Re: [PATCH] sched: correct testing need_resched in mutex_spin_on_owner()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2011-06-07 at 22:47 +0800, Hillf Danton wrote:
> Oh, it looks you are willing to rethink about testing need_resched?

Dude, however did you come up with that deduction?

mutex_spin_on_owner() does two things:

 - it validates that owner is in fact still running (if so it must be on
another cpu, since we're running on this one).

 - it ensures we play nice and reschedule when we need to, so we don't
hog our cpu.

Testing TIF_NEED_RESCHED on owner like you propose is wrong because:

 - owner is not a stable pointer you can dereference, see
owner_running(), you first need to validate that its still a valid
pointer and then keep it valid while dereferencing it.

 - if owner were to reschedule, it would leave the cpu and we'd break
out of the loop anyway by means of owner_running() failing, so its
superfluous.

Please, get a grip on reality and stop sending endless streams of
patches based on wrong assumptions and mis-understandings.