Schedulers on the default cid mapping treat [0, nr_online_cids) as the
online set and restart on hotplug. Schedulers that install their own mapping
with scx_bpf_cid_override() have no way to learn which cids are online: the
count no longer identifies members and the CPU-form cpumask is unusable from
cid programs. This is an obvious hole in the cid API.
Add scx_bpf_online_cmask(), a kernel-maintained cmask in the scheduler's
arena, allocated alongside the per-CPU scratch masks and populated after the
cid mapping is finalized and before ops.init(), for child schedulers too.
The pointer stays valid through ops.exit() with no reference to take. It is
the arena offset as a void pointer, the same form struct_ops arena arguments
arrive in. The verifier types the void return as a scalar for the program's
arena cast.
The mask follows the SCX hotplug notifications: seeded from cpu_active_mask
and updated before ops.cid_online/offline() runs, so it lags cpu_online_mask
only inside a hotplug transition. Updates walk the scheduler list under the
lock that also serializes unlinking. Reads are live, not atomic snapshots.
Root initialization excludes hotplug.
v2: Reworded the getter kerneldoc (Andrea Righi).
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
scx_qmap tracks idle cids itself. pick_direct_dispatch_cid() claims a cid by
clearing its bit and the task is inserted into that cid's local DSQ, which
kicks the CPU. When the task does not arrive, for example because the insert
fell back to the global DSQ after an affinity change, the CPU wakes, finds
nothing and picks idle again. That is not an idle transition, so
ops.update_idle() is not called and the cid stays marked busy until an
unrelated task runs on it.
Restore the claim from ops.dispatch(). The kick guarantees a dispatch on the
kicked CPU, and when it finds nothing to run with a NULL @prev, the CPU is
going back to idle. Document the pattern in ops.update_idle(), which reports
only actual transitions.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Cc: Andrea Righi <arighi@nvidia.com>
qmap can leave partition work pending with no runner. The effective-cap
callback publishes its request after failing to acquire part_busy, while
redistribute() checks for pending work before releasing it. Either ordering
can miss a request arriving as the current runner finishes, delaying the
update until the round-robin timer runs.
Publish requests before trying to become the runner and release part_busy
before checking for more work. Have all holders drain pending requests after
releasing it, including the stats flush. Distinguish mask refreshes from
repartitions so an effective-cap update only rebuilds the partition when a
repartition was also requested.
Fixes: e9151ed5c9 ("tools/sched_ext: scx_qmap - Expand hierarchical sub-scheduling")
Reported-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
qmap decides placements from self_cids, which redistribute() derives from
the caps view at ops.sub_caps_updated() time. That view runs ahead of the
cpus: a granted cid can be in self_cids before its cpu has reported the caps
in effect through ops.sub_ecaps_updated(). ops.update_idle() only comes once
BASE is in effect, so the idle-gated placements reach such a cid only
through an idle bit left over from an earlier hold. The highpri scan has no
gate at all:
parent cpu Y, qmap cpu X
grants ENQ on X to qmap
sub_caps_updated() adds X
to self_cids
highpri scan moves a task
to X with PREEMPT
caps not in effect,
move denied, task
bounced with REENQ_CAP
reject drain, enqueue
the scan moves it to X
again
denied again
dispatch syncs ecaps,
sub_ecaps_updated(X)
Every highpri move to X in that window is denied and bounced. The two
callbacks are meant to split the roles: ops.sub_caps_updated() tracks what
the node holds and drives what it delegates to its children, while
ops.sub_ecaps_updated() says whether a task can run on a cpu now. qmap used
the first for both. Track the caps in effect from ops.sub_ecaps_updated() as
avail_cids and place only on self_cids & avail_cids, so that self_cids stays
the delegation split and avail_cids gates the placement.
The stranded tests keep self_cids, as they ask whether the split gives the
task anywhere at all. A highpri task whose self_cids lack caps in effect
waits for them instead of being moved and bounced.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
qmap's stranded fallback forces a task that can run on none of its self cids
onto its first allowed cid with SCX_ENQ_RESCUE, and adds SCX_ENQ_IMMED when
that cid is a time-share it holds. On such a cid the insert stops being a
rescue request:
1. A task is enqueued while none of its allowed cids is in self_cids. At
attach self_cids is still empty.
2. qmap inserts it into cid 0's local DSQ with SCX_ENQ_RESCUE |
SCX_ENQ_IMMED.
3. The kernel finds ENQ_IMMED held on cid 0, admits the insert and skips the
rescue diversion.
4. cid 0's cpu is busy, so the IMMED task is bounced back to qmap with
SCX_ENQ_REENQ.
5. qmap's enqueue sees the same inputs and repeats step 2. Nothing runs in
between.
6. The reenqueue limit ejects qmap with SCX_EXIT_ERROR_REENQ.
The caps granted during the parent's ops.sub_attach() are delivered after
the sub already holds its tasks, while the per-cid effective caps that mark
the time-shares are delivered from the first dispatch after bypass lifts, so
every attach that receives a time-share on a task's first allowed cid starts
the loop. Drop IMMED from the rescue inserts so that step 3 diverts to the
rescue path.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
dispatch_one() tests ops flags and bypass state against the root scheduler
in both places where it decides to keep running @prev: the early keep of a
@prev with slice left tests the root's bypass state, and the keep-last at
the end tests the root's SCX_OPS_ENQ_LAST and bypass state. Both are
properties of the scheduler @prev belongs to, and put_prev_task_scx(), which
acts on the outcome, reads them from that scheduler. When @prev belongs to a
sub-scheduler the two sides disagree.
