bcachefs: add counters for failed shrinker reclaim

This adds distinct counters for every reason the btree node shrinker can
fail to free an object - if our shrinker isn't making progress, this
will tell us why.

Signed-off-by: Daniel Hill <daniel@gluo.nz>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/btree_cache.c

@@ -16,6 +16,12 @@
 #include <linux/prefetch.h>
 #include <linux/sched/mm.h>
 
+#define BTREE_CACHE_NOT_FREED_INCREMENT(counter) \
+do {						 \
+	if (shrinker_counter)			 \
+		bc->not_freed_##counter++;	 \
+} while (0)
+
 const char * const bch2_btree_node_flags[] = {
 #define x(f)	#f,
 	BTREE_FLAGS()
@@ -238,7 +244,7 @@ static inline struct btree *btree_cache_find(struct btree_cache *bc,
  * this version is for btree nodes that have already been freed (we're not
  * reaping a real btree node)
  */
-static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
+static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush, bool shrinker_counter)
 {
 	struct btree_cache *bc = &c->btree_cache;
 	int ret = 0;
@@ -260,38 +266,64 @@ wait_on_io:
 	if (b->flags & ((1U << BTREE_NODE_dirty)|
 			(1U << BTREE_NODE_read_in_flight)|
 			(1U << BTREE_NODE_write_in_flight))) {
-		if (!flush)
+		if (!flush) {
+			if (btree_node_dirty(b))
+				BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
+			else if (btree_node_read_in_flight(b))
+				BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
+			else if (btree_node_write_in_flight(b))
+				BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 			return -BCH_ERR_ENOMEM_btree_node_reclaim;
+		}
 
 		/* XXX: waiting on IO with btree cache lock held */
 		bch2_btree_node_wait_on_read(b);
 		bch2_btree_node_wait_on_write(b);
 	}
 
-	if (!six_trylock_intent(&b->c.lock))
+	if (!six_trylock_intent(&b->c.lock)) {
+		BTREE_CACHE_NOT_FREED_INCREMENT(lock_intent);
 		return -BCH_ERR_ENOMEM_btree_node_reclaim;
+	}
 
-	if (!six_trylock_write(&b->c.lock))
+	if (!six_trylock_write(&b->c.lock)) {
+		BTREE_CACHE_NOT_FREED_INCREMENT(lock_write);
 		goto out_unlock_intent;
+	}
 
 	/* recheck under lock */
 	if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
 			(1U << BTREE_NODE_write_in_flight))) {
-		if (!flush)
+		if (!flush) {
+			if (btree_node_read_in_flight(b))
+				BTREE_CACHE_NOT_FREED_INCREMENT(read_in_flight);
+			else if (btree_node_write_in_flight(b))
+				BTREE_CACHE_NOT_FREED_INCREMENT(write_in_flight);
 			goto out_unlock;
+		}
 		six_unlock_write(&b->c.lock);
 		six_unlock_intent(&b->c.lock);
 		goto wait_on_io;
 	}
 
-	if (btree_node_noevict(b) ||
-	    btree_node_write_blocked(b) ||
-	    btree_node_will_make_reachable(b))
+	if (btree_node_noevict(b)) {
+		BTREE_CACHE_NOT_FREED_INCREMENT(noevict);
 		goto out_unlock;
+	}
+	if (btree_node_write_blocked(b)) {
+		BTREE_CACHE_NOT_FREED_INCREMENT(write_blocked);
+		goto out_unlock;
+	}
+	if (btree_node_will_make_reachable(b)) {
+		BTREE_CACHE_NOT_FREED_INCREMENT(will_make_reachable);
+		goto out_unlock;
+	}
 
 	if (btree_node_dirty(b)) {
-		if (!flush)
+		if (!flush) {
+			BTREE_CACHE_NOT_FREED_INCREMENT(dirty);
 			goto out_unlock;
+		}
 		/*
 		 * Using the underscore version because we don't want to compact
 		 * bsets after the write, since this node is about to be evicted
@@ -321,14 +353,14 @@ out_unlock_intent:
 	goto out;
 }
 
-static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
+static int btree_node_reclaim(struct bch_fs *c, struct btree *b, bool shrinker_counter)
 {
-	return __btree_node_reclaim(c, b, false);
+	return __btree_node_reclaim(c, b, false, shrinker_counter);
 }
 
 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
 {
-	return __btree_node_reclaim(c, b, true);
+	return __btree_node_reclaim(c, b, true, false);
 }
 
 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
@@ -376,11 +408,12 @@ static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
 		if (touched >= nr)
 			goto out;
 
