
|
|
|
|
* Level Offset Granularity Range
|
|
|
|
* 0 0 1 ms 0 ms - 63 ms
|
|
|
|
* 1 64 8 ms 64 ms - 511 ms
|
|
|
|
* 2 128 64 ms 512 ms - 4095 ms (512ms - ~4s)
|
|
|
|
* 3 192 512 ms 4096 ms - 32767 ms (~4s - ~32s)
|
|
|
|
* 4 256 4096 ms (~4s) 32768 ms - 262143 ms (~32s - ~4m)
|
|
|
|
* 5 320 32768 ms (~32s) 262144 ms - 2097151 ms (~4m - ~34m)
|
|
|
|
* 6 384 262144 ms (~4m) 2097152 ms - 16777215 ms (~34m - ~4h)
|
|
|
|
* 7 448 2097152 ms (~34m) 16777216 ms - 134217727 ms (~4h - ~1d)
|
|
|
|
* 8 512 16777216 ms (~4h) 134217728 ms - 1073741822 ms (~1d - ~12d)
|
|
|
|
*
|
|
|
|
* HZ 300
|
|
|
|
* Level Offset Granularity Range
|
|
|
|
* 0 0 3 ms 0 ms - 210 ms
|
|
|
|
* 1 64 26 ms 213 ms - 1703 ms (213ms - ~1s)
|
|
|
|
* 2 128 213 ms 1706 ms - 13650 ms (~1s - ~13s)
|
|
|
|
* 3 192 1706 ms (~1s) 13653 ms - 109223 ms (~13s - ~1m)
|
|
|
|
* 4 256 13653 ms (~13s) 109226 ms - 873810 ms (~1m - ~14m)
|
|
|
|
* 5 320 109226 ms (~1m) 873813 ms - 6990503 ms (~14m - ~1h)
|
|
|
|
* 6 384 873813 ms (~14m) 6990506 ms - 55924050 ms (~1h - ~15h)
|
|
|
|
* 7 448 6990506 ms (~1h) 55924053 ms - 447392423 ms (~15h - ~5d)
|
|
|
|
* 8 512 55924053 ms (~15h) 447392426 ms - 3579139406 ms (~5d - ~41d)
|
|
|
|
*
|
|
|
|
* HZ 250
|
|
|
|
* Level Offset Granularity Range
|
|
|
|
* 0 0 4 ms 0 ms - 255 ms
|
|
|
|
* 1 64 32 ms 256 ms - 2047 ms (256ms - ~2s)
|
|
|
|
* 2 128 256 ms 2048 ms - 16383 ms (~2s - ~16s)
|
|
|
|
* 3 192 2048 ms (~2s) 16384 ms - 131071 ms (~16s - ~2m)
|
|
|
|
* 4 256 16384 ms (~16s) 131072 ms - 1048575 ms (~2m - ~17m)
|
|
|
|
* 5 320 131072 ms (~2m) 1048576 ms - 8388607 ms (~17m - ~2h)
|
|
|
|
* 6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
|
|
|
|
* 7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
|
|
|
|
* 8 512 67108864 ms (~18h) 536870912 ms - 4294967288 ms (~6d - ~49d)
|
|
|
|
*
|
|
|
|
* HZ 100
|
|
|
|
* Level Offset Granularity Range
|
|
|
|
* 0 0 10 ms 0 ms - 630 ms
|
|
|
|
* 1 64 80 ms 640 ms - 5110 ms (640ms - ~5s)
|
|
|
|
* 2 128 640 ms 5120 ms - 40950 ms (~5s - ~40s)
|
|
|
|
* 3 192 5120 ms (~5s) 40960 ms - 327670 ms (~40s - ~5m)
|
|
|
|
* 4 256 40960 ms (~40s) 327680 ms - 2621430 ms (~5m - ~43m)
|
|
|
|
* 5 320 327680 ms (~5m) 2621440 ms - 20971510 ms (~43m - ~5h)
|
|
|
|
* 6 384 2621440 ms (~43m) 20971520 ms - 167772150 ms (~5h - ~1d)
|
|
|
|
* 7 448 20971520 ms (~5h) 167772160 ms - 1342177270 ms (~1d - ~15d)
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/* Clock divisor for the next level */
|
|
|
|
#define LVL_CLK_SHIFT 3
|
|
|
|
#define LVL_CLK_DIV (1UL << LVL_CLK_SHIFT)
|
|
|
|
#define LVL_CLK_MASK (LVL_CLK_DIV - 1)
|
|
|
|
#define LVL_SHIFT(n) ((n) * LVL_CLK_SHIFT)
|
|
|
|
#define LVL_GRAN(n) (1UL << LVL_SHIFT(n))
|
2005-04-16 22:20:36 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/*
|
|
|
|
* The time start value for each level to select the bucket at enqueue
|
|
|
|
* time.
|
|
|
|
*/
|
|
|
|
#define LVL_START(n) ((LVL_SIZE - 1) << (((n) - 1) * LVL_CLK_SHIFT))
|
|
|
|
|
|
|
|
/* Size of each clock level */
|
|
|
|
#define LVL_BITS 6
|
|
|
|
#define LVL_SIZE (1UL << LVL_BITS)
|
|
|
|
#define LVL_MASK (LVL_SIZE - 1)
|
|
|
|
#define LVL_OFFS(n) ((n) * LVL_SIZE)
|
|
|
|
|
|
|
|
/* Level depth */
|
|
|
|
#if HZ > 100
|
|
|
|
# define LVL_DEPTH 9
|
|
|
|
# else
|
|
|
|
# define LVL_DEPTH 8
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* The cutoff (max. capacity of the wheel) */
|
|
|
|
#define WHEEL_TIMEOUT_CUTOFF (LVL_START(LVL_DEPTH))
|
|
|
|
#define WHEEL_TIMEOUT_MAX (WHEEL_TIMEOUT_CUTOFF - LVL_GRAN(LVL_DEPTH - 1))
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The resulting wheel size. If NOHZ is configured we allocate two
|
|
|
|
* wheels so we have a separate storage for the deferrable timers.
|
|
|
|
*/
|
|
|
|
#define WHEEL_SIZE (LVL_SIZE * LVL_DEPTH)
|
|
|
|
|
|
|
|
#ifdef CONFIG_NO_HZ_COMMON
|
|
|
|
# define NR_BASES 2
|
|
|
|
# define BASE_STD 0
|
|
|
|
# define BASE_DEF 1
|
|
|
|
#else
|
|
|
|
# define NR_BASES 1
|
|
|
|
# define BASE_STD 0
|
|
|
|
# define BASE_DEF 0
|
|
|
|
#endif
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base {
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spinlock_t lock;
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
struct timer_list *running_timer;
|
2019-07-26 18:31:00 +00:00
|
|
|
#ifdef CONFIG_PREEMPT_RT
|
|
|
|
spinlock_t expiry_lock;
|
|
|
|
atomic_t timer_waiters;
|
|
|
|
#endif
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned long clk;
|
2016-07-04 09:50:36 +00:00
|
|
|
unsigned long next_expiry;
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned int cpu;
|
2016-07-04 09:50:36 +00:00
|
|
|
bool is_idle;
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
bool must_forward_clk;
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
DECLARE_BITMAP(pending_map, WHEEL_SIZE);
|
|
|
|
struct hlist_head vectors[WHEEL_SIZE];
|
2007-05-08 07:27:44 +00:00
|
|
|
} ____cacheline_aligned;
|
2012-08-08 18:10:25 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static DEFINE_PER_CPU(struct timer_base, timer_bases[NR_BASES]);
|
2007-05-08 07:27:44 +00:00
|
|
|
|
2018-01-14 22:30:51 +00:00
|
|
|
#ifdef CONFIG_NO_HZ_COMMON
|
|
|
|
|
2017-12-21 10:41:49 +00:00
|
|
|
static DEFINE_STATIC_KEY_FALSE(timers_nohz_active);
|
2018-01-14 22:30:51 +00:00
|
|
|
static DEFINE_MUTEX(timer_keys_mutex);
|
|
|
|
|
|
|
|
static void timer_update_keys(struct work_struct *work);
|
|
|
|
static DECLARE_WORK(timer_update_work, timer_update_keys);
|
|
|
|
|
|
|
|
#ifdef CONFIG_SMP
|
2015-05-26 22:50:33 +00:00
|
|
|
unsigned int sysctl_timer_migration = 1;
|
|
|
|
|
2018-01-14 22:30:51 +00:00
|
|
|
DEFINE_STATIC_KEY_FALSE(timers_migration_enabled);
|
|
|
|
|
|
|
|
static void timers_update_migration(void)
|
2015-05-26 22:50:33 +00:00
|
|
|
{
|
2018-01-14 22:30:51 +00:00
|
|
|
if (sysctl_timer_migration && tick_nohz_active)
|
|
|
|
static_branch_enable(&timers_migration_enabled);
|
|
|
|
else
|
|
|
|
static_branch_disable(&timers_migration_enabled);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static inline void timers_update_migration(void) { }
|
|
|
|
#endif /* !CONFIG_SMP */
|
2015-05-26 22:50:33 +00:00
|
|
|
|
2018-01-14 22:30:51 +00:00
|
|
|
static void timer_update_keys(struct work_struct *work)
|
|
|
|
{
|
|
|
|
mutex_lock(&timer_keys_mutex);
|
|
|
|
timers_update_migration();
|
|
|
|
static_branch_enable(&timers_nohz_active);
|
|
|
|
mutex_unlock(&timer_keys_mutex);
|
|
|
|
}
|
2015-05-26 22:50:33 +00:00
|
|
|
|
2018-01-14 22:30:51 +00:00
|
|
|
void timers_update_nohz(void)
|
|
|
|
{
|
|
|
|
schedule_work(&timer_update_work);
|
2015-05-26 22:50:33 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
int timer_migration_handler(struct ctl_table *table, int write,
|
2020-04-24 06:43:38 +00:00
|
|
|
void *buffer, size_t *lenp, loff_t *ppos)
|
2015-05-26 22:50:33 +00:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2018-01-14 22:30:51 +00:00
|
|
|
mutex_lock(&timer_keys_mutex);
|
2017-04-19 22:24:50 +00:00
|
|
|
ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
2015-05-26 22:50:33 +00:00
|
|
|
if (!ret && write)
|
2018-01-14 22:30:51 +00:00
|
|
|
timers_update_migration();
|
|
|
|
mutex_unlock(&timer_keys_mutex);
|
2015-05-26 22:50:33 +00:00
|
|
|
return ret;
|
|
|
|
}
|
2017-12-21 10:41:49 +00:00
|
|
|
|
|
|
|
static inline bool is_timers_nohz_active(void)
|
|
|
|
{
|
|
|
|
return static_branch_unlikely(&timers_nohz_active);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static inline bool is_timers_nohz_active(void) { return false; }
|
2018-01-14 22:30:51 +00:00
|
|
|
#endif /* NO_HZ_COMMON */
|
2015-05-26 22:50:33 +00:00
|
|
|
|
2008-11-06 07:42:48 +00:00
|
|
|
static unsigned long round_jiffies_common(unsigned long j, int cpu,
|
|
|
|
bool force_up)
|
2006-12-10 10:21:24 +00:00
|
|
|
{
|
|
|
|
int rem;
|
|
|
|
unsigned long original = j;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We don't want all cpus firing their timers at once hitting the
|
|
|
|
* same lock or cachelines, so we skew each extra cpu with an extra
|
|
|
|
* 3 jiffies. This 3 jiffies came originally from the mm/ code which
|
|
|
|
* already did this.
|
|
|
|
* The skew is done by adding 3*cpunr, then round, then subtract this
|
|
|
|
* extra offset again.
|
|
|
|
*/
|
|
|
|
j += cpu * 3;
|
|
|
|
|
|
|
|
rem = j % HZ;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the target jiffie is just after a whole second (which can happen
|
|
|
|
* due to delays of the timer irq, long irq off times etc etc) then
|
|
|
|
* we should round down to the whole second, not up. Use 1/4th second
|
|
|
|
* as cutoff for this rounding as an extreme upper bound for this.
|
2008-11-06 07:42:48 +00:00
|
|
|
* But never round down if @force_up is set.
|
2006-12-10 10:21:24 +00:00
|
|
|
*/
|
2008-11-06 07:42:48 +00:00
|
|
|
if (rem < HZ/4 && !force_up) /* round down */
|
2006-12-10 10:21:24 +00:00
|
|
|
j = j - rem;
|
|
|
|
else /* round up */
|
|
|
|
j = j - rem + HZ;
|
|
|
|
|
|
|
|
/* now that we have rounded, subtract the extra skew again */
|
|
|
|
j -= cpu * 3;
|
|
|
|
|
2013-05-21 18:43:50 +00:00
|
|
|
/*
|
|
|
|
* Make sure j is still in the future. Otherwise return the
|
|
|
|
* unmodified value.
|
|
|
|
*/
|
|
|
|
return time_is_after_jiffies(j) ? j : original;
|
2006-12-10 10:21:24 +00:00
|
|
|
}
|
2008-11-06 07:42:48 +00:00
|
|
|
|
|
|
|
/**
|
|
|
|
* __round_jiffies - function to round jiffies to a full second
|
|
|
|
* @j: the time in (absolute) jiffies that should be rounded
|
|
|
|
* @cpu: the processor number on which the timeout will happen
|
|
|
|
*
|
|
|
|
* __round_jiffies() rounds an absolute time in the future (in jiffies)
|
|
|
|
* up or down to (approximately) full seconds. This is useful for timers
|
|
|
|
* for which the exact time they fire does not matter too much, as long as
|
|
|
|
* they fire approximately every X seconds.
|
|
|
|
*
|
|
|
|
* By rounding these timers to whole seconds, all such timers will fire
|
|
|
|
* at the same time, rather than at various times spread out. The goal
|
|
|
|
* of this is to have the CPU wake up less, which saves power.
|
|
|
|
*
|
|
|
|
* The exact rounding is skewed for each processor to avoid all
|
|
|
|
* processors firing at the exact same time, which could lead
|
|
|
|
* to lock contention or spurious cache line bouncing.
|
|
|
|
*
|
|
|
|
* The return value is the rounded version of the @j parameter.
|
|
|
|
*/
|
|
|
|
unsigned long __round_jiffies(unsigned long j, int cpu)
|
|
|
|
{
|
|
|
|
return round_jiffies_common(j, cpu, false);
|
|
|
|
}
|
2006-12-10 10:21:24 +00:00
|
|
|
EXPORT_SYMBOL_GPL(__round_jiffies);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* __round_jiffies_relative - function to round jiffies to a full second
|
|
|
|
* @j: the time in (relative) jiffies that should be rounded
|
|
|
|
* @cpu: the processor number on which the timeout will happen
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* __round_jiffies_relative() rounds a time delta in the future (in jiffies)
|
2006-12-10 10:21:24 +00:00
|
|
|
* up or down to (approximately) full seconds. This is useful for timers
|
|
|
|
* for which the exact time they fire does not matter too much, as long as
|
|
|
|
* they fire approximately every X seconds.
|
|
|
|
*
|
|
|
|
* By rounding these timers to whole seconds, all such timers will fire
|
|
|
|
* at the same time, rather than at various times spread out. The goal
|
|
|
|
* of this is to have the CPU wake up less, which saves power.
|
|
|
|
*
|
|
|
|
* The exact rounding is skewed for each processor to avoid all
|
|
|
|
* processors firing at the exact same time, which could lead
|
|
|
|
* to lock contention or spurious cache line bouncing.
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* The return value is the rounded version of the @j parameter.
|
2006-12-10 10:21:24 +00:00
|
|
|
*/
|
|
|
|
unsigned long __round_jiffies_relative(unsigned long j, int cpu)
|
|
|
|
{
|
2008-11-06 07:42:48 +00:00
|
|
|
unsigned long j0 = jiffies;
|
|
|
|
|
|
|
|
/* Use j0 because jiffies might change while we run */
|
|
|
|
return round_jiffies_common(j + j0, cpu, false) - j0;
|
2006-12-10 10:21:24 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(__round_jiffies_relative);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* round_jiffies - function to round jiffies to a full second
|
|
|
|
* @j: the time in (absolute) jiffies that should be rounded
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* round_jiffies() rounds an absolute time in the future (in jiffies)
|
2006-12-10 10:21:24 +00:00
|
|
|
* up or down to (approximately) full seconds. This is useful for timers
|
|
|
|
* for which the exact time they fire does not matter too much, as long as
|
|
|
|
* they fire approximately every X seconds.
|
|
|
|
*
|
|
|
|
* By rounding these timers to whole seconds, all such timers will fire
|
|
|
|
* at the same time, rather than at various times spread out. The goal
|
|
|
|
* of this is to have the CPU wake up less, which saves power.
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* The return value is the rounded version of the @j parameter.
|
2006-12-10 10:21:24 +00:00
|
|
|
*/
|
|
|
|
unsigned long round_jiffies(unsigned long j)
|
|
|
|
{
|
2008-11-06 07:42:48 +00:00
|
|
|
return round_jiffies_common(j, raw_smp_processor_id(), false);
|
2006-12-10 10:21:24 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(round_jiffies);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* round_jiffies_relative - function to round jiffies to a full second
|
|
|
|
* @j: the time in (relative) jiffies that should be rounded
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* round_jiffies_relative() rounds a time delta in the future (in jiffies)
|
2006-12-10 10:21:24 +00:00
|
|
|
* up or down to (approximately) full seconds. This is useful for timers
|
|
|
|
* for which the exact time they fire does not matter too much, as long as
|
|
|
|
* they fire approximately every X seconds.
|
|
|
|
*
|
|
|
|
* By rounding these timers to whole seconds, all such timers will fire
|
|
|
|
* at the same time, rather than at various times spread out. The goal
|
|
|
|
* of this is to have the CPU wake up less, which saves power.
