commit 3af7524b14 upstream.
Running N CPU-bound tasks on an N CPUs platform:
- with asymmetric CPU capacity
- not being a DynamIq system (i.e. having a PKG level sched domain
without the SD_SHARE_PKG_RESOURCES flag set)
.. might result in a task placement where two tasks run on a big CPU
and none on a little CPU. This placement could be more optimal by
using all CPUs.
Testing platform:
Juno-r2:
- 2 big CPUs (1-2), maximum capacity of 1024
- 4 little CPUs (0,3-5), maximum capacity of 383
Testing workload ([1]):
Spawn 6 CPU-bound tasks. During the first 100ms (step 1), each tasks
is affine to a CPU, except for:
- one little CPU which is left idle.
- one big CPU which has 2 tasks affine.
After the 100ms (step 2), remove the cpumask affinity.
Behavior before the patch:
During step 2, the load balancer running from the idle CPU tags sched
domains as:
- little CPUs: 'group_has_spare'. Cf. group_has_capacity() and
group_is_overloaded(), 3 CPU-bound tasks run on a 4 CPUs
sched-domain, and the idle CPU provides enough spare capacity
regarding the imbalance_pct
- big CPUs: 'group_overloaded'. Indeed, 3 tasks run on a 2 CPUs
sched-domain, so the following path is used:
group_is_overloaded()
\-if (sgs->sum_nr_running <= sgs->group_weight) return true;
The following path which would change the migration type to
'migrate_task' is not taken:
calculate_imbalance()
\-if (env->idle != CPU_NOT_IDLE && env->imbalance == 0)
as the local group has some spare capacity, so the imbalance
is not 0.
The migration type requested is 'migrate_util' and the busiest
runqueue is the big CPU's runqueue having 2 tasks (each having a
utilization of 512). The idle little CPU cannot pull one of these
task as its capacity is too small for the task. The following path
is used:
detach_tasks()
\-case migrate_util:
\-if (util > env->imbalance) goto next;
After the patch:
As the number of failed balancing attempts grows (with
'nr_balance_failed'), progressively make it easier to migrate
a big task to the idling little CPU. A similar mechanism is
used for the 'migrate_load' migration type.
Improvement:
Running the testing workload [1] with the step 2 representing
a ~10s load for a big CPU:
Before patch: ~19.3s
After patch: ~18s (-6.7%)
Similar issue reported at:
https://lore.kernel.org/lkml/20230716014125.139577-1-qyousef@layalina.io/
Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Pierre Gondois <pierre.gondois@arm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Acked-by: Qais Yousef <qyousef@layalina.io>
Link: https://lore.kernel.org/r/20231206090043.634697-1-pierre.gondois@arm.com
Cc: John Stultz <jstultz@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 50181c0cff upstream.
Lukasz Luba reported that a thread's util_est can significantly decrease as
a result of sharing the CPU with other threads.
The use case can be easily reproduced with a periodic task TA that runs 1ms
and sleeps 100us. When the task is alone on the CPU, its max utilization and
its util_est is around 888. If another similar task starts to run on the
same CPU, TA will have to share the CPU runtime and its maximum utilization
will decrease around half the CPU capacity (512) then TA's util_est will
follow this new maximum trend which is only the result of sharing the CPU
with others tasks.
Such situation can be detected with runnable_avg wich is close or
equal to util_avg when TA is alone, but increases above util_avg when TA
shares the CPU with other threads and wait on the runqueue.
[ We prefer an util_est that overestimate rather than under estimate
because in 1st case we will not provide enough performance to the
task which will remain under-provisioned, whereas in the other case we
will create some idle time which will enable to reduce contention and
as a result reduces the util_est so the overestimate will be transient
whereas the underestimate will remain. ]
[ mingo: Refined the changelog, added comments from the LKML discussion. ]
Reported-by: Lukasz Luba <lukasz.luba@arm.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/lkml/CAKfTPtDd-HhF-YiNTtL9i5k0PfJbF819Yxu4YquzfXgwi7voyw@mail.gmail.com/#t
Link: https://lore.kernel.org/r/20231122140119.472110-1-vincent.guittot@linaro.org
Cc: Hongyan Xia <hongyan.xia2@arm.com>
Cc: John Stultz <jstultz@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 53abe3e1c1 ]
Clang is not happy with set but unused variable (this is visible
with `make W=1` build:
kernel/sched/sched.h:3744:18: error: variable 'cpumask' set but not used [-Werror,-Wunused-but-set-variable]
It seems like the variable was never used along with the assignment
that does not have side effects as far as I can see. Remove those
altogether.
Fixes: 223baf9d17 ("sched: Fix performance regression introduced by mm_cid")
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Tested-by: Eric Biggers <ebiggers@kernel.org>
Reviewed-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 17e3e88ed0 ]
The check for some lost idle pelt time should be always done when
pick_next_task_fair() fails to pick a task and not only when we call it
from the fair fast-path.
The case happens when the last running task on rq is a RT or DL task. When
the latter goes to sleep and the /Sum of util_sum of the rq is at the max
value, we don't account the lost of idle time whereas we should.
Fixes: 67692435c4 ("sched: Rework pick_next_task() slow-path")
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 8fd5485fb4 upstream.
When a CPU chooses to call push_dl_task and picks a task to push to
another CPU's runqueue then it will call find_lock_later_rq method
which would take a double lock on both CPUs' runqueues. If one of the
locks aren't readily available, it may lead to dropping the current
runqueue lock and reacquiring both the locks at once. During this window
it is possible that the task is already migrated and is running on some
other CPU. These cases are already handled. However, if the task is
migrated and has already been executed and another CPU is now trying to
wake it up (ttwu) such that it is queued again on the runqeue
(on_rq is 1) and also if the task was run by the same CPU, then the
current checks will pass even though the task was migrated out and is no
longer in the pushable tasks list.
Please go through the original rt change for more details on the issue.
To fix this, after the lock is obtained inside the find_lock_later_rq,
it ensures that the task is still at the head of pushable tasks list.
Also removed some checks that are no longer needed with the addition of
this new check.
However, the new check of pushable tasks list only applies when
find_lock_later_rq is called by push_dl_task. For the other caller i.e.
dl_task_offline_migration, existing checks are used.
Signed-off-by: Harshit Agarwal <harshit@nutanix.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20250408045021.3283624-1-harshit@nutanix.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit e37617c8e5 upstream.
Linus reported a ~50% performance regression on single-threaded
workloads on his AMD Ryzen system, and bisected it to:
9c0b4bb7f6 ("sched/cpufreq: Rework schedutil governor performance estimation")
When frequency invariance is not enabled, get_capacity_ref_freq(policy)
is supposed to return the current frequency and the performance margin
applied by map_util_perf(), enabling the utilization to go above the
maximum compute capacity and to select a higher frequency than the current one.
After the changes in 9c0b4bb7f6, the performance margin was applied
earlier in the path to take into account utilization clampings and
we couldn't get a utilization higher than the maximum compute capacity,
and the CPU remained 'stuck' at lower frequencies.
To fix this, we must use a frequency above the current frequency to
get a chance to select a higher OPP when the current one becomes fully used.
Apply the same margin and return a frequency 25% higher than the current
one in order to switch to the next OPP before we fully use the CPU
at the current one.
[ mingo: Clarified the changelog. ]
Fixes: 9c0b4bb7f6 ("sched/cpufreq: Rework schedutil governor performance estimation")
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Bisected-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: Wyes Karny <wkarny@gmail.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Wyes Karny <wkarny@gmail.com>
Link: https://lore.kernel.org/r/20240114183600.135316-1-vincent.guittot@linaro.org
Signed-off-by: Wentao Guan <guanwentao@uniontech.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 155213a2ae ]
schbench (https://github.com/masoncl/schbench.git) is showing a
regression from previous production kernels that bisected down to:
sched/fair: Remove sysctl_sched_migration_cost condition (c5b0a7eefc)
The schbench command line was:
schbench -L -m 4 -M auto -t 256 -n 0 -r 0 -s 0
This creates 4 message threads pinned to CPUs 0-3, and 256x4 worker
threads spread across the rest of the CPUs. Neither the worker threads
or the message threads do any work, they just wake each other up and go
back to sleep as soon as possible.
The end result is the first 4 CPUs are pegged waking up those 1024
workers, and the rest of the CPUs are constantly banging in and out of
idle. If I take a v6.9 Linus kernel and revert that one commit,
performance goes from 3.4M RPS to 5.4M RPS.
schedstat shows there are ~100x more new idle balance operations, and
profiling shows the worker threads are spending ~20% of their CPU time
on new idle balance. schedstats also shows that almost all of these new
idle balance attemps are failing to find busy groups.
