The reason for the clause in try_to_wake_up() is, per its comment, that
find_proxy_task()'s proxy_deactivate() is not always called with a cleared
p->blocked_on.
However, that seems silly and easily cured. Make sure to always call
proxy_deactivate() with a cleared p->blocked_on such that we might remove this
clause from the common wake-up path.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: John Stultz <jstultz@google.com>
Link: https://patch.msgid.link/20260526113322.244729903%40infradead.org
Add a new is_blocked flag to the task struct. This flag is set
by try_to_block_task() and cleared by ttwu_do_wakeup() and
tracks if the task is blocked.
Traditionally this would mirror !p->on_rq, however due things
like DELAY_DEQUEUE and PROXY_EXEC, this can diverge, so its
useful to manage separately.
Additionally with this, we might be able to get rid of the
p->se.sched_delayed (ab)use in the core code (eventually).
Taken whole cloth from Peter's email:
https://lore.kernel.org/lkml/20260501132143.GC1026330@noisy.programming.kicks-ass.net/
With a few additional p->is_blocked = 0 in a few cases where
we return current if blocked_on gets zeroed or there is
no owner. This may hint that these current special cases
might be dropped eventually.
This change also helps resolve wait-queue stalls seen with
proxy-execution. See previous patch attempts for details:
https://lore.kernel.org/lkml/20260430215103.2978955-2-jstultz@google.com/
Reported-by: Vineeth Pillai <vineethrp@google.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260512025635.2840817-7-jstultz@google.com
This patch adds logic so try_to_wake_up() will notice if we are
waking a task where blocked_on == PROXY_WAKING, and if necessary
dequeue the task so the wakeup will naturally return-migrate the
donor task back to a cpu it can run on.
This helps performance as we do the dequeue and wakeup under the
locks normally taken in the try_to_wake_up() and avoids having
to do proxy_force_return() from __schedule(), which has to
re-take similar locks and then force a pick again loop.
This was split out from the larger proxy patch, and
significantly reworked.
Credits for the original patch go to:
Peter Zijlstra (Intel) <peterz@infradead.org>
Juri Lelli <juri.lelli@redhat.com>
Valentin Schneider <valentin.schneider@arm.com>
Connor O'Brien <connoro@google.com>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260512025635.2840817-6-jstultz@google.com
The DL scheduler keeps the current task in the rbtree, since
the deadline value isn't usually chagned while the task is
runnable.
This results in set_next_task() and put_prev_task() being
simpler, but unfortunately this causes complexity elsewhere.
Specifically when update_curr_dl() updates the deadline, it has
to dequeue and then enqueue the task.
From put_prev_task_dl(), we first call update_curr_dl(), and
then call enqueue_pushable_dl_task().
However, with Proxy Exec this goes awry. Since when a mutex is
released, we might wake the waiting rq->donor. This will cause
put_prev_task() to be called on the donor to take it off the
cpu for return migration.
At that point, from put_prev_task_dl() the update_curr_dl()
logic will dequeue & enqueue the task, and the enqueue function
will call enqueue_pushable_dl_task() (since the task_current()
check won't prevent it). Then back up the callstack in
put_prev_task_dl() we'll end up calling
enqueue_pushable_dl_task() again, tripping the
!RB_EMPTY_NODE(&p->pushable_dl_tasks) warning.
So to avoid this, use Peter's suggested[1] approach, and add a
dl_rq->curr pointer that is set/cleared from set_next_task()/
put_prev_task(), which effectively tracks the rq->donor. We can
then use this to avoid adding the active donor to the pushable
list from enqueue_task_dl().
[1]: https://lore.kernel.org/lkml/20260304095123.GP606826@noisy.programming.kicks-ass.net/
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260512025635.2840817-4-jstultz@google.com
Historically, the prev value from __schedule() was the rq->curr.
This prev value is passed down through numerous functions, and
used in the class scheduler implementations. The fact that
prev was on_cpu until the end of __schedule(), meant it was
stable across the rq lock drops that the class->balance()
implementations often do.
However, with proxy-exec, the prev passed to functions called
by __schedule() is rq->donor, which may not be the same as
rq->curr and may not be on_cpu, this makes the prev value
potentially unstable across rq lock drops.
A recently found issue with proxy-exec, is when we begin doing
return migration from try_to_wake_up(), its possible we may be
waking up the rq->donor. When we do this, we proxy_resched_idle()
to put_prev_set_next() setting the rq->donor to rq->idle, allowing
the rq->donor to be return migrated and allowed to run.
This however runs into trouble, as on another cpu we might be in
the middle of calling __schedule(). Conceptually the rq lock is
held for the majority of the time, but in calling prev_balance()
its possible the class->balance() handler call may briefly drop the rq lock.
This opens a window for try_to_wake_up() to wake and return migrate the
rq->donor before the class logic reacquires the rq lock.
Unfortunately prev_balance() pass in a prev argument, to which we pass
rq->donor. However this prev value can now become stale and incorrect across a
rq lock drop.
