direct_write_fallback(): on error revert the ->ki_pos update from buffered write

If we fail filemap_write_and_wait_range() on the range the buffered write went
into, we only report the "number of bytes which we direct-written", to quote
the comment in there.  Which is fine, but buffered write has already advanced
iocb->ki_pos, so we need to roll that back.  Otherwise we end up with e.g.
write(2) advancing position by more than the amount it reports having written.

Fixes: 182c25e9 "filemap: update ki_pos in generic_perform_write"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Message-Id: <20230827214518.GU3390869@ZenIV>
Signed-off-by: Christian Brauner <brauner@kernel.org>

fs/libfs.c

@@ -1903,6 +1903,7 @@ ssize_t direct_write_fallback(struct kiocb *iocb, struct iov_iter *iter,
 		 * We don't know how much we wrote, so just return the number of
 		 * bytes which were direct-written
 		 */
+		iocb->ki_pos -= buffered_written;
 		if (direct_written)
 			return direct_written;
 		return err;