Under core scheduling, pick_next_task() selects for all SMT siblings under
one continuous hold of the shared core-wide rq lock, and sched_ext's
dispatch can release that lock from inside the pick. In preparation for
making the core-sched pick detect the releases and retry, add
rq->scx.lock_drop_seq and bump it at every site that can release an rq lock
while a dispatch may be in flight. The counter is only maintained while core
scheduling is enabled. No functional changes.
Fixes: 4c95380701 ("sched/ext: Fold balance_scx() into pick_task_scx()")
Cc: stable@vger.kernel.org # v6.19+
Signed-off-by: Tejun Heo <tj@kernel.org>
Under core scheduling, dispatch runs from within the core-wide pick and can
target a sibling rq, so ops.dispatch() may execute on a CPU different from
the dispatched rq's. Several kfunc paths assumed the two always coincide:
- scx_dsq_move() decided whether an rq lock is held by testing this_rq()'s
rq flags and lock-danced accordingly. A dispatch for a sibling took the
unlocked-context branch and acquired the source rq lock on top of the
already held dispatched rq lock which could deadlock.
- scx_bpf_sub_dispatch() dispatched this_rq() with its stashed
sub_dispatch_prev, which is NULL when dispatching for a sibling.
- finish_dispatch(), scx_bpf_dsq_reenq() and scx_bpf_dsq_nr_queued()
resolved SCX_DSQ_LOCAL to this CPU's local DSQ rather than the dispatched
rq's. The latter two are callable from other rq-locked operations too,
where SCX_DSQ_LOCAL now likewise resolves to the op's rq. This changes
behavior also without core scheduling, e.g. for ops.enqueue() running a
remote wakeup on the waking CPU, and is intended: which CPU happens to
execute an operation is incidental, the op's rq is what it is operating
on, and the resolution now matches the insert side where SCX_DSQ_LOCAL
dispatches land on the task's rq.
Use the rq tracked by scx_locked_rq(), which is set to the dispatched rq
around ops invocations and NULL in unlocked contexts.
Fixes: 4c95380701 ("sched/ext: Fold balance_scx() into pick_task_scx()")
Cc: stable@vger.kernel.org # v6.19+
Signed-off-by: Tejun Heo <tj@kernel.org>
SCX_RQ_BAL_KEEP tells the pick to keep running the previous task, a leftover
from when balancing and picking were separate operations. An rq-level flag
only works while dispatches and picks pair up one to one, which core
scheduling breaks: selections interleave through dispatch's lock drops and a
pick can consume a stale flag, keeping a task that has since been dequeued.
Fixing core scheduling support requires the decision to travel with the
dispatch that made it. Make scx_dispatch_sched() and balance_one() return an
explicit verdict instead and drop the flag's plumbing from the tools autogen
enum headers.
Also factor the pick-side invocation, its follow-up queueing and the
post-dispatch checks out of do_pick_task_scx() into dispatch_pick(). No
functional changes intended.
v2: Drop the SCX_RQ_BAL_KEEP plumbing from the tools autogen enum headers
as well (Andrea).
Fixes: 4c95380701 ("sched/ext: Fold balance_scx() into pick_task_scx()")
Cc: stable@vger.kernel.org # v6.19+
Signed-off-by: Tejun Heo <tj@kernel.org>
Core scheduling's pick_next_task() operates on all sibling rqs under one
acquisition of the shared core-wide lock. A ->pick_task() that releases the
rq lock leaves every sibling __lock momentarily free, letting
__sched_core_flip(false) complete mid-selection and rebind rq_lockp() under
it. The selection resumes on the split locks, touching sibling state it no
longer protects, and __schedule() finally releases a lock that was never
taken while leaking the one that was.
Count in-flight core-wide selections in the leader's rq->core_pick_in_flight
and make __sched_core_flip() wait for the count to drain. The count only
changes under the shared lock, which the flip holds while sampling, so no
other ordering is needed. The wait can repeat while selections overlap, but
the flip backs off between samples and flips are rare cookie-lifetime
events.
sched_core_cpu_deactivate() moves the count to the new leader - a stale copy
left behind would bias it forever if that CPU later returns as its own
leader.
Fixes: 539f65125d ("sched: Add core wide task selection and scheduling")
Cc: stable@vger.kernel.org # v5.14+
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Core scheduling's pick_next_task() breaks when a ->pick_task()
implementation can release the rq lock. The selection state derived on entry
is only valid while the lock is held continuously. Once a pick can drop the
lock, an interleaving selection can invalidate all of it: the single-CPU
fast path can commit an uncookied pick although the core went cookied during
the release, and forceidle committed by the interleaving selection skews the
restarted pass's accounting.
Fix it by restarting the whole selection when a pick returns RETRY_TASK
after releasing the lock: a single restart point above the state derivation
replaces the per-loop restart labels, so a retry picks up state committed by
interleaving selections and accounts and resets forceidle like a fresh
selection would.
need_sync and fi_before latch across retries. Clock validity can't be
re-derived - there is no program-ordered way to tell whether the own and
core rq clocks are still updated after the lock was released, as other
lockers' pin cycles may or may not have invalidated them. When restarting,
clear core_clock_updated so that the sibling loop re-updates the core rq,
and update the own rq clock if invalidated.
Fixes: 4c95380701 ("sched/ext: Fold balance_scx() into pick_task_scx()")
Cc: stable@vger.kernel.org # v6.19+
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Pull bpf-next d114bb9893 ("Merge branch
'add-arena-argument-support-to-kfuncs-and-struct_ops'") to make the __arena
and __arena__nullable kfunc and struct_ops argument suffixes available. The
suffixed arguments will be used to convert sched_ext kfuncs and struct_ops
callbacks that currently pass arena pointers as scalars and rebase them by
hand.
With the changes that enable preempt count to track IRQ disabling
nesting, we don't have enough bits in 32-bit preempt count
implementation, as a result we move NMI nesting bits out of the 32-bit
preempt count. However on the architectures that can support 64-bit
preempt count implementation, we can keep the NMI nesting bits in the
32-bit preempt count and avoid maintaining NMI nesting bits outside of
the same cache line.
Therefore HAS_SEPARATE_PREEMPT_RESCHED_BITS is introduced to allow
architectures to select this. Note that under this Kconfig, preempt
count is maintained in a 64-bit word however preempt_count() still
remains as an int because all the effective bits still fit in
(previously we mask out NEED_RESCHED bit in preempt_count()). This
should make no functional changes for existing preempt_count() users.
Enable this for x86_64 along with the introduction of the Kconfig.
[boqun: Undo the __preempt_count_{add,sub}() optimization in 32-bit
preempt count since it may introduce {over,under}flow]
Originally-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260804161447.84806-11-boqun@kernel.org
Currently preempt_count() is always a non-negative int on all archs
(PREEMPT_NEED_RESCHED archs will mask out the MSB when returning
preempt_count()), hence the checking in __cant_migrate() is in fact just
checking whether preempt_count() is 0 or not. In a future change, we are
going to use all the 32 bits of preempt_count(), which would make
negative int values possible from preempt_count(). Therefore convert the
"> 0" comparison into a zero check to prepare for the future change.
No functional changes are intended.
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260804161447.84806-10-boqun@kernel.org
The preempt_offset is always 0 in all the callsites of __cant_sleep(),
hence remove it. It also allows us to clear up the code a bit by
no longer using a "preempt_count() > .." comparison.
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260804161447.84806-9-boqun@kernel.org
The semantics of various IRQ disabling guards match what
*_irq_{disable,enable}() provide, i.e. the interrupt disabling is
properly nested, therefore it's OK to switch to use
*_irq_{disable,enable}() primitives.
[boqun: Adjust the user-side changes in do_sched_cfs_*_timer() provided
by Peter and Lyude]
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260804161447.84806-8-boqun@kernel.org
Merge cpufreq updates for 7.3-rc1:
- Minor fixes and cleanups in assorted cpufreq drivers (Dan Carpenter,
Guru Das Srinagesh, Haoxiang Li, Karl Mehltretter, Sasha Finkelstein,
and Pan Chuang)
- Fix cpufreq table creation and bios_limits() callback in the Rust
bindings (Priya Bala Govindasamy)
- Add IPQ5210 support to qcom-nvmem driver (Varadarajan Narayanan)
- Adjust the .adjust_perf() cpufreq driver callback to allow the
maximum performance value to be passed to drivers and update the
intel_pstate driver to use it (Rafael Wysocki)
- Set policy->cur to the actual requested frequency in the intel_pstate
driver when the performance policy is used (Rafael Wysocki)
- Simplify HWP handling on Broadwell processors in intel_pstate (Rafael
Wysocki)
- Fix setting minimum P-state at init time in intel_pstate (Rafael
Wysocki)
- Consolidate frequency values computation in intel_pstate and clean up
code in that driver (Rafael Wysocki)
- Add missing kernel-doc desciptions for structure and union members in
the amd-pstate driver (David Vernet)
- Handle missing policy in dynamic EPP callbacks in the amd-pstate
driver (EDAMAMEX)
- Introduce EXPORT_SYMBOL_FOR_PSTATE_UT() to export amd-pstate driver
symbols to the amd-pstate-ut subdriver (K Prateek Nayak)
- Add dynamic EPP as an "energy_performance_preference" mode in
amd-pstate, remove the "amd_dynamic_epp" kernel command line option
and the "dynamic_epp" sysfs attribute, and update the dynamic_epp
documentation accordingly (K Prateek Nayak)
- Add unit tests for CPPC Performance Priority and the "dynamic" EPP
mode in the amd-pstate driver (K Prateek Nayak)
- Set min_limit_freq based on bios_min_perf in amd-pstate and remove
the defensive check for bios_min_perf from it (K Prateek Nayak)
- Fix EPP return type and handle errors in amd-pstate during
initialization, toggle auto_sel in active mode on shared memory
systems, and cache the firmware programmed EPP value (Marco Scardovi)
- Skip tests in amd-pstate-ut if the amd-pstate driver is not in active
use (Qianheng Peng)
- Replace sprintf() with sysfs_emit() in sysfs show in the cpufreq
schedutil governor and fix a self-contradictory comment in
sugov_iowait_apply() (Zhongqiu Han)
- Fix the usage example for the sampling_rate tunable of the ondemand
cpufreq governor in admin-guide (wangxiaodong)
* pm-cpufreq: (40 commits)
cpufreq: imx6q: fix out-of-bounds write when probed more than once
cpufreq: imx6q: fix devres accumulation across driver rebind
rust: cpufreq: Fix temporary write in Registration::bios_limit_callback
rust: cpufreq: Add CPUFREQ_TABLE_END as last table entry in TableBuilder::to_table
cpufreq: intel_pstate: Adjust policy->cur in active mode to policy
cpufreq/amd-pstate: Document missing kernel-doc members
cpufreq/amd-pstate-ut: Add unit test for CPPC Performance Priority
cpufreq/amd-pstate-ut: Add unit test for "dynamic" EPP mode
cpufreq/amd-pstate: Reduce the scope of exported symbols
Documentation/amd-pstate: Update dynamic_epp documentation with new behavior
cpufreq/amd-pstate: Remove "amd_dynamic_epp" cmdline and "dynamic_epp" sysfs
cpufreq/amd-pstate: Add dynamic EPP as an "energy_performance_preference" mode
cpufreq/amd-pstate: Extract platform profile to EPP conversion into a helper
cpufreq/amd-pstate: Remove the defensive check for bios_min_perf
cpufreq/amd-pstate: Set min_limit_freq based on bios_min_perf
cpufreq: apple-soc: Calculate frequency as a 64-bit value
kselftest: cpufreq: Backup and restore governor for sptests
selftests/cpufreq: Remove unnecessary sudo from quick_shuffle()
selftests/cpufreq: Remove unused local variables from switch_show_governor()
cpufreq/amd-pstate: handle missing policy in dynamic EPP callbacks
...
Commit 9c63e84db2 ("sched/core: Disable SD_PREFER_SIBLING on asymmetric
CPU capacity domains") removed the SD_PREFER_SIBLING from the domains with
asymmetric capacity. This was done to avoid spreading tasks to sibling
scheduling groups with less capacity, but this does not happen: checks for
capacity in update_sd_pick_busiest(), sched_balance_find_src_group(), and
sched_balance_find_src_rq() prevent migrations from high- to low-capacity
CPUs if the busiest group is not overloaded.
The cluster topology is a notable example: some systems have scheduling
domains spanning CPUs of asymmetric capacity, grouped into two or more
equal-capacity clusters sharing an L2 cache. When CONFIG_SCHED_CLUSTER is
enabled, SD_PREFER_SIBLING is needed in the domain to spread load across
these clusters.
CPUs with spare capacity, big or small, have always helped overloaded
groups. Once the overloading condition disappears, misfit load will still
be used to move high-utilization tasks to bigger CPUs if they have spare
capacity.
Adding the SD_PREFER_SIBLING flag shifts load balancing in shared-LLC
domains from equalizing the number of idle CPUs to equalizing the number
of running tasks. This enables migrations among clusters from newly-idle
load balance, where the outgoing task is already dequeued but the CPU
has not yet transitioned to idle.
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-6-bb500bf4afd4@linux.intel.com
sched_balance_find_src_rq() avoids selecting a runqueue with a single
running task as busiest if doing so results in migrating the task to a
CPU with less than ~5% of extra capacity. It also unintentionally
prevents migrations between CPUs of identical capacity.
When CONFIG_SCHED_CLUSTER is enabled, load should be balanced across
clusters of CPUs with the same capacity. Allowing migration between CPUs
of identical capacity is necessary to meet this goal.
Use get_actual_cpu_capacity() to reflect architectural capacity as well
as diminished capacity due to hardware or cpufreq pressure. Guard this
check with the sched_cluster_active static key so that systems without
cluster topology are unaffected.
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-5-bb500bf4afd4@linux.intel.com
In domains with asymmetric capacity, identifying misfit load in a
scheduling group is not useful when the destination CPU cannot help (i.e.,
its capacity exceeds the group's maximum CPU capacity by less than ~5%). In
such cases, it also prevents load balance among clusters of equal capacity
when CONFIG_SCHED_CLUSTER is enabled. This happens because
update_sd_pick_busiest() skips candidate groups of type misfit_task if the
destination CPU has similar capacity.
Skipping misfit load accounting in this situation allows the group to be
classified as has_spare or fully_busy and lets load balancing proceed. Keep
marking scheduling groups as overloaded when misfit tasks are present. The
sg_overloaded flag propagates to the root domain and allows bigger CPUs in
it to help via newly idle balance.
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-4-bb500bf4afd4@linux.intel.com
update_sd_pick_busiest() may incorrectly select a fully_busy group as the
busiest group when its per-CPU capacity exceeds that of the destination
CPU. This happens because the type of busiest group is initialized to
group_has_spare and allows the fully_busy group to win the type comparison.
update_sd_pick_busiest() should not choose a candidate scheduling group
with at most one runnable task if its per-CPU capacity is greater than that
of the destination CPU. Such a check already exists, but it is done too
late: after the type comparison, preventing a subsequent fully_busy group
of equal per-CPU capacity from being correctly selected.
Move this check to occur before comparing group types.
