Ext4 needs to serialize unaligned direct AIO because the zeroing of partial blocks of two competing unaligned AIOs can result in data corruption.

However it decides not to serialize if the potentially unaligned aio is past i_size with the rationale that no pending writes are possible past i_size. Unfortunately if the i_size is not block aligned and the second unaligned write lands past i_size, but still into the same block, it has the potential of corrupting the previous unaligned write to the same block.

This is (very simplified) reproducer from Frank

// 41472 = (10 * 4096) + 512 // 37376 = 41472 - 4096

ftruncate(fd, 41472); io_prep_pwrite(iocbs[0], fd, buf[0], 4096, 37376); io_prep_pwrite(iocbs[1], fd, buf[1], 4096, 41472);

io_submit(io_ctx, 1, &iocbs[1]); io_submit(io_ctx, 1, &iocbs[2]);

io_getevents(io_ctx, 2, 2, events, NULL);

Without this patch the 512B range from 40960 up to the start of the second unaligned write (41472) is going to be zeroed overwriting the data written by the first write. This is a data corruption.

00000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 00009200 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 * 0000a000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 0000a200 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31

With this patch the data corruption is avoided because we will recognize the unaligned_aio and wait for the unwritten extent conversion.

00000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 * 00009200 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 30 * 0000a200 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 31 * 0000b200

Reported-by: Frank Sorenson fsorenso@redhat.com Signed-off-by: Lukas Czerner lczerner@redhat.com Fixes: e9e3bcecf44c ("ext4: serialize unaligned asynchronous DIO") Cc: stable@vger.kernel.org --- fs/ext4/file.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/ext4/file.c b/fs/ext4/file.c index 69d65d49837b..98ec11f69cd4 100644 --- a/fs/ext4/file.c +++ b/fs/ext4/file.c @@ -125,7 +125,7 @@ ext4_unaligned_aio(struct inode *inode, struct iov_iter *from, loff_t pos) struct super_block *sb = inode->i_sb; int blockmask = sb->s_blocksize - 1;

- if (pos >= i_size_read(inode)) + if (pos >= ALIGN(i_size_read(inode), sb->s_blocksize)) return 0;

if ((pos | iov_iter_alignment(from)) & blockmask)