So, to correct this, rework the prev_balance() call so that it does not take a
"prev" argument.
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260512025635.2840817-2-jstultz@google.com
Vineeth found came up with a test driver that could trip up
workqueue stalls. After fixing one issue this test found,
Vineeth reported the test was still failing.
Greatly simplified, a task that tries to take a mutex already
owned by another task that is sleeping, can hit a edge case in
the mutex_lock_common() case.
If the task fails to get the lock, calls into schedule, but gets
a spurious wakeup, it will find that it is first waiter, and
go into the mutex_optimistic_spin() logic. Though before calling
mutex_optimistic_spin(), we clear task blocked_on state, since
mutex_optimistic_spin() may call schedule() if need_resched() is
set.
After mutex_optimistic_spin() fails, we set blocked_on again,
restart the main mutex loop, try to take the lock and call into
schedule_preempt_disabled().
From there, with proxy-execution, we'll see the task is
blocked_on, follow the chain, see the owner is sleeping and
dequeue the waiting task from the runqueue.
This all sounds fine and reasonable. But what I had missed is
that in mutex_optimistic_spin(), not only do we call schedule()
but we set TASK_RUNNABLE right before doing so.
This is ok for that invocation of schedule(). But when we come
back we re-set the blocked_on we had just cleared, but we do not
re-set the task state to TASK_INTERRUPTIBLE/UNINTERRUPTIBLE.
This means we have a task that is blocked_on & TASK_RUNNABLE,
so when the proxy execution code dequeues the task, we are
in trouble since future wakeups will be shortcut by the
ttwu_state_match() check.
Thus, to avoid this, after mutex_optimistic_spin(), set the task
state back when we set blocked_on.
Many many thanks again to Vineeth for his very useful testing
driver that uncovered this long hidden bug, that I hadn't
tripped in all my testing! Very impressed with the problems he's
uncovered!
Reported-by: Vineeth Pillai <vineethrp@google.com>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Vineeth Pillai <vineethrp@google.com>
Link: https://patch.msgid.link/20260430215103.2978955-3-jstultz@google.com
update_tg_load_avg() is called once per leaf cfs_rq from the
__update_blocked_fair() walk that runs inside the NOHZ idle-balance
softirq, and again from update_load_avg() with UPDATE_TG. Its first
operation after the trivial early-outs is unconditionally:
now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
return;
Jakub ran into a system where nohz_idle_balance() was taking 75%
of a CPU (which is handling network traffic and doing many irq_exit_cpu
calls), with 35% of that CPU spent in update_load_avg, and 17% of the
CPU in sched_clock_cpu(), reading the TSC.
In a quick synthetic test, it looks like this patch reduces the
CPU use of sched_balance_update_blocked_averages by about 20%.
Switch the rate-limit to read rq_clock(rq_of(cfs_rq)) instead.
This eliminates the rdtsc, and uses a fairly fresh timestamp,
because all callers of update_tg_load_avg() and clear_tg_load_avg()
hold rq->lock and have called update_rq_clock(rq) within microseconds:
caller pre-state
__update_blocked_fair encloser did update_rq_clock(rq)
update_load_avg's three UPDATE_TG sites under rq->lock after enqueue/dequeue/update_curr
attach_/detach_entity_cfs_rq preceded by update_load_avg(...)
clear_tg_load_avg via offline path rq_clock_start_loop_update(rq) upfront
so rq->clock is fresh at every call. Since cfs_rqs are per-CPU
per-task_group, cfs_rq->last_update_tg_load_avg is always compared
against the same rq's clock; no cross-rq drift.
Signed-off-by: Rik van Riel <riel@surriel.com>
Assisted-by: Claude (Anthropic)
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260527110250.6a91718d@fangorn
Commit cd959a3562 ("sched_ext: Add a DL server for sched_ext tasks")
introduced an ext_server deadline server to protect sched_ext tasks from
fair/RT starvation, mirroring the existing fair_server.
Currently, both servers reserve their 50ms/1000ms bandwidth at boot,
regardless of whether a BPF scheduler is loaded. Unused bandwidth is
still reclaimed at runtime by other classes, but the static reservation
prevents the RT class from implicitly using that headroom when one of
the two classes is guaranteed to be empty.
A sysadmin can work around this by writing
/sys/kernel/debug/sched/{fair,ext}_server/cpu*/runtime, but that
requires manual action and not all systems expose debugfs.
A better approach is to make server bandwidth reservations dynamic: only
the scheduling policy that is currently active should register its
reservation, while the inactive one should not artificially hold
capacity (keeping both reservations only when the BPF scheduler is
running in partial mode):
+---------------------------------------------+-------------+------------+
| BPF scheduler state | fair server | ext server |
+---------------------------------------------+-------------+------------+
| not loaded (default boot) | reserved | none |
| loaded full mode (!SCX_OPS_SWITCH_PARTIAL) | none | reserved |
| loaded partial mode (SCX_OPS_SWITCH_PARTIAL)| reserved | reserved |
+---------------------------------------------+-------------+------------+
To achieve this, introduce an "attached/detached" state for each
deadline server, so the kernel can decide whether a server's bandwidth
should be accounted in global bandwidth tracking.
At boot, the system starts with only the fair server contributing to
bandwidth accounting. When a BPF scheduler is enabled, the ext server is
attached and may replace or complement the fair server depending on
whether full or partial mode is used. When sched_ext is disabled, the
system restores the previous deadline bandwidth values and behavior.
The transition logic ensures that switching between scheduling modes is
consistent and reversible, without losing runtime configuration or
requiring manual intervention.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://patch.msgid.link/20260526164420.638711-2-arighi@nvidia.com
Writing runtime or period via the per-CPU dl_server debugfs files
(/sys/kernel/debug/sched/{fair,ext}_server/cpu*/{runtime,period}) on an
offline CPU can trigger two distinct kernel issues:
1) Divide-by-zero in dl_server_apply_params():
Oops: divide error: 0000 [#1] SMP NOPTI
RIP: 0010:dl_server_apply_params+0x239/0x3a0
Call Trace:
sched_server_write_common.isra.0+0x21a/0x3c0
full_proxy_write+0x78/0xd0
vfs_write+0xe7/0x6e0
Both __dl_sub() and __dl_add() divide by cpus internally, which can be
0 once the CPU has been removed from any active root-domain span (this
has been latent since the debugfs interface was introduced).
2) WARN_ON_ONCE in dl_server_start():
WARNING: kernel/sched/deadline.c:1805 at dl_server_start+0x232/0x270
Commit ee6e44dfe6 ("sched/deadline: Stop dl_server before CPU goes
offline") added this check to catch enqueueing the server on an
offline rq.
There's no meaningful semantics for re-configuring the per-CPU dl_server
bandwidth while the CPU is offline, so simply reject the write with
-EBUSY so userspace gets a clear error.
Closes: https://lore.kernel.org/all/20260526092228.3B6891F00A3A@smtp.kernel.org/
Fixes: d741f297bc ("sched/fair: Fair server interface")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Tested-by: abaci-kreproducer <abaci@linux.alibaba.com>
Link: https://patch.msgid.link/20260526100502.575774-1-arighi@nvidia.com
Venkat Reported a boot kernel panic next-20260522. Git bisect pointed to
b5ea300a17 ("sched/cache: Make LLC id continuous")
Stacktrace points to llc_mask being null.
NIP [c000000000e58504] _find_first_bit+0x44/0x130
LR [c000000000e58500] _find_first_bit+0x40/0x130
Call Trace:
build_sched_domains+0xad8/0xe50
sched_init_smp+0xa8/0x164
kernel_init_freeable+0x250/0x370
ret_from_kernel_user_thread+0x14/0x1c
On powerpc, cpu_coregroup_mask is available only when the underlying
hardware support coregroup. In shared LPAR, QEMU guest or power9 etc
coregroup isn't supported. In such cases llc_mask was being referenced
when it was null leading to panic.
On powerpc, LLC is at SMT core level. So assumption that coregroup(MC)
domain point to LLC is wrong. Provide a way for archs to say where its
LLC is if it not at MC domain.
Fixes: b5ea300a17 ("sched/cache: Make LLC id continuous")
Closes: https://lore.kernel.org/all/51154de7-3700-4cb4-82f2-1b3a8fa427f7@linux.ibm.com/
Reported-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Co-developed-by: Chen, Yu C <yu.c.chen@intel.com>
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Tested-by: Venkat Rao Bagalkote <venkat88@linux.ibm.com>
Tested-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Link: https://patch.msgid.link/20260529075712.1181039-1-sshegde@linux.ibm.com
The reason for pick_next_task_fair() is the put/set optimization that
avoids touching the common ancestors. However, it is possible to
implement this in the put_prev_task() and set_next_task() calls as
used in put_prev_set_next_task().
Notably, put_prev_set_next_task() is the only site that:
- calls put_prev_task() with a .next argument;
- calls set_next_task() with .first = true.
This means that put_prev_task() can determine the common hierarchy and
stop there, and then set_next_task() can terminate where put_prev_task
stopped.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260511120628.057634261@infradead.org
With commit 50653216e4 ("sched: Add support to pick functions to
take rf") removing the balance callback, the pick_task() callback is
in charge of newidle balancing.
This means pick_task_fair() should do so too. This hasn't been a
problem in practise because pick_next_task_fair() is used. However,
since we'll be removing that one shortly, make sure pick_next_task()
is up to scratch.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260511120627.944705718@infradead.org
Commit c9d93a73ce ("sched/fair: Drop redundant RCU read lock in NOHZ
kick path") removed the rcu_read_lock()/unlock() pair from
set_cpu_sd_state_busy() and set_cpu_sd_state_idle() on the assumption
that all callers run in a safe context for rcu_dereference_all(): IRQs
disabled or cpus_write_lock() held.
