Btrfs: fix infinite path build loops in incremental send

The send operation processes inodes by their ascending number, and assumes
that any rename/move operation can be successfully performed (sent to the
caller) once all previous inodes (those with a smaller inode number than the
one we're currently processing) were processed.

This is not true when an incremental send had to process an hierarchical change
between 2 snapshots where the parent-children relationship between directory
inodes was reversed - that is, parents became children and children became
parents. This situation made the path building code go into an infinite loop,
which kept allocating more and more memory that eventually lead to a krealloc
warning being displayed in dmesg:

  WARNING: CPU: 1 PID: 5705 at mm/page_alloc.c:2477 __alloc_pages_nodemask+0x365/0xad0()
  Modules linked in: btrfs raid6_pq xor pci_stub vboxpci(O) vboxnetadp(O) vboxnetflt(O) vboxdrv(O) snd_hda_codec_hdmi snd_hda_codec_realtek joydev radeon snd_hda_intel snd_hda_codec snd_hwdep snd_seq_midi snd_pcm psmouse i915 snd_rawmidi serio_raw snd_seq_midi_event lpc_ich snd_seq snd_timer ttm snd_seq_device rfcomm drm_kms_helper parport_pc bnep bluetooth drm ppdev snd soundcore i2c_algo_bit snd_page_alloc binfmt_misc video lp parport r8169 mii hid_generic usbhid hid
  CPU: 1 PID: 5705 Comm: btrfs Tainted: G           O 3.13.0-rc7-fdm-btrfs-next-18+ #3
  Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./Z77 Pro4, BIOS P1.50 09/04/2012
  [ 5381.660441]  00000000000009ad ffff8806f6f2f4e8 ffffffff81777434 0000000000000007
  [ 5381.660447]  0000000000000000 ffff8806f6f2f528 ffffffff8104a9ec ffff8807038f36f0
  [ 5381.660452]  0000000000000000 0000000000000206 ffff8807038f2490 ffff8807038f36f0
  [ 5381.660457] Call Trace:
  [ 5381.660464]  [<ffffffff81777434>] dump_stack+0x4e/0x68
  [ 5381.660471]  [<ffffffff8104a9ec>] warn_slowpath_common+0x8c/0xc0
  [ 5381.660476]  [<ffffffff8104aa3a>] warn_slowpath_null+0x1a/0x20
  [ 5381.660480]  [<ffffffff81144995>] __alloc_pages_nodemask+0x365/0xad0
  [ 5381.660487]  [<ffffffff8108313f>] ? local_clock+0x4f/0x60
  [ 5381.660491]  [<ffffffff811430e8>] ? free_one_page+0x98/0x440
  [ 5381.660495]  [<ffffffff8108313f>] ? local_clock+0x4f/0x60
  [ 5381.660502]  [<ffffffff8113fae4>] ? __get_free_pages+0x14/0x50
  [ 5381.660508]  [<ffffffff81095fb8>] ? trace_hardirqs_off_caller+0x28/0xd0
  [ 5381.660515]  [<ffffffff81183caf>] alloc_pages_current+0x10f/0x1f0
  [ 5381.660520]  [<ffffffff8113fae4>] ? __get_free_pages+0x14/0x50
  [ 5381.660524]  [<ffffffff8113fae4>] __get_free_pages+0x14/0x50
  [ 5381.660530]  [<ffffffff8115dace>] kmalloc_order_trace+0x3e/0x100
  [ 5381.660536]  [<ffffffff81191ea0>] __kmalloc_track_caller+0x220/0x230
  [ 5381.660560]  [<ffffffffa0729fdb>] ? fs_path_ensure_buf.part.12+0x6b/0x200 [btrfs]
  [ 5381.660564]  [<ffffffff8178085c>] ? retint_restore_args+0xe/0xe
  [ 5381.660569]  [<ffffffff811580ef>] krealloc+0x6f/0xb0
  [ 5381.660586]  [<ffffffffa0729fdb>] fs_path_ensure_buf.part.12+0x6b/0x200 [btrfs]
  [ 5381.660601]  [<ffffffffa072a208>] fs_path_prepare_for_add+0x98/0xb0 [btrfs]
  [ 5381.660615]  [<ffffffffa072a2bc>] fs_path_add_path+0x2c/0x60 [btrfs]
  [ 5381.660628]  [<ffffffffa072c55c>] get_cur_path+0x7c/0x1c0 [btrfs]

