Re: Latency: allowing resheduling while holding spin_locks

[george anzinger]

Nigel Gamble wrote:
> 
> On Sat, 13 Jan 2001, Roger Larsson wrote:
> > A rethinking of the rescheduling strategy...
> 
> Actually, I think you have more-or-less described how successful
> preemptible kernels have already been developed, given that your
> "sleeping spin locks" are really just sleeping mutexes (or binary
> semaphores).
> 
> 1.  Short critical regions are protected by spin_lock_irq().  The maximum
> value of "short" is therefore bounded by the maximum time we are happy
> to disable (local) interrupts - ideally ~100us.
> 
> 2.  Longer regions are protected by sleeping mutexes.
> 
> 3.  Algorithms are rearchitected until all of the highly contended locks
> are of type 1, and only low contention locks are of type 2.
> 
> This approach has the advantage that we don't need to use a no-preempt
> count, and test it on exit from every spinlock to see if a preempting
> interrupt that has caused a need_resched has occurred, since we won't
> see the interrupt until it's safe to do the preemptive resched.

I agree that this was true in days of yore.  But these days the irq
instructions introduce serialization points and, me thinks, may be much
more time consuming than the "++, --, if (false)" that a preemption
count implemtation introduces.  Could some one with a knowledge of the
hardware comment on this?

I am not suggesting that the "++, --, if (false)" is faster than an
interrupt, but that it is faster than cli, sti.  Of course we are
assuming that there is <stuff> between the cli and the sti as there is
between the ++ and the -- if (false).

George

> 
> Nigel Gamble                                    [e-mail]
> Mountain View, CA, USA.                         http://www.nrg.org/