[PATCHSET v8 sched_ext/for-7.3] sched: Make proxy execution compatible with sched_ext

[PATCHSET v8 sched_ext/for-7.3] sched: Make proxy execution compatible with sched_ext

[Andrea Righi]

This series enables using proxy execution with sched_ext and is based on early
work by John Stultz [1].

Background
==========

Proxy execution (proxy-exec) lets a waiting task ("donor") donate its scheduling
context to a mutex owner, so the owner can run while the donor stays eligible on
the runqueue.

Currently, proxy-exec and sched_ext are mutually exclusive at build time: we
can't enable CONFIG_SCHED_PROXY_EXEC=y and CONFIG_SCHED_CLASS_EXT=y in the same
kernel.

This restriction can be problematic for Linux distributions and for anyone who
wants to ship one kernel and choose features at runtime.

Why are they mutually exclusive?
================================

sched_ext schedulers drive dispatch through their own interfaces. A proxy-exec
handoff can run a task that the BPF scheduler never dispatched through that
path. sched_ext callbacks then observe a "current" task that does not match what
the BPF side considers running, so kfuncs and helper state can see an
inconsistent view of the executing task.

sched_ext also tracks runnable work through Dispatch Queues (DSQs) and BPF
chosen dispatch rules, while the core scheduler still maintains classic per-CPU
runqueues and pick paths. A proxy handoff can therefore switch the CPU to a task
that the BPF scheduler never inserted or ordered through its DSQ interface.

DSQ state, vtime, and "who is running" bookkeeping inside the BPF program can
then disagree with what the core actually executes, so helpers and kfuncs that
assume their dispatched task is current may observe stale or inconsistent state.

Design: supporting proxy-exec with sched_ext
============================================

Provide proxy-exec in sched_ext as optional per-scheduler capability: a BPF
scheduler can set the ops flag SCX_OPS_ENQ_BLOCKED to keep mutex-blocked donors
runnable and receive them through ops.enqueue() with SCX_ENQ_BLOCKED set.

This flag requires an ops.enqueue() callback and gives BPF control over whether,
where, and in what order to dispatch each donor. The core then walks the
mutex-owner chain and, when needed, migrates the donor's scheduling context to
the owner's rq before executing the owner.

Without the SCX_OPS_ENQ_BLOCKED flag, mutex waiters block normally and do not
participate in proxy-exec while owned by that scheduler.

The donor-to-owner handoff is modeled like a "function call" from the
scheduler's perspective. The donor remains the running scheduling entity
selected by BPF: its scheduling context, runtime and slice are consumed while
the core temporarily invokes the mutex owner's code to make the critical section
progress. It is not a scheduler-visible switch to the owner.

Accordingly, the donor remains the scheduling context presented to sched_ext,
while the mutex owner is treated solely as the execution context selected
internally by the core scheduler. Scheduling state is accounted against
rq->donor where appropriate, while rq->curr identifies the execution context.
The internal owner substitution does not generate synthetic sched_ext callbacks
for a task that BPF did not dispatch.

Scheduler-facing current-task kfuncs follow the same rule:
scx_bpf_task_running(), scx_bpf_cpu_curr() and scx_bpf_cid_curr() report the
donor selected by the scheduler rather than the mutex owner whose code happens
to execute on its behalf.

The sched_ext callback bookkeeping is adjusted accordingly. Blocked proxy
donors do not generate spurious ops.running() callbacks and ops.stopping() is
only called for tasks that entered a tracked ops.running()/stopping() session.
This state is independent of physical rq->curr execution, which may refer to
the mutex owner instead of the donor.

NOHZ CFS bandwidth checks also follow rq->donor rather than rq->curr, so a
bandwidth-constrained FAIR donor keeps the tick running while an owner executes
on its behalf.

Scheduler ownership changes need special handling because a donor may already be
blocked when a root scheduler is enabled, when a task moves between root and
sub-schedulers, or when PI de-boosting moves it into EXT. If the incoming
scheduler does not set SCX_OPS_ENQ_BLOCKED, the retained donor is removed from
the runqueue and stays on the normal mutex wait path. Otherwise it remains
eligible for BPF admission.

scx_qmap has been modified with a -B option to enable queueing mutex-blocked
tasks for proxy-exec. Blocked donors receive a fresh slice, are inserted at the
head of their current cid's local DSQ and request immediate preemption. This is
an intentionally aggressive policy for making proxy-exec easy to observe.
Capability-rejection re-enqueues bypass this blocked-donor fast path and fall
back to the shared DSQ, avoiding repeated rejection on a cid that no longer has
the required capability. A new statistic reports blocked-donor dispatches.

A new kselftest (enq_blocked) is also introduced to validate proxy-exec with
sched_ext. The test creates a priority inversion with a low-priority owner
(nice +19), a high-priority donor (nice -20) and one nice 0 contender per
available CPU. It exercises a same-CPU topology and a cross-CPU topology, where
the donor and owner run on different CPUs. Each configuration runs with
SCX_OPS_ENQ_BLOCKED first disabled and then enabled, counts blocked-donor
enqueues by CPU and reports mutex hold/wait-time deltas. Access to the mutex is
provided by a loadable kernel module built via TEST_GEN_MODS_DIR, with the test
responsible for its lifecycle.

Example kselftest run:

===== START =====
TEST: enq_blocked
DESCRIPTION: Verify proxy donor admission under CPU-wide contention
OUTPUT:

[topology=same-cpu SCX_OPS_ENQ_BLOCKED=disabled]
  proxy_exec=enabled
  donor_cpu=0
  owner_cpu=0
  nr_contenders=16
  measured_trials=10
  owner_nice=19
  donor_nice=-20
  contender_nice=0
  mutex_hold_avg_ns=263587544 (263.587 ms, samples=10)
  mutex_wait_avg_ns=263597934 (263.597 ms, samples=10)
  nr_blocked_enqueues=0
  nr_blocked_enqueues_donor_cpu=0
  nr_blocked_enqueues_owner_cpu=0
  nr_blocked_enqueues_other_cpu=0
  nr_blocked_wakeups=0

[topology=same-cpu SCX_OPS_ENQ_BLOCKED=enabled]
  proxy_exec=enabled
  donor_cpu=0
  owner_cpu=0
  nr_contenders=16
  measured_trials=10
  owner_nice=19
  donor_nice=-20
  contender_nice=0
  mutex_hold_avg_ns=251789178 (251.789 ms, samples=10)
  mutex_wait_avg_ns=209808599 (209.808 ms, samples=10)
  nr_blocked_enqueues=60
  nr_blocked_enqueues_donor_cpu=60
  nr_blocked_enqueues_owner_cpu=0
  nr_blocked_enqueues_other_cpu=0
  nr_blocked_wakeups=0

[topology=same-cpu delta: enabled - disabled]
  mutex_hold_delta_ns=-11798366 (-4.48%)
  mutex_wait_delta_ns=-53789335 (-20.41%)

[topology=cross-cpu SCX_OPS_ENQ_BLOCKED=disabled]
  proxy_exec=enabled
  donor_cpu=0
  owner_cpu=1
  nr_contenders=16
  measured_trials=10
  owner_nice=19
  donor_nice=-20
  contender_nice=0
  mutex_hold_avg_ns=246794454 (246.794 ms, samples=10)
  mutex_wait_avg_ns=247298497 (247.298 ms, samples=10)
  nr_blocked_enqueues=0
  nr_blocked_enqueues_donor_cpu=0
  nr_blocked_enqueues_owner_cpu=0
  nr_blocked_enqueues_other_cpu=0
  nr_blocked_wakeups=0

[topology=cross-cpu SCX_OPS_ENQ_BLOCKED=enabled]
  proxy_exec=enabled
  donor_cpu=0
  owner_cpu=1
  nr_contenders=16
  measured_trials=10
  owner_nice=19
  donor_nice=-20
  contender_nice=0
  mutex_hold_avg_ns=215486809 (215.486 ms, samples=10)
  mutex_wait_avg_ns=215498090 (215.498 ms, samples=10)
  nr_blocked_enqueues=80
  nr_blocked_enqueues_donor_cpu=20
  nr_blocked_enqueues_owner_cpu=60
  nr_blocked_enqueues_other_cpu=0
  nr_blocked_wakeups=0

[topology=cross-cpu delta: enabled - disabled]
  mutex_hold_delta_ns=-31307645 (-12.69%)
  mutex_wait_delta_ns=-31800407 (-12.86%)
ok 1 enq_blocked #
=====  END  =====


=============================

RESULTS:

PASSED:  1
SKIPPED: 0
FAILED:  0

References
==========

[1] https://lore.kernel.org/all/20251206001451.1418225-1-jstultz@google.com

Git tree: git://git.kernel.org/pub/scm/linux/kernel/git/arighi/linux.git scx-proxy-exec

Changes in v8:
 - Rework remote-DSQ migration check so the locked re-check handles only
   proxy-exec state changes (Tejun Heo)
 - Use the current-donor helper when preventing migration of an active donor
   (sashiko)
 - Handle blocked-donor reactivation into scx schedulers without
   SCX_OPS_ENQ_BLOCKED (John Stultz)
 - Document sched_ext callback behavior under proxy execution (sashiko)
 - Avoid redispatching a rejected blocked donor to the same local DSQ in
   scx_qmap (sashiko)
 - Add a preparatory fix to avoid false migration warnings when proxy-exec moves
   a migration-disabled donor
 - Skip the enq_blocked kselftest when proxy execution is unavailable
 - Link to v7: https://lore.kernel.org/all/20260716132229.61603-1-arighi@nvidia.com/

Changes in v7:
 - Move the remote CPU check out of the lockless remote-DSQ scan and perform it
   only after locking the source rq, retain the locked migration recheck
   immediately before moving the task (sashiko)
 - Complete the scheduler-context conversion from rq->curr to rq->donor for
   capability revocation and preemption checks, make scx_bpf_task_running(),
   scx_bpf_cpu_curr() and scx_bpf_cid_curr() expose the donor selected by the
   scheduler (sashiko)
 - Handle SCX_ENQ_REENQ before scx_qmap's blocked-donor fast path to avoid an
   infinite reject/re-enqueue loop (sashiko)
 - Add scx_qmap stat for SCX_ENQ_BLOCKED dispatches
 - Link to v6: https://lore.kernel.org/all/20260715205622.276220-1-arighi@nvidia.com/

Changes in v6:
 - Add a scheduler-core fix to make NOHZ CFS bandwidth checks follow the proxy
   donor instead of the physical execution context (sashiko)
 - Block retained donors when entering sched_ext through PI de-boosting or
   global scheduler activation (sashiko)
 - Distinguish ordinary wakeup activations from retained-donor admissions so
   SCX_ENQ_BLOCKED is not reported for normal wakeups; extend the selftest to
   detect this regression (sashiko)
 - Validate unexpected-CPU blocked-donor enqueues in both same-CPU and
   cross-CPU test topologies (sashiko)
 - Link to v5: https://lore.kernel.org/all/20260713162112.26785-1-arighi@nvidia.com/

Changes in v5:
 - Split retained-donor deactivation and sched_ext's default rejection into
   preparatory patches so the scheduler-core changes can be routed separately
 - Drop the proxy destination query kfuncs and the preparatory mutex lock-scope
   change (John Stultz)
 - Use p->is_blocked instead of task_is_blocked() to fix a WARN triggered during
   scx_pair testing (John Stultz)
 - Rename SCX_TASK_IS_RUNNING to SCX_TASK_RUN_TRACKED and document that it
   tracks an ops.running()/stopping() session rather than physical rq->curr
   execution
 - Keep a proxy-migrated blocked donor on the owner's rq until wakeup instead of
   allowing BPF-directed migration to pull it back to wake_cpu and cause
   repeated donor migration
 - Extend kselftest to cover same-CPU and cross-CPU proxy-exec switches
 - Make scx_qmap insert blocked donors at the head of their current cid's local
   DSQ and request immediate preemption
 - Link to v4: https://lore.kernel.org/all/20260710083913.30573-1-arighi@nvidia.com/

Changes in v4:
 - Harden scx_bpf_task_proxy_cpu() locking and blocked-state validation, return
   -ENOENT for unrunnable owners and drop unnecessary donor-affinity checks
   (K Prateek Nayak, sashiko)
 - Reschedule remote CPUs after donor deactivation, handle sched_setscheduler()
   admission and assert scheduler-change locking (sashiko)
 - Avoid a potential KCSAN data-race report in the lockless blocked-donor
   migration check by using rcu_access_pointer() for rq->donor (sashiko)
 - Fix tick dependency updates for incoming EXT contexts and keep the tick
   enabled for blocked donors (sashiko)
 - Dump EXT donors in scx_dump_state() (sashiko)
 - Reordered the preparatory sched_ext changes so real running-state tracking
   is established before donor-based accounting, and split the proxy
   destination query kfuncs into a separate patch
 - Add compatibility wrappers for the proxy CPU/cid kfuncs and document their
   results as scheduling hints
 - Link to v3: https://lore.kernel.org/all/20260706070410.282826-1-arighi@nvidia.com/

Changes in v3:
 - Dropped the core restrictions on proxy-migrating migration-disabled and
   single-CPU donors: proxy execution moves the scheduling context, not the
   task's execution context. (Peter Zijlstra, K Prateek Nayak)
 - Dropped the sched_ext-specific put_prev_task()/set_next_task() exception and
   fixed ops.running()/ops.stopping() pairing inside sched_ext instead; track a
   real running transition even when either callback is absent. (Peter Zijlstra)
 - Dropped the kf_tasks[] nesting and nested ops.runnable() patches from v2;
   extensive proxy-exec testing did not reproduce task-op re-entry, so retain
   the existing non-nesting invariant.
 - Replaced scx_bpf_task_is_blocked() with SCX_ENQ_BLOCKED in ops.enqueue()
   flags, identifying blocked-donor admission directly at enqueue time.
 - Expanded the curr/donor description to cover mixed scheduling classes and
   fixed local preemption to expire rq->donor's slice. (Aiqun Maria Yu)
 - Allow inactive blocked donors to be placed on a remote local DSQ while
   preventing normal migration of an active rq donor. (Aiqun Maria Yu)
 - Deactivate retained donors when ownership changes to a root or sub-scheduler
   without SCX_OPS_ENQ_BLOCKED, and extend the selftest to cover attaching a
   scheduler after the donor blocks. (K Prateek Nayak)
 - Harden remote DSQ consumption by rejecting active tasks and rechecking
   migration eligibility after locking the source rq. Fall back to the global
   DSQ without treating an eligibility change during the lock handoff as a BPF
   scheduler error.
 - scx_qmap has a command line option (-B) to enable blocked-donor queueing
 - Link to v2: https://lore.kernel.org/all/20260702171909.1994478-1-arighi@nvidia.com/

Changes in v2:
 - Rebased onto sched_ext/for-7.3 and adapted the series to the split
   sched_ext implementation and cid-form scheduler interfaces.
 - Replaced the global sched_proxy_exec_scx boot-time opt-in with the
   per-scheduler SCX_OPS_ENQ_BLOCKED capability, allowing BPF to control donor
   admission and ordering through ops.enqueue().
 - Added scx_bpf_task_is_blocked(), scx_bpf_task_proxy_cpu(), and
   scx_bpf_task_proxy_cid(); enforce CPU/cid API separation for cid-form
   schedulers.
 - Added proxy exec support to scx_qmap, including optional owner-cid steering,
   affinity validation, and fallback to the donor's current cid.
 - Added a kselftest with a kernel mutex test and a three-task priority
   inversion workload, test is executed with blocked task admission disabled and
   enabled, validates the behavior, and reports hold/wait-time deltas.
 - Link to v1: https://lore.kernel.org/all/20260506174639.535232-1-arighi@nvidia.com/

Andrea Righi (11):
      sched/core: Avoid false migration warning for proxy donors
      sched: Make NOHZ CFS bandwidth checks follow proxy donor
      sched: Add helper to block retained proxy donors
      sched_ext: Block proxy donors across scheduler transitions
      sched_ext: Fix ops.running/stopping() pairing for proxy-exec donors
      sched_ext: Fix proxy-exec race in consume_remote_task()
      sched_ext: Split curr|donor references properly
      sched_ext: Delegate proxy donor admission to BPF schedulers
      sched_ext: Add selftest for blocked donor admission
      sched_ext: scx_qmap: Add proxy execution support
      sched: Allow enabling proxy exec with sched_ext

John Stultz (1):
      sched_ext: Handle blocked donor migration with proxy execution

 Documentation/scheduler/sched-ext.rst              |   6 +
 include/linux/sched/ext.h                          |   3 +
 init/Kconfig                                       |   2 -
 kernel/sched/core.c                                |  59 +-
 kernel/sched/ext/ext.c                             | 374 +++++++--
 kernel/sched/ext/ext.h                             |   6 +
 kernel/sched/ext/internal.h                        |  29 +-
 kernel/sched/ext/sub.c                             |   8 +-
 kernel/sched/ext/sub.h                             |   8 +-
 kernel/sched/sched.h                               |   6 +
 kernel/sched/syscalls.c                            |   3 +
 tools/sched_ext/include/scx/compat.h               |   1 +
 tools/sched_ext/include/scx/enum_defs.autogen.h    |   1 +
 tools/sched_ext/include/scx/enums.autogen.bpf.h    |   3 +
 tools/sched_ext/include/scx/enums.autogen.h        |   1 +
 tools/sched_ext/scx_qmap.bpf.c                     |  34 +
 tools/sched_ext/scx_qmap.c                         |  13 +-
 tools/sched_ext/scx_qmap.h                         |   1 +
 tools/testing/selftests/sched_ext/.gitignore       |   4 +
 tools/testing/selftests/sched_ext/Makefile         |   2 +
 tools/testing/selftests/sched_ext/config           |   2 +
 .../testing/selftests/sched_ext/enq_blocked.bpf.c  | 116 +++
 tools/testing/selftests/sched_ext/enq_blocked.c    | 918 +++++++++++++++++++++
 tools/testing/selftests/sched_ext/enq_blocked.h    |  28 +
 .../selftests/sched_ext/test_modules/Makefile      |  13 +
 .../sched_ext/test_modules/scx_enq_blocked_test.c  | 195 +++++
 26 files changed, 1753 insertions(+), 83 deletions(-)
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.bpf.c
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.c
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.h
 create mode 100644 tools/testing/selftests/sched_ext/test_modules/Makefile
 create mode 100644 tools/testing/selftests/sched_ext/test_modules/scx_enq_blocked_test.c

[PATCH 01/12] sched/core: Avoid false migration warning for proxy donors

[Andrea Righi]

Proxy execution can move a blocked donor's scheduling context to the
lock owner's CPU even when the donor is migration-disabled. The donor
does not execute there, and its original execution CPU remains recorded
in wake_cpu.

set_task_cpu() warns unconditionally for migration-disabled tasks, so a
subsequent proxy migration or the wakeup path returning the donor home
triggers a false positive. Moving a blocked scheduling context does not
violate the migration-disabled execution context.

Exclude blocked proxy donors from the warning. The proxy wakeup path
restores an executable placement before clearing the blocked state.

This is a preparatory fix for enabling proxy execution together with
sched_ext.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/core.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 3cc6fb1d20547..a0e813abddc28 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3376,7 +3376,13 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 	 */
 	WARN_ON_ONCE(!cpu_online(new_cpu));
 
-	WARN_ON_ONCE(is_migration_disabled(p));
+	/*
+	 * Proxy execution can move a blocked task's scheduling context to any
+	 * CPU without moving its migration-disabled execution context. The
+	 * wakeup path will return the task to a CPU where it can execute.
+	 */
+	WARN_ON_ONCE(is_migration_disabled(p) &&
+		     !(sched_proxy_exec() && p->is_blocked));
 
 	trace_sched_migrate_task(p, new_cpu);
 
-- 
2.55.0

[PATCH 02/12] sched: Make NOHZ CFS bandwidth checks follow proxy donor

[Andrea Righi]

Proxy execution separates the scheduling context in rq->donor from the
physical execution context in rq->curr. sched_can_stop_tick() checks the
latter for CFS bandwidth constraints and only does so when nr_running is
one.

A retained proxy donor keeps both the donor and mutex owner queued. The
check therefore misses a constrained FAIR donor and may stop the tick
while its runtime still needs to be enforced.

Check the selected donor instead and remove the nr_running restriction.
The donor being a queued FAIR task is sufficient to require bandwidth
accounting regardless of other runnable tasks.

Fixes: af0c8b2bf67b ("sched: Split scheduler and execution contexts")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Link: https://lore.kernel.org/r/20260713164807.E5ED21F00A3A@smtp.kernel.org
Acked-by: John Stultz <jstultz@google.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/core.c | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index a0e813abddc28..074dc14c22d5d 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1409,11 +1409,8 @@ static void nohz_csd_func(void *info)
 #endif /* CONFIG_NO_HZ_COMMON */
 
 #ifdef CONFIG_NO_HZ_FULL
-static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
+static inline bool __need_bw_check(struct task_struct *p)
 {
-	if (rq->nr_running != 1)
-		return false;
-
 	if (p->sched_class != &fair_sched_class)
 		return false;
 
@@ -1462,14 +1459,17 @@ bool sched_can_stop_tick(struct rq *rq)
 		return false;
 
 	/*
-	 * If there is one task and it has CFS runtime bandwidth constraints
-	 * and it's on the cpu now we don't want to stop the tick.
+	 * If the selected scheduling context has CFS runtime bandwidth
+	 * constraints, don't stop the tick. With proxy execution, rq->donor is
+	 * the selected scheduling context while rq->curr is the task physically
+	 * executing on its behalf.
+	 *
 	 * This check prevents clearing the bit if a newly enqueued task here is
-	 * dequeued by migrating while the constrained task continues to run.
+	 * dequeued by migrating while the constrained donor continues to run.
 	 * E.g. going from 2->1 without going through pick_next_task().
 	 */
-	if (__need_bw_check(rq, rq->curr)) {
-		if (cfs_task_bw_constrained(rq->curr))
+	if (__need_bw_check(rq->donor)) {
+		if (cfs_task_bw_constrained(rq->donor))
 			return false;
 	}
 
-- 
2.55.0

[PATCH 03/12] sched: Add helper to block retained proxy donors

[Andrea Righi]

Scheduler ownership changes may need to turn a mutex-blocked task
retained on the runqueue for proxy execution back into a normally
blocked task.

