1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2012 Alexander Block. All rights reserved.
6 #include <linux/bsearch.h>
8 #include <linux/file.h>
9 #include <linux/sort.h>
10 #include <linux/mount.h>
11 #include <linux/xattr.h>
12 #include <linux/posix_acl_xattr.h>
13 #include <linux/radix-tree.h>
14 #include <linux/vmalloc.h>
15 #include <linux/string.h>
16 #include <linux/compat.h>
17 #include <linux/crc32c.h>
18 #include <linux/fsverity.h>
25 #include "btrfs_inode.h"
26 #include "transaction.h"
27 #include "compression.h"
28 #include "print-tree.h"
29 #include "accessors.h"
31 #include "file-item.h"
34 #include "lru_cache.h"
37 * Maximum number of references an extent can have in order for us to attempt to
38 * issue clone operations instead of write operations. This currently exists to
39 * avoid hitting limitations of the backreference walking code (taking a lot of
40 * time and using too much memory for extents with large number of references).
42 #define SEND_MAX_EXTENT_REFS 1024
45 * A fs_path is a helper to dynamically build path names with unknown size.
46 * It reallocates the internal buffer on demand.
47 * It allows fast adding of path elements on the right side (normal path) and
48 * fast adding to the left side (reversed path). A reversed path can also be
49 * unreversed if needed.
58 unsigned short buf_len:15;
59 unsigned short reversed:1;
63 * Average path length does not exceed 200 bytes, we'll have
64 * better packing in the slab and higher chance to satisfy
65 * a allocation later during send.
70 #define FS_PATH_INLINE_SIZE \
71 (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
74 /* reused for each extent */
76 struct btrfs_root *root;
83 #define SEND_MAX_NAME_CACHE_SIZE 256
86 * Limit the root_ids array of struct backref_cache_entry to 17 elements.
87 * This makes the size of a cache entry to be exactly 192 bytes on x86_64, which
88 * can be satisfied from the kmalloc-192 slab, without wasting any space.
89 * The most common case is to have a single root for cloning, which corresponds
90 * to the send root. Having the user specify more than 16 clone roots is not
91 * common, and in such rare cases we simply don't use caching if the number of
92 * cloning roots that lead down to a leaf is more than 17.
94 #define SEND_MAX_BACKREF_CACHE_ROOTS 17
97 * Max number of entries in the cache.
98 * With SEND_MAX_BACKREF_CACHE_ROOTS as 17, the size in bytes, excluding
99 * maple tree's internal nodes, is 24K.
101 #define SEND_MAX_BACKREF_CACHE_SIZE 128
104 * A backref cache entry maps a leaf to a list of IDs of roots from which the
105 * leaf is accessible and we can use for clone operations.
106 * With SEND_MAX_BACKREF_CACHE_ROOTS as 12, each cache entry is 128 bytes (on
109 struct backref_cache_entry {
110 struct btrfs_lru_cache_entry entry;
111 u64 root_ids[SEND_MAX_BACKREF_CACHE_ROOTS];
112 /* Number of valid elements in the root_ids array. */
116 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
117 static_assert(offsetof(struct backref_cache_entry, entry) == 0);
120 * Max number of entries in the cache that stores directories that were already
121 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
122 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
123 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
125 #define SEND_MAX_DIR_CREATED_CACHE_SIZE 64
128 * Max number of entries in the cache that stores directories that were already
129 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
130 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
131 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
133 #define SEND_MAX_DIR_UTIMES_CACHE_SIZE 64
136 struct file *send_filp;
142 * Whether BTRFS_SEND_A_DATA attribute was already added to current
143 * command (since protocol v2, data must be the last attribute).
146 struct page **send_buf_pages;
147 u64 flags; /* 'flags' member of btrfs_ioctl_send_args is u64 */
148 /* Protocol version compatibility requested */
151 struct btrfs_root *send_root;
152 struct btrfs_root *parent_root;
153 struct clone_root *clone_roots;
156 /* current state of the compare_tree call */
157 struct btrfs_path *left_path;
158 struct btrfs_path *right_path;
159 struct btrfs_key *cmp_key;
162 * Keep track of the generation of the last transaction that was used
163 * for relocating a block group. This is periodically checked in order
164 * to detect if a relocation happened since the last check, so that we
165 * don't operate on stale extent buffers for nodes (level >= 1) or on
166 * stale disk_bytenr values of file extent items.
168 u64 last_reloc_trans;
171 * infos of the currently processed inode. In case of deleted inodes,
172 * these are the values from the deleted inode.
179 u64 cur_inode_last_extent;
180 u64 cur_inode_next_write_offset;
182 bool cur_inode_new_gen;
183 bool cur_inode_deleted;
184 bool ignore_cur_inode;
185 bool cur_inode_needs_verity;
186 void *verity_descriptor;
190 struct list_head new_refs;
191 struct list_head deleted_refs;
193 struct btrfs_lru_cache name_cache;
196 * The inode we are currently processing. It's not NULL only when we
197 * need to issue write commands for data extents from this inode.
199 struct inode *cur_inode;
200 struct file_ra_state ra;
201 u64 page_cache_clear_start;
202 bool clean_page_cache;
205 * We process inodes by their increasing order, so if before an
206 * incremental send we reverse the parent/child relationship of
207 * directories such that a directory with a lower inode number was
208 * the parent of a directory with a higher inode number, and the one
209 * becoming the new parent got renamed too, we can't rename/move the
210 * directory with lower inode number when we finish processing it - we
211 * must process the directory with higher inode number first, then
212 * rename/move it and then rename/move the directory with lower inode
213 * number. Example follows.
215 * Tree state when the first send was performed:
227 * Tree state when the second (incremental) send is performed:
236 * The sequence of steps that lead to the second state was:
238 * mv /a/b/c/d /a/b/c2/d2
239 * mv /a/b/c /a/b/c2/d2/cc
241 * "c" has lower inode number, but we can't move it (2nd mv operation)
242 * before we move "d", which has higher inode number.
244 * So we just memorize which move/rename operations must be performed
245 * later when their respective parent is processed and moved/renamed.
248 /* Indexed by parent directory inode number. */
249 struct rb_root pending_dir_moves;
252 * Reverse index, indexed by the inode number of a directory that
253 * is waiting for the move/rename of its immediate parent before its
254 * own move/rename can be performed.
256 struct rb_root waiting_dir_moves;
259 * A directory that is going to be rm'ed might have a child directory
260 * which is in the pending directory moves index above. In this case,
261 * the directory can only be removed after the move/rename of its child
262 * is performed. Example:
282 * Sequence of steps that lead to the send snapshot:
283 * rm -f /a/b/c/foo.txt
285 * mv /a/b/c/x /a/b/YY
288 * When the child is processed, its move/rename is delayed until its
289 * parent is processed (as explained above), but all other operations
290 * like update utimes, chown, chgrp, etc, are performed and the paths
291 * that it uses for those operations must use the orphanized name of
292 * its parent (the directory we're going to rm later), so we need to
293 * memorize that name.
295 * Indexed by the inode number of the directory to be deleted.
297 struct rb_root orphan_dirs;
299 struct rb_root rbtree_new_refs;
300 struct rb_root rbtree_deleted_refs;
302 struct btrfs_lru_cache backref_cache;
303 u64 backref_cache_last_reloc_trans;
305 struct btrfs_lru_cache dir_created_cache;
306 struct btrfs_lru_cache dir_utimes_cache;
309 struct pending_dir_move {
311 struct list_head list;
315 struct list_head update_refs;
318 struct waiting_dir_move {
322 * There might be some directory that could not be removed because it
323 * was waiting for this directory inode to be moved first. Therefore
324 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
331 struct orphan_dir_info {
335 u64 last_dir_index_offset;
336 u64 dir_high_seq_ino;
339 struct name_cache_entry {
341 * The key in the entry is an inode number, and the generation matches
342 * the inode's generation.
344 struct btrfs_lru_cache_entry entry;
348 int need_later_update;
353 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
354 static_assert(offsetof(struct name_cache_entry, entry) == 0);
357 #define ADVANCE_ONLY_NEXT -1
359 enum btrfs_compare_tree_result {
360 BTRFS_COMPARE_TREE_NEW,
361 BTRFS_COMPARE_TREE_DELETED,
362 BTRFS_COMPARE_TREE_CHANGED,
363 BTRFS_COMPARE_TREE_SAME,
367 static void inconsistent_snapshot_error(struct send_ctx *sctx,
368 enum btrfs_compare_tree_result result,
371 const char *result_string;
374 case BTRFS_COMPARE_TREE_NEW:
375 result_string = "new";
377 case BTRFS_COMPARE_TREE_DELETED:
378 result_string = "deleted";
380 case BTRFS_COMPARE_TREE_CHANGED:
381 result_string = "updated";
383 case BTRFS_COMPARE_TREE_SAME:
385 result_string = "unchanged";
389 result_string = "unexpected";
392 btrfs_err(sctx->send_root->fs_info,
393 "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
394 result_string, what, sctx->cmp_key->objectid,
395 sctx->send_root->root_key.objectid,
397 sctx->parent_root->root_key.objectid : 0));
401 static bool proto_cmd_ok(const struct send_ctx *sctx, int cmd)
403 switch (sctx->proto) {
404 case 1: return cmd <= BTRFS_SEND_C_MAX_V1;
405 case 2: return cmd <= BTRFS_SEND_C_MAX_V2;
406 case 3: return cmd <= BTRFS_SEND_C_MAX_V3;
407 default: return false;
411 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
413 static struct waiting_dir_move *
414 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
416 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);
418 static int need_send_hole(struct send_ctx *sctx)
420 return (sctx->parent_root && !sctx->cur_inode_new &&
421 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
422 S_ISREG(sctx->cur_inode_mode));
425 static void fs_path_reset(struct fs_path *p)
428 p->start = p->buf + p->buf_len - 1;
438 static struct fs_path *fs_path_alloc(void)
442 p = kmalloc(sizeof(*p), GFP_KERNEL);
446 p->buf = p->inline_buf;
447 p->buf_len = FS_PATH_INLINE_SIZE;
452 static struct fs_path *fs_path_alloc_reversed(void)
464 static void fs_path_free(struct fs_path *p)
468 if (p->buf != p->inline_buf)
473 static int fs_path_len(struct fs_path *p)
475 return p->end - p->start;
478 static int fs_path_ensure_buf(struct fs_path *p, int len)
486 if (p->buf_len >= len)
489 if (len > PATH_MAX) {
494 path_len = p->end - p->start;
495 old_buf_len = p->buf_len;
498 * Allocate to the next largest kmalloc bucket size, to let
499 * the fast path happen most of the time.
501 len = kmalloc_size_roundup(len);
503 * First time the inline_buf does not suffice
505 if (p->buf == p->inline_buf) {
506 tmp_buf = kmalloc(len, GFP_KERNEL);
508 memcpy(tmp_buf, p->buf, old_buf_len);
510 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
518 tmp_buf = p->buf + old_buf_len - path_len - 1;
519 p->end = p->buf + p->buf_len - 1;
520 p->start = p->end - path_len;
521 memmove(p->start, tmp_buf, path_len + 1);
524 p->end = p->start + path_len;
529 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
535 new_len = p->end - p->start + name_len;
536 if (p->start != p->end)
538 ret = fs_path_ensure_buf(p, new_len);
543 if (p->start != p->end)
545 p->start -= name_len;
546 *prepared = p->start;
548 if (p->start != p->end)
559 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
564 ret = fs_path_prepare_for_add(p, name_len, &prepared);
567 memcpy(prepared, name, name_len);
573 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
578 ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
581 memcpy(prepared, p2->start, p2->end - p2->start);
587 static int fs_path_add_from_extent_buffer(struct fs_path *p,
588 struct extent_buffer *eb,
589 unsigned long off, int len)
594 ret = fs_path_prepare_for_add(p, len, &prepared);
598 read_extent_buffer(eb, prepared, off, len);
604 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
606 p->reversed = from->reversed;
609 return fs_path_add_path(p, from);
612 static void fs_path_unreverse(struct fs_path *p)
621 len = p->end - p->start;
623 p->end = p->start + len;
624 memmove(p->start, tmp, len + 1);
628 static struct btrfs_path *alloc_path_for_send(void)
630 struct btrfs_path *path;
632 path = btrfs_alloc_path();
635 path->search_commit_root = 1;
636 path->skip_locking = 1;
637 path->need_commit_sem = 1;
641 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
647 ret = kernel_write(filp, buf + pos, len - pos, off);
658 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
660 struct btrfs_tlv_header *hdr;
661 int total_len = sizeof(*hdr) + len;
662 int left = sctx->send_max_size - sctx->send_size;
664 if (WARN_ON_ONCE(sctx->put_data))
667 if (unlikely(left < total_len))
670 hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
671 put_unaligned_le16(attr, &hdr->tlv_type);
672 put_unaligned_le16(len, &hdr->tlv_len);
673 memcpy(hdr + 1, data, len);
674 sctx->send_size += total_len;
679 #define TLV_PUT_DEFINE_INT(bits) \
680 static int tlv_put_u##bits(struct send_ctx *sctx, \
681 u##bits attr, u##bits value) \
683 __le##bits __tmp = cpu_to_le##bits(value); \
684 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp)); \
687 TLV_PUT_DEFINE_INT(8)
688 TLV_PUT_DEFINE_INT(32)
689 TLV_PUT_DEFINE_INT(64)
691 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
692 const char *str, int len)
696 return tlv_put(sctx, attr, str, len);
699 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
702 return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
705 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
706 struct extent_buffer *eb,
707 struct btrfs_timespec *ts)
709 struct btrfs_timespec bts;
710 read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
711 return tlv_put(sctx, attr, &bts, sizeof(bts));
715 #define TLV_PUT(sctx, attrtype, data, attrlen) \
717 ret = tlv_put(sctx, attrtype, data, attrlen); \
719 goto tlv_put_failure; \
722 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
724 ret = tlv_put_u##bits(sctx, attrtype, value); \
726 goto tlv_put_failure; \
729 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
730 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
731 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
732 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
733 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
735 ret = tlv_put_string(sctx, attrtype, str, len); \
737 goto tlv_put_failure; \
739 #define TLV_PUT_PATH(sctx, attrtype, p) \
741 ret = tlv_put_string(sctx, attrtype, p->start, \
742 p->end - p->start); \
744 goto tlv_put_failure; \
746 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
748 ret = tlv_put_uuid(sctx, attrtype, uuid); \
750 goto tlv_put_failure; \
752 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
754 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
756 goto tlv_put_failure; \
759 static int send_header(struct send_ctx *sctx)
761 struct btrfs_stream_header hdr;
763 strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
764 hdr.version = cpu_to_le32(sctx->proto);
765 return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
770 * For each command/item we want to send to userspace, we call this function.
772 static int begin_cmd(struct send_ctx *sctx, int cmd)
774 struct btrfs_cmd_header *hdr;
776 if (WARN_ON(!sctx->send_buf))
779 if (unlikely(sctx->send_size != 0)) {
780 btrfs_err(sctx->send_root->fs_info,
781 "send: command header buffer not empty cmd %d offset %llu",
782 cmd, sctx->send_off);
786 sctx->send_size += sizeof(*hdr);
787 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
788 put_unaligned_le16(cmd, &hdr->cmd);
793 static int send_cmd(struct send_ctx *sctx)
796 struct btrfs_cmd_header *hdr;
799 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
800 put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
801 put_unaligned_le32(0, &hdr->crc);
803 crc = crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
804 put_unaligned_le32(crc, &hdr->crc);
806 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
810 sctx->put_data = false;
816 * Sends a move instruction to user space
818 static int send_rename(struct send_ctx *sctx,
819 struct fs_path *from, struct fs_path *to)
821 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
824 btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
826 ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
830 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
831 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
833 ret = send_cmd(sctx);
841 * Sends a link instruction to user space
843 static int send_link(struct send_ctx *sctx,
844 struct fs_path *path, struct fs_path *lnk)
846 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
849 btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
851 ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
855 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
856 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
858 ret = send_cmd(sctx);
866 * Sends an unlink instruction to user space
868 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
870 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
873 btrfs_debug(fs_info, "send_unlink %s", path->start);
875 ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
879 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
881 ret = send_cmd(sctx);
889 * Sends a rmdir instruction to user space
891 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
893 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
896 btrfs_debug(fs_info, "send_rmdir %s", path->start);
898 ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
902 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
904 ret = send_cmd(sctx);
911 struct btrfs_inode_info {
923 * Helper function to retrieve some fields from an inode item.
925 static int get_inode_info(struct btrfs_root *root, u64 ino,
926 struct btrfs_inode_info *info)
929 struct btrfs_path *path;
930 struct btrfs_inode_item *ii;
931 struct btrfs_key key;
933 path = alloc_path_for_send();
938 key.type = BTRFS_INODE_ITEM_KEY;
940 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
950 ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
951 struct btrfs_inode_item);
952 info->size = btrfs_inode_size(path->nodes[0], ii);
953 info->gen = btrfs_inode_generation(path->nodes[0], ii);
954 info->mode = btrfs_inode_mode(path->nodes[0], ii);
955 info->uid = btrfs_inode_uid(path->nodes[0], ii);
956 info->gid = btrfs_inode_gid(path->nodes[0], ii);
957 info->rdev = btrfs_inode_rdev(path->nodes[0], ii);
958 info->nlink = btrfs_inode_nlink(path->nodes[0], ii);
960 * Transfer the unchanged u64 value of btrfs_inode_item::flags, that's
961 * otherwise logically split to 32/32 parts.
963 info->fileattr = btrfs_inode_flags(path->nodes[0], ii);
966 btrfs_free_path(path);
970 static int get_inode_gen(struct btrfs_root *root, u64 ino, u64 *gen)
973 struct btrfs_inode_info info = { 0 };
977 ret = get_inode_info(root, ino, &info);
982 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
987 * Helper function to iterate the entries in ONE btrfs_inode_ref or
988 * btrfs_inode_extref.
989 * The iterate callback may return a non zero value to stop iteration. This can
990 * be a negative value for error codes or 1 to simply stop it.
992 * path must point to the INODE_REF or INODE_EXTREF when called.
994 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
995 struct btrfs_key *found_key, int resolve,
996 iterate_inode_ref_t iterate, void *ctx)
998 struct extent_buffer *eb = path->nodes[0];
999 struct btrfs_inode_ref *iref;
1000 struct btrfs_inode_extref *extref;
1001 struct btrfs_path *tmp_path;
1005 int slot = path->slots[0];
1012 unsigned long name_off;
1013 unsigned long elem_size;
1016 p = fs_path_alloc_reversed();
1020 tmp_path = alloc_path_for_send();
1027 if (found_key->type == BTRFS_INODE_REF_KEY) {
1028 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
1029 struct btrfs_inode_ref);
1030 total = btrfs_item_size(eb, slot);
1031 elem_size = sizeof(*iref);
1033 ptr = btrfs_item_ptr_offset(eb, slot);
1034 total = btrfs_item_size(eb, slot);
1035 elem_size = sizeof(*extref);
1038 while (cur < total) {
1041 if (found_key->type == BTRFS_INODE_REF_KEY) {
1042 iref = (struct btrfs_inode_ref *)(ptr + cur);
1043 name_len = btrfs_inode_ref_name_len(eb, iref);
1044 name_off = (unsigned long)(iref + 1);
1045 index = btrfs_inode_ref_index(eb, iref);
1046 dir = found_key->offset;
1048 extref = (struct btrfs_inode_extref *)(ptr + cur);
1049 name_len = btrfs_inode_extref_name_len(eb, extref);
1050 name_off = (unsigned long)&extref->name;
1051 index = btrfs_inode_extref_index(eb, extref);
1052 dir = btrfs_inode_extref_parent(eb, extref);
1056 start = btrfs_ref_to_path(root, tmp_path, name_len,
1058 p->buf, p->buf_len);
1059 if (IS_ERR(start)) {
1060 ret = PTR_ERR(start);
1063 if (start < p->buf) {
1064 /* overflow , try again with larger buffer */
1065 ret = fs_path_ensure_buf(p,
1066 p->buf_len + p->buf - start);
1069 start = btrfs_ref_to_path(root, tmp_path,
1072 p->buf, p->buf_len);
1073 if (IS_ERR(start)) {
1074 ret = PTR_ERR(start);
1077 if (unlikely(start < p->buf)) {
1078 btrfs_err(root->fs_info,
1079 "send: path ref buffer underflow for key (%llu %u %llu)",
1080 found_key->objectid,
1089 ret = fs_path_add_from_extent_buffer(p, eb, name_off,
1095 cur += elem_size + name_len;
1096 ret = iterate(num, dir, index, p, ctx);
1103 btrfs_free_path(tmp_path);
1108 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
1109 const char *name, int name_len,
1110 const char *data, int data_len,
1114 * Helper function to iterate the entries in ONE btrfs_dir_item.
1115 * The iterate callback may return a non zero value to stop iteration. This can
1116 * be a negative value for error codes or 1 to simply stop it.
1118 * path must point to the dir item when called.
1120 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1121 iterate_dir_item_t iterate, void *ctx)
1124 struct extent_buffer *eb;
1125 struct btrfs_dir_item *di;
1126 struct btrfs_key di_key;
1138 * Start with a small buffer (1 page). If later we end up needing more
1139 * space, which can happen for xattrs on a fs with a leaf size greater
1140 * then the page size, attempt to increase the buffer. Typically xattr
1144 buf = kmalloc(buf_len, GFP_KERNEL);
1150 eb = path->nodes[0];
1151 slot = path->slots[0];
1152 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1155 total = btrfs_item_size(eb, slot);
1158 while (cur < total) {
1159 name_len = btrfs_dir_name_len(eb, di);
1160 data_len = btrfs_dir_data_len(eb, di);
1161 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1163 if (btrfs_dir_ftype(eb, di) == BTRFS_FT_XATTR) {
1164 if (name_len > XATTR_NAME_MAX) {
1165 ret = -ENAMETOOLONG;
1168 if (name_len + data_len >
1169 BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1177 if (name_len + data_len > PATH_MAX) {
1178 ret = -ENAMETOOLONG;
1183 if (name_len + data_len > buf_len) {
1184 buf_len = name_len + data_len;
1185 if (is_vmalloc_addr(buf)) {
1189 char *tmp = krealloc(buf, buf_len,
1190 GFP_KERNEL | __GFP_NOWARN);
1197 buf = kvmalloc(buf_len, GFP_KERNEL);
1205 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1206 name_len + data_len);
1208 len = sizeof(*di) + name_len + data_len;
1209 di = (struct btrfs_dir_item *)((char *)di + len);
1212 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1229 static int __copy_first_ref(int num, u64 dir, int index,
1230 struct fs_path *p, void *ctx)
1233 struct fs_path *pt = ctx;
1235 ret = fs_path_copy(pt, p);
1239 /* we want the first only */
1244 * Retrieve the first path of an inode. If an inode has more then one
1245 * ref/hardlink, this is ignored.