The keep-last case is visible. The root set SCX_OPS_ENQ_LAST, so a lone
@prev of a sub-scheduler is not kept and is enqueued with SCX_ENQ_LAST to a
sub-scheduler that never opted in. This trips the WARN_ON_ONCE in
put_prev_task_scx() for the missing flag, and the sub-scheduler queues the
task like any other and triggers no follow-up scheduling event, which can
lead to stalls.
Test SCX_OPS_ENQ_LAST and bypass state on @prev's sched in both places and
charge SCX_EV_DISPATCH_KEEP_LAST to it. Read the sched at each decision, as
the dispatch in between can drop the rq lock.
Fixes: 88234b075c ("sched_ext: Introduce scx_task_sched[_rcu]()")
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
dispatch_one() uses the root scheduler for everything it does, including the
two decisions to keep running @prev, which are wrong when @prev belongs to a
sub-scheduler. The function has to deal with @prev's scheduler too. Rename
the root's local from sch to root_sch for clarity and to make room for it.
No functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
alpha requires percpu variables in modules to be defined as weak so that the
compiler generates GOT based external references for them. This puts two
extra restrictions on percpu variable definitions. The symbol must be
globally unique even when static and a static percpu variable can't be
defined inside a function. DEBUG_FORCE_WEAK_PER_CPU exists to give generic
code build coverage for these restrictions without building for alpha.
MEM_ALLOC_PROFILING defines a static percpu counter at each allocation call
site and thus can't be built with weak percpu definitions, so it depends on
!DEBUG_FORCE_WEAK_PER_CPU. As allmodconfig enables DEBUG_FORCE_WEAK_PER_CPU,
this knocks MEM_ALLOC_PROFILING out of allmodconfig build coverage.
allmodconfig coverage for MEM_ALLOC_PROFILING is worth more than build
coverage for restrictions which only matter to alpha module builds. Drop
DEBUG_FORCE_WEAK_PER_CPU. Restriction violations will now show up only on
alpha builds.
Link: https://lore.kernel.org/178656406317.2437052.7257990869957704195@slm.duckdns.org
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Suren Baghdasaryan <surenb@google.com>
Acked-by: Gabriele Monaco <gmonaco@redhat.com> [include/rv/da_monitor.h]
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
scx_dsq_move() verifies that the task belongs to the calling scheduler
before taking any locks and aborts the scheduler on mismatch. The task can
lose the sched association at any point: It can run and fully exit, which
clears the association, or get rehomed to a different sub-sched. Both are
benign races, but the early ownership check escalates them into scheduler
aborts.
Move the ownership check below the cursor-lost check. Every ownership change
dequeues the task first, so a task that is still on the iterated DSQ under
the lock while owned elsewhere indicates a genuine violation and should
abort.
Also fix two stale comments still referencing sched_ext_free(), which has
been renamed to sched_ext_dead().
Fixes: bb4d9fd551 ("sched_ext: scx_dsq_move() should validate the task belongs to the right scheduler")
Signed-off-by: Tejun Heo <tj@kernel.org>
Sync common.bpf.h, compat.bpf.h and compat.h with the scx repo, which
accumulated the following:
- __COMPAT_read_enum() can now recover 64-bit scx enum values from kernel
BTF generated without BTF_KIND_ENUM64 support (pahole < 1.24 or
--skip_encoding_btf_enum64, e.g. COS/GKE kernels), substituting values
from the build-time vmlinux.h cross-checked against the low 32 bits the
kernel does provide.
- is_migration_disabled() no longer assumes the BPF prolog always disables
migration. Since 8e4f0b1ebc ("bpf: use rcu_read_lock_dont_migrate() for
trampoline.c") the prolog only does so under CONFIG_PREEMPT_RCU, so the
old current-task test under-reported on v6.18+ !PREEMPT_RCU kernels. A
runtime probe on bpf_scx_reg() handles older kernels with backported
trampoline behavior.
- __COMPAT_scx_bpf_dsq_peek() is gated behind kernel v7.1 where 2f2ea77092
("sched_ext: Use dsq->first_task instead of list_empty() in
dispatch_enqueue() FIFO-tail") fixed the kfunc spuriously returning NULL
on non-empty FIFO DSQs, and the new
scx_bpf_reenqueue_local_from_anywhere() provides a callable-from-anywhere
reenqueue which prefers the generic scx_bpf_dsq_reenq(). Both were first
posted by Gavin Guo and Changwoo Min and are picked up here with the
review feedback folded in.
- __COMPAT_scx_bpf_cpu_curr() and the scx_bpf_cpu_rq() declaration are
restored. Schedulers built from these headers still run on pre-v6.18
kernels where scx_bpf_cpu_curr() does not resolve and the scx_bpf_cpu_rq()
fallback still exists.
- scx_clock_task() and scx_clock_pelt() document their stale-read behavior
for remote idle CPUs under NO_HZ_IDLE.
Link: https://lore.kernel.org/all/20260817143126.562923-1-changwoo@igalia.com
Signed-off-by: Tejun Heo <tj@kernel.org>
Regenerate enum_defs.autogen.h against the current tree, picking up the
dispatch verdict enums and dropping the marker for the removed
SCX_RQ_IN_BALANCE. Add enums_abi.autogen.h, a table of 64-bit scx enumerator
values generated from vmlinux.h, used as the substitution source when the
running kernel's BTF truncates 64-bit enum values to 32 bits.
