bcachefs: Whiteouts for snapshots

This patch adds KEY_TYPE_whiteout, a new type of whiteout for snapshots,
when we're deleting and the key being deleted is in an ancestor
snapshot - and updates the transaction update/commit path to use it.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

fs/bcachefs/bcachefs_format.h

@@ -327,7 +327,7 @@ static inline void bkey_init(struct bkey *k)
 */
 #define BCH_BKEY_TYPES()				\
 	x(deleted,		0)			\
-	x(discard,		1)			\
+	x(whiteout,		1)			\
 	x(error,		2)			\
 	x(cookie,		3)			\
 	x(hash_whiteout,	4)			\
@@ -361,7 +361,7 @@ struct bch_deleted {
 	struct bch_val		v;
 };
 
-struct bch_discard {
+struct bch_whiteout {
 	struct bch_val		v;
 };
 

fs/bcachefs/bkey.h

@@ -63,7 +63,7 @@ static inline void set_bkey_val_bytes(struct bkey *k, unsigned bytes)
 #define bkey_deleted(_k)	((_k)->type == KEY_TYPE_deleted)
 
 #define bkey_whiteout(_k)				\
-	((_k)->type == KEY_TYPE_deleted || (_k)->type == KEY_TYPE_discard)
+	((_k)->type == KEY_TYPE_deleted || (_k)->type == KEY_TYPE_whiteout)
 
 enum bkey_lr_packed {
 	BKEY_PACKED_BOTH,

fs/bcachefs/bkey_methods.c

@@ -31,7 +31,7 @@ static const char *deleted_key_invalid(const struct bch_fs *c,
 	.key_invalid = deleted_key_invalid,		\
 }
 
-#define bch2_bkey_ops_discard (struct bkey_ops) {	\
+#define bch2_bkey_ops_whiteout (struct bkey_ops) {	\
 	.key_invalid = deleted_key_invalid,		\
 }
 
@@ -101,6 +101,8 @@ const char *bch2_bkey_val_invalid(struct bch_fs *c, struct bkey_s_c k)
 
 static unsigned bch2_key_types_allowed[] = {
 	[BKEY_TYPE_extents] =
+		(1U << KEY_TYPE_deleted)|
+		(1U << KEY_TYPE_whiteout)|
 		(1U << KEY_TYPE_error)|
 		(1U << KEY_TYPE_cookie)|
 		(1U << KEY_TYPE_extent)|
@@ -108,30 +110,43 @@ static unsigned bch2_key_types_allowed[] = {
 		(1U << KEY_TYPE_reflink_p)|
 		(1U << KEY_TYPE_inline_data),
 	[BKEY_TYPE_inodes] =
+		(1U << KEY_TYPE_deleted)|
+		(1U << KEY_TYPE_whiteout)|
 		(1U << KEY_TYPE_inode)|
 		(1U << KEY_TYPE_inode_generation),
 	[BKEY_TYPE_dirents] =
+		(1U << KEY_TYPE_deleted)|
+		(1U << KEY_TYPE_whiteout)|
 		(1U << KEY_TYPE_hash_whiteout)|
 		(1U << KEY_TYPE_dirent),
 	[BKEY_TYPE_xattrs] =
+		(1U << KEY_TYPE_deleted)|
+		(1U << KEY_TYPE_whiteout)|
 		(1U << KEY_TYPE_cookie)|
 		(1U << KEY_TYPE_hash_whiteout)|
 		(1U << KEY_TYPE_xattr),
 	[BKEY_TYPE_alloc] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_alloc)|
 		(1U << KEY_TYPE_alloc_v2),
 	[BKEY_TYPE_quotas] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_quota),
 	[BKEY_TYPE_stripes] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_stripe),
 	[BKEY_TYPE_reflink] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_reflink_v)|
 		(1U << KEY_TYPE_indirect_inline_data),
 	[BKEY_TYPE_subvolumes] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_subvolume),
 	[BKEY_TYPE_snapshots] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_snapshot),
 	[BKEY_TYPE_btree] =
+		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_btree_ptr)|
 		(1U << KEY_TYPE_btree_ptr_v2),
 };
@@ -139,21 +154,18 @@ static unsigned bch2_key_types_allowed[] = {
 const char *__bch2_bkey_invalid(struct bch_fs *c, struct bkey_s_c k,
 				enum btree_node_type type)
 {
-	unsigned key_types_allowed = (1U << KEY_TYPE_deleted)|
-		bch2_key_types_allowed[type] ;
-
 	if (k.k->u64s < BKEY_U64s)
 		return "u64s too small";
 
-	if (!(key_types_allowed & (1U << k.k->type)))
+	if (!(bch2_key_types_allowed[type] & (1U << k.k->type)))
 		return "invalid key type for this btree";
 
 	if (type == BKEY_TYPE_btree &&
 	    bkey_val_u64s(k.k) > BKEY_BTREE_PTR_VAL_U64s_MAX)
 		return "value too big";
 
