Update bcachefs sources to 446219cb11 bcachefs: Dirent repair code

.bcachefs_revision

@@ -1 +1 @@
-15f6e66e86a97245d967fedcb2f33598c174fd96
+446219cb11af8b6a4c6b837e336bac45f43854c9

libbcachefs/alloc.c

@@ -610,7 +610,7 @@ static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 
 	return  (data_wantness << 9) |
 		(needs_journal_commit << 8) |
-		bucket_gc_gen(ca, b);
+		(bucket_gc_gen(ca, b) / 16);
 }
 
 static inline int bucket_alloc_cmp(alloc_heap *h,

libbcachefs/bset.c

@@ -155,7 +155,7 @@ static void bch2_btree_node_iter_next_check(struct btree_node_iter *_iter,
 	bkey_unpack_key(b, k);
 
 	if (n &&
-	    __btree_node_iter_cmp(b, k, n) > 0) {
+	    bkey_iter_cmp(b, k, n) > 0) {
 		struct btree_node_iter_set *set;
 		struct bkey ku = bkey_unpack_key(b, k);
 		struct bkey nu = bkey_unpack_key(b, n);
@@ -214,10 +214,10 @@ void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 	struct bkey_packed *next = (void *) (where->_data + clobber_u64s);
 #if 0
 	BUG_ON(prev &&
-	       __btree_node_iter_cmp(b, prev, insert) > 0);
+	       bkey_iter_cmp(b, prev, insert) > 0);
 #else
 	if (prev &&
-	    __btree_node_iter_cmp(b, prev, insert) > 0) {
+	    bkey_iter_cmp(b, prev, insert) > 0) {
 		struct bkey k1 = bkey_unpack_key(b, prev);
 		struct bkey k2 = bkey_unpack_key(b, insert);
 		char buf1[100];
@@ -236,10 +236,10 @@ void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
 #endif
 #if 0
 	BUG_ON(next != btree_bkey_last(b, t) &&
-	       __btree_node_iter_cmp(b, insert, next) > 0);
+	       bkey_iter_cmp(b, insert, next) > 0);
 #else
 	if (next != btree_bkey_last(b, t) &&
-	    __btree_node_iter_cmp(b, insert, next) > 0) {
+	    bkey_iter_cmp(b, insert, next) > 0) {
 		struct bkey k1 = bkey_unpack_key(b, insert);
 		struct bkey k2 = bkey_unpack_key(b, next);
 		char buf1[100];
@@ -1289,7 +1289,7 @@ void bch2_bset_delete(struct btree *b,
 __flatten
 static struct bkey_packed *bset_search_write_set(const struct btree *b,
 				struct bset_tree *t,
-				struct bpos search,
+				struct bpos *search,
 				const struct bkey_packed *packed_search)
 {
 	unsigned l = 0, r = t->size;
@@ -1297,7 +1297,7 @@ static struct bkey_packed *bset_search_write_set(const struct btree *b,
 	while (l + 1 != r) {
 		unsigned m = (l + r) >> 1;
 
-		if (bkey_cmp(rw_aux_tree(b, t)[m].k, search) < 0)
+		if (bkey_cmp(rw_aux_tree(b, t)[m].k, *search) < 0)
 			l = m;
 		else
 			r = m;
@@ -1319,7 +1319,7 @@ static int bset_search_tree_slowpath(const struct btree *b,
 __flatten
 static struct bkey_packed *bset_search_tree(const struct btree *b,
 				struct bset_tree *t,
-				struct bpos search,
+				struct bpos *search,
 				const struct bkey_packed *packed_search)
 {
 	struct ro_aux_tree *base = ro_aux_tree_base(b, t);
@@ -1360,7 +1360,7 @@ static struct bkey_packed *bset_search_tree(const struct btree *b,
 				     bkey_mantissa(packed_search, f, n));
 		else
 			n = n * 2 + bset_search_tree_slowpath(b, t,
-						&search, packed_search, n);
+						search, packed_search, n);
 	} while (n < t->size);
 
 	inorder = __eytzinger1_to_inorder(n >> 1, t->size, t->extra);
@@ -1387,10 +1387,9 @@ static struct bkey_packed *bset_search_tree(const struct btree *b,
 __always_inline __flatten
 static struct bkey_packed *bch2_bset_search(struct btree *b,
 				struct bset_tree *t,
-				struct bpos search,
+				struct bpos *search,
 				struct bkey_packed *packed_search,
-				const struct bkey_packed *lossy_packed_search,
+				const struct bkey_packed *lossy_packed_search)
-				bool strictly_greater)
 {
 	struct bkey_packed *m;
 
@@ -1424,7 +1423,7 @@ static struct bkey_packed *bch2_bset_search(struct btree *b,
 		 * start and end - handle that here:
 		 */
 
-		if (bkey_cmp(search, t->max_key) > 0)
+		if (bkey_cmp(*search, t->max_key) > 0)
 			return btree_bkey_last(b, t);
 
 		m = bset_search_tree(b, t, search, lossy_packed_search);
@@ -1433,21 +1432,21 @@ static struct bkey_packed *bch2_bset_search(struct btree *b,
 
 	if (lossy_packed_search)
 		while (m != btree_bkey_last(b, t) &&
-		       !btree_iter_pos_cmp_p_or_unp(b, search, lossy_packed_search,
+		       bkey_iter_cmp_p_or_unp(b, search, lossy_packed_search,
-						    m, strictly_greater))
+					      m) > 0)
 			m = bkey_next(m);
 
 	if (!packed_search)
 		while (m != btree_bkey_last(b, t) &&
-		       !btree_iter_pos_cmp_packed(b, &search, m, strictly_greater))
+		       bkey_iter_pos_cmp(b, search, m) > 0)
 			m = bkey_next(m);
 
 	if (btree_keys_expensive_checks(b)) {
 		struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
 
 		BUG_ON(prev &&
-		       btree_iter_pos_cmp_p_or_unp(b, search, packed_search,
+		       bkey_iter_cmp_p_or_unp(b, search, packed_search,
-						   prev, strictly_greater));
+					      prev) <= 0);
 	}
 
 	return m;
@@ -1455,6 +1454,25 @@ static struct bkey_packed *bch2_bset_search(struct btree *b,
 
 /* Btree node iterator */
 
+static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
+			      struct btree *b,
+			      const struct bkey_packed *k,
+			      const struct bkey_packed *end)
+{
+	if (k != end) {
+		struct btree_node_iter_set *pos;
+
+		btree_node_iter_for_each(iter, pos)
+			;
+
+		BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
+		*pos = (struct btree_node_iter_set) {
+			__btree_node_key_to_offset(b, k),
+			__btree_node_key_to_offset(b, end)
+		};
+	}
+}
+
 void bch2_btree_node_iter_push(struct btree_node_iter *iter,
 			       struct btree *b,
 			       const struct bkey_packed *k,
@@ -1466,17 +1484,15 @@ void bch2_btree_node_iter_push(struct btree_node_iter *iter,
 
 noinline __flatten __attribute__((cold))
 static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
-			      struct btree *b, struct bpos search,
+			      struct btree *b, struct bpos *search)
-			      bool strictly_greater)
 {
 	struct bset_tree *t;
 
-	trace_bkey_pack_pos_fail(&search);
+	trace_bkey_pack_pos_fail(search);
 
 	for_each_bset(b, t)
 		__bch2_btree_node_iter_push(iter, b,
-			bch2_bset_search(b, t, search, NULL, NULL,
+			bch2_bset_search(b, t, search, NULL, NULL),
-					strictly_greater),
 			btree_bkey_last(b, t));
 
 	bch2_btree_node_iter_sort(iter, b);
@@ -1523,18 +1539,17 @@ static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
  *    past any extents that compare equal to the position we searched for.
  */
 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
-			       struct btree *b, struct bpos search,
+			       struct btree *b, struct bpos *search)
-			       bool strictly_greater)
 {
 	struct bset_tree *t;
 	struct bkey_packed p, *packed_search = NULL;
 
-	EBUG_ON(bkey_cmp(search, b->data->min_key) < 0);
+	EBUG_ON(bkey_cmp(*search, b->data->min_key) < 0);
 	bset_aux_tree_verify(b);
 
 	memset(iter, 0, sizeof(*iter));
 
-	switch (bch2_bkey_pack_pos_lossy(&p, search, b)) {
+	switch (bch2_bkey_pack_pos_lossy(&p, *search, b)) {
 	case BKEY_PACK_POS_EXACT:
 		packed_search = &p;
 		break;
@@ -1542,16 +1557,14 @@ void bch2_btree_node_iter_init(struct btree_node_iter *iter,
 		packed_search = NULL;
 		break;
 	case BKEY_PACK_POS_FAIL:
-		btree_node_iter_init_pack_failed(iter, b, search,
+		btree_node_iter_init_pack_failed(iter, b, search);
-						 strictly_greater);
 		return;
 	}
 
 	for_each_bset(b, t)
 		__bch2_btree_node_iter_push(iter, b,
 					   bch2_bset_search(b, t, search,
-							   packed_search, &p,
+							    packed_search, &p),
-							   strictly_greater),
 					   btree_bkey_last(b, t));
 
 	bch2_btree_node_iter_sort(iter, b);
@@ -1685,7 +1698,7 @@ struct bkey_packed *bch2_btree_node_iter_prev_filter(struct btree_node_iter *ite
 			bch2_btree_node_iter_bset_pos(iter, b, t),
 			min_key_type);
 		if (k &&
-		    (!prev || __btree_node_iter_cmp(b, k, prev) > 0)) {
+		    (!prev || bkey_iter_cmp(b, k, prev) > 0)) {
 			prev = k;
 			end = t->end_offset;
 		}

libbcachefs/bset.h

@@ -367,41 +367,6 @@ static inline int bkey_cmp_p_or_unp(const struct btree *b,
 	return __bch2_bkey_cmp_left_packed_format_checked(b, l, r);
 }
 
-/* Returns true if @k is after iterator position @pos */
-static inline bool btree_iter_pos_cmp(struct btree_iter *iter,
-				      const struct bkey *k)
-{
-	int cmp = bkey_cmp(k->p, iter->pos);
-
-	return cmp > 0 ||
-		(cmp == 0 &&
-		 !(iter->flags & BTREE_ITER_IS_EXTENTS) && !bkey_deleted(k));
-}
-
-/* Returns true if @k is after iterator position @pos */
-static inline bool btree_iter_pos_cmp_packed(const struct btree *b,
-					     struct bpos *pos,
-					     const struct bkey_packed *k,
-					     bool strictly_greater)
-{
-	int cmp = bkey_cmp_left_packed(b, k, pos);
-
-	return cmp > 0 ||
-		(cmp == 0 && !strictly_greater && !bkey_deleted(k));
-}
-
-static inline bool btree_iter_pos_cmp_p_or_unp(const struct btree *b,
-					struct bpos pos,
-					const struct bkey_packed *pos_packed,
-					const struct bkey_packed *k,
-					bool strictly_greater)
-{
-	int cmp = bkey_cmp_p_or_unp(b, k, pos_packed, &pos);
-
-	return cmp > 0 ||
-		(cmp == 0 && !strictly_greater && !bkey_deleted(k));
-}
-
 struct bset_tree *bch2_bkey_to_bset(struct btree *, struct bkey_packed *);
 
 struct bkey_packed *bch2_bkey_prev_filter(struct btree *, struct bset_tree *,
@@ -443,7 +408,7 @@ void bch2_btree_node_iter_push(struct btree_node_iter *, struct btree *,
 			      const struct bkey_packed *,
 			      const struct bkey_packed *);
 void bch2_btree_node_iter_init(struct btree_node_iter *, struct btree *,
-			       struct bpos, bool);
+			       struct bpos *);
 void bch2_btree_node_iter_init_from_start(struct btree_node_iter *,
 					  struct btree *);
 struct bkey_packed *bch2_btree_node_iter_bset_pos(struct btree_node_iter *,
@@ -472,11 +437,16 @@ static inline bool bch2_btree_node_iter_end(struct btree_node_iter *iter)
 	return __btree_node_iter_set_end(iter, 0);
 }
 
-static inline int __btree_node_iter_cmp(struct btree *b,
+/*
+ * When keys compare equal, deleted keys compare first:
+ *
+ * XXX: only need to compare pointers for keys that are both within a
+ * btree_node_iterator - we need to break ties for prev() to work correctly
+ */
+static inline int bkey_iter_cmp(struct btree *b,
 				const struct bkey_packed *l,
 				const struct bkey_packed *r)
 {
-	/* When keys compare equal deleted keys come first */
 	return bkey_cmp_packed(b, l, r)
 		?: (int) bkey_deleted(r) - (int) bkey_deleted(l)
 		?: (l > r) - (l < r);
@@ -486,28 +456,27 @@ static inline int btree_node_iter_cmp(struct btree *b,
 				      struct btree_node_iter_set l,
 				      struct btree_node_iter_set r)
 {
-	return __btree_node_iter_cmp(b,
+	return bkey_iter_cmp(b,
 			__btree_node_offset_to_key(b, l.k),
 			__btree_node_offset_to_key(b, r.k));
 }
 
-static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
+/* These assume l (the search key) is not a deleted key: */
-			      struct btree *b,
+static inline int bkey_iter_pos_cmp(struct btree *b,
-			      const struct bkey_packed *k,
+			struct bpos *l,
-			      const struct bkey_packed *end)
+			const struct bkey_packed *r)
 {
-	if (k != end) {
+	return -bkey_cmp_left_packed(b, r, l)
-		struct btree_node_iter_set *pos;
+		?: (int) bkey_deleted(r);
-
-		btree_node_iter_for_each(iter, pos)
-			;
-
-		BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
-		*pos = (struct btree_node_iter_set) {
-			__btree_node_key_to_offset(b, k),
-			__btree_node_key_to_offset(b, end)
-		};
 }
+
+static inline int bkey_iter_cmp_p_or_unp(struct btree *b,
+			struct bpos *l,
+			const struct bkey_packed *l_packed,
+			const struct bkey_packed *r)
+{
+	return -bkey_cmp_p_or_unp(b, r, l_packed, l)
+		?: (int) bkey_deleted(r);
 }
 
 static inline struct bkey_packed *

libbcachefs/btree_io.c

@@ -21,10 +21,7 @@
 
 /* btree_node_iter_large: */
 
-#define btree_node_iter_cmp_heap(h, _l, _r)				\
+#define btree_node_iter_cmp_heap(h, _l, _r) btree_node_iter_cmp(b, _l, _r)
-	__btree_node_iter_cmp(b,					\
-			       __btree_node_offset_to_key(b, (_l).k),	\
-			       __btree_node_offset_to_key(b, (_r).k))
 
 void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
 				     struct btree *b,

libbcachefs/btree_iter.c

@@ -23,6 +23,39 @@ static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
 		iter->l[l].b != BTREE_ITER_NOT_END;
 }
 
+/* Returns < 0 if @k is before iter pos, > 0 if @k is after */
+static inline int __btree_iter_pos_cmp(struct btree_iter *iter,
+				       const struct btree *b,
+				       const struct bkey_packed *k,
+				       bool interior_node)
+{
+	int cmp = bkey_cmp_left_packed(b, k, &iter->pos);
+
+	if (cmp)
+		return cmp;
+	if (bkey_deleted(k))
+		return -1;
+
+	/*
+	 * Normally, for extents we want the first key strictly greater than
+	 * the iterator position - with the exception that for interior nodes,
+	 * we don't want to advance past the last key if the iterator position
+	 * is POS_MAX:
+	 */
+	if (iter->flags & BTREE_ITER_IS_EXTENTS &&
+	    (!interior_node ||
+	     bkey_cmp_left_packed_byval(b, k, POS_MAX)))
+		return -1;
+	return 1;
+}
+
+static inline int btree_iter_pos_cmp(struct btree_iter *iter,
+				     const struct btree *b,
+				     const struct bkey_packed *k)
+{
+	return __btree_iter_pos_cmp(iter, b, k, b->level != 0);
+}
+
 /* Btree node locking: */
 
 /*
@@ -389,8 +422,7 @@ static void __bch2_btree_iter_verify(struct btree_iter *iter,
 	k = b->level || iter->flags & BTREE_ITER_IS_EXTENTS
 		? bch2_btree_node_iter_prev_filter(&tmp, b, KEY_TYPE_DISCARD)
 		: bch2_btree_node_iter_prev_all(&tmp, b);
-	if (k && btree_iter_pos_cmp_packed(b, &iter->pos, k,
+	if (k && btree_iter_pos_cmp(iter, b, k) > 0) {
-				iter->flags & BTREE_ITER_IS_EXTENTS)) {
 		char buf[100];
 		struct bkey uk = bkey_unpack_key(b, k);
 
@@ -400,8 +432,7 @@ static void __bch2_btree_iter_verify(struct btree_iter *iter,
 	}
 
 	k = bch2_btree_node_iter_peek_all(&l->iter, b);
-	if (k && !btree_iter_pos_cmp_packed(b, &iter->pos, k,
+	if (k && btree_iter_pos_cmp(iter, b, k) < 0) {
-				iter->flags & BTREE_ITER_IS_EXTENTS)) {
 		char buf[100];
 		struct bkey uk = bkey_unpack_key(b, k);
 
@@ -453,8 +484,7 @@ static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 
 	/* didn't find the bset in the iterator - might have to readd it: */
 	if (new_u64s &&
-	    btree_iter_pos_cmp_packed(b, &iter->pos, where,
+	    btree_iter_pos_cmp(iter, b, where) > 0) {
-				      iter->flags & BTREE_ITER_IS_EXTENTS)) {
 		btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 
 		bch2_btree_node_iter_push(node_iter, b, where, end);
@@ -474,8 +504,7 @@ found:
 		return;
 
 	if (new_u64s &&
-	    btree_iter_pos_cmp_packed(b, &iter->pos, where,
+	    btree_iter_pos_cmp(iter, b, where) > 0) {
-				iter->flags & BTREE_ITER_IS_EXTENTS)) {
 		set->k = offset;
 	} else if (set->k < offset + clobber_u64s) {
 		set->k = offset + new_u64s;
@@ -515,9 +544,8 @@ iter_current_key_not_modified:
 	 * always point to the key for the child node the btree iterator points
 	 * to.
 	 */
-	if (b->level && new_u64s && !bkey_deleted(where) &&
+	if (b->level && new_u64s &&
-	    btree_iter_pos_cmp_packed(b, &iter->pos, where,
+	    btree_iter_pos_cmp(iter, b, where) > 0) {
-				iter->flags & BTREE_ITER_IS_EXTENTS)) {
 		struct bset_tree *t;
 		struct bkey_packed *k;
 
