bcachefs: added lock held time stats

We now record the length of time btree locks are held and expose this in debugfs.

Enabled via CONFIG_BCACHEFS_LOCK_TIME_STATS.

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

7 files changed, 145 insertions, 15 deletions

diff --git a/fs/bcachefs/Kconfig b/fs/bcachefs/Kconfig
index d2eb65e9032b..7ae85900e5b4 100644
--- a/fs/bcachefs/Kconfig
+++ b/fs/bcachefs/Kconfig
@@ -59,6 +59,12 @@ config BCACHEFS_TESTS
 	help
 	Include some unit and performance tests for the core btree code
 
+config BCACHEFS_LOCK_TIME_STATS
+       bool "bcachefs lock time statistics"
+       depends on BCACHEFS_FS
+       help
+       Expose statistics for how long we held a lock in debugfs
+
 config BCACHEFS_NO_LATENCY_ACCT
 	bool "disable latency accounting and time stats"
 	depends on BCACHEFS_FS
diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index 7020eee5de21..9cd6f840b71a 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -319,6 +319,8 @@ BCH_DEBUG_PARAMS_DEBUG()
 #undef BCH_DEBUG_PARAM
 #endif
 
+#define BCH_LOCK_TIME_NR 128
+
 #define BCH_TIME_STATS()			\
 	x(btree_node_mem_alloc)			\
 	x(btree_node_split)			\
@@ -527,6 +529,11 @@ struct btree_debug {
 	unsigned		id;
 };
 
+struct lock_held_stats {
+	struct bch2_time_stats	times[BCH_LOCK_TIME_NR];
+	const char		*names[BCH_LOCK_TIME_NR];
+};
+
 struct bch_fs_pcpu {
 	u64			sectors_available;
 };
@@ -920,6 +927,8 @@ mempool_t		bio_bounce_pages;
 	bool			promote_whole_extents;
 
 	struct bch2_time_stats	times[BCH_TIME_STAT_NR];
+
+	struct lock_held_stats	lock_held_stats;
 };
 
 static inline void bch2_set_ra_pages(struct bch_fs *c, unsigned ra_pages)
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index d708bf32d408..30958cbb9532 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -177,7 +177,7 @@ bool __bch2_btree_node_relock(struct btree_trans *trans,
 	if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 	    (btree_node_lock_seq_matches(path, b, level) &&
 	     btree_node_lock_increment(trans, b, level, want))) {
-		mark_btree_node_locked(path, level, want);
+		mark_btree_node_locked(trans, path, level, want);
 		return true;
 	}
 fail:
@@ -230,7 +230,7 @@ bool bch2_btree_node_upgrade(struct btree_trans *trans,
 
 	return false;
 success:
-	mark_btree_node_intent_locked(path, level);
+	mark_btree_node_intent_locked(trans, path, level);
 	return true;
 }
 
@@ -1161,7 +1161,7 @@ void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 			    t != BTREE_NODE_UNLOCKED) {
 				btree_node_unlock(trans, path, b->c.level);
 				six_lock_increment(&b->c.lock, (enum six_lock_type) t);
-				mark_btree_node_locked(path, b->c.level, (enum six_lock_type) t);
+				mark_btree_node_locked(trans, path, b->c.level, (enum six_lock_type) t);
 			}
 
 			btree_path_level_init(trans, path, b);
@@ -1238,7 +1238,7 @@ static inline int btree_path_lock_root(struct btree_trans *trans,
 			for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 				path->l[i].b = NULL;
 
-			mark_btree_node_locked(path, path->level, lock_type);
+			mark_btree_node_locked(trans, path, path->level, lock_type);
 			btree_path_level_init(trans, path, b);
 			return 0;
 		}
@@ -1402,7 +1402,7 @@ static __always_inline int btree_path_down(struct btree_trans *trans,
 	if (unlikely(ret))
 		goto err;
 
-	mark_btree_node_locked(path, level, lock_type);
+	mark_btree_node_locked(trans, path, level, lock_type);
 	btree_path_level_init(trans, path, b);
 
 	if (likely(!trans->journal_replay_not_finished &&
@@ -3273,6 +3273,15 @@ void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
 	trans->journal_replay_not_finished =
 		!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
 
