Re: BUG: ext3 hang in transaction commit

[Marcelo Tosatti <mtosatti@redhat.com>]

On Fri, Apr 04, 2008 at 12:34:50PM +0200, Jan Kara wrote:

> >         Max throughput per process                      =   66393.62 KB/sec
> >         Avg throughput per process                      =   26518.38 KB/sec
> >         Min xfer                                        =   44424.00 KB
> > 
> > And this is when fsync becomes nasty.
>   Could you get for me movies from Seekwatcher (taken on the host) -
> http://oss.oracle.com/~mason/seekwatcher/
>   That should confirm my suspicion. Actually, if writing data in the bad
> order is really the problem than my rewrite of ordered mode in JBD can
> substantially help this workload (I can send you the patches if you dare
> to try something really experimental ;).

OK, will record the animation.

Will be pleased to try your patches. Where can they be found?

> > blktrace output shows that the maximum latency for a single write
> > request io complete is 1.5 seconds, which is similar to what is seen
> > under "writeback" mode.
> > 
> > I reduced hung_task_timeout_secs to 30 for this report, but vim and
> > rsyslogd have been seen hung up to 120 seconds.
> > 
> 
> <snip traces of processes waiting for commit>
> 
> > As Peter mentioned it eventually gets out of this state (after several
> > minutes) and fsync instances complete.
>   Yes, that is just a combined effect of *lots* of ordered data
> accumulated in one transaction (we don't limit amount of ordered data in
> a transaction in any way) and writing them out during transaction order
> in a random order.