1247 static int get_inode_path(struct btrfs_root *root,
1248 u64 ino, struct fs_path *path)
1251 struct btrfs_key key, found_key;
1252 struct btrfs_path *p;
1254 p = alloc_path_for_send();
1258 fs_path_reset(path);
1261 key.type = BTRFS_INODE_REF_KEY;
1264 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1271 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1272 if (found_key.objectid != ino ||
1273 (found_key.type != BTRFS_INODE_REF_KEY &&
1274 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1279 ret = iterate_inode_ref(root, p, &found_key, 1,
1280 __copy_first_ref, path);
1290 struct backref_ctx {
1291 struct send_ctx *sctx;
1293 /* number of total found references */
1297 * used for clones found in send_root. clones found behind cur_objectid
1298 * and cur_offset are not considered as allowed clones.
1303 /* may be truncated in case it's the last extent in a file */
1306 /* The bytenr the file extent item we are processing refers to. */
1308 /* The owner (root id) of the data backref for the current extent. */
1310 /* The offset of the data backref for the current extent. */
1314 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1316 u64 root = (u64)(uintptr_t)key;
1317 const struct clone_root *cr = elt;
1319 if (root < cr->root->root_key.objectid)
1321 if (root > cr->root->root_key.objectid)
1326 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1328 const struct clone_root *cr1 = e1;
1329 const struct clone_root *cr2 = e2;
1331 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1333 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1339 * Called for every backref that is found for the current extent.
1340 * Results are collected in sctx->clone_roots->ino/offset.
1342 static int iterate_backrefs(u64 ino, u64 offset, u64 num_bytes, u64 root_id,
1345 struct backref_ctx *bctx = ctx_;
1346 struct clone_root *clone_root;
1348 /* First check if the root is in the list of accepted clone sources */
1349 clone_root = bsearch((void *)(uintptr_t)root_id, bctx->sctx->clone_roots,
1350 bctx->sctx->clone_roots_cnt,
1351 sizeof(struct clone_root),
1352 __clone_root_cmp_bsearch);
1356 /* This is our own reference, bail out as we can't clone from it. */
1357 if (clone_root->root == bctx->sctx->send_root &&
1358 ino == bctx->cur_objectid &&
1359 offset == bctx->cur_offset)
1363 * Make sure we don't consider clones from send_root that are
1364 * behind the current inode/offset.
1366 if (clone_root->root == bctx->sctx->send_root) {
1368 * If the source inode was not yet processed we can't issue a
1369 * clone operation, as the source extent does not exist yet at
1370 * the destination of the stream.
1372 if (ino > bctx->cur_objectid)
1375 * We clone from the inode currently being sent as long as the
1376 * source extent is already processed, otherwise we could try
1377 * to clone from an extent that does not exist yet at the
1378 * destination of the stream.
1380 if (ino == bctx->cur_objectid &&
1381 offset + bctx->extent_len >
1382 bctx->sctx->cur_inode_next_write_offset)
1387 clone_root->found_ref = true;
1390 * If the given backref refers to a file extent item with a larger
1391 * number of bytes than what we found before, use the new one so that
1392 * we clone more optimally and end up doing less writes and getting
1393 * less exclusive, non-shared extents at the destination.
1395 if (num_bytes > clone_root->num_bytes) {
1396 clone_root->ino = ino;
1397 clone_root->offset = offset;
1398 clone_root->num_bytes = num_bytes;
1401 * Found a perfect candidate, so there's no need to continue
1404 if (num_bytes >= bctx->extent_len)
1405 return BTRFS_ITERATE_EXTENT_INODES_STOP;
1411 static bool lookup_backref_cache(u64 leaf_bytenr, void *ctx,
1412 const u64 **root_ids_ret, int *root_count_ret)
1414 struct backref_ctx *bctx = ctx;
1415 struct send_ctx *sctx = bctx->sctx;
1416 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1417 const u64 key = leaf_bytenr >> fs_info->sectorsize_bits;
1418 struct btrfs_lru_cache_entry *raw_entry;
1419 struct backref_cache_entry *entry;
1421 if (sctx->backref_cache.size == 0)
1425 * If relocation happened since we first filled the cache, then we must
1426 * empty the cache and can not use it, because even though we operate on
1427 * read-only roots, their leaves and nodes may have been reallocated and
1428 * now be used for different nodes/leaves of the same tree or some other
1431 * We are called from iterate_extent_inodes() while either holding a
1432 * transaction handle or holding fs_info->commit_root_sem, so no need
1433 * to take any lock here.
1435 if (fs_info->last_reloc_trans > sctx->backref_cache_last_reloc_trans) {
1436 btrfs_lru_cache_clear(&sctx->backref_cache);
1440 raw_entry = btrfs_lru_cache_lookup(&sctx->backref_cache, key, 0);
1444 entry = container_of(raw_entry, struct backref_cache_entry, entry);
1445 *root_ids_ret = entry->root_ids;
1446 *root_count_ret = entry->num_roots;
1451 static void store_backref_cache(u64 leaf_bytenr, const struct ulist *root_ids,
1454 struct backref_ctx *bctx = ctx;
1455 struct send_ctx *sctx = bctx->sctx;
1456 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1457 struct backref_cache_entry *new_entry;
1458 struct ulist_iterator uiter;
1459 struct ulist_node *node;
1463 * We're called while holding a transaction handle or while holding
1464 * fs_info->commit_root_sem (at iterate_extent_inodes()), so must do a
1467 new_entry = kmalloc(sizeof(struct backref_cache_entry), GFP_NOFS);
1468 /* No worries, cache is optional. */
1472 new_entry->entry.key = leaf_bytenr >> fs_info->sectorsize_bits;
1473 new_entry->entry.gen = 0;
1474 new_entry->num_roots = 0;
1475 ULIST_ITER_INIT(&uiter);
1476 while ((node = ulist_next(root_ids, &uiter)) != NULL) {
1477 const u64 root_id = node->val;
1478 struct clone_root *root;
1480 root = bsearch((void *)(uintptr_t)root_id, sctx->clone_roots,
1481 sctx->clone_roots_cnt, sizeof(struct clone_root),
1482 __clone_root_cmp_bsearch);
1486 /* Too many roots, just exit, no worries as caching is optional. */
1487 if (new_entry->num_roots >= SEND_MAX_BACKREF_CACHE_ROOTS) {
1492 new_entry->root_ids[new_entry->num_roots] = root_id;
1493 new_entry->num_roots++;
1497 * We may have not added any roots to the new cache entry, which means
1498 * none of the roots is part of the list of roots from which we are
1499 * allowed to clone. Cache the new entry as it's still useful to avoid
1500 * backref walking to determine which roots have a path to the leaf.
1502 * Also use GFP_NOFS because we're called while holding a transaction
1503 * handle or while holding fs_info->commit_root_sem.
1505 ret = btrfs_lru_cache_store(&sctx->backref_cache, &new_entry->entry,
1507 ASSERT(ret == 0 || ret == -ENOMEM);
1509 /* Caching is optional, no worries. */
1515 * We are called from iterate_extent_inodes() while either holding a
1516 * transaction handle or holding fs_info->commit_root_sem, so no need
1517 * to take any lock here.
1519 if (sctx->backref_cache.size == 1)
1520 sctx->backref_cache_last_reloc_trans = fs_info->last_reloc_trans;
1523 static int check_extent_item(u64 bytenr, const struct btrfs_extent_item *ei,
1524 const struct extent_buffer *leaf, void *ctx)
1526 const u64 refs = btrfs_extent_refs(leaf, ei);
1527 const struct backref_ctx *bctx = ctx;
1528 const struct send_ctx *sctx = bctx->sctx;
1530 if (bytenr == bctx->bytenr) {
1531 const u64 flags = btrfs_extent_flags(leaf, ei);
1533 if (WARN_ON(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
1537 * If we have only one reference and only the send root as a
1538 * clone source - meaning no clone roots were given in the
1539 * struct btrfs_ioctl_send_args passed to the send ioctl - then
1540 * it's our reference and there's no point in doing backref
1541 * walking which is expensive, so exit early.
1543 if (refs == 1 && sctx->clone_roots_cnt == 1)
1548 * Backreference walking (iterate_extent_inodes() below) is currently
1549 * too expensive when an extent has a large number of references, both
1550 * in time spent and used memory. So for now just fallback to write
1551 * operations instead of clone operations when an extent has more than
1552 * a certain amount of references.
1554 if (refs > SEND_MAX_EXTENT_REFS)
1560 static bool skip_self_data_ref(u64 root, u64 ino, u64 offset, void *ctx)
1562 const struct backref_ctx *bctx = ctx;
1564 if (ino == bctx->cur_objectid &&
1565 root == bctx->backref_owner &&
1566 offset == bctx->backref_offset)
1573 * Given an inode, offset and extent item, it finds a good clone for a clone
1574 * instruction. Returns -ENOENT when none could be found. The function makes
1575 * sure that the returned clone is usable at the point where sending is at the
1576 * moment. This means, that no clones are accepted which lie behind the current
1579 * path must point to the extent item when called.
1581 static int find_extent_clone(struct send_ctx *sctx,
1582 struct btrfs_path *path,
1583 u64 ino, u64 data_offset,
1585 struct clone_root **found)
1587 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1593 struct btrfs_file_extent_item *fi;
1594 struct extent_buffer *eb = path->nodes[0];
1595 struct backref_ctx backref_ctx = { 0 };
1596 struct btrfs_backref_walk_ctx backref_walk_ctx = { 0 };
1597 struct clone_root *cur_clone_root;
1602 * With fallocate we can get prealloc extents beyond the inode's i_size,
1603 * so we don't do anything here because clone operations can not clone
1604 * to a range beyond i_size without increasing the i_size of the
1605 * destination inode.
1607 if (data_offset >= ino_size)
1610 fi = btrfs_item_ptr(eb, path->slots[0], struct btrfs_file_extent_item);
1611 extent_type = btrfs_file_extent_type(eb, fi);
1612 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1615 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1619 compressed = btrfs_file_extent_compression(eb, fi);
1620 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1621 logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1624 * Setup the clone roots.
1626 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1627 cur_clone_root = sctx->clone_roots + i;
1628 cur_clone_root->ino = (u64)-1;
1629 cur_clone_root->offset = 0;
1630 cur_clone_root->num_bytes = 0;
1631 cur_clone_root->found_ref = false;
1634 backref_ctx.sctx = sctx;
1635 backref_ctx.cur_objectid = ino;
1636 backref_ctx.cur_offset = data_offset;
1637 backref_ctx.bytenr = disk_byte;
1639 * Use the header owner and not the send root's id, because in case of a
1640 * snapshot we can have shared subtrees.
1642 backref_ctx.backref_owner = btrfs_header_owner(eb);
1643 backref_ctx.backref_offset = data_offset - btrfs_file_extent_offset(eb, fi);
1646 * The last extent of a file may be too large due to page alignment.
1647 * We need to adjust extent_len in this case so that the checks in
1648 * iterate_backrefs() work.
1650 if (data_offset + num_bytes >= ino_size)
1651 backref_ctx.extent_len = ino_size - data_offset;
1653 backref_ctx.extent_len = num_bytes;
1656 * Now collect all backrefs.
1658 backref_walk_ctx.bytenr = disk_byte;
1659 if (compressed == BTRFS_COMPRESS_NONE)
1660 backref_walk_ctx.extent_item_pos = btrfs_file_extent_offset(eb, fi);
1661 backref_walk_ctx.fs_info = fs_info;
1662 backref_walk_ctx.cache_lookup = lookup_backref_cache;
1663 backref_walk_ctx.cache_store = store_backref_cache;
1664 backref_walk_ctx.indirect_ref_iterator = iterate_backrefs;
1665 backref_walk_ctx.check_extent_item = check_extent_item;
1666 backref_walk_ctx.user_ctx = &backref_ctx;
1669 * If have a single clone root, then it's the send root and we can tell
1670 * the backref walking code to skip our own backref and not resolve it,
1671 * since we can not use it for cloning - the source and destination
1672 * ranges can't overlap and in case the leaf is shared through a subtree
1673 * due to snapshots, we can't use those other roots since they are not
1674 * in the list of clone roots.
1676 if (sctx->clone_roots_cnt == 1)
1677 backref_walk_ctx.skip_data_ref = skip_self_data_ref;
1679 ret = iterate_extent_inodes(&backref_walk_ctx, true, iterate_backrefs,
1684 down_read(&fs_info->commit_root_sem);
1685 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
1687 * A transaction commit for a transaction in which block group
1688 * relocation was done just happened.
1689 * The disk_bytenr of the file extent item we processed is
1690 * possibly stale, referring to the extent's location before
1691 * relocation. So act as if we haven't found any clone sources
1692 * and fallback to write commands, which will read the correct
1693 * data from the new extent location. Otherwise we will fail
1694 * below because we haven't found our own back reference or we
1695 * could be getting incorrect sources in case the old extent
1696 * was already reallocated after the relocation.
1698 up_read(&fs_info->commit_root_sem);
1701 up_read(&fs_info->commit_root_sem);
1703 btrfs_debug(fs_info,
1704 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1705 data_offset, ino, num_bytes, logical);
1707 if (!backref_ctx.found) {
1708 btrfs_debug(fs_info, "no clones found");
1712 cur_clone_root = NULL;
1713 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1714 struct clone_root *clone_root = &sctx->clone_roots[i];
1716 if (!clone_root->found_ref)
1720 * Choose the root from which we can clone more bytes, to
1721 * minimize write operations and therefore have more extent
1722 * sharing at the destination (the same as in the source).
1724 if (!cur_clone_root ||
1725 clone_root->num_bytes > cur_clone_root->num_bytes) {
1726 cur_clone_root = clone_root;
1729 * We found an optimal clone candidate (any inode from
1730 * any root is fine), so we're done.
1732 if (clone_root->num_bytes >= backref_ctx.extent_len)
1737 if (cur_clone_root) {
1738 *found = cur_clone_root;
1747 static int read_symlink(struct btrfs_root *root,
1749 struct fs_path *dest)
1752 struct btrfs_path *path;
1753 struct btrfs_key key;
1754 struct btrfs_file_extent_item *ei;
1760 path = alloc_path_for_send();
1765 key.type = BTRFS_EXTENT_DATA_KEY;
1767 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1772 * An empty symlink inode. Can happen in rare error paths when
1773 * creating a symlink (transaction committed before the inode
1774 * eviction handler removed the symlink inode items and a crash
1775 * happened in between or the subvol was snapshoted in between).
1776 * Print an informative message to dmesg/syslog so that the user
1777 * can delete the symlink.
1779 btrfs_err(root->fs_info,
1780 "Found empty symlink inode %llu at root %llu",
1781 ino, root->root_key.objectid);
1786 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1787 struct btrfs_file_extent_item);
1788 type = btrfs_file_extent_type(path->nodes[0], ei);
1789 if (unlikely(type != BTRFS_FILE_EXTENT_INLINE)) {
1791 btrfs_crit(root->fs_info,
1792 "send: found symlink extent that is not inline, ino %llu root %llu extent type %d",
1793 ino, btrfs_root_id(root), type);
1796 compression = btrfs_file_extent_compression(path->nodes[0], ei);
1797 if (unlikely(compression != BTRFS_COMPRESS_NONE)) {
1799 btrfs_crit(root->fs_info,
1800 "send: found symlink extent with compression, ino %llu root %llu compression type %d",
1801 ino, btrfs_root_id(root), compression);
1805 off = btrfs_file_extent_inline_start(ei);
1806 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1808 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1811 btrfs_free_path(path);
1816 * Helper function to generate a file name that is unique in the root of
1817 * send_root and parent_root. This is used to generate names for orphan inodes.
1819 static int gen_unique_name(struct send_ctx *sctx,
1821 struct fs_path *dest)
1824 struct btrfs_path *path;
1825 struct btrfs_dir_item *di;
1830 path = alloc_path_for_send();
1835 struct fscrypt_str tmp_name;
1837 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1839 ASSERT(len < sizeof(tmp));
1840 tmp_name.name = tmp;
1841 tmp_name.len = strlen(tmp);
1843 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1844 path, BTRFS_FIRST_FREE_OBJECTID,
1846 btrfs_release_path(path);
1852 /* not unique, try again */
1857 if (!sctx->parent_root) {
1863 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1864 path, BTRFS_FIRST_FREE_OBJECTID,
1866 btrfs_release_path(path);
1872 /* not unique, try again */
1880 ret = fs_path_add(dest, tmp, strlen(tmp));
1883 btrfs_free_path(path);
1888 inode_state_no_change,
1889 inode_state_will_create,
1890 inode_state_did_create,
1891 inode_state_will_delete,
1892 inode_state_did_delete,
1895 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen,
1896 u64 *send_gen, u64 *parent_gen)
1903 struct btrfs_inode_info info;
1905 ret = get_inode_info(sctx->send_root, ino, &info);
1906 if (ret < 0 && ret != -ENOENT)
1908 left_ret = (info.nlink == 0) ? -ENOENT : ret;
1909 left_gen = info.gen;
1911 *send_gen = ((left_ret == -ENOENT) ? 0 : info.gen);
1913 if (!sctx->parent_root) {
1914 right_ret = -ENOENT;
1916 ret = get_inode_info(sctx->parent_root, ino, &info);
1917 if (ret < 0 && ret != -ENOENT)
1919 right_ret = (info.nlink == 0) ? -ENOENT : ret;
1920 right_gen = info.gen;
1922 *parent_gen = ((right_ret == -ENOENT) ? 0 : info.gen);
1925 if (!left_ret && !right_ret) {
1926 if (left_gen == gen && right_gen == gen) {
1927 ret = inode_state_no_change;
1928 } else if (left_gen == gen) {
1929 if (ino < sctx->send_progress)
1930 ret = inode_state_did_create;
1932 ret = inode_state_will_create;
1933 } else if (right_gen == gen) {
1934 if (ino < sctx->send_progress)
1935 ret = inode_state_did_delete;
1937 ret = inode_state_will_delete;
1941 } else if (!left_ret) {
1942 if (left_gen == gen) {
1943 if (ino < sctx->send_progress)
1944 ret = inode_state_did_create;
1946 ret = inode_state_will_create;
1950 } else if (!right_ret) {
1951 if (right_gen == gen) {
1952 if (ino < sctx->send_progress)
1953 ret = inode_state_did_delete;
1955 ret = inode_state_will_delete;
1967 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen,
1968 u64 *send_gen, u64 *parent_gen)
1972 if (ino == BTRFS_FIRST_FREE_OBJECTID)
1975 ret = get_cur_inode_state(sctx, ino, gen, send_gen, parent_gen);
1979 if (ret == inode_state_no_change ||
1980 ret == inode_state_did_create ||
1981 ret == inode_state_will_delete)
1991 * Helper function to lookup a dir item in a dir.
1993 static int lookup_dir_item_inode(struct btrfs_root *root,
1994 u64 dir, const char *name, int name_len,
1998 struct btrfs_dir_item *di;
1999 struct btrfs_key key;
2000 struct btrfs_path *path;
2001 struct fscrypt_str name_str = FSTR_INIT((char *)name, name_len);
2003 path = alloc_path_for_send();
2007 di = btrfs_lookup_dir_item(NULL, root, path, dir, &name_str, 0);
2008 if (IS_ERR_OR_NULL(di)) {
2009 ret = di ? PTR_ERR(di) : -ENOENT;
2012 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
2013 if (key.type == BTRFS_ROOT_ITEM_KEY) {
2017 *found_inode = key.objectid;
2020 btrfs_free_path(path);
2025 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
2026 * generation of the parent dir and the name of the dir entry.
2028 static int get_first_ref(struct btrfs_root *root, u64 ino,
2029 u64 *dir, u64 *dir_gen, struct fs_path *name)
2032 struct btrfs_key key;
2033 struct btrfs_key found_key;
2034 struct btrfs_path *path;
2038 path = alloc_path_for_send();
2043 key.type = BTRFS_INODE_REF_KEY;
2046 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
2050 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2052 if (ret || found_key.objectid != ino ||
2053 (found_key.type != BTRFS_INODE_REF_KEY &&
2054 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
2059 if (found_key.type == BTRFS_INODE_REF_KEY) {
2060 struct btrfs_inode_ref *iref;
2061 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2062 struct btrfs_inode_ref);
2063 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
2064 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2065 (unsigned long)(iref + 1),
2067 parent_dir = found_key.offset;
2069 struct btrfs_inode_extref *extref;
2070 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2071 struct btrfs_inode_extref);
2072 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
2073 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2074 (unsigned long)&extref->name, len);
2075 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
2079 btrfs_release_path(path);
2082 ret = get_inode_gen(root, parent_dir, dir_gen);
2090 btrfs_free_path(path);
2094 static int is_first_ref(struct btrfs_root *root,
2096 const char *name, int name_len)
2099 struct fs_path *tmp_name;
2102 tmp_name = fs_path_alloc();
2106 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
2110 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
2115 ret = !memcmp(tmp_name->start, name, name_len);
2118 fs_path_free(tmp_name);
2123 * Used by process_recorded_refs to determine if a new ref would overwrite an
2124 * already existing ref. In case it detects an overwrite, it returns the
2125 * inode/gen in who_ino/who_gen.
2126 * When an overwrite is detected, process_recorded_refs does proper orphanizing
2127 * to make sure later references to the overwritten inode are possible.
2128 * Orphanizing is however only required for the first ref of an inode.
2129 * process_recorded_refs does an additional is_first_ref check to see if
2130 * orphanizing is really required.
2132 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2133 const char *name, int name_len,
2134 u64 *who_ino, u64 *who_gen, u64 *who_mode)
2137 u64 parent_root_dir_gen;
2138 u64 other_inode = 0;
2139 struct btrfs_inode_info info;
2141 if (!sctx->parent_root)
2144 ret = is_inode_existent(sctx, dir, dir_gen, NULL, &parent_root_dir_gen);
2149 * If we have a parent root we need to verify that the parent dir was
2150 * not deleted and then re-created, if it was then we have no overwrite
2151 * and we can just unlink this entry.
2153 * @parent_root_dir_gen was set to 0 if the inode does not exist in the
2156 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID &&
2157 parent_root_dir_gen != dir_gen)
2160 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
2168 * Check if the overwritten ref was already processed. If yes, the ref
2169 * was already unlinked/moved, so we can safely assume that we will not
2170 * overwrite anything at this point in time.
2172 if (other_inode > sctx->send_progress ||
2173 is_waiting_for_move(sctx, other_inode)) {
2174 ret = get_inode_info(sctx->parent_root, other_inode, &info);
2178 *who_ino = other_inode;
2179 *who_gen = info.gen;
2180 *who_mode = info.mode;
2188 * Checks if the ref was overwritten by an already processed inode. This is
2189 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
2190 * thus the orphan name needs be used.
