Kernel test robot reported that
tools/testing/selftests/kvm/hardware_disable_test was failing due to
commit 704069649b ("sched/core: Rework sched_class::wakeup_preempt()
and rq_modified_*()")
It turns out there were two related problems that could lead to a
missed preemption:
- when hitting newidle balance from the idle thread, it would elevate
rb->next_class from &idle_sched_class to &fair_sched_class, causing
later wakeup_preempt() calls to not hit the sched_class_above()
case, and not issue resched_curr().
Notably, this modification pattern should only lower the
next_class, and never raise it. Create two new helper functions to
wrap this.
- when doing schedule_idle(), it was possible to miss (re)setting
rq->next_class to &idle_sched_class, leading to the very same
problem.
Cc: Sean Christopherson <seanjc@google.com>
Fixes: 704069649b ("sched/core: Rework sched_class::wakeup_preempt() and rq_modified_*()")
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202602122157.4e861298-lkp@intel.com
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260218163329.GQ1395416@noisy.programming.kicks-ass.net
Shinichiro reported a KASAN UAF, which is actually an out of bounds access
in the MMCID management code.
CPU0 CPU1
T1 runs in userspace
T0: fork(T4) -> Switch to per CPU CID mode
fixup() set MM_CID_TRANSIT on T1/CPU1
T4 exit()
T3 exit()
T2 exit()
T1 exit() switch to per task mode
---> Out of bounds access.
As T1 has not scheduled after T0 set the TRANSIT bit, it exits with the
TRANSIT bit set. sched_mm_cid_remove_user() clears the TRANSIT bit in
the task and drops the CID, but it does not touch the per CPU storage.
That's functionally correct because a CID is only owned by the CPU when
the ONCPU bit is set, which is mutually exclusive with the TRANSIT flag.
Now sched_mm_cid_exit() assumes that the CID is CPU owned because the
prior mode was per CPU. It invokes mm_drop_cid_on_cpu() which clears the
not set ONCPU bit and then invokes clear_bit() with an insanely large
bit number because TRANSIT is set (bit 29).
Prevent that by actually validating that the CID is CPU owned in
mm_drop_cid_on_cpu().
Fixes: 007d84287c ("sched/mmcid: Drop per CPU CID immediately when switching to per task mode")
Reported-by: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Tested-by: Shinichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Cc: stable@vger.kernel.org
Closes: https://lore.kernel.org/aYsZrixn9b6s_2zL@shinmob
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull x86 paravirt updates from Borislav Petkov:
- A nice cleanup to the paravirt code containing a unification of the
paravirt clock interface, taming the include hell by splitting the
pv_ops structure and removing of a bunch of obsolete code (Juergen
Gross)
* tag 'x86_paravirt_for_v7.0_rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (23 commits)
x86/paravirt: Use XOR r32,r32 to clear register in pv_vcpu_is_preempted()
x86/paravirt: Remove trailing semicolons from alternative asm templates
x86/pvlocks: Move paravirt spinlock functions into own header
x86/paravirt: Specify pv_ops array in paravirt macros
x86/paravirt: Allow pv-calls outside paravirt.h
objtool: Allow multiple pv_ops arrays
x86/xen: Drop xen_mmu_ops
x86/xen: Drop xen_cpu_ops
x86/xen: Drop xen_irq_ops
x86/paravirt: Move pv_native_*() prototypes to paravirt.c
x86/paravirt: Introduce new paravirt-base.h header
x86/paravirt: Move paravirt_sched_clock() related code into tsc.c
x86/paravirt: Use common code for paravirt_steal_clock()
riscv/paravirt: Use common code for paravirt_steal_clock()
loongarch/paravirt: Use common code for paravirt_steal_clock()
arm64/paravirt: Use common code for paravirt_steal_clock()
arm/paravirt: Use common code for paravirt_steal_clock()
sched: Move clock related paravirt code to kernel/sched
paravirt: Remove asm/paravirt_api_clock.h
x86/paravirt: Move thunk macros to paravirt_types.h
...
Pull scheduler updates from Ingo Molnar:
"Scheduler Kconfig space updates:
- Further consolidate configurable preemption modes (Peter Zijlstra)
Reduce the number of architectures that are allowed to offer
PREEMPT_NONE and PREEMPT_VOLUNTARY, reducing the number of
preemption models from four to just two: 'full' and 'lazy' on
up-to-date architectures (arm64, loongarch, powerpc, riscv, s390,
x86).
None and voluntary are only available as legacy features on
platforms that don't implement lazy preemption yet, or which don't
even support preemption.
The goal is to eventually remove cond_resched() and voluntary
preemption altogether.
RSEQ based 'scheduler time slice extension' support (Thomas Gleixner
and Peter Zijlstra):
This allows a thread to request a time slice extension when it enters
a critical section to avoid contention on a resource when the thread
is scheduled out inside of the critical section.
- Add fields and constants for time slice extension
- Provide static branch for time slice extensions
- Add statistics for time slice extensions
- Add prctl() to enable time slice extensions
- Implement sys_rseq_slice_yield()
- Implement syscall entry work for time slice extensions
- Implement time slice extension enforcement timer
- Reset slice extension when scheduled
- Implement rseq_grant_slice_extension()
- entry: Hook up rseq time slice extension
- selftests: Implement time slice extension test
- Allow registering RSEQ with slice extension
- Move slice_ext_nsec to debugfs
- Lower default slice extension
- selftests/rseq: Add rseq slice histogram script
Scheduler performance/scalability improvements:
- Update rq->avg_idle when a task is moved to an idle CPU, which
improves the scalability of various workloads (Shubhang Kaushik)
- Reorder fields in 'struct rq' for better caching (Blake Jones)
- Fair scheduler SMP NOHZ balancing code speedups (Shrikanth Hegde):
- Move checking for nohz cpus after time check
- Change likelyhood of nohz.nr_cpus
- Remove nohz.nr_cpus and use weight of cpumask instead
- Avoid false sharing for sched_clock_irqtime (Wangyang Guo)
- Cleanups (Yury Norov):
- Drop useless cpumask_empty() in find_energy_efficient_cpu()
- Simplify task_numa_find_cpu()
- Use cpumask_weight_and() in sched_balance_find_dst_group()
DL scheduler updates:
- Add a deadline server for sched_ext tasks (by Andrea Righi and Joel
Fernandes, with fixes by Peter Zijlstra)
RT scheduler updates:
- Skip currently executing CPU in rto_next_cpu() (Chen Jinghuang)
Entry code updates and performance improvements (Jinjie Ruan)
This is part of the scheduler tree in this cycle due to inter-
dependencies with the RSEQ based time slice extension work:
- Remove unused syscall argument from syscall_trace_enter()
- Rework syscall_exit_to_user_mode_work() for architecture reuse
- Add arch_ptrace_report_syscall_entry/exit()
- Inline syscall_exit_work() and syscall_trace_enter()
Scheduler core updates (Peter Zijlstra):
- Rework sched_class::wakeup_preempt() and rq_modified_*()
- Avoid rq->lock bouncing in sched_balance_newidle()
- Rename rcu_dereference_check_sched_domain() =>
rcu_dereference_sched_domain()
- <linux/compiler_types.h>: Add the __signed_scalar_typeof() helper
