Re: [PATCH] sched: Add logic to handle parallel try_to_wake_up() of the same task

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jul 15, 2013 at 06:14:34PM +0400, Kirill Tkhai wrote:
> >>
> >>   #ifdef CONFIG_SMP
> >>  + p->state = TASK_WAKING;
> >>  + smp_wmb();
> >>  +
> >
> > This too is broken; the loop below needs to be completed first,
> > otherwise we change p->state while the task is still on the CPU and it
> > might read the wrong p->state.
> 
> This place is below (on_rq && ttwu_remote) check, so the task
> either 'dequeued and on_cpu == 0'
> or it's in the middle of schedule() on arch, which wants unlocked
> context switch.
> 
> Nobody scheduler's probes p->state between prepare_lock_switch() and
> finish_lock_switch(). Archs with unlocked ctx switch (mips and ia64)
> don't change or probe state of previous process during context_switch.

It means its after deactivate_task(), but before context_switch(). It so
happens that
context_switch()->prepare_task_switch()->trace_sched_switch() inspects
p->state.

Even if this was not the case, touching a task that is 'life' on another
CPU is very _very_ bad practise.