Even without this loop, the incremental send couldn't succeed, because it would attempt
to send a rename/move operation for the lower inode before the highest inode number was
renamed/move. This issue is easy to trigger with the following steps:

  $ mkfs.btrfs -f /dev/sdb3
  $ mount /dev/sdb3 /mnt/btrfs
  $ mkdir -p /mnt/btrfs/a/b/c/d
  $ mkdir /mnt/btrfs/a/b/c2
  $ btrfs subvol snapshot -r /mnt/btrfs /mnt/btrfs/snap1
  $ mv /mnt/btrfs/a/b/c/d /mnt/btrfs/a/b/c2/d2
  $ mv /mnt/btrfs/a/b/c /mnt/btrfs/a/b/c2/d2/cc
  $ btrfs subvol snapshot -r /mnt/btrfs /mnt/btrfs/snap2
  $ btrfs send -p /mnt/btrfs/snap1 /mnt/btrfs/snap2 > /tmp/incremental.send

The structure of the filesystem when the first snapshot is taken is:

	 .                       (ino 256)
	 |-- a                   (ino 257)
	     |-- b               (ino 258)
	         |-- c           (ino 259)
	         |   |-- d       (ino 260)
                 |
	         |-- c2          (ino 261)

And its structure when the second snapshot is taken is:

	 .                       (ino 256)
	 |-- a                   (ino 257)
	     |-- b               (ino 258)
	         |-- c2          (ino 261)
	             |-- d2      (ino 260)
	                 |-- cc  (ino 259)

Before the move/rename operation is performed for the inode 259, the
move/rename for inode 260 must be performed, since 259 is now a child
of 260.

A test case for xfstests, with a more complex scenario, will follow soon.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>

fs/btrfs/send.c

@@ -121,6 +121,74 @@ struct send_ctx {
 	int name_cache_size;
 
 	char *read_buf;
+
+	/*
+	 * We process inodes by their increasing order, so if before an
+	 * incremental send we reverse the parent/child relationship of
+	 * directories such that a directory with a lower inode number was
+	 * the parent of a directory with a higher inode number, and the one
+	 * becoming the new parent got renamed too, we can't rename/move the
+	 * directory with lower inode number when we finish processing it - we
+	 * must process the directory with higher inode number first, then
+	 * rename/move it and then rename/move the directory with lower inode
+	 * number. Example follows.
+	 *
+	 * Tree state when the first send was performed:
+	 *
+	 * .
+	 * |-- a                   (ino 257)
+	 *     |-- b               (ino 258)
+	 *         |
+	 *         |
+	 *         |-- c           (ino 259)
+	 *         |   |-- d       (ino 260)
+	 *         |
+	 *         |-- c2          (ino 261)
+	 *
+	 * Tree state when the second (incremental) send is performed:
+	 *
+	 * .
+	 * |-- a                   (ino 257)
+	 *     |-- b               (ino 258)
+	 *         |-- c2          (ino 261)
+	 *             |-- d2      (ino 260)
+	 *                 |-- cc  (ino 259)
+	 *
+	 * The sequence of steps that lead to the second state was:
+	 *
+	 * mv /a/b/c/d /a/b/c2/d2
+	 * mv /a/b/c /a/b/c2/d2/cc
+	 *
+	 * "c" has lower inode number, but we can't move it (2nd mv operation)
+	 * before we move "d", which has higher inode number.
+	 *
+	 * So we just memorize which move/rename operations must be performed
+	 * later when their respective parent is processed and moved/renamed.
+	 */
+
+	/* Indexed by parent directory inode number. */
+	struct rb_root pending_dir_moves;
+
+	/*
+	 * Reverse index, indexed by the inode number of a directory that
+	 * is waiting for the move/rename of its immediate parent before its
+	 * own move/rename can be performed.
+	 */
+	struct rb_root waiting_dir_moves;
+};
+
+struct pending_dir_move {
+	struct rb_node node;
+	struct list_head list;
+	u64 parent_ino;
+	u64 ino;
+	u64 gen;
+	struct list_head update_refs;
+};
+
+struct waiting_dir_move {
+	struct rb_node node;
+	u64 ino;
 };
 
