Re: [PATCH] sched_ext: Fix spurious WARN on stale ops_state in ops_dequeue()

[Samuele Mariotti <smariotti@disroot.org>]

Hi Andrea,

On 13/05/2026 16:26, Andrea Righi wrote:
>Hi Samuele,
>
>On Wed, May 13, 2026 at 11:53:29AM +0200, Samuele Mariotti wrote:
>> ops_dequeue() can race with finish_dispatch() and spuriously trigger the
>> "queued task must be in BPF scheduler's custody" warning.
>>
>> ops_dequeue() snapshots p->scx.ops_state via atomic_long_read_acquire()
>> and then, in the SCX_OPSS_QUEUED arm, asserts that SCX_TASK_IN_CUSTODY
>> is set. The two reads are not atomic w.r.t. a concurrent
>> finish_dispatch() running on another CPU:
>>
>> CPU 1                                    CPU 2
>> =====                                    =====
>>                                          dequeue_task_scx()
>>                                            ops_dequeue()
>>                                              opss = read_acquire(ops_state)
>>                                                   = SCX_OPSS_QUEUED
>> finish_dispatch()
>>   cmpxchg ops_state:
>>     SCX_OPSS_QUEUED -> SCX_OPSS_DISPATCHING  [succeeds]
>>   dispatch_enqueue(SCX_DSQ_GLOBAL,
>>                    SCX_ENQ_CLEAR_OPSS)
>>     call_task_dequeue()
>>       p->scx.flags &= ~SCX_TASK_IN_CUSTODY
>>                                              WARN_ON_ONCE(!(p->scx.flags &
>>                                                      SCX_TASK_IN_CUSTODY))
>>                                             /* opss is stale: QUEUED,
>>                                              * but task already claimed */
>>     set_release(ops_state, SCX_OPSS_NONE)
>>
>> The race has been observed via two distinct call chains: the most common
>> goes through sched_setaffinity(), a rarer variant through
>> sched_change_begin().
>>
>> For SCX_DSQ_GLOBAL / SCX_DSQ_BYPASS, dispatch_enqueue() clears
>> SCX_TASK_IN_CUSTODY before clearing ops_state to SCX_OPSS_NONE
>> (intentional, to avoid concurrent non-atomic RMW of p->scx.flags against
>> ops_dequeue()). The window between those two writes is exactly what
>> ops_dequeue() observes as "QUEUED without custody".
>>
>> The observed state is not actually inconsistent, it just means CPU 1 has
>> already claimed the task and the QUEUED value held by CPU 2 is stale.
>> Re-read ops_state in that case; the next read is guaranteed to return
>> SCX_OPSS_DISPATCHING or SCX_OPSS_NONE, both of which exit the switch
>> cleanly. The retry is bounded: once IN_CUSTODY is cleared, ops_state has
>> already advanced past QUEUED for this dispatch cycle, and a fresh QUEUED
>> would require re-enqueue under p's rq lock, which CPU 2 holds.
>>
>> Fixes: ebf1ccff79c4 ("sched_ext: Fix ops.dequeue() semantics")
>> Suggested-by: Andrea Righi <arighi@nvidia.com>
>> Signed-off-by: Samuele Mariotti <smariotti@disroot.org>
>> Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
>> ---
>>  kernel/sched/ext.c | 5 ++++-
>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
>> index 23f7b3f63b09..d285e37f2177 100644
>> --- a/kernel/sched/ext.c
>> +++ b/kernel/sched/ext.c
>> @@ -2078,6 +2078,7 @@ static void ops_dequeue(struct rq *rq, struct task_struct *p, u64 deq_flags)
>>      /* dequeue is always temporary, don't reset runnable_at */
>>      clr_task_runnable(p, false);
>>
>> +retry:
>>      /* acquire ensures that we see the preceding updates on QUEUED */
>>      opss = atomic_long_read_acquire(&p->scx.ops_state);
>>
>> @@ -2092,7 +2093,9 @@ static void ops_dequeue(struct rq *rq, struct task_struct *p, u64 deq_flags)
>>          BUG();
>>      case SCX_OPSS_QUEUED:
>>          /* A queued task must always be in BPF scheduler's custody */
>> -        WARN_ON_ONCE(!(p->scx.flags & SCX_TASK_IN_CUSTODY));
>> +        if (!(p->scx.flags & SCX_TASK_IN_CUSTODY))
>> +            goto retry;
>
>Can we add a cpu_relax() before the goto? A hot spin polling two cachelines from
>another CPU could be very unkind to SMT siblings and bus traffic.
>
>Moreover, we completely lose the original WARN_ON_ONCE(), so we don't catch the
>case where the invariant QUEUED -> IN_CUSTODY is violated by a realy bug. How
>about adding a max retries as well, i.e., something like this:
>
>    int retries = 0;
>
>    ...
>retry:
>    ...
>    if (!(p->scx.flags & SCX_TASK_IN_CUSTODY) &&
>        !WARN_ON_ONCE(retries++ >= 128)) {
>        cpu_relax();
>        goto retry;
>    }
>
>> +
>>          if (atomic_long_try_cmpxchg(&p->scx.ops_state, &opss,
>>                          SCX_OPSS_NONE))
>>              break;
>> --
>> 2.54.0
>>
>
>Thanks,
>-Andrea

Thanks for the suggestion. I agree with adding cpu_relax() and the
retry limit to preserve the original WARN_ON_ONCE() as a safety net
for real bugs.

Given the improvements to efficiency, I would also improve the non-atomic
read of p->scx.flags by using READ_ONCE(), preventing the compiler from
caching the value across retries and ensuring each iteration observes the
latest value written by the concurrent finish_dispatch(). I would also
lower the retry limit from 128 to 4, since the maximum number of retries
observed empirically is 1, so 4 gives a reasonable safety margin without
spinning unnecessarily long. 

Something like this: 

if (!(READ_ONCE(p->scx.flags) & SCX_TASK_IN_CUSTODY) &&
     !WARN_ON_ONCE(retries++ >= 4)) {
         cpu_relax();
         goto retry;
}

Let me know if this looks good to you.

Thanks,
Samuele