@@ -528,7 +556,7 @@ iter_current_key_not_modified:
 			k = bch2_bkey_prev_all(b, t,
 				bch2_btree_node_iter_bset_pos(node_iter, b, t));
 			if (k &&
-			    __btree_node_iter_cmp(b, k, where) > 0) {
+			    bkey_iter_cmp(b, k, where) > 0) {
 				struct btree_node_iter_set *set;
 				unsigned offset =
 					__btree_node_key_to_offset(b, bkey_next(k));
@@ -609,9 +637,23 @@ static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter,
 			bch2_btree_node_iter_peek(&l->iter, l->b));
 }
 
-static inline void __btree_iter_advance(struct btree_iter_level *l)
+static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
+					     struct btree_iter_level *l,
+					     int max_advance)
 {
+	struct bkey_packed *k;
+	int nr_advanced = 0;
+
+	while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
+	       btree_iter_pos_cmp(iter, l->b, k) < 0) {
+		if (max_advance > 0 && nr_advanced >= max_advance)
+			return false;
+
 		bch2_btree_node_iter_advance(&l->iter, l->b);
+		nr_advanced++;
+	}
+
+	return true;
 }
 
 /*
@@ -656,8 +698,8 @@ static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
 					     struct btree *b)
 {
-	return !btree_iter_pos_cmp(iter, &b->key.k) &&
+	return __btree_iter_pos_cmp(iter, NULL,
-		bkey_cmp(b->key.k.p, POS_MAX);
+			bkey_to_packed(&b->key), true) < 0;
 }
 
 static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
@@ -669,16 +711,18 @@ static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
 }
 
 static inline void __btree_iter_init(struct btree_iter *iter,
-				     struct btree *b)
+				     unsigned level)
 {
-	struct btree_iter_level *l = &iter->l[b->level];
+	struct btree_iter_level *l = &iter->l[level];
 
-	bch2_btree_node_iter_init(&l->iter, b, iter->pos,
+	bch2_btree_node_iter_init(&l->iter, l->b, &iter->pos);
-				  iter->flags & BTREE_ITER_IS_EXTENTS);
+
+	if (iter->flags & BTREE_ITER_IS_EXTENTS)
+		btree_iter_advance_to_pos(iter, l, -1);
 
 	/* Skip to first non whiteout: */
-	if (b->level)
+	if (level)
-		bch2_btree_node_iter_peek(&l->iter, b);
+		bch2_btree_node_iter_peek(&l->iter, l->b);
 
 	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 }
@@ -693,7 +737,7 @@ static inline void btree_iter_node_set(struct btree_iter *iter,
 
 	iter->l[b->level].lock_seq = b->lock.state.seq;
 	iter->l[b->level].b = b;
-	__btree_iter_init(iter, b);
+	__btree_iter_init(iter, b->level);
 }
 
 /*
@@ -747,7 +791,7 @@ void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
 	struct btree_iter *linked;
 
 	for_each_btree_iter_with_node(iter, b, linked)
-		__btree_iter_init(linked, b);
+		__btree_iter_init(linked, b->level);
 }
 
 static inline int btree_iter_lock_root(struct btree_iter *iter,
@@ -986,15 +1030,8 @@ int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
 	 *
 	 * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
 	 */
-	if (btree_iter_node(iter, iter->level)) {
+	if (btree_iter_node(iter, iter->level))
-		struct btree_iter_level *l = &iter->l[iter->level];
+		btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
-		struct bkey_s_c k;
-		struct bkey u;
-
-		while ((k = __btree_iter_peek_all(iter, l, &u)).k &&
-		       !btree_iter_pos_cmp(iter, k.k))
-			__btree_iter_advance(l);
-	}
 
 	/*
 	 * Note: iter->nodes[iter->level] may be temporarily NULL here - that
@@ -1137,7 +1174,6 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth)
 void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
 {
 	struct btree_iter_level *l = &iter->l[0];
-	struct bkey_packed *k;
 