Fixes: 0b0695f2b3 ("sched/fair: Rework load_balance()")
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-3-bb500bf4afd4@linux.intel.com
The argument sg_overloaded of update_sg_lb_stats() is only consumed when
balancing at the root domain. It only makes sense to update it in such a
case. Commit 3229adbe78 ("sched/fair: Do not compute overloaded status
unnecessarily during lb") updated the logic accordingly but missed the case
in which the root domain has the SD_ASYM_CPUCAPACITY flag. Fix this.
Fixes: 3229adbe78 ("sched/fair: Do not compute overloaded status unnecessarily during lb")
Reported-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Tested-by: Christian Loehle <christian.loehle@arm.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-2-bb500bf4afd4@linux.intel.com
When picking a busiest CPU with only one running task, the function
sched_balance_find_src_rq() skips candidate CPUs if the destination CPU has
less than ~5% extra capacity. This condition only holds if all the SMT
siblings of a CPU are idle.
SMT siblings share the computing resources of a physical core and this
results in reduced capacity if more than one sibling is busy.
Skipping a CPU as described would prevent the load balancer from pulling
tasks from a scheduling group previously and correctly identified as
group_smt_balance (i.e., one with more than one task running).
Do not skip a candidate CPU of similar capacity if it has busy SMT
siblings.
Signed-off-by: Ricardo Neri <ricardo.neri-calderon@linux.intel.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>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260720-rneri-fix-cas-clusters-v6-1-bb500bf4afd4@linux.intel.com
find_new_ilb() selects the first idle housekeeping CPU without
considering whether another thread is running on the same physical core.
On an SMT system, the idle load balancer can therefore activate both
siblings even when another housekeeping CPU has an entirely idle core.
On most SMT systems, this is not problematic because the idle load
balancer is a short-lived activity and the transient wakeup of a sibling
has negligible performance impact.
However, this can be particularly costly on NVIDIA Olympus cores used in
Vera. Briefly activating an otherwise idle sibling can reduce the
performance available to the other sibling and this effect does not
necessarily end once the activated sibling becomes idle: after the ILB
finishes and its CPU enters WFI, full single-thread performance is
restored only after the sibling has remained idle for a qualification
interval (10 Ki cycles on the tested Vera system). Repeated short
sibling wakeups can therefore sustain the interference even with little
actual overlap.
Prevent this by preferring an idle housekeeping CPU whose entire SMT
core is idle. Retain the first idle CPU as a fallback when no fully idle
core is available, so NOHZ balancing continues to make forward progress.
Once a partially busy core has been examined, skip its remaining SMT
siblings to avoid repeating the core-idle check on wide SMT systems.
Tests performed using an ad hoc GEMM benchmark running one CPU-intensive
task per SMT core within its CPU affinity mask improved from
approximately 6.2 TFLOP/s to 9.4 TFLOP/s.
Note that this preference may wake a fully idle physical core instead of
using an idle sibling of an active core, potentially increasing ILB
wakeup latency or energy consumption on some architectures. It may also
scan additional CPUs before selecting the one to run the ILB. The
selection falls back to the first idle CPU when no fully idle SMT core
is available. Non-SMT systems continue to select the first idle
housekeeping CPU.
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mete Durlu <meted@linux.ibm.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://patch.msgid.link/20260804151324.918020-1-arighi@nvidia.com
CONFIG_SYSCTL used to make sense as a separate hidden bool before commit
61a47c1ad3 ("sysctl: Remove the sysctl system call"); it was selected
by both CONFIG_SYSCTL_SYSCALL and CONFIG_PROC_SYSCTL.
Today CONFIG_PROC_SYSCTL is the only selector, so the two are always
equal. Kill the hidden bool, rename the PROC_SYSCTL prompt to SYSCTL,
and s/CONFIG_PROC_SYSCTL/CONFIG_SYSCTL/ tree-wide.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Joel Granados <joel.granados@kernel.org>
blkdev.h gets included in various places outside the block layer just
for struct blk_plug and related plugging functions.
Split blk_plug into a separate helper to reduce the amount of code
that needs to get rebuilt when blkdev.h changes and to slightly
reduce compile times.
In io_uring this requires pulling in a few other headers explicitly that
previously were implicitly included through blkdev.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christian Brauner (Amutable) <brauner@kernel.org>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Link: https://patch.msgid.link/20260804125524.740996-1-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When rescue demand on a cpu persistently exceeds the configured bandwidth,
tasks age on that cpu's rescue DSQ until the stall watchdog fires. The
watchdog blames the waiting task's owner, but the misbehaving party is
whoever floods the queue, not whoever happens to time out.
Track each sched's recent rescue consumption per cpu as a decaying average.
Once the oldest waiter on a cpu's rescue DSQ has been queued past a
threshold derived from the rescue knobs (4s at the defaults), the rescue
timer ejects the sub with the highest recent consumption on that cpu with
SCX_EXIT_ERROR_RESCUE. With no recent consumer there is no victim and
nothing is ejected - the generic stall watchdog eventually blames the
waiter's owner instead. Ejections on a cpu are spaced one threshold apart so
the freed bandwidth can drain the backlog before another sub is judged.
The overload check only wins the race against the stall watchdog when the
watchdog timeout clears the threshold, and a single in-budget wait must not
cross the trigger on its own. Warn on a scheduler whose timeout doesn't fit
and on knobs whose funding period exceeds half the threshold.
v2: - Track kill_at in jiffies_64 - on 32-bit, the time_before() grace check
wraps 2^31 ticks after the last ejection and suppresses ejections.
(sashiko AI)
- Track rescue_avg_at in jiffies_64 likewise - the unsigned long decay
delta truncates mod 2^32 on 32-bit and can revive a weeks-old usage
average in the victim pick.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A local DSQ insert lacking the needed caps is diverted to the reject DSQ and
bounced back through ops.enqueue() so the scheduler can re-decide. That
recovery assumes the scheduler has somewhere legal to send the task. When it
doesn't, e.g. when the task's affinity is restricted to cids delegated away,
the task starves until the stall watchdog ejects the scheduler. An exiting
task is worse - it skips ops.enqueue() and the rejection becomes a
self-requeuing cycle that burns the CPU until the watchdog fires.
Add SCX_ENQ_RESCUE, a fallback modifier on local DSQ inserts. When the
insert would be rejected for missing caps, the kernel takes over and runs
the task on the target CPU without consulting the owning scheduler. The
kernel sets the flag itself when enqueueing an exiting task.
Rescue is a last-resort forward-progress backstop with a persistent
disadvantage, not a way around cap enforcement. A per-CPU token bucket
accrues rescue_bandwidth_ppt (default 2%) of CPU time and rescues run one at
a time in arrival order. Each is granted a slice of the rescue_quantum_us
(default 5ms) quantum divided across the waiters, waits at the tail of the
local DSQ claiming no priority, and rejoins its scheduler as a fresh arrival
once the slice is served.
The schedulers keep their normal control over an admitted rescuee and may
preempt or reslice it. Service is measured on CPU time actually received, so
neither shortens the rescue. Prolonged denial escalates - the remaining
slice turns into protected execution (SCX_TASK_PROTECTED) and the rescuee
preempts the current task. Escalation is paced by the same bucket, and
delivered service converges on the configured bandwidth no matter how
aggressively the schedulers dispatch.
Both knobs are root-only and SCX_RESCUE_DISABLE turns rescue off, making
SCX_ENQ_RESCUE inserts reject as usual.
v2: - Add SCX_OPS_OPEN() fix-ups for the new ops fields so cpu-form
schedulers setting them still load on older kernels. (Andrea)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A BPF scheduler can displace any of its tasks at will - cut a running one's
slice with an SCX_ENQ_PREEMPT dispatch, an SCX_KICK_PREEMPT kick or a direct
shortening, and jump a queued one with HEAD insertions. Sometimes the kernel
needs a slice and a DSQ position to stick regardless.
Add SCX_TASK_PROTECTED, guarding both:
- The slice becomes immutable. Every scheduler-reachable write is refused
and counted as SCX_EV_SLICE_DENIED. Higher scheduling classes are
unaffected. PREEMPT|IMMED can't preempt a running protected task and gets
reenqueued.
- A protected task that reached the head of its DSQ keeps it - HEAD
insertions land behind the leading run of protected tasks and reenqueue
sweeps skip them. Only rq-owned DSQs can hold protected tasks, so the walk
runs only for them.
The bit lives in p->scx.flags so that both the refusal and the head walk
read it under the rq lock that protects it.
Protection ends when the slice is consumed, when the task leaves the rq
except for a save/restore on the running task, on a yield, when the
scheduler enters bypass, and when the task leaves scx. The flag is
kernel-internal and not used yet.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
p->scx.slice and p->scx.dsq_vtime writes have no synchronization rules. The
dsq insert kfuncs write both fields synchronously from whatever context
they're called in - a direct dispatch from ops.select_cpu() writes with only
pi_lock held - and, as the kfuncs are safe to call spuriously with the
invalid dispatch discarded later, a scheduler can modify any task's slice by
spuriously calling them. The latter stands in the way of an upcoming patch
which adds kernel-granted slices that the schedulers must not be able to
modify.
Give both fields explicit rules. While the task is running, sleeping or
queued on an rq-owned DSQ, the rq lock protects them - these are the states
where the kernel consumes the slice. While queued on a user DSQ or on the
BPF side, the kernel neither consumes nor decides on the fields and every
writer acts for the BPF scheduler - synchronizing the writers is the
scheduler's responsibility and whichever write lands last wins.
To conform, an insert kfunc no longer writes the fields when called. The
values travel with the dispatch and take effect when the task is inserted. A
discarded dispatch has no side effects. The rq lock rule is asserted at the
slice store.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
SCX_ENQ_IGNORE_CAPS is kernel-internal and marks a placement the kernel
forces. scx_caps_for_enq() waives the enqueue cap for it, but a PREEMPT
insert still picks up the preemption cap requirement from
scx_caps_for_preempt(). Update scx_caps_for_preempt() to take enq_flags and
require nothing when SCX_ENQ_IGNORE_CAPS is set.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The dsq insert kfuncs reject __SCX_ENQ_INTERNAL_MASK bits in
scx_dsq_insert_preamble() instead of scx_vet_enq_flags(). A scheduler can
smuggle internal flags such as SCX_ENQ_CLEAR_OPSS through the dsq move
kfuncs and corrupt the dispatch protocol. Move the rejection into
scx_vet_enq_flags(). The vtime move wrapper OR'd the internal
SCX_ENQ_DSQ_PRIQ bit into enq_flags before the vet; the bit now goes in
inside scx_dsq_move() after the vet.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_bpf_now() couples the valid-or-fresh rq clock read to the current rq.
The read is useful for kernel-internal timing against a specific rq,
including a remotely locked one. Factor it out into __scx_bpf_now().
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
set_task_slice(), task_unlink_from_dsq(), move_local_task_to_local_dsq(),
init_dsq() and dump_line() will be used outside ext.c. Add the scx_ prefix
and declare them in internal.h. The scx_sched_all list will also be used
outside ext.c, drop its static. No functional changes.
v2: Declare scx_sched_all outside the CONFIG_EXT_SUB_SCHED block - the
definition is unconditional. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The following rescue execution addition gives the function a third possible
destination, making a name that enumerates the outcomes a poor fit. Rename
to the destination-neutral scx_resolve_local_dsq(). No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The built-in idle masks are reset with all online CPUs marked idle
before sched_ext is enabled. Busy CPUs can therefore be incorrectly
advertised as idle until their next idle transition.
Initialize the masks empty so that the initial state is conservative.
When bypass is lifted, every CPU is rescheduled and idle-to-idle
re-picks populate the masks with CPUs that are actually idle. Later
idle transitions keep the masks up to date.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Reviewed-by: Kuba Piecuch <jpiecuch@google.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
- More lifecycle fixes for the new sub-scheduler support: a failed enable
could tear down a never-linked sub-scheduler in a way that races the root
scheduler's disable and leads to a use-after-free, tasks that were not on
the ext class could still get the enable callback, and a policy-rejection
path silently rewrote a running task's scheduling policy instead of
aborting the scheduler.
- Scheduler enable/disable could deadlock with cgroup removal and a
concurrent cgroup weight write through kernfs. Fixed by reordering lock
acquisition.
- Sync wakeups could leave the waker CPU incorrectly marked idle in the
built-in idle-CPU tracking.
- A selftest fix for sleeping tasks whose CPU affinity changes before
wakeup.
-----BEGIN PGP SIGNATURE-----
iIQEABYKACwWIQTfIjM1kS57o3GsC/uxYfJx3gVYGQUCam/mog4cdGpAa2VybmVs
Lm9yZwAKCRCxYfJx3gVYGe0NAQC2UrrXeSm97RTCv+5HBfom/8sjPdB2tiAwGiUt
fUUQ+AD9FP30urJaoiALUiL+EGWpLFejwUCXNXlT+0kN2TiFxQM=
=EuWf
-----END PGP SIGNATURE-----
Merge tag 'sched_ext-for-7.2-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext
Pull sched_ext fixes from Tejun Heo:
- More lifecycle fixes for the new sub-scheduler support: a failed
enable could tear down a never-linked sub-scheduler in a way that
races the root scheduler's disable and leads to a use-after-free,
tasks that were not on the ext class could still get the enable
callback, and a policy-rejection path silently rewrote a running
task's scheduling policy instead of aborting the scheduler.
- Scheduler enable/disable could deadlock with cgroup removal and a
concurrent cgroup weight write through kernfs. Fixed by reordering
lock acquisition.
- Sync wakeups could leave the waker CPU incorrectly marked idle in the
built-in idle-CPU tracking.
- A selftest fix for sleeping tasks whose CPU affinity changes before
wakeup.
* tag 'sched_ext-for-7.2-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
selftests/sched_ext: Handle sleeping task affinity changes in numa test
sched_ext: Mark waker CPU busy when selected in WAKE_SYNC case
sched_ext: Don't enable non-ext tasks in the sub-sched task loops
sched_ext: Skip sub-disable teardown for never-linked sub-schedulers
sched_ext: Take cgroup_lock() first in scx_cgroup_lock()
sched_ext: Reject setting disallow from init_task outside the enable path
- A pressure trigger's poll timer could be re-armed while the last trigger
was being torn down and then fire after the cgroup was freed. Tie the
timer to the cgroup's lifetime and shut it down when the cgroup is freed.
- Writing to a pressure file forked a worker kthread while holding the
cgroup mutex, creating lock dependencies from the mutex to the whole fork
path. A pressure write racing a sched_ext scheduler enable, which blocks
forks before grabbing the mutex, deadlocked. Fork the worker with the
mutex dropped.
- Documentation fix for io.latency behavior on non-rotational devices.
-----BEGIN PGP SIGNATURE-----
iIQEABYKACwWIQTfIjM1kS57o3GsC/uxYfJx3gVYGQUCam/mlw4cdGpAa2VybmVs
Lm9yZwAKCRCxYfJx3gVYGVhRAQCyNBHCHpaY9erKTezenGDK6D+LqbfLWiWuIntB
swiwLQEA6h6Rgob2GDDRHOey7+XwF6PHh6xoh4FhSYboeEZd1wc=
=Re+7
-----END PGP SIGNATURE-----
Merge tag 'cgroup-for-7.2-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup
Pull cgroup fixes from Tejun Heo:
- A pressure trigger's poll timer could be re-armed while the last
trigger was being torn down and then fire after the cgroup was freed.
Tie the timer to the cgroup's lifetime and shut it down when the
cgroup is freed.
- Writing to a pressure file forked a worker kthread while holding the
cgroup mutex, creating lock dependencies from the mutex to the whole
fork path. A pressure write racing a sched_ext scheduler enable,
which blocks forks before grabbing the mutex, deadlocked.
Fork the worker with the mutex dropped.
- Documentation fix for io.latency behavior on non-rotational devices.
* tag 'cgroup-for-7.2-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
Docs/admin-guide/cgroup-v2: document io.latency rotational vs non-rotational behavior
sched/psi: Shut down rtpoll_timer in psi_cgroup_free()
sched/psi: Create the psimon kthread outside of cgroup_mutex
If the SCX_ENABLING -> SCX_ENABLED cmpxchg at the tail of
scx_root_enable_workfn() fails, the function jumps to err_disable
without setting ret. At that point ret still holds the return value
of the last successful __scx_init_task() call, which is 0, so the
err_disable fallback reports the meaningless message:
scx_root_enable() failed (0)
Set ret = -EBUSY, consistent with the other enable-state guards at
the top of the same function, so the fallback always reports a real
errno.
Signed-off-by: Liang Luo <luoliang@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
The kernel-doc comment for sched_ext_ops::sub_cgroup_id uses the old
@cgroup_id name, which no longer matches the struct member. This
produces two kernel-doc warnings:
Warning: struct member sub_cgroup_id not described in sched_ext_ops
Warning: Excess struct member cgroup_id description in sched_ext_ops
Update the @param name to match the actual member.
Signed-off-by: Liang Luo <luoliang@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
* pm-cpufreq:
cpufreq/amd-pstate: handle missing policy in dynamic EPP callbacks
cpufreq/amd-pstate: Cache the firmware programmed EPP value
cpufreq/amd-pstate: Toggle auto_sel in active mode on shared memory systems
cpufreq/amd-pstate: Fix EPP return type and handle errors during initialization
cpufreq: amd-pstate-ut: Skip tests when amd-pstate driver is not active
cpufreq: schedutil: Replace sprintf() with sysfs_emit() in sysfs show
cpufreq: schedutil: Fix self-contradictory comment in sugov_iowait_apply()
Documentation: admin-guide: cpufreq: fix sampling_rate example command
cpufreq: intel_pstate: Move two functions closer to callers
cpufreq: intel_pstate: Consolidate frequency values computation
cpufreq: intel_pstate: Introduce intel_pstate_update_freq_limits()
cpufreq: intel_pstate: Fix setting minimum P-state at init time
cpufreq: intel_pstate: Rename INTEL_PSTATE_HWP_BROADWELL
cpufreq: intel_pstate: Simplify HWP handling on Broadwell
cpufreq: intel_pstate: Adjust the .adjust_perf() driver callback
cpufreq: intel_pstate: Rearrange checks in hybrid_get_cost()
Commit 14a8570564 ("sched/deadline: Use revised wakeup rule for
dl_server") applies the revised wakeup rule to any server, as a result
servers that are not running (dl_defer_running == 0) and start with a
deadline overflow get enqueued and can boost tasks as if they were
running, invalidating the defer rule and the documented state model.
Apply the revised wakeup rule only for deferrable servers that are
marked as running.
Fixes: 14a8570564 ("sched/deadline: Use revised wakeup rule for dl_server")
Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Juri Lelli <juri.lelli@redhat.com>
Tested-by: Andrea Righi <arighi@nvidia.com>
Link: https://patch.msgid.link/20260522125833.264145-1-gmonaco@redhat.com
scx_link_sched() carries each failure out of the locked section through
err_msg and ret because scx_error() used to take scx_sched_lock and couldn't
be called under it. That restriction is gone, so report each failure at the
site it's detected and return directly. The scx_error() here claims the exit
on the sched being linked, which has no descendants yet, and the locked
propagation walk is deferred, so nothing reacquires scx_sched_lock inline.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_hardlockup() defers the abort to an irq_work because exit claiming used
to take scx_sched_lock and couldn't run from NMI. The deferral is now
unnecessary - claiming is NMI-safe and asserting ->aborting is exactly what
breaks the live-locks that hard-lock CPUs. Call handle_lockup() directly and
drop the irq_work. This also makes the self-detected case recoverable: the
perf watchdog fires on the hard-locked CPU itself, where a queued irq_work
never runs with IRQs off.
Also fix the return value: %true used to be returned whenever sched_ext was
loaded, suppressing the kernel's hardlockup report even when the abort was
refused. Return %true only when this call initiated the abort.
Fixes: bd2d76455b ("sched_ext: Defer scx_hardlockup() out of NMI")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The per-cpu kick lists are protected by IRQ masking which doesn't stop NMIs,
so scx_bpf_kick_cpu() from NMI silently drops the kick after a one-time
warning. A dropped kick can leave a CPU idle when the scheduler believes it
was woken, which is a correctness problem for the scheduler even if the
kernel is fine. Now that scx_error() works from NMI, abort the scheduler
instead so that the bug is surfaced deterministically. The warned_nmi_kick
tracking is no longer needed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The bstr exit kfuncs format the message into a shared static buffer under a
raw spinlock before initiating the exit. The lock can't be taken from NMI
and needlessly serializes all bstr exits system-wide.
Now that exit claiming is lock-free, reverse the order: claim the exit first
and format directly into the exit_info message buffer which the claim winner
owns exclusively. The new scx_exit_bstr() implements the sequence, replacing
scx_bstr_format(), and the shared buffer and lock are deleted; the formatter
itself is what bpf_trace_printk() already runs from NMI. scx_prog_sched()
callers were relying on the lock for RCU protection, which is now provided
explicitly.
A malformed format no longer changes or fails the requested operation:
scx_bpf_exit_bstr() keeps its graceful exit kind and scx_bpf_sub_kill_bstr()
still kills the child, with a fallback message carrying the formatting
errno, while the sched that supplied the bad format is aborted for its bug.
Before this and the previous patch, an "any" category kfunc called from NMI
context could trigger scx_error() and deadlock - e.g. a tracing prog
attached to a function running in NMI calling scx_bpf_dsq_peek() on a
non-existent DSQ would try to grab scx_sched_lock, which may be held by the
interrupted CPU. This and the previous patch fix the deadlock: scx_error()
and the bstr exit kfuncs, and thus scx_bpf_error() and scx_bpf_exit(), are
now safe to call from any context including NMI.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_claim_exit() claims descendants' exits by walking the subtree under
scx_sched_lock, making exit claiming, and thus scx_error(), unusable from
NMI and from under scx_sched_lock. However, kfuncs raising errors can run
from NMI-attached BPF progs, the hardlockup handler runs in NMI, and
scx_link_sched() wants to report failures under the lock.
The walk does two things with different urgencies: ->aborting must be
asserted synchronously to break IRQs-off dispatch-path live-locks, while the
descendants' exit_kind claims can happen later. Split them: sweep ->aborting
locklessly under RCU to unwedge the system and defer the locked
SCX_EXIT_PARENT walk to a new irq_work, both of which are NMI-safe.
The sweep stores each node's ->aborting and then reads its children list
while scx_link_sched() inserts and then checks the parent's ->aborting, the
two sides paired by full barriers - one side always sees the other. A link
that sees ->aborting undoes its insert and fails. As the undo's
list_del_rcu() leaves ->sibling non-empty, list_empty() can no longer
identify a never-linked sched during teardown - add sch->linked instead.
trace_sched_ext_exit can now fire from NMI and is called after the
->aborting stores so that its callbacks don't hold up live-lock recovery.
The exit backtrace is skipped for NMI exits as stack_trace_save()'s
NMI-safety is arch-dependent and undocumented.
v2: Move trace_sched_ext_exit() after the ->aborting stores (Andrea).
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Unlike local reenqueues, cap rejections have no repeat limit. A
malfunctioning scheduler can keep re-inserting a task to a cid it lacks caps
on, cycling the task through reject and reenqueue. This was assumed safe
because a task that never runs trips the stall watchdog. However, the
reenqueue irq_work re-arms itself and outranks the timer vector, blocking
everything else on the CPU including stall detection and recovery, until the
NMI hardlockup detector fires.
Local reenqueues already have a repeat cap, SCX_REENQ_LOCAL_MAX_REPEAT,
which needs generalizing to cover all reenqueues. It also has an attribution
problem. Counted per-cpu on root, it tears down the whole hierarchy even
when a sub-scheduler caused the repeated reenqueues.
Generalize by bounding every reenqueue with one per-task counter. reenq_cnt
is bumped in scx_do_enqueue_task() on each SCX_ENQ_REENQ, the single path
every reenqueue producer passes through, and cleared in clr_task_runnable()
when the task is picked to run and in scx_disable_task() when it leaves the
scheduler's control. Past SCX_REENQ_MAX_REPEAT the task's owning scheduler
is ejected with a new SCX_EXIT_ERROR_REENQ and the task is left stranded to
be picked up during sched exit.
The SCX_EV_REENQ_LOCAL_REPEAT event becomes SCX_EV_REENQ_REPEAT, counting
repeat reenqueues from all sources.
v2: Count SCX_EV_REENQ_REPEAT only when a reenqueue leads to another
reenqueue, not on every reenqueue.
v3: - Also clear reenq_cnt in scx_disable_task() so that the count doesn't
carry over to the next owner across sched class switches, scheduler
replacement or sub-scheduler rehoming (Andrea Righi).
- Update the stale SCX_EV_REENQ_LOCAL_REPEAT references in sched-ext.rst
(Andrea Righi).
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
dsq->first_task is __rcu for the lockless scx_bpf_dsq_peek(). The task
removal path compares it against the departing task with a plain load, which
sparse flags. The comparison runs under the dsq lock and only tests
identity, so rcu_access_pointer() is the fit.
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_root is __rcu and naked accesses were left as transitional markers for
the multi-scheduler transition, to be converted to accesses through the
associated scheduler instances. Most accesses have since been converted to
resolve the sched from the program or task at hand. The remaining naked
sites divide into ones that semantically always want the root sched, which
this patch resolves, and one that is left to a later patch.
The resolved sites:
- The SCX_OPS_TID_TO_TASK validation and the ecaps sync kick already hold a
sched whose ancestors[] pins the root as entry 0 with plain pointers
stable for the sched's lifetime. Reach the root through the sched at hand.
- The dispatch entry, class switch, idle notification and fork init paths
only execute while the scheduler is live and scx_root never changes inside
the live window, so no update can race them. Add scx_root_protected_live()
which documents that invariant and resolves with a plain load.
- The hotplug path, including the ecaps reseeds, runs with the hotplug lock
held, which excludes the scx_root writers. Add scx_root_protected(), which
accepts either the hotplug lock or scx_enable_mutex.
- Is-root tests use a zero level instead of comparing against the global.
touch_core_sched_dispatch() stays naked, to be resolved by a later patch.
Signed-off-by: Tejun Heo <tj@kernel.org>
cgrp->scx_sched is __rcu and published with rcu_assign_pointer() but every
reader loads it with a plain access, so sparse flags all of them. The reads
are lock-protected: enable/disable paths rewrite the field under all of
scx_enable_mutex, scx_fork_rwsem and cgroup_mutex, and cgroup creation
inherits the parent's sched under cgroup_mutex before the new cgroup is
reachable, so holding any one of the three locks makes the read stable.
Add scx_cgroup_sched() which states the protection with
rcu_dereference_check() and convert the readers. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_bpf_cidperf_set() reaches cpufreq with no cap check, so any cid-form
sub-sched can steer the frequency of any cid in its view, including ones it
holds nothing on.
Gate it behind a new SCX_CAP_PERF rather than SCX_CAP_BASE: hardware control
is a separate axis from queue access - a parent may well delegate scheduling
on a cid without handing over its frequency. PERF neither implies nor is
implied by the other caps. The check runs under the target rq's lock, which
ecaps updates are also folded under, so it is authoritative - a write can
never land after a revoke has taken effect. Denials are counted in
SCX_EV_SUB_CIDPERF_DENIED.
The operation is synchronous and the outcome is reported to the caller:
scx_bpf_cidperf_set() now returns 0 or -errno, -EACCES on denial. The
cid-form interface is still under initial development, so the signature is
changed in place without versioning.
scx_qmap grants PERF alongside its existing cid grants so the cpuperf demo
keeps working in sub-scheds.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Factor the cpuperf target write out of scx_bpf_cpuperf_set() into
scx_cpuperf_set() which takes the acting sched and returns 0 or -errno, and
flatten the nested validation into early returns. No functional change.
Prep for gating the write behind a cap and reporting the outcome from the
cid-form kfunc.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
kick_one_cpu() silently skips a kick when the kicking sub-sched lacks
SCX_CAP_BASE on the target cid, as does kick_one_cpu_if_idle() for idle
kicks. The skips are sound with the same logic as the reenq gate but are
invisible today, unlike the preempt degradation counted in
SCX_EV_SUB_PREEMPT_DENIED. Count them in a new SCX_EV_SUB_KICK_DENIED event
so every cap denial is observable.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_bpf_dsq_reenq() with an SCX_DSQ_LOCAL_ON target schedules deferred reenq
work on the cid's cpu, raising an IPI when the target rq isn't the locked
one. Nothing checks caps along the way, so a sub-sched holding no cap at all
on a cid can force its cpu to take IPIs and rq lock cycles at will. The
analogous scx_bpf_kick_cid() path gates delivery on SCX_CAP_BASE in
kick_one_cpu() to prevent exactly this.
Apply the same rule at the reenq scheduling point: if the calling sched
lacks SCX_CAP_BASE on the target cid, drop the reenq and count it in the new
SCX_EV_SUB_REENQ_DENIED event. The check is lockless, which is fine: a reenq
slipping through right after a revoke is harmless, and a wrong denial can't
happen - if the caller has seen its ownership of the cpu, the check sees it
too.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The flags passed to the pick-idle kfuncs are values from the
scx_pick_idle_cpu_flags enum, whose members are prefixed
SCX_PICK_IDLE_ (SCX_PICK_IDLE_CORE, SCX_PICK_IDLE_IN_NODE).
Three kernel-doc comments in idle.c erroneously used
%SCX_PICK_IDLE_CPU_* which does not correspond to any defined flag
name, while the adjacent scx_bpf_pick_idle_cpu_node() correctly
documents %SCX_PICK_IDLE_*.
Fix the three occurrences to use the correct SCX_PICK_IDLE_* prefix.
Signed-off-by: Liang Luo <luoliang@kylinos.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
Add missing function parameter descriptions and use the correct
function name in kernel-doc comments to avoid kernel-doc warnings:
Warning: kernel/sched/ext/ext.c:2692 function parameter 'sch' not described in 'finish_dispatch'
Warning: kernel/sched/ext/ext.c:5309 function parameter 'stalled_mask' not described in 'scx_rcu_cpu_stall'
Warning: kernel/sched/ext/ext.c:5405 function parameter 'cpu' not described in 'scx_hardlockup'
Warning: kernel/sched/ext/ext.c:8470 expecting prototype for scx_bpf_dsq_insert(). Prototype was for scx_bpf_dsq_insert___v2() instead
Warning: kernel/sched/ext/ext.c:8784 expecting prototype for scx_bpf_dsq_move_to_local(). Prototype was for scx_bpf_dsq_move_to_local___v2() instead
Warning: kernel/sched/ext/ext.c:9498 expecting prototype for scx_bpf_reenqueue_local(). Prototype was for scx_bpf_reenqueue_local___v2() instead
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
* Avoid running unit tests without amd-pstate
* Fixes for EPP on shared memory systems
* Fixes for dynamic EPP callbacks
* Avoid loading on guests
* Allow lowest nonlinear == minimum freq
-----BEGIN PGP SIGNATURE-----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=rxGv
-----END PGP SIGNATURE-----
Merge tag 'amd-pstate-v7.3-2026-07-22' of ssh://gitolite.kernel.org/pub/scm/linux/kernel/git/superm1/linux
Merge amd-pstate content for 7.3 (07/22/26) from Mario Limonciello:
"* Avoid running unit tests without amd-pstate
* Fixes for EPP on shared memory systems
* Fixes for dynamic EPP callbacks
* Avoid loading on guests
* Allow lowest nonlinear == minimum freq"
* tag 'amd-pstate-v7.3-2026-07-22' of ssh://gitolite.kernel.org/pub/scm/linux/kernel/git/superm1/linux: (923 commits)
cpufreq/amd-pstate: handle missing policy in dynamic EPP callbacks
cpufreq/amd-pstate: Cache the firmware programmed EPP value
cpufreq/amd-pstate: Toggle auto_sel in active mode on shared memory systems
cpufreq/amd-pstate: Fix EPP return type and handle errors during initialization
cpufreq: amd-pstate-ut: Skip tests when amd-pstate driver is not active
cpufreq/amd-pstate: Prevent the driver from loading on unsupported hardware
cpufreq/amd-pstate: Loosen requirement on lowest nonlinear frequency != min freq
Linux 7.2-rc4
Revert "drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs"
drm/amd/display: check GRPH_FLIP status before sending event
drm/amd/display: consolidate DCN vblank/flip handling onto vupdate_no_lock
drm/amd: Create a device link between APU display and XHCI devices
drm/amd/display: wire DCN42B mcache programming callback
drm/amd/display: set new_stream to NULL after release
drm/amd/display: Force PWM backlight on Lenovo Legion 5 15ARH05
drm/amdkfd: free MQD managers on DQM init failures
drm/amdgpu/ttm: Consider concurrent VM flushes for buffer entities
drm/amd/pm/smu7: Fix AC/DC switch notification
drm/amdgpu: Disable PCIe dynamic speed switching on Ryzen Pinnacle Ridge
drm/amdgpu: always emit the job vm fence
...
The nesting depth check and the cgroup online check in
scx_sub_enable_workfn() reach err_disable without setting ret, so
the fallback error added by commit db4e9defd2 ("sched_ext: Record
an error on errno-only sub-enable failure") reports
"scx_sub_enable() failed (0)".
This is currently harmless because both paths record their own
scx_error() first and the first error wins, but it leaves the
fallback broken for these paths. Set -EINVAL and -ENODEV there
so the fallback always reports a real errno.
v2: The validate_ops() path from v1 is already fixed in for-7.3
(sub.c already has ret = scx_validate_ops()), so only the two
remaining paths are addressed.
Signed-off-by: Cui Jian <cjian720@163.com>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
SCX's built-in idle CPU tracking is imperfect and can be out-of-sync
with the actual idle state of CPUs, especially immediately after
enabling SCX due to scx_idle_enable() marking all online CPUs idle.
scx_select_cpu_dfl() skips marking the selected CPU as busy if
the selected CPU is the waker CPU in the SCX_WAKE_SYNC case.
If the waker CPU was marked idle by SCX, it will still be marked idle
after CPU selection and potentially even after switching to the wakee.
In the allowed_cpus selftest, this can manifest as the test failing with
the following message in dmesg:
allowed_cpus.bpf.c:21: CPU 0 should be marked as busy
This patch explicitly marks the waker CPU as busy. With this patch,
the test failure no longer reproduces. There are still some pretty
unlikely races that could make the test fail (e.g. pick_task_idle()
marking the selected CPU idle between selection and validation), but
these can't be fixed easily.
Signed-off-by: Kuba Piecuch <jpiecuch@google.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Use sysfs_emit() instead of sprintf() in rate_limit_us_show().
sysfs_emit() is the preferred API for sysfs output as it provides
PAGE_SIZE bounds checking and ensures proper sysfs formatting.
No functional change intended.
Signed-off-by: Zhongqiu Han <zhongqiu.han@oss.qualcomm.com>
Link: https://patch.msgid.link/20260716131546.1159644-1-zhongqiu.han@oss.qualcomm.com
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Commit 16a03c71bb ("cpufreq: schedutil: Merge initialization code of
sg_cpu in single loop") merged the per-CPU initialization and the
utilization-hook registration into a single loop in sugov_start().
For a shared cpufreq policy this re-introduces the race originally fixed
by commit ab2f7cf141 ("cpufreq: schedutil: Fix sugov_start() versus
sugov_update_shared() race").
The scheduler's util path reaches the hook under RCU-sched and never takes
policy->rwsem, so the rwsem held across sugov_start() cannot serialize the
two. Once the first CPU's hook is published, sugov_update_shared() may run
and, via sugov_next_freq_shared(), read/write each sibling sugov_cpu
(iowait_boost, util, bw_min, ...) concurrently with the memset() still
initializing them, with no lock common to both sides: the update side holds
sg_policy->update_lock while the init side holds only policy->rwsem, which
the scheduler's util path never takes.
The walk only accesses scalar members, never a pointer like ->sg_policy,
so it does not crash today; it merely uses stale (or zero on first start)
values that skew the frequency selection and tracepoints. It is still a
genuine data race, and a latent crash once any pointer member is
dereferenced there.
Restore the two-phase approach: initialize all per-CPU structures first,
and only then publish the per-CPU utilization update hooks.
Fixes: 16a03c71bb ("cpufreq: schedutil: Merge initialization code of sg_cpu in single loop")
Cc: stable@vger.kernel.org
Signed-off-by: Zhongqiu Han <zhongqiu.han@oss.qualcomm.com>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/20260716115159.848403-1-zhongqiu.han@oss.qualcomm.com
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The cid tables are visible to the cid kfuncs while being modified: the
first enable publishes the global pointers before filling them,
ops.init_cids() overrides rewrite them in place, and re-enables rebuild
them in place. A racing TRACING or SYSCALL program can read unfilled
entries, including uninitialized memory in the kmalloc'd tables, or torn
topo updates.
Tie the tables' lifetimes to the root sched instead: each root enable
builds a fresh set privately and publishes the per-table __rcu globals once
the layout is final, and root disable unpublishes and RCU-frees the set. A
non-NULL global is now always a fully built table which stays valid for the
reader's RCU read section, and lookups stay two loads. Kfuncs treat NULL as
no-mapping, also after the scheduler exits instead of reporting the stale
last mapping.
The cid kfuncs are available whether the root scheduler is cid-form or
cpu-form, the latter to allow gradual migration to cids. Every root
therefore builds and publishes a default mapping.
Every reader must either be gated on scheduler liveness or NULL-check
inside an RCU read section. Fix the two kfuncs that were neither:
scx_bpf_this_cid() read the table with no RCU or preemption protection and
scx_bpf_task_cid() relied on KF_RCU, which doesn't put a sleepable program
in an RCU read section. The hotplug callbacks are instead serialized by
retiring the tables inside the cpus_read_lock() section that clears
scx_root.
v2: Document why every root builds the tables (desc + cid.c comment).
Reported-by: Andrea Righi <arighi@nvidia.com>
Closes: https://lore.kernel.org/r/al3tLtPZZkFjMveK@gpd4
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
select_task_rq_scx() falls into the default path when the scheduler has no
ops.select_cpu or is bypassing. There it calls scx_select_cpu_dfl() and
direct-dispatches to the picked CPU's local DSQ.
While bypassing, neither does anything: the enqueue path routes the task to
a bypass DSQ before consulting the direct-dispatch target, so the direct
dispatch never happens, and the CPU pick at most shifts which CPU's bypass
DSQ receives the task. Worse, when the scheduler does its own idle tracking,
the built-in idle cpumasks the pick consults are not even updated, so it
doesn't work anyway.
Return prev_cpu without the default selection while bypassing and let the
bypass enqueue place the task.
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
check_rq_for_timeouts() blames a runnable stall on the task's owner. Under
a sub-scheduler hierarchy the stalled task can be sitting on a DSQ that a
different scheduler has to drain, e.g. an ancestor's bypass DSQ while the
owner is bypassing. The drainer then escapes blame while the owner is
exited, and when the owner's exit is already claimed, nothing actionable is
reported at all.
Blame the DSQ's owning scheduler instead. The local DSQ is consumed by the
cpu itself and keeps blame on the owner. Detection keeps the owner's timeout
and single-scheduler behavior is unchanged.
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Two unrelated things go by "cgroup" in the cid form. Sub-schedulers attach
to cgroups, and the cgroup_*() ops deliver cpu controller events. While the
ops names suggest cgroup2 hierarchy, they actually operate on the cpu
controller.
Rename them to cpuctl_* in struct sched_ext_ops_cid, which has no users
outside scx_qmap yet. The cpu form is deployed ABI and keeps the old names.
The layout is unchanged and the kernel keeps calling through the cpu-form
union view.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Sub-schedulers don't get cgroups yet: every task_group is inited on the root
sched and the routing added by the previous patches always resolves to it.
Add the handover: an enabling sub-scheduler takes over the cgroups in its
subtree and a disabling one returns them to its parent.
scx_cgroup_claim_subtree() runs while the sub enables, after the subtree's
cgrp->scx_sched's are set and before any task is claimed. It inits each
subtree task_group on the sub, exits it from the parent and updates
tg->scx.sched. A failed ops.cgroup_init() unwinds the sub-side inits and
aborts the enable with the parent untouched.
Disabling reverses it with scx_cgroup_return_subtree(): exit each cgroup
from the sub, then re-init it on the parent with the current tg->scx.*
values, resyncing weight and bandwidth changes made while the sub had it.
When a re-init fails, the parent is failed and the remaining task_groups
still transfer uninited and get no cgroup ops - the same punting done for
tasks. The dying parent's own disable moves them onward.
The handover walks include dying but not yet offlined task_groups, the same
as root's bulk walks: a removed cgroup keeps hosting scheduling events until
its dying tasks finish their final context switches, and its
ops.cgroup_exit() must follow the last of them. tg on/offlining is excluded
through cgroup_lock(), so either ordering against an rmdir of a subtree
cgroup delivers balanced init/exit pairs.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
With sub-schedulers claiming cgroup subtrees, cgroup ops must be delivered
to each task_group's sched rather than always to root. Add tg->scx.sched to
track which sched initialized the task_group. It is set and cleared together
with SCX_TG_INITED.
Deliver the ops accordingly:
- ops.cgroup_exit() goes to the sched whose ops.cgroup_init() it pairs with.
- ops.cgroup_prep_move/move/cancel_move() go to the task's sched, and only
for moves that don't re-home the task. A re-homing move is reported
through the ops.exit_task/init_task() pair instead. The cgroups passed to
the move ops can be outside the sched's inited set as the cpu controller
can be coarser than the sub-scheduler topology.
- Knobs of a cgroup belong to the parent, so ops.set_weight/idle/bandwidth()
go to the parent task_group's sched.
All task_groups currently resolve to the root sched, so no behavior changes
until sub-schedulers start claiming cgroups.
While at it, scx_cgroup_init() is restructured so both paths share the
recording.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A task's sched (p->scx.sched) must match its cgroup's owner
(cgrp->scx_sched). cgroup migration breaks the invariant:
scx_cgroup_move_task() only fires root's ops.cgroup_move() and never
re-homes the task, leading to wrong-sched scheduling and, once the stale
sched is freed, a use-after-free.
Hook into the new cgroup task migration events and re-home each task whose
destination cgroup is owned by a different sched. The events map naturally
to the transfer: MIGRATING runs the fallible init for the destination sched,
letting it reject the migration the same way ops.cgroup_prep_move() can,
MIGRATED does the re-home, which can't fail, and CANCELED undoes the init
when the migration falls through.
Pre-commit, the task's task_group still reflects the source, so
__scx_init_task() grows an explicit cgroup argument for the migration path
to hand ops.init_task() the destination cgroup.
Signed-off-by: Tejun Heo <tj@kernel.org>
Closes: https://lore.kernel.org/r/alnxrsexEe_nQwqL@gpd4
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_cgroup_enabled is in the CONFIG_EXT_GROUP_SCHED block. The upcoming
cgroup migration re-homing needs the gate outside the block. Move the
definition and flag flips outside CONFIG_EXT_GROUP_SCHED. No functional
changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Factor out scx_rehome_task() and scx_punt_task() from the sub-disable
re-home loop and scx_fail_parent(). The upcoming cgroup migration re-homing
also needs scx_rehome_task(). No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Pull to receive:
477869bfaf ("sched_ext: Reject setting disallow from init_task outside the enable path")
5f8b69642d ("sched_ext: Take cgroup_lock() first in scx_cgroup_lock()")
8c13364db9 ("sched_ext: Skip sub-disable teardown for never-linked sub-schedulers")
5cdc928598 ("sched_ext: Don't enable non-ext tasks in the sub-sched task loops")
as dependencies for the upcoming cgroup migration patchset and to
resolve the conflicts with the ext.c/sub.c split on for-7.3.
5f8b69642d comments scx_cgroup_lock() which for-7.3 exported for
sub.c. Resolved by keeping the exported version with the comment.
8c13364db9 and 5cdc928598 patch the pre-split sub-sched enable and
disable paths in ext.c which for-7.3 moved to sub.c. Resolved by
applying the never-linked teardown skip and the class gates to sub.c.
Signed-off-by: Tejun Heo <tj@kernel.org>
Root enable and scx_post_fork() enable a task only if it's on the ext class.
Tasks on other classes, possible under an SCX_OPS_SWITCH_PARTIAL root, are
left READY and enabled by switching_to_scx() when they switch over. The sub
enable-commit pass and the sub-disable re-home loop enable unconditionally,
so a fair-class READY task in the subtree becomes ENABLED while not on
sched_ext. A later switch to SCHED_EXT then trips the task state validation
WARN (ENABLED with the previous state not READY) and calls ops.enable() a
second time.
Gate scx_enable_task() on the task's class in both loops.
Fixes: 337ec00b1d ("sched_ext: Implement cgroup sub-sched enabling and disabling")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A sub-scheduler enable can fail before scx_link_sched() links the sched into
the hierarchy, e.g. when the parent is already being disabled, and cleanup
still runs the full scx_sub_disable().
That is racy against root disable: drain_descendants() is the only ordering
between a sub's disable-time task walk and root disable's all-task teardown,
and an unlinked sub is invisible to it. Root's teardown can thus run between
the never-linked sub's drain and its walk, exiting every task to no
scheduler.
The walk then trips the membership WARN and re-homes the exited tasks onto
the dying hierarchy, a use-after-free.
Skip the cgroup ownership reset and the task walk if @sch was never linked,
indicated by the empty ->sibling as unlinking only happens later in the same
function. The membership WARN remains valid: a linked sub is always waited
on by an ancestor's drain.