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* The return value is the rounded version of the @j parameter.
|
2006-12-10 10:21:24 +00:00
|
|
|
*/
|
|
|
|
unsigned long round_jiffies_relative(unsigned long j)
|
|
|
|
{
|
|
|
|
return __round_jiffies_relative(j, raw_smp_processor_id());
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(round_jiffies_relative);
|
|
|
|
|
2008-11-06 07:42:48 +00:00
|
|
|
/**
|
|
|
|
* __round_jiffies_up - function to round jiffies up to a full second
|
|
|
|
* @j: the time in (absolute) jiffies that should be rounded
|
|
|
|
* @cpu: the processor number on which the timeout will happen
|
|
|
|
*
|
|
|
|
* This is the same as __round_jiffies() except that it will never
|
|
|
|
* round down. This is useful for timeouts for which the exact time
|
|
|
|
* of firing does not matter too much, as long as they don't fire too
|
|
|
|
* early.
|
|
|
|
*/
|
|
|
|
unsigned long __round_jiffies_up(unsigned long j, int cpu)
|
|
|
|
{
|
|
|
|
return round_jiffies_common(j, cpu, true);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(__round_jiffies_up);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* __round_jiffies_up_relative - function to round jiffies up to a full second
|
|
|
|
* @j: the time in (relative) jiffies that should be rounded
|
|
|
|
* @cpu: the processor number on which the timeout will happen
|
|
|
|
*
|
|
|
|
* This is the same as __round_jiffies_relative() except that it will never
|
|
|
|
* round down. This is useful for timeouts for which the exact time
|
|
|
|
* of firing does not matter too much, as long as they don't fire too
|
|
|
|
* early.
|
|
|
|
*/
|
|
|
|
unsigned long __round_jiffies_up_relative(unsigned long j, int cpu)
|
|
|
|
{
|
|
|
|
unsigned long j0 = jiffies;
|
|
|
|
|
|
|
|
/* Use j0 because jiffies might change while we run */
|
|
|
|
return round_jiffies_common(j + j0, cpu, true) - j0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(__round_jiffies_up_relative);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* round_jiffies_up - function to round jiffies up to a full second
|
|
|
|
* @j: the time in (absolute) jiffies that should be rounded
|
|
|
|
*
|
|
|
|
* This is the same as round_jiffies() except that it will never
|
|
|
|
* round down. This is useful for timeouts for which the exact time
|
|
|
|
* of firing does not matter too much, as long as they don't fire too
|
|
|
|
* early.
|
|
|
|
*/
|
|
|
|
unsigned long round_jiffies_up(unsigned long j)
|
|
|
|
{
|
|
|
|
return round_jiffies_common(j, raw_smp_processor_id(), true);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(round_jiffies_up);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* round_jiffies_up_relative - function to round jiffies up to a full second
|
|
|
|
* @j: the time in (relative) jiffies that should be rounded
|
|
|
|
*
|
|
|
|
* This is the same as round_jiffies_relative() except that it will never
|
|
|
|
* round down. This is useful for timeouts for which the exact time
|
|
|
|
* of firing does not matter too much, as long as they don't fire too
|
|
|
|
* early.
|
|
|
|
*/
|
|
|
|
unsigned long round_jiffies_up_relative(unsigned long j)
|
|
|
|
{
|
|
|
|
return __round_jiffies_up_relative(j, raw_smp_processor_id());
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(round_jiffies_up_relative);
|
|
|
|
|
2010-03-11 22:04:36 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static inline unsigned int timer_get_idx(struct timer_list *timer)
|
2010-03-11 22:04:36 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
return (timer->flags & TIMER_ARRAYMASK) >> TIMER_ARRAYSHIFT;
|
2010-03-11 22:04:36 +00:00
|
|
|
}
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static inline void timer_set_idx(struct timer_list *timer, unsigned int idx)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
timer->flags = (timer->flags & ~TIMER_ARRAYMASK) |
|
|
|
|
idx << TIMER_ARRAYSHIFT;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/*
|
|
|
|
* Helper function to calculate the array index for a given expiry
|
|
|
|
* time.
|
|
|
|
*/
|
2020-07-17 14:05:41 +00:00
|
|
|
static inline unsigned calc_index(unsigned long expires, unsigned lvl)
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
{
|
|
|
|
expires = (expires + LVL_GRAN(lvl)) >> LVL_SHIFT(lvl);
|
|
|
|
return LVL_OFFS(lvl) + (expires & LVL_MASK);
|
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:39 +00:00
|
|
|
static int calc_wheel_index(unsigned long expires, unsigned long clk)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2016-07-04 09:50:39 +00:00
|
|
|
unsigned long delta = expires - clk;
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned int idx;
|
|
|
|
|
|
|
|
if (delta < LVL_START(1)) {
|
|
|
|
idx = calc_index(expires, 0);
|
|
|
|
} else if (delta < LVL_START(2)) {
|
|
|
|
idx = calc_index(expires, 1);
|
|
|
|
} else if (delta < LVL_START(3)) {
|
|
|
|
idx = calc_index(expires, 2);
|
|
|
|
} else if (delta < LVL_START(4)) {
|
|
|
|
idx = calc_index(expires, 3);
|
|
|
|
} else if (delta < LVL_START(5)) {
|
|
|
|
idx = calc_index(expires, 4);
|
|
|
|
} else if (delta < LVL_START(6)) {
|
|
|
|
idx = calc_index(expires, 5);
|
|
|
|
} else if (delta < LVL_START(7)) {
|
|
|
|
idx = calc_index(expires, 6);
|
|
|
|
} else if (LVL_DEPTH > 8 && delta < LVL_START(8)) {
|
|
|
|
idx = calc_index(expires, 7);
|
|
|
|
} else if ((long) delta < 0) {
|
2016-07-04 09:50:39 +00:00
|
|
|
idx = clk & LVL_MASK;
|
2005-04-16 22:20:36 +00:00
|
|
|
} else {
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/*
|
|
|
|
* Force expire obscene large timeouts to expire at the
|
|
|
|
* capacity limit of the wheel.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2020-07-17 14:05:40 +00:00
|
|
|
if (delta >= WHEEL_TIMEOUT_CUTOFF)
|
|
|
|
expires = clk + WHEEL_TIMEOUT_MAX;
|
2015-05-26 22:50:26 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
idx = calc_index(expires, LVL_DEPTH - 1);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2016-07-04 09:50:39 +00:00
|
|
|
return idx;
|
|
|
|
}
|
2015-05-26 22:50:26 +00:00
|
|
|
|
2016-07-04 09:50:39 +00:00
|
|
|
/*
|
|
|
|
* Enqueue the timer into the hash bucket, mark it pending in
|
|
|
|
* the bitmap and store the index in the timer flags.
|
|
|
|
*/
|
|
|
|
static void enqueue_timer(struct timer_base *base, struct timer_list *timer,
|
|
|
|
unsigned int idx)
|
|
|
|
{
|
|
|
|
hlist_add_head(&timer->entry, base->vectors + idx);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
__set_bit(idx, base->pending_map);
|
|
|
|
timer_set_idx(timer, idx);
|
2019-03-21 12:09:19 +00:00
|
|
|
|
|
|
|
trace_timer_start(timer, timer->expires, timer->flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:39 +00:00
|
|
|
static void
|
|
|
|
__internal_add_timer(struct timer_base *base, struct timer_list *timer)
|
2012-05-25 22:08:57 +00:00
|
|
|
{
|
2016-07-04 09:50:39 +00:00
|
|
|
unsigned int idx;
|
|
|
|
|
|
|
|
idx = calc_wheel_index(timer->expires, base->clk);
|
|
|
|
enqueue_timer(base, timer, idx);
|
|
|
|
}
|
2014-06-21 23:29:14 +00:00
|
|
|
|
2016-07-04 09:50:39 +00:00
|
|
|
static void
|
|
|
|
trigger_dyntick_cpu(struct timer_base *base, struct timer_list *timer)
|
|
|
|
{
|
2018-01-14 22:30:51 +00:00
|
|
|
if (!is_timers_nohz_active())
|
2016-07-04 09:50:36 +00:00
|
|
|
return;
|
2015-05-26 22:50:24 +00:00
|
|
|
|
2012-05-25 22:08:57 +00:00
|
|
|
/*
|
2016-07-04 09:50:36 +00:00
|
|
|
* TODO: This wants some optimizing similar to the code below, but we
|
|
|
|
* will do that when we switch from push to pull for deferrable timers.
|
2012-05-25 22:08:57 +00:00
|
|
|
*/
|
2016-07-04 09:50:36 +00:00
|
|
|
if (timer->flags & TIMER_DEFERRABLE) {
|
|
|
|
if (tick_nohz_full_cpu(base->cpu))
|
2015-05-26 22:50:35 +00:00
|
|
|
wake_up_nohz_cpu(base->cpu);
|
2016-07-04 09:50:36 +00:00
|
|
|
return;
|
2012-05-25 22:08:58 +00:00
|
|
|
}
|
2014-06-21 23:29:14 +00:00
|
|
|
|
|
|
|
/*
|
2016-07-04 09:50:36 +00:00
|
|
|
* We might have to IPI the remote CPU if the base is idle and the
|
|
|
|
* timer is not deferrable. If the other CPU is on the way to idle
|
|
|
|
* then it can't set base->is_idle as we hold the base lock:
|
2014-06-21 23:29:14 +00:00
|
|
|
*/
|
2016-07-04 09:50:36 +00:00
|
|
|
if (!base->is_idle)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* Check whether this is the new first expiring timer: */
|
|
|
|
if (time_after_eq(timer->expires, base->next_expiry))
|
|
|
|
return;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Set the next expiry time and kick the CPU so it can reevaluate the
|
|
|
|
* wheel:
|
|
|
|
*/
|
2020-07-03 01:06:57 +00:00
|
|
|
if (time_before(timer->expires, base->clk)) {
|
|
|
|
/*
|
|
|
|
* Prevent from forward_timer_base() moving the base->clk
|
|
|
|
* backward
|
|
|
|
*/
|
|
|
|
base->next_expiry = base->clk;
|
|
|
|
} else {
|
|
|
|
base->next_expiry = timer->expires;
|
|
|
|
}
|
2018-07-16 06:08:57 +00:00
|
|
|
wake_up_nohz_cpu(base->cpu);
|
2016-07-04 09:50:39 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
internal_add_timer(struct timer_base *base, struct timer_list *timer)
|
|
|
|
{
|
|
|
|
__internal_add_timer(base, timer);
|
|
|
|
trigger_dyntick_cpu(base, timer);
|
2012-05-25 22:08:57 +00:00
|
|
|
}
|
|
|
|
|
2008-04-30 07:55:03 +00:00
|
|
|
#ifdef CONFIG_DEBUG_OBJECTS_TIMERS
|
|
|
|
|
|
|
|
static struct debug_obj_descr timer_debug_descr;
|
|
|
|
|
2011-03-07 08:58:33 +00:00
|
|
|
static void *timer_debug_hint(void *addr)
|
|
|
|
{
|
|
|
|
return ((struct timer_list *) addr)->function;
|
|
|
|
}
|
|
|
|
|
2016-05-20 00:09:41 +00:00
|
|
|
static bool timer_is_static_object(void *addr)
|
|
|
|
{
|
|
|
|
struct timer_list *timer = addr;
|
|
|
|
|
|
|
|
return (timer->entry.pprev == NULL &&
|
|
|
|
timer->entry.next == TIMER_ENTRY_STATIC);
|
|
|
|
}
|
|
|
|
|
2008-04-30 07:55:03 +00:00
|
|
|
/*
|
|
|
|
* fixup_init is called when:
|
|
|
|
* - an active object is initialized
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
*/
|
2016-05-20 00:09:29 +00:00
|
|
|
static bool timer_fixup_init(void *addr, enum debug_obj_state state)
|
2008-04-30 07:55:03 +00:00
|
|
|
{
|
|
|
|
struct timer_list *timer = addr;
|
|
|
|
|
|
|
|
switch (state) {
|
|
|
|
case ODEBUG_STATE_ACTIVE:
|
|
|
|
del_timer_sync(timer);
|
|
|
|
debug_object_init(timer, &timer_debug_descr);
|
2016-05-20 00:09:29 +00:00
|
|
|
return true;
|
2008-04-30 07:55:03 +00:00
|
|
|
default:
|
2016-05-20 00:09:29 +00:00
|
|
|
return false;
|
2008-04-30 07:55:03 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-11-08 03:48:26 +00:00
|
|
|
/* Stub timer callback for improperly used timers. */
|
2017-10-18 14:10:19 +00:00
|
|
|
static void stub_timer(struct timer_list *unused)
|
2011-11-08 03:48:26 +00:00
|
|
|
{
|
|
|
|
WARN_ON(1);
|
|
|
|
}
|
|
|
|
|
2008-04-30 07:55:03 +00:00
|
|
|
/*
|
|
|
|
* fixup_activate is called when:
|
|
|
|
* - an active object is activated
|
2016-05-20 00:09:41 +00:00
|
|
|
* - an unknown non-static object is activated
|
2008-04-30 07:55:03 +00:00
|
|
|
*/
|
2016-05-20 00:09:29 +00:00
|
|
|
static bool timer_fixup_activate(void *addr, enum debug_obj_state state)
|
2008-04-30 07:55:03 +00:00
|
|
|
{
|
|
|
|
struct timer_list *timer = addr;
|
|
|
|
|
|
|
|
switch (state) {
|
|
|
|
case ODEBUG_STATE_NOTAVAILABLE:
|
2017-10-18 14:10:19 +00:00
|
|
|
timer_setup(timer, stub_timer, 0);
|
2016-05-20 00:09:41 +00:00
|
|
|
return true;
|
2008-04-30 07:55:03 +00:00
|
|
|
|
|
|
|
case ODEBUG_STATE_ACTIVE:
|
|
|
|
WARN_ON(1);
|
2019-01-23 08:14:13 +00:00
|
|
|
/* fall through */
|
2008-04-30 07:55:03 +00:00
|
|
|
default:
|
2016-05-20 00:09:29 +00:00
|
|
|
return false;
|
2008-04-30 07:55:03 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* fixup_free is called when:
|
|
|
|
* - an active object is freed
|
|
|
|
*/
|
2016-05-20 00:09:29 +00:00
|
|
|
static bool timer_fixup_free(void *addr, enum debug_obj_state state)
|
2008-04-30 07:55:03 +00:00
|
|
|
{
|
|
|
|
struct timer_list *timer = addr;
|
|
|
|
|
|
|
|
switch (state) {
|
|
|
|
case ODEBUG_STATE_ACTIVE:
|
|
|
|
del_timer_sync(timer);
|
|
|
|
debug_object_free(timer, &timer_debug_descr);
|
2016-05-20 00:09:29 +00:00
|
|
|
return true;
|
2008-04-30 07:55:03 +00:00
|
|
|
default:
|
2016-05-20 00:09:29 +00:00
|
|
|
return false;
|
2008-04-30 07:55:03 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-11-08 03:48:28 +00:00
|
|
|
/*
|
|
|
|
* fixup_assert_init is called when:
|
|
|
|
* - an untracked/uninit-ed object is found
|
|
|
|
*/
|
2016-05-20 00:09:29 +00:00
|
|
|
static bool timer_fixup_assert_init(void *addr, enum debug_obj_state state)
|
2011-11-08 03:48:28 +00:00
|
|
|
{
|
|
|
|
struct timer_list *timer = addr;
|
|
|
|
|
|
|
|
switch (state) {
|
|
|
|
case ODEBUG_STATE_NOTAVAILABLE:
|
2017-10-18 14:10:19 +00:00
|
|
|
timer_setup(timer, stub_timer, 0);
|
2016-05-20 00:09:41 +00:00
|
|
|
return true;
|
2011-11-08 03:48:28 +00:00
|
|
|
default:
|
2016-05-20 00:09:29 +00:00
|
|
|
return false;
|
2011-11-08 03:48:28 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-04-30 07:55:03 +00:00
|
|
|
static struct debug_obj_descr timer_debug_descr = {
|
2011-11-08 03:48:28 +00:00
|
|
|
.name = "timer_list",
|
|
|
|
.debug_hint = timer_debug_hint,
|
2016-05-20 00:09:41 +00:00
|
|
|
.is_static_object = timer_is_static_object,
|
2011-11-08 03:48:28 +00:00
|
|
|
.fixup_init = timer_fixup_init,
|
|
|
|
.fixup_activate = timer_fixup_activate,
|
|
|
|
.fixup_free = timer_fixup_free,
|
|
|
|
.fixup_assert_init = timer_fixup_assert_init,
|
2008-04-30 07:55:03 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
static inline void debug_timer_init(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_init(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void debug_timer_activate(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_activate(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void debug_timer_deactivate(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_deactivate(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void debug_timer_free(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_free(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
|
2011-11-08 03:48:28 +00:00
|
|
|
static inline void debug_timer_assert_init(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_assert_init(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
|
2017-10-23 01:14:46 +00:00
|
|
|
static void do_init_timer(struct timer_list *timer,
|
|
|
|
void (*func)(struct timer_list *),
|
|
|
|
unsigned int flags,
|
2012-08-08 18:10:27 +00:00
|
|
|
const char *name, struct lock_class_key *key);
|
2008-04-30 07:55:03 +00:00
|
|
|
|
2017-10-23 01:14:46 +00:00
|
|
|
void init_timer_on_stack_key(struct timer_list *timer,
|
|
|
|
void (*func)(struct timer_list *),
|
|
|
|
unsigned int flags,
|
2012-08-08 18:10:27 +00:00
|
|
|
const char *name, struct lock_class_key *key)
|
2008-04-30 07:55:03 +00:00
|
|
|
{
|
|
|
|
debug_object_init_on_stack(timer, &timer_debug_descr);
|
2017-10-23 01:14:46 +00:00
|
|
|
do_init_timer(timer, func, flags, name, key);
|
2008-04-30 07:55:03 +00:00
|
|
|
}
|
2009-01-29 15:03:20 +00:00
|
|
|
EXPORT_SYMBOL_GPL(init_timer_on_stack_key);
|
2008-04-30 07:55:03 +00:00
|
|
|
|
|
|
|
void destroy_timer_on_stack(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_object_free(timer, &timer_debug_descr);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_GPL(destroy_timer_on_stack);
|
|
|
|
|
|
|
|
#else
|
|
|
|
static inline void debug_timer_init(struct timer_list *timer) { }
|
|
|
|
static inline void debug_timer_activate(struct timer_list *timer) { }
|
|
|
|
static inline void debug_timer_deactivate(struct timer_list *timer) { }
|
2011-11-08 03:48:28 +00:00
|
|
|
static inline void debug_timer_assert_init(struct timer_list *timer) { }
|
2008-04-30 07:55:03 +00:00
|
|
|
#endif
|
|
|
|
|
2009-08-10 02:48:59 +00:00
|
|
|
static inline void debug_init(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_timer_init(timer);
|
|
|
|
trace_timer_init(timer);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void debug_deactivate(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_timer_deactivate(timer);
|
|
|
|
trace_timer_cancel(timer);
|
|
|
|
}
|
|
|
|
|
2011-11-08 03:48:28 +00:00
|
|
|
static inline void debug_assert_init(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
debug_timer_assert_init(timer);
|
|
|
|
}
|
|
|
|
|
2017-10-23 01:14:46 +00:00
|
|
|
static void do_init_timer(struct timer_list *timer,
|
|
|
|
void (*func)(struct timer_list *),
|
|
|
|
unsigned int flags,
|
2012-08-08 18:10:27 +00:00
|
|
|
const char *name, struct lock_class_key *key)
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
{
|
2015-05-26 22:50:28 +00:00
|
|
|
timer->entry.pprev = NULL;
|
2017-10-23 01:14:46 +00:00
|
|
|
timer->function = func;
|
2015-05-26 22:50:29 +00:00
|
|
|
timer->flags = flags | raw_smp_processor_id();
|
2009-01-29 15:03:20 +00:00
|
|
|
lockdep_init_map(&timer->lockdep_map, name, key, 0);
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
}
|
2008-04-30 07:55:03 +00:00
|
|
|
|
|
|
|
/**
|
2009-04-02 00:47:23 +00:00
|
|
|
* init_timer_key - initialize a timer
|
2008-04-30 07:55:03 +00:00
|
|
|
* @timer: the timer to be initialized
|
2017-10-23 01:14:46 +00:00
|
|
|
* @func: timer callback function
|
2012-08-08 18:10:27 +00:00
|
|
|
* @flags: timer flags
|
2009-04-02 00:47:23 +00:00
|
|
|
* @name: name of the timer
|
|
|
|
* @key: lockdep class key of the fake lock used for tracking timer
|
|
|
|
* sync lock dependencies
|
2008-04-30 07:55:03 +00:00
|
|
|
*
|
2009-04-02 00:47:23 +00:00
|
|
|
* init_timer_key() must be done to a timer prior calling *any* of the
|
2008-04-30 07:55:03 +00:00
|
|
|
* other timer functions.
|
|
|
|
*/
|
2017-10-23 01:14:46 +00:00
|
|
|
void init_timer_key(struct timer_list *timer,
|
|
|
|
void (*func)(struct timer_list *), unsigned int flags,
|
2012-08-08 18:10:27 +00:00
|
|
|
const char *name, struct lock_class_key *key)
|
2008-04-30 07:55:03 +00:00
|
|
|
{
|
2009-08-10 02:48:59 +00:00
|
|
|
debug_init(timer);
|
2017-10-23 01:14:46 +00:00
|
|
|
do_init_timer(timer, func, flags, name, key);
|
2008-04-30 07:55:03 +00:00
|
|
|
}
|
2009-01-29 15:03:20 +00:00
|
|
|
EXPORT_SYMBOL(init_timer_key);
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
|
2012-05-25 22:08:57 +00:00
|
|
|
static inline void detach_timer(struct timer_list *timer, bool clear_pending)
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
{
|
2015-05-26 22:50:28 +00:00
|
|
|
struct hlist_node *entry = &timer->entry;
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
|
2009-08-10 02:48:59 +00:00
|
|
|
debug_deactivate(timer);
|
2008-04-30 07:55:03 +00:00
|
|
|
|
2015-05-26 22:50:28 +00:00
|
|
|
__hlist_del(entry);
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
if (clear_pending)
|
2015-05-26 22:50:28 +00:00
|
|
|
entry->pprev = NULL;
|
|
|
|
entry->next = LIST_POISON2;
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:28 +00:00
|
|
|
static int detach_if_pending(struct timer_list *timer, struct timer_base *base,
|
2012-05-25 22:08:57 +00:00
|
|
|
bool clear_pending)
|
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned idx = timer_get_idx(timer);
|
|
|
|
|
2012-05-25 22:08:57 +00:00
|
|
|
if (!timer_pending(timer))
|
|
|
|
return 0;
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
if (hlist_is_singular_node(&timer->entry, base->vectors + idx))
|
|
|
|
__clear_bit(idx, base->pending_map);
|
|
|
|
|
2012-05-25 22:08:57 +00:00
|
|
|
detach_timer(timer, clear_pending);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static inline struct timer_base *get_timer_cpu_base(u32 tflags, u32 cpu)
|
|
|
|
{
|
|
|
|
struct timer_base *base = per_cpu_ptr(&timer_bases[BASE_STD], cpu);
|
|
|
|
|
|
|
|
/*
|
2017-12-22 14:51:12 +00:00
|
|
|
* If the timer is deferrable and NO_HZ_COMMON is set then we need
|
|
|
|
* to use the deferrable base.
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
*/
|
2017-12-22 14:51:12 +00:00
|
|
|
if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
base = per_cpu_ptr(&timer_bases[BASE_DEF], cpu);
|
|
|
|
return base;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct timer_base *get_timer_this_cpu_base(u32 tflags)
|
|
|
|
{
|
|
|
|
struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
|
|
|
|
|
|
|
|
/*
|
2017-12-22 14:51:12 +00:00
|
|
|
* If the timer is deferrable and NO_HZ_COMMON is set then we need
|
|
|
|
* to use the deferrable base.
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
*/
|
2017-12-22 14:51:12 +00:00
|
|
|
if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && (tflags & TIMER_DEFERRABLE))
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
base = this_cpu_ptr(&timer_bases[BASE_DEF]);
|
|
|
|
return base;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct timer_base *get_timer_base(u32 tflags)
|
|
|
|
{
|
|
|
|
return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK);
|
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:36 +00:00
|
|
|
static inline struct timer_base *
|
2016-10-22 11:07:37 +00:00
|
|
|
get_target_base(struct timer_base *base, unsigned tflags)
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
{
|
2018-01-14 22:30:51 +00:00
|
|
|
#if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
|
|
|
|
if (static_branch_likely(&timers_migration_enabled) &&
|
|
|
|
!(tflags & TIMER_PINNED))
|
|
|
|
return get_timer_cpu_base(tflags, get_nohz_timer_target());
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
#endif
|
2018-01-14 22:30:51 +00:00
|
|
|
return get_timer_this_cpu_base(tflags);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:36 +00:00
|
|
|
static inline void forward_timer_base(struct timer_base *base)
|
|
|
|
{
|
2018-01-14 22:30:51 +00:00
|
|
|
#ifdef CONFIG_NO_HZ_COMMON
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
unsigned long jnow;
|
2016-10-22 11:07:37 +00:00
|
|
|
|
2016-07-04 09:50:36 +00:00
|
|
|
/*
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
* We only forward the base when we are idle or have just come out of
|
|
|
|
* idle (must_forward_clk logic), and have a delta between base clock
|
|
|
|
* and jiffies. In the common case, run_timers will take care of it.
|
2016-07-04 09:50:36 +00:00
|
|
|
*/
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
if (likely(!base->must_forward_clk))
|
|
|
|
return;
|
|
|
|
|
|
|
|
jnow = READ_ONCE(jiffies);
|
|
|
|
base->must_forward_clk = base->is_idle;
|
|
|
|
if ((long)(jnow - base->clk) < 2)
|
2016-07-04 09:50:36 +00:00
|
|
|
return;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the next expiry value is > jiffies, then we fast forward to
|
|
|
|
* jiffies otherwise we forward to the next expiry value.
|
|
|
|
*/
|
2020-07-03 01:06:57 +00:00
|
|
|
if (time_after(base->next_expiry, jnow)) {
|
2016-10-22 11:07:37 +00:00
|
|
|
base->clk = jnow;
|
2020-07-03 01:06:57 +00:00
|
|
|
} else {
|
|
|
|
if (WARN_ON_ONCE(time_before(base->next_expiry, base->clk)))
|
|
|
|
return;
|
2016-07-04 09:50:36 +00:00
|
|
|
base->clk = base->next_expiry;
|
2020-07-03 01:06:57 +00:00
|
|
|
}
|
2016-07-04 09:50:36 +00:00
|
|
|
#endif
|
2018-01-14 22:30:51 +00:00
|
|
|
}
|
2016-07-04 09:50:36 +00:00
|
|
|
|
|
|
|
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
/*
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means
|
|
|
|
* that all timers which are tied to this base are locked, and the base itself
|
|
|
|
* is locked too.
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
*
|
|
|
|
* So __run_timers/migrate_timers can safely modify all timers which could
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* be found in the base->vectors array.
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
*
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* When a timer is migrating then the TIMER_MIGRATING flag is set and we need
|
|
|
|
* to wait until the migration is done.
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
*/
|
2016-07-04 09:50:28 +00:00
|
|
|
static struct timer_base *lock_timer_base(struct timer_list *timer,
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned long *flags)
|
2006-09-29 08:59:36 +00:00
|
|
|
__acquires(timer->base->lock)
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
{
|
|
|
|
for (;;) {
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base *base;
|
2016-10-24 09:41:56 +00:00
|
|
|
u32 tf;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We need to use READ_ONCE() here, otherwise the compiler
|
|
|
|
* might re-read @tf between the check for TIMER_MIGRATING
|
|
|
|
* and spin_lock().
|
|
|
|
*/
|
|
|
|
tf = READ_ONCE(timer->flags);
|
2015-05-26 22:50:29 +00:00
|
|
|
|
|
|
|
if (!(tf & TIMER_MIGRATING)) {
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
base = get_timer_base(tf);
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock_irqsave(&base->lock, *flags);
|
2015-05-26 22:50:29 +00:00
|
|
|
if (timer->flags == tf)
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
return base;
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irqrestore(&base->lock, *flags);
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
}
|
|
|
|
cpu_relax();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
#define MOD_TIMER_PENDING_ONLY 0x01
|
|
|
|
#define MOD_TIMER_REDUCE 0x02
|
2019-11-07 19:37:38 +00:00
|
|
|
#define MOD_TIMER_NOTPENDING 0x04
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
|
2009-02-18 11:23:29 +00:00
|
|
|
static inline int
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
__mod_timer(struct timer_list *timer, unsigned long expires, unsigned int options)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base *base, *new_base;
|
2016-07-04 09:50:40 +00:00
|
|
|
unsigned int idx = UINT_MAX;
|
|
|
|
unsigned long clk = 0, flags;
|
2015-05-26 22:50:33 +00:00
|
|
|
int ret = 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2016-10-24 09:55:10 +00:00
|
|
|
BUG_ON(!timer->function);
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/*
|
2016-07-04 09:50:40 +00:00
|
|
|
* This is a common optimization triggered by the networking code - if
|
|
|
|
* the timer is re-modified to have the same timeout or ends up in the
|
|
|
|
* same array bucket then just return:
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
*/
|
2019-11-07 19:37:38 +00:00
|
|
|
if (!(options & MOD_TIMER_NOTPENDING) && timer_pending(timer)) {
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
/*
|
|
|
|
* The downside of this optimization is that it can result in
|
|
|
|
* larger granularity than you would get from adding a new
|
|
|
|
* timer with this expiry.
|
|
|
|
*/
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
long diff = timer->expires - expires;
|
|
|
|
|
|
|
|
if (!diff)
|
|
|
|
return 1;
|
|
|
|
if (options & MOD_TIMER_REDUCE && diff <= 0)
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
return 1;
|
2016-10-24 09:55:10 +00:00
|
|
|
|
2016-07-04 09:50:40 +00:00
|
|
|
/*
|
2016-10-24 09:55:10 +00:00
|
|
|
* We lock timer base and calculate the bucket index right
|
|
|
|
* here. If the timer ends up in the same bucket, then we
|
|
|
|
* just update the expiry time and avoid the whole
|
|
|
|
* dequeue/enqueue dance.
|
2016-07-04 09:50:40 +00:00
|
|
|
*/
|
2016-10-24 09:55:10 +00:00
|
|
|
base = lock_timer_base(timer, &flags);
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
forward_timer_base(base);
|
2016-07-04 09:50:40 +00:00
|
|
|
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
if (timer_pending(timer) && (options & MOD_TIMER_REDUCE) &&
|
|
|
|
time_before_eq(timer->expires, expires)) {
|
|
|
|
ret = 1;
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2016-10-24 09:55:10 +00:00
|
|
|
clk = base->clk;
|
2016-07-04 09:50:40 +00:00
|
|
|
idx = calc_wheel_index(expires, clk);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Retrieve and compare the array index of the pending
|
|
|
|
* timer. If it matches set the expiry to the new value so a
|
|
|
|
* subsequent call will exit in the expires check above.
|
|
|
|
*/
|
|
|
|
if (idx == timer_get_idx(timer)) {
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
if (!(options & MOD_TIMER_REDUCE))
|
|
|
|
timer->expires = expires;
|
|
|
|
else if (time_after(timer->expires, expires))
|
|
|
|
timer->expires = expires;
|
2016-10-24 09:55:10 +00:00
|
|
|
ret = 1;
|
|
|
|
goto out_unlock;
|
2016-07-04 09:50:40 +00:00
|
|
|
}
|
2016-10-24 09:55:10 +00:00
|
|
|
} else {
|
|
|
|
base = lock_timer_base(timer, &flags);
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
forward_timer_base(base);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
}
|
|
|
|
|
2012-05-25 22:08:57 +00:00
|
|
|
ret = detach_if_pending(timer, base, false);
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
if (!ret && (options & MOD_TIMER_PENDING_ONLY))
|
2012-05-25 22:08:57 +00:00
|
|
|
goto out_unlock;
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
new_base = get_target_base(base, timer->flags);
|
2009-04-16 06:46:41 +00:00
|
|
|
|
2006-03-31 10:30:30 +00:00
|
|
|
if (base != new_base) {
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* We are trying to schedule the timer on the new base.
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
* However we can't change timer's base while it is running,
|
|
|
|
* otherwise del_timer_sync() can't detect that the timer's
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
* handler yet has not finished. This also guarantees that the
|
|
|
|
* timer is serialized wrt itself.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2006-03-31 10:30:31 +00:00
|
|
|
if (likely(base->running_timer != timer)) {
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
/* See the comment in lock_timer_base() */
|
2015-05-26 22:50:29 +00:00
|
|
|
timer->flags |= TIMER_MIGRATING;
|
|
|
|
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock(&base->lock);
|
2006-03-31 10:30:31 +00:00
|
|
|
base = new_base;
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock(&base->lock);
|
2015-08-17 17:18:48 +00:00
|
|
|
WRITE_ONCE(timer->flags,
|
|
|
|
(timer->flags & ~TIMER_BASEMASK) | base->cpu);
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
forward_timer_base(base);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-03-21 12:09:19 +00:00
|
|
|
debug_timer_activate(timer);
|
2017-12-22 14:51:14 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
timer->expires = expires;
|
2016-07-04 09:50:40 +00:00
|
|
|
/*
|
|
|
|
* If 'idx' was calculated above and the base time did not advance
|
2016-10-24 09:55:10 +00:00
|
|
|
* between calculating 'idx' and possibly switching the base, only
|
|
|
|
* enqueue_timer() and trigger_dyntick_cpu() is required. Otherwise
|
|
|
|
* we need to (re)calculate the wheel index via
|
|
|
|
* internal_add_timer().
|
2016-07-04 09:50:40 +00:00
|
|
|
*/
|
|
|
|
if (idx != UINT_MAX && clk == base->clk) {
|
|
|
|
enqueue_timer(base, timer, idx);
|
|
|
|
trigger_dyntick_cpu(base, timer);
|
|
|
|
} else {
|
|
|
|
internal_add_timer(base, timer);
|
|
|
|
}
|
2009-02-18 11:23:29 +00:00
|
|
|
|
|
|
|
out_unlock:
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irqrestore(&base->lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2006-09-29 08:59:46 +00:00
|
|
|
/**
|
2009-02-18 11:23:29 +00:00
|
|
|
* mod_timer_pending - modify a pending timer's timeout
|
|
|
|
* @timer: the pending timer to be modified
|
|
|
|
* @expires: new timeout in jiffies
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2009-02-18 11:23:29 +00:00
|
|
|
* mod_timer_pending() is the same for pending timers as mod_timer(),
|
|
|
|
* but will not re-activate and modify already deleted timers.
|
|
|
|
*
|
|
|
|
* It is useful for unserialized use of timers.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2009-02-18 11:23:29 +00:00
|
|
|
int mod_timer_pending(struct timer_list *timer, unsigned long expires)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
return __mod_timer(timer, expires, MOD_TIMER_PENDING_ONLY);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2009-02-18 11:23:29 +00:00
|
|
|
EXPORT_SYMBOL(mod_timer_pending);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2006-09-29 08:59:46 +00:00
|
|
|
/**
|
2005-04-16 22:20:36 +00:00
|
|
|
* mod_timer - modify a timer's timeout
|
|
|
|
* @timer: the timer to be modified
|
2006-09-29 08:59:46 +00:00
|
|
|
* @expires: new timeout in jiffies
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* mod_timer() is a more efficient way to update the expire field of an
|
2005-04-16 22:20:36 +00:00
|
|
|
* active timer (if the timer is inactive it will be activated)
|
|
|
|
*
|
|
|
|
* mod_timer(timer, expires) is equivalent to:
|
|
|
|
*
|
|
|
|
* del_timer(timer); timer->expires = expires; add_timer(timer);
|
|
|
|
*
|
|
|
|
* Note that if there are multiple unserialized concurrent users of the
|
|
|
|
* same timer, then mod_timer() is the only safe way to modify the timeout,
|
|
|
|
* since add_timer() cannot modify an already running timer.
|
|
|
|
*
|
|
|
|
* The function returns whether it has modified a pending timer or not.
|
|
|
|
* (ie. mod_timer() of an inactive timer returns 0, mod_timer() of an
|
|
|
|
* active timer returns 1.)
|
|
|
|
*/
|
|
|
|
int mod_timer(struct timer_list *timer, unsigned long expires)
|
|
|
|
{
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
return __mod_timer(timer, expires, 0);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(mod_timer);
|
|
|
|
|
timers: Add a function to start/reduce a timer
Add a function, similar to mod_timer(), that will start a timer if it isn't
running and will modify it if it is running and has an expiry time longer
than the new time. If the timer is running with an expiry time that's the
same or sooner, no change is made.
The function looks like:
int timer_reduce(struct timer_list *timer, unsigned long expires);
This can be used by code such as networking code to make it easier to share
a timer for multiple timeouts. For instance, in upcoming AF_RXRPC code,
the rxrpc_call struct will maintain a number of timeouts:
unsigned long ack_at;
unsigned long resend_at;
unsigned long ping_at;
unsigned long expect_rx_by;
unsigned long expect_req_by;
unsigned long expect_term_by;
each of which is set independently of the others. With timer reduction
available, when the code needs to set one of the timeouts, it only needs to
look at that timeout and then call timer_reduce() to modify the timer,
starting it or bringing it forward if necessary. There is no need to refer
to the other timeouts to see which is earliest and no need to take any lock
other than, potentially, the timer lock inside timer_reduce().
Note, that this does not protect against concurrent invocations of any of
the timer functions.
As an example, the expect_rx_by timeout above, which terminates a call if
we don't get a packet from the server within a certain time window, would
be set something like this:
unsigned long now = jiffies;
unsigned long expect_rx_by = now + packet_receive_timeout;
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
timer_reduce(&call->timer, expect_rx_by);
The timer service code (which might, say, be in a work function) would then
check all the timeouts to see which, if any, had triggered, deal with
those:
t = READ_ONCE(call->ack_at);
if (time_after_eq(now, t)) {
cmpxchg(&call->ack_at, t, now + MAX_JIFFY_OFFSET);
set_bit(RXRPC_CALL_EV_ACK, &call->events);
}
and then restart the timer if necessary by finding the soonest timeout that
hasn't yet passed and then calling timer_reduce().
The disadvantage of doing things this way rather than comparing the timers
each time and calling mod_timer() is that you *will* take timer events
unless you can finish what you're doing and delete the timer in time.
The advantage of doing things this way is that you don't need to use a lock
to work out when the next timer should be set, other than the timer's own
lock - which you might not have to take.
[ tglx: Fixed weird formatting and adopted it to pending changes ]
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: keyrings@vger.kernel.org
Cc: linux-afs@lists.infradead.org
Link: https://lkml.kernel.org/r/151023090769.23050.1801643667223880753.stgit@warthog.procyon.org.uk
2017-11-09 12:35:07 +00:00
|
|
|
/**
|
|
|
|
* timer_reduce - Modify a timer's timeout if it would reduce the timeout
|
|
|
|
* @timer: The timer to be modified
|
|
|
|
* @expires: New timeout in jiffies
|
|
|
|
*
|
|
|
|
* timer_reduce() is very similar to mod_timer(), except that it will only
|
|
|
|
* modify a running timer if that would reduce the expiration time (it will
|
|
|
|
* start a timer that isn't running).
|
|
|
|
*/
|
|
|
|
int timer_reduce(struct timer_list *timer, unsigned long expires)
|
|
|
|
{
|
|
|
|
return __mod_timer(timer, expires, MOD_TIMER_REDUCE);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(timer_reduce);
|
|
|
|
|
2009-02-18 11:23:29 +00:00
|
|
|
/**
|
|
|
|
* add_timer - start a timer
|
|
|
|
* @timer: the timer to be added
|
|
|
|
*
|
2017-10-23 01:18:19 +00:00
|
|
|
* The kernel will do a ->function(@timer) callback from the
|
2009-02-18 11:23:29 +00:00
|
|
|
* timer interrupt at the ->expires point in the future. The
|
|
|
|
* current time is 'jiffies'.
|
|
|
|
*
|
2017-10-23 01:18:19 +00:00
|
|
|
* The timer's ->expires, ->function fields must be set prior calling this
|
|
|
|
* function.
|
2009-02-18 11:23:29 +00:00
|
|
|
*
|
|
|
|
* Timers with an ->expires field in the past will be executed in the next
|
|
|
|
* timer tick.
|
|
|
|
*/
|
|
|
|
void add_timer(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
BUG_ON(timer_pending(timer));
|
2019-11-07 19:37:38 +00:00
|
|
|
__mod_timer(timer, timer->expires, MOD_TIMER_NOTPENDING);
|
2009-02-18 11:23:29 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(add_timer);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* add_timer_on - start a timer on a particular CPU
|
|
|
|
* @timer: the timer to be added
|
|
|
|
* @cpu: the CPU to start it on
|
|
|
|
*
|
|
|
|
* This is not very scalable on SMP. Double adds are not possible.
|
|
|
|
*/
|
|
|
|
void add_timer_on(struct timer_list *timer, int cpu)
|
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
struct timer_base *new_base, *base;
|
2009-02-18 11:23:29 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
BUG_ON(timer_pending(timer) || !timer->function);
|
2015-11-04 17:15:33 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
new_base = get_timer_cpu_base(timer->flags, cpu);
|
|
|
|
|
2015-11-04 17:15:33 +00:00
|
|
|
/*
|
|
|
|
* If @timer was on a different CPU, it should be migrated with the
|
|
|
|
* old base locked to prevent other operations proceeding with the
|
|
|
|
* wrong base locked. See lock_timer_base().
|
|
|
|
*/
|
|
|
|
base = lock_timer_base(timer, &flags);
|
|
|
|
if (base != new_base) {
|
|
|
|
timer->flags |= TIMER_MIGRATING;
|
|
|
|
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock(&base->lock);
|
2015-11-04 17:15:33 +00:00
|
|
|
base = new_base;
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock(&base->lock);
|
2015-11-04 17:15:33 +00:00
|
|
|
WRITE_ONCE(timer->flags,
|
|
|
|
(timer->flags & ~TIMER_BASEMASK) | cpu);
|
|
|
|
}
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
forward_timer_base(base);
|
2015-11-04 17:15:33 +00:00
|
|
|
|
2019-03-21 12:09:19 +00:00
|
|
|
debug_timer_activate(timer);
|
2009-02-18 11:23:29 +00:00
|
|
|
internal_add_timer(base, timer);
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irqrestore(&base->lock, flags);
|
2009-02-18 11:23:29 +00:00
|
|
|
}
|
2009-05-19 20:49:07 +00:00
|
|
|
EXPORT_SYMBOL_GPL(add_timer_on);
|
2009-02-18 11:23:29 +00:00
|
|
|
|
2006-09-29 08:59:46 +00:00
|
|
|
/**
|
2017-03-07 11:48:02 +00:00
|
|
|
* del_timer - deactivate a timer.
|
2005-04-16 22:20:36 +00:00
|
|
|
* @timer: the timer to be deactivated
|
|
|
|
*
|
|
|
|
* del_timer() deactivates a timer - this works on both active and inactive
|
|
|
|
* timers.
|
|
|
|
*
|
|
|
|
* The function returns whether it has deactivated a pending timer or not.
|
|
|
|
* (ie. del_timer() of an inactive timer returns 0, del_timer() of an
|
|
|
|
* active timer returns 1.)
|
|
|
|
*/
|
|
|
|
int del_timer(struct timer_list *timer)
|
|
|
|
{
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base *base;
|
2005-04-16 22:20:36 +00:00
|
|
|
unsigned long flags;
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
int ret = 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2011-11-08 03:48:28 +00:00
|
|
|
debug_assert_init(timer);
|
|
|
|
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
if (timer_pending(timer)) {
|
|
|
|
base = lock_timer_base(timer, &flags);
|
2012-05-25 22:08:57 +00:00
|
|
|
ret = detach_if_pending(timer, base, true);
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irqrestore(&base->lock, flags);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
return ret;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(del_timer);
|
|
|
|
|
2006-09-29 08:59:46 +00:00
|
|
|
/**
|
|
|
|
* try_to_del_timer_sync - Try to deactivate a timer
|
2017-05-30 19:41:03 +00:00
|
|
|
* @timer: timer to delete
|
2006-09-29 08:59:46 +00:00
|
|
|
*
|
2005-06-23 07:08:59 +00:00
|
|
|
* This function tries to deactivate a timer. Upon successful (ret >= 0)
|
|
|
|
* exit the timer is not queued and the handler is not running on any CPU.
|
|
|
|
*/
|
|
|
|
int try_to_del_timer_sync(struct timer_list *timer)
|
|
|
|
{
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base *base;
|
2005-06-23 07:08:59 +00:00
|
|
|
unsigned long flags;
|
|
|
|
int ret = -1;
|
|
|
|
|
2011-11-08 03:48:28 +00:00
|
|
|
debug_assert_init(timer);
|
|
|
|
|
2005-06-23 07:08:59 +00:00
|
|
|
base = lock_timer_base(timer, &flags);
|
|
|
|
|
2017-02-08 19:26:59 +00:00
|
|
|
if (base->running_timer != timer)
|
2012-05-25 22:08:57 +00:00
|
|
|
ret = detach_if_pending(timer, base, true);
|
2017-02-08 19:26:59 +00:00
|
|
|
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irqrestore(&base->lock, flags);
|
2005-06-23 07:08:59 +00:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
2007-04-26 22:46:56 +00:00
|
|
|
EXPORT_SYMBOL(try_to_del_timer_sync);
|
|
|
|
|
2019-07-26 18:31:00 +00:00
|
|
|
#ifdef CONFIG_PREEMPT_RT
|
|
|
|
static __init void timer_base_init_expiry_lock(struct timer_base *base)
|
|
|
|
{
|
|
|
|
spin_lock_init(&base->expiry_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void timer_base_lock_expiry(struct timer_base *base)
|
|
|
|
{
|
|
|
|
spin_lock(&base->expiry_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void timer_base_unlock_expiry(struct timer_base *base)
|
|
|
|
{
|
|
|
|
spin_unlock(&base->expiry_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The counterpart to del_timer_wait_running().
|
|
|
|
*
|
|
|
|
* If there is a waiter for base->expiry_lock, then it was waiting for the
|
|
|
|
* timer callback to finish. Drop expiry_lock and reaquire it. That allows
|
|
|
|
* the waiter to acquire the lock and make progress.
|
|
|
|
*/
|
|
|
|
static void timer_sync_wait_running(struct timer_base *base)
|
|
|
|
{
|
|
|
|
if (atomic_read(&base->timer_waiters)) {
|
|
|
|
spin_unlock(&base->expiry_lock);
|
|
|
|
spin_lock(&base->expiry_lock);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This function is called on PREEMPT_RT kernels when the fast path
|
|
|
|
* deletion of a timer failed because the timer callback function was
|
|
|
|
* running.
|
|
|
|
*
|
|
|
|
* This prevents priority inversion, if the softirq thread on a remote CPU
|
|
|
|
* got preempted, and it prevents a life lock when the task which tries to
|
|
|
|
* delete a timer preempted the softirq thread running the timer callback
|
|
|
|
* function.
|
|
|
|
*/
|
|
|
|
static void del_timer_wait_running(struct timer_list *timer)
|
|
|
|
{
|
|
|
|
u32 tf;
|
|
|
|
|
|
|
|
tf = READ_ONCE(timer->flags);
|
|
|
|
if (!(tf & TIMER_MIGRATING)) {
|
|
|
|
struct timer_base *base = get_timer_base(tf);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Mark the base as contended and grab the expiry lock,
|
|
|
|
* which is held by the softirq across the timer
|
|
|
|
* callback. Drop the lock immediately so the softirq can
|
|
|
|
* expire the next timer. In theory the timer could already
|
|
|
|
* be running again, but that's more than unlikely and just
|
|
|
|
* causes another wait loop.
|
|
|
|
*/
|
|
|
|
atomic_inc(&base->timer_waiters);
|
|
|
|
spin_lock_bh(&base->expiry_lock);
|
|
|
|
atomic_dec(&base->timer_waiters);
|
|
|
|
spin_unlock_bh(&base->expiry_lock);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static inline void timer_base_init_expiry_lock(struct timer_base *base) { }
|
|
|
|
static inline void timer_base_lock_expiry(struct timer_base *base) { }
|
|
|
|
static inline void timer_base_unlock_expiry(struct timer_base *base) { }
|
|
|
|
static inline void timer_sync_wait_running(struct timer_base *base) { }
|
|
|
|
static inline void del_timer_wait_running(struct timer_list *timer) { }
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
|
2006-09-29 08:59:46 +00:00
|
|
|
/**
|
2005-04-16 22:20:36 +00:00
|
|
|
* del_timer_sync - deactivate a timer and wait for the handler to finish.
|
|
|
|
* @timer: the timer to be deactivated
|
|
|
|
*
|
|
|
|
* This function only differs from del_timer() on SMP: besides deactivating
|
|
|
|
* the timer it also makes sure the handler has finished executing on other
|
|
|
|
* CPUs.
|
|
|
|
*
|
2007-02-10 09:45:59 +00:00
|
|
|
* Synchronization rules: Callers must prevent restarting of the timer,
|
2005-04-16 22:20:36 +00:00
|
|
|
* otherwise this function is meaningless. It must not be called from
|
2012-08-08 18:10:28 +00:00
|
|
|
* interrupt contexts unless the timer is an irqsafe one. The caller must
|
|
|
|
* not hold locks which would prevent completion of the timer's
|
|
|
|
* handler. The timer's handler must not call add_timer_on(). Upon exit the
|
|
|
|
* timer is not queued and the handler is not running on any CPU.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2012-08-08 18:10:28 +00:00
|
|
|
* Note: For !irqsafe timers, you must not hold locks that are held in
|
|
|
|
* interrupt context while calling this function. Even if the lock has
|
2018-05-07 09:35:48 +00:00
|
|
|
* nothing to do with the timer in question. Here's why::
|
2011-02-08 17:39:54 +00:00
|
|
|
*
|
|
|
|
* CPU0 CPU1
|
|
|
|
* ---- ----
|
2018-05-07 09:35:48 +00:00
|
|
|
* <SOFTIRQ>
|
|
|
|
* call_timer_fn();
|
|
|
|
* base->running_timer = mytimer;
|
|
|
|
* spin_lock_irq(somelock);
|
2011-02-08 17:39:54 +00:00
|
|
|
* <IRQ>
|
|
|
|
* spin_lock(somelock);
|
2018-05-07 09:35:48 +00:00
|
|
|
* del_timer_sync(mytimer);
|
|
|
|
* while (base->running_timer == mytimer);
|
2011-02-08 17:39:54 +00:00
|
|
|
*
|
|
|
|
* Now del_timer_sync() will never return and never release somelock.
|
|
|
|
* The interrupt on the other CPU is waiting to grab somelock but
|
|
|
|
* it has interrupted the softirq that CPU0 is waiting to finish.
|
|
|
|
*
|
2005-04-16 22:20:36 +00:00
|
|
|
* The function returns whether it has deactivated a pending timer or not.
|
|
|
|
*/
|
|
|
|
int del_timer_sync(struct timer_list *timer)
|
|
|
|
{
|
2019-07-26 18:31:00 +00:00
|
|
|
int ret;
|
|
|
|
|
2009-01-29 15:03:20 +00:00
|
|
|
#ifdef CONFIG_LOCKDEP
|
2011-02-03 14:09:41 +00:00
|
|
|
unsigned long flags;
|
|
|
|
|
2011-02-08 17:39:54 +00:00
|
|
|
/*
|
|
|
|
* If lockdep gives a backtrace here, please reference
|
|
|
|
* the synchronization rules above.
|
|
|
|
*/
|
2011-02-08 14:18:00 +00:00
|
|
|
local_irq_save(flags);
|
2009-01-29 15:03:20 +00:00
|
|
|
lock_map_acquire(&timer->lockdep_map);
|
|
|
|
lock_map_release(&timer->lockdep_map);
|
2011-02-08 14:18:00 +00:00
|
|
|
local_irq_restore(flags);
|
2009-01-29 15:03:20 +00:00
|
|
|
#endif
|
2010-10-20 22:57:33 +00:00
|
|
|
/*
|
|
|
|
* don't use it in hardirq context, because it
|
|
|
|
* could lead to deadlock.
|
|
|
|
*/
|
2015-05-26 22:50:29 +00:00
|
|
|
WARN_ON(in_irq() && !(timer->flags & TIMER_IRQSAFE));
|
2019-07-26 18:31:00 +00:00
|
|
|
|
|
|
|
do {
|
|
|
|
ret = try_to_del_timer_sync(timer);
|
|
|
|
|
|
|
|
if (unlikely(ret < 0)) {
|
|
|
|
del_timer_wait_running(timer);
|
|
|
|
cpu_relax();
|
|
|
|
}
|
|
|
|
} while (ret < 0);
|
|
|
|
|
|
|
|
return ret;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
EXPORT_SYMBOL(del_timer_sync);
|
2005-04-16 22:20:36 +00:00
|
|
|
#endif
|
|
|
|
|
2019-03-21 12:09:21 +00:00
|
|
|
static void call_timer_fn(struct timer_list *timer,
|
|
|
|
void (*fn)(struct timer_list *),
|
|
|
|
unsigned long baseclk)
|
2010-03-12 20:10:29 +00:00
|
|
|
{
|
2013-08-14 12:55:24 +00:00
|
|
|
int count = preempt_count();
|
2010-03-12 20:10:29 +00:00
|
|
|
|
|
|
|
#ifdef CONFIG_LOCKDEP
|
|
|
|
/*
|
|
|
|
* It is permissible to free the timer 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
|
|
|
|
* timer->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, &timer->lockdep_map);
|
2010-03-12 20:10:29 +00:00
|
|
|
#endif
|
|
|
|
/*
|
|
|
|
* Couple the lock chain with the lock chain at
|
|
|
|
* del_timer_sync() by acquiring the lock_map around the fn()
|
|
|
|
* call here and in del_timer_sync().
|
|
|
|
*/
|
|
|
|
lock_map_acquire(&lockdep_map);
|
|
|
|
|
2019-03-21 12:09:21 +00:00
|
|
|
trace_timer_expire_entry(timer, baseclk);
|
2017-10-23 02:15:40 +00:00
|
|
|
fn(timer);
|
2010-03-12 20:10:29 +00:00
|
|
|
trace_timer_expire_exit(timer);
|
|
|
|
|
|
|
|
lock_map_release(&lockdep_map);
|
|
|
|
|
2013-08-14 12:55:24 +00:00
|
|
|
if (count != preempt_count()) {
|
2019-03-25 19:32:28 +00:00
|
|
|
WARN_ONCE(1, "timer: %pS preempt leak: %08x -> %08x\n",
|
2013-08-14 12:55:24 +00:00
|
|
|
fn, count, preempt_count());
|
2010-03-12 19:13:23 +00:00
|
|
|
/*
|
|
|
|
* Restore the preempt count. That gives us a decent
|
|
|
|
* chance to survive and extract information. If the
|
|
|
|
* callback kept a lock held, bad luck, but not worse
|
|
|
|
* than the BUG() we had.
|
|
|
|
*/
|
2013-08-14 12:55:24 +00:00
|
|
|
preempt_count_set(count);
|
2010-03-12 20:10:29 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static void expire_timers(struct timer_base *base, struct hlist_head *head)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2019-03-21 12:09:21 +00:00
|
|
|
/*
|
|
|
|
* This value is required only for tracing. base->clk was
|
|
|
|
* incremented directly before expire_timers was called. But expiry
|
|
|
|
* is related to the old base->clk value.