 	EBUG_ON(iter->level != 0);
 	EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
@@ -1147,12 +1183,10 @@ void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_
 	iter->pos = new_pos;
 	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 
-	while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
+	btree_iter_advance_to_pos(iter, l, -1);
-	       !btree_iter_pos_cmp_packed(l->b, &iter->pos, k,
-					  iter->flags & BTREE_ITER_IS_EXTENTS))
-		__btree_iter_advance(l);
 
-	if (!k && btree_iter_pos_after_node(iter, l->b))
+	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);
 }
 
@@ -1169,30 +1203,15 @@ void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 	level = btree_iter_up_until_locked(iter, true);
 
 	if (btree_iter_node(iter, level)) {
-		unsigned nr_advanced = 0;
-		struct btree_iter_level *l = &iter->l[level];
-		struct bkey_s_c k;
-		struct bkey u;
-
 		/*
 		 * We might have to skip over many keys, or just a few: try
 		 * advancing the node iterator, and if we have to skip over too
 		 * many keys just reinit it (or if we're rewinding, since that
 		 * is expensive).
 		 */
-		if (cmp > 0) {
+		if (cmp < 0 ||
-			while ((k = __btree_iter_peek_all(iter, l, &u)).k &&
+		    !btree_iter_advance_to_pos(iter, &iter->l[level], 8))
-			       !btree_iter_pos_cmp(iter, k.k)) {
+			__btree_iter_init(iter, level);
-				if (nr_advanced > 8)
-					goto reinit_node;
-
-				__btree_iter_advance(l);
-				nr_advanced++;
-			}
-		} else {
-reinit_node:
-			__btree_iter_init(iter, iter->l[level].b);
-		}
 
 		/* Don't leave it locked if we're not supposed to: */
 		if (btree_lock_want(iter, level) == BTREE_NODE_UNLOCKED)
@@ -1295,7 +1314,7 @@ struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
 	}
 
 	do {
-		__btree_iter_advance(l);
+		bch2_btree_node_iter_advance(&l->iter, l->b);
 		p = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 		if (unlikely(!p))
 			return bch2_btree_iter_peek_next_leaf(iter);
@@ -1366,7 +1385,7 @@ recheck:
 	while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
 	       bkey_deleted(k.k) &&
 	       bkey_cmp(bkey_start_pos(k.k), iter->pos) == 0)
-		__btree_iter_advance(l);
+		bch2_btree_node_iter_advance(&l->iter, l->b);
 
 	/*
 	 * iterator is now at the correct position for inserting at iter->pos,
@@ -1463,7 +1482,7 @@ recheck:
 	while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
 	       bkey_deleted(k.k) &&
 	       bkey_cmp(k.k->p, iter->pos) == 0)
-		__btree_iter_advance(l);
+		bch2_btree_node_iter_advance(&l->iter, l->b);
 
 	/*
 	 * If we got to the end of the node, check if we need to traverse to the
@@ -1527,7 +1546,7 @@ struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
 	}
 
 	if (!bkey_deleted(&iter->k))
-		__btree_iter_advance(&iter->l[0]);
+		bch2_btree_node_iter_advance(&iter->l[0].iter, iter->l[0].b);
 
 	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 

libbcachefs/btree_update_interior.c

@@ -34,7 +34,7 @@ static void btree_node_interior_verify(struct btree *b)
 
 	BUG_ON(!b->level);
 
-	bch2_btree_node_iter_init(&iter, b, b->key.k.p, false);
+	bch2_btree_node_iter_init(&iter, b, &b->key.k.p);
 #if 1
 	BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) ||
 	       bkey_cmp_left_packed(b, k, &b->key.k.p));
@@ -1190,7 +1190,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 			     gc_pos_btree_node(b), &stats, 0, 0);
 
 	while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
-	       !btree_iter_pos_cmp_packed(b, &insert->k.p, k, false))
+	       bkey_iter_pos_cmp(b, &insert->k.p, k) > 0)
 		bch2_btree_node_iter_advance(node_iter, b);
 
 	/*
@@ -1321,7 +1321,7 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
 
 	BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
 
-	bch2_btree_node_iter_init(&node_iter, b, k->k.p, false);
+	bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
 
 	while (!bch2_keylist_empty(keys)) {
 		k = bch2_keylist_front(keys);

libbcachefs/dirent.c

@@ -15,16 +15,7 @@ unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent d)
 	unsigned len = bkey_val_bytes(d.k) -
 		offsetof(struct bch_dirent, d_name);
 
-	while (len && !d.v->d_name[len - 1])
+	return strnlen(d.v->d_name, len);
-		--len;
-
-	return len;
-}
-
-static unsigned dirent_val_u64s(unsigned len)
-{
-	return DIV_ROUND_UP(offsetof(struct bch_dirent, d_name) + len,
-			    sizeof(u64));
 }
 
 static u64 bch2_dirent_hash(const struct bch_hash_info *info,
@@ -107,9 +98,6 @@ const char *bch2_dirent_invalid(const struct bch_fs *c, struct bkey_s_c k)
 		if (len > BCH_NAME_MAX)
 			return "dirent name too big";
 
-		if (memchr(d.v->d_name, '/', len))
-			return "dirent name has invalid characters";
-
 		return NULL;
 	case BCH_DIRENT_WHITEOUT:
 		return bkey_val_bytes(k.k) != 0

libbcachefs/dirent.h

@@ -22,6 +22,12 @@ struct bch_inode_info;
 
 unsigned bch2_dirent_name_bytes(struct bkey_s_c_dirent);
 
+static inline unsigned dirent_val_u64s(unsigned len)
+{
+	return DIV_ROUND_UP(offsetof(struct bch_dirent, d_name) + len,
+			    sizeof(u64));
+}
+
 int __bch2_dirent_create(struct btree_trans *, u64,
 			 const struct bch_hash_info *, u8,
 			 const struct qstr *, u64, int);

libbcachefs/fsck.c

@@ -247,6 +247,29 @@ fsck_err:
 	return ret;
 }
 
+static bool key_has_correct_hash(const struct bch_hash_desc desc,
+				 struct hash_check *h, struct bch_fs *c,
+				 struct btree_iter *k_iter, struct bkey_s_c k)
+{
+	u64 hash;
+
+	if (k.k->type != desc.whiteout_type &&
+	    k.k->type != desc.key_type)
+		return true;
+
+	if (k.k->p.offset != h->next)
+		bch2_btree_iter_copy(h->chain, k_iter);
+	h->next = k.k->p.offset + 1;
+
+	if (k.k->type != desc.key_type)
+		return true;
+
+	hash = desc.hash_bkey(&h->info, k);
+
+	return hash >= h->chain->pos.offset &&
+		hash <= k.k->p.offset;
+}
+
 static int hash_check_key(const struct bch_hash_desc desc,
 			  struct hash_check *h, struct bch_fs *c,
 			  struct btree_iter *k_iter, struct bkey_s_c k)
@@ -270,9 +293,10 @@ static int hash_check_key(const struct bch_hash_desc desc,
 
 	if (fsck_err_on(hashed < h->chain->pos.offset ||
 			hashed > k.k->p.offset, c,
-			"hash table key at wrong offset: %llu, "
+			"hash table key at wrong offset: btree %u, %llu, "
 			"hashed to %llu chain starts at %llu\n%s",
-			k.k->p.offset, hashed, h->chain->pos.offset,
+			desc.btree_id, k.k->p.offset,
+			hashed, h->chain->pos.offset,
 			(bch2_bkey_val_to_text(c, bkey_type(0, desc.btree_id),
 					       buf, sizeof(buf), k), buf))) {
 		ret = hash_redo_key(desc, h, c, k_iter, k, hashed);
@@ -288,6 +312,90 @@ fsck_err:
 	return ret;
 }
 