Fair scheduler updates/refactoring (Peter Zijlstra and Ingo Molnar):
- Fold the sched_avg update
- Change rcu_dereference_check_sched_domain() to rcu-sched
- Switch to rcu_dereference_all()
- Remove superfluous rcu_read_lock()
- Limit hrtick work
- Join two #ifdef CONFIG_FAIR_GROUP_SCHED blocks
- Clean up comments in 'struct cfs_rq'
- Separate se->vlag from se->vprot
- Rename cfs_rq::avg_load to cfs_rq::sum_weight
- Rename cfs_rq::avg_vruntime to ::sum_w_vruntime & helper functions
- Introduce and use the vruntime_cmp() and vruntime_op() wrappers for
wrapped-signed aritmetics
- Sort out 'blocked_load*' namespace noise
Scheduler debugging code updates:
- Export hidden tracepoints to modules (Gabriele Monaco)
- Convert copy_from_user() + kstrtouint() to kstrtouint_from_user()
(Fushuai Wang)
- Add assertions to QUEUE_CLASS (Peter Zijlstra)
- hrtimer: Fix tracing oddity (Thomas Gleixner)
Misc fixes and cleanups:
- Re-evaluate scheduling when migrating queued tasks out of throttled
cgroups (Zicheng Qu)
- Remove task_struct->faults_disabled_mapping (Christoph Hellwig)
- Fix math notation errors in avg_vruntime comment (Zhan Xusheng)
- sched/cpufreq: Use %pe format for PTR_ERR() printing
(zenghongling)"
* tag 'sched-core-2026-02-09' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (64 commits)
sched: Re-evaluate scheduling when migrating queued tasks out of throttled cgroups
sched/cpufreq: Use %pe format for PTR_ERR() printing
sched/rt: Skip currently executing CPU in rto_next_cpu()
sched/clock: Avoid false sharing for sched_clock_irqtime
selftests/sched_ext: Add test for DL server total_bw consistency
selftests/sched_ext: Add test for sched_ext dl_server
sched/debug: Fix dl_server (re)start conditions
sched/debug: Add support to change sched_ext server params
sched_ext: Add a DL server for sched_ext tasks
sched/debug: Stop and start server based on if it was active
sched/debug: Fix updating of ppos on server write ops
sched/deadline: Clear the defer params
entry: Inline syscall_exit_work() and syscall_trace_enter()
entry: Add arch_ptrace_report_syscall_entry/exit()
entry: Rework syscall_exit_to_user_mode_work() for architecture reuse
entry: Remove unused syscall argument from syscall_trace_enter()
sched: remove task_struct->faults_disabled_mapping
sched: Update rq->avg_idle when a task is moved to an idle CPU
selftests/rseq: Add rseq slice histogram script
hrtimer: Fix trace oddity
...
Pull locking updates from Ingo Molnar:
"Lock debugging:
- Implement compiler-driven static analysis locking context checking,
using the upcoming Clang 22 compiler's context analysis features
(Marco Elver)
We removed Sparse context analysis support, because prior to
removal even a defconfig kernel produced 1,700+ context tracking
Sparse warnings, the overwhelming majority of which are false
positives. On an allmodconfig kernel the number of false positive
context tracking Sparse warnings grows to over 5,200... On the plus
side of the balance actual locking bugs found by Sparse context
analysis is also rather ... sparse: I found only 3 such commits in
the last 3 years. So the rate of false positives and the
maintenance overhead is rather high and there appears to be no
active policy in place to achieve a zero-warnings baseline to move
the annotations & fixers to developers who introduce new code.
Clang context analysis is more complete and more aggressive in
trying to find bugs, at least in principle. Plus it has a different
model to enabling it: it's enabled subsystem by subsystem, which
results in zero warnings on all relevant kernel builds (as far as
our testing managed to cover it). Which allowed us to enable it by
default, similar to other compiler warnings, with the expectation
that there are no warnings going forward. This enforces a
zero-warnings baseline on clang-22+ builds (Which are still limited
in distribution, admittedly)
Hopefully the Clang approach can lead to a more maintainable
zero-warnings status quo and policy, with more and more subsystems
and drivers enabling the feature. Context tracking can be enabled
for all kernel code via WARN_CONTEXT_ANALYSIS_ALL=y (default
disabled), but this will generate a lot of false positives.
( Having said that, Sparse support could still be added back,
if anyone is interested - the removal patch is still
relatively straightforward to revert at this stage. )
Rust integration updates: (Alice Ryhl, Fujita Tomonori, Boqun Feng)
- Add support for Atomic<i8/i16/bool> and replace most Rust native
AtomicBool usages with Atomic<bool>
- Clean up LockClassKey and improve its documentation
- Add missing Send and Sync trait implementation for SetOnce
- Make ARef Unpin as it is supposed to be
- Add __rust_helper to a few Rust helpers as a preparation for
helper LTO
- Inline various lock related functions to avoid additional function
calls
WW mutexes:
- Extend ww_mutex tests and other test-ww_mutex updates (John
Stultz)
Misc fixes and cleanups:
- rcu: Mark lockdep_assert_rcu_helper() __always_inline (Arnd
Bergmann)
- locking/local_lock: Include more missing headers (Peter Zijlstra)
- seqlock: fix scoped_seqlock_read kernel-doc (Randy Dunlap)
- rust: sync: Replace `kernel::c_str!` with C-Strings (Tamir
Duberstein)"
* tag 'locking-core-2026-02-08' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (90 commits)
locking/rwlock: Fix write_trylock_irqsave() with CONFIG_INLINE_WRITE_TRYLOCK
rcu: Mark lockdep_assert_rcu_helper() __always_inline
compiler-context-analysis: Remove __assume_ctx_lock from initializers
tomoyo: Use scoped init guard
crypto: Use scoped init guard
kcov: Use scoped init guard
compiler-context-analysis: Introduce scoped init guards
cleanup: Make __DEFINE_LOCK_GUARD handle commas in initializers
seqlock: fix scoped_seqlock_read kernel-doc
tools: Update context analysis macros in compiler_types.h
rust: sync: Replace `kernel::c_str!` with C-Strings
rust: sync: Inline various lock related methods
rust: helpers: Move #define __rust_helper out of atomic.c
rust: wait: Add __rust_helper to helpers
rust: time: Add __rust_helper to helpers
rust: task: Add __rust_helper to helpers
rust: sync: Add __rust_helper to helpers
rust: refcount: Add __rust_helper to helpers
rust: rcu: Add __rust_helper to helpers
rust: processor: Add __rust_helper to helpers
...
Pull kthread updates from Frederic Weisbecker:
"The kthread code provides an infrastructure which manages the
preferred affinity of unbound kthreads (node or custom cpumask)
against housekeeping (CPU isolation) constraints and CPU hotplug
events.
One crucial missing piece is the handling of cpuset: when an isolated
partition is created, deleted, or its CPUs updated, all the unbound
kthreads in the top cpuset become indifferently affine to _all_ the
non-isolated CPUs, possibly breaking their preferred affinity along
the way.
Solve this with performing the kthreads affinity update from cpuset to
the kthreads consolidated relevant code instead so that preferred
affinities are honoured and applied against the updated cpuset
isolated partitions.
The dispatch of the new isolated cpumasks to timers, workqueues and
kthreads is performed by housekeeping, as per the nice Tejun's
suggestion.
As a welcome side effect, HK_TYPE_DOMAIN then integrates both the set
from boot defined domain isolation (through isolcpus=) and cpuset
isolated partitions. Housekeeping cpumasks are now modifiable with a
specific RCU based synchronization. A big step toward making
nohz_full= also mutable through cpuset in the future"
* tag 'kthread-for-7.0' of git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks: (33 commits)
doc: Add housekeeping documentation
kthread: Document kthread_affine_preferred()
kthread: Comment on the purpose and placement of kthread_affine_node() call
kthread: Honour kthreads preferred affinity after cpuset changes
sched/arm64: Move fallback task cpumask to HK_TYPE_DOMAIN
sched: Switch the fallback task allowed cpumask to HK_TYPE_DOMAIN
kthread: Rely on HK_TYPE_DOMAIN for preferred affinity management
kthread: Include kthreadd to the managed affinity list
kthread: Include unbound kthreads in the managed affinity list
kthread: Refine naming of affinity related fields
PCI: Remove superfluous HK_TYPE_WQ check
sched/isolation: Remove HK_TYPE_TICK test from cpu_is_isolated()
cpuset: Remove cpuset_cpu_is_isolated()
timers/migration: Remove superfluous cpuset isolation test
cpuset: Propagate cpuset isolation update to timers through housekeeping
cpuset: Propagate cpuset isolation update to workqueue through housekeeping
PCI: Flush PCI probe workqueue on cpuset isolated partition change
sched/isolation: Flush vmstat workqueues on cpuset isolated partition change
sched/isolation: Flush memcg workqueues on cpuset isolated partition change
cpuset: Update HK_TYPE_DOMAIN cpumask from cpuset
...
During the investigation of the various transition mode issues
instrumentation revealed that the amount of bitmap operations can be
significantly reduced when a task with a transitional CID schedules out
after the fixup function completed and disabled the transition mode.
At that point the mode is stable and therefore it is not required to drop
the transitional CID back into the pool. As the fixup is complete the
potential exhaustion of the CID pool is not longer possible, so the CID can
be transferred to the scheduling out task or to the CPU depending on the
current ownership mode.
The racy snapshot of mm_cid::mode which contains both the ownership state
and the transition bit is valid because runqueue lock is held and the fixup
function of a concurrent mode switch is serialized.
Assigning the ownership right there not only spares the bitmap access for
dropping the CID it also avoids it when the task is scheduled back in as it
directly hits the fast path in both modes when the CID is within the
optimal range. If it's outside the range the next schedule in will need to
converge so dropping it right away is sensible. In the good case this also
allows to go into the fast path on the next schedule in operation.
With a thread pool benchmark which is configured to cross the mode switch
boundaries frequently this reduces the number of bitmap operations by about
30% and increases the fastpath utilization in the low single digit
percentage range.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20260201192835.100194627@kernel.org
Shrikanth reported a hard lockup which he observed once. The stack trace
shows the following CID related participants:
watchdog: CPU 23 self-detected hard LOCKUP @ mm_get_cid+0xe8/0x188
NIP: mm_get_cid+0xe8/0x188
LR: mm_get_cid+0x108/0x188
mm_cid_switch_to+0x3c4/0x52c
__schedule+0x47c/0x700
schedule_idle+0x3c/0x64
do_idle+0x160/0x1b0
cpu_startup_entry+0x48/0x50
start_secondary+0x284/0x288
start_secondary_prolog+0x10/0x14
watchdog: CPU 11 self-detected hard LOCKUP @ plpar_hcall_norets_notrace+0x18/0x2c
NIP: plpar_hcall_norets_notrace+0x18/0x2c
LR: queued_spin_lock_slowpath+0xd88/0x15d0
_raw_spin_lock+0x80/0xa0
raw_spin_rq_lock_nested+0x3c/0xf8
mm_cid_fixup_cpus_to_tasks+0xc8/0x28c
sched_mm_cid_exit+0x108/0x22c
do_exit+0xf4/0x5d0
make_task_dead+0x0/0x178
system_call_exception+0x128/0x390
system_call_vectored_common+0x15c/0x2ec
The task on CPU11 is running the CID ownership mode change fixup function
and is stuck on a runqueue lock. The task on CPU23 is trying to get a CID
from the pool with the same runqueue lock held, but the pool is empty.
After decoding a similar issue in the opposite direction switching from per
task to per CPU mode the tool which models the possible scenarios failed to
come up with a similar loop hole.
This showed up only once, was not reproducible and according to tooling not
related to a overlooked scheduling scenario permutation. But the fact that
it was observed on a PowerPC system gave the right hint: PowerPC is a
weakly ordered architecture.
The transition mechanism does:
WRITE_ONCE(mm->mm_cid.transit, MM_CID_TRANSIT);
WRITE_ONCE(mm->mm_cid.percpu, new_mode);
fixup()
WRITE_ONCE(mm->mm_cid.transit, 0);
mm_cid_schedin() does:
if (!READ_ONCE(mm->mm_cid.percpu))
...
cid |= READ_ONCE(mm->mm_cid.transit);
so weakly ordered systems can observe percpu == false and transit == 0 even
if the fixup function has not yet completed. As a consequence the task will
not drop the CID when scheduling out before the fixup is completed, which
means the CID space can be exhausted and the next task scheduling in will
loop in mm_get_cid() and the fixup thread can livelock on the held runqueue
lock as above.
This could obviously be solved by using:
smp_store_release(&mm->mm_cid.percpu, true);
and
smp_load_acquire(&mm->mm_cid.percpu);
but that brings a memory barrier back into the scheduler hotpath, which was
just designed out by the CID rewrite.
That can be completely avoided by combining the per CPU mode and the
transit storage into a single mm_cid::mode member and ordering the stores
against the fixup functions to prevent the CPU from reordering them.
That makes the update of both states atomic and a concurrent read observes
always consistent state.
The price is an additional AND operation in mm_cid_schedin() to evaluate
the per CPU or the per task path, but that's in the noise even on strongly
ordered architectures as the actual load can be significantly more
expensive and the conditional branch evaluation is there anyway.
Fixes: fbd0e71dc3 ("sched/mmcid: Provide CID ownership mode fixup functions")
Closes: https://lore.kernel.org/bdfea828-4585-40e8-8835-247c6a8a76b0@linux.ibm.com
Reported-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20260201192834.965217106@kernel.org
Ihor reported a BPF CI failure which turned out to be a live lock in the
MM_CID management. The scenario is:
A test program creates the 5th thread, which means the MM_CID users become
more than the number of CPUs (four in this example), so it switches to per
CPU ownership mode.
At this point each live task of the program has a CID associated. Assume
thread creation order assignment for simplicity.
T0 CID0 runs fork() and creates T4
T1 CID1
T2 CID2
T3 CID3
T4 --- not visible yet
T0 sets mm_cid::percpu = true and transfers its own CID to CPU0 where it
runs on and then starts the fixup which walks through the threads to
transfer the per task CIDs either to the CPU the task is running on or drop
it back into the pool if the task is not on a CPU.
