Re: aio stress panic on 2.6.11-mm1

[Suparna Bhattacharya <suparna@in.ibm.com>]

On Wed, Mar 09, 2005 at 12:21:01PM +0100, Andi Kleen wrote:
> Arjan van de Ven <arjan@infradead.org> writes:
> 
> > On Wed, 2005-03-09 at 16:34 +0530, Suparna Bhattacharya wrote:
> >> Any sense of how costly it is to use spin_lock_irq's vs spin_lock
> >> (across different architectures) ? Isn't rwsem used very widely ?
>
> On P4s cli/sti is quite costly, let's says 100+ cycles. That is mostly
> because it synchronizes the CPU partly. The Intel tables say 26/36 cycles
> latency, but in practice it seems to take longer because of the
> synchronization.
>
> I would assume this is the worst case, everywhere else it should
> be cheaper (except perhaps in some virtualized environments)
> On P-M and AMD K7/K8 it is only a few cycles difference.
> >
> > oh also rwsems aren't used all that much simply because they are quite
> > more expensive than regular semaphores, so that you need a HUGE bias in
> > reader/writer ratio to make it even worth it...
> 
> I agree. I think in fact once Christopher L's lockless page fault fast path
> goes in it would be a good idea to reevaluate if mmap_sem should
> be really a rwsem and possibly change it back to a normal semaphore
> that perhaps gets dropped only on a page cache miss.


OK - makes sense. Thanks !

Regards
Suparna

-- 
Suparna Bhattacharya (suparna@in.ibm.com)
Linux Technology Center
IBM Software Lab, India