+	while (c->lock_held_stats.names[trans->lock_name_idx] != fn
+	       && c->lock_held_stats.names[trans->lock_name_idx] != 0)
+		trans->lock_name_idx++;
+
+	if (trans->lock_name_idx >= BCH_LOCK_TIME_NR)
+		pr_warn_once("lock_times array not big enough!");
+	else
+		c->lock_held_stats.names[trans->lock_name_idx] = fn;
+
 	bch2_trans_alloc_paths(trans, c);
 
 	if (expected_mem_bytes) {
diff --git a/fs/bcachefs/btree_key_cache.c b/fs/bcachefs/btree_key_cache.c
index baf1f25b91ca..e5a29240bbcc 100644
--- a/fs/bcachefs/btree_key_cache.c
+++ b/fs/bcachefs/btree_key_cache.c
@@ -382,7 +382,7 @@ retry:
 		if (!ck)
 			goto retry;
 
-		mark_btree_node_locked(path, 0, SIX_LOCK_intent);
+		mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 		path->locks_want = 1;
 	} else {
 		enum six_lock_type lock_want = __btree_lock_want(path, 0);
@@ -403,7 +403,7 @@ retry:
 			goto retry;
 		}
 
-		mark_btree_node_locked(path, 0, lock_want);
+		mark_btree_node_locked(trans, path, 0, lock_want);
 	}
 
 	path->l[0].lock_seq	= ck->c.lock.state.seq;
diff --git a/fs/bcachefs/btree_locking.h b/fs/bcachefs/btree_locking.h
index 230f376993ae..b8708466c4e3 100644
--- a/fs/bcachefs/btree_locking.h
+++ b/fs/bcachefs/btree_locking.h
@@ -57,7 +57,8 @@ static inline void mark_btree_node_unlocked(struct btree_path *path,
 	path->nodes_intent_locked &= ~(1 << level);
 }
 
-static inline void mark_btree_node_locked(struct btree_path *path,
+static inline void mark_btree_node_locked_noreset(struct btree_trans *trans,
+					  struct btree_path *path,
 					  unsigned level,
 					  enum six_lock_type type)
 {
@@ -69,10 +70,22 @@ static inline void mark_btree_node_locked(struct btree_path *path,
 	path->nodes_intent_locked |= type << level;
 }
 
-static inline void mark_btree_node_intent_locked(struct btree_path *path,
+static inline void mark_btree_node_locked(struct btree_trans *trans,
+					  struct btree_path *path,
+					  unsigned level,
+					  enum six_lock_type type)
+{
+	mark_btree_node_locked_noreset(trans, path, level, type);
+#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
+	path->l[level].lock_taken_time = ktime_get_ns();
+#endif
+}
+
+static inline void mark_btree_node_intent_locked(struct btree_trans *trans,
+						 struct btree_path *path,
 						 unsigned level)
 {
-	mark_btree_node_locked(path, level, SIX_LOCK_intent);
+	mark_btree_node_locked_noreset(trans, path, level, SIX_LOCK_intent);
 }
 
 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
@@ -101,8 +114,18 @@ static inline void btree_node_unlock(struct btree_trans *trans,
 
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
 
-	if (lock_type != BTREE_NODE_UNLOCKED)
+	if (lock_type != BTREE_NODE_UNLOCKED) {
 		six_unlock_type(&path->l[level].b->c.lock, lock_type);
+#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
+		if (trans->lock_name_idx < BCH_LOCK_TIME_NR) {
+			struct bch_fs *c = trans->c;
+
+			__bch2_time_stats_update(&c->lock_held_stats.times[trans->lock_name_idx],
+					       path->l[level].lock_taken_time,
+						 ktime_get_ns());
+		}
+#endif
+	}
 	mark_btree_node_unlocked(path, level);
 }
 
@@ -196,10 +219,17 @@ static inline bool btree_node_lock(struct btree_trans *trans,
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
 	EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
 
-	return likely(six_trylock_type(&b->c.lock, type)) ||
+	if (likely(six_trylock_type(&b->c.lock, type)) ||
 		btree_node_lock_increment(trans, b, level, type) ||
 		__bch2_btree_node_lock(trans, path, b, pos, level, type,
-				       should_sleep_fn, p, ip);
+				       should_sleep_fn, p, ip)) {
+#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
+		path->l[b->c.level].lock_taken_time = ktime_get_ns();
+#endif
+		return true;
+	} else {
+		return false;
+	}
 }
 