2191 * process_recorded_refs also uses it to avoid unlinking of refs that were
2194 static int did_overwrite_ref(struct send_ctx *sctx,
2195 u64 dir, u64 dir_gen,
2196 u64 ino, u64 ino_gen,
2197 const char *name, int name_len)
2202 u64 send_root_dir_gen;
2204 if (!sctx->parent_root)
2207 ret = is_inode_existent(sctx, dir, dir_gen, &send_root_dir_gen, NULL);
2212 * @send_root_dir_gen was set to 0 if the inode does not exist in the
2215 if (dir != BTRFS_FIRST_FREE_OBJECTID && send_root_dir_gen != dir_gen)
2218 /* check if the ref was overwritten by another ref */
2219 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
2221 if (ret == -ENOENT) {
2222 /* was never and will never be overwritten */
2224 } else if (ret < 0) {
2228 if (ow_inode == ino) {
2229 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2233 /* It's the same inode, so no overwrite happened. */
2234 if (ow_gen == ino_gen)
2239 * We know that it is or will be overwritten. Check this now.
2240 * The current inode being processed might have been the one that caused
2241 * inode 'ino' to be orphanized, therefore check if ow_inode matches
2242 * the current inode being processed.
2244 if (ow_inode < sctx->send_progress)
2247 if (ino != sctx->cur_ino && ow_inode == sctx->cur_ino) {
2249 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2253 if (ow_gen == sctx->cur_inode_gen)
2261 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
2262 * that got overwritten. This is used by process_recorded_refs to determine
2263 * if it has to use the path as returned by get_cur_path or the orphan name.
2265 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
2268 struct fs_path *name = NULL;
2272 if (!sctx->parent_root)
2275 name = fs_path_alloc();
2279 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2283 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2284 name->start, fs_path_len(name));
2291 static inline struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2294 struct btrfs_lru_cache_entry *entry;
2296 entry = btrfs_lru_cache_lookup(&sctx->name_cache, ino, gen);
2300 return container_of(entry, struct name_cache_entry, entry);
2304 * Used by get_cur_path for each ref up to the root.
2305 * Returns 0 if it succeeded.
2306 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2307 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2308 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2309 * Returns <0 in case of error.
2311 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2315 struct fs_path *dest)
2319 struct name_cache_entry *nce;
2322 * First check if we already did a call to this function with the same
2323 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2324 * return the cached result.
2326 nce = name_cache_search(sctx, ino, gen);
2328 if (ino < sctx->send_progress && nce->need_later_update) {
2329 btrfs_lru_cache_remove(&sctx->name_cache, &nce->entry);
2332 *parent_ino = nce->parent_ino;
2333 *parent_gen = nce->parent_gen;
2334 ret = fs_path_add(dest, nce->name, nce->name_len);
2343 * If the inode is not existent yet, add the orphan name and return 1.
2344 * This should only happen for the parent dir that we determine in
2345 * record_new_ref_if_needed().
2347 ret = is_inode_existent(sctx, ino, gen, NULL, NULL);
2352 ret = gen_unique_name(sctx, ino, gen, dest);
2360 * Depending on whether the inode was already processed or not, use
2361 * send_root or parent_root for ref lookup.
2363 if (ino < sctx->send_progress)
2364 ret = get_first_ref(sctx->send_root, ino,
2365 parent_ino, parent_gen, dest);
2367 ret = get_first_ref(sctx->parent_root, ino,
2368 parent_ino, parent_gen, dest);
2373 * Check if the ref was overwritten by an inode's ref that was processed
2374 * earlier. If yes, treat as orphan and return 1.
2376 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2377 dest->start, dest->end - dest->start);
2381 fs_path_reset(dest);
2382 ret = gen_unique_name(sctx, ino, gen, dest);
2390 * Store the result of the lookup in the name cache.
2392 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2398 nce->entry.key = ino;
2399 nce->entry.gen = gen;
2400 nce->parent_ino = *parent_ino;
2401 nce->parent_gen = *parent_gen;
2402 nce->name_len = fs_path_len(dest);
2404 strcpy(nce->name, dest->start);
2406 if (ino < sctx->send_progress)
2407 nce->need_later_update = 0;
2409 nce->need_later_update = 1;
2411 nce_ret = btrfs_lru_cache_store(&sctx->name_cache, &nce->entry, GFP_KERNEL);
2422 * Magic happens here. This function returns the first ref to an inode as it
2423 * would look like while receiving the stream at this point in time.
2424 * We walk the path up to the root. For every inode in between, we check if it
2425 * was already processed/sent. If yes, we continue with the parent as found
2426 * in send_root. If not, we continue with the parent as found in parent_root.
2427 * If we encounter an inode that was deleted at this point in time, we use the
2428 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2429 * that were not created yet and overwritten inodes/refs.
2431 * When do we have orphan inodes:
2432 * 1. When an inode is freshly created and thus no valid refs are available yet
2433 * 2. When a directory lost all it's refs (deleted) but still has dir items
2434 * inside which were not processed yet (pending for move/delete). If anyone
2435 * tried to get the path to the dir items, it would get a path inside that
2437 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2438 * of an unprocessed inode. If in that case the first ref would be
2439 * overwritten, the overwritten inode gets "orphanized". Later when we
2440 * process this overwritten inode, it is restored at a new place by moving
2443 * sctx->send_progress tells this function at which point in time receiving
2446 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2447 struct fs_path *dest)
2450 struct fs_path *name = NULL;
2451 u64 parent_inode = 0;
2455 name = fs_path_alloc();
2462 fs_path_reset(dest);
2464 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2465 struct waiting_dir_move *wdm;
2467 fs_path_reset(name);
2469 if (is_waiting_for_rm(sctx, ino, gen)) {
2470 ret = gen_unique_name(sctx, ino, gen, name);
2473 ret = fs_path_add_path(dest, name);
2477 wdm = get_waiting_dir_move(sctx, ino);
2478 if (wdm && wdm->orphanized) {
2479 ret = gen_unique_name(sctx, ino, gen, name);
2482 ret = get_first_ref(sctx->parent_root, ino,
2483 &parent_inode, &parent_gen, name);
2485 ret = __get_cur_name_and_parent(sctx, ino, gen,
2495 ret = fs_path_add_path(dest, name);
2506 fs_path_unreverse(dest);
2511 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2513 static int send_subvol_begin(struct send_ctx *sctx)
2516 struct btrfs_root *send_root = sctx->send_root;
2517 struct btrfs_root *parent_root = sctx->parent_root;
2518 struct btrfs_path *path;
2519 struct btrfs_key key;
2520 struct btrfs_root_ref *ref;
2521 struct extent_buffer *leaf;
2525 path = btrfs_alloc_path();
2529 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2531 btrfs_free_path(path);
2535 key.objectid = send_root->root_key.objectid;
2536 key.type = BTRFS_ROOT_BACKREF_KEY;
2539 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2548 leaf = path->nodes[0];
2549 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2550 if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2551 key.objectid != send_root->root_key.objectid) {
2555 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2556 namelen = btrfs_root_ref_name_len(leaf, ref);
2557 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2558 btrfs_release_path(path);
2561 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2565 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2570 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2572 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2573 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2574 sctx->send_root->root_item.received_uuid);
2576 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2577 sctx->send_root->root_item.uuid);
2579 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2580 btrfs_root_ctransid(&sctx->send_root->root_item));
2582 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2583 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2584 parent_root->root_item.received_uuid);
2586 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2587 parent_root->root_item.uuid);
2588 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2589 btrfs_root_ctransid(&sctx->parent_root->root_item));
2592 ret = send_cmd(sctx);
2596 btrfs_free_path(path);
2601 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2603 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2607 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2609 p = fs_path_alloc();
2613 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2617 ret = get_cur_path(sctx, ino, gen, p);
2620 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2621 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2623 ret = send_cmd(sctx);
2631 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2633 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2637 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2639 p = fs_path_alloc();
2643 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2647 ret = get_cur_path(sctx, ino, gen, p);
2650 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2651 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2653 ret = send_cmd(sctx);
2661 static int send_fileattr(struct send_ctx *sctx, u64 ino, u64 gen, u64 fileattr)
2663 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2667 if (sctx->proto < 2)
2670 btrfs_debug(fs_info, "send_fileattr %llu fileattr=%llu", ino, fileattr);
2672 p = fs_path_alloc();
2676 ret = begin_cmd(sctx, BTRFS_SEND_C_FILEATTR);
2680 ret = get_cur_path(sctx, ino, gen, p);
2683 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2684 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILEATTR, fileattr);
2686 ret = send_cmd(sctx);
2694 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2696 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2700 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2703 p = fs_path_alloc();
2707 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2711 ret = get_cur_path(sctx, ino, gen, p);
2714 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2715 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2716 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2718 ret = send_cmd(sctx);
2726 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2728 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2730 struct fs_path *p = NULL;
2731 struct btrfs_inode_item *ii;
2732 struct btrfs_path *path = NULL;
2733 struct extent_buffer *eb;
2734 struct btrfs_key key;
2737 btrfs_debug(fs_info, "send_utimes %llu", ino);
2739 p = fs_path_alloc();
2743 path = alloc_path_for_send();
2750 key.type = BTRFS_INODE_ITEM_KEY;
2752 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2758 eb = path->nodes[0];
2759 slot = path->slots[0];
2760 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2762 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2766 ret = get_cur_path(sctx, ino, gen, p);
2769 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2770 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2771 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2772 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2773 if (sctx->proto >= 2)
2774 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_OTIME, eb, &ii->otime);
2776 ret = send_cmd(sctx);
2781 btrfs_free_path(path);
2786 * If the cache is full, we can't remove entries from it and do a call to
2787 * send_utimes() for each respective inode, because we might be finishing
2788 * processing an inode that is a directory and it just got renamed, and existing
2789 * entries in the cache may refer to inodes that have the directory in their
2790 * full path - in which case we would generate outdated paths (pre-rename)
2791 * for the inodes that the cache entries point to. Instead of prunning the
2792 * cache when inserting, do it after we finish processing each inode at
2793 * finish_inode_if_needed().
2795 static int cache_dir_utimes(struct send_ctx *sctx, u64 dir, u64 gen)
2797 struct btrfs_lru_cache_entry *entry;
2800 entry = btrfs_lru_cache_lookup(&sctx->dir_utimes_cache, dir, gen);
2804 /* Caching is optional, don't fail if we can't allocate memory. */
2805 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2807 return send_utimes(sctx, dir, gen);
2812 ret = btrfs_lru_cache_store(&sctx->dir_utimes_cache, entry, GFP_KERNEL);
2813 ASSERT(ret != -EEXIST);
2816 return send_utimes(sctx, dir, gen);
2822 static int trim_dir_utimes_cache(struct send_ctx *sctx)
2824 while (sctx->dir_utimes_cache.size > SEND_MAX_DIR_UTIMES_CACHE_SIZE) {
2825 struct btrfs_lru_cache_entry *lru;
2828 lru = btrfs_lru_cache_lru_entry(&sctx->dir_utimes_cache);
2829 ASSERT(lru != NULL);
2831 ret = send_utimes(sctx, lru->key, lru->gen);
2835 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, lru);
2842 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2843 * a valid path yet because we did not process the refs yet. So, the inode
2844 * is created as orphan.
2846 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2848 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2852 struct btrfs_inode_info info;
2857 btrfs_debug(fs_info, "send_create_inode %llu", ino);
2859 p = fs_path_alloc();
2863 if (ino != sctx->cur_ino) {
2864 ret = get_inode_info(sctx->send_root, ino, &info);
2871 gen = sctx->cur_inode_gen;
2872 mode = sctx->cur_inode_mode;
2873 rdev = sctx->cur_inode_rdev;
2876 if (S_ISREG(mode)) {
2877 cmd = BTRFS_SEND_C_MKFILE;
2878 } else if (S_ISDIR(mode)) {
2879 cmd = BTRFS_SEND_C_MKDIR;
2880 } else if (S_ISLNK(mode)) {
2881 cmd = BTRFS_SEND_C_SYMLINK;
2882 } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2883 cmd = BTRFS_SEND_C_MKNOD;
2884 } else if (S_ISFIFO(mode)) {
2885 cmd = BTRFS_SEND_C_MKFIFO;
2886 } else if (S_ISSOCK(mode)) {
2887 cmd = BTRFS_SEND_C_MKSOCK;
2889 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2890 (int)(mode & S_IFMT));
2895 ret = begin_cmd(sctx, cmd);
2899 ret = gen_unique_name(sctx, ino, gen, p);
2903 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2904 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2906 if (S_ISLNK(mode)) {
2908 ret = read_symlink(sctx->send_root, ino, p);
2911 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2912 } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2913 S_ISFIFO(mode) || S_ISSOCK(mode)) {
2914 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2915 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2918 ret = send_cmd(sctx);
2929 static void cache_dir_created(struct send_ctx *sctx, u64 dir)
2931 struct btrfs_lru_cache_entry *entry;
2934 /* Caching is optional, ignore any failures. */
2935 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2941 ret = btrfs_lru_cache_store(&sctx->dir_created_cache, entry, GFP_KERNEL);
2947 * We need some special handling for inodes that get processed before the parent
2948 * directory got created. See process_recorded_refs for details.
2949 * This function does the check if we already created the dir out of order.
2951 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2955 struct btrfs_path *path = NULL;
2956 struct btrfs_key key;
2957 struct btrfs_key found_key;
2958 struct btrfs_key di_key;
2959 struct btrfs_dir_item *di;
2961 if (btrfs_lru_cache_lookup(&sctx->dir_created_cache, dir, 0))
2964 path = alloc_path_for_send();
2969 key.type = BTRFS_DIR_INDEX_KEY;
2972 btrfs_for_each_slot(sctx->send_root, &key, &found_key, path, iter_ret) {
2973 struct extent_buffer *eb = path->nodes[0];
2975 if (found_key.objectid != key.objectid ||
2976 found_key.type != key.type) {
2981 di = btrfs_item_ptr(eb, path->slots[0], struct btrfs_dir_item);
2982 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2984 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2985 di_key.objectid < sctx->send_progress) {
2987 cache_dir_created(sctx, dir);
2991 /* Catch error found during iteration */
2995 btrfs_free_path(path);
3000 * Only creates the inode if it is:
3001 * 1. Not a directory
3002 * 2. Or a directory which was not created already due to out of order
3003 * directories. See did_create_dir and process_recorded_refs for details.
3005 static int send_create_inode_if_needed(struct send_ctx *sctx)
3009 if (S_ISDIR(sctx->cur_inode_mode)) {
3010 ret = did_create_dir(sctx, sctx->cur_ino);
3017 ret = send_create_inode(sctx, sctx->cur_ino);
3019 if (ret == 0 && S_ISDIR(sctx->cur_inode_mode))
3020 cache_dir_created(sctx, sctx->cur_ino);
3025 struct recorded_ref {
3026 struct list_head list;
3028 struct fs_path *full_path;
3032 struct rb_node node;
3033 struct rb_root *root;
3036 static struct recorded_ref *recorded_ref_alloc(void)
3038 struct recorded_ref *ref;
3040 ref = kzalloc(sizeof(*ref), GFP_KERNEL);
3043 RB_CLEAR_NODE(&ref->node);
3044 INIT_LIST_HEAD(&ref->list);
3048 static void recorded_ref_free(struct recorded_ref *ref)
3052 if (!RB_EMPTY_NODE(&ref->node))
3053 rb_erase(&ref->node, ref->root);
3054 list_del(&ref->list);
3055 fs_path_free(ref->full_path);
3059 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
3061 ref->full_path = path;
3062 ref->name = (char *)kbasename(ref->full_path->start);
3063 ref->name_len = ref->full_path->end - ref->name;
3066 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
3068 struct recorded_ref *new;
3070 new = recorded_ref_alloc();
3074 new->dir = ref->dir;
3075 new->dir_gen = ref->dir_gen;
3076 list_add_tail(&new->list, list);
3080 static void __free_recorded_refs(struct list_head *head)
3082 struct recorded_ref *cur;
3084 while (!list_empty(head)) {
3085 cur = list_entry(head->next, struct recorded_ref, list);
3086 recorded_ref_free(cur);
3090 static void free_recorded_refs(struct send_ctx *sctx)
3092 __free_recorded_refs(&sctx->new_refs);
3093 __free_recorded_refs(&sctx->deleted_refs);
3097 * Renames/moves a file/dir to its orphan name. Used when the first
3098 * ref of an unprocessed inode gets overwritten and for all non empty
3101 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
3102 struct fs_path *path)
3105 struct fs_path *orphan;
3107 orphan = fs_path_alloc();
3111 ret = gen_unique_name(sctx, ino, gen, orphan);
3115 ret = send_rename(sctx, path, orphan);
3118 fs_path_free(orphan);
3122 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
3123 u64 dir_ino, u64 dir_gen)
3125 struct rb_node **p = &sctx->orphan_dirs.rb_node;
3126 struct rb_node *parent = NULL;
3127 struct orphan_dir_info *entry, *odi;
3131 entry = rb_entry(parent, struct orphan_dir_info, node);
3132 if (dir_ino < entry->ino)
3134 else if (dir_ino > entry->ino)
3135 p = &(*p)->rb_right;
3136 else if (dir_gen < entry->gen)
3138 else if (dir_gen > entry->gen)
3139 p = &(*p)->rb_right;
3144 odi = kmalloc(sizeof(*odi), GFP_KERNEL);
3146 return ERR_PTR(-ENOMEM);
3149 odi->last_dir_index_offset = 0;
3150 odi->dir_high_seq_ino = 0;
3152 rb_link_node(&odi->node, parent, p);
3153 rb_insert_color(&odi->node, &sctx->orphan_dirs);
3157 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
3158 u64 dir_ino, u64 gen)
3160 struct rb_node *n = sctx->orphan_dirs.rb_node;
3161 struct orphan_dir_info *entry;
3164 entry = rb_entry(n, struct orphan_dir_info, node);
3165 if (dir_ino < entry->ino)
3167 else if (dir_ino > entry->ino)
3169 else if (gen < entry->gen)
3171 else if (gen > entry->gen)
3179 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
3181 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
3186 static void free_orphan_dir_info(struct send_ctx *sctx,
3187 struct orphan_dir_info *odi)
3191 rb_erase(&odi->node, &sctx->orphan_dirs);
3196 * Returns 1 if a directory can be removed at this point in time.
3197 * We check this by iterating all dir items and checking if the inode behind
3198 * the dir item was already processed.
3200 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen)
3204 struct btrfs_root *root = sctx->parent_root;
3205 struct btrfs_path *path;
3206 struct btrfs_key key;
3207 struct btrfs_key found_key;
3208 struct btrfs_key loc;
3209 struct btrfs_dir_item *di;
3210 struct orphan_dir_info *odi = NULL;
3211 u64 dir_high_seq_ino = 0;
3212 u64 last_dir_index_offset = 0;
3215 * Don't try to rmdir the top/root subvolume dir.
3217 if (dir == BTRFS_FIRST_FREE_OBJECTID)
3220 odi = get_orphan_dir_info(sctx, dir, dir_gen);
3221 if (odi && sctx->cur_ino < odi->dir_high_seq_ino)
3224 path = alloc_path_for_send();
3230 * Find the inode number associated with the last dir index
3231 * entry. This is very likely the inode with the highest number
3232 * of all inodes that have an entry in the directory. We can
3233 * then use it to avoid future calls to can_rmdir(), when
3234 * processing inodes with a lower number, from having to search
3235 * the parent root b+tree for dir index keys.
3238 key.type = BTRFS_DIR_INDEX_KEY;
3239 key.offset = (u64)-1;
3241 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3244 } else if (ret > 0) {
3245 /* Can't happen, the root is never empty. */
3246 ASSERT(path->slots[0] > 0);
3247 if (WARN_ON(path->slots[0] == 0)) {
3254 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3255 if (key.objectid != dir || key.type != BTRFS_DIR_INDEX_KEY) {
3256 /* No index keys, dir can be removed. */
3261 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3262 struct btrfs_dir_item);
3263 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3264 dir_high_seq_ino = loc.objectid;
3265 if (sctx->cur_ino < dir_high_seq_ino) {
3270 btrfs_release_path(path);
3274 key.type = BTRFS_DIR_INDEX_KEY;
3275 key.offset = (odi ? odi->last_dir_index_offset : 0);
3277 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
3278 struct waiting_dir_move *dm;
3280 if (found_key.objectid != key.objectid ||
3281 found_key.type != key.type)
3284 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3285 struct btrfs_dir_item);
3286 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3288 dir_high_seq_ino = max(dir_high_seq_ino, loc.objectid);
3289 last_dir_index_offset = found_key.offset;
3291 dm = get_waiting_dir_move(sctx, loc.objectid);
3293 dm->rmdir_ino = dir;
3294 dm->rmdir_gen = dir_gen;
3299 if (loc.objectid > sctx->cur_ino) {
3308 free_orphan_dir_info(sctx, odi);
3313 btrfs_free_path(path);
3319 odi = add_orphan_dir_info(sctx, dir, dir_gen);
3321 return PTR_ERR(odi);
3326 odi->last_dir_index_offset = last_dir_index_offset;
3327 odi->dir_high_seq_ino = max(odi->dir_high_seq_ino, dir_high_seq_ino);
3332 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3334 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3336 return entry != NULL;
3339 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3341 struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3342 struct rb_node *parent = NULL;
3343 struct waiting_dir_move *entry, *dm;
3345 dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3351 dm->orphanized = orphanized;
3355 entry = rb_entry(parent, struct waiting_dir_move, node);
3356 if (ino < entry->ino) {
3358 } else if (ino > entry->ino) {
3359 p = &(*p)->rb_right;
3366 rb_link_node(&dm->node, parent, p);
3367 rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3371 static struct waiting_dir_move *
3372 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3374 struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3375 struct waiting_dir_move *entry;
3378 entry = rb_entry(n, struct waiting_dir_move, node);
3379 if (ino < entry->ino)
3381 else if (ino > entry->ino)
3389 static void free_waiting_dir_move(struct send_ctx *sctx,
3390 struct waiting_dir_move *dm)
3394 rb_erase(&dm->node, &sctx->waiting_dir_moves);
3398 static int add_pending_dir_move(struct send_ctx *sctx,
3402 struct list_head *new_refs,
3403 struct list_head *deleted_refs,
3404 const bool is_orphan)
3406 struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3407 struct rb_node *parent = NULL;
3408 struct pending_dir_move *entry = NULL, *pm;
3409 struct recorded_ref *cur;
3413 pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3416 pm->parent_ino = parent_ino;
3419 INIT_LIST_HEAD(&pm->list);
3420 INIT_LIST_HEAD(&pm->update_refs);
3421 RB_CLEAR_NODE(&pm->node);
3425 entry = rb_entry(parent, struct pending_dir_move, node);
3426 if (parent_ino < entry->parent_ino) {
3428 } else if (parent_ino > entry->parent_ino) {
3429 p = &(*p)->rb_right;
3436 list_for_each_entry(cur, deleted_refs, list) {
3437 ret = dup_ref(cur, &pm->update_refs);
3441 list_for_each_entry(cur, new_refs, list) {
3442 ret = dup_ref(cur, &pm->update_refs);
3447 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3452 list_add_tail(&pm->list, &entry->list);
3454 rb_link_node(&pm->node, parent, p);
3455 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3460 __free_recorded_refs(&pm->update_refs);
3466 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3469 struct rb_node *n = sctx->pending_dir_moves.rb_node;
3470 struct pending_dir_move *entry;
3473 entry = rb_entry(n, struct pending_dir_move, node);
3474 if (parent_ino < entry->parent_ino)
3476 else if (parent_ino > entry->parent_ino)
3484 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3485 u64 ino, u64 gen, u64 *ancestor_ino)
3488 u64 parent_inode = 0;
3490 u64 start_ino = ino;
3493 while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3494 fs_path_reset(name);
3496 if (is_waiting_for_rm(sctx, ino, gen))
3498 if (is_waiting_for_move(sctx, ino)) {
3499 if (*ancestor_ino == 0)
3500 *ancestor_ino = ino;
3501 ret = get_first_ref(sctx->parent_root, ino,
3502 &parent_inode, &parent_gen, name);
3504 ret = __get_cur_name_and_parent(sctx, ino, gen,
3514 if (parent_inode == start_ino) {
3516 if (*ancestor_ino == 0)
3517 *ancestor_ino = ino;
3526 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3528 struct fs_path *from_path = NULL;
3529 struct fs_path *to_path = NULL;
3530 struct fs_path *name = NULL;
3531 u64 orig_progress = sctx->send_progress;
3532 struct recorded_ref *cur;
3533 u64 parent_ino, parent_gen;
3534 struct waiting_dir_move *dm = NULL;
3541 name = fs_path_alloc();
3542 from_path = fs_path_alloc();
3543 if (!name || !from_path) {
3548 dm = get_waiting_dir_move(sctx, pm->ino);
3550 rmdir_ino = dm->rmdir_ino;
3551 rmdir_gen = dm->rmdir_gen;
3552 is_orphan = dm->orphanized;
3553 free_waiting_dir_move(sctx, dm);
3556 ret = gen_unique_name(sctx, pm->ino,
3557 pm->gen, from_path);
3559 ret = get_first_ref(sctx->parent_root, pm->ino,
3560 &parent_ino, &parent_gen, name);
3563 ret = get_cur_path(sctx, parent_ino, parent_gen,
3567 ret = fs_path_add_path(from_path, name);
3572 sctx->send_progress = sctx->cur_ino + 1;
3573 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3577 LIST_HEAD(deleted_refs);
3578 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3579 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3580 &pm->update_refs, &deleted_refs,
3585 dm = get_waiting_dir_move(sctx, pm->ino);
3587 dm->rmdir_ino = rmdir_ino;
3588 dm->rmdir_gen = rmdir_gen;
3592 fs_path_reset(name);
3595 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3599 ret = send_rename(sctx, from_path, to_path);
3604 struct orphan_dir_info *odi;
3607 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3609 /* already deleted */
3614 ret = can_rmdir(sctx, rmdir_ino, gen);
3620 name = fs_path_alloc();
3625 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3628 ret = send_rmdir(sctx, name);
3634 ret = cache_dir_utimes(sctx, pm->ino, pm->gen);
3639 * After rename/move, need to update the utimes of both new parent(s)
3640 * and old parent(s).
