Barrier support.

Barriers are submitted to a worker thread that issues them in-order.

__process_bio functions is modified that when it sees a barrier request,
it waits for all pending IO before the request, then submits the barrier
and waits for it.

DM_ENDIO_REQUEUE doesn't work and I doubt that it can be made work with
barriers.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>

---
 drivers/md/dm.c |   77 +++++++++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 62 insertions(+), 15 deletions(-)

Index: linux-2.6.28-clean/drivers/md/dm.c
===================================================================
--- linux-2.6.28-clean.orig/drivers/md/dm.c	2009-01-12 15:17:18.000000000 +0100
+++ linux-2.6.28-clean/drivers/md/dm.c	2009-01-12 15:19:40.000000000 +0100
@@ -100,6 +100,11 @@ struct mapped_device {
 	struct bio_list pushback;
 
 	/*
+	 * An error from the barrier request currently being processed.
+	 */
+	int barrier_error;
+
+	/*
 	 * Processing queue (flush/barriers)
 	 */
 	struct workqueue_struct *wq;
@@ -381,6 +386,10 @@ static void end_io_acct(struct dm_io *io
 	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
+	 * barrier bio
+	 */
 	dm_disk(md)->part0.in_flight = pending =
 		atomic_dec_return(&md->pending);
 
@@ -480,7 +489,8 @@ static void dec_pending(struct dm_io *io
 			 * suspend queue merges the pushback list.
 			 */
 			spin_lock_irqsave(&io->md->pushback_lock, flags);
-			if (__noflush_suspending(io->md))
+			/* TODO: no barriers for noflush are supported now */
+			if (__noflush_suspending(io->md) && !bio_barrier(io->bio))
 				bio_list_add(&io->md->pushback, io->bio);
 			else
 				/* noflush suspend was interrupted. */
@@ -488,13 +498,23 @@ static void dec_pending(struct dm_io *io
 			spin_unlock_irqrestore(&io->md->pushback_lock, flags);
 		}
 
-		end_io_acct(io);
-
-		if (io->error != DM_ENDIO_REQUEUE) {
-			blk_add_trace_bio(io->md->queue, io->bio,
-					  BLK_TA_COMPLETE);
+		if (bio_barrier(io->bio)) {
+			/*
+			 * There could be just one barrier request, so we use
+			 * per-device variable for error reporting is OK.
+			 * Note that you can't touch the bio after end_io_acct
+			 */
+			io->md->barrier_error = io->error;
+			end_io_acct(io);
+		} else {
+			end_io_acct(io);
+
+			if (io->error != DM_ENDIO_REQUEUE) {
+				blk_add_trace_bio(io->md->queue, io->bio,
+						  BLK_TA_COMPLETE);
 
-			bio_endio(io->bio, io->error);
+				bio_endio(io->bio, io->error);
+			}
 		}
 
 		free_io(io->md, io);
@@ -660,6 +680,7 @@ static struct bio *clone_bio(struct bio 
 
 	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;
@@ -780,7 +801,10 @@ static void __process_bio(struct mapped_
 
 	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;
 	}
 
@@ -868,11 +892,6 @@ static int dm_request(struct request_que
 	 * 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();
@@ -884,7 +903,8 @@ static int dm_request(struct request_que
 	 * If we're suspended we have to queue
 	 * this io for later.
 	 */
-	if (unlikely(test_bit(DMF_BLOCK_IO, &md->flags))) {
+	if (unlikely(test_bit(DMF_BLOCK_IO, &md->flags)) ||
+	    unlikely(bio_barrier(bio))) {
 		up_read(&md->io_lock);
 
 		queue_io(md, bio);
@@ -1341,6 +1361,12 @@ static int dm_wait_for_completion(struct
 	return r;
 }
 
+static int dm_flush(struct mapped_device *md)
+{
+	dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
+	return 0;
+}
+
 static void __merge_pushback_list(struct mapped_device *md)
 {
 	unsigned long flags;
@@ -1368,8 +1394,29 @@ static void dm_wq_work(struct work_struc
 		}
 		up_write(&md->io_lock);
 
-		__process_bio(md, c);
+		if (!bio_barrier(c))
+			__process_bio(md, c);
+		else {
+			int error = dm_flush(md);
+			if (unlikely(error)) {
+				bio_endio(c, error);
+				goto next_bio;
+			}
+			if (bio_empty_barrier(c)) {
+				bio_endio(c, 0);
+				goto next_bio;
+			}
+
+			__process_bio(md, c);
+
+			error = dm_flush(md);
+			if (!error && md->barrier_error)
+				error = md->barrier_error;
+
+			bio_endio(c, error);
+		}
 
+next_bio:
 		down_write(&md->io_lock);
 	}