dm snapshot: flush merged data before committing metadata

If the origin device has a volatile write-back cache and the following
events occur:

1: After finishing merge operation of one set of exceptions,
merge_callback() is invoked.
2: Update the metadata in COW device tracking the merge completion.
This update to COW device is flushed cleanly.
3: System crashes and the origin device's cache where the recent
merge was completed has not been flushed.

During the next cycle when we read the metadata from the COW device,
we will skip reading those metadata whose merge was completed in
step (1). This will lead to data loss/corruption.

To address this, flush the origin device post merge IO before
updating the metadata.

Cc: [e-mail]
Signed-off-by: Akilesh Kailash <[e-mail]>
Signed-off-by: Mike Snitzer <[e-mail]>

drivers/md/dm-snap.c

···
 	 * for them to be committed.
 	 */
 	struct bio_list bios_queued_during_merge;
+
+	/*
+	 * Flush data after merge.
+	 */
+	struct bio flush_bio;
 };
 
 /*
···
 
 static void error_bios(struct bio *bio);
 
+static int flush_data(struct dm_snapshot *s)
+{
+	struct bio *flush_bio = &s->flush_bio;
+
+	bio_reset(flush_bio);
+	bio_set_dev(flush_bio, s->origin->bdev);
+	flush_bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
+
+	return submit_bio_wait(flush_bio);
+}
+
 static void merge_callback(int read_err, unsigned long write_err, void *context)
 {
 	struct dm_snapshot *s = context;
···
 			DMERR("Read error: shutting down merge.");
 		else
 			DMERR("Write error: shutting down merge.");
+		goto shut;
+	}
+
+	if (flush_data(s) < 0) {
+		DMERR("Flush after merge failed: shutting down merge");
 		goto shut;
 	}
 
···
 	s->first_merging_chunk = 0;
 	s->num_merging_chunks = 0;
 	bio_list_init(&s->bios_queued_during_merge);
+	bio_init(&s->flush_bio, NULL, 0);
 
 	/* Allocate hash table for COW data */
 	if (init_hash_tables(s)) {
···
 	mempool_exit(&s->pending_pool);
 
 	dm_exception_store_destroy(s->store);
+
+	bio_uninit(&s->flush_bio);
 
 	dm_put_device(ti, s->cow);