Signed-off-by: Tejun Heo <tj@kernel.org>
On PREEMPT_RT, bh_worker() wraps work item execution in pool->cb_lock to
provide a handshake for canceling BH work items. When a CPU goes down,
drain_dead_softirq_workfn() runs the dead pool's bh_worker() nested inside
the local pool's bh_worker(), acquiring the cb_locks of two different pools
without a nesting annotation. lockdep reports possible recursive locking:
============================================
WARNING: possible recursive locking detected
--------------------------------------------
ktimers/0/16 is trying to acquire lock:
ffff8880b873a990 (&pool->cb_lock){+...}-{3:3}, at: bh_worker+0x7d/0x880
but task is already holding lock:
ffff8880b863a990 (&pool->cb_lock){+...}-{3:3}, at: bh_worker+0x7d/0x880
Call Trace:
bh_worker+0x7d/0x880 kernel/workqueue.c:3688
drain_dead_softirq_workfn+0x95/0x220 kernel/workqueue.c:3763
process_scheduled_works+0xa8e/0x14e0 kernel/workqueue.c:3405
bh_worker+0x46a/0x880 kernel/workqueue.c:3708
tasklet_action+0xc/0x70 kernel/softirq.c:965
The nesting can't deadlock. A pool's bh_worker() runs nested only while the
pool's CPU is dead, entered from a live pool's bh_worker() on the draining
CPU, so the ordering is always live to dead. CPU hotplug operations are
serialized and the drain is synchronous, so the nesting depth never exceeds
two. Annotate the inner acquisition with SINGLE_DEPTH_NESTING.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: syzbot+1bd20115328f8254ed62@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=1bd20115328f8254ed62
Fixes: ad7c7f4b9c ("workqueue: Provide a handshake for canceling BH workers")
Cc: stable@vger.kernel.org # v6.18+
Pull to receive the __arena argument conversion:
67f1f4a48c ("sched_ext: Pass kernel arena pointers to ops_cid callbacks")
a8dc810968 ("sched_ext: Convert sub-cap kfuncs to __arena cmask arguments")
a05c5b5cb5 ("sched_ext: Convert scx_bpf_cid_override() to __arena array arguments")
along with the bpf-next branch carrying the __arena argument support they
depend on.
Conflict in kernel/sched/ext/ext.c between:
c384ab8a0b ("sched_ext: Move the config-off sub-cap kfunc stubs into sub.c")
and:
a8dc810968 ("sched_ext: Convert sub-cap kfuncs to __arena cmask arguments")
which updated the stubs in their old ext.c location. Resolved by keeping
ext.c without the stubs and applying the prototype conversion to the
relocated stubs in sub.c.
Signed-off-by: Tejun Heo <tj@kernel.org>
dequeue_task_scx() masks SCX_DEQ_CORE_SCHED_EXEC out of the
SCX_DEQ_SCHED_CHANGE decision, but the test can never fire: the incoming
flags are an int of generic DEQUEUE_* bits while the flag is bit 32, and the
core-sched execute path never goes through class dequeue anyway -
set_next_task_scx() calls ops_dequeue() with the flag directly. The test was
live when the SCX_DEQ_SCHED_CHANGE computation sat in ops_dequeue() and
became dead when 03f5304aad ("sched_ext: Pass full dequeue flags to
ops.quiescent()") moved the computation here. Drop it.
Signed-off-by: Tejun Heo <tj@kernel.org>
With sub-schedulers, tasks of different schedulers routinely share rqs and
SMT siblings, but scx_prio_less() consults ops.core_sched_before() only when
both tasks belong to the same scheduler. Every pair spanning two schedulers
falls back to the default ordering, so no scheduler can express ordering
across a scheduler boundary, including a root over its sub-schedulers'
tasks.
Order a pair spanning schedulers by the nearest common ancestor that
implements ops.core_sched_before(): both tasks are in its subtree, making
this the one op where a scheduler is called on tasks it delegated to its
sub-schedulers and may not be scheduling anymore. Same-scheduler pairs keep
using the owning scheduler's op so a parent never orders inside a subtree it
delegated. The op is skipped when the deciding scheduler is bypassing on
either task's CPU.
Update scx_qmap to fall back to the kernel's default ordering when handed a
delegated task it has no task_ctx for.
Signed-off-by: Tejun Heo <tj@kernel.org>
The default core-sched ordering runs the longest waiting task first by
comparing p->scx.core_sched_at stamps. The stamp is maintained under two
rules. touch_core_sched() stamps when a task starts waiting for a CPU and
when its slice runs out. If the scheduler implements
ops.core_sched_before(), touch_core_sched_dispatch() re-stamps on every
dispatch.
A comparison can see one stamp taken under each rule, which isn't a
meaningful ordering. The dispatch rule also buys little - it only aligns
bypass-mode comparisons with the local DSQ order. Multiple schedulers make
the mixed comparisons more common.
Wait time is what p->scx.runnable_at already tracks for the stall watchdog.
Delete core_sched_at with both touch functions and compare runnable_at in
the scx_prio_less() fallback.
runnable_at is refreshed only on enqueue and goes stale while a task keeps
occupying its CPU. Instead of re-stamping, order a running task after every
waiting task as it is the most recently serviced.
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_prio_less() implements prio_less() semantics - %true means that @a is
the lower priority and should run after @b. ops.core_sched_before() is
documented to return %true when @a should run before @b. scx_prio_less()
returns the op's value as-is, inverting the documented semantics at runtime.
