dm cache: fix race when issuing a POLICY_REPLACE operation

There is a race between a policy deciding to replace a cache entry,
the core target writing back any dirty data from this block, and other
IO threads doing IO to the same block.

This sort of problem is avoided most of the time by the core target
grabbing a bio prison cell before making the request to the policy.
But for a demotion the core target doesn't know which block will be
demoted, so can't do this in advance.

Fix this demotion race by introducing a callback to the policy interface
that allows the policy to grab the cell on behalf of the core target.

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

1 files changed, 38 insertions, 20 deletions

diff --git a/drivers/md/dm-cache-target.c b/drivers/md/dm-cache-target.c
index 41b2594a80c6..d5982480630b 100644
--- a/drivers/md/dm-cache-target.c
+++ b/drivers/md/dm-cache-target.c
@@ -1439,16 +1439,43 @@ static void inc_miss_counter(struct cache *cache, struct bio *bio)
 		   &cache->stats.read_miss : &cache->stats.write_miss);
 }
 
+/*----------------------------------------------------------------*/
+
+struct old_oblock_lock {
+	struct policy_locker locker;
+	struct cache *cache;
+	struct prealloc *structs;
+	struct dm_bio_prison_cell *cell;
+};
+
+static int null_locker(struct policy_locker *locker, dm_oblock_t b)
+{
+	/* This should never be called */
+	BUG();
+	return 0;
+}
+
+static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
+{
+	struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
+	struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
+
+	return bio_detain(l->cache, b, NULL, cell_prealloc,
+			  (cell_free_fn) prealloc_put_cell,
+			  l->structs, &l->cell);
+}
+
 static void process_bio(struct cache *cache, struct prealloc *structs,
 			struct bio *bio)
 {
 	int r;
 	bool release_cell = true;
 	dm_oblock_t block = get_bio_block(cache, bio);
-	struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
+	struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
 	struct policy_result lookup_result;
 	bool passthrough = passthrough_mode(&cache->features);
 	bool discarded_block, can_migrate;
+	struct old_oblock_lock ool;
 
 	/*
 	 * Check to see if that block is currently migrating.
@@ -1463,8 +1490,12 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
 	discarded_block = is_discarded_oblock(cache, block);
 	can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
 
+	ool.locker.fn = cell_locker;
+	ool.cache = cache;
+	ool.structs = structs;
+	ool.cell = NULL;
 	r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
-		       bio, &lookup_result);
+		       bio, &ool.locker, &lookup_result);
 
 	if (r == -EWOULDBLOCK)
 		/* migration has been denied */
@@ -1521,27 +1552,11 @@ static void process_bio(struct cache *cache, struct prealloc *structs,
 		break;
 
 	case POLICY_REPLACE:
-		cell_prealloc = prealloc_get_cell(structs);
-		r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
-			       (cell_free_fn) prealloc_put_cell,
-			       structs, &old_ocell);
-		if (r > 0) {
-			/*
-			 * We have to be careful to avoid lock inversion of
-			 * the cells.  So we back off, and wait for the
-			 * old_ocell to become free.
-			 */
-			policy_force_mapping(cache->policy, block,
-					     lookup_result.old_oblock);
-			atomic_inc(&cache->stats.cache_cell_clash);
-			break;
-		}
 		atomic_inc(&cache->stats.demotion);
 		atomic_inc(&cache->stats.promotion);
-
 		demote_then_promote(cache, structs, lookup_result.old_oblock,
 				    block, lookup_result.cblock,
-				    old_ocell, new_ocell);
+				    ool.cell, new_ocell);
 		release_cell = false;
 		break;
 
@@ -2589,6 +2604,9 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
 	bool discarded_block;
 	struct policy_result lookup_result;
 	struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
+	struct old_oblock_lock ool;
+
+	ool.locker.fn = null_locker;
 
 	if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
 		/*
@@ -2627,7 +2645,7 @@ static int __cache_map(struct cache *cache, struct bio *bio, struct dm_bio_priso
 	discarded_block = is_discarded_oblock(cache, block);
 
 	r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
-		       bio, &lookup_result);
+		       bio, &ool.locker, &lookup_result);
 	if (r == -EWOULDBLOCK) {
 		cell_defer(cache, *cell, true);
 		return DM_MAPIO_SUBMITTED;