-	if (btree_node_type_is_extents(type)) {
-		if ((k.k->size == 0) != bkey_deleted(k.k))
+	if (btree_node_type_is_extents(type) && !bkey_whiteout(k.k)) {
+		if (k.k->size == 0)
 			return "bad size field";
 
 		if (k.k->size > k.k->p.offset)

fs/bcachefs/btree_update_leaf.c

@@ -1002,21 +1002,24 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 		goto next;
 	}
 
-	if (!bkey_cmp(k.k->p, bkey_start_pos(&insert->k)))
+	if (!bkey_cmp(k.k->p, start))
 		goto next;
 
 	while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
+		bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
+		bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
+
 		/*
 		 * If we're going to be splitting a compressed extent, note it
 		 * so that __bch2_trans_commit() can increase our disk
 		 * reservation:
 		 */
-		if (bkey_cmp(bkey_start_pos(k.k), start) < 0 &&
-		    bkey_cmp(k.k->p, insert->k.p) > 0 &&
+		if (((front_split && back_split) ||
+		     ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
 		    (compressed_sectors = bch2_bkey_sectors_compressed(k)))
 			trans->extra_journal_res += compressed_sectors;
 
-		if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
+		if (front_split) {
 			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 			if ((ret = PTR_ERR_OR_ZERO(update)))
 				goto err;
@@ -1027,6 +1030,32 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 
 			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
 					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_ALL_SNAPSHOTS|
+					     BTREE_ITER_INTENT);
+			ret   = bch2_btree_iter_traverse(&update_iter) ?:
+				bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+
+			if (ret)
+				goto err;
+		}
+
+		if (k.k->p.snapshot != insert->k.p.snapshot &&
+		    (front_split || back_split)) {
+			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+
+			bkey_reassemble(update, k);
+
+			bch2_cut_front(start, update);
+			bch2_cut_back(insert->k.p, update);
+
+			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_ALL_SNAPSHOTS|
 					     BTREE_ITER_INTENT);
 			ret   = bch2_btree_iter_traverse(&update_iter) ?:
 				bch2_trans_update(trans, &update_iter, update,
@@ -1038,12 +1067,32 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 		}
 
 		if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
-			ret = bch2_btree_delete_at(trans, &iter, flags);
+			update = bch2_trans_kmalloc(trans, sizeof(*update));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+
+			bkey_init(&update->k);
+			update->k.p = k.k->p;
+
+			if (insert->k.p.snapshot != k.k->p.snapshot) {
+				update->k.p.snapshot = insert->k.p.snapshot;
+				update->k.type = KEY_TYPE_whiteout;
+			}
+
+			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_INTENT);
+			ret   = bch2_btree_iter_traverse(&update_iter) ?:
+				bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+
 			if (ret)
 				goto err;
 		}
 
-		if (bkey_cmp(k.k->p, insert->k.p) > 0) {
+		if (back_split) {
 			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 			if ((ret = PTR_ERR_OR_ZERO(update)))
 				goto err;
@@ -1051,10 +1100,15 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 			bkey_reassemble(update, k);
 			bch2_cut_front(insert->k.p, update);
 
-			ret = bch2_trans_update(trans, &iter, update, flags);
+			bch2_trans_copy_iter(&update_iter, &iter);
+			update_iter.pos = update->k.p;
+			ret   = bch2_trans_update(trans, &update_iter, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
+						  flags);
+			bch2_trans_iter_exit(trans, &update_iter);
+
 			if (ret)
 				goto err;
-
 			goto out;
 		}
 next:
@@ -1086,6 +1140,39 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 	return ret;
 }
 
+/*
+ * When deleting, check if we need to emit a whiteout (because we're overwriting
+ * something in an ancestor snapshot)
+ */
+static int need_whiteout_for_snapshot(struct btree_trans *trans,
+				      enum btree_id btree_id, struct bpos pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u32 snapshot = pos.snapshot;
+	int ret;
+
+	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+		return 0;
+
+	pos.snapshot++;
+
+	for_each_btree_key(trans, iter, btree_id, pos,
+			   BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
+		if (bkey_cmp(k.k->p, pos))
+			break;
+
+		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+					      k.k->p.snapshot)) {
+			ret = !bkey_whiteout(k.k);
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
+
 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 		      struct bkey_i *k, enum btree_update_flags flags)
 {
@@ -1118,6 +1205,16 @@ int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 		       btree_insert_entry_cmp(i - 1, i) >= 0);
 #endif
 
+	if (bkey_deleted(&n.k->k) &&
+	    (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
+		int ret = need_whiteout_for_snapshot(trans, n.btree_id, n.k->k.p);
+		if (unlikely(ret < 0))
+			return ret;
+
+		if (ret)
+			n.k->k.type = KEY_TYPE_whiteout;
+	}
+
 	/*
 	 * Pending updates are kept sorted: first, find position of new update,
 	 * then delete/trim any updates the new update overwrites: