1 files changed, 174 insertions, 105 deletions

diff --git a/libbcachefs/btree_locking.h b/libbcachefs/btree_locking.h
index 205c6b59..3bc490bc 100644
--- a/libbcachefs/btree_locking.h
+++ b/libbcachefs/btree_locking.h
@@ -14,68 +14,71 @@
 
 #include "btree_iter.h"
 
+extern struct lock_class_key bch2_btree_node_lock_key;
+
 static inline bool is_btree_node(struct btree_path *path, unsigned l)
 {
 	return l < BTREE_MAX_DEPTH && !IS_ERR_OR_NULL(path->l[l].b);
 }
 
+static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
+{
+	return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
+		? &trans->c->btree_transaction_stats[trans->fn_idx]
+		: NULL;
+}
+
 /* matches six lock types */
 enum btree_node_locked_type {
 	BTREE_NODE_UNLOCKED		= -1,
 	BTREE_NODE_READ_LOCKED		= SIX_LOCK_read,
 	BTREE_NODE_INTENT_LOCKED	= SIX_LOCK_intent,
+	BTREE_NODE_WRITE_LOCKED		= SIX_LOCK_write,
 };
 
 static inline int btree_node_locked_type(struct btree_path *path,
 					 unsigned level)
 {
-	/*
-	 * We're relying on the fact that if nodes_intent_locked is set
-	 * nodes_locked must be set as well, so that we can compute without
-	 * branches:
-	 */
-	return BTREE_NODE_UNLOCKED +
-		((path->nodes_locked >> level) & 1) +
-		((path->nodes_intent_locked >> level) & 1);
+	return BTREE_NODE_UNLOCKED + ((path->nodes_locked >> (level << 1)) & 3);
 }
 
-static inline bool btree_node_intent_locked(struct btree_path *path,
-					    unsigned level)
+static inline bool btree_node_write_locked(struct btree_path *path, unsigned l)
 {
-	return btree_node_locked_type(path, level) == BTREE_NODE_INTENT_LOCKED;
+	return btree_node_locked_type(path, l) == BTREE_NODE_WRITE_LOCKED;
 }
 
-static inline bool btree_node_read_locked(struct btree_path *path,
-					  unsigned level)
+static inline bool btree_node_intent_locked(struct btree_path *path, unsigned l)
 {
-	return btree_node_locked_type(path, level) == BTREE_NODE_READ_LOCKED;
+	return btree_node_locked_type(path, l) == BTREE_NODE_INTENT_LOCKED;
 }
 
-static inline bool btree_node_locked(struct btree_path *path, unsigned level)
+static inline bool btree_node_read_locked(struct btree_path *path, unsigned l)
 {
-	return path->nodes_locked & (1 << level);
+	return btree_node_locked_type(path, l) == BTREE_NODE_READ_LOCKED;
 }
 
-static inline void mark_btree_node_unlocked(struct btree_path *path,
-					    unsigned level)
+static inline bool btree_node_locked(struct btree_path *path, unsigned level)
 {
-	path->nodes_locked &= ~(1 << level);
-	path->nodes_intent_locked &= ~(1 << level);
+	return btree_node_locked_type(path, level) != BTREE_NODE_UNLOCKED;
 }
 
-static inline void mark_btree_node_locked_noreset(struct btree_trans *trans,
-					  struct btree_path *path,
-					  unsigned level,
-					  enum six_lock_type type)
+static inline void mark_btree_node_locked_noreset(struct btree_path *path,
+						  unsigned level,
+						  enum btree_node_locked_type type)
 {
 	/* relying on this to avoid a branch */
 	BUILD_BUG_ON(SIX_LOCK_read   != 0);
 	BUILD_BUG_ON(SIX_LOCK_intent != 1);
 
-	BUG_ON(trans->in_traverse_all && path->sorted_idx > trans->traverse_all_idx);
+	path->nodes_locked &= ~(3U << (level << 1));
+	path->nodes_locked |= (type + 1) << (level << 1);
+}
 
-	path->nodes_locked |= 1 << level;
-	path->nodes_intent_locked |= type << level;
+static inline void mark_btree_node_unlocked(struct btree_path *path,
+					    unsigned level)
+{
+	EBUG_ON(btree_node_write_locked(path, level));
+	mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
 }
 
 static inline void mark_btree_node_locked(struct btree_trans *trans,
@@ -83,19 +86,12 @@ static inline void mark_btree_node_locked(struct btree_trans *trans,
 					  unsigned level,
 					  enum six_lock_type type)
 {
-	mark_btree_node_locked_noreset(trans, path, level, type);
+	mark_btree_node_locked_noreset(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_noreset(trans, path, level, SIX_LOCK_intent);
-}
-
 static inline enum six_lock_type __btree_lock_want(struct btree_path *path, int level)
 {
 	return level < path->locks_want
@@ -115,13 +111,6 @@ btree_lock_want(struct btree_path *path, int level)
 	return BTREE_NODE_UNLOCKED;
 }
 
-static inline struct btree_transaction_stats *btree_trans_stats(struct btree_trans *trans)
-{
-	return trans->fn_idx < ARRAY_SIZE(trans->c->btree_transaction_stats)
-		? &trans->c->btree_transaction_stats[trans->fn_idx]
-		: NULL;
-}
-
 static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 					      struct btree_path *path, unsigned level)
 {
@@ -135,6 +124,8 @@ static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 #endif
 }
 
+/* unlock: */
+
 static inline void btree_node_unlock(struct btree_trans *trans,
 				     struct btree_path *path, unsigned level)
 {
@@ -149,59 +140,91 @@ static inline void btree_node_unlock(struct btree_trans *trans,
 	mark_btree_node_unlocked(path, level);
 }
 
+static inline int btree_path_lowest_level_locked(struct btree_path *path)
+{
+	return __ffs(path->nodes_locked) >> 1;
+}
+
+static inline int btree_path_highest_level_locked(struct btree_path *path)
+{
+	return __fls(path->nodes_locked) >> 1;
+}
+
 static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
 					    struct btree_path *path)
 {
 	btree_path_set_dirty(path, BTREE_ITER_NEED_RELOCK);
 
 	while (path->nodes_locked)
-		btree_node_unlock(trans, path, __ffs(path->nodes_locked));
+		btree_node_unlock(trans, path, btree_path_lowest_level_locked(path));
 }
 
-static inline enum bch_time_stats lock_to_time_stat(enum six_lock_type type)
+/*
+ * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
+ * succeed:
+ */
+static inline void
+bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
+				     struct btree *b)
 {
-	switch (type) {
-	case SIX_LOCK_read:
-		return BCH_TIME_btree_lock_contended_read;
-	case SIX_LOCK_intent:
-		return BCH_TIME_btree_lock_contended_intent;
-	case SIX_LOCK_write:
-		return BCH_TIME_btree_lock_contended_write;
-	default:
-		BUG();
-	}
+	struct btree_path *linked;
+
+	EBUG_ON(path->l[b->c.level].b != b);
+	EBUG_ON(path->l[b->c.level].lock_seq + 1 != b->c.lock.state.seq);
+	EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
+
+	mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
+
+	trans_for_each_path_with_node(trans, b, linked)
+		linked->l[b->c.level].lock_seq += 2;
+
+	six_unlock_write(&b->c.lock);
+}
+
+void bch2_btree_node_unlock_write(struct btree_trans *,
+			struct btree_path *, struct btree *);
+
+/* lock: */
+
+static inline int __must_check
+btree_node_lock_nopath(struct btree_trans *trans,
+		       struct btree_bkey_cached_common *b,
+		       enum six_lock_type type)
+{
+	six_lock_type(&b->lock, type, NULL, NULL);
+	return 0;
+}
+
+static inline void btree_node_lock_nopath_nofail(struct btree_trans *trans,
+					 struct btree_bkey_cached_common *b,
+					 enum six_lock_type type)
+{
+	int ret = btree_node_lock_nopath(trans, b, type);
+
+	BUG_ON(ret);
 }
 
 static inline int btree_node_lock_type(struct btree_trans *trans,
 				       struct btree_path *path,
-				       struct btree *b,
+				       struct btree_bkey_cached_common *b,
 				       struct bpos pos, unsigned level,
 				       enum six_lock_type type,
 				       six_lock_should_sleep_fn should_sleep_fn, void *p)
 {
-	struct bch_fs *c = trans->c;
-	u64 start_time;
 	int ret;
 
-	if (six_trylock_type(&b->c.lock, type))
+	if (six_trylock_type(&b->lock, type))
 		return 0;
 
-	start_time = local_clock();
-
 	trans->locking_path_idx = path->idx;
 	trans->locking_pos	= pos;
 	trans->locking_btree_id	= path->btree_id;
 	trans->locking_level	= level;
 	trans->locking_lock_type = type;
-	trans->locking		= &b->c;
-	ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p);
+	trans->locking		= b;
+	ret = six_lock_type(&b->lock, type, should_sleep_fn, p);
 	trans->locking = NULL;
-
-	if (ret)
-		return ret;
-
-	bch2_time_stats_update(&c->times[lock_to_time_stat(type)], start_time);
-	return 0;
+	return ret;
 }
 
 /*
@@ -209,15 +232,16 @@ static inline int btree_node_lock_type(struct btree_trans *trans,
  * iterators:
  */
 static inline bool btree_node_lock_increment(struct btree_trans *trans,
-					     struct btree *b, unsigned level,
+					     struct btree_bkey_cached_common *b,
+					     unsigned level,
 					     enum btree_node_locked_type want)
 {
 	struct btree_path *path;
 
 	trans_for_each_path(trans, path)
-		if (path->l[level].b == b &&
+		if (&path->l[level].b->c == b &&
 		    btree_node_locked_type(path, level) >= want) {
-			six_lock_increment(&b->c.lock, want);
+			six_lock_increment(&b->lock, want);
 			return true;
 		}
 
@@ -225,14 +249,16 @@ static inline bool btree_node_lock_increment(struct btree_trans *trans,
 }
 
 int __bch2_btree_node_lock(struct btree_trans *, struct btree_path *,
-			   struct btree *, struct bpos, unsigned,
+			   struct btree_bkey_cached_common *,
+			   struct bpos, unsigned,
 			   enum six_lock_type,
 			   six_lock_should_sleep_fn, void *,
 			   unsigned long);
 
 static inline int btree_node_lock(struct btree_trans *trans,
 			struct btree_path *path,
-			struct btree *b, struct bpos pos, unsigned level,
+			struct btree_bkey_cached_common *b,
+			struct bpos pos, unsigned level,
 			enum six_lock_type type,
 			six_lock_should_sleep_fn should_sleep_fn, void *p,
 			unsigned long ip)
@@ -242,67 +268,97 @@ static inline int btree_node_lock(struct btree_trans *trans,
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
 	EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
 
-	if (likely(six_trylock_type(&b->c.lock, type)) ||
+	if (likely(six_trylock_type(&b->lock, type)) ||
 	    btree_node_lock_increment(trans, b, level, type) ||
 	    !(ret = __bch2_btree_node_lock(trans, path, b, pos, level, type,
 					   should_sleep_fn, p, ip))) {
 #ifdef CONFIG_BCACHEFS_LOCK_TIME_STATS
-		path->l[b->c.level].lock_taken_time = ktime_get_ns();
+		path->l[b->level].lock_taken_time = ktime_get_ns();
 #endif
 	}
 
 	return ret;
 }
 
+void __bch2_btree_node_lock_write(struct btree_trans *, struct btree_bkey_cached_common *);
+
+static inline void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
+					      struct btree_path *path,
+					      struct btree_bkey_cached_common *b)
+{
+	EBUG_ON(&path->l[b->level].b->c != b);
+	EBUG_ON(path->l[b->level].lock_seq != b->lock.state.seq);
+	EBUG_ON(!btree_node_intent_locked(path, b->level));
+
+	/*
+	 * six locks are unfair, and read locks block while a thread wants a
+	 * write lock: thus, we need to tell the cycle detector we have a write
+	 * lock _before_ taking the lock:
+	 */
+	mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_write);
+
+	if (unlikely(!six_trylock_write(&b->lock)))
+		__bch2_btree_node_lock_write(trans, b);
+}
+
+static inline int __must_check
+bch2_btree_node_lock_write(struct btree_trans *trans,
+			   struct btree_path *path,
+			   struct btree_bkey_cached_common *b)
+{
+	bch2_btree_node_lock_write_nofail(trans, path, b);
+	return 0;
+}
+
+/* relock: */
+
+bool bch2_btree_path_relock_norestart(struct btree_trans *,
+				      struct btree_path *, unsigned long);
 bool __bch2_btree_node_relock(struct btree_trans *, struct btree_path *, unsigned);
 
 static inline bool bch2_btree_node_relock(struct btree_trans *trans,
 					  struct btree_path *path, unsigned level)
 {
 	EBUG_ON(btree_node_locked(path, level) &&
-		btree_node_locked_type(path, level) !=
-		__btree_lock_want(path, level));
+		!btree_node_write_locked(path, level) &&
+		btree_node_locked_type(path, level) != __btree_lock_want(path, level));
 
 	return likely(btree_node_locked(path, level)) ||
-		__bch2_btree_node_relock(trans, path, level);
+		(!IS_ERR_OR_NULL(path->l[level].b) &&
+		 __bch2_btree_node_relock(trans, path, level));
 }
 
-/*
- * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
- * succeed:
- */
-static inline void
-bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
-				     struct btree *b)
+static inline int bch2_btree_path_relock(struct btree_trans *trans,
+				struct btree_path *path, unsigned long trace_ip)
 {
-	struct btree_path *linked;
-
-	EBUG_ON(path->l[b->c.level].b != b);
-	EBUG_ON(path->l[b->c.level].lock_seq + 1 != b->c.lock.state.seq);
-
-	trans_for_each_path_with_node(trans, b, linked)
-		linked->l[b->c.level].lock_seq += 2;
+	if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
+		trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
+		return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
+	}
 
-	six_unlock_write(&b->c.lock);
+	return 0;
 }
 
-void bch2_btree_node_unlock_write(struct btree_trans *,
-			struct btree_path *, struct btree *);
+/* upgrade */
 
-void __bch2_btree_node_lock_write(struct btree_trans *, struct btree *);
+bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *,
+			       struct btree_path *, unsigned);
+bool __bch2_btree_path_upgrade(struct btree_trans *,
+			       struct btree_path *, unsigned);
 
-static inline void bch2_btree_node_lock_write(struct btree_trans *trans,
-					      struct btree_path *path,
-					      struct btree *b)
+static inline bool bch2_btree_path_upgrade(struct btree_trans *trans,
+					   struct btree_path *path,
+					   unsigned new_locks_want)
 {
-	EBUG_ON(path->l[b->c.level].b != b);
-	EBUG_ON(path->l[b->c.level].lock_seq != b->c.lock.state.seq);
-	EBUG_ON(!btree_node_intent_locked(path, b->c.level));
+	new_locks_want = min(new_locks_want, BTREE_MAX_DEPTH);
 
-	if (unlikely(!six_trylock_write(&b->c.lock)))
-		__bch2_btree_node_lock_write(trans, b);
+	return path->locks_want < new_locks_want
+		? __bch2_btree_path_upgrade(trans, path, new_locks_want)
+		: path->uptodate == BTREE_ITER_UPTODATE;
 }
 
+/* misc: */
+
 static inline void btree_path_set_should_be_locked(struct btree_path *path)
 {
 	EBUG_ON(!btree_node_locked(path, path->level));
@@ -326,7 +382,20 @@ static inline void btree_path_set_level_up(struct btree_trans *trans,
 	btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 }
 
+/* debug */
+
 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *,
-				struct btree_path *, struct btree *, unsigned);
+				struct btree_path *,
+				struct btree_bkey_cached_common *b,
+				unsigned);
+
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+void bch2_btree_path_verify_locks(struct btree_path *);
+void bch2_trans_verify_locks(struct btree_trans *);
+#else
+static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
+static inline void bch2_trans_verify_locks(struct btree_trans *trans) {}
+#endif
 
 #endif /* _BCACHEFS_BTREE_LOCKING_H */