During that T1 - T3 are free to schedule in and out before the fixup caught
up with them. Going through all possible permutations with a python script
revealed a few problematic cases. The most trivial one is:
T1 schedules in on CPU1 and observes percpu == true, so it transfers
its CID to CPU1
T1 is migrated to CPU2 and schedule in observes percpu == true, but
CPU2 does not have a CID associated and T1 transferred its own to
CPU1
So it has to allocate one with CPU2 runqueue lock held, but the
pool is empty, so it keeps looping in mm_get_cid().
Now T0 reaches T1 in the thread walk and tries to lock the corresponding
runqueue lock, which is held causing a full live lock.
There is a similar scenario in the reverse direction of switching from per
CPU to task mode which is way more obvious and got therefore addressed by
an intermediate mode. In this mode the CIDs are marked with MM_CID_TRANSIT,
which means that they are neither owned by the CPU nor by the task. When a
task schedules out with a transit CID it drops the CID back into the pool
making it available for others to use temporarily. Once the task which
initiated the mode switch finished the fixup it clears the transit mode and
the process goes back into per task ownership mode.
Unfortunately this insight was not mapped back to the task to CPU mode
switch as the above described scenario was not considered in the analysis.
Apply the same transit mechanism to the task to CPU mode switch to handle
these problematic cases correctly.
As with the CPU to task transition this results in a potential temporary
contention on the CID bitmap, but that's only for the time it takes to
complete the transition. After that it stays in steady mode which does not
touch the bitmap at all.
Fixes: fbd0e71dc3 ("sched/mmcid: Provide CID ownership mode fixup functions")
Closes: https://lore.kernel.org/2b7463d7-0f58-4e34-9775-6e2115cfb971@linux.dev
Reported-by: Ihor Solodrai <ihor.solodrai@linux.dev>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20260201192834.897115238@kernel.org
The HK_TYPE_DOMAIN housekeeping cpumask is now modifiable at runtime.
In order to synchronize against vmstat workqueue to make sure
that no asynchronous vmstat work is still pending or executing on a
newly made isolated CPU, the housekeeping susbsystem must flush the
vmstat workqueues.
This involves flushing the whole mm_percpu_wq workqueue, shared with
LRU drain, introducing here a welcome side effect.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Marco Crivellari <marco.crivellari@suse.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Waiman Long <longman@redhat.com>
Cc: linux-mm@kvack.org
Until now, HK_TYPE_DOMAIN used to only include boot defined isolated
CPUs passed through isolcpus= boot option. Users interested in also
knowing the runtime defined isolated CPUs through cpuset must use
different APIs: cpuset_cpu_is_isolated(), cpu_is_isolated(), etc...
There are many drawbacks to that approach:
1) Most interested subsystems want to know about all isolated CPUs, not
just those defined on boot time.
2) cpuset_cpu_is_isolated() / cpu_is_isolated() are not synchronized with
concurrent cpuset changes.
3) Further cpuset modifications are not propagated to subsystems
Solve 1) and 2) and centralize all isolated CPUs within the
HK_TYPE_DOMAIN housekeeping cpumask.
Subsystems can rely on RCU to synchronize against concurrent changes.
The propagation mentioned in 3) will be handled in further patches.
[Chen Ridong: Fix cpu_hotplug_lock deadlock and use correct static
branch API]
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Waiman Long <longman@redhat.com>
Reviewed-by: Chen Ridong <chenridong@huawei.com>
Signed-off-by: Chen Ridong <chenridong@huawei.com>
Cc: "Michal Koutný" <mkoutny@suse.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Marco Crivellari <marco.crivellari@suse.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Waiman Long <longman@redhat.com>
Cc: cgroups@vger.kernel.org
HK_TYPE_DOMAIN's cpumask will soon be made modifiable by cpuset.
A synchronization mechanism is then needed to synchronize the updates
with the housekeeping cpumask readers.
Turn the housekeeping cpumasks into RCU pointers. Once a housekeeping
cpumask will be modified, the update side will wait for an RCU grace
period and propagate the change to interested subsystem when deemed
necessary.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Marco Crivellari <marco.crivellari@suse.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Waiman Long <longman@redhat.com>
Read-mostly sched_clock_irqtime may share the same cacheline with
frequently updated nohz struct. Make it as static_key to avoid
false sharing issue.
The only user of disable_sched_clock_irqtime()
is tsc_.*mark_unstable() which may be invoked under atomic context
and require a workqueue to disable static_key. But both of them
calls clear_sched_clock_stable() just before doing
disable_sched_clock_irqtime(). We can reuse
"sched_clock_work" to also disable sched_clock_irqtime().
One additional case need to handle is if the tsc is marked unstable
before late_initcall() phase, sched_clock_work will not be invoked
and sched_clock_irqtime will stay enabled although clock is unstable:
tsc_init()
enable_sched_clock_irqtime() # irqtime accounting is enabled here
...
if (unsynchronized_tsc()) # true
mark_tsc_unstable()
clear_sched_clock_stable()
__sched_clock_stable_early = 0;
...
if (static_key_count(&sched_clock_running.key) == 2)
# Only happens at sched_clock_init_late()
__clear_sched_clock_stable(); # Never executed
...
# late_initcall() phase
sched_clock_init_late()
if (__sched_clock_stable_early) # Already false
__set_sched_clock_stable(); # sched_clock is never marked stable
# TSC unstable, but sched_clock_work won't run to disable irqtime
So we need to disable_sched_clock_irqtime() in sched_clock_init_late()
if clock is unstable.
Reported-by: Benjamin Lei <benjamin.lei@intel.com>
Suggested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Suggested-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Wangyang Guo <wangyang.guo@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Tianyou Li <tianyou.li@intel.com>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260127072509.2627346-1-wangyang.guo@intel.com
sched_ext currently suffers starvation due to RT. The same workload when
converted to EXT can get zero runtime if RT is 100% running, causing EXT
processes to stall. Fix it by adding a DL server for EXT.
A kselftest is also included later to confirm that both DL servers are
functioning correctly:
# ./runner -t rt_stall
===== START =====
TEST: rt_stall
DESCRIPTION: Verify that RT tasks cannot stall SCHED_EXT tasks
OUTPUT:
TAP version 13
1..1
# Runtime of FAIR task (PID 1511) is 0.250000 seconds
# Runtime of RT task (PID 1512) is 4.750000 seconds
# FAIR task got 5.00% of total runtime
ok 1 PASS: FAIR task got more than 4.00% of runtime
TAP version 13
1..1
# Runtime of EXT task (PID 1514) is 0.250000 seconds
# Runtime of RT task (PID 1515) is 4.750000 seconds
# EXT task got 5.00% of total runtime
ok 2 PASS: EXT task got more than 4.00% of runtime
TAP version 13
1..1
# Runtime of FAIR task (PID 1517) is 0.250000 seconds
# Runtime of RT task (PID 1518) is 4.750000 seconds
# FAIR task got 5.00% of total runtime
ok 3 PASS: FAIR task got more than 4.00% of runtime
TAP version 13
1..1
# Runtime of EXT task (PID 1521) is 0.250000 seconds
# Runtime of RT task (PID 1522) is 4.750000 seconds
# EXT task got 5.00% of total runtime
ok 4 PASS: EXT task got more than 4.00% of runtime
ok 1 rt_stall #
===== END =====
Co-developed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
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: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/20260126100050.3854740-5-arighi@nvidia.com
Currently, rq->idle_stamp is only used to calculate avg_idle during
wakeups. This means other paths that move a task to an idle CPU such as
fork/clone, execve, or migrations, do not end the CPU's idle status in
the scheduler's eyes, leading to an inaccurate avg_idle.