 struct name_cache_entry {
@@ -144,6 +212,8 @@ struct name_cache_entry {
 	char name[];
 };
 
+static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
+
 static int need_send_hole(struct send_ctx *sctx)
 {
 	return (sctx->parent_root && !sctx->cur_inode_new &&
@@ -1897,6 +1967,7 @@ static void name_cache_free(struct send_ctx *sctx)
  */
 static int __get_cur_name_and_parent(struct send_ctx *sctx,
 				     u64 ino, u64 gen,
+				     int skip_name_cache,
 				     u64 *parent_ino,
 				     u64 *parent_gen,
 				     struct fs_path *dest)
@@ -1906,6 +1977,8 @@ static int __get_cur_name_and_parent(struct send_ctx *sctx,
 	struct btrfs_path *path = NULL;
 	struct name_cache_entry *nce = NULL;
 
+	if (skip_name_cache)
+		goto get_ref;
 	/*
 	 * First check if we already did a call to this function with the same
 	 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
@@ -1950,11 +2023,12 @@ static int __get_cur_name_and_parent(struct send_ctx *sctx,
 		goto out_cache;
 	}
 
+get_ref:
 	/*
 	 * Depending on whether the inode was already processed or not, use
 	 * send_root or parent_root for ref lookup.
 	 */
-	if (ino < sctx->send_progress)
+	if (ino < sctx->send_progress && !skip_name_cache)
 		ret = get_first_ref(sctx->send_root, ino,
 				    parent_ino, parent_gen, dest);
 	else
@@ -1978,6 +2052,8 @@ static int __get_cur_name_and_parent(struct send_ctx *sctx,
 			goto out;
 		ret = 1;
 	}
+	if (skip_name_cache)
+		goto out;
 
 out_cache:
 	/*
@@ -2045,6 +2121,9 @@ static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
 	u64 parent_inode = 0;
 	u64 parent_gen = 0;
 	int stop = 0;
+	u64 start_ino = ino;
+	u64 start_gen = gen;
+	int skip_name_cache = 0;
 
 	name = fs_path_alloc();
 	if (!name) {
@@ -2052,19 +2131,32 @@ static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
 		goto out;
 	}
 
+	if (is_waiting_for_move(sctx, ino))
+		skip_name_cache = 1;
+
+again:
 	dest->reversed = 1;
 	fs_path_reset(dest);
 
 	while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
 		fs_path_reset(name);
 
-		ret = __get_cur_name_and_parent(sctx, ino, gen,
+		ret = __get_cur_name_and_parent(sctx, ino, gen, skip_name_cache,
 				&parent_inode, &parent_gen, name);
 		if (ret < 0)
 			goto out;
 		if (ret)
 			stop = 1;
 
+		if (!skip_name_cache &&
+		    is_waiting_for_move(sctx, parent_inode)) {
+			ino = start_ino;
+			gen = start_gen;
+			stop = 0;
+			skip_name_cache = 1;
+			goto again;
+		}
+
 		ret = fs_path_add_path(dest, name);
 		if (ret < 0)
 			goto out;
@@ -2636,10 +2728,349 @@ static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 send_progress)
 	return ret;
 }
 