|
|
|
|
*/
|
|
|
|
unsigned long baseclk = base->clk - 1;
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
while (!hlist_empty(head)) {
|
|
|
|
struct timer_list *timer;
|
2017-10-23 02:15:40 +00:00
|
|
|
void (*fn)(struct timer_list *);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
timer = hlist_entry(head->first, struct timer_list, entry);
|
2015-05-26 22:50:24 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
base->running_timer = timer;
|
|
|
|
detach_timer(timer, true);
|
2015-05-26 22:50:24 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
fn = timer->function;
|
|
|
|
|
|
|
|
if (timer->flags & TIMER_IRQSAFE) {
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock(&base->lock);
|
2019-03-21 12:09:21 +00:00
|
|
|
call_timer_fn(timer, fn, baseclk);
|
2019-07-26 18:31:00 +00:00
|
|
|
base->running_timer = NULL;
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock(&base->lock);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
} else {
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irq(&base->lock);
|
2019-03-21 12:09:21 +00:00
|
|
|
call_timer_fn(timer, fn, baseclk);
|
2019-07-26 18:31:00 +00:00
|
|
|
base->running_timer = NULL;
|
|
|
|
timer_sync_wait_running(base);
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock_irq(&base->lock);
|
2015-05-26 22:50:24 +00:00
|
|
|
}
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
}
|
|
|
|
}
|
2015-05-26 22:50:24 +00:00
|
|
|
|
2016-07-04 09:50:34 +00:00
|
|
|
static int __collect_expired_timers(struct timer_base *base,
|
|
|
|
struct hlist_head *heads)
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
{
|
|
|
|
unsigned long clk = base->clk;
|
|
|
|
struct hlist_head *vec;
|
|
|
|
int i, levels = 0;
|
|
|
|
unsigned int idx;
|
2006-06-23 09:05:55 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
for (i = 0; i < LVL_DEPTH; i++) {
|
|
|
|
idx = (clk & LVL_MASK) + i * LVL_SIZE;
|
|
|
|
|
|
|
|
if (__test_and_clear_bit(idx, base->pending_map)) {
|
|
|
|
vec = base->vectors + idx;
|
|
|
|
hlist_move_list(vec, heads++);
|
|
|
|
levels++;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/* Is it time to look at the next level? */
|
|
|
|
if (clk & LVL_CLK_MASK)
|
|
|
|
break;
|
|
|
|
/* Shift clock for the next level granularity */
|
|
|
|
clk >>= LVL_CLK_SHIFT;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
return levels;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2011-08-10 21:21:01 +00:00
|
|
|
#ifdef CONFIG_NO_HZ_COMMON
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
2016-07-04 09:50:34 +00:00
|
|
|
* Find the next pending bucket of a level. Search from level start (@offset)
|
|
|
|
* + @clk upwards and if nothing there, search from start of the level
|
|
|
|
* (@offset) up to @offset + clk.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
static int next_pending_bucket(struct timer_base *base, unsigned offset,
|
|
|
|
unsigned clk)
|
|
|
|
{
|
|
|
|
unsigned pos, start = offset + clk;
|
|
|
|
unsigned end = offset + LVL_SIZE;
|
|
|
|
|
|
|
|
pos = find_next_bit(base->pending_map, end, start);
|
|
|
|
if (pos < end)
|
|
|
|
return pos - start;
|
|
|
|
|
|
|
|
pos = find_next_bit(base->pending_map, start, offset);
|
|
|
|
return pos < start ? pos + LVL_SIZE - start : -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2016-07-04 09:50:34 +00:00
|
|
|
* Search the first expiring timer in the various clock levels. Caller must
|
|
|
|
* hold base->lock.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2016-07-04 09:50:28 +00:00
|
|
|
static unsigned long __next_timer_interrupt(struct timer_base *base)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
unsigned long clk, next, adj;
|
|
|
|
unsigned lvl, offset = 0;
|
|
|
|
|
|
|
|
next = base->clk + NEXT_TIMER_MAX_DELTA;
|
|
|
|
clk = base->clk;
|
|
|
|
for (lvl = 0; lvl < LVL_DEPTH; lvl++, offset += LVL_SIZE) {
|
|
|
|
int pos = next_pending_bucket(base, offset, clk & LVL_MASK);
|
|
|
|
|
|
|
|
if (pos >= 0) {
|
|
|
|
unsigned long tmp = clk + (unsigned long) pos;
|
|
|
|
|
|
|
|
tmp <<= LVL_SHIFT(lvl);
|
|
|
|
if (time_before(tmp, next))
|
|
|
|
next = tmp;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
/*
|
|
|
|
* Clock for the next level. If the current level clock lower
|
|
|
|
* bits are zero, we look at the next level as is. If not we
|
|
|
|
* need to advance it by one because that's going to be the
|
|
|
|
* next expiring bucket in that level. base->clk is the next
|
|
|
|
* expiring jiffie. So in case of:
|
|
|
|
*
|
|
|
|
* LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
|
|
|
|
* 0 0 0 0 0 0
|
|
|
|
*
|
|
|
|
* we have to look at all levels @index 0. With
|
|
|
|
*
|
|
|
|
* LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
|
|
|
|
* 0 0 0 0 0 2
|
|
|
|
*
|
|
|
|
* LVL0 has the next expiring bucket @index 2. The upper
|
|
|
|
* levels have the next expiring bucket @index 1.
|
|
|
|
*
|
|
|
|
* In case that the propagation wraps the next level the same
|
|
|
|
* rules apply:
|
|
|
|
*
|
|
|
|
* LVL5 LVL4 LVL3 LVL2 LVL1 LVL0
|
|
|
|
* 0 0 0 0 F 2
|
|
|
|
*
|
|
|
|
* So after looking at LVL0 we get:
|
|
|
|
*
|
|
|
|
* LVL5 LVL4 LVL3 LVL2 LVL1
|
|
|
|
* 0 0 0 1 0
|
|
|
|
*
|
|
|
|
* So no propagation from LVL1 to LVL2 because that happened
|
|
|
|
* with the add already, but then we need to propagate further
|
|
|
|
* from LVL2 to LVL3.
|
|
|
|
*
|
|
|
|
* So the simple check whether the lower bits of the current
|
|
|
|
* level are 0 or not is sufficient for all cases.
|
|
|
|
*/
|
|
|
|
adj = clk & LVL_CLK_MASK ? 1 : 0;
|
|
|
|
clk >>= LVL_CLK_SHIFT;
|
|
|
|
clk += adj;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
return next;
|
2007-02-16 09:27:46 +00:00
|
|
|
}
|
2006-03-06 23:42:45 +00:00
|
|
|
|
2007-02-16 09:27:46 +00:00
|
|
|
/*
|
|
|
|
* Check, if the next hrtimer event is before the next timer wheel
|
|
|
|
* event:
|
|
|
|
*/
|
2015-04-14 21:08:58 +00:00
|
|
|
static u64 cmp_next_hrtimer_event(u64 basem, u64 expires)
|
2007-02-16 09:27:46 +00:00
|
|
|
{
|
2015-04-14 21:08:58 +00:00
|
|
|
u64 nextevt = hrtimer_get_next_event();
|
2006-05-20 22:00:24 +00:00
|
|
|
|
2007-03-25 12:31:17 +00:00
|
|
|
/*
|
2015-04-14 21:08:58 +00:00
|
|
|
* If high resolution timers are enabled
|
|
|
|
* hrtimer_get_next_event() returns KTIME_MAX.
|
2007-03-25 12:31:17 +00:00
|
|
|
*/
|
2015-04-14 21:08:58 +00:00
|
|
|
if (expires <= nextevt)
|
|
|
|
return expires;
|
2007-05-29 21:47:39 +00:00
|
|
|
|
|
|
|
/*
|
2015-04-14 21:08:58 +00:00
|
|
|
* If the next timer is already expired, return the tick base
|
|
|
|
* time so the tick is fired immediately.
|
2007-05-29 21:47:39 +00:00
|
|
|
*/
|
2015-04-14 21:08:58 +00:00
|
|
|
if (nextevt <= basem)
|
|
|
|
return basem;
|
2007-05-29 21:47:39 +00:00
|
|
|
|
2007-03-25 12:31:17 +00:00
|
|
|
/*
|
2015-04-14 21:08:58 +00:00
|
|
|
* Round up to the next jiffie. High resolution timers are
|
|
|
|
* off, so the hrtimers are expired in the tick and we need to
|
|
|
|
* make sure that this tick really expires the timer to avoid
|
|
|
|
* a ping pong of the nohz stop code.
|
|
|
|
*
|
|
|
|
* Use DIV_ROUND_UP_ULL to prevent gcc calling __divdi3
|
2007-03-25 12:31:17 +00:00
|
|
|
*/
|
2015-04-14 21:08:58 +00:00
|
|
|
return DIV_ROUND_UP_ULL(nextevt, TICK_NSEC) * TICK_NSEC;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
2007-02-16 09:27:46 +00:00
|
|
|
|
|
|
|
/**
|
2015-04-14 21:08:58 +00:00
|
|
|
* get_next_timer_interrupt - return the time (clock mono) of the next timer
|
|
|
|
* @basej: base time jiffies
|
|
|
|
* @basem: base time clock monotonic
|
|
|
|
*
|
|
|
|
* Returns the tick aligned clock monotonic time of the next pending
|
|
|
|
* timer or KTIME_MAX if no timer is pending.
|
2007-02-16 09:27:46 +00:00
|
|
|
*/
|
2015-04-14 21:08:58 +00:00
|
|
|
u64 get_next_timer_interrupt(unsigned long basej, u64 basem)
|
2007-02-16 09:27:46 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
|
2015-04-14 21:08:58 +00:00
|
|
|
u64 expires = KTIME_MAX;
|
|
|
|
unsigned long nextevt;
|
timers: Fix get_next_timer_interrupt() computation
The tick_nohz_stop_sched_tick() routine is not properly
canceling the sched timer when nothing is pending, because
get_next_timer_interrupt() is no longer returning KTIME_MAX in
that case. This causes periodic interrupts when none are needed.
When determining the next interrupt time, we first use
__next_timer_interrupt() to get the first expiring timer in the
timer wheel. If no timer is found, we return the base clock value
plus NEXT_TIMER_MAX_DELTA to indicate there is no timer in the
timer wheel.
Back in get_next_timer_interrupt(), we set the "expires" value
by converting the timer wheel expiry (in ticks) to a nsec value.
But we don't want to do this if the timer wheel expiry value
indicates no timer; we want to return KTIME_MAX.
Prior to commit 500462a9de65 ("timers: Switch to a non-cascading
wheel") we checked base->active_timers to see if any timers
were active, and if not, we didn't touch the expiry value and so
properly returned KTIME_MAX. Now we don't have active_timers.
To fix this, we now just check the timer wheel expiry value to
see if it is "now + NEXT_TIMER_MAX_DELTA", and if it is, we don't
try to compute a new value based on it, but instead simply let the
KTIME_MAX value in expires remain.
Fixes: 500462a9de65 "timers: Switch to a non-cascading wheel"
Signed-off-by: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: John Stultz <john.stultz@linaro.org>
Link: http://lkml.kernel.org/r/1470688147-22287-1-git-send-email-cmetcalf@mellanox.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-08-08 20:29:07 +00:00
|
|
|
bool is_max_delta;
|
2007-02-16 09:27:46 +00:00
|
|
|
|
2010-12-01 09:11:09 +00:00
|
|
|
/*
|
|
|
|
* Pretend that there is no timer pending if the cpu is offline.
|
|
|
|
* Possible pending timers will be migrated later to an active cpu.
|
|
|
|
*/
|
|
|
|
if (cpu_is_offline(smp_processor_id()))
|
2012-05-25 22:08:59 +00:00
|
|
|
return expires;
|
|
|
|
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock(&base->lock);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
nextevt = __next_timer_interrupt(base);
|
timers: Fix get_next_timer_interrupt() computation
The tick_nohz_stop_sched_tick() routine is not properly
canceling the sched timer when nothing is pending, because
get_next_timer_interrupt() is no longer returning KTIME_MAX in
that case. This causes periodic interrupts when none are needed.
When determining the next interrupt time, we first use
__next_timer_interrupt() to get the first expiring timer in the
timer wheel. If no timer is found, we return the base clock value
plus NEXT_TIMER_MAX_DELTA to indicate there is no timer in the
timer wheel.
Back in get_next_timer_interrupt(), we set the "expires" value
by converting the timer wheel expiry (in ticks) to a nsec value.
But we don't want to do this if the timer wheel expiry value
indicates no timer; we want to return KTIME_MAX.
Prior to commit 500462a9de65 ("timers: Switch to a non-cascading
wheel") we checked base->active_timers to see if any timers
were active, and if not, we didn't touch the expiry value and so
properly returned KTIME_MAX. Now we don't have active_timers.
To fix this, we now just check the timer wheel expiry value to
see if it is "now + NEXT_TIMER_MAX_DELTA", and if it is, we don't
try to compute a new value based on it, but instead simply let the
KTIME_MAX value in expires remain.
Fixes: 500462a9de65 "timers: Switch to a non-cascading wheel"
Signed-off-by: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: John Stultz <john.stultz@linaro.org>
Link: http://lkml.kernel.org/r/1470688147-22287-1-git-send-email-cmetcalf@mellanox.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-08-08 20:29:07 +00:00
|
|
|
is_max_delta = (nextevt == base->clk + NEXT_TIMER_MAX_DELTA);
|
2016-07-04 09:50:36 +00:00
|
|
|
base->next_expiry = nextevt;
|
|
|
|
/*
|
timers: Prevent base clock rewind when forwarding clock
Ashton and Michael reported, that kernel versions 4.8 and later suffer from
USB timeouts which are caused by the timer wheel rework.
This is caused by a bug in the base clock forwarding mechanism, which leads
to timers expiring early. The scenario which leads to this is:
run_timers()
while (jiffies >= base->clk) {
collect_expired_timers();
base->clk++;
expire_timers();
}
So base->clk = jiffies + 1. Now the cpu goes idle:
idle()
get_next_timer_interrupt()
nextevt = __next_time_interrupt();
if (time_after(nextevt, base->clk))
base->clk = jiffies;
jiffies has not advanced since run_timers(), so this assignment effectively
decrements base->clk by one.
base->clk is the index into the timer wheel arrays. So let's assume the
following state after the base->clk increment in run_timers():
jiffies = 0
base->clk = 1
A timer gets enqueued with an expiry delta of 63 ticks (which is the case
with the USB timeout and HZ=250) so the resulting bucket index is:
base->clk + delta = 1 + 63 = 64
The timer goes into the first wheel level. The array size is 64 so it ends
up in bucket 0, which is correct as it takes 63 ticks to advance base->clk
to index into bucket 0 again.
If the cpu goes idle before jiffies advance, then the bug in the forwarding
mechanism sets base->clk back to 0, so the next invocation of run_timers()
at the next tick will index into bucket 0 and therefore expire the timer 62
ticks too early.
Instead of blindly setting base->clk to jiffies we must make the forwarding
conditional on jiffies > base->clk, but we cannot use jiffies for this as
we might run into the following issue:
if (time_after(jiffies, base->clk) {
if (time_after(nextevt, base->clk))
base->clk = jiffies;
jiffies can increment between the check and the assigment far enough to
advance beyond nextevt. So we need to use a stable value for checking.
get_next_timer_interrupt() has the basej argument which is the jiffies
value snapshot taken in the calling code. So we can just that.
Thanks to Ashton for bisecting and providing trace data!
Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Reported-by: Ashton Holmes <scoopta@gmail.com>
Reported-by: Michael Thayer <michael.thayer@oracle.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Michal Necasek <michal.necasek@oracle.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: knut.osmundsen@oracle.com
Cc: stable@vger.kernel.org
Cc: stern@rowland.harvard.edu
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20161022110552.175308322@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-10-22 11:07:35 +00:00
|
|
|
* We have a fresh next event. Check whether we can forward the
|
|
|
|
* base. We can only do that when @basej is past base->clk
|
|
|
|
* otherwise we might rewind base->clk.
|
2016-07-04 09:50:36 +00:00
|
|
|
*/
|
timers: Prevent base clock rewind when forwarding clock
Ashton and Michael reported, that kernel versions 4.8 and later suffer from
USB timeouts which are caused by the timer wheel rework.
This is caused by a bug in the base clock forwarding mechanism, which leads
to timers expiring early. The scenario which leads to this is:
run_timers()
while (jiffies >= base->clk) {
collect_expired_timers();
base->clk++;
expire_timers();
}
So base->clk = jiffies + 1. Now the cpu goes idle:
idle()
get_next_timer_interrupt()
nextevt = __next_time_interrupt();
if (time_after(nextevt, base->clk))
base->clk = jiffies;
jiffies has not advanced since run_timers(), so this assignment effectively
decrements base->clk by one.
base->clk is the index into the timer wheel arrays. So let's assume the
following state after the base->clk increment in run_timers():
jiffies = 0
base->clk = 1
A timer gets enqueued with an expiry delta of 63 ticks (which is the case
with the USB timeout and HZ=250) so the resulting bucket index is:
base->clk + delta = 1 + 63 = 64
The timer goes into the first wheel level. The array size is 64 so it ends
up in bucket 0, which is correct as it takes 63 ticks to advance base->clk
to index into bucket 0 again.
If the cpu goes idle before jiffies advance, then the bug in the forwarding
mechanism sets base->clk back to 0, so the next invocation of run_timers()
at the next tick will index into bucket 0 and therefore expire the timer 62
ticks too early.
Instead of blindly setting base->clk to jiffies we must make the forwarding
conditional on jiffies > base->clk, but we cannot use jiffies for this as
we might run into the following issue:
if (time_after(jiffies, base->clk) {
if (time_after(nextevt, base->clk))
base->clk = jiffies;
jiffies can increment between the check and the assigment far enough to
advance beyond nextevt. So we need to use a stable value for checking.
get_next_timer_interrupt() has the basej argument which is the jiffies
value snapshot taken in the calling code. So we can just that.
Thanks to Ashton for bisecting and providing trace data!
Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Reported-by: Ashton Holmes <scoopta@gmail.com>
Reported-by: Michael Thayer <michael.thayer@oracle.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Michal Necasek <michal.necasek@oracle.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: knut.osmundsen@oracle.com
Cc: stable@vger.kernel.org
Cc: stern@rowland.harvard.edu
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20161022110552.175308322@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-10-22 11:07:35 +00:00
|
|
|
if (time_after(basej, base->clk)) {
|
|
|
|
if (time_after(nextevt, basej))
|
|
|
|
base->clk = basej;
|
|
|
|
else if (time_after(nextevt, base->clk))
|
|
|
|
base->clk = nextevt;
|
|
|
|
}
|
2016-07-04 09:50:34 +00:00
|
|
|
|
2016-07-04 09:50:36 +00:00
|
|
|
if (time_before_eq(nextevt, basej)) {
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
expires = basem;
|
2016-07-04 09:50:36 +00:00
|
|
|
base->is_idle = false;
|
|
|
|
} else {
|
timers: Fix get_next_timer_interrupt() computation
The tick_nohz_stop_sched_tick() routine is not properly
canceling the sched timer when nothing is pending, because
get_next_timer_interrupt() is no longer returning KTIME_MAX in
that case. This causes periodic interrupts when none are needed.
When determining the next interrupt time, we first use
__next_timer_interrupt() to get the first expiring timer in the
timer wheel. If no timer is found, we return the base clock value
plus NEXT_TIMER_MAX_DELTA to indicate there is no timer in the
timer wheel.
Back in get_next_timer_interrupt(), we set the "expires" value
by converting the timer wheel expiry (in ticks) to a nsec value.
But we don't want to do this if the timer wheel expiry value
indicates no timer; we want to return KTIME_MAX.
Prior to commit 500462a9de65 ("timers: Switch to a non-cascading
wheel") we checked base->active_timers to see if any timers
were active, and if not, we didn't touch the expiry value and so
properly returned KTIME_MAX. Now we don't have active_timers.
To fix this, we now just check the timer wheel expiry value to
see if it is "now + NEXT_TIMER_MAX_DELTA", and if it is, we don't
try to compute a new value based on it, but instead simply let the
KTIME_MAX value in expires remain.
Fixes: 500462a9de65 "timers: Switch to a non-cascading wheel"
Signed-off-by: Chris Metcalf <cmetcalf@mellanox.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: John Stultz <john.stultz@linaro.org>
Link: http://lkml.kernel.org/r/1470688147-22287-1-git-send-email-cmetcalf@mellanox.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-08-08 20:29:07 +00:00
|
|
|
if (!is_max_delta)
|
2017-08-01 07:11:52 +00:00
|
|
|
expires = basem + (u64)(nextevt - basej) * TICK_NSEC;
|
2016-07-04 09:50:36 +00:00
|
|
|
/*
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
* If we expect to sleep more than a tick, mark the base idle.
|
|
|
|
* Also the tick is stopped so any added timer must forward
|
|
|
|
* the base clk itself to keep granularity small. This idle
|
|
|
|
* logic is only maintained for the BASE_STD base, deferrable
|
|
|
|
* timers may still see large granularity skew (by design).
|
2016-07-04 09:50:36 +00:00
|
|
|
*/
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
if ((expires - basem) > TICK_NSEC) {
|
|
|
|
base->must_forward_clk = true;
|
2016-07-04 09:50:36 +00:00
|
|
|
base->is_idle = true;
|
timers: Fix excessive granularity of new timers after a nohz idle
When a timer base is idle, it is forwarded when a new timer is added
to ensure that granularity does not become excessive. When not idle,
the timer tick is expected to increment the base.
