Re: [PATCH 1/2] sched: Skip class callbacks with SCHED_FLAG_KEEP_PARAMS

[Andrea Righi <arighi@nvidia.com>]

Hi Peter,

On Thu, Jul 30, 2026 at 11:57:37AM +0200, Peter Zijlstra wrote:
> On Thu, Jul 30, 2026 at 12:59:37PM +0530, K Prateek Nayak wrote:
> > Hello Andrea,
> > 
> > On 7/30/2026 11:20 AM, Andrea Righi wrote:
> > > SCHED_FLAG_KEEP_PARAMS suppresses scheduler parameter and class changes,
> > > but __sched_setscheduler() can still set DEQUEUE_CLASS when the class
> > > computed from the requested policy differs from the task's current
> > > class.
> > > 
> > > This causes the switching_from(), switched_from(), switching_to() and
> > > switched_to() callbacks to run even though p->sched_class remains
> > > unchanged.
> > > 
> > > Set DEQUEUE_CLASS only when the sched class is allowed to change.
> > > 
> > > Fixes: 637b0682821b ("sched: Fold sched_class::switch{ing,ed}_{to,from}() into the change pattern")
> > > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> > > ---
> > >  kernel/sched/syscalls.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
> > > index b215b0ead9a60..903b47f5d0b74 100644
> > > --- a/kernel/sched/syscalls.c
> > > +++ b/kernel/sched/syscalls.c
> > > @@ -675,7 +675,8 @@ int __sched_setscheduler(struct task_struct *p,
> > >  	prev_class = p->sched_class;
> > >  	next_class = __setscheduler_class(policy, newprio);
> > >  
> > > -	if (prev_class != next_class)
> > > +	if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS) &&
> > > +	    prev_class != next_class)
> > >  		queue_flags |= DEQUEUE_CLASS;
> > 
> > It seems there also exists a SCHED_FLAG_KEEP_POLICY which should be used
> > to retain the policy. By a quick look at the syscall path,
> > SCHED_FLAG_KEEP_PARAMS should just do an equivalent of:
> > 
> >     get_params(p, &attr, 0)
> > 
> > and match the some properties of sched_attr with that of the task.
> > 
> > The task is still free to switch classes if the attr->sched_policy
> > differs given SCHED_FLAG_KEEP_PARAMS and SCHED_FLAG_KEEP_POLICY are two
> > distinct flags and there is a separate SCHED_FLAG_KEEP_ALL to combines
> > them both.
> > 
> > I'm having a sneaky suspicion that the bit in sched_change guard below
> > should actually check for SCHED_FLAG_KEEP_ALL like:
> > 
> > diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
> > index b215b0ead9a6..baf1fc59069e 100644
> > --- a/kernel/sched/syscalls.c
> > +++ b/kernel/sched/syscalls.c
> > @@ -680,7 +680,7 @@ int __sched_setscheduler(struct task_struct *p,
> >  
> >  	scoped_guard (sched_change, p, queue_flags) {
> >  
> > -		if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
> > +		if ((attr->sched_flags & SCHED_FLAG_KEEP_ALL) != SCHED_FLAG_KEEP_ALL) {
> >  			__setscheduler_params(p, attr);
> >  			p->sched_class = next_class;
> >  			p->prio = newprio;
> 
> Hmm, but if we have KEEP_PARAMS then params are loaded with the current
> set and actually re-setting them doesn't matter. So perhaps we should
> just remove that conditional entirely?
> 
> ---
> diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
> index b215b0ead9a6..1893938c76db 100644
> --- a/kernel/sched/syscalls.c
> +++ b/kernel/sched/syscalls.c
> @@ -680,12 +680,10 @@ int __sched_setscheduler(struct task_struct *p,
>  
>  	scoped_guard (sched_change, p, queue_flags) {
>  
> -		if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
> -			__setscheduler_params(p, attr);
> -			p->sched_class = next_class;
> -			p->prio = newprio;
> -			__setscheduler_dl_pi(newprio, policy, p, scope);
> -		}
> +		__setscheduler_params(p, attr);
> +		p->sched_class = next_class;
> +		p->prio = newprio;
> +		__setscheduler_dl_pi(newprio, policy, p, scope);
>  		__setscheduler_uclamp(p, attr);
>  
>  		if (scope->queued) {
> @@ -979,8 +977,11 @@ SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
>  	if (!p)
>  		return -ESRCH;
>  
> -	if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
> +	if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS) {
> +		if (attr.sched_policy != SETPARAM_POLICY && attr.sched_policy != p->policy)
> +			return -EINVAL;
>  		get_params(p, &attr, 0);
> +	}
>  
>  	return sched_setattr(p, &attr);
>  }

Rejecting a policy change when KEEP_PARAMS is set makes sense to me.

However, I don't think we can remove the conditional and reapply the values
returned by get_params(). That snapshot is taken before the rq lock is acquired,
so a concurrent scheduler update could change the parameters in between.

What about retaining the existing update conditional and rejecting KEEP_PARAMS
with policy != p->policy under the rq lock, after the existing policy recheck?

KEEP_POLICY will already have been resolved to the current policy there. This
should prevent both the spurious class callbacks and the deadline bandwidth
accounting issue at their source.

Thanks,
-Andrea