Re: [patch] Latency Tracer, voluntary-preempt-2.6.8-rc4-O6

[Lee Revell]

On Thu, 2004-08-19 at 18:32, Lee Revell wrote:
> On Thu, 2004-08-19 at 15:30, Theodore Ts'o wrote:
> > On Thu, Aug 19, 2004 at 07:19:58AM -0400, Lee Revell wrote:
> > > > I doubt SHA_CODE_SIZE will make a sufficient difference to avoid the
> > > > latency problems.  What we would need to do is to change the code so
> > > > that the rekey operation in __check_and_rekey takes place in a
> > > > workqueue.  Say, something like this (warning, I haven't tested this
> > > > patch; if it breaks, you get to keep both pieces):
> > > > 
> > > 
> > > Tested, works for me.  This should probably be pushed upstream, as well
> > > as added to -P5, correct?  Is there any disadvantage to doing it this
> > > way?
> > 
> > Great, I will be pushing this upstream very shortly.
> > 
> 
> Hmm, turns out that this does not fix the problem:
> 
> http://krustophenia.net/testresults.php?dataset=2.6.8.1-P4#/var/www/2.6.8.1-P4/extract_entropy_latency_trace.txt
> 

Looking at the code, it's obvious why this doesn't fix the problem - you
are still doing 380 usecs (with SHA_CODE_SIZE 3) to 780 usecs (with
SHA_CODE_SIZE 0) worth of work inside the spinlock, it is just being
deferred via a workqueue, rather than executing right away.  So this
improves the average case, but does not help the worst case scenario at
all.  380 usecs is too long to hold a spinlock.

Lee





-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/