[ Upstream commit 39afe5d6fc ]
There are scenarios where non-affine wakeups are incorrectly counted as
affine wakeups by schedstats.
When wake_affine_idle() returns prev_cpu which doesn't equal to
nr_cpumask_bits, it will slip through the check: target == nr_cpumask_bits
in wake_affine() and be counted as if target == this_cpu in schedstats.
Replace target == nr_cpumask_bits with target != this_cpu to make sure
affine wakeups are accurately tallied.
Fixes: 806486c377 (sched/fair: Do not migrate if the prev_cpu is idle)
Suggested-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Signed-off-by: Libo Chen <libo.chen@oracle.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Gautham R. Shenoy <gautham.shenoy@amd.com>
Link: https://lore.kernel.org/r/20220810223313.386614-1-libo.chen@oracle.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ceeadb83ae ]
If we want to use the schedstats facility to trace other sched classes, we
should make it independent of fair sched class. The struct sched_statistics
is the schedular statistics of a task_struct or a task_group. So we can
move it into struct task_struct and struct task_group to achieve the goal.
After the patch, schestats are orgnized as follows,
struct task_struct {
...
struct sched_entity se;
struct sched_rt_entity rt;
struct sched_dl_entity dl;
...
struct sched_statistics stats;
...
};
Regarding the task group, schedstats is only supported for fair group
sched, and a new struct sched_entity_stats is introduced, suggested by
Peter -
struct sched_entity_stats {
struct sched_entity se;
struct sched_statistics stats;
} __no_randomize_layout;
Then with the se in a task_group, we can easily get the stats.
The sched_statistics members may be frequently modified when schedstats is
enabled, in order to avoid impacting on random data which may in the same
cacheline with them, the struct sched_statistics is defined as cacheline
aligned.
As this patch changes the core struct of scheduler, so I verified the
performance it may impact on the scheduler with 'perf bench sched
pipe', suggested by Mel. Below is the result, in which all the values
are in usecs/op.
Before After
kernel.sched_schedstats=0 5.2~5.4 5.2~5.4
kernel.sched_schedstats=1 5.3~5.5 5.3~5.5
[These data is a little difference with the earlier version, that is
because my old test machine is destroyed so I have to use a new
different test machine.]
Almost no impact on the sched performance.
No functional change.
[lkp@intel.com: reported build failure in earlier version]
Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/20210905143547.4668-3-laoar.shao@gmail.com
Stable-dep-of: 39afe5d6fc ("sched/fair: Fix inaccurate tally of ttwu_move_affine")
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit da07d2f9c1 upstream.
Traversing the Perf Domains requires rcu_read_lock() to be held and is
conditional on sched_energy_enabled(). Ensure right protections applied.
Also skip capacity inversion detection for our own pd; which was an
error.
Fixes: 44c7b80bff ("sched/fair: Detect capacity inversion")
Reported-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20230112122708.330667-3-qyousef@layalina.io
(cherry picked from commit da07d2f9c1)
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit aa69c36f31 upstream.
We do consider thermal pressure in util_fits_cpu() for uclamp_min only.
With the exception of the biggest cores which by definition are the max
performance point of the system and all tasks by definition should fit.
Even under thermal pressure, the capacity of the biggest CPU is the
highest in the system and should still fit every task. Except when it
reaches capacity inversion point, then this is no longer true.
We can handle this by using the inverted capacity as capacity_orig in
util_fits_cpu(). Which not only addresses the problem above, but also
ensure uclamp_max now considers the inverted capacity. Force fitting
a task when a CPU is in this adverse state will contribute to making the
thermal throttling last longer.
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-10-qais.yousef@arm.com
(cherry picked from commit aa69c36f31)
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 44c7b80bff upstream.
Check each performance domain to see if thermal pressure is causing its
capacity to be lower than another performance domain.
We assume that each performance domain has CPUs with the same
capacities, which is similar to an assumption made in energy_model.c
We also assume that thermal pressure impacts all CPUs in a performance
domain equally.
If there're multiple performance domains with the same capacity_orig, we
will trigger a capacity inversion if the domain is under thermal
pressure.
The new cpu_in_capacity_inversion() should help users to know when
information about capacity_orig are not reliable and can opt in to use
the inverted capacity as the 'actual' capacity_orig.
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-9-qais.yousef@arm.com
(cherry picked from commit 44c7b80bff)
[fix trivial conflict in kernel/sched/sched.h due to code shuffling]
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit d81304bc61 upstream.
If the utilization of the woken up task is 0, we skip the energy
calculation because it has no impact.
But if the task is boosted (uclamp_min != 0) will have an impact on task
placement and frequency selection. Only skip if the util is truly
0 after applying uclamp values.
Change uclamp_task_cpu() signature to avoid unnecessary additional calls
to uclamp_eff_get(). feec() is the only user now.
Fixes: 732cd75b8c ("sched/fair: Select an energy-efficient CPU on task wake-up")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-8-qais.yousef@arm.com
(cherry picked from commit d81304bc61)
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit c56ab1b350 upstream.
So that it is now uclamp aware.
This fixes a major problem of busy tasks capped with UCLAMP_MAX keeping
the system in overutilized state which disables EAS and leads to wasting
energy in the long run.
Without this patch running a busy background activity like JIT
compilation on Pixel 6 causes the system to be in overutilized state
74.5% of the time.
With this patch this goes down to 9.79%.
It also fixes another problem when long running tasks that have their
UCLAMP_MIN changed while running such that they need to upmigrate to
honour the new UCLAMP_MIN value. The upmigration doesn't get triggered
because overutilized state never gets set in this state, hence misfit
migration never happens at tick in this case until the task wakes up
again.
Fixes: af24bde8df ("sched/uclamp: Add uclamp support to energy_compute()")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-7-qais.yousef@arm.com
(cherry picked from commit c56ab1b350)
[Fixed trivial conflict in cpu_overutilized() - use cpu_util() instead
of cpu_util_cfs()]
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 244226035a upstream.
As reported by Yun Hsiang [1], if a task has its uclamp_min >= 0.8 * 1024,
it'll always pick the previous CPU because fits_capacity() will always
return false in this case.
The new util_fits_cpu() logic should handle this correctly for us beside
more corner cases where similar failures could occur, like when using
UCLAMP_MAX.
We open code uclamp_rq_util_with() except for the clamp() part,
util_fits_cpu() needs the 'raw' values to be passed to it.
Also introduce uclamp_rq_{set, get}() shorthand accessors to get uclamp
value for the rq. Makes the code more readable and ensures the right
rules (use READ_ONCE/WRITE_ONCE) are respected transparently.
[1] https://lists.linaro.org/pipermail/eas-dev/2020-July/001488.html
Fixes: 1d42509e47 ("sched/fair: Make EAS wakeup placement consider uclamp restrictions")
Reported-by: Yun Hsiang <hsiang023167@gmail.com>
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-4-qais.yousef@arm.com
(cherry picked from commit 244226035a)
[Conflict in kernel/sched/fair.c mainly due to new automatic variables
being added on master vs 5.15]
Signed-off-by: Qais Yousef (Google) <qyousef@layalina.io>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 91dcf1e806 ]
When local group is fully busy but its average load is above system load,
computing the imbalance will overflow and local group is not the best
target for pulling this load.
Fixes: 0b0695f2b3 ("sched/fair: Rework load_balance()")
Reported-by: Tingjia Cao <tjcao980311@gmail.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Tingjia Cao <tjcao980311@gmail.com>
Link: https://lore.kernel.org/lkml/CABcWv9_DAhVBOq2=W=2ypKE9dKM5s2DvoV8-U0+GDwwuKZ89jQ@mail.gmail.com/T/
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 0635490078 ]
In calculate_imbalance function, when the value of local->avg_load is
greater than or equal to busiest->avg_load, the calculated sds->avg_load is
not used. So this calculation can be placed in a more appropriate position.
Signed-off-by: zgpeng <zgpeng@tencent.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Samuel Liao <samuelliao@tencent.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/1649239025-10010-1-git-send-email-zgpeng@tencent.com
Stable-dep-of: 91dcf1e806 ("sched/fair: Fix imbalance overflow")
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit a53ce18cac upstream.
Commit 829c1651e9 ("sched/fair: sanitize vruntime of entity being placed")
fixes an overflowing bug, but ignore a case that se->exec_start is reset
after a migration.
For fixing this case, we delay the reset of se->exec_start after
placing the entity which se->exec_start to detect long sleeping task.
In order to take into account a possible divergence between the clock_task
of 2 rqs, we increase the threshold to around 104 days.
Fixes: 829c1651e9 ("sched/fair: sanitize vruntime of entity being placed")
Originally-by: Zhang Qiao <zhangqiao22@huawei.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Zhang Qiao <zhangqiao22@huawei.com>
Link: https://lore.kernel.org/r/20230317160810.107988-1-vincent.guittot@linaro.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 829c1651e9 upstream.
When a scheduling entity is placed onto cfs_rq, its vruntime is pulled
to the base level (around cfs_rq->min_vruntime), so that the entity
doesn't gain extra boost when placed backwards.
