Re: AIM7 40% regression with 2.6.26-rc1

[Ingo Molnar <mingo@elte.hu>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> 
> 
> On Wed, 7 May 2008, Ingo Molnar wrote:
> > 
> > it was removed by me in the course of this discussion:
> > 
> >    http://lkml.org/lkml/2008/1/2/58
> > 
> > the whole discussion started IIRC because !CONFIG_PREEMPT_BKL [the 
> > spinlock version] was broken for a longer period of time (it crashed 
> > trivially), because nobody apparently used it.
> 
> Hmm. I've generally used PREEMPT_NONE, and always thought PREEMPT_BKL 
> was the known-flaky one.
> 
> The thread you point to also says that it's PREEMPT_BKL=y that was the 
> problem (ie "I've seen 1s+ desktop latencies due to PREEMPT_BKL when I 
> was still using reiserfs."), not the plain spinlock approach.

no, there was another problem (which i couldnt immediately find because 
lkml.org only indexes part of the threads, i'll research it some more), 
which was some cond_resched() thing in the !PREEMPT_BKL case.

> But it would definitely be interesting to see the crash reports. And 
> the help message always said "Say N if you are unsure." even if it 
> ended up being marked 'y' by default at some point (and then in 
> January was made first unconditional, and then removed entirely)
> 
> Because in many ways, the non-preempt BKL is the *much* simpler case. 
> I don't see why it would crash - it just turns the BKL into a trivial 
> counting spinlock that can sleep.

yeah. The latencies are a different problem, and indeed were reported 
against PREEMPT_BKL, and believed to be due to reiser3 and the tty code. 
(reiser3 runs almost all of its code under the BKL)

The !PREEMPT_BKL crash was some simple screwup on my part of getting 
atomicity checks wrong in cond_resched() - and it went unnoticed for a 
long time - or something like that. I'll try to find that discussion.

	Ingo