Always trigger a resched after a protected period even if the entity is
still eligible. It can happen that an entity remains eligible at the end
of the protected period but must let an entity with a shorter slice to run
in order to keep its lag shorter than slice. This is particulalry true
with run to parity which tries to maximize the lag.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20250708165630.1948751-7-vincent.guittot@linaro.org
Run to parity ensures that current will get a chance to run its full
slice in one go but this can create large latency and/or lag for
entities with shorter slice that have exhausted their previous slice
and wait to run their next slice.
Clamp the run to parity to the shortest slice of all enqueued entities.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20250708165630.1948751-5-vincent.guittot@linaro.org
EEVDF expects the scheduler to allocate a time quantum to the selected
entity and then pick a new entity for next quantum.
Although this notion of time quantum is not strictly doable in our case,
we can ensure a minimum runtime for each task most of the time and pick a
new entity after a minimum time has elapsed.
Reuse the slice protection of run to parity to ensure such runtime
quantum.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20250708165630.1948751-3-vincent.guittot@linaro.org
Chris reported that commit 5f6bd380c7 ("sched/rt: Remove default
bandwidth control") caused a significant dip in his favourite
benchmark of the day. Simply disabling dl_server cured things.
His workload hammers the 0->1, 1->0 transitions, and the
dl_server_{start,stop}() overhead kills it -- fairly obviously a bad
idea in hind sight and all that.
Change things around to only disable the dl_server when there has not
been a fair task around for a whole period. Since the default period
is 1 second, this ensures the benchmark never trips this, overhead
gone.
Fixes: 557a6bfc66 ("sched/fair: Add trivial fair server")
Reported-by: Chris Mason <clm@meta.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lkml.kernel.org/r/20250702121158.465086194@infradead.org
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
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
In our testing with 6.12 based kernel on a big.LITTLE system, we were
seeing instances of RT tasks being blocked from running on the LITTLE
cpus for multiple seconds of time, apparently by the dl_server. This
far exceeds the default configured 50ms per second runtime.
This is due to the fair dl_server runtime calculation being scaled
for frequency & capacity of the cpu.
Consider the following case under a Big.LITTLE architecture:
Assume the runtime is: 50,000,000 ns, and Frequency/capacity
scale-invariance defined as below:
Frequency scale-invariance: 100
Capacity scale-invariance: 50
First by Frequency scale-invariance,
the runtime is scaled to 50,000,000 * 100 >> 10 = 4,882,812
Then by capacity scale-invariance,
it is further scaled to 4,882,812 * 50 >> 10 = 238,418.
So it will scaled to 238,418 ns.
This smaller "accounted runtime" value is what ends up being
subtracted against the fair-server's runtime for the current period.
Thus after 50ms of real time, we've only accounted ~238us against the
fair servers runtime. This 209:1 ratio in this example means that on
the smaller cpu the fair server is allowed to continue running,
blocking RT tasks, for over 10 seconds before it exhausts its supposed
50ms of runtime. And on other hardware configurations it can be even
worse.
For the fair deadline_server, to prevent realtime tasks from being
unexpectedly delayed, we really do want to use fixed time, and not
scaled time for smaller capacity/frequency cpus. So remove the scaling
from the fair server's accounting to fix this.
Fixes: a110a81c52 ("sched/deadline: Deferrable dl server")
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Suggested-by: John Stultz <jstultz@google.com>
Signed-off-by: kuyo chang <kuyo.chang@mediatek.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Acked-by: John Stultz <jstultz@google.com>
Tested-by: John Stultz <jstultz@google.com>
Link: https://lore.kernel.org/r/20250702021440.2594736-1-kuyo.chang@mediatek.com
On Mon, Jun 02, 2025 at 03:22:13PM +0800, Kuyo Chang wrote:
> So, the potential race scenario is:
>
> CPU0 CPU1
> // doing migrate_swap(cpu0/cpu1)
> stop_two_cpus()
> ...
> // doing _cpu_down()
> sched_cpu_deactivate()
> set_cpu_active(cpu, false);
> balance_push_set(cpu, true);
> cpu_stop_queue_two_works
> __cpu_stop_queue_work(stopper1,...);
> __cpu_stop_queue_work(stopper2,..);
> stop_cpus_in_progress -> true
> preempt_enable();
> ...
> 1st balance_push
> stop_one_cpu_nowait
> cpu_stop_queue_work
> __cpu_stop_queue_work
> list_add_tail -> 1st add push_work
> wake_up_q(&wakeq); -> "wakeq is empty.
> This implies that the stopper is at wakeq@migrate_swap."
> preempt_disable
> wake_up_q(&wakeq);
> wake_up_process // wakeup migrate/0
> try_to_wake_up
> ttwu_queue
> ttwu_queue_cond ->meet below case
> if (cpu == smp_processor_id())
> return false;
> ttwu_do_activate
> //migrate/0 wakeup done
> wake_up_process // wakeup migrate/1
> try_to_wake_up
> ttwu_queue
> ttwu_queue_cond
> ttwu_queue_wakelist
> __ttwu_queue_wakelist
> __smp_call_single_queue
> preempt_enable();
>
> 2nd balance_push
> stop_one_cpu_nowait
> cpu_stop_queue_work
> __cpu_stop_queue_work
> list_add_tail -> 2nd add push_work, so the double list add is detected
> ...
> ...
> cpu1 get ipi, do sched_ttwu_pending, wakeup migrate/1
>
So this balance_push() is part of schedule(), and schedule() is supposed
to switch to stopper task, but because of this race condition, stopper
task is stuck in WAKING state and not actually visible to be picked.
Therefore CPU1 can do another schedule() and end up doing another
balance_push() even though the last one hasn't been done yet.
This is a confluence of fail, where both wake_q and ttwu_wakelist can
cause crucial wakeups to be delayed, resulting in the malfunction of
balance_push.
Since there is only a single stopper thread to be woken, the wake_q
doesn't really add anything here, and can be removed in favour of
direct wakeups of the stopper thread.
Then add a clause to ttwu_queue_cond() to ensure the stopper threads
are never queued / delayed.
Of all 3 moving parts, the last addition was the balance_push()
machinery, so pick that as the point the bug was introduced.
Fixes: 2558aacff8 ("sched/hotplug: Ensure only per-cpu kthreads run during hotplug")
Reported-by: Kuyo Chang <kuyo.chang@mediatek.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Kuyo Chang <kuyo.chang@mediatek.com>
Link: https://lkml.kernel.org/r/20250605100009.GO39944@noisy.programming.kicks-ass.net
sched_ext performed a kfunc renaming pass in 6.13 and kept the old names
around for compatibility with old binaries. These were scheduled for
cleanup in 6.15 but were missed. Submitting for cleanup in for-next.
Removed the kfuncs, their flags, and any references I could find to them
in doc comments. Left the entries in include/scx/compat.bpf.h as they're
still useful to make new binaries compatible with old kernels.
Tested by applying to my kernel. It builds and a modern version of
scx_lavd loads fine.
Signed-off-by: Jake Hillion <jake@hillion.co.uk>
Signed-off-by: Tejun Heo <tj@kernel.org>
For systems using a sched_ext scheduler and has panic_on_rcu_stall
enabled, try kicking out the current scheduler before issuing a panic.
While there are numerous reasons for RCU CPU stalls that are not
directly attributed to the scheduler, deferring the panic gives
sched_ext an opportunity to provide additional debug info when ejecting
the current scheduler. Also, handling the event more gracefully allows
us to potentially recover the system instead of incurring additional
down time.
Suggested-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: David Dai <david.dai@linux.dev>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add a waitqueue helper to add a priority waiter that requires exclusive
wakeups, i.e. that requires that it be the _only_ priority waiter. The
API will be used by KVM to ensure that at most one of KVM's irqfds is
bound to a single eventfd (across the entire kernel).
Open code the helper instead of using __add_wait_queue() so that the
common path doesn't need to "handle" impossible failures.
Cc: K Prateek Nayak <kprateek.nayak@amd.com>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250522235223.3178519-9-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Drop the setting of WQ_FLAG_EXCLUSIVE from add_wait_queue_priority() and
instead have callers manually add the flag prior to adding their structure
to the queue. Blindly setting WQ_FLAG_EXCLUSIVE is flawed, as the nature
of exclusive, priority waiters means that only the first waiter added will
ever receive notifications.
Pushing the flawed behavior to callers will allow fixing the problem one
hypervisor at a time (KVM added the flawed API, and then KVM's code was
copy+pasted nearly verbatim by Xen and Hyper-V), and will also allow for
adding an API that provides true exclusivity, i.e. that guarantees at most
one priority waiter is in the queue.
Opportunistically add a comment in Hyper-V to call out the mess. Xen
privcmd's irqfd_wakefup() doesn't actually operate in exclusive mode, i.e.
can be "fixed" simply by dropping WQ_FLAG_EXCLUSIVE. And KVM is primed to
switch to the aforementioned fully exclusive API, i.e. won't be carrying
the flawed code for long.
No functional change intended.
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250522235223.3178519-7-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
From 077814f57f8acce13f91dc34bbd2b7e4911fbf25 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Fri, 13 Jun 2025 15:06:47 -1000
- Add CONFIG_GROUP_SCHED_BANDWIDTH which is selected by both
CONFIG_CFS_BANDWIDTH and EXT_GROUP_SCHED.
- Put bandwidth control interface files for both cgroup v1 and v2 under
CONFIG_GROUP_SCHED_BANDWIDTH.
- Update tg_bandwidth() to fetch configuration parameters from fair if
CONFIG_CFS_BANDWIDTH, SCX otherwise.
