bcachefs: Rework btree node pinning

In backpointers fsck, we do a seqential scan of one btree, and check
references to another: extents <-> backpointers

Checking references generates random lookups, so we want to pin that
btree in memory (or only a range, if it doesn't fit in ram).

Previously, this was done with a simple check in the shrinker - "if
btree node is in range being pinned, don't free it" - but this generated
OOMs, as our shrinker wasn't well behaved if there was less memory
available than expected.

Instead, we now have two different shrinkers and lru lists; the second
shrinker being for pinned nodes, with seeks set much higher than normal
- so they can still be freed if necessary, but we'll prefer not to.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/backpointers.c

@@ -752,10 +752,12 @@ static int bch2_get_btree_in_memory_pos(struct btree_trans *trans,
 	s64 mem_may_pin = mem_may_pin_bytes(c);
 	int ret = 0;
 
+	bch2_btree_cache_unpin(c);
+
 	btree_interior_mask |= btree_leaf_mask;
 
-	c->btree_cache.pinned_nodes_leaf_mask		= btree_leaf_mask;
-	c->btree_cache.pinned_nodes_interior_mask	= btree_interior_mask;
+	c->btree_cache.pinned_nodes_mask[0]		= btree_leaf_mask;
+	c->btree_cache.pinned_nodes_mask[1]		= btree_interior_mask;
 	c->btree_cache.pinned_nodes_start		= start;
 	c->btree_cache.pinned_nodes_end			= *end = BBPOS_MAX;
 
@@ -777,6 +779,7 @@ static int bch2_get_btree_in_memory_pos(struct btree_trans *trans,
 					BBPOS(btree, b->key.k.p);
 				break;
 			}
+			bch2_node_pin(c, b);
 			0;
 		}));
 	}
@@ -936,8 +939,7 @@ int bch2_check_extents_to_backpointers(struct bch_fs *c)
 	bch2_trans_put(trans);
 	bch2_bkey_buf_exit(&s.last_flushed, c);
 
-	c->btree_cache.pinned_nodes_leaf_mask = 0;
-	c->btree_cache.pinned_nodes_interior_mask = 0;
+	bch2_btree_cache_unpin(c);
 
 	bch_err_fn(c, ret);
 	return ret;
@@ -1053,8 +1055,7 @@ int bch2_check_backpointers_to_extents(struct bch_fs *c)
 	}
 	bch2_trans_put(trans);
 
-	c->btree_cache.pinned_nodes_leaf_mask = 0;
-	c->btree_cache.pinned_nodes_interior_mask = 0;
+	bch2_btree_cache_unpin(c);
 
 	bch_err_fn(c, ret);
 	return ret;

fs/bcachefs/btree_cache.h

@@ -19,6 +19,9 @@ int __bch2_btree_node_hash_insert(struct btree_cache *, struct btree *);
 int bch2_btree_node_hash_insert(struct btree_cache *, struct btree *,
 				unsigned, enum btree_id);
 
+void bch2_node_pin(struct bch_fs *, struct btree *);
+void bch2_btree_cache_unpin(struct bch_fs *);
+
 void bch2_btree_node_update_key_early(struct btree_trans *, enum btree_id, unsigned,
 				      struct bkey_s_c, struct bkey_i *);
 

fs/bcachefs/btree_types.h

@@ -147,8 +147,7 @@ struct btree {
 	x(noevict)				\
 	x(write_blocked)			\
 	x(will_make_reachable)			\
-	x(access_bit)				\
-	x(pinned)				\
+	x(access_bit)
 
 enum bch_btree_cache_not_freed_reasons {
 #define x(n) BCH_BTREE_CACHE_NOT_FREED_##n,
@@ -157,6 +156,13 @@ enum bch_btree_cache_not_freed_reasons {
 	BCH_BTREE_CACHE_NOT_FREED_REASONS_NR,
 };
 
+struct btree_cache_list {
+	unsigned		idx;
+	struct shrinker		*shrink;
+	struct list_head	list;
+	size_t			nr;
+};
+
 struct btree_cache {
 	struct rhashtable	table;
 	bool			table_init_done;
@@ -174,12 +180,11 @@ struct btree_cache {
 	 * should never grow past ~2-3 nodes in practice.
 	 */
 	struct mutex		lock;
-	struct list_head	live;
 	struct list_head	freeable;
 	struct list_head	freed_pcpu;
 	struct list_head	freed_nonpcpu;
+	struct btree_cache_list	live[2];
 
-	size_t			nr_live;
 	size_t			nr_freeable;
 	size_t			nr_reserve;
 	size_t			nr_by_btree[BTREE_ID_NR];
@@ -188,7 +193,6 @@ struct btree_cache {
 	/* shrinker stats */
 	size_t			nr_freed;
 	u64			not_freed[BCH_BTREE_CACHE_NOT_FREED_REASONS_NR];
-	struct shrinker		*shrink;
 
 	/*
 	 * If we need to allocate memory for a new btree node and that
@@ -201,8 +205,8 @@ struct btree_cache {
 
 	struct bbpos		pinned_nodes_start;
 	struct bbpos		pinned_nodes_end;
-	u64			pinned_nodes_leaf_mask;
-	u64			pinned_nodes_interior_mask;
+	/* btree id mask: 0 for leaves, 1 for interior */
+	u64			pinned_nodes_mask[2];
 };
 
 struct btree_node_iter {
@@ -594,7 +598,8 @@ enum btree_write_type {
 	x(dying)							\
 	x(fake)								\
 	x(need_rewrite)							\
-	x(never_write)
+	x(never_write)							\
+	x(pinned)
 
 enum btree_flags {
 	/* First bits for btree node write type */

fs/bcachefs/btree_update_interior.c

@@ -1904,7 +1904,7 @@ static void __btree_increase_depth(struct btree_update *as, struct btree_trans *
 	six_unlock_intent(&n->c.lock);
 
 	mutex_lock(&c->btree_cache.lock);
-	list_add_tail(&b->list, &c->btree_cache.live);
+	list_add_tail(&b->list, &c->btree_cache.live[btree_node_pinned(b)].list);
 	mutex_unlock(&c->btree_cache.lock);
 
 	bch2_trans_verify_locks(trans);

fs/bcachefs/journal_reclaim.c

@@ -641,6 +641,7 @@ static u64 journal_seq_to_flush(struct journal *j)
 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+	struct btree_cache *bc = &c->btree_cache;
 	bool kthread = (current->flags & PF_KTHREAD) != 0;
 	u64 seq_to_flush;
 	size_t min_nr, min_key_cache, nr_flushed;
@@ -681,7 +682,8 @@ static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 		if (j->watermark != BCH_WATERMARK_stripe)
 			min_nr = 1;
 
-		if (atomic_long_read(&c->btree_cache.nr_dirty) * 2 > c->btree_cache.nr_live)
+		size_t btree_cache_live = bc->live[0].nr + bc->live[1].nr;
+		if (atomic_long_read(&bc->nr_dirty) * 2 > btree_cache_live)
 			min_nr = 1;
 
 		min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
@@ -689,8 +691,7 @@ static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 		trace_and_count(c, journal_reclaim_start, c,
 				direct, kicked,
 				min_nr, min_key_cache,
-				atomic_long_read(&c->btree_cache.nr_dirty),
-				c->btree_cache.nr_live,
+				atomic_long_read(&bc->nr_dirty), btree_cache_live,
 				atomic_long_read(&c->btree_key_cache.nr_dirty),
 				atomic_long_read(&c->btree_key_cache.nr_keys));
 

fs/bcachefs/sysfs.c

@@ -244,14 +244,18 @@ static struct attribute sysfs_state_rw = {
 
 static size_t bch2_btree_cache_size(struct bch_fs *c)
 {
+	struct btree_cache *bc = &c->btree_cache;
 	size_t ret = 0;
 	struct btree *b;
 
-	mutex_lock(&c->btree_cache.lock);
-	list_for_each_entry(b, &c->btree_cache.live, list)
+	mutex_lock(&bc->lock);
+	list_for_each_entry(b, &bc->live[0].list, list)
 		ret += btree_buf_bytes(b);
-
-	mutex_unlock(&c->btree_cache.lock);
+	list_for_each_entry(b, &bc->live[1].list, list)
+		ret += btree_buf_bytes(b);
+	list_for_each_entry(b, &bc->freeable, list)
+		ret += btree_buf_bytes(b);
+	mutex_unlock(&bc->lock);
 	return ret;
 }
 
@@ -444,11 +448,12 @@ STORE(bch2_fs)
 		return -EROFS;
 
 	if (attr == &sysfs_trigger_btree_cache_shrink) {
+		struct btree_cache *bc = &c->btree_cache;
 		struct shrink_control sc;
 
 		sc.gfp_mask = GFP_KERNEL;
 		sc.nr_to_scan = strtoul_or_return(buf);
-		c->btree_cache.shrink->scan_objects(c->btree_cache.shrink, &sc);
+		bc->live[0].shrink->scan_objects(bc->live[0].shrink, &sc);
 	}
 
 	if (attr == &sysfs_trigger_btree_key_cache_shrink) {