+static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
+{
+	struct rb_node *n = sctx->waiting_dir_moves.rb_node;
+	struct waiting_dir_move *entry;
+
+	while (n) {
+		entry = rb_entry(n, struct waiting_dir_move, node);
+		if (ino < entry->ino)
+			n = n->rb_left;
+		else if (ino > entry->ino)
+			n = n->rb_right;
+		else
+			return 1;
+	}
+	return 0;
+}
+
+static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino)
+{
+	struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
+	struct rb_node *parent = NULL;
+	struct waiting_dir_move *entry, *dm;
+
+	dm = kmalloc(sizeof(*dm), GFP_NOFS);
+	if (!dm)
+		return -ENOMEM;
+	dm->ino = ino;
+
+	while (*p) {
+		parent = *p;
+		entry = rb_entry(parent, struct waiting_dir_move, node);
+		if (ino < entry->ino) {
+			p = &(*p)->rb_left;
+		} else if (ino > entry->ino) {
+			p = &(*p)->rb_right;
+		} else {
+			kfree(dm);
+			return -EEXIST;
+		}
+	}
+
+	rb_link_node(&dm->node, parent, p);
+	rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
+	return 0;
+}
+
+#ifdef CONFIG_BTRFS_ASSERT
+
+static int del_waiting_dir_move(struct send_ctx *sctx, u64 ino)
+{
+	struct rb_node *n = sctx->waiting_dir_moves.rb_node;
+	struct waiting_dir_move *entry;
+
+	while (n) {
+		entry = rb_entry(n, struct waiting_dir_move, node);
+		if (ino < entry->ino) {
+			n = n->rb_left;
+		} else if (ino > entry->ino) {
+			n = n->rb_right;
+		} else {
+			rb_erase(&entry->node, &sctx->waiting_dir_moves);
+			kfree(entry);
+			return 0;
+		}
+	}
+	return -ENOENT;
+}
+
+#endif
+
+static int add_pending_dir_move(struct send_ctx *sctx, u64 parent_ino)
+{
+	struct rb_node **p = &sctx->pending_dir_moves.rb_node;
+	struct rb_node *parent = NULL;
+	struct pending_dir_move *entry, *pm;
+	struct recorded_ref *cur;
+	int exists = 0;
+	int ret;
+
+	pm = kmalloc(sizeof(*pm), GFP_NOFS);
+	if (!pm)
+		return -ENOMEM;
+	pm->parent_ino = parent_ino;
+	pm->ino = sctx->cur_ino;
+	pm->gen = sctx->cur_inode_gen;
+	INIT_LIST_HEAD(&pm->list);
+	INIT_LIST_HEAD(&pm->update_refs);
+	RB_CLEAR_NODE(&pm->node);
+
+	while (*p) {
+		parent = *p;
+		entry = rb_entry(parent, struct pending_dir_move, node);
+		if (parent_ino < entry->parent_ino) {
+			p = &(*p)->rb_left;
+		} else if (parent_ino > entry->parent_ino) {
+			p = &(*p)->rb_right;
+		} else {
+			exists = 1;
+			break;
+		}
+	}
+
+	list_for_each_entry(cur, &sctx->deleted_refs, list) {
+		ret = dup_ref(cur, &pm->update_refs);
+		if (ret < 0)
+			goto out;
+	}
+	list_for_each_entry(cur, &sctx->new_refs, list) {
+		ret = dup_ref(cur, &pm->update_refs);
+		if (ret < 0)
+			goto out;
+	}
+
+	ret = add_waiting_dir_move(sctx, pm->ino);
+	if (ret)
+		goto out;
+
+	if (exists) {
+		list_add_tail(&pm->list, &entry->list);
+	} else {
+		rb_link_node(&pm->node, parent, p);
+		rb_insert_color(&pm->node, &sctx->pending_dir_moves);
+	}
+	ret = 0;
+out:
+	if (ret) {
+		__free_recorded_refs(&pm->update_refs);
+		kfree(pm);
+	}
+	return ret;
+}
+
+static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
+						      u64 parent_ino)
+{
+	struct rb_node *n = sctx->pending_dir_moves.rb_node;
+	struct pending_dir_move *entry;
+
+	while (n) {
+		entry = rb_entry(n, struct pending_dir_move, node);
+		if (parent_ino < entry->parent_ino)
+			n = n->rb_left;
+		else if (parent_ino > entry->parent_ino)
+			n = n->rb_right;
+		else
+			return entry;
+	}
+	return NULL;
+}
+
+static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
+{
+	struct fs_path *from_path = NULL;
+	struct fs_path *to_path = NULL;
+	u64 orig_progress = sctx->send_progress;
+	struct recorded_ref *cur;
+	int ret;
+
+	from_path = fs_path_alloc();
+	if (!from_path)
+		return -ENOMEM;
+
+	sctx->send_progress = pm->ino;
+	ret = get_cur_path(sctx, pm->ino, pm->gen, from_path);
+	if (ret < 0)
+		goto out;
+
+	to_path = fs_path_alloc();
+	if (!to_path) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	sctx->send_progress = sctx->cur_ino + 1;
+	ASSERT(del_waiting_dir_move(sctx, pm->ino) == 0);
+	ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
+	if (ret < 0)
+		goto out;
+
+	ret = send_rename(sctx, from_path, to_path);
+	if (ret < 0)
+		goto out;
+
+	ret = send_utimes(sctx, pm->ino, pm->gen);
+	if (ret < 0)
+		goto out;
+
+	/*
+	 * After rename/move, need to update the utimes of both new parent(s)
+	 * and old parent(s).
+	 */
+	list_for_each_entry(cur, &pm->update_refs, list) {
+		ret = send_utimes(sctx, cur->dir, cur->dir_gen);
+		if (ret < 0)
+			goto out;
+	}
+
+out:
+	fs_path_free(from_path);
+	fs_path_free(to_path);
+	sctx->send_progress = orig_progress;
+
+	return ret;
+}
+
+static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
+{
+	if (!list_empty(&m->list))
+		list_del(&m->list);
+	if (!RB_EMPTY_NODE(&m->node))
+		rb_erase(&m->node, &sctx->pending_dir_moves);
+	__free_recorded_refs(&m->update_refs);
+	kfree(m);
+}
+
+static void tail_append_pending_moves(struct pending_dir_move *moves,
+				      struct list_head *stack)
+{
+	if (list_empty(&moves->list)) {
+		list_add_tail(&moves->list, stack);
+	} else {
+		LIST_HEAD(list);
+		list_splice_init(&moves->list, &list);
+		list_add_tail(&moves->list, stack);
+		list_splice_tail(&list, stack);
+	}
+}
+
+static int apply_children_dir_moves(struct send_ctx *sctx)
+{
+	struct pending_dir_move *pm;
+	struct list_head stack;
+	u64 parent_ino = sctx->cur_ino;
+	int ret = 0;
+
+	pm = get_pending_dir_moves(sctx, parent_ino);
+	if (!pm)
+		return 0;
+
+	INIT_LIST_HEAD(&stack);
+	tail_append_pending_moves(pm, &stack);
+
+	while (!list_empty(&stack)) {
+		pm = list_first_entry(&stack, struct pending_dir_move, list);
+		parent_ino = pm->ino;
+		ret = apply_dir_move(sctx, pm);
+		free_pending_move(sctx, pm);
+		if (ret)
+			goto out;
+		pm = get_pending_dir_moves(sctx, parent_ino);
