Re: What to do when... xfs_repair hangs?

[Sean Caron <scaron@xxxxxxxxx>]

Sorry, all, I was a little out-of-it on Friday afternoon, of course I had kicked off xfs_repair actually in the background with all output sent to a file, and I was just doing 'tail -f' on that file.

So I kill the 'tail -f' and jump back to the command line, it appears that xfs_repair segfaulted and died.

That line of text:

disconnected inode 1109099673,

was indeed the last thing that it printed before it crashed.

If I look in dmesg, I just see -

xfs_repair[6770]: segfault at 28 ip 000000000042307b sp 00007fffef61bad0 error 4 in xfs_repair[400000+72000]

and that's it.

I checked with 'df' and there's plenty of space everywhere; I don't see why it would have faulted out trying to connect something to lost+found.

Underlying storage should be good; this is basically a RAID 60 built on top of a bunch of JBODs with LSI SAS9200 cards. MD sees all strings as started and running OK; no problems getting the array assembled at all.

Since Dave is saying it's OK to try re-running xfs_repair; it'll just pick up where it left off; let me give it another pass and see if it manages to complete, or if it segfaults out again. I guess it it poops out a second time, maybe we'll just want to consider rebuilding the filesystem and restoring from our copies?

Thanks for the feedback,

Sean

On Fri, May 30, 2014 at 8:01 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:

On Fri, May 30, 2014 at 03:49:13PM -0400, Sean Caron wrote:

> Hi all,
>
> Long story short, we have a big array formatted as XFS, we had a machine go
> down hard maybe a month, month and a half ago... when it came back up, XFS
> faulted out when we attempted to mount the filesystem; it complained the
> log was bad or something... I did a dry run of xfs_repair (-L) and it
> looked pretty bad, so we mounted up the filesystem read-only, ran a
> backup... I think we got pretty much everything out OK except maybe files
> that were open at the time of the crash.
>
> Now with a backup in hand, we kicked off xfs_repair "for real"... it ran
> for a while and did its thing, but now it appears to be stuck at the stage -
>
> - agno = 436
> rebuilding directory inode ...
> rebuilding directory inode ...
> rebuilding directory inode ...
> ...
> - traversal finished ...
> - moving disconected inodes to lost+found ...
> disconnected inode 1109099673,
>
> and then it just stops. I don't know how long its been sitting like that,
> but it hasn't moved in the last hour or two. I assume that's not good...

Is that the total of the last line of output? If so, it's likely
stuck creating the lost+found directory. It's possible there's a
corruption in the inode AVL tree (e.g. endless loop) that is causing
it to spin doing an inode record lookup, but otherwise I can't see
any reason for it getting stuck here.

The information that Brian asked for will be a good start in
tracking this down, as will the complete output of xfs_repair...

> Interestingly when we ran a dry run of xfs_repair (-L) it got all the way
> through; it never hung up at any point. Not sure why it would start to hang
> up, once it gets run "for real".

That's because a dry-run skips the "move to lost_found" phase.

> This machine is in single-user-mode, I have exactly 24 lines of console
> with no scrollback buffer, no other tty available besides that which I'm
> running xfs_repair on, the system console.

$ man script

or

$ man tee

> Running Linux kernel 3.4.61, Ubuntu 12.04 LTS 64-bit with whatever their
> current xfsprogs is.

Upgrading xfsprogs to 3.2.0 would be a good idea.

> This is a bit of an exceptional situation for me; I've never seen
> xfs_repair just hang outright. I hoped I could maybe get some feedback from
> the experts here... what should I do?
>
> Try to Control-C out of the xfs_repair and ... re-run it?

That's fine - the next time repair runs it will start again and
repair anything that wasn't repaired in the last run.

Cheers,

Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx