diff --git a/fs/bcachefs/btree_cache.c b/fs/bcachefs/btree_cache.c
index 775b3e8468da..f32fc45c85d2 100644
--- a/fs/bcachefs/btree_cache.c
+++ b/fs/bcachefs/btree_cache.c
@@ -147,6 +147,11 @@ int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
 	b->c.level	= level;
 	b->c.btree_id	= id;
 
+	if (level)
+		six_lock_pcpu_alloc(&b->c.lock);
+	else
+		six_lock_pcpu_free_rcu(&b->c.lock);
+
 	mutex_lock(&bc->lock);
 	ret = __bch2_btree_node_hash_insert(bc, b);
 	if (!ret)
@@ -393,6 +398,7 @@ void bch2_fs_btree_cache_exit(struct bch_fs *c)
 	while (!list_empty(&bc->freed)) {
 		b = list_first_entry(&bc->freed, struct btree, list);
 		list_del(&b->list);
+		six_lock_pcpu_free(&b->c.lock);
 		kfree(b);
 	}
 
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index 5f30626d1852..cf41ece0d66e 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -79,11 +79,19 @@ void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
 	 * goes to 0, and it's safe because we have the node intent
 	 * locked:
 	 */
-	atomic64_sub(__SIX_VAL(read_lock, readers),
-		     &b->c.lock.state.counter);
+	if (!b->c.lock.readers)
+		atomic64_sub(__SIX_VAL(read_lock, readers),
+			     &b->c.lock.state.counter);
+	else
+		this_cpu_sub(*b->c.lock.readers, readers);
+
 	btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
-	atomic64_add(__SIX_VAL(read_lock, readers),
-		     &b->c.lock.state.counter);
+
+	if (!b->c.lock.readers)
+		atomic64_add(__SIX_VAL(read_lock, readers),
+			     &b->c.lock.state.counter);
+	else
+		this_cpu_add(*b->c.lock.readers, readers);
 }
 
 bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
diff --git a/fs/bcachefs/btree_update_interior.c b/fs/bcachefs/btree_update_interior.c
index 4ad8084714f9..2c202dd01766 100644
--- a/fs/bcachefs/btree_update_interior.c
+++ b/fs/bcachefs/btree_update_interior.c
@@ -988,6 +988,11 @@ static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 	list_del_init(&b->list);
 	mutex_unlock(&c->btree_cache.lock);
 
+	if (b->c.level)
+		six_lock_pcpu_alloc(&b->c.lock);
+	else
+		six_lock_pcpu_free(&b->c.lock);
+
 	mutex_lock(&c->btree_root_lock);
 	BUG_ON(btree_node_root(c, b) &&
 	       (b->c.level < btree_node_root(c, b)->c.level ||