Re: sched_ext: Fix SCX_KICK_WAIT to work reliably

[Andrea Righi <arighi@nvidia.com>]

On Wed, Oct 22, 2025 at 08:37:50AM -1000, Tejun Heo wrote:
> Hello, Andrea.
> 
> On Wed, Oct 22, 2025 at 09:43:25AM +0200, Andrea Righi wrote:
> > > @@ -5208,12 +5214,11 @@ static void kick_cpus_irq_workfn(struct
> > >  
> > >  		if (cpu != cpu_of(this_rq)) {
> > 
> > It's probably fine anyway, but should we check for cpu_online(cpu) here?
> 
> This block gets activated iff kick_one_cpu() returns true and that is gated
> by the CPU being online && the current task being on SCX. For the CPU to go
> offline, that task has to go off CPU and thus increment the sequence
> counter.

I was thinking if the CPU goes offline after kick_one_cpu() returns and
before reaching this loop, but even in this case we're not accessing
anything unsafe, so we should be fine.

> 
> > >  			while (smp_load_acquire(wait_pnt_seq) == pseqs[cpu])
> > >  				cpu_relax();
> > 
> > I'm wondering if we can break the semantic if cpu_rq(cpu)->curr->scx.slice
> > is refilled concurrently between kick_one_cpu() and this busy wait. In this
> > case we return, because wait_pnt_seq is incremented, but we keep running
> > the same task.
> > 
> > Should we introduce a flag (or something similar) to force the re-enqueue
> > of the prev task in this case?
> 
> Ah, right, that's a hole. There's another hole. The BPF scheduler can choose
> to run the same task and put_prev_task_scx() won't be called. I think we
> need to bump the seq count on entry to pick_task_scx() too. That should
> solve both problems. All that we're guaranteeing is that we wait until the
> task enters scheduling path. If a higher class task gets picked,
> put_prev_task_scx() will be called. Otherwise, we break the wait when
> pick_task_scx() is entered.

Yeah, that sounds reasonable to me.

Thanks,
-Andrea