However, if the entity being placed wasn't executed for a long time, its
vruntime may get too far behind (e.g. while cfs_rq was executing a
low-weight hog), which can inverse the vruntime comparison due to s64
overflow. This results in the entity being placed with its original
vruntime way forwards, so that it will effectively never get to the cpu.
To prevent that, ignore the vruntime of the entity being placed if it
didn't execute for much longer than the characteristic sheduler time
scale.
[rkagan: formatted, adjusted commit log, comments, cutoff value]
Signed-off-by: Zhang Qiao <zhangqiao22@huawei.com>
Co-developed-by: Roman Kagan <rkagan@amazon.de>
Signed-off-by: Roman Kagan <rkagan@amazon.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20230130122216.3555094-1-rkagan@amazon.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit a2e7f03ed2 ]
Use the new util_fits_cpu() to ensure migration margin and capacity
pressure are taken into account correctly when uclamp is being used
otherwise we will fail to consider CPUs as fitting in scenarios where
they should.
s/asym_fits_capacity/asym_fits_cpu/ to better reflect what it does now.
Fixes: b4c9c9f156 ("sched/fair: Prefer prev cpu in asymmetric wakeup path")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-6-qais.yousef@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 740cf8a760 ]
Create an inline helper for conditional code to be only executed on
asymmetric CPU capacity systems. This makes these (currently ~10 and
future) conditions a lot more readable.
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20220729111305.1275158-2-dietmar.eggemann@arm.com
Stable-dep-of: a2e7f03ed2 ("sched/uclamp: Make asym_fits_capacity() use util_fits_cpu()")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit a7ba894821 ]
Since commit 89aafd67f2 ("sched/fair: Use prev instead of new target as recent_used_cpu"),
p->recent_used_cpu is unconditionnaly set with prev.
Fixes: 89aafd67f2 ("sched/fair: Use prev instead of new target as recent_used_cpu")
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lkml.kernel.org/r/20210928103544.27489-1-vincent.guittot@linaro.org
Stable-dep-of: a2e7f03ed2 ("sched/uclamp: Make asym_fits_capacity() use util_fits_cpu()")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit b759caa1d9 ]
Use the new util_fits_cpu() to ensure migration margin and capacity
pressure are taken into account correctly when uclamp is being used
otherwise we will fail to consider CPUs as fitting in scenarios where
they should.
Fixes: b4c9c9f156 ("sched/fair: Prefer prev cpu in asymmetric wakeup path")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-5-qais.yousef@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit b48e16a697 ]
So that the new uclamp rules in regard to migration margin and capacity
pressure are taken into account correctly.
Fixes: a7008c07a5 ("sched/fair: Make task_fits_capacity() consider uclamp restrictions")
Co-developed-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-3-qais.yousef@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 48d5e9daa8 ]
fits_capacity() verifies that a util is within 20% margin of the
capacity of a CPU, which is an attempt to speed up upmigration.
But when uclamp is used, this 20% margin is problematic because for
example if a task is boosted to 1024, then it will not fit on any CPU
according to fits_capacity() logic.
Or if a task is boosted to capacity_orig_of(medium_cpu). The task will
end up on big instead on the desired medium CPU.
Similar corner cases exist for uclamp and usage of capacity_of().
Slightest irq pressure on biggest CPU for example will make a 1024
boosted task look like it can't fit.
What we really want is for uclamp comparisons to ignore the migration
margin and capacity pressure, yet retain them for when checking the
_actual_ util signal.
For example, task p:
p->util_avg = 300
p->uclamp[UCLAMP_MIN] = 1024
Will fit a big CPU. But
p->util_avg = 900
p->uclamp[UCLAMP_MIN] = 1024
will not, this should trigger overutilized state because the big CPU is
now *actually* being saturated.
Similar reasoning applies to capping tasks with UCLAMP_MAX. For example:
p->util_avg = 1024
p->uclamp[UCLAMP_MAX] = capacity_orig_of(medium_cpu)
Should fit the task on medium cpus without triggering overutilized
state.
Inlined comments expand more on desired behavior in more scenarios.
Introduce new util_fits_cpu() function which encapsulates the new logic.
The new function is not used anywhere yet, but will be used to update
various users of fits_capacity() in later patches.
Fixes: af24bde8df ("sched/uclamp: Add uclamp support to energy_compute()")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20220804143609.515789-2-qais.yousef@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ef8df9798d ]
task_util and capacity are comparable unsigned long values. There is no
need for an intermidiate implicit signed cast.
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20211207095755.859972-1-vincent.donnefort@arm.com
Stable-dep-of: 48d5e9daa8 ("sched/uclamp: Fix relationship between uclamp and migration margin")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 70fb5ccf2e ]
[Problem Statement]
select_idle_cpu() might spend too much time searching for an idle CPU,
when the system is overloaded.
The following histogram is the time spent in select_idle_cpu(),
when running 224 instances of netperf on a system with 112 CPUs
per LLC domain:
@usecs:
[0] 533 | |
[1] 5495 | |
[2, 4) 12008 | |
[4, 8) 239252 | |
[8, 16) 4041924 |@@@@@@@@@@@@@@ |
[16, 32) 12357398 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 14820255 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[64, 128) 13047682 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 8235013 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 4507667 |@@@@@@@@@@@@@@@ |
[512, 1K) 2600472 |@@@@@@@@@ |
[1K, 2K) 927912 |@@@ |
[2K, 4K) 218720 | |
[4K, 8K) 98161 | |
[8K, 16K) 37722 | |
[16K, 32K) 6715 | |
[32K, 64K) 477 | |
[64K, 128K) 7 | |
netperf latency usecs:
=======
case load Lat_99th std%
TCP_RR thread-224 257.39 ( 0.21)
The time spent in select_idle_cpu() is visible to netperf and might have a negative
impact.
[Symptom analysis]
The patch [1] from Mel Gorman has been applied to track the efficiency
of select_idle_sibling. Copy the indicators here:
SIS Search Efficiency(se_eff%):
A ratio expressed as a percentage of runqueues scanned versus
idle CPUs found. A 100% efficiency indicates that the target,
prev or recent CPU of a task was idle at wakeup. The lower the
efficiency, the more runqueues were scanned before an idle CPU
was found.
SIS Domain Search Efficiency(dom_eff%):
Similar, except only for the slower SIS
patch.
SIS Fast Success Rate(fast_rate%):
Percentage of SIS that used target, prev or
recent CPUs.
SIS Success rate(success_rate%):
Percentage of scans that found an idle CPU.
The test is based on Aubrey's schedtests tool, including netperf, hackbench,
schbench and tbench.
