From: Zheng Liu <gnehzuil.liu@gmail.com>
To: Jan Kara <jack@suse.cz>, Dave Chinner <david@fromorbit.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>,
linux-fsdevel@vger.kernel.org,
Christoph Hellwig <hch@infradead.org>,
Dmitry Monakhov <dmonakhov@openvz.org>,
Alexander Viro <viro@zeniv.linux.org.uk>,
Zheng Liu <wenqing.lz@taobao.com>,
Lukas Czerner <lczerner@redhat.com>
Subject: Re: [PATCH v3] vfs: fix a bug when we do some dio reads with append dio writes
Date: Wed, 15 Jan 2014 15:19:33 +0800 [thread overview]
Message-ID: <20140115071933.GA3449@gmail.com> (raw)
In-Reply-To: <20140114191901.GC27863@quack.suse.cz>
Hi all,
Sorry for the delay because of holiday and other issues.
On Tue, Jan 14, 2014 at 08:19:01PM +0100, Jan Kara wrote:
[...]
> > Sorry for the delay - I have been working on this and I've now figured
> > out whats really going on here. I'll describe the GFS2 case first, but
> > it seems that it is something that will affect other fs too, including
> > local fs most likely.
> >
> > So we have one thread doing writes to a file, 1M at a time, using
> > O_DIRECT. Now with GFS2 that results in more or less a fallback to a
> > buffered write - there is one important difference however, and that is
> > that the direct i/o page invalidations still occur, just as if it were a
> > direct i/o write.
> >
> > Thread two is doing direct i/o reads. When this results in calling
> > ->direct_IO, all is well. GFS2 does all the correct locking to ensure
> > that this is coherent with the buffered writes. There is a problem
> > though - in generic_file_aio_read() we have:
> >
> > /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
> > if (filp->f_flags & O_DIRECT) {
> > loff_t size;
> > struct address_space *mapping;
> > struct inode *inode;
> >
> > mapping = filp->f_mapping;
> > inode = mapping->host;
> > if (!count)
> > goto out; /* skip atime */
> > size = i_size_read(inode);
> > if (pos < size) {
> > retval = filemap_write_and_wait_range(mapping, pos,
> > pos + iov_length(iov, nr_segs) - 1);
> > if (!retval) {
> > retval = mapping->a_ops->direct_IO(READ, iocb,
> > iov, pos, nr_segs);
> > }
> >
> > The important thing here is that it is checking the i_size with no
> > locking! This means that it can check i_size, find its the same as the
> > request read position (if the read gets in first before the next block
> > is written to the file) and then fall back to buffered read.
> >
> > That buffered read then gets a page which is invalidated by the page
> > invalidation after the (DIO, but fallen back to buffered) write has
> > occurred. I added some trace_printk() calls to the read path and it is
> > fairly clear that is the path thats being taken when the failure occurs.
> > Also, it explains why, despite the disk having been used and reused many
> > times, the page that is returned from the failure is always 0, so its
> > not a page thats been read from disk. It also explains why I've only
> > ever seen this for the first page of the read, and the rest of the read
> > is ok.
> Yes, race like this is something Zheng and I have been talking about from
> the beginning. Thanks for tracking down the details.
>
> > Changing the test "if (pos < size) {" to "if (1) {" results in zero
> > failures from the test program, since it takes the ->direct_IO path each
> > time which then checks the i_size under the glock, in a race free
> > manner.
> >
> > So far as I can tell, this can be just as much a problem for local
> > filesystems, since they may update i_size at any time too. I note that
> > XFS appears to wrap this bit of code under its iolock, so that probably
> > explains why XFS doesn't suffer from this issue.
> >
> > So what is the correct fix... we have several options I think:
> >
> > a) Ignore it on the basis that its not important for normal workloads
> Zheng has a real world usecase which hits this race and it does look
> rather normal. So don't like a) very much.
Yes, as far as I know, at least there are two applications that has met
this problem in our product system at Taobao. So that is why I dig into
this bug and hope it can be fixed.
>
> > b) Simply remove the "if (pos < size) {" test and leave the fs to get
> > on with it (I know that leaves the btrfs test which also uses i_size
> > later on, but that doesn't appear to be an issue since its after we've
> > called ->direct_IO). Is it an issue that this may call an extra
> > filemap_write_and_wait_range() in some cases where it didn't before? I
> > suspect not since the extra calls would be for cases where the pages
> > were beyond the end of file anyway (according to i_size), so a no-op in
> > most cases.
I try this approach and I can confirm that it is ok for ext4.
Dave, I have send out a patch for xfstests to add a new testcase
according to this bug but I don't get any response [1]. Could you
please review it? Thanks.
1. http://comments.gmane.org/gmane.comp.file-systems.xfs.general/58428
Regards,
- Zheng
next prev parent reply other threads:[~2014-01-15 7:15 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2013-12-07 10:55 [PATCH v3] vfs: fix a bug when we do some dio reads with append dio writes Zheng Liu
2013-12-16 9:37 ` Steven Whitehouse
2013-12-16 10:01 ` Jan Kara
2013-12-17 9:43 ` Steven Whitehouse
2013-12-17 11:16 ` Zheng Liu
2013-12-17 12:17 ` Steven Whitehouse
2013-12-17 16:41 ` Zheng Liu
2013-12-19 12:27 ` Steven Whitehouse
2013-12-19 22:44 ` Dave Chinner
2013-12-20 9:28 ` Steven Whitehouse
2013-12-23 3:00 ` Dave Chinner
2014-01-14 15:22 ` Steven Whitehouse
2014-01-14 19:19 ` Jan Kara
2014-01-15 7:19 ` Zheng Liu [this message]
2014-01-16 15:35 ` [RFC] [PATCH] Fix race when checking i_size on direct i/o read Steven Whitehouse
2014-01-17 10:20 ` Miklos Szeredi
2014-01-24 14:42 ` Steven Whitehouse
2014-01-27 10:13 ` Jan Kara
2013-12-17 14:01 ` [PATCH v3] vfs: fix a bug when we do some dio reads with append dio writes Jan Kara
2013-12-17 15:32 ` Steven Whitehouse
2013-12-17 16:39 ` Jan Kara
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20140115071933.GA3449@gmail.com \
--to=gnehzuil.liu@gmail.com \
--cc=david@fromorbit.com \
--cc=dmonakhov@openvz.org \
--cc=hch@infradead.org \
--cc=jack@suse.cz \
--cc=lczerner@redhat.com \
--cc=linux-fsdevel@vger.kernel.org \
--cc=swhiteho@redhat.com \
--cc=viro@zeniv.linux.org.uk \
--cc=wenqing.lz@taobao.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).