bcachefs: More bset.c microoptimization

Improve a few paper cuts that've shown up during profiling.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/bkey.c

@@ -329,7 +329,7 @@ bool bch2_bkey_pack_key(struct bkey_packed *out, const struct bkey *in,
 void bch2_bkey_unpack(const struct btree *b, struct bkey_i *dst,
 		 const struct bkey_packed *src)
 {
-	dst->k = bkey_unpack_key(b, src);
+	__bkey_unpack_key(b, &dst->k, src);
 
 	memcpy_u64s(&dst->v,
 		    bkeyp_val(&b->format, src),

fs/bcachefs/bset.c

@@ -1548,11 +1548,13 @@ static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
  *    So we've got to search for start_of_range, then after the lookup iterate
  *    past any extents that compare equal to the position we searched for.
  */
+__flatten
 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
 			       struct btree *b, struct bpos *search)
 {
 	struct bset_tree *t;
 	struct bkey_packed p, *packed_search = NULL;
+	struct btree_node_iter_set *pos = iter->data;
 
 	EBUG_ON(bkey_cmp(*search, b->data->min_key) < 0);
 	bset_aux_tree_verify(b);
@@ -1571,11 +1573,17 @@ void bch2_btree_node_iter_init(struct btree_node_iter *iter,
 		return;
 	}
 
-	for_each_bset(b, t)
-		__bch2_btree_node_iter_push(iter, b,
-					   bch2_bset_search(b, t, search,
-							    packed_search, &p),
-					   btree_bkey_last(b, t));
+	for_each_bset(b, t) {
+		struct bkey_packed *k = bch2_bset_search(b, t, search,
+							 packed_search, &p);
+		struct bkey_packed *end = btree_bkey_last(b, t);
+
+		if (k != end)
+			*pos++ = (struct btree_node_iter_set) {
+				__btree_node_key_to_offset(b, k),
+				__btree_node_key_to_offset(b, end)
+			};
+	}
 
 	bch2_btree_node_iter_sort(iter, b);
 }

fs/bcachefs/btree_iter.c

@@ -939,7 +939,7 @@ static void btree_iter_prefetch(struct btree_iter *iter)
 		btree_node_unlock(iter, iter->level);
 }
 
-static inline int btree_iter_down(struct btree_iter *iter)
+static __always_inline int btree_iter_down(struct btree_iter *iter)
 {
 	struct bch_fs *c = iter->trans->c;
 	struct btree_iter_level *l = &iter->l[iter->level];
@@ -948,7 +948,7 @@ static inline int btree_iter_down(struct btree_iter *iter)
 	enum six_lock_type lock_type = __btree_lock_want(iter, level);
 	BKEY_PADDED(k) tmp;
 
-	BUG_ON(!btree_node_locked(iter, iter->level));
+	EBUG_ON(!btree_node_locked(iter, iter->level));
 
 	bch2_bkey_unpack(l->b, &tmp.k,
 			 bch2_btree_node_iter_peek(&l->iter, l->b));