Test on vanilla kernel:
schedstat_parse.py -f netperf_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
TCP_RR 28 threads 99.978 18.535 99.995 100.000
TCP_RR 56 threads 99.397 5.671 99.964 100.000
TCP_RR 84 threads 21.721 6.818 73.632 100.000
TCP_RR 112 threads 12.500 5.533 59.000 100.000
TCP_RR 140 threads 8.524 4.535 49.020 100.000
TCP_RR 168 threads 6.438 3.945 40.309 99.999
TCP_RR 196 threads 5.397 3.718 32.320 99.982
TCP_RR 224 threads 4.874 3.661 25.775 99.767
UDP_RR 28 threads 99.988 17.704 99.997 100.000
UDP_RR 56 threads 99.528 5.977 99.970 100.000
UDP_RR 84 threads 24.219 6.992 76.479 100.000
UDP_RR 112 threads 13.907 5.706 62.538 100.000
UDP_RR 140 threads 9.408 4.699 52.519 100.000
UDP_RR 168 threads 7.095 4.077 44.352 100.000
UDP_RR 196 threads 5.757 3.775 35.764 99.991
UDP_RR 224 threads 5.124 3.704 28.748 99.860
schedstat_parse.py -f schbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
normal 1 mthread 99.152 6.400 99.941 100.000
normal 2 mthreads 97.844 4.003 99.908 100.000
normal 3 mthreads 96.395 2.118 99.917 99.998
normal 4 mthreads 55.288 1.451 98.615 99.804
normal 5 mthreads 7.004 1.870 45.597 61.036
normal 6 mthreads 3.354 1.346 20.777 34.230
normal 7 mthreads 2.183 1.028 11.257 21.055
normal 8 mthreads 1.653 0.825 7.849 15.549
schedstat_parse.py -f hackbench_vanilla.log
(each group has 28 tasks)
case load se_eff% dom_eff% fast_rate% success_rate%
process-pipe 1 group 99.991 7.692 99.999 100.000
process-pipe 2 groups 99.934 4.615 99.997 100.000
process-pipe 3 groups 99.597 3.198 99.987 100.000
process-pipe 4 groups 98.378 2.464 99.958 100.000
process-pipe 5 groups 27.474 3.653 89.811 99.800
process-pipe 6 groups 20.201 4.098 82.763 99.570
process-pipe 7 groups 16.423 4.156 77.398 99.316
process-pipe 8 groups 13.165 3.920 72.232 98.828
process-sockets 1 group 99.977 5.882 99.999 100.000
process-sockets 2 groups 99.927 5.505 99.996 100.000
process-sockets 3 groups 99.397 3.250 99.980 100.000
process-sockets 4 groups 79.680 4.258 98.864 99.998
process-sockets 5 groups 7.673 2.503 63.659 92.115
process-sockets 6 groups 4.642 1.584 58.946 88.048
process-sockets 7 groups 3.493 1.379 49.816 81.164
process-sockets 8 groups 3.015 1.407 40.845 75.500
threads-pipe 1 group 99.997 0.000 100.000 100.000
threads-pipe 2 groups 99.894 2.932 99.997 100.000
threads-pipe 3 groups 99.611 4.117 99.983 100.000
threads-pipe 4 groups 97.703 2.624 99.937 100.000
threads-pipe 5 groups 22.919 3.623 87.150 99.764
threads-pipe 6 groups 18.016 4.038 80.491 99.557
threads-pipe 7 groups 14.663 3.991 75.239 99.247
threads-pipe 8 groups 12.242 3.808 70.651 98.644
threads-sockets 1 group 99.990 6.667 99.999 100.000
threads-sockets 2 groups 99.940 5.114 99.997 100.000
threads-sockets 3 groups 99.469 4.115 99.977 100.000
threads-sockets 4 groups 87.528 4.038 99.400 100.000
threads-sockets 5 groups 6.942 2.398 59.244 88.337
threads-sockets 6 groups 4.359 1.954 49.448 87.860
threads-sockets 7 groups 2.845 1.345 41.198 77.102
threads-sockets 8 groups 2.871 1.404 38.512 74.312
schedstat_parse.py -f tbench_vanilla.log
case load se_eff% dom_eff% fast_rate% success_rate%
loopback 28 threads 99.976 18.369 99.995 100.000
loopback 56 threads 99.222 7.799 99.934 100.000
loopback 84 threads 19.723 6.819 70.215 100.000
loopback 112 threads 11.283 5.371 55.371 99.999
loopback 140 threads 0.000 0.000 0.000 0.000
loopback 168 threads 0.000 0.000 0.000 0.000
loopback 196 threads 0.000 0.000 0.000 0.000
loopback 224 threads 0.000 0.000 0.000 0.000
According to the test above, if the system becomes busy, the
SIS Search Efficiency(se_eff%) drops significantly. Although some
benchmarks would finally find an idle CPU(success_rate% = 100%), it is
doubtful whether it is worth it to search the whole LLC domain.
[Proposal]
It would be ideal to have a crystal ball to answer this question:
How many CPUs must a wakeup path walk down, before it can find an idle
CPU? Many potential metrics could be used to predict the number.
One candidate is the sum of util_avg in this LLC domain. The benefit
of choosing util_avg is that it is a metric of accumulated historic
activity, which seems to be smoother than instantaneous metrics
(such as rq->nr_running). Besides, choosing the sum of util_avg
would help predict the load of the LLC domain more precisely, because
SIS_PROP uses one CPU's idle time to estimate the total LLC domain idle
time.
In summary, the lower the util_avg is, the more select_idle_cpu()
should scan for idle CPU, and vice versa. When the sum of util_avg
in this LLC domain hits 85% or above, the scan stops. The reason to
choose 85% as the threshold is that this is the imbalance_pct(117)
when a LLC sched group is overloaded.
Introduce the quadratic function:
y = SCHED_CAPACITY_SCALE - p * x^2
and y'= y / SCHED_CAPACITY_SCALE
x is the ratio of sum_util compared to the CPU capacity:
x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
y' is the ratio of CPUs to be scanned in the LLC domain,
and the number of CPUs to scan is calculated by:
nr_scan = llc_weight * y'
Choosing quadratic function is because:
[1] Compared to the linear function, it scans more aggressively when the
sum_util is low.
[2] Compared to the exponential function, it is easier to calculate.
[3] It seems that there is no accurate mapping between the sum of util_avg
and the number of CPUs to be scanned. Use heuristic scan for now.
For a platform with 112 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 112 111 108 102 93 81 65 47 25 1 0 ...
For a platform with 16 CPUs per LLC, the number of CPUs to scan is:
sum_util% 0 5 15 25 35 45 55 65 75 85 86 ...
scan_nr 16 15 15 14 13 11 9 6 3 0 0 ...
Furthermore, to minimize the overhead of calculating the metrics in
select_idle_cpu(), borrow the statistics from periodic load balance.
As mentioned by Abel, on a platform with 112 CPUs per LLC, the
sum_util calculated by periodic load balance after 112 ms would
decay to about 0.5 * 0.5 * 0.5 * 0.7 = 8.75%, thus bringing a delay
in reflecting the latest utilization. But it is a trade-off.
Checking the util_avg in newidle load balance would be more frequent,
but it brings overhead - multiple CPUs write/read the per-LLC shared
variable and introduces cache contention. Tim also mentioned that,
it is allowed to be non-optimal in terms of scheduling for the
short-term variations, but if there is a long-term trend in the load
behavior, the scheduler can adjust for that.
When SIS_UTIL is enabled, the select_idle_cpu() uses the nr_scan
calculated by SIS_UTIL instead of the one from SIS_PROP. As Peter and
Mel suggested, SIS_UTIL should be enabled by default.
This patch is based on the util_avg, which is very sensitive to the
CPU frequency invariance. There is an issue that, when the max frequency
has been clamp, the util_avg would decay insanely fast when
the CPU is idle. Commit addca28512 ("cpufreq: intel_pstate: Handle no_turbo
in frequency invariance") could be used to mitigate this symptom, by adjusting
the arch_max_freq_ratio when turbo is disabled. But this issue is still
not thoroughly fixed, because the current code is unaware of the user-specified
max CPU frequency.
[Test result]
netperf and tbench were launched with 25% 50% 75% 100% 125% 150%
175% 200% of CPU number respectively. Hackbench and schbench were launched
by 1, 2 ,4, 8 groups. Each test lasts for 100 seconds and repeats 3 times.
The following is the benchmark result comparison between
baseline:vanilla v5.19-rc1 and compare:patched kernel. Positive compare%
indicates better performance.
Each netperf test is a:
netperf -4 -H 127.0.1 -t TCP/UDP_RR -c -C -l 100
netperf.throughput
=======
case load baseline(std%) compare%( std%)
TCP_RR 28 threads 1.00 ( 0.34) -0.16 ( 0.40)
TCP_RR 56 threads 1.00 ( 0.19) -0.02 ( 0.20)
TCP_RR 84 threads 1.00 ( 0.39) -0.47 ( 0.40)
TCP_RR 112 threads 1.00 ( 0.21) -0.66 ( 0.22)
TCP_RR 140 threads 1.00 ( 0.19) -0.69 ( 0.19)
TCP_RR 168 threads 1.00 ( 0.18) -0.48 ( 0.18)
TCP_RR 196 threads 1.00 ( 0.16) +194.70 ( 16.43)
TCP_RR 224 threads 1.00 ( 0.16) +197.30 ( 7.85)
UDP_RR 28 threads 1.00 ( 0.37) +0.35 ( 0.33)
UDP_RR 56 threads 1.00 ( 11.18) -0.32 ( 0.21)
UDP_RR 84 threads 1.00 ( 1.46) -0.98 ( 0.32)
UDP_RR 112 threads 1.00 ( 28.85) -2.48 ( 19.61)
UDP_RR 140 threads 1.00 ( 0.70) -0.71 ( 14.04)
UDP_RR 168 threads 1.00 ( 14.33) -0.26 ( 11.16)
UDP_RR 196 threads 1.00 ( 12.92) +186.92 ( 20.93)
UDP_RR 224 threads 1.00 ( 11.74) +196.79 ( 18.62)
Take the 224 threads as an example, the SIS search metrics changes are
illustrated below:
vanilla patched
4544492 +237.5% 15338634 sched_debug.cpu.sis_domain_search.avg
38539 +39686.8% 15333634 sched_debug.cpu.sis_failed.avg
128300000 -87.9% 15551326 sched_debug.cpu.sis_scanned.avg
5842896 +162.7% 15347978 sched_debug.cpu.sis_search.avg
There is -87.9% less CPU scans after patched, which indicates lower overhead.
Besides, with this patch applied, there is -13% less rq lock contention
in perf-profile.calltrace.cycles-pp._raw_spin_lock.raw_spin_rq_lock_nested
.try_to_wake_up.default_wake_function.woken_wake_function.
This might help explain the performance improvement - Because this patch allows
the waking task to remain on the previous CPU, rather than grabbing other CPUs'
lock.
