diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index db022e5..8a994be 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -125,6 +125,11 @@
 	spinlock_t deferred_lock;
 
 	/*
+	 * An error from the barrier request currently being processed.
+	 */
+	int barrier_error;
+
+	/*
 	 * Processing queue (flush/barriers)
 	 */
 	struct workqueue_struct *wq;
@@ -425,6 +430,10 @@
 	part_stat_add(cpu, &dm_disk(md)->part0, ticks[rw], duration);
 	part_stat_unlock();
 
+	/*
+	 * After this is decremented the bio must not be touched if it is
+	 * a barrier.
+	 */
 	dm_disk(md)->part0.in_flight = pending =
 		atomic_dec_return(&md->pending);
 
@@ -531,25 +540,35 @@
 			 */
 			spin_lock_irqsave(&md->deferred_lock, flags);
 			if (__noflush_suspending(md))
-				bio_list_add(&md->deferred, io->bio);
+				bio_list_add_head(&md->deferred, io->bio);
 			else
 				/* noflush suspend was interrupted. */
 				io->error = -EIO;
 			spin_unlock_irqrestore(&md->deferred_lock, flags);
 		}
 
-		end_io_acct(io);
-
 		io_error = io->error;
 		bio = io->bio;
 
-		free_io(md, io);
+		if (bio_barrier(bio)) {
+			/*
+			 * There can be just one barrier request so we use
+			 * a per-device variable for error reporting.
+			 * Note that you can't touch the bio after end_io_acct
+			 */
+			md->barrier_error = io_error;
+			end_io_acct(io);
+		} else {
+			end_io_acct(io);
 
-		if (io_error != DM_ENDIO_REQUEUE) {
-			trace_block_bio_complete(md->queue, bio);
+			if (io_error != DM_ENDIO_REQUEUE) {
+				trace_block_bio_complete(md->queue, bio);
 
-			bio_endio(bio, io_error);
+				bio_endio(bio, io_error);
+			}
 		}
+
+		free_io(md, io);
 	}
 }
 
@@ -691,7 +710,7 @@
 
 	clone->bi_sector = sector;
 	clone->bi_bdev = bio->bi_bdev;
-	clone->bi_rw = bio->bi_rw;
+	clone->bi_rw = bio->bi_rw & ~(1 << BIO_RW_BARRIER);
 	clone->bi_vcnt = 1;
 	clone->bi_size = to_bytes(len);
 	clone->bi_io_vec->bv_offset = offset;
@@ -718,6 +737,7 @@
 
 	clone = bio_alloc_bioset(GFP_NOIO, bio->bi_max_vecs, bs);
 	__bio_clone(clone, bio);
+	clone->bi_rw &= ~(1 << BIO_RW_BARRIER);
 	clone->bi_destructor = dm_bio_destructor;
 	clone->bi_sector = sector;
 	clone->bi_idx = idx;
@@ -846,7 +866,10 @@
 
 	ci.map = dm_get_table(md);
 	if (unlikely(!ci.map)) {
-		bio_io_error(bio);
+		if (!bio_barrier(bio))
+			bio_io_error(bio);
+		else
+			md->barrier_error = -EIO;
 		return;
 	}
 
@@ -930,15 +953,6 @@
 	struct mapped_device *md = q->queuedata;
 	int cpu;
 
-	/*
-	 * There is no use in forwarding any barrier request since we can't
-	 * guarantee it is (or can be) handled by the targets correctly.
-	 */
-	if (unlikely(bio_barrier(bio))) {
-		bio_endio(bio, -EOPNOTSUPP);
-		return 0;
-	}
-
 	down_read(&md->io_lock);
 
 	cpu = part_stat_lock();
@@ -950,7 +964,8 @@
 	 * If we're suspended or the thread is processing barriers
 	 * we have to queue this io for later.
 	 */
-	if (unlikely(test_bit(DMF_QUEUE_IO_TO_THREAD, &md->flags))) {
+	if (unlikely(test_bit(DMF_QUEUE_IO_TO_THREAD, &md->flags)) ||
+	    unlikely(bio_barrier(bio))) {
 		up_read(&md->io_lock);
 
 		if (unlikely(test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) &&
@@ -1415,6 +1430,36 @@
 	return r;
 }
 
+static int dm_flush(struct mapped_device *md)
+{
+	dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
+	return 0;
+}
+
+static void process_barrier(struct mapped_device *md, struct bio *bio)
+{
+	int error = dm_flush(md);
+
+	if (unlikely(error)) {
+		bio_endio(bio, error);
+		return;
+	}
+	if (bio_empty_barrier(bio)) {
+		bio_endio(bio, 0);
+		return;
+	}
+
+	__split_and_process_bio(md, bio);
+
+	error = dm_flush(md);
+
+	if (!error && md->barrier_error)
+		error = md->barrier_error;
+
+	if (md->barrier_error != DM_ENDIO_REQUEUE)
+		bio_endio(bio, error);
+}
+
 /*
  * Process the deferred bios
  */
@@ -1438,7 +1483,10 @@
 
 		up_write(&md->io_lock);
 
-		__split_and_process_bio(md, c);
+		if (bio_barrier(c))
+			process_barrier(md, c);
+		else
+			__split_and_process_bio(md, c);
 
 		down_write(&md->io_lock);
 	}