3642 list_for_each_entry(cur, &pm->update_refs, list) {
3644 * The parent inode might have been deleted in the send snapshot
3646 ret = get_inode_info(sctx->send_root, cur->dir, NULL);
3647 if (ret == -ENOENT) {
3654 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
3661 fs_path_free(from_path);
3662 fs_path_free(to_path);
3663 sctx->send_progress = orig_progress;
3668 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3670 if (!list_empty(&m->list))
3672 if (!RB_EMPTY_NODE(&m->node))
3673 rb_erase(&m->node, &sctx->pending_dir_moves);
3674 __free_recorded_refs(&m->update_refs);
3678 static void tail_append_pending_moves(struct send_ctx *sctx,
3679 struct pending_dir_move *moves,
3680 struct list_head *stack)
3682 if (list_empty(&moves->list)) {
3683 list_add_tail(&moves->list, stack);
3686 list_splice_init(&moves->list, &list);
3687 list_add_tail(&moves->list, stack);
3688 list_splice_tail(&list, stack);
3690 if (!RB_EMPTY_NODE(&moves->node)) {
3691 rb_erase(&moves->node, &sctx->pending_dir_moves);
3692 RB_CLEAR_NODE(&moves->node);
3696 static int apply_children_dir_moves(struct send_ctx *sctx)
3698 struct pending_dir_move *pm;
3700 u64 parent_ino = sctx->cur_ino;
3703 pm = get_pending_dir_moves(sctx, parent_ino);
3707 tail_append_pending_moves(sctx, pm, &stack);
3709 while (!list_empty(&stack)) {
3710 pm = list_first_entry(&stack, struct pending_dir_move, list);
3711 parent_ino = pm->ino;
3712 ret = apply_dir_move(sctx, pm);
3713 free_pending_move(sctx, pm);
3716 pm = get_pending_dir_moves(sctx, parent_ino);
3718 tail_append_pending_moves(sctx, pm, &stack);
3723 while (!list_empty(&stack)) {
3724 pm = list_first_entry(&stack, struct pending_dir_move, list);
3725 free_pending_move(sctx, pm);
3731 * We might need to delay a directory rename even when no ancestor directory
3732 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3733 * renamed. This happens when we rename a directory to the old name (the name
3734 * in the parent root) of some other unrelated directory that got its rename
3735 * delayed due to some ancestor with higher number that got renamed.
3741 * |---- a/ (ino 257)
3742 * | |---- file (ino 260)
3744 * |---- b/ (ino 258)
3745 * |---- c/ (ino 259)
3749 * |---- a/ (ino 258)
3750 * |---- x/ (ino 259)
3751 * |---- y/ (ino 257)
3752 * |----- file (ino 260)
3754 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3755 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3756 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3759 * 1 - rename 259 from 'c' to 'x'
3760 * 2 - rename 257 from 'a' to 'x/y'
3761 * 3 - rename 258 from 'b' to 'a'
3763 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3764 * be done right away and < 0 on error.
3766 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3767 struct recorded_ref *parent_ref,
3768 const bool is_orphan)
3770 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3771 struct btrfs_path *path;
3772 struct btrfs_key key;
3773 struct btrfs_key di_key;
3774 struct btrfs_dir_item *di;
3778 struct waiting_dir_move *wdm;
3780 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3783 path = alloc_path_for_send();
3787 key.objectid = parent_ref->dir;
3788 key.type = BTRFS_DIR_ITEM_KEY;
3789 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3791 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3794 } else if (ret > 0) {
3799 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3800 parent_ref->name_len);
3806 * di_key.objectid has the number of the inode that has a dentry in the
3807 * parent directory with the same name that sctx->cur_ino is being
3808 * renamed to. We need to check if that inode is in the send root as
3809 * well and if it is currently marked as an inode with a pending rename,
3810 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3811 * that it happens after that other inode is renamed.
3813 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3814 if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3819 ret = get_inode_gen(sctx->parent_root, di_key.objectid, &left_gen);
3822 ret = get_inode_gen(sctx->send_root, di_key.objectid, &right_gen);
3829 /* Different inode, no need to delay the rename of sctx->cur_ino */
3830 if (right_gen != left_gen) {
3835 wdm = get_waiting_dir_move(sctx, di_key.objectid);
3836 if (wdm && !wdm->orphanized) {
3837 ret = add_pending_dir_move(sctx,
3839 sctx->cur_inode_gen,
3842 &sctx->deleted_refs,
3848 btrfs_free_path(path);
3853 * Check if inode ino2, or any of its ancestors, is inode ino1.
3854 * Return 1 if true, 0 if false and < 0 on error.
3856 static int check_ino_in_path(struct btrfs_root *root,
3861 struct fs_path *fs_path)
3866 return ino1_gen == ino2_gen;
3868 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3873 fs_path_reset(fs_path);
3874 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3878 return parent_gen == ino1_gen;
3885 * Check if inode ino1 is an ancestor of inode ino2 in the given root for any
3886 * possible path (in case ino2 is not a directory and has multiple hard links).
3887 * Return 1 if true, 0 if false and < 0 on error.
3889 static int is_ancestor(struct btrfs_root *root,
3893 struct fs_path *fs_path)
3895 bool free_fs_path = false;
3898 struct btrfs_path *path = NULL;
3899 struct btrfs_key key;
3902 fs_path = fs_path_alloc();
3905 free_fs_path = true;
3908 path = alloc_path_for_send();
3914 key.objectid = ino2;
3915 key.type = BTRFS_INODE_REF_KEY;
3918 btrfs_for_each_slot(root, &key, &key, path, iter_ret) {
3919 struct extent_buffer *leaf = path->nodes[0];
3920 int slot = path->slots[0];
3924 if (key.objectid != ino2)
3926 if (key.type != BTRFS_INODE_REF_KEY &&
3927 key.type != BTRFS_INODE_EXTREF_KEY)
3930 item_size = btrfs_item_size(leaf, slot);
3931 while (cur_offset < item_size) {
3935 if (key.type == BTRFS_INODE_EXTREF_KEY) {
3937 struct btrfs_inode_extref *extref;
3939 ptr = btrfs_item_ptr_offset(leaf, slot);
3940 extref = (struct btrfs_inode_extref *)
3942 parent = btrfs_inode_extref_parent(leaf,
3944 cur_offset += sizeof(*extref);
3945 cur_offset += btrfs_inode_extref_name_len(leaf,
3948 parent = key.offset;
3949 cur_offset = item_size;
3952 ret = get_inode_gen(root, parent, &parent_gen);
3955 ret = check_ino_in_path(root, ino1, ino1_gen,
3956 parent, parent_gen, fs_path);
3966 btrfs_free_path(path);
3968 fs_path_free(fs_path);
3972 static int wait_for_parent_move(struct send_ctx *sctx,
3973 struct recorded_ref *parent_ref,
3974 const bool is_orphan)
3977 u64 ino = parent_ref->dir;
3978 u64 ino_gen = parent_ref->dir_gen;
3979 u64 parent_ino_before, parent_ino_after;
3980 struct fs_path *path_before = NULL;
3981 struct fs_path *path_after = NULL;
3984 path_after = fs_path_alloc();
3985 path_before = fs_path_alloc();
3986 if (!path_after || !path_before) {
3992 * Our current directory inode may not yet be renamed/moved because some
3993 * ancestor (immediate or not) has to be renamed/moved first. So find if
3994 * such ancestor exists and make sure our own rename/move happens after
3995 * that ancestor is processed to avoid path build infinite loops (done
3996 * at get_cur_path()).
3998 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3999 u64 parent_ino_after_gen;
4001 if (is_waiting_for_move(sctx, ino)) {
4003 * If the current inode is an ancestor of ino in the
4004 * parent root, we need to delay the rename of the
4005 * current inode, otherwise don't delayed the rename
4006 * because we can end up with a circular dependency
4007 * of renames, resulting in some directories never
4008 * getting the respective rename operations issued in
4009 * the send stream or getting into infinite path build
4012 ret = is_ancestor(sctx->parent_root,
4013 sctx->cur_ino, sctx->cur_inode_gen,
4019 fs_path_reset(path_before);
4020 fs_path_reset(path_after);
4022 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
4023 &parent_ino_after_gen, path_after);
4026 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
4028 if (ret < 0 && ret != -ENOENT) {
4030 } else if (ret == -ENOENT) {
4035 len1 = fs_path_len(path_before);
4036 len2 = fs_path_len(path_after);
4037 if (ino > sctx->cur_ino &&
4038 (parent_ino_before != parent_ino_after || len1 != len2 ||
4039 memcmp(path_before->start, path_after->start, len1))) {
4042 ret = get_inode_gen(sctx->parent_root, ino, &parent_ino_gen);
4045 if (ino_gen == parent_ino_gen) {
4050 ino = parent_ino_after;
4051 ino_gen = parent_ino_after_gen;
4055 fs_path_free(path_before);
4056 fs_path_free(path_after);
4059 ret = add_pending_dir_move(sctx,
4061 sctx->cur_inode_gen,
4064 &sctx->deleted_refs,
4073 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4076 struct fs_path *new_path;
4079 * Our reference's name member points to its full_path member string, so
4080 * we use here a new path.
4082 new_path = fs_path_alloc();
4086 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
4088 fs_path_free(new_path);
4091 ret = fs_path_add(new_path, ref->name, ref->name_len);
4093 fs_path_free(new_path);
4097 fs_path_free(ref->full_path);
4098 set_ref_path(ref, new_path);
4104 * When processing the new references for an inode we may orphanize an existing
4105 * directory inode because its old name conflicts with one of the new references
4106 * of the current inode. Later, when processing another new reference of our
4107 * inode, we might need to orphanize another inode, but the path we have in the
4108 * reference reflects the pre-orphanization name of the directory we previously
4109 * orphanized. For example:
4111 * parent snapshot looks like:
4114 * |----- f1 (ino 257)
4115 * |----- f2 (ino 258)
4116 * |----- d1/ (ino 259)
4117 * |----- d2/ (ino 260)
4119 * send snapshot looks like:
4122 * |----- d1 (ino 258)
4123 * |----- f2/ (ino 259)
4124 * |----- f2_link/ (ino 260)
4125 * | |----- f1 (ino 257)
4127 * |----- d2 (ino 258)
4129 * When processing inode 257 we compute the name for inode 259 as "d1", and we
4130 * cache it in the name cache. Later when we start processing inode 258, when
4131 * collecting all its new references we set a full path of "d1/d2" for its new
4132 * reference with name "d2". When we start processing the new references we
4133 * start by processing the new reference with name "d1", and this results in
4134 * orphanizing inode 259, since its old reference causes a conflict. Then we
4135 * move on the next new reference, with name "d2", and we find out we must
4136 * orphanize inode 260, as its old reference conflicts with ours - but for the
4137 * orphanization we use a source path corresponding to the path we stored in the
4138 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
4139 * receiver fail since the path component "d1/" no longer exists, it was renamed
4140 * to "o259-6-0/" when processing the previous new reference. So in this case we
4141 * must recompute the path in the new reference and use it for the new
4142 * orphanization operation.
4144 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4149 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
4153 fs_path_reset(ref->full_path);
4154 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
4158 ret = fs_path_add(ref->full_path, name, ref->name_len);
4162 /* Update the reference's base name pointer. */
4163 set_ref_path(ref, ref->full_path);
4170 * This does all the move/link/unlink/rmdir magic.
4172 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
4174 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4176 struct recorded_ref *cur;
4177 struct recorded_ref *cur2;
4178 LIST_HEAD(check_dirs);
4179 struct fs_path *valid_path = NULL;
4183 int did_overwrite = 0;
4185 u64 last_dir_ino_rm = 0;
4186 bool can_rename = true;
4187 bool orphanized_dir = false;
4188 bool orphanized_ancestor = false;
4190 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
4193 * This should never happen as the root dir always has the same ref
4194 * which is always '..'
4196 if (unlikely(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID)) {
4198 "send: unexpected inode %llu in process_recorded_refs()",
4204 valid_path = fs_path_alloc();
4211 * First, check if the first ref of the current inode was overwritten
4212 * before. If yes, we know that the current inode was already orphanized
4213 * and thus use the orphan name. If not, we can use get_cur_path to
4214 * get the path of the first ref as it would like while receiving at
4215 * this point in time.
4216 * New inodes are always orphan at the beginning, so force to use the
4217 * orphan name in this case.
4218 * The first ref is stored in valid_path and will be updated if it
4219 * gets moved around.
4221 if (!sctx->cur_inode_new) {
4222 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
4223 sctx->cur_inode_gen);
4229 if (sctx->cur_inode_new || did_overwrite) {
4230 ret = gen_unique_name(sctx, sctx->cur_ino,
4231 sctx->cur_inode_gen, valid_path);
4236 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4243 * Before doing any rename and link operations, do a first pass on the
4244 * new references to orphanize any unprocessed inodes that may have a
4245 * reference that conflicts with one of the new references of the current
4246 * inode. This needs to happen first because a new reference may conflict
4247 * with the old reference of a parent directory, so we must make sure
4248 * that the path used for link and rename commands don't use an
4249 * orphanized name when an ancestor was not yet orphanized.
4256 * |----- testdir/ (ino 259)
4257 * | |----- a (ino 257)
4259 * |----- b (ino 258)
4264 * |----- testdir_2/ (ino 259)
4265 * | |----- a (ino 260)
4267 * |----- testdir (ino 257)
4268 * |----- b (ino 257)
4269 * |----- b2 (ino 258)
4271 * Processing the new reference for inode 257 with name "b" may happen
4272 * before processing the new reference with name "testdir". If so, we
4273 * must make sure that by the time we send a link command to create the
4274 * hard link "b", inode 259 was already orphanized, since the generated
4275 * path in "valid_path" already contains the orphanized name for 259.
4276 * We are processing inode 257, so only later when processing 259 we do
4277 * the rename operation to change its temporary (orphanized) name to
4280 list_for_each_entry(cur, &sctx->new_refs, list) {
4281 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4284 if (ret == inode_state_will_create)
4288 * Check if this new ref would overwrite the first ref of another
4289 * unprocessed inode. If yes, orphanize the overwritten inode.
4290 * If we find an overwritten ref that is not the first ref,
4293 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4294 cur->name, cur->name_len,
4295 &ow_inode, &ow_gen, &ow_mode);
4299 ret = is_first_ref(sctx->parent_root,
4300 ow_inode, cur->dir, cur->name,
4305 struct name_cache_entry *nce;
4306 struct waiting_dir_move *wdm;
4308 if (orphanized_dir) {
4309 ret = refresh_ref_path(sctx, cur);
4314 ret = orphanize_inode(sctx, ow_inode, ow_gen,
4318 if (S_ISDIR(ow_mode))
4319 orphanized_dir = true;
4322 * If ow_inode has its rename operation delayed
4323 * make sure that its orphanized name is used in
4324 * the source path when performing its rename
4327 wdm = get_waiting_dir_move(sctx, ow_inode);
4329 wdm->orphanized = true;
4332 * Make sure we clear our orphanized inode's
4333 * name from the name cache. This is because the
4334 * inode ow_inode might be an ancestor of some
4335 * other inode that will be orphanized as well
4336 * later and has an inode number greater than
4337 * sctx->send_progress. We need to prevent
4338 * future name lookups from using the old name
4339 * and get instead the orphan name.
4341 nce = name_cache_search(sctx, ow_inode, ow_gen);
4343 btrfs_lru_cache_remove(&sctx->name_cache,
4347 * ow_inode might currently be an ancestor of
4348 * cur_ino, therefore compute valid_path (the
4349 * current path of cur_ino) again because it
4350 * might contain the pre-orphanization name of
4351 * ow_inode, which is no longer valid.
4353 ret = is_ancestor(sctx->parent_root,
4355 sctx->cur_ino, NULL);
4357 orphanized_ancestor = true;
4358 fs_path_reset(valid_path);
4359 ret = get_cur_path(sctx, sctx->cur_ino,
4360 sctx->cur_inode_gen,
4367 * If we previously orphanized a directory that
4368 * collided with a new reference that we already
4369 * processed, recompute the current path because
4370 * that directory may be part of the path.
4372 if (orphanized_dir) {
4373 ret = refresh_ref_path(sctx, cur);
4377 ret = send_unlink(sctx, cur->full_path);
4385 list_for_each_entry(cur, &sctx->new_refs, list) {
4387 * We may have refs where the parent directory does not exist
4388 * yet. This happens if the parent directories inum is higher
4389 * than the current inum. To handle this case, we create the
4390 * parent directory out of order. But we need to check if this
4391 * did already happen before due to other refs in the same dir.
4393 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4396 if (ret == inode_state_will_create) {
4399 * First check if any of the current inodes refs did
4400 * already create the dir.
4402 list_for_each_entry(cur2, &sctx->new_refs, list) {
4405 if (cur2->dir == cur->dir) {
4412 * If that did not happen, check if a previous inode
4413 * did already create the dir.
4416 ret = did_create_dir(sctx, cur->dir);
4420 ret = send_create_inode(sctx, cur->dir);
4423 cache_dir_created(sctx, cur->dir);
4427 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4428 ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4437 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4439 ret = wait_for_parent_move(sctx, cur, is_orphan);
4449 * link/move the ref to the new place. If we have an orphan
4450 * inode, move it and update valid_path. If not, link or move
4451 * it depending on the inode mode.
4453 if (is_orphan && can_rename) {
4454 ret = send_rename(sctx, valid_path, cur->full_path);
4458 ret = fs_path_copy(valid_path, cur->full_path);
4461 } else if (can_rename) {
4462 if (S_ISDIR(sctx->cur_inode_mode)) {
4464 * Dirs can't be linked, so move it. For moved
4465 * dirs, we always have one new and one deleted
4466 * ref. The deleted ref is ignored later.
4468 ret = send_rename(sctx, valid_path,
4471 ret = fs_path_copy(valid_path,
4477 * We might have previously orphanized an inode
4478 * which is an ancestor of our current inode,
4479 * so our reference's full path, which was
4480 * computed before any such orphanizations, must
4483 if (orphanized_dir) {
4484 ret = update_ref_path(sctx, cur);
4488 ret = send_link(sctx, cur->full_path,
4494 ret = dup_ref(cur, &check_dirs);
4499 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4501 * Check if we can already rmdir the directory. If not,
4502 * orphanize it. For every dir item inside that gets deleted
4503 * later, we do this check again and rmdir it then if possible.
4504 * See the use of check_dirs for more details.
4506 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen);
4510 ret = send_rmdir(sctx, valid_path);
4513 } else if (!is_orphan) {
4514 ret = orphanize_inode(sctx, sctx->cur_ino,
4515 sctx->cur_inode_gen, valid_path);
4521 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4522 ret = dup_ref(cur, &check_dirs);
4526 } else if (S_ISDIR(sctx->cur_inode_mode) &&
4527 !list_empty(&sctx->deleted_refs)) {
4529 * We have a moved dir. Add the old parent to check_dirs
4531 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4533 ret = dup_ref(cur, &check_dirs);
4536 } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4538 * We have a non dir inode. Go through all deleted refs and
4539 * unlink them if they were not already overwritten by other
4542 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4543 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4544 sctx->cur_ino, sctx->cur_inode_gen,
4545 cur->name, cur->name_len);
4550 * If we orphanized any ancestor before, we need
4551 * to recompute the full path for deleted names,
4552 * since any such path was computed before we
4553 * processed any references and orphanized any
4556 if (orphanized_ancestor) {
4557 ret = update_ref_path(sctx, cur);
4561 ret = send_unlink(sctx, cur->full_path);
4565 ret = dup_ref(cur, &check_dirs);
4570 * If the inode is still orphan, unlink the orphan. This may
4571 * happen when a previous inode did overwrite the first ref
4572 * of this inode and no new refs were added for the current
4573 * inode. Unlinking does not mean that the inode is deleted in
4574 * all cases. There may still be links to this inode in other
4578 ret = send_unlink(sctx, valid_path);
4585 * We did collect all parent dirs where cur_inode was once located. We
4586 * now go through all these dirs and check if they are pending for
4587 * deletion and if it's finally possible to perform the rmdir now.
4588 * We also update the inode stats of the parent dirs here.
