iomap: iomap that extends beyond EOF should be marked dirty

When doing a direct IO that spans the current EOF, and there are
written blocks beyond EOF that extend beyond the current write, the
only metadata update that needs to be done is a file size extension.

However, we don't mark such iomaps as IOMAP_F_DIRTY to indicate that
there is IO completion metadata updates required, and hence we may
fail to correctly sync file size extensions made in IO completion
when O_DSYNC writes are being used and the hardware supports FUA.

Hence when setting IOMAP_F_DIRTY, we need to also take into account
whether the iomap spans the current EOF. If it does, then we need to
mark it dirty so that IO completion will call generic_write_sync()
to flush the inode size update to stable storage correctly.

Fixes: 3460cac ("iomap: Use FUA for pure data O_DSYNC DIO writes")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: removed the ext4 part; they'll handle it separately]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs/xfs/xfs_iomap.c

@@ -1049,6 +1049,13 @@ xfs_file_iomap_begin(
 	trace_xfs_iomap_alloc(ip, offset, length, XFS_DATA_FORK, &imap);
 
 out_finish:
+	/*
+	 * Writes that span EOF might trigger an IO size update on completion,
+	 * so consider them to be dirty for the purposes of O_DSYNC even if
+	 * there is no other metadata changes pending or have been made here.
+	 */
+	if ((flags & IOMAP_WRITE) && offset + length > i_size_read(inode))
+		iomap->flags |= IOMAP_F_DIRTY;
 	return xfs_bmbt_to_iomap(ip, iomap, &imap, shared);
 
 out_found:

include/linux/iomap.h

@@ -32,6 +32,8 @@ struct vm_fault;
  *
  * IOMAP_F_DIRTY indicates the inode has uncommitted metadata needed to access
  * written data and requires fdatasync to commit them to persistent storage.
+ * This needs to take into account metadata changes that *may* be made at IO
+ * completion, such as file size updates from direct IO.
  */
 #define IOMAP_F_NEW		0x01	/* blocks have been newly allocated */
 #define IOMAP_F_DIRTY		0x02	/* uncommitted metadata */