This patch introduces update_rq_avg_idle() to provide a more accurate
measurement of CPU idle duration. By invoking this helper in
put_prev_task_idle(), we ensure avg_idle is updated whenever a CPU
stops being idle, regardless of how the new task arrived.
Testing on an 80-core Ampere Altra (ARMv8) with 6.19-rc5 baseline:
- Hackbench : +7.2% performance gain at 16 threads.
- Schbench: Reduced p99.9 tail latencies at high concurrency.
Signed-off-by: Shubhang Kaushik <shubhang@os.amperecomputing.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Shubhang Kaushik <shubhang@os.amperecomputing.com>
Link: https://patch.msgid.link/20260121-v8-patch-series-v8-1-b7f1cbee5055@os.amperecomputing.com
A fix for the dl_server 'requires' idle_cpu() usage, which made me
note that it and available_idle_cpu() are extern function calls.
And while idle_cpu() is used outside of kernel/sched/,
available_idle_cpu() is not.
This makes it hard to make idle_cpu() an inline helper, so provide
idle_rq() and implement idle_cpu() and available_idle_cpu() using
that.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Paravirt clock related functions are available in multiple archs.
In order to share the common parts, move the common static keys
to kernel/sched/ and remove them from the arch specific files.
Make a common paravirt_steal_clock() implementation available in
kernel/sched/cputime.c, guarding it with a new config option
CONFIG_HAVE_PV_STEAL_CLOCK_GEN, which can be selected by an arch
in case it wants to use that common variant.
Signed-off-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260105110520.21356-7-jgross@suse.com
All architectures supporting CONFIG_PARAVIRT share the same contents
of asm/paravirt_api_clock.h:
#include <asm/paravirt.h>
So remove all incarnations of asm/paravirt_api_clock.h and remove the
only place where it is included, as there asm/paravirt.h is included
anyway.
Signed-off-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com> # powerpc, scheduler bits
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260105110520.21356-6-jgross@suse.com
This colocates some hot fields in "struct rq" to be on the same cache line
as others that are often accessed at the same time or in similar ways.
Using data from a Google-internal fleet-scale profiler, I found three
distinct groups of hot fields in struct rq:
- (1) The runqueue lock: __lock.
- (2) Those accessed from hot code in pick_next_task_fair():
nr_running, nr_numa_running, nr_preferred_running,
ttwu_pending, cpu_capacity, curr, idle.
- (3) Those accessed from some other hot codepaths, e.g.
update_curr(), update_rq_clock(), and scheduler_tick():
clock_task, clock_pelt, clock, lost_idle_time,
clock_update_flags, clock_pelt_idle, clock_idle.
The cycles spent on accessing these different groups of fields broke down
roughly as follows:
- 50% on group (1) (the runqueue lock, always read-write)
- 39% on group (2) (load:store ratio around 38:1)
- 8% on group (3) (load:store ratio around 5:1)
- 3% on all the other fields
Most of the fields in group (3) are already in a cache line grouping; this
patch just adds "clock" and "clock_update_flags" to that group. The fields
in group (2) are scattered across several cache lines; the main effect of
this patch is to group them together, on a single line at the beginning of
the structure. A few other less performance-critical fields (nr_switches,
numa_migrate_on, has_blocked_load, nohz_csd, last_blocked_load_update_tick)
were also reordered to reduce holes in the data structure.
Since the runqueue lock is acquired from so many different contexts, and is
basically always accessed using an atomic operation, putting it on either
of the cache lines for groups (2) or (3) would slow down accesses to those
fields dramatically, since those groups are read-mostly accesses.
To test this, I wrote a focused load test that would put load on the
pick_next_task_fair() path. A parent process would fork many child
processes, and each child would nanosleep() for 1 msec many times in a
loop. The load test was monitored with "perf", and I looked at the amount
of cycles that were spent with sched_balance_rq() on the stack. The test
was reliably spending ~5% of all of its cycles there. I ran it 100 times
on a pair of 2-socket Intel Haswell machines (72 vCPUs per machine) - one
running the tip of sched/core, the other running this change - using 360
child processes and 8192 1-msec sleeps per child. The mean cycle count
dropped from 5.14B to 4.91B, or a *4.6% decrease* in relevant scheduler
cycles.
Given that this change reduces cache misses in a very hot kernel codepath,
there's likely to be additional application performance improvement due to
reduced cache conflicts from kernel data structures.
On a Power11 system with 128-byte cache lines, my test showed a ~5%
decrease in relevant scheduler cycles, along with a slight increase in user
time - both positive indicators. This data comes from
https://lore.kernel.org/lkml/affdc6b1-9980-44d1-89db-d90730c1e384@linux.ibm.com/
This is the case even though the additional "____cacheline_aligned" that
puts the runqueue lock on the next cache line adds an additional 64 bytes
of padding on those machines. This patch does not change the size of
"struct rq" on machines with 64-byte cache lines.
I also ran "hackbench" to try to test this change, but it didn't show
conclusive results. Looking at a CPU cycle profile of the hackbench run,
it was spending 95% of its cycles inside __alloc_skb(), __kfree_skb(), or
kmem_cache_free() - almost all of which was spent updating memcg counters
or contending on the list_lock in kmem_cache_node. And it spent less than
0.5% of its cycles inside either schedule() or try_to_wake_up(). So it's
not surprising that it didn't show useful results here.
The "__no_randomize_layout" was added to reflect the fact that performance
of code that references this data structure is unusually sensitive to
placement of its members.
Signed-off-by: Blake Jones <blakejones@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com>
Reviewed-by: Josh Don <joshdon@google.com>
Tested-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com>
Link: https://patch.msgid.link/20251202023743.1524247-1-blakejones@google.com
This demonstrates a larger conversion to use Clang's context
analysis. The benefit is additional static checking of locking rules,
along with better documentation.
Notably, kernel/sched contains sufficiently complex synchronization
patterns, and application to core.c & fair.c demonstrates that the
latest Clang version has become powerful enough to start applying this
to more complex subsystems (with some modest annotations and changes).
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20251219154418.3592607-37-elver@google.com
Change sched_class::wakeup_preempt() to also get called for
cross-class wakeups, specifically those where the woken task
is of a higher class than the previous highest class.
In order to do this, track the current highest class of the runqueue
in rq::next_class and have wakeup_preempt() track this upwards for
each new wakeup. Additionally have schedule() re-set the value on
pick.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://patch.msgid.link/20251127154725.901391274@infradead.org
The ::avg_vruntime field is a misnomer: it says it's an
'average vruntime', but in reality it's the momentary sum
of the weighted vruntimes of all queued tasks, which is
at least a division away from being an average.
This is clear from comments about the math of fair scheduling:
* \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
This confusion is increased by the cfs_avg_vruntime() function,
which does perform the division and returns a true average.
