Re: yield() in i2c non-happy paths hits BUG under -rt patch

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2009-11-19 at 15:00 +0100, Jean Delvare wrote:

> > 	cond_resched();
> 
> Are you saying that most calls to yield() should be replaced with calls
> to cond_resched()?

No, depends on the reason yield() is used. Some cases can be replaced by
locking constructs, such as a condition variable.

> I admit I a little skeptical. While the description of cond_resched()
> ("latency reduction via explicit rescheduling in places that are safe")
> sounds promising, following the calls leads me to:
> 
> static inline int need_resched(void)
> {
> 	return unlikely(test_thread_flag(TIF_NEED_RESCHED));
> }
> 
> So apparently the condition for need_resched() to do anything is
> considered unlikely... suggesting that cond_resched() is a no-op in
> most cases? I don't quite get the point of moving away from sched()
> because it is a no-op, if we end up with a no-op under a different name.

TIF_NEED_RESCHED gets set by the scheduler whenever it decides current
needs to get preempted, its unlikely() because that reduces the code
impact of cond_resched() and similar in the case we don't schedule, if
we do schedule() a mis-predicted branch isn't going to be noticed on the
overhead of scheduling.

So there's a few cases,

1) PREEMPT=n
2) Voluntary preempt
3) PREEMPT=y


1) non of this has any effect, if the scheduler wants to reschedule a
task that's in the kernel, it'll have to wait until it gets back to
user-space.

2) uses cond_resched() and similar to have explicit preemption points,
so we don't need to wait as long as 1).

3) preempts directly when !preempt_count(), when IRQs are disabled, the
IPI that will accompany TIF_NEED_RESCHED will be delayed and
local_irq_enable()/restore() will effect a reschedule due to the pending
IPI. If preemption was disabled while the IPI hit nothing will happen,
but preempt_enable() will do the reschedule once preempt_count reaches
0.