Call the op with the arguments swapped.
scx_qmap followed the wiring instead of the documentation and returned %true
for the younger task, so the two inversions canceled out and it behaved as
intended. Flip its comparison to match. scx_qmap is likely the only current
user in or out of the kernel tree. Any scheduler written the same way needs
the same flip, while schedulers following the documentation are fixed by
this change.
Fixes: 7b0888b7cc ("sched_ext: Implement core-sched support")
Cc: stable@vger.kernel.org # v6.12+
Signed-off-by: Tejun Heo <tj@kernel.org>
The EOPNOTSUPP stubs for the sub-cap kfuncs live in ext.c under #ifndef
CONFIG_EXT_SUB_SCHED while the real definitions live in sub.c. Move the
stubs into sub.c so all sub kfunc definitions live in one file. Pure code
move, no functional change.
Signed-off-by: Tejun Heo <tj@kernel.org>
sched_class->balance() is gone from sched_ext and what balance_one() does is
run dispatch to produce something pickable. Update the balance-era names to
dispatch terms:
- balance_one() -> dispatch_one()
- SCX_RQ_IN_BALANCE -> SCX_RQ_IN_DISPATCH
No BPF scheduler reads the flag. The enum autogen headers gain the new name
with the old entry retained like other removed enumerators, zero-filling at
load time. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
The fixup demoting a keep verdict when @prev is not on ext_sched_class
guarded against the rq-level SCX_RQ_BAL_KEEP flag going stale back when
balancing and picking were separate operations.
The verdict now travels in the return value, created and consumed in one
invocation against the @prev it evaluated, and every keep decision tests
SCX_TASK_QUEUED under the rq lock, which implies ext_sched_class as a class
switch dequeues first. Drop the fixup along with dispatch_core_pick()'s
copy.
Signed-off-by: Tejun Heo <tj@kernel.org>
kick_sync_wait_bal_cb() assumes it runs on the rq's CPU from the
__schedule() tail: the snapshots it compares against live in that CPU's
percpu area and the busy-wait runs with the rq lock dropped and IRQs
enabled.
However, dispatch can now drop the rq lock while the callback sits queued,
and rq lock takers in that window (the sched class change paths, the scx
task iterator) flush pending balance callbacks on release, running the
callback on a foreign CPU. Such a run compares against unrelated snapshots
and can deadlock when the executing CPU is itself a wait target.
Bail on a foreign CPU and leave the wait state alone. The wait only observes
progress that the resched kicks already guarantee and the rq's next wait
picks up the stale cpus_to_sync bits.
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_bpf_events() always reads the root scheduler's event counters, so a
sub-scheduler program querying its own events silently gets the root's
instead and has no BPF-visible way to read its own (the per-scheduler sysfs
"events" file is the only interface). Resolve the scheduler from the calling
program with scx_prog_sched(). Unassociated programs follow the usual
scx_prog_sched() resolution: the root scheduler under a pre-sub-attach
compat root and zeroed counters otherwise.
Also fix up the malformed comment into proper kerneldoc.
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_root_disable() invalidates each rq's clock before taking the rq lock.
scx_rq_clock_invalidate() is a plain read-modify-write of rq->scx.flags and
every other writer of the word runs under the rq lock, so the unlocked
update can race a concurrent flags update and lose one side's bits.
The invalidation doesn't matter in the first place. The cached clock is read
only by scx_bpf_now() from a loaded scheduler's BPF programs, nothing can
re-validate the clock while sched_ext is disabled as scx_rq_clock_update()
is gated on scx_enabled() too, and the usual rq lock cycles under the next
scheduler refresh or invalidate it before it's practically observable. Drop
the invalidation instead of fixing the locking.
v2: Description and comment updated - the invalidation is unnecessary rather
than subsumed by the rq lock cycle below.
Fixes: 3a9910b590 ("sched_ext: Implement scx_bpf_now()")
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Changwoo Min <changwoo@igalia.com>
scx_bpf_dsq_reenq() is registered with KF_IMPLICIT_ARGS and its kernel BTF
prototype omits the trailing bpf_prog_aux argument. The ___compat extern
declares the argument explicitly, so libbpf never matches the prototype and
the weak extern silently stays unresolved on every kernel. The wrapper
always takes the old fallback path, which disables generic reenq users like
scx_qmap's lowpri mechanism and fails non-local reenq with "kernel too old"
even on kernels that have the kfunc.
Drop the explicit aux argument. Also correct the stale v6.20 reference, the
kfunc was added in v7.1.
Fixes: 9c34c5074d ("sched_ext: Introduce scx_bpf_dsq_reenq() for remote local DSQ reenqueue")
Cc: stable@vger.kernel.org # v7.1+
Signed-off-by: Tejun Heo <tj@kernel.org>
scx_process_sync_ecaps() sets up the dispatch context for
ops.sub_ecaps_updated() in the target cpu's pcpu context recovered from the
llist node. However, the context is per executing cpu: the dispatch kfuncs
resolve it with this_cpu_ptr() and the dispatch buffer lives in it. What the
dispatches target is determined by the rq recorded in the context, not by
which cpu's context it is. Under core scheduling the pick runs balance_one()
for sibling rqs, so a sync processed for a sibling invokes the op with the
executing cpu's context not set up and its dispatch kfuncs misoperate on a
NULL or stale rq.
Set up the executing cpu's dsp_ctx instead, matching scx_dispatch_sched().
The recorded rq keeps the dispatches targeting the synced cpu.
