lkml.org 
[lkml]   [2016]   [Jul]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: cleanup_net()/net_mutex hung tasks + kobject release debugging
Dave Jones <davej@codemonkey.org.uk> writes:

> On Thu, Jul 21, 2016 at 02:43:34PM +0200, Vegard Nossum wrote:
>
> > The rules for net_mutex are very simple, it's used in very few places so
> > I don't see how the locking could get messed up there. I'll buy your
> > theory that the lock is held for a long time if there are a lot of
> > namespaces to iterate over. I decided to time it myself and it seems
> > that cleanup_net() can hold the mutex for 30-40 seconds at a time, which
> > is surely wrong.
>
> > so on a hunch I disabled DEBUG_KOBJECT_RELEASE, and that does indeed
> > solve the problem -- cleanup_net() still holds the mutex for fairly
> > long, but only up to max ~5 seconds at a time as opposed to 30-40.
>
> Yeah, I never ran with that option enabled (it used to cause my testbox
> to not boot, and I never got around to debugging why). I thought five seconds
> was painful enough. I guess we have different thresholds for acceptable
> behaviour here :-)
>
> Could be one of the other debug options I had enabled exacerbates the
> cleanup_net problem in a similar way though.
>
> > There's maybe a case for cleanup_net() to release the mutex every now
> > and again during cleanup, but I was also seeing a few other hung tasks
> > unrelated to net_mutex when I disabled the unshare() system call in
> > trinity, which makes me wonder if we need a more general solution.
>
> Not sure. We may have to just look at these on a case by case basis.

The best you can easily do in cleanup_net with net_mutex is to reduce
the number of net namespaces you free at once. Which sounds attractive
except that last I looked most of the time was spent in syncrhonize_rcu.

Because the namespaces can share those synchronize_rcu calls cleaning up
a bunch of network namespaces all at once is actually a pretty big
optimization in terms of system performance.

Though if someone wants to dig in and point out non-shared
synchronize_rcu calls or other obvious sillies happening in cleanup_net
I will be happy to see what we can do.

Eric

\
 
 \ /
  Last update: 2016-07-30 16:01    [W:0.034 / U:0.812 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site