Re: [Bug 18632] "INFO: task" dpkg "blocked for more than 120 seconds.

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, Jun 09, 2011 at 11:32:17AM +0800, Wu Fengguang wrote:
> On Thu, Jun 09, 2011 at 06:02:41AM +0800, Andrew Morton wrote:
> > (udate by email because bugzilla is broken)
> > 
> > On Wed, 8 Jun 2011 21:38:43 GMT
> > bugzilla-daemon@bugzilla.kernel.org wrote:
> > 
> > > https://bugzilla.kernel.org/show_bug.cgi?id=18632
> > 
> > I always get a bit worried when I see dm-crypt involved - it doesn't
> > seem terribly well-behaved or reliable.
> > 
> > That being said, some of the other reports here indicate that we are
> > hitting a sync() livelock, sometimes fsync().
> 
> These comments all mention "blocked sync(1) on concurrent heavy writes".
> 
>         Comment #6 From Douglas Augusto
>         Comment #12 From Nathan Grennan 
>         Comment #16 From justincase@yopmail.com 

>         Comment #23 From Carlos Alberto Lopez Perez

Oh it's not sync(1) in Carlos' case, but random commands like man/dd
as well as

        task xfssyncd:451 blocked for more than 120 seconds.

It looks more like XFS related livelock, because I ran into similar
problem in a very simple XFS setup on some plain disk partition.

Thanks,
Fengguang

> Comment #17 From Justin Wheeler is special in that fio is doing
> read-only IO. Is it updating the mtime and the mtime writes are in
> turn starved by the higher-priority continuous reads? (or at least the
> write speed cannot keep up with the new mtime dirties)
> 
> > A pile of patches from
> > Wu Fengguang which could/should address these issues is about to hit
> > the linux-next tree.
> 
> Some livelock fixes will show up in tomorrow's linux-next tree. As for
> now the fixes are available here:
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/wfg/writeback.git fs-writeback
> 
> Thanks,
> Fengguang