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

[Thomas Gleixner <tglx-hfZtesqFncYOwBW4kG4KsQ@public.gmane.org>]

On Thu, 19 Nov 2009, Jean Delvare wrote:
> > That still does not explain why yield() is necessary _between_ the
> > transaction attempts.
> 
> It is not _necessary_. We are just trying to be fair to other kernel
> threads, because bit-banging is expensive and this is the only case
> where we know we can tolerate a delay.
> 
> Just to clarify things... does (or did) yield() have anything to do
> with CONFIG_PREEMPT_VOLUNTARY?

No.
 
> > That code is fully preemptible, otherwise you could not call
> > yield().
> 
> How does one know what code is preemtible and what code is not? The
> rest of the i2c-algo-bit code should definitely _not_ be preemtible, as
> it is highly timing sensitive.

Code is preemptible when preempt_count() == 0 and interrupts are
enabled. spin_lock() implicitely disables preemption.
 
> > And as I said before nobody even noticed that the yield()
> > default implementation was changed to a NOOP by default in the
> > scheduler.
> 
> Well, I guess only people monitoring system latency would notice, as
> this is the only thing yield() was supposed to help with in the first
> place.
> 
> You say "NOOP by default", does this imply there is a way to change
> this?

There is a sysctl: sysctl_sched_compat_yield
 
> Was yield() turned into NOOP by design, or was it a bug?

By design. The semantics of yield and the fairness approach of CFS are
not really working well together. Also yield() for SCHED_OTHER is not
really specified.

Thanks,

	tglx