Re: preempt timing violations

[Lee Revell <rlrevell@joe-job.com>]

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