Fix race when checking i_size on direct i/o read
authorSteven Whitehouse <swhiteho@redhat.com>
Fri, 24 Jan 2014 14:42:22 +0000 (14:42 +0000)
committerAl Viro <viro@zeniv.linux.org.uk>
Sun, 26 Jan 2014 13:26:42 +0000 (08:26 -0500)
So far I've had one ACK for this, and no other comments. So I think it
is probably time to send this via some suitable tree. I'm guessing that
the vfs tree would be the most appropriate route, but not sure that
there is one at the moment (don't see anything recent at kernel.org)
so in that case I think -mm is the "back up plan". Al, please let me
know if you will take this?

Steve.

---------------------

Following on from the "Re: [PATCH v3] vfs: fix a bug when we do some dio
reads with append dio writes" thread on linux-fsdevel, this patch is my
current version of the fix proposed as option (b) in that thread.

Removing the i_size test from the direct i/o read path at vfs level
means that filesystems now have to deal with requests which are beyond
i_size themselves. These I've divided into three sets:

 a) Those with "no op" ->direct_IO (9p, cifs, ceph)
These are obviously not going to be an issue

 b) Those with "home brew" ->direct_IO (nfs, fuse)
I've been told that NFS should not have any problem with the larger
i_size, however I've added an extra test to FUSE to duplicate the
original behaviour just to be on the safe side.

 c) Those using __blockdev_direct_IO()
These call through to ->get_block() which should deal with the EOF
condition correctly. I've verified that with GFS2 and I believe that
Zheng has verified it for ext4. I've also run the test on XFS and it
passes both before and after this change.

The part of the patch in filemap.c looks a lot larger than it really is
- there are only two lines of real change. The rest is just indentation
of the contained code.

There remains a test of i_size though, which was added for btrfs. It
doesn't cause the other filesystems a problem as the test is performed
after ->direct_IO has been called. It is possible that there is a race
that does matter to btrfs, however this patch doesn't change that, so
its still an overall improvement.

Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-by: Zheng Liu <gnehzuil.liu@gmail.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Dave Chinner <david@fromorbit.com>
Acked-by: Miklos Szeredi <miklos@szeredi.hu>
Cc: Chris Mason <clm@fb.com>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fs/fuse/file.c
mm/filemap.c

index 7e70506297bc4f6ca4c7afd61ca457627a9d1b3b..89fdfd1919af90a87cd8e3e607e13cb507d1f4a3 100644 (file)
@@ -2710,6 +2710,9 @@ fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
        inode = file->f_mapping->host;
        i_size = i_size_read(inode);
 
+       if ((rw == READ) && (offset > i_size))
+               return 0;
+
        /* optimization for short read */
        if (async_dio && rw != WRITE && offset + count > i_size) {
                if (offset >= i_size)
index b7749a92021c39b93907b94baf352f58f9db0359..01842867c9d27bbb42ffabaffcb7ed9596164e7b 100644 (file)
@@ -1428,30 +1428,28 @@ generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                if (!count)
                        goto out; /* skip atime */
                size = i_size_read(inode);
-               if (pos < size) {
-                       retval = filemap_write_and_wait_range(mapping, pos,
+               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);
-                       }
-                       if (retval > 0) {
-                               *ppos = pos + retval;
-                               count -= retval;
-                       }
+               if (!retval) {
+                       retval = mapping->a_ops->direct_IO(READ, iocb,
+                                                          iov, pos, nr_segs);
+               }
+               if (retval > 0) {
+                       *ppos = pos + retval;
+                       count -= retval;
+               }
 
-                       /*
-                        * Btrfs can have a short DIO read if we encounter
-                        * compressed extents, so if there was an error, or if
-                        * we've already read everything we wanted to, or if
-                        * there was a short read because we hit EOF, go ahead
-                        * and return.  Otherwise fallthrough to buffered io for
-                        * the rest of the read.
-                        */
-                       if (retval < 0 || !count || *ppos >= size) {
-                               file_accessed(filp);
-                               goto out;
-                       }
+               /*
+                * Btrfs can have a short DIO read if we encounter
+                * compressed extents, so if there was an error, or if
+                * we've already read everything we wanted to, or if
+                * there was a short read because we hit EOF, go ahead
+                * and return.  Otherwise fallthrough to buffered io for
+                * the rest of the read.
+                */
+               if (retval < 0 || !count || *ppos >= size) {
+                       file_accessed(filp);
+                       goto out;
                }
        }