Fixes: b81a6c018c ("sched_ext: Add sub_ecaps_updated() effective-cap change notifier")
Reported-by: David Carlier <devnexen@gmail.com>
Link: https://lore.kernel.org/all/20260813045931.8691-1-devnexen@gmail.com/
Signed-off-by: Tejun Heo <tj@kernel.org>
The insn library encodes the immediate and shifted-register forms of
ADD/SUB but not the extended-register form. The BPF JIT wants it to
rebase a 32-bit arena offset onto the arena kernel base in a single
instruction, add xN, xBASE, wN, uxtw, instead of a separate zero-extend
followed by a plain add.
Add aarch64_insn_gen_add_sub_extended_reg(), modeled on the
shifted-register generator. The option and imm3 fields occupy the same
bits as the shifted form's shift amount, so they are encoded through the
existing IMM_6 field type. The opt field in bits 23:22 is part of the
opcode here rather than a shift type, and any value other than 00 is
unallocated, so the decode masks cover it.
Note that register 31 does not mean the same thing in the two forms: in
the extended-register encoding it is SP for Rn, and for Rd unless the
instruction sets the flags, while it stays XZR for Rm. Callers porting a
shifted-register site that passes A64_ZR need to be aware of that, so
say so above the function.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Puranjay Mohan <puranjay@kernel.org>
Reviewed-by: Xu Kuohai <xukuohai@huawei.com>
Link: https://lore.kernel.org/bpf/20260813190356.335181-3-puranjay@kernel.org
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
scx_bpf_cid_override() predates the cid-form arena transition and takes its
arrays as verifier-checked mem+size buffers, forcing scx_qmap to keep the
cpu_to_cid and shard_start arrays in writable bss while the rest of its
state lives in the arena. Unify on arena arguments before cid-form
schedulers start seeing real use.
BPF now translates between BPF and kernel arena addresses for __arena
arguments. Take the arrays as __arena arguments, with the counts passed in
entries. The counts now size the snapshot copies and are bounds-checked
before them.
scx_qmap moves the arrays into struct qmap_arena. As the arena is mmapped at
load, the loader populates them between load and attach instead of before
load.
The arena argument address translation is currently implemented only on
x86-64. Schedulers calling this kfunc load only there for now.
Signed-off-by: Tejun Heo <tj@kernel.org>
The sub-cap kfuncs take their cmask arguments as __ign pointers. The values
cross the kfunc boundary as unchecked scalars and scx_cmask_ref_init()
rebases them into the arena by hand.
BPF now translates between BPF and kernel arena addresses for __arena
arguments. Tag the cmask arguments __arena so the kfuncs receive kernel
addresses and scx_cmask_ref_init() loses the hand-rolled conversion. The
optional denied_out keeps its NULL not-provided signal via
__arena__nullable. The mandatory masks use plain __arena.
scx_qmap's call sites drop the (void *)(long) casts since the BPF-side
declarations type the cmask arguments __arena and take arena pointers
directly.
The arena argument address translation is currently implemented only on
x86-64. Schedulers calling these kfuncs load only there for now.
Signed-off-by: Tejun Heo <tj@kernel.org>
The cid-form set_cmask() and sub_caps_updated() callbacks receive cmasks
that the kernel builds in the arena, and the kernel converts the kernel
addresses to the BPF arena pointer form by hand before each call.
BPF now translates between BPF and kernel arena addresses for __arena
arguments. Tag the arguments __arena in the cfi stubs and the ops_cid member
declarations and pass the kernel arena addresses directly, dropping the
manual scx_kaddr_to_arena() conversions and the now-unused helper. The
delivered value is unchanged and existing BPF-side code works as before.
The arena argument address translation is currently implemented only on
x86-64. cid-form schedulers implementing these callbacks load only there for
now.
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull to receive:
c10b216a07 ("sched/core: Handle pick_task() releasing the rq lock")
f3629c63a4 ("sched/core: Make core-sched flips wait for in-flight selections")
ffaab58d21 ("sched_ext: Replace SCX_RQ_BAL_KEEP with a dispatch verdict return")
3dd52416e4 ("sched_ext: Fix this_rq() assumptions in dispatch kfuncs")
f2da958711 ("sched_ext: Count rq lock releases in rq->scx.lock_drop_seq")
d954004205 ("sched_ext: Fix rq->core_pick corruption under core scheduling")
for the pending core scheduling follow-ups and to resolve the conflicts
with the code reorganization and cap gate work on for-7.3.
ffaab58d21 converts scx_dispatch_sched() to a dispatch verdict return
which for-7.3 moved from ext.c into inlines.h. Resolved by applying the
conversion to the relocated copy and combining balance_one()'s verdict
returns with the scx_task_can_stay_on_cpu() gate from the cap work.
ffaab58d21 and 3dd52416e4 update scx_bpf_sub_dispatch() which
for-7.3 moved into sub.c. Resolved by applying the scx_locked_rq()
switch and the verdict test to the sub.c copy.
f2da958711 instruments the open-coded lock releases in
consume_remote_task() which for-7.3 folded into switch_rq_lock().
Resolved by keeping the accounting in switch_rq_lock() which covers all
its callers.
d954004205 widens the put_prev_task_scx() WARN suppression to all
core-sched rqs on the same condition that for-7.3 gated with
scx_task_can_stay_on_cpu(). Resolved by combining both.
Signed-off-by: Tejun Heo <tj@kernel.org>
Core scheduling's pick_next_task() picks what to run on every SMT sibling of
the core in a single pass under the shared core-wide rq lock. The selection
state is consistent only while the lock is held continuously, so
->pick_task() originally could not release it. However, since 4c95380701
("sched/ext: Fold balance_scx() into pick_task_scx()"), sched_ext runs
dispatch from inside the pick and dispatching can drop the rq lock. To
support this, pick_next_task() has been updated to restart the whole
selection when a pick returns RETRY_TASK after releasing the lock.
When selections on the same core interleave through the dropped lock, they
corrupt each other's state: one clears the other's rq->core_pick leading to
a NULL deref, or invalidates its keep-the-previous-task decision leaving a
dequeued task running, which deadlocks the next wakeup and matches the
reported hard hangs. A cookied ping-pong load on an SMT machine makes the
interleavings frequent and kills the kernel within seconds.
Fix it by making the pick return RETRY_TASK whenever dispatch released the
rq lock, so that a selection only ever commits picks made under a
continuously held lock. The previous patch's rq->scx.lock_drop_seq counts
the releases. A dispatch that touched nothing never releases the lock and
its verdict, including "nothing to run", stands: retries are bounded, each
following a dispatch that actually did something, and an idle CPU does not
loop.
If another dispatch is already in flight on the rq, skip dispatching and
pick from what is already queued locally - the in-flight dispatch has
released the lock, so its own selection will retry and re-pick this rq,
while returning RETRY_TASK here would only spin on the lock that dispatch
needs to finish.
Balance callbacks must run in the context that queued them, so they can only
be queued on the CPU's own rq. When dispatching for another rq, run the
deferred work directly instead - that rq may consume all its picks through
the core-sched fast path and never queue the callback itself.
The put_prev_task_scx() warning about a runnable task being left behind
assumed that dispatch ran as part of the very pick that is switching away.
That now only holds on the non-core path, so gate it and drop the
cookie-match test, which is always true without core scheduling, from its
condition.
Fixes: 4c95380701 ("sched/ext: Fold balance_scx() into pick_task_scx()")
Cc: stable@vger.kernel.org # v6.19+
Reported-by: ElXreno <elxreno@gmail.com>
Link: https://github.com/sched-ext/scx/issues/3715
Signed-off-by: Tejun Heo <tj@kernel.org>
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.
Add a test_arena_stack member with eight leading scalar arguments so the
arena pointer is passed on the stack.
The callback validates the first and last scalar ctx slots before
dereferencing the pointer in ctx[8]. This exercises the indirect
trampoline stack layout and arena conversion together, and prevents a
regression where stack arguments are read one slot late.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Tested-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-13-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
save_args() reads stack-passed arguments relative to rbp assuming two
return addresses sit between the saved rbp and the arguments, which
holds when the trampoline is entered through the fentry call from a
traced function. An indirect trampoline is called through a function
pointer, so only the caller's return address is on the stack and the
arguments start at rbp + 16, not rbp + 24. Every stack-passed argument
of a struct_ops callback with more than six argument slots is read one
slot off.
This has gone unnoticed because no in-tree struct_ops member passes
arguments on the stack. The jmp-entry form already accounts for having
a single return address; treat BPF_TRAMP_F_INDIRECT the same way.
Fixes: 473e3150e3 ("bpf, x86: allow function arguments up to 12 for TRACING")
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Tested-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-12-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Add test_arena and test_arena_nullable members to bpf_testmod_ops3 with
arena-tagged stub arguments and kfuncs that forward a caller-provided
pointer to them. The kfuncs take arena-tagged arguments, so each round
trip exercises both conversion directions end to end: the kfunc receives
a kernel arena address and the trampoline converts it back to an arena
pointer for the callback.
The non-nullable callback dereferences its argument with no NULL branch
and captures the raw ctx value, which the trigger program compares
against the arena offset of the passed object, pinning the exact
(u32)(kaddr - kern_vm_start) conversion. The nullable callback verifies
that only a true kernel NULL arrives as NULL. Failure coverage: a
program with no arena is rejected when it loads. The tests run on x86-64
and skip elsewhere, as the programs fail verification where the JIT
lacks arena argument support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-11-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Pin the exact rebase sequences the JITs emit for __arena and
__arena__nullable kfunc arguments with __jited assertions on x86-64: the
unconditional truncate-and-add, the nullable test-and-skip variant, and
all five argument registers in one call, which also covers the
REX-prefixed encoding of r8 on x86. The capture kfuncs take the argument
without dereferencing, so only the emitted code is under test. The
tests skip without LLVM disassembler support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-10-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Add arena-argument kfuncs to bpf_testmod, which also exercises the
argument rebasing on module kfuncs, and tests covering the accepted
argument forms (arena pointer, low 32 bits as a scalar, full user
address as a scalar), the exact rebase semantics via capture kfuncs
returning the raw argument (zero low 32 bits arrive as the arena kernel
base under __arena and as NULL under __arena__nullable), five arena
arguments in one call, a mixed __arena plus __arena__nullable call
exercising both bitmasks on one call site, a kernel-side dereference of
an unpopulated page recovering through the scratch page, and the
rejections (no arena in the program, incompatible register type).
The tests run on x86-64 and skip elsewhere, as programs with
arena-tagged kfunc args fail verification where the JIT lacks support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-9-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Implement the struct_ops arena argument conversion on x86. save_args()
receives the arena base from bpf_tramp_arena_base() and consults the
btf_func_model argument flags as it copies each native argument into the
BPF ctx, routing a marked argument through RAX:
movl %esrc, %eax /* truncate and clear the upper 32 bits */
subl $base_lo, %eax
movq %rax, ctx_slot
A nullable argument tests the full 64-bit kernel pointer first:
movq %rsrc, %rax
testq %rax, %rax
jz 1f
subl $base_lo, %eax
1:
movq %rax, ctx_slot
The 32-bit subtraction is sufficient since (u32)(kaddr - base) ==
(u32)kaddr - (u32)base, and it clears the upper half as the JITs require
of arena pointer registers. Stack-passed arguments already reload
through RAX, so only the subtraction (and the NULL test) is inserted
there.
