Re: sched: wakeup setting TIF_NEED_RESCHED too frequently

From: Peter Zijlstra
Date: Wed Apr 28 2021 - 04:30:47 EST


On Tue, Apr 27, 2021 at 02:16:50PM -0700, Jörn Engel wrote:
> Anyway, trying to find a cause, I noticed the following call chain:
> set_nr_if_polling()
> ttwu_queue_remote()
> ttwu_queue()
> try_to_wake_up()
> default_wake_function()
> curr->func()
> __wake_up_common()
> __wake_up_common_lock()
> __wake_up()
> wake_up()
>
> Call chain above is manually created from source code. Closest sample I
> caught with instrumentation is missing the leaf calls after
> try_to_wake_up():
> _raw_spin_unlock_irqrestore+0x1f/0x40
> try_to_wake_up+0x425/0x5e0
> wake_up_q+0x3f/0x80
> futex_wake+0x159/0x180
> do_futex+0xcd/0xba0
>
> Afaics, the result is us setting TIF_NEED_RESCHED on any wakeup, unless
> wake_list is already populated. Is that actually intentional? And is
> that useful for performance of latency? I think it isn't, but I am
> probably missing something here.

set_nr_if_polling() only sets TIF_NEED_RESCHED IFF TIF_POLLING_NRFLAG is
set. TIF_POLLING_NRFLAG is only set when idle and is then sufficient to
wake that cpu and reschedule.

If not TIF_POLLING_NRFLAG, then we send an IPI to deal with the
non-empty wake_list.