Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <b2e18765bc22ea851c2293c15a8aa4c3cec0fde5.camel@kernel.org>
Subject: Re: [PATCH v3 1/3] ext4: don't increase iversion counter for
 ea_inodes
From:   Jeff Layton <jlayton@kernel.org>
To:     Lukas Czerner <lczerner@redhat.com>, linux-ext4@vger.kernel.org
Cc:     tytso@mit.edu, jack@suse.cz, linux-fsdevel@vger.kernel.org,
        ebiggers@kernel.org, david@fromorbit.com
Date:   Fri, 12 Aug 2022 14:42:36 -0400
In-Reply-To: <20220812123727.46397-1-lczerner@redhat.com>
References: <20220812123727.46397-1-lczerner@redhat.com>
Content-Type: text/plain; charset="ISO-8859-15"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.44.4 (3.44.4-1.fc36) 
MIME-Version: 1.0
Precedence: bulk

On Fri, 2022-08-12 at 14:37 +0200, Lukas Czerner wrote:
> ea_inodes are using i_version for storing part of the reference count so
> we really need to leave it alone.
>=20
> The problem can be reproduced by xfstest ext4/026 when iversion is
> enabled. Fix it by not calling inode_inc_iversion() for EXT4_EA_INODE_FL
> inodes in ext4_mark_iloc_dirty().
>=20
> Signed-off-by: Lukas Czerner <lczerner@redhat.com>
> Reviewed-by: Jan Kara <jack@suse.cz>
> Reviewed-by: Jeff Layton <jlayton@kernel.org>
> ---
> v2, v3: no change
>=20
>  fs/ext4/inode.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
>=20
> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> index 601214453c3a..2a220be34caa 100644
> --- a/fs/ext4/inode.c
> +++ b/fs/ext4/inode.c
> @@ -5731,7 +5731,12 @@ int ext4_mark_iloc_dirty(handle_t *handle,
>  	}
>  	ext4_fc_track_inode(handle, inode);
> =20
> -	if (IS_I_VERSION(inode))
> +	/*
> +	 * ea_inodes are using i_version for storing reference count, don't
> +	 * mess with it
> +	 */
> +	if (IS_I_VERSION(inode) &&
> +	    !(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL))
>  		inode_inc_iversion(inode);
> =20
>  	/* the do_update_inode consumes one bh->b_count */


I've spent some time writing tests for the i_version counter (still
quite rough right now), and what I've found is that this particular
inode_inc_iversion results in the counter being bumped on _reads_ as
well as writes, due to the atime changing. This call to
inode_inc_iversion seems to make no sense, as we aren't bumping the
mtime here.

I'm still working on and testing this, but I think we'll probably just
want to remove this inode_inc_iversion entirely, and leave the i_version
bumping for normal files to happen when the timestamps are updated. So
far, my testing seems to indicate that that does the right thing.

Hopefully I'll have some testcases + patches for this next week
sometime.

Cheers,
--=20
Jeff Layton <jlayton@kernel.org>