The fix used here is to crank up the cost of the newidle balance whenever it
fails. Since we don't want sd->max_newidle_lb_cost to grow out of
control, this also changes update_newidle_cost() to use
sysctl_sched_migration_cost as the upper limit on max_newidle_lb_cost.
Signed-off-by: Chris Mason <clm@fb.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20250626144017.1510594-2-clm@fb.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 8f0eed4a78 upstream.
After commit f5d39b0208 ("freezer,sched: Rewrite core freezer logic"),
tasks that transition directly from TASK_FREEZABLE to TASK_FROZEN are
always woken up on the thaw path. Prior to that commit, tasks could ask
freezer to consider them "frozen enough" via freezer_do_not_count(). The
commit replaced freezer_do_not_count() with a TASK_FREEZABLE state which
allows freezer to immediately mark the task as TASK_FROZEN without
waking up the task. This is efficient for the suspend path, but on the
thaw path, the task is always woken up even if the task didn't need to
wake up and goes back to its TASK_(UN)INTERRUPTIBLE state. Although
these tasks are capable of handling of the wakeup, we can observe a
power/perf impact from the extra wakeup.
We observed on Android many tasks wait in the TASK_FREEZABLE state
(particularly due to many of them being binder clients). We observed
nearly 4x the number of tasks and a corresponding linear increase in
latency and power consumption when thawing the system. The latency
increased from ~15ms to ~50ms.
Avoid the spurious wakeups by saving the state of TASK_FREEZABLE tasks.
If the task was running before entering TASK_FROZEN state
(__refrigerator()) or if the task received a wake up for the saved
state, then the task is woken on thaw. saved_state from PREEMPT_RT locks
can be re-used because freezer would not stomp on the rtlock wait flow:
TASK_RTLOCK_WAIT isn't considered freezable.
Reported-by: Prakash Viswalingam <quic_prakashv@quicinc.com>
Signed-off-by: Elliot Berman <quic_eberman@quicinc.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Chen Ridong <chenridong@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit fbaa6a181a upstream.
In preparation for freezer to also use saved_state, remove the
CONFIG_PREEMPT_RT compilation guard around saved_state.
On the arm64 platform I tested which did not have CONFIG_PREEMPT_RT,
there was no statistically significant deviation by applying this patch.
Test methodology:
perf bench sched message -g 40 -l 40
Signed-off-by: Elliot Berman <quic_eberman@quicinc.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Chen Ridong <chenridong@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 99b773d720 ]
With the seqcount moved out of the group into a global psi_seq,
re-initializing the seqcount on group creation is causing seqcount
corruption.
Fixes: 570c8efd5e ("sched/psi: Optimize psi_group_change() cpu_clock() usage")
Reported-by: Chris Mason <clm@meta.com>
Suggested-by: Beata Michalska <beata.michalska@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 570c8efd5e ]
Dietmar reported that commit 3840cbe24c ("sched: psi: fix bogus
pressure spikes from aggregation race") caused a regression for him on
a high context switch rate benchmark (schbench) due to the now
repeating cpu_clock() calls.
In particular the problem is that get_recent_times() will extrapolate
the current state to 'now'. But if an update uses a timestamp from
before the start of the update, it is possible to get two reads
with inconsistent results. It is effectively back-dating an update.
(note that this all hard-relies on the clock being synchronized across
CPUs -- if this is not the case, all bets are off).
Combine this problem with the fact that there are per-group-per-cpu
seqcounts, the commit in question pushed the clock read into the group
iteration, causing tree-depth cpu_clock() calls. On architectures
where cpu_clock() has appreciable overhead, this hurts.
Instead move to a per-cpu seqcount, which allows us to have a single
clock read for all group updates, increasing internal consistency and
lowering update overhead. This comes at the cost of a longer update
side (proportional to the tree depth) which can cause the read side to
retry more often.
Fixes: 3840cbe24c ("sched: psi: fix bogus pressure spikes from aggregation race")
Reported-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com>,
Link: https://lkml.kernel.org/20250522084844.GC31726@noisy.programming.kicks-ass.net
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 36569780b0 upstream.
The commit e6fe3f422b ("sched: Make multiple runqueue task counters
32-bit") changed nr_uninterruptible to an unsigned int. But the
nr_uninterruptible values for each of the CPU runqueues can grow to
large numbers, sometimes exceeding INT_MAX. This is valid, if, over
time, a large number of tasks are migrated off of one CPU after going
into an uninterruptible state. Only the sum of all nr_interruptible
values across all CPUs yields the correct result, as explained in a
comment in kernel/sched/loadavg.c.
Change the type of nr_uninterruptible back to unsigned long to prevent
overflows, and thus the miscalculation of load average.
Fixes: e6fe3f422b ("sched: Make multiple runqueue task counters 32-bit")
Signed-off-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20250709173328.606794-1-aruna.ramakrishna@oracle.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 2ae891b826 ]
The old default value for slice is 0.75 msec * (1 + ilog(ncpus)) which
means that we have a default slice of:
0.75 for 1 cpu
1.50 up to 3 cpus
2.25 up to 7 cpus
3.00 for 8 cpus and above.
For HZ=250 and HZ=100, because of the tick accuracy, the runtime of
tasks is far higher than their slice.
For HZ=1000 with 8 cpus or more, the accuracy of tick is already
satisfactory, but there is still an issue that tasks will get an extra
tick because the tick often arrives a little faster than expected. In
this case, the task can only wait until the next tick to consider that it
has reached its deadline, and will run 1ms longer.
vruntime + sysctl_sched_base_slice = deadline
|-----------|-----------|-----------|-----------|
1ms 1ms 1ms 1ms
^ ^ ^ ^
tick1 tick2 tick3 tick4(nearly 4ms)
There are two reasons for tick error: clockevent precision and the
CONFIG_IRQ_TIME_ACCOUNTING/CONFIG_PARAVIRT_TIME_ACCOUNTING. with
CONFIG_IRQ_TIME_ACCOUNTING every tick will be less than 1ms, but even
without it, because of clockevent precision, tick still often less than
1ms.
In order to make scheduling more precise, we changed 0.75 to 0.70,
Using 0.70 instead of 0.75 should not change much for other configs
and would fix this issue:
0.70 for 1 cpu
1.40 up to 3 cpus
2.10 up to 7 cpus
2.8 for 8 cpus and above.
This does not guarantee that tasks can run the slice time accurately
every time, but occasionally running an extra tick has little impact.
Signed-off-by: zihan zhou <15645113830zzh@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20250208075322.13139-1-15645113830zzh@gmail.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 79443a7e9d ]
The handling of the limits_changed flag in struct sugov_policy needs to
be explicitly synchronized to ensure that cpufreq policy limits updates
will not be missed in some cases.
Without that synchronization it is theoretically possible that
the limits_changed update in sugov_should_update_freq() will be
reordered with respect to the reads of the policy limits in
cpufreq_driver_resolve_freq() and in that case, if the limits_changed
update in sugov_limits() clobbers the one in sugov_should_update_freq(),
the new policy limits may not take effect for a long time.
Likewise, the limits_changed update in sugov_limits() may theoretically
get reordered with respect to the updates of the policy limits in
cpufreq_set_policy() and if sugov_should_update_freq() runs between
them, the policy limits change may be missed.
To ensure that the above situations will not take place, add memory
barriers preventing the reordering in question from taking place and
add READ_ONCE() and WRITE_ONCE() annotations around all of the
limits_changed flag updates to prevent the compiler from messing up
with that code.
Fixes: 600f5badb7 ("cpufreq: schedutil: Don't skip freq update when limits change")
Cc: 5.3+ <stable@vger.kernel.org> # 5.3+
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/3376719.44csPzL39Z@rjwysocki.net
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9c0b4bb7f6 ]
The current method to take into account uclamp hints when estimating the
target frequency can end in a situation where the selected target
frequency is finally higher than uclamp hints, whereas there are no real
needs. Such cases mainly happen because we are currently mixing the
traditional scheduler utilization signal with the uclamp performance
hints. By adding these 2 metrics, we loose an important information when
it comes to select the target frequency, and we have to make some
assumptions which can't fit all cases.
Rework the interface between the scheduler and schedutil governor in order
to propagate all information down to the cpufreq governor.
effective_cpu_util() interface changes and now returns the actual
utilization of the CPU with 2 optional inputs:
- The minimum performance for this CPU; typically the capacity to handle
the deadline task and the interrupt pressure. But also uclamp_min
request when available.