Each hackbench test is a:
hackbench -g $job --process/threads --pipe/sockets -l 1000000 -s 100
hackbench.throughput
=========
case load baseline(std%) compare%( std%)
process-pipe 1 group 1.00 ( 1.29) +0.57 ( 0.47)
process-pipe 2 groups 1.00 ( 0.27) +0.77 ( 0.81)
process-pipe 4 groups 1.00 ( 0.26) +1.17 ( 0.02)
process-pipe 8 groups 1.00 ( 0.15) -4.79 ( 0.02)
process-sockets 1 group 1.00 ( 0.63) -0.92 ( 0.13)
process-sockets 2 groups 1.00 ( 0.03) -0.83 ( 0.14)
process-sockets 4 groups 1.00 ( 0.40) +5.20 ( 0.26)
process-sockets 8 groups 1.00 ( 0.04) +3.52 ( 0.03)
threads-pipe 1 group 1.00 ( 1.28) +0.07 ( 0.14)
threads-pipe 2 groups 1.00 ( 0.22) -0.49 ( 0.74)
threads-pipe 4 groups 1.00 ( 0.05) +1.88 ( 0.13)
threads-pipe 8 groups 1.00 ( 0.09) -4.90 ( 0.06)
threads-sockets 1 group 1.00 ( 0.25) -0.70 ( 0.53)
threads-sockets 2 groups 1.00 ( 0.10) -0.63 ( 0.26)
threads-sockets 4 groups 1.00 ( 0.19) +11.92 ( 0.24)
threads-sockets 8 groups 1.00 ( 0.08) +4.31 ( 0.11)
Each tbench test is a:
tbench -t 100 $job 127.0.0.1
tbench.throughput
======
case load baseline(std%) compare%( std%)
loopback 28 threads 1.00 ( 0.06) -0.14 ( 0.09)
loopback 56 threads 1.00 ( 0.03) -0.04 ( 0.17)
loopback 84 threads 1.00 ( 0.05) +0.36 ( 0.13)
loopback 112 threads 1.00 ( 0.03) +0.51 ( 0.03)
loopback 140 threads 1.00 ( 0.02) -1.67 ( 0.19)
loopback 168 threads 1.00 ( 0.38) +1.27 ( 0.27)
loopback 196 threads 1.00 ( 0.11) +1.34 ( 0.17)
loopback 224 threads 1.00 ( 0.11) +1.67 ( 0.22)
Each schbench test is a:
schbench -m $job -t 28 -r 100 -s 30000 -c 30000
schbench.latency_90%_us
========
case load baseline(std%) compare%( std%)
normal 1 mthread 1.00 ( 31.22) -7.36 ( 20.25)*
normal 2 mthreads 1.00 ( 2.45) -0.48 ( 1.79)
normal 4 mthreads 1.00 ( 1.69) +0.45 ( 0.64)
normal 8 mthreads 1.00 ( 5.47) +9.81 ( 14.28)
*Consider the Standard Deviation, this -7.36% regression might not be valid.
Also, a OLTP workload with a commercial RDBMS has been tested, and there
is no significant change.
There were concerns that unbalanced tasks among CPUs would cause problems.
For example, suppose the LLC domain is composed of 8 CPUs, and 7 tasks are
bound to CPU0~CPU6, while CPU7 is idle:
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5 CPU6 CPU7
util_avg 1024 1024 1024 1024 1024 1024 1024 0
Since the util_avg ratio is 87.5%( = 7/8 ), which is higher than 85%,
select_idle_cpu() will not scan, thus CPU7 is undetected during scan.
But according to Mel, it is unlikely the CPU7 will be idle all the time
because CPU7 could pull some tasks via CPU_NEWLY_IDLE.
lkp(kernel test robot) has reported a regression on stress-ng.sock on a
very busy system. According to the sched_debug statistics, it might be caused
by SIS_UTIL terminates the scan and chooses a previous CPU earlier, and this
might introduce more context switch, especially involuntary preemption, which
impacts a busy stress-ng. This regression has shown that, not all benchmarks
in every scenario benefit from idle CPU scan limit, and it needs further
investigation.
Besides, there is slight regression in hackbench's 16 groups case when the
LLC domain has 16 CPUs. Prateek mentioned that we should scan aggressively
in an LLC domain with 16 CPUs. Because the cost to search for an idle one
among 16 CPUs is negligible. The current patch aims to propose a generic
solution and only considers the util_avg. Something like the below could
be applied on top of the current patch to fulfill the requirement:
if (llc_weight <= 16)
nr_scan = nr_scan * 32 / llc_weight;
For LLC domain with 16 CPUs, the nr_scan will be expanded to 2 times large.
The smaller the CPU number this LLC domain has, the larger nr_scan will be
expanded. This needs further investigation.
There is also ongoing work[2] from Abel to filter out the busy CPUs during
wakeup, to further speed up the idle CPU scan. And it could be a following-up
optimization on top of this change.
Suggested-by: Tim Chen <tim.c.chen@intel.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Yicong Yang <yangyicong@hisilicon.com>
Tested-by: Mohini Narkhede <mohini.narkhede@intel.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://lore.kernel.org/r/20220612163428.849378-1-yu.c.chen@intel.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 64eaf50731 ]
Since commit 2312729688 ("sched/fair: Update scale invariance of PELT")
change to use rq_clock_pelt() instead of rq_clock_task(), we should also
use rq_clock_pelt() for throttled_clock_task_time and throttled_clock_task
accounting to get correct cfs_rq_clock_pelt() of throttled cfs_rq. And
rename throttled_clock_task(_time) to be clock_pelt rather than clock_task.
Fixes: 2312729688 ("sched/fair: Update scale invariance of PELT")
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20220408115309.81603-1-zhouchengming@bytedance.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 40f5aa4c5e ]
The warning in cfs_rq_is_decayed() triggered:
SCHED_WARN_ON(cfs_rq->avg.load_avg ||
cfs_rq->avg.util_avg ||
cfs_rq->avg.runnable_avg)
There exists a corner case in attach_entity_load_avg() which will
cause load_sum to be zero while load_avg will not be.
Consider se_weight is 88761 as per the sched_prio_to_weight[] table.
Further assume the get_pelt_divider() is 47742, this gives:
se->avg.load_avg is 1.
However, calculating load_sum:
se->avg.load_sum = div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
se->avg.load_sum = 1*47742/88761 = 0.
Then enqueue_load_avg() adds this to the cfs_rq totals:
cfs_rq->avg.load_avg += se->avg.load_avg;
cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
Resulting in load_avg being 1 with load_sum is 0, which will trigger
the WARN.
Fixes: f207934fb7 ("sched/fair: Align PELT windows between cfs_rq and its se")
Signed-off-by: kuyo chang <kuyo.chang@mediatek.com>
[peterz: massage changelog]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lkml.kernel.org/r/20220414090229.342-1-kuyo.chang@mediatek.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 2cfb7a1b03 ]
There are inconsistencies when determining if a NUMA imbalance is allowed
that should be corrected.
o allow_numa_imbalance changes types and is not always examining
the destination group so both the type should be corrected as
well as the naming.
o find_idlest_group uses the sched_domain's weight instead of the
group weight which is different to find_busiest_group
o find_busiest_group uses the source group instead of the destination
which is different to task_numa_find_cpu
o Both find_idlest_group and find_busiest_group should account
for the number of running tasks if a move was allowed to be
consistent with task_numa_find_cpu
Fixes: 7d2b5dd0bc ("sched/numa: Allow a floating imbalance between NUMA nodes")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Gautham R. Shenoy <gautham.shenoy@amd.com>
Link: https://lore.kernel.org/r/20220208094334.16379-2-mgorman@techsingularity.net
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 98b0d89022 ]
Rick reported performance regressions in bugzilla because of cpu frequency
being lower than before:
https://bugzilla.kernel.org/show_bug.cgi?id=215045
He bisected the problem to:
commit 1c35b07e6d ("sched/fair: Ensure _sum and _avg values stay consistent")
This commit forces util_sum to be synced with the new util_avg after
removing the contribution of a task and before the next periodic sync. By
doing so util_sum is rounded to its lower bound and might lost up to
LOAD_AVG_MAX-1 of accumulated contribution which has not yet been
reflected in util_avg.
Instead of always setting util_sum to the low bound of util_avg, which can
significantly lower the utilization of root cfs_rq after propagating the
change down into the hierarchy, we revert the change of util_sum and
propagate the difference.
In addition, we also check that cfs's util_sum always stays above the
lower bound for a given util_avg as it has been observed that
sched_entity's util_sum is sometimes above cfs one.
Fixes: 1c35b07e6d ("sched/fair: Ensure _sum and _avg values stay consistent")
Reported-by: Rick Yiu <rickyiu@google.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Tested-by: Sachin Sant <sachinp@linux.ibm.com>
Link: https://lkml.kernel.org/r/20220111134659.24961-2-vincent.guittot@linaro.org
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 014ba44e81 ]
select_idle_sibling() has a special case for tasks woken up by a per-CPU
kthread where the selected CPU is the previous one. For asymmetric CPU
capacity systems, the assumption was that the wakee couldn't have a
bigger utilization during task placement than it used to have during the
last activation. That was not considering uclamp.min which can completely
change between two task activations and as a consequence mandates the
fitness criterion asym_fits_capacity(), even for the exit path described
above.
