bcache: btree node reserve for tiering

If the source tier is full, btree_check_reserve() would block. This could
prevent tiering from making progress in copying to the destination tier.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

6 files changed, 35 insertions, 4 deletions

diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
index 300aa6d7a7b0..9308c053e9fc 100644
--- a/drivers/md/bcache/alloc.c
+++ b/drivers/md/bcache/alloc.c
@@ -518,7 +518,7 @@ out:
 	BUG_ON(ca->set->gc_mark_valid &&
 	       GC_MARK(b) != GC_MARK_DIRTY);
 
-	if (reserve <= RESERVE_MOVINGGC_BTREE)
+	if (reserve <= RESERVE_METADATA_LAST)
 		SET_GC_MARK(b, GC_MARK_METADATA);
 	else
 		SET_GC_MARK(b, GC_MARK_DIRTY);
diff --git a/drivers/md/bcache/bcache.h b/drivers/md/bcache/bcache.h
index 46524c1edeb3..b5db867a8c35 100644
--- a/drivers/md/bcache/bcache.h
+++ b/drivers/md/bcache/bcache.h
@@ -443,6 +443,8 @@ struct cached_dev {
 enum alloc_reserve {
 	RESERVE_PRIO	= BTREE_ID_NR,
 	RESERVE_MOVINGGC_BTREE,
+	RESERVE_TIERING_BTREE,
+	RESERVE_METADATA_LAST = RESERVE_TIERING_BTREE,
 	RESERVE_MOVINGGC,
 	RESERVE_NONE,
 	RESERVE_NR,
diff --git a/drivers/md/bcache/request.c b/drivers/md/bcache/request.c
index 8ad057d6a84f..579cc9d289b0 100644
--- a/drivers/md/bcache/request.c
+++ b/drivers/md/bcache/request.c
@@ -48,6 +48,31 @@ static void bio_csum(struct bio *bio, struct bkey *k)
 
 /* Insert data into cache */
 
+static enum alloc_reserve bch_btree_reserve(struct data_insert_op *op)
+{
+	if (op->moving_gc) {
+		/*
+		 * free_inc.size buckets are set aside for moving GC
+		 * btree node allocations. This means that if moving GC
+		 * runs out of new buckets for btree nodes, it will have
+		 * put back at least free_inc.size buckets back on
+		 * free_inc, preventing a deadlock.
+		 *
+		 * XXX: figure out a less stupid way of achieving this
+		 */
+		return RESERVE_MOVINGGC_BTREE;
+	} else if (op->tiering) {
+		/*
+		 * Tiering needs a btree node reserve because of how
+		 * btree_check_reserve() works -- if the cache tier is
+		 * full, we don't want tiering to block forever.
+		 */
+		return RESERVE_TIERING_BTREE;
+	}
+
+	return BTREE_ID_EXTENTS;
+}
+
 /**
  * bch_data_insert_keys - insert extent btree keys for a write
  */
@@ -56,12 +81,11 @@ static void bch_data_insert_keys(struct closure *cl)
 	struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
 	struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
 	enum btree_id id = BTREE_ID_EXTENTS;
-	enum alloc_reserve reserve = id;
+	enum alloc_reserve reserve;
 	unsigned i;
 	int ret;
 
-	if (op->moving_gc)
-		reserve = RESERVE_MOVINGGC_BTREE;
+	reserve = bch_btree_reserve(op);
 
 	ret = bch_btree_insert(op->c, id, reserve, &op->insert_keys,
 			       replace_key, cl, op->flush);
diff --git a/drivers/md/bcache/request.h b/drivers/md/bcache/request.h
index 94631d33378a..38477c763884 100644
--- a/drivers/md/bcache/request.h
+++ b/drivers/md/bcache/request.h
@@ -37,6 +37,8 @@ struct data_insert_op {
 		/* Use moving GC reserves for buckets, btree nodes and
 		 * open buckets? */
 		unsigned	moving_gc:1;
+		/* Use tiering reserves for btree nodes? */
+		unsigned	tiering:1;
 		/* Set on completion */
 		unsigned	replace_collision:1;
 		/* Internal */
diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
index 52879765c6d1..596d87f5e652 100644
--- a/drivers/md/bcache/super.c
+++ b/drivers/md/bcache/super.c
@@ -1820,6 +1820,8 @@ static int cache_alloc(struct cache *ca)
 	if (!init_fifo(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
 	    !init_fifo(&ca->free[RESERVE_MOVINGGC_BTREE],
 		       free_inc_reserve, GFP_KERNEL) ||
+	    !init_fifo(&ca->free[RESERVE_TIERING_BTREE],
+		       BTREE_NODE_RESERVE, GFP_KERNEL) ||
 	    !init_fifo(&ca->free[RESERVE_MOVINGGC],
 		       movinggc_reserve, GFP_KERNEL) ||
 	    !init_fifo(&ca->free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
diff --git a/drivers/md/bcache/tier.c b/drivers/md/bcache/tier.c
index 6435f321d470..5ff73532996b 100644
--- a/drivers/md/bcache/tier.c
+++ b/drivers/md/bcache/tier.c
@@ -104,6 +104,7 @@ static void read_tiering(struct cache_set *c)
 					&io->bio.bio, 0,
 					false, false, false,
 					&io->w->key, &io->w->key);
+		io->op.tiering	= 1;
 
 		io->op.tier	= PTR_TIER(c, &w->key,
 					   bch_extent_ptrs(&w->key) - 1) + 1;