However there are several problems:
- If an existing timer is modified, the base is forwarded only after
the index is calculated.
- The base is not forwarded by add_timer_on.
- There is a window after a timer is restarted from a nohz idle, after
it is marked not-idle and before the timer tick on this CPU, where a
timer may be added but the ancient base does not get forwarded.
These result in excessive granularity (a 1 jiffy timeout can blow out
to 100s of jiffies), which cause the rcu lockup detector to trigger,
among other things.
Fix this by keeping track of whether the timer base has been idle
since it was last run or forwarded, and if so then forward it before
adding a new timer.
There is still a case where mod_timer optimises the case of a pending
timer mod with the same expiry time, where the timer can see excessive
granularity relative to the new, shorter interval. A comment is added,
but it's not changed because it is an important fastpath for
networking.
This has been tested and found to fix the RCU softlockup messages.
Testing was also done with tracing to measure requested versus
achieved wakeup latencies for all non-deferrable timers in an idle
system (with no lockup watchdogs running). Wakeup latency relative to
absolute latency is calculated (note this suffers from round-up skew
at low absolute times) and analysed:
max avg std
upstream 506.0 1.20 4.68
patched 2.0 1.08 0.15
The bug was noticed due to the lockup detector Kconfig changes
dropping it out of people's .configs and resulting in larger base
clk skew When the lockup detectors are enabled, no CPU can go idle for
longer than 4 seconds, which limits the granularity errors.
Sub-optimal timer behaviour is observable on a smaller scale in that
case:
max avg std
upstream 9.0 1.05 0.19
patched 2.0 1.04 0.11
Fixes: Fixes: a683f390b93f ("timers: Forward the wheel clock whenever possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: David Miller <davem@davemloft.net>
Cc: dzickus@redhat.com
Cc: sfr@canb.auug.org.au
Cc: mpe@ellerman.id.au
Cc: Stephen Boyd <sboyd@codeaurora.org>
Cc: linuxarm@huawei.com
Cc: abdhalee@linux.vnet.ibm.com
Cc: John Stultz <john.stultz@linaro.org>
Cc: akpm@linux-foundation.org
Cc: paulmck@linux.vnet.ibm.com
Cc: torvalds@linux-foundation.org
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/20170822084348.21436-1-npiggin@gmail.com
2017-08-22 08:43:48 +00:00
|
|
|
}
|
2012-05-25 22:08:59 +00:00
|
|
|
}
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock(&base->lock);
|
2007-02-16 09:27:46 +00:00
|
|
|
|
2015-04-14 21:08:58 +00:00
|
|
|
return cmp_next_hrtimer_event(basem, expires);
|
2007-02-16 09:27:46 +00:00
|
|
|
}
|
2016-07-04 09:50:34 +00:00
|
|
|
|
2016-07-04 09:50:36 +00:00
|
|
|
/**
|
|
|
|
* timer_clear_idle - Clear the idle state of the timer base
|
|
|
|
*
|
|
|
|
* Called with interrupts disabled
|
|
|
|
*/
|
|
|
|
void timer_clear_idle(void)
|
|
|
|
{
|
|
|
|
struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We do this unlocked. The worst outcome is a remote enqueue sending
|
|
|
|
* a pointless IPI, but taking the lock would just make the window for
|
|
|
|
* sending the IPI a few instructions smaller for the cost of taking
|
|
|
|
* the lock in the exit from idle path.
|
|
|
|
*/
|
|
|
|
base->is_idle = false;
|
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:34 +00:00
|
|
|
static int collect_expired_timers(struct timer_base *base,
|
|
|
|
struct hlist_head *heads)
|
|
|
|
{
|
2019-09-19 12:04:47 +00:00
|
|
|
unsigned long now = READ_ONCE(jiffies);
|
|
|
|
|
2016-07-04 09:50:34 +00:00
|
|
|
/*
|
|
|
|
* NOHZ optimization. After a long idle sleep we need to forward the
|
|
|
|
* base to current jiffies. Avoid a loop by searching the bitfield for
|
|
|
|
* the next expiring timer.
|
|
|
|
*/
|
2019-09-19 12:04:47 +00:00
|
|
|
if ((long)(now - base->clk) > 2) {
|
2016-07-04 09:50:34 +00:00
|
|
|
unsigned long next = __next_timer_interrupt(base);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the next timer is ahead of time forward to current
|
2016-07-04 09:50:36 +00:00
|
|
|
* jiffies, otherwise forward to the next expiry time:
|
2016-07-04 09:50:34 +00:00
|
|
|
*/
|
2019-09-19 12:04:47 +00:00
|
|
|
if (time_after(next, now)) {
|
2017-10-09 03:55:59 +00:00
|
|
|
/*
|
|
|
|
* The call site will increment base->clk and then
|
|
|
|
* terminate the expiry loop immediately.
|
|
|
|
*/
|
2019-09-19 12:04:47 +00:00
|
|
|
base->clk = now;
|
2016-07-04 09:50:34 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
base->clk = next;
|
|
|
|
}
|
|
|
|
return __collect_expired_timers(base, heads);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static inline int collect_expired_timers(struct timer_base *base,
|
|
|
|
struct hlist_head *heads)
|
|
|
|
{
|
|
|
|
return __collect_expired_timers(base, heads);
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
2007-10-18 10:06:11 +00:00
|
|
|
* Called from the timer interrupt handler to charge one tick to the current
|
2005-04-16 22:20:36 +00:00
|
|
|
* process. user_tick is 1 if the tick is user time, 0 for system.
|
|
|
|
*/
|
|
|
|
void update_process_times(int user_tick)
|
|
|
|
{
|
|
|
|
struct task_struct *p = current;
|
|
|
|
|
|
|
|
/* Note: this timer irq context must be accounted for as well. */
|
2007-11-09 21:39:38 +00:00
|
|
|
account_process_tick(p, user_tick);
|
2005-04-16 22:20:36 +00:00
|
|
|
run_local_timers();
|
2018-11-21 19:35:03 +00:00
|
|
|
rcu_sched_clock_irq(user_tick);
|
2010-10-14 06:01:34 +00:00
|
|
|
#ifdef CONFIG_IRQ_WORK
|
|
|
|
if (in_irq())
|
2014-08-16 16:37:19 +00:00
|
|
|
irq_work_tick();
|
2010-10-14 06:01:34 +00:00
|
|
|
#endif
|
2005-04-16 22:20:36 +00:00
|
|
|
scheduler_tick();
|
posix-timers: Make them configurable
Some embedded systems have no use for them. This removes about
25KB from the kernel binary size when configured out.
Corresponding syscalls are routed to a stub logging the attempt to
use those syscalls which should be enough of a clue if they were
disabled without proper consideration. They are: timer_create,
timer_gettime: timer_getoverrun, timer_settime, timer_delete,
clock_adjtime, setitimer, getitimer, alarm.
The clock_settime, clock_gettime, clock_getres and clock_nanosleep
syscalls are replaced by simple wrappers compatible with CLOCK_REALTIME,
CLOCK_MONOTONIC and CLOCK_BOOTTIME only which should cover the vast
majority of use cases with very little code.
Signed-off-by: Nicolas Pitre <nico@linaro.org>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: John Stultz <john.stultz@linaro.org>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
Cc: Paul Bolle <pebolle@tiscali.nl>
Cc: linux-kbuild@vger.kernel.org
Cc: netdev@vger.kernel.org
Cc: Michal Marek <mmarek@suse.com>
Cc: Edward Cree <ecree@solarflare.com>
Link: http://lkml.kernel.org/r/1478841010-28605-7-git-send-email-nicolas.pitre@linaro.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2016-11-11 05:10:10 +00:00
|
|
|
if (IS_ENABLED(CONFIG_POSIX_TIMERS))
|
2019-08-19 14:31:47 +00:00
|
|
|
run_posix_cpu_timers();
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2016-07-04 09:50:33 +00:00
|
|
|
/**
|
|
|
|
* __run_timers - run all expired timers (if any) on this CPU.
|
|
|
|
* @base: the timer vector to be processed.
|
|
|
|
*/
|
|
|
|
static inline void __run_timers(struct timer_base *base)
|
|
|
|
{
|
|
|
|
struct hlist_head heads[LVL_DEPTH];
|
|
|
|
int levels;
|
|
|
|
|
|
|
|
if (!time_after_eq(jiffies, base->clk))
|
|
|
|
return;
|
|
|
|
|
2019-07-26 18:31:00 +00:00
|
|
|
timer_base_lock_expiry(base);
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock_irq(&base->lock);
|
2016-07-04 09:50:33 +00:00
|
|
|
|
2018-08-02 08:51:03 +00:00
|
|
|
/*
|
|
|
|
* timer_base::must_forward_clk must be cleared before running
|
|
|
|
* timers so that any timer functions that call mod_timer() will
|
|
|
|
* not try to forward the base. Idle tracking / clock forwarding
|
|
|
|
* logic is only used with BASE_STD timers.
|
|
|
|
*
|
|
|
|
* The must_forward_clk flag is cleared unconditionally also for
|
|
|
|
* the deferrable base. The deferrable base is not affected by idle
|
|
|
|
* tracking and never forwarded, so clearing the flag is a NOOP.
|
|
|
|
*
|
|
|
|
* The fact that the deferrable base is never forwarded can cause
|
|
|
|
* large variations in granularity for deferrable timers, but they
|
|
|
|
* can be deferred for long periods due to idle anyway.
|
|
|
|
*/
|
|
|
|
base->must_forward_clk = false;
|
|
|
|
|
2016-07-04 09:50:33 +00:00
|
|
|
while (time_after_eq(jiffies, base->clk)) {
|
|
|
|
|
|
|
|
levels = collect_expired_timers(base, heads);
|
|
|
|
base->clk++;
|
|
|
|
|
|
|
|
while (levels--)
|
|
|
|
expire_timers(base, heads + levels);
|
|
|
|
}
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock_irq(&base->lock);
|
2019-07-26 18:31:00 +00:00
|
|
|
timer_base_unlock_expiry(base);
|
2016-07-04 09:50:33 +00:00
|
|
|
}
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
/*
|
|
|
|
* This function runs timers and the timer-tq in bottom half context.
|
|
|
|
*/
|
2016-06-20 18:42:34 +00:00
|
|
|
static __latent_entropy void run_timer_softirq(struct softirq_action *h)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
__run_timers(base);
|
2017-12-22 14:51:12 +00:00
|
|
|
if (IS_ENABLED(CONFIG_NO_HZ_COMMON))
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
__run_timers(this_cpu_ptr(&timer_bases[BASE_DEF]));
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Called by the local, per-CPU timer interrupt on SMP.
|
|
|
|
*/
|
|
|
|
void run_local_timers(void)
|
|
|
|
{
|
2016-07-04 09:50:37 +00:00
|
|
|
struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
|
|
|
|
|
2008-01-25 20:08:31 +00:00
|
|
|
hrtimer_run_queues();
|
2016-07-04 09:50:37 +00:00
|
|
|
/* Raise the softirq only if required. */
|
|
|
|
if (time_before(jiffies, base->clk)) {
|
2018-01-14 22:19:49 +00:00
|
|
|
if (!IS_ENABLED(CONFIG_NO_HZ_COMMON))
|
2016-07-04 09:50:37 +00:00
|
|
|
return;
|
|
|
|
/* CPU is awake, so check the deferrable base. */
|
|
|
|
base++;
|
|
|
|
if (time_before(jiffies, base->clk))
|
|
|
|
return;
|
|
|
|
}
|
2005-04-16 22:20:36 +00:00
|
|
|
raise_softirq(TIMER_SOFTIRQ);
|
|
|
|
}
|
|
|
|
|
2017-10-04 23:26:55 +00:00
|
|
|
/*
|
|
|
|
* Since schedule_timeout()'s timer is defined on the stack, it must store
|
|
|
|
* the target task on the stack as well.
|
|
|
|
*/
|
|
|
|
struct process_timer {
|
|
|
|
struct timer_list timer;
|
|
|
|
struct task_struct *task;
|
|
|
|
};
|
|
|
|
|
|
|
|
static void process_timeout(struct timer_list *t)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2017-10-04 23:26:55 +00:00
|
|
|
struct process_timer *timeout = from_timer(timeout, t, timer);
|
|
|
|
|
|
|
|
wake_up_process(timeout->task);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* schedule_timeout - sleep until timeout
|
|
|
|
* @timeout: timeout value in jiffies
|
|
|
|
*
|
2020-01-17 22:59:00 +00:00
|
|
|
* Make the current task sleep until @timeout jiffies have elapsed.
|
|
|
|
* The function behavior depends on the current task state
|
|
|
|
* (see also set_current_state() description):
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2020-01-17 22:59:00 +00:00
|
|
|
* %TASK_RUNNING - the scheduler is called, but the task does not sleep
|
|
|
|
* at all. That happens because sched_submit_work() does nothing for
|
|
|
|
* tasks in %TASK_RUNNING state.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* %TASK_UNINTERRUPTIBLE - at least @timeout jiffies are guaranteed to
|
2016-10-21 15:58:51 +00:00
|
|
|
* pass before the routine returns unless the current task is explicitly
|
2020-01-17 22:59:00 +00:00
|
|
|
* woken up, (e.g. by wake_up_process()).
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
|
|
|
* %TASK_INTERRUPTIBLE - the routine may return early if a signal is
|
2016-10-21 15:58:51 +00:00
|
|
|
* delivered to the current task or the current task is explicitly woken
|
|
|
|
* up.
|
2005-04-16 22:20:36 +00:00
|
|
|
*
|
2020-01-17 22:59:00 +00:00
|
|
|
* The current task state is guaranteed to be %TASK_RUNNING when this
|
2005-04-16 22:20:36 +00:00
|
|
|
* routine returns.
|
|
|
|
*
|
|
|
|
* Specifying a @timeout value of %MAX_SCHEDULE_TIMEOUT will schedule
|
|
|
|
* the CPU away without a bound on the timeout. In this case the return
|
|
|
|
* value will be %MAX_SCHEDULE_TIMEOUT.
|
|
|
|
*
|
2016-10-21 15:58:51 +00:00
|
|
|
* Returns 0 when the timer has expired otherwise the remaining time in
|
2020-01-17 22:59:00 +00:00
|
|
|
* jiffies will be returned. In all cases the return value is guaranteed
|
2016-10-21 15:58:51 +00:00
|
|
|
* to be non-negative.
|
2005-04-16 22:20:36 +00:00
|
|
|
*/
|
2008-02-08 12:19:53 +00:00
|
|
|
signed long __sched schedule_timeout(signed long timeout)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2017-10-04 23:26:55 +00:00
|
|
|
struct process_timer timer;
|
2005-04-16 22:20:36 +00:00
|
|
|
unsigned long expire;
|
|
|
|
|
|
|
|
switch (timeout)
|
|
|
|
{
|
|
|
|
case MAX_SCHEDULE_TIMEOUT:
|
|
|
|
/*
|
|
|
|
* These two special cases are useful to be comfortable
|
|
|
|
* in the caller. Nothing more. We could take
|
|
|
|
* MAX_SCHEDULE_TIMEOUT from one of the negative value
|
|
|
|
* but I' d like to return a valid offset (>=0) to allow
|
|
|
|
* the caller to do everything it want with the retval.
|
|
|
|
*/
|
|
|
|
schedule();
|
|
|
|
goto out;
|
|
|
|
default:
|
|
|
|
/*
|
|
|
|
* Another bit of PARANOID. Note that the retval will be
|
|
|
|
* 0 since no piece of kernel is supposed to do a check
|
|
|
|
* for a negative retval of schedule_timeout() (since it
|
|
|
|
* should never happens anyway). You just have the printk()
|
|
|
|
* that will tell you if something is gone wrong and where.
|
|
|
|
*/
|
2006-12-22 09:10:14 +00:00
|
|
|
if (timeout < 0) {
|
2005-04-16 22:20:36 +00:00
|
|
|
printk(KERN_ERR "schedule_timeout: wrong timeout "
|
2006-12-22 09:10:14 +00:00
|
|
|
"value %lx\n", timeout);
|
|
|
|
dump_stack();
|
2005-04-16 22:20:36 +00:00
|
|
|
current->state = TASK_RUNNING;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
expire = timeout + jiffies;
|
|
|
|
|
2017-10-04 23:26:55 +00:00
|
|
|
timer.task = current;
|
|
|
|
timer_setup_on_stack(&timer.timer, process_timeout, 0);
|
2019-11-07 19:37:38 +00:00
|
|
|
__mod_timer(&timer.timer, expire, MOD_TIMER_NOTPENDING);
|
2005-04-16 22:20:36 +00:00
|
|
|
schedule();
|
2017-10-04 23:26:55 +00:00
|
|
|
del_singleshot_timer_sync(&timer.timer);
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2008-04-30 07:55:03 +00:00
|
|
|
/* Remove the timer from the object tracker */
|
2017-10-04 23:26:55 +00:00
|
|
|
destroy_timer_on_stack(&timer.timer);
|
2008-04-30 07:55:03 +00:00
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
timeout = expire - jiffies;
|
|
|
|
|
|
|
|
out:
|
|
|
|
return timeout < 0 ? 0 : timeout;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(schedule_timeout);
|
|
|
|
|
2005-09-13 08:25:15 +00:00
|
|
|
/*
|
|
|
|
* We can use __set_current_state() here because schedule_timeout() calls
|
|
|
|
* schedule() unconditionally.
|
|
|
|
*/
|
2005-09-10 07:27:21 +00:00
|
|
|
signed long __sched schedule_timeout_interruptible(signed long timeout)
|
|
|
|
{
|
2005-10-30 23:01:42 +00:00
|
|
|
__set_current_state(TASK_INTERRUPTIBLE);
|
|
|
|
return schedule_timeout(timeout);
|
2005-09-10 07:27:21 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(schedule_timeout_interruptible);
|
|
|
|
|
2007-12-06 16:59:46 +00:00
|
|
|
signed long __sched schedule_timeout_killable(signed long timeout)
|
|
|
|
{
|
|
|
|
__set_current_state(TASK_KILLABLE);
|
|
|
|
return schedule_timeout(timeout);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(schedule_timeout_killable);
|
|
|
|
|
2005-09-10 07:27:21 +00:00
|
|
|
signed long __sched schedule_timeout_uninterruptible(signed long timeout)
|
|
|
|
{
|
2005-10-30 23:01:42 +00:00
|
|
|
__set_current_state(TASK_UNINTERRUPTIBLE);
|
|
|
|
return schedule_timeout(timeout);
|
2005-09-10 07:27:21 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(schedule_timeout_uninterruptible);
|
|
|
|
|
2016-03-25 21:20:21 +00:00
|
|
|
/*
|
|
|
|
* Like schedule_timeout_uninterruptible(), except this task will not contribute
|
|
|
|
* to load average.
|
|
|
|
*/
|
|
|
|
signed long __sched schedule_timeout_idle(signed long timeout)
|
|
|
|
{
|
|
|
|
__set_current_state(TASK_IDLE);
|
|
|
|
return schedule_timeout(timeout);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(schedule_timeout_idle);
|
|
|
|
|
2005-04-16 22:20:36 +00:00
|
|
|
#ifdef CONFIG_HOTPLUG_CPU
|
2016-07-04 09:50:28 +00:00
|
|
|
static void migrate_timer_list(struct timer_base *new_base, struct hlist_head *head)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
|
|
|
struct timer_list *timer;
|
2015-05-26 22:50:29 +00:00
|
|
|
int cpu = new_base->cpu;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2015-05-26 22:50:28 +00:00
|
|
|
while (!hlist_empty(head)) {
|
|
|
|
timer = hlist_entry(head->first, struct timer_list, entry);
|
2012-05-25 22:08:57 +00:00
|
|
|
detach_timer(timer, false);
|
2015-05-26 22:50:29 +00:00
|
|
|
timer->flags = (timer->flags & ~TIMER_BASEMASK) | cpu;
|
2005-04-16 22:20:36 +00:00
|
|
|
internal_add_timer(new_base, timer);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-12-27 20:37:25 +00:00
|
|
|
int timers_prepare_cpu(unsigned int cpu)
|
|
|
|
{
|
|
|
|
struct timer_base *base;
|
|
|
|
int b;
|
|
|
|
|
|
|
|
for (b = 0; b < NR_BASES; b++) {
|
|
|
|
base = per_cpu_ptr(&timer_bases[b], cpu);
|
|
|
|
base->clk = jiffies;
|
|
|
|
base->next_expiry = base->clk + NEXT_TIMER_MAX_DELTA;
|
|
|
|
base->is_idle = false;
|
|
|
|
base->must_forward_clk = true;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-07-13 17:16:59 +00:00
|
|
|
int timers_dead_cpu(unsigned int cpu)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2016-07-04 09:50:28 +00:00
|
|
|
struct timer_base *old_base;
|
|
|
|
struct timer_base *new_base;
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
int b, i;
|
2005-04-16 22:20:36 +00:00
|
|
|
|
|
|
|
BUG_ON(cpu_online(cpu));
|
[PATCH] timers fixes/improvements
This patch tries to solve following problems:
1. del_timer_sync() is racy. The timer can be fired again after
del_timer_sync have checked all cpus and before it will recheck
timer_pending().