+static int check_dirent_hash(struct hash_check *h, struct bch_fs *c,
+			     struct btree_iter *iter, struct bkey_s_c *k)
+{
+	struct bkey_i_dirent *d = NULL;
+	int ret = -EINVAL;
+	char buf[200];
+	unsigned len;
+	u64 hash;
+
+	if (key_has_correct_hash(bch2_dirent_hash_desc, h, c, iter, *k))
+		return 0;
+
+	len = bch2_dirent_name_bytes(bkey_s_c_to_dirent(*k));
+	BUG_ON(!len);
+
+	memcpy(buf, bkey_s_c_to_dirent(*k).v->d_name, len);
+	buf[len] = '\0';
+
+	d = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
+	if (!d) {
+		bch_err(c, "memory allocation failure");
+		return -ENOMEM;
+	}
+
+	bkey_reassemble(&d->k_i, *k);
+
+	do {
+		--len;
+		if (!len)
+			goto err_redo;
+
+		d->k.u64s = BKEY_U64s + dirent_val_u64s(len);
+
+		BUG_ON(bkey_val_bytes(&d->k) <
+		       offsetof(struct bch_dirent, d_name) + len);
+
+		memset(d->v.d_name + len, 0,
+		       bkey_val_bytes(&d->k) -
+		       offsetof(struct bch_dirent, d_name) - len);
+
+		hash = bch2_dirent_hash_desc.hash_bkey(&h->info,
+						bkey_i_to_s_c(&d->k_i));
+	} while (hash < h->chain->pos.offset ||
+		 hash > k->k->p.offset);
+
+	if (fsck_err(c, "dirent with junk at end, was %s (%zu) now %s (%u)",
+		     buf, strlen(buf), d->v.d_name, len)) {
+		ret = bch2_btree_insert_at(c, NULL, NULL,
+					   BTREE_INSERT_NOFAIL,
+					   BTREE_INSERT_ENTRY(iter, &d->k_i));
+		if (ret)
+			goto err;
+
+		*k = bch2_btree_iter_peek(iter);
+
+		BUG_ON(k->k->type != BCH_DIRENT);
+	}
+err:
+fsck_err:
+	kfree(d);
+	return ret;
+err_redo:
+	bch_err(c, "cannot fix dirent by removing trailing garbage %s (%zu)",
+		buf, strlen(buf));
+
+	hash = bch2_dirent_hash_desc.hash_bkey(&h->info, *k);
+
+	if (fsck_err(c, "hash table key at wrong offset: btree %u, offset %llu, "
+			"hashed to %llu chain starts at %llu\n%s",
+			BTREE_ID_DIRENTS,
+			k->k->p.offset, hash, h->chain->pos.offset,
+			(bch2_bkey_val_to_text(c, bkey_type(0, BTREE_ID_DIRENTS),
+					       buf, sizeof(buf), *k), buf))) {
+		ret = hash_redo_key(bch2_dirent_hash_desc,
+				    h, c, iter, *k, hash);
+		if (ret)
+			bch_err(c, "hash_redo_key err %i", ret);
+		else
+			ret = 1;
+	}
+
+	goto err;
+}
+
 static int bch2_inode_truncate(struct bch_fs *c, u64 inode_nr, u64 new_size)
 {
 	return bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
@@ -434,11 +542,13 @@ static int check_dirents(struct bch_fs *c)
 		if (w.first_this_inode && w.have_inode)
 			hash_check_set_inode(&h, c, &w.inode);
 
-		ret = hash_check_key(bch2_dirent_hash_desc, &h, c, iter, k);
+		ret = check_dirent_hash(&h, c, iter, &k);
 		if (ret > 0) {
 			ret = 0;
 			continue;
 		}
+		if (ret)
+			goto fsck_err;
 
 		if (ret)
 			goto fsck_err;
@@ -457,7 +567,12 @@ static int check_dirents(struct bch_fs *c)
 				". dirent") ||
 		    fsck_err_on(name_len == 2 &&
 				!memcmp(d.v->d_name, "..", 2), c,
-				".. dirent")) {
+				".. dirent") ||
+		    fsck_err_on(name_len == 2 &&
+				!memcmp(d.v->d_name, "..", 2), c,
+				".. dirent") ||
+		    fsck_err_on(memchr(d.v->d_name, '/', name_len), c,
+				"dirent name has invalid chars")) {
 			ret = remove_dirent(c, iter, d);
 			if (ret)
 				goto err;

libbcachefs/tests.c

@@ -270,6 +270,42 @@ static void test_iterate_slots_extents(struct bch_fs *c, u64 nr)
 	bch2_btree_iter_unlock(&iter);
 }
 
+/*
+ * XXX: we really want to make sure we've got a btree with depth > 0 for these
+ * tests
+ */
+static void test_peek_end(struct bch_fs *c, u64 nr)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_DIRENTS, POS_MIN, 0);
+
+	k = bch2_btree_iter_peek(&iter);
+	BUG_ON(k.k);
+
+	k = bch2_btree_iter_peek(&iter);
+	BUG_ON(k.k);
+
+	bch2_btree_iter_unlock(&iter);
+}
+
+static void test_peek_end_extents(struct bch_fs *c, u64 nr)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+
+	bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN, 0);
+
+	k = bch2_btree_iter_peek(&iter);
+	BUG_ON(k.k);
+
+	k = bch2_btree_iter_peek(&iter);
+	BUG_ON(k.k);
+
+	bch2_btree_iter_unlock(&iter);
+}
+
 /* extent unit tests */
 
 u64 test_version;
@@ -554,6 +590,8 @@ void bch2_btree_perf_test(struct bch_fs *c, const char *testname,
 	perf_test(test_iterate_extents);
 	perf_test(test_iterate_slots);
 	perf_test(test_iterate_slots_extents);
+	perf_test(test_peek_end);
+	perf_test(test_peek_end_extents);
 
 	perf_test(test_extent_overwrite_front);
 	perf_test(test_extent_overwrite_back);