linux-ext4 - Re: [PATCH 1/1] quota: flush quota_release

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20241118125344.a3n3kn6crvrixglb@quack3>
Date: Mon, 18 Nov 2024 13:53:44 +0100
From: Jan Kara <jack@...e.cz>
To: Baokun Li <libaokun1@...wei.com>
Cc: Ojaswin Mujoo <ojaswin@...ux.ibm.com>,
	Ritesh Harjani <ritesh.list@...il.com>, linux-ext4@...r.kernel.org,
	Jan Kara <jack@...e.com>, linux-kernel@...r.kernel.org,
	linux-fsdevel@...r.kernel.org, Disha Goel <disgoel@...ux.ibm.com>,
	Yang Erkun <yangerkun@...wei.com>
Subject: Re: [PATCH 1/1] quota: flush quota_release_work upon quota writeback

On Mon 18-11-24 09:29:19, Baokun Li wrote:
> On 2024/11/17 1:59, Ojaswin Mujoo wrote:
> > On Sat, Nov 16, 2024 at 02:20:26AM +0530, Ritesh Harjani wrote:
> > > Ojaswin Mujoo <ojaswin@...ux.ibm.com> writes:
> > > 
> > > > One of the paths quota writeback is called from is:
> > > > 
> > > > freeze_super()
> > > >    sync_filesystem()
> > > >      ext4_sync_fs()
> > > >        dquot_writeback_dquots()
> > > > 
> > > > Since we currently don't always flush the quota_release_work queue in
> > > > this path, we can end up with the following race:
> > > > 
> > > >   1. dquot are added to releasing_dquots list during regular operations.
> > > >   2. FS freeze starts, however, this does not flush the quota_release_work queue.
> > > >   3. Freeze completes.
> > > >   4. Kernel eventually tries to flush the workqueue while FS is frozen which
> > > >      hits a WARN_ON since transaction gets started during frozen state:
> > > > 
> > > >    ext4_journal_check_start+0x28/0x110 [ext4] (unreliable)
> > > >    __ext4_journal_start_sb+0x64/0x1c0 [ext4]
> > > >    ext4_release_dquot+0x90/0x1d0 [ext4]
> > > >    quota_release_workfn+0x43c/0x4d0
> > > > 
> > > > Which is the following line:
> > > > 
> > > >    WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > > 
> > > > Which ultimately results in generic/390 failing due to dmesg
> > > > noise. This was detected on powerpc machine 15 cores.
> > > > 
> > > > To avoid this, make sure to flush the workqueue during
> > > > dquot_writeback_dquots() so we dont have any pending workitems after
> > > > freeze.
> > > Not just that, sync_filesystem can also be called from other places and
> > > quota_release_workfn() could write out and and release the dquot
> > > structures if such are found during processing of releasing_dquots list.
> > > IIUC, this was earlier done in the same dqput() context but had races
> > > with dquot_mark_dquot_dirty(). Hence the final dqput() will now add the
> > > dquot structures to releasing_dquots list and will schedule a delayed
> > > workfn which will process the releasing_dquots list.
> > Hi Ritesh,
> > 
> > Ohh right, thanks for the context. I see this was done here:
> > 
> >    dabc8b207566 quota: fix dqput() to follow the guarantees dquot_srcu
> >    should provide

Yup.

> Nice catch! Thanks for fixing this up!
> 
> Have you tested the performance impact of this patch? It looks like the
> unconditional call to flush_delayed_work() in dquot_writeback_dquots()
> may have some performance impact for frequent chown/sync scenarios.

Well, but sync(2) or so is expensive anyway. Also dquot_writeback_dquots()
should persist all pending quota modifications and it is true that pending
dquot_release() calls can remove quota structures from the quota file and
thus are by definition pending modifications. So I agree with Ojaswin that
putting the workqueue flush there makes sense and is practically required
for data consistency guarantees.

> When calling release_dquot(), we will only remove the quota of an object
> (user/group/project) from disk if it is not quota-limited and does not
> use any inode or block.
> 
> Asynchronous removal is now much more performance friendly, not only does
> it make full use of the multi-core, but for scenarios where we have to
> repeatedly chown between two objects, delayed release avoids the need to
> repeatedly allocate/free space in memory and on disk.

True, but unless you call sync(2) in between these two calls this is going
to still hold.

> Overall, since the actual dirty data is already on the disk, there is no
> consistency issue here as it is just clearing unreferenced quota on the
> disk, so I thought maybe it would be better to call flush_delayed_work()
> in the freeze context.

To summarise, I don't think real-life workloads are going to observe the
benefit and conceptually the call really belongs more to
dquot_writeback_dquots().

								Honza
-- 
Jan Kara <jack@...e.com>
SUSE Labs, CR