Fixes: b4c9c9f156 ("sched/fair: Prefer prev cpu in asymmetric wakeup path")
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lkml.kernel.org/r/20211129173115.4006346-1-vincent.donnefort@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8b4e74ccb5 ]
select_idle_sibling() has a special case for tasks woken up by a per-CPU
kthread, where the selected CPU is the previous one. However, the current
condition for this exit path is incomplete. A task can wake up from an
interrupt context (e.g. hrtimer), while a per-CPU kthread is running. A
such scenario would spuriously trigger the special case described above.
Also, a recent change made the idle task like a regular per-CPU kthread,
hence making that situation more likely to happen
(is_per_cpu_kthread(swapper) being true now).
Checking for task context makes sure select_idle_sibling() will not
interpret a wake up from any other context as a wake up by a per-CPU
kthread.
Fixes: 52262ee567 ("sched/fair: Allow a per-CPU kthread waking a task to stack on the same CPU, to fix XFS performance regression")
Signed-off-by: Vincent Donnefort <vincent.donnefort@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Link: https://lore.kernel.org/r/20211201143450.479472-1-vincent.donnefort@arm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit b027789e5e ]
Kevin is reporting crashes which point to a use-after-free of a cfs_rq
in update_blocked_averages(). Initial debugging revealed that we've
live cfs_rq's (on_list=1) in an about to be kfree()'d task group in
free_fair_sched_group(). However, it was unclear how that can happen.
His kernel config happened to lead to a layout of struct sched_entity
that put the 'my_q' member directly into the middle of the object
which makes it incidentally overlap with SLUB's freelist pointer.
That, in combination with SLAB_FREELIST_HARDENED's freelist pointer
mangling, leads to a reliable access violation in form of a #GP which
made the UAF fail fast.
Michal seems to have run into the same issue[1]. He already correctly
diagnosed that commit a7b359fc6a ("sched/fair: Correctly insert
cfs_rq's to list on unthrottle") is causing the preconditions for the
UAF to happen by re-adding cfs_rq's also to task groups that have no
more running tasks, i.e. also to dead ones. His analysis, however,
misses the real root cause and it cannot be seen from the crash
backtrace only, as the real offender is tg_unthrottle_up() getting
called via sched_cfs_period_timer() via the timer interrupt at an
inconvenient time.
When unregister_fair_sched_group() unlinks all cfs_rq's from the dying
task group, it doesn't protect itself from getting interrupted. If the
timer interrupt triggers while we iterate over all CPUs or after
unregister_fair_sched_group() has finished but prior to unlinking the
task group, sched_cfs_period_timer() will execute and walk the list of
task groups, trying to unthrottle cfs_rq's, i.e. re-add them to the
dying task group. These will later -- in free_fair_sched_group() -- be
kfree()'ed while still being linked, leading to the fireworks Kevin
and Michal are seeing.
To fix this race, ensure the dying task group gets unlinked first.
However, simply switching the order of unregistering and unlinking the
task group isn't sufficient, as concurrent RCU walkers might still see
it, as can be seen below:
CPU1: CPU2:
: timer IRQ:
: do_sched_cfs_period_timer():
: :
: distribute_cfs_runtime():
: rcu_read_lock();
: :
: unthrottle_cfs_rq():
sched_offline_group(): :
: walk_tg_tree_from(…,tg_unthrottle_up,…):
list_del_rcu(&tg->list); :
(1) : list_for_each_entry_rcu(child, &parent->children, siblings)
: :
(2) list_del_rcu(&tg->siblings); :
: tg_unthrottle_up():
unregister_fair_sched_group(): struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
: :
list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); :
: :
: if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running)
(3) : list_add_leaf_cfs_rq(cfs_rq);
: :
: :
: :
: :
: :
(4) : rcu_read_unlock();
CPU 2 walks the task group list in parallel to sched_offline_group(),
specifically, it'll read the soon to be unlinked task group entry at
(1). Unlinking it on CPU 1 at (2) therefore won't prevent CPU 2 from
still passing it on to tg_unthrottle_up(). CPU 1 now tries to unlink
all cfs_rq's via list_del_leaf_cfs_rq() in
unregister_fair_sched_group(). Meanwhile CPU 2 will re-add some of
these at (3), which is the cause of the UAF later on.
To prevent this additional race from happening, we need to wait until
walk_tg_tree_from() has finished traversing the task groups, i.e.
after the RCU read critical section ends in (4). Afterwards we're safe
to call unregister_fair_sched_group(), as each new walk won't see the
dying task group any more.
On top of that, we need to wait yet another RCU grace period after
unregister_fair_sched_group() to ensure print_cfs_stats(), which might
run concurrently, always sees valid objects, i.e. not already free'd
ones.
This patch survives Michal's reproducer[2] for 8h+ now, which used to
trigger within minutes before.
[1] https://lore.kernel.org/lkml/20211011172236.11223-1-mkoutny@suse.com/
[2] https://lore.kernel.org/lkml/20211102160228.GA57072@blackbody.suse.cz/
Fixes: a7b359fc6a ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
[peterz: shuffle code around a bit]
Reported-by: Kevin Tanguy <kevin.tanguy@corp.ovh.com>
Signed-off-by: Mathias Krause <minipli@grsecurity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
Since commit a7b359fc6a ("sched/fair: Correctly insert cfs_rq's to
list on unthrottle") we add cfs_rqs with no runnable tasks but not fully
decayed into the load (leaf) list. We may ignore adding some ancestors
and therefore breaking tmp_alone_branch invariant. This broke LTP test
cfs_bandwidth01 and it was partially fixed in commit fdaba61ef8
("sched/fair: Ensure that the CFS parent is added after unthrottling").
I noticed the named test still fails even with the fix (but with low
probability, 1 in ~1000 executions of the test). The reason is when
bailing out of unthrottle_cfs_rq early, we may miss adding ancestors of
the unthrottled cfs_rq, thus, not joining tmp_alone_branch properly.
Fix this by adding ancestors if we notice the unthrottled cfs_rq was
added to the load list.
Fixes: a7b359fc6a ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Odin Ugedal <odin@uged.al>
Link: https://lore.kernel.org/r/20210917153037.11176-1-mkoutny@suse.com
- The biggest change in this cycle is scheduler support for asymmetric
scheduling affinity, to support the execution of legacy 32-bit tasks on
AArch32 systems that also have 64-bit-only CPUs.
Architectures can fill in this functionality by defining their
own task_cpu_possible_mask(p). When this is done, the scheduler will
make sure the task will only be scheduled on CPUs that support it.
(The actual arm64 specific changes are not part of this tree.)
For other architectures there will be no change in functionality.
- Add cgroup SCHED_IDLE support
- Increase node-distance flexibility & delay determining it until a CPU
is brought online. (This enables platforms where node distance isn't
final until the CPU is only.)
- Deadline scheduler enhancements & fixes
- Misc fixes & cleanups.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
-----BEGIN PGP SIGNATURE-----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=VdGE
-----END PGP SIGNATURE-----
Merge tag 'sched-core-2021-08-30' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull scheduler updates from Ingo Molnar:
- The biggest change in this cycle is scheduler support for asymmetric
scheduling affinity, to support the execution of legacy 32-bit tasks
on AArch32 systems that also have 64-bit-only CPUs.
Architectures can fill in this functionality by defining their own
task_cpu_possible_mask(p). When this is done, the scheduler will make
sure the task will only be scheduled on CPUs that support it.
(The actual arm64 specific changes are not part of this tree.)
For other architectures there will be no change in functionality.
- Add cgroup SCHED_IDLE support
- Increase node-distance flexibility & delay determining it until a CPU
is brought online. (This enables platforms where node distance isn't
final until the CPU is only.)
- Deadline scheduler enhancements & fixes
- Misc fixes & cleanups.
* tag 'sched-core-2021-08-30' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (27 commits)
eventfd: Make signal recursion protection a task bit
sched/fair: Mark tg_is_idle() an inline in the !CONFIG_FAIR_GROUP_SCHED case
sched: Introduce dl_task_check_affinity() to check proposed affinity
sched: Allow task CPU affinity to be restricted on asymmetric systems
sched: Split the guts of sched_setaffinity() into a helper function
sched: Introduce task_struct::user_cpus_ptr to track requested affinity
sched: Reject CPU affinity changes based on task_cpu_possible_mask()
cpuset: Cleanup cpuset_cpus_allowed_fallback() use in select_fallback_rq()
cpuset: Honour task_cpu_possible_mask() in guarantee_online_cpus()
cpuset: Don't use the cpu_possible_mask as a last resort for cgroup v1
sched: Introduce task_cpu_possible_mask() to limit fallback rq selection
sched: Cgroup SCHED_IDLE support
sched/topology: Skip updating masks for non-online nodes
sched: Replace deprecated CPU-hotplug functions.
sched: Skip priority checks with SCHED_FLAG_KEEP_PARAMS
sched: Fix UCLAMP_FLAG_IDLE setting
sched/deadline: Fix missing clock update in migrate_task_rq_dl()
sched/fair: Avoid a second scan of target in select_idle_cpu
sched/fair: Use prev instead of new target as recent_used_cpu
sched: Don't report SCHED_FLAG_SUGOV in sched_getattr()
...
