Re: [RFC PATCH 6/9] sched: Treat try_to_block_task with pending signal as wakeup

[Gabriele Monaco <gmonaco@redhat.com>]

On Wed, 2025-04-16 at 11:20 +0200, Nam Cao wrote:
> On Fri, Apr 04, 2025 at 10:45:19AM +0200, Gabriele Monaco wrote:
> > If a task sets itself to interruptible and schedules, the
> > __schedule
> > function checks whether there's a pending signal and, if that's the
> > case, updates the state of the task to runnable instead of
> > dequeuing.
> > By looking at the tracepoints, we see the task enters the scheduler
> > while sleepable but exits as runnable. From a modelling
> > perspective,
> > this is equivalent to a wakeup and the tracepoints should reflect
> > that.
> > 
> > Add the waking/wakeup tracepoints in the try_to_block_task function
> > and
> > set the prev_state used by the sched_switch tracepoint to
> > TASK_RUNNING
> > if the task had a pending signal and was not blocked.
> > 
> > Signed-off-by: Gabriele Monaco <gmonaco@redhat.com>
> > ---
> >  kernel/sched/core.c | 11 +++++++++--
> >  1 file changed, 9 insertions(+), 2 deletions(-)
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index f2f79236d5811..48cb32abce01a 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -6584,7 +6584,12 @@ static bool try_to_block_task(struct rq *rq,
> > struct task_struct *p,
> >  	int flags = DEQUEUE_NOCLOCK;
> >  
> >  	if (signal_pending_state(task_state, p)) {
> > -		WRITE_ONCE(p->__state, TASK_RUNNING);
> > +		/*
> > +		 * From a modelling perspective, this is
> > equivalent to a wakeup
> > +		 * before dequeuing the task: trace accordingly.
> > +		 */
> > +		trace_sched_waking(p);
> > +		ttwu_do_wakeup(p);
> >  		return false;
> >  	}
> >  
> > @@ -6721,7 +6726,9 @@ static void __sched notrace __schedule(int
> > sched_mode)
> >  			goto picked;
> >  		}
> >  	} else if (!preempt && prev_state) {
> > -		try_to_block_task(rq, prev, prev_state);
> > +		/* Task was not blocked due to a signal and is
> > again runnable */
> > +		if (!try_to_block_task(rq, prev, prev_state))
> > +			prev_state = TASK_RUNNING;
> >  		switch_count = &prev->nvcsw;
> >  	}
> 
> I couldn't reproduce the problem that this patch is solving. But
> staring at
> the srs monitor, I made an educated guess that this is to accomodate
> the
> transition "sleepable x wakeup -> running"?
> 
> But for this transition, no real wakeup happens, just the task's
> state is
> changed to "sleepable" then back to "running", right? Sleep hasn't
> actually
> happened yet?
> 
> If that is the case, would the patch below also solves it? It would
> turn
> the transition into "sleepable x set_runnable -> running", which I
> think
> describe it more accurately.

Yeah that's pretty much it, there are a few problems though:
1. set_state should occur in task context and not while scheduling
2. set_state doesn't expect a task switch to occur

One way to solve this is to do like you said but add a flag to the
tracepoint to tell the model this set state is a special one happening
while scheduling, after that one, we may be scheduled out.

I didn't really like adding another state so I dropped that.

However, a task can be woken up before being scheduled out (I'd agree
with you it's not quite a wakeup as it wasn't yet sleeping, but it
happens, e.g. p == current in try_to_wake_up).
This case with the signal is, in that sense, a wakeup. We can even see
the tracepoint at times.

Anyway, that issue was mostly hypothetical, the patch also fixes the
prev_state (there's a patch by Peter on tip doing the same thing) and I
need to make sure it's really possible to see the issue after that too.

Thanks for looking into it,
Gabriele