- The maximum targeting performance for this CPU which reflects the
maximum level that we would like to not exceed. By default it will be
the CPU capacity but can be reduced because of some performance hints
set with uclamp. The value can be lower than actual utilization and/or
min performance level.
A new sugov_effective_cpu_perf() interface is also available to compute
the final performance level that is targeted for the CPU, after applying
some cpufreq headroom and taking into account all inputs.
With these 2 functions, schedutil is now able to decide when it must go
above uclamp hints. It now also has a generic way to get the min
performance level.
The dependency between energy model and cpufreq governor and its headroom
policy doesn't exist anymore.
eenv_pd_max_util() asks schedutil for the targeted performance after
applying the impact of the waking task.
[ mingo: Refined the changelog & C comments. ]
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Rafael J. Wysocki <rafael@kernel.org>
Link: https://lore.kernel.org/r/20231122133904.446032-2-vincent.guittot@linaro.org
Stable-dep-of: 79443a7e9d ("cpufreq/sched: Explicitly synchronize limits_changed flag handling")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 7bc263840b ]
Remove the rq::cpu_capacity_orig field and use arch_scale_cpu_capacity()
instead.
The scheduler uses 3 methods to get access to a CPU's max compute capacity:
- arch_scale_cpu_capacity(cpu) which is the default way to get a CPU's capacity.
- cpu_capacity_orig field which is periodically updated with
arch_scale_cpu_capacity().
- capacity_orig_of(cpu) which encapsulates rq->cpu_capacity_orig.
There is no real need to save the value returned by arch_scale_cpu_capacity()
in struct rq. arch_scale_cpu_capacity() returns:
- either a per_cpu variable.
- or a const value for systems which have only one capacity.
Remove rq::cpu_capacity_orig and use arch_scale_cpu_capacity() everywhere.
No functional changes.
Some performance tests on Arm64:
- small SMP device (hikey): no noticeable changes
- HMP device (RB5): hackbench shows minor improvement (1-2%)
- large smp (thx2): hackbench and tbench shows minor improvement (1%)
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lore.kernel.org/r/20231009103621.374412-2-vincent.guittot@linaro.org
Stable-dep-of: 79443a7e9d ("cpufreq/sched: Explicitly synchronize limits_changed flag handling")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit cfde542df7 ]
Commit 8e461a1cb4 ("cpufreq: schedutil: Fix superfluous updates caused
by need_freq_update") modified sugov_should_update_freq() to set the
need_freq_update flag only for drivers with CPUFREQ_NEED_UPDATE_LIMITS
set, but that flag generally needs to be set when the policy limits
change because the driver callback may need to be invoked for the new
limits to take effect.
However, if the return value of cpufreq_driver_resolve_freq() after
applying the new limits is still equal to the previously selected
frequency, the driver callback needs to be invoked only in the case
when CPUFREQ_NEED_UPDATE_LIMITS is set (which means that the driver
specifically wants its callback to be invoked every time the policy
limits change).
Update the code accordingly to avoid missing policy limits changes for
drivers without CPUFREQ_NEED_UPDATE_LIMITS.
Fixes: 8e461a1cb4 ("cpufreq: schedutil: Fix superfluous updates caused by need_freq_update")
Closes: https://lore.kernel.org/lkml/Z_Tlc6Qs-tYpxWYb@linaro.org/
Reported-by: Stephan Gerhold <stephan.gerhold@linaro.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/3010358.e9J7NaK4W3@rjwysocki.net
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 14672f059d ]
The ftrace selftest reported a failure because writing -1 to
sched_rt_runtime_us returns -EBUSY. This happens when the possible
CPUs are different from active CPUs.
Active CPUs are part of one root domain, while remaining CPUs are part
of def_root_domain. Since active cpumask is being used, this results in
cpus=0 when a non active CPUs is used in the loop.
Fix it by looping over the online CPUs instead for validating the
bandwidth calculations.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/20250306052954.452005-2-sshegde@linux.ibm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 76f970ce51 upstream.
This reverts commit eff6c8ce8d.
Hazem reported a 30% drop in UnixBench spawn test with commit
eff6c8ce8d ("sched/core: Reduce cost of sched_move_task when config
autogroup") on a m6g.xlarge AWS EC2 instance with 4 vCPUs and 16 GiB RAM
(aarch64) (single level MC sched domain):
https://lkml.kernel.org/r/20250205151026.13061-1-hagarhem@amazon.com
There is an early bail from sched_move_task() if p->sched_task_group is
equal to p's 'cpu cgroup' (sched_get_task_group()). E.g. both are
pointing to taskgroup '/user.slice/user-1000.slice/session-1.scope'
(Ubuntu '22.04.5 LTS').
So in:
do_exit()
sched_autogroup_exit_task()
sched_move_task()
if sched_get_task_group(p) == p->sched_task_group
return
/* p is enqueued */
dequeue_task() \
sched_change_group() |
task_change_group_fair() |
detach_task_cfs_rq() | (1)
set_task_rq() |
attach_task_cfs_rq() |
enqueue_task() /
(1) isn't called for p anymore.
Turns out that the regression is related to sgs->group_util in
group_is_overloaded() and group_has_capacity(). If (1) isn't called for
all the 'spawn' tasks then sgs->group_util is ~900 and
sgs->group_capacity = 1024 (single CPU sched domain) and this leads to
group_is_overloaded() returning true (2) and group_has_capacity() false
(3) much more often compared to the case when (1) is called.
I.e. there are much more cases of 'group_is_overloaded' and
'group_fully_busy' in WF_FORK wakeup sched_balance_find_dst_cpu() which
then returns much more often a CPU != smp_processor_id() (5).
This isn't good for these extremely short running tasks (FORK + EXIT)
and also involves calling sched_balance_find_dst_group_cpu() unnecessary
(single CPU sched domain).
Instead if (1) is called for 'p->flags & PF_EXITING' then the path
(4),(6) is taken much more often.
select_task_rq_fair(..., wake_flags = WF_FORK)
cpu = smp_processor_id()
new_cpu = sched_balance_find_dst_cpu(..., cpu, ...)
group = sched_balance_find_dst_group(..., cpu)
do {
update_sg_wakeup_stats()
sgs->group_type = group_classify()
if group_is_overloaded() (2)
return group_overloaded
if !group_has_capacity() (3)
return group_fully_busy
return group_has_spare (4)
} while group
if local_sgs.group_type > idlest_sgs.group_type
return idlest (5)
case group_has_spare:
if local_sgs.idle_cpus >= idlest_sgs.idle_cpus
return NULL (6)
Unixbench Tests './Run -c 4 spawn' on:
(a) VM AWS instance (m7gd.16xlarge) with v6.13 ('maxcpus=4 nr_cpus=4')
and Ubuntu 22.04.5 LTS (aarch64).
Shell & test run in '/user.slice/user-1000.slice/session-1.scope'.
w/o patch w/ patch
21005 27120
(b) i7-13700K with tip/sched/core ('nosmt maxcpus=8 nr_cpus=8') and
Ubuntu 22.04.5 LTS (x86_64).
Shell & test run in '/A'.
w/o patch w/ patch
67675 88806
CONFIG_SCHED_AUTOGROUP=y & /sys/proc/kernel/sched_autogroup_enabled equal
0 or 1.
Reported-by: Hazem Mohamed Abuelfotoh <abuehaze@amazon.com>
Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Hagar Hemdan <hagarhem@amazon.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/r/20250314151345.275739-1-dietmar.eggemann@arm.com
[Hagar: clean revert of eff6c8ce8dd7 to make it work on 6.6]
Signed-off-by: Hagar Hemdan <hagarhem@amazon.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit ed4fb6d7ef upstream.
The timerslack_ns setting is used to specify how much the hardware
timers should be delayed, to potentially dispatch multiple timers in a
single interrupt. This is a performance optimization. Timers of
realtime tasks (having a realtime scheduling policy) should not be
delayed.
This logic was inconsitently applied to the hrtimers, leading to delays
of realtime tasks which used timed waits for events (e.g. condition
variables). Due to the downstream override of the slack for rt tasks,
the procfs reported incorrect (non-zero) timerslack_ns values.
This is changed by setting the timer_slack_ns task attribute to 0 for
all tasks with a rt policy. By that, downstream users do not need to
specially handle rt tasks (w.r.t. the slack), and the procfs entry
shows the correct value of "0". Setting non-zero slack values (either
via procfs or PR_SET_TIMERSLACK) on tasks with a rt policy is ignored,
as stated in "man 2 PR_SET_TIMERSLACK":
Timer slack is not applied to threads that are scheduled under a
real-time scheduling policy (see sched_setscheduler(2)).
The special handling of timerslack on rt tasks in downstream users
is removed as well.
Signed-off-by: Felix Moessbauer <felix.moessbauer@siemens.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/all/20240814121032.368444-2-felix.moessbauer@siemens.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 3b4035ddbf ]
child_cfs_rq_on_list attempts to convert a 'prev' pointer to a cfs_rq.
This 'prev' pointer can originate from struct rq's leaf_cfs_rq_list,
making the conversion invalid and potentially leading to memory
corruption. Depending on the relative positions of leaf_cfs_rq_list and
the task group (tg) pointer within the struct, this can cause a memory
fault or access garbage data.
The issue arises in list_add_leaf_cfs_rq, where both
cfs_rq->leaf_cfs_rq_list and rq->leaf_cfs_rq_list are added to the same
leaf list. Also, rq->tmp_alone_branch can be set to rq->leaf_cfs_rq_list.
This adds a check `if (prev == &rq->leaf_cfs_rq_list)` after the main
conditional in child_cfs_rq_on_list. This ensures that the container_of
operation will convert a correct cfs_rq struct.
This check is sufficient because only cfs_rqs on the same CPU are added
to the list, so verifying the 'prev' pointer against the current rq's list
head is enough.
Fixes a potential memory corruption issue that due to current struct
layout might not be manifesting as a crash but could lead to unpredictable
behavior when the layout changes.
Fixes: fdaba61ef8 ("sched/fair: Ensure that the CFS parent is added after unthrottling")
Signed-off-by: Zecheng Li <zecheng@google.com>
Reviewed-and-tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20250304214031.2882646-1-zecheng@google.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 82c387ef75 upstream.
David reported a warning observed while loop testing kexec jump:
Interrupts enabled after irqrouter_resume+0x0/0x50
WARNING: CPU: 0 PID: 560 at drivers/base/syscore.c:103 syscore_resume+0x18a/0x220
kernel_kexec+0xf6/0x180
__do_sys_reboot+0x206/0x250
do_syscall_64+0x95/0x180
The corresponding interrupt flag trace:
hardirqs last enabled at (15573): [<ffffffffa8281b8e>] __up_console_sem+0x7e/0x90
hardirqs last disabled at (15580): [<ffffffffa8281b73>] __up_console_sem+0x63/0x90
That means __up_console_sem() was invoked with interrupts enabled. Further
instrumentation revealed that in the interrupt disabled section of kexec
jump one of the syscore_suspend() callbacks woke up a task, which set the
NEED_RESCHED flag. A later callback in the resume path invoked
cond_resched() which in turn led to the invocation of the scheduler:
__cond_resched+0x21/0x60
down_timeout+0x18/0x60
acpi_os_wait_semaphore+0x4c/0x80
acpi_ut_acquire_mutex+0x3d/0x100
acpi_ns_get_node+0x27/0x60
acpi_ns_evaluate+0x1cb/0x2d0
acpi_rs_set_srs_method_data+0x156/0x190
acpi_pci_link_set+0x11c/0x290
irqrouter_resume+0x54/0x60
syscore_resume+0x6a/0x200
kernel_kexec+0x145/0x1c0
__do_sys_reboot+0xeb/0x240
do_syscall_64+0x95/0x180
This is a long standing problem, which probably got more visible with
the recent printk changes. Something does a task wakeup and the
scheduler sets the NEED_RESCHED flag. cond_resched() sees it set and
invokes schedule() from a completely bogus context. The scheduler
enables interrupts after context switching, which causes the above
warning at the end.
Quite some of the code paths in syscore_suspend()/resume() can result in
triggering a wakeup with the exactly same consequences. They might not
have done so yet, but as they share a lot of code with normal operations
it's just a question of time.
The problem only affects the PREEMPT_NONE and PREEMPT_VOLUNTARY scheduling
models. Full preemption is not affected as cond_resched() is disabled and
the preemption check preemptible() takes the interrupt disabled flag into
account.
Cure the problem by adding a corresponding check into cond_resched().
Reported-by: David Woodhouse <dwmw@amazon.co.uk>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: David Woodhouse <dwmw@amazon.co.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: stable@vger.kernel.org
Closes: https://lore.kernel.org/all/7717fe2ac0ce5f0a2c43fdab8b11f4483d54a2a4.camel@infradead.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 108ad09990 ]
When steal time exceeds the measured delta when updating clock_task, we
currently try to catch up the excess in future updates.
However, this results in inaccurate run times for the future things using
clock_task, in some situations, as they end up getting additional steal
time that did not actually happen.
This is because there is a window between reading the elapsed time in
update_rq_clock() and sampling the steal time in update_rq_clock_task().
If the VCPU gets preempted between those two points, any additional
steal time is accounted to the outgoing task even though the calculated
delta did not actually contain any of that "stolen" time.
When this race happens, we can end up with steal time that exceeds the
calculated delta, and the previous code would try to catch up that excess
steal time in future clock updates, which is given to the next,
incoming task, even though it did not actually have any time stolen.
This behavior is particularly bad when steal time can be very long,
which we've seen when trying to extend steal time to contain the duration
that the host was suspended [0]. When this happens, clock_task stays
frozen, during which the running task stays running for the whole
duration, since its run time doesn't increase.
However the race can happen even under normal operation.
Ideally we would read the elapsed cpu time and the steal time atomically,
to prevent this race from happening in the first place, but doing so
is non-trivial.
Since the time between those two points isn't otherwise accounted anywhere,
neither to the outgoing task nor the incoming task (because the "end of
outgoing task" and "start of incoming task" timestamps are the same),
I would argue that the right thing to do is to simply drop any excess steal
time, in order to prevent these issues.
[0] https://lore.kernel.org/kvm/20240820043543.837914-1-suleiman@google.com/
Signed-off-by: Suleiman Souhlal <suleiman@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241118043745.1857272-1-suleiman@google.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8e461a1cb4 ]
A redundant frequency update is only truly needed when there is a policy
limits change with a driver that specifies CPUFREQ_NEED_UPDATE_LIMITS.
In spite of that, drivers specifying CPUFREQ_NEED_UPDATE_LIMITS receive a
frequency update _all the time_, not just for a policy limits change,
because need_freq_update is never cleared.
Furthermore, ignore_dl_rate_limit()'s usage of need_freq_update also leads
to a redundant frequency update, regardless of whether or not the driver
specifies CPUFREQ_NEED_UPDATE_LIMITS, when the next chosen frequency is the
same as the current one.
Fix the superfluous updates by only honoring CPUFREQ_NEED_UPDATE_LIMITS
when there's a policy limits change, and clearing need_freq_update when a
requisite redundant update occurs.
This is neatly achieved by moving up the CPUFREQ_NEED_UPDATE_LIMITS test
and instead setting need_freq_update to false in sugov_update_next_freq().
Fixes: 600f5badb7 ("cpufreq: schedutil: Don't skip freq update when limits change")
Signed-off-by: Sultan Alsawaf (unemployed) <sultan@kerneltoast.com>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/20241212015734.41241-2-sultan@kerneltoast.com
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit a430d99e34 ]
In /proc/schedstat, lb_hot_gained reports the number hot tasks pulled
during load balance. This value is incremented in can_migrate_task()
if the task is migratable and hot. After incrementing the value,
load balancer can still decide not to migrate this task leading to wrong
accounting. Fix this by incrementing stats when hot tasks are detached.
This issue only exists in detach_tasks() where we can decide to not
migrate hot task even if it is migratable. However, in detach_one_task(),
we migrate it unconditionally.
[Swapnil: Handled the case where nr_failed_migrations_hot was not accounted properly and wrote commit log]
Fixes: d31980846f ("sched: Move up affinity check to mitigate useless redoing overhead")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reported-by: "Gautham R. Shenoy" <gautham.shenoy@amd.com>
Not-yet-signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Swapnil Sapkal <swapnil.sapkal@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241220063224.17767-2-swapnil.sapkal@amd.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit b23decf8ac ]
Idle tasks are initialized via __sched_fork() twice:
fork_idle()
copy_process()
sched_fork()
__sched_fork()
init_idle()
__sched_fork()
Instead of cleaning this up, sched_ext hacked around it. Even when analyis
and solution were provided in a discussion, nobody cared to clean this up.
init_idle() is also invoked from sched_init() to initialize the boot CPU's
idle task, which requires the __sched_fork() invocation. But this can be
trivially solved by invoking __sched_fork() before init_idle() in
sched_init() and removing the __sched_fork() invocation from init_idle().