Fixes: 337ec00b1d ("sched_ext: Implement cgroup sub-sched enabling and disabling")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_cgroup_lock() write-locks scx_cgroup_ops_rwsem and then takes
cgroup_lock(), which can deadlock through kernfs:
scx enable/disable cgroup rmdir cpu.weight write
------------------ ------------ ----------------
cgroup_lock()
percpu_down_write(rwsem)
cgroup_lock()
kernfs_get_active()
percpu_down_read(rwsem)
kernfs_drain()
The enable path waits for the rmdir to release cgroup_mutex. The rmdir,
deactivating the cpu controller's files, waits in kernfs_drain() for the
write's active reference. The write, in scx_group_set_weight(), waits for
the rwsem behind the pending writer.
Take cgroup_lock() first. The set_* paths take no cgroup locks inside the
read side, so a pending write-lock then only waits for read sections that
always run to completion, and no dependency from the rwsem back to
cgroup_mutex remains.
Fixes: a5bd6ba30b ("sched_ext: Use cgroup_lock/unlock() to synchronize against cgroup operations")
Cc: stable@vger.kernel.org # v6.18+
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The p->scx.disallow revert assumes the root enable path, where the switching
loop reads the reverted policy right afterwards and leaves the task off SCX.
The sub-scheduler disable path also reaches it when re-initializing the
returned tasks on a root parent. Nothing reads the policy there: the task is
enabled on root anyway and keeps running on the ext class with a silently
rewritten policy.
Kill the sched instead, matching the fork and non-root branches, and update
the disallow documentation, which equated !fork with the load path and
pointed at a stale debugfs path for nr_rejected.
Fixes: 337ec00b1d ("sched_ext: Implement cgroup sub-sched enabling and disabling")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
With CONFIG_EXT_SUB_SCHED=y but no sub-scheduler attached - the common case
- hot paths still pay for sub-sched bookkeeping. Gate it behind
__scx_has_subs, a static key counting live sub-schedulers, so that a
root-only system stops paying.
Most conversions are simple skip-if-no-sub tests. scx_idle_notify() is
special - it's a hierarchy walk, so give it a fast path which notifies the
root directly using the same tests as the walk. A pending
SCX_RQ_SUB_IDLE_RENOTIFY can be ignored as no sub can be owed one and the
caller clears the flag either way.
Suggested-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
rq->scx.sub_dispatch_prev is sub-sched-only but was left unconditional. Move
it into the CONFIG_EXT_SUB_SCHED block next to ecaps_to_sync and gate its
updates.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_dispatch_sched() is common dispatch machinery and looks out of place in
sub.h, but it needs scx_cpu_arg() from cid.h and can't move into internal.h
without creating a circular include. Add inlines.h on top of internal.h and
cid.h, and move the function there. The function was sub.h's only cid.h
user, so drop that include. Pure code move, no functional change.
v2: Host the function in a new inlines.h instead of at internal.h's tail,
which formed a circular include with cid.h. Drop sub.h's now-unused
cid.h include. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_discard_ecaps_to_sync() waited for balance_one() to consume a dying
sched's queued ecaps sync, polling with resched_cpu() + msleep(). The wait
is unbounded - the ext dl_server forces picks through sustained fair or RT
load only while ext tasks are queued, so an ext-idle cpu monopolized by a
higher class can stall the teardown indefinitely.
Remove the node directly instead: take all queued nodes, drop the dying
sched's and resplice the rest. Consumption runs under the rq lock and batch
nodes read as on-list throughout, so the producer-side dedup stays correct.
A node that an in-flight scx_process_sync_ecaps() batch holds across a
dispatch-induced rq unlock still needs a wait, but one bounded by that batch
completing rather than by a future balance.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The kerneldoc of sugov_iowait_apply() says the IO boost value is increased
in sugov_iowait_apply() and, in the same sentence, that it is decreased by
the same function. That is self-contradictory, and the first part is wrong:
sugov_iowait_apply() only decreases the boost.
The boost is actually increased in sugov_iowait_boost(). Fix the comment to
name sugov_iowait_boost() as the place where the boost is increased, so it
matches the code.
No functional change.
Fixes: fd7d5287fd ("cpufreq: schedutil: Cleanup and document iowait boost")
Signed-off-by: Zhongqiu Han <zhongqiu.han@oss.qualcomm.com>
Reviewed-by: Christian Loehle <christian.loehle@arm.com>
Link: https://patch.msgid.link/20260703092433.4080165-1-zhongqiu.han@oss.qualcomm.com
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
In some cases, the processor may not actually stick to the "desired"
performance level programmed through the driver's .adjust_perf()
callback and may go above it, which may not be desirable (for instance,
there may be a UCLAMP_MAX limit set for the task currently running on
the given CPU which should be respected).
Address that by adjusting the .adjust_perf() callback to take an
additional argument, max_perf, representing the maximum allowed
performance level of the CPU and update the intel_pstate driver to
take that argument into account as appropriate.
Accordingly, adjust cpufreq_driver_adjust_perf() and the other existing
user of .adjust_perf(), which is the amd-pstate driver (but the behavior
of that driver is not changed).
While at it, also update the cpufreq_driver_adjust_perf()
documentation to reflect this change and some previous code
changes that have not been taken into account in it.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Zhongqiu Han <zhongqiu.han@oss.qualcomm.com>
Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>
Link: https://patch.msgid.link/6277654.lOV4Wx5bFT@rafael.j.wysocki
[ rjw: Adjusted Rust function formatting ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
psi_schedule_rtpoll_work() is called locklessly from the scheduler hotpath
and can race psi_trigger_destroy() taking down the last rtpoll trigger under
rtpoll_trigger_lock:
psi_schedule_rtpoll_work() psi_trigger_destroy()
rcu_read_lock();
task = rcu_dereference(rtpoll_task);
rcu_assign_pointer(rtpoll_task, NULL);
timer_delete(&rtpoll_timer);
mod_timer(&rtpoll_timer, ...);
rcu_read_unlock();
synchronize_rcu();
kthread_stop(task_to_destroy);
The group can then be freed with the re-armed timer still pending, and
poll_timer_fn() runs on freed memory.
461daba06b ("psi: eliminate kthread_worker from psi trigger scheduling
mechanism") deleted the timer synchronously after the synchronize_rcu(),
which prevented this but raced trigger creation instead: the deletion could
cancel the timer that a new trigger set armed during the grace period and,
as creation also reinitialized the timer at the time, corrupt it.
8f91efd870 ("psi: Fix race between psi_trigger_create/destroy") moved the
initialization into group_init() and the deletion into the locked section,
trading the creation races for the window above.
Neither placement in the destruction path works. A pending timer firing
while the group is alive is harmless though. poll_timer_fn() just wakes the
rtpoll waitqueue and doesn't re-arm itself. Bind the timer to the group's
lifetime instead and shut it down in psi_cgroup_free(). Nothing can arm it
by then. timer_shutdown_sync() because the timer is never armed again.
Fixes: 8f91efd870 ("psi: Fix race between psi_trigger_create/destroy")
Cc: stable@vger.kernel.org # v5.10+
Reported-by: Sashiko AI <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/all/20260711000434.36C4A1F000E9@smtp.kernel.org/
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Matt Fleming <mfleming@cloudflare.com>
Acked-by: Suren Baghdasaryan <surenb@google.com>
a5b98009f1 ("sched/psi: fix race between file release and pressure write")
made pressure_write() hold cgroup_mutex across psi_trigger_create(), which
forks the psimon kthread for the first rtpoll trigger. As kthread creation
depends on the whole fork path, the commit inadvertently created a lot of
unwanted locking dependencies from cgroup_mutex.
sched_ext got hit by one: its enable path blocks forks and then grabs
cgroup_mutex, so a pressure write racing a scheduler enable deadlocks, with
every other fork piling up behind.
Fix it by splitting trigger creation so that the worker is forked with
cgroup_mutex dropped and the kernfs active reference left broken. The latter
matters because rmdir and cgroup.pressure writes drain active references
under cgroup_mutex. Publishing the trigger last keeps error reporting
synchronous and preserves the of->priv lifetime rules.
The trigger registered in the first stage pins the group's rtpoll machinery
across the unlocked window, leaving only creation races to resolve. The
catch-up poll on installation covers scheduling attempts dropped while there
was no worker.
v2: Retagged sched/psi (was cgroup).
Fixes: a5b98009f1 ("sched/psi: fix race between file release and pressure write")
Cc: stable@vger.kernel.org
Cc: Edward Adam Davis <eadavis@qq.com>
Cc: Chen Ridong <chenridong@huaweicloud.com>
Reported-by: Matt Fleming <mfleming@cloudflare.com>
Closes: https://lore.kernel.org/all/20260710100441.2653477-1-matt@readmodwrite.com/
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Matt Fleming <mfleming@cloudflare.com>
Acked-by: Suren Baghdasaryan <surenb@google.com>
A cid-form scheduler can grant caps to and revoke them from its child
sub-schedulers but has no way to tear one down. Add scx_bpf_sub_kill() to
evict a direct child with a printf-style reason that reaches the child's
scx_exit_info. No exit code is taken because the child is a separate
scheduler whose exit-code semantics the parent cannot know. The child and
its subtree are disabled through the usual async path under a new exit kind,
SCX_EXIT_PARENT_KILL.
The bstr formatting infrastructure in ext.c is exposed through internal.h
with scx_ prefixes so the kfunc, which lives in sub.c, can format the
reason.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_process_sync_ecaps() consumes ecaps syncs while the sched is bypassing
without delivering ops.sub_ecaps_updated(), leaving reported_ecaps stale.
Nothing re-queued a sync when bypass lifted, so a cid whose caps never
change again would never be notified. Attach-time initial grants hit this
every time: they are consumed during the enable bypass window, so a sched
never learned its initial effective caps through the callback.
Re-queue a sync for every (sched, cpu) with an undelivered delta at the
per-cpu bypass exit in scx_bypass(), next to the idle renotify catch-up. The
next balance on the cpu then delivers the pending delta with proper dispatch
context.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
__scx_update_idle() notified only the root scheduler. A sub-scheduler that
holds a cid needs that cid's idle state to place and kick on it.
Deliver ops.update_idle() to every scheduler that holds SCX_CAP_BASE on the
transitioning cid. The root holds every cap, so a real transition always
reaches it.
Real transitions are not enough on their own. A cid that is already idle
when a sub-sched gains baseline access produces no transition, so the new
holder would never learn it is idle. The ecaps sync arms a re-notify on the
gain, and the next idle pick delivers ops.update_idle() to just that sched,
leaving holders that already track the cpu untouched. A matching loss of
baseline access drops any pending re-notify.
Bypass suppresses ops.update_idle() too, so a cpu that goes idle during a
bypass window and stays idle yields no transition to re-deliver on
un-bypass. Arm the same re-notify for every sched leaving bypass. The acute
case is a child granted cids during its own ops.sub_attach(). The grant
lands while the child is bypassed and the notify walk skips it, so on
un-bypass it holds cids it never saw go idle. The root is owed the same and
is armed through a separate per-rq flag, which keeps this working when
sub-schedulers are compiled out.
v2: Gate the idle catch-up in pick_task_idle() to avoid a double ops.update_idle(). (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_local_or_reject_dsq() authorizes a local-DSQ insert against the caps
of the scheduler doing the insert. On the consume/dispatch paths that is
the scheduler running balance_one(), passed down through
scx_consume_dispatch_q() and move_local_task_to_local_dsq(), so the check
is correct.
The remote-move path loses it. move_remote_task_to_local_dsq()
re-activates @p on the destination rq through enqueue_task_scx(), which
reconstructs the scheduler from the task, i.e. @p's owner. When an
ancestor places a descendant's task - e.g. draining a bypassed
sub-scheduler - the owner is a sub-scheduler of the placer, so
authorizing against the owner checks a narrower cap set and can
spuriously reject a task the placer is entitled to run.
Carry the placing scheduler across the activate_task() boundary the same
way enq_flags already are, via a per-rq field set only for the duration
of the re-activation, and have scx_local_or_reject_dsq() authorize
against it. The placer's caps are a superset of the owner's, so this
admits what the placer may run and keeps rejecting what it may not.
v2: Document @sch in move_remote_task_to_local_dsq()'s kerneldoc. (Andrea)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A kick forces a scheduling event on the target cpu, and a preemption also
evicts the running task. Gate both on caps. Any kick requires baseline
access on the cid, and preempting a task the sub-sched does not own -
whether by a SCX_ENQ_PREEMPT insert or a SCX_KICK_PREEMPT kick - requires
the new SCX_CAP_PREEMPT. Gating either alone would leave a hole - the
weakest cap authorizing preempting kicks, or plain kicks disturbing cpus the
kicker has no access to.
Preempting the sched's own subtree is always allowed, and the cap extends
the right to any task on the cid. PREEMPT implies ENQ, and so ENQ_IMMED.
A preempting insert tests the running task under the target rq lock and is
rejected and reenqueued unless the victim is in the inserter's subtree or it
holds PREEMPT. A migration-disabled task is admitted regardless, but with
SCX_ENQ_PREEMPT stripped.
Kicks are enforced on the delivery path, where the effective caps can be
read coherently under the target rq's lock. A kick from a sub-sched lacking
SCX_CAP_BASE on the cid is dropped, and a SCX_KICK_PREEMPT kick without
PREEMPT for a task outside the kicker's subtree degrades to a plain
reschedule.
Unlike the enqueue caps, PREEMPT is checked only at the instant of the
insert or kick, never as a standing property of a queued task.
v2: Clear SCX_ENQ_PREEMPT on the offline and migration_pending force-admits.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Add SCX_CAP_ENQ, which gates inserting tasks onto a cid's local DSQ. Unlike
IMMED enqueue, plain enqueues can pile up, so ENQ is the stronger cap and
implies ENQ_IMMED. Losing ENQ also triggers the reenq scan. The scan tests
each queued task and the running task against the cap each needs via
scx_caps_for_task(), so an ENQ-only loss reenqueues plain tasks, evicting a
running one, while IMMED tasks, which need only ENQ_IMMED, stay put.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A task's slice grants it cpu occupancy - how long it holds its cpu. In a
sub-scheduler hierarchy cpu access is delegated through revocable
capabilities, so a task's occupancy must follow them. Only its own scheduler
sets its slice, and extending the slice is allowed only while that scheduler
holds baseline cpu access (SCX_CAP_BASE) on the cpu. Otherwise a scheduler
could keep occupying a cpu it has been denied simply by handing out long
slices.
The cap check reads effective caps, which are coherent only under the task's
rq lock, and the kernel decrements the slice under that lock as the task
runs, so a running task's slice can be changed only there while a queued
task's can be set directly. Make scx_bpf_task_set_slice() apply the slice
under the rq lock. Synchronously when the caller already holds it, otherwise
by stashing it in the new p->scx.slice_oob, tagged with the scheduler's id
so a request that outlived a reassignment is dropped. Whether the caller
holds @p's current rq lock is tested with p->scx.runnable_cpu.