4590 list_for_each_entry(cur, &check_dirs, list) {
4592 * In case we had refs into dirs that were not processed yet,
4593 * we don't need to do the utime and rmdir logic for these dirs.
4594 * The dir will be processed later.
4596 if (cur->dir > sctx->cur_ino)
4599 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4603 if (ret == inode_state_did_create ||
4604 ret == inode_state_no_change) {
4605 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
4608 } else if (ret == inode_state_did_delete &&
4609 cur->dir != last_dir_ino_rm) {
4610 ret = can_rmdir(sctx, cur->dir, cur->dir_gen);
4614 ret = get_cur_path(sctx, cur->dir,
4615 cur->dir_gen, valid_path);
4618 ret = send_rmdir(sctx, valid_path);
4621 last_dir_ino_rm = cur->dir;
4629 __free_recorded_refs(&check_dirs);
4630 free_recorded_refs(sctx);
4631 fs_path_free(valid_path);
4635 static int rbtree_ref_comp(const void *k, const struct rb_node *node)
4637 const struct recorded_ref *data = k;
4638 const struct recorded_ref *ref = rb_entry(node, struct recorded_ref, node);
4641 if (data->dir > ref->dir)
4643 if (data->dir < ref->dir)
4645 if (data->dir_gen > ref->dir_gen)
4647 if (data->dir_gen < ref->dir_gen)
4649 if (data->name_len > ref->name_len)
4651 if (data->name_len < ref->name_len)
4653 result = strcmp(data->name, ref->name);
4661 static bool rbtree_ref_less(struct rb_node *node, const struct rb_node *parent)
4663 const struct recorded_ref *entry = rb_entry(node, struct recorded_ref, node);
4665 return rbtree_ref_comp(entry, parent) < 0;
4668 static int record_ref_in_tree(struct rb_root *root, struct list_head *refs,
4669 struct fs_path *name, u64 dir, u64 dir_gen,
4670 struct send_ctx *sctx)
4673 struct fs_path *path = NULL;
4674 struct recorded_ref *ref = NULL;
4676 path = fs_path_alloc();
4682 ref = recorded_ref_alloc();
4688 ret = get_cur_path(sctx, dir, dir_gen, path);
4691 ret = fs_path_add_path(path, name);
4696 ref->dir_gen = dir_gen;
4697 set_ref_path(ref, path);
4698 list_add_tail(&ref->list, refs);
4699 rb_add(&ref->node, root, rbtree_ref_less);
4703 if (path && (!ref || !ref->full_path))
4705 recorded_ref_free(ref);
4710 static int record_new_ref_if_needed(int num, u64 dir, int index,
4711 struct fs_path *name, void *ctx)
4714 struct send_ctx *sctx = ctx;
4715 struct rb_node *node = NULL;
4716 struct recorded_ref data;
4717 struct recorded_ref *ref;
4720 ret = get_inode_gen(sctx->send_root, dir, &dir_gen);
4725 data.dir_gen = dir_gen;
4726 set_ref_path(&data, name);
4727 node = rb_find(&data, &sctx->rbtree_deleted_refs, rbtree_ref_comp);
4729 ref = rb_entry(node, struct recorded_ref, node);
4730 recorded_ref_free(ref);
4732 ret = record_ref_in_tree(&sctx->rbtree_new_refs,
4733 &sctx->new_refs, name, dir, dir_gen,
4740 static int record_deleted_ref_if_needed(int num, u64 dir, int index,
4741 struct fs_path *name, void *ctx)
4744 struct send_ctx *sctx = ctx;
4745 struct rb_node *node = NULL;
4746 struct recorded_ref data;
4747 struct recorded_ref *ref;
4750 ret = get_inode_gen(sctx->parent_root, dir, &dir_gen);
4755 data.dir_gen = dir_gen;
4756 set_ref_path(&data, name);
4757 node = rb_find(&data, &sctx->rbtree_new_refs, rbtree_ref_comp);
4759 ref = rb_entry(node, struct recorded_ref, node);
4760 recorded_ref_free(ref);
4762 ret = record_ref_in_tree(&sctx->rbtree_deleted_refs,
4763 &sctx->deleted_refs, name, dir,
4770 static int record_new_ref(struct send_ctx *sctx)
4774 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4775 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4784 static int record_deleted_ref(struct send_ctx *sctx)
4788 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4789 sctx->cmp_key, 0, record_deleted_ref_if_needed,
4799 static int record_changed_ref(struct send_ctx *sctx)
4803 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4804 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4807 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4808 sctx->cmp_key, 0, record_deleted_ref_if_needed, sctx);
4818 * Record and process all refs at once. Needed when an inode changes the
4819 * generation number, which means that it was deleted and recreated.
4821 static int process_all_refs(struct send_ctx *sctx,
4822 enum btrfs_compare_tree_result cmd)
4826 struct btrfs_root *root;
4827 struct btrfs_path *path;
4828 struct btrfs_key key;
4829 struct btrfs_key found_key;
4830 iterate_inode_ref_t cb;
4831 int pending_move = 0;
4833 path = alloc_path_for_send();
4837 if (cmd == BTRFS_COMPARE_TREE_NEW) {
4838 root = sctx->send_root;
4839 cb = record_new_ref_if_needed;
4840 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4841 root = sctx->parent_root;
4842 cb = record_deleted_ref_if_needed;
4844 btrfs_err(sctx->send_root->fs_info,
4845 "Wrong command %d in process_all_refs", cmd);
4850 key.objectid = sctx->cmp_key->objectid;
4851 key.type = BTRFS_INODE_REF_KEY;
4853 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
4854 if (found_key.objectid != key.objectid ||
4855 (found_key.type != BTRFS_INODE_REF_KEY &&
4856 found_key.type != BTRFS_INODE_EXTREF_KEY))
4859 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4863 /* Catch error found during iteration */
4868 btrfs_release_path(path);
4871 * We don't actually care about pending_move as we are simply
4872 * re-creating this inode and will be rename'ing it into place once we
4873 * rename the parent directory.
4875 ret = process_recorded_refs(sctx, &pending_move);
4877 btrfs_free_path(path);
4881 static int send_set_xattr(struct send_ctx *sctx,
4882 struct fs_path *path,
4883 const char *name, int name_len,
4884 const char *data, int data_len)
4888 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4892 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4893 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4894 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4896 ret = send_cmd(sctx);
4903 static int send_remove_xattr(struct send_ctx *sctx,
4904 struct fs_path *path,
4905 const char *name, int name_len)
4909 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4913 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4914 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4916 ret = send_cmd(sctx);
4923 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4924 const char *name, int name_len, const char *data,
4925 int data_len, void *ctx)
4928 struct send_ctx *sctx = ctx;
4930 struct posix_acl_xattr_header dummy_acl;
4932 /* Capabilities are emitted by finish_inode_if_needed */
4933 if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4936 p = fs_path_alloc();
4941 * This hack is needed because empty acls are stored as zero byte
4942 * data in xattrs. Problem with that is, that receiving these zero byte
4943 * acls will fail later. To fix this, we send a dummy acl list that
4944 * only contains the version number and no entries.
4946 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4947 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4948 if (data_len == 0) {
4949 dummy_acl.a_version =
4950 cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4951 data = (char *)&dummy_acl;
4952 data_len = sizeof(dummy_acl);
4956 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4960 ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4967 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4968 const char *name, int name_len,
4969 const char *data, int data_len, void *ctx)
4972 struct send_ctx *sctx = ctx;
4975 p = fs_path_alloc();
4979 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4983 ret = send_remove_xattr(sctx, p, name, name_len);
4990 static int process_new_xattr(struct send_ctx *sctx)
4994 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4995 __process_new_xattr, sctx);
5000 static int process_deleted_xattr(struct send_ctx *sctx)
5002 return iterate_dir_item(sctx->parent_root, sctx->right_path,
5003 __process_deleted_xattr, sctx);
5006 struct find_xattr_ctx {
5014 static int __find_xattr(int num, struct btrfs_key *di_key, const char *name,
5015 int name_len, const char *data, int data_len, void *vctx)
5017 struct find_xattr_ctx *ctx = vctx;
5019 if (name_len == ctx->name_len &&
5020 strncmp(name, ctx->name, name_len) == 0) {
5021 ctx->found_idx = num;
5022 ctx->found_data_len = data_len;
5023 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
5024 if (!ctx->found_data)
5031 static int find_xattr(struct btrfs_root *root,
5032 struct btrfs_path *path,
5033 struct btrfs_key *key,
5034 const char *name, int name_len,
5035 char **data, int *data_len)
5038 struct find_xattr_ctx ctx;
5041 ctx.name_len = name_len;
5043 ctx.found_data = NULL;
5044 ctx.found_data_len = 0;
5046 ret = iterate_dir_item(root, path, __find_xattr, &ctx);
5050 if (ctx.found_idx == -1)
5053 *data = ctx.found_data;
5054 *data_len = ctx.found_data_len;
5056 kfree(ctx.found_data);
5058 return ctx.found_idx;
5062 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
5063 const char *name, int name_len,
5064 const char *data, int data_len,
5068 struct send_ctx *sctx = ctx;
5069 char *found_data = NULL;
5070 int found_data_len = 0;
5072 ret = find_xattr(sctx->parent_root, sctx->right_path,
5073 sctx->cmp_key, name, name_len, &found_data,
5075 if (ret == -ENOENT) {
5076 ret = __process_new_xattr(num, di_key, name, name_len, data,
5078 } else if (ret >= 0) {
5079 if (data_len != found_data_len ||
5080 memcmp(data, found_data, data_len)) {
5081 ret = __process_new_xattr(num, di_key, name, name_len,
5082 data, data_len, ctx);
5092 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
5093 const char *name, int name_len,
5094 const char *data, int data_len,
5098 struct send_ctx *sctx = ctx;
5100 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
5101 name, name_len, NULL, NULL);
5103 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
5111 static int process_changed_xattr(struct send_ctx *sctx)
5115 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
5116 __process_changed_new_xattr, sctx);
5119 ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
5120 __process_changed_deleted_xattr, sctx);
5126 static int process_all_new_xattrs(struct send_ctx *sctx)
5130 struct btrfs_root *root;
5131 struct btrfs_path *path;
5132 struct btrfs_key key;
5133 struct btrfs_key found_key;
5135 path = alloc_path_for_send();
5139 root = sctx->send_root;
5141 key.objectid = sctx->cmp_key->objectid;
5142 key.type = BTRFS_XATTR_ITEM_KEY;
5144 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
5145 if (found_key.objectid != key.objectid ||
5146 found_key.type != key.type) {
5151 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
5155 /* Catch error found during iteration */
5159 btrfs_free_path(path);
5163 static int send_verity(struct send_ctx *sctx, struct fs_path *path,
5164 struct fsverity_descriptor *desc)
5168 ret = begin_cmd(sctx, BTRFS_SEND_C_ENABLE_VERITY);
5172 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
5173 TLV_PUT_U8(sctx, BTRFS_SEND_A_VERITY_ALGORITHM,
5174 le8_to_cpu(desc->hash_algorithm));
5175 TLV_PUT_U32(sctx, BTRFS_SEND_A_VERITY_BLOCK_SIZE,
5176 1U << le8_to_cpu(desc->log_blocksize));
5177 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SALT_DATA, desc->salt,
5178 le8_to_cpu(desc->salt_size));
5179 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SIG_DATA, desc->signature,
5180 le32_to_cpu(desc->sig_size));
5182 ret = send_cmd(sctx);
5189 static int process_verity(struct send_ctx *sctx)
5192 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5193 struct inode *inode;
5196 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, sctx->send_root);
5198 return PTR_ERR(inode);
5200 ret = btrfs_get_verity_descriptor(inode, NULL, 0);
5204 if (ret > FS_VERITY_MAX_DESCRIPTOR_SIZE) {
5208 if (!sctx->verity_descriptor) {
5209 sctx->verity_descriptor = kvmalloc(FS_VERITY_MAX_DESCRIPTOR_SIZE,
5211 if (!sctx->verity_descriptor) {
5217 ret = btrfs_get_verity_descriptor(inode, sctx->verity_descriptor, ret);
5221 p = fs_path_alloc();
5226 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5230 ret = send_verity(sctx, p, sctx->verity_descriptor);
5241 static inline u64 max_send_read_size(const struct send_ctx *sctx)
5243 return sctx->send_max_size - SZ_16K;
5246 static int put_data_header(struct send_ctx *sctx, u32 len)
5248 if (WARN_ON_ONCE(sctx->put_data))
5250 sctx->put_data = true;
5251 if (sctx->proto >= 2) {
5253 * Since v2, the data attribute header doesn't include a length,
5254 * it is implicitly to the end of the command.
5256 if (sctx->send_max_size - sctx->send_size < sizeof(__le16) + len)
5258 put_unaligned_le16(BTRFS_SEND_A_DATA, sctx->send_buf + sctx->send_size);
5259 sctx->send_size += sizeof(__le16);
5261 struct btrfs_tlv_header *hdr;
5263 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
5265 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
5266 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
5267 put_unaligned_le16(len, &hdr->tlv_len);
5268 sctx->send_size += sizeof(*hdr);
5273 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
5275 struct btrfs_root *root = sctx->send_root;
5276 struct btrfs_fs_info *fs_info = root->fs_info;
5278 pgoff_t index = offset >> PAGE_SHIFT;
5280 unsigned pg_offset = offset_in_page(offset);
5283 ret = put_data_header(sctx, len);
5287 last_index = (offset + len - 1) >> PAGE_SHIFT;
5289 while (index <= last_index) {
5290 unsigned cur_len = min_t(unsigned, len,
5291 PAGE_SIZE - pg_offset);
5293 page = find_lock_page(sctx->cur_inode->i_mapping, index);
5295 page_cache_sync_readahead(sctx->cur_inode->i_mapping,
5296 &sctx->ra, NULL, index,
5297 last_index + 1 - index);
5299 page = find_or_create_page(sctx->cur_inode->i_mapping,
5307 if (PageReadahead(page))
5308 page_cache_async_readahead(sctx->cur_inode->i_mapping,
5309 &sctx->ra, NULL, page_folio(page),
5310 index, last_index + 1 - index);
5312 if (!PageUptodate(page)) {
5313 btrfs_read_folio(NULL, page_folio(page));
5315 if (!PageUptodate(page)) {
5318 "send: IO error at offset %llu for inode %llu root %llu",
5319 page_offset(page), sctx->cur_ino,
5320 sctx->send_root->root_key.objectid);
5327 memcpy_from_page(sctx->send_buf + sctx->send_size, page,
5328 pg_offset, cur_len);
5334 sctx->send_size += cur_len;
5341 * Read some bytes from the current inode/file and send a write command to
5344 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5346 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5350 p = fs_path_alloc();
5354 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5356 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5360 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5364 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5365 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5366 ret = put_file_data(sctx, offset, len);
5370 ret = send_cmd(sctx);
5379 * Send a clone command to user space.
5381 static int send_clone(struct send_ctx *sctx,
5382 u64 offset, u32 len,
5383 struct clone_root *clone_root)
5389 btrfs_debug(sctx->send_root->fs_info,
5390 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5391 offset, len, clone_root->root->root_key.objectid,
5392 clone_root->ino, clone_root->offset);
5394 p = fs_path_alloc();
5398 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5402 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5406 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5407 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5408 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5410 if (clone_root->root == sctx->send_root) {
5411 ret = get_inode_gen(sctx->send_root, clone_root->ino, &gen);
5414 ret = get_cur_path(sctx, clone_root->ino, gen, p);
5416 ret = get_inode_path(clone_root->root, clone_root->ino, p);
5422 * If the parent we're using has a received_uuid set then use that as
5423 * our clone source as that is what we will look for when doing a
5426 * This covers the case that we create a snapshot off of a received
5427 * subvolume and then use that as the parent and try to receive on a
5430 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5431 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5432 clone_root->root->root_item.received_uuid);
5434 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5435 clone_root->root->root_item.uuid);
5436 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5437 btrfs_root_ctransid(&clone_root->root->root_item));
5438 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5439 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5440 clone_root->offset);
5442 ret = send_cmd(sctx);
5451 * Send an update extent command to user space.
5453 static int send_update_extent(struct send_ctx *sctx,
5454 u64 offset, u32 len)
5459 p = fs_path_alloc();
5463 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5467 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5471 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5472 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5473 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5475 ret = send_cmd(sctx);
5483 static int send_hole(struct send_ctx *sctx, u64 end)
5485 struct fs_path *p = NULL;
5486 u64 read_size = max_send_read_size(sctx);
5487 u64 offset = sctx->cur_inode_last_extent;
5491 * A hole that starts at EOF or beyond it. Since we do not yet support
5492 * fallocate (for extent preallocation and hole punching), sending a
5493 * write of zeroes starting at EOF or beyond would later require issuing
5494 * a truncate operation which would undo the write and achieve nothing.
5496 if (offset >= sctx->cur_inode_size)
5500 * Don't go beyond the inode's i_size due to prealloc extents that start
5503 end = min_t(u64, end, sctx->cur_inode_size);
5505 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5506 return send_update_extent(sctx, offset, end - offset);
5508 p = fs_path_alloc();
5511 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5513 goto tlv_put_failure;
5514 while (offset < end) {
5515 u64 len = min(end - offset, read_size);
5517 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5520 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5521 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5522 ret = put_data_header(sctx, len);
5525 memset(sctx->send_buf + sctx->send_size, 0, len);
5526 sctx->send_size += len;
5527 ret = send_cmd(sctx);
5532 sctx->cur_inode_next_write_offset = offset;
5538 static int send_encoded_inline_extent(struct send_ctx *sctx,
5539 struct btrfs_path *path, u64 offset,
5542 struct btrfs_root *root = sctx->send_root;
5543 struct btrfs_fs_info *fs_info = root->fs_info;
5544 struct inode *inode;
5545 struct fs_path *fspath;
5546 struct extent_buffer *leaf = path->nodes[0];
5547 struct btrfs_key key;
5548 struct btrfs_file_extent_item *ei;
5553 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5555 return PTR_ERR(inode);
5557 fspath = fs_path_alloc();
5563 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5567 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5571 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5572 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5573 ram_bytes = btrfs_file_extent_ram_bytes(leaf, ei);
5574 inline_size = btrfs_file_extent_inline_item_len(leaf, path->slots[0]);
5576 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5577 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5578 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5579 min(key.offset + ram_bytes - offset, len));
5580 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN, ram_bytes);
5581 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET, offset - key.offset);
5582 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5583 btrfs_file_extent_compression(leaf, ei));
5586 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5588 ret = put_data_header(sctx, inline_size);
5591 read_extent_buffer(leaf, sctx->send_buf + sctx->send_size,
5592 btrfs_file_extent_inline_start(ei), inline_size);
5593 sctx->send_size += inline_size;
5595 ret = send_cmd(sctx);
5599 fs_path_free(fspath);
5604 static int send_encoded_extent(struct send_ctx *sctx, struct btrfs_path *path,
5605 u64 offset, u64 len)
5607 struct btrfs_root *root = sctx->send_root;
5608 struct btrfs_fs_info *fs_info = root->fs_info;
5609 struct inode *inode;
5610 struct fs_path *fspath;
5611 struct extent_buffer *leaf = path->nodes[0];
5612 struct btrfs_key key;
5613 struct btrfs_file_extent_item *ei;
5614 u64 disk_bytenr, disk_num_bytes;
5616 struct btrfs_cmd_header *hdr;
5620 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5622 return PTR_ERR(inode);
5624 fspath = fs_path_alloc();
5630 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5634 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5638 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5639 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5640 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
5641 disk_num_bytes = btrfs_file_extent_disk_num_bytes(leaf, ei);
5643 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5644 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5645 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5646 min(key.offset + btrfs_file_extent_num_bytes(leaf, ei) - offset,
5648 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN,
5649 btrfs_file_extent_ram_bytes(leaf, ei));
5650 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET,
5651 offset - key.offset + btrfs_file_extent_offset(leaf, ei));
5652 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5653 btrfs_file_extent_compression(leaf, ei));
5656 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5657 TLV_PUT_U32(sctx, BTRFS_SEND_A_ENCRYPTION, 0);
5659 ret = put_data_header(sctx, disk_num_bytes);
5664 * We want to do I/O directly into the send buffer, so get the next page
5665 * boundary in the send buffer. This means that there may be a gap
5666 * between the beginning of the command and the file data.
5668 data_offset = PAGE_ALIGN(sctx->send_size);
5669 if (data_offset > sctx->send_max_size ||
5670 sctx->send_max_size - data_offset < disk_num_bytes) {
5676 * Note that send_buf is a mapping of send_buf_pages, so this is really
5677 * reading into send_buf.
5679 ret = btrfs_encoded_read_regular_fill_pages(BTRFS_I(inode), offset,
5680 disk_bytenr, disk_num_bytes,
5681 sctx->send_buf_pages +
5682 (data_offset >> PAGE_SHIFT));
5686 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
5687 hdr->len = cpu_to_le32(sctx->send_size + disk_num_bytes - sizeof(*hdr));
5689 crc = crc32c(0, sctx->send_buf, sctx->send_size);
5690 crc = crc32c(crc, sctx->send_buf + data_offset, disk_num_bytes);
5691 hdr->crc = cpu_to_le32(crc);
5693 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
5696 ret = write_buf(sctx->send_filp, sctx->send_buf + data_offset,
5697 disk_num_bytes, &sctx->send_off);
5699 sctx->send_size = 0;
5700 sctx->put_data = false;
5704 fs_path_free(fspath);
5709 static int send_extent_data(struct send_ctx *sctx, struct btrfs_path *path,
5710 const u64 offset, const u64 len)
5712 const u64 end = offset + len;
5713 struct extent_buffer *leaf = path->nodes[0];
5714 struct btrfs_file_extent_item *ei;
5715 u64 read_size = max_send_read_size(sctx);
5718 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5719 return send_update_extent(sctx, offset, len);
5721 ei = btrfs_item_ptr(leaf, path->slots[0],
5722 struct btrfs_file_extent_item);
5723 if ((sctx->flags & BTRFS_SEND_FLAG_COMPRESSED) &&
5724 btrfs_file_extent_compression(leaf, ei) != BTRFS_COMPRESS_NONE) {
5725 bool is_inline = (btrfs_file_extent_type(leaf, ei) ==
5726 BTRFS_FILE_EXTENT_INLINE);
5729 * Send the compressed extent unless the compressed data is
5730 * larger than the decompressed data. This can happen if we're
5731 * not sending the entire extent, either because it has been
5732 * partially overwritten/truncated or because this is a part of
5733 * the extent that we couldn't clone in clone_range().
5736 btrfs_file_extent_inline_item_len(leaf,
5737 path->slots[0]) <= len) {
5738 return send_encoded_inline_extent(sctx, path, offset,
5740 } else if (!is_inline &&
5741 btrfs_file_extent_disk_num_bytes(leaf, ei) <= len) {
5742 return send_encoded_extent(sctx, path, offset, len);
5746 if (sctx->cur_inode == NULL) {
5747 struct btrfs_root *root = sctx->send_root;
5749 sctx->cur_inode = btrfs_iget(root->fs_info->sb, sctx->cur_ino, root);
5750 if (IS_ERR(sctx->cur_inode)) {
5751 int err = PTR_ERR(sctx->cur_inode);
5753 sctx->cur_inode = NULL;
5756 memset(&sctx->ra, 0, sizeof(struct file_ra_state));
5757 file_ra_state_init(&sctx->ra, sctx->cur_inode->i_mapping);
5760 * It's very likely there are no pages from this inode in the page
5761 * cache, so after reading extents and sending their data, we clean
5762 * the page cache to avoid trashing the page cache (adding pressure
5763 * to the page cache and forcing eviction of other data more useful
5764 * for applications).
5766 * We decide if we should clean the page cache simply by checking
5767 * if the inode's mapping nrpages is 0 when we first open it, and
5768 * not by using something like filemap_range_has_page() before
5769 * reading an extent because when we ask the readahead code to
5770 * read a given file range, it may (and almost always does) read
5771 * pages from beyond that range (see the documentation for
5772 * page_cache_sync_readahead()), so it would not be reliable,
5773 * because after reading the first extent future calls to
5774 * filemap_range_has_page() would return true because the readahead
5775 * on the previous extent resulted in reading pages of the current
5778 sctx->clean_page_cache = (sctx->cur_inode->i_mapping->nrpages == 0);
5779 sctx->page_cache_clear_start = round_down(offset, PAGE_SIZE);
5782 while (sent < len) {
5783 u64 size = min(len - sent, read_size);
5786 ret = send_write(sctx, offset + sent, size);
5792 if (sctx->clean_page_cache && PAGE_ALIGNED(end)) {
5794 * Always operate only on ranges that are a multiple of the page
5795 * size. This is not only to prevent zeroing parts of a page in
5796 * the case of subpage sector size, but also to guarantee we evict
5797 * pages, as passing a range that is smaller than page size does
5798 * not evict the respective page (only zeroes part of its content).
5800 * Always start from the end offset of the last range cleared.
5801 * This is because the readahead code may (and very often does)
5802 * reads pages beyond the range we request for readahead. So if
5803 * we have an extent layout like this:
5805 * [ extent A ] [ extent B ] [ extent C ]
5807 * When we ask page_cache_sync_readahead() to read extent A, it
5808 * may also trigger reads for pages of extent B. If we are doing
5809 * an incremental send and extent B has not changed between the
5810 * parent and send snapshots, some or all of its pages may end
5811 * up being read and placed in the page cache. So when truncating
5812 * the page cache we always start from the end offset of the
5813 * previously processed extent up to the end of the current
5816 truncate_inode_pages_range(&sctx->cur_inode->i_data,
5817 sctx->page_cache_clear_start,
5819 sctx->page_cache_clear_start = end;
5826 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5827 * found, call send_set_xattr function to emit it.
5829 * Return 0 if there isn't a capability, or when the capability was emitted
5830 * successfully, or < 0 if an error occurred.
5832 static int send_capabilities(struct send_ctx *sctx)
5834 struct fs_path *fspath = NULL;
5835 struct btrfs_path *path;
5836 struct btrfs_dir_item *di;
5837 struct extent_buffer *leaf;
5838 unsigned long data_ptr;
5843 path = alloc_path_for_send();
5847 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5848 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5850 /* There is no xattr for this inode */
5852 } else if (IS_ERR(di)) {
5857 leaf = path->nodes[0];
5858 buf_len = btrfs_dir_data_len(leaf, di);
5860 fspath = fs_path_alloc();
5861 buf = kmalloc(buf_len, GFP_KERNEL);
5862 if (!fspath || !buf) {
5867 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5871 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5872 read_extent_buffer(leaf, buf, data_ptr, buf_len);
5874 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5875 strlen(XATTR_NAME_CAPS), buf, buf_len);
5878 fs_path_free(fspath);
5879 btrfs_free_path(path);
5883 static int clone_range(struct send_ctx *sctx, struct btrfs_path *dst_path,
5884 struct clone_root *clone_root, const u64 disk_byte,
5885 u64 data_offset, u64 offset, u64 len)
5887 struct btrfs_path *path;
5888 struct btrfs_key key;
5890 struct btrfs_inode_info info;
5891 u64 clone_src_i_size = 0;
5894 * Prevent cloning from a zero offset with a length matching the sector
5895 * size because in some scenarios this will make the receiver fail.
5897 * For example, if in the source filesystem the extent at offset 0
5898 * has a length of sectorsize and it was written using direct IO, then
5899 * it can never be an inline extent (even if compression is enabled).
5900 * Then this extent can be cloned in the original filesystem to a non
5901 * zero file offset, but it may not be possible to clone in the
5902 * destination filesystem because it can be inlined due to compression
5903 * on the destination filesystem (as the receiver's write operations are
5904 * always done using buffered IO). The same happens when the original
5905 * filesystem does not have compression enabled but the destination
5908 if (clone_root->offset == 0 &&
5909 len == sctx->send_root->fs_info->sectorsize)
5910 return send_extent_data(sctx, dst_path, offset, len);
5912 path = alloc_path_for_send();
5917 * There are inodes that have extents that lie behind its i_size. Don't
5918 * accept clones from these extents.
5920 ret = get_inode_info(clone_root->root, clone_root->ino, &info);
5921 btrfs_release_path(path);
5924 clone_src_i_size = info.size;
5927 * We can't send a clone operation for the entire range if we find
5928 * extent items in the respective range in the source file that
5929 * refer to different extents or if we find holes.
5930 * So check for that and do a mix of clone and regular write/copy
5931 * operations if needed.
5935 * mkfs.btrfs -f /dev/sda
5936 * mount /dev/sda /mnt
5937 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5938 * cp --reflink=always /mnt/foo /mnt/bar
5939 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5940 * btrfs subvolume snapshot -r /mnt /mnt/snap
5942 * If when we send the snapshot and we are processing file bar (which
5943 * has a higher inode number than foo) we blindly send a clone operation
5944 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5945 * a file bar that matches the content of file foo - iow, doesn't match
5946 * the content from bar in the original filesystem.
5948 key.objectid = clone_root->ino;
5949 key.type = BTRFS_EXTENT_DATA_KEY;
5950 key.offset = clone_root->offset;
5951 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5954 if (ret > 0 && path->slots[0] > 0) {
5955 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5956 if (key.objectid == clone_root->ino &&
5957 key.type == BTRFS_EXTENT_DATA_KEY)
5962 struct extent_buffer *leaf = path->nodes[0];
5963 int slot = path->slots[0];
5964 struct btrfs_file_extent_item *ei;
5968 u64 clone_data_offset;
5969 bool crossed_src_i_size = false;
5971 if (slot >= btrfs_header_nritems(leaf)) {
5972 ret = btrfs_next_leaf(clone_root->root, path);
5980 btrfs_item_key_to_cpu(leaf, &key, slot);
5983 * We might have an implicit trailing hole (NO_HOLES feature
5984 * enabled). We deal with it after leaving this loop.
5986 if (key.objectid != clone_root->ino ||
5987 key.type != BTRFS_EXTENT_DATA_KEY)
5990 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5991 type = btrfs_file_extent_type(leaf, ei);
5992 if (type == BTRFS_FILE_EXTENT_INLINE) {
5993 ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5994 ext_len = PAGE_ALIGN(ext_len);
5996 ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5999 if (key.offset + ext_len <= clone_root->offset)
6002 if (key.offset > clone_root->offset) {
6003 /* Implicit hole, NO_HOLES feature enabled. */
6004 u64 hole_len = key.offset - clone_root->offset;
6008 ret = send_extent_data(sctx, dst_path, offset,
6017 clone_root->offset += hole_len;
6018 data_offset += hole_len;
6021 if (key.offset >= clone_root->offset + len)
6024 if (key.offset >= clone_src_i_size)
6027 if (key.offset + ext_len > clone_src_i_size) {
6028 ext_len = clone_src_i_size - key.offset;
6029 crossed_src_i_size = true;
6032 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
6033 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
6034 clone_root->offset = key.offset;
6035 if (clone_data_offset < data_offset &&
6036 clone_data_offset + ext_len > data_offset) {
6039 extent_offset = data_offset - clone_data_offset;
6040 ext_len -= extent_offset;
6041 clone_data_offset += extent_offset;
6042 clone_root->offset += extent_offset;
6046 clone_len = min_t(u64, ext_len, len);
6048 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
6049 clone_data_offset == data_offset) {
6050 const u64 src_end = clone_root->offset + clone_len;
6051 const u64 sectorsize = SZ_64K;
6054 * We can't clone the last block, when its size is not
6055 * sector size aligned, into the middle of a file. If we
6056 * do so, the receiver will get a failure (-EINVAL) when
6057 * trying to clone or will silently corrupt the data in
6058 * the destination file if it's on a kernel without the
6059 * fix introduced by commit ac765f83f1397646
6060 * ("Btrfs: fix data corruption due to cloning of eof
6063 * So issue a clone of the aligned down range plus a
6064 * regular write for the eof block, if we hit that case.
6066 * Also, we use the maximum possible sector size, 64K,
6067 * because we don't know what's the sector size of the
6068 * filesystem that receives the stream, so we have to
6069 * assume the largest possible sector size.
6071 if (src_end == clone_src_i_size &&
6072 !IS_ALIGNED(src_end, sectorsize) &&
6073 offset + clone_len < sctx->cur_inode_size) {
6076 slen = ALIGN_DOWN(src_end - clone_root->offset,
6079 ret = send_clone(sctx, offset, slen,
6084 ret = send_extent_data(sctx, dst_path,
6088 ret = send_clone(sctx, offset, clone_len,
6091 } else if (crossed_src_i_size && clone_len < len) {
6093 * If we are at i_size of the clone source inode and we
6094 * can not clone from it, terminate the loop. This is
6095 * to avoid sending two write operations, one with a
6096 * length matching clone_len and the final one after
6097 * this loop with a length of len - clone_len.
6099 * When using encoded writes (BTRFS_SEND_FLAG_COMPRESSED
6100 * was passed to the send ioctl), this helps avoid
6101 * sending an encoded write for an offset that is not
6102 * sector size aligned, in case the i_size of the source
6103 * inode is not sector size aligned. That will make the
6104 * receiver fallback to decompression of the data and
6105 * writing it using regular buffered IO, therefore while
6106 * not incorrect, it's not optimal due decompression and
6107 * possible re-compression at the receiver.
6111 ret = send_extent_data(sctx, dst_path, offset,
6121 offset += clone_len;
6122 clone_root->offset += clone_len;
6125 * If we are cloning from the file we are currently processing,
6126 * and using the send root as the clone root, we must stop once
6127 * the current clone offset reaches the current eof of the file
6128 * at the receiver, otherwise we would issue an invalid clone
6129 * operation (source range going beyond eof) and cause the
6130 * receiver to fail. So if we reach the current eof, bail out
6131 * and fallback to a regular write.
6133 if (clone_root->root == sctx->send_root &&
6134 clone_root->ino == sctx->cur_ino &&
6135 clone_root->offset >= sctx->cur_inode_next_write_offset)
6138 data_offset += clone_len;
6144 ret = send_extent_data(sctx, dst_path, offset, len);
6148 btrfs_free_path(path);
6152 static int send_write_or_clone(struct send_ctx *sctx,
6153 struct btrfs_path *path,
6154 struct btrfs_key *key,
6155 struct clone_root *clone_root)
6158 u64 offset = key->offset;
6160 u64 bs = sctx->send_root->fs_info->sectorsize;
6162 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
6166 if (clone_root && IS_ALIGNED(end, bs)) {
6167 struct btrfs_file_extent_item *ei;
6171 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6172 struct btrfs_file_extent_item);
6173 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
6174 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
6175 ret = clone_range(sctx, path, clone_root, disk_byte,
6176 data_offset, offset, end - offset);
6178 ret = send_extent_data(sctx, path, offset, end - offset);
6180 sctx->cur_inode_next_write_offset = end;
6184 static int is_extent_unchanged(struct send_ctx *sctx,
6185 struct btrfs_path *left_path,
6186 struct btrfs_key *ekey)
6189 struct btrfs_key key;
6190 struct btrfs_path *path = NULL;
6191 struct extent_buffer *eb;
6193 struct btrfs_key found_key;
6194 struct btrfs_file_extent_item *ei;
6199 u64 left_offset_fixed;
6207 path = alloc_path_for_send();
6211 eb = left_path->nodes[0];
6212 slot = left_path->slots[0];
6213 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6214 left_type = btrfs_file_extent_type(eb, ei);
6216 if (left_type != BTRFS_FILE_EXTENT_REG) {
6220 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6221 left_len = btrfs_file_extent_num_bytes(eb, ei);
6222 left_offset = btrfs_file_extent_offset(eb, ei);
6223 left_gen = btrfs_file_extent_generation(eb, ei);
6226 * Following comments will refer to these graphics. L is the left
6227 * extents which we are checking at the moment. 1-8 are the right
6228 * extents that we iterate.
6231 * |-1-|-2a-|-3-|-4-|-5-|-6-|
6234 * |--1--|-2b-|...(same as above)
6236 * Alternative situation. Happens on files where extents got split.
6238 * |-----------7-----------|-6-|
6240 * Alternative situation. Happens on files which got larger.
6243 * Nothing follows after 8.
6246 key.objectid = ekey->objectid;
6247 key.type = BTRFS_EXTENT_DATA_KEY;
6248 key.offset = ekey->offset;
6249 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
6258 * Handle special case where the right side has no extents at all.
6260 eb = path->nodes[0];
6261 slot = path->slots[0];
6262 btrfs_item_key_to_cpu(eb, &found_key, slot);
6263 if (found_key.objectid != key.objectid ||
6264 found_key.type != key.type) {
6265 /* If we're a hole then just pretend nothing changed */
6266 ret = (left_disknr) ? 0 : 1;
6271 * We're now on 2a, 2b or 7.
6274 while (key.offset < ekey->offset + left_len) {
6275 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6276 right_type = btrfs_file_extent_type(eb, ei);
6277 if (right_type != BTRFS_FILE_EXTENT_REG &&
6278 right_type != BTRFS_FILE_EXTENT_INLINE) {
6283 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6284 right_len = btrfs_file_extent_ram_bytes(eb, ei);
6285 right_len = PAGE_ALIGN(right_len);
6287 right_len = btrfs_file_extent_num_bytes(eb, ei);
6291 * Are we at extent 8? If yes, we know the extent is changed.
6292 * This may only happen on the first iteration.
6294 if (found_key.offset + right_len <= ekey->offset) {
6295 /* If we're a hole just pretend nothing changed */
6296 ret = (left_disknr) ? 0 : 1;
6301 * We just wanted to see if when we have an inline extent, what
6302 * follows it is a regular extent (wanted to check the above
6303 * condition for inline extents too). This should normally not
6304 * happen but it's possible for example when we have an inline
6305 * compressed extent representing data with a size matching
6306 * the page size (currently the same as sector size).
6308 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6313 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6314 right_offset = btrfs_file_extent_offset(eb, ei);
6315 right_gen = btrfs_file_extent_generation(eb, ei);
6317 left_offset_fixed = left_offset;
6318 if (key.offset < ekey->offset) {
6319 /* Fix the right offset for 2a and 7. */
6320 right_offset += ekey->offset - key.offset;
6322 /* Fix the left offset for all behind 2a and 2b */
6323 left_offset_fixed += key.offset - ekey->offset;
6327 * Check if we have the same extent.
6329 if (left_disknr != right_disknr ||
6330 left_offset_fixed != right_offset ||
6331 left_gen != right_gen) {
6337 * Go to the next extent.
6339 ret = btrfs_next_item(sctx->parent_root, path);
6343 eb = path->nodes[0];
6344 slot = path->slots[0];
6345 btrfs_item_key_to_cpu(eb, &found_key, slot);
6347 if (ret || found_key.objectid != key.objectid ||
6348 found_key.type != key.type) {
6349 key.offset += right_len;
6352 if (found_key.offset != key.offset + right_len) {
6360 * We're now behind the left extent (treat as unchanged) or at the end
6361 * of the right side (treat as changed).
6363 if (key.offset >= ekey->offset + left_len)
6370 btrfs_free_path(path);
6374 static int get_last_extent(struct send_ctx *sctx, u64 offset)
6376 struct btrfs_path *path;
6377 struct btrfs_root *root = sctx->send_root;
6378 struct btrfs_key key;
6381 path = alloc_path_for_send();
6385 sctx->cur_inode_last_extent = 0;
6387 key.objectid = sctx->cur_ino;
6388 key.type = BTRFS_EXTENT_DATA_KEY;
6389 key.offset = offset;
6390 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
6394 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
6395 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
6398 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6400 btrfs_free_path(path);
6404 static int range_is_hole_in_parent(struct send_ctx *sctx,
6408 struct btrfs_path *path;
6409 struct btrfs_key key;
6410 struct btrfs_root *root = sctx->parent_root;
6411 u64 search_start = start;
6414 path = alloc_path_for_send();
6418 key.objectid = sctx->cur_ino;
6419 key.type = BTRFS_EXTENT_DATA_KEY;
6420 key.offset = search_start;
6421 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6424 if (ret > 0 && path->slots[0] > 0)
6427 while (search_start < end) {
6428 struct extent_buffer *leaf = path->nodes[0];
6429 int slot = path->slots[0];
6430 struct btrfs_file_extent_item *fi;
6433 if (slot >= btrfs_header_nritems(leaf)) {
6434 ret = btrfs_next_leaf(root, path);
6442 btrfs_item_key_to_cpu(leaf, &key, slot);
6443 if (key.objectid < sctx->cur_ino ||
6444 key.type < BTRFS_EXTENT_DATA_KEY)
6446 if (key.objectid > sctx->cur_ino ||
6447 key.type > BTRFS_EXTENT_DATA_KEY ||
6451 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6452 extent_end = btrfs_file_extent_end(path);
6453 if (extent_end <= start)
6455 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
6456 search_start = extent_end;
6466 btrfs_free_path(path);
6470 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
6471 struct btrfs_key *key)
6475 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
6479 * Get last extent's end offset (exclusive) if we haven't determined it
6480 * yet (we're processing the first file extent item that is new), or if
6481 * we're at the first slot of a leaf and the last extent's end is less
6482 * than the current extent's offset, because we might have skipped
6483 * entire leaves that contained only file extent items for our current
6484 * inode. These leaves have a generation number smaller (older) than the
6485 * one in the current leaf and the leaf our last extent came from, and
6486 * are located between these 2 leaves.
6488 if ((sctx->cur_inode_last_extent == (u64)-1) ||
6489 (path->slots[0] == 0 && sctx->cur_inode_last_extent < key->offset)) {
6490 ret = get_last_extent(sctx, key->offset - 1);
6495 if (sctx->cur_inode_last_extent < key->offset) {
6496 ret = range_is_hole_in_parent(sctx,
6497 sctx->cur_inode_last_extent,
6502 ret = send_hole(sctx, key->offset);
6506 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6510 static int process_extent(struct send_ctx *sctx,
6511 struct btrfs_path *path,
6512 struct btrfs_key *key)
6514 struct clone_root *found_clone = NULL;
6517 if (S_ISLNK(sctx->cur_inode_mode))
6520 if (sctx->parent_root && !sctx->cur_inode_new) {
6521 ret = is_extent_unchanged(sctx, path, key);
6529 struct btrfs_file_extent_item *ei;
6532 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6533 struct btrfs_file_extent_item);
6534 type = btrfs_file_extent_type(path->nodes[0], ei);
6535 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
6536 type == BTRFS_FILE_EXTENT_REG) {
6538 * The send spec does not have a prealloc command yet,
6539 * so just leave a hole for prealloc'ed extents until
6540 * we have enough commands queued up to justify rev'ing
6543 if (type == BTRFS_FILE_EXTENT_PREALLOC) {
6548 /* Have a hole, just skip it. */
6549 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
6556 ret = find_extent_clone(sctx, path, key->objectid, key->offset,
6557 sctx->cur_inode_size, &found_clone);
6558 if (ret != -ENOENT && ret < 0)
6561 ret = send_write_or_clone(sctx, path, key, found_clone);
6565 ret = maybe_send_hole(sctx, path, key);
6570 static int process_all_extents(struct send_ctx *sctx)
6574 struct btrfs_root *root;
6575 struct btrfs_path *path;
6576 struct btrfs_key key;
6577 struct btrfs_key found_key;
6579 root = sctx->send_root;
6580 path = alloc_path_for_send();
6584 key.objectid = sctx->cmp_key->objectid;
6585 key.type = BTRFS_EXTENT_DATA_KEY;
6587 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
6588 if (found_key.objectid != key.objectid ||
6589 found_key.type != key.type) {
6594 ret = process_extent(sctx, path, &found_key);
6598 /* Catch error found during iteration */
6602 btrfs_free_path(path);
6606 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6608 int *refs_processed)
6612 if (sctx->cur_ino == 0)
6614 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6615 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6617 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6620 ret = process_recorded_refs(sctx, pending_move);
6624 *refs_processed = 1;
6629 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6632 struct btrfs_inode_info info;
6643 bool need_fileattr = false;
6644 int need_truncate = 1;
6645 int pending_move = 0;
6646 int refs_processed = 0;
6648 if (sctx->ignore_cur_inode)
6651 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6657 * We have processed the refs and thus need to advance send_progress.
6658 * Now, calls to get_cur_xxx will take the updated refs of the current
6659 * inode into account.
6661 * On the other hand, if our current inode is a directory and couldn't
6662 * be moved/renamed because its parent was renamed/moved too and it has
6663 * a higher inode number, we can only move/rename our current inode
6664 * after we moved/renamed its parent. Therefore in this case operate on
6665 * the old path (pre move/rename) of our current inode, and the
6666 * move/rename will be performed later.
6668 if (refs_processed && !pending_move)
6669 sctx->send_progress = sctx->cur_ino + 1;
6671 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6673 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6675 ret = get_inode_info(sctx->send_root, sctx->cur_ino, &info);
6678 left_mode = info.mode;
6679 left_uid = info.uid;
6680 left_gid = info.gid;
6681 left_fileattr = info.fileattr;
6683 if (!sctx->parent_root || sctx->cur_inode_new) {
6685 if (!S_ISLNK(sctx->cur_inode_mode))
6687 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6692 ret = get_inode_info(sctx->parent_root, sctx->cur_ino, &info);
6695 old_size = info.size;
6696 right_mode = info.mode;
6697 right_uid = info.uid;
6698 right_gid = info.gid;
6699 right_fileattr = info.fileattr;
6701 if (left_uid != right_uid || left_gid != right_gid)
6703 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6705 if (!S_ISLNK(sctx->cur_inode_mode) && left_fileattr != right_fileattr)
6706 need_fileattr = true;
6707 if ((old_size == sctx->cur_inode_size) ||
6708 (sctx->cur_inode_size > old_size &&
6709 sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6713 if (S_ISREG(sctx->cur_inode_mode)) {
6714 if (need_send_hole(sctx)) {
6715 if (sctx->cur_inode_last_extent == (u64)-1 ||
6716 sctx->cur_inode_last_extent <
6717 sctx->cur_inode_size) {
6718 ret = get_last_extent(sctx, (u64)-1);
6722 if (sctx->cur_inode_last_extent < sctx->cur_inode_size) {
6723 ret = range_is_hole_in_parent(sctx,
6724 sctx->cur_inode_last_extent,
6725 sctx->cur_inode_size);
6728 } else if (ret == 0) {
6729 ret = send_hole(sctx, sctx->cur_inode_size);
6733 /* Range is already a hole, skip. */
6738 if (need_truncate) {
6739 ret = send_truncate(sctx, sctx->cur_ino,
6740 sctx->cur_inode_gen,
6741 sctx->cur_inode_size);
6748 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6749 left_uid, left_gid);
6754 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6759 if (need_fileattr) {
6760 ret = send_fileattr(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6766 if (proto_cmd_ok(sctx, BTRFS_SEND_C_ENABLE_VERITY)
6767 && sctx->cur_inode_needs_verity) {
6768 ret = process_verity(sctx);
6773 ret = send_capabilities(sctx);
6778 * If other directory inodes depended on our current directory
6779 * inode's move/rename, now do their move/rename operations.
