Hi Andrea,

On Thu Mar 19, 2026 at 8:35 AM UTC, Andrea Righi wrote:
> A BPF scheduler may rely on p->cpus_ptr from ops.dispatch() to select a
> target CPU. However, task affinity can change between the dispatch
> decision and its finalization in finish_dispatch(). When this happens,
> the scheduler may attempt to dispatch a task to a CPU that is no longer
> allowed, resulting in fatal errors such as:
>
>  EXIT: runtime error (SCX_DSQ_LOCAL[_ON] target CPU 10 not allowed for stress-ng-race-[13565])
>
> This race exists because ops.dispatch() runs without holding the task's
> run queue lock, allowing a concurrent set_cpus_allowed() to update
> p->cpus_ptr while the BPF scheduler is still using it. The dispatch is
> then finalized using stale affinity information.
>
> Example timeline:
>
>   CPU0                                      CPU1
>   ----                                      ----
>                                             task_rq_lock(p)
>   if (cpumask_test_cpu(cpu, p->cpus_ptr))
>                                             set_cpus_allowed_scx(p, new_mask)
>                                             task_rq_unlock(p)
>       scx_bpf_dsq_insert(p,
>               SCX_DSQ_LOCAL_ON | cpu, 0)
>
> With commit ebf1ccff79c4 ("sched_ext: Fix ops.dequeue() semantics"), BPF
> schedulers can avoid the affinity race by tracking task state and
> handling %SCX_DEQ_SCHED_CHANGE in ops.dequeue(): when a task is dequeued
> due to a property change, the scheduler can update the task state and
> skip the direct dispatch from ops.dispatch() for non-queued tasks.
>
> However, schedulers that do not implement task state tracking and
> dispatch directly to a local DSQ directly from ops.dispatch() may
> trigger the scx_error() condition when the kernel validates the
> destination in dispatch_to_local_dsq().

The two paragraphs above mention "direct dispatch from ops.dispatch()"
and "dispatch directly to a local DSQ directly from ops.dispatch()".
My understanding is that a "direct dispatch" can only happen from
ops.select_cpu() or ops.enqueue(), not from ops.dispatch(). Is this just
an unfortunate choice of words?
Would "dispatch to a local DSQ" be a more accurate phrase here?

Thanks,
Kuba