dm bio prison v1: improve concurrent IO performance

Split the bio prison into multiple regions, with a separate rbtree and
associated lock for each region.

To get fast bio prison locking and not damage the performance of
discards too much the bio-prison now stipulates that discards should
not cross a BIO_PRISON_MAX_RANGE boundary.

Because the range of a key (block_end - block_begin) must not exceed
BIO_PRISON_MAX_RANGE: break_up_discard_bio() now ensures the data
range reflected in PHYSICAL key doesn't exceed BIO_PRISON_MAX_RANGE.
And splitting the thin target's discards (handled with VIRTUAL key) is
achieved by updating dm-thin.c to set limits->max_discard_sectors in
terms of BIO_PRISON_MAX_RANGE _and_ setting the thin and thin-pool
targets' max_discard_granularity to true.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>

1 files changed, 54 insertions, 38 deletions

diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
index 00323428919e..33ad5695f959 100644
--- a/drivers/md/dm-thin.c
+++ b/drivers/md/dm-thin.c
@@ -1674,54 +1674,69 @@ static void break_up_discard_bio(struct thin_c *tc, dm_block_t begin, dm_block_t
 	struct dm_cell_key data_key;
 	struct dm_bio_prison_cell *data_cell;
 	struct dm_thin_new_mapping *m;
-	dm_block_t virt_begin, virt_end, data_begin;
+	dm_block_t virt_begin, virt_end, data_begin, data_end;
+	dm_block_t len, next_boundary;
 
 	while (begin != end) {
-		r = ensure_next_mapping(pool);
-		if (r)
-			/* we did our best */
-			return;
-
 		r = dm_thin_find_mapped_range(tc->td, begin, end, &virt_begin, &virt_end,
 					      &data_begin, &maybe_shared);
-		if (r)
+		if (r) {
 			/*
 			 * Silently fail, letting any mappings we've
 			 * created complete.
 			 */
 			break;
-
-		build_key(tc->td, PHYSICAL, data_begin, data_begin + (virt_end - virt_begin), &data_key);
-		if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
-			/* contention, we'll give up with this range */
-			begin = virt_end;
-			continue;
 		}
 
-		/*
-		 * IO may still be going to the destination block.  We must
-		 * quiesce before we can do the removal.
-		 */
-		m = get_next_mapping(pool);
-		m->tc = tc;
-		m->maybe_shared = maybe_shared;
-		m->virt_begin = virt_begin;
-		m->virt_end = virt_end;
-		m->data_block = data_begin;
-		m->cell = data_cell;
-		m->bio = bio;
+		data_end = data_begin + (virt_end - virt_begin);
 
 		/*
-		 * The parent bio must not complete before sub discard bios are
-		 * chained to it (see end_discard's bio_chain)!
-		 *
-		 * This per-mapping bi_remaining increment is paired with
-		 * the implicit decrement that occurs via bio_endio() in
-		 * end_discard().
+		 * Make sure the data region obeys the bio prison restrictions.
 		 */
-		bio_inc_remaining(bio);
-		if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
-			pool->process_prepared_discard(m);
+		while (data_begin < data_end) {
+			r = ensure_next_mapping(pool);
+			if (r)
+				return; /* we did our best */
+
+			next_boundary = ((data_begin >> BIO_PRISON_MAX_RANGE_SHIFT) + 1)
+				<< BIO_PRISON_MAX_RANGE_SHIFT;
+			len = min_t(sector_t, data_end - data_begin, next_boundary - data_begin);
+
+			build_key(tc->td, PHYSICAL, data_begin, data_begin + len, &data_key);
+			if (bio_detain(tc->pool, &data_key, NULL, &data_cell)) {
+				/* contention, we'll give up with this range */
+				data_begin += len;
+				continue;
+			}
+
+			/*
+			 * IO may still be going to the destination block.  We must
+			 * quiesce before we can do the removal.
+			 */
+			m = get_next_mapping(pool);
+			m->tc = tc;
+			m->maybe_shared = maybe_shared;
+			m->virt_begin = virt_begin;
+			m->virt_end = virt_begin + len;
+			m->data_block = data_begin;
+			m->cell = data_cell;
+			m->bio = bio;
+
+			/*
+			 * The parent bio must not complete before sub discard bios are
+			 * chained to it (see end_discard's bio_chain)!
+			 *
+			 * This per-mapping bi_remaining increment is paired with
+			 * the implicit decrement that occurs via bio_endio() in
+			 * end_discard().
+			 */
+			bio_inc_remaining(bio);
+			if (!dm_deferred_set_add_work(pool->all_io_ds, &m->list))
+				pool->process_prepared_discard(m);
+
+			virt_begin += len;
+			data_begin += len;
+		}
 
 		begin = virt_end;
 	}
@@ -3380,13 +3395,13 @@ static int pool_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	 */
 	if (pf.discard_enabled && pf.discard_passdown) {
 		ti->num_discard_bios = 1;
-
 		/*
 		 * Setting 'discards_supported' circumvents the normal
 		 * stacking of discard limits (this keeps the pool and
 		 * thin devices' discard limits consistent).
 		 */
 		ti->discards_supported = true;
+		ti->max_discard_granularity = true;
 	}
 	ti->private = pt;
 
@@ -4096,7 +4111,7 @@ static struct target_type pool_target = {
 	.name = "thin-pool",
 	.features = DM_TARGET_SINGLETON | DM_TARGET_ALWAYS_WRITEABLE |
 		    DM_TARGET_IMMUTABLE,
-	.version = {1, 22, 0},
+	.version = {1, 23, 0},
 	.module = THIS_MODULE,
 	.ctr = pool_ctr,
 	.dtr = pool_dtr,
@@ -4261,6 +4276,7 @@ static int thin_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	if (tc->pool->pf.discard_enabled) {
 		ti->discards_supported = true;
 		ti->num_discard_bios = 1;
+		ti->max_discard_granularity = true;
 	}
 
 	mutex_unlock(&dm_thin_pool_table.mutex);
@@ -4476,12 +4492,12 @@ static void thin_io_hints(struct dm_target *ti, struct queue_limits *limits)
 		return;
 
 	limits->discard_granularity = pool->sectors_per_block << SECTOR_SHIFT;
-	limits->max_discard_sectors = 2048 * 1024 * 16; /* 16G */
+	limits->max_discard_sectors = pool->sectors_per_block * BIO_PRISON_MAX_RANGE;
 }
 
 static struct target_type thin_target = {
 	.name = "thin",
-	.version = {1, 22, 0},
+	.version = {1, 23, 0},
 	.module	= THIS_MODULE,
 	.ctr = thin_ctr,
 	.dtr = thin_dtr,