6781 if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6782 ret = apply_children_dir_moves(sctx);
6786 * Need to send that every time, no matter if it actually
6787 * changed between the two trees as we have done changes to
6788 * the inode before. If our inode is a directory and it's
6789 * waiting to be moved/renamed, we will send its utimes when
6790 * it's moved/renamed, therefore we don't need to do it here.
6792 sctx->send_progress = sctx->cur_ino + 1;
6795 * If the current inode is a non-empty directory, delay issuing
6796 * the utimes command for it, as it's very likely we have inodes
6797 * with an higher number inside it. We want to issue the utimes
6798 * command only after adding all dentries to it.
6800 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_size > 0)
6801 ret = cache_dir_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6803 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6811 ret = trim_dir_utimes_cache(sctx);
6816 static void close_current_inode(struct send_ctx *sctx)
6820 if (sctx->cur_inode == NULL)
6823 i_size = i_size_read(sctx->cur_inode);
6826 * If we are doing an incremental send, we may have extents between the
6827 * last processed extent and the i_size that have not been processed
6828 * because they haven't changed but we may have read some of their pages
6829 * through readahead, see the comments at send_extent_data().
6831 if (sctx->clean_page_cache && sctx->page_cache_clear_start < i_size)
6832 truncate_inode_pages_range(&sctx->cur_inode->i_data,
6833 sctx->page_cache_clear_start,
6834 round_up(i_size, PAGE_SIZE) - 1);
6836 iput(sctx->cur_inode);
6837 sctx->cur_inode = NULL;
6840 static int changed_inode(struct send_ctx *sctx,
6841 enum btrfs_compare_tree_result result)
6844 struct btrfs_key *key = sctx->cmp_key;
6845 struct btrfs_inode_item *left_ii = NULL;
6846 struct btrfs_inode_item *right_ii = NULL;
6850 close_current_inode(sctx);
6852 sctx->cur_ino = key->objectid;
6853 sctx->cur_inode_new_gen = false;
6854 sctx->cur_inode_last_extent = (u64)-1;
6855 sctx->cur_inode_next_write_offset = 0;
6856 sctx->ignore_cur_inode = false;
6859 * Set send_progress to current inode. This will tell all get_cur_xxx
6860 * functions that the current inode's refs are not updated yet. Later,
6861 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6863 sctx->send_progress = sctx->cur_ino;
6865 if (result == BTRFS_COMPARE_TREE_NEW ||
6866 result == BTRFS_COMPARE_TREE_CHANGED) {
6867 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6868 sctx->left_path->slots[0],
6869 struct btrfs_inode_item);
6870 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6873 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6874 sctx->right_path->slots[0],
6875 struct btrfs_inode_item);
6876 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6879 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6880 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6881 sctx->right_path->slots[0],
6882 struct btrfs_inode_item);
6884 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6888 * The cur_ino = root dir case is special here. We can't treat
6889 * the inode as deleted+reused because it would generate a
6890 * stream that tries to delete/mkdir the root dir.
6892 if (left_gen != right_gen &&
6893 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6894 sctx->cur_inode_new_gen = true;
6898 * Normally we do not find inodes with a link count of zero (orphans)
6899 * because the most common case is to create a snapshot and use it
6900 * for a send operation. However other less common use cases involve
6901 * using a subvolume and send it after turning it to RO mode just
6902 * after deleting all hard links of a file while holding an open
6903 * file descriptor against it or turning a RO snapshot into RW mode,
6904 * keep an open file descriptor against a file, delete it and then
6905 * turn the snapshot back to RO mode before using it for a send
6906 * operation. The former is what the receiver operation does.
6907 * Therefore, if we want to send these snapshots soon after they're
6908 * received, we need to handle orphan inodes as well. Moreover, orphans
6909 * can appear not only in the send snapshot but also in the parent
6910 * snapshot. Here are several cases:
6912 * Case 1: BTRFS_COMPARE_TREE_NEW
6913 * | send snapshot | action
6914 * --------------------------------
6915 * nlink | 0 | ignore
6917 * Case 2: BTRFS_COMPARE_TREE_DELETED
6918 * | parent snapshot | action
6919 * ----------------------------------
6920 * nlink | 0 | as usual
6921 * Note: No unlinks will be sent because there're no paths for it.
6923 * Case 3: BTRFS_COMPARE_TREE_CHANGED
6924 * | | parent snapshot | send snapshot | action
6925 * -----------------------------------------------------------------------
6926 * subcase 1 | nlink | 0 | 0 | ignore
6927 * subcase 2 | nlink | >0 | 0 | new_gen(deletion)
6928 * subcase 3 | nlink | 0 | >0 | new_gen(creation)
6931 if (result == BTRFS_COMPARE_TREE_NEW) {
6932 if (btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii) == 0) {
6933 sctx->ignore_cur_inode = true;
6936 sctx->cur_inode_gen = left_gen;
6937 sctx->cur_inode_new = true;
6938 sctx->cur_inode_deleted = false;
6939 sctx->cur_inode_size = btrfs_inode_size(
6940 sctx->left_path->nodes[0], left_ii);
6941 sctx->cur_inode_mode = btrfs_inode_mode(
6942 sctx->left_path->nodes[0], left_ii);
6943 sctx->cur_inode_rdev = btrfs_inode_rdev(
6944 sctx->left_path->nodes[0], left_ii);
6945 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6946 ret = send_create_inode_if_needed(sctx);
6947 } else if (result == BTRFS_COMPARE_TREE_DELETED) {
6948 sctx->cur_inode_gen = right_gen;
6949 sctx->cur_inode_new = false;
6950 sctx->cur_inode_deleted = true;
6951 sctx->cur_inode_size = btrfs_inode_size(
6952 sctx->right_path->nodes[0], right_ii);
6953 sctx->cur_inode_mode = btrfs_inode_mode(
6954 sctx->right_path->nodes[0], right_ii);
6955 } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6956 u32 new_nlinks, old_nlinks;
6958 new_nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6959 old_nlinks = btrfs_inode_nlink(sctx->right_path->nodes[0], right_ii);
6960 if (new_nlinks == 0 && old_nlinks == 0) {
6961 sctx->ignore_cur_inode = true;
6963 } else if (new_nlinks == 0 || old_nlinks == 0) {
6964 sctx->cur_inode_new_gen = 1;
6967 * We need to do some special handling in case the inode was
6968 * reported as changed with a changed generation number. This
6969 * means that the original inode was deleted and new inode
6970 * reused the same inum. So we have to treat the old inode as
6971 * deleted and the new one as new.
6973 if (sctx->cur_inode_new_gen) {
6975 * First, process the inode as if it was deleted.
6977 if (old_nlinks > 0) {
6978 sctx->cur_inode_gen = right_gen;
6979 sctx->cur_inode_new = false;
6980 sctx->cur_inode_deleted = true;
6981 sctx->cur_inode_size = btrfs_inode_size(
6982 sctx->right_path->nodes[0], right_ii);
6983 sctx->cur_inode_mode = btrfs_inode_mode(
6984 sctx->right_path->nodes[0], right_ii);
6985 ret = process_all_refs(sctx,
6986 BTRFS_COMPARE_TREE_DELETED);
6992 * Now process the inode as if it was new.
6994 if (new_nlinks > 0) {
6995 sctx->cur_inode_gen = left_gen;
6996 sctx->cur_inode_new = true;
6997 sctx->cur_inode_deleted = false;
6998 sctx->cur_inode_size = btrfs_inode_size(
6999 sctx->left_path->nodes[0],
7001 sctx->cur_inode_mode = btrfs_inode_mode(
7002 sctx->left_path->nodes[0],
7004 sctx->cur_inode_rdev = btrfs_inode_rdev(
7005 sctx->left_path->nodes[0],
7007 ret = send_create_inode_if_needed(sctx);
7011 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
7015 * Advance send_progress now as we did not get
7016 * into process_recorded_refs_if_needed in the
7019 sctx->send_progress = sctx->cur_ino + 1;
7022 * Now process all extents and xattrs of the
7023 * inode as if they were all new.
7025 ret = process_all_extents(sctx);
7028 ret = process_all_new_xattrs(sctx);
7033 sctx->cur_inode_gen = left_gen;
7034 sctx->cur_inode_new = false;
7035 sctx->cur_inode_new_gen = false;
7036 sctx->cur_inode_deleted = false;
7037 sctx->cur_inode_size = btrfs_inode_size(
7038 sctx->left_path->nodes[0], left_ii);
7039 sctx->cur_inode_mode = btrfs_inode_mode(
7040 sctx->left_path->nodes[0], left_ii);
7049 * We have to process new refs before deleted refs, but compare_trees gives us
7050 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
7051 * first and later process them in process_recorded_refs.
7052 * For the cur_inode_new_gen case, we skip recording completely because
7053 * changed_inode did already initiate processing of refs. The reason for this is
7054 * that in this case, compare_tree actually compares the refs of 2 different
7055 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
7056 * refs of the right tree as deleted and all refs of the left tree as new.
7058 static int changed_ref(struct send_ctx *sctx,
7059 enum btrfs_compare_tree_result result)
7063 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7064 inconsistent_snapshot_error(sctx, result, "reference");
7068 if (!sctx->cur_inode_new_gen &&
7069 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
7070 if (result == BTRFS_COMPARE_TREE_NEW)
7071 ret = record_new_ref(sctx);
7072 else if (result == BTRFS_COMPARE_TREE_DELETED)
7073 ret = record_deleted_ref(sctx);
7074 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7075 ret = record_changed_ref(sctx);
7082 * Process new/deleted/changed xattrs. We skip processing in the
7083 * cur_inode_new_gen case because changed_inode did already initiate processing
7084 * of xattrs. The reason is the same as in changed_ref
7086 static int changed_xattr(struct send_ctx *sctx,
7087 enum btrfs_compare_tree_result result)
7091 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7092 inconsistent_snapshot_error(sctx, result, "xattr");
7096 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7097 if (result == BTRFS_COMPARE_TREE_NEW)
7098 ret = process_new_xattr(sctx);
7099 else if (result == BTRFS_COMPARE_TREE_DELETED)
7100 ret = process_deleted_xattr(sctx);
7101 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7102 ret = process_changed_xattr(sctx);
7109 * Process new/deleted/changed extents. We skip processing in the
7110 * cur_inode_new_gen case because changed_inode did already initiate processing
7111 * of extents. The reason is the same as in changed_ref
7113 static int changed_extent(struct send_ctx *sctx,
7114 enum btrfs_compare_tree_result result)
7119 * We have found an extent item that changed without the inode item
7120 * having changed. This can happen either after relocation (where the
7121 * disk_bytenr of an extent item is replaced at
7122 * relocation.c:replace_file_extents()) or after deduplication into a
7123 * file in both the parent and send snapshots (where an extent item can
7124 * get modified or replaced with a new one). Note that deduplication
7125 * updates the inode item, but it only changes the iversion (sequence
7126 * field in the inode item) of the inode, so if a file is deduplicated
7127 * the same amount of times in both the parent and send snapshots, its
7128 * iversion becomes the same in both snapshots, whence the inode item is
7129 * the same on both snapshots.
7131 if (sctx->cur_ino != sctx->cmp_key->objectid)
7134 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7135 if (result != BTRFS_COMPARE_TREE_DELETED)
7136 ret = process_extent(sctx, sctx->left_path,
7143 static int changed_verity(struct send_ctx *sctx, enum btrfs_compare_tree_result result)
7147 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7148 if (result == BTRFS_COMPARE_TREE_NEW)
7149 sctx->cur_inode_needs_verity = true;
7154 static int dir_changed(struct send_ctx *sctx, u64 dir)
7156 u64 orig_gen, new_gen;
7159 ret = get_inode_gen(sctx->send_root, dir, &new_gen);
7163 ret = get_inode_gen(sctx->parent_root, dir, &orig_gen);
7167 return (orig_gen != new_gen) ? 1 : 0;
7170 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
7171 struct btrfs_key *key)
7173 struct btrfs_inode_extref *extref;
7174 struct extent_buffer *leaf;
7175 u64 dirid = 0, last_dirid = 0;
7182 /* Easy case, just check this one dirid */
7183 if (key->type == BTRFS_INODE_REF_KEY) {
7184 dirid = key->offset;
7186 ret = dir_changed(sctx, dirid);
7190 leaf = path->nodes[0];
7191 item_size = btrfs_item_size(leaf, path->slots[0]);
7192 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
7193 while (cur_offset < item_size) {
7194 extref = (struct btrfs_inode_extref *)(ptr +
7196 dirid = btrfs_inode_extref_parent(leaf, extref);
7197 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
7198 cur_offset += ref_name_len + sizeof(*extref);
7199 if (dirid == last_dirid)
7201 ret = dir_changed(sctx, dirid);
7211 * Updates compare related fields in sctx and simply forwards to the actual
7212 * changed_xxx functions.
7214 static int changed_cb(struct btrfs_path *left_path,
7215 struct btrfs_path *right_path,
7216 struct btrfs_key *key,
7217 enum btrfs_compare_tree_result result,
7218 struct send_ctx *sctx)
7223 * We can not hold the commit root semaphore here. This is because in
7224 * the case of sending and receiving to the same filesystem, using a
7225 * pipe, could result in a deadlock:
7227 * 1) The task running send blocks on the pipe because it's full;
7229 * 2) The task running receive, which is the only consumer of the pipe,
7230 * is waiting for a transaction commit (for example due to a space
7231 * reservation when doing a write or triggering a transaction commit
7232 * when creating a subvolume);
7234 * 3) The transaction is waiting to write lock the commit root semaphore,
7235 * but can not acquire it since it's being held at 1).
7237 * Down this call chain we write to the pipe through kernel_write().
7238 * The same type of problem can also happen when sending to a file that
7239 * is stored in the same filesystem - when reserving space for a write
7240 * into the file, we can trigger a transaction commit.
7242 * Our caller has supplied us with clones of leaves from the send and
7243 * parent roots, so we're safe here from a concurrent relocation and
7244 * further reallocation of metadata extents while we are here. Below we
7245 * also assert that the leaves are clones.
7247 lockdep_assert_not_held(&sctx->send_root->fs_info->commit_root_sem);
7250 * We always have a send root, so left_path is never NULL. We will not
7251 * have a leaf when we have reached the end of the send root but have
7252 * not yet reached the end of the parent root.
7254 if (left_path->nodes[0])
7255 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7256 &left_path->nodes[0]->bflags));
7258 * When doing a full send we don't have a parent root, so right_path is
7259 * NULL. When doing an incremental send, we may have reached the end of
7260 * the parent root already, so we don't have a leaf at right_path.
7262 if (right_path && right_path->nodes[0])
7263 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7264 &right_path->nodes[0]->bflags));
7266 if (result == BTRFS_COMPARE_TREE_SAME) {
7267 if (key->type == BTRFS_INODE_REF_KEY ||
7268 key->type == BTRFS_INODE_EXTREF_KEY) {
7269 ret = compare_refs(sctx, left_path, key);
7274 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
7275 return maybe_send_hole(sctx, left_path, key);
7279 result = BTRFS_COMPARE_TREE_CHANGED;
7283 sctx->left_path = left_path;
7284 sctx->right_path = right_path;
7285 sctx->cmp_key = key;
7287 ret = finish_inode_if_needed(sctx, 0);
7291 /* Ignore non-FS objects */
7292 if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
7293 key->objectid == BTRFS_FREE_SPACE_OBJECTID)
7296 if (key->type == BTRFS_INODE_ITEM_KEY) {
7297 ret = changed_inode(sctx, result);
7298 } else if (!sctx->ignore_cur_inode) {
7299 if (key->type == BTRFS_INODE_REF_KEY ||
7300 key->type == BTRFS_INODE_EXTREF_KEY)
7301 ret = changed_ref(sctx, result);
7302 else if (key->type == BTRFS_XATTR_ITEM_KEY)
7303 ret = changed_xattr(sctx, result);
7304 else if (key->type == BTRFS_EXTENT_DATA_KEY)
7305 ret = changed_extent(sctx, result);
7306 else if (key->type == BTRFS_VERITY_DESC_ITEM_KEY &&
7308 ret = changed_verity(sctx, result);
7315 static int search_key_again(const struct send_ctx *sctx,
7316 struct btrfs_root *root,
7317 struct btrfs_path *path,
7318 const struct btrfs_key *key)
7322 if (!path->need_commit_sem)
7323 lockdep_assert_held_read(&root->fs_info->commit_root_sem);
7326 * Roots used for send operations are readonly and no one can add,
7327 * update or remove keys from them, so we should be able to find our
7328 * key again. The only exception is deduplication, which can operate on
7329 * readonly roots and add, update or remove keys to/from them - but at
7330 * the moment we don't allow it to run in parallel with send.
7332 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7335 btrfs_print_tree(path->nodes[path->lowest_level], false);
7336 btrfs_err(root->fs_info,
7337 "send: key (%llu %u %llu) not found in %s root %llu, lowest_level %d, slot %d",
7338 key->objectid, key->type, key->offset,
7339 (root == sctx->parent_root ? "parent" : "send"),
7340 root->root_key.objectid, path->lowest_level,
7341 path->slots[path->lowest_level]);
7348 static int full_send_tree(struct send_ctx *sctx)
7351 struct btrfs_root *send_root = sctx->send_root;
7352 struct btrfs_key key;
7353 struct btrfs_fs_info *fs_info = send_root->fs_info;
7354 struct btrfs_path *path;
7356 path = alloc_path_for_send();
7359 path->reada = READA_FORWARD_ALWAYS;
7361 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
7362 key.type = BTRFS_INODE_ITEM_KEY;
7365 down_read(&fs_info->commit_root_sem);
7366 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7367 up_read(&fs_info->commit_root_sem);
7369 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
7376 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
7378 ret = changed_cb(path, NULL, &key,
7379 BTRFS_COMPARE_TREE_NEW, sctx);
7383 down_read(&fs_info->commit_root_sem);
7384 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7385 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7386 up_read(&fs_info->commit_root_sem);
7388 * A transaction used for relocating a block group was
7389 * committed or is about to finish its commit. Release
7390 * our path (leaf) and restart the search, so that we
7391 * avoid operating on any file extent items that are
7392 * stale, with a disk_bytenr that reflects a pre
7393 * relocation value. This way we avoid as much as
7394 * possible to fallback to regular writes when checking
7395 * if we can clone file ranges.
7397 btrfs_release_path(path);
7398 ret = search_key_again(sctx, send_root, path, &key);
7402 up_read(&fs_info->commit_root_sem);
7405 ret = btrfs_next_item(send_root, path);
7415 ret = finish_inode_if_needed(sctx, 1);
7418 btrfs_free_path(path);
7422 static int replace_node_with_clone(struct btrfs_path *path, int level)
7424 struct extent_buffer *clone;
7426 clone = btrfs_clone_extent_buffer(path->nodes[level]);
7430 free_extent_buffer(path->nodes[level]);
7431 path->nodes[level] = clone;
7436 static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
7438 struct extent_buffer *eb;
7439 struct extent_buffer *parent = path->nodes[*level];
7440 int slot = path->slots[*level];
7441 const int nritems = btrfs_header_nritems(parent);
7445 lockdep_assert_held_read(&parent->fs_info->commit_root_sem);
7446 ASSERT(*level != 0);
7448 eb = btrfs_read_node_slot(parent, slot);
7453 * Trigger readahead for the next leaves we will process, so that it is
7454 * very likely that when we need them they are already in memory and we
7455 * will not block on disk IO. For nodes we only do readahead for one,
7456 * since the time window between processing nodes is typically larger.
7458 reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
7460 for (slot++; slot < nritems && reada_done < reada_max; slot++) {
7461 if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
7462 btrfs_readahead_node_child(parent, slot);
7463 reada_done += eb->fs_info->nodesize;
7467 path->nodes[*level - 1] = eb;
7468 path->slots[*level - 1] = 0;
7472 return replace_node_with_clone(path, 0);
7477 static int tree_move_next_or_upnext(struct btrfs_path *path,
7478 int *level, int root_level)
7482 nritems = btrfs_header_nritems(path->nodes[*level]);
7484 path->slots[*level]++;
7486 while (path->slots[*level] >= nritems) {
7487 if (*level == root_level) {
7488 path->slots[*level] = nritems - 1;
7493 path->slots[*level] = 0;
7494 free_extent_buffer(path->nodes[*level]);
7495 path->nodes[*level] = NULL;
7497 path->slots[*level]++;
7499 nritems = btrfs_header_nritems(path->nodes[*level]);
7506 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
7509 static int tree_advance(struct btrfs_path *path,
7510 int *level, int root_level,
7512 struct btrfs_key *key,
7517 if (*level == 0 || !allow_down) {
7518 ret = tree_move_next_or_upnext(path, level, root_level);
7520 ret = tree_move_down(path, level, reada_min_gen);
7524 * Even if we have reached the end of a tree, ret is -1, update the key
7525 * anyway, so that in case we need to restart due to a block group
7526 * relocation, we can assert that the last key of the root node still
7527 * exists in the tree.
7530 btrfs_item_key_to_cpu(path->nodes[*level], key,
7531 path->slots[*level]);
7533 btrfs_node_key_to_cpu(path->nodes[*level], key,
7534 path->slots[*level]);
7539 static int tree_compare_item(struct btrfs_path *left_path,
7540 struct btrfs_path *right_path,
7545 unsigned long off1, off2;
7547 len1 = btrfs_item_size(left_path->nodes[0], left_path->slots[0]);
7548 len2 = btrfs_item_size(right_path->nodes[0], right_path->slots[0]);
7552 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
7553 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
7554 right_path->slots[0]);
7556 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
7558 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
7565 * A transaction used for relocating a block group was committed or is about to
7566 * finish its commit. Release our paths and restart the search, so that we are
7567 * not using stale extent buffers:
7569 * 1) For levels > 0, we are only holding references of extent buffers, without
7570 * any locks on them, which does not prevent them from having been relocated
7571 * and reallocated after the last time we released the commit root semaphore.
