dm: optimize flushes

Device mapper sends flush bios to all the targets and the targets send it
to the underlying device. That may be inefficient, for example if a table
contains 10 linear targets pointing to the same physical device, then
device mapper would send 10 flush bios to that device - despite the fact
that only one bio would be sufficient.

This commit optimizes the flush behavior. It introduces a per-target
variable flush_bypasses_map - it is set when the target supports flush
optimization - currently, the dm-linear and dm-stripe targets support it.
When all the targets in a table have flush_bypasses_map,
flush_bypasses_map on the table is set. __send_empty_flush tests if the
table has flush_bypasses_map - and if it has, no flush bios are sent to
the targets via the "map" method and the list dm_table->devices is
iterated and the flush bios are sent to each member of the list.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Reviewed-by: Mike Snitzer <snitzer@kernel.org>
Suggested-by: Yang Yang <yang.yang@vivo.com>

drivers/md/dm-core.h

@@ -206,6 +206,8 @@ struct dm_table {
 
 	bool integrity_supported:1;
 	bool singleton:1;
+	/* set if all the targets in the table have "flush_bypasses_map" set */
+	bool flush_bypasses_map:1;
 
 	/*
 	 * Indicates the rw permissions for the new logical device.  This

drivers/md/dm-linear.c

@@ -62,6 +62,7 @@ static int linear_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	ti->num_discard_bios = 1;
 	ti->num_secure_erase_bios = 1;
 	ti->num_write_zeroes_bios = 1;
+	ti->flush_bypasses_map = true;
 	ti->private = lc;
 	return 0;
 

drivers/md/dm-stripe.c

@@ -157,6 +157,7 @@ static int stripe_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	ti->num_discard_bios = stripes;
 	ti->num_secure_erase_bios = stripes;
 	ti->num_write_zeroes_bios = stripes;
+	ti->flush_bypasses_map = true;
 
 	sc->chunk_size = chunk_size;
 	if (chunk_size & (chunk_size - 1))

drivers/md/dm-table.c

@@ -160,6 +160,7 @@ int dm_table_create(struct dm_table **result, blk_mode_t mode,
 	t->type = DM_TYPE_NONE;
 	t->mode = mode;
 	t->md = md;
+	t->flush_bypasses_map = true;
 	*result = t;
 	return 0;
 }
@@ -748,6 +749,9 @@ int dm_table_add_target(struct dm_table *t, const char *type,
 	if (ti->limit_swap_bios && !static_key_enabled(&swap_bios_enabled.key))
 		static_branch_enable(&swap_bios_enabled);
 
+	if (!ti->flush_bypasses_map)
+		t->flush_bypasses_map = false;
+
 	return 0;
 
  bad:

drivers/md/dm.c

@@ -645,7 +645,7 @@ static struct bio *alloc_tio(struct clone_info *ci, struct dm_target *ti,
 
 	/* Set default bdev, but target must bio_set_dev() before issuing IO */
 	clone->bi_bdev = md->disk->part0;
-	if (unlikely(ti->needs_bio_set_dev))
+	if (likely(ti != NULL) && unlikely(ti->needs_bio_set_dev))
 		bio_set_dev(clone, md->disk->part0);
 
 	if (len) {
@@ -1107,7 +1107,7 @@ static void clone_endio(struct bio *bio)
 	blk_status_t error = bio->bi_status;
 	struct dm_target_io *tio = clone_to_tio(bio);
 	struct dm_target *ti = tio->ti;
-	dm_endio_fn endio = ti->type->end_io;
+	dm_endio_fn endio = likely(ti != NULL) ? ti->type->end_io : NULL;
 	struct dm_io *io = tio->io;
 	struct mapped_device *md = io->md;
 
@@ -1154,7 +1154,7 @@ static void clone_endio(struct bio *bio)
 	}
 
 	if (static_branch_unlikely(&swap_bios_enabled) &&
-	    unlikely(swap_bios_limit(ti, bio)))
+	    likely(ti != NULL) && unlikely(swap_bios_limit(ti, bio)))
 		up(&md->swap_bios_semaphore);
 
 	free_tio(bio);
@@ -1553,6 +1553,7 @@ static void __send_empty_flush(struct clone_info *ci)
 	ci->sector_count = 0;
 	ci->io->tio.clone.bi_iter.bi_size = 0;
 
+	if (!t->flush_bypasses_map) {
 		for (unsigned int i = 0; i < t->num_targets; i++) {
 			unsigned int bios;
 			struct dm_target *ti = dm_table_get_target(t, i);
@@ -1565,6 +1566,31 @@ static void __send_empty_flush(struct clone_info *ci)
 						     NULL, GFP_NOWAIT);
 			atomic_sub(ti->num_flush_bios - bios, &ci->io->io_count);
 		}
+	} else {
+		/*
+		 * Note that there's no need to grab t->devices_lock here
+		 * because the targets that support flush optimization don't
+		 * modify the list of devices.
+		 */
+		struct list_head *devices = dm_table_get_devices(t);
+		unsigned int len = 0;
+		struct dm_dev_internal *dd;
+		list_for_each_entry(dd, devices, list) {
+			struct bio *clone;
+			/*
+			 * Note that the structure dm_target_io is not
+			 * associated with any target (because the device may be
+			 * used by multiple targets), so we set tio->ti = NULL.
+			 * We must check for NULL in the I/O processing path, to
+			 * avoid NULL pointer dereference.
+			 */
+			clone = alloc_tio(ci, NULL, 0, &len, GFP_NOIO);
+			atomic_add(1, &ci->io->io_count);
+			bio_set_dev(clone, dd->dm_dev->bdev);
+			clone->bi_end_io = clone_endio;
+			dm_submit_bio_remap(clone, NULL);
+		}
+	}
 
 	/*
 	 * alloc_io() takes one extra reference for submission, so the

include/linux/device-mapper.h

@@ -397,6 +397,21 @@ struct dm_target {
 	 * bio_set_dev(). NOTE: ideally a target should _not_ need this.
 	 */
 	bool needs_bio_set_dev:1;
+
+	/*
+	 * Set if the target supports flush optimization. If all the targets in
+	 * a table have flush_bypasses_map set, the dm core will not send
+	 * flushes to the targets via a ->map method. It will iterate over
+	 * dm_table->devices and send flushes to the devices directly. This
+	 * optimization reduces the number of flushes being sent when multiple
+	 * targets in a table use the same underlying device.
+	 *
+	 * This optimization may be enabled on targets that just pass the
+	 * flushes to the underlying devices without performing any other
+	 * actions on the flush request. Currently, dm-linear and dm-stripe
+	 * support it.
+	 */
+	bool flush_bypasses_map:1;
 };
 
 void *dm_per_bio_data(struct bio *bio, size_t data_size);