Date: Sat, 20 Dec 2014 06:51:33 +0000
From: Al Viro <viro@ZenIV.linux.org.uk>
To: Christoph Hellwig <hch@infradead.org>
Cc: Jan Kara <jack@suse.cz>, Omar Sandoval <osandov@osandov.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Trond Myklebust <trond.myklebust@primarydata.com>,
        David Sterba <dsterba@suse.cz>, linux-fsdevel@vger.kernel.org,
        linux-mm@kvack.org, linux-nfs@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH 2/8] swap: lock i_mutex for swap_writepage direct_IO
Message-ID: <20141220065133.GC22149@ZenIV.linux.org.uk>
References: <cover.1418618044.git.osandov@osandov.com>
 <a59510f4552a5d3557958cdb0ce1b23b3abfc75b.1418618044.git.osandov@osandov.com>
 <20141215162705.GA23887@quack.suse.cz>
 <20141215165615.GA19041@infradead.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
In-Reply-To: <20141215165615.GA19041@infradead.org>
Sender: linux-nfs-owner@vger.kernel.org

On Mon, Dec 15, 2014 at 08:56:15AM -0800, Christoph Hellwig wrote:
> On Mon, Dec 15, 2014 at 05:27:05PM +0100, Jan Kara wrote:
> > On Sun 14-12-14 21:26:56, Omar Sandoval wrote:
> > > The generic write code locks i_mutex for a direct_IO. Swap-over-NFS
> > > doesn't grab the mutex because nfs_direct_IO doesn't expect i_mutex to
> > > be held, but most direct_IO implementations do.
> >   I think you are speaking about direct IO writes only, aren't you? For DIO
> > reads we don't hold i_mutex AFAICS. And also for DIO writes we don't
> > necessarily hold i_mutex - see for example XFS which doesn't take i_mutex
> > for direct IO writes. It uses it's internal rwlock for this (see
> > xfs_file_dio_aio_write()). So I think this is just wrong.
> 
> The problem is that the use of ->direct_IO by the swap code is a gross
> layering violation.  ->direct_IO is a callback for the filesystem, and
> the swap code need to call ->read_iter instead of ->readpage and
> ->write_tier instead of ->direct_IO, and leave the locking to the
> filesystem.

The thing is, ->read_iter() and ->write_iter() might decide to fall back to 
buffered IO path.  XFS is unusual in that respect - there O_DIRECT ends up
with short write in such case.  Other filesystems, OTOH...