Re: [PATCH sched_ext/for-7.4-fixes] sched_ext: Don't route tasks to SCX before __scx_enabled is fully patched

From: Andrea Righi

Date: Wed Nov 18 2026 - 09:20:59 EST


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@xxxxxxxxxxxxxxx # v6.12+
> Signed-off-by: Kuba Piecuch <jpiecuch@xxxxxxxxxx>

Looks good, good catch.

Reviewed-by: Andrea Righi <arighi@xxxxxxxxxx>

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
>