Re: [PATCH v2 sched_ext/for-7.1] sched_ext: Invalidate dispatch decisions on CPU affinity changes

[Kuba Piecuch <jpiecuch@google.com>]

Hi Cheng-Yang,

On Wed Apr 22, 2026 at 6:33 AM UTC, Cheng-Yang Chou wrote:
> Hi Tejun, Andrea, and Kuba
>
> On Mon, Mar 23, 2026 at 01:13:20PM -1000, Tejun Heo wrote:
>> > The simple way to do this is to do scx_bpf_dsq_insert() at the very beginning,
>> > once we know which task we would like to dispatch, and cancel the pending
>> > dispatch via scx_bpf_dispatch_cancel() if any of the pre-dispatch checks fail
>> > on the BPF side. This way, the "critical section" includes BPF-side checks, and
>> > SCX will ignore the dispatch if there was a dequeue/enqueue racing with the
>> > critical section.
>> > 
>> > With this solution, we can throw an error if task_can_run_on_remote_rq() is
>> > false, because we know that there was no racing cpumask change (if there was,
>> > it would have been caught earlier, in finish_dispatch()).
>> 
>> Yeah, I think this makes more sense. qseq is already there to provide
>> protection against these events. It's just that the capturing of qseq is too
>> late. If insert/cancel is too ugly, we can introduce another kfunc to
>> capture the qseq - scx_bpf_dsq_insert_begin() or something like that - and
>> stash it in a per-cpu variable. That way, qseq would be cover the "current"
>> queued instance and the existing qseq mechanism would be able to reliably
>> ignore the ones that lost race to dequeue.
>
> Since this has been stale for a while, I prepared a patch to implement
> scx_bpf_dsq_insert_begin() as suggested.
>
> Is anyone else working on this? If not, I'm happy to send the formal
> patch to fix this.

Thanks for creating the patch. A couple of thoughts:

1. Do we have a use case that requires dsq_insert_begin() that isn't
   satisfied using the "insert and then cancel if needed" approach?

2. Do we want to restrict ourselves through the one qseq slot provided by
   dsq_insert_begin()? The most flexible approach IMO would be to simply
   allow BPF to read the qseq directly via a kfunc and then supply it to
   dsq_insert() later. With this, we can have multiple qseqs saved at the
   same time, and we can even pass them between CPUs, e.g. if one CPU
   dequeues a task for a sibling CPU, but we want the checks to be made inside
   the sibling's ops.dispatch() (I just made this use case it up, it may not
   be practical.)
   That said, exposing an internal thing like qseq to BPF may be a step too far.
   Let me know what you think.

Thanks,
Kuba