bcachefs: Add accounting for dirty btree nodes/keys

This lets us improve journal reclaim, so that it now tries to make sure
no more than 3/4s of the btree node cache and btree key cache are dirty
- ensuring the shrinkers can free memory.

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

fs/bcachefs/btree_cache.c

@@ -382,11 +382,13 @@ void bch2_fs_btree_cache_exit(struct bch_fs *c)
 
 		if (btree_node_dirty(b))
 			bch2_btree_complete_write(c, b, btree_current_write(b));
-		clear_btree_node_dirty(b);
+		clear_btree_node_dirty(c, b);
 
 		btree_node_data_free(c, b);
 	}
 
+	BUG_ON(atomic_read(&c->btree_cache.dirty));
+
 	while (!list_empty(&bc->freed)) {
 		b = list_first_entry(&bc->freed, struct btree, list);
 		list_del(&b->list);

fs/bcachefs/btree_io.c

@@ -1498,6 +1498,8 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 		new ^=  (1 << BTREE_NODE_write_idx);
 	} while (cmpxchg_acquire(&b->flags, old, new) != old);
 
+	atomic_dec(&c->btree_cache.dirty);
+
 	BUG_ON(btree_node_fake(b));
 	BUG_ON((b->will_make_reachable != 0) != !b->written);
 

fs/bcachefs/btree_io.h

@@ -14,6 +14,23 @@ struct btree_write;
 struct btree;
 struct btree_iter;
 
+static inline bool btree_node_dirty(struct btree *b)
+{
+	return test_bit(BTREE_NODE_dirty, &b->flags);
+}
+
+static inline void set_btree_node_dirty(struct bch_fs *c, struct btree *b)
+{
+	if (!test_and_set_bit(BTREE_NODE_dirty, &b->flags))
+		atomic_inc(&c->btree_cache.dirty);
+}
+
+static inline void clear_btree_node_dirty(struct bch_fs *c, struct btree *b)
+{
+	if (test_and_clear_bit(BTREE_NODE_dirty, &b->flags))
+		atomic_dec(&c->btree_cache.dirty);
+}
+
 struct btree_read_bio {
 	struct bch_fs		*c;
 	u64			start_time;

fs/bcachefs/btree_key_cache.c

@@ -65,6 +65,8 @@ static void bkey_cached_evict(struct btree_key_cache *c,
 	BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
 				      bch2_btree_key_cache_params));
 	memset(&ck->key, ~0, sizeof(ck->key));
+
+	c->nr_keys--;
 }
 
 static void bkey_cached_free(struct btree_key_cache *c,
@@ -135,6 +137,8 @@ btree_key_cache_create(struct btree_key_cache *c,
 		return NULL;
 	}
 
+	c->nr_keys++;
+
 	list_move(&ck->list, &c->clean);
 	six_unlock_write(&ck->c.lock);
 
@@ -355,10 +359,14 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 
 	bch2_journal_pin_drop(j, &ck->journal);
 	bch2_journal_preres_put(j, &ck->res);
-	clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 
 	if (!evict) {
 		mutex_lock(&c->btree_key_cache.lock);
+		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
+			c->btree_key_cache.nr_dirty--;
+		}
+
 		list_move_tail(&ck->list, &c->btree_key_cache.clean);
 		mutex_unlock(&c->btree_key_cache.lock);
 	} else {
@@ -371,6 +379,11 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 		six_lock_write(&ck->c.lock, NULL, NULL);
 
 		mutex_lock(&c->btree_key_cache.lock);
+		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
+			c->btree_key_cache.nr_dirty--;
+		}
+
 		bkey_cached_evict(&c->btree_key_cache, ck);
 		bkey_cached_free(&c->btree_key_cache, ck);
 		mutex_unlock(&c->btree_key_cache.lock);
@@ -448,9 +461,10 @@ bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 
 	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 		mutex_lock(&c->btree_key_cache.lock);
-		list_del_init(&ck->list);
+		list_move(&ck->list, &c->btree_key_cache.dirty);
 
 		set_bit(BKEY_CACHED_DIRTY, &ck->flags);
+		c->btree_key_cache.nr_dirty++;
 		mutex_unlock(&c->btree_key_cache.lock);
 	}
 
@@ -467,20 +481,28 @@ void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 }
 #endif
 
-void bch2_fs_btree_key_cache_exit(struct btree_key_cache *c)
+void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 {
+	struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 	struct bkey_cached *ck, *n;
 
-	mutex_lock(&c->lock);
-	list_for_each_entry_safe(ck, n, &c->clean, list) {
+	mutex_lock(&bc->lock);
+	list_splice(&bc->dirty, &bc->clean);
+
+	list_for_each_entry_safe(ck, n, &bc->clean, list) {
 		kfree(ck->k);
 		kfree(ck);
+		bc->nr_keys--;
 	}
-	list_for_each_entry_safe(ck, n, &c->freed, list)
-		kfree(ck);
-	mutex_unlock(&c->lock);
 
-	rhashtable_destroy(&c->table);
+	BUG_ON(bc->nr_dirty && !bch2_journal_error(&c->journal));
+	BUG_ON(bc->nr_keys);
+
+	list_for_each_entry_safe(ck, n, &bc->freed, list)
+		kfree(ck);
+	mutex_unlock(&bc->lock);
+
+	rhashtable_destroy(&bc->table);
 }
 
 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
@@ -488,6 +510,7 @@ void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 	mutex_init(&c->lock);
 	INIT_LIST_HEAD(&c->freed);
 	INIT_LIST_HEAD(&c->clean);
+	INIT_LIST_HEAD(&c->dirty);
 }
 
 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)

fs/bcachefs/btree_types.h

@@ -158,6 +158,7 @@ struct btree_cache {
 	/* Number of elements in live + freeable lists */
 	unsigned		used;
 	unsigned		reserve;
+	atomic_t		dirty;
 	struct shrinker		shrink;
 
 	/*
@@ -294,6 +295,10 @@ struct btree_key_cache {
 	struct rhashtable	table;
 	struct list_head	freed;
 	struct list_head	clean;
+	struct list_head	dirty;
+
+	size_t			nr_keys;
+	size_t			nr_dirty;
 };
 
 struct bkey_cached_key {
@@ -411,7 +416,6 @@ enum btree_flags {
 
 BTREE_FLAG(read_in_flight);
 BTREE_FLAG(read_error);
-BTREE_FLAG(dirty);
 BTREE_FLAG(need_write);
 BTREE_FLAG(noevict);
 BTREE_FLAG(write_idx);

fs/bcachefs/btree_update_interior.c

@@ -149,7 +149,7 @@ void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 
 	b->ob.nr = 0;
 
-	clear_btree_node_dirty(b);
+	clear_btree_node_dirty(c, b);
 
 	btree_node_lock_type(c, b, SIX_LOCK_write);
 	__btree_node_free(c, b);
@@ -264,7 +264,7 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev
 	b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 
 	set_btree_node_accessed(b);
-	set_btree_node_dirty(b);
+	set_btree_node_dirty(c, b);
 	set_btree_node_need_write(b);
 
 	bch2_bset_init_first(b, &b->data->keys);
@@ -827,7 +827,7 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 		closure_wake_up(&c->btree_interior_update_wait);
 	}
 
-	clear_btree_node_dirty(b);
+	clear_btree_node_dirty(c, b);
 	clear_btree_node_need_write(b);
 
 	/*
@@ -1034,7 +1034,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 		bch2_btree_node_iter_advance(node_iter, b);
 
 	bch2_btree_bset_insert_key(iter, b, node_iter, insert);
-	set_btree_node_dirty(b);
+	set_btree_node_dirty(as->c, b);
 	set_btree_node_need_write(b);
 }
 

fs/bcachefs/btree_update_leaf.c

@@ -191,7 +191,7 @@ static bool btree_insert_key_leaf(struct btree_trans *trans,
 	bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 
 	if (unlikely(!btree_node_dirty(b)))
-		set_btree_node_dirty(b);
+		set_btree_node_dirty(c, b);
 
 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 	u64s_added = (int) bset_u64s(t) - old_u64s;

fs/bcachefs/journal_reclaim.c

@@ -547,6 +547,12 @@ void bch2_journal_reclaim(struct journal *j)
 
 		if (j->prereserved.reserved * 2 > j->prereserved.remaining)
 			min_nr = 1;
+
+		if ((atomic_read(&c->btree_cache.dirty) * 4 >
+		     c->btree_cache.used  * 3) ||
+		    (c->btree_key_cache.nr_dirty * 4 >
+		     c->btree_key_cache.nr_keys))
+			min_nr = 1;
 	} while (journal_flush_pins(j, seq_to_flush, min_nr));
 
 	if (!bch2_journal_error(j))