Re: Quota-enabled XFS hangs during mount

[Martin Svec <martin.svec@zoner.cz>]

Hello,

Dne 25.1.2017 v 23:17 Brian Foster napsal(a):
> On Tue, Jan 24, 2017 at 02:17:36PM +0100, Martin Svec wrote:
>> Hello,
>>
>> Dne 23.1.2017 v 14:44 Brian Foster napsal(a):
>>> On Mon, Jan 23, 2017 at 10:44:20AM +0100, Martin Svec wrote:
>>>> Hello Dave,
>>>>
>>>> Any updates on this? It's a bit annoying to workaround the bug by increasing RAM just because of the
>>>> initial quotacheck.
>>>>
>>> Note that Dave is away on a bit of an extended vacation[1]. It looks
>>> like he was in the process of fishing through the code to spot any
>>> potential problems related to quotacheck+reclaim. I see you've cc'd him
>>> directly so we'll see if we get a response wrt to if he got anywhere
>>> with that...
>>>
>>> Skimming back through this thread, it looks like we have an issue where
>>> quota check is not quite reliable in the event of reclaim, and you
>>> appear to be reproducing this due to a probably unique combination of
>>> large inode count and low memory.
>>>
>>> Is my understanding correct that you've reproduced this on more recent
>>> kernels than the original report? 
>> Yes, I repeated the tests using 4.9.3 kernel on another VM where we hit this issue.
>>
>> Configuration:
>> * vSphere 5.5 virtual machine, 2 vCPUs, virtual disks residing on iSCSI VMFS datastore
>> * Debian Jessie 64 bit webserver, vanilla kernel 4.9.3
>> * 180 GB XFS data disk mounted as /www
>>
>> Quotacheck behavior depends on assigned RAM:
>> * 2 or less GiB: mount /www leads to a storm of OOM kills including shell, ttys etc., so the system
>> becomes unusable.
>> * 3 GiB: mount /www task hangs in the same way as I reported in earlier in this thread.
>> * 4 or more GiB: mount /www succeeds.
>>
> I was able to reproduce the quotacheck OOM situation on latest kernels.
> This problem actually looks like a regression as of commit 17c12bcd3
> ("xfs: when replaying bmap operations, don't let unlinked inodes get
> reaped"), but I don't think that patch is the core problem. That patch
> pulled up setting MS_ACTIVE on the superblock from after XFS runs
> quotacheck to before it (for other reasons), which has a side effect of
> causing inodes to be placed onto the lru once they are released. Before
> this change, all inodes were immediately marked for reclaim once
> released from quotacheck because the superblock had not been set active.
>
> The problem here is first that quotacheck issues a bulkstat and thus
> grabs and releases every inode in the fs. The quotacheck occurs at mount
> time, which means we still hold the s_umount lock and thus the shrinker
> cannot run even though it is registered. Therefore, we basically just
> populate the lru until we've consumed too much memory and blow up.
>
> I think the solution here is to preserve the quotacheck behavior prior
> to commit 17c12bcd3 via something like the following:
>
> --- a/fs/xfs/xfs_qm.c
> +++ b/fs/xfs/xfs_qm.c
> @@ -1177,7 +1177,7 @@ xfs_qm_dqusage_adjust(
>  	 * the case in all other instances. It's OK that we do this because
>  	 * quotacheck is done only at mount time.
>  	 */
> -	error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip);
> +	error = xfs_iget(mp, NULL, ino, XFS_IGET_DONTCACHE, XFS_ILOCK_EXCL, &ip);
>  	if (error) {
>  		*res = BULKSTAT_RV_NOTHING;
>  		return error;
>
> ... which allows quotacheck to run as normal in my quick tests. Could
> you try this on your more recent kernel tests and see whether you still
> reproduce any problems?

The above patch fixes OOM issues and reduces overall memory consumption during quotacheck. However,
it does not fix the original xfs_qm_flush_one() freezing. I'm still able to reproduce it with 1 GB
of RAM or lower. Tested with 4.9.5 kernel.

If it makes sense to you, I can rsync the whole filesystem to a new XFS volume and repeat the tests.
At least, that could tell us if the problem depends on a particular state of on-disk metadata
structures or it's a general property of the given filesystem tree.

Martin