-		if (!btree_node_reclaim(c, b)) {
+		if (!btree_node_reclaim(c, b, true)) {
 			btree_node_data_free(c, b);
 			six_unlock_write(&b->c.lock);
 			six_unlock_intent(&b->c.lock);
 			freed++;
+			bc->freed++;
 		}
 	}
 restart:
@@ -389,9 +422,11 @@ restart:
 
 		if (btree_node_accessed(b)) {
 			clear_btree_node_accessed(b);
-		} else if (!btree_node_reclaim(c, b)) {
+			bc->not_freed_access_bit++;
+		} else if (!btree_node_reclaim(c, b, true)) {
 			freed++;
 			btree_node_data_free(c, b);
+			bc->freed++;
 
 			bch2_btree_node_hash_remove(bc, b);
 			six_unlock_write(&b->c.lock);
@@ -599,7 +634,7 @@ static struct btree *btree_node_cannibalize(struct bch_fs *c)
 	struct btree *b;
 
 	list_for_each_entry_reverse(b, &bc->live, list)
-		if (!btree_node_reclaim(c, b))
+		if (!btree_node_reclaim(c, b, false))
 			return b;
 
 	while (1) {
@@ -635,7 +670,7 @@ struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_rea
 	 * disk node. Check the freed list before allocating a new one:
 	 */
 	list_for_each_entry(b, freed, list)
-		if (!btree_node_reclaim(c, b)) {
+		if (!btree_node_reclaim(c, b, false)) {
 			list_del_init(&b->list);
 			goto got_node;
 		}
@@ -661,7 +696,7 @@ got_node:
 	 * the list. Check if there's any freed nodes there:
 	 */
 	list_for_each_entry(b2, &bc->freeable, list)
-		if (!btree_node_reclaim(c, b2)) {
+		if (!btree_node_reclaim(c, b2, false)) {
 			swap(b->data, b2->data);
 			swap(b->aux_data, b2->aux_data);
 			btree_node_to_freedlist(bc, b2);
@@ -1280,12 +1315,12 @@ static void prt_btree_cache_line(struct printbuf *out, const struct bch_fs *c,
 	prt_printf(out, " (%u)\n", nr);
 }
 
-void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
+void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc)
 {
-	const struct btree_cache *bc = &c->btree_cache;
+	struct bch_fs *c = container_of(bc, struct bch_fs, btree_cache);
 
 	if (!out->nr_tabstops)
-		printbuf_tabstop_push(out, 24);
+		printbuf_tabstop_push(out, 32);
 
 	prt_btree_cache_line(out, c, "total:",		bc->used);
 	prt_btree_cache_line(out, c, "nr dirty:",	atomic_read(&bc->dirty));
@@ -1294,4 +1329,17 @@ void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
 
 	for (unsigned i = 0; i < ARRAY_SIZE(bc->used_by_btree); i++)
 		prt_btree_cache_line(out, c, bch2_btree_id_str(i), bc->used_by_btree[i]);
+
+	prt_newline(out);
+	prt_printf(out, "freed:\t%u\n", bc->freed);
+	prt_printf(out, "not freed:\n");
+	prt_printf(out, "  dirty\t%u\n", bc->not_freed_dirty);
+	prt_printf(out, "  write in flight\t%u\n", bc->not_freed_write_in_flight);
+	prt_printf(out, "  read in flight\t%u\n", bc->not_freed_read_in_flight);
+	prt_printf(out, "  lock intent failed\t%u\n", bc->not_freed_lock_intent);
+	prt_printf(out, "  lock write failed\t%u\n", bc->not_freed_lock_write);
+	prt_printf(out, "  access bit\t%u\n", bc->not_freed_access_bit);
+	prt_printf(out, "  no evict failed\t%u\n", bc->not_freed_noevict);
+	prt_printf(out, "  write blocked\t%u\n", bc->not_freed_write_blocked);
+	prt_printf(out, "  will make reachable\t%u\n", bc->not_freed_will_make_reachable);
 }

fs/bcachefs/btree_cache.h

@@ -134,6 +134,6 @@ static inline struct btree *btree_node_root(struct bch_fs *c, struct btree *b)
 const char *bch2_btree_id_str(enum btree_id);
 void bch2_btree_pos_to_text(struct printbuf *, struct bch_fs *, const struct btree *);
 void bch2_btree_node_to_text(struct printbuf *, struct bch_fs *, const struct btree *);
-void bch2_btree_cache_to_text(struct printbuf *, const struct bch_fs *);
+void bch2_btree_cache_to_text(struct printbuf *, const struct btree_cache *);
 
 #endif /* _BCACHEFS_BTREE_CACHE_H */

fs/bcachefs/btree_types.h

@@ -163,6 +163,16 @@ struct btree_cache {
 	/* Number of elements in live + freeable lists */
 	unsigned		used;
 	unsigned		reserve;
+	unsigned		freed;
+	unsigned		not_freed_lock_intent;
+	unsigned		not_freed_lock_write;
+	unsigned		not_freed_dirty;
+	unsigned		not_freed_read_in_flight;
+	unsigned		not_freed_write_in_flight;
+	unsigned		not_freed_noevict;
+	unsigned		not_freed_write_blocked;
+	unsigned		not_freed_will_make_reachable;
+	unsigned		not_freed_access_bit;
 	atomic_t		dirty;
 	struct shrinker		*shrink;
 

fs/bcachefs/sysfs.c

@@ -383,7 +383,7 @@ SHOW(bch2_fs)
 		bch2_journal_debug_to_text(out, &c->journal);
 
 	if (attr == &sysfs_btree_cache)
-		bch2_btree_cache_to_text(out, c);
+		bch2_btree_cache_to_text(out, &c->btree_cache);
 
 	if (attr == &sysfs_btree_key_cache)
 		bch2_btree_key_cache_to_text(out, &c->btree_key_cache);