7572 * The exception are the root nodes, for which we always have a clone, see
7573 * the comment at btrfs_compare_trees();
7575 * 2) For leaves, level 0, we are holding copies (clones) of extent buffers, so
7576 * we are safe from the concurrent relocation and reallocation. However they
7577 * can have file extent items with a pre relocation disk_bytenr value, so we
7578 * restart the start from the current commit roots and clone the new leaves so
7579 * that we get the post relocation disk_bytenr values. Not doing so, could
7580 * make us clone the wrong data in case there are new extents using the old
7581 * disk_bytenr that happen to be shared.
7583 static int restart_after_relocation(struct btrfs_path *left_path,
7584 struct btrfs_path *right_path,
7585 const struct btrfs_key *left_key,
7586 const struct btrfs_key *right_key,
7589 const struct send_ctx *sctx)
7594 lockdep_assert_held_read(&sctx->send_root->fs_info->commit_root_sem);
7596 btrfs_release_path(left_path);
7597 btrfs_release_path(right_path);
7600 * Since keys can not be added or removed to/from our roots because they
7601 * are readonly and we do not allow deduplication to run in parallel
7602 * (which can add, remove or change keys), the layout of the trees should
7605 left_path->lowest_level = left_level;
7606 ret = search_key_again(sctx, sctx->send_root, left_path, left_key);
7610 right_path->lowest_level = right_level;
7611 ret = search_key_again(sctx, sctx->parent_root, right_path, right_key);
7616 * If the lowest level nodes are leaves, clone them so that they can be
7617 * safely used by changed_cb() while not under the protection of the
7618 * commit root semaphore, even if relocation and reallocation happens in
7621 if (left_level == 0) {
7622 ret = replace_node_with_clone(left_path, 0);
7627 if (right_level == 0) {
7628 ret = replace_node_with_clone(right_path, 0);
7634 * Now clone the root nodes (unless they happen to be the leaves we have
7635 * already cloned). This is to protect against concurrent snapshotting of
7636 * the send and parent roots (see the comment at btrfs_compare_trees()).
7638 root_level = btrfs_header_level(sctx->send_root->commit_root);
7639 if (root_level > 0) {
7640 ret = replace_node_with_clone(left_path, root_level);
7645 root_level = btrfs_header_level(sctx->parent_root->commit_root);
7646 if (root_level > 0) {
7647 ret = replace_node_with_clone(right_path, root_level);
7656 * This function compares two trees and calls the provided callback for
7657 * every changed/new/deleted item it finds.
7658 * If shared tree blocks are encountered, whole subtrees are skipped, making
7659 * the compare pretty fast on snapshotted subvolumes.
7661 * This currently works on commit roots only. As commit roots are read only,
7662 * we don't do any locking. The commit roots are protected with transactions.
7663 * Transactions are ended and rejoined when a commit is tried in between.
7665 * This function checks for modifications done to the trees while comparing.
7666 * If it detects a change, it aborts immediately.
7668 static int btrfs_compare_trees(struct btrfs_root *left_root,
7669 struct btrfs_root *right_root, struct send_ctx *sctx)
7671 struct btrfs_fs_info *fs_info = left_root->fs_info;
7674 struct btrfs_path *left_path = NULL;
7675 struct btrfs_path *right_path = NULL;
7676 struct btrfs_key left_key;
7677 struct btrfs_key right_key;
7678 char *tmp_buf = NULL;
7679 int left_root_level;
7680 int right_root_level;
7683 int left_end_reached = 0;
7684 int right_end_reached = 0;
7685 int advance_left = 0;
7686 int advance_right = 0;
7693 left_path = btrfs_alloc_path();
7698 right_path = btrfs_alloc_path();
7704 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
7710 left_path->search_commit_root = 1;
7711 left_path->skip_locking = 1;
7712 right_path->search_commit_root = 1;
7713 right_path->skip_locking = 1;
7716 * Strategy: Go to the first items of both trees. Then do
7718 * If both trees are at level 0
7719 * Compare keys of current items
7720 * If left < right treat left item as new, advance left tree
7722 * If left > right treat right item as deleted, advance right tree
7724 * If left == right do deep compare of items, treat as changed if
7725 * needed, advance both trees and repeat
7726 * If both trees are at the same level but not at level 0
7727 * Compare keys of current nodes/leafs
7728 * If left < right advance left tree and repeat
7729 * If left > right advance right tree and repeat
7730 * If left == right compare blockptrs of the next nodes/leafs
7731 * If they match advance both trees but stay at the same level
7733 * If they don't match advance both trees while allowing to go
7735 * If tree levels are different
7736 * Advance the tree that needs it and repeat
7738 * Advancing a tree means:
7739 * If we are at level 0, try to go to the next slot. If that's not
7740 * possible, go one level up and repeat. Stop when we found a level
7741 * where we could go to the next slot. We may at this point be on a
7744 * If we are not at level 0 and not on shared tree blocks, go one
7747 * If we are not at level 0 and on shared tree blocks, go one slot to
7748 * the right if possible or go up and right.
7751 down_read(&fs_info->commit_root_sem);
7752 left_level = btrfs_header_level(left_root->commit_root);
7753 left_root_level = left_level;
7755 * We clone the root node of the send and parent roots to prevent races
7756 * with snapshot creation of these roots. Snapshot creation COWs the
7757 * root node of a tree, so after the transaction is committed the old
7758 * extent can be reallocated while this send operation is still ongoing.
7759 * So we clone them, under the commit root semaphore, to be race free.
7761 left_path->nodes[left_level] =
7762 btrfs_clone_extent_buffer(left_root->commit_root);
7763 if (!left_path->nodes[left_level]) {
7768 right_level = btrfs_header_level(right_root->commit_root);
7769 right_root_level = right_level;
7770 right_path->nodes[right_level] =
7771 btrfs_clone_extent_buffer(right_root->commit_root);
7772 if (!right_path->nodes[right_level]) {
7777 * Our right root is the parent root, while the left root is the "send"
7778 * root. We know that all new nodes/leaves in the left root must have
7779 * a generation greater than the right root's generation, so we trigger
7780 * readahead for those nodes and leaves of the left root, as we know we
7781 * will need to read them at some point.
7783 reada_min_gen = btrfs_header_generation(right_root->commit_root);
7785 if (left_level == 0)
7786 btrfs_item_key_to_cpu(left_path->nodes[left_level],
7787 &left_key, left_path->slots[left_level]);
7789 btrfs_node_key_to_cpu(left_path->nodes[left_level],
7790 &left_key, left_path->slots[left_level]);
7791 if (right_level == 0)
7792 btrfs_item_key_to_cpu(right_path->nodes[right_level],
7793 &right_key, right_path->slots[right_level]);
7795 btrfs_node_key_to_cpu(right_path->nodes[right_level],
7796 &right_key, right_path->slots[right_level]);
7798 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7801 if (need_resched() ||
7802 rwsem_is_contended(&fs_info->commit_root_sem)) {
7803 up_read(&fs_info->commit_root_sem);
7805 down_read(&fs_info->commit_root_sem);
7808 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7809 ret = restart_after_relocation(left_path, right_path,
7810 &left_key, &right_key,
7811 left_level, right_level,
7815 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7818 if (advance_left && !left_end_reached) {
7819 ret = tree_advance(left_path, &left_level,
7821 advance_left != ADVANCE_ONLY_NEXT,
7822 &left_key, reada_min_gen);
7824 left_end_reached = ADVANCE;
7829 if (advance_right && !right_end_reached) {
7830 ret = tree_advance(right_path, &right_level,
7832 advance_right != ADVANCE_ONLY_NEXT,
7833 &right_key, reada_min_gen);
7835 right_end_reached = ADVANCE;
7841 if (left_end_reached && right_end_reached) {
7844 } else if (left_end_reached) {
7845 if (right_level == 0) {
7846 up_read(&fs_info->commit_root_sem);
7847 ret = changed_cb(left_path, right_path,
7849 BTRFS_COMPARE_TREE_DELETED,
7853 down_read(&fs_info->commit_root_sem);
7855 advance_right = ADVANCE;
7857 } else if (right_end_reached) {
7858 if (left_level == 0) {
7859 up_read(&fs_info->commit_root_sem);
7860 ret = changed_cb(left_path, right_path,
7862 BTRFS_COMPARE_TREE_NEW,
7866 down_read(&fs_info->commit_root_sem);
7868 advance_left = ADVANCE;
7872 if (left_level == 0 && right_level == 0) {
7873 up_read(&fs_info->commit_root_sem);
7874 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7876 ret = changed_cb(left_path, right_path,
7878 BTRFS_COMPARE_TREE_NEW,
7880 advance_left = ADVANCE;
7881 } else if (cmp > 0) {
7882 ret = changed_cb(left_path, right_path,
7884 BTRFS_COMPARE_TREE_DELETED,
7886 advance_right = ADVANCE;
7888 enum btrfs_compare_tree_result result;
7890 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7891 ret = tree_compare_item(left_path, right_path,
7894 result = BTRFS_COMPARE_TREE_CHANGED;
7896 result = BTRFS_COMPARE_TREE_SAME;
7897 ret = changed_cb(left_path, right_path,
7898 &left_key, result, sctx);
7899 advance_left = ADVANCE;
7900 advance_right = ADVANCE;
7905 down_read(&fs_info->commit_root_sem);
7906 } else if (left_level == right_level) {
7907 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7909 advance_left = ADVANCE;
7910 } else if (cmp > 0) {
7911 advance_right = ADVANCE;
7913 left_blockptr = btrfs_node_blockptr(
7914 left_path->nodes[left_level],
7915 left_path->slots[left_level]);
7916 right_blockptr = btrfs_node_blockptr(
7917 right_path->nodes[right_level],
7918 right_path->slots[right_level]);
7919 left_gen = btrfs_node_ptr_generation(
7920 left_path->nodes[left_level],
7921 left_path->slots[left_level]);
7922 right_gen = btrfs_node_ptr_generation(
7923 right_path->nodes[right_level],
7924 right_path->slots[right_level]);
7925 if (left_blockptr == right_blockptr &&
7926 left_gen == right_gen) {
7928 * As we're on a shared block, don't
7929 * allow to go deeper.
7931 advance_left = ADVANCE_ONLY_NEXT;
7932 advance_right = ADVANCE_ONLY_NEXT;
7934 advance_left = ADVANCE;
7935 advance_right = ADVANCE;
7938 } else if (left_level < right_level) {
7939 advance_right = ADVANCE;
7941 advance_left = ADVANCE;
7946 up_read(&fs_info->commit_root_sem);
7948 btrfs_free_path(left_path);
7949 btrfs_free_path(right_path);
7954 static int send_subvol(struct send_ctx *sctx)
7958 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7959 ret = send_header(sctx);
7964 ret = send_subvol_begin(sctx);
7968 if (sctx->parent_root) {
7969 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7972 ret = finish_inode_if_needed(sctx, 1);
7976 ret = full_send_tree(sctx);
7982 free_recorded_refs(sctx);
7987 * If orphan cleanup did remove any orphans from a root, it means the tree
7988 * was modified and therefore the commit root is not the same as the current
7989 * root anymore. This is a problem, because send uses the commit root and
7990 * therefore can see inode items that don't exist in the current root anymore,
7991 * and for example make calls to btrfs_iget, which will do tree lookups based
7992 * on the current root and not on the commit root. Those lookups will fail,
7993 * returning a -ESTALE error, and making send fail with that error. So make
7994 * sure a send does not see any orphans we have just removed, and that it will
7995 * see the same inodes regardless of whether a transaction commit happened
7996 * before it started (meaning that the commit root will be the same as the
7997 * current root) or not.
7999 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
8002 struct btrfs_trans_handle *trans = NULL;
8005 if (sctx->parent_root &&
8006 sctx->parent_root->node != sctx->parent_root->commit_root)
8009 for (i = 0; i < sctx->clone_roots_cnt; i++)
8010 if (sctx->clone_roots[i].root->node !=
8011 sctx->clone_roots[i].root->commit_root)
8015 return btrfs_end_transaction(trans);
8020 /* Use any root, all fs roots will get their commit roots updated. */
8022 trans = btrfs_join_transaction(sctx->send_root);
8024 return PTR_ERR(trans);
8028 return btrfs_commit_transaction(trans);
8032 * Make sure any existing dellaloc is flushed for any root used by a send
8033 * operation so that we do not miss any data and we do not race with writeback
8034 * finishing and changing a tree while send is using the tree. This could
8035 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
8036 * a send operation then uses the subvolume.
8037 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
8039 static int flush_delalloc_roots(struct send_ctx *sctx)
8041 struct btrfs_root *root = sctx->parent_root;
8046 ret = btrfs_start_delalloc_snapshot(root, false);
8049 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8052 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8053 root = sctx->clone_roots[i].root;
8054 ret = btrfs_start_delalloc_snapshot(root, false);
8057 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8063 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
8065 spin_lock(&root->root_item_lock);
8066 root->send_in_progress--;
8068 * Not much left to do, we don't know why it's unbalanced and
8069 * can't blindly reset it to 0.
8071 if (root->send_in_progress < 0)
8072 btrfs_err(root->fs_info,
8073 "send_in_progress unbalanced %d root %llu",
8074 root->send_in_progress, root->root_key.objectid);
8075 spin_unlock(&root->root_item_lock);
8078 static void dedupe_in_progress_warn(const struct btrfs_root *root)
8080 btrfs_warn_rl(root->fs_info,
8081 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
8082 root->root_key.objectid, root->dedupe_in_progress);
8085 long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
8088 struct btrfs_root *send_root = BTRFS_I(inode)->root;
8089 struct btrfs_fs_info *fs_info = send_root->fs_info;
8090 struct btrfs_root *clone_root;
8091 struct send_ctx *sctx = NULL;
8093 u64 *clone_sources_tmp = NULL;
8094 int clone_sources_to_rollback = 0;
8096 int sort_clone_roots = 0;
8097 struct btrfs_lru_cache_entry *entry;
8098 struct btrfs_lru_cache_entry *tmp;
8100 if (!capable(CAP_SYS_ADMIN))
8104 * The subvolume must remain read-only during send, protect against
8105 * making it RW. This also protects against deletion.
8107 spin_lock(&send_root->root_item_lock);
8108 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
8109 dedupe_in_progress_warn(send_root);
8110 spin_unlock(&send_root->root_item_lock);
8113 send_root->send_in_progress++;
8114 spin_unlock(&send_root->root_item_lock);
8117 * Userspace tools do the checks and warn the user if it's
8120 if (!btrfs_root_readonly(send_root)) {
8126 * Check that we don't overflow at later allocations, we request
8127 * clone_sources_count + 1 items, and compare to unsigned long inside
8128 * access_ok. Also set an upper limit for allocation size so this can't
8129 * easily exhaust memory. Max number of clone sources is about 200K.
8131 if (arg->clone_sources_count > SZ_8M / sizeof(struct clone_root)) {
8136 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
8141 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
8147 INIT_LIST_HEAD(&sctx->new_refs);
8148 INIT_LIST_HEAD(&sctx->deleted_refs);
8150 btrfs_lru_cache_init(&sctx->name_cache, SEND_MAX_NAME_CACHE_SIZE);
8151 btrfs_lru_cache_init(&sctx->backref_cache, SEND_MAX_BACKREF_CACHE_SIZE);
8152 btrfs_lru_cache_init(&sctx->dir_created_cache,
8153 SEND_MAX_DIR_CREATED_CACHE_SIZE);
8155 * This cache is periodically trimmed to a fixed size elsewhere, see
8156 * cache_dir_utimes() and trim_dir_utimes_cache().
8158 btrfs_lru_cache_init(&sctx->dir_utimes_cache, 0);
8160 sctx->pending_dir_moves = RB_ROOT;
8161 sctx->waiting_dir_moves = RB_ROOT;
8162 sctx->orphan_dirs = RB_ROOT;
8163 sctx->rbtree_new_refs = RB_ROOT;
8164 sctx->rbtree_deleted_refs = RB_ROOT;
8166 sctx->flags = arg->flags;
8168 if (arg->flags & BTRFS_SEND_FLAG_VERSION) {
8169 if (arg->version > BTRFS_SEND_STREAM_VERSION) {
8173 /* Zero means "use the highest version" */
8174 sctx->proto = arg->version ?: BTRFS_SEND_STREAM_VERSION;
8178 if ((arg->flags & BTRFS_SEND_FLAG_COMPRESSED) && sctx->proto < 2) {
8183 sctx->send_filp = fget(arg->send_fd);
8184 if (!sctx->send_filp || !(sctx->send_filp->f_mode & FMODE_WRITE)) {
8189 sctx->send_root = send_root;
8191 * Unlikely but possible, if the subvolume is marked for deletion but
8192 * is slow to remove the directory entry, send can still be started
8194 if (btrfs_root_dead(sctx->send_root)) {
8199 sctx->clone_roots_cnt = arg->clone_sources_count;
8201 if (sctx->proto >= 2) {
8202 u32 send_buf_num_pages;
8204 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V2;
8205 sctx->send_buf = vmalloc(sctx->send_max_size);
8206 if (!sctx->send_buf) {
8210 send_buf_num_pages = sctx->send_max_size >> PAGE_SHIFT;
8211 sctx->send_buf_pages = kcalloc(send_buf_num_pages,
8212 sizeof(*sctx->send_buf_pages),
8214 if (!sctx->send_buf_pages) {
8218 for (i = 0; i < send_buf_num_pages; i++) {
8219 sctx->send_buf_pages[i] =
8220 vmalloc_to_page(sctx->send_buf + (i << PAGE_SHIFT));
8223 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
8224 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
8226 if (!sctx->send_buf) {
8231 sctx->clone_roots = kvcalloc(arg->clone_sources_count + 1,
8232 sizeof(*sctx->clone_roots),
8234 if (!sctx->clone_roots) {
8239 alloc_size = array_size(sizeof(*arg->clone_sources),
8240 arg->clone_sources_count);
8242 if (arg->clone_sources_count) {
8243 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
8244 if (!clone_sources_tmp) {
8249 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
8256 for (i = 0; i < arg->clone_sources_count; i++) {
8257 clone_root = btrfs_get_fs_root(fs_info,
8258 clone_sources_tmp[i], true);
8259 if (IS_ERR(clone_root)) {
8260 ret = PTR_ERR(clone_root);
8263 spin_lock(&clone_root->root_item_lock);
8264 if (!btrfs_root_readonly(clone_root) ||
8265 btrfs_root_dead(clone_root)) {
8266 spin_unlock(&clone_root->root_item_lock);
8267 btrfs_put_root(clone_root);
8271 if (clone_root->dedupe_in_progress) {
8272 dedupe_in_progress_warn(clone_root);
8273 spin_unlock(&clone_root->root_item_lock);
8274 btrfs_put_root(clone_root);
8278 clone_root->send_in_progress++;
8279 spin_unlock(&clone_root->root_item_lock);
8281 sctx->clone_roots[i].root = clone_root;
8282 clone_sources_to_rollback = i + 1;
8284 kvfree(clone_sources_tmp);
8285 clone_sources_tmp = NULL;
8288 if (arg->parent_root) {
8289 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
8291 if (IS_ERR(sctx->parent_root)) {
8292 ret = PTR_ERR(sctx->parent_root);
8296 spin_lock(&sctx->parent_root->root_item_lock);
8297 sctx->parent_root->send_in_progress++;
8298 if (!btrfs_root_readonly(sctx->parent_root) ||
8299 btrfs_root_dead(sctx->parent_root)) {
8300 spin_unlock(&sctx->parent_root->root_item_lock);
8304 if (sctx->parent_root->dedupe_in_progress) {
8305 dedupe_in_progress_warn(sctx->parent_root);
8306 spin_unlock(&sctx->parent_root->root_item_lock);
8310 spin_unlock(&sctx->parent_root->root_item_lock);
8314 * Clones from send_root are allowed, but only if the clone source
8315 * is behind the current send position. This is checked while searching
8316 * for possible clone sources.
8318 sctx->clone_roots[sctx->clone_roots_cnt++].root =
8319 btrfs_grab_root(sctx->send_root);
8321 /* We do a bsearch later */
8322 sort(sctx->clone_roots, sctx->clone_roots_cnt,
8323 sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
8325 sort_clone_roots = 1;
8327 ret = flush_delalloc_roots(sctx);
8331 ret = ensure_commit_roots_uptodate(sctx);
8335 ret = send_subvol(sctx);
8339 btrfs_lru_cache_for_each_entry_safe(&sctx->dir_utimes_cache, entry, tmp) {
8340 ret = send_utimes(sctx, entry->key, entry->gen);
8343 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, entry);
8346 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
8347 ret = begin_cmd(sctx, BTRFS_SEND_C_END);
8350 ret = send_cmd(sctx);
8356 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
8357 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
8359 struct pending_dir_move *pm;
8361 n = rb_first(&sctx->pending_dir_moves);
8362 pm = rb_entry(n, struct pending_dir_move, node);
8363 while (!list_empty(&pm->list)) {
8364 struct pending_dir_move *pm2;
8366 pm2 = list_first_entry(&pm->list,
8367 struct pending_dir_move, list);
8368 free_pending_move(sctx, pm2);
8370 free_pending_move(sctx, pm);
8373 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
8374 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
8376 struct waiting_dir_move *dm;
8378 n = rb_first(&sctx->waiting_dir_moves);
8379 dm = rb_entry(n, struct waiting_dir_move, node);
8380 rb_erase(&dm->node, &sctx->waiting_dir_moves);
8384 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
8385 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
8387 struct orphan_dir_info *odi;
8389 n = rb_first(&sctx->orphan_dirs);
8390 odi = rb_entry(n, struct orphan_dir_info, node);
8391 free_orphan_dir_info(sctx, odi);
8394 if (sort_clone_roots) {
8395 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8396 btrfs_root_dec_send_in_progress(
8397 sctx->clone_roots[i].root);
8398 btrfs_put_root(sctx->clone_roots[i].root);
8401 for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
8402 btrfs_root_dec_send_in_progress(
8403 sctx->clone_roots[i].root);
8404 btrfs_put_root(sctx->clone_roots[i].root);
8407 btrfs_root_dec_send_in_progress(send_root);
8409 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
8410 btrfs_root_dec_send_in_progress(sctx->parent_root);
8411 btrfs_put_root(sctx->parent_root);
8414 kvfree(clone_sources_tmp);
8417 if (sctx->send_filp)
8418 fput(sctx->send_filp);
8420 kvfree(sctx->clone_roots);
8421 kfree(sctx->send_buf_pages);
8422 kvfree(sctx->send_buf);
8423 kvfree(sctx->verity_descriptor);
8425 close_current_inode(sctx);
8427 btrfs_lru_cache_clear(&sctx->name_cache);
8428 btrfs_lru_cache_clear(&sctx->backref_cache);
8429 btrfs_lru_cache_clear(&sctx->dir_created_cache);
8430 btrfs_lru_cache_clear(&sctx->dir_utimes_cache);