Add sched_proxy_block_task() to perform that transition while holding
p->pi_lock and the task's runqueue lock. Reset an active donor to the
physical current task before blocking it. If DELAY_DEQUEUE keeps a FAIR
donor queued, the immediately following sched_change_begin() completes
the dequeue through switching_from_fair() before recording its queued
state.

This is a preparatory change to support proxy execution with sched_ext.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/core.c  | 28 ++++++++++++++++++++++++++++
 kernel/sched/sched.h |  6 ++++++
 2 files changed, 34 insertions(+)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 074dc14c22d5d..b1a320365bc16 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -6770,6 +6770,34 @@ static void proxy_deactivate(struct rq *rq, struct task_struct *donor)
 	block_task(rq, donor, state);
 }
 
+/*
+ * Remove a retained proxy donor before changing its scheduler ownership.
+ * The caller holds p->pi_lock, so p cannot wake and migrate if block_task()
+ * drops it from the runqueue. If DELAY_DEQUEUE keeps a blocked fair task
+ * queued, switching_from_fair() completes the dequeue in the immediately
+ * following sched_change_begin().
+ */
+void sched_proxy_block_task(struct rq *rq, struct task_struct *p)
+{
+	unsigned long state = READ_ONCE(p->__state);
+
+	lockdep_assert_held(&p->pi_lock);
+	lockdep_assert_rq_held(rq);
+
+	if (!p->is_blocked || !task_on_rq_queued(p))
+		return;
+	if (WARN_ON_ONCE(state == TASK_RUNNING))
+		return;
+
+	if (task_current_donor(rq, p))
+		proxy_reset_donor(rq);
+
+	if (!p->se.sched_delayed)
+		block_task(rq, p, state);
+
+	WARN_ON_ONCE(task_on_rq_queued(p) && !p->se.sched_delayed);
+}
+
 static inline void proxy_release_rq_lock(struct rq *rq, struct rq_flags *rf)
 	__releases(__rq_lockp(rq))
 {
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 64d79e9efc3d1..437c4f68c9e53 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -2477,6 +2477,12 @@ static inline bool task_is_blocked(struct task_struct *p)
 	return !!p->blocked_on;
 }
 
+#ifdef CONFIG_SCHED_PROXY_EXEC
+void sched_proxy_block_task(struct rq *rq, struct task_struct *p);
+#else
+static inline void sched_proxy_block_task(struct rq *rq, struct task_struct *p) {}
+#endif
+
 static inline int task_on_cpu(struct rq *rq, struct task_struct *p)
 {
 	return p->on_cpu;
-- 
2.55.0

[PATCH 04/12] sched_ext: Block proxy donors across scheduler transitions

[Andrea Righi]

Proxy execution retains mutex-blocked donors on the runqueue so their
scheduling context can execute a lock owner. sched_ext cannot safely
retain such donors unless the BPF scheduler explicitly participates in
their admission and ordering.

Make sched_ext reject retained donors by default. Force blocked EXT
tasks through the regular block path in schedule(), and fully deactivate
a retained donor before sched_setscheduler(), PI de-boosting, or global
activation moves it into the EXT class.

This is a preparatory change to support proxy execution with sched_ext.

Acked-by: John Stultz <jstultz@google.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/core.c     |  5 ++++-
 kernel/sched/ext/ext.c  | 44 +++++++++++++++++++++++++++++++++++++++++
 kernel/sched/ext/ext.h  |  6 ++++++
 kernel/sched/syscalls.c |  3 +++
 4 files changed, 57 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index b1a320365bc16..e6ea4f88153d9 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -7168,7 +7168,8 @@ static void __sched notrace __schedule(int sched_mode)
 		 * task_is_blocked() will always be false).
 		 */
 		try_to_block_task(rq, prev, &prev_state,
-				  !task_is_blocked(prev));
+				  !task_is_blocked(prev) ||
+				  !scx_allow_proxy_exec(prev));
 		switch_count = &prev->nvcsw;
 	}
 
@@ -7721,6 +7722,8 @@ void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
 	if (prev_class != next_class)
 		queue_flag |= DEQUEUE_CLASS;
 
+	scx_prepare_setscheduler(p, next_class);
+
 	scoped_guard (sched_change, p, queue_flag) {
 		/*
 		 * Boosting condition are:
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index b730eac4b13fe..fd1f55c53662b 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -24,6 +24,46 @@
 
 DEFINE_RAW_SPINLOCK(scx_sched_lock);
 
+bool scx_allow_proxy_exec(const struct task_struct *p)
+{
+	return p->sched_class != &ext_sched_class;
+}
+
+/*
+ * Called after sched_setscheduler() validation and immediately before
+ * sched_change_begin(), with @p's pi and rq locks held.
+ */
+void scx_prepare_setscheduler(struct task_struct *p,
+			      const struct sched_class *next_class)
+{
+	lockdep_assert_held(&p->pi_lock);
+	lockdep_assert_rq_held(task_rq(p));
+
+	/*
+	 * Retained proxy donors need admission only when entering EXT. A PI
+	 * boost moves an EXT task to RT/DL and may keep it queued; the matching
+	 * de-boost moves it back to EXT and therefore falls through below.
+	 */
+	if (p->sched_class == next_class || next_class != &ext_sched_class)
+		return;
+
+	sched_proxy_block_task(task_rq(p), p);
+}
+
+/*
+ * Called with @p's pi and rq locks held immediately before
+ * sched_change_begin(). The caller must pass DEQUEUE_NOCLOCK so the rq clock
+ * is updated only once.
+ */
+static void scx_prepare_task_sched_change(struct task_struct *p)
+{
+	lockdep_assert_held(&p->pi_lock);
+	lockdep_assert_rq_held(task_rq(p));
+
+	update_rq_clock(task_rq(p));
+	sched_proxy_block_task(task_rq(p), p);
+}
+
 /*
  * NOTE: sched_ext is in the process of growing multiple scheduler support and
  * scx_root usage is in a transitional state. Naked dereferences are safe if the
@@ -7431,6 +7471,10 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 
 		if (old_class != new_class)
 			queue_flags |= DEQUEUE_CLASS;
+		if (new_class == &ext_sched_class) {
+			scx_prepare_task_sched_change(p);
+			queue_flags |= DEQUEUE_NOCLOCK;
+		}
 
 		scoped_guard (sched_change, p, queue_flags) {
 			set_task_slice(p, READ_ONCE(sch->slice_dfl));
diff --git a/kernel/sched/ext/ext.h b/kernel/sched/ext/ext.h
index 0b7fc46aee08c..d708abf2c3bb8 100644
--- a/kernel/sched/ext/ext.h
+++ b/kernel/sched/ext/ext.h
@@ -18,8 +18,11 @@ bool scx_can_stop_tick(struct rq *rq);
 void scx_rq_activate(struct rq *rq);
 void scx_rq_deactivate(struct rq *rq);
 int scx_check_setscheduler(struct task_struct *p, int policy);
+void scx_prepare_setscheduler(struct task_struct *p,
+			      const struct sched_class *next_class);
 bool task_should_scx(int policy);
 bool scx_allow_ttwu_queue(const struct task_struct *p);
+bool scx_allow_proxy_exec(const struct task_struct *p);
 void init_sched_ext_class(void);
 
 static inline u32 scx_cpuperf_target(s32 cpu)
@@ -52,8 +55,11 @@ static inline bool scx_can_stop_tick(struct rq *rq) { return true; }
 static inline void scx_rq_activate(struct rq *rq) {}
 static inline void scx_rq_deactivate(struct rq *rq) {}
 static inline int scx_check_setscheduler(struct task_struct *p, int policy) { return 0; }
+static inline void scx_prepare_setscheduler(struct task_struct *p,
+					    const struct sched_class *next_class) {}
 static inline bool task_on_scx(const struct task_struct *p) { return false; }
 static inline bool scx_allow_ttwu_queue(const struct task_struct *p) { return true; }
+static inline bool scx_allow_proxy_exec(const struct task_struct *p) { return true; }
 static inline void init_sched_ext_class(void) {}
 
 #endif	/* CONFIG_SCHED_CLASS_EXT */
diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
index b215b0ead9a60..2bbba3dc8c890 100644
--- a/kernel/sched/syscalls.c
+++ b/kernel/sched/syscalls.c
@@ -678,6 +678,9 @@ int __sched_setscheduler(struct task_struct *p,
 	if (prev_class != next_class)
 		queue_flags |= DEQUEUE_CLASS;
 
+	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS))
+		scx_prepare_setscheduler(p, next_class);
+
 	scoped_guard (sched_change, p, queue_flags) {
 
 		if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
-- 
2.55.0

[PATCH 05/12] sched_ext: Fix ops.running/stopping() pairing for proxy-exec donors

[Andrea Righi]

With proxy-exec, pick_next_task() can return a task with p->is_blocked
set (a proxy donor). put_prev_set_next_task() then calls
set_next_task_scx() on this "ghost" task even though the task only
provides scheduling context and never actually runs.

Calling ops.running() for such a donor produces a spurious running
event. Simply suppressing ops.running() is not sufficient because the
following put_prev_task_scx() would still invoke ops.stopping(),
resulting in an unpaired stopping event.

Introduce SCX_TASK_RUN_TRACKED to track whether a task entered an
ops.running()/stopping() session. Set and clear the flag independently
of ops.running() and ops.stopping(), as the callbacks are independently
optional. Invoke ops.running() only for non-blocked tasks and invoke
ops.stopping() only after entering a tracked session. This keeps the
callbacks paired for proxy donors while preserving stopping
notifications for schedulers which only implement ops.stopping().

This is a preparatory change for enabling proxy execution together with
sched_ext. The explicit running-state tracking is also required by later
donor-based accounting: it prevents an EXT owner executing for a non-EXT
donor from being treated as the active EXT scheduling context when it is
dequeued.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 include/linux/sched/ext.h |  2 ++
 kernel/sched/ext/ext.c    | 33 +++++++++++++++++++++++++--------
 2 files changed, 27 insertions(+), 8 deletions(-)

diff --git a/include/linux/sched/ext.h b/include/linux/sched/ext.h
index 78b2f289cb986..901772d8ec15d 100644
--- a/include/linux/sched/ext.h
+++ b/include/linux/sched/ext.h
@@ -103,6 +103,8 @@ enum scx_ent_flags {
 	SCX_TASK_SUB_INIT	= 1 << 4, /* task being initialized for a sub sched */
 	SCX_TASK_IMMED		= 1 << 5, /* task is on local DSQ with %SCX_ENQ_IMMED */
 
+	SCX_TASK_RUN_TRACKED	= 1 << 6, /* task is in an ops.running()/stopping() session */
+
 	/*
 	 * Bits 8 to 10 are used to carry task state:
 	 *
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index fd1f55c53662b..30d5965d09108 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -2217,9 +2217,13 @@ static bool dequeue_task_scx(struct rq *rq, struct task_struct *p, int core_deq_
 	 * information meaningful to the BPF scheduler and can be suppressed by
 	 * skipping the callbacks if the task is !QUEUED.
 	 */
-	if (SCX_HAS_OP(sch, stopping) && task_current(rq, p)) {
-		update_curr_scx(rq);
-		SCX_CALL_OP_TASK(sch, stopping, rq, p, false);
+	if (task_current(rq, p) &&
+	    (p->scx.flags & SCX_TASK_RUN_TRACKED)) {
+		if (SCX_HAS_OP(sch, stopping)) {
+			update_curr_scx(rq);
+			SCX_CALL_OP_TASK(sch, stopping, rq, p, false);
+		}
+		p->scx.flags &= ~SCX_TASK_RUN_TRACKED;
 	}
 
 	if (SCX_HAS_OP(sch, quiescent) && !task_on_rq_migrating(p))
@@ -2921,9 +2925,17 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 
 	p->se.exec_start = rq_clock_task(rq);
 
-	/* see dequeue_task_scx() on why we skip when !QUEUED */
-	if (SCX_HAS_OP(sch, running) && (p->scx.flags & SCX_TASK_QUEUED))
-		SCX_CALL_OP_TASK(sch, running, rq, p);
+	/*
+	 * See dequeue_task_scx() for why we skip when !QUEUED. A blocked task can
+	 * remain queued here only as a retained proxy donor. Skip it because it
+	 * provides scheduling context but never runs itself.
+	 */
+	if ((p->scx.flags & SCX_TASK_QUEUED) && !p->is_blocked) {
+		if (SCX_HAS_OP(sch, running))
+			SCX_CALL_OP_TASK(sch, running, rq, p);
+
+		p->scx.flags |= SCX_TASK_RUN_TRACKED;
+	}
 
 	clr_task_runnable(p, true);
 
@@ -3029,8 +3041,13 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 	update_curr_scx(rq);
 
 	/* see dequeue_task_scx() on why we skip when !QUEUED */
-	if (SCX_HAS_OP(sch, stopping) && (p->scx.flags & SCX_TASK_QUEUED))
-		SCX_CALL_OP_TASK(sch, stopping, rq, p, true);
+	if ((p->scx.flags & SCX_TASK_QUEUED) &&
+	    (p->scx.flags & SCX_TASK_RUN_TRACKED)) {
+		if (SCX_HAS_OP(sch, stopping))
+			SCX_CALL_OP_TASK(sch, stopping, rq, p, true);
+
+		p->scx.flags &= ~SCX_TASK_RUN_TRACKED;
+	}
 
 	if (p->scx.flags & SCX_TASK_QUEUED) {
 		set_task_runnable(rq, p);
-- 
2.55.0

[PATCH 06/12] sched_ext: Fix proxy-exec race in consume_remote_task()

[Andrea Righi]

Without proxy execution, the DSQ lock and holding_cpu handshake ensure
that a task cannot be dequeued or start running during the rq-lock
handoff without clearing holding_cpu.

Proxy execution is an exception. A task can start physically executing
as a lock owner while its own scheduling context remains on the DSQ. Its
on-CPU or migration-disabled state can therefore change without clearing
holding_cpu.

Keep the existing lockless destination validation unchanged. After
switching to the source rq lock, recheck only the on-CPU and
migration-disabled states when proxy execution is enabled. If either
changed, cancel the transfer and requeue the task on the global DSQ
without treating the race as a BPF scheduler error.

A rejected task may be requeued onto the DSQ currently being consumed.
Snapshot the DSQ sequence and skip tasks added after the consume attempt
started to prevent repeatedly selecting the same task. Keep this retry
bounding confined to proxy execution so the non-proxy path remains
unchanged.

With proxy execution enabled, stress-ng --pipeherd can trigger this race
and migrate an active execution context, leading to
sleeping-while-atomic warnings and subsequent lockdep corruption.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/ext/ext.c | 79 ++++++++++++++++++++++++++++++++++++------
 1 file changed, 69 insertions(+), 10 deletions(-)

diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index 30d5965d09108..b5fd91d9bca4f 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -2372,8 +2372,10 @@ static void move_remote_task_to_local_dsq(struct scx_sched *sch,
  * - The BPF scheduler is bypassed while the rq is offline and we can always say
  *   no to the BPF scheduler initiated migrations while offline.
  *
- * The caller must ensure that @p and @rq are on different CPUs.
- * If enforce == true, caller must hold @p's rq lock.
+ * The caller must ensure that @p and @rq are on different CPUs. If @enforce is
+ * true, report violations attributable to BPF-directed migrations. The caller
+ * must hold @p's rq lock to avoid reporting a transient race as a scheduler
+ * error.
  */
 static bool task_can_run_on_remote_rq(struct scx_sched *sch,
 				      struct task_struct *p, struct rq *rq,
@@ -2381,11 +2383,6 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
 {
 	s32 cpu = cpu_of(rq);
 
-	/*
-	 * To prevent races with @p still running on its old CPU while switching
-	 * out, make sure we're holding @p's rq lock so as not to risk
-	 * erroneously killing the BPF scheduler.
-	 */
 	if (enforce)
 		lockdep_assert_rq_held(task_rq(p));
 
@@ -2432,6 +2429,28 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
 	return true;
 }
 
+/*
+ * Proxy execution can change @p's execution and migration-disabled state
+ * without touching its DSQ entry or clearing holding_cpu. Check those states
+ * with @p's rq locked. Without proxy execution, the holding_cpu handshake is
+ * sufficient and this must not affect the existing migration path.
+ */
+static bool task_can_move_from_locked_rq(struct task_struct *p)
+{
+	struct rq *src_rq = task_rq(p);
+
+	lockdep_assert_rq_held(src_rq);
+
+	if (!sched_proxy_exec())
+		return true;
+
+	/* @p may be rq->curr under another task's proxy scheduling context. */
+	if (task_on_cpu(src_rq, p))
+		return false;
+
+	return !is_migration_disabled(p);
+}
+
 /**
  * unlink_dsq_and_switch_rq_lock() - Unlink task and switch to its rq lock
  * @p: target task
@@ -2489,6 +2508,27 @@ static bool consume_remote_task(struct scx_sched *sch, struct rq *this_rq,
 				struct scx_dispatch_q *dsq, struct rq *src_rq)
 {
 	if (unlink_dsq_and_switch_rq_lock(p, dsq, this_rq, src_rq)) {
+		/*
+		 * Proxy execution may have changed @p's running or migration-disabled
+		 * state while switching rq locks without clearing holding_cpu. This is
+		 * a kernel-side race, not an invalid BPF placement request, so don't
+		 * abort the scheduler on failure. Fall back to the global DSQ, where
+		 * normal consumption filters can select a valid destination.
+		 */
+		if (unlikely(!task_can_move_from_locked_rq(p))) {
+			p->scx.dsq = NULL;
+			p->scx.holding_cpu = -1;
+			scx_dispatch_enqueue(sch, src_rq,
+					     find_global_dsq(sch, task_cpu(p)), p,
+					     enq_flags | SCX_ENQ_CLEAR_OPSS |
+					     SCX_ENQ_GDSQ_FALLBACK);
+			if (sched_class_above(p->sched_class,
+					      src_rq->donor->sched_class))
+				resched_curr(src_rq);
+			switch_rq_lock(src_rq, this_rq);
+			return false;
+		}
+
 		move_remote_task_to_local_dsq(sch, p, enq_flags, src_rq, this_rq);
 		return true;
 	} else {
@@ -2527,7 +2567,8 @@ static struct rq *move_task_between_dsqs(struct scx_sched *sch,
 	if (dst_dsq->id == SCX_DSQ_LOCAL) {
 		dst_rq = container_of(dst_dsq, struct rq, scx.local_dsq);
 		if (src_rq != dst_rq &&
-		    unlikely(!task_can_run_on_remote_rq(sch, p, dst_rq, true))) {
+		    unlikely(!task_can_move_from_locked_rq(p) ||
+			     !task_can_run_on_remote_rq(sch, p, dst_rq, true))) {
 			dst_dsq = find_global_dsq(sch, task_cpu(p));
 			dst_rq = src_rq;
 			enq_flags |= SCX_ENQ_GDSQ_FALLBACK;
@@ -2570,7 +2611,9 @@ bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
 			    struct scx_dispatch_q *dsq, u64 enq_flags)
 {
 	struct task_struct *p;
-retry:
+	bool proxy_exec = sched_proxy_exec();
+	u32 scan_seq = 0;
+
 	/*
 	 * The caller can't expect to successfully consume a task if the task's
 	 * addition to @dsq isn't guaranteed to be visible somehow. Test
@@ -2580,10 +2623,23 @@ bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
 		return false;
 
 	raw_spin_lock(&dsq->lock);
+	if (proxy_exec)
+		scan_seq = READ_ONCE(dsq->seq);
 
+retry:
 	nldsq_for_each_task(p, dsq) {
 		struct rq *task_rq = task_rq(p);
 
+		/*
+		 * consume_remote_task() may drop @dsq->lock and requeue a
+		 * rejected task onto this same DSQ. Don't reconsider tasks queued
+		 * after this consume attempt started. This bounds the retry loop
+		 * even when the rejected task remains on-CPU.
+		 */
+		if (proxy_exec &&
+		    unlikely(u32_before(scan_seq, p->scx.dsq_seq)))
+			continue;
+
 		/*
 		 * This loop can lead to multiple lockup scenarios, e.g. the BPF
 		 * scheduler can put an enormous number of affinitized tasks into
@@ -2605,6 +2661,8 @@ bool scx_consume_dispatch_q(struct scx_sched *sch, struct rq *rq,
 		if (task_can_run_on_remote_rq(sch, p, rq, false)) {
 			if (likely(consume_remote_task(sch, rq, p, enq_flags, dsq, task_rq)))
 				return true;
+			/* consume_remote_task() returns with @dsq unlocked. */
+			raw_spin_lock(&dsq->lock);
 			goto retry;
 		}
 	}
@@ -2690,7 +2748,8 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 			p->scx.holding_cpu = -1;
 			scx_dispatch_enqueue(sch, dst_rq, &dst_rq->scx.local_dsq, p,
 					     enq_flags);
-		} else if (unlikely(!task_can_run_on_remote_rq(sch, p, dst_rq, true))) {
+		} else if (unlikely(!task_can_move_from_locked_rq(p) ||
+				    !task_can_run_on_remote_rq(sch, p, dst_rq, true))) {
 			p->scx.holding_cpu = -1;
 			fallback = true;
 			scx_dispatch_enqueue(sch, src_rq, find_global_dsq(sch, task_cpu(p)),
-- 
2.55.0

[PATCH 07/12] sched_ext: Split curr|donor references properly

[Andrea Righi]

With proxy execution, the task selected by the scheduler and the task
physically executing can differ. A blocked mutex waiter donates its
scheduling context to the lock owner:

  D -----------------> M -------------> O ----------------> T
  [donor] blocked on [mutex] owned by [owner] preempted by [task]
     \_________________________________^
          donates scheduling context

where:

  D = blocked donor
  M = mutex
  O = mutex owner
  T = competing runnable task

During a proxy execution switch, D supplies the scheduling class,
priority, and runtime budget, while O supplies the execution context: O
is the task whose code physically executes. T is a competing runnable
task which may preempt the D/O proxy execution.

Consider FAIR and EXT tasks with sched_ext running in partial mode. FAIR
can be replaced with a higher scheduling class such as RT or deadline
without changing the class interaction described here. The possible
combinations are:

  1. D is EXT, O is EXT, T is EXT

     D can interrupt T according to BPF scheduling policy. O executes
     with D's EXT priority and runtime budget, while T waits in EXT.

  2. D is EXT, O is EXT, T is FAIR

     D is visible to the BPF scheduler, but cannot preempt T because
     EXT is below FAIR. Once T stops, BPF can dispatch D and O executes
     with D's EXT priority and runtime budget. If T becomes runnable
     again, it preempts the D/O proxy execution.

  3. D is EXT, O is FAIR, T is EXT

     This cannot represent T preempting O because EXT is below FAIR.

  4. D is EXT, O is FAIR, T is FAIR

     D cannot boost O above T because EXT is below FAIR. O and T
     continue competing under FAIR. Once O releases M, D wakes and
     resumes normal EXT scheduling.

  5. D is FAIR, O is EXT, T is EXT

     D preempts T as the higher-class scheduling context. O executes
     with D's FAIR priority and runtime budget, while T waits in EXT.
     D is not visible to the BPF scheduler.

  6. D is FAIR, O is EXT, T is FAIR

     D competes with T according to its FAIR deadline. When D is
     selected, O executes with D's FAIR priority and runtime budget.
     D is not visible to the BPF scheduler.

  7. D is FAIR, O is FAIR, T is EXT

     This cannot represent T preempting O because EXT is below FAIR.

  8. D is FAIR, O is FAIR, T is FAIR

     O, T, and D all have FAIR scheduling contexts. D remains runnable
     as a blocked proxy donor. When CFS selects D, O executes using D's
     FAIR scheduling context. When CFS selects O, O executes using its
     own FAIR context, and when CFS selects T, T executes normally. D
     is not visible to the BPF scheduler.

Thus, sched_ext policy and accounting must generally use rq->donor, the
scheduler-selected task which supplies the scheduling context, rather
than rq->curr, the task whose code physically executes. Without proxy
execution they are the same task.

On nohz_full CPUs, a blocked proxy donor must retain the scheduler tick
even when it has an infinite slice. Otherwise, a full dynticks CPU could
stop the tick while rq->curr and rq->donor differ, violating assumptions
made by the remote NOHZ tick path.

This is a conservative compromise that keeps the change local to
sched_ext, at the cost of a periodic tick while a blocked proxy donor is
selected. Allowing blocked proxy donors to run tickless would require
making the core scheduler's remote tick handling aware that rq->curr and
rq->donor can differ.

Moreover, extend scx_dump_state() to report both contexts. Each CPU
record now includes a donor= line. If an EXT donor differs from
rq->curr, also emit its detailed task record. The existing '*' marker
continues to identify rq->curr, while the donor= line identifies the
otherwise unmarked donor record.

Note that at this point in the series, CONFIG_SCHED_PROXY_EXEC still
depends on !CONFIG_SCHED_CLASS_EXT, so proxy execution and sched_ext
cannot be enabled together. The scheduling changes are therefore
preparatory. A later patch removes this restriction.

Co-developed-by: John Stultz <jstultz@google.com>
Signed-off-by: John Stultz <jstultz@google.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 Documentation/scheduler/sched-ext.rst |   6 ++
 kernel/sched/ext/ext.c                | 107 +++++++++++++++++---------
 kernel/sched/ext/internal.h           |   6 ++
 kernel/sched/ext/sub.h                |   8 +-
 4 files changed, 87 insertions(+), 40 deletions(-)

diff --git a/Documentation/scheduler/sched-ext.rst b/Documentation/scheduler/sched-ext.rst
index 2771ea4cc14af..4d8bcbdacb9fc 100644
--- a/Documentation/scheduler/sched-ext.rst
+++ b/Documentation/scheduler/sched-ext.rst
@@ -487,6 +487,12 @@ and edge cases, to name a few examples:
   class, in which case it will exit the tick-dispatch loop even though it is runnable
   and has a non-zero slice.
 
+* Under proxy execution, sched_ext continues to observe the donor as the current
+  scheduling context. A blocked donor does not enter an ``ops.running()`` /
+  ``ops.stopping()`` session because it does not execute itself, and the lock
+  owner executing on its behalf is intentionally not reported through these
+  callbacks.
+
 See the "Scheduling Cycle" section for a more detailed description of how
 a freshly woken up task gets on a CPU.
 
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index b5fd91d9bca4f..a466adf8e1548 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -1341,20 +1341,27 @@ static void apply_task_slice_oob(struct rq *rq, struct task_struct *p)
 
 static void update_curr_scx(struct rq *rq)
 {
-	struct task_struct *curr = rq->curr;
+	struct task_struct *donor;
 	s64 delta_exec;
 
+	/*
+	 * update_curr_scx() is selected through rq->donor->sched_class, not
+	 * rq->curr->sched_class, so @donor is always an EXT task here. If an EXT
+	 * owner executes for a FAIR donor, FAIR's update_curr() runs instead.
+	 */
+	donor = rq->donor;
+
 	/* apply even on 0 delta_exec, callers may still act on the slice */
-	apply_task_slice_oob(rq, curr);
+	apply_task_slice_oob(rq, donor);
 
 	delta_exec = update_curr_common(rq);
 	if (unlikely(delta_exec <= 0))
 		return;
 
-	if (curr->scx.slice != SCX_SLICE_INF) {
-		curr->scx.slice -= min_t(u64, curr->scx.slice, delta_exec);
-		if (!curr->scx.slice)
-			touch_core_sched(rq, curr);
+	if (donor->scx.slice != SCX_SLICE_INF) {
+		donor->scx.slice -= min_t(u64, donor->scx.slice, delta_exec);
+		if (!donor->scx.slice)
+			touch_core_sched(rq, donor);
 	}
 
 	dl_server_update(&rq->ext_server, delta_exec);
@@ -1524,9 +1531,9 @@ static void rq_owned_post_enq(struct scx_sched *sch, struct rq *rq,
 	if (rq->scx.flags & SCX_RQ_IN_BALANCE)
 		return;
 
-	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->curr &&
-	    rq->curr->sched_class == &ext_sched_class) {
-		set_task_slice(rq->curr, 0);
+	if ((enq_flags & SCX_ENQ_PREEMPT) && p != rq->donor &&
+	    rq->donor->sched_class == &ext_sched_class) {
+		set_task_slice(rq->donor, 0);
 		resched_curr(rq);
 	}
 }
@@ -2761,7 +2768,8 @@ static void dispatch_to_local_dsq(struct scx_sched *sch, struct rq *rq,
 		}
 
 		/* if the destination CPU is idle, wake it up */
-		if (!fallback && sched_class_above(p->sched_class, dst_rq->curr->sched_class))
+		if (!fallback && sched_class_above(p->sched_class,
+						      dst_rq->donor->sched_class))
 			resched_curr(dst_rq);
 	}
 
@@ -2972,6 +2980,7 @@ static int balance_one(struct rq *rq, struct task_struct *prev)
 static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 {
 	struct scx_sched *sch = scx_task_sched(p);
+	bool can_stop_tick;
 
 	if (p->scx.flags & SCX_TASK_QUEUED) {
 		/*
@@ -3000,6 +3009,7 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 
 	/* apply any pending out-of-band slice request before the tick decision */
 	apply_task_slice_oob(rq, p);
+	can_stop_tick = p->scx.slice == SCX_SLICE_INF && !p->is_blocked;
 
 	/*
 	 * @p is getting newly scheduled or got kicked after someone updated its
@@ -3010,7 +3020,7 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 	 * nohz. In the future, we might want to add a mechanism to update
 	 * load_avgs periodically on tick-stopped CPUs.
 	 */
-	if (p->scx.slice == SCX_SLICE_INF) {
+	if (can_stop_tick) {
 		if (!(rq->scx.flags & SCX_RQ_CAN_STOP_TICK)) {
 			/*
 			 * Bypass mode always assigns finite slices, so @p
@@ -3031,7 +3041,8 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 
 		/*
 		 * @rq still references the outgoing scheduling context. A finite
-		 * slice is sufficient by itself to require the tick.
+		 * slice or a blocked proxy donor is sufficient by itself to require
+		 * the tick.
 		 */
 		if (tick_nohz_full_cpu(cpu_of(rq)))
 			tick_nohz_dep_set_cpu(cpu_of(rq), TICK_DEP_BIT_SCHED);
@@ -3206,7 +3217,7 @@ static struct task_struct *first_local_task(struct rq *rq)
 static struct task_struct *
 do_pick_task_scx(struct rq *rq, struct rq_flags *rf, bool force_scx)
 {
-	struct task_struct *prev = rq->curr;
+	struct task_struct *prev = rq->donor;
 	bool keep_prev;
 	struct task_struct *p;
 
@@ -3566,9 +3577,9 @@ void scx_tick(struct rq *rq)
 	update_other_load_avgs(rq);
 }
 
-static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
+static void task_tick_scx(struct rq *rq, struct task_struct *donor, int queued)
 {
-	struct scx_sched *sch = scx_task_sched(curr);
+	struct scx_sched *sch = scx_task_sched(donor);
 
 	update_curr_scx(rq);
 
@@ -3577,13 +3588,13 @@ static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
 	 * we can't trust the slice management or ops.core_sched_before().
 	 */
 	if (scx_bypassing(sch, cpu_of(rq))) {
-		set_task_slice(curr, 0);
-		touch_core_sched(rq, curr);
+		set_task_slice(donor, 0);
+		touch_core_sched(rq, donor);
 	} else if (SCX_HAS_OP(sch, tick)) {
-		SCX_CALL_OP_TASK(sch, tick, rq, curr);
+		SCX_CALL_OP_TASK(sch, tick, rq, donor);
 	}
 
-	if (!curr->scx.slice)
+	if (!donor->scx.slice)
 		resched_curr(rq);
 }
 
@@ -4221,16 +4232,16 @@ static u32 reenq_local(struct scx_sched *sch, struct rq *rq, u64 reenq_flags)
 	}
 
 	/*
-	 * The revoke that scheduled this scan may have raced the pick: curr
+	 * The revoke that scheduled this scan may have raced the pick: donor
 	 * may be a now-capless task, either one that kept running or one
 	 * promoted off the local DSQ between the ecaps sync and this scan.
 	 * Zero the slice to evict it. The enqueue gate blocks new capless
 	 * inserts, so no later pick can slip through after the scan.
 	 */
 	if ((reenq_flags & SCX_REENQ_CAP_REVOKE) &&
-	    rq->curr->sched_class == &ext_sched_class &&
-	    scx_task_reenq_on_cap_revoke(rq, rq->curr)) {
-		set_task_slice(rq->curr, 0);
+	    rq->donor->sched_class == &ext_sched_class &&
+	    scx_task_reenq_on_cap_revoke(rq, rq->donor)) {
+		set_task_slice(rq->donor, 0);
 		resched_curr(rq);
 	}
 
@@ -4416,14 +4427,18 @@ static void run_deferred(struct rq *rq)
 #ifdef CONFIG_NO_HZ_FULL
 bool scx_can_stop_tick(struct rq *rq)
 {
-	struct task_struct *p = rq->curr;
+	struct task_struct *p = rq->donor;
 	struct scx_sched *sch = scx_task_sched(p);
 
+	/* The remote tick path assumes that proxy execution is not active. */
+	if (rq->curr != rq->donor)
+		return false;
+
 	if (p->sched_class != &ext_sched_class)
 		return true;
 
 	/*
-	 * @rq->curr may still reference an outgoing EXT task after it has been
+	 * @rq->donor may still reference an outgoing EXT task after it has been
 	 * dequeued. If no EXT tasks are accounted on @rq, ignore its stale
 	 * slice state. If another task is dispatched from a DSQ,
 	 * set_next_task_scx() will update the dependency for the incoming task.
@@ -4437,7 +4452,8 @@ bool scx_can_stop_tick(struct rq *rq)
 	/*
 	 * @rq can dispatch from different DSQs, so we can't tell whether it
 	 * needs the tick or not by looking at nr_running. Allow stopping ticks
-	 * iff the BPF scheduler indicated so. See set_next_task_scx().
+	 * iff set_next_task_scx() determined that the selected scheduling context
+	 * can run tickless.
 	 */
 	return rq->scx.flags & SCX_RQ_CAN_STOP_TICK;
 }
@@ -6626,6 +6642,9 @@ static void scx_dump_cpu(struct scx_sched *sch, struct seq_buf *s,
 	dump_line(&ns, "          curr=%s[%d] class=%ps",
 		  rq->curr->comm, rq->curr->pid,
 		  rq->curr->sched_class);
+	dump_line(&ns, "          donor=%s[%d] class=%ps",
+		  rq->donor->comm, rq->donor->pid,
+		  rq->donor->sched_class);
 	if (!cpumask_empty(pcpu->cpus_to_kick))
 		dump_line(&ns, "  cpus_to_kick   : %*pb",
 			  cpumask_pr_args(pcpu->cpus_to_kick));
@@ -6669,6 +6688,10 @@ static void scx_dump_cpu(struct scx_sched *sch, struct seq_buf *s,
 	if (rq->curr->sched_class == &ext_sched_class &&
 	    (dump_all_tasks || scx_task_on_sched(sch, rq->curr)))
 		scx_dump_task(sch, s, dctx, rq, rq->curr, '*');
+	if (rq->donor != rq->curr &&
+	    rq->donor->sched_class == &ext_sched_class &&
+	    (dump_all_tasks || scx_task_on_sched(sch, rq->donor)))
+		scx_dump_task(sch, s, dctx, rq, rq->donor, ' ');
 
 	list_for_each_entry(p, &rq->scx.runnable_list, scx.runnable_node)
 		if (dump_all_tasks || scx_task_on_sched(sch, p))
@@ -8184,7 +8207,7 @@ static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_r
 	unsigned long flags;
 
 	raw_spin_rq_lock_irqsave(rq, flags);
-	cur_class = rq->curr->sched_class;
+	cur_class = rq->donor->sched_class;
 
 	/*
 	 * During CPU hotplug, a CPU may depend on kicking itself to make
@@ -8201,7 +8224,7 @@ static bool kick_one_cpu(s32 cpu, struct scx_sched_pcpu *pcpu, struct rq *this_r
 			if (cur_class == &ext_sched_class) {
 				if (likely(!scx_missing_caps(pcpu->sch, cpu,
 							     scx_caps_for_preempt(pcpu->sch, rq))))
-					set_task_slice(rq->curr, 0);
+					set_task_slice(rq->donor, 0);
 				else
 					__scx_add_event(pcpu->sch,
 							SCX_EV_SUB_PREEMPT_DENIED, 1);
@@ -9172,13 +9195,15 @@ __bpf_kfunc bool scx_bpf_task_set_slice(struct task_struct *p, u64 slice,
 		return false;
 
 	/*
-	 * Directly write only when we hold the lock of the rq @p is queued or
-	 * running on. See the slice write rules above.
+	 * Directly write only when we hold the lock of the rq @p is queued on or
+	 * provides the current scheduling context for. Under proxy execution,
+	 * rq->donor owns and consumes the slice while rq->curr executes on its
+	 * behalf. See the slice write rules above.
 	 */
 	locked_rq = scx_locked_rq();
 	if (!locked_rq ||
 	    (READ_ONCE(p->scx.runnable_cpu) != cpu_of(locked_rq) &&
-	     !task_current(locked_rq, p))) {
+	     !task_current_donor(locked_rq, p))) {
 		set_task_slice_oob(sch, p, slice);
 		return true;
 	}
@@ -10013,12 +10038,17 @@ __bpf_kfunc void scx_bpf_put_cpumask(const struct cpumask *cpumask)
 }
 
 /**
- * scx_bpf_task_running - Is task currently running?
+ * scx_bpf_task_running - Is task the current scheduling context?
  * @p: task of interest
+ *
+ * Under proxy execution, this reports the donor rather than the task whose
+ * code is physically executing. The physical execution context is intentionally
+ * not exposed to the BPF scheduler, which continues to observe the donor as the
+ * running scheduling context.
  */
 __bpf_kfunc bool scx_bpf_task_running(const struct task_struct *p)
 {
-	return task_rq(p)->curr == p;
+	return rcu_access_pointer(task_rq(p)->donor) == p;
 }
 
 /**
@@ -10075,10 +10105,15 @@ __bpf_kfunc struct rq *scx_bpf_locked_rq(const struct bpf_prog_aux *aux)
 }
 
 /**
- * scx_bpf_cpu_curr - Return remote CPU's curr task
+ * scx_bpf_cpu_curr - Return remote CPU's current scheduling context
  * @cpu: CPU of interest
  * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
  *
+ * Under proxy execution, this returns the donor, which supplies the scheduling
+ * policy and runtime budget, rather than the task whose code is physically
+ * executing. The physical execution context is intentionally not exposed to
+ * the BPF scheduler.
+ *
  * Callers must hold RCU read lock (KF_RCU).
  */
 __bpf_kfunc struct task_struct *scx_bpf_cpu_curr(s32 cpu, const struct bpf_prog_aux *aux)
@@ -10094,7 +10129,7 @@ __bpf_kfunc struct task_struct *scx_bpf_cpu_curr(s32 cpu, const struct bpf_prog_
 	if (!scx_cpu_valid(sch, cpu, NULL))
 		return NULL;
 
-	return rcu_dereference(cpu_rq(cpu)->curr);
+	return rcu_dereference(cpu_rq(cpu)->donor);
 }
 
 /**
@@ -10118,7 +10153,7 @@ __bpf_kfunc struct task_struct *scx_bpf_cid_curr(s32 cid, const struct bpf_prog_
 	cpu = scx_cid_to_cpu(sch, cid);
 	if (cpu < 0)
 		return NULL;
-	return rcu_dereference(cpu_rq(cpu)->curr);
+	return rcu_dereference(cpu_rq(cpu)->donor);
 }
 
 /**
diff --git a/kernel/sched/ext/internal.h b/kernel/sched/ext/internal.h
index 26bfda216524e..5dfb7e466108f 100644
--- a/kernel/sched/ext/internal.h
+++ b/kernel/sched/ext/internal.h
@@ -447,6 +447,12 @@ struct sched_ext_ops {
 	 * Therefore, always use scx_bpf_task_cpu(@p) to determine the
 	 * target CPU the task is going to use.
 	 *
+	 * Under proxy execution, the BPF scheduler continues to observe the
+	 * donor as the current scheduling context. A blocked donor does not
+	 * enter a ->running()/->stopping() session because it does not execute
+	 * itself, and the lock owner executing on its behalf is intentionally
+	 * not reported through these callbacks.
+	 *
 	 * See ->runnable() for explanation on the task state notifiers.
 	 */
 	void (*running)(struct task_struct *p);
diff --git a/kernel/sched/ext/sub.h b/kernel/sched/ext/sub.h
index 625d7ce334aa8..a164e6b2f2562 100644
--- a/kernel/sched/ext/sub.h
+++ b/kernel/sched/ext/sub.h
@@ -141,14 +141,14 @@ static inline u64 scx_caps_for_task(struct task_struct *p)
 	return SCX_CAP_ENQ;
 }
 
-/* the cap @sch needs to preempt @rq's current task, 0 if none */
+/* the cap @sch needs to preempt @rq's current scheduling context, 0 if none */
 static inline u64 scx_caps_for_preempt(struct scx_sched *sch, struct rq *rq)
 {
-	struct task_struct *curr = rq->curr;
+	struct task_struct *donor = rq->donor;
 
 	/* a non-ext task can't be preempted by ext, own-subtree needs no cap */
-	if (curr->sched_class != &ext_sched_class ||
-	    scx_is_descendant(scx_task_sched(curr), sch))
+	if (donor->sched_class != &ext_sched_class ||
+	    scx_is_descendant(scx_task_sched(donor), sch))
 		return 0;
 	return SCX_CAP_PREEMPT;
 }
-- 
2.55.0

[PATCH 08/12] sched_ext: Handle blocked donor migration with proxy execution

[Andrea Righi]

From: John Stultz <jstultz@google.com>

With proxy execution enabled, mutex-blocked donors stay runnable so
their scheduling context can execute the lock owner.

sched_ext may normally relocate a queued blocked donor: set_task_cpu()
updates both task_cpu() and wake_cpu, making the destination rq the
donor's new callback home. This remains valid if the donor was
previously proxy-migrated: the BPF scheduler is intentionally selecting
a new return destination, which the next proxy migration will preserve.
The usual migration-disabled, affinity, and rq-online checks apply.

An active donor cannot be moved normally because the source rq still
references it through rq->donor for scheduling-class operations and
runtime accounting. Moving it would associate the task with a different
rq while leaving those source-rq references in place. The core proxy
migration path avoids this by switching the rq donor to idle before
moving the scheduling context.

Check active-donor state in task_can_move_from_locked_rq(), where the
source rq is held and the result is authoritative. Keep it out of the
preliminary lockless DSQ filter, so a donor transition during the
rq-lock handoff is resolved by the locked recheck rather than consuming
a dispatch kick and skipping the task.

Allow normal sched_ext migration of blocked donors unless the donor is
active on its rq. In that case, defer migration until it is switched out
or leave it to the proxy machinery.

Keep blocked donors on the local DSQ so they remain visible to the proxy
pick path. This is a temporary solution: in a later commit, proxy donor
admission will be delegated to the BPF scheduler and donors will instead
be routed through ops.enqueue().

