The Linux Kernel Mailing List
 help / color / mirror / Atom feed
* [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched
@ 2026-11-18 10:24 Kuba Piecuch
  2026-11-18 13:25 ` Tao Cui
  2026-11-18 14:20 ` Andrea Righi
  0 siblings, 2 replies; 3+ messages in thread
From: Kuba Piecuch @ 2026-11-18 10:24 UTC (permalink / raw)
  To: Tejun Heo, David Vernet, Andrea Righi, Changwoo Min
  Cc: Kuba Piecuch, sched-ext, linux-kernel

scx_root_enable_workfn() sets scx_switching_all and calls
static_branch_enable(&__scx_enabled) before switching tasks, and
task_should_scx() gates only on scx_enabled(). static_branch_enable()
patches jump-label sites in address order rather than atomically:
task_should_scx() may be patched before __schedule().

If a concurrent fork or sched_setscheduler() runs in that window,
task_should_scx() observes scx_enabled() and moves the running task
from fair_sched_class to ext_sched_class, setting need_resched on the
class degradation. The following __schedule() can still observe
!scx_enabled() in next_active_class(), skip ext_sched_class, and pick
rq->idle while the runnable SCX task is still rq->donor:

  WARNING: kernel/sched/ext/ext.c:3840 at put_prev_task_scx+0x405/0x7a0
  Sched_ext: enq_select_cpu (enabling+all), task: runnable_at=+0ms
  Call Trace:
   __schedule+0x5cc/0x1100
   preempt_schedule_common+0x12/0x40
   preempt_schedule_thunk+0x16/0x40
   __sched_setscheduler+0x8d8/0x920

Gate task_should_scx() on a scx_task_switch_enabled flag that is set with
release semantics only after static_branch_enable(&__scx_enabled) has
finished patching all call sites and scx_switching_all is updated, and
cleared on disable alongside scx_switching_all.

This bug was found and the patch was written with the help of an AI
coding assistant.

Fixes: f0e1a0643a59 ("sched_ext: Implement BPF extensible scheduler class")
Cc: stable@vger.kernel.org # v6.12+
Signed-off-by: Kuba Piecuch <jpiecuch@google.com>
---
 kernel/sched/ext/ext.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
index 10de6bd865e8..e8c30a21e03e 100644
--- a/kernel/sched/ext/ext.c
+++ b/kernel/sched/ext/ext.c
@@ -134,6 +134,7 @@ DEFINE_PERCPU_RWSEM(scx_fork_rwsem);
 static atomic_t scx_enable_state_var = ATOMIC_INIT(SCX_DISABLED);
 static DEFINE_RAW_SPINLOCK(scx_bypass_lock);
 static bool scx_init_task_enabled;
+static bool scx_task_switch_enabled;
 static bool scx_switching_all;
 DEFINE_STATIC_KEY_FALSE(__scx_switched_all);
 static DEFINE_STATIC_KEY_FALSE(__scx_tid_to_task_enabled);
@@ -6493,6 +6494,15 @@ bool task_should_scx(int policy)
 	if (!scx_enabled())
 		return false;
 
+	/*
+	 * static_branch_enable(&__scx_enabled) patches call sites one by one.
+	 * Wait until all sites are patched before routing tasks to SCX, so
+	 * __schedule() never sees !scx_enabled() when a task is on SCX. Pairs
+	 * with smp_store_release() in scx_root_enable_workfn().
+	 */
+	if (unlikely(!smp_load_acquire(&scx_task_switch_enabled)))
+		return false;
+
 	/* scx is taking over all SCHED_OTHER and SCHED_EXT tasks */
 	if (READ_ONCE(scx_switching_all))
 		return true;
@@ -7568,6 +7578,7 @@ static void scx_root_disable(struct scx_sched *sch)
 
 	static_branch_disable(&__scx_switched_all);
 	WRITE_ONCE(scx_switching_all, false);
+	WRITE_ONCE(scx_task_switch_enabled, false);
 
 	/*
 	 * Shut down cgroup support before tasks so that the cgroup attach and
@@ -8965,11 +8976,14 @@ static void scx_root_enable_workfn(struct kthread_work *work)
 	percpu_up_write(&scx_fork_rwsem);
 
 	/*
-	 * All tasks are READY. It's safe to turn on scx_enabled() and switch
-	 * all eligible tasks.
+	 * All tasks are READY. Finish patching all scx_enabled() sites before
+	 * opening task_should_scx() and switching eligible tasks.
 	 */
-	WRITE_ONCE(scx_switching_all, !(ops->flags & SCX_OPS_SWITCH_PARTIAL));
+	WARN_ON_ONCE(scx_task_switch_enabled);
 	static_branch_enable(&__scx_enabled);
+	WRITE_ONCE(scx_switching_all, !(ops->flags & SCX_OPS_SWITCH_PARTIAL));
+	/* pairs with smp_load_acquire() in task_should_scx() */
+	smp_store_release(&scx_task_switch_enabled, true);
 