Do so and clean up the comments explaining this historical leftover.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241028103142.359584747@linutronix.de
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 0664e2c311 ]
When running the following command:
while true; do
stress-ng --cyclic 30 --timeout 30s --minimize --quiet
done
a warning is eventually triggered:
WARNING: CPU: 43 PID: 2848 at kernel/sched/deadline.c:794
setup_new_dl_entity+0x13e/0x180
...
Call Trace:
<TASK>
? show_trace_log_lvl+0x1c4/0x2df
? enqueue_dl_entity+0x631/0x6e0
? setup_new_dl_entity+0x13e/0x180
? __warn+0x7e/0xd0
? report_bug+0x11a/0x1a0
? handle_bug+0x3c/0x70
? exc_invalid_op+0x14/0x70
? asm_exc_invalid_op+0x16/0x20
enqueue_dl_entity+0x631/0x6e0
enqueue_task_dl+0x7d/0x120
__do_set_cpus_allowed+0xe3/0x280
__set_cpus_allowed_ptr_locked+0x140/0x1d0
__set_cpus_allowed_ptr+0x54/0xa0
migrate_enable+0x7e/0x150
rt_spin_unlock+0x1c/0x90
group_send_sig_info+0xf7/0x1a0
? kill_pid_info+0x1f/0x1d0
kill_pid_info+0x78/0x1d0
kill_proc_info+0x5b/0x110
__x64_sys_kill+0x93/0xc0
do_syscall_64+0x5c/0xf0
entry_SYSCALL_64_after_hwframe+0x6e/0x76
RIP: 0033:0x7f0dab31f92b
This warning occurs because set_cpus_allowed dequeues and enqueues tasks
with the ENQUEUE_RESTORE flag set. If the task is boosted, the warning
is triggered. A boosted task already had its parameters set by
rt_mutex_setprio, and a new call to setup_new_dl_entity is unnecessary,
hence the WARN_ON call.
Check if we are requeueing a boosted task and avoid calling
setup_new_dl_entity if that's the case.
Fixes: 295d6d5e37 ("sched/deadline: Fix switching to -deadline")
Signed-off-by: Wander Lairson Costa <wander@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/20240724142253.27145-2-wander@redhat.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit c708a4dc5a ]
Now that trace_sched_stat_runtime() no longer takes a vruntime
argument, the task specific bits are identical between
update_curr_common() and update_curr().
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Stable-dep-of: 0664e2c311 ("sched/deadline: Fix warning in migrate_enable for boosted tasks")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 5fe6ec8f6a ]
Tracing the runtime delta makes sense, observer can sum over time.
Tracing the absolute vruntime makes less sense, inconsistent:
absolute-vs-delta, but also vruntime delta can be computed from
runtime delta.
Removing the vruntime thing also makes the two tracepoint sites
identical, allowing to unify the code in a later patch.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Stable-dep-of: 0664e2c311 ("sched/deadline: Fix warning in migrate_enable for boosted tasks")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e23edc86b0 ]
The name is a bit opaque - make it clear that this is about wakeup
preemption.
Also rename the ->check_preempt_curr() methods similarly.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Stable-dep-of: 0664e2c311 ("sched/deadline: Fix warning in migrate_enable for boosted tasks")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 82845683ca ]
Other scheduling classes already postfix their similar methods
with the class name.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Stable-dep-of: 0664e2c311 ("sched/deadline: Fix warning in migrate_enable for boosted tasks")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e932c4ab38 ]
Scheduler raises a SCHED_SOFTIRQ to trigger a load balancing event on
from the IPI handler on the idle CPU. If the SMP function is invoked
from an idle CPU via flush_smp_call_function_queue() then the HARD-IRQ
flag is not set and raise_softirq_irqoff() needlessly wakes ksoftirqd
because soft interrupts are handled before ksoftirqd get on the CPU.
Adding a trace_printk() in nohz_csd_func() at the spot of raising
SCHED_SOFTIRQ and enabling trace events for sched_switch, sched_wakeup,
and softirq_entry (for SCHED_SOFTIRQ vector alone) helps observing the
current behavior:
<idle>-0 [000] dN.1.: nohz_csd_func: Raising SCHED_SOFTIRQ from nohz_csd_func
<idle>-0 [000] dN.4.: sched_wakeup: comm=ksoftirqd/0 pid=16 prio=120 target_cpu=000
<idle>-0 [000] .Ns1.: softirq_entry: vec=7 [action=SCHED]
<idle>-0 [000] .Ns1.: softirq_exit: vec=7 [action=SCHED]
<idle>-0 [000] d..2.: sched_switch: prev_comm=swapper/0 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=ksoftirqd/0 next_pid=16 next_prio=120
ksoftirqd/0-16 [000] d..2.: sched_switch: prev_comm=ksoftirqd/0 prev_pid=16 prev_prio=120 prev_state=S ==> next_comm=swapper/0 next_pid=0 next_prio=120
...
Use __raise_softirq_irqoff() to raise the softirq. The SMP function call
is always invoked on the requested CPU in an interrupt handler. It is
guaranteed that soft interrupts are handled at the end.
Following are the observations with the changes when enabling the same
set of events:
<idle>-0 [000] dN.1.: nohz_csd_func: Raising SCHED_SOFTIRQ for nohz_idle_balance
<idle>-0 [000] dN.1.: softirq_raise: vec=7 [action=SCHED]
<idle>-0 [000] .Ns1.: softirq_entry: vec=7 [action=SCHED]
No unnecessary ksoftirqd wakeups are seen from idle task's context to
service the softirq.
Fixes: b2a02fc43a ("smp: Optimize send_call_function_single_ipi()")
Closes: https://lore.kernel.org/lkml/fcf823f-195e-6c9a-eac3-25f870cb35ac@inria.fr/ [1]
Reported-by: Julia Lawall <julia.lawall@inria.fr>
Suggested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lore.kernel.org/r/20241119054432.6405-5-kprateek.nayak@amd.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ff47a0acfc ]
Commit b2a02fc43a ("smp: Optimize send_call_function_single_ipi()")
optimizes IPIs to idle CPUs in TIF_POLLING_NRFLAG mode by setting the
TIF_NEED_RESCHED flag in idle task's thread info and relying on
flush_smp_call_function_queue() in idle exit path to run the
call-function. A softirq raised by the call-function is handled shortly
after in do_softirq_post_smp_call_flush() but the TIF_NEED_RESCHED flag
remains set and is only cleared later when schedule_idle() calls
__schedule().
need_resched() check in _nohz_idle_balance() exists to bail out of load
balancing if another task has woken up on the CPU currently in-charge of
idle load balancing which is being processed in SCHED_SOFTIRQ context.
Since the optimization mentioned above overloads the interpretation of
TIF_NEED_RESCHED, check for idle_cpu() before going with the existing
need_resched() check which can catch a genuine task wakeup on an idle
CPU processing SCHED_SOFTIRQ from do_softirq_post_smp_call_flush(), as
well as the case where ksoftirqd needs to be preempted as a result of
new task wakeup or slice expiry.
In case of PREEMPT_RT or threadirqs, although the idle load balancing
may be inhibited in some cases on the ilb CPU, the fact that ksoftirqd
is the only fair task going back to sleep will trigger a newidle balance
on the CPU which will alleviate some imbalance if it exists if idle
balance fails to do so.
Fixes: b2a02fc43a ("smp: Optimize send_call_function_single_ipi()")
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241119054432.6405-4-kprateek.nayak@amd.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ea9cffc0a1 ]
The need_resched() check currently in nohz_csd_func() can be tracked
to have been added in scheduler_ipi() back in 2011 via commit
ca38062e57 ("sched: Use resched IPI to kick off the nohz idle balance")
Since then, it has travelled quite a bit but it seems like an idle_cpu()
check currently is sufficient to detect the need to bail out from an
idle load balancing. To justify this removal, consider all the following
case where an idle load balancing could race with a task wakeup:
o Since commit f3dd3f6745 ("sched: Remove the limitation of WF_ON_CPU
on wakelist if wakee cpu is idle") a target perceived to be idle
(target_rq->nr_running == 0) will return true for
ttwu_queue_cond(target) which will offload the task wakeup to the idle
target via an IPI.
In all such cases target_rq->ttwu_pending will be set to 1 before
queuing the wake function.