Signed-off-by: John Stultz <jstultz@google.com>
Co-developed-by: Andrea Righi <arighi@nvidia.com>
Signed-off-by: Andrea Righi <arighi@nvidia.com>

 interactive rebase in progress; onto c0414e1387386
---
 kernel/sched/ext/ext.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index a466adf8e1548..bf53f2be77f05 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -2455,6 +2455,10 @@ static bool task_can_move_from_locked_rq(struct task_struct *p)
 	if (task_on_cpu(src_rq, p))
 		return false;
 
+	/* Don't move the current scheduling context off its source rq. */
+	if (task_current_donor(src_rq, p))
+		return false;
+
 	return !is_migration_disabled(p);
 }
 
@@ -3122,6 +3126,25 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 	if (p->scx.flags & SCX_TASK_QUEUED) {
 		set_task_runnable(rq, p);
 
+		/*
+		 * Mutex-blocked donors stay queued on the runqueue under proxy
+		 * execution, but the donor never runs as itself, proxy-exec
+		 * walks the blocked_on chain on the next __schedule() and runs
+		 * the lock owner in its place.
+		 *
+		 * Put the donor on the local DSQ directly so pick_next_task()
+		 * can still see it. find_proxy_task() will either run the chain
+		 * owner or deactivate the donor so the wakeup path can return it
+		 * and let BPF make a new dispatch decision once it is unblocked.
+		 *
+		 * This is preparatory code: a later patch will delegate blocked-donor
+		 * admission to the BPF scheduler.
+		 */
+		if (p->is_blocked) {
+			scx_dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, 0);
+			goto switch_class;
+		}
+
 		/*
 		 * If @p has slice left and is being put, @p is getting
 		 * preempted by a higher priority scheduler class or core-sched
-- 
2.55.0

[PATCH 09/12] sched_ext: Delegate proxy donor admission to BPF schedulers

[Andrea Righi]

Proxy execution keeps a blocked donor runnable so its scheduling context
can execute the mutex owner. Dispatching sched_ext donors on a local
DSQ bypasses the BPF scheduler ordering policy and can give donors
more CPU priority than intended to perform the proxy execution handoff.

Add SCX_OPS_ENQ_BLOCKED as an explicit proxy execution capability. Tasks
owned by schedulers without the flag block normally. Schedulers with the
flag receive blocked donors through ops.enqueue() with SCX_ENQ_BLOCKED
set in enq_flags and can apply their own admission policy.

From a high-level perspective, the resulting flow for a BPF scheduler
with SCX_OPS_ENQ_BLOCKED is:

    D ------ blocked on -----> M ------ owned by -----> O
 [donor]                    [mutex]                [owner]
    |
    | ops.enqueue(D, SCX_ENQ_BLOCKED)
    | BPF dispatches D to CPUi
    v
 +-----------------+
 | CPUi local DSQ  |
 +-------+---------+
         |
         | pick_next_task() selects D
         v
 +-----------------+
 |   proxy exec    | move D to O's rq
 +-------+---------+
         |
         | run O using D's scheduling context
         v
    rq->curr  = O
    rq->donor = D
         |
         | O releases M
         v
 +-----------------+
 |   proxy exec    | return D to CPUi via wakeup
 +-----------------+

A proxy migration preserves the donor's wake_cpu while moving its
scheduling context to the owner's rq, so task_cpu() differs from wake_cpu
until the donor wakes. Prevent BPF-directed migrations from pulling a
blocked donor off this proxy rq. Otherwise each blocked re-enqueue may
move the donor back to its callback rq only for proxy execution to move
it to the owner again.

Scheduler ownership can change after a donor has already blocked. Since
sched_change preserves queued state, handle both iterator-driven moves
to root or sub-schedulers and sched_setscheduler() transitions into EXT.
Before entering a scheduler without the flag, deactivate the retained
donor. It remains on the mutex wait path and wakes normally when the
mutex becomes available.

The global sched_ext enable state can change while __schedule() holds
the runqueue lock. Once an EXT task has an assigned scheduler, consult
it directly so the task cannot retain a donor during the transition
unless the scheduler opted in. Tasks without an assigned scheduler keep
the generic proxy execution behavior.

The preparation helpers require and assert that p->pi_lock and p's rq
lock are held before updating rq state.

The donor starts associated with its original CPU. A BPF scheduler may
dispatch it directly into that CPU local DSQ to let the core resolve the
mutex owner and execute it with the donor scheduling context.

A normal sleeper also has p->is_blocked set until after
wakeup_preempt(). Exclude full wakeup activations from SCX_ENQ_BLOCKED
and carry a transient task flag from enqueue_task_scx() to
wakeup_preempt_scx(). Retained donor wakeups skip enqueue_task_scx(), so
they remain distinguishable without changing core scheduler wake flags.

Reschedule a retained donor when its mutex wakes it so ops.dispatch()
can reconsider the now-unblocked task. Make SCX_OPS_ENQ_BLOCKED
override the exiting and migration-disabled enqueue fallbacks so
opted-in schedulers receive all eligible donor requests.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 include/linux/sched/ext.h                     |   1 +
 kernel/sched/ext/ext.c                        | 122 ++++++++++++++----
 kernel/sched/ext/internal.h                   |  23 +++-
 kernel/sched/ext/sub.c                        |   8 +-
 tools/sched_ext/include/scx/compat.h          |   1 +
 .../sched_ext/include/scx/enum_defs.autogen.h |   1 +
 .../sched_ext/include/scx/enums.autogen.bpf.h |   3 +
 tools/sched_ext/include/scx/enums.autogen.h   |   1 +
 8 files changed, 134 insertions(+), 26 deletions(-)

diff --git a/include/linux/sched/ext.h b/include/linux/sched/ext.h
index 901772d8ec15d..946659749d58a 100644
--- a/include/linux/sched/ext.h
+++ b/include/linux/sched/ext.h
@@ -104,6 +104,7 @@ enum scx_ent_flags {
 	SCX_TASK_IMMED		= 1 << 5, /* task is on local DSQ with %SCX_ENQ_IMMED */
 
 	SCX_TASK_RUN_TRACKED	= 1 << 6, /* task is in an ops.running()/stopping() session */
+	SCX_TASK_ENQ_WAKEUP	= 1 << 7, /* wakeup enqueue awaiting wakeup_preempt() */
 
 	/*
 	 * Bits 8 to 10 are used to carry task state:
diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index bf53f2be77f05..1e91f5e71cbdd 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -26,7 +26,18 @@ DEFINE_RAW_SPINLOCK(scx_sched_lock);
 
 bool scx_allow_proxy_exec(const struct task_struct *p)
 {
-	return p->sched_class != &ext_sched_class;
+	struct scx_sched *sch;
+
+	if (p->sched_class != &ext_sched_class)
+		return true;
+
+	/*
+	 * scx_enabled() may change while __schedule() holds only @p's rq lock.
+	 * Once @p is associated with a scheduler, use that scheduler's policy
+	 * even while the global enable state is transitioning.
+	 */
+	sch = scx_task_sched(p);
+	return !sch || (sch->ops.flags & SCX_OPS_ENQ_BLOCKED);
 }
 
 /*
@@ -36,6 +47,8 @@ bool scx_allow_proxy_exec(const struct task_struct *p)
 void scx_prepare_setscheduler(struct task_struct *p,
 			      const struct sched_class *next_class)
 {
+	struct scx_sched *sch;
+
 	lockdep_assert_held(&p->pi_lock);
 	lockdep_assert_rq_held(task_rq(p));
 
@@ -47,7 +60,13 @@ void scx_prepare_setscheduler(struct task_struct *p,
 	if (p->sched_class == next_class || next_class != &ext_sched_class)
 		return;
 
-	sched_proxy_block_task(task_rq(p), p);
+	sch = scx_task_sched(p);
+	if (WARN_ON_ONCE(!sch))
+		return;
+
+	/* Block retained donors that the incoming scheduler cannot manage. */
+	if (!(sch->ops.flags & SCX_OPS_ENQ_BLOCKED))
+		sched_proxy_block_task(task_rq(p), p);
 }
 
 /*
@@ -55,13 +74,16 @@ void scx_prepare_setscheduler(struct task_struct *p,
  * sched_change_begin(). The caller must pass DEQUEUE_NOCLOCK so the rq clock
  * is updated only once.
  */
-static void scx_prepare_task_sched_change(struct task_struct *p)
+void scx_prepare_task_sched_change(struct task_struct *p, struct scx_sched *sch)
 {
 	lockdep_assert_held(&p->pi_lock);
 	lockdep_assert_rq_held(task_rq(p));
 
 	update_rq_clock(task_rq(p));
-	sched_proxy_block_task(task_rq(p), p);
+
+	/* Block retained donors that the incoming scheduler cannot manage. */
+	if (!(sch->ops.flags & SCX_OPS_ENQ_BLOCKED))
+		sched_proxy_block_task(task_rq(p), p);
 }
 
 /*
@@ -1922,6 +1944,7 @@ void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
 	struct scx_sched *sch = scx_task_sched(p);
 	struct task_struct **ddsp_taskp;
 	struct scx_dispatch_q *dsq;
+	bool enq_blocked;
 	unsigned long qseq;
 
 	WARN_ON_ONCE(!(p->scx.flags & SCX_TASK_QUEUED));
@@ -1954,15 +1977,22 @@ void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
 	if (p->scx.ddsp_dsq_id != SCX_DSQ_INVALID)
 		goto direct;
 
+	/* %SCX_OPS_ENQ_BLOCKED takes precedence over the fallbacks below. */
+	enq_blocked = (sch->ops.flags & SCX_OPS_ENQ_BLOCKED) &&
+		      p->is_blocked && !(enq_flags & SCX_ENQ_WAKEUP);
+	if (enq_blocked)
+		enq_flags |= SCX_ENQ_BLOCKED;
+
 	/* see %SCX_OPS_ENQ_EXITING */
-	if (!(sch->ops.flags & SCX_OPS_ENQ_EXITING) &&
+	if (!enq_blocked && !(sch->ops.flags & SCX_OPS_ENQ_EXITING) &&
 	    unlikely(p->flags & PF_EXITING)) {
 		__scx_add_event(sch, SCX_EV_ENQ_SKIP_EXITING, 1);
 		goto local;
 	}
 
 	/* see %SCX_OPS_ENQ_MIGRATION_DISABLED */
-	if (!(sch->ops.flags & SCX_OPS_ENQ_MIGRATION_DISABLED) &&
+	if (!enq_blocked &&
+	    !(sch->ops.flags & SCX_OPS_ENQ_MIGRATION_DISABLED) &&
 	    is_migration_disabled(p)) {
 		__scx_add_event(sch, SCX_EV_ENQ_SKIP_MIGRATION_DISABLED, 1);
 		goto local;
@@ -2070,8 +2100,17 @@ static void enqueue_task_scx(struct rq *rq, struct task_struct *p, int core_enq_
 	int sticky_cpu = p->scx.sticky_cpu;
 	u64 enq_flags = core_enq_flags | rq->scx.remote_activate_enq_flags;
 
-	if (enq_flags & ENQUEUE_WAKEUP)
+	/*
+	 * p->is_blocked is cleared after wakeup_preempt(), so remember whether
+	 * this is a full wakeup activation. If wakeup_preempt_scx() isn't called,
+	 * set_next_task_scx() or a subsequent non-wakeup enqueue clears the flag.
+	 */
+	if (enq_flags & ENQUEUE_WAKEUP) {
 		rq->scx.flags |= SCX_RQ_IN_WAKEUP;
+		p->scx.flags |= SCX_TASK_ENQ_WAKEUP;
+	} else {
+		p->scx.flags &= ~SCX_TASK_ENQ_WAKEUP;
+	}
 
 	/*
 	 * Restoring a running task will be immediately followed by
@@ -2276,11 +2315,24 @@ static void wakeup_preempt_scx(struct rq *rq, struct task_struct *p, int wake_fl
 {
 	/*
 	 * Preemption between SCX tasks is implemented by resetting the victim
-	 * task's slice to 0 and triggering reschedule on the target CPU.
-	 * Nothing to do.
+	 * task's slice to 0 and triggering reschedule on the target CPU. A
+	 * mutex-blocked task is kept queued for proxy execution, so its wakeup
+	 * doesn't go through enqueue_task_scx(). If the BPF scheduler manages
+	 * blocked donors, reschedule explicitly so that it can reconsider a
+	 * donor it declined to dispatch while blocked.
 	 */
-	if (p->sched_class == &ext_sched_class)
+	if (p->sched_class == &ext_sched_class) {
+		bool enq_wakeup = p->scx.flags & SCX_TASK_ENQ_WAKEUP;
+
+		p->scx.flags &= ~SCX_TASK_ENQ_WAKEUP;
+		if (!enq_wakeup && p->is_blocked) {
+			struct scx_sched *sch = scx_task_sched(p);
+
+			if (sch && (sch->ops.flags & SCX_OPS_ENQ_BLOCKED))
+				resched_curr(rq);
+		}
 		return;
+	}
 
 	/*
 	 * Getting preempted by a higher-priority class. Reenqueue IMMED tasks.
@@ -2395,6 +2447,20 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
 
 	WARN_ON_ONCE(task_cpu(p) == cpu);
 
+	/*
+	 * A blocked donor may be moved normally to select a new callback rq.
+	 * set_task_cpu() updates wake_cpu and makes the destination rq its new
+	 * callback home.
+	 *
+	 * proxy_set_task_cpu() instead preserves wake_cpu when moving a donor to
+	 * its lock owner's CPU. Keep such a donor on the proxy rq until it wakes;
+	 * otherwise normal BPF placement may repeatedly pull it back to its
+	 * callback rq only for proxy execution to move it to the owner again.
+	 */
+	if (sched_proxy_exec() && p->is_blocked &&
+	    task_cpu(p) != p->wake_cpu)
+		return false;
+
 	/*
 	 * If @p has migration disabled, @p->cpus_ptr is updated to contain only
 	 * the pinned CPU in migrate_disable_switch() while @p is being switched
@@ -2986,6 +3052,8 @@ static void set_next_task_scx(struct rq *rq, struct task_struct *p, bool first)
 	struct scx_sched *sch = scx_task_sched(p);
 	bool can_stop_tick;
 
+	p->scx.flags &= ~SCX_TASK_ENQ_WAKEUP;
+
 	if (p->scx.flags & SCX_TASK_QUEUED) {
 		/*
 		 * Core-sched might decide to execute @p before it is
@@ -3127,21 +3195,24 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
 		set_task_runnable(rq, p);
 
 		/*
-		 * Mutex-blocked donors stay queued on the runqueue under proxy
-		 * execution, but the donor never runs as itself, proxy-exec
-		 * walks the blocked_on chain on the next __schedule() and runs
-		 * the lock owner in its place.
+		 * The rq lock has remained held since scx_allow_proxy_exec(), so
+		 * @p's scheduler association cannot have changed. An associated
+		 * donor stays queued only when its BPF scheduler enables
+		 * %SCX_OPS_ENQ_BLOCKED; delegate its admission to that scheduler.
 		 *
-		 * Put the donor on the local DSQ directly so pick_next_task()
-		 * can still see it. find_proxy_task() will either run the chain
-		 * owner or deactivate the donor so the wakeup path can return it
-		 * and let BPF make a new dispatch decision once it is unblocked.
-		 *
-		 * This is preparatory code: a later patch will delegate blocked-donor
-		 * admission to the BPF scheduler.
+		 * If @sch is NULL, @p is transitioning into the root scheduler. The
+		 * root is published before tasks enter EXT and cannot be cleared while
+		 * this rq is locked. Preserve generic proxy execution by placing the
+		 * donor directly on the local DSQ.
 		 */
 		if (p->is_blocked) {
-			scx_dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, 0);
+			if (sch) {
+				WARN_ON_ONCE(!(sch->ops.flags & SCX_OPS_ENQ_BLOCKED));
+				scx_do_enqueue_task(rq, p, 0, -1);
+			} else {
+				scx_dispatch_enqueue(scx_root, rq, &rq->scx.local_dsq,
+						     p, 0);
+			}
 			goto switch_class;
 		}
 
@@ -7209,6 +7280,11 @@ int scx_validate_ops(struct scx_sched *sch, const struct sched_ext_ops *ops)
 		return -EINVAL;
 	}
 
+	if ((ops->flags & SCX_OPS_ENQ_BLOCKED) && !ops->enqueue) {
+		scx_error(sch, "SCX_OPS_ENQ_BLOCKED requires ops.enqueue() to be implemented");
+		return -EINVAL;
+	}
+
 	/*
 	 * SCX_OPS_TID_TO_TASK is enabled by the root scheduler. A sub-sched
 	 * may set it to declare a dependency; reject if the root hasn't
@@ -7594,7 +7670,7 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 		if (old_class != new_class)
 			queue_flags |= DEQUEUE_CLASS;
 		if (new_class == &ext_sched_class) {
-			scx_prepare_task_sched_change(p);
+			scx_prepare_task_sched_change(p, sch);
 			queue_flags |= DEQUEUE_NOCLOCK;
 		}
 
diff --git a/kernel/sched/ext/internal.h b/kernel/sched/ext/internal.h
index 5dfb7e466108f..47bf8e8c34590 100644
--- a/kernel/sched/ext/internal.h
+++ b/kernel/sched/ext/internal.h
@@ -213,6 +213,19 @@ enum scx_ops_flags {
 	 */
 	SCX_OPS_TID_TO_TASK		= 1LLU << 8,
 
+	/*
+	 * If set, mutex-blocked tasks remain runnable as proxy donors and are
+	 * passed to ops.enqueue() with %SCX_ENQ_BLOCKED. The BPF scheduler controls
+	 * when donors are dispatched and whether they should preempt other work.
+	 *
+	 * If clear, mutex-blocked tasks are removed from the runqueue normally
+	 * and cannot donate their scheduling context through proxy execution.
+	 *
+	 * For blocked donors, this flag takes precedence over
+	 * %SCX_OPS_ENQ_EXITING and %SCX_OPS_ENQ_MIGRATION_DISABLED.
+	 */
+	SCX_OPS_ENQ_BLOCKED		= 1LLU << 9,
+
 	SCX_OPS_ALL_FLAGS		= SCX_OPS_KEEP_BUILTIN_IDLE |
 					  SCX_OPS_ENQ_LAST |
 					  SCX_OPS_ENQ_EXITING |
@@ -221,7 +234,8 @@ enum scx_ops_flags {
 					  SCX_OPS_SWITCH_PARTIAL |
 					  SCX_OPS_BUILTIN_IDLE_PER_NODE |
 					  SCX_OPS_ALWAYS_ENQ_IMMED |
-					  SCX_OPS_TID_TO_TASK,
+					  SCX_OPS_TID_TO_TASK |
+					  SCX_OPS_ENQ_BLOCKED,
 
 	/* high 8 bits are internal, don't include in SCX_OPS_ALL_FLAGS */
 	__SCX_OPS_INTERNAL_MASK		= 0xffLLU << 56,
@@ -1650,6 +1664,12 @@ enum scx_enq_flags {
 	 */
 	SCX_ENQ_LAST		= 1LLU << 41,
 
+	/*
+	 * The task is blocked on a mutex and is being kept runnable as a proxy
+	 * donor. Only passed to ops.enqueue() when %SCX_OPS_ENQ_BLOCKED is set.
+	 */
+	SCX_ENQ_BLOCKED		= 1LLU << 42,
+
 	/* high 8 bits are internal */
 	__SCX_ENQ_INTERNAL_MASK	= 0xffLLU << 56,
 
@@ -1946,6 +1966,7 @@ void scx_task_iter_start(struct scx_task_iter *iter, struct cgroup *cgrp);
 void scx_task_iter_unlock(struct scx_task_iter *iter);
 void scx_task_iter_stop(struct scx_task_iter *iter);
 struct task_struct *scx_task_iter_next_locked(struct scx_task_iter *iter);
+void scx_prepare_task_sched_change(struct task_struct *p, struct scx_sched *sch);
 void scx_dispatch_dequeue(struct rq *rq, struct task_struct *p);
 void scx_do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags,
 			 int sticky_cpu);
diff --git a/kernel/sched/ext/sub.c b/kernel/sched/ext/sub.c
index 8d8737149bc07..e167bb7b40cf4 100644
--- a/kernel/sched/ext/sub.c
+++ b/kernel/sched/ext/sub.c
@@ -773,7 +773,9 @@ static void scx_rehome_task(struct scx_sched *to, struct task_struct *p)
 	lockdep_assert_held(&p->pi_lock);
 	lockdep_assert_rq_held(task_rq(p));
 
-	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
+	scx_prepare_task_sched_change(p, to);
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE |
+		      DEQUEUE_NOCLOCK) {
 		scx_disable_and_exit_task(scx_task_sched(p), p);
 		scx_set_task_state(p, SCX_TASK_INIT_BEGIN);
 		scx_set_task_state(p, SCX_TASK_INIT);
@@ -803,7 +805,9 @@ static void scx_punt_task(struct scx_sched *to, struct task_struct *p)
 	lockdep_assert_rq_held(task_rq(p));
 	WARN_ON_ONCE(!READ_ONCE(to->bypass_depth));
 
-	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE) {
+	scx_prepare_task_sched_change(p, to);
+	scoped_guard (sched_change, p, DEQUEUE_SAVE | DEQUEUE_MOVE |
+		      DEQUEUE_NOCLOCK) {
 		scx_disable_and_exit_task(scx_task_sched(p), p);
 		scx_set_task_sched(p, to);
 	}
diff --git a/tools/sched_ext/include/scx/compat.h b/tools/sched_ext/include/scx/compat.h
index 7757252d52e21..0e83ebfbee2b5 100644
--- a/tools/sched_ext/include/scx/compat.h
+++ b/tools/sched_ext/include/scx/compat.h
@@ -117,6 +117,7 @@ static inline bool __COMPAT_struct_has_field(const char *type, const char *field
 #define SCX_OPS_ALLOW_QUEUED_WAKEUP SCX_OPS_FLAG(SCX_OPS_ALLOW_QUEUED_WAKEUP)
 #define SCX_OPS_BUILTIN_IDLE_PER_NODE SCX_OPS_FLAG(SCX_OPS_BUILTIN_IDLE_PER_NODE)
 #define SCX_OPS_ALWAYS_ENQ_IMMED SCX_OPS_FLAG(SCX_OPS_ALWAYS_ENQ_IMMED)
+#define SCX_OPS_ENQ_BLOCKED SCX_OPS_FLAG(SCX_OPS_ENQ_BLOCKED)
 
 #define SCX_PICK_IDLE_FLAG(name) __COMPAT_ENUM_OR_ZERO("scx_pick_idle_cpu_flags", #name)
 
diff --git a/tools/sched_ext/include/scx/enum_defs.autogen.h b/tools/sched_ext/include/scx/enum_defs.autogen.h
index da4b459820fdd..79b31eb7db7cb 100644
--- a/tools/sched_ext/include/scx/enum_defs.autogen.h
+++ b/tools/sched_ext/include/scx/enum_defs.autogen.h
@@ -55,6 +55,7 @@
 #define HAVE_SCX_ENQ_IMMED
 #define HAVE_SCX_ENQ_REENQ
 #define HAVE_SCX_ENQ_LAST
+#define HAVE_SCX_ENQ_BLOCKED
 #define HAVE___SCX_ENQ_INTERNAL_MASK
 #define HAVE_SCX_ENQ_CLEAR_OPSS
 #define HAVE_SCX_ENQ_DSQ_PRIQ
diff --git a/tools/sched_ext/include/scx/enums.autogen.bpf.h b/tools/sched_ext/include/scx/enums.autogen.bpf.h
index dafccbb6b69d2..7efe7b9346b49 100644
--- a/tools/sched_ext/include/scx/enums.autogen.bpf.h
+++ b/tools/sched_ext/include/scx/enums.autogen.bpf.h
@@ -130,6 +130,9 @@ const volatile u64 __SCX_ENQ_REENQ __weak;
 const volatile u64 __SCX_ENQ_LAST __weak;
 #define SCX_ENQ_LAST __SCX_ENQ_LAST
 
+const volatile u64 __SCX_ENQ_BLOCKED __weak;
+#define SCX_ENQ_BLOCKED __SCX_ENQ_BLOCKED
+
 const volatile u64 __SCX_ENQ_CLEAR_OPSS __weak;
 #define SCX_ENQ_CLEAR_OPSS __SCX_ENQ_CLEAR_OPSS
 
diff --git a/tools/sched_ext/include/scx/enums.autogen.h b/tools/sched_ext/include/scx/enums.autogen.h
index bbd4901f4fce3..f8fbeb7fbf95b 100644
--- a/tools/sched_ext/include/scx/enums.autogen.h
+++ b/tools/sched_ext/include/scx/enums.autogen.h
@@ -47,6 +47,7 @@
 	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_IMMED); \
 	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_REENQ); \
 	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_LAST); \
+	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_BLOCKED); \
 	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_CLEAR_OPSS); \
 	SCX_ENUM_SET(skel, scx_enq_flags, SCX_ENQ_DSQ_PRIQ); \
 	SCX_ENUM_SET(skel, scx_deq_flags, SCX_DEQ_SCHED_CHANGE); \
-- 
2.55.0

[PATCH 10/12] sched_ext: Add selftest for blocked donor admission

[Andrea Righi]

SCX_OPS_ENQ_BLOCKED allows BPF schedulers to receive blocked proxy
donors through ops.enqueue(). SCX_ENQ_BLOCKED identifies blocked-donor
admission requests. Add selftest coverage for this interface.