Revocation is enforced through the same grant. When a cpu's effective caps
lose SCX_CAP_BASE, the cap-revoke reenq scan also checks the running task
and zeroes its slice to evict it. The scan runs as a balance callback after
the pick, so this catches both the task that was running when the revoke
landed and a capless task the pick just promoted off the local DSQ. The
paths that keep a task on its cpu - holding on to the last runnable task in
balance, the ENQ_LAST reinsertion and the slice refill on pick - skip tasks
lacking baseline access. A migration-disabled task is exempt, mirroring its
capless admission on insert.
v4: Test rq ownership with p->scx.runnable_cpu, closing a remote-wakeup TOCTOU. (sashiko AI)
v3: Keep a pending out-of-band slice request across refill and preserve. (sashiko AI)
v2: Only write slice directly when @p is queued on the held rq. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Add p->scx.runnable_cpu, the cpu @p is runnable on, or -1 when it is not.
It is stamped as @p joins the runnable_list (set_task_runnable()) and
cleared as it leaves (clr_task_runnable()), both under the rq lock.
task_cpu() can't answer "is @p on this rq" reliably: a remote wakeup changes
it under @p's pi_lock alone, without the source rq lock, so it can read as
the locked rq while @p is really elsewhere. runnable_cpu changes only under
the rq lock, so a caller holding an rq lock can compare against it to know
whether that is @p's current rq.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A later change makes set_task_slice() also drop a pending out-of-band slice
request, so the BPF-triggered writes to p->scx.slice need to funnel through
one place. Introduce set_task_slice() and route those writes through it.
update_curr_scx() decrements curr->scx.slice directly for accounting and is
left alone. No functional change - the helper only assigns p->scx.slice.
v2: Reword the set_task_slice comment to "BPF-triggered writes". (Andrea)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Neither a scx_sched pointer nor its cgroup id uniquely identifies a
scheduler instance. A freed sched's memory can be reallocated, and a cgroup
can detach one sched and attach another. Add a monotonic, never-reused u64
id. A later patch compares it to drop a slice request that outlived a change
of a task's owning scheduler.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Replace the __SCX_CAP_DUMMY placeholder with SCX_CAP_ENQ_IMMED, which gates
inserting IMMED tasks onto a cid's local DSQ. An IMMED enqueue is guaranteed
to either get its task running on the cpu at once or hand it back to the
scheduler, so IMMED work can never pile up on the cpu's queue and a cpu can
be shared across sub-scheds through IMMED access without any of them
swamping it.
That makes ENQ_IMMED the natural baseline, the minimal cap to make any use
of a cpu. SCX_CAP_BASE aliases it so gates on basic cpu access can state the
intention instead of naming ENQ_IMMED.
Enforcement covers inserts and queued tasks. An insert without the cap is
diverted to the reject DSQ, and queued tasks are reenqueued when the cap is
lost. scx_bpf_sub_dispatch() skips a child that lacks the cap on the cpu, as
its inserts would only be rejected. Vacating the running task on cap loss
lands in a later patch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A SAVE/RESTORE requeue re-inserts a running task in place and is immediately
followed by set_next_task_scx(). It is not a real scheduling event: the task
is already admitted to its cid and must return to the local DSQ
unconditionally.
scx_caps_for_enq() maps an enqueue to the cap its local-DSQ insert requires.
Add SCX_ENQ_IGNORE_CAPS, set it on the RESTORE-in-place branch of
enqueue_task_scx(), and have scx_caps_for_enq() require no caps for it, so
the cid admission gate never diverts an in-place restore to the reject DSQ.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
When a sub-scheduler dispatches a task to a CPU it lacks the required
capability on, the task must be rejected rather than allowed to run.
Add the machinery for that. Each rq gets a reject DSQ, a kernel-internal
holding queue that is never run and that the BPF scheduler cannot reach. An
insert that must be refused is diverted there instead of the local DSQ, and
a deferred requeue then hands the parked tasks back to the BPF scheduler to
re-decide. A cap revoke extends this to already-queued tasks. When the
revoke reaches the cpu's effective caps, the cpu scans its local DSQ and
reenqueues the tasks that no longer qualify.
A migration-disabled task must run on its cpu, so a capless one is admitted
anyway and counted in the new SCX_EV_SUB_FORCED_ADMIT event.
This is preparation for the actual sub-sched cap enforcement. The divert is
wired but inert here.
v2: Admit offline-rq and migration_pending inserts to local, not reject. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The local DSQ is synchronized by the containing rq lock rather than its own
dsq->lock. A later patch adds a second such DSQ. In preparation, factor the
"rq owns the lock" test into dsq_is_rq_owned() and rename
local_dsq_post_enq() to rq_owned_post_enq(), taking @rq explicitly.
No behavior change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A sub-scheduler that gains or loses effective caps on a cpu may want to act
on it right away - e.g. place or preempt on a newly usable cpu. The existing
ops.sub_caps_updated() doesn't fit as it is delivered asynchronously to
scheduling operations and can arrive before the per-cpu effective caps go
live.
Add ops.sub_ecaps_updated(cid, before, after), a cid-form callback fired
from scx_process_sync_ecaps() when a sub-sched's effective caps on a cid
change. It runs in dispatch context so the sched can insert, kick or preempt
on the cid directly. @before is the caps as of the last delivery.
Cpu hotplug rides the same machinery. Going down zeroes each sched's ecaps
on the cpu's cid, with queued syncs discarded at consumption while the cpu
is inactive. Coming back up queues a sync for every sched. reported_ecaps is
kept across the down/up cycle, so the resync fires the callback only if
ownership actually changed while the cpu was down.
v2: Compute cid below the active-cpu guard; discard queued syncs on !cpu_active(). (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Checking a sched's caps on a cid would need to test several cap bits against
caps[] to account for implied caps. Also, caps[] modifications aren't
synchronized against scheduling operations on each cpu, which can lead to
awkward race conditions.
Collect them per cpu instead. caps[] under pshard->lock stays the target
configuration. scx_sched_pcpu->ecaps is added, the transposed effective
copy: the set of cap bits the sched holds on that cpu which can be accessed
with a single read. It is stable under the rq lock. It can also be read
locklessly with READ_ONCE().
Grant and revoke only mutate caps[]. They queue a sync request on the target
cpu's rq->scx.ecaps_to_sync and kick it, and the cpu recomputes the queued
scheds' ecaps from caps[] in balance_one() under its own rq lock. A dying
sched runs the sync directly to retire its queued request before freeing. As
held references can defer the freeing past the enclosing root scheduler's
lifetime, root enable discards leftover sync requests before going live.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Wire up ops_cid.sub_caps_updated() to notify sub-scheds of cap changes.
Three constraints shape the design:
1. Static memory. Deliveries use a fixed-size buffer, both for runtime
efficiency and so notifications can't be lost under memory pressure.
2. High-frequency updates. Grant/revoke can mutate caps in bursts, and the
notifier path must absorb that without amplifying it.
3. Recursive grant/revoke from the callback. A child receiving a
notification can call grant/revoke on its own children, which can
cascade recursively down its subtree.
(1) and (2) lead to coalescing into a fixed payload. Each delivery carries a
single (cmask, caps) pair covering every change since the previous one.
Direction (set vs cleared) isn't encoded as it doesn't fit in the fixed-size
summary. The callback queries scx_bpf_sub_caps() for current state. Only one
delivery is in flight per shard. Further changes fold into the same buffer
and ship as the next callback, so a shard's callbacks fire in order.
(3) leads to deferred delivery. Events accumulate during grant/revoke and
are delivered after the shard lock is released.
v2:
- Request a private stack for ops.sub_caps_updated(). (sashiko AI)
- Build cmask_arena_out via scx_cmask_ref, not by re-reading its header.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Caps are per-cid permissions parents delegate to direct children via
scx_bpf_sub_grant() / scx_bpf_sub_revoke(). A child's cap set is always a
subset of its parent's. Sub-scheds check their caps locally, and cross-sched
communication is needed only when the delegation set itself changes.
Caps will be used to implement sub-sched scheduling on the enqueue path.
Picking a cid for a task at a leaf depends on which cids the leaf is allowed
to use, and resolving that programmatically on every enqueue would mean a
cross-sched round-trip call chain, possibly retrying if the request can't be
granted as-is. The dispatch path is different - it runs as top-down
recursion via scx_bpf_sub_dispatch().
Locking is per shard. cid space is split into shards, and each sub-sched has
its own pshard->lock for each shard. Operations are broken up on shard
boundaries. Different shards never contend. Shards are expected to be
topology-aligned and likely to serve as the locality unit when cids are
allocated to schedulers, so per-shard lock granularity scales naturally with
the allocation pattern.
This patch adds the framework with a single dummy cap. Real caps land in
later patches.
The enable path is reordered for pshards. scx_arena_pool_init() moves ahead
of scx_link_sched() so the pshards are allocated before the sched becomes
reachable - scx_alloc_pshards() skips allocation when the arena pool isn't
initialized.
- scx_bpf_sub_grant(): Per-cid all-or-nothing grant to direct child.
- scx_bpf_sub_revoke(): Clear caps on @cmask across @child and its subtree.
- scx_bpf_sub_caps(): Lockless snapshot of caps on a cid range.
/sys/kernel/sched_ext/SCHED/caps shows the caps each scheduler currently
holds.
v4: Move the pshard[] full build/publish and the err_disable scx_error() recording to earlier patches. (sashiko AI)
v3: Build pshard[] fully before publishing it, read it with READ_ONCE. (sashiko AI)
v2: Validate ops before scx_link_sched() publishes the sub. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A scheduler's BPF programs can outlive it. A timer it armed or a tracing
program it loaded can fire after ops.exit() has run, before the programs are
unloaded, and scx_prog_sched() still resolves the program to its scheduler
through ops->priv. Harmless while kfuncs touch only lifetime-stable state,
but a hazard once a kfunc reads global state a newly loaded scheduler can
change underneath it.
Add scx_sched->dead, set right after ops.exit() and drained with
synchronize_rcu(). It follows exit() rather than preceding it so exit()'s own
kfunc calls still resolve to @sch. scx_prog_sched() returns NULL for a dead
scheduler, so every kfunc's existing !sch bail rejects it at one choke
point.
v2: Check dead in the CONFIG_EXT_SUB_SCHED=n scx_prog_sched() too. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Factor the sibling/ancestor portion of scx_next_descendant_pre() out as
scx_skip_subtree_pre(), a pre-order walk primitive that skips @pos's
subtree, and call it from scx_next_descendant_pre(). Same locking rules as
the existing primitive.
Used in a follow-up to fast-skip subtrees that have nothing to do during a
descendant walk.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Future kfuncs need to walk descendants without scx_sched_lock. Make the
walker RCU-safe so that they can. A sub-sched's fields are initialized
before it is linked, so a walk that observes a linked node also observes its
setup. In-place changes after linking carry their own ordering.
Switch the children/sibling list ops to RCU and expand the descendant walker
to accept rcu_read_lock as a valid read-side context. Walkers that mutate
keep scx_sched_lock.
A sub-sched can be linked while an ancestor is bypassing, after the bypass
walk that propagates the depth has passed its parent. Bypass state is a
per-cpu flag plus a depth count and can't be established atomically at link
time, so refuse to link under a bypassing ancestor. Take scx_bypass_lock
across linking to check the parent's bypass state coherently.
v3: Reject linking under a bypassing ancestor instead of inheriting bypass_depth. (sashiko AI)
v2: Inherit bypass_depth before publishing @sch on the RCU sibling list.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_call_op_set_cpumask() builds a per-cpu cmask in the set_cmask scratch,
which lives in BPF-writable arena. A scheduler can corrupt the scratch's
inline header (base, nr_cids, alloc_words) from another cpu, so sizing and
indexing the write from it risks an out-of-bounds write.
Drive the build from kernel-known geometry instead.
scx_cmask_ref_init_kern() imposes base and nr_cids rather than reading them,
and scx_cmask_ref_from_cpumask() fills the scratch from the ref. Neither
reads the header back.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
kfuncs taking struct scx_cmask * from BPF arena memory have two problems.
The pointer can be any value the BPF prog hands in, and the header (@base,
@nr_cids, @alloc_words) can be mutated by the prog concurrently with kernel
access.
Add scx_cmask_ref, a validated handle. _init() normalizes the input pointer
into the arena's kern_vm range via scx_arena_to_kaddr() and snapshots the
header, rejecting a range outside the machine or a nr_cids whose words
exceed the declared @alloc_words. Downstream sizing uses the snapshot, not
the live header. _shard() reads slices while _or() and _copy() write back,
all bounded by the snapshot. No callers yet.
struct scx_cmask's bits[] carried __counted_by(alloc_words), so
UBSAN_BOUNDS and FORTIFY_SOURCE bound accesses to the array. That bound is
read from @alloc_words at the access. For an arena cmask @alloc_words is
BPF-writable. A prog that sets it larger than the real allocation makes the
check pass on a genuine overrun, so the annotation catches nothing, and it
only runs under those debug configs. Drop it - _init() validates
@alloc_words explicitly, and kernel-owned cmasks set it themselves.
v2: Validate @alloc_words in _init(), drop __counted_by. (Andrea, sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Add struct scx_pshard and sch->pshard[] indexed by shard_idx, each entry
allocated on its shard's NUMA node from scx_shard_node[si]. The struct
starts empty (one dummy field). Follow-up patches will grow it as
shard-local state lands. Only cid-type schedulers with an arena pool get
pshards.
Allocation happens after ops.init_cids() returns so any
scx_bpf_cid_override() it issues has finalized scx_nr_cid_shards and
scx_shard_node[]. sch->nr_pshards records the array size for the async RCU
free path, which may run after a later scheduler's scx_cid_init() has
rewritten the global.
v3: Build and publish pshard[] fully-formed here rather than a later patch.
v2: Free the partially-allocated pshard array on alloc failure. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Split kobject_init_and_add() in scx_alloc_and_add_sched(): only
kobject_init() runs there. A new scx_sched_sysfs_add() helper does
kobject_add() (and creates sub_kset when the scheduler implements
ops.sub_attach), called by both enable workfns once @sch is linked and its
sysfs-visible state is initialized. Prep so a future caps attribute can rely
on @sch being fully built by the time it's sysfs-visible. Add early enough
that a stall later in enable still leaves sysfs inspectable.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
An overridden cid mapping invalidates the auto-generated shard layout, so
the override call has to provide both. Extend scx_bpf_cid_override() with a
shard_start[] array that lists the first cid of each shard (starting at 0,
strictly increasing, last shard implicitly extends to num_possible_cpus()).
A scheduler that wants only custom shards with the auto-generated cid
mapping can read the current mapping and pass it back unchanged.
Overridden shards can span NUMA nodes, so scx_shard_node[] is rebuilt by
majority count: each shard is assigned to the node that owns the most cpus
in it.
v2: Snapshot the caller's cpu_to_cid/shard_start arrays before validating. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Sub-sched operations need a scalable locking / work domain smaller than the
whole cid space. Carve the cid space into topology-respecting shards: each
shard is a contiguous cid range that stays within one LLC, and LLCs larger
than the per-shard cap (default 24 cids, configurable via
ops.cid_shard_size) split into enough shards to fit. A hard cap of
SCX_CID_SHARD_MAX_CPUS prevents pathological sizes under custom
configurations.
No-topo cids pack into their own shards so every cid has a shard assignment.
Also build scx_cid_shard_ranges[] for O(1) shard-to-cid-range lookup and
scx_shard_node[] so callers can size or place work by NUMA without walking
cids. Auto-built shards inherit their LLC's node. No-topo shards carry
NUMA_NO_NODE.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A cid-form scheduler that calls scx_bpf_cid_override() to install a custom
cid layout can only do so from ops.init(). Enable-path setup that depends on
the cid layout thus has to run after ops.init(), and ops.init() itself can't
use anything derived from the final layout, which turned out to be too
restrictive.
