ext4: flush any pending end_io requests before DIO reads w/dioread_nolock
Jiaying Zhang [Fri, 19 Aug 2011 23:13:32 +0000 (19:13 -0400)]
There is a race between ext4 buffer write and direct_IO read with
dioread_nolock mount option enabled. The problem is that we clear
PageWriteback flag during end_io time but will do
uninitialized-to-initialized extent conversion later with dioread_nolock.
If an O_direct read request comes in during this period, ext4 will return
zero instead of the recently written data.

This patch checks whether there are any pending uninitialized-to-initialized
extent conversion requests before doing O_direct read to close the race.
Note that this is just a bandaid fix. The fundamental issue is that we
clear PageWriteback flag before we really complete an IO, which is
problem-prone. To fix the fundamental issue, we may need to implement an
extent tree cache that we can use to look up pending to-be-converted extents.

Signed-off-by: Jiaying Zhang <jiayingz@google.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org

fs/ext4/indirect.c

index b8602cd..0962642 100644 (file)
@@ -800,12 +800,17 @@ ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
        }
 
 retry:
-       if (rw == READ && ext4_should_dioread_nolock(inode))
+       if (rw == READ && ext4_should_dioread_nolock(inode)) {
+               if (unlikely(!list_empty(&ei->i_completed_io_list))) {
+                       mutex_lock(&inode->i_mutex);
+                       ext4_flush_completed_IO(inode);
+                       mutex_unlock(&inode->i_mutex);
+               }
                ret = __blockdev_direct_IO(rw, iocb, inode,
                                 inode->i_sb->s_bdev, iov,
                                 offset, nr_segs,
                                 ext4_get_block, NULL, NULL, 0);
-       else {
+       } else {
                ret = blockdev_direct_IO(rw, iocb, inode, iov,
                                 offset, nr_segs, ext4_get_block);