Re: [PATCH 08/17] sched_ext: Block proxy donors across scheduler transitions

[Andrea Righi]

Hi Tejun,

On Sun, Aug 16, 2026 at 11:32:20AM -1000, Tejun Heo wrote:
> Hello,
> 
> On Sun, Aug 16, 2026 at 07:35:06PM +0200, Andrea Righi wrote:
> > @@ -11295,6 +11295,15 @@ sched_change_begin(struct task_struct *p, const struct sched_class *next_class,
> >  		flags |= DEQUEUE_NOCLOCK;
> >  	}
> >  
> > +	/*
> > +	 * Don't carry retained proxy state across scheduling class changes.
> > +	 * Compatible RT/DL PI transitions could preserve the session so that a
> > +	 * boosted donor continues proxy-executing its lock owner. Defining which
> > +	 * class transitions can safely retain that state is left for future work.
> > +	 */
> > +	if ((flags & DEQUEUE_CLASS) && next_class != p->sched_class)
> > +		sched_proxy_block_task(rq, p);
> > +
> 
> I suppose this prevents RT PI propagation through existing proxy execution
> chain regardless of sched class in use? That sounds pretty convoluted and
> overlapping anyway but John does this look okay to you?

rt_mutex PI propagation itself still takes place, what this prevents is carrying
the boosted scheduling context through an existing proxy chain when the boost
changes scheduling class. Same-class RT priority changes are preserved.

The reset was intended as a conservative way to avoid carrying retained state
into or out of sched_ext, but this also affects fair/RT/DL proxy transitions.

Maybe we can narrow the condition to transitions into or out of sched_ext, like:

#ifdef CONFIG_SCHED_CLASS_EXT
	/*
	 * Don't carry retained proxy state into or out of sched_ext. The BPF
	 * scheduler must explicitly admit blocked donors and cannot inherit a
	 * proxy session owned by another scheduling class.
  	 */
  	if ((flags & DEQUEUE_CLASS) &&
	    (p->sched_class == &ext_sched_class || next_class == &ext_sched_class))
		sched_proxy_block_task(rq, p);
#endif

This would preserve proxy execution across fair/RT/DL PI transitions while still
ending it before entering or leaving sched_ext. WDYT?

Thanks,
-Andrea