The sum of all weighted vruntimes should be named thusly,
so rename the field to ::sum_w_vruntime. (As arguably
::sum_weighted_vruntime would be a bit of a mouthful.)
Understanding the scheduler is hard enough already, without
extra layers of obfuscated naming. ;-)
Also rename related helper functions:
sum_vruntime_add() => sum_w_vruntime_add()
sum_vruntime_sub() => sum_w_vruntime_sub()
sum_vruntime_update() => sum_w_vruntime_update()
With the notable exception of cfs_avg_vruntime(), which
was named accurately.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://patch.msgid.link/20251201064647.1851919-7-mingo@kernel.org
The ::avg_load field is a long-standing misnomer: it says it's an
'average load', but in reality it's the momentary sum of the load
of all currently runnable tasks. We'd have to also perform a
division by nr_running (or use time-decay) to arrive at any sort
of average value.
This is clear from comments about the math of fair scheduling:
* \Sum w_i := cfs_rq->avg_load
The sum of all weights is ... the sum of all weights, not
the average of all weights.
To make it doubly confusing, there's also an ::avg_load
in the load-balancing struct sg_lb_stats, which *is* a
true average.
The second part of the field's name is a minor misnomer
as well: it says 'load', and it is indeed a load_weight
structure as it shares code with the load-balancer - but
it's only in an SMP load-balancing context where
load = weight, in the fair scheduling context the primary
purpose is the weighting of different nice levels.
So rename the field to ::sum_weight instead, which makes
the terminology of the EEVDF math match up with our
implementation of it:
* \Sum w_i := cfs_rq->sum_weight
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://patch.msgid.link/20251201064647.1851919-6-mingo@kernel.org
Join two identical #ifdef blocks:
#ifdef CONFIG_FAIR_GROUP_SCHED
...
#endif
#ifdef CONFIG_FAIR_GROUP_SCHED
...
#endif
Also mark nested #ifdef blocks in the usual fashion, to make
it more apparent where in a nested hierarchy of #ifdefs we
are at a glance.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Link: https://patch.msgid.link/20251201064647.1851919-2-mingo@kernel.org
With the {rcu,sched,bh} RCU flavours being unified, it doesn't really
make sense to check for just the rcu one. Switch to the _all family of
verification which includes all 3 of the listed flavours.
Notably, this will enable us to remove some superfluous
rcu_read_lock() regions when we know they are inside preempt/IRQ
disabled regions.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Remove check from the name for being surplus to requirements.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull scheduler fixes from Ingo Molnar:
"Miscellaneous scheduler fixes/cleanups:
- Fix psi_dequeue() for Proxy Execution
- Fix hrtick() vs. scheduling context bug
- Fix unfairness caused by stalled tg_load_avg_contrib when the last
task migrates out
- Fix whitespace noise in headers
- Remove a preempt-disable section in rt_mutex_setprio()"
* tag 'sched-urgent-2025-12-06' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/core: Fix psi_dequeue() for Proxy Execution
sched/fair: Fix unfairness caused by stalled tg_load_avg_contrib when the last task migrates out
sched/rt: Remove a preempt-disable section in rt_mutex_setprio()
sched/hrtick: Fix hrtick() vs. scheduling context
sched/headers: Remove whitespace noise from kernel/sched/sched.h
Pull sched_ext updates from Tejun Heo:
- Improve recovery from misbehaving BPF schedulers.
When a scheduler puts many tasks with varying affinity restrictions
on a shared DSQ, CPUs scanning through tasks they cannot run can
overwhelm the system, causing lockups.
Bypass mode now uses per-CPU DSQs with a load balancer to avoid this,
and hooks into the hardlockup detector to attempt recovery.
Add scx_cpu0 example scheduler to demonstrate this scenario.
- Add lockless peek operation for DSQs to reduce lock contention for
schedulers that need to query queue state during load balancing.
- Allow scx_bpf_reenqueue_local() to be called from anywhere in
preparation for deprecating cpu_acquire/release() callbacks in favor
of generic BPF hooks.
- Prepare for hierarchical scheduler support: add
scx_bpf_task_set_slice() and scx_bpf_task_set_dsq_vtime() kfuncs,
make scx_bpf_dsq_insert*() return bool, and wrap kfunc args in
structs for future aux__prog parameter.
- Implement cgroup_set_idle() callback to notify BPF schedulers when a
cgroup's idle state changes.
- Fix migration tasks being incorrectly downgraded from
stop_sched_class to rt_sched_class across sched_ext enable/disable.
Applied late as the fix is low risk and the bug subtle but needs
stable backporting.
- Various fixes and cleanups including cgroup exit ordering,
SCX_KICK_WAIT reliability, and backward compatibility improvements.
* tag 'sched_ext-for-6.19' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext: (44 commits)
sched_ext: Fix incorrect sched_class settings for per-cpu migration tasks
sched_ext: tools: Removing duplicate targets during non-cross compilation
sched_ext: Use kvfree_rcu() to release per-cpu ksyncs object
sched_ext: Pass locked CPU parameter to scx_hardlockup() and add docs
sched_ext: Update comments replacing breather with aborting mechanism
sched_ext: Implement load balancer for bypass mode
sched_ext: Factor out abbreviated dispatch dequeue into dispatch_dequeue_locked()
sched_ext: Factor out scx_dsq_list_node cursor initialization into INIT_DSQ_LIST_CURSOR
sched_ext: Add scx_cpu0 example scheduler
sched_ext: Hook up hardlockup detector
sched_ext: Make handle_lockup() propagate scx_verror() result
sched_ext: Refactor lockup handlers into handle_lockup()
sched_ext: Make scx_exit() and scx_vexit() return bool
sched_ext: Exit dispatch and move operations immediately when aborting
sched_ext: Simplify breather mechanism with scx_aborting flag
sched_ext: Use per-CPU DSQs instead of per-node global DSQs in bypass mode
sched_ext: Refactor do_enqueue_task() local and global DSQ paths
sched_ext: Use shorter slice in bypass mode
sched_ext: Mark racy bitfields to prevent adding fields that can't tolerate races
sched_ext: Minor cleanups to scx_task_iter
...
Pull rseq updates from Thomas Gleixner:
"A large overhaul of the restartable sequences and CID management:
The recent enablement of RSEQ in glibc resulted in regressions which
are caused by the related overhead. It turned out that the decision to
invoke the exit to user work was not really a decision. More or less
each context switch caused that. There is a long list of small issues
which sums up nicely and results in a 3-4% regression in I/O
benchmarks.
The other detail which caused issues due to extra work in context
switch and task migration is the CID (memory context ID) management.
It also requires to use a task work to consolidate the CID space,
which is executed in the context of an arbitrary task and results in
sporadic uncontrolled exit latencies.
The rewrite addresses this by:
- Removing deprecated and long unsupported functionality
- Moving the related data into dedicated data structures which are
optimized for fast path processing.
- Caching values so actual decisions can be made
- Replacing the current implementation with a optimized inlined
variant.
- Separating fast and slow path for architectures which use the
generic entry code, so that only fault and error handling goes into
the TIF_NOTIFY_RESUME handler.