Keep arena and nullable classification in btf_func_model.
bpf_tramp_arena_base() returns a base only for a single-program
struct_ops indirect trampoline; other trampolines pass zero and perform
no conversion. The size probe reruns the same emission with the same
model and nodes, so the image size matches by construction.
With both the kfunc and struct_ops directions implemented, flip
bpf_jit_supports_arena_args() on for x86.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-8-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Implement arena argument rebasing for kfunc calls on x86. R12 already
holds kern_vm_start whenever the prog has an arena, so each tagged
argument costs two instructions emitted right before the call:
movl %eN, %eN /* truncate, clear the upper 32 bits */
addq %r12, %rN
A nullable argument tests the truncated value and jumps over the add:
movl %eN, %eN
testl %eN, %eN
jz 1f
addq %r12, %rN
1:
addq carries a REX prefix for every argument register and is always
three bytes, so the jz displacement is constant. The sequence is native
code generated after constant blinding has run on the BPF instruction
stream, so blinding never sees the rebase and needs no special handling.
bpf_jit_supports_arena_args() is not flipped yet; that happens when the
struct_ops trampoline side is in place as well.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-7-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
A struct_ops callback cannot receive an arena pointer directly, so
passing one takes two steps. The pointer arrives as a bare u64 that the
callback casts, and because the two sides address the arena through
different bases it also has to be rebased by hand on the way in.
Add the __arena and __arena__nullable stub argument suffixes to make this
convenient. The callback declares the parameter as an arena pointer,
receives it as a PTR_TO_ARENA register, and dereferences it directly,
while the kernel caller just passes the natural kernel arena address
(kaddr). The trampoline converts the value while saving the arguments
into the BPF ctx, ctx[slot] = (u32)(kaddr - kern_vm_start), so the
program never sees a kernel address and nothing rewrites the ctx after
the fact. The converted value keeps the upper 32 bits clear as the JITs
require of arena pointer registers and behaves like any cast_kern'ed
arena pointer, so cast_user recovers the full user-visible address.
__arena converts unconditionally and the kernel caller must not pass
NULL. __arena__nullable preserves NULL, tested on the full 64-bit kernel
pointer, and surfaces to the verifier as PTR_TO_ARENA (but not as a
PTR_TO_ARENA | PTR_MAYBE_NULL). The reason is that PTR_TO_ARENA in the
program's type state already encompasses NULL-ness, so it is not
meaningful to force a NULL check for the program.
The composite suffix intentionally ends in __nullable. Classify
__arena__nullable before the generic suffix so scalar arena pointees do
not take the generic nullable BTF pointer path.
This patch adds the generic side. prepare_arg_info() records arena and
nullable argument flags in the struct_ops function model, and
bpf_tramp_arena_base() returns the arena base for a single-program
struct_ops indirect trampoline. Only that trampoline converts: its
program's arena is fixed at generation time. Generic trampolines can mix
programs with different arenas and reject arena context arguments
defensively, which is unreachable today as only struct_ops programs
carry them. Architectures that do not implement the conversion are
gated out at verification time with bpf_jit_supports_arena_args().
Co-developed-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-6-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Passing an arena pointer to a kfunc takes two steps today. There is no
arena pointer argument type, so the pointer crosses the boundary as a
bare scalar, and the kfunc then offsets it by the arena base and casts
it before it can touch the memory. Every such kfunc open-codes the same
translation.
Add the __arena and __arena__nullable argument suffixes to make this more
convenient. The kfunc declares the parameter by its real pointer type
and dereferences it directly, with the JIT rebasing the value at the
call site, rN = kern_vm_start + (u32)rN. No bounds check is needed: the
u32 offset stays within the guard-padded arena kernel mapping, and a
fault on an unpopulated page recovers through the per-arena scratch
page. A suffixed argument accepts a PTR_TO_ARENA or scalar register,
matching global subprog arena arguments.
__arena rebases unconditionally, so the kfunc never sees NULL and a
value with zero in the low 32 bits arrives as the arena base.
__arena__nullable preserves NULL for optional arguments by skipping the
rebase when the truncated value, arena offset 0, is zero. Keeping the
plain form NULL-free saves the NULL test on every call.
The double separator makes the annotations composable:
__arena__nullable also ends in __nullable and naturally follows the
common nullable argument path. Plain __arena follows that path too for
verifier type checking because both forms accept a constant zero; the
function-model flag still determines whether the JIT preserves NULL or
rebases it to the arena base.
This patch adds the verifier side: the suffixes are recognized in
check_kfunc_args() and distilled into argument flags in the function
model stored in the kfunc descriptor. JITs retrieve the model while
emitting the call, avoiding per-call state in insn_aux_data.
JITs declare support with bpf_jit_supports_arena_args() and verification
fails with -ENOTSUPP elsewhere.
Co-developed-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://patch.msgid.link/20260808003938.3486067-5-memxor@gmail.com
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
A sched holds only the cids its parent granted and nothing guarantees that
they cover its tasks' affinities. A task that can run on none of them has
nowhere to go and qmap stalls out: it force-inserts the task onto its first
allowed cid, but the kernel bounces the insert back and the task parks in
SHARED_DSQ, which is drained only on self cids it can't run on.
