bcachefs: btree_iter->should_be_locked

Add a field to struct btree_iter for tracking whether it should be
locked - this fixes spurious transaction restarts in
bch2_trans_relock().

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

3 files changed, 34 insertions, 3 deletions

diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index 8840531266d7..683a2e37f23a 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -456,12 +456,20 @@ void bch2_trans_downgrade(struct btree_trans *trans)
 
 /* Btree transaction locking: */
 
+static inline bool btree_iter_should_be_locked(struct btree_trans *trans,
+					       struct btree_iter *iter)
+{
+	return (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) ||
+		iter->should_be_locked;
+}
+
 bool bch2_trans_relock(struct btree_trans *trans)
 {
 	struct btree_iter *iter;
 
 	trans_for_each_iter(trans, iter)
-		if (!bch2_btree_iter_relock(iter, _RET_IP_)) {
+		if (!bch2_btree_iter_relock(iter, _RET_IP_) &&
+		    btree_iter_should_be_locked(trans, iter)) {
 			trace_trans_restart_relock(trans->ip, _RET_IP_,
 					iter->btree_id, &iter->real_pos);
 			return false;
@@ -1420,9 +1428,16 @@ btree_iter_traverse(struct btree_iter *iter)
 int __must_check
 bch2_btree_iter_traverse(struct btree_iter *iter)
 {
+	int ret;
+
 	btree_iter_set_search_pos(iter, btree_iter_search_key(iter));
 
-	return btree_iter_traverse(iter);
+	ret = btree_iter_traverse(iter);
+	if (ret)
+		return ret;
+
+	iter->should_be_locked = true;
+	return 0;
 }
 
 /* Iterate across nodes (leaf and interior nodes) */
@@ -1448,6 +1463,7 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
 	iter->pos = iter->real_pos = b->key.k.p;
 
 	bch2_btree_iter_verify(iter);
+	iter->should_be_locked = true;
 
 	return b;
 }
@@ -1504,6 +1520,7 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
 	iter->pos = iter->real_pos = b->key.k.p;
 
 	bch2_btree_iter_verify(iter);
+	iter->should_be_locked = true;
 
 	return b;
 }
@@ -1520,6 +1537,7 @@ static void btree_iter_set_search_pos(struct btree_iter *iter, struct bpos new_p
 		goto out;
 
 	iter->real_pos = new_pos;
+	iter->should_be_locked = false;
 
 	if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
 		btree_node_unlock(iter, 0);
@@ -1679,6 +1697,7 @@ start:
 
 	bch2_btree_iter_verify_entry_exit(iter);
 	bch2_btree_iter_verify(iter);
+	iter->should_be_locked = true;
 	return k;
 }
 
@@ -1763,6 +1782,7 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
 out:
 	bch2_btree_iter_verify_entry_exit(iter);
 	bch2_btree_iter_verify(iter);
+	iter->should_be_locked = true;
 	return k;
 no_key:
 	/*
@@ -1862,6 +1882,8 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 
 	bch2_btree_iter_verify_entry_exit(iter);
 	bch2_btree_iter_verify(iter);
+	iter->should_be_locked = true;
+
 	return k;
 }
 
@@ -1899,6 +1921,8 @@ struct bkey_s_c bch2_btree_iter_peek_cached(struct btree_iter *iter)
 		bkey_cmp(iter->pos, ck->key.pos));
 	BUG_ON(!ck->valid);
 
+	iter->should_be_locked = true;
+
 	return bkey_i_to_s_c(ck->k);
 }
 
diff --git a/fs/bcachefs/btree_iter.h b/fs/bcachefs/btree_iter.h
index 01b834bf79f7..a2ce711fd61f 100644
--- a/fs/bcachefs/btree_iter.h
+++ b/fs/bcachefs/btree_iter.h
@@ -178,6 +178,7 @@ static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos
 	iter->k.p.offset	= iter->pos.offset	= new_pos.offset;
 	iter->k.p.snapshot	= iter->pos.snapshot	= new_pos.snapshot;
 	iter->k.size = 0;
+	iter->should_be_locked = false;
 }
 
 /* Sort order for locking btree iterators: */
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index 172a90f4db4f..bc0f482b53d2 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -252,7 +252,13 @@ struct btree_iter {
 	u8			idx;
 
 	enum btree_id		btree_id:4;
-	enum btree_iter_uptodate uptodate:4;
+	enum btree_iter_uptodate uptodate:3;
+	/*
+	 * True if we've returned a key (and thus are expected to keep it
+	 * locked), false after set_pos - for avoiding spurious transaction
+	 * restarts in bch2_trans_relock():
+	 */
+	bool			should_be_locked:1;
 	unsigned		level:4,
 				min_depth:4,
 				locks_want:4,