- Rewriting the CID management so that it becomes mostly invisible in
the context switch path. That moves the work of switching modes
into the fork/exit path, which is a reasonable tradeoff. That work
is only required when a process creates more threads than the
cpuset it is allowed to run on or when enough threads exit after
that. An artificial thread pool benchmarks which triggers this did
not degrade, it actually improved significantly.
The main effect in migration heavy scenarios is that runqueue lock
held time and therefore contention goes down significantly"
* tag 'core-rseq-2025-11-30' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (54 commits)
sched/mmcid: Switch over to the new mechanism
sched/mmcid: Implement deferred mode change
irqwork: Move data struct to a types header
sched/mmcid: Provide CID ownership mode fixup functions
sched/mmcid: Provide new scheduler CID mechanism
sched/mmcid: Introduce per task/CPU ownership infrastructure
sched/mmcid: Serialize sched_mm_cid_fork()/exit() with a mutex
sched/mmcid: Provide precomputed maximal value
sched/mmcid: Move initialization out of line
signal: Move MMCID exit out of sighand lock
sched/mmcid: Convert mm CID mask to a bitmap
cpumask: Cache num_possible_cpus()
sched/mmcid: Use cpumask_weighted_or()
cpumask: Introduce cpumask_weighted_or()
sched/mmcid: Prevent pointless work in mm_update_cpus_allowed()
sched/mmcid: Move scheduler code out of global header
sched: Fixup whitespace damage
sched/mmcid: Cacheline align MM CID storage
sched/mmcid: Use proper data structures
sched/mmcid: Revert the complex CID management
...
Now that all pieces are in place, change the implementations of
sched_mm_cid_fork() and sched_mm_cid_exit() to adhere to the new strict
ownership scheme and switch context_switch() over to use the new
mm_cid_schedin() functionality.
The common case is that there is no mode change required, which makes
fork() and exit() just update the user count and the constraints.
In case that a new user would exceed the CID space limit the fork() context
handles the transition to per CPU mode with mm::mm_cid::mutex held. exit()
handles the transition back to per task mode when the user count drops
below the switch back threshold. fork() might also be forced to handle a
deferred switch back to per task mode, when a affinity change increased the
number of allowed CPUs enough.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20251119172550.280380631@linutronix.de
The MM CID management has two fundamental requirements:
1) It has to guarantee that at no given point in time the same CID is
used by concurrent tasks in userspace.
2) The CID space must not exceed the number of possible CPUs in a
system. While most allocators (glibc, tcmalloc, jemalloc) do not
care about that, there seems to be at least some LTTng library
depending on it.
The CID space compaction itself is not a functional correctness
requirement, it is only a useful optimization mechanism to reduce the
memory foot print in unused user space pools.
The optimal CID space is:
min(nr_tasks, nr_cpus_allowed);
Where @nr_tasks is the number of actual user space threads associated to
the mm and @nr_cpus_allowed is the superset of all task affinities. It is
growth only as it would be insane to take a racy snapshot of all task
affinities when the affinity of one task changes just do redo it 2
milliseconds later when the next task changes it's affinity.
That means that as long as the number of tasks is lower or equal than the
number of CPUs allowed, each task owns a CID. If the number of tasks
exceeds the number of CPUs allowed it switches to per CPU mode, where the
CPUs own the CIDs and the tasks borrow them as long as they are scheduled
in.
For transition periods CIDs can go beyond the optimal space as long as they
don't go beyond the number of possible CPUs.
The current upstream implementation adds overhead into task migration to
keep the CID with the task. It also has to do the CID space consolidation
work from a task work in the exit to user space path. As that work is
assigned to a random task related to a MM this can inflict unwanted exit
latencies.
Implement the context switch parts of a strict ownership mechanism to
address this.
This removes most of the work from the task which schedules out. Only
during transitioning from per CPU to per task ownership it is required to
drop the CID when leaving the CPU to prevent CID space exhaustion. Other
than that scheduling out is just a single check and branch.
The task which schedules in has to check whether:
1) The ownership mode changed
2) The CID is within the optimal CID space
In stable situations this results in zero work. The only short disruption
is when ownership mode changes or when the associated CID is not in the
optimal CID space. The latter only happens when tasks exit and therefore
the optimal CID space shrinks.
That mechanism is strictly optimized for the common case where no change
happens. The only case where it actually causes a temporary one time spike
is on mode changes when and only when a lot of tasks related to a MM
schedule exactly at the same time and have eventually to compete on
allocating a CID from the bitmap.
In the sysbench test case which triggered the spinlock contention in the
initial CID code, __schedule() drops significantly in perf top on a 128
Core (256 threads) machine when running sysbench with 255 threads, which
fits into the task mode limit of 256 together with the parent thread:
Upstream rseq/perf branch +CID rework
0.42% 0.37% 0.32% [k] __schedule
Increasing the number of threads to 256, which puts the test process into
per CPU mode looks about the same.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20251119172550.023984859@linutronix.de
The MM CID management has two fundamental requirements:
1) It has to guarantee that at no given point in time the same CID is
used by concurrent tasks in userspace.
2) The CID space must not exceed the number of possible CPUs in a
system. While most allocators (glibc, tcmalloc, jemalloc) do not care
about that, there seems to be at least librseq depending on it.
The CID space compaction itself is not a functional correctness
requirement, it is only a useful optimization mechanism to reduce the
memory foot print in unused user space pools.
The optimal CID space is:
min(nr_tasks, nr_cpus_allowed);
Where @nr_tasks is the number of actual user space threads associated to
the mm and @nr_cpus_allowed is the superset of all task affinities. It is
growth only as it would be insane to take a racy snapshot of all task
affinities when the affinity of one task changes just do redo it 2
milliseconds later when the next task changes its affinity.
That means that as long as the number of tasks is lower or equal than the
number of CPUs allowed, each task owns a CID. If the number of tasks
exceeds the number of CPUs allowed it switches to per CPU mode, where the
CPUs own the CIDs and the tasks borrow them as long as they are scheduled
in.
For transition periods CIDs can go beyond the optimal space as long as they
don't go beyond the number of possible CPUs.
The current upstream implementation adds overhead into task migration to
keep the CID with the task. It also has to do the CID space consolidation
work from a task work in the exit to user space path. As that work is
assigned to a random task related to a MM this can inflict unwanted exit
latencies.
This can be done differently by implementing a strict CID ownership
mechanism. Either the CIDs are owned by the tasks or by the CPUs. The
latter provides less locality when tasks are heavily migrating, but there
is no justification to optimize for overcommit scenarios and thereby
penalizing everyone else.
Provide the basic infrastructure to implement this:
- Change the UNSET marker to BIT(31) from ~0U
- Add the ONCPU marker as BIT(30)
- Add the TRANSIT marker as BIT(29)
That allows to check for ownership trivially and provides a simple check for
UNSET as well. The TRANSIT marker is required to prevent CID space
exhaustion when switching from per CPU to per task mode.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://patch.msgid.link/20251119172549.960252358@linutronix.de
Reading mm::mm_users and mm:::mm_cid::nr_cpus_allowed every time to compute
the maximal CID value is just wasteful as that value is only changing on
fork(), exit() and eventually when the affinity changes.
So it can be easily precomputed at those points and provided in mm::mm_cid
for consumption in the hot path.
