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

[Mike Galbraith <bitbucket@online.de>]

On Sat, 2013-07-13 at 19:45 +0400, Kirill Tkhai wrote: 
> This patch adds optimization of try_to_wake_up() function
> for cases when the system is doing parallel wake_up
> of the same task on the different cpus. Also it adds
> accounting the statistics of these situations.
> 
> We check the status of the task we want to wake up.
> If it is TASK_WAKING then the task is manipulated
> by try_to_wake_up() on another cpu. And after this
> check it will be a moment when the task is queued
> and his status is TASK_RUNNING. We just return
> earlier when we are sure the task will be TASK_RUNNING
> in the future (maybe right after the check). The profit is
> we don't loop while we are waiting the spinlock.

Hm, you're adding cycles to the common case to shave spin cycles in the
very rare case, then spending some to note that a collision happened.
What makes recording worth even 1 cycle?  What am I gonna do with the
knowledge that $dinkynum wakeups intersected at a some random task?

-Mike