Re: [RFC PATCH] sched: Introduce per-mm/cpu concurrency id state

[Mathieu Desnoyers]

On 2023-03-31 04:38, Aaron Lu wrote:
> On Thu, Mar 30, 2023 at 07:09:11PM -0400, Mathieu Desnoyers wrote:
>
> >   void sched_mm_cid_exit_signals(struct task_struct *t)
> >   {
> >   	struct mm_struct *mm = t->mm;
> > -	unsigned long flags;
> > +	struct rq *rq = this_rq();
>
> Got many below messages due to the above line:
>
> [   19.294089] BUG: using smp_processor_id() in preemptible [00000000] code: kworker/u449:0/1621
>
> > +	struct rq_flags rf;
> >   
> >   	if (!mm)
> >   		return;
> > -	local_irq_save(flags);
> > +	rq_lock_irqsave(rq, &rf);
> >   	mm_cid_put(mm, t->mm_cid);
> >   	t->mm_cid = -1;
> >   	t->mm_cid_active = 0;
> > -	local_irq_restore(flags);
> > +	rq_unlock_irqrestore(rq, &rf);
> >   }
> >   
> >   void sched_mm_cid_before_execve(struct task_struct *t)
> >   {
> >   	struct mm_struct *mm = t->mm;
> > -	unsigned long flags;
> > +	struct rq *rq = this_rq();
>
> Also here;
>
> > +	struct rq_flags rf;
> >   
> >   	if (!mm)
> >   		return;
> > -	local_irq_save(flags);
> > +	rq_lock_irqsave(rq, &rf);
> >   	mm_cid_put(mm, t->mm_cid);
> >   	t->mm_cid = -1;
> >   	t->mm_cid_active = 0;
> > -	local_irq_restore(flags);
> > +	rq_unlock_irqrestore(rq, &rf);
> >   }
> >   
> >   void sched_mm_cid_after_execve(struct task_struct *t)
> >   {
> >   	struct mm_struct *mm = t->mm;
> > -	unsigned long flags;
> > +	struct rq *rq = this_rq();
>
> And here.
>
> > +	struct rq_flags rf;
> >   
> >   	if (!mm)
> >   		return;
> > -	local_irq_save(flags);
> > +	rq_lock_irqsave(rq, &rf);
> >   	t->mm_cid = mm_cid_get(mm);
> >   	t->mm_cid_active = 1;
> > -	local_irq_restore(flags);
> > +	rq_unlock_irqrestore(rq, &rf);
> >   	rseq_set_notify_resume(t);
> >   }
>
> I used below diff to get rid of these messages without understanding the
> purpose of these functions:

I'll fold this fix into the next round, thanks!

Mathieu

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index f07b87d155bd..7194c29f3c91 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -11444,45 +11444,57 @@ void sched_mm_cid_migrate_to(struct rq *dst_rq, struct task_struct *t, int src_c
>   void sched_mm_cid_exit_signals(struct task_struct *t)
>   {
>   	struct mm_struct *mm = t->mm;
> -	struct rq *rq = this_rq();
>   	struct rq_flags rf;
> +	struct rq *rq;
>   
>   	if (!mm)
>   		return;
> +
> +	preempt_disable();
> +	rq = this_rq();
>   	rq_lock_irqsave(rq, &rf);
>   	mm_cid_put(mm, t->mm_cid);
>   	t->mm_cid = -1;
>   	t->mm_cid_active = 0;
>   	rq_unlock_irqrestore(rq, &rf);
> +	preempt_enable();
>   }
>   
>   void sched_mm_cid_before_execve(struct task_struct *t)
>   {
>   	struct mm_struct *mm = t->mm;
> -	struct rq *rq = this_rq();
>   	struct rq_flags rf;
> +	struct rq *rq;
>   
>   	if (!mm)
>   		return;
> +
> +	preempt_disable();
> +	rq = this_rq();
>   	rq_lock_irqsave(rq, &rf);
>   	mm_cid_put(mm, t->mm_cid);
>   	t->mm_cid = -1;
>   	t->mm_cid_active = 0;
>   	rq_unlock_irqrestore(rq, &rf);
> +	preempt_enable();
>   }
>   
>   void sched_mm_cid_after_execve(struct task_struct *t)
>   {
>   	struct mm_struct *mm = t->mm;
> -	struct rq *rq = this_rq();
>   	struct rq_flags rf;
> +	struct rq *rq;
>   
>   	if (!mm)
>   		return;
> +
> +	preempt_disable();
> +	rq = this_rq();
>   	rq_lock_irqsave(rq, &rf);
>   	t->mm_cid = mm_cid_get(mm);
>   	t->mm_cid_active = 1;
>   	rq_unlock_irqrestore(rq, &rf);
> +	preempt_enable();
>   	rseq_set_notify_resume(t);
>   }
>   

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com