Re: Quota-enabled XFS hangs during mount

From: Dave Chinner <david@fromorbit.com>
To: Martin Svec <martin.svec@zoner.cz>
Cc: linux-xfs@vger.kernel.org
Subject: Re: Quota-enabled XFS hangs during mount
Date: Fri, 4 Nov 2016 07:40:49 +1100	[thread overview]
Message-ID: <20161103204049.GA28177@dastard> (raw)
In-Reply-To: <7993e9b8-6eb8-6a0d-aa72-01346cca1b63@zoner.cz>

On Thu, Nov 03, 2016 at 01:04:39PM +0100, Martin Svec wrote:
> Dne 3.11.2016 v 2:31 Dave Chinner napsal(a):
> > On Wed, Nov 02, 2016 at 05:31:00PM +0100, Martin Svec wrote:
> >>> How many inodes? How much RAM?
> >> orthosie:~# df -i
> >> Filesystem        Inodes   IUsed     IFree IUse% Mounted on
> >> /dev/sdd1      173746096 5214637 168531459    4% /www
> >>
> >> The virtual machine has 2 virtual cores and 2 GB RAM. None of it is a bottleneck, I think.
> > Even though you think this is irrelevant and not important, it
> > actually points me directly at a potential vector and a reason as to
> > why this is not a comonly seen problem.
> >
> > i.e. 5.2 million inodes with only 2GB RAM is enough to cause memory
> > pressure during a quotacheck. inode buffers alone will require
> > a minimum of 1.5GB RAM over the course of the quotacheck, and memory
> > reclaim will iterate cached dquots and try to flush them, thereby
> > exercising the flush lock /before/ the quotacheck scan completion
> > dquot writeback tries to take it.
> >
> > Now I need to go read code....
> 
> Yes, that makes sense, I didn't know that quotacheck requires all inodes to be loaded in memory at
> the same time.

It doesn't require all inodes to be loaded into memory. Indeed, the
/inodes/ don't get cached during a quotacheck. What does get cached
is the metadata buffers that are traversed - th eproblem comes when
the caches are being reclaimed....

> I temporarily increased virtual machine's RAM to 3GB and the
> problem is gone! Setting RAM back to 2GB reproducibly causes the
> quotacheck to freeze again, 1GB of RAM results in OOM... So you're
> right that the flush deadlock is triggered by a memory pressure.

Good to know.

> Sorry, I forgot to attach sysrq-w output to the previous response. Here it is:

<snip>

ok, nothing else apparently stuck, so this is likely a leaked lock
or completion.

Ok, I'll keep looking.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com