Set SCX_ENQ_RESCUE on these inserts so the kernel diverts such tasks to its
rescue path instead of bouncing them. The force-insert covers scheds with
and without children and fires on re-enqueues, and the SHARED_DSQ scan on
every dispatch rescues tasks stranded there - the enqueue-time check misses
a task whose cids were lost while it was already queued. The wrong-cid fault
injection carries the flag too and doubles as a deterministic rescue-traffic
generator.
-B and -q set the root-only rescue bandwidth and quantum ops knobs. -B 0
maps to SCX_RESCUE_DISABLE and turns rescue off kernel-side. Rescue inserts
are counted and reported in the hier stats line.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
pick_direct_dispatch_cid() sent a pinned task straight to its only cpu
without an idle check. An insert onto a time-shared cid is forced IMMED,
which may not wait on a busy cpu - the kernel bounces the task back to
ops.enqueue(), the shortcut re-inserts it, and the loop runs into the
reenqueue repeat limit.
Run pinned tasks through the same idle test as everyone else and queue them
when the cpu is busy. always_enq_immed forces IMMED on every local insert
and already skipped the shortcut for the same reason. The check was the last
reader of the always_enq_immed rodata mirror, so drop the variable. The -I
option still sets SCX_OPS_ALWAYS_ENQ_IMMED.
v2: Drop the now-unread always_enq_immed rodata mirror. (sashiko AI)
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Mechanical regeneration for the enumerators added by the slice write rework
and the rescue machinery. BPF schedulers reference SCX_ENQ_RESCUE, the rest
are enum_defs presence markers.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
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>
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>
Regenerate enum_defs.autogen.h from the current vmlinux.h to pick up the SCX
enum changes accumulated since the last regeneration, including the
SCX_REENQ_LOCAL_MAX_REPEAT to SCX_REENQ_MAX_REPEAT rename.
Reported-by: Andrea Righi <arighi@nvidia.com>
Link: https://lore.kernel.org/all/amZsEbZJdDgjstPF@gpd4/
Signed-off-by: Tejun Heo <tj@kernel.org>
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 tools restart when the kernel exits the scheduler with
SCX_ECODE_ACT_RESTART. The restart decision doesn't consult exit_req, so an
exit request arriving while the restart condition persists is ignored and
the tool reloads in a tight loop. Test exit_req before restarting.
scx_userland needs more: its main loop never watches the kernel-side exit
and exit_req doubles as the stats printer's stop signal, set by the teardown
and reset on each restart. Add the missing UEI_EXITED() test and give the
printer its own stop flag so that exit_req only means an exit request and
stays latched like in the other tools.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.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>
SCX_OPS_OPEN() clears compat-gated ops fields which the running kernel
lacks. The clears dereference cpu-form member names and compile for
cid-form skeletons only because both ops structs currently name their
cgroup ops identically, which an upcoming rename will end. No load-time
fix-up can apply to a cid-form scheduler anyway as the cid form postdates
every compat-gated op.
Factor the skeleton open path out of SCX_OPS_OPEN() and add
SCX_OPS_CID_OPEN() which uses only that shared part. Switch scx_qmap, the
only cid-form scheduler, over.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
Add -J init-fail which makes ops.init_task() fail with -ENOMEM for tasks
whose comm starts with "qmfail", and -J cgrp-init-fail which does the same
in ops.cgroup_init() for cgroups named "qmfail*".
The former exercises the migration veto path: the cgroup.procs write must
fail with the injected errno while the destination sched stays up and the
task stays put. The latter exercises the ownership-return failure path: a
parent failing to re-init a returned cgroup leaves it unowned, and moves and
set_* ops against it must be skipped instead of dereferencing the missing
owner. Matching on "qmfail" names keeps the injecting scheduler's own enable
unaffected.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Andrea Righi <arighi@nvidia.com>
With the set_* ops delivered to the parent's sched, a parent qmap instance
now receives cgroup_set_weight for its child subs' attach points. Update
the matching sub_sched_ctx weight and redistribute() in-kernel, and drop
the userspace feed_weights() polling. This exercises the knob routing end
to end.
The self weight is fixed at 100: a cgroup's weight is its parent's knob and
not the scheduler's own business. This drops the self-weight polling and the
repartition PROG_RUN poke with it.
sub_attach seeds the slot with the cgroup's current weight, read through
bpf_cgroup_from_id(), so a weight set before the sub attaches is picked up.
A write racing the attach can still be lost until the next value-changing
cpu.weight write. Acceptable for a demo.
While at it, add a traced ops.cgroup_move() so tests can observe move
delivery.
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>
A subsystem can attach to the cgroup hierarchy itself, independent of which
controllers are enabled where - BPF hooks already behave this way and
sched_ext sub-schedulers do too. Controller callbacks can't track task
migrations for them: sched_ext must re-home a task whose migration crosses a
sub-scheduler boundary, but the cpu controller's attach callbacks fire only
when the task_group changes and miss moves whenever the controller topology
is coarser than the sub-scheduler topology.
Add cgroup_task_notifier with per-task migration events mirroring the
can_attach/attach/cancel_attach phases so that a consumer which prepares
per-task state can also veto a migration: CGROUP_TASK_MIGRATING fires
pre-commit, CGROUP_TASK_MIGRATED post-commit and
CGROUP_TASK_MIGRATE_CANCELED unwinds a failed migration. Only migrations
that change a task's dfl cgroup are reported.
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>
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>
Tejun Heo