+		if (pm)
+			tail_append_pending_moves(pm, &stack);
+	}
+	return 0;
+
+out:
+	while (!list_empty(&stack)) {
+		pm = list_first_entry(&stack, struct pending_dir_move, list);
+		free_pending_move(sctx, pm);
+	}
+	return ret;
+}
+
+static int wait_for_parent_move(struct send_ctx *sctx,
+				struct recorded_ref *parent_ref)
+{
+	int ret;
+	u64 ino = parent_ref->dir;
+	u64 parent_ino_before, parent_ino_after;
+	u64 new_gen, old_gen;
+	struct fs_path *path_before = NULL;
+	struct fs_path *path_after = NULL;
+	int len1, len2;
+
+	if (parent_ref->dir <= sctx->cur_ino)
+		return 0;
+
+	if (is_waiting_for_move(sctx, ino))
+		return 1;
+
+	ret = get_inode_info(sctx->parent_root, ino, NULL, &old_gen,
+			     NULL, NULL, NULL, NULL);
+	if (ret == -ENOENT)
+		return 0;
+	else if (ret < 0)
+		return ret;
+
+	ret = get_inode_info(sctx->send_root, ino, NULL, &new_gen,
+			     NULL, NULL, NULL, NULL);
+	if (ret < 0)
+		return ret;
+
+	if (new_gen != old_gen)
+		return 0;
+
+	path_before = fs_path_alloc();
+	if (!path_before)
+		return -ENOMEM;
+
+	ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
+			    NULL, path_before);
+	if (ret == -ENOENT) {
+		ret = 0;
+		goto out;
+	} else if (ret < 0) {
+		goto out;
+	}
+
+	path_after = fs_path_alloc();
+	if (!path_after) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
+			    NULL, path_after);
+	if (ret == -ENOENT) {
+		ret = 0;
+		goto out;
+	} else if (ret < 0) {
+		goto out;
+	}
+
+	len1 = fs_path_len(path_before);
+	len2 = fs_path_len(path_after);
+	if ((parent_ino_before != parent_ino_after) && (len1 != len2 ||
+	     memcmp(path_before->start, path_after->start, len1))) {
+		ret = 1;
+		goto out;
+	}
+	ret = 0;
+
+out:
+	fs_path_free(path_before);
+	fs_path_free(path_after);
+
+	return ret;
+}
+
 /*
  * This does all the move/link/unlink/rmdir magic.
  */
-static int process_recorded_refs(struct send_ctx *sctx)
+static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
 {
 	int ret = 0;
 	struct recorded_ref *cur;
@@ -2788,11 +3219,17 @@ verbose_printk("btrfs: process_recorded_refs %llu\n", sctx->cur_ino);
 				 * dirs, we always have one new and one deleted
 				 * ref. The deleted ref is ignored later.
 				 */
-				ret = send_rename(sctx, valid_path,
-						cur->full_path);
-				if (ret < 0)
-					goto out;
-				ret = fs_path_copy(valid_path, cur->full_path);
+				if (wait_for_parent_move(sctx, cur)) {
+					ret = add_pending_dir_move(sctx,
+								   cur->dir);
+					*pending_move = 1;
+				} else {
+					ret = send_rename(sctx, valid_path,
+							  cur->full_path);
+					if (!ret)
+						ret = fs_path_copy(valid_path,
+							       cur->full_path);
+				}
 				if (ret < 0)
 					goto out;
 			} else {
@@ -3161,6 +3598,7 @@ static int process_all_refs(struct send_ctx *sctx,
 	struct extent_buffer *eb;
 	int slot;
 	iterate_inode_ref_t cb;
+	int pending_move = 0;
 