If an idle load balance races here, following scenarios are possible:
- The CPU is not in TIF_POLLING_NRFLAG mode in which case an actual
IPI is sent to the CPU to wake it out of idle. If the
nohz_csd_func() queues before sched_ttwu_pending(), the idle load
balance will bail out since idle_cpu(target) returns 0 since
target_rq->ttwu_pending is 1. If the nohz_csd_func() is queued after
sched_ttwu_pending() it should see rq->nr_running to be non-zero and
bail out of idle load balancing.
- The CPU is in TIF_POLLING_NRFLAG mode and instead of an actual IPI,
the sender will simply set TIF_NEED_RESCHED for the target to put it
out of idle and flush_smp_call_function_queue() in do_idle() will
execute the call function. Depending on the ordering of the queuing
of nohz_csd_func() and sched_ttwu_pending(), the idle_cpu() check in
nohz_csd_func() should either see target_rq->ttwu_pending = 1 or
target_rq->nr_running to be non-zero if there is a genuine task
wakeup racing with the idle load balance kick.
o The waker CPU perceives the target CPU to be busy
(targer_rq->nr_running != 0) but the CPU is in fact going idle and due
to a series of unfortunate events, the system reaches a case where the
waker CPU decides to perform the wakeup by itself in ttwu_queue() on
the target CPU but target is concurrently selected for idle load
balance (XXX: Can this happen? I'm not sure, but we'll consider the
mother of all coincidences to estimate the worst case scenario).
ttwu_do_activate() calls enqueue_task() which would increment
"rq->nr_running" post which it calls wakeup_preempt() which is
responsible for setting TIF_NEED_RESCHED (via a resched IPI or by
setting TIF_NEED_RESCHED on a TIF_POLLING_NRFLAG idle CPU) The key
thing to note in this case is that rq->nr_running is already non-zero
in case of a wakeup before TIF_NEED_RESCHED is set which would
lead to idle_cpu() check returning false.
In all cases, it seems that need_resched() check is unnecessary when
checking for idle_cpu() first since an impending wakeup racing with idle
load balancer will either set the "rq->ttwu_pending" or indicate a newly
woken task via "rq->nr_running".
Chasing the reason why this check might have existed in the first place,
I came across Peter's suggestion on the fist iteration of Suresh's
patch from 2011 [1] where the condition to raise the SCHED_SOFTIRQ was:
sched_ttwu_do_pending(list);
if (unlikely((rq->idle == current) &&
rq->nohz_balance_kick &&
!need_resched()))
raise_softirq_irqoff(SCHED_SOFTIRQ);
Since the condition to raise the SCHED_SOFIRQ was preceded by
sched_ttwu_do_pending() (which is equivalent of sched_ttwu_pending()) in
the current upstream kernel, the need_resched() check was necessary to
catch a newly queued task. Peter suggested modifying it to:
if (idle_cpu() && rq->nohz_balance_kick && !need_resched())
raise_softirq_irqoff(SCHED_SOFTIRQ);
where idle_cpu() seems to have replaced "rq->idle == current" check.
Even back then, the idle_cpu() check would have been sufficient to catch
a new task being enqueued. Since commit b2a02fc43a ("smp: Optimize
send_call_function_single_ipi()") overloads the interpretation of
TIF_NEED_RESCHED for TIF_POLLING_NRFLAG idling, remove the
need_resched() check in nohz_csd_func() to raise SCHED_SOFTIRQ based
on Peter's suggestion.
Fixes: b2a02fc43a ("smp: Optimize send_call_function_single_ipi()")
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241119054432.6405-3-kprateek.nayak@amd.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 5f1b64e9a9 ]
[Problem Description]
When running the hackbench program of LTP, the following memory leak is
reported by kmemleak.
# /opt/ltp/testcases/bin/hackbench 20 thread 1000
Running with 20*40 (== 800) tasks.
# dmesg | grep kmemleak
...
kmemleak: 480 new suspected memory leaks (see /sys/kernel/debug/kmemleak)
kmemleak: 665 new suspected memory leaks (see /sys/kernel/debug/kmemleak)
# cat /sys/kernel/debug/kmemleak
unreferenced object 0xffff888cd8ca2c40 (size 64):
comm "hackbench", pid 17142, jiffies 4299780315
hex dump (first 32 bytes):
ac 74 49 00 01 00 00 00 4c 84 49 00 01 00 00 00 .tI.....L.I.....
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace (crc bff18fd4):
[<ffffffff81419a89>] __kmalloc_cache_noprof+0x2f9/0x3f0
[<ffffffff8113f715>] task_numa_work+0x725/0xa00
[<ffffffff8110f878>] task_work_run+0x58/0x90
[<ffffffff81ddd9f8>] syscall_exit_to_user_mode+0x1c8/0x1e0
[<ffffffff81dd78d5>] do_syscall_64+0x85/0x150
[<ffffffff81e0012b>] entry_SYSCALL_64_after_hwframe+0x76/0x7e
...
This issue can be consistently reproduced on three different servers:
* a 448-core server
* a 256-core server
* a 192-core server
[Root Cause]
Since multiple threads are created by the hackbench program (along with
the command argument 'thread'), a shared vma might be accessed by two or
more cores simultaneously. When two or more cores observe that
vma->numab_state is NULL at the same time, vma->numab_state will be
overwritten.
Although current code ensures that only one thread scans the VMAs in a
single 'numa_scan_period', there might be a chance for another thread
to enter in the next 'numa_scan_period' while we have not gotten till
numab_state allocation [1].
Note that the command `/opt/ltp/testcases/bin/hackbench 50 process 1000`
cannot the reproduce the issue. It is verified with 200+ test runs.
[Solution]
Use the cmpxchg atomic operation to ensure that only one thread executes
the vma->numab_state assignment.
[1] https://lore.kernel.org/lkml/1794be3c-358c-4cdc-a43d-a1f841d91ef7@amd.com/
Link: https://lkml.kernel.org/r/20241113102146.2384-1-ahuang12@lenovo.com
Fixes: ef6a22b70f ("sched/numa: apply the scan delay to every new vma")
Signed-off-by: Adrian Huang <ahuang12@lenovo.com>
Reported-by: Jiwei Sun <sunjw10@lenovo.com>
Reviewed-by: Raghavendra K T <raghavendra.kt@amd.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Ben Segall <bsegall@google.com>
Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 84db47ca71 ]
Since commit fc137c0dda ("sched/numa: enhance vma scanning logic")
NUMA Balancing allows updating PTEs to trap NUMA hinting faults if the
task had previously accessed VMA. However unconditional scan of VMAs are
allowed during initial phase of VMA creation until process's
mm numa_scan_seq reaches 2 even though current task had not accessed VMA.
Rationale:
- Without initial scan subsequent PTE update may never happen.
- Give fair opportunity to all the VMAs to be scanned and subsequently
understand the access pattern of all the VMAs.
But it has a corner case where, if a VMA is created after some time,
process's mm numa_scan_seq could be already greater than 2.
For e.g., values of mm numa_scan_seq when VMAs are created by running
mmtest autonuma benchmark briefly looks like:
start_seq=0 : 459
start_seq=2 : 138
start_seq=3 : 144
start_seq=4 : 8
start_seq=8 : 1
start_seq=9 : 1
This results in no unconditional PTE updates for those VMAs created after
some time.
Fix:
- Note down the initial value of mm numa_scan_seq in per VMA start_seq.
- Allow unconditional scan till start_seq + 2.
Result:
SUT: AMD EPYC Milan with 2 NUMA nodes 256 cpus.
base kernel: upstream 6.6-rc6 with Mels patches [1] applied.
kernbench
========== base patched %gain
Amean elsp-128 165.09 ( 0.00%) 164.78 * 0.19%*
Duration User 41404.28 41375.08
Duration System 9862.22 9768.48
Duration Elapsed 519.87 518.72
Ops NUMA PTE updates 1041416.00 831536.00
Ops NUMA hint faults 263296.00 220966.00
Ops NUMA pages migrated 258021.00 212769.00
Ops AutoNUMA cost 1328.67 1114.69
autonumabench
NUMA01_THREADLOCAL
==================
Amean elsp-NUMA01_THREADLOCAL 81.79 (0.00%) 67.74 * 17.18%*
Duration User 54832.73 47379.67
Duration System 75.00 185.75
Duration Elapsed 576.72 476.09
Ops NUMA PTE updates 394429.00 11121044.00
Ops NUMA hint faults 1001.00 8906404.00
Ops NUMA pages migrated 288.00 2998694.00
Ops AutoNUMA cost 7.77 44666.84
Signed-off-by: Raghavendra K T <raghavendra.kt@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Link: https://lore.kernel.org/r/2ea7cbce80ac7c62e90cbfb9653a7972f902439f.1697816692.git.raghavendra.kt@amd.com
Stable-dep-of: 5f1b64e9a9 ("sched/numa: fix memory leak due to the overwritten vma->numab_state")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9c70b2a33c ]
When running stress-ng-vm-segv test, we found a null pointer dereference
error in task_numa_work(). Here is the backtrace:
[323676.066985] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000020
......
[323676.067108] CPU: 35 PID: 2694524 Comm: stress-ng-vm-se
......
[323676.067113] pstate: 23401009 (nzCv daif +PAN -UAO +TCO +DIT +SSBS BTYPE=--)
[323676.067115] pc : vma_migratable+0x1c/0xd0
[323676.067122] lr : task_numa_work+0x1ec/0x4e0
[323676.067127] sp : ffff8000ada73d20
[323676.067128] x29: ffff8000ada73d20 x28: 0000000000000000 x27: 000000003e89f010
[323676.067130] x26: 0000000000080000 x25: ffff800081b5c0d8 x24: ffff800081b27000
[323676.067133] x23: 0000000000010000 x22: 0000000104d18cc0 x21: ffff0009f7158000
[323676.067135] x20: 0000000000000000 x19: 0000000000000000 x18: ffff8000ada73db8
[323676.067138] x17: 0001400000000000 x16: ffff800080df40b0 x15: 0000000000000035
[323676.067140] x14: ffff8000ada73cc8 x13: 1fffe0017cc72001 x12: ffff8000ada73cc8
[323676.067142] x11: ffff80008001160c x10: ffff000be639000c x9 : ffff8000800f4ba4
[323676.067145] x8 : ffff000810375000 x7 : ffff8000ada73974 x6 : 0000000000000001
[323676.067147] x5 : 0068000b33e26707 x4 : 0000000000000001 x3 : ffff0009f7158000
[323676.067149] x2 : 0000000000000041 x1 : 0000000000004400 x0 : 0000000000000000
[323676.067152] Call trace:
[323676.067153] vma_migratable+0x1c/0xd0
[323676.067155] task_numa_work+0x1ec/0x4e0
[323676.067157] task_work_run+0x78/0xd8
[323676.067161] do_notify_resume+0x1ec/0x290
[323676.067163] el0_svc+0x150/0x160
[323676.067167] el0t_64_sync_handler+0xf8/0x128
[323676.067170] el0t_64_sync+0x17c/0x180
[323676.067173] Code: d2888001 910003fd f9000bf3 aa0003f3 (f9401000)
[323676.067177] SMP: stopping secondary CPUs
[323676.070184] Starting crashdump kernel...
stress-ng-vm-segv in stress-ng is used to stress test the SIGSEGV error
handling function of the system, which tries to cause a SIGSEGV error on
return from unmapping the whole address space of the child process.
Normally this program will not cause kernel crashes. But before the
munmap system call returns to user mode, a potential task_numa_work()
for numa balancing could be added and executed. In this scenario, since the
child process has no vma after munmap, the vma_next() in task_numa_work()
will return a null pointer even if the vma iterator restarts from 0.
Recheck the vma pointer before dereferencing it in task_numa_work().
Fixes: 214dbc4281 ("sched: convert to vma iterator")
Signed-off-by: Shawn Wang <shawnwang@linux.alibaba.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org # v6.2+
Link: https://lkml.kernel.org/r/20241025022208.125527-1-shawnwang@linux.alibaba.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 73ab05aa46 ]
With KASAN and PREEMPT_RT enabled, calling task_work_add() in
task_tick_mm_cid() may cause the following splat.
[ 63.696416] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
[ 63.696416] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 610, name: modprobe
[ 63.696416] preempt_count: 10001, expected: 0
[ 63.696416] RCU nest depth: 1, expected: 1
This problem is caused by the following call trace.
sched_tick() [ acquire rq->__lock ]
-> task_tick_mm_cid()
-> task_work_add()
-> __kasan_record_aux_stack()
-> kasan_save_stack()
-> stack_depot_save_flags()
-> alloc_pages_mpol_noprof()
-> __alloc_pages_noprof()
-> get_page_from_freelist()
-> rmqueue()
-> rmqueue_pcplist()
-> __rmqueue_pcplist()
-> rmqueue_bulk()
-> rt_spin_lock()
The rq lock is a raw_spinlock_t. We can't sleep while holding
it. IOW, we can't call alloc_pages() in stack_depot_save_flags().
The task_tick_mm_cid() function with its task_work_add() call was
introduced by commit 223baf9d17 ("sched: Fix performance regression
introduced by mm_cid") in v6.4 kernel.
Fortunately, there is a kasan_record_aux_stack_noalloc() variant that
calls stack_depot_save_flags() while not allowing it to allocate
new pages. To allow task_tick_mm_cid() to use task_work without
page allocation, a new TWAF_NO_ALLOC flag is added to enable calling
kasan_record_aux_stack_noalloc() instead of kasan_record_aux_stack()
if set. The task_tick_mm_cid() function is modified to add this new flag.
The possible downside is the missing stack trace in a KASAN report due
to new page allocation required when task_work_add_noallloc() is called
which should be rare.
Fixes: 223baf9d17 ("sched: Fix performance regression introduced by mm_cid")
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20241010014432.194742-1-longman@redhat.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3840cbe24c ]
Brandon reports sporadic, non-sensical spikes in cumulative pressure
time (total=) when reading cpu.pressure at a high rate. This is due to
a race condition between reader aggregation and tasks changing states.
While it affects all states and all resources captured by PSI, in
practice it most likely triggers with CPU pressure, since scheduling
events are so frequent compared to other resource events.
The race context is the live snooping of ongoing stalls during a
pressure read. The read aggregates per-cpu records for stalls that
have concluded, but will also incorporate ad-hoc the duration of any
active state that hasn't been recorded yet. This is important to get
timely measurements of ongoing stalls. Those ad-hoc samples are
calculated on-the-fly up to the current time on that CPU; since the
stall hasn't concluded, it's expected that this is the minimum amount
of stall time that will enter the per-cpu records once it does.
The problem is that the path that concludes the state uses a CPU clock
read that is not synchronized against aggregators; the clock is read
outside of the seqlock protection. This allows aggregators to race and
snoop a stall with a longer duration than will actually be recorded.
With the recorded stall time being less than the last snapshot
remembered by the aggregator, a subsequent sample will underflow and
observe a bogus delta value, resulting in an erratic jump in pressure.
Fix this by moving the clock read of the state change into the seqlock
protection. This ensures no aggregation can snoop live stalls past the
time that's recorded when the state concludes.
Reported-by: Brandon Duffany <brandon@buildbuddy.io>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=219194
Link: https://lore.kernel.org/lkml/20240827121851.GB438928@cmpxchg.org/
Fixes: df77430639 ("psi: Reduce calls to sched_clock() in psi")
Cc: stable@vger.kernel.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit f22cde4371 ]
Problem statement:
Since commit fc137c0dda ("sched/numa: enhance vma scanning logic"), the
Numa vma scan overhead has been reduced a lot. Meanwhile, the reducing of
the vma scan might create less Numa page fault information. The
insufficient information makes it harder for the Numa balancer to make
decision. Later, commit b7a5b537c5 ("sched/numa: Complete scanning of
partial VMAs regardless of PID activity") and commit 84db47ca71
("sched/numa: Fix mm numa_scan_seq based unconditional scan") are found to
bring back part of the performance.
Recently when running SPECcpu omnetpp_r on a 320 CPUs/2 Sockets system, a
long duration of remote Numa node read was observed by PMU events: A few
cores having ~500MB/s remote memory access for ~20 seconds. It causes
high core-to-core variance and performance penalty. After the
investigation, it is found that many vmas are skipped due to the active
PID check. According to the trace events, in most cases,
vma_is_accessed() returns false because the history access info stored in
pids_active array has been cleared.
Proposal:
The main idea is to adjust vma_is_accessed() to let it return true easier.
Thus compare the diff between mm->numa_scan_seq and
vma->numab_state->prev_scan_seq. If the diff has exceeded the threshold,
scan the vma.
This patch especially helps the cases where there are small number of
threads, like the process-based SPECcpu. Without this patch, if the
SPECcpu process access the vma at the beginning, then sleeps for a long
time, the pid_active array will be cleared. A a result, if this process
is woken up again, it never has a chance to set prot_none anymore.