That assumption is wrong for the CPU hotplug teardown path. When CPUs
are taken offline, set_cpu_sd_state_busy() is invoked via:
cpuhp/N kthread
cpuhp_thread_fun()
cpuhp_invoke_callback()
sched_cpu_deactivate()
nohz_balance_exit_idle()
set_cpu_sd_state_busy()
rcu_dereference_all(per_cpu(sd_llc, cpu))
The cpuhp kthread holds cpu_hotplug_lock (percpu-rwsem) but runs with
preemption and IRQs enabled. As a result, lockdep correctly reports a
suspicious RCU usage on CPU offline, e.g.:
# echo 0 > /sys/devices/system/cpu/cpu1/online
=============================
WARNING: suspicious RCU usage
-----------------------------
kernel/sched/fair.c:12793 suspicious rcu_dereference_check() usage!
...
2 locks held by cpuhp/1/20:
#0: (cpu_hotplug_lock){++++}-{0:0}, at: cpuhp_thread_fun+0x42/0x1ae
#1: (cpuhp_state-down){+.+.}-{0:0}, at: cpuhp_thread_fun+0x72/0x1ae
Call Trace:
lockdep_rcu_suspicious
nohz_balance_exit_idle
sched_cpu_deactivate
cpuhp_invoke_callback
cpuhp_thread_fun
smpboot_thread_fn
Fix this by adding RCU read lock coverage to the one caller that lacks
it: nohz_balance_exit_idle() in the CPU hotplug teardown.
The other callers (nohz_balancer_kick() and nohz_balance_enter_idle())
genuinely run with IRQs disabled, so they remain unchanged.
Fixes: c9d93a73ce ("sched/fair: Drop redundant RCU read lock in NOHZ kick path")
Closes: https://lore.kernel.org/all/38fe0a1d-1a48-435a-910a-c278024d9ac9@samsung.com/
Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260522092523.2046095-1-arighi@nvidia.com
Recent optimizations of sd->shared assignment moved to allocating a
single instance of per-CPU sched_domain_shared objects per s_data.
Recent optimizations to select_idle_capacity() moved the sd->shared
assignments to "sd_asym" domain when ASYM_CPUCAPACITY is detected but
cache-aware scheduling mandates the presence of "sd_llc_shared" to
compute and cache per-LLC statistics.
Use an "alloc_flags" union in sched_domain_shared to claim a
sched_domain_shared object per sched_domain. Allocation starts searching
for an available / matching sched_domain_shared instance from the first
CPU of sched_domain_span(sd) (sd can be sd_llc, or sd_asym). If the
shared object is claimed by another domain, the instance corresponding
to next CPU in the domain span is explored until a matching / available
instance is found.
In case of a single CPU in sched_domain_span(), the domain will be
degenerated and a temporary overlap of ->shared objects across different
domains is acceptable.
"alloc_flags" forms a union with "nr_idle_scan" and the stale flags are
left as is when the sd->shared is published. The expectation is for the
first load balancing instance to correct the value just like the current
behavior, except the initial value is no longer 0.
Originally-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>
Tested-by: Andrea Righi <arighi@nvidia.com>
RT migration is done aggressively. When a CPU schedules out a high
priority RT task for a lower priority task, it will look to see if there's
any RT tasks that are waiting to run on another CPU that is of higher
priority than the task this CPU is about to run. If it finds one, it will
pull that task over to the CPU and allow it to run there instead.
Normally, this pulling is done by looking at the RT overloaded mask (rto)
which contains all the CPUs in the scheduler domain with RT tasks that are
waiting to run due to a higher priority RT task currently running on their
CPU. The CPU that is about to schedule a lower priority task will grab the
rq lock of the overloaded CPU and move the RT task from that CPU's runqueue
to the local one and schedule the higher priority RT task.
This caused issues when a lot of CPUs would schedule a lower priority task
at the same time. They would all try to grab the same runqueue lock of
the CPU with the overloaded RT tasks. Only the first CPU that got in will
get that task. All the others would wait until they got the runqueue lock
and see there's nothing to pull and do nothing. On systems with lots of
CPUs, this caused a large latency (up to 500us) which is beyond what
PREEMPT_RT is to allow.
The solution to that was to create an RT_PUSH_IPI logic. When any CPU
wanted to pull a task, instead of grabbing the runqueue lock of the
overloaded CPU, it would start by sending an IPI to the overloaded CPU,
and that IPI handler would have the CPU with the waiting RT task do a push
instead. Then that handler would send an IPI to the next CPU with
overloaded RT tasks, and so on. Note, after the first CPU starts this
process, if another CPU wanted to do a pull, it would see that the process
has already begun and would only increment a counter to have the IPIs
continue again.