 	path = alloc_path_for_send();
 	if (!path)
@@ -3204,7 +3642,9 @@ static int process_all_refs(struct send_ctx *sctx,
 	}
 	btrfs_release_path(path);
 
-	ret = process_recorded_refs(sctx);
+	ret = process_recorded_refs(sctx, &pending_move);
+	/* Only applicable to an incremental send. */
+	ASSERT(pending_move == 0);
 
 out:
 	btrfs_free_path(path);
@@ -4165,7 +4605,9 @@ static int process_all_extents(struct send_ctx *sctx)
 	return ret;
 }
 
-static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end)
+static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
+					   int *pending_move,
+					   int *refs_processed)
 {
 	int ret = 0;
 
@@ -4177,17 +4619,11 @@ static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end)
 	if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
 		goto out;
 
-	ret = process_recorded_refs(sctx);
+	ret = process_recorded_refs(sctx, pending_move);
 	if (ret < 0)
 		goto out;
 
-	/*
-	 * We have processed the refs and thus need to advance send_progress.
-	 * Now, calls to get_cur_xxx will take the updated refs of the current
-	 * inode into account.
-	 */
-	sctx->send_progress = sctx->cur_ino + 1;
-
+	*refs_processed = 1;
 out:
 	return ret;
 }
@@ -4203,11 +4639,29 @@ static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
 	u64 right_gid;
 	int need_chmod = 0;
 	int need_chown = 0;
+	int pending_move = 0;
+	int refs_processed = 0;
 
-	ret = process_recorded_refs_if_needed(sctx, at_end);
+	ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
+					      &refs_processed);
 	if (ret < 0)
 		goto out;
 
+	/*
+	 * We have processed the refs and thus need to advance send_progress.
+	 * Now, calls to get_cur_xxx will take the updated refs of the current
+	 * inode into account.
+	 *
+	 * On the other hand, if our current inode is a directory and couldn't
+	 * be moved/renamed because its parent was renamed/moved too and it has
+	 * a higher inode number, we can only move/rename our current inode
+	 * after we moved/renamed its parent. Therefore in this case operate on
+	 * the old path (pre move/rename) of our current inode, and the
+	 * move/rename will be performed later.
+	 */
+	if (refs_processed && !pending_move)
+		sctx->send_progress = sctx->cur_ino + 1;
+
 	if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
 		goto out;
 	if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
@@ -4269,9 +4723,21 @@ static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
 	}
 
 	/*
-	 * Need to send that every time, no matter if it actually changed
-	 * between the two trees as we have done changes to the inode before.
+	 * If other directory inodes depended on our current directory
+	 * inode's move/rename, now do their move/rename operations.
 	 */
+	if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
+		ret = apply_children_dir_moves(sctx);
+		if (ret)
+			goto out;
+	}
+
+	/*
+	 * Need to send that every time, no matter if it actually
+	 * changed between the two trees as we have done changes to
+	 * the inode before.
+	 */
+	sctx->send_progress = sctx->cur_ino + 1;
 	ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
 	if (ret < 0)
 		goto out;
@@ -4839,6 +5305,9 @@ long btrfs_ioctl_send(struct file *mnt_file, void __user *arg_)
 		goto out;
 	}
 
+	sctx->pending_dir_moves = RB_ROOT;
+	sctx->waiting_dir_moves = RB_ROOT;
+
 	sctx->clone_roots = vzalloc(sizeof(struct clone_root) *
 			(arg->clone_sources_count + 1));
 	if (!sctx->clone_roots) {
@@ -4947,6 +5416,34 @@ long btrfs_ioctl_send(struct file *mnt_file, void __user *arg_)
 	}
 
 out:
+	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
+	while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
+		struct rb_node *n;
+		struct pending_dir_move *pm;
+
+		n = rb_first(&sctx->pending_dir_moves);
+		pm = rb_entry(n, struct pending_dir_move, node);
+		while (!list_empty(&pm->list)) {
+			struct pending_dir_move *pm2;
+
+			pm2 = list_first_entry(&pm->list,
+					       struct pending_dir_move, list);
+			free_pending_move(sctx, pm2);
+		}
+		free_pending_move(sctx, pm);
+	}
+
+	WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
+	while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
+		struct rb_node *n;
+		struct waiting_dir_move *dm;
+
+		n = rb_first(&sctx->waiting_dir_moves);
+		dm = rb_entry(n, struct waiting_dir_move, node);
+		rb_erase(&dm->node, &sctx->waiting_dir_moves);
+		kfree(dm);
+	}
+
 	if (sort_clone_roots) {
 		for (i = 0; i < sctx->clone_roots_cnt; i++)
 			btrfs_root_dec_send_in_progress(