- Update tg_set_bandwidth() to update the parameters for both fair and SCX.
- Add bandwidth control parameters to struct scx_cgroup_init_args.
- Add sched_ext_ops.cgroup_set_bandwidth() which is invoked on bandwidth
control parameter updates.
- Update scx_qmap and maximal selftest to test the new feature.
Signed-off-by: Tejun Heo <tj@kernel.org>
More sched_ext fields will be added to struct task_group. In preparation,
factor out sched_ext fields into struct scx_task_group to reduce clutter in
the common header. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull sched_ext/for-6.16-fixes to receive:
c50784e99f ("sched_ext: Make scx_group_set_weight() always update tg->scx.weight")
33796b9187 ("sched_ext, sched/core: Don't call scx_group_set_weight() prematurely from sched_create_group()")
which are needed to implement CPU bandwidth control interface support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull tip/sched/core to receive the following commits:
d403a3689a ("sched/fair: Move max_cfs_quota_period decl and default_cfs_period() def from fair.c to sched.h")
de4c80c696 ("sched/core: Relocate tg_get_cfs_*() and cpu_cfs_*_read_*()")
43e33f53e2 ("sched/core: Reorganize cgroup bandwidth control interface file reads")
5bc34be478 ("sched/core: Reorganize cgroup bandwidth control interface file writes")
These will be used to implement CPU bandwidth interface support in
sched_ext.
Signed-off-by: Tejun Heo <tj@kernel.org>
- Move input parameter validation from tg_set_cfs_bandwidth() to the new
outer function tg_set_bandwidth(). The outer function handles parameters
in usecs, validates them and calls tg_set_cfs_bandwidth() which converts
them into nsecs. This matches tg_bandwidth() on the read side.
- max/min_cfs_* consts are now used by tg_set_bandwidth(). Relocate, convert
into usecs and drop "cfs" from the names.
- Reimplement cpu_cfs_{period|quote|burst}_write_*() using tg_bandwidth()
and tg_set_bandwidth() and replace "cfs" in the names with "bw".
- Update cpu_max_write() to use tg_set_bandiwdth(). cpu_period_quota_parse()
is updated to drop nsec conversion accordingly. This aligns the behavior
with cfs_period_quota_print().
- Drop now unused tg_set_cfs_{period|quota|burst}().
- While at it, for consistency, rename default_cfs_period() to
default_bw_period_us() and make it return usecs.
This is to prepare for adding bandwidth control support to sched_ext.
tg_set_bandwidth() will be used as the muxing point. No functional changes
intended.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250614012346.2358261-5-tj@kernel.org
- Update tg_get_cfs_*() to return u64 values. These are now used as the low
level accessors to the fair's bandwidth configuration parameters.
Translation to usecs takes place in these functions.
- Add tg_bandwidth() which reads all three bandwidth parameters using
tg_get_cfs_*().
- Reimplement cgroup interface read functions using tg_bandwidth(). Drop cfs
from the function names.
This is to prepare for adding bandwidth control support to sched_ext.
tg_bandwidth() will be used as the muxing point similar to tg_weight(). No
functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250614012346.2358261-4-tj@kernel.org
During task_group creation, sched_create_group() calls
scx_group_set_weight() with CGROUP_WEIGHT_DFL to initialize the sched_ext
portion. This is premature and ends up calling ops.cgroup_set_weight() with
an incorrect @cgrp before ops.cgroup_init() is called.
sched_create_group() should just initialize SCX related fields in the new
task_group. Fix it by factoring out scx_tg_init() from sched_init() and
making sched_create_group() call that function instead of
scx_group_set_weight().
v2: Retain CONFIG_EXT_GROUP_SCHED ifdef in sched_init() as removing it leads
to build failures on !CONFIG_GROUP_SCHED configs.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 8195136669 ("sched_ext: Add cgroup support")
Cc: stable@vger.kernel.org # v6.12+
Otherwise, tg->scx.weight can go out of sync while scx_cgroup is not enabled
and ops.cgroup_init() may be called with a stale weight value.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 8195136669 ("sched_ext: Add cgroup support")
Cc: stable@vger.kernel.org # v6.12+
Use NULL instead of 0 to signal no LLC domain, matching numa_span() and
the function comment.
No functional change.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Simplify the scheduler by making formerly SMP-only primitives and data
structures unconditional.
tj: Updated subject for clarity.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Simplify the scheduler by making formerly SMP-only primitives and data
structures unconditional.
tj: Updated subject for clarity. Fixed stray #else block which wasn't
removed causing build failure.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Simplify the scheduler by making formerly SMP-only primitives and data
structures unconditional.
tj: Updated subject for clarity. Replace #if defined() with #ifdef.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Simplify the scheduler by making formerly SMP-only primitives and data
structures unconditional.
tj: Updated subject for clarity.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