 	/*
 	 * We're fully committed and can't fail. The task READY -> ENABLED

base-commit: 700a332c3fe33014bbd5ae852d71ba679fc29e0a
-- 
2.56.0.679.gc527c65306-goog


^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched
  2026-11-18 10:24 [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched Kuba Piecuch
@ 2026-11-18 13:25 ` Tao Cui
  2026-11-18 14:20 ` Andrea Righi
  1 sibling, 0 replies; 3+ messages in thread
From: Tao Cui @ 2026-11-18 13:25 UTC (permalink / raw)
  To: Kuba Piecuch, Tejun Heo, David Vernet, Andrea Righi, Changwoo Min
  Cc: cui.tao, sched-ext, linux-kernel

Hello, Kuba,

在 2026/11/18 18:24, Kuba Piecuch 写道:
> scx_root_enable_workfn() sets scx_switching_all and calls
> static_branch_enable(&__scx_enabled) before switching tasks, and
> task_should_scx() gates only on scx_enabled(). static_branch_enable()
> patches jump-label sites in address order rather than atomically:
> task_should_scx() may be patched before __schedule().
> 
> If a concurrent fork or sched_setscheduler() runs in that window,
> task_should_scx() observes scx_enabled() and moves the running task
> from fair_sched_class to ext_sched_class, setting need_resched on the
> class degradation. The following __schedule() can still observe
> !scx_enabled() in next_active_class(), skip ext_sched_class, and pick
> rq->idle while the runnable SCX task is still rq->donor:
> 
>   WARNING: kernel/sched/ext/ext.c:3840 at put_prev_task_scx+0x405/0x7a0
>   Sched_ext: enq_select_cpu (enabling+all), task: runnable_at=+0ms
>   Call Trace:
>    __schedule+0x5cc/0x1100
>    preempt_schedule_common+0x12/0x40
>    preempt_schedule_thunk+0x16/0x40
>    __sched_setscheduler+0x8d8/0x920
> 
> Gate task_should_scx() on a scx_task_switch_enabled flag that is set with
> release semantics only after static_branch_enable(&__scx_enabled) has
> finished patching all call sites and scx_switching_all is updated, and
> cleared on disable alongside scx_switching_all.
> 
> This bug was found and the patch was written with the help of an AI
> coding assistant.
> 
> Fixes: f0e1a0643a59 ("sched_ext: Implement BPF extensible scheduler class")
> Cc: stable@vger.kernel.org # v6.12+
> Signed-off-by: Kuba Piecuch <jpiecuch@google.com>

The release/acquire gate looks right on both the
enable and disable paths.

Applied on top of 700a332c3fe3 (for-7.4-fixes HEAD) and ran 25
scx_simple load/unload cycles in qemu under a fork and
sched_setscheduler() storm. The baseline hit the reported warning on
its own 14s into the run, log below. The patched kernel went through
the identical run with zero warnings.

[   14.225125] CYCLE-7-start
[   14.352153] ------------[ cut here ]------------
[   14.352157] WARNING: kernel/sched/ext/ext.c:3840 at put_prev_task_scx+0x4e4/0x750, CPU#2: storm/91
[   14.354627] Sched_ext: simple (enabling+all), task: runnable_at=+0ms
[   14.372849] Call Trace:
[   14.373019]  <TASK>
[   14.373164]  __schedule+0x52d/0xe60
[   14.373395]  ? bpf__sched_ext_ops_running+0x62/0xac
[   14.373709]  preempt_schedule_common+0x12/0x40
[   14.380061]  preempt_schedule_thunk+0x16/0x40
[   14.380628]  __sched_setscheduler+0x815/0x8a0
[   14.380918]  do_sched_setscheduler+0x154/0x1f0
[   14.381220]  __x64_sys_sched_setscheduler+0x1c/0x40

Tested-by: Tao Cui <cuitao@kylinos.cn>
Reviewed-by: Tao Cui <cuitao@kylinos.cn>

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched
  2026-11-18 10:24 [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched Kuba Piecuch
  2026-11-18 13:25 ` Tao Cui
@ 2026-11-18 14:20 ` Andrea Righi
  1 sibling, 0 replies; 3+ messages in thread
From: Andrea Righi @ 2026-11-18 14:20 UTC (permalink / raw)
  To: Kuba Piecuch
  Cc: Tejun Heo, David Vernet, Changwoo Min, sched-ext, linux-kernel

Hi Kuba,

On Wed, Nov 18, 2026 at 10:24:52AM +0000, Kuba Piecuch wrote:
> scx_root_enable_workfn() sets scx_switching_all and calls
> static_branch_enable(&__scx_enabled) before switching tasks, and
> task_should_scx() gates only on scx_enabled(). static_branch_enable()
> patches jump-label sites in address order rather than atomically:
> task_should_scx() may be patched before __schedule().
> 
> If a concurrent fork or sched_setscheduler() runs in that window,
> task_should_scx() observes scx_enabled() and moves the running task
> from fair_sched_class to ext_sched_class, setting need_resched on the
> class degradation. The following __schedule() can still observe
> !scx_enabled() in next_active_class(), skip ext_sched_class, and pick
> rq->idle while the runnable SCX task is still rq->donor:
> 
>   WARNING: kernel/sched/ext/ext.c:3840 at put_prev_task_scx+0x405/0x7a0
>   Sched_ext: enq_select_cpu (enabling+all), task: runnable_at=+0ms
>   Call Trace:
>    __schedule+0x5cc/0x1100
>    preempt_schedule_common+0x12/0x40
>    preempt_schedule_thunk+0x16/0x40
>    __sched_setscheduler+0x8d8/0x920
> 
> Gate task_should_scx() on a scx_task_switch_enabled flag that is set with
> release semantics only after static_branch_enable(&__scx_enabled) has
> finished patching all call sites and scx_switching_all is updated, and
> cleared on disable alongside scx_switching_all.
> 
> This bug was found and the patch was written with the help of an AI
> coding assistant.
> 
> Fixes: f0e1a0643a59 ("sched_ext: Implement BPF extensible scheduler class")
> Cc: stable@vger.kernel.org # v6.12+
> Signed-off-by: Kuba Piecuch <jpiecuch@google.com>

Looks good, good catch.

Reviewed-by: Andrea Righi <arighi@nvidia.com>

Thanks,
-Andrea

> ---
>  kernel/sched/ext/ext.c | 20 +++++++++++++++++---
>  1 file changed, 17 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched/ext/ext.c b/kernel/sched/ext/ext.c
> index 10de6bd865e8..e8c30a21e03e 100644
> --- a/kernel/sched/ext/ext.c
> +++ b/kernel/sched/ext/ext.c
> @@ -134,6 +134,7 @@ DEFINE_PERCPU_RWSEM(scx_fork_rwsem);
>  static atomic_t scx_enable_state_var = ATOMIC_INIT(SCX_DISABLED);
>  static DEFINE_RAW_SPINLOCK(scx_bypass_lock);
>  static bool scx_init_task_enabled;
> +static bool scx_task_switch_enabled;
>  static bool scx_switching_all;
>  DEFINE_STATIC_KEY_FALSE(__scx_switched_all);
>  static DEFINE_STATIC_KEY_FALSE(__scx_tid_to_task_enabled);
> @@ -6493,6 +6494,15 @@ bool task_should_scx(int policy)
>  	if (!scx_enabled())
>  		return false;
>  
> +	/*
> +	 * static_branch_enable(&__scx_enabled) patches call sites one by one.
> +	 * Wait until all sites are patched before routing tasks to SCX, so
> +	 * __schedule() never sees !scx_enabled() when a task is on SCX. Pairs
> +	 * with smp_store_release() in scx_root_enable_workfn().
> +	 */
> +	if (unlikely(!smp_load_acquire(&scx_task_switch_enabled)))
> +		return false;
> +
>  	/* scx is taking over all SCHED_OTHER and SCHED_EXT tasks */
>  	if (READ_ONCE(scx_switching_all))
>  		return true;
> @@ -7568,6 +7578,7 @@ static void scx_root_disable(struct scx_sched *sch)
>  
>  	static_branch_disable(&__scx_switched_all);
>  	WRITE_ONCE(scx_switching_all, false);
> +	WRITE_ONCE(scx_task_switch_enabled, false);
>  
>  	/*
>  	 * Shut down cgroup support before tasks so that the cgroup attach and
> @@ -8965,11 +8976,14 @@ static void scx_root_enable_workfn(struct kthread_work *work)
>  	percpu_up_write(&scx_fork_rwsem);
>  
>  	/*
> -	 * All tasks are READY. It's safe to turn on scx_enabled() and switch
> -	 * all eligible tasks.
> +	 * All tasks are READY. Finish patching all scx_enabled() sites before
> +	 * opening task_should_scx() and switching eligible tasks.
>  	 */
> -	WRITE_ONCE(scx_switching_all, !(ops->flags & SCX_OPS_SWITCH_PARTIAL));
> +	WARN_ON_ONCE(scx_task_switch_enabled);
>  	static_branch_enable(&__scx_enabled);
> +	WRITE_ONCE(scx_switching_all, !(ops->flags & SCX_OPS_SWITCH_PARTIAL));
> +	/* pairs with smp_load_acquire() in task_should_scx() */
> +	smp_store_release(&scx_task_switch_enabled, true);
>  
>  	/*
>  	 * We're fully committed and can't fail. The task READY -> ENABLED
> 
> base-commit: 700a332c3fe33014bbd5ae852d71ba679fc29e0a
> -- 
> 2.56.0.679.gc527c65306-goog
> 

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2026-11-18 14:20 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 10:24 [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched Kuba Piecuch
2026-11-18 13:25 ` Tao Cui
2026-11-18 14:20 ` Andrea Righi

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox