Re: [tip:x86/asm] x86: Speed up ___preempt_schedule*() by using THUNK helpers

[Oleg Nesterov]

On 10/03, Linus Torvalds wrote:
>
> On Fri, Oct 3, 2014 at 4:26 PM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> >
> > And I _think_ that preempt_schedule_context() should be fixed anyway,
> > although I am not sure there is no something else. It does:
> >
> >
> >         preempt_disable_notrace();
> >         prev_ctx = exception_enter();
> >         preempt_enable_no_resched_notrace();
> >
> >         preempt_schedule();
> >
> >         preempt_disable_notrace();
> >         exception_exit(prev_ctx);
> >         preempt_enable_notrace();
> >
> > but exception_exit() is heavy, it is quite possible that TIF_NEED_RESCHED
> > and thus set_preempt_need_resched() can be set again when we call
> > preempt_enable_notrace(). And in this case preempt_schedule_context()
> > will be called recursively.
>
> Why the hell is it using "preempt_enable_notrace()" in the first
> place? Shouldn't it use "preempt_enable_no_resched_notrace()",

Yes, this this is the main problem.

> > Frederic, how about the patch below?
>
> Why do it multiple times? The whole concept is fundamentally racy
> anyway, in it doesn't guarantee that any *new* "need_resched()" would
> be reacted to (since they could happen *after* the test),

But in this case we rely on scheduler_ipi() and return-from-irq path?

> The real fix would appear to be to use
> "preempt_enable_no_resched_notrace()", which your patch did, but
> without the loop.

Not sure... preempt_schedule() does the same and afaics for good reason.

But perhaps you are right. I am already sleeping, will try to recheck
tomorrow. And in fact I got lost in preempt.h files... I can't even
understand why __preempt_schedule_context() is only called by
preempt_enable_notrace().

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/