 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned);
@@ -252,5 +282,3 @@ static inline void bch2_btree_node_lock_write(struct btree_trans *trans,
 }
 
 #endif /* _BCACHEFS_BTREE_LOCKING_H */
-
-
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index b184ec512499..8cf3ef749020 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -251,6 +251,9 @@ struct btree_path {
 		struct btree	*b;
 		struct btree_node_iter iter;
 		u32		lock_seq;
+#ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
+		u64             lock_taken_time;
+#endif
 	}			l[BTREE_MAX_DEPTH];
 #ifdef CONFIG_BCACHEFS_DEBUG
 	unsigned long		ip_allocated;
@@ -436,6 +439,7 @@ struct btree_trans {
 	unsigned		journal_u64s;
 	unsigned		journal_preres_u64s;
 	struct replicas_delta_list *fs_usage_deltas;
+	int                      lock_name_idx;
 };
 
 #define BTREE_FLAGS()							\
diff --git a/fs/bcachefs/debug.c b/fs/bcachefs/debug.c
index 0f25b75e3de7..45f5229f20eb 100644
--- a/fs/bcachefs/debug.c
+++ b/fs/bcachefs/debug.c
@@ -638,6 +638,75 @@ static const struct file_operations journal_pins_ops = {
 	.read		= bch2_journal_pins_read,
 };
 
+static int lock_held_stats_open(struct inode *inode, struct file *file)
+{
+	struct bch_fs *c = inode->i_private;
+	struct dump_iter *i;
+
+	i = kzalloc(sizeof(struct dump_iter), GFP_KERNEL);
+
+	if (!i)
+		return -ENOMEM;
+
+	i->iter = 0;
+	i->c    = c;
+	i->buf  = PRINTBUF;
+	file->private_data = i;
+
+	return 0;
+}
+
+static int lock_held_stats_release(struct inode *inode, struct file *file)
+{
+	struct dump_iter *i = file->private_data;
+
+	printbuf_exit(&i->buf);
+	kfree(i);
+
+	return 0;
+}
+
+static ssize_t lock_held_stats_read(struct file *file, char __user *buf,
+				      size_t size, loff_t *ppos)
+{
+	struct dump_iter        *i = file->private_data;
+	struct lock_held_stats *lhs = &i->c->lock_held_stats;
+	int err;
+
+	i->ubuf = buf;
+	i->size = size;
+	i->ret  = 0;
+
+	while (lhs->names[i->iter] != 0 && i->iter < BCH_LOCK_TIME_NR) {
+		err = flush_buf(i);
+		if (err)
+			return err;
+
+		if (!i->size)
+			break;
+
+		prt_printf(&i->buf, "%s:", lhs->names[i->iter]);
+		prt_newline(&i->buf);
+		printbuf_indent_add(&i->buf, 8);
+		bch2_time_stats_to_text(&i->buf, &lhs->times[i->iter]);
+		printbuf_indent_sub(&i->buf, 8);
+		prt_newline(&i->buf);
+		i->iter++;
+	}
+
+	if (i->buf.allocation_failure)
+		return -ENOMEM;
+
+	return i->ret;
+}
+
+static const struct file_operations lock_held_stats_op = {
+	.owner = THIS_MODULE,
+	.open = lock_held_stats_open,
+	.release = lock_held_stats_release,
+	.read = lock_held_stats_read,
+};
+
 void bch2_fs_debug_exit(struct bch_fs *c)
 {
 	if (!IS_ERR_OR_NULL(c->fs_debug_dir))
@@ -668,6 +737,11 @@ void bch2_fs_debug_init(struct bch_fs *c)
 	debugfs_create_file("journal_pins", 0400, c->fs_debug_dir,
 			    c->btree_debug, &journal_pins_ops);
 
+	if (IS_ENABLED(CONFIG_BCACHEFS_LOCK_TIME_STATS)) {
+		debugfs_create_file("lock_held_stats", 0400, c->fs_debug_dir,
+				c, &lock_held_stats_op);
+	}
+
 	c->btree_debug_dir = debugfs_create_dir("btrees", c->fs_debug_dir);
 	if (IS_ERR_OR_NULL(c->btree_debug_dir))
 		return;