lkml.org 
[lkml]   [2004]   [Jul]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: preempt timing violations
From
Date
On Mon, 2004-07-26 at 04:50, Ingo Molnar wrote:
> * Lee Revell <rlrevell@joe-job.com> wrote:
>
> > Latency with 2.6.8-rc2 + voluntary-preempt-I4 is the best so far.
> > After extended testing there only seem to be a few hot spots. In
> > several minutes I saw an 11ms violation, a 14ms violation, and several
> > 2ms violations.
> >
> > get_user_pages() is much better in 2.6.8-rc1-mm1 than 2.6.8-rc2. Is
> > there any chance of getting the fix into mainline?
>
> in -J3 i've added a cond_resched to the latency-generating point of
> get_user_pages(). (The biggest latencies happen via
> make_pages_present(), which gets triggered by mlockall()/MAP_LOCKED.)
>

OK, great, this is the biggest remaing issue with -I4, because jackd
does that a lot.

> > 2ms non-preemptible critical section violated 1 ms preempt threshold
> > starting at unmap_vmas+0x1ff/0x210 and ending at
> > unmap_vmas+0x1f5/0x210
>
> this is the normal sys_exit()->exit_mmap()->unmap_vmas() path. It's
> weird that it generated a 2ms latency. What are the values of
> voluntary_preemption and kernel_preemption on your current kernel? With
> a 2:0 setting we ought to have a reschedule point every 32 pages. Do
> you know which application triggers this latency and is it easy to
> reproduce?
>

2 and 1. Now that I think about it, this could have happened during
bootup, before my rc.local set these. I will try passing them on the
kernel command line.

Not sure I understand the difference between 2:1 and 2:0. Would the
latter make the kernel only preemptible at the voluntary preemption
points?

> > 14ms non-preemptible critical section violated 1 ms preempt threshold
> > starting at tty_write+0x1b6/0x290 and ending at schedule+0x2fd/0x5b0
>
> does this one trigger when you are using the VGA console? (or fbcon)?
>
> it's not immediately obvious to me precisely where this latency comes
> from, it would be nice to know how to reproduce it.

It think this one was caused by switching virtual consoles. At one
point Andrew Morton suggested I remove the (un)lock_kernel calls from
do_tty_write. This fixed the problem, with no detectable side effects.
Maybe this could be incorporated into voluntary-preempt, it would be
useful to have more than one person to test it.

Lee

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

\
 
 \ /
  Last update: 2005-03-22 14:04    [from the cache]
©2003-2011 Jasper Spaans