Re: BTRFS hang with 3.16-rc5 (and also with 3.16-rc4)

[Liu Bo <bo.li.liu@oracle.com>]

On Tue, Jul 29, 2014 at 05:07:31PM +0200, Torbjørn wrote:
> On 07/29/2014 12:18 PM, Liu Bo wrote:
> >On Mon, Jul 28, 2014 at 01:11:19PM +0200, Torbjørn wrote:
> >>On 28. juli 2014 12:00, Liu Bo wrote:
> >><snip>
> >>>This seems to be incomplete(Looks like dmesg has reached its buffer size limit),
> >>>does /var/log/message have the whole stack info?
> >>>
> >>>thanks,
> >>>-liubo
> >>Hi,
> >>
> >>Complete log was over 40MB. I uploaded everything from boot until
> >>"blocked for 120 seconds" started to appear.
> >>If you want all the trailing log as well, let me know.
> >>
> >>https://gist.github.com/anonymous/7958d8917967f727f324
> >Sorry...still don't get why it's locked up, io_ctl_prepare_pages() has several
> >callers, and they are properly released from the code level.  And the warnings
> >printed in the log belong to other btrfs partitions, not the hanged btrfs one,
> >and we're still not able to know which one holds the free space cache inode page.
> >
> >Maybe we'd better resort to a bisect between 3.14 and 3.15(I know it'd be a lot
> >of time though).
> >
> >Here, doing rsync on compress=lzo full btrfs never hit that problem, shrug...
> >
> >thanks,
> >-liubo
> >
> >>--
> >>Torbjørn
> That's too bad.
> 
> My reproducer is not 100% guaranteed to trigger the hang, so doing a
> bisect might lead us to some innocent commit.
> I have run the rsync + snapshot job several times here now, and no hang.

Good news! I've reproduced it with my xfstests config, will dig into it closer.

thanks,
-liubo