It's not actually used in the !CONFIG_FAIR_GROUP_SCHED case:
kernel/sched/fair.c:488:12: warning: ‘tg_is_idle’ defined but not used [-Wunused-function]
Keep around a placeholder nevertheless, for API completeness. Mark it inline,
so the compiler doesn't think it must be used.
Fixes: 304000390f88: ("sched: Cgroup SCHED_IDLE support")
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Josh Don <joshdon@google.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
This extends SCHED_IDLE to cgroups.
Interface: cgroup/cpu.idle.
0: default behavior
1: SCHED_IDLE
Extending SCHED_IDLE to cgroups means that we incorporate the existing
aspects of SCHED_IDLE; a SCHED_IDLE cgroup will count all of its
descendant threads towards the idle_h_nr_running count of all of its
ancestor cgroups. Thus, sched_idle_rq() will work properly.
Additionally, SCHED_IDLE cgroups are configured with minimum weight.
There are two key differences between the per-task and per-cgroup
SCHED_IDLE interface:
- The cgroup interface allows tasks within a SCHED_IDLE hierarchy to
maintain their relative weights. The entity that is "idle" is the
cgroup, not the tasks themselves.
- Since the idle entity is the cgroup, our SCHED_IDLE wakeup preemption
decision is not made by comparing the current task with the woken
task, but rather by comparing their matching sched_entity.
A typical use-case for this is a user that creates an idle and a
non-idle subtree. The non-idle subtree will dominate competition vs
the idle subtree, but the idle subtree will still be high priority vs
other users on the system. The latter is accomplished via comparing
matching sched_entity in the waken preemption path (this could also be
improved by making the sched_idle_rq() decision dependent on the
perspective of a specific task).
For now, we maintain the existing SCHED_IDLE semantics. Future patches
may make improvements that extend how we treat SCHED_IDLE entities.
The per-task_group idle field is an integer that currently only holds
either a 0 or a 1. This is explicitly typed as an integer to allow for
further extensions to this API. For example, a negative value may
indicate a highly latency-sensitive cgroup that should be preferred
for preemption/placement/etc.
Signed-off-by: Josh Don <joshdon@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20210730020019.1487127-2-joshdon@google.com
When select_idle_cpu starts scanning for an idle CPU, it starts with
a target CPU that has already been checked by select_idle_sibling.
This patch starts with the next CPU instead.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210804115857.6253-3-mgorman@techsingularity.net
After select_idle_sibling, p->recent_used_cpu is set to the
new target. However on the next wakeup, prev will be the same as
recent_used_cpu unless the load balancer has moved the task since the
last wakeup. It still works, but is less efficient than it could be.
This patch preserves recent_used_cpu for longer.
The impact on SIS efficiency is tiny so the SIS statistic patches were
used to track the hit rate for using recent_used_cpu. With perf bench
pipe on a 2-socket Cascadelake machine, the hit rate went from 57.14%
to 85.32%. For more intensive wakeup loads like hackbench, the hit rate
is almost negligible but rose from 0.21% to 6.64%. For scaling loads
like tbench, the hit rate goes from almost 0% to 25.42% overall. Broadly
speaking, on tbench, the success rate is much higher for lower thread
counts and drops to almost 0 as the workload scales to towards saturation.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210804115857.6253-2-mgorman@techsingularity.net
Use the loop variable instead of the function argument to test the
other SMT siblings for idle.
Fixes: ff7db0bf24 ("sched/numa: Prefer using an idle CPU as a migration target instead of comparing tasks")
Signed-off-by: Mika Penttilä <mika.penttila@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Pankaj Gupta <pankaj.gupta@ionos.com>
Link: https://lkml.kernel.org/r/20210722063946.28951-1-mika.penttila@gmail.com
The time remaining until expiry of the refresh_timer can be negative.
Casting the type to an unsigned 64-bit value will cause integer
underflow, making the runtime_refresh_within return false instead of
true. These situations are rare, but they do happen.
This does not cause user-facing issues or errors; other than
possibly unthrottling cfs_rq's using runtime from the previous period(s),
making the CFS bandwidth enforcement less strict in those (special)
situations.
Signed-off-by: Odin Ugedal <odin@uged.al>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Link: https://lore.kernel.org/r/20210629121452.18429-1-odin@uged.al
commit 9e077b52d8 ("sched/pelt: Check that *_avg are null when *_sum are")
reported some inconsitencies between *_avg and *_sum.
commit 1c35b07e6d ("sched/fair: Ensure _sum and _avg values stay consistent")
fixed some but one remains when dequeuing load.
sync the cfs's load_sum with its load_avg after dequeuing the load of a
sched_entity.
Fixes: 9e077b52d8 ("sched/pelt: Check that *_avg are null when *_sum are")
Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Odin Ugedal <odin@uged.al>
Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Link: https://lore.kernel.org/r/20210701171837.32156-1-vincent.guittot@linaro.org
new warning that several people reported.
- Flip CONFIG_SCHED_CORE to default-disabled, and update the
Kconfig help text.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
-----BEGIN PGP SIGNATURE-----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=Jgwc
-----END PGP SIGNATURE-----
Merge tag 'sched-urgent-2021-06-30' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull scheduler fixes from Ingo Molnar:
- Fix a small inconsistency (bug) in load tracking, caught by a new
warning that several people reported.
- Flip CONFIG_SCHED_CORE to default-disabled, and update the Kconfig
help text.
* tag 'sched-urgent-2021-06-30' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/core: Disable CONFIG_SCHED_CORE by default
sched/fair: Ensure _sum and _avg values stay consistent
- Changes to core scheduling facilities:
- Add "Core Scheduling" via CONFIG_SCHED_CORE=y, which enables
coordinated scheduling across SMT siblings. This is a much
requested feature for cloud computing platforms, to allow
the flexible utilization of SMT siblings, without exposing
untrusted domains to information leaks & side channels, plus
to ensure more deterministic computing performance on SMT
systems used by heterogenous workloads.
There's new prctls to set core scheduling groups, which
allows more flexible management of workloads that can share
siblings.
- Fix task->state access anti-patterns that may result in missed
wakeups and rename it to ->__state in the process to catch new
abuses.
- Load-balancing changes:
- Tweak newidle_balance for fair-sched, to improve
'memcache'-like workloads.
- "Age" (decay) average idle time, to better track & improve workloads
such as 'tbench'.
- Fix & improve energy-aware (EAS) balancing logic & metrics.
- Fix & improve the uclamp metrics.
- Fix task migration (taskset) corner case on !CONFIG_CPUSET.
- Fix RT and deadline utilization tracking across policy changes
- Introduce a "burstable" CFS controller via cgroups, which allows
bursty CPU-bound workloads to borrow a bit against their future
quota to improve overall latencies & batching. Can be tweaked
via /sys/fs/cgroup/cpu/<X>/cpu.cfs_burst_us.
- Rework assymetric topology/capacity detection & handling.
- Scheduler statistics & tooling:
- Disable delayacct by default, but add a sysctl to enable
it at runtime if tooling needs it. Use static keys and
other optimizations to make it more palatable.
- Use sched_clock() in delayacct, instead of ktime_get_ns().
- Misc cleanups and fixes.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
-----BEGIN PGP SIGNATURE-----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=3VDr
-----END PGP SIGNATURE-----
Merge tag 'sched-core-2021-06-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull scheduler udpates from Ingo Molnar:
- Changes to core scheduling facilities:
- Add "Core Scheduling" via CONFIG_SCHED_CORE=y, which enables
coordinated scheduling across SMT siblings. This is a much
requested feature for cloud computing platforms, to allow the
flexible utilization of SMT siblings, without exposing untrusted
domains to information leaks & side channels, plus to ensure more
deterministic computing performance on SMT systems used by
heterogenous workloads.
There are new prctls to set core scheduling groups, which allows
more flexible management of workloads that can share siblings.
- Fix task->state access anti-patterns that may result in missed
wakeups and rename it to ->__state in the process to catch new
abuses.
- Load-balancing changes:
- Tweak newidle_balance for fair-sched, to improve 'memcache'-like
workloads.
- "Age" (decay) average idle time, to better track & improve
workloads such as 'tbench'.
- Fix & improve energy-aware (EAS) balancing logic & metrics.
- Fix & improve the uclamp metrics.
- Fix task migration (taskset) corner case on !CONFIG_CPUSET.
- Fix RT and deadline utilization tracking across policy changes
- Introduce a "burstable" CFS controller via cgroups, which allows
bursty CPU-bound workloads to borrow a bit against their future
quota to improve overall latencies & batching. Can be tweaked via
/sys/fs/cgroup/cpu/<X>/cpu.cfs_burst_us.
- Rework assymetric topology/capacity detection & handling.
- Scheduler statistics & tooling:
- Disable delayacct by default, but add a sysctl to enable it at
runtime if tooling needs it. Use static keys and other
optimizations to make it more palatable.
- Use sched_clock() in delayacct, instead of ktime_get_ns().