Exercise a priority inversion using a weighted-vruntime BPF scheduler.
A nice +19 owner holds a shared mutex, a nice -20 donor blocks on it,
and nice 0 CPU contenders, one per allowed CPU, keep the system busy.
Test both a same-CPU topology and a cross-CPU topology with the donor
and owner on different CPUs.

Treat blocked donors according to the normal BPF ordering policy and
assign the default slice on every enqueue, as for other tasks. Run each
CPU placement configuration with SCX_OPS_ENQ_BLOCKED first disabled and
then enabled, count blocked-donor enqueues by CPU and report average
mutex hold and wait times. Verify that full wakeups are never reported
as blocked-donor admissions by checking that SCX_ENQ_WAKEUP and
SCX_ENQ_BLOCKED are not set together.

Proxy execution coverage requires CONFIG_SCHED_PROXY_EXEC=y, which the
selftest config selects. Access to the kernel mutex is provided via a
loadable kernel module, built through TEST_GEN_MODS_DIR and managed by
the test.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 tools/testing/selftests/sched_ext/.gitignore  |   4 +
 tools/testing/selftests/sched_ext/Makefile    |   2 +
 tools/testing/selftests/sched_ext/config      |   2 +
 .../selftests/sched_ext/enq_blocked.bpf.c     | 116 +++
 .../testing/selftests/sched_ext/enq_blocked.c | 918 ++++++++++++++++++
 .../testing/selftests/sched_ext/enq_blocked.h |  28 +
 .../selftests/sched_ext/test_modules/Makefile |  13 +
 .../test_modules/scx_enq_blocked_test.c       | 195 ++++
 8 files changed, 1278 insertions(+)
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.bpf.c
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.c
 create mode 100644 tools/testing/selftests/sched_ext/enq_blocked.h
 create mode 100644 tools/testing/selftests/sched_ext/test_modules/Makefile
 create mode 100644 tools/testing/selftests/sched_ext/test_modules/scx_enq_blocked_test.c

diff --git a/tools/testing/selftests/sched_ext/.gitignore b/tools/testing/selftests/sched_ext/.gitignore
index ae5491a114c09..54a1fd2af713d 100644
--- a/tools/testing/selftests/sched_ext/.gitignore
+++ b/tools/testing/selftests/sched_ext/.gitignore
@@ -4,3 +4,7 @@
 !Makefile
 !.gitignore
 !config
+!test_modules/
+!test_modules/scx_enq_blocked_test.c
+!test_modules/Makefile
+test_modules/*.mod.c
diff --git a/tools/testing/selftests/sched_ext/Makefile b/tools/testing/selftests/sched_ext/Makefile
index 3cfe90e0f34fa..51a16b3d32d9b 100644
--- a/tools/testing/selftests/sched_ext/Makefile
+++ b/tools/testing/selftests/sched_ext/Makefile
@@ -5,6 +5,7 @@ include ../../../scripts/Makefile.arch
 include ../../../scripts/Makefile.include
 
 TEST_GEN_PROGS := runner
+TEST_GEN_MODS_DIR := test_modules
 
 # override lib.mk's default rules
 OVERRIDE_TARGETS := 1
@@ -164,6 +165,7 @@ all_test_bpfprogs := $(foreach prog,$(wildcard *.bpf.c),$(INCLUDE_DIR)/$(patsubs
 auto-test-targets :=			\
 	create_dsq			\
 	dequeue				\
+	enq_blocked			\
 	enq_last_no_enq_fails		\
 	ddsp_bogus_dsq_fail		\
 	ddsp_vtimelocal_fail		\
diff --git a/tools/testing/selftests/sched_ext/config b/tools/testing/selftests/sched_ext/config
index aa901b05c8ad6..affa3cf33470a 100644
--- a/tools/testing/selftests/sched_ext/config
+++ b/tools/testing/selftests/sched_ext/config
@@ -6,3 +6,5 @@ CONFIG_BPF=y
 CONFIG_BPF_SYSCALL=y
 CONFIG_DEBUG_INFO=y
 CONFIG_DEBUG_INFO_BTF=y
+CONFIG_EXPERT=y
+CONFIG_SCHED_PROXY_EXEC=y
diff --git a/tools/testing/selftests/sched_ext/enq_blocked.bpf.c b/tools/testing/selftests/sched_ext/enq_blocked.bpf.c
new file mode 100644
index 0000000000000..212690bf4e07b
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/enq_blocked.bpf.c
@@ -0,0 +1,116 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+ *
+ * Verify that SCX_OPS_ENQ_BLOCKED passes blocked proxy donors through
+ * ops.enqueue() and record whether callbacks occur on the donor or owner CPU.
+ */
+
+#include <scx/common.bpf.h>
+
+#define SHARED_DSQ 0
+
+char _license[] SEC("license") = "GPL";
+
+s32 donor_pid;
+s32 donor_cpu = -1;
+s32 owner_cpu = -1;
+u64 nr_blocked_enqueues;
+u64 nr_blocked_enqueues_donor_cpu;
+u64 nr_blocked_enqueues_owner_cpu;
+u64 nr_blocked_enqueues_other_cpu;
+u64 nr_blocked_wakeups;
+static u64 vtime_now;
+
+UEI_DEFINE(uei);
+
+s32 BPF_STRUCT_OPS(enq_blocked_select_cpu,
+		   struct task_struct *p, s32 prev_cpu, u64 wake_flags)
+{
+	return prev_cpu;
+}
+
+void BPF_STRUCT_OPS(enq_blocked_enqueue, struct task_struct *p, u64 enq_flags)
+{
+	u64 vtime = p->scx.dsq_vtime;
+
+	if (enq_flags & SCX_ENQ_BLOCKED) {
+		int cpu = scx_bpf_task_cpu(p);
+
+		if (enq_flags & SCX_ENQ_WAKEUP)
+			__sync_fetch_and_add(&nr_blocked_wakeups, 1);
+
+		if (p->pid == donor_pid) {
+			__sync_fetch_and_add(&nr_blocked_enqueues, 1);
+			if (cpu == donor_cpu)
+				__sync_fetch_and_add(&nr_blocked_enqueues_donor_cpu, 1);
+			else if (cpu == owner_cpu)
+				__sync_fetch_and_add(&nr_blocked_enqueues_owner_cpu, 1);
+			else
+				__sync_fetch_and_add(&nr_blocked_enqueues_other_cpu, 1);
+		}
+	}
+
+	/* Limit the amount of budget an idling task can accumulate. */
+	if (time_before(vtime, vtime_now - SCX_SLICE_DFL))
+		vtime = vtime_now - SCX_SLICE_DFL;
+
+	scx_bpf_dsq_insert_vtime(p, SHARED_DSQ, SCX_SLICE_DFL, vtime,
+				 enq_flags);
+	scx_bpf_kick_cpu(scx_bpf_task_cpu(p), SCX_KICK_IDLE);
+}
+
+void BPF_STRUCT_OPS(enq_blocked_dispatch, s32 cpu, struct task_struct *prev)
+{
+	scx_bpf_dsq_move_to_local(SHARED_DSQ, 0);
+}
+
+void BPF_STRUCT_OPS(enq_blocked_running, struct task_struct *p)
+{
+	if (time_before(vtime_now, p->scx.dsq_vtime))
+		vtime_now = p->scx.dsq_vtime;
+}
+
+void BPF_STRUCT_OPS(enq_blocked_stopping, struct task_struct *p, bool runnable)
+{
+	u64 delta = scale_by_task_weight_inverse(p,
+					 SCX_SLICE_DFL - p->scx.slice);
+
+	scx_bpf_task_set_dsq_vtime(p, p->scx.dsq_vtime + delta);
+}
+
+void BPF_STRUCT_OPS(enq_blocked_enable, struct task_struct *p)
+{
+	scx_bpf_task_set_dsq_vtime(p, vtime_now);
+}
+
+s32 BPF_STRUCT_OPS_SLEEPABLE(enq_blocked_init)
+{
+	int ret;
+
+	ret = scx_bpf_create_dsq(SHARED_DSQ, -1);
+	if (ret) {
+		scx_bpf_error("failed to create DSQ %d (%d)", SHARED_DSQ, ret);
+		return ret;
+	}
+
+	return 0;
+}
+
+void BPF_STRUCT_OPS(enq_blocked_exit, struct scx_exit_info *ei)
+{
+	UEI_RECORD(uei, ei);
+}
+
+SEC(".struct_ops.link")
+struct sched_ext_ops enq_blocked_ops = {
+	.select_cpu		= (void *)enq_blocked_select_cpu,
+	.enqueue		= (void *)enq_blocked_enqueue,
+	.dispatch		= (void *)enq_blocked_dispatch,
+	.running		= (void *)enq_blocked_running,
+	.stopping		= (void *)enq_blocked_stopping,
+	.enable			= (void *)enq_blocked_enable,
+	.init			= (void *)enq_blocked_init,
+	.exit			= (void *)enq_blocked_exit,
+	.name			= "enq_blocked",
+};
diff --git a/tools/testing/selftests/sched_ext/enq_blocked.c b/tools/testing/selftests/sched_ext/enq_blocked.c
new file mode 100644
index 0000000000000..896c66820a744
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/enq_blocked.c
@@ -0,0 +1,918 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+ *
+ * Exercise a priority inversion with the owner and donor first pinned to the
+ * same CPU, then with each on a different CPU. A high-priority donor blocks on
+ * a mutex held by a low-priority owner while one medium-priority contender per
+ * available CPU keeps the system busy. A weighted-vruntime BPF scheduler runs
+ * both CPU placement configurations with SCX_OPS_ENQ_BLOCKED first disabled
+ * and then enabled. The test validates blocked-donor admission and reports the
+ * average mutex hold and wait times, plus their enabled-minus-disabled deltas,
+ * for each configuration. The timing data is informational.
+ *
+ * CONFIG_SCHED_PROXY_EXEC=y is required to exercise the proxy-execution paths.
+ */
+#define _GNU_SOURCE
+
+#include <bpf/bpf.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <limits.h>
+#include <pthread.h>
+#include <sched.h>
+#include <scx/common.h>
+#include <stdatomic.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/resource.h>
+#include <sys/syscall.h>
+#include <time.h>
+#include <unistd.h>
+
+#include "enq_blocked.bpf.skel.h"
+#include "enq_blocked.h"
+#include "scx_test.h"
+
+#define MODULE_NAME	"scx_enq_blocked_test"
+#define MODULE_FILE	"test_modules/" MODULE_NAME ".ko"
+#define DEVICE_PATH	"/dev/scx_enq_blocked"
+#define WAIT_STEP_US	1000
+#define WAIT_TIMEOUT_MS	2000
+#define NR_WARMUP_TRIALS	1
+#define NR_MEASURED_TRIALS	10
+#define NR_TRIALS	(NR_WARMUP_TRIALS + NR_MEASURED_TRIALS)
+#define JOIN_TIMEOUT_MS	((NR_TRIALS + 1) * WAIT_TIMEOUT_MS)
+#define OWNER_NICE	19
+#define DONOR_NICE	-20
+#define CONTENDER_NICE	0
+
+struct thread_ctx {
+	atomic_bool start_donor;
+	atomic_bool abort;
+	atomic_bool stop_contender;
+	atomic_bool measurement_ready;
+	atomic_int donor_pid;
+	atomic_int donor_completed;
+	int fd;
+	int donor_cpu;
+	int owner_cpu;
+};
+
+struct contender_ctx {
+	struct thread_ctx *thread_ctx;
+	atomic_int status;
+	int cpu;
+};
+
+struct run_result {
+	struct enq_blocked_stats stats;
+	u64 nr_blocked_enqueues;
+	u64 nr_blocked_enqueues_donor_cpu;
+	u64 nr_blocked_enqueues_owner_cpu;
+	u64 nr_blocked_enqueues_other_cpu;
+	u64 nr_blocked_wakeups;
+};
+
+static bool parse_bool(const char *value, bool *result)
+{
+	if (!strcasecmp(value, "1") || !strcasecmp(value, "y") ||
+	    !strcasecmp(value, "yes") || !strcasecmp(value, "on") ||
+	    !strcasecmp(value, "true")) {
+		*result = true;
+		return true;
+	}
+
+	if (!strcasecmp(value, "0") || !strcasecmp(value, "n") ||
+	    !strcasecmp(value, "no") || !strcasecmp(value, "off") ||
+	    !strcasecmp(value, "false")) {
+		*result = false;
+		return true;
+	}
+
+	return false;
+}
+
+static bool cmdline_bool(const char *name, bool default_value)
+{
+	char cmdline[4096], *newline, *saveptr = NULL, *token;
+	size_t name_len = strlen(name);
+	bool value = default_value;
+	FILE *file;
+
+	file = fopen("/proc/cmdline", "r");
+	if (!file)
+		return default_value;
+
+	if (!fgets(cmdline, sizeof(cmdline), file)) {
+		fclose(file);
+		return default_value;
+	}
+	fclose(file);
+	newline = strchr(cmdline, '\n');
+	if (newline)
+		*newline = '\0';
+
+	for (token = strtok_r(cmdline, " ", &saveptr); token;
+	     token = strtok_r(NULL, " ", &saveptr)) {
+		bool parsed;
+
+		if (strncmp(token, name, name_len) || token[name_len] != '=')
+			continue;
+		if (parse_bool(token + name_len + 1, &parsed))
+			value = parsed;
+	}
+
+	return value;
+}
+
+static int module_path(char *path, size_t size)
+{
+	ssize_t len;
+	char *slash;
+
+	len = readlink("/proc/self/exe", path, size - 1);
+	if (len < 0)
+		return -errno;
+	path[len] = '\0';
+
+	slash = strrchr(path, '/');
+	if (!slash)
+		return -EINVAL;
+	*slash = '\0';
+
+	if (snprintf(slash, size - (slash - path), "/%s", MODULE_FILE) >=
+	    size - (slash - path))
+		return -ENAMETOOLONG;
+
+	return 0;
+}
+
+static int load_test_module(bool *loaded_here)
+{
+	char path[PATH_MAX];
+	int fd, err;
+
+	err = module_path(path, sizeof(path));
+	if (err)
+		return err;
+
+	fd = open(path, O_RDONLY | O_CLOEXEC);
+	if (fd < 0)
+		return -errno;
+
+	if (syscall(SYS_finit_module, fd, "", 0)) {
+		err = errno;
+		close(fd);
+		if (err == EEXIST)
+			return 0;
+		return -err;
+	}
+
+	close(fd);
+	*loaded_here = true;
+	return 0;
+}
+
+static void unload_test_module(bool loaded_here)
+{
+	if (loaded_here && syscall(SYS_delete_module, MODULE_NAME, O_NONBLOCK))
+		SCX_ERR("Failed to unload %s (%d)", MODULE_NAME, errno);
+}
+
+static int pin_to_cpu(int cpu)
+{
+	cpu_set_t mask;
+
+	CPU_ZERO(&mask);
+	CPU_SET(cpu, &mask);
+	return sched_setaffinity(0, sizeof(mask), &mask) ? errno : 0;
+}
+
+static int select_test_cpus(bool cross_cpu, cpu_set_t *mask, int *donor_cpu,
+			    int *owner_cpu)
+{
+	int cpu, first = -1;
+
+	if (sched_getaffinity(0, sizeof(*mask), mask))
+		return -errno;
+
+	for (cpu = 0; cpu < CPU_SETSIZE; cpu++) {
+		if (!CPU_ISSET(cpu, mask))
+			continue;
+		if (first < 0) {
+			first = cpu;
+			if (!cross_cpu)
+				break;
+		} else {
+			*donor_cpu = first;
+			*owner_cpu = cpu;
+			return 0;
+		}
+	}
+
+	if (first < 0)
+		return -ENODEV;
+	if (cross_cpu)
+		return -EAGAIN;
+
+	*donor_cpu = first;
+	*owner_cpu = first;
+	return 0;
+}
+
+static int set_nice(int nice)
+{
+	return setpriority(PRIO_PROCESS, 0, nice) ? errno : 0;
+}
+
+static bool wait_for_pid(atomic_int *pid)
+{
+	int waited_ms;
+
+	for (waited_ms = 0; waited_ms < WAIT_TIMEOUT_MS; waited_ms++) {
+		if (atomic_load_explicit(pid, memory_order_acquire) > 0)
+			return true;
+		usleep(WAIT_STEP_US);
+	}
+
+	return false;
+}
+
+static int wait_for_contenders(struct contender_ctx *contenders,
+			       size_t nr_contenders)
+{
+	size_t i, nr_ready;
+	int status, waited_ms;
+
+	for (waited_ms = 0; waited_ms < WAIT_TIMEOUT_MS; waited_ms++) {
+		nr_ready = 0;
+		for (i = 0; i < nr_contenders; i++) {
+			status = atomic_load_explicit(&contenders[i].status,
+						      memory_order_acquire);
+			if (status < 0)
+				return status;
+			if (status > 0)
+				nr_ready++;
+		}
+		if (nr_ready == nr_contenders)
+			return 1;
+		usleep(WAIT_STEP_US);
+	}
+
+	return -ETIMEDOUT;
+}
+
+static int wait_for_donor_state(struct thread_ctx *ctx, int expected)
+{
+	int state, waited_ms;
+
+	for (waited_ms = 0; waited_ms < WAIT_TIMEOUT_MS; waited_ms++) {
+		state = ioctl(ctx->fd, ENQ_BLOCKED_IOCTL_DONOR_STATE);
+		if (state == expected)
+			return state;
+		if (state < 0 && errno != ENOENT)
+			return -errno;
+		usleep(WAIT_STEP_US);
+	}
+
+	return -ETIMEDOUT;
+}
+
+static bool wait_for_donor(struct thread_ctx *ctx, int trial)
+{
+	int waited_ms;
+
+	for (waited_ms = 0; waited_ms < WAIT_TIMEOUT_MS; waited_ms++) {
+		if (atomic_load_explicit(&ctx->donor_completed,
+					 memory_order_acquire) >= trial)
+			return true;
+		if (atomic_load_explicit(&ctx->abort, memory_order_relaxed))
+			return false;
+		usleep(WAIT_STEP_US);
+	}
+
+	return false;
+}
+
+static bool wait_for_measurement(struct thread_ctx *ctx)
+{
+	while (!atomic_load_explicit(&ctx->measurement_ready,
+				     memory_order_acquire) &&
+	       !atomic_load_explicit(&ctx->abort, memory_order_relaxed))
+		sched_yield();
+
+	return !atomic_load_explicit(&ctx->abort, memory_order_relaxed);
+}
+
+static void *contender_fn(void *arg)
+{
+	struct contender_ctx *contender = arg;
+	struct thread_ctx *ctx = contender->thread_ctx;
+	int err;
+
+	err = pin_to_cpu(contender->cpu);
+	if (!err)
+		err = set_nice(CONTENDER_NICE);
+	atomic_store_explicit(&contender->status, err ? -err : 1,
+			      memory_order_release);
+	if (err)
+		return (void *)(uintptr_t)err;
+
+	while (!atomic_load_explicit(&ctx->stop_contender,
+				     memory_order_relaxed))
+		;
+
+	return NULL;
+}
+
+static void *owner_fn(void *arg)
+{
+	struct thread_ctx *ctx = arg;
+	int err, i;
+
+	err = pin_to_cpu(ctx->owner_cpu);
+	if (err)
+		return (void *)(uintptr_t)err;
+	err = set_nice(OWNER_NICE);
+	if (err)
+		return (void *)(uintptr_t)err;
+
+	for (i = 0; i < NR_TRIALS; i++) {
+		if (ioctl(ctx->fd, ENQ_BLOCKED_IOCTL_OWNER))
+			return (void *)(uintptr_t)errno;
+		if (!wait_for_donor(ctx, i + 1))
+			return (void *)(uintptr_t)ETIMEDOUT;
+
+		if (i + 1 == NR_WARMUP_TRIALS && !wait_for_measurement(ctx))
+			return NULL;
+	}
+
+	return NULL;
+}
+
+static int run_donor_trial(struct thread_ctx *ctx)
+{
+	int waited_ms;
+
+	for (waited_ms = 0; waited_ms < WAIT_TIMEOUT_MS; waited_ms++) {
+		if (!ioctl(ctx->fd, ENQ_BLOCKED_IOCTL_DONOR))
+			return 0;
+		if (errno != EAGAIN)
+			return -errno;
+		usleep(WAIT_STEP_US);
+	}
+
+	return -ETIMEDOUT;
+}
+
+static void *donor_fn(void *arg)
+{
+	struct thread_ctx *ctx = arg;
+	int err, i;
+
+	err = pin_to_cpu(ctx->donor_cpu);
+	if (err)
+		return (void *)(uintptr_t)err;
+	err = set_nice(DONOR_NICE);
+	if (err)
+		return (void *)(uintptr_t)err;
+
+	atomic_store_explicit(&ctx->donor_pid, syscall(SYS_gettid),
+			      memory_order_release);
+	while (!atomic_load_explicit(&ctx->start_donor, memory_order_acquire) &&
+	       !atomic_load_explicit(&ctx->abort, memory_order_relaxed))
+		sched_yield();
+
+	if (atomic_load_explicit(&ctx->abort, memory_order_relaxed))
+		return NULL;
+
+	for (i = 0; i < NR_TRIALS; i++) {
+		err = run_donor_trial(ctx);
+		if (err)
+			return (void *)(uintptr_t)-err;
+		atomic_store_explicit(&ctx->donor_completed, i + 1,
+				      memory_order_release);
+	}
+
+	return NULL;
+}
+
+static void print_avg_time(const char *name, u64 total_ns, u64 samples)
+{
+	u64 avg_ns = samples ? total_ns / samples : 0;
+
+	printf("  %s_avg_ns=%llu (%llu.%03llu ms, samples=%llu)\n", name,
+	       (unsigned long long)avg_ns,
+	       (unsigned long long)(avg_ns / 1000000),
+	       (unsigned long long)((avg_ns / 1000) % 1000),
+	       (unsigned long long)samples);
+}
+
+static void print_avg_delta(const char *name, u64 disabled_total,
+			    u64 disabled_samples, u64 enabled_total,
+			    u64 enabled_samples)
+{
+	u64 disabled_avg, enabled_avg;
+	s64 delta_ns;
+	double delta_pct;
+
+	if (!disabled_samples || !enabled_samples)
+		return;
+
+	disabled_avg = disabled_total / disabled_samples;
+	enabled_avg = enabled_total / enabled_samples;
+	delta_ns = (s64)enabled_avg - (s64)disabled_avg;
+	delta_pct = disabled_avg ? 100.0 * delta_ns / disabled_avg : 0.0;
+
+	printf("  %s_delta_ns=%+lld (%+.2f%%)\n", name,
+	       (long long)delta_ns, delta_pct);
+}
+
+static int join_thread(pthread_t thread, const struct timespec *deadline,
+		       int *thread_err)
+{
+	void *result;
+	int err;
+
+	err = pthread_timedjoin_np(thread, &result, deadline);
+	if (err)
+		return err;
+
+	*thread_err = (int)(uintptr_t)result;
+	return 0;
+}
+
+static void set_join_deadline(struct timespec *deadline)
+{
+	clock_gettime(CLOCK_REALTIME, deadline);
+	deadline->tv_sec += JOIN_TIMEOUT_MS / 1000;
+	deadline->tv_nsec += (JOIN_TIMEOUT_MS % 1000) * 1000000;
+	if (deadline->tv_nsec >= 1000000000) {
+		deadline->tv_sec++;
+		deadline->tv_nsec -= 1000000000;
+	}
+}
+
+static enum scx_test_status setup(void **ctx)
+{
+	struct enq_blocked *skel;
+	u64 flag;
+
+	skel = enq_blocked__open();
+	SCX_FAIL_IF(!skel, "Failed to open skel");
+	SCX_ENUM_INIT(skel);
+
+	flag = SCX_OPS_ENQ_BLOCKED;
+	if (!flag) {
+		enq_blocked__destroy(skel);
+		fprintf(stderr, "SKIP: SCX_OPS_ENQ_BLOCKED is unavailable\n");
+		return SCX_TEST_SKIP;
+	}
+
+	enq_blocked__destroy(skel);
+	*ctx = NULL;
+	return SCX_TEST_PASS;
+}
+
+static enum scx_test_status run_one(bool enq_blocked, bool cross_cpu,
+				    struct run_result *result)
+{
+	struct enq_blocked *skel;
+	struct thread_ctx thread_ctx = {};
+	struct contender_ctx *contender_ctxs = NULL;
+	struct bpf_link *link = NULL;
+	pthread_t owner, donor, *contenders = NULL;
+	struct timespec join_deadline;
+	cpu_set_t allowed_cpus;
+	bool module_loaded = false;
+	bool owner_started = false, donor_started = false;
+	bool join_timed_out = false;
+	bool proxy_enabled;
+	enum scx_test_status status = SCX_TEST_PASS;
+	int cpu, donor_pid, donor_state, err, thread_err;
+	size_t i, nr_contenders, nr_contenders_started = 0;
+	size_t nr_contenders_joined = 0;
+	u64 nr_blocked, nr_blocked_donor_cpu, nr_blocked_owner_cpu;
+	u64 nr_blocked_other_cpu, nr_blocked_wakeups;
+	struct enq_blocked_stats stats;
+
+	err = select_test_cpus(cross_cpu, &allowed_cpus, &thread_ctx.donor_cpu,
+			       &thread_ctx.owner_cpu);
+	if (err == -EAGAIN) {
+		fprintf(stderr, "SKIP: cross-CPU case requires two allowed CPUs\n");
+		return SCX_TEST_SKIP;
+	}
+	if (err) {
+		SCX_ERR("Failed to select test CPUs (%d)", -err);
+		return SCX_TEST_FAIL;
+	}
+	nr_contenders = CPU_COUNT(&allowed_cpus);
+	contenders = calloc(nr_contenders, sizeof(*contenders));
+	contender_ctxs = calloc(nr_contenders, sizeof(*contender_ctxs));
+	if (!contenders || !contender_ctxs) {
+		SCX_ERR("Failed to allocate %zu contender threads", nr_contenders);
+		status = SCX_TEST_FAIL;
+		goto out_contenders;
+	}
+
+	skel = enq_blocked__open();
+	if (!skel) {
+		SCX_ERR("Failed to open skel");
+		status = SCX_TEST_FAIL;
+		goto out_contenders;
+	}
+	SCX_ENUM_INIT(skel);
+	skel->struct_ops.enq_blocked_ops->flags =
+		SCX_OPS_ENQ_LAST |
+		(enq_blocked ? SCX_OPS_ENQ_BLOCKED : 0);
+	if (enq_blocked__load(skel)) {
+		SCX_ERR("Failed to load skel");
+		status = SCX_TEST_FAIL;
+		goto out_skel;
+	}
+
+	err = load_test_module(&module_loaded);
+	if (err == -EPERM || err == -ENOENT) {
+		fprintf(stderr, "SKIP: cannot load mutex fixture (%d)\n", -err);
+		status = SCX_TEST_SKIP;
+		goto out_skel;
+	}
+	if (err) {
+		SCX_ERR("Failed to load mutex fixture (%d)", -err);
+		status = SCX_TEST_FAIL;
+		goto out_skel;
+	}
+
+	thread_ctx.fd = open(DEVICE_PATH, O_RDONLY | O_CLOEXEC);
+	if (thread_ctx.fd < 0) {
+		SCX_ERR("Failed to open %s (%d)", DEVICE_PATH, errno);
+		status = SCX_TEST_FAIL;
+		goto out_module;
+	}
+	err = ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_PROXY_SUPPORTED);
+	if (err < 0) {
+		SCX_ERR("Failed to query proxy-exec support (%d)", errno);
+		status = SCX_TEST_FAIL;
+		goto out_fd;
+	}
+	proxy_enabled = err && cmdline_bool("sched_proxy_exec", true);
+	if (!proxy_enabled) {
+		fprintf(stderr, "SKIP: proxy execution is not enabled\n");
+		status = SCX_TEST_SKIP;
+		goto out_fd;
+	}
+	if (ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_RESET_STATS)) {
+		SCX_ERR("Failed to reset mutex statistics (%d)", errno);
+		status = SCX_TEST_FAIL;
+		goto out_fd;
+	}
+
+	if (ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_PREP_ATTACH)) {
+		SCX_ERR("Failed to prepare scheduler attachment (%d)", errno);
+		status = SCX_TEST_FAIL;
+		goto out_fd;
+	}
+
+	err = pthread_create(&owner, NULL, owner_fn, &thread_ctx);
+	if (err) {
+		SCX_ERR("Failed to create owner thread (%d)", err);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+	owner_started = true;
+
+	err = pthread_create(&donor, NULL, donor_fn, &thread_ctx);
+	if (err) {
+		SCX_ERR("Failed to create donor thread (%d)", err);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+	donor_started = true;
+
+	if (!wait_for_pid(&thread_ctx.donor_pid)) {
+		SCX_ERR("Timed out waiting for donor thread");
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	donor_pid = atomic_load_explicit(&thread_ctx.donor_pid,
+					 memory_order_acquire);
+	skel->bss->donor_pid = donor_pid;
+	skel->data->donor_cpu = thread_ctx.donor_cpu;
+	skel->data->owner_cpu = thread_ctx.owner_cpu;
+	atomic_store_explicit(&thread_ctx.start_donor, true,
+			      memory_order_release);
+
+	donor_state = ENQ_BLOCKED_DONOR_SLEEPING;
+	if (proxy_enabled)
+		donor_state |= ENQ_BLOCKED_DONOR_ON_RQ;
+	err = wait_for_donor_state(&thread_ctx, donor_state);
+	if (err < 0) {
+		SCX_ERR("Donor did not block before scheduler attachment (%d)", -err);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	link = bpf_map__attach_struct_ops(skel->maps.enq_blocked_ops);
+	if (!link) {
+		SCX_ERR("Failed to attach scheduler");
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	donor_state = ENQ_BLOCKED_DONOR_SLEEPING;
+	if (proxy_enabled && enq_blocked)
+		donor_state |= ENQ_BLOCKED_DONOR_ON_RQ;
+	err = wait_for_donor_state(&thread_ctx, donor_state);
+	if (err < 0) {
+		SCX_ERR("Unexpected donor state after scheduler attachment (%d)",
+			-err);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	if (ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_ATTACH_DONE)) {
+		SCX_ERR("Failed to complete scheduler attachment (%d)", errno);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	i = 0;
+	for (cpu = 0; cpu < CPU_SETSIZE; cpu++) {
+		if (!CPU_ISSET(cpu, &allowed_cpus))
+			continue;
+
+		contender_ctxs[i].thread_ctx = &thread_ctx;
+		contender_ctxs[i].cpu = cpu;
+		atomic_init(&contender_ctxs[i].status, 0);
+		err = pthread_create(&contenders[i], NULL, contender_fn,
+				     &contender_ctxs[i]);
+		if (err) {
+			SCX_ERR("Failed to create contender for CPU %d (%d)",
+				cpu, err);
+			status = SCX_TEST_FAIL;
+			goto out;
+		}
+		nr_contenders_started++;
+		i++;
+	}
+
+	err = wait_for_contenders(contender_ctxs, nr_contenders);
+	if (err != 1) {
+		SCX_ERR("Contender threads failed (%d)", -err);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+
+	/*
+	 * The first trial spans scheduler attachment and validates the state
+	 * transition, but including it would skew scheduling latency. Exclude it
+	 * from both the mutex and BPF enqueue measurements.
+	 */
+	if (!wait_for_donor(&thread_ctx, NR_WARMUP_TRIALS)) {
+		SCX_ERR("Timed out waiting for warm-up trial");
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+	if (ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_RESET_STATS)) {
+		SCX_ERR("Failed to reset mutex statistics after warm-up (%d)",
+			errno);
+		status = SCX_TEST_FAIL;
+		goto out;
+	}
+	skel->bss->nr_blocked_enqueues = 0;
+	skel->bss->nr_blocked_enqueues_donor_cpu = 0;
+	skel->bss->nr_blocked_enqueues_owner_cpu = 0;
+	skel->bss->nr_blocked_enqueues_other_cpu = 0;
+	skel->bss->nr_blocked_wakeups = 0;
+	atomic_store_explicit(&thread_ctx.measurement_ready, true,
+			      memory_order_release);
+
+out:
+	ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_ATTACH_DONE);
+	if (status != SCX_TEST_PASS) {
+		atomic_store_explicit(&thread_ctx.abort, true, memory_order_release);
+		atomic_store_explicit(&thread_ctx.start_donor, true,
+				      memory_order_release);
+		atomic_store_explicit(&thread_ctx.measurement_ready, true,
+				      memory_order_release);
+	}
+
+	set_join_deadline(&join_deadline);
+	if (donor_started) {
+		err = join_thread(donor, &join_deadline, &thread_err);
+		if (err == ETIMEDOUT) {
+			SCX_ERR("Timed out waiting for donor thread");
+			join_timed_out = true;
+			status = SCX_TEST_FAIL;
+		} else if (err) {
+			SCX_ERR("Failed to join donor thread (%d)", err);
+			status = SCX_TEST_FAIL;
+		} else {
+			donor_started = false;
+			if (thread_err) {
+				SCX_ERR("Donor thread failed (%d)", thread_err);
+				status = SCX_TEST_FAIL;
+			}
+		}
+	}
+	if (!join_timed_out && owner_started) {
+		err = join_thread(owner, &join_deadline, &thread_err);
+		if (err == ETIMEDOUT) {
+			SCX_ERR("Timed out waiting for owner thread");
+			join_timed_out = true;
+			status = SCX_TEST_FAIL;
+		} else if (err) {
+			SCX_ERR("Failed to join owner thread (%d)", err);
+			status = SCX_TEST_FAIL;
+		} else {
+			owner_started = false;
+			if (thread_err) {
+				SCX_ERR("Owner thread failed (%d)", thread_err);
+				status = SCX_TEST_FAIL;
+			}
+		}
+	}
+	atomic_store_explicit(&thread_ctx.stop_contender, true,
+			      memory_order_release);
+	for (i = 0; !join_timed_out && i < nr_contenders_started; i++) {
+		err = join_thread(contenders[i], &join_deadline, &thread_err);
+		if (err == ETIMEDOUT) {
+			SCX_ERR("Timed out waiting for contender on CPU %d",
+				contender_ctxs[i].cpu);
+			join_timed_out = true;
+			status = SCX_TEST_FAIL;
+		} else if (err) {
+			SCX_ERR("Failed to join contender on CPU %d (%d)",
+				contender_ctxs[i].cpu, err);
+			status = SCX_TEST_FAIL;
+		} else {
+			nr_contenders_joined++;
+			if (thread_err) {
+				SCX_ERR("Contender on CPU %d failed (%d)",
+					contender_ctxs[i].cpu, thread_err);
+				status = SCX_TEST_FAIL;
+			}
+		}
+	}
+
+	/* Restore the fair scheduler before waiting for any stranded thread. */
+	if (join_timed_out) {
+		atomic_store_explicit(&thread_ctx.abort, true,
+				      memory_order_release);
+		if (link) {
+			bpf_link__destroy(link);
+			link = NULL;
+		}
+		if (donor_started)
+			pthread_join(donor, NULL);
+		if (owner_started)
+			pthread_join(owner, NULL);
+		for (i = nr_contenders_joined;
+		     i < nr_contenders_started; i++)
+			pthread_join(contenders[i], NULL);
+	}
+
+	if (ioctl(thread_ctx.fd, ENQ_BLOCKED_IOCTL_GET_STATS, &stats)) {
+		SCX_ERR("Failed to read mutex statistics (%d)", errno);
+		status = SCX_TEST_FAIL;
+	} else {
+		result->stats = stats;
+		printf("\n[topology=%s SCX_OPS_ENQ_BLOCKED=%s]\n",
+		       cross_cpu ? "cross-cpu" : "same-cpu",
+		       enq_blocked ? "enabled" : "disabled");
+		printf("  proxy_exec=%s\n",
+		       proxy_enabled ? "enabled" : "disabled");
+		printf("  donor_cpu=%d\n", thread_ctx.donor_cpu);
+		printf("  owner_cpu=%d\n", thread_ctx.owner_cpu);
+		printf("  nr_contenders=%zu\n", nr_contenders);
+		printf("  measured_trials=%d\n", NR_MEASURED_TRIALS);
+		printf("  owner_nice=%d\n", OWNER_NICE);
+		printf("  donor_nice=%d\n", DONOR_NICE);
+		printf("  contender_nice=%d\n", CONTENDER_NICE);
+		print_avg_time("mutex_hold", stats.hold_time_ns, stats.nr_holds);
+		print_avg_time("mutex_wait", stats.wait_time_ns, stats.nr_waits);
+		if (stats.nr_holds != NR_MEASURED_TRIALS ||
+		    stats.nr_waits != NR_MEASURED_TRIALS) {
+			SCX_ERR("Expected %d measured trials, got %llu holds and %llu waits",
+				NR_MEASURED_TRIALS,
+				(unsigned long long)stats.nr_holds,
+				(unsigned long long)stats.nr_waits);
+			status = SCX_TEST_FAIL;
+		}
+	}
+
+	nr_blocked = skel->bss->nr_blocked_enqueues;
+	nr_blocked_donor_cpu = skel->bss->nr_blocked_enqueues_donor_cpu;
+	nr_blocked_owner_cpu = skel->bss->nr_blocked_enqueues_owner_cpu;
+	nr_blocked_other_cpu = skel->bss->nr_blocked_enqueues_other_cpu;
+	nr_blocked_wakeups = skel->bss->nr_blocked_wakeups;
+	result->nr_blocked_enqueues = nr_blocked;
+	result->nr_blocked_enqueues_donor_cpu = nr_blocked_donor_cpu;
+	result->nr_blocked_enqueues_owner_cpu = nr_blocked_owner_cpu;
+	result->nr_blocked_enqueues_other_cpu = nr_blocked_other_cpu;
+	result->nr_blocked_wakeups = nr_blocked_wakeups;
+	printf("  nr_blocked_enqueues=%llu\n",
+	       (unsigned long long)nr_blocked);
+	printf("  nr_blocked_enqueues_donor_cpu=%llu\n",
+	       (unsigned long long)nr_blocked_donor_cpu);
+	printf("  nr_blocked_enqueues_owner_cpu=%llu\n",
+	       (unsigned long long)nr_blocked_owner_cpu);
+	printf("  nr_blocked_enqueues_other_cpu=%llu\n",
+	       (unsigned long long)nr_blocked_other_cpu);
+	printf("  nr_blocked_wakeups=%llu\n",
+	       (unsigned long long)nr_blocked_wakeups);
+	if (status == SCX_TEST_PASS) {
+		if (enq_blocked && proxy_enabled && !nr_blocked) {
+			SCX_ERR("ops.enqueue() did not receive the blocked donor");
+			status = SCX_TEST_FAIL;
+		} else if ((!enq_blocked || !proxy_enabled) && nr_blocked) {
+			SCX_ERR("ops.enqueue() unexpectedly received %llu blocked donors",
+				(unsigned long long)nr_blocked);
+			status = SCX_TEST_FAIL;
+		} else if (nr_blocked_wakeups) {
+			SCX_ERR("Ordinary wakeups received %llu blocked enqueue flags",
+				(unsigned long long)nr_blocked_wakeups);
+			status = SCX_TEST_FAIL;
+		} else if (nr_blocked_other_cpu) {
+			SCX_ERR("Blocked donor had %llu enqueues on unexpected CPUs",
+				(unsigned long long)nr_blocked_other_cpu);
+			status = SCX_TEST_FAIL;
+		}
+	}
+
+	if (skel->data->uei.kind != EXIT_KIND(SCX_EXIT_NONE)) {
+		SCX_ERR("Scheduler exited unexpectedly (kind=%llu code=%lld)",
+			(unsigned long long)skel->data->uei.kind,
+			(long long)skel->data->uei.exit_code);
+		status = SCX_TEST_FAIL;
+	}
+
+	if (link)
+		bpf_link__destroy(link);
+out_fd:
+	close(thread_ctx.fd);
+out_module:
+	unload_test_module(module_loaded);
+out_skel:
+	enq_blocked__destroy(skel);
+out_contenders:
+	free(contender_ctxs);
+	free(contenders);
+	return status;
+}
+
+static enum scx_test_status run_topology(bool cross_cpu)
+{
+	struct run_result disabled = {}, enabled = {};
+	enum scx_test_status status;
+
+	status = run_one(false, cross_cpu, &disabled);
+	if (status != SCX_TEST_PASS)
+		return status;
+
+	status = run_one(true, cross_cpu, &enabled);
+	if (status != SCX_TEST_PASS)
+		return status;
+
+	printf("\n[topology=%s delta: enabled - disabled]\n",
+	       cross_cpu ? "cross-cpu" : "same-cpu");
+	print_avg_delta("mutex_hold", disabled.stats.hold_time_ns,
+			disabled.stats.nr_holds, enabled.stats.hold_time_ns,
+			enabled.stats.nr_holds);
+	print_avg_delta("mutex_wait", disabled.stats.wait_time_ns,
+			disabled.stats.nr_waits, enabled.stats.wait_time_ns,
+			enabled.stats.nr_waits);
+
+	return SCX_TEST_PASS;
+}
+
+static enum scx_test_status run(void *ctx)
+{
+	enum scx_test_status status;
+
+	(void)ctx;
+
+	status = run_topology(false);
+	if (status != SCX_TEST_PASS)
+		return status;
+
+	status = run_topology(true);
+	if (status == SCX_TEST_SKIP)
+		return SCX_TEST_PASS;
+
+	return status;
+}
+
+struct scx_test enq_blocked = {
+	.name = "enq_blocked",
+	.description = "Verify proxy donor admission under CPU-wide contention",
+	.setup = setup,
+	.run = run,
+};
+
+REGISTER_SCX_TEST(&enq_blocked)
diff --git a/tools/testing/selftests/sched_ext/enq_blocked.h b/tools/testing/selftests/sched_ext/enq_blocked.h
new file mode 100644
index 0000000000000..ef1eb97feebe6
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/enq_blocked.h
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES */
+#ifndef __ENQ_BLOCKED_H
+#define __ENQ_BLOCKED_H
+
+#include <linux/ioctl.h>
+#include <linux/types.h>
+
+struct enq_blocked_stats {
+	__u64 hold_time_ns;
+	__u64 wait_time_ns;
+	__u64 nr_holds;
+	__u64 nr_waits;
+};
+
+#define ENQ_BLOCKED_IOCTL_OWNER	_IO('s', 1)
+#define ENQ_BLOCKED_IOCTL_DONOR	_IO('s', 2)
+#define ENQ_BLOCKED_IOCTL_RESET_STATS	_IO('s', 3)
+#define ENQ_BLOCKED_IOCTL_GET_STATS	_IOR('s', 4, struct enq_blocked_stats)
+#define ENQ_BLOCKED_IOCTL_PREP_ATTACH	_IO('s', 5)
+#define ENQ_BLOCKED_IOCTL_ATTACH_DONE	_IO('s', 6)
+#define ENQ_BLOCKED_IOCTL_DONOR_STATE	_IO('s', 7)
+#define ENQ_BLOCKED_IOCTL_PROXY_SUPPORTED _IO('s', 8)
+
+#define ENQ_BLOCKED_DONOR_SLEEPING	(1U << 0)
+#define ENQ_BLOCKED_DONOR_ON_RQ		(1U << 1)
+
+#endif /* __ENQ_BLOCKED_H */
diff --git a/tools/testing/selftests/sched_ext/test_modules/Makefile b/tools/testing/selftests/sched_ext/test_modules/Makefile
new file mode 100644
index 0000000000000..a0e9e9401ead6
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/test_modules/Makefile
@@ -0,0 +1,13 @@
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+
+TESTMODS_DIR := $(realpath $(dir $(abspath $(lastword $(MAKEFILE_LIST)))))
+KDIR ?= $(if $(O),$(O),$(realpath ../../../../..))
+
+obj-m += scx_enq_blocked_test.o
+
+all:
+	+$(Q)$(MAKE) -C $(KDIR) M=$(TESTMODS_DIR) modules
+
+clean:
+	+$(Q)$(MAKE) -C $(KDIR) M=$(TESTMODS_DIR) clean
diff --git a/tools/testing/selftests/sched_ext/test_modules/scx_enq_blocked_test.c b/tools/testing/selftests/sched_ext/test_modules/scx_enq_blocked_test.c
new file mode 100644
index 0000000000000..908689ed55786
--- /dev/null
+++ b/tools/testing/selftests/sched_ext/test_modules/scx_enq_blocked_test.c
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
+ *
+ * Kernel mutex fixture for the sched_ext SCX_OPS_ENQ_BLOCKED selftest.
+ */
+
+#include <linux/atomic.h>
+#include <linux/fs.h>
+#include <linux/jiffies.h>
+#include <linux/ktime.h>
+#include <linux/miscdevice.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/sched.h>
+#include <linux/uaccess.h>
+
+#include "../enq_blocked.h"
+
+#define DONOR_WAIT_TIMEOUT	msecs_to_jiffies(2000)
+#define ATTACH_WAIT_TIMEOUT	msecs_to_jiffies(10000)
+#define MUTEX_HOLD_TIME		msecs_to_jiffies(200)
+
+static DEFINE_MUTEX(test_mutex);
+static DEFINE_SPINLOCK(donor_lock);
+static struct task_struct *donor_task;
+static atomic_t owner_ready = ATOMIC_INIT(0);
+static atomic_t donor_started = ATOMIC_INIT(0);
+static atomic_t attach_pending = ATOMIC_INIT(0);
+static atomic_t attach_done = ATOMIC_INIT(0);
+static atomic64_t hold_time_ns = ATOMIC64_INIT(0);
+static atomic64_t wait_time_ns = ATOMIC64_INIT(0);
+static atomic64_t nr_holds = ATOMIC64_INIT(0);
+static atomic64_t nr_waits = ATOMIC64_INIT(0);
+
+static long run_owner(void)
+{
+	unsigned long timeout;
+	u64 start_ns;
+	long ret = 0;
+
+	atomic_set(&donor_started, 0);
+	mutex_lock(&test_mutex);
+	start_ns = ktime_get_ns();
+	atomic_set(&owner_ready, 1);
+
+	timeout = jiffies + DONOR_WAIT_TIMEOUT;
+	while (!atomic_read(&donor_started)) {
+		if (time_after(jiffies, timeout)) {
+			ret = -ETIMEDOUT;
+			goto out;
+		}
+		cond_resched();
+	}
+	if (atomic_xchg(&attach_pending, 0)) {
+		timeout = jiffies + ATTACH_WAIT_TIMEOUT;
+		while (!atomic_read(&attach_done)) {
+			if (time_after(jiffies, timeout)) {
+				ret = -ETIMEDOUT;
+				goto out;
+			}
+			cond_resched();
+		}
+	}
+
+	/* Keep yielding while the donor blocks on test_mutex. */
+	timeout = jiffies + MUTEX_HOLD_TIME;
+	while (time_before(jiffies, timeout))
+		cond_resched();
+
+out:
+	atomic_set(&owner_ready, 0);
+	atomic64_add(ktime_get_ns() - start_ns, &hold_time_ns);
+	atomic64_inc(&nr_holds);
+	mutex_unlock(&test_mutex);
+	return ret;
+}
+
+static long run_donor(void)
+{
+	unsigned long flags;
+	u64 start_ns;
+
+	if (!atomic_read(&owner_ready))
+		return -EAGAIN;
+
+	get_task_struct(current);
+	spin_lock_irqsave(&donor_lock, flags);
+	WARN_ON_ONCE(donor_task);
+	donor_task = current;
+	spin_unlock_irqrestore(&donor_lock, flags);
+
+	atomic_set(&donor_started, 1);
+	start_ns = ktime_get_ns();
+	mutex_lock(&test_mutex);
+
+	spin_lock_irqsave(&donor_lock, flags);
+	donor_task = NULL;
+	spin_unlock_irqrestore(&donor_lock, flags);
+	put_task_struct(current);
+
+	atomic64_add(ktime_get_ns() - start_ns, &wait_time_ns);
+	atomic64_inc(&nr_waits);
+	mutex_unlock(&test_mutex);
+	return 0;
+}
+
+static long get_donor_state(void)
+{
+	struct task_struct *task;
+	unsigned long flags;
+	long state = 0;
+
+	spin_lock_irqsave(&donor_lock, flags);
+	task = donor_task;
+	if (task)
+		get_task_struct(task);
+	spin_unlock_irqrestore(&donor_lock, flags);
+	if (!task)
+		return -ENOENT;
+
+	if (READ_ONCE(task->__state) != TASK_RUNNING)
+		state |= ENQ_BLOCKED_DONOR_SLEEPING;
+	if (READ_ONCE(task->on_rq))
+		state |= ENQ_BLOCKED_DONOR_ON_RQ;
+	put_task_struct(task);
+	return state;
+}
+
+static void reset_stats(void)
+{
+	atomic64_set(&hold_time_ns, 0);
+	atomic64_set(&wait_time_ns, 0);
+	atomic64_set(&nr_holds, 0);
+	atomic64_set(&nr_waits, 0);
+}
+
+static long get_stats(unsigned long arg)
+{
+	struct enq_blocked_stats stats = {
+		.hold_time_ns = atomic64_read(&hold_time_ns),
+		.wait_time_ns = atomic64_read(&wait_time_ns),
+		.nr_holds = atomic64_read(&nr_holds),
+		.nr_waits = atomic64_read(&nr_waits),
+	};
+
+	return copy_to_user((void __user *)arg, &stats, sizeof(stats)) ?
+		-EFAULT : 0;
+}
+
+static long enq_blocked_ioctl(struct file *file, unsigned int cmd,
+			      unsigned long arg)
+{
+	switch (cmd) {
+	case ENQ_BLOCKED_IOCTL_OWNER:
+		return run_owner();
+	case ENQ_BLOCKED_IOCTL_DONOR:
+		return run_donor();
+	case ENQ_BLOCKED_IOCTL_RESET_STATS:
+		reset_stats();
+		return 0;
+	case ENQ_BLOCKED_IOCTL_GET_STATS:
+		return get_stats(arg);
+	case ENQ_BLOCKED_IOCTL_PREP_ATTACH:
+		atomic_set(&attach_done, 0);
+		atomic_set(&attach_pending, 1);
+		return 0;
+	case ENQ_BLOCKED_IOCTL_ATTACH_DONE:
+		atomic_set(&attach_done, 1);
+		return 0;
+	case ENQ_BLOCKED_IOCTL_DONOR_STATE:
+		return get_donor_state();
+	case ENQ_BLOCKED_IOCTL_PROXY_SUPPORTED:
+		return IS_ENABLED(CONFIG_SCHED_PROXY_EXEC);
+	default:
+		return -EINVAL;
+	}
+}
+
+static const struct file_operations enq_blocked_fops = {
+	.owner			= THIS_MODULE,
+	.unlocked_ioctl		= enq_blocked_ioctl,
+};
+
+static struct miscdevice enq_blocked_device = {
+	.minor	= MISC_DYNAMIC_MINOR,
+	.name	= "scx_enq_blocked",
+	.fops	= &enq_blocked_fops,
+	.mode	= 0600,
+};
+
+module_misc_device(enq_blocked_device);
+MODULE_AUTHOR("Andrea Righi <arighi@nvidia.com>");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("sched_ext blocked donor test module");
-- 
2.55.0

[PATCH 11/12] sched_ext: scx_qmap: Add proxy execution support

[Andrea Righi]

Add a -B option to opt scx_qmap into queueing mutex-blocked tasks for
proxy execution. Without the option, SCX_OPS_ENQ_BLOCKED remains clear
and mutex waiters block normally. With -B, blocked donors are passed to
qmap_enqueue() with SCX_ENQ_BLOCKED.

When scx_qmap receives a blocked donor, dispatch it directly to the
local DSQ of its current cid with a fresh slice and SCX_ENQ_PREEMPT.
This places the donor at the head of the DSQ and requests an immediate
reschedule, allowing the core proxy-exec path to run the mutex owner
using the donor's scheduling context as soon as the donor is selected.

The blocked policy is intentionally unfair and can strongly prioritize
tasks using contended mutexes, but scx_qmap is a demo scheduler and such
aggressive behavior makes proxy-exec support easy to observe. Count
these dispatch attempts in nr_enq_blocked and report their per-interval
delta.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 tools/sched_ext/scx_qmap.bpf.c | 34 ++++++++++++++++++++++++++++++++++
 tools/sched_ext/scx_qmap.c     | 13 ++++++++++---
 tools/sched_ext/scx_qmap.h     |  1 +
 3 files changed, 45 insertions(+), 3 deletions(-)

diff --git a/tools/sched_ext/scx_qmap.bpf.c b/tools/sched_ext/scx_qmap.bpf.c
index aead17658573b..5ca1467c1325a 100644
--- a/tools/sched_ext/scx_qmap.bpf.c
+++ b/tools/sched_ext/scx_qmap.bpf.c
@@ -451,6 +451,40 @@ void BPF_STRUCT_OPS(qmap_enqueue, struct task_struct *p, u64 enq_flags)
 	 */
 	taskc->core_sched_seq = qa.core_sched_tail_seqs[idx]++;
 
+	/*
+	 * A blocked donor rejected by a local DSQ must not be dispatched back
+	 * to the same DSQ immediately. Move it to the global DSQ so another CPU
+	 * can pick it up and kick an idle cid to accelerate its consumption.
+	 */
+	if ((enq_flags & (SCX_ENQ_BLOCKED | SCX_ENQ_REENQ)) ==
+	    (SCX_ENQ_BLOCKED | SCX_ENQ_REENQ)) {
+		taskc->force_local = false;
+		scx_bpf_dsq_insert(p, SHARED_DSQ, 0, enq_flags);
+		cid = cmask_next_and2_set_wrap(&taskc->cpus_allowed,
+					       &qa.idle_cids.mask,
+					       &qa.self_cids.mask, 0);
+		if (cid < scx_bpf_nr_cids())
+			scx_bpf_kick_cid(cid, SCX_KICK_IDLE);
+		return;
+	}
+
+	/*
+	 * Insert a blocked mutex donor at the head of its current cid's local
+	 * DSQ with a fresh slice and %SCX_ENQ_PREEMPT, requesting an immediate
+	 * reschedule. Once selected, the core proxy-exec path can immediately
+	 * run the mutex owner using the donor's scheduling context.
+	 *
+	 * This policy is intentionally unfair and can strongly prioritize tasks
+	 * using contended mutexes; scx_qmap is a demonstration scheduler and
+	 * this behavior makes proxy-exec support easy to observe.
+	 */
+	if (enq_flags & SCX_ENQ_BLOCKED) {
+		__sync_fetch_and_add(&qa.nr_enq_blocked, 1);
+		scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL_ON | scx_bpf_task_cid(p),
+				   slice_ns, enq_flags | SCX_ENQ_PREEMPT);
+		return;
+	}
+
 	/*
 	 * A node with children delegates most cids. A task of ours that can run
 	 * on none of our self cids (e.g. a per-NUMA kthread pinned to delegated
diff --git a/tools/sched_ext/scx_qmap.c b/tools/sched_ext/scx_qmap.c
index 27ffda1c519e8..8e943e8cb202d 100644
--- a/tools/sched_ext/scx_qmap.c
+++ b/tools/sched_ext/scx_qmap.c
@@ -46,7 +46,7 @@ const char help_fmt[] =
 "See the top-of-file comment in .bpf.c for the design.\n"
 "\n"
 "Usage: %s [-s SLICE_US] [-e COUNT] [-t COUNT] [-T COUNT] [-l COUNT] [-b COUNT]\n"
-"       [-N COUNT] [-P] [-M] [-H] [-c CG_PATH] [-d PID] [-D LEN] [-S] [-p] [-I]\n"
+"       [-N COUNT] [-P] [-M] [-H] [-c CG_PATH] [-d PID] [-D LEN] [-S] [-p] [-I] [-B]\n"
 "       [-F COUNT] [-i SEC] [-R MS] [-J MODE] [-v]\n"
 "\n"
 "  -s SLICE_US   Override slice duration\n"
@@ -65,6 +65,7 @@ const char help_fmt[] =
 "  -S            Suppress qmap-specific debug dump\n"
 "  -p            Switch only tasks on SCHED_EXT policy instead of all\n"
 "  -I            Turn on SCX_OPS_ALWAYS_ENQ_IMMED\n"
+"  -B            Turn on SCX_OPS_ENQ_BLOCKED\n"
 "  -F COUNT      IMMED stress: force every COUNT'th enqueue to a busy local DSQ (use with -I)\n"
 "  -C MODE       cid-override test (shuffle|bad-dup|bad-range|bad-mono)\n"
 "  -i SEC        Stats interval, seconds (default 5)\n"
@@ -105,6 +106,7 @@ struct hier_prev {
 	u64 nr_dsps[MAX_SUB_SCHEDS];
 	u64 nr_reenq_cap;
 	u64 nr_reenq_immed;
+	u64 nr_enq_blocked;
 	u64 nr_inject_attempts;
 };
 
@@ -187,13 +189,15 @@ static void print_hier(struct qmap_arena *qa, struct hier_prev *prev, u64 own_cg
 	}
 
 	format_cid_ranges(qa, CID_SHARED, ranges, sizeof(ranges));
-	printf("hier   : nsub=%llu excl=%u shared=%s rr=%s reenq cap/immed +%llu/+%llu inj=+%llu\n",
+	printf("hier   : nsub=%llu excl=%u shared=%s rr=%s reenq cap/immed +%llu/+%llu blocked=+%llu inj=+%llu\n",
 	       (unsigned long long)qa->nr_sub_scheds, qa->part.nr_excl, ranges, rr,
 	       (unsigned long long)(qa->nr_reenq_cap - prev->nr_reenq_cap),
 	       (unsigned long long)(qa->nr_reenq_immed - prev->nr_reenq_immed),
+	       (unsigned long long)(qa->nr_enq_blocked - prev->nr_enq_blocked),
 	       (unsigned long long)(qa->nr_inject_attempts - prev->nr_inject_attempts));
 	prev->nr_reenq_cap = qa->nr_reenq_cap;
 	prev->nr_reenq_immed = qa->nr_reenq_immed;
+	prev->nr_enq_blocked = qa->nr_enq_blocked;
 	prev->nr_inject_attempts = qa->nr_inject_attempts;
 
 	printf("hier   : %-4s %10s %4s %6s %8s  %s\n",
@@ -256,7 +260,7 @@ int main(int argc, char **argv)
 	skel->rodata->slice_ns = __COMPAT_ENUM_OR_ZERO("scx_public_consts", "SCX_SLICE_DFL");
 	skel->rodata->max_tasks = 16384;
 
-	while ((opt = getopt(argc, argv, "s:e:t:T:l:b:N:PMHc:d:D:SpIF:C:i:R:J:vh")) != -1) {
+	while ((opt = getopt(argc, argv, "s:e:t:T:l:b:N:PMHc:d:D:SpIBF:C:i:R:J:vh")) != -1) {
 		switch (opt) {
 		case 's':
 			skel->rodata->slice_ns = strtoull(optarg, NULL, 0) * 1000;
@@ -318,6 +322,9 @@ int main(int argc, char **argv)
 			skel->rodata->always_enq_immed = true;
 			skel->struct_ops.qmap_ops->flags |= SCX_OPS_ALWAYS_ENQ_IMMED;
 			break;
+		case 'B':
+			skel->struct_ops.qmap_ops->flags |= SCX_OPS_ENQ_BLOCKED;
+			break;
 		case 'F':
 			skel->rodata->immed_stress_nth = strtoul(optarg, NULL, 0);
 			break;
diff --git a/tools/sched_ext/scx_qmap.h b/tools/sched_ext/scx_qmap.h
index c42f7ef74b894..7f4d3b33e6133 100644
--- a/tools/sched_ext/scx_qmap.h
+++ b/tools/sched_ext/scx_qmap.h
@@ -174,6 +174,7 @@ struct qmap_arena {
 	/* bpf -> userspace: stats */
 	u64 nr_reenq_cap;		/* SCX_TASK_REENQ_CAP bounces */
 	u64 nr_reenq_immed;		/* SCX_TASK_REENQ_IMMED bounces */
+	u64 nr_enq_blocked;		/* SCX_ENQ_BLOCKED dispatches */
 	u64 nr_inject_attempts;		/* fault-injection: dispatches to an unheld cid */
 	u32 inject_mode;		/* fault-injection mode (QMAP_INJ_*) */
 };
-- 
2.55.0

[PATCH 12/12] sched: Allow enabling proxy exec with sched_ext

[Andrea Righi]

Now that sched_ext can handle proxy donors and BPF schedulers can opt in
to blocked-donor enqueueing with SCX_OPS_ENQ_BLOCKED, remove the
!SCHED_CLASS_EXT dependency from SCHED_PROXY_EXEC and allow both options
to be enabled together.

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 init/Kconfig | 2 --
 1 file changed, 2 deletions(-)

diff --git a/init/Kconfig b/init/Kconfig
index 5230d4879b1c8..0817e62266e03 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -934,8 +934,6 @@ config SCHED_PROXY_EXEC
 	bool "Proxy Execution"
 	# Avoid some build failures w/ PREEMPT_RT until it can be fixed
 	depends on !PREEMPT_RT
-	# Need to investigate how to inform sched_ext of split contexts
-	depends on !SCHED_CLASS_EXT
 	# Not particularly useful until we get to multi-rq proxying
 	depends on EXPERT
 	help
-- 
2.55.0

Re: [PATCH 01/12] sched/core: Avoid false migration warning for proxy donors

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The warning for migrating a migration-disabled task is incorrectly suppressed for all sleeping tasks when proxy execution is enabled, instead of just for blocked proxy donors.
--

--- Patch [1]: [PATCH 01/12] sched/core: Avoid false migration warning for proxy donors ---
commit 3b0fd432ea71e1ba27a8e56bb387fb372b053daf
Author: Andrea Righi <arighi@nvidia.com>

sched/core: Avoid false migration warning for proxy donors

This commit updates set_task_cpu to suppress a false positive warning when
migrating a proxy donor. It allows moving a blocked scheduling context to
another CPU even when the donor task has migration disabled.

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
[ ... ]
> @@ -3376,7 +3376,13 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
>  	 */
>  	WARN_ON_ONCE(!cpu_online(new_cpu));
>  
> -	WARN_ON_ONCE(is_migration_disabled(p));
> +	/*
> +	 * Proxy execution can move a blocked task's scheduling context to any
> +	 * CPU without moving its migration-disabled execution context. The
> +	 * wakeup path will return the task to a CPU where it can execute.
> +	 */
> +	WARN_ON_ONCE(is_migration_disabled(p) &&
> +		     !(sched_proxy_exec() && p->is_blocked));

[Severity: Medium]
Does this logic inadvertently suppress the warning for all sleeping tasks
rather than just proxy donors?

Because p->is_blocked is set unconditionally for all sleeping tasks when
__schedule calls try_to_block_task, this condition in set_task_cpu appears
to suppress the warning for any migration-disabled sleeping task whenever
proxy execution is globally enabled.

If another scheduler component, such as a load balancer or a BPF sched_ext
extension, incorrectly migrates a normal sleeping task that has migration
disabled, this overly broad check might hide the invalid migration and mask
further issues.

>  
>  	trace_sched_migrate_task(p, new_cpu);

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260721063242.552774-1-arighi@nvidia.com?part=1

Re: [PATCH 01/12] sched/core: Avoid false migration warning for proxy donors

[John Stultz]

On Mon, Jul 20, 2026 at 11:33 PM Andrea Righi <arighi@nvidia.com> wrote:
>
> Proxy execution can move a blocked donor's scheduling context to the
> lock owner's CPU even when the donor is migration-disabled. The donor
> does not execute there, and its original execution CPU remains recorded
> in wake_cpu.
>
> set_task_cpu() warns unconditionally for migration-disabled tasks, so a
> subsequent proxy migration or the wakeup path returning the donor home
> triggers a false positive. Moving a blocked scheduling context does not
> violate the migration-disabled execution context.
>
> Exclude blocked proxy donors from the warning. The proxy wakeup path
> restores an executable placement before clearing the blocked state.
>
> This is a preparatory fix for enabling proxy execution together with
> sched_ext.
>
> Signed-off-by: Andrea Righi <arighi@nvidia.com>

Nice catch on this!

I'm wondering why I've not tripped this myself yet. Is
migration_disabled set more often in the sched_ext paths? Or was there
a specific workload that you hit this with?

Acked-by: John Stultz <jstultz@google.com>

thanks
-john

Re: [PATCH 03/12] sched: Add helper to block retained proxy donors

[John Stultz]

On Mon, Jul 20, 2026 at 11:33 PM Andrea Righi <arighi@nvidia.com> wrote:
>
> Scheduler ownership changes may need to turn a mutex-blocked task
> retained on the runqueue for proxy execution back into a normally
> blocked task.
>
> Add sched_proxy_block_task() to perform that transition while holding
> p->pi_lock and the task's runqueue lock. Reset an active donor to the
> physical current task before blocking it. If DELAY_DEQUEUE keeps a FAIR
> donor queued, the immediately following sched_change_begin() completes
> the dequeue through switching_from_fair() before recording its queued
> state.
>
> This is a preparatory change to support proxy execution with sched_ext.
>
> Signed-off-by: Andrea Righi <arighi@nvidia.com>

Acked-by: John Stultz <jstultz@google.com>

Re: [PATCH 12/12] sched: Allow enabling proxy exec with sched_ext

[John Stultz]

On Mon, Jul 20, 2026 at 11:34 PM Andrea Righi <arighi@nvidia.com> wrote:
>
> Now that sched_ext can handle proxy donors and BPF schedulers can opt in
> to blocked-donor enqueueing with SCX_OPS_ENQ_BLOCKED, remove the
> !SCHED_CLASS_EXT dependency from SCHED_PROXY_EXEC and allow both options
> to be enabled together.
>
> Signed-off-by: Andrea Righi <arighi@nvidia.com>

Acked-by: John Stultz <jstultz@google.com>

So far things have held up pretty well in my testing.
I've seen a few cases where the scx scheduler gets disabled due to
stalls under stress-ng pressure or locktorture+other workloads, but I
suspect that may be separate from proxy-exec (I'll continue to do some
testing both with and without to confirm).

Huge thanks for this effort!
-john

Re: [PATCH 06/12] sched_ext: Fix proxy-exec race in consume_remote_task()

[John Stultz]

On Mon, Jul 20, 2026 at 11:34 PM Andrea Righi <arighi@nvidia.com> wrote:
>
> Without proxy execution, the DSQ lock and holding_cpu handshake ensure
> that a task cannot be dequeued or start running during the rq-lock
> handoff without clearing holding_cpu.
>
> Proxy execution is an exception. A task can start physically executing
> as a lock owner while its own scheduling context remains on the DSQ. Its
> on-CPU or migration-disabled state can therefore change without clearing
> holding_cpu.
>
> Keep the existing lockless destination validation unchanged. After
> switching to the source rq lock, recheck only the on-CPU and
> migration-disabled states when proxy execution is enabled. If either
> changed, cancel the transfer and requeue the task on the global DSQ
> without treating the race as a BPF scheduler error.
>
> A rejected task may be requeued onto the DSQ currently being consumed.
> Snapshot the DSQ sequence and skip tasks added after the consume attempt
> started to prevent repeatedly selecting the same task. Keep this retry
> bounding confined to proxy execution so the non-proxy path remains
> unchanged.
>
> With proxy execution enabled, stress-ng --pipeherd can trigger this race
> and migrate an active execution context, leading to
> sleeping-while-atomic warnings and subsequent lockdep corruption.
>
> Signed-off-by: Andrea Righi <arighi@nvidia.com>

So, I've acked the core scheduler changes, but I'm not sure I know the
sched_ext logic well enough for my ack to be very meaningful.

However, as a sign of general positivity the series, if you feel its
useful, feel free to add "Acked-by: John Stultz <jstultz@google.com>"
to the entire series.

thanks
-john

Re: [PATCH 02/12] sched: Make NOHZ CFS bandwidth checks follow proxy donor

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch fails to remove the `rq->nr_running != 1` restriction from `sched_fair_update_stop_tick()`, leaving proxy execution bandwidth enforcement broken during context switches.
--

--- Patch [2]: [PATCH 02/12] sched: Make NOHZ CFS bandwidth checks follow proxy donor ---
commit 4af18b2dbe3d9116ee4049aa70ca4c8662c0104c
Author: Andrea Righi <arighi@nvidia.com>

sched: Make NOHZ CFS bandwidth checks follow proxy donor

Proxy execution separates the scheduling context in rq->donor from the
physical execution context in rq->curr. sched_can_stop_tick() checks the
latter for CFS bandwidth constraints and only does so when nr_running is
one. A retained proxy donor keeps both the donor and mutex owner queued,
so this patch checks the selected donor instead and removes the
nr_running restriction.

Link: https://lore.kernel.org/r/20260713164807.E5ED21F00A3A@smtp.kernel.org

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index f788b72d1bfff..35d41d45e365e 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1409,11 +1409,8 @@ static void nohz_csd_func(void *info)
>  #endif /* CONFIG_NO_HZ_COMMON */
>  
>  #ifdef CONFIG_NO_HZ_FULL
> -static inline bool __need_bw_check(struct rq *rq, struct task_struct *p)
> +static inline bool __need_bw_check(struct task_struct *p)
>  {
> -	if (rq->nr_running != 1)
> -		return false;
> -

[Severity: High]
This is a pre-existing issue, but does removing this restriction here fully
resolve the proxy execution bandwidth enforcement bug?

Looking at sched_fair_update_stop_tick() in kernel/sched/fair.c, there is an
identical check that prevents tick updates when nr_running != 1:

kernel/sched/fair.c:sched_fair_update_stop_tick() {
	if (rq->nr_running != 1)
		return;
	...
}

If a constrained FAIR task donates its scheduling context to a lower-priority
task (like an SCX task) via proxy execution, nr_running becomes >= 2 while
h_nr_queued remains 1.

During enqueue, sched_can_stop_tick() evaluates against the old rq->donor
(e.g., the idle task) and may clear the tick dependency. When the scheduler
picks the FAIR donor, sched_fair_update_stop_tick() returns early because
rq->nr_running != 1.

Can this result in the tick remaining cleared, allowing a constrained FAIR
task to bypass its CFS bandwidth quota?

Should the nr_running restriction in sched_fair_update_stop_tick() also be
removed or updated to properly account for proxy execution?

>  	if (p->sched_class != &fair_sched_class)
>  		return false;
>

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260721063242.552774-1-arighi@nvidia.com?part=2

Re: [PATCH 01/12] sched/core: Avoid false migration warning for proxy donors

[Andrea Righi]

On Tue, Jul 21, 2026 at 12:00:45AM -0700, John Stultz wrote:
> On Mon, Jul 20, 2026 at 11:33 PM Andrea Righi <arighi@nvidia.com> wrote:
> >
> > Proxy execution can move a blocked donor's scheduling context to the
> > lock owner's CPU even when the donor is migration-disabled. The donor
> > does not execute there, and its original execution CPU remains recorded
> > in wake_cpu.
> >
> > set_task_cpu() warns unconditionally for migration-disabled tasks, so a
> > subsequent proxy migration or the wakeup path returning the donor home
> > triggers a false positive. Moving a blocked scheduling context does not
> > violate the migration-disabled execution context.
> >
> > Exclude blocked proxy donors from the warning. The proxy wakeup path
> > restores an executable placement before clearing the blocked state.
> >
> > This is a preparatory fix for enabling proxy execution together with
> > sched_ext.
> >
> > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> 
> Nice catch on this!
> 
> I'm wondering why I've not tripped this myself yet. Is
> migration_disabled set more often in the sched_ext paths? Or was there
> a specific workload that you hit this with?

I triggered this with a modified scx_cosmos with proxy-exec support (option
--proxy-exec): https://github.com/sched-ext/scx/tree/cosmos-proxy-exec

This was the trace, triggered when scx_cosmos was stopped:

  [   12.992201] WARNING: kernel/sched/core.c:3444 at set_task_cpu+0x1d3/0x280, CPU#6: kworker/u65:1/103
  [   12.992387] Modules linked in:
  [   12.992593] CPU: 6 UID: 0 PID: 103 Comm: kworker/u65:1 Not tainted 7.1.0-virtme #3 PREEMPT(full)
  [   12.993012] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
  [   12.993075] Workqueue: events_unbound bpf_map_free_deferred
  [   12.993139] RIP: 0010:set_task_cpu+0x1d3/0x280
  [   12.993285] Code: 14 89 6b 64 c6 83 09 17 00 00 01 5b 41 5e 41 5f 5d c3 cc cc cc cc cc 90 0f 0b 90 66 83 bb 80 05 00 00 00 0f 84 f8 fe
  ff ff 90 <0f> 0b 90 e9 ef fe ff ff 48 8d 3d ce 12 4c 02 67 48 0f b9 3a e9 03
  [   12.993656] RSP: 0018:ffffc900003f7cf8 EFLAGS: 00010002
  [   12.994642] RAX: 0000000000000000 RBX: ffff888100c6c800 RCX: 0000000000000046
  [   12.994756] RDX: ffffc900003f7d2c RSI: ffff888100c6d338 RDI: ffff888100c454d0
  [   12.994860] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000000
  [   12.994956] R10: 000000000000003b R11: ffffffff813f5c30 R12: 0000000000000000
  [   12.995055] R13: ffff888100c6d320 R14: ffff888100c44800 R15: 0000000000000206
  [   12.995167] FS:  0000000000000000(0000) GS:ffff8882b380a000(0000) knlGS:0000000000000000
  [   12.995273] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  [   12.995354] CR2: 00007fe19b902000 CR3: 000000000364c000 CR4: 0000000000750ef0
  [   12.995462] PKRU: 55555554
  [   12.995502] Call Trace:
  [   12.995540]  <TASK>
  [   12.995580]  try_to_wake_up+0x360/0x6f0
  [   12.995640]  __mutex_unlock_slowpath+0x32f/0x540
  [   12.995716]  ? bpf_local_storage_map_free+0x84/0x280
  [   12.995789]  ? process_scheduled_works+0x225/0x520
  [   12.995861]  bpf_local_storage_map_free+0x257/0x280
  [   12.995934]  bpf_map_free+0x84/0x110
  [   12.995992]  process_scheduled_works+0x271/0x520
  [   12.996071]  worker_thread+0x1e1/0x300
  [   12.996133]  ? __pfx_worker_thread+0x10/0x10
  [   12.996206]  kthread+0x10c/0x140
  [   12.996266]  ? __pfx_kthread+0x10/0x10
  [   12.996319]  ret_from_fork+0x189/0x330
  [   12.996378]  ? __pfx_kthread+0x10/0x10
  [   12.996435]  ret_from_fork_asm+0x1a/0x30
  [   12.996500]  </TASK>

Looking at bpf_local_storage_map_free(), I think I triggered this warning,
because bpf_map_free() calls migrate_disable() around map->ops->map_free() and
the deferred map-free workers can block on mutexes while migration is disabled.
Then proxy-exec moves the worker to the mutex owner's CPU and when the mutex is
released, ttwu returns the worker to its pinned CPU and trips the set_task_cpu()
warning.

> 
> Acked-by: John Stultz <jstultz@google.com>
> 
> thanks
> -john

Thanks,
-Andrea

Re: [PATCH 06/12] sched_ext: Fix proxy-exec race in consume_remote_task()

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] task_can_move_from_locked_rq() fails to check if the task is the active donor on the source runqueue under proxy execution.
--

--- Patch [6]: [PATCH 06/12] sched_ext: Fix proxy-exec race in consume_remote_task() ---
commit 681e47b7800ef6cd6756031dfb833542e391c7df
Author: Andrea Righi <arighi@nvidia.com>

sched_ext: Fix proxy-exec race in consume_remote_task()

This patch addresses a proxy execution race where a task's state can change
without clearing holding_cpu. It introduces task_can_move_from_locked_rq()
to recheck the on-CPU and migration-disabled states after switching to the
source runqueue lock. If the state changed, the task is requeued on the
global DSQ instead of aborting the BPF scheduler.

> diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
> index 30d5965d09108..b5fd91d9bca4f 100644
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c

[ ... ]

> @@ -2432,6 +2429,28 @@ static bool task_can_run_on_remote_rq(struct scx_sched *sch,
>  	return true;
>  }
>  
> +/*
> + * Proxy execution can change @p's execution and migration-disabled state
> + * without touching its DSQ entry or clearing holding_cpu. Check those states
> + * with @p's rq locked. Without proxy execution, the holding_cpu handshake is
> + * sufficient and this must not affect the existing migration path.
> + */
> +static bool task_can_move_from_locked_rq(struct task_struct *p)
> +{
> +	struct rq *src_rq = task_rq(p);
> +
> +	lockdep_assert_rq_held(src_rq);
> +
> +	if (!sched_proxy_exec())
> +		return true;
> +
> +	/* @p may be rq->curr under another task's proxy scheduling context. */
> +	if (task_on_cpu(src_rq, p))
> +		return false;
> +

[Severity: High]
Is it possible for the active scheduling context (the donor) to be migrated
here?

When a task blocks on a mutex and becomes a proxy donor, it is retained on
the runqueue (p->on_rq == 1) without being dequeued. put_prev_task_scx()
sees SCX_TASK_QUEUED and enqueues this blocked task onto a DSQ.

When another CPU attempts to consume from this DSQ, consume_remote_task()
invokes this new task_can_move_from_locked_rq(). Since the donor is
blocked, task_on_cpu(src_rq, p) evaluates to false.

Without a check for task_current_donor(src_rq, p), could this return true
and allow move_remote_task_to_local_dsq() to migrate the active donor,
corrupting the proxy execution PI chain?

> +	return !is_migration_disabled(p);
> +}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260721063242.552774-1-arighi@nvidia.com?part=6

Re: [PATCH 07/12] sched_ext: Split curr|donor references properly

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] task_tick_scx() unconditionally invokes ops.tick() for blocked proxy donors, violating the BPF state machine contract.
- [Medium] Missing task_current_donor() check in enqueue_task_scx() violates DSQ invariants for proxy donors.
--

--- Patch [7]: [PATCH 07/12] sched_ext: Split curr|donor references properly ---
commit 0a1e9539e4e14034ca89a7b3b4988f413e2e416b
Author: Andrea Righi <arighi@nvidia.com>

sched_ext: Split curr|donor references properly

This commit updates the sched_ext scheduler class to distinguish between the
physical execution context and the scheduling context during proxy execution.
It ensures that sched_ext policy, accounting, and various BPF kfuncs observe
the donor task instead of the physically executing task.

> diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
> index b5fd91d9bca4f..a466adf8e1548 100644
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c

[ ... ]

> @@ -3566,9 +3577,9 @@ void scx_tick(struct rq *rq)
>  	update_other_load_avgs(rq);
>  }
>  
> -static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
> +static void task_tick_scx(struct rq *rq, struct task_struct *donor, int queued)
>  {
> -	struct scx_sched *sch = scx_task_sched(curr);
> +	struct scx_sched *sch = scx_task_sched(donor);
>  
>  	update_curr_scx(rq);
>  
> @@ -3577,13 +3588,13 @@ static void task_tick_scx(struct rq *rq, struct task_struct *curr, int queued)
>  	 * we can't trust the slice management or ops.core_sched_before().
>  	 */
>  	if (scx_bypassing(sch, cpu_of(rq))) {
> -		set_task_slice(curr, 0);
> -		touch_core_sched(rq, curr);
> +		set_task_slice(donor, 0);
> +		touch_core_sched(rq, donor);
>  	} else if (SCX_HAS_OP(sch, tick)) {
> -		SCX_CALL_OP_TASK(sch, tick, rq, curr);
> +		SCX_CALL_OP_TASK(sch, tick, rq, donor);
>  	}

[Severity: High]
Does this code incorrectly call ops.tick() for a blocked donor?

The commit message explicitly states that a blocked donor does not enter an
ops.running() or ops.stopping() session. However, this unconditionally invokes
ops.tick() for the donor without verifying if it is blocked or tracked as
running.

Since BPF schedulers may allocate per-CPU state upon ops.running(), could this
crash or leak state when ops.tick() fires for a task the scheduler never
tracked as running?

[ ... ]

> @@ -9172,13 +9195,15 @@ __bpf_kfunc bool scx_bpf_task_set_slice(struct task_struct *p, u64 slice,
>  		return false;
>  
>  	/*
> -	 * Directly write only when we hold the lock of the rq @p is queued or
> -	 * running on. See the slice write rules above.
> +	 * Directly write only when we hold the lock of the rq @p is queued on or
> +	 * provides the current scheduling context for. Under proxy execution,
> +	 * rq->donor owns and consumes the slice while rq->curr executes on its
> +	 * behalf. See the slice write rules above.
>  	 */
>  	locked_rq = scx_locked_rq();
>  	if (!locked_rq ||
>  	    (READ_ONCE(p->scx.runnable_cpu) != cpu_of(locked_rq) &&
> -	     !task_current(locked_rq, p))) {
> +	     !task_current_donor(locked_rq, p))) {
>  		set_task_slice_oob(sch, p, slice);
>  		return true;
>  	}

This patch updates several task_current() checks to task_current_donor(),
but it appears one was missed in enqueue_task_scx():

kernel/sched/ext/ext.c:enqueue_task_scx() {
    ...
	/*
	 * Restoring a running task will be immediately followed by
	 * set_next_task_scx() which expects the task to not be on the BPF
	 * scheduler as tasks can only start running through local DSQs. Force
	 * direct-dispatch into the local DSQ by setting the sticky_cpu. Mark
	 * IGNORE_CAPS to force entry into the local DSQ.
	 */
	if (unlikely(enq_flags & ENQUEUE_RESTORE) && task_current(rq, p)) {
		sticky_cpu = cpu_of(rq);
		enq_flags |= SCX_ENQ_IGNORE_CAPS;
	}
    ...
}

[Severity: Medium]
Should this use task_current_donor(rq, p) instead of task_current(rq, p)?

If a blocked proxy donor undergoes priority inheritance, it could be
dequeued and re-enqueued with ENQUEUE_RESTORE. Since the donor is not
physically rq->curr, task_current(rq, p) evaluates to false.

Would this bypass the local DSQ assignment and erroneously push the active
proxy donor into the BPF scheduler instead of the local DSQ?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260721063242.552774-1-arighi@nvidia.com?part=7

Re: [PATCH 08/12] sched_ext: Handle blocked donor migration with proxy execution

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] The newly added proxy-execution handling for blocked donors in put_prev_task_scx() is unreachable for sched_ext tasks, completely breaking proxy execution for this class.
- [Medium] Blocked proxy donors are enqueued to the tail of the local DSQ instead of the head, causing priority inversion when other tasks are batched.
--

--- Patch [8]: [PATCH 08/12] sched_ext: Handle blocked donor migration with proxy execution ---
commit 6380b59c7e35fb635b7bc11e5fa121a8296ed2b7
Author: John Stultz <jstultz@google.com>

sched_ext: Handle blocked donor migration with proxy execution

With proxy execution enabled, mutex-blocked donors stay runnable so
their scheduling context can execute the lock owner.

> diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
> index a466adf8e1548..bf53f2be77f05 100644
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c
[ ... ]
> @@ -3122,6 +3126,25 @@ static void put_prev_task_scx(struct rq *rq, struct task_struct *p,
>  	if (p->scx.flags & SCX_TASK_QUEUED) {
>  		set_task_runnable(rq, p);
>  
> +		/*
> +		 * Mutex-blocked donors stay queued on the runqueue under proxy
> +		 * execution, but the donor never runs as itself, proxy-exec
> +		 * walks the blocked_on chain on the next __schedule() and runs
> +		 * the lock owner in its place.
> +		 *
> +		 * Put the donor on the local DSQ directly so pick_next_task()
> +		 * can still see it. find_proxy_task() will either run the chain
> +		 * owner or deactivate the donor so the wakeup path can return it
> +		 * and let BPF make a new dispatch decision once it is unblocked.
> +		 *
> +		 * This is preparatory code: a later patch will delegate blocked-donor
> +		 * admission to the BPF scheduler.
> +		 */
> +		if (p->is_blocked) {
> +			scx_dispatch_enqueue(sch, rq, &rq->scx.local_dsq, p, 0);

[Severity: High]
Is this newly added proxy execution handling for blocked donors reachable for
sched_ext tasks?

When a sched_ext task blocks on a mutex, try_to_block_task() checks
!task_is_blocked(prev) || !scx_allow_proxy_exec(prev) to decide if it
should be fully dequeued. scx_allow_proxy_exec() unconditionally returns
false for ext_sched_class:

kernel/sched/ext/ext.c:scx_allow_proxy_exec() {
        return p->sched_class != &ext_sched_class;
}

Since should_block evaluates to true, the task is passed to block_task(),
which unconditionally dequeues it via dequeue_task_scx(). Dequeuing clears
SCX_TASK_QUEUED.

When put_prev_task_scx() is then called by __schedule() during the task
switch, the task is no longer queued (p->scx.flags & SCX_TASK_QUEUED is
false), meaning this entire block of newly added proxy donor retention
logic seems to be bypassed as dead code.

[Severity: Medium]
Will enqueuing the blocked proxy donor to the tail of the local DSQ cause a
priority inversion?

If the BPF scheduler has previously dispatched a batch of tasks to the local
DSQ, using 0 (tail enqueue) places the blocked donor behind them.
pick_next_task() will pop and execute the batched tasks before the donor,
delaying the execution of the lock owner.

Should this use SCX_ENQ_HEAD instead, similar to slice-preempted tasks, so
the lock owner can execute immediately to release the contended mutex?

> +			goto switch_class;
> +		}
> +
>  		/*
>  		 * If @p has slice left and is being put, @p is getting
>  		 * preempted by a higher priority scheduler class or core-sched

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260721063242.552774-1-arighi@nvidia.com?part=8