dm: fix failure when empty flush's bi_sector points beyond the device end

An empty flush bio can have arbitrary bi_sector. The commit 2b1c6d7
introduced a regression that device mapper would fail an empty flush bio
with -EIO if the sector pointed beyond the end of the device.

The commit introduced an optimization, that optimization would pass
flushes to __split_and_process_bio and __split_and_process_bio is not
prepared to handle empty bios. Fix this bug by passing only non-empty
flushes to __split_and_process_bio - non-empty flushes must have valid
bi_sector. Empty bios will go through __send_empty_flush, as they did
before the optimization.

This problem can be reproduced by running the lvm2 test:
make check_local T=lvconvert-thin.sh LVM_TEST_PREFER_BRD=0

Signed-off-by: Mikulas Patocka <[email protected]>
Fixes: 2b1c6d7 ("dm: optimize REQ_PREFLUSH with data when using the linear target")
Reported-by: Zdenek Kabelac <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>

Author: Mikulas Patocka

drivers/md/dm.c

@@ -2005,7 +2005,7 @@ static void dm_split_and_process_bio(struct mapped_device *md,
 		 * linear target or multiple linear targets pointing to the same
 		 * device), we can send the flush with data directly to it.
 		 */
-		if (map->flush_bypasses_map) {
+		if (bio->bi_iter.bi_size && map->flush_bypasses_map) {
 			struct list_head *devices = dm_table_get_devices(map);
 			if (devices->next == devices->prev)
 				goto send_preflush_with_data;