- Misc cleanups and fixes.
* tag 'sched-core-2021-06-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (72 commits)
sched/doc: Update the CPU capacity asymmetry bits
sched/topology: Rework CPU capacity asymmetry detection
sched/core: Introduce SD_ASYM_CPUCAPACITY_FULL sched_domain flag
psi: Fix race between psi_trigger_create/destroy
sched/fair: Introduce the burstable CFS controller
sched/uclamp: Fix uclamp_tg_restrict()
sched/rt: Fix Deadline utilization tracking during policy change
sched/rt: Fix RT utilization tracking during policy change
sched: Change task_struct::state
sched,arch: Remove unused TASK_STATE offsets
sched,timer: Use __set_current_state()
sched: Add get_current_state()
sched,perf,kvm: Fix preemption condition
sched: Introduce task_is_running()
sched: Unbreak wakeups
sched/fair: Age the average idle time
sched/cpufreq: Consider reduced CPU capacity in energy calculation
sched/fair: Take thermal pressure into account while estimating energy
thermal/cpufreq_cooling: Update offline CPUs per-cpu thermal_pressure
sched/fair: Return early from update_tg_cfs_load() if delta == 0
...
On a 128 cores AMD machine, there are 8 cores in nohz_full mode, and
the others are used for housekeeping. When many housekeeping cpus are
in idle state, we can observe huge time burn in the loop for searching
nearest busy housekeeper cpu by ftrace.
9) | get_nohz_timer_target() {
9) | housekeeping_test_cpu() {
9) 0.390 us | housekeeping_get_mask.part.1();
9) 0.561 us | }
9) 0.090 us | __rcu_read_lock();
9) 0.090 us | housekeeping_cpumask();
9) 0.521 us | housekeeping_cpumask();
9) 0.140 us | housekeeping_cpumask();
...
9) 0.500 us | housekeeping_cpumask();
9) | housekeeping_any_cpu() {
9) 0.090 us | housekeeping_get_mask.part.1();
9) 0.100 us | sched_numa_find_closest();
9) 0.491 us | }
9) 0.100 us | __rcu_read_unlock();
9) + 76.163 us | }
for_each_cpu_and() is a micro function, so in get_nohz_timer_target()
function the
for_each_cpu_and(i, sched_domain_span(sd),
housekeeping_cpumask(HK_FLAG_TIMER))
equals to below:
for (i = -1; i = cpumask_next_and(i, sched_domain_span(sd),
housekeeping_cpumask(HK_FLAG_TIMER)), i < nr_cpu_ids;)
That will cause that housekeeping_cpumask() will be invoked many times.
The housekeeping_cpumask() function returns a const value, so it is
unnecessary to invoke it every time. This patch can minimize the worst
searching time from ~76us to ~16us in my testing.
Similarly, the find_new_ilb() function has the same problem.
Co-developed-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Yuan ZhaoXiong <yuanzhaoxiong@baidu.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/1622985115-51007-1-git-send-email-yuanzhaoxiong@baidu.com
The _sum and _avg values are in general sync together with the PELT
divider. They are however not always completely in perfect sync,
resulting in situations where _sum gets to zero while _avg stays
positive. Such situations are undesirable.
This comes from the fact that PELT will increase period_contrib, also
increasing the PELT divider, without updating _sum and _avg values to
stay in perfect sync where (_sum == _avg * divider). However, such PELT
change will never lower _sum, making it impossible to end up in a
situation where _sum is zero and _avg is not.
Therefore, we need to ensure that when subtracting load outside PELT,
that when _sum is zero, _avg is also set to zero. This occurs when
(_sum < _avg * divider), and the subtracted (_avg * divider) is bigger
or equal to the current _sum, while the subtracted _avg is smaller than
the current _avg.
Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org>
Signed-off-by: Odin Ugedal <odin@uged.al>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Link: https://lore.kernel.org/r/20210624111815.57937-1-odin@uged.al
The CFS bandwidth controller limits CPU requests of a task group to
quota during each period. However, parallel workloads might be bursty
so that they get throttled even when their average utilization is under
quota. And they are latency sensitive at the same time so that
throttling them is undesired.
We borrow time now against our future underrun, at the cost of increased
interference against the other system users. All nicely bounded.
Traditional (UP-EDF) bandwidth control is something like:
(U = \Sum u_i) <= 1
This guaranteeds both that every deadline is met and that the system is
stable. After all, if U were > 1, then for every second of walltime,
we'd have to run more than a second of program time, and obviously miss
our deadline, but the next deadline will be further out still, there is
never time to catch up, unbounded fail.
This work observes that a workload doesn't always executes the full
quota; this enables one to describe u_i as a statistical distribution.
For example, have u_i = {x,e}_i, where x is the p(95) and x+e p(100)
(the traditional WCET). This effectively allows u to be smaller,
increasing the efficiency (we can pack more tasks in the system), but at
the cost of missing deadlines when all the odds line up. However, it
does maintain stability, since every overrun must be paired with an
underrun as long as our x is above the average.
That is, suppose we have 2 tasks, both specify a p(95) value, then we
have a p(95)*p(95) = 90.25% chance both tasks are within their quota and
everything is good. At the same time we have a p(5)p(5) = 0.25% chance
both tasks will exceed their quota at the same time (guaranteed deadline
fail). Somewhere in between there's a threshold where one exceeds and
the other doesn't underrun enough to compensate; this depends on the
specific CDFs.
At the same time, we can say that the worst case deadline miss, will be
\Sum e_i; that is, there is a bounded tardiness (under the assumption
that x+e is indeed WCET).
The benefit of burst is seen when testing with schbench. Default value of
kernel.sched_cfs_bandwidth_slice_us(5ms) and CONFIG_HZ(1000) is used.
mkdir /sys/fs/cgroup/cpu/test
echo $$ > /sys/fs/cgroup/cpu/test/cgroup.procs
echo 100000 > /sys/fs/cgroup/cpu/test/cpu.cfs_quota_us
echo 100000 > /sys/fs/cgroup/cpu/test/cpu.cfs_burst_us
./schbench -m 1 -t 3 -r 20 -c 80000 -R 10
The average CPU usage is at 80%. I run this for 10 times, and got long tail
latency for 6 times and got throttled for 8 times.
Tail latencies are shown below, and it wasn't the worst case.
Latency percentiles (usec)
50.0000th: 19872
75.0000th: 21344
90.0000th: 22176
95.0000th: 22496
*99.0000th: 22752
99.5000th: 22752
99.9000th: 22752
min=0, max=22727
rps: 9.90 p95 (usec) 22496 p99 (usec) 22752 p95/cputime 28.12% p99/cputime 28.44%
The interferenece when using burst is valued by the possibilities for
missing the deadline and the average WCET. Test results showed that when
there many cgroups or CPU is under utilized, the interference is
limited. More details are shown in:
https://lore.kernel.org/lkml/5371BD36-55AE-4F71-B9D7-B86DC32E3D2B@linux.alibaba.com/
Co-developed-by: Shanpei Chen <shanpeic@linux.alibaba.com>
Signed-off-by: Shanpei Chen <shanpeic@linux.alibaba.com>
Co-developed-by: Tianchen Ding <dtcccc@linux.alibaba.com>
Signed-off-by: Tianchen Ding <dtcccc@linux.alibaba.com>
Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Link: https://lore.kernel.org/r/20210621092800.23714-2-changhuaixin@linux.alibaba.com
Ensure that a CFS parent will be in the list whenever one of its children is also
in the list.
A warning on rq->tmp_alone_branch != &rq->leaf_cfs_rq_list has been
reported while running LTP test cfs_bandwidth01.
Odin Ugedal found the root cause:
$ tree /sys/fs/cgroup/ltp/ -d --charset=ascii
/sys/fs/cgroup/ltp/
|-- drain
`-- test-6851
`-- level2
|-- level3a
| |-- worker1
| `-- worker2
`-- level3b
`-- worker3
Timeline (ish):
- worker3 gets throttled
- level3b is decayed, since it has no more load
- level2 get throttled
- worker3 get unthrottled
- level2 get unthrottled
- worker3 is added to list
- level3b is not added to list, since nr_running==0 and is decayed
[ Vincent Guittot: Rebased and updated to fix for the reported warning. ]
Fixes: a7b359fc6a ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
Reported-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Rik van Riel <riel@surriel.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Sachin Sant <sachinp@linux.vnet.ibm.com>
Acked-by: Odin Ugedal <odin@uged.al>
Link: https://lore.kernel.org/r/20210621174330.11258-1-vincent.guittot@linaro.org
Change the type and name of task_struct::state. Drop the volatile and
shrink it to an 'unsigned int'. Rename it in order to find all uses
such that we can use READ_ONCE/WRITE_ONCE as appropriate.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Acked-by: Will Deacon <will@kernel.org>
Acked-by: Daniel Thompson <daniel.thompson@linaro.org>
Link: https://lore.kernel.org/r/20210611082838.550736351@infradead.org
This commit in sched/urgent moved the cfs_rq_is_decayed() function:
a7b359fc6a37: ("sched/fair: Correctly insert cfs_rq's to list on unthrottle")
and this fresh commit in sched/core modified it in the old location:
9e077b52d86a: ("sched/pelt: Check that *_avg are null when *_sum are")
Merge the two variants.