Because only the first 2 times of access is granted for vma scan:
(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2 to be
worse, no other threads within the task can help set the prot_none. This
causes information lost.
Raghavendra helped test current patch and got the positive result
on the AMD platform:
autonumabench NUMA01
base patched
Amean syst-NUMA01 194.05 ( 0.00%) 165.11 * 14.92%*
Amean elsp-NUMA01 324.86 ( 0.00%) 315.58 * 2.86%*
Duration User 380345.36 368252.04
Duration System 1358.89 1156.23
Duration Elapsed 2277.45 2213.25
autonumabench NUMA02
Amean syst-NUMA02 1.12 ( 0.00%) 1.09 * 2.93%*
Amean elsp-NUMA02 3.50 ( 0.00%) 3.56 * -1.84%*
Duration User 1513.23 1575.48
Duration System 8.33 8.13
Duration Elapsed 28.59 29.71
kernbench
Amean user-256 22935.42 ( 0.00%) 22535.19 * 1.75%*
Amean syst-256 7284.16 ( 0.00%) 7608.72 * -4.46%*
Amean elsp-256 159.01 ( 0.00%) 158.17 * 0.53%*
Duration User 68816.41 67615.74
Duration System 21873.94 22848.08
Duration Elapsed 506.66 504.55
Intel 256 CPUs/2 Sockets:
autonuma benchmark also shows improvements:
v6.10-rc5 v6.10-rc5
+patch
Amean syst-NUMA01 245.85 ( 0.00%) 230.84 * 6.11%*
Amean syst-NUMA01_THREADLOCAL 205.27 ( 0.00%) 191.86 * 6.53%*
Amean syst-NUMA02 18.57 ( 0.00%) 18.09 * 2.58%*
Amean syst-NUMA02_SMT 2.63 ( 0.00%) 2.54 * 3.47%*
Amean elsp-NUMA01 517.17 ( 0.00%) 526.34 * -1.77%*
Amean elsp-NUMA01_THREADLOCAL 99.92 ( 0.00%) 100.59 * -0.67%*
Amean elsp-NUMA02 15.81 ( 0.00%) 15.72 * 0.59%*
Amean elsp-NUMA02_SMT 13.23 ( 0.00%) 12.89 * 2.53%*
v6.10-rc5 v6.10-rc5
+patch
Duration User 1064010.16 1075416.23
Duration System 3307.64 3104.66
Duration Elapsed 4537.54 4604.73
The SPECcpu remote node access issue disappears with the patch applied.
Link: https://lkml.kernel.org/r/20240827112958.181388-1-yu.c.chen@intel.com
Fixes: fc137c0dda ("sched/numa: enhance vma scanning logic")
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Yujie Liu <yujie.liu@intel.com>
Reported-by: Xiaoping Zhou <xiaoping.zhou@intel.com>
Reviewed-and-tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: "Chen, Tim C" <tim.c.chen@intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Juri Lelli <juri.lelli@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raghavendra K T <raghavendra.kt@amd.com>
Cc: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit f169c62ff7 ]
VMAs are skipped if there is no recent fault activity but this represents
a chicken-and-egg problem as there may be no fault activity if the PTEs
are never updated to trap NUMA hints. There is an indirect reliance on
scanning to be forced early in the lifetime of a task but this may fail
to detect changes in phase behaviour. Force inactive VMAs to be scanned
when all other eligible VMAs have been updated within the same scan
sequence.
Test results in general look good with some changes in performance, both
negative and positive, depending on whether the additional scanning and
faulting was beneficial or not to the workload. The autonuma benchmark
workload NUMA01_THREADLOCAL was picked for closer examination. The workload
creates two processes with numerous threads and thread-local storage that
is zero-filled in a loop. It exercises the corner case where unrelated
threads may skip VMAs that are thread-local to another thread and still
has some VMAs that inactive while the workload executes.
The VMA skipping activity frequency with and without the patch:
6.6.0-rc2-sched-numabtrace-v1
=============================
649 reason=scan_delay
9,094 reason=unsuitable
48,915 reason=shared_ro
143,919 reason=inaccessible
193,050 reason=pid_inactive
6.6.0-rc2-sched-numabselective-v1
=============================
146 reason=seq_completed
622 reason=ignore_pid_inactive
624 reason=scan_delay
6,570 reason=unsuitable
16,101 reason=shared_ro
27,608 reason=inaccessible
41,939 reason=pid_inactive
Note that with the patch applied, the PID activity is ignored
(ignore_pid_inactive) to ensure a VMA with some activity is completely
scanned. In addition, a small number of VMAs are scanned when no other
eligible VMA is available during a single scan window (seq_completed).
The number of times a VMA is skipped due to no PID activity from the
scanning task (pid_inactive) drops dramatically. It is expected that
this will increase the number of PTEs updated for NUMA hinting faults
as well as hinting faults but these represent PTEs that would otherwise
have been missed. The tradeoff is scan+fault overhead versus improving
locality due to migration.
On a 2-socket Cascade Lake test machine, the time to complete the
workload is as follows;
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1
Min elsp-NUMA01_THREADLOCAL 174.22 ( 0.00%) 117.64 ( 32.48%)
Amean elsp-NUMA01_THREADLOCAL 175.68 ( 0.00%) 123.34 * 29.79%*
Stddev elsp-NUMA01_THREADLOCAL 1.20 ( 0.00%) 4.06 (-238.20%)
CoeffVar elsp-NUMA01_THREADLOCAL 0.68 ( 0.00%) 3.29 (-381.70%)
Max elsp-NUMA01_THREADLOCAL 177.18 ( 0.00%) 128.03 ( 27.74%)
The time to complete the workload is reduced by almost 30%:
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1 /
Duration User 91201.80 63506.64
Duration System 2015.53 1819.78
Duration Elapsed 1234.77 868.37
In this specific case, system CPU time was not increased but it's not
universally true.
From vmstat, the NUMA scanning and fault activity is as follows;
6.6.0-rc2 6.6.0-rc2
sched-numabtrace-v1 sched-numabselective-v1
Ops NUMA base-page range updates 64272.00 26374386.00
Ops NUMA PTE updates 36624.00 55538.00
Ops NUMA PMD updates 54.00 51404.00
Ops NUMA hint faults 15504.00 75786.00
Ops NUMA hint local faults % 14860.00 56763.00
Ops NUMA hint local percent 95.85 74.90
Ops NUMA pages migrated 1629.00 6469222.00
Both the number of PTE updates and hint faults is dramatically
increased. While this is superficially unfortunate, it represents
ranges that were simply skipped without the patch. As a result
of the scanning and hinting faults, many more pages were also
migrated but as the time to completion is reduced, the overhead
is offset by the gain.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Link: https://lore.kernel.org/r/20231010083143.19593-7-mgorman@techsingularity.net
Stable-dep-of: f22cde4371 ("sched/numa: Fix the vma scan starving issue")
Signed-off-by: Sasha Levin <sashal@kernel.org>
Pierre GondoisandGreg Kroah-Hartman
Vincent GuittotandGreg Kroah-Hartman
Andy ShevchenkoandGreg Kroah-Hartman
Ingo MolnarandGreg Kroah-Hartman
Harshit AgarwalandGreg Kroah-Hartman
Christian LoehleandGreg Kroah-Hartman
Chris MasonandGreg Kroah-Hartman
Elliot BermanandGreg Kroah-Hartman
Peter ZijlstraandGreg Kroah-Hartman
Aruna RamakrishnaandGreg Kroah-Hartman
zihan zhouandGreg Kroah-Hartman
Rafael J. WysockiandGreg Kroah-Hartman
Shrikanth HegdeandGreg Kroah-Hartman
Dietmar EggemannandGreg Kroah-Hartman
Jann HornandGreg Kroah-Hartman
Felix MoessbauerandGreg Kroah-Hartman
Zecheng LiandGreg Kroah-Hartman
Thomas GleixnerandGreg Kroah-Hartman
Suleiman SouhlalandGreg Kroah-Hartman
Sultan Alsawaf (unemployed)andGreg Kroah-Hartman
Wander Lairson CostaandGreg Kroah-Hartman
K Prateek NayakandGreg Kroah-Hartman
Adrian HuangandGreg Kroah-Hartman
Raghavendra K TandGreg Kroah-Hartman
Shawn WangandGreg Kroah-Hartman
Waiman LongandGreg Kroah-Hartman
Johannes WeinerandGreg Kroah-Hartman
Yujie LiuandGreg Kroah-Hartman
Mel GormanandGreg Kroah-Hartman