2. It has scalability problems. All cpus are scanned to determine
if the timer is running on that cpu.
With this patch del_timer_sync is O(1) and no slower than plain
del_timer(pending_timer), unless it has to actually wait for
completion of the currently running timer.
The only restriction is that the recurring timer should not use
add_timer_on().
3. The timers are not serialized wrt to itself.
If CPU_0 does mod_timer(jiffies+1) while the timer is currently
running on CPU 1, it is quite possible that local interrupt on
CPU_0 will start that timer before it finished on CPU_1.
4. The timers locking is suboptimal. __mod_timer() takes 3 locks
at once and still requires wmb() in del_timer/run_timers.
The new implementation takes 2 locks sequentially and does not
need memory barriers.
Currently ->base != NULL means that the timer is pending. In that case
->base.lock is used to lock the timer. __mod_timer also takes timer->lock
because ->base can be == NULL.
This patch uses timer->entry.next != NULL as indication that the timer is
pending. So it does __list_del(), entry->next = NULL instead of list_del()
when the timer is deleted.
The ->base field is used for hashed locking only, it is initialized
in init_timer() which sets ->base = per_cpu(tvec_bases). When the
tvec_bases.lock is locked, it means that all timers which are tied
to this base via timer->base are locked, and the base itself is locked
too.
So __run_timers/migrate_timers can safely modify all timers which could
be found on ->tvX lists (pending timers).
When the timer's base is locked, and the timer removed from ->entry list
(which means that _run_timers/migrate_timers can't see this timer), it is
possible to set timer->base = NULL and drop the lock: the timer remains
locked.
This patch adds lock_timer_base() helper, which waits for ->base != NULL,
locks the ->base, and checks it is still the same.
__mod_timer() schedules the timer on the local CPU and changes it's base.
However, it does not lock both old and new bases at once. It locks the
timer via lock_timer_base(), deletes the timer, sets ->base = NULL, and
unlocks old base. Then __mod_timer() locks new_base, sets ->base = new_base,
and adds this timer. This simplifies the code, because AB-BA deadlock is not
possible. __mod_timer() also ensures that the timer's base is not changed
while the timer's handler is running on the old base.
__run_timers(), del_timer() do not change ->base anymore, they only clear
pending flag.
So del_timer_sync() can test timer->base->running_timer == timer to detect
whether it is running or not.
We don't need timer_list->lock anymore, this patch kills it.
We also don't need barriers. del_timer() and __run_timers() used smp_wmb()
before clearing timer's pending flag. It was needed because __mod_timer()
did not lock old_base if the timer is not pending, so __mod_timer()->list_add()
could race with del_timer()->list_del(). With this patch these functions are
serialized through base->lock.
One problem. TIMER_INITIALIZER can't use per_cpu(tvec_bases). So this patch
adds global
struct timer_base_s {
spinlock_t lock;
struct timer_list *running_timer;
} __init_timer_base;
which is used by TIMER_INITIALIZER. The corresponding fields in tvec_t_base_s
struct are replaced by struct timer_base_s t_base.
It is indeed ugly. But this can't have scalability problems. The global
__init_timer_base.lock is used only when __mod_timer() is called for the first
time AND the timer was compile time initialized. After that the timer migrates
to the local CPU.
Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Renaud Lienhart <renaud.lienhart@free.fr>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:08:56 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
for (b = 0; b < NR_BASES; b++) {
|
|
|
|
old_base = per_cpu_ptr(&timer_bases[b], cpu);
|
|
|
|
new_base = get_cpu_ptr(&timer_bases[b]);
|
|
|
|
/*
|
|
|
|
* The caller is globally serialized and nobody else
|
|
|
|
* takes two locks at once, deadlock is not possible.
|
|
|
|
*/
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock_irq(&new_base->lock);
|
|
|
|
raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
|
2018-01-18 11:50:22 +00:00
|
|
|
/*
|
|
|
|
* The current CPUs base clock might be stale. Update it
|
|
|
|
* before moving the timers over.
|
|
|
|
*/
|
|
|
|
forward_timer_base(new_base);
|
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
BUG_ON(old_base->running_timer);
|
|
|
|
|
|
|
|
for (i = 0; i < WHEEL_SIZE; i++)
|
|
|
|
migrate_timer_list(new_base, old_base->vectors + i);
|
2015-03-31 15:19:01 +00:00
|
|
|
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_unlock(&old_base->lock);
|
|
|
|
raw_spin_unlock_irq(&new_base->lock);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
put_cpu_ptr(&timer_bases);
|
|
|
|
}
|
2016-07-13 17:16:59 +00:00
|
|
|
return 0;
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
2015-03-31 15:19:02 +00:00
|
|
|
#endif /* CONFIG_HOTPLUG_CPU */
|
2005-04-16 22:20:36 +00:00
|
|
|
|
2015-05-26 22:50:29 +00:00
|
|
|
static void __init init_timer_cpu(int cpu)
|
2015-03-31 15:19:01 +00:00
|
|
|
{
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
struct timer_base *base;
|
|
|
|
int i;
|
2015-03-31 15:19:01 +00:00
|
|
|
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
for (i = 0; i < NR_BASES; i++) {
|
|
|
|
base = per_cpu_ptr(&timer_bases[i], cpu);
|
|
|
|
base->cpu = cpu;
|
2017-06-27 16:15:38 +00:00
|
|
|
raw_spin_lock_init(&base->lock);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
base->clk = jiffies;
|
2019-07-26 18:31:00 +00:00
|
|
|
timer_base_init_expiry_lock(base);
|
timers: Switch to a non-cascading wheel
The current timer wheel has some drawbacks:
1) Cascading:
Cascading can be an unbound operation and is completely pointless in most
cases because the vast majority of the timer wheel timers are canceled or
rearmed before expiration. (They are used as timeout safeguards, not as
real timers to measure time.)
2) No fast lookup of the next expiring timer:
In NOHZ scenarios the first timer soft interrupt after a long NOHZ period
must fast forward the base time to the current value of jiffies. As we
have no way to find the next expiring timer fast, the code loops linearly
and increments the base time one by one and checks for expired timers
in each step. This causes unbound overhead spikes exactly in the moment
when we should wake up as fast as possible.
After a thorough analysis of real world data gathered on laptops,
workstations, webservers and other machines (thanks Chris!) I came to the
conclusion that the current 'classic' timer wheel implementation can be
modified to address the above issues.
The vast majority of timer wheel timers is canceled or rearmed before
expiry. Most of them are timeouts for networking and other I/O tasks. The
nature of timeouts is to catch the exception from normal operation (TCP ack
timed out, disk does not respond, etc.). For these kinds of timeouts the
accuracy of the timeout is not really a concern. Timeouts are very often
approximate worst-case values and in case the timeout fires, we already
waited for a long time and performance is down the drain already.
The few timers which actually expire can be split into two categories:
1) Short expiry times which expect halfways accurate expiry
2) Long term expiry times are inaccurate today already due to the
batching which is done for NOHZ automatically and also via the
set_timer_slack() API.
So for long term expiry timers we can avoid the cascading property and just
leave them in the less granular outer wheels until expiry or
cancelation. Timers which are armed with a timeout larger than the wheel
capacity are no longer cascaded. We expire them with the longest possible
timeout (6+ days). We have not observed such timeouts in our data collection,
but at least we handle them, applying the rule of the least surprise.
To avoid extending the wheel levels for HZ=1000 so we can accomodate the
longest observed timeouts (5 days in the network conntrack code) we reduce the
first level granularity on HZ=1000 to 4ms, which effectively is the same as
the HZ=250 behaviour. From our data analysis there is nothing which relies on
that 1ms granularity and as a side effect we get better batching and timer
locality for the networking code as well.
Contrary to the classic wheel the granularity of the next wheel is not the
capacity of the first wheel. The granularities of the wheels are in the
currently chosen setting 8 times the granularity of the previous wheel.
So for HZ=250 we end up with the following granularity levels:
Level Offset Granularity Range
0 0 4 ms 0 ms - 252 ms
1 64 32 ms 256 ms - 2044 ms (256ms - ~2s)
2 128 256 ms 2048 ms - 16380 ms (~2s - ~16s)
3 192 2048 ms (~2s) 16384 ms - 131068 ms (~16s - ~2m)
4 256 16384 ms (~16s) 131072 ms - 1048572 ms (~2m - ~17m)
5 320 131072 ms (~2m) 1048576 ms - 8388604 ms (~17m - ~2h)
6 384 1048576 ms (~17m) 8388608 ms - 67108863 ms (~2h - ~18h)
7 448 8388608 ms (~2h) 67108864 ms - 536870911 ms (~18h - ~6d)
That's a worst case inaccuracy of 12.5% for the timers which are queued at the
beginning of a level.
So the new wheel concept addresses the old issues:
1) Cascading is avoided completely
2) By keeping the timers in the bucket until expiry/cancelation we can track
the buckets which have timers enqueued in a bucket bitmap and therefore can
look up the next expiring timer very fast and O(1).
A further benefit of the concept is that the slack calculation which is done
on every timer start is no longer necessary because the granularity levels
provide natural batching already.
Our extensive testing with various loads did not show any performance
degradation vs. the current wheel implementation.
This patch does not address the 'fast lookup' issue as we wanted to make sure
that there is no regression introduced by the wheel redesign. The
optimizations are in follow up patches.
This patch contains fixes from Anna-Maria Gleixner and Richard Cochran.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
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@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.108621834@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-07-04 09:50:30 +00:00
|
|
|
}
|
2015-03-31 15:19:01 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void __init init_timer_cpus(void)
|
2005-04-16 22:20:36 +00:00
|
|
|
{
|
2015-03-31 15:19:01 +00:00
|
|
|
int cpu;
|
|
|
|
|
2015-05-26 22:50:29 +00:00
|
|
|
for_each_possible_cpu(cpu)
|
|
|
|
init_timer_cpu(cpu);
|
2015-03-31 15:19:01 +00:00
|
|
|
}
|
2012-08-08 18:10:25 +00:00
|
|
|
|
2015-03-31 15:19:01 +00:00
|
|
|
void __init init_timers(void)
|
|
|
|
{
|
|
|
|
init_timer_cpus();
|
Remove argument from open_softirq which is always NULL
As git-grep shows, open_softirq() is always called with the last argument
being NULL
block/blk-core.c: open_softirq(BLOCK_SOFTIRQ, blk_done_softirq, NULL);
kernel/hrtimer.c: open_softirq(HRTIMER_SOFTIRQ, run_hrtimer_softirq, NULL);
kernel/rcuclassic.c: open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
kernel/rcupreempt.c: open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
kernel/sched.c: open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
kernel/softirq.c: open_softirq(TASKLET_SOFTIRQ, tasklet_action, NULL);
kernel/softirq.c: open_softirq(HI_SOFTIRQ, tasklet_hi_action, NULL);
kernel/timer.c: open_softirq(TIMER_SOFTIRQ, run_timer_softirq, NULL);
net/core/dev.c: open_softirq(NET_TX_SOFTIRQ, net_tx_action, NULL);
net/core/dev.c: open_softirq(NET_RX_SOFTIRQ, net_rx_action, NULL);
This observation has already been made by Matthew Wilcox in June 2002
(http://www.cs.helsinki.fi/linux/linux-kernel/2002-25/0687.html)
"I notice that none of the current softirq routines use the data element
passed to them."
and the situation hasn't changed since them. So it appears we can safely
remove that extra argument to save 128 (54) bytes of kernel data (text).
Signed-off-by: Carlos R. Mafra <crmafra@ift.unesp.br>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2008-05-15 14:15:37 +00:00
|
|
|
open_softirq(TIMER_SOFTIRQ, run_timer_softirq);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* msleep - sleep safely even with waitqueue interruptions
|
|
|
|
* @msecs: Time in milliseconds to sleep for
|
|
|
|
*/
|
|
|
|
void msleep(unsigned int msecs)
|
|
|
|
{
|
|
|
|
unsigned long timeout = msecs_to_jiffies(msecs) + 1;
|
|
|
|
|
2005-09-10 07:27:24 +00:00
|
|
|
while (timeout)
|
|
|
|
timeout = schedule_timeout_uninterruptible(timeout);
|
2005-04-16 22:20:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
EXPORT_SYMBOL(msleep);
|
|
|
|
|
|
|
|
/**
|
2005-06-25 21:58:43 +00:00
|
|
|
* msleep_interruptible - sleep waiting for signals
|
2005-04-16 22:20:36 +00:00
|
|
|
* @msecs: Time in milliseconds to sleep for
|
|
|
|
*/
|
|
|
|
unsigned long msleep_interruptible(unsigned int msecs)
|
|
|
|
{
|
|
|
|
unsigned long timeout = msecs_to_jiffies(msecs) + 1;
|
|
|
|
|
2005-09-10 07:27:24 +00:00
|
|
|
while (timeout && !signal_pending(current))
|
|
|
|
timeout = schedule_timeout_interruptible(timeout);
|
2005-04-16 22:20:36 +00:00
|
|
|
return jiffies_to_msecs(timeout);
|
|
|
|
}
|
|
|
|
|
|
|
|
EXPORT_SYMBOL(msleep_interruptible);
|
timer: Added usleep_range timer
usleep_range is a finer precision implementations of msleep
and is designed to be a drop-in replacement for udelay where
a precise sleep / busy-wait is unnecessary.
Since an easy interface to hrtimers could lead to an undesired
proliferation of interrupts, we provide only a "range" API,
forcing the caller to think about an acceptable tolerance on
both ends and hopefully avoiding introducing another interrupt.
INTRO
As discussed here ( http://lkml.org/lkml/2007/8/3/250 ), msleep(1) is not
precise enough for many drivers (yes, sleep precision is an unfair notion,
but consistently sleeping for ~an order of magnitude greater than requested
is worth fixing). This patch adds a usleep API so that udelay does not have
to be used. Obviously not every udelay can be replaced (those in atomic
contexts or being used for simple bitbanging come to mind), but there are
many, many examples of
mydriver_write(...)
/* Wait for hardware to latch */
udelay(100)
in various drivers where a busy-wait loop is neither beneficial nor
necessary, but msleep simply does not provide enough precision and people
are using a busy-wait loop instead.
CONCERNS FROM THE RFC
Why is udelay a problem / necessary? Most callers of udelay are in device/
driver initialization code, which is serial...
As I see it, there is only benefit to sleeping over a delay; the
notion of "refactoring" areas that use udelay was presented, but
I see usleep as the refactoring. Consider i2c, if the bus is busy,
you need to wait a bit (say 100us) before trying again, your
current options are:
* udelay(100)
* msleep(1) <-- As noted above, actually as high as ~20ms
on some platforms, so not really an option
* Manually set up an hrtimer to try again in 100us (which
is what usleep does anyway...)
People choose the udelay route because it is EASY; we need to
provide a better easy route.
Device / driver / boot code is *currently* serial, but every few
months someone makes noise about parallelizing boot, and IMHO, a
little forward-thinking now is one less thing to worry about
if/when that ever happens
udelay's could be preempted
Sure, but if udelay plans on looping 1000 times, and it gets
preempted on loop 200, whenever it's scheduled again, it is
going to do the next 800 loops.
Is the interruptible case needed?
Probably not, but I see usleep as a very logical parallel to msleep,
so it made sense to include the "full" API. Processors are getting
faster (albeit not as quickly as they are becoming more parallel),
so if someone wanted to be interruptible for a few usecs, why not
let them? If this is a contentious point, I'm happy to remove it.
OTHER THOUGHTS
I believe there is also value in exposing the usleep_range option; it gives
the scheduler a lot more flexibility and allows the programmer to express
his intent much more clearly; it's something I would hope future driver
writers will take advantage of.
To get the results in the NUMBERS section below, I literally s/udelay/usleep
the kernel tree; I had to go in and undo the changes to the USB drivers, but
everything else booted successfully; I find that extremely telling in and
of itself -- many people are using a delay API where a sleep will suit them
just fine.
SOME ATTEMPTS AT NUMBERS
It turns out that calculating quantifiable benefit on this is challenging,
so instead I will simply present the current state of things, and I hope
this to be sufficient:
How many udelay calls are there in 2.6.35-rc5?
udealy(ARG) >= | COUNT
1000 | 319
500 | 414
100 | 1146
20 | 1832
I am working on Android, so that is my focus for this. The following table
is a modified usleep that simply printk's the amount of time requested to
sleep; these tests were run on a kernel with udelay >= 20 --> usleep
"boot" is power-on to lock screen
"power collapse" is when the power button is pushed and the device suspends
"resume" is when the power button is pushed and the lock screen is displayed
(no touchscreen events or anything, just turning on the display)
"use device" is from the unlock swipe to clicking around a bit; there is no
sd card in this phone, so fail loading music, video, camera
ACTION | TOTAL NUMBER OF USLEEP CALLS | NET TIME (us)
boot | 22 | 1250
power-collapse | 9 | 1200
resume | 5 | 500
use device | 59 | 7700
The most interesting category to me is the "use device" field; 7700us of
busy-wait time that could be put towards better responsiveness, or at the
least less power usage.
Signed-off-by: Patrick Pannuto <ppannuto@codeaurora.org>
Cc: apw@canonical.com
Cc: corbet@lwn.net
Cc: arjan@linux.intel.com
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-08-02 22:01:04 +00:00
|
|
|
|
|
|
|
/**
|
2016-05-31 21:23:02 +00:00
|
|
|
* usleep_range - Sleep for an approximate time
|
timer: Added usleep_range timer
usleep_range is a finer precision implementations of msleep
and is designed to be a drop-in replacement for udelay where
a precise sleep / busy-wait is unnecessary.
Since an easy interface to hrtimers could lead to an undesired
proliferation of interrupts, we provide only a "range" API,
forcing the caller to think about an acceptable tolerance on
both ends and hopefully avoiding introducing another interrupt.
INTRO
As discussed here ( http://lkml.org/lkml/2007/8/3/250 ), msleep(1) is not
precise enough for many drivers (yes, sleep precision is an unfair notion,
but consistently sleeping for ~an order of magnitude greater than requested
is worth fixing). This patch adds a usleep API so that udelay does not have
to be used. Obviously not every udelay can be replaced (those in atomic
contexts or being used for simple bitbanging come to mind), but there are
many, many examples of
mydriver_write(...)
/* Wait for hardware to latch */
udelay(100)
in various drivers where a busy-wait loop is neither beneficial nor
necessary, but msleep simply does not provide enough precision and people
are using a busy-wait loop instead.
CONCERNS FROM THE RFC
Why is udelay a problem / necessary? Most callers of udelay are in device/
driver initialization code, which is serial...
As I see it, there is only benefit to sleeping over a delay; the
notion of "refactoring" areas that use udelay was presented, but
I see usleep as the refactoring. Consider i2c, if the bus is busy,
you need to wait a bit (say 100us) before trying again, your
current options are:
* udelay(100)
* msleep(1) <-- As noted above, actually as high as ~20ms
on some platforms, so not really an option
* Manually set up an hrtimer to try again in 100us (which
is what usleep does anyway...)
People choose the udelay route because it is EASY; we need to
provide a better easy route.
Device / driver / boot code is *currently* serial, but every few
months someone makes noise about parallelizing boot, and IMHO, a
little forward-thinking now is one less thing to worry about
if/when that ever happens
udelay's could be preempted
Sure, but if udelay plans on looping 1000 times, and it gets
preempted on loop 200, whenever it's scheduled again, it is
going to do the next 800 loops.
Is the interruptible case needed?
Probably not, but I see usleep as a very logical parallel to msleep,
so it made sense to include the "full" API. Processors are getting
faster (albeit not as quickly as they are becoming more parallel),
so if someone wanted to be interruptible for a few usecs, why not
let them? If this is a contentious point, I'm happy to remove it.
OTHER THOUGHTS
I believe there is also value in exposing the usleep_range option; it gives
the scheduler a lot more flexibility and allows the programmer to express
his intent much more clearly; it's something I would hope future driver
writers will take advantage of.
To get the results in the NUMBERS section below, I literally s/udelay/usleep
the kernel tree; I had to go in and undo the changes to the USB drivers, but
everything else booted successfully; I find that extremely telling in and
of itself -- many people are using a delay API where a sleep will suit them
just fine.
SOME ATTEMPTS AT NUMBERS
It turns out that calculating quantifiable benefit on this is challenging,
so instead I will simply present the current state of things, and I hope
this to be sufficient:
How many udelay calls are there in 2.6.35-rc5?
udealy(ARG) >= | COUNT
1000 | 319
500 | 414
100 | 1146
20 | 1832
I am working on Android, so that is my focus for this. The following table
is a modified usleep that simply printk's the amount of time requested to
sleep; these tests were run on a kernel with udelay >= 20 --> usleep
"boot" is power-on to lock screen
"power collapse" is when the power button is pushed and the device suspends
"resume" is when the power button is pushed and the lock screen is displayed
(no touchscreen events or anything, just turning on the display)
"use device" is from the unlock swipe to clicking around a bit; there is no
sd card in this phone, so fail loading music, video, camera
ACTION | TOTAL NUMBER OF USLEEP CALLS | NET TIME (us)
boot | 22 | 1250
power-collapse | 9 | 1200
resume | 5 | 500
use device | 59 | 7700
The most interesting category to me is the "use device" field; 7700us of
busy-wait time that could be put towards better responsiveness, or at the
least less power usage.
Signed-off-by: Patrick Pannuto <ppannuto@codeaurora.org>
Cc: apw@canonical.com
Cc: corbet@lwn.net
Cc: arjan@linux.intel.com
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-08-02 22:01:04 +00:00
|
|
|
* @min: Minimum time in usecs to sleep
|
|
|
|
* @max: Maximum time in usecs to sleep
|
2016-05-31 21:23:02 +00:00
|
|
|
*
|
|
|
|
* In non-atomic context where the exact wakeup time is flexible, use
|
|
|
|
* usleep_range() instead of udelay(). The sleep improves responsiveness
|
|
|
|
* by avoiding the CPU-hogging busy-wait of udelay(), and the range reduces
|
|
|
|
* power usage by allowing hrtimers to take advantage of an already-
|
|
|
|
* scheduled interrupt instead of scheduling a new one just for this sleep.
|
timer: Added usleep_range timer
usleep_range is a finer precision implementations of msleep
and is designed to be a drop-in replacement for udelay where
a precise sleep / busy-wait is unnecessary.
Since an easy interface to hrtimers could lead to an undesired
proliferation of interrupts, we provide only a "range" API,
forcing the caller to think about an acceptable tolerance on
both ends and hopefully avoiding introducing another interrupt.
INTRO
As discussed here ( http://lkml.org/lkml/2007/8/3/250 ), msleep(1) is not
precise enough for many drivers (yes, sleep precision is an unfair notion,
but consistently sleeping for ~an order of magnitude greater than requested
is worth fixing). This patch adds a usleep API so that udelay does not have
to be used. Obviously not every udelay can be replaced (those in atomic
contexts or being used for simple bitbanging come to mind), but there are
many, many examples of
mydriver_write(...)
/* Wait for hardware to latch */
udelay(100)
in various drivers where a busy-wait loop is neither beneficial nor
necessary, but msleep simply does not provide enough precision and people
are using a busy-wait loop instead.
CONCERNS FROM THE RFC
Why is udelay a problem / necessary? Most callers of udelay are in device/
driver initialization code, which is serial...
As I see it, there is only benefit to sleeping over a delay; the
notion of "refactoring" areas that use udelay was presented, but
I see usleep as the refactoring. Consider i2c, if the bus is busy,
you need to wait a bit (say 100us) before trying again, your
current options are:
* udelay(100)
* msleep(1) <-- As noted above, actually as high as ~20ms
on some platforms, so not really an option
* Manually set up an hrtimer to try again in 100us (which
is what usleep does anyway...)
People choose the udelay route because it is EASY; we need to
provide a better easy route.
Device / driver / boot code is *currently* serial, but every few
months someone makes noise about parallelizing boot, and IMHO, a
little forward-thinking now is one less thing to worry about
if/when that ever happens
udelay's could be preempted
Sure, but if udelay plans on looping 1000 times, and it gets
preempted on loop 200, whenever it's scheduled again, it is
going to do the next 800 loops.
Is the interruptible case needed?
Probably not, but I see usleep as a very logical parallel to msleep,
so it made sense to include the "full" API. Processors are getting
faster (albeit not as quickly as they are becoming more parallel),
so if someone wanted to be interruptible for a few usecs, why not
let them? If this is a contentious point, I'm happy to remove it.
OTHER THOUGHTS
I believe there is also value in exposing the usleep_range option; it gives
the scheduler a lot more flexibility and allows the programmer to express
his intent much more clearly; it's something I would hope future driver
writers will take advantage of.
To get the results in the NUMBERS section below, I literally s/udelay/usleep
the kernel tree; I had to go in and undo the changes to the USB drivers, but
everything else booted successfully; I find that extremely telling in and
of itself -- many people are using a delay API where a sleep will suit them
just fine.
SOME ATTEMPTS AT NUMBERS
It turns out that calculating quantifiable benefit on this is challenging,
so instead I will simply present the current state of things, and I hope
this to be sufficient:
How many udelay calls are there in 2.6.35-rc5?
udealy(ARG) >= | COUNT
1000 | 319
500 | 414
100 | 1146
20 | 1832
I am working on Android, so that is my focus for this. The following table
is a modified usleep that simply printk's the amount of time requested to
sleep; these tests were run on a kernel with udelay >= 20 --> usleep
"boot" is power-on to lock screen
"power collapse" is when the power button is pushed and the device suspends
"resume" is when the power button is pushed and the lock screen is displayed
(no touchscreen events or anything, just turning on the display)
"use device" is from the unlock swipe to clicking around a bit; there is no
sd card in this phone, so fail loading music, video, camera
ACTION | TOTAL NUMBER OF USLEEP CALLS | NET TIME (us)
boot | 22 | 1250
power-collapse | 9 | 1200
resume | 5 | 500
use device | 59 | 7700
The most interesting category to me is the "use device" field; 7700us of
busy-wait time that could be put towards better responsiveness, or at the
least less power usage.
Signed-off-by: Patrick Pannuto <ppannuto@codeaurora.org>
Cc: apw@canonical.com
Cc: corbet@lwn.net
Cc: arjan@linux.intel.com
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-08-02 22:01:04 +00:00
|
|
|
*/
|
2015-04-14 21:09:30 +00:00
|
|
|
void __sched usleep_range(unsigned long min, unsigned long max)
|
timer: Added usleep_range timer
usleep_range is a finer precision implementations of msleep
and is designed to be a drop-in replacement for udelay where
a precise sleep / busy-wait is unnecessary.
Since an easy interface to hrtimers could lead to an undesired
proliferation of interrupts, we provide only a "range" API,
forcing the caller to think about an acceptable tolerance on
both ends and hopefully avoiding introducing another interrupt.
INTRO
As discussed here ( http://lkml.org/lkml/2007/8/3/250 ), msleep(1) is not
precise enough for many drivers (yes, sleep precision is an unfair notion,
but consistently sleeping for ~an order of magnitude greater than requested
is worth fixing). This patch adds a usleep API so that udelay does not have
to be used. Obviously not every udelay can be replaced (those in atomic
contexts or being used for simple bitbanging come to mind), but there are
many, many examples of
mydriver_write(...)
/* Wait for hardware to latch */
udelay(100)
in various drivers where a busy-wait loop is neither beneficial nor
necessary, but msleep simply does not provide enough precision and people
are using a busy-wait loop instead.
CONCERNS FROM THE RFC
Why is udelay a problem / necessary? Most callers of udelay are in device/
driver initialization code, which is serial...
As I see it, there is only benefit to sleeping over a delay; the
notion of "refactoring" areas that use udelay was presented, but
I see usleep as the refactoring. Consider i2c, if the bus is busy,
you need to wait a bit (say 100us) before trying again, your
current options are:
* udelay(100)
* msleep(1) <-- As noted above, actually as high as ~20ms
on some platforms, so not really an option
* Manually set up an hrtimer to try again in 100us (which
is what usleep does anyway...)
People choose the udelay route because it is EASY; we need to
provide a better easy route.
Device / driver / boot code is *currently* serial, but every few
months someone makes noise about parallelizing boot, and IMHO, a
little forward-thinking now is one less thing to worry about
if/when that ever happens
udelay's could be preempted
Sure, but if udelay plans on looping 1000 times, and it gets
preempted on loop 200, whenever it's scheduled again, it is
going to do the next 800 loops.
Is the interruptible case needed?
Probably not, but I see usleep as a very logical parallel to msleep,
so it made sense to include the "full" API. Processors are getting
faster (albeit not as quickly as they are becoming more parallel),
so if someone wanted to be interruptible for a few usecs, why not
let them? If this is a contentious point, I'm happy to remove it.
OTHER THOUGHTS
I believe there is also value in exposing the usleep_range option; it gives
the scheduler a lot more flexibility and allows the programmer to express
his intent much more clearly; it's something I would hope future driver
writers will take advantage of.
To get the results in the NUMBERS section below, I literally s/udelay/usleep
the kernel tree; I had to go in and undo the changes to the USB drivers, but
everything else booted successfully; I find that extremely telling in and
of itself -- many people are using a delay API where a sleep will suit them
just fine.
SOME ATTEMPTS AT NUMBERS
It turns out that calculating quantifiable benefit on this is challenging,
so instead I will simply present the current state of things, and I hope
this to be sufficient:
How many udelay calls are there in 2.6.35-rc5?
udealy(ARG) >= | COUNT
1000 | 319
500 | 414
100 | 1146
20 | 1832
I am working on Android, so that is my focus for this. The following table
is a modified usleep that simply printk's the amount of time requested to
sleep; these tests were run on a kernel with udelay >= 20 --> usleep
"boot" is power-on to lock screen
"power collapse" is when the power button is pushed and the device suspends
"resume" is when the power button is pushed and the lock screen is displayed
(no touchscreen events or anything, just turning on the display)
"use device" is from the unlock swipe to clicking around a bit; there is no
sd card in this phone, so fail loading music, video, camera
ACTION | TOTAL NUMBER OF USLEEP CALLS | NET TIME (us)
boot | 22 | 1250
power-collapse | 9 | 1200
resume | 5 | 500
use device | 59 | 7700
The most interesting category to me is the "use device" field; 7700us of
busy-wait time that could be put towards better responsiveness, or at the
least less power usage.
Signed-off-by: Patrick Pannuto <ppannuto@codeaurora.org>
Cc: apw@canonical.com
Cc: corbet@lwn.net
Cc: arjan@linux.intel.com
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-08-02 22:01:04 +00:00
|
|
|
{
|
2016-10-21 15:58:50 +00:00
|
|
|
ktime_t exp = ktime_add_us(ktime_get(), min);
|
|
|
|
u64 delta = (u64)(max - min) * NSEC_PER_USEC;
|
|
|
|
|
|
|
|
for (;;) {
|
|
|
|
__set_current_state(TASK_UNINTERRUPTIBLE);
|
|
|
|
/* Do not return before the requested sleep time has elapsed */
|
|
|
|
if (!schedule_hrtimeout_range(&exp, delta, HRTIMER_MODE_ABS))
|
|
|
|
break;
|
|
|
|
}
|
timer: Added usleep_range timer
usleep_range is a finer precision implementations of msleep
and is designed to be a drop-in replacement for udelay where
a precise sleep / busy-wait is unnecessary.
Since an easy interface to hrtimers could lead to an undesired
proliferation of interrupts, we provide only a "range" API,
forcing the caller to think about an acceptable tolerance on
both ends and hopefully avoiding introducing another interrupt.
INTRO
As discussed here ( http://lkml.org/lkml/2007/8/3/250 ), msleep(1) is not
precise enough for many drivers (yes, sleep precision is an unfair notion,
but consistently sleeping for ~an order of magnitude greater than requested
is worth fixing). This patch adds a usleep API so that udelay does not have
to be used. Obviously not every udelay can be replaced (those in atomic
contexts or being used for simple bitbanging come to mind), but there are
many, many examples of
mydriver_write(...)
/* Wait for hardware to latch */
udelay(100)
in various drivers where a busy-wait loop is neither beneficial nor
necessary, but msleep simply does not provide enough precision and people
are using a busy-wait loop instead.
CONCERNS FROM THE RFC
Why is udelay a problem / necessary? Most callers of udelay are in device/
driver initialization code, which is serial...
As I see it, there is only benefit to sleeping over a delay; the
notion of "refactoring" areas that use udelay was presented, but
I see usleep as the refactoring. Consider i2c, if the bus is busy,
you need to wait a bit (say 100us) before trying again, your
current options are:
* udelay(100)
* msleep(1) <-- As noted above, actually as high as ~20ms
on some platforms, so not really an option
* Manually set up an hrtimer to try again in 100us (which
is what usleep does anyway...)
People choose the udelay route because it is EASY; we need to
provide a better easy route.
Device / driver / boot code is *currently* serial, but every few
months someone makes noise about parallelizing boot, and IMHO, a
little forward-thinking now is one less thing to worry about
if/when that ever happens
udelay's could be preempted
Sure, but if udelay plans on looping 1000 times, and it gets
preempted on loop 200, whenever it's scheduled again, it is
going to do the next 800 loops.
Is the interruptible case needed?
Probably not, but I see usleep as a very logical parallel to msleep,
so it made sense to include the "full" API. Processors are getting
faster (albeit not as quickly as they are becoming more parallel),
so if someone wanted to be interruptible for a few usecs, why not
let them? If this is a contentious point, I'm happy to remove it.
OTHER THOUGHTS
I believe there is also value in exposing the usleep_range option; it gives
the scheduler a lot more flexibility and allows the programmer to express
his intent much more clearly; it's something I would hope future driver
writers will take advantage of.
To get the results in the NUMBERS section below, I literally s/udelay/usleep
the kernel tree; I had to go in and undo the changes to the USB drivers, but
everything else booted successfully; I find that extremely telling in and
of itself -- many people are using a delay API where a sleep will suit them
just fine.
SOME ATTEMPTS AT NUMBERS
It turns out that calculating quantifiable benefit on this is challenging,
so instead I will simply present the current state of things, and I hope
this to be sufficient:
How many udelay calls are there in 2.6.35-rc5?
udealy(ARG) >= | COUNT
1000 | 319
500 | 414
100 | 1146
20 | 1832
I am working on Android, so that is my focus for this. The following table
is a modified usleep that simply printk's the amount of time requested to
sleep; these tests were run on a kernel with udelay >= 20 --> usleep
"boot" is power-on to lock screen
"power collapse" is when the power button is pushed and the device suspends
"resume" is when the power button is pushed and the lock screen is displayed
(no touchscreen events or anything, just turning on the display)
"use device" is from the unlock swipe to clicking around a bit; there is no
sd card in this phone, so fail loading music, video, camera
ACTION | TOTAL NUMBER OF USLEEP CALLS | NET TIME (us)
boot | 22 | 1250
power-collapse | 9 | 1200
resume | 5 | 500
use device | 59 | 7700
The most interesting category to me is the "use device" field; 7700us of
busy-wait time that could be put towards better responsiveness, or at the
least less power usage.
Signed-off-by: Patrick Pannuto <ppannuto@codeaurora.org>
Cc: apw@canonical.com
Cc: corbet@lwn.net
Cc: arjan@linux.intel.com
Cc: Randy Dunlap <rdunlap@xenotime.net>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2010-08-02 22:01:04 +00:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(usleep_range);
|