bcachefs: Fix a deadlock in bch2_btree_node_get_sibling()

There was a bad interaction with bch2_btree_iter_set_pos_same_leaf(),
which can leave a btree node locked that is just outside iter->pos,
breaking the lock ordering checks in __bch2_btree_node_lock(). Ideally
we should get rid of this corner case, but for now fix it locally with
verbose comments.

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

4 files changed, 29 insertions, 11 deletions

diff --git a/fs/bcachefs/btree_cache.c b/fs/bcachefs/btree_cache.c
index 732bbd377e12..fc69f6855d29 100644
--- a/fs/bcachefs/btree_cache.c
+++ b/fs/bcachefs/btree_cache.c
@@ -850,6 +850,18 @@ struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 	if (!parent)
 		return NULL;
 
+	/*
+	 * There's a corner case where a btree_iter might have a node locked
+	 * that is just outside its current pos - when
+	 * bch2_btree_iter_set_pos_same_leaf() gets to the end of the node.
+	 *
+	 * But the lock ordering checks in __bch2_btree_node_lock() go off of
+	 * iter->pos, not the node's key: so if the iterator is marked as
+	 * needing to be traversed, we risk deadlock if we don't bail out here:
+	 */
+	if (iter->uptodate >= BTREE_ITER_NEED_TRAVERSE)
+		return ERR_PTR(-EINTR);
+
 	if (!bch2_btree_node_relock(iter, level + 1)) {
 		ret = ERR_PTR(-EINTR);
 		goto out;
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index 29929298a1a9..6abcbe3debe5 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -205,8 +205,9 @@ bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 		if (!linked->nodes_locked)
 			continue;
 
-		/* * Must lock btree nodes in key order: */
-		if (__btree_iter_cmp(iter->btree_id, pos, linked) < 0)
+		/* Must lock btree nodes in key order: */
+		if ((cmp_int(iter->btree_id, linked->btree_id) ?:
+		     bkey_cmp(pos, linked->pos)) < 0)
 			ret = false;
 
 		/*
@@ -1320,6 +1321,16 @@ void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_
 
 	btree_iter_advance_to_pos(iter, l, -1);
 
+	/*
+	 * XXX:
+	 * keeping a node locked that's outside (even just outside) iter->pos
+	 * breaks __bch2_btree_node_lock(). This seems to only affect
+	 * bch2_btree_node_get_sibling so for now it's fixed there, but we
+	 * should try to get rid of this corner case.
+	 *
+	 * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
+	 */
+
 	if (bch2_btree_node_iter_end(&l->iter) &&
 	    btree_iter_pos_after_node(iter, l->b))
 		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
@@ -2194,6 +2205,7 @@ void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
 		bch2_trans_preload_mem(trans, expected_mem_bytes);
 
 #ifdef CONFIG_BCACHEFS_DEBUG
+	trans->pid = current->pid;
 	mutex_lock(&c->btree_trans_lock);
 	list_add(&trans->list, &c->btree_trans_list);
 	mutex_unlock(&c->btree_trans_lock);
@@ -2232,7 +2244,7 @@ void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
 
 	mutex_lock(&c->btree_trans_lock);
 	list_for_each_entry(trans, &c->btree_trans_list, list) {
-		pr_buf(out, "%ps\n", (void *) trans->ip);
+		pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
 
 		trans_for_each_iter(trans, iter) {
 			if (!iter->nodes_locked)
diff --git a/fs/bcachefs/btree_iter.h b/fs/bcachefs/btree_iter.h
index 841a5834f1a8..ab35fcd8b8b4 100644
--- a/fs/bcachefs/btree_iter.h
+++ b/fs/bcachefs/btree_iter.h
@@ -172,17 +172,10 @@ void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *, struct bpos);
 void __bch2_btree_iter_set_pos(struct btree_iter *, struct bpos, bool);
 void bch2_btree_iter_set_pos(struct btree_iter *, struct bpos);
 
-static inline int __btree_iter_cmp(enum btree_id id,
-				   struct bpos pos,
-				   const struct btree_iter *r)
-{
-	return cmp_int(id, r->btree_id) ?: bkey_cmp(pos, r->pos);
-}
-
 static inline int btree_iter_cmp(const struct btree_iter *l,
 				 const struct btree_iter *r)
 {
-	return __btree_iter_cmp(l->btree_id, l->pos, r);
+	return cmp_int(l->btree_id, r->btree_id) ?: bkey_cmp(l->pos, r->pos);
 }
 
 /*
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index b86d7369eb2d..e97248ca3aa2 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -284,6 +284,7 @@ struct btree_trans {
 #ifdef CONFIG_BCACHEFS_DEBUG
 	struct list_head	list;
 	struct btree		*locking;
+	pid_t			pid;
 #endif
 	unsigned long		ip;