Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Tue, 16 May 2023 10:12:36 +1000
From:   Dave Chinner <david@fromorbit.com>
To:     Tycho Andersen <tycho@tycho.pizza>
Cc:     "Darrick J. Wong" <djwong@kernel.org>, linux-xfs@vger.kernel.org,
        linux-kernel@vger.kernel.org,
        Tycho Andersen <tandersen@netflix.com>,
        "Eric W. Biederman" <ebiederm@xmission.com>
Subject: Re: [PATCH] xfs: don't do inodgc work if task is exiting
Message-ID: <20230516001236.GJ2651828@dread.disaster.area>
References: <20230511151702.14704-1-tycho@tycho.pizza>
 <20230512000724.GH858799@frogsfrogsfrogs>
 <ZF5X/8RQSzGMn3aZ@tycho.pizza>
 <20230512225414.GE3223426@dread.disaster.area>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230512225414.GE3223426@dread.disaster.area>
Precedence: bulk

On Sat, May 13, 2023 at 08:54:14AM +1000, Dave Chinner wrote:
> On Fri, May 12, 2023 at 09:15:11AM -0600, Tycho Andersen wrote:
> I also recently found evidence of another AGI -> inode buffer -> AGF
> -> inode buffer deadlock issue debugging another hang from a machine
> making heavy use of O_TMPFILEs via OVL.  Without a kernel code to
> examine in intricate detail, I can't say if this is what you are
> seeing.
> 
> I also don't have a fix for that yet because, well, it's complex and
> I can really only accomodate one complex high priority bug at a time
> in my head and it hasn't been my highiest priority. As it is, I
> don't think the fix for it will be trivially backportable to 5.15,
> either....

Just FYI: Having paged this back into my brain looking at this
issue, it occurred to me yesterday that there is a relatively simple
fix for this issue: using ->iop_precommit() to move adding the inode
to the inode cluster buffer from xfs_trans_log_inode() to inside
xfs_trans_commit(). This ensures the inode cluster buffers don't get
locked until xfs_trans_commit() is called, as we already do with the
xfs_iunlink_item processing for unlinked list operations.

I've just written the patch to do this, but it also isn't a 5.15
backport candidate because the infrastructure and xfs_iunlink_item
processing wasn't introduced until 5.19....

-Dave.
-- 
Dave Chinner
david@fromorbit.com