The RT_PUSH_IPI solved the latency problem with PREEMPT_RT but could cause
a new issue with non PREEMPT_RT. Namely, softirqs run in a threaded
context on PREEMPT_RT but they can run in an interrupt context in non-RT.
If an IPI lands on a CPU that has just woken up multiple RT tasks and the
current CPU is running a non RT or a low priority RT task, instead of
doing a push, it would simply do a schedule on that CPU. But if a softirq
was also executing on this CPU, the schedule would need to wait until the
softirq finished. Until then, the CPU would still be considered overloaded
as there are RT tasks still waiting to run on it.
A live lock occurred on a workload that was doing heavy networking traffic
on a large machine where the softirqs would run 500us out of 750us. And it
would also be waking up RT tasks, causing the RT pull logic to be
constantly executed.
When a softirq triggered on a CPU with RT tasks queued but not running
yet, and the other CPUs would see this CPU as being overloaded, they would
send an IPI over to it. The CPU would notice that the waiting RT tasks are
of higher priority than the currently running task and simply schedule
that CPU instead. But because the softirq was executing, before it could
schedule, it would receive another IPI to do the same. The amount of IPIs
would slow down the currently running softirq so much that before it could
return back to task context, it would execute another softirq never
allowing the CPU to schedule. This live locked that CPU.
As RT_PUSH_IPI was created to help PREEMPT_RT, make it default off if
PREEMPT_RT is not enabled.
Fixes: b6366f048e ("sched/rt: Use IPI to trigger RT task push migration instead of pulling")
Closes: https://lore.kernel.org/all/20260506235716.2530720-1-tj@kernel.org/
Reported-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20260515103740.25ccbed8@gandalf.local.home
K Prateek noticed we weren't setting the rq->next_class in
proxy_resched_idle(), when I was debugging an issue seen with
CONFIG_SCHED_PROXY_EXEC and some of Peter's new patches, and
suggested this fix.
So set rq->next_class when we temporarily switch the donor to
idle, so we don't accidentally call wakeup_preempt_fair()
with idle as the donor.
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260514234732.3170197-1-jstultz@google.com
Add to select_idle_capacity() the same SIS_UTIL-controlled idle-scan
mechanism, already used by select_idle_cpu(): when sched_feat(SIS_UTIL)
is enabled and the LLC domain has sched_domain_shared data, derive the
per-attempt scan limit from sd->shared->nr_idle_scan.
That bounds the walk on large LLCs: once nr_idle_scan is exhausted,
return the best CPU seen so far. The early exit is gated on
!has_idle_core so an active idle-core search (SMT with idle cores
reported by test_idle_cores()) isn't cut short before it gets a chance
to find one.
Co-developed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-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://patch.msgid.link/20260509180955.1840064-6-arighi@nvidia.com
When SD_ASYM_CPUCAPACITY load balancing considers pulling a misfit task,
capacity_of(dst_cpu) can overstate available compute if the SMT sibling is
busy: the core does not deliver its full nominal capacity.
If SMT is active and dst_cpu is not on a fully idle core, skip this
destination so we do not migrate a misfit expecting a capacity upgrade we
cannot actually provide.
Reported-by: Felix Abecassis <fabecassis@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260509180955.1840064-5-arighi@nvidia.com
On systems with asymmetric CPU capacity (e.g., ACPI/CPPC reporting
different per-core frequencies), the wakeup path uses
select_idle_capacity() and prioritizes idle CPUs with higher capacity
for better task placement. However, when those CPUs belong to SMT cores,
their effective capacity can be much lower than the nominal capacity
when the sibling thread is busy: SMT siblings compete for shared
resources, so a "high capacity" CPU that is idle but whose sibling is
busy does not deliver its full capacity. This effective capacity
reduction cannot be modeled by the static capacity value alone.
Introduce SMT awareness in the asym-capacity idle selection policy: when
SMT is active, always prefer fully-idle SMT cores over partially-idle
ones.
Prioritizing fully-idle SMT cores yields better task placement because
the effective capacity of partially-idle SMT cores is reduced; always
preferring them when available leads to more accurate capacity usage on
task wakeup.
On an SMT system with asymmetric CPU capacities (NVIDIA Vera Rubin),
SMT-aware idle selection has been shown to improve throughput by around
15-18% over NO_ASYM mainline and by around 60% over ASYM mainline, for
CPU-bound workloads (NVBLAS) running an amount of tasks equal to the
amount of SMT cores.
Reported-by: Felix Abecassis <fabecassis@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260511142502.3873984-1-arighi@nvidia.com
On asymmetric CPU capacity systems, the wakeup path uses
select_idle_capacity(), which scans the span of sd_asym_cpucapacity
rather than sd_llc.
The has_idle_cores hint however lives on sd_llc->shared, so the
wakeup-time read of has_idle_cores operates on an LLC-scoped blob while
the actual scan/decision spans the asym domain; nr_busy_cpus also lives
in the same shared sched_domain data, but it's never used in the asym
CPU capacity scenario.
Therefore, move the sched_domain_shared object to sd_asym_cpucapacity
whenever the CPU has a SD_ASYM_CPUCAPACITY_FULL ancestor and that
ancestor is non-overlapping (i.e., not built from SD_NUMA). In that case
the scope of has_idle_cores matches the scope of the wakeup scan.
Fall back to attaching the shared object to sd_llc in three cases:
1) plain symmetric systems (no SD_ASYM_CPUCAPACITY_FULL anywhere);
2) CPUs in an exclusive cpuset that carves out a symmetric capacity
island: has_asym is system-wide but those CPUs have no
SD_ASYM_CPUCAPACITY_FULL ancestor in their hierarchy and follow
the symmetric LLC path in select_idle_sibling();
3) exotic topologies where SD_ASYM_CPUCAPACITY_FULL lands on an
SD_NUMA-built domain. init_sched_domain_shared() keys the shared
blob off cpumask_first(span), which on overlapping NUMA domains
would alias unrelated spans onto the same blob. Keep the shared
object on the LLC there; select_idle_capacity() gracefully skips
the has_idle_cores preference when sd->shared is NULL.
While at it, also rename the per-CPU sd_llc_shared to sd_balance_shared,
as it is no longer strictly tied to the LLC.
Co-developed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Acked-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260516055850.1345932-1-arighi@nvidia.com
nohz_balancer_kick() is reached from sched_balance_trigger(), which is
called from sched_tick(). sched_tick() runs with IRQs disabled, so the
additional rcu_read_lock/unlock() used around sched_domain accesses in
this path is redundant. Rely on the existing IRQ-disabled context (and
the rcu_dereference_all() checking) instead.
The same applies to set_cpu_sd_state_idle(), called from the idle entry
path with IRQs disabled, and to set_cpu_sd_state_busy(), reachable via
nohz_balance_exit_idle() from two contexts: nohz_balancer_kick() (IRQs
disabled, as above) and sched_cpu_deactivate() (the CPUHP_AP_ACTIVE
teardown, which runs under cpus_write_lock(), so it cannot race with
sched-domain rebuilds). In both cases the rcu_dereference_all()
validation is sufficient.
No functional change intended.
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260509180955.1840064-2-arighi@nvidia.com
There is a use of sched_smt_active() and explicit use of sched_smt_present.
Remove the explicit usage for better code maintenance and readability.
Note that this differs slightly for update_idle_core. It used to call
static_branch_unlikely earlier and now it will call static_branch_likely.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Link: https://patch.msgid.link/20260515172456.542799-5-sshegde@linux.ibm.com
For fastpaths such as wakeup and load balance even minimal code additions
can add up. is_core_idle is accessed during load balance.
Other callsites of is_core_idle make sched_smt_active() check first.
Make the same check in should_we_balance.
Rest of access to cpu_smt_mask isn't in fastpath.
Note: Remove the stale comment above is_core_idle. Enqueue methods
of fair aren't close to it anymore.
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Link: https://patch.msgid.link/20260515172456.542799-4-sshegde@linux.ibm.com
Now, that cpu_smt_mask is defined as cpumask_of(cpu) for
CONFIG_SCHED_SMT=n, it is possible to get rid of the ifdeffery.
Effectively,
- This makes sched_smt_present is defined always
- cpumask_weight(cpumask_of(cpu)) == 1. So sched_smt_present_inc/dec
will never enable the sched_smt_present. Which is expected.
- Paths that were compile-time eliminated become runtime guarded
using static keys.
- Defines set_idle_cores, test_idle_cores, etc which could likely benefit
the CONFIG_SCHED_SMT=n systems to use the same optimizations within the
LLC at wakeups.
- This will expose sched_smt_present symbol for CONFIG_SCHED_SMT=n.
Likely not a concern.
- There is a bloat of code CONFIG_SCHED_SMT=n. (NR_CPUS=2048)
add/remove: 24/18 grow/shrink: 26/28 up/down: 6396/-3188 (3208)
Total: Before=30629880, After=30633088, chg +0.01%
- No code bloat for CONFIG_SCHED_SMT=y, which is expected.
- Add comments around stop_core_cpuslocked on why ifdefs are not
removed.
- This leaves the remaining uses of CONFIG_SCHED_SMT mainly for
topology building bits which has a policy based decision.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Phil Auld <pauld@redhat.com>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Acked-by: Tejun Heo <tj@kernel.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260515172456.542799-3-sshegde@linux.ibm.com
Commit 77baa5bafc ("sched/cputime: Fix mul_u64_u64_div_u64() precision
for cputime") added a clamp in cputime_adjust():
if (unlikely(stime > rtime))
stime = rtime;
The justification was that mul_u64_u64_div_u64() could over-approximate
on some architectures (notably arm64 and the old 32-bit fallback), so
the mathematically impossible stime > rtime was nevertheless reachable
and would underflow utime = rtime - stime.
That premise no longer holds. Commit b29a62d87c ("mul_u64_u64_div_u64:
make it precise always") replaced the fallback implementation with an
exact 128-bit long division, and the x86_64 inline asm already produced
exact results. The helper now returns the mathematically correct
floor(a*b/d) on every architecture, so stime <= rtime is guaranteed by
stime <= stime + utime and the clamp is dead code.
Remove it along with its stale comment.
Signed-off-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260514202629.673539-1-nico@fluxnic.net
Set the default total bandwidth for SCHED_DEADLINE tasks and servers
to ONE. FIFO/RR tasks are already throttled by fair-servers and
ext-servers, and the sysctl_sched_rt_runtime parameter now only
defines the total bw that is allowed to deadline entities.
Signed-off-by: Yuri Andriaccio <yurand2000@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260430213835.62217-22-yurand2000@gmail.com
sched_cache_present is a global static key, but build_sched_domains()
is called per partition from the "Build new domains" loop in
partition_sched_domains_locked(). Each call unconditionally sets the
key based solely on the has_multi_llcs local variable for that partition.
The call to the last partition set the value even when there
are previous partitions with multiple LLCs.
If partition A (multi-LLC) is built first, the key is enabled. Then
when partition B (single-LLC) is built, the key is disabled. The
multi-LLC partition A is still active but the key is now off.
Fix it by doing a similar thing as sched_energy_present: check the
multi-LLCs during the iteration over all the partitions rather than
checking it on a single partition.
This bug was reported by sashiko.
Fixes: d59f4fd1d3 ("sched/cache: Enable cache aware scheduling for multi LLCs NUMA node")
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/c541af2547d54509fbfd3b3a1e8072e2e5c7ff68.1778703694.git.tim.c.chen@linux.intel.com
If there are multiple LLCs in the system, cache aware scheduling
should be enabled. However, there is a corner case where, if there
is a single NUMA node and a single LLC per node, cache aware
scheduling will be turned on in the current implementation -
because at this moment, the parent domain has not yet been
degenerated, and it is possible that the current domain has the
same cpu span as its parent. There is no need to turn cache aware
scheduling on in this scenario.
Fix it by iterating the parent domains to find a domain that is
a superset of the current sd_llc, so that later, after the duplicated
parent domains have been degenerated, cache aware scheduling will
take effect.
For example, the expected behavior would be:
2 sockets, 1 LLC per socket: MC span=0-3, PKG span=0-7, has_multi_llcs=true
1 socket, 2 LLCs per socket: MC span=0-3, PKG span=0-7, has_multi_llcs=true
2 sockets, 2 LLCs per socket: MC span=0-3, PKG span=0-7, has_multi_llcs=true
1 socket, 1 LLC per socket: MC span=0-3, PKG span=0-3, has_multi_llcs=false
This bug was reported by sashiko.
Fixes: d59f4fd1d3 ("sched/cache: Enable cache aware scheduling for multi LLCs NUMA node")
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/6328a8a7f40925cec2a712d81ee58128a4c4444a.1778703694.git.tim.c.chen@linux.intel.com
sched_cache_active_set_unlocked() checks hardware support without
locks:
static void sched_cache_active_set(bool locked)
{
/* hardware does not support */
if (!static_branch_likely(&sched_cache_present)) {
_sched_cache_active_set(false, locked);
return;
}
...
If build_sched_domains() runs concurrently during CPU hotplug,
it can disable sched_cache_present under sched_domains_mutex
and the CPU hotplug lock. If a debugfs write thread evaluates
sched_cache_present as true right before that, and then blocks
or gets preempted, it might proceed to enable sched_cache_active
after the hardware support has been marked as absent. Make it
safer by acquiring cpus_read_lock() and sched_domains_mutex_lock()
when the user changes sched_cache_active via debugfs.
This bug was reported by sashiko.
Fixes: 067a313581 ("sched/cache: Allow the user space to turn on and off cache aware scheduling")
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/9afddf439687f04bb56b46625bd9f153eb8abad5.1778703694.git.tim.c.chen@linux.intel.com
There is a race condition that, after a task is enqueued
on a runqueue, task_llc(p) may change due to CPU hotplug,
because the llc_id is dynamically allocated and adjusted
at runtime.
Therefore, checking task_llc(p) to determine whether the
task is being dequeued from its preferred LLC is unreliable
and can cause inconsistent values.
To fix this problem, record whether p is enqueued on its
preferred LLC, in order to pair with account_llc_dequeue()
to maintain a consistent nr_pref_llc_running per runqueue.
This bug was reported by sashiko, and the solution was once
suggested by Prateek.
Fixes: 46afe3af7e ("sched/cache: Track LLC-preferred tasks per runqueue")
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/0c8c6a1571d66792a4d2ff0103ba3cc13e059046.1778703694.git.tim.c.chen@linux.intel.com
mm->sc_stat.cpu is written by task_cache_work() and could be read
locklessly by several functions on other CPUs. Use READ_ONCE and
WRITE_ONCE on mm->sc_stat.cpu access and write to prevent inconsistent
values from compiler optimizations when there are multiple accesses.
For example in get_pref_llc(), if the writer updated the field between
two compiler-generated loads, the validation (e.g., cpu != -1) and
subsequent use (e.g., llc_id(cpu)) could operate on different values,
allowing a negative CPU ID to be used as an index.
Leave plain write in mm_init_sched(), where the mm is not
yet visible to other CPUs.
This bug was reported by sashiko.
Fixes: 47d8696b95 ("sched/cache: Assign preferred LLC ID to processes")
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/63ea494f12efcf265d7134400a06cd75d7f2c310.1778703694.git.tim.c.chen@linux.intel.com
Introduce a set of debugfs knobs to control how aggressively the
cache aware scheduling does the task aggregation.
(1) aggr_tolerance
With sched_cache enabled, the scheduler uses a process's footprint
as a proxy for its LLC footprint to determine if aggregating tasks
on the preferred LLC could cause cache contention. If the footprint
exceeds the LLC size, aggregation is skipped. Since the kernel
cannot efficiently track per-task cache usage (resctrl is
user-space only), userspace can provide a more accurate hint.
Introduce /sys/kernel/debug/sched/llc_balancing/aggr_tolerance to
let users control how strictly footprint limits aggregation. Values
range from 0 to 100:
- 0: Cache-aware scheduling is disabled.
- 1: Strict; tasks with footprint larger than LLC size are skipped.
- >=100: Aggressive; tasks are aggregated regardless of footprint.
For example, with a 32MB L3 cache:
- aggr_tolerance=1 -> tasks with footprint > 32MB are skipped.
- aggr_tolerance=99 -> tasks with footprint > 784GB are skipped
(784GB = (1 + (99 - 1) * 256) * 32MB).
Similarly, /sys/kernel/debug/sched/llc_balancing/aggr_tolerance also
controls how strictly the number of active threads is considered when
doing cache aware load balance. The number of SMTs is also considered.
High SMT counts reduce the aggregation capacity, preventing excessive
task aggregation on SMT-heavy systems like Power10/Power11.
Yangyu suggested introducing separate aggregation controls for the
number of active threads and memory footprint checks. Since there are
plans to add per-process/task group controls, fine-grained tunables are
deferred to that implementation.
(2) epoch_period, epoch_affinity_timeout,
imb_pct, overaggr_pct are also turned into tunables.
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Suggested-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com>
Suggested-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Suggested-by: Tingyin Duan <tingyin.duan@gmail.com>
Suggested-by: Jianyong Wu <jianyong.wu@outlook.com>
Suggested-by: Yangyu Chen <cyy@cyyself.name>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Tingyin Duan <tingyin.duan@gmail.com>
Link: https://patch.msgid.link/1c62cc060ba2b33d7b1f0ed98b3390128edbae93.1778703694.git.tim.c.chen@linux.intel.com
Prateek and Tingyin reported that memory-intensive workloads (such as
stream) can saturate memory bandwidth and caches on the preferred LLC
when sched_cache aggregates too many threads.
To mitigate this, estimate a process's memory footprint by comparing
its NUMA balancing fault statistics to the size of the LLC. If the
footprint exceeds the LLC size, skip cache-aware scheduling.
Note that footprint is only an approximation of the memory footprint,
since the kernel lacks suitable metrics to estimate the real working
set. If a user-provided hint is available in the future, it would be
more accurate. A later patch will allow users to provide a hint to
adjust this threshold.
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Suggested-by: Vern Hao <vernhao@tencent.com>
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Tingyin Duan <tingyin.duan@gmail.com>
Link: https://patch.msgid.link/95cf64a385bcc12f18dcebe9d59e8d3ba8bb318f.1778703694.git.tim.c.chen@linux.intel.com
For a single thread, the current wakeup path tends to place it
on the same LLC where it was previously running with cache-hot
data. There is no need to enable cache-aware scheduling for
single-threaded processes for the following reasons:
1. Cache-aware scheduling primarily benefits multi-threaded
processes where threads share data. Single-threaded processes
typically have no inter-thread data sharing and thus gain little.
2. Enabling it incurs the additional overhead of tracking the
thread's residency in the LLCs.
3. Bypassing single-threaded processes avoids excessive
concentration of such tasks on a single LLC.
Nevertheless, this check can be omitted if users explicitly
provide hints for such single-threaded workloads where different
processes have shared memory, e.g., via prctl() or other interfaces
to be added in the future.
Signed-off-by: Chen Yu <yu.c.chen@intel.com>
Co-developed-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Tingyin Duan <tingyin.duan@gmail.com>
Link: https://patch.msgid.link/8a59a13aa58fdb48e410ecb2aabd97fe3ea5d256.1778703694.git.tim.c.chen@linux.intel.com