Add an ops.init_cids() callback dedicated to finalizing the cid layout. It
runs before the rest of the enable-path setup, so the final layout is in
effect for everything that follows including ops.init(), which now runs
after the arena pool and cmask scratch allocations.
scx_bpf_cid_override() is restricted to ops.init_cids() at load time. It
sits in a kfunc set gated by SCX_KF_ALLOW_INIT_CIDS, a flag set only on the
init_cids op, so the verifier rejects a call from any other context. The
runtime root-only check is dropped as ops.init_cids() only runs during root
enable.
The qmap demo moves its override into a dedicated qmap_init_cids() and,
while at it, introduces an enum for the cid override modes instead of
hard-coded integers.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The kick machinery kept its targets in rq->scx shared by every sched on the
cpu. A preempt kick carried no record of which scheduler requested it.
A later patch needs preempt kicks scoped to the requesting scheduler so a
sub-scheduler can preempt only tasks in its own subtree. Move the kick masks
into the per-sched per-cpu scx_sched_pcpu and have scx_kick_cpu() link the
sched onto a per-cpu list (rq->scx.sched_pcpus_to_kick). The cpu's single
kick irq_work walks that list and kicks each sched's targets on its behalf,
so a kick stays attributed to its scheduler.
The SCX_KICK_WAIT sync set (cpus_to_sync, the kick_sync snapshot and the
balance-callback trigger) stays in rq->scx: the waiter is the cpu, not the
scheduler, and its only writers, the kick irq_work and the wait balance
callback, are cpu-local.
On disable, free_kick_syncs() flushes each cpu's pending kick irq_work
before clearing @ksyncs, so a late kick unlinks its to_kick_node instead of
early-returning on a NULL @ksyncs and leaving the node linked at free.
v3: Flush the kick irq_work in free_kick_syncs() before clearing @ksyncs. (sashiko AI)
v2: Warn once per sched on scx_bpf_kick_cpu() from NMI.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A per-task op must be dispatched on the scheduler that owns the task.
SCX_CALL_OP_TASK() and its _RET twin take @sch explicitly, and a caller that
passes the wrong scheduler would silently run the op on it. Add a
WARN_ON_ONCE() that @sch matches the task's owner so such a mismatch is
caught rather than hidden.
Two sites legitimately target a scheduler other than the task's owner:
cgroup_move() runs on the root sched, and scx_sub_init_cancel_task() fires
exit_task() on a task not yet associated with @sch. Both switch to the inner
__SCX_CALL_OP_TASK(), which dispatches on the explicit @sch without the
assert.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
- Lifecycle fixes for the new sub-scheduler support: two use-after-frees
and an enable-failure path that left a half-initialized sub-scheduler
linked.
- Two dispatch-path locking bugs: a spurious scheduler abort from a
migration race, and a lockdep splat from stale runqueue-lock tracking.
- Callback and task-state fixes: stale scheduler-owned state on a task
leaving SCX, a weight callback running after disable, and a bogus
warning on core-scheduling forced idle.
- On nohz_full, finite-slice tasks could miss the tick that expires their
slice. Enable it when such a task is picked, with a selftest.
- Smaller fixes: userspace CPU-mask helpers, ratelimited deprecation
warnings, docs and a sparse annotation.
-----BEGIN PGP SIGNATURE-----
iIQEABYKACwWIQTfIjM1kS57o3GsC/uxYfJx3gVYGQUCalVWgw4cdGpAa2VybmVs
Lm9yZwAKCRCxYfJx3gVYGTmUAP0QHX7Ae/g7yMPEB4slURpnSj/wE7hRCI5jTmay
2iIEJgD/RAhpwUUAp4Abozt0mjQMWh9UyEVjOxPCVNKaNh1XfwU=
=EBCm
-----END PGP SIGNATURE-----
Merge tag 'sched_ext-for-7.2-rc3-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext
Pull sched_ext fixes from Tejun Heo:
- Lifecycle fixes for the new sub-scheduler support: two
use-after-frees and an enable-failure path that left a
half-initialized sub-scheduler linked.
- Two dispatch-path locking bugs: a spurious scheduler abort from a
migration race, and a lockdep splat from stale runqueue-lock
tracking.
- Callback and task-state fixes: stale scheduler-owned state on a task
leaving SCX, a weight callback running after disable, and a bogus
warning on core-scheduling forced idle.
- On nohz_full, finite-slice tasks could miss the tick that expires
their slice. Enable it when such a task is picked, with a selftest.
- Smaller fixes: userspace CPU-mask helpers, ratelimited deprecation
warnings, docs and a sparse annotation.
* tag 'sched_ext-for-7.2-rc3-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext:
sched_ext: Skip ops.set_weight() for disabled tasks
tools/sched_ext: scx - Fix cmask_subset(), cmask_equal() and cmask_weight()
sched_ext: Fix premature ops->priv publication in scx_alloc_and_add_sched()
sched_ext: Record an error on errno-only sub-enable failure
selftests/sched_ext: Verify nohz_full tick behavior
sched_ext: Enable tick for finite slices on nohz_full
sched_ext: Preserve rq tracking across local DSQ dispatch
sched_ext: Documentation: Fix ops table header reference
sched_ext: Don't warn on core-sched forced idle in put_prev_task_scx()
sched_ext: Pin parent scx_sched across a child sub-scheduler's lifetime
sched_ext: Annotate ksyncs with __rcu in alloc/free_kick_syncs()
sched_ext: Check remote rq eligibility under task's rq lock
sched_ext: Reset dsq_vtime and slice when a task leaves SCX
sched_ext: Avoid flooding the log with deprecation warnings
When switching a task's sched_class away from sched_ext, we get the
following sequence of events in __sched_setscheduler():
sched_change_begin()
switched_from_scx()
scx_disable_task(p)
ops.disable(p)
__setscheduler_params()
set_load_weight()
reweight_task_scx(p)
ops.set_weight(p)
p->sched_class = next_class;
sched_change_end()
...
Notably, ops.set_weight() is called _after_ ops.disable().
This violates the expected semantics of the callbacks, the expectation
being that ops.disable() can only be followed by ops.exit_task() or
ops.enable().
Skipping the weight adjustment for disabled tasks should be harmless
since the weight will be recalculated in scx_enable_task() if the task
ever rejoins SCX.
Fixes: 637b068282 ("sched: Fold sched_class::switch{ing,ed}_{to,from}() into the change pattern")
Cc: stable@vger.kernel.org # v6.19+
Signed-off-by: Kuba Piecuch <jpiecuch@google.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
sched_ext schedulers have state in BPF programs and kernel. scx_dump
provides kernel state and BPF program state on error, but this is static
in what it can provide.
Add a sched_ext_exit tracepoint in scx_claim_exit() so that BPF programs
can dynamically inspect scheduler specific state at the moment of exit.
Pass the exiting scx_sched so attached programs can read its state, and,
since exits propagate through a hierarchy of sub-schedulers, identify
which scheduler each event belongs to.
Signed-off-by: Pat Somaru <patso@likewhatevs.io>
Signed-off-by: Tejun Heo <tj@kernel.org>
consume_remote_task() drops this_rq before unlinking a remote task from
its DSQ and locking src_rq. When called from ops.dispatch(),
scx_locked_rq() keeps pointing to this_rq across the lock dance. The
subsequent switch_rq_lock(src_rq, this_rq) cannot update the tracking
because its guard does not match, leaving the tracking stale while
src_rq is held.
Keep this_rq locked until the task has been unlinked and the DSQ lock
released, then use switch_rq_lock() to switch directly to src_rq. Use
the same helper to restore this_rq after losing the dequeue race. The
successful path already switches back through
move_remote_task_to_local_dsq(), so scx_locked_rq() now follows the
actually held rq throughout the consume path.
Suggested-by: Tejun Heo <tj@kernel.org>
Link: https://lore.kernel.org/all/455e701bca66bdecde530d225f4dba0a@kernel.org
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_bpf_kick_cid() returned an error code, but the value conveys nothing
actionable and no caller consumes it. The kick is asynchronous, so a
successful return only means it was queued. An invalid @cid is already
reported through scx_error() by scx_cid_to_cpu(), and a missing scheduler
leaves nothing to kick.
Make scx_bpf_kick_cid() return void to match scx_bpf_kick_cpu(). The
cid-form kfuncs are not in practical use yet, so the ABI change is safe.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Direct writes to p->scx.slice and p->scx.dsq_vtime bypass
scx_bpf_task_set_slice/dsq_vtime() and the authority checks they carry.
Those checks exist for sub-schedulers, which attach only through the
cid-form struct_ops, so the direct writes only need to be closed there.
Give sched_ext_ops_cid its own verifier ops that reject the two fields.
cid-form is a new interface with no legacy users, so there is no
compatibility to keep. The cpu-form keeps direct writes, and the deprecation
warning they carried is dropped.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_rq.extra_enq_flags carries scx-specific enqueue flags across the
activate_task() boundary during a cross-rq task move in
move_remote_task_to_local_dsq(). Rename it to remote_activate_enq_flags to
name that role, and fix the stale comment that referenced the old
move_task_to_local_dsq() name.
No functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The cid-form ops overlay their cpu-form siblings at the same struct slot.
Ops whose signature matches the sibling are invoked through the cpu-form
call sites unchanged, but set_cmask() takes an arena cmask address rather
than a cpumask, so scx_call_op_set_cpumask() calls ops_cid.set_cmask()
directly and hand-rolled the kf_tasks[] and locked_rq bracket that
SCX_CALL_OP_TASK() provides. The hand-rolled bracket reset locked_rq to
NULL on exit instead of restoring the saved value, so a nested call would
clobber the outer op's locked-rq tracking.
Parameterize the dispatch macros by the ops-table member and add
SCX_CALL_CID_OP_TASK(), which routes through sch->ops_cid. Convert
scx_call_op_set_cpumask() to it and drop the hand-rolled bracket. The only
behavioral change is that locked_rq is now saved and restored like every
other op call site.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The cmask ops can operate on BPF-arena cmasks which BPF programs may read
and write concurrently. The _RACY op variants existed to make such lockless
reads explicit but this turned out to be too restrictive. Mark the word
accesses in all the two-cmask ops with READ_ONCE/WRITE_ONCE instead and drop
the _RACY variants.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Pull to receive:
db4e9defd2 ("sched_ext: Record an error on errno-only sub-enable failure")
49b3378a75 ("sched_ext: Fix premature ops->priv publication in scx_alloc_and_add_sched()")
e6979d05c6 ("tools/sched_ext: scx - Fix cmask_subset(), cmask_equal() and cmask_weight()")
for further sub-sched changes and to resolve the conflicts with the
sub-sched updates on for-7.3.
db4e9defd2 adds scx_error() to the sub-enable err_disable sink which
for-7.3 moved from ext.c into sub.c. Resolved by applying the fix to
scx_sub_enable_workfn() in sub.c.
49b3378a75 drops RCU_INIT_POINTER() from an scx_alloc_and_add_sched()
unwind label whose body changed with for-7.3's stall_cpus addition.
Resolved by dropping the line from the updated unwind.
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_alloc_and_add_sched() publishes @sch through ops->priv before allocating
the cgroup path. If that allocation fails, the unwind path clears ops->priv
and frees @sch immediately. scx_prog_sched() callers can dereference
ops->priv from RCU context the moment it is set, so freeing without a grace
period can use-after-free a concurrent kfunc caller.
Move the publication below the cgroup path allocation so that every failure
path after publication frees @sch through kobject_put(), whose release path
defers the freeing by a grace period.
Fixes: 105dcd005b ("sched_ext: Introduce scx_prog_sched()")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_sub_enable_workfn() has several failure paths that only return an errno
(e.g. -ENOMEM from an allocation) and jump to err_disable without calling
scx_error(). scx_flush_disable_work() runs the disable, and thus ops.exit(),
only when an error has been recorded, so an errno-only failure leaves the
half-initialized sub-scheduler linked.
Record an error at the err_disable sink so every errno-only failure runs the
disable path.
Fixes: ebeca1f930 ("sched_ext: Introduce cgroup sub-sched support")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The comment for scx_bpf_dsq_insert() references "@dsp_id" in the
description body, but the parameter is named "@dsq_id" in both the
parameter list and the function signature.
Signed-off-by: Liang Luo <luoliang@kylinos.cn>
Acked-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
set_next_task_scx() updates the tick dependency before __schedule()
updates rq->curr. When switching from a non-EXT task, such as idle, to
an EXT task with a finite slice, sched_update_tick_dependency() checks
the outgoing task and can allow the tick to remain stopped.
The dependency can also be lost without a slice-type transition. After a
finite-slice task leaves the CPU idle, the enqueue path can clear the
dependency against the idle rq->curr. SCX_RQ_CAN_STOP_TICK still records
a finite slice, so another finite task skips the transition block and
can run without the ticks needed to expire its slice.
The reverse mismatch can also happen when the last finite-slice EXT task
is dequeued: sub_nr_running() updates the dependency before rq->curr
changes, so the outgoing task state can keep the dependency set after
the CPU goes idle.
Fix this by unconditionally enabling the scheduler tick whenever a
finite-slice EXT task is selected on a nohz_full CPU. Moreover, when the
last runnable EXT task leaves, ignore the outgoing EXT slice state so
the generic scheduler can correctly re-evaluate and clear the tick
dependency.
Fixes: 22a920209a ("sched_ext: Implement tickless support")
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
dispatch_to_local_dsq() can run from scx_bpf_dsq_move_to_local() while
ops.dispatch() has recorded the current rq. Moving a task to a local DSQ
may switch to the source or destination rq before synchronously invoking
ops.dequeue() through the following path:
SCX_CALL_OP(dispatch, rq)
ops.dispatch()
scx_bpf_dsq_move_to_local()
scx_flush_dispatch_buf()
finish_dispatch()
dispatch_to_local_dsq()
scx_dispatch_enqueue()
local_dsq_post_enq()
call_task_dequeue()
SCX_CALL_OP_TASK(dequeue, locked_rq, ...)
The nested callback saves the recorded rq and restores it on return. If
the rq tracking does not follow the lock switch, update_locked_rq() can
trigger the following lockdep assertion while restoring an rq which is
no longer held:
WARNING: kernel/sched/sched.h:1641 at call_task_dequeue+0x160/0x170
Call Trace:
scx_dispatch_enqueue+0x2b0/0x460
dispatch_to_local_dsq+0x138/0x230
scx_flush_dispatch_buf+0x1af/0x220
scx_bpf_dsq_move_to_local___v2+0xe2/0x1c0
bpf__sched_ext_ops_dispatch+0x4b/0xa7
do_pick_task_scx+0x3b6/0x910
__pick_next_task+0x105/0x1f0
__schedule+0x3e7/0x1980
Introduce switch_rq_lock() to update the tracking state together with
each rq lock handoff. Use it in dispatch_to_local_dsq(),
move_remote_task_to_local_dsq() and the in-balance paths of
scx_dsq_move(), ensuring that scx_locked_rq() consistently refers to the
rq whose lock is actually held throughout the lock dance.
Fixes: 7fb39e4eb4 ("sched_ext: Save and restore scx_locked_rq across SCX_CALL_OP")
Cc: stable@vger.kernel.org # 7.1+
Signed-off-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The usage of the hrtimer base introduces a dependency on the timer base
structure definitions from the widely-used hrtimer.h.
Move the helper to hrtimer_bases.h to trim this dependency.
Also adapt the two only callers to now include hrtimer_bases.h.
Signed-off-by: Thomas Weißschuh (Schneider Electric) <thomas.weissschuh@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Link: https://patch.msgid.link/20260702-hrtimer-header-dependencies-v1-2-c50b19bda473@linutronix.de
Load balancing can migrate delayed dequeue tasks to even the load between
CPUs.
sched_balance_rq()
-> detach_task()
-> deactivate_task(DEQUEUE_NOCLOCK)
-> set_task_cpu(dst_cpu)
-> attach_task()
-> activate_task(ENQUEUE_NOCLOCK)
-> wakeup_preempt()
A delayed task with shorter slice can be dequeued during pick_next_entity()
but then jump to preempt because eligible.
Always jump to update for delayed dequeue task even with shorter slice.
Fixes: ba0d3bf5f9 ("sched/eevdf: Cancel slice protection if short slice task is eligible")
Reported-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260701164920.1571352-1-vincent.guittot@linaro.org
The sub-scheduler implementation has grown and will continue to expand. Move
the sub-scheduler functions from ext.c into a new kernel/sched/ext/sub.c.
sub.h holds the prototypes and the !CONFIG_EXT_SUB_SCHED no-op stubs.
scx_dispatch_sched() is shared: balance_one() in ext.c and the
scx_bpf_sub_dispatch() kfunc in sub.c both call it, and the latter re-enters
it as sub-scheduler dispatch nests. It moves into sub.h as a static
__always_inline so both callers keep it inlined and per-level stack stays
bounded across the recursion. The event macros it uses move to internal.h.
No functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The following trivial helpers in ext.c are called from both ext.c and the
sub-scheduler code. Define them as static inline in internal.h.
- scx_bypass_dsq()
- scx_bypass_dsp_enabled()
- scx_ops_sanitize_err()
- scx_schedule_reenq_local()
No functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
The sub-scheduler implementation is about to move into its own sub.c, from
where it calls a set of ext.c helpers and shares a few ext.c globals. Make
those reachable across the new file boundary ahead of the move.
No functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A later change moves the sub-scheduler implementation out of ext.c into its
own file, from where it calls a number of file-local ext.c helpers. Give
those helpers the scx_ prefix that cross-file sched_ext symbols carry, ahead
of the move so the mechanical rename stays out of the code-motion patch. No
functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Fix checkpatch "Macros with complex values should be enclosed in
parentheses" error.
Suggested-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Three sites enumerate every scx_event_stats field by name:
scx_read_events(), scx_attr_events_show(), and scx_dump_state().
Adding a new SCX_EV_* event requires three separate manual edits.
Missing any one silently drops counts or omits the field from
diagnostic output with no compile-time indication.
SCX_EVENTS_LIST(SCX_EVENT) in internal.h becomes the single
authoritative list of event names, so all three sites stay in sync.
No functional changes.
Signed-off-by: Cheng-Yang Chou <yphbchou0911@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
put_prev_task_scx() warns when a runnable task drops to a lower sched_class
without SCX_OPS_ENQ_LAST, on the assumption that balance_one() would have
kept it running. Core scheduling breaks that: a forced-idle SMT sibling
reschedules through the core_pick fast path in pick_next_task(), which skips
pick_task_scx() and thus balance_one(), so a runnable task can drop to idle
with ENQ_LAST unset.
Gate the warning on sched_cpu_cookie_match(): a cookie mismatch means core
scheduling forced the idle, while a match (or core scheduling off) still
catches a genuine missing-ENQ_LAST drop.
Fixes: 7c65ae81ea ("sched_ext: Don't call put_prev_task_scx() before picking the next task")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
A few comments still describe the pre-EEVDF CFS world:
- sysctl_sched_base_slice is documented as "Minimal preemption
granularity for CPU-bound tasks". That was the wording of the old
sysctl_sched_min_granularity, renamed in commit e4ec3318a1
("sched/debug: Rename sysctl_sched_min_granularity to
sysctl_sched_base_slice"). Under EEVDF it is the default base time
slice / request size (r_i) used to compute the virtual deadline, as
documented in update_deadline().
- Two comments still mention sched_slice(), which was removed when the
fair class committed to EEVDF in commit 5e963f2bd4 ("sched/fair:
Commit to EEVDF"). The dequeue-path comment should simply refer to
the task's slice (se->slice); the forced-idle comment describes the
slice accounting now performed by __entity_slice_used(), which is the
function actually used right below it.
No functional changes.
[ mingo: Ported to a more recent scheduler base ]
Signed-off-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://patch.msgid.link/20260629030200.3165589-1-zhanxusheng@xiaomi.com
nr_migrations_cold, nr_wakeups_passive and nr_wakeups_idle are not
being updated anywhere. So remove them.
These are per process stats. So updating sched stats version isn't
necessary.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260625124648.802832-2-sshegde@linux.ibm.com
psi_account_irqtime() reads irq_time_read() into a per-rq cumulative
counter and only bails out when the delta vs. the previously accounted
amount is negative. A delta of exactly zero is treated as "do the
work": psi_write_begin() is taken, cpu_clock(cpu) is read (which on
x86 ends up in native_sched_clock() / rdtsc) and the cgroup ancestor
chain is walked to add zero to every group's PSI_IRQ_FULL bucket.
The zero-delta case is common in practice -- it fires every time a
context switch crosses a PSI group boundary on a CPU that hasn't
serviced an interrupt between the two switches.
Measured on a 176-thread AMD EPYC 9D64 server running a compute
intensive production workload, instrumented with bpftrace over a 30s
window (irq_time_read() read directly from the per-CPU cpu_irqtime so
that delta == 0 and delta < 0 could be separated):
@total 17,229,311 (100.0%)
@ret_curr_swapper 7,864,195 ( 45.6%) curr->pid == 0
@ret_samegrp 323,299 ( 1.9%) same cgroup as prev
@reached_delta 9,041,817 ( 52.5%)
@delta_positive 6,358,192 ( 36.9%) real work
@delta_zero 2,683,625 ( 15.6%) work wasted (this patch)
@delta_negative (0) ( 0.0%) monotonic clock
So 15.6 % of all psi_account_irqtime() calls - and 29.7 % of the
calls that get past the early returns - hit the delta == 0 case;
delta < 0 did not occur once in the 30 s window. Under the current
code each of those ~89 k calls per second performs the full seqcount
write + cpu_clock() read + cgroup-chain walk just to add 0 to every
group's PSI_IRQ_FULL counter.
Extend the early-return to also cover delta == 0. rq->psi_irq_time
does not need updating in that case (it would store the same value
back) and no PSI bucket would change. The existing behaviour for
delta > 0 is untouched.
Signed-off-by: Usama Arif <usama.arif@linux.dev>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Link: https://patch.msgid.link/20260617175219.2494857-2-usama.arif@linux.dev
Because active_balance is initialized to 0, and need_active_balance()
is a pre-condition for setting it to 1, the condition '!active_balance
|| need_active_balance()' is a truism and can be removed.
Signed-off-by: Xin Zhao <jackzxcui1989@163.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260617072151.1173416-3-jackzxcui1989@163.com
Active load balancing relies on migration threads, which temporarily preempt
tasks on the source runqueue (src_rq). This preemption can negatively impact
overall system performance. The active balancing logic includes a check to
verify whether the current task (curr) on src_rq can actually run on the
destination runqueue (dst_rq). We have observed that when curr is a CFS task
and its on_rq flag is 0, the active balancing failure rate is exceptionally
high. The following table summarizes test data collected over 300 seconds on an
18-CPU platform under a specific fillback task scenario:
fair: busiest->curr->sched_class == &fair_sched_class
on_rq: busiest->curr->on_rq
total: active balance count triggered of correspondent type
fail: fail to migrate one task in active_load_balance_cpu_stop()
fair && !on_rq !fair && !on_rq
domain total fail total fail
cpu0 0x00003 0 0 0 0
cpu0 0x3ffff 33 33 1 1
cpu1 0x00003 0 0 0 0
cpu1 0x3ffff 42 42 0 0
cpu2 0x0003c 4 4 0 0
cpu2 0x3ffff 12 12 0 0
cpu3 0x0003c 3 3 0 0
cpu3 0x3ffff 8 7 0 0
cpu4 0x0003c 2 2 0 0
cpu4 0x3ffff 5 4 0 0
cpu5 0x0003c 4 4 0 0
cpu5 0x3ffff 8 8 0 0
cpu6 0x003c0 60 60 0 0
cpu6 0x3ffff 28 27 0 0
cpu7 0x003c0 194 184 0 0
cpu7 0x3ffff 35 35 1 1
cpu8 0x003c0 240 228 0 0
cpu8 0x3ffff 28 28 0 0
cpu9 0x003c0 0 0 0 0
cpu9 0x3ffff 10 10 0 0
cpu10 0x03c00 52 50 0 0
cpu10 0x3ffff 0 0 0 0
cpu11 0x03c00 70 68 0 0
cpu11 0x3ffff 1 1 0 0
cpu12 0x03c00 73 72 0 0
cpu12 0x3ffff 0 0 0 0
cpu13 0x03c00 79 76 0 0
cpu13 0x3ffff 0 0 0 0
cpu14 0x3c000 0 0 0 0
cpu14 0x3ffff 57 55 1 0
cpu15 0x3c000 53 52 1 0
cpu15 0x3ffff 30 29 0 0
cpu16 0x3c000 344 341 10 6
cpu16 0x3ffff 103 100 2 1
cpu17 0x3c000 183 179 2 2
cpu17 0x3ffff 78 77 0 0
sum 1839 1791 18 11
In __schedule(), before curr is updated to next, pick_next_task() invokes
sched_balance_rq(). This function temporarily unlocks and relocks the runqueue,
creating a window where other CPUs may observe rq->curr->on_rq as 0.
We can safely skip active balancing when src_rq->curr->on_rq == 0, as other
eligible tasks have likely already been evaluated.
We retain the affinity check on dst_rq to trigger active balancing, since such
tasks are often woken by (or wake up) tasks on src_rq that share similar
affinity constraints. Furthermore, detach_tasks() releases the runqueue lock;
any tasks awakened during this window may preempt the previous CFS task. My
testing (data not shown) indicates that active balancing succeeds in 98.4% of
cases where !fair && on_rq.
This scenario does not require a stop-work callback, but would necessitate an
additional detach/attach path. As Valentin and Vincent have already discussed,
this addition does not appear justified at this time (see [1]).
Since can_migrate_task() already checks on_cpu during the cfs_tasks traversal,
adding an on_rq check will have negligible performance overhead due to cache
locality.
There are two reasons for not combining the on_rq check with the
cpumask_test_cpu() check:
- Avoiding new scenarios that would skip the logic for resetting
balance_interval to min_interval.
- The existing check for whether the busiest CPU recently triggered active
load balancing already filters more cases than the on_rq check.
[1]: https://lore.kernel.org/lkml/20190815145107.5318-5-valentin.schneider@arm.com/
Signed-off-by: Xin Zhao <jackzxcui1989@163.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Link: https://patch.msgid.link/20260617072151.1173416-2-jackzxcui1989@163.com
When a task with a shorter slice is enqueued, we protect the running
task which has a longer slice until it becomes ineligible instead of a
full slice in order to speedup the switch to other tasks until the task
with the shortest slice is scheduled. This helps to the task to not wait
too many full slices before running.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-7-vincent.guittot@linaro.org
Even if p will not preempt current, it modifies the avg_vruntime and
possibly the min slice. Make sure to update the slice protection with the
updated figures. As an example, Batch and Sched Idle tasks can otherwise
get a larger lag than their slice and finaly delay the scheduling of a
normal task, which deadline will be a later.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-6-vincent.guittot@linaro.org
If a short slice task will not be the next to be picked but is eligible,
we cancel the slice protection to speedup the time when the short slice
task will be the next to run.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-5-vincent.guittot@linaro.org
Even if resched is already set, we might want to update or even cancel
the slice protection and ensure that the newly waking task will be the
next one to run.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-4-vincent.guittot@linaro.org
Take into account the lag of current task when updating the slice
protection in order to ensure that the absolute value of lags will remain
in the range [0 : slice+tick]
A task that already has a negative lag will see its protection reduced
whereas a task with positive lag will keep a full slice protection.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-3-vincent.guittot@linaro.org
If a shorter slice task can preempt current at wakeup, we make sure that
the decision will not be overwritten in between by setting the task as the
next buddy. This still implies that the waking task remains eligible when
the scheduler will actually pick the next task to run.
Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Link: https://patch.msgid.link/20260624151229.1710703-2-vincent.guittot@linaro.org
On a tangential note, I just noticed set_preempt_buddy() has two unused
parameters. Seems to have been like that since it was introduced in
commit e837456fdc ("sched/fair: Reimplement NEXT_BUDDY to align with
EEVDF goals").
Signed-off-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Change fair/cgroup to a single runqueue.
Infamously fair/cgroup isn't working for a number of people; typically
the complaint is latencies and/or overhead. The latency issue is due
to the intermediate entries that represent a combination of tasks and
thereby obfuscate the runnability of tasks.
The approach here is to leave the cgroup hierarchy as is; including
the intermediate enqueue/dequeue but move the actual EEVDF runqueue
outside. This means things like the shares_weight approximation are
fully preserved.
That is, given a hierarchy like:
R
|
se--G1
/ \
G2--se se--G3
/ \ |
T1--se se--T2 se--T3
This is fully maintained for load tracking, however the EEVDF parts of
cfs_rq/se go unused for the intermediates and are instead connected
like:
_R_
/ | \
T1 T2 T3
Since the effective weight of the entities is determined by the
hierarchy, this gets recomputed on enqueue,set_next_task and tick.
Notably, the effective weight (se->h_load) is computed from the
hierarchical fraction: se->load / cfs_rq->load.
Since EEVDF is now exclusively operating on rq->cfs, it needs to
consider cfs_rq->h_nr_queued rather than cfs_rq->nr_queued. Similarly,
only tasks can get delayed, simplifying some of the cgroup cleanup.
One place where additional information was required was
set_next_task() / put_prev_task(), where we need to track 'current'
both in the hierarchical sense (cfs_rq->h_curr) and in the flat sense
(cfs_rq->curr).
As a result of only having a single level to pick from, much of the
complications in pick_next_task() and preemption go away.
Since many of the hierarchical operations are still there, this won't
immediately fix the performance issues, but hopefully it will fix some
of the latency issues.
TODO: split struct cfs_rq / struct sched_entity
TODO: try and get rid of h_curr
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260605124052.227463677%40infradead.org
For all the reasons described in the preceding patches, the way cgroup
weight is computed is problematic. However, changing it is bound to
also lead to trouble. Esp. since people might have taken to inflating
the weight value where they can.
Since things are configurable, change the default and hope this serves
more people than it hurts, esp. in the longer run.
Specifically, this prepares for a flattened runqueue, where the hierarchical
weight becomes far more important (F_g^d terms), so getting rid of small F_g is
imperative.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260605124052.080482755%40infradead.org