Conflicts:
kernel/sched/fair.c
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This is a partial forward-port of Peter Ziljstra's work first posted
at:
https://lore.kernel.org/lkml/20180530142236.667774973@infradead.org/
Currently select_idle_cpu()'s proportional scheme uses the average idle
time *for when we are idle*, that is temporally challenged. When a CPU
is not at all idle, we'll happily continue using whatever value we did
see when the CPU goes idle. To fix this, introduce a separate average
idle and age it (the existing value still makes sense for things like
new-idle balancing, which happens when we do go idle).
The overall goal is to not spend more time scanning for idle CPUs than
we're idle for. Otherwise we're inhibiting work. This means that we need to
consider the cost over all the wake-ups between consecutive idle periods.
To track this, the scan cost is subtracted from the estimated average
idle time.
The impact of this patch is related to workloads that have domains that
are fully busy or overloaded. Without the patch, the scan depth may be
too high because a CPU is not reaching idle.
Due to the nature of the patch, this is a regression magnet. It
potentially wins when domains are almost fully busy or overloaded --
at that point searches are likely to fail but idle is not being aged
as CPUs are active so search depth is too large and useless. It will
potentially show regressions when there are idle CPUs and a deep search is
beneficial. This tbench result on a 2-socket broadwell machine partially
illustates the problem
5.13.0-rc2 5.13.0-rc2
vanilla sched-avgidle-v1r5
Hmean 1 445.02 ( 0.00%) 451.36 * 1.42%*
Hmean 2 830.69 ( 0.00%) 846.03 * 1.85%*
Hmean 4 1350.80 ( 0.00%) 1505.56 * 11.46%*
Hmean 8 2888.88 ( 0.00%) 2586.40 * -10.47%*
Hmean 16 5248.18 ( 0.00%) 5305.26 * 1.09%*
Hmean 32 8914.03 ( 0.00%) 9191.35 * 3.11%*
Hmean 64 10663.10 ( 0.00%) 10192.65 * -4.41%*
Hmean 128 18043.89 ( 0.00%) 18478.92 * 2.41%*
Hmean 256 16530.89 ( 0.00%) 17637.16 * 6.69%*
Hmean 320 16451.13 ( 0.00%) 17270.97 * 4.98%*
Note that 8 was a regression point where a deeper search would have helped
but it gains for high thread counts when searches are useless. Hackbench
is a more extreme example although not perfect as the tasks idle rapidly
hackbench-process-pipes
5.13.0-rc2 5.13.0-rc2
vanilla sched-avgidle-v1r5
Amean 1 0.3950 ( 0.00%) 0.3887 ( 1.60%)
Amean 4 0.9450 ( 0.00%) 0.9677 ( -2.40%)
Amean 7 1.4737 ( 0.00%) 1.4890 ( -1.04%)
Amean 12 2.3507 ( 0.00%) 2.3360 * 0.62%*
Amean 21 4.0807 ( 0.00%) 4.0993 * -0.46%*
Amean 30 5.6820 ( 0.00%) 5.7510 * -1.21%*
Amean 48 8.7913 ( 0.00%) 8.7383 ( 0.60%)
Amean 79 14.3880 ( 0.00%) 13.9343 * 3.15%*
Amean 110 21.2233 ( 0.00%) 19.4263 * 8.47%*
Amean 141 28.2930 ( 0.00%) 25.1003 * 11.28%*
Amean 172 34.7570 ( 0.00%) 30.7527 * 11.52%*
Amean 203 41.0083 ( 0.00%) 36.4267 * 11.17%*
Amean 234 47.7133 ( 0.00%) 42.0623 * 11.84%*
Amean 265 53.0353 ( 0.00%) 47.7720 * 9.92%*
Amean 296 60.0170 ( 0.00%) 53.4273 * 10.98%*
Stddev 1 0.0052 ( 0.00%) 0.0025 ( 51.57%)
Stddev 4 0.0357 ( 0.00%) 0.0370 ( -3.75%)
Stddev 7 0.0190 ( 0.00%) 0.0298 ( -56.64%)
Stddev 12 0.0064 ( 0.00%) 0.0095 ( -48.38%)
Stddev 21 0.0065 ( 0.00%) 0.0097 ( -49.28%)
Stddev 30 0.0185 ( 0.00%) 0.0295 ( -59.54%)
Stddev 48 0.0559 ( 0.00%) 0.0168 ( 69.92%)
Stddev 79 0.1559 ( 0.00%) 0.0278 ( 82.17%)
Stddev 110 1.1728 ( 0.00%) 0.0532 ( 95.47%)
Stddev 141 0.7867 ( 0.00%) 0.0968 ( 87.69%)
Stddev 172 1.0255 ( 0.00%) 0.0420 ( 95.91%)
Stddev 203 0.8106 ( 0.00%) 0.1384 ( 82.92%)
Stddev 234 1.1949 ( 0.00%) 0.1328 ( 88.89%)
Stddev 265 0.9231 ( 0.00%) 0.0820 ( 91.11%)
Stddev 296 1.0456 ( 0.00%) 0.1327 ( 87.31%)
Again, higher thread counts benefit and the standard deviation
shows that results are also a lot more stable when the idle
time is aged.
The patch potentially matters when a socket was multiple LLCs as the
maximum search depth is lower. However, some of the test results were
suspiciously good (e.g. specjbb2005 gaining 50% on a Zen1 machine) and
other results were not dramatically different to other mcahines.
Given the nature of the patch, Peter's full series is not being forward
ported as each part should stand on its own. Preferably they would be
merged at different times to reduce the risk of false bisections.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210615111611.GH30378@techsingularity.net
Energy Aware Scheduling (EAS) needs to predict the decisions made by
SchedUtil. The map_util_freq() exists to do that.
There are corner cases where the max allowed frequency might be reduced
(due to thermal). SchedUtil as a CPUFreq governor, is aware of that
but EAS is not. This patch aims to address it.
SchedUtil stores the maximum allowed frequency in
'sugov_policy::next_freq' field. EAS has to predict that value, which is
the real used frequency. That value is made after a call to
cpufreq_driver_resolve_freq() which clamps to the CPUFreq policy limits.
In the existing code EAS is not able to predict that real frequency.
This leads to energy estimation errors.
To avoid wrong energy estimation in EAS (due to frequency miss prediction)
make sure that the step which calculates Performance Domain frequency,
is also aware of the allowed CPU capacity.
Furthermore, modify map_util_freq() to not extend the frequency value.
Instead, use map_util_perf() to extend the util value in both places:
SchedUtil and EAS, but for EAS clamp it to max allowed CPU capacity.
In the end, we achieve the same desirable behavior for both subsystems
and alignment in regards to the real CPU frequency.
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> (For the schedutil part)
Link: https://lore.kernel.org/r/20210614191238.23224-1-lukasz.luba@arm.com
Energy Aware Scheduling (EAS) needs to be able to predict the frequency
requests made by the SchedUtil governor to properly estimate energy used
in the future. It has to take into account CPUs utilization and forecast
Performance Domain (PD) frequency. There is a corner case when the max
allowed frequency might be reduced due to thermal. SchedUtil is aware of
that reduced frequency, so it should be taken into account also in EAS
estimations.
SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of
a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping
to 'policy::max'. SchedUtil is responsible to respect that upper limit
while setting the frequency through CPUFreq drivers. This effective
frequency is stored internally in 'sugov_policy::next_freq' and EAS has
to predict that value.
In the existing code the raw value of arch_scale_cpu_capacity() is used
for clamping the returned CPU utilization from effective_cpu_util().
This patch fixes issue with too big single CPU utilization, by introducing
clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU
capacity reduced by thermal pressure raw value.
Thanks to knowledge about allowed CPU capacity, we don't get too big value
for a single CPU utilization, which is then added to the util sum. The
util sum is used as a source of information for estimating whole PD energy.
To avoid wrong energy estimation in EAS (due to capped frequency), make
sure that the calculation of util sum is aware of allowed CPU capacity.
This thermal pressure might be visible in scenarios where the CPUs are not
heavily loaded, but some other component (like GPU) drastically reduced
available power budget and increased the SoC temperature. Thus, we still
use EAS for task placement and CPUs are not over-utilized.
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lore.kernel.org/r/20210614191128.22735-1-lukasz.luba@arm.com
Libo Chen
Yafang Shao
Qais Yousef
Qais Yousef
Vincent Guittot
zgpeng
Zhang Qiao
Dietmar Eggemann
Vincent Donnefort
Chen Yu
Chengming Zhou
kuyo chang
Mel Gorman
Mathias Krause
Michal Koutný
Linus Torvalds
Ingo Molnar
Josh Don
Mika Penttilä
Odin Ugedal
Yuan ZhaoXiong
Huaixin Chang
Rik van Riel
Peter Zijlstra
Lukasz Luba