xfs: trim IO to found COW extent limit

A bug existed in the XFS reflink code between v5.1 and v5.5 in which
the mapping for a COW IO was not trimmed to the mapping of the COW
extent that was found.  This resulted in a too-short copy, and
corruption of other files which shared the original extent.

(This happened only when extent size hints were set, which bypasses
delalloc and led to this code path.)

This was (inadvertently) fixed upstream with

36adcbace2 "xfs: fill out the srcmap in iomap_begin"

and related patches which moved lots of this functionality to
the iomap subsystem.

Hence, this is a -stable only patch, targeted to fix this
corruption vector without other major code changes.

Fixes: 78f0cc9d55 ("xfs: don't use delalloc extents for COW on files with extsize hints")
Cc: <stable@vger.kernel.org> # 5.4.x
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>

fs/xfs/xfs_iomap.c

@@ -1002,9 +1002,15 @@ xfs_file_iomap_begin(
 		 * I/O, which must be block aligned, we need to report the
 		 * newly allocated address.  If the data fork has a hole, copy
 		 * the COW fork mapping to avoid allocating to the data fork.
+		 *
+		 * Otherwise, ensure that the imap range does not extend past
+		 * the range allocated/found in cmap.
 		 */
 		if (directio || imap.br_startblock == HOLESTARTBLOCK)
 			imap = cmap;
+		else
+			xfs_trim_extent(&imap, cmap.br_startoff,
+					cmap.br_blockcount);
 
 		end_fsb = imap.br_startoff + imap.br_blockcount;
 		length = XFS_FSB_TO_B(mp, end_fsb) - offset;