But there is an issue with using mm::mm_users for accounting because that
does not necessarily reflect the number of user space tasks as other kernel
code can take temporary references on the MM which skew the picture.
Solve that by adding a users counter to struct mm_mm_cid, which is modified
by fork() and exit() and used for precomputing under mm_mm_cid::lock.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20251119172549.832764634@linutronix.de
The CID management is a complex beast, which affects both scheduling and
task migration. The compaction mechanism forces random tasks of a process
into task work on exit to user space causing latency spikes.
Revert back to the initial simple bitmap allocating mechanics, which are
known to have scalability issues as that allows to gradually build up a
replacement functionality in a reviewable way.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20251119172549.068197830@linutronix.de
Bypass mode routes tasks through fallback dispatch queues. Originally a single
global DSQ, b7b3b2dbae ("sched_ext: Split the global DSQ per NUMA node")
changed this to per-node DSQs to resolve NUMA-related livelocks.
Dan Schatzberg found per-node DSQs can still livelock when many threads are
pinned to different small CPU subsets: each CPU must scan many incompatible
tasks to find runnable ones, causing severe contention with high CPU counts.
Switch to per-CPU bypass DSQs. Each task queues on its current CPU. Default
idle CPU selection and direct dispatch handle most cases well.
This introduces a failure mode when tasks concentrate on one CPU in
over-saturated systems. If the BPF scheduler severely skews placement before
triggering bypass, that CPU's queue may be too long to drain, causing RCU
stalls. A load balancer in a future patch will address this. The bypass DSQ is
separate from local DSQ to enable load balancing: local DSQs use rq locks,
preventing efficient scanning and transfer across CPUs, especially problematic
when systems are already contended.
v2: Clarified why bypass DSQ is separate from local DSQ (Andrea Righi).
Reported-by: Dan Schatzberg <schatzberg.dan@gmail.com>
Reviewed-by: Dan Schatzberg <schatzberg.dan@gmail.com>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The dl_server time accounting code is a little odd. The normal scheduler
pattern is to update curr before doing something, such that the old state is
fully accounted before changing state.
Notably, the dl_server_timer() needs to propagate the current time accounting
since the current task could be ran by dl_server and thus this can affect
dl_se->runtime. Similarly for dl_server_start().
And since the (deferred) dl_server wants idle time accounted, rework
sched_idle_class time accounting to be more like all the others.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20251020141130.GJ3245006@noisy.programming.kicks-ass.net
Basically, from the constraint that the sum of lag is zero, you can
infer that the 0-lag point is the weighted average of the individual
vruntime, which is what we're trying to compute:
\Sum w_i * v_i
avg = --------------
\Sum w_i
Now, since vruntime takes the whole u64 (worse, it wraps), this
multiplication term in the numerator is not something we can compute;
instead we do the min_vruntime (v0 henceforth) thing like:
v_i = (v_i - v0) + v0
This does two things:
- it keeps the key: (v_i - v0) 'small';
- it creates a relative 0-point in the modular space.
If you do that subtitution and work it all out, you end up with:
\Sum w_i * (v_i - v0)
avg = --------------------- + v0
\Sum w_i
Since you cannot very well track a ratio like that (and not suffer
terrible numerical problems) we simpy track the numerator and
denominator individually and only perform the division when strictly
needed.
Notably, the numerator lives in cfs_rq->avg_vruntime and the denominator
lives in cfs_rq->avg_load.
The one extra 'funny' is that these numbers track the entities in the
tree, and current is typically outside of the tree, so avg_vruntime()
adds current when needed before doing the division.
(vruntime_eligible() elides the division by cross-wise multiplication)
Anyway, as mentioned above, we currently use the CFS era min_vruntime
for this purpose. However, this thing can only move forward, while the
above avg can in fact move backward (when a non-eligible task leaves,
the average becomes smaller), this can cause trouble when through
happenstance (or construction) these values drift far enough apart to
wreck the game.
Replace cfs_rq::min_vruntime with cfs_rq::zero_vruntime which is kept
near/at avg_vruntime, following its motion.
The down-side is that this requires computing the avg more often.
Fixes: 147f3efaa2 ("sched/fair: Implement an EEVDF-like scheduling policy")
Reported-by: Zicheng Qu <quzicheng@huawei.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20251106111741.GC4068168@noisy.programming.kicks-ass.net
Cc: stable@vger.kernel.org
After removing the various condition bits earlier it turns out that one
extra information is needed to avoid setting event::sched_switch and
TIF_NOTIFY_RESUME unconditionally on every context switch.
The update of the RSEQ user space memory is only required, when either
the task was interrupted in user space and schedules
or
the CPU or MM CID changes in schedule() independent of the entry mode
Right now only the interrupt from user information is available.
Add an event flag, which is set when the CPU or MM CID or both change.
Evaluate this event in the scheduler to decide whether the sched_switch
event and the TIF bit need to be set.
It's an extra conditional in context_switch(), but the downside of
unconditionally handling RSEQ after a context switch to user is way more
significant. The utilized boolean logic minimizes this to a single
conditional branch.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Link: https://patch.msgid.link/20251027084307.578058898@linutronix.de
The ops.cpu_acquire/release() callbacks miss events under multiple conditions.
There are two distinct task dispatch gaps that can cause cpu_released flag
desynchronization:
1. balance-to-pick_task gap: This is what was originally reported. balance_scx()
can enqueue a task, but during consume_remote_task() when the rq lock is
released, a higher priority task can be enqueued and ultimately picked while
cpu_released remains false. This gap is closeable via RETRY_TASK handling.
2. ttwu-to-pick_task gap: ttwu() can directly dispatch a task to a CPU's local
DSQ. By the time the sched path runs on the target CPU, higher class tasks may
already be queued. In such cases, nothing on sched_ext side will be invoked,
and the only solution would be a hook invoked regardless of sched class, which
isn't desirable.
Rather than adding invasive core hooks, BPF schedulers can use generic BPF
mechanisms like tracepoints. From SCX scheduler's perspective, this is congruent
with other mechanisms it already uses and doesn't add further friction.
The main use case for cpu_release() was calling scx_bpf_reenqueue_local() when
a CPU gets preempted by a higher priority scheduling class. However, the old
scx_bpf_reenqueue_local() could only be called from cpu_release() context.
Add a new version of scx_bpf_reenqueue_local() that can be called from any
context by deferring the actual re-enqueue operation. This eliminates the need
for cpu_acquire/release() ops entirely. Schedulers can now use standard BPF
mechanisms like the sched_switch tracepoint to detect and handle CPU preemption.
Update scx_qmap to demonstrate the new approach using sched_switch instead of
cpu_release, with compat support for older kernels. Mark cpu_acquire/release()
as deprecated. The old scx_bpf_reenqueue_local() variant will be removed in
v6.23.
Reported-by: Wen-Fang Liu <liuwenfang@honor.com>
Link: https://lore.kernel.org/all/8d64c74118c6440f81bcf5a4ac6b9f00@honor.com/
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Peter Zijlstra
Thomas Gleixner
Linus Torvalds
Frederic Weisbecker
Wangyang Guo
Andrea Righi
Shubhang Kaushik
Juergen Gross
Blake Jones
Marco Elver
Ingo Molnar
Thomas Gleixner
Tejun Heo
Fernand Sieber