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"
29 #include "print-tree.h"
30 #include "accessors.h"
32 #include "file-item.h"
35 #include "lru_cache.h"
38 * Maximum number of references an extent can have in order for us to attempt to
39 * issue clone operations instead of write operations. This currently exists to
40 * avoid hitting limitations of the backreference walking code (taking a lot of
41 * time and using too much memory for extents with large number of references).
43 #define SEND_MAX_EXTENT_REFS 1024
46 * A fs_path is a helper to dynamically build path names with unknown size.
47 * It reallocates the internal buffer on demand.
48 * It allows fast adding of path elements on the right side (normal path) and
49 * fast adding to the left side (reversed path). A reversed path can also be
50 * unreversed if needed.
59 unsigned short buf_len:15;
60 unsigned short reversed:1;
64 * Average path length does not exceed 200 bytes, we'll have
65 * better packing in the slab and higher chance to satisfy
66 * a allocation later during send.
71 #define FS_PATH_INLINE_SIZE \
72 (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
75 /* reused for each extent */
77 struct btrfs_root *root;
84 #define SEND_MAX_NAME_CACHE_SIZE 256
87 * Limit the root_ids array of struct backref_cache_entry to 17 elements.
88 * This makes the size of a cache entry to be exactly 192 bytes on x86_64, which
89 * can be satisfied from the kmalloc-192 slab, without wasting any space.
90 * The most common case is to have a single root for cloning, which corresponds
91 * to the send root. Having the user specify more than 16 clone roots is not
92 * common, and in such rare cases we simply don't use caching if the number of
93 * cloning roots that lead down to a leaf is more than 17.
95 #define SEND_MAX_BACKREF_CACHE_ROOTS 17
98 * Max number of entries in the cache.
99 * With SEND_MAX_BACKREF_CACHE_ROOTS as 17, the size in bytes, excluding
100 * maple tree's internal nodes, is 24K.
102 #define SEND_MAX_BACKREF_CACHE_SIZE 128
105 * A backref cache entry maps a leaf to a list of IDs of roots from which the
106 * leaf is accessible and we can use for clone operations.
107 * With SEND_MAX_BACKREF_CACHE_ROOTS as 12, each cache entry is 128 bytes (on
110 struct backref_cache_entry {
111 struct btrfs_lru_cache_entry entry;
112 u64 root_ids[SEND_MAX_BACKREF_CACHE_ROOTS];
113 /* Number of valid elements in the root_ids array. */
117 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
118 static_assert(offsetof(struct backref_cache_entry, entry) == 0);
121 * Max number of entries in the cache that stores directories that were already
122 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
123 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
124 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
126 #define SEND_MAX_DIR_CREATED_CACHE_SIZE 64
129 * Max number of entries in the cache that stores directories that were already
130 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
131 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
132 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
134 #define SEND_MAX_DIR_UTIMES_CACHE_SIZE 64
137 struct file *send_filp;
143 * Whether BTRFS_SEND_A_DATA attribute was already added to current
144 * command (since protocol v2, data must be the last attribute).
147 struct page **send_buf_pages;
148 u64 flags; /* 'flags' member of btrfs_ioctl_send_args is u64 */
149 /* Protocol version compatibility requested */
152 struct btrfs_root *send_root;
153 struct btrfs_root *parent_root;
154 struct clone_root *clone_roots;
157 /* current state of the compare_tree call */
158 struct btrfs_path *left_path;
159 struct btrfs_path *right_path;
160 struct btrfs_key *cmp_key;
163 * Keep track of the generation of the last transaction that was used
164 * for relocating a block group. This is periodically checked in order
165 * to detect if a relocation happened since the last check, so that we
166 * don't operate on stale extent buffers for nodes (level >= 1) or on
167 * stale disk_bytenr values of file extent items.
169 u64 last_reloc_trans;
172 * infos of the currently processed inode. In case of deleted inodes,
173 * these are the values from the deleted inode.
180 u64 cur_inode_last_extent;
181 u64 cur_inode_next_write_offset;
183 bool cur_inode_new_gen;
184 bool cur_inode_deleted;
185 bool ignore_cur_inode;
186 bool cur_inode_needs_verity;
187 void *verity_descriptor;
191 struct list_head new_refs;
192 struct list_head deleted_refs;
194 struct btrfs_lru_cache name_cache;
197 * The inode we are currently processing. It's not NULL only when we
198 * need to issue write commands for data extents from this inode.
200 struct inode *cur_inode;
201 struct file_ra_state ra;
202 u64 page_cache_clear_start;
203 bool clean_page_cache;
206 * We process inodes by their increasing order, so if before an
207 * incremental send we reverse the parent/child relationship of
208 * directories such that a directory with a lower inode number was
209 * the parent of a directory with a higher inode number, and the one
210 * becoming the new parent got renamed too, we can't rename/move the
211 * directory with lower inode number when we finish processing it - we
212 * must process the directory with higher inode number first, then
213 * rename/move it and then rename/move the directory with lower inode
214 * number. Example follows.
216 * Tree state when the first send was performed:
228 * Tree state when the second (incremental) send is performed:
237 * The sequence of steps that lead to the second state was:
239 * mv /a/b/c/d /a/b/c2/d2
240 * mv /a/b/c /a/b/c2/d2/cc
242 * "c" has lower inode number, but we can't move it (2nd mv operation)
243 * before we move "d", which has higher inode number.
245 * So we just memorize which move/rename operations must be performed
246 * later when their respective parent is processed and moved/renamed.
249 /* Indexed by parent directory inode number. */
250 struct rb_root pending_dir_moves;
253 * Reverse index, indexed by the inode number of a directory that
254 * is waiting for the move/rename of its immediate parent before its
255 * own move/rename can be performed.
257 struct rb_root waiting_dir_moves;
260 * A directory that is going to be rm'ed might have a child directory
261 * which is in the pending directory moves index above. In this case,
262 * the directory can only be removed after the move/rename of its child
263 * is performed. Example:
283 * Sequence of steps that lead to the send snapshot:
284 * rm -f /a/b/c/foo.txt
286 * mv /a/b/c/x /a/b/YY
289 * When the child is processed, its move/rename is delayed until its
290 * parent is processed (as explained above), but all other operations
291 * like update utimes, chown, chgrp, etc, are performed and the paths
292 * that it uses for those operations must use the orphanized name of
293 * its parent (the directory we're going to rm later), so we need to
294 * memorize that name.
296 * Indexed by the inode number of the directory to be deleted.
298 struct rb_root orphan_dirs;
300 struct rb_root rbtree_new_refs;
301 struct rb_root rbtree_deleted_refs;
303 struct btrfs_lru_cache backref_cache;
304 u64 backref_cache_last_reloc_trans;
306 struct btrfs_lru_cache dir_created_cache;
307 struct btrfs_lru_cache dir_utimes_cache;
310 struct pending_dir_move {
312 struct list_head list;
316 struct list_head update_refs;
319 struct waiting_dir_move {
323 * There might be some directory that could not be removed because it
324 * was waiting for this directory inode to be moved first. Therefore
325 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
332 struct orphan_dir_info {
336 u64 last_dir_index_offset;
337 u64 dir_high_seq_ino;
340 struct name_cache_entry {
342 * The key in the entry is an inode number, and the generation matches
343 * the inode's generation.
345 struct btrfs_lru_cache_entry entry;
349 int need_later_update;
354 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
355 static_assert(offsetof(struct name_cache_entry, entry) == 0);
358 #define ADVANCE_ONLY_NEXT -1
360 enum btrfs_compare_tree_result {
361 BTRFS_COMPARE_TREE_NEW,
362 BTRFS_COMPARE_TREE_DELETED,
363 BTRFS_COMPARE_TREE_CHANGED,
364 BTRFS_COMPARE_TREE_SAME,
368 static void inconsistent_snapshot_error(struct send_ctx *sctx,
369 enum btrfs_compare_tree_result result,
372 const char *result_string;
375 case BTRFS_COMPARE_TREE_NEW:
376 result_string = "new";
378 case BTRFS_COMPARE_TREE_DELETED:
379 result_string = "deleted";
381 case BTRFS_COMPARE_TREE_CHANGED:
382 result_string = "updated";
384 case BTRFS_COMPARE_TREE_SAME:
386 result_string = "unchanged";
390 result_string = "unexpected";
393 btrfs_err(sctx->send_root->fs_info,
394 "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
395 result_string, what, sctx->cmp_key->objectid,
396 sctx->send_root->root_key.objectid,
398 sctx->parent_root->root_key.objectid : 0));
402 static bool proto_cmd_ok(const struct send_ctx *sctx, int cmd)
404 switch (sctx->proto) {
405 case 1: return cmd <= BTRFS_SEND_C_MAX_V1;
406 case 2: return cmd <= BTRFS_SEND_C_MAX_V2;
407 case 3: return cmd <= BTRFS_SEND_C_MAX_V3;
408 default: return false;
412 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
414 static struct waiting_dir_move *
415 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
417 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);
419 static int need_send_hole(struct send_ctx *sctx)
421 return (sctx->parent_root && !sctx->cur_inode_new &&
422 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
423 S_ISREG(sctx->cur_inode_mode));
426 static void fs_path_reset(struct fs_path *p)
429 p->start = p->buf + p->buf_len - 1;
439 static struct fs_path *fs_path_alloc(void)
443 p = kmalloc(sizeof(*p), GFP_KERNEL);
447 p->buf = p->inline_buf;
448 p->buf_len = FS_PATH_INLINE_SIZE;
453 static struct fs_path *fs_path_alloc_reversed(void)
465 static void fs_path_free(struct fs_path *p)
469 if (p->buf != p->inline_buf)
474 static int fs_path_len(struct fs_path *p)
476 return p->end - p->start;
479 static int fs_path_ensure_buf(struct fs_path *p, int len)
487 if (p->buf_len >= len)
490 if (len > PATH_MAX) {
495 path_len = p->end - p->start;
496 old_buf_len = p->buf_len;
499 * Allocate to the next largest kmalloc bucket size, to let
500 * the fast path happen most of the time.
502 len = kmalloc_size_roundup(len);
504 * First time the inline_buf does not suffice
506 if (p->buf == p->inline_buf) {
507 tmp_buf = kmalloc(len, GFP_KERNEL);
509 memcpy(tmp_buf, p->buf, old_buf_len);
511 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
519 tmp_buf = p->buf + old_buf_len - path_len - 1;
520 p->end = p->buf + p->buf_len - 1;
521 p->start = p->end - path_len;
522 memmove(p->start, tmp_buf, path_len + 1);
525 p->end = p->start + path_len;
530 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
536 new_len = p->end - p->start + name_len;
537 if (p->start != p->end)
539 ret = fs_path_ensure_buf(p, new_len);
544 if (p->start != p->end)
546 p->start -= name_len;
547 *prepared = p->start;
549 if (p->start != p->end)
560 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
565 ret = fs_path_prepare_for_add(p, name_len, &prepared);
568 memcpy(prepared, name, name_len);
574 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
579 ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
582 memcpy(prepared, p2->start, p2->end - p2->start);
588 static int fs_path_add_from_extent_buffer(struct fs_path *p,
589 struct extent_buffer *eb,
590 unsigned long off, int len)
595 ret = fs_path_prepare_for_add(p, len, &prepared);
599 read_extent_buffer(eb, prepared, off, len);
605 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
607 p->reversed = from->reversed;
610 return fs_path_add_path(p, from);
613 static void fs_path_unreverse(struct fs_path *p)
622 len = p->end - p->start;
624 p->end = p->start + len;
625 memmove(p->start, tmp, len + 1);
629 static struct btrfs_path *alloc_path_for_send(void)
631 struct btrfs_path *path;
633 path = btrfs_alloc_path();
636 path->search_commit_root = 1;
637 path->skip_locking = 1;
638 path->need_commit_sem = 1;
642 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
648 ret = kernel_write(filp, buf + pos, len - pos, off);
659 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
661 struct btrfs_tlv_header *hdr;
662 int total_len = sizeof(*hdr) + len;
663 int left = sctx->send_max_size - sctx->send_size;
665 if (WARN_ON_ONCE(sctx->put_data))
668 if (unlikely(left < total_len))
671 hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
672 put_unaligned_le16(attr, &hdr->tlv_type);
673 put_unaligned_le16(len, &hdr->tlv_len);
674 memcpy(hdr + 1, data, len);
675 sctx->send_size += total_len;
680 #define TLV_PUT_DEFINE_INT(bits) \
681 static int tlv_put_u##bits(struct send_ctx *sctx, \
682 u##bits attr, u##bits value) \
684 __le##bits __tmp = cpu_to_le##bits(value); \
685 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp)); \
688 TLV_PUT_DEFINE_INT(8)
689 TLV_PUT_DEFINE_INT(32)
690 TLV_PUT_DEFINE_INT(64)
692 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
693 const char *str, int len)
697 return tlv_put(sctx, attr, str, len);
700 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
703 return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
706 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
707 struct extent_buffer *eb,
708 struct btrfs_timespec *ts)
710 struct btrfs_timespec bts;
711 read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
712 return tlv_put(sctx, attr, &bts, sizeof(bts));
716 #define TLV_PUT(sctx, attrtype, data, attrlen) \
718 ret = tlv_put(sctx, attrtype, data, attrlen); \
720 goto tlv_put_failure; \
723 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
725 ret = tlv_put_u##bits(sctx, attrtype, value); \
727 goto tlv_put_failure; \
730 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
731 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
732 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
733 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
734 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
736 ret = tlv_put_string(sctx, attrtype, str, len); \
738 goto tlv_put_failure; \
740 #define TLV_PUT_PATH(sctx, attrtype, p) \
742 ret = tlv_put_string(sctx, attrtype, p->start, \
743 p->end - p->start); \
745 goto tlv_put_failure; \
747 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
749 ret = tlv_put_uuid(sctx, attrtype, uuid); \
751 goto tlv_put_failure; \
753 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
755 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
757 goto tlv_put_failure; \
760 static int send_header(struct send_ctx *sctx)
762 struct btrfs_stream_header hdr;
764 strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
765 hdr.version = cpu_to_le32(sctx->proto);
766 return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
771 * For each command/item we want to send to userspace, we call this function.
773 static int begin_cmd(struct send_ctx *sctx, int cmd)
775 struct btrfs_cmd_header *hdr;
777 if (WARN_ON(!sctx->send_buf))
780 BUG_ON(sctx->send_size);
782 sctx->send_size += sizeof(*hdr);
783 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
784 put_unaligned_le16(cmd, &hdr->cmd);
789 static int send_cmd(struct send_ctx *sctx)
792 struct btrfs_cmd_header *hdr;
795 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
796 put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
797 put_unaligned_le32(0, &hdr->crc);
799 crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
800 put_unaligned_le32(crc, &hdr->crc);
802 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
806 sctx->put_data = false;
812 * Sends a move instruction to user space
814 static int send_rename(struct send_ctx *sctx,
815 struct fs_path *from, struct fs_path *to)
817 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
820 btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
822 ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
826 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
827 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
829 ret = send_cmd(sctx);
837 * Sends a link instruction to user space
839 static int send_link(struct send_ctx *sctx,
840 struct fs_path *path, struct fs_path *lnk)
842 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
845 btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
847 ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
851 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
852 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
854 ret = send_cmd(sctx);
862 * Sends an unlink instruction to user space
864 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
866 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
869 btrfs_debug(fs_info, "send_unlink %s", path->start);
871 ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
875 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
877 ret = send_cmd(sctx);
885 * Sends a rmdir instruction to user space
887 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
889 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
892 btrfs_debug(fs_info, "send_rmdir %s", path->start);
894 ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
898 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
900 ret = send_cmd(sctx);
907 struct btrfs_inode_info {
919 * Helper function to retrieve some fields from an inode item.
921 static int get_inode_info(struct btrfs_root *root, u64 ino,
922 struct btrfs_inode_info *info)
925 struct btrfs_path *path;
926 struct btrfs_inode_item *ii;
927 struct btrfs_key key;
929 path = alloc_path_for_send();
934 key.type = BTRFS_INODE_ITEM_KEY;
936 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
946 ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
947 struct btrfs_inode_item);
948 info->size = btrfs_inode_size(path->nodes[0], ii);
949 info->gen = btrfs_inode_generation(path->nodes[0], ii);
950 info->mode = btrfs_inode_mode(path->nodes[0], ii);
951 info->uid = btrfs_inode_uid(path->nodes[0], ii);
952 info->gid = btrfs_inode_gid(path->nodes[0], ii);
953 info->rdev = btrfs_inode_rdev(path->nodes[0], ii);
954 info->nlink = btrfs_inode_nlink(path->nodes[0], ii);
956 * Transfer the unchanged u64 value of btrfs_inode_item::flags, that's
957 * otherwise logically split to 32/32 parts.
959 info->fileattr = btrfs_inode_flags(path->nodes[0], ii);
962 btrfs_free_path(path);
966 static int get_inode_gen(struct btrfs_root *root, u64 ino, u64 *gen)
969 struct btrfs_inode_info info = { 0 };
973 ret = get_inode_info(root, ino, &info);
978 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
983 * Helper function to iterate the entries in ONE btrfs_inode_ref or
984 * btrfs_inode_extref.
985 * The iterate callback may return a non zero value to stop iteration. This can
986 * be a negative value for error codes or 1 to simply stop it.
988 * path must point to the INODE_REF or INODE_EXTREF when called.
990 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
991 struct btrfs_key *found_key, int resolve,
992 iterate_inode_ref_t iterate, void *ctx)
994 struct extent_buffer *eb = path->nodes[0];
995 struct btrfs_inode_ref *iref;
996 struct btrfs_inode_extref *extref;
997 struct btrfs_path *tmp_path;
1001 int slot = path->slots[0];
1008 unsigned long name_off;
1009 unsigned long elem_size;
1012 p = fs_path_alloc_reversed();
1016 tmp_path = alloc_path_for_send();
1023 if (found_key->type == BTRFS_INODE_REF_KEY) {
1024 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
1025 struct btrfs_inode_ref);
1026 total = btrfs_item_size(eb, slot);
1027 elem_size = sizeof(*iref);
1029 ptr = btrfs_item_ptr_offset(eb, slot);
1030 total = btrfs_item_size(eb, slot);
1031 elem_size = sizeof(*extref);
1034 while (cur < total) {
1037 if (found_key->type == BTRFS_INODE_REF_KEY) {
1038 iref = (struct btrfs_inode_ref *)(ptr + cur);
1039 name_len = btrfs_inode_ref_name_len(eb, iref);
1040 name_off = (unsigned long)(iref + 1);
1041 index = btrfs_inode_ref_index(eb, iref);
1042 dir = found_key->offset;
1044 extref = (struct btrfs_inode_extref *)(ptr + cur);
1045 name_len = btrfs_inode_extref_name_len(eb, extref);
1046 name_off = (unsigned long)&extref->name;
1047 index = btrfs_inode_extref_index(eb, extref);
1048 dir = btrfs_inode_extref_parent(eb, extref);
1052 start = btrfs_ref_to_path(root, tmp_path, name_len,
1054 p->buf, p->buf_len);
1055 if (IS_ERR(start)) {
1056 ret = PTR_ERR(start);
1059 if (start < p->buf) {
1060 /* overflow , try again with larger buffer */
1061 ret = fs_path_ensure_buf(p,
1062 p->buf_len + p->buf - start);
1065 start = btrfs_ref_to_path(root, tmp_path,
1068 p->buf, p->buf_len);
1069 if (IS_ERR(start)) {
1070 ret = PTR_ERR(start);
1073 BUG_ON(start < p->buf);
1077 ret = fs_path_add_from_extent_buffer(p, eb, name_off,
1083 cur += elem_size + name_len;
1084 ret = iterate(num, dir, index, p, ctx);
1091 btrfs_free_path(tmp_path);
1096 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
1097 const char *name, int name_len,
1098 const char *data, int data_len,
1102 * Helper function to iterate the entries in ONE btrfs_dir_item.
1103 * The iterate callback may return a non zero value to stop iteration. This can
1104 * be a negative value for error codes or 1 to simply stop it.
1106 * path must point to the dir item when called.
1108 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1109 iterate_dir_item_t iterate, void *ctx)
1112 struct extent_buffer *eb;
1113 struct btrfs_dir_item *di;
1114 struct btrfs_key di_key;
1126 * Start with a small buffer (1 page). If later we end up needing more
1127 * space, which can happen for xattrs on a fs with a leaf size greater
1128 * then the page size, attempt to increase the buffer. Typically xattr
1132 buf = kmalloc(buf_len, GFP_KERNEL);
1138 eb = path->nodes[0];
1139 slot = path->slots[0];
1140 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1143 total = btrfs_item_size(eb, slot);
1146 while (cur < total) {
1147 name_len = btrfs_dir_name_len(eb, di);
1148 data_len = btrfs_dir_data_len(eb, di);
1149 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1151 if (btrfs_dir_ftype(eb, di) == BTRFS_FT_XATTR) {
1152 if (name_len > XATTR_NAME_MAX) {
1153 ret = -ENAMETOOLONG;
1156 if (name_len + data_len >
1157 BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1165 if (name_len + data_len > PATH_MAX) {
1166 ret = -ENAMETOOLONG;
1171 if (name_len + data_len > buf_len) {
1172 buf_len = name_len + data_len;
1173 if (is_vmalloc_addr(buf)) {
1177 char *tmp = krealloc(buf, buf_len,
1178 GFP_KERNEL | __GFP_NOWARN);
1185 buf = kvmalloc(buf_len, GFP_KERNEL);
1193 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1194 name_len + data_len);
1196 len = sizeof(*di) + name_len + data_len;
1197 di = (struct btrfs_dir_item *)((char *)di + len);
1200 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1217 static int __copy_first_ref(int num, u64 dir, int index,
1218 struct fs_path *p, void *ctx)
1221 struct fs_path *pt = ctx;
1223 ret = fs_path_copy(pt, p);
1227 /* we want the first only */
1232 * Retrieve the first path of an inode. If an inode has more then one
1233 * ref/hardlink, this is ignored.
1235 static int get_inode_path(struct btrfs_root *root,
1236 u64 ino, struct fs_path *path)
1239 struct btrfs_key key, found_key;
1240 struct btrfs_path *p;
1242 p = alloc_path_for_send();
1246 fs_path_reset(path);
1249 key.type = BTRFS_INODE_REF_KEY;
1252 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1259 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1260 if (found_key.objectid != ino ||
1261 (found_key.type != BTRFS_INODE_REF_KEY &&
1262 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1267 ret = iterate_inode_ref(root, p, &found_key, 1,
1268 __copy_first_ref, path);
1278 struct backref_ctx {
1279 struct send_ctx *sctx;
1281 /* number of total found references */
1285 * used for clones found in send_root. clones found behind cur_objectid
1286 * and cur_offset are not considered as allowed clones.
1291 /* may be truncated in case it's the last extent in a file */
1294 /* The bytenr the file extent item we are processing refers to. */
1296 /* The owner (root id) of the data backref for the current extent. */
1298 /* The offset of the data backref for the current extent. */
1302 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1304 u64 root = (u64)(uintptr_t)key;
1305 const struct clone_root *cr = elt;
1307 if (root < cr->root->root_key.objectid)
1309 if (root > cr->root->root_key.objectid)
1314 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1316 const struct clone_root *cr1 = e1;
1317 const struct clone_root *cr2 = e2;
1319 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1321 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1327 * Called for every backref that is found for the current extent.
1328 * Results are collected in sctx->clone_roots->ino/offset.
1330 static int iterate_backrefs(u64 ino, u64 offset, u64 num_bytes, u64 root_id,
1333 struct backref_ctx *bctx = ctx_;
1334 struct clone_root *clone_root;
1336 /* First check if the root is in the list of accepted clone sources */
1337 clone_root = bsearch((void *)(uintptr_t)root_id, bctx->sctx->clone_roots,
1338 bctx->sctx->clone_roots_cnt,
1339 sizeof(struct clone_root),
1340 __clone_root_cmp_bsearch);
1344 /* This is our own reference, bail out as we can't clone from it. */
1345 if (clone_root->root == bctx->sctx->send_root &&
1346 ino == bctx->cur_objectid &&
1347 offset == bctx->cur_offset)
1351 * Make sure we don't consider clones from send_root that are
1352 * behind the current inode/offset.
1354 if (clone_root->root == bctx->sctx->send_root) {
1356 * If the source inode was not yet processed we can't issue a
1357 * clone operation, as the source extent does not exist yet at
1358 * the destination of the stream.
1360 if (ino > bctx->cur_objectid)
1363 * We clone from the inode currently being sent as long as the
1364 * source extent is already processed, otherwise we could try
1365 * to clone from an extent that does not exist yet at the
1366 * destination of the stream.
1368 if (ino == bctx->cur_objectid &&
1369 offset + bctx->extent_len >
1370 bctx->sctx->cur_inode_next_write_offset)
1375 clone_root->found_ref = true;
1378 * If the given backref refers to a file extent item with a larger
1379 * number of bytes than what we found before, use the new one so that
1380 * we clone more optimally and end up doing less writes and getting
1381 * less exclusive, non-shared extents at the destination.
1383 if (num_bytes > clone_root->num_bytes) {
1384 clone_root->ino = ino;
1385 clone_root->offset = offset;
1386 clone_root->num_bytes = num_bytes;
1389 * Found a perfect candidate, so there's no need to continue
1392 if (num_bytes >= bctx->extent_len)
1393 return BTRFS_ITERATE_EXTENT_INODES_STOP;
1399 static bool lookup_backref_cache(u64 leaf_bytenr, void *ctx,
1400 const u64 **root_ids_ret, int *root_count_ret)
1402 struct backref_ctx *bctx = ctx;
1403 struct send_ctx *sctx = bctx->sctx;
1404 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1405 const u64 key = leaf_bytenr >> fs_info->sectorsize_bits;
1406 struct btrfs_lru_cache_entry *raw_entry;
1407 struct backref_cache_entry *entry;
1409 if (btrfs_lru_cache_size(&sctx->backref_cache) == 0)
1413 * If relocation happened since we first filled the cache, then we must
1414 * empty the cache and can not use it, because even though we operate on
1415 * read-only roots, their leaves and nodes may have been reallocated and
1416 * now be used for different nodes/leaves of the same tree or some other
1419 * We are called from iterate_extent_inodes() while either holding a
1420 * transaction handle or holding fs_info->commit_root_sem, so no need
1421 * to take any lock here.
1423 if (fs_info->last_reloc_trans > sctx->backref_cache_last_reloc_trans) {
1424 btrfs_lru_cache_clear(&sctx->backref_cache);
1428 raw_entry = btrfs_lru_cache_lookup(&sctx->backref_cache, key, 0);
1432 entry = container_of(raw_entry, struct backref_cache_entry, entry);
1433 *root_ids_ret = entry->root_ids;
1434 *root_count_ret = entry->num_roots;
1439 static void store_backref_cache(u64 leaf_bytenr, const struct ulist *root_ids,
1442 struct backref_ctx *bctx = ctx;
1443 struct send_ctx *sctx = bctx->sctx;
1444 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1445 struct backref_cache_entry *new_entry;
1446 struct ulist_iterator uiter;
1447 struct ulist_node *node;
1451 * We're called while holding a transaction handle or while holding
1452 * fs_info->commit_root_sem (at iterate_extent_inodes()), so must do a
1455 new_entry = kmalloc(sizeof(struct backref_cache_entry), GFP_NOFS);
1456 /* No worries, cache is optional. */
1460 new_entry->entry.key = leaf_bytenr >> fs_info->sectorsize_bits;
1461 new_entry->entry.gen = 0;
1462 new_entry->num_roots = 0;
1463 ULIST_ITER_INIT(&uiter);
1464 while ((node = ulist_next(root_ids, &uiter)) != NULL) {
1465 const u64 root_id = node->val;
1466 struct clone_root *root;
1468 root = bsearch((void *)(uintptr_t)root_id, sctx->clone_roots,
1469 sctx->clone_roots_cnt, sizeof(struct clone_root),
1470 __clone_root_cmp_bsearch);
1474 /* Too many roots, just exit, no worries as caching is optional. */
1475 if (new_entry->num_roots >= SEND_MAX_BACKREF_CACHE_ROOTS) {
1480 new_entry->root_ids[new_entry->num_roots] = root_id;
1481 new_entry->num_roots++;
1485 * We may have not added any roots to the new cache entry, which means
1486 * none of the roots is part of the list of roots from which we are
1487 * allowed to clone. Cache the new entry as it's still useful to avoid
1488 * backref walking to determine which roots have a path to the leaf.
1490 * Also use GFP_NOFS because we're called while holding a transaction
1491 * handle or while holding fs_info->commit_root_sem.
1493 ret = btrfs_lru_cache_store(&sctx->backref_cache, &new_entry->entry,
1495 ASSERT(ret == 0 || ret == -ENOMEM);
1497 /* Caching is optional, no worries. */
1503 * We are called from iterate_extent_inodes() while either holding a
1504 * transaction handle or holding fs_info->commit_root_sem, so no need
1505 * to take any lock here.
1507 if (btrfs_lru_cache_size(&sctx->backref_cache) == 1)
1508 sctx->backref_cache_last_reloc_trans = fs_info->last_reloc_trans;
1511 static int check_extent_item(u64 bytenr, const struct btrfs_extent_item *ei,
1512 const struct extent_buffer *leaf, void *ctx)
1514 const u64 refs = btrfs_extent_refs(leaf, ei);
1515 const struct backref_ctx *bctx = ctx;
1516 const struct send_ctx *sctx = bctx->sctx;
1518 if (bytenr == bctx->bytenr) {
1519 const u64 flags = btrfs_extent_flags(leaf, ei);
1521 if (WARN_ON(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
1525 * If we have only one reference and only the send root as a
1526 * clone source - meaning no clone roots were given in the
1527 * struct btrfs_ioctl_send_args passed to the send ioctl - then
1528 * it's our reference and there's no point in doing backref
1529 * walking which is expensive, so exit early.
1531 if (refs == 1 && sctx->clone_roots_cnt == 1)
1536 * Backreference walking (iterate_extent_inodes() below) is currently
1537 * too expensive when an extent has a large number of references, both
1538 * in time spent and used memory. So for now just fallback to write
1539 * operations instead of clone operations when an extent has more than
1540 * a certain amount of references.
1542 if (refs > SEND_MAX_EXTENT_REFS)
1548 static bool skip_self_data_ref(u64 root, u64 ino, u64 offset, void *ctx)
1550 const struct backref_ctx *bctx = ctx;
1552 if (ino == bctx->cur_objectid &&
1553 root == bctx->backref_owner &&
1554 offset == bctx->backref_offset)
1561 * Given an inode, offset and extent item, it finds a good clone for a clone
1562 * instruction. Returns -ENOENT when none could be found. The function makes
1563 * sure that the returned clone is usable at the point where sending is at the
1564 * moment. This means, that no clones are accepted which lie behind the current
1567 * path must point to the extent item when called.
1569 static int find_extent_clone(struct send_ctx *sctx,
1570 struct btrfs_path *path,
1571 u64 ino, u64 data_offset,
1573 struct clone_root **found)
1575 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1581 struct btrfs_file_extent_item *fi;
1582 struct extent_buffer *eb = path->nodes[0];
1583 struct backref_ctx backref_ctx = { 0 };
1584 struct btrfs_backref_walk_ctx backref_walk_ctx = { 0 };
1585 struct clone_root *cur_clone_root;
1590 * With fallocate we can get prealloc extents beyond the inode's i_size,
1591 * so we don't do anything here because clone operations can not clone
1592 * to a range beyond i_size without increasing the i_size of the
1593 * destination inode.
1595 if (data_offset >= ino_size)
1598 fi = btrfs_item_ptr(eb, path->slots[0], struct btrfs_file_extent_item);
1599 extent_type = btrfs_file_extent_type(eb, fi);
1600 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1603 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1607 compressed = btrfs_file_extent_compression(eb, fi);
1608 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1609 logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1612 * Setup the clone roots.
1614 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1615 cur_clone_root = sctx->clone_roots + i;
1616 cur_clone_root->ino = (u64)-1;
1617 cur_clone_root->offset = 0;
1618 cur_clone_root->num_bytes = 0;
1619 cur_clone_root->found_ref = false;
1622 backref_ctx.sctx = sctx;
1623 backref_ctx.cur_objectid = ino;
1624 backref_ctx.cur_offset = data_offset;
1625 backref_ctx.bytenr = disk_byte;
1627 * Use the header owner and not the send root's id, because in case of a
1628 * snapshot we can have shared subtrees.
1630 backref_ctx.backref_owner = btrfs_header_owner(eb);
1631 backref_ctx.backref_offset = data_offset - btrfs_file_extent_offset(eb, fi);
1634 * The last extent of a file may be too large due to page alignment.
1635 * We need to adjust extent_len in this case so that the checks in
1636 * iterate_backrefs() work.
1638 if (data_offset + num_bytes >= ino_size)
1639 backref_ctx.extent_len = ino_size - data_offset;
1641 backref_ctx.extent_len = num_bytes;
1644 * Now collect all backrefs.
1646 backref_walk_ctx.bytenr = disk_byte;
1647 if (compressed == BTRFS_COMPRESS_NONE)
1648 backref_walk_ctx.extent_item_pos = btrfs_file_extent_offset(eb, fi);
1649 backref_walk_ctx.fs_info = fs_info;
1650 backref_walk_ctx.cache_lookup = lookup_backref_cache;
1651 backref_walk_ctx.cache_store = store_backref_cache;
1652 backref_walk_ctx.indirect_ref_iterator = iterate_backrefs;
1653 backref_walk_ctx.check_extent_item = check_extent_item;
1654 backref_walk_ctx.user_ctx = &backref_ctx;
1657 * If have a single clone root, then it's the send root and we can tell
1658 * the backref walking code to skip our own backref and not resolve it,
1659 * since we can not use it for cloning - the source and destination
1660 * ranges can't overlap and in case the leaf is shared through a subtree
1661 * due to snapshots, we can't use those other roots since they are not
1662 * in the list of clone roots.
1664 if (sctx->clone_roots_cnt == 1)
1665 backref_walk_ctx.skip_data_ref = skip_self_data_ref;
1667 ret = iterate_extent_inodes(&backref_walk_ctx, true, iterate_backrefs,
1672 down_read(&fs_info->commit_root_sem);
1673 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
1675 * A transaction commit for a transaction in which block group
1676 * relocation was done just happened.
1677 * The disk_bytenr of the file extent item we processed is
1678 * possibly stale, referring to the extent's location before
1679 * relocation. So act as if we haven't found any clone sources
1680 * and fallback to write commands, which will read the correct
1681 * data from the new extent location. Otherwise we will fail
1682 * below because we haven't found our own back reference or we
1683 * could be getting incorrect sources in case the old extent
1684 * was already reallocated after the relocation.
1686 up_read(&fs_info->commit_root_sem);
1689 up_read(&fs_info->commit_root_sem);
1691 btrfs_debug(fs_info,
1692 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1693 data_offset, ino, num_bytes, logical);
1695 if (!backref_ctx.found) {
1696 btrfs_debug(fs_info, "no clones found");
1700 cur_clone_root = NULL;
1701 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1702 struct clone_root *clone_root = &sctx->clone_roots[i];
1704 if (!clone_root->found_ref)
1708 * Choose the root from which we can clone more bytes, to
1709 * minimize write operations and therefore have more extent
1710 * sharing at the destination (the same as in the source).
1712 if (!cur_clone_root ||
1713 clone_root->num_bytes > cur_clone_root->num_bytes) {
1714 cur_clone_root = clone_root;
1717 * We found an optimal clone candidate (any inode from
1718 * any root is fine), so we're done.
1720 if (clone_root->num_bytes >= backref_ctx.extent_len)
1725 if (cur_clone_root) {
1726 *found = cur_clone_root;
1735 static int read_symlink(struct btrfs_root *root,
1737 struct fs_path *dest)
1740 struct btrfs_path *path;
1741 struct btrfs_key key;
1742 struct btrfs_file_extent_item *ei;
1748 path = alloc_path_for_send();
1753 key.type = BTRFS_EXTENT_DATA_KEY;
1755 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1760 * An empty symlink inode. Can happen in rare error paths when
1761 * creating a symlink (transaction committed before the inode
1762 * eviction handler removed the symlink inode items and a crash
1763 * happened in between or the subvol was snapshoted in between).
1764 * Print an informative message to dmesg/syslog so that the user
1765 * can delete the symlink.
1767 btrfs_err(root->fs_info,
1768 "Found empty symlink inode %llu at root %llu",
1769 ino, root->root_key.objectid);
1774 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1775 struct btrfs_file_extent_item);
1776 type = btrfs_file_extent_type(path->nodes[0], ei);
1777 compression = btrfs_file_extent_compression(path->nodes[0], ei);
1778 BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1779 BUG_ON(compression);
1781 off = btrfs_file_extent_inline_start(ei);
1782 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1784 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1787 btrfs_free_path(path);
1792 * Helper function to generate a file name that is unique in the root of
1793 * send_root and parent_root. This is used to generate names for orphan inodes.
1795 static int gen_unique_name(struct send_ctx *sctx,
1797 struct fs_path *dest)
1800 struct btrfs_path *path;
1801 struct btrfs_dir_item *di;
1806 path = alloc_path_for_send();
1811 struct fscrypt_str tmp_name;
1813 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1815 ASSERT(len < sizeof(tmp));
1816 tmp_name.name = tmp;
1817 tmp_name.len = strlen(tmp);
1819 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1820 path, BTRFS_FIRST_FREE_OBJECTID,
1822 btrfs_release_path(path);
1828 /* not unique, try again */
1833 if (!sctx->parent_root) {
1839 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1840 path, BTRFS_FIRST_FREE_OBJECTID,
1842 btrfs_release_path(path);
1848 /* not unique, try again */
1856 ret = fs_path_add(dest, tmp, strlen(tmp));
1859 btrfs_free_path(path);
1864 inode_state_no_change,
1865 inode_state_will_create,
1866 inode_state_did_create,
1867 inode_state_will_delete,
1868 inode_state_did_delete,
1871 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen,
1872 u64 *send_gen, u64 *parent_gen)
1879 struct btrfs_inode_info info;
1881 ret = get_inode_info(sctx->send_root, ino, &info);
1882 if (ret < 0 && ret != -ENOENT)
1884 left_ret = (info.nlink == 0) ? -ENOENT : ret;
1885 left_gen = info.gen;
1887 *send_gen = ((left_ret == -ENOENT) ? 0 : info.gen);
1889 if (!sctx->parent_root) {
1890 right_ret = -ENOENT;
1892 ret = get_inode_info(sctx->parent_root, ino, &info);
1893 if (ret < 0 && ret != -ENOENT)
1895 right_ret = (info.nlink == 0) ? -ENOENT : ret;
1896 right_gen = info.gen;
1898 *parent_gen = ((right_ret == -ENOENT) ? 0 : info.gen);
1901 if (!left_ret && !right_ret) {
1902 if (left_gen == gen && right_gen == gen) {
1903 ret = inode_state_no_change;
1904 } else if (left_gen == gen) {
1905 if (ino < sctx->send_progress)
1906 ret = inode_state_did_create;
1908 ret = inode_state_will_create;
1909 } else if (right_gen == gen) {
1910 if (ino < sctx->send_progress)
1911 ret = inode_state_did_delete;
1913 ret = inode_state_will_delete;
1917 } else if (!left_ret) {
1918 if (left_gen == gen) {
1919 if (ino < sctx->send_progress)
1920 ret = inode_state_did_create;
1922 ret = inode_state_will_create;
1926 } else if (!right_ret) {
1927 if (right_gen == gen) {
1928 if (ino < sctx->send_progress)
1929 ret = inode_state_did_delete;
1931 ret = inode_state_will_delete;
1943 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen,
1944 u64 *send_gen, u64 *parent_gen)
1948 if (ino == BTRFS_FIRST_FREE_OBJECTID)
1951 ret = get_cur_inode_state(sctx, ino, gen, send_gen, parent_gen);
1955 if (ret == inode_state_no_change ||
1956 ret == inode_state_did_create ||
1957 ret == inode_state_will_delete)
1967 * Helper function to lookup a dir item in a dir.
1969 static int lookup_dir_item_inode(struct btrfs_root *root,
1970 u64 dir, const char *name, int name_len,
1974 struct btrfs_dir_item *di;
1975 struct btrfs_key key;
1976 struct btrfs_path *path;
1977 struct fscrypt_str name_str = FSTR_INIT((char *)name, name_len);
1979 path = alloc_path_for_send();
1983 di = btrfs_lookup_dir_item(NULL, root, path, dir, &name_str, 0);
1984 if (IS_ERR_OR_NULL(di)) {
1985 ret = di ? PTR_ERR(di) : -ENOENT;
1988 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1989 if (key.type == BTRFS_ROOT_ITEM_KEY) {
1993 *found_inode = key.objectid;
1996 btrfs_free_path(path);
2001 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
2002 * generation of the parent dir and the name of the dir entry.
2004 static int get_first_ref(struct btrfs_root *root, u64 ino,
2005 u64 *dir, u64 *dir_gen, struct fs_path *name)
2008 struct btrfs_key key;
2009 struct btrfs_key found_key;
2010 struct btrfs_path *path;
2014 path = alloc_path_for_send();
2019 key.type = BTRFS_INODE_REF_KEY;
2022 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
2026 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2028 if (ret || found_key.objectid != ino ||
2029 (found_key.type != BTRFS_INODE_REF_KEY &&
2030 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
2035 if (found_key.type == BTRFS_INODE_REF_KEY) {
2036 struct btrfs_inode_ref *iref;
2037 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2038 struct btrfs_inode_ref);
2039 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
2040 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2041 (unsigned long)(iref + 1),
2043 parent_dir = found_key.offset;
2045 struct btrfs_inode_extref *extref;
2046 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2047 struct btrfs_inode_extref);
2048 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
2049 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2050 (unsigned long)&extref->name, len);
2051 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
2055 btrfs_release_path(path);
2058 ret = get_inode_gen(root, parent_dir, dir_gen);
2066 btrfs_free_path(path);
2070 static int is_first_ref(struct btrfs_root *root,
2072 const char *name, int name_len)
2075 struct fs_path *tmp_name;
2078 tmp_name = fs_path_alloc();
2082 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
2086 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
2091 ret = !memcmp(tmp_name->start, name, name_len);
2094 fs_path_free(tmp_name);
2099 * Used by process_recorded_refs to determine if a new ref would overwrite an
2100 * already existing ref. In case it detects an overwrite, it returns the
2101 * inode/gen in who_ino/who_gen.
2102 * When an overwrite is detected, process_recorded_refs does proper orphanizing
2103 * to make sure later references to the overwritten inode are possible.
2104 * Orphanizing is however only required for the first ref of an inode.
2105 * process_recorded_refs does an additional is_first_ref check to see if
2106 * orphanizing is really required.
2108 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2109 const char *name, int name_len,
2110 u64 *who_ino, u64 *who_gen, u64 *who_mode)
2113 u64 parent_root_dir_gen;
2114 u64 other_inode = 0;
2115 struct btrfs_inode_info info;
2117 if (!sctx->parent_root)
2120 ret = is_inode_existent(sctx, dir, dir_gen, NULL, &parent_root_dir_gen);
2125 * If we have a parent root we need to verify that the parent dir was
2126 * not deleted and then re-created, if it was then we have no overwrite
2127 * and we can just unlink this entry.
2129 * @parent_root_dir_gen was set to 0 if the inode does not exist in the
2132 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID &&
2133 parent_root_dir_gen != dir_gen)
2136 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
2144 * Check if the overwritten ref was already processed. If yes, the ref
2145 * was already unlinked/moved, so we can safely assume that we will not
2146 * overwrite anything at this point in time.
2148 if (other_inode > sctx->send_progress ||
2149 is_waiting_for_move(sctx, other_inode)) {
2150 ret = get_inode_info(sctx->parent_root, other_inode, &info);
2154 *who_ino = other_inode;
2155 *who_gen = info.gen;
2156 *who_mode = info.mode;
2164 * Checks if the ref was overwritten by an already processed inode. This is
2165 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
2166 * thus the orphan name needs be used.
2167 * process_recorded_refs also uses it to avoid unlinking of refs that were
2170 static int did_overwrite_ref(struct send_ctx *sctx,
2171 u64 dir, u64 dir_gen,
2172 u64 ino, u64 ino_gen,
2173 const char *name, int name_len)
2178 u64 send_root_dir_gen;
2180 if (!sctx->parent_root)
2183 ret = is_inode_existent(sctx, dir, dir_gen, &send_root_dir_gen, NULL);
2188 * @send_root_dir_gen was set to 0 if the inode does not exist in the
2191 if (dir != BTRFS_FIRST_FREE_OBJECTID && send_root_dir_gen != dir_gen)
2194 /* check if the ref was overwritten by another ref */
2195 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
2197 if (ret == -ENOENT) {
2198 /* was never and will never be overwritten */
2200 } else if (ret < 0) {
2204 if (ow_inode == ino) {
2205 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2209 /* It's the same inode, so no overwrite happened. */
2210 if (ow_gen == ino_gen)
2215 * We know that it is or will be overwritten. Check this now.
2216 * The current inode being processed might have been the one that caused
2217 * inode 'ino' to be orphanized, therefore check if ow_inode matches
2218 * the current inode being processed.
2220 if (ow_inode < sctx->send_progress)
2223 if (ino != sctx->cur_ino && ow_inode == sctx->cur_ino) {
2225 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2229 if (ow_gen == sctx->cur_inode_gen)
2237 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
2238 * that got overwritten. This is used by process_recorded_refs to determine
2239 * if it has to use the path as returned by get_cur_path or the orphan name.
2241 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
2244 struct fs_path *name = NULL;
2248 if (!sctx->parent_root)
2251 name = fs_path_alloc();
2255 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2259 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2260 name->start, fs_path_len(name));
2267 static inline struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2270 struct btrfs_lru_cache_entry *entry;
2272 entry = btrfs_lru_cache_lookup(&sctx->name_cache, ino, gen);
2276 return container_of(entry, struct name_cache_entry, entry);
2280 * Used by get_cur_path for each ref up to the root.
2281 * Returns 0 if it succeeded.
2282 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2283 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2284 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2285 * Returns <0 in case of error.
2287 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2291 struct fs_path *dest)
2295 struct name_cache_entry *nce;
2298 * First check if we already did a call to this function with the same
2299 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2300 * return the cached result.
2302 nce = name_cache_search(sctx, ino, gen);
2304 if (ino < sctx->send_progress && nce->need_later_update) {
2305 btrfs_lru_cache_remove(&sctx->name_cache, &nce->entry);
2308 *parent_ino = nce->parent_ino;
2309 *parent_gen = nce->parent_gen;
2310 ret = fs_path_add(dest, nce->name, nce->name_len);
2319 * If the inode is not existent yet, add the orphan name and return 1.
2320 * This should only happen for the parent dir that we determine in
2321 * record_new_ref_if_needed().
2323 ret = is_inode_existent(sctx, ino, gen, NULL, NULL);
2328 ret = gen_unique_name(sctx, ino, gen, dest);
2336 * Depending on whether the inode was already processed or not, use
2337 * send_root or parent_root for ref lookup.
2339 if (ino < sctx->send_progress)
2340 ret = get_first_ref(sctx->send_root, ino,
2341 parent_ino, parent_gen, dest);
2343 ret = get_first_ref(sctx->parent_root, ino,
2344 parent_ino, parent_gen, dest);
2349 * Check if the ref was overwritten by an inode's ref that was processed
2350 * earlier. If yes, treat as orphan and return 1.
2352 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2353 dest->start, dest->end - dest->start);
2357 fs_path_reset(dest);
2358 ret = gen_unique_name(sctx, ino, gen, dest);
2366 * Store the result of the lookup in the name cache.
2368 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2374 nce->entry.key = ino;
2375 nce->entry.gen = gen;
2376 nce->parent_ino = *parent_ino;
2377 nce->parent_gen = *parent_gen;
2378 nce->name_len = fs_path_len(dest);
2380 strcpy(nce->name, dest->start);
2382 if (ino < sctx->send_progress)
2383 nce->need_later_update = 0;
2385 nce->need_later_update = 1;
2387 nce_ret = btrfs_lru_cache_store(&sctx->name_cache, &nce->entry, GFP_KERNEL);
2398 * Magic happens here. This function returns the first ref to an inode as it
2399 * would look like while receiving the stream at this point in time.
2400 * We walk the path up to the root. For every inode in between, we check if it
2401 * was already processed/sent. If yes, we continue with the parent as found
2402 * in send_root. If not, we continue with the parent as found in parent_root.
2403 * If we encounter an inode that was deleted at this point in time, we use the
2404 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2405 * that were not created yet and overwritten inodes/refs.
2407 * When do we have orphan inodes:
2408 * 1. When an inode is freshly created and thus no valid refs are available yet
2409 * 2. When a directory lost all it's refs (deleted) but still has dir items
2410 * inside which were not processed yet (pending for move/delete). If anyone
2411 * tried to get the path to the dir items, it would get a path inside that
2413 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2414 * of an unprocessed inode. If in that case the first ref would be
2415 * overwritten, the overwritten inode gets "orphanized". Later when we
2416 * process this overwritten inode, it is restored at a new place by moving
2419 * sctx->send_progress tells this function at which point in time receiving
2422 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2423 struct fs_path *dest)
2426 struct fs_path *name = NULL;
2427 u64 parent_inode = 0;
2431 name = fs_path_alloc();
2438 fs_path_reset(dest);
2440 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2441 struct waiting_dir_move *wdm;
2443 fs_path_reset(name);
2445 if (is_waiting_for_rm(sctx, ino, gen)) {
2446 ret = gen_unique_name(sctx, ino, gen, name);
2449 ret = fs_path_add_path(dest, name);
2453 wdm = get_waiting_dir_move(sctx, ino);
2454 if (wdm && wdm->orphanized) {
2455 ret = gen_unique_name(sctx, ino, gen, name);
2458 ret = get_first_ref(sctx->parent_root, ino,
2459 &parent_inode, &parent_gen, name);
2461 ret = __get_cur_name_and_parent(sctx, ino, gen,
2471 ret = fs_path_add_path(dest, name);
2482 fs_path_unreverse(dest);
2487 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2489 static int send_subvol_begin(struct send_ctx *sctx)
2492 struct btrfs_root *send_root = sctx->send_root;
2493 struct btrfs_root *parent_root = sctx->parent_root;
2494 struct btrfs_path *path;
2495 struct btrfs_key key;
2496 struct btrfs_root_ref *ref;
2497 struct extent_buffer *leaf;
2501 path = btrfs_alloc_path();
2505 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2507 btrfs_free_path(path);
2511 key.objectid = send_root->root_key.objectid;
2512 key.type = BTRFS_ROOT_BACKREF_KEY;
2515 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2524 leaf = path->nodes[0];
2525 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2526 if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2527 key.objectid != send_root->root_key.objectid) {
2531 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2532 namelen = btrfs_root_ref_name_len(leaf, ref);
2533 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2534 btrfs_release_path(path);
2537 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2541 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2546 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2548 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2549 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2550 sctx->send_root->root_item.received_uuid);
2552 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2553 sctx->send_root->root_item.uuid);
2555 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2556 btrfs_root_ctransid(&sctx->send_root->root_item));
2558 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2559 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2560 parent_root->root_item.received_uuid);
2562 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2563 parent_root->root_item.uuid);
2564 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2565 btrfs_root_ctransid(&sctx->parent_root->root_item));
2568 ret = send_cmd(sctx);
2572 btrfs_free_path(path);
2577 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2579 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2583 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2585 p = fs_path_alloc();
2589 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2593 ret = get_cur_path(sctx, ino, gen, p);
2596 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2597 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2599 ret = send_cmd(sctx);
2607 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2609 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2613 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2615 p = fs_path_alloc();
2619 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2623 ret = get_cur_path(sctx, ino, gen, p);
2626 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2627 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2629 ret = send_cmd(sctx);
2637 static int send_fileattr(struct send_ctx *sctx, u64 ino, u64 gen, u64 fileattr)
2639 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2643 if (sctx->proto < 2)
2646 btrfs_debug(fs_info, "send_fileattr %llu fileattr=%llu", ino, fileattr);
2648 p = fs_path_alloc();
2652 ret = begin_cmd(sctx, BTRFS_SEND_C_FILEATTR);
2656 ret = get_cur_path(sctx, ino, gen, p);
2659 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2660 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILEATTR, fileattr);
2662 ret = send_cmd(sctx);
2670 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2672 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2676 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2679 p = fs_path_alloc();
2683 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2687 ret = get_cur_path(sctx, ino, gen, p);
2690 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2691 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2692 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2694 ret = send_cmd(sctx);
2702 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2704 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2706 struct fs_path *p = NULL;
2707 struct btrfs_inode_item *ii;
2708 struct btrfs_path *path = NULL;
2709 struct extent_buffer *eb;
2710 struct btrfs_key key;
2713 btrfs_debug(fs_info, "send_utimes %llu", ino);
2715 p = fs_path_alloc();
2719 path = alloc_path_for_send();
2726 key.type = BTRFS_INODE_ITEM_KEY;
2728 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2734 eb = path->nodes[0];
2735 slot = path->slots[0];
2736 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2738 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2742 ret = get_cur_path(sctx, ino, gen, p);
2745 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2746 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2747 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2748 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2749 if (sctx->proto >= 2)
2750 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_OTIME, eb, &ii->otime);
2752 ret = send_cmd(sctx);
2757 btrfs_free_path(path);
2762 * If the cache is full, we can't remove entries from it and do a call to
2763 * send_utimes() for each respective inode, because we might be finishing
2764 * processing an inode that is a directory and it just got renamed, and existing
2765 * entries in the cache may refer to inodes that have the directory in their
2766 * full path - in which case we would generate outdated paths (pre-rename)
2767 * for the inodes that the cache entries point to. Instead of prunning the
2768 * cache when inserting, do it after we finish processing each inode at
2769 * finish_inode_if_needed().
2771 static int cache_dir_utimes(struct send_ctx *sctx, u64 dir, u64 gen)
2773 struct btrfs_lru_cache_entry *entry;
2776 entry = btrfs_lru_cache_lookup(&sctx->dir_utimes_cache, dir, gen);
2780 /* Caching is optional, don't fail if we can't allocate memory. */
2781 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2783 return send_utimes(sctx, dir, gen);
2788 ret = btrfs_lru_cache_store(&sctx->dir_utimes_cache, entry, GFP_KERNEL);
2789 ASSERT(ret != -EEXIST);
2792 return send_utimes(sctx, dir, gen);
2798 static int trim_dir_utimes_cache(struct send_ctx *sctx)
2800 while (btrfs_lru_cache_size(&sctx->dir_utimes_cache) >
2801 SEND_MAX_DIR_UTIMES_CACHE_SIZE) {
2802 struct btrfs_lru_cache_entry *lru;
2805 lru = btrfs_lru_cache_lru_entry(&sctx->dir_utimes_cache);
2806 ASSERT(lru != NULL);
2808 ret = send_utimes(sctx, lru->key, lru->gen);
2812 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, lru);
2819 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2820 * a valid path yet because we did not process the refs yet. So, the inode
2821 * is created as orphan.
2823 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2825 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2829 struct btrfs_inode_info info;
2834 btrfs_debug(fs_info, "send_create_inode %llu", ino);
2836 p = fs_path_alloc();
2840 if (ino != sctx->cur_ino) {
2841 ret = get_inode_info(sctx->send_root, ino, &info);
2848 gen = sctx->cur_inode_gen;
2849 mode = sctx->cur_inode_mode;
2850 rdev = sctx->cur_inode_rdev;
2853 if (S_ISREG(mode)) {
2854 cmd = BTRFS_SEND_C_MKFILE;
2855 } else if (S_ISDIR(mode)) {
2856 cmd = BTRFS_SEND_C_MKDIR;
2857 } else if (S_ISLNK(mode)) {
2858 cmd = BTRFS_SEND_C_SYMLINK;
2859 } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2860 cmd = BTRFS_SEND_C_MKNOD;
2861 } else if (S_ISFIFO(mode)) {
2862 cmd = BTRFS_SEND_C_MKFIFO;
2863 } else if (S_ISSOCK(mode)) {
2864 cmd = BTRFS_SEND_C_MKSOCK;
2866 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2867 (int)(mode & S_IFMT));
2872 ret = begin_cmd(sctx, cmd);
2876 ret = gen_unique_name(sctx, ino, gen, p);
2880 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2881 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2883 if (S_ISLNK(mode)) {
2885 ret = read_symlink(sctx->send_root, ino, p);
2888 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2889 } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2890 S_ISFIFO(mode) || S_ISSOCK(mode)) {
2891 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2892 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2895 ret = send_cmd(sctx);
2906 static void cache_dir_created(struct send_ctx *sctx, u64 dir)
2908 struct btrfs_lru_cache_entry *entry;
2911 /* Caching is optional, ignore any failures. */
2912 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2918 ret = btrfs_lru_cache_store(&sctx->dir_created_cache, entry, GFP_KERNEL);
2924 * We need some special handling for inodes that get processed before the parent
2925 * directory got created. See process_recorded_refs for details.
2926 * This function does the check if we already created the dir out of order.
2928 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2932 struct btrfs_path *path = NULL;
2933 struct btrfs_key key;
2934 struct btrfs_key found_key;
2935 struct btrfs_key di_key;
2936 struct btrfs_dir_item *di;
2938 if (btrfs_lru_cache_lookup(&sctx->dir_created_cache, dir, 0))
2941 path = alloc_path_for_send();
2946 key.type = BTRFS_DIR_INDEX_KEY;
2949 btrfs_for_each_slot(sctx->send_root, &key, &found_key, path, iter_ret) {
2950 struct extent_buffer *eb = path->nodes[0];
2952 if (found_key.objectid != key.objectid ||
2953 found_key.type != key.type) {
2958 di = btrfs_item_ptr(eb, path->slots[0], struct btrfs_dir_item);
2959 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2961 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2962 di_key.objectid < sctx->send_progress) {
2964 cache_dir_created(sctx, dir);
2968 /* Catch error found during iteration */
2972 btrfs_free_path(path);
2977 * Only creates the inode if it is:
2978 * 1. Not a directory
2979 * 2. Or a directory which was not created already due to out of order
2980 * directories. See did_create_dir and process_recorded_refs for details.
2982 static int send_create_inode_if_needed(struct send_ctx *sctx)
2986 if (S_ISDIR(sctx->cur_inode_mode)) {
2987 ret = did_create_dir(sctx, sctx->cur_ino);
2994 ret = send_create_inode(sctx, sctx->cur_ino);
2996 if (ret == 0 && S_ISDIR(sctx->cur_inode_mode))
2997 cache_dir_created(sctx, sctx->cur_ino);
3002 struct recorded_ref {
3003 struct list_head list;
3005 struct fs_path *full_path;
3009 struct rb_node node;
3010 struct rb_root *root;
3013 static struct recorded_ref *recorded_ref_alloc(void)
3015 struct recorded_ref *ref;
3017 ref = kzalloc(sizeof(*ref), GFP_KERNEL);
3020 RB_CLEAR_NODE(&ref->node);
3021 INIT_LIST_HEAD(&ref->list);
3025 static void recorded_ref_free(struct recorded_ref *ref)
3029 if (!RB_EMPTY_NODE(&ref->node))
3030 rb_erase(&ref->node, ref->root);
3031 list_del(&ref->list);
3032 fs_path_free(ref->full_path);
3036 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
3038 ref->full_path = path;
3039 ref->name = (char *)kbasename(ref->full_path->start);
3040 ref->name_len = ref->full_path->end - ref->name;
3043 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
3045 struct recorded_ref *new;
3047 new = recorded_ref_alloc();
3051 new->dir = ref->dir;
3052 new->dir_gen = ref->dir_gen;
3053 list_add_tail(&new->list, list);
3057 static void __free_recorded_refs(struct list_head *head)
3059 struct recorded_ref *cur;
3061 while (!list_empty(head)) {
3062 cur = list_entry(head->next, struct recorded_ref, list);
3063 recorded_ref_free(cur);
3067 static void free_recorded_refs(struct send_ctx *sctx)
3069 __free_recorded_refs(&sctx->new_refs);
3070 __free_recorded_refs(&sctx->deleted_refs);
3074 * Renames/moves a file/dir to its orphan name. Used when the first
3075 * ref of an unprocessed inode gets overwritten and for all non empty
3078 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
3079 struct fs_path *path)
3082 struct fs_path *orphan;
3084 orphan = fs_path_alloc();
3088 ret = gen_unique_name(sctx, ino, gen, orphan);
3092 ret = send_rename(sctx, path, orphan);
3095 fs_path_free(orphan);
3099 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
3100 u64 dir_ino, u64 dir_gen)
3102 struct rb_node **p = &sctx->orphan_dirs.rb_node;
3103 struct rb_node *parent = NULL;
3104 struct orphan_dir_info *entry, *odi;
3108 entry = rb_entry(parent, struct orphan_dir_info, node);
3109 if (dir_ino < entry->ino)
3111 else if (dir_ino > entry->ino)
3112 p = &(*p)->rb_right;
3113 else if (dir_gen < entry->gen)
3115 else if (dir_gen > entry->gen)
3116 p = &(*p)->rb_right;
3121 odi = kmalloc(sizeof(*odi), GFP_KERNEL);
3123 return ERR_PTR(-ENOMEM);
3126 odi->last_dir_index_offset = 0;
3127 odi->dir_high_seq_ino = 0;
3129 rb_link_node(&odi->node, parent, p);
3130 rb_insert_color(&odi->node, &sctx->orphan_dirs);
3134 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
3135 u64 dir_ino, u64 gen)
3137 struct rb_node *n = sctx->orphan_dirs.rb_node;
3138 struct orphan_dir_info *entry;
3141 entry = rb_entry(n, struct orphan_dir_info, node);
3142 if (dir_ino < entry->ino)
3144 else if (dir_ino > entry->ino)
3146 else if (gen < entry->gen)
3148 else if (gen > entry->gen)
3156 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
3158 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
3163 static void free_orphan_dir_info(struct send_ctx *sctx,
3164 struct orphan_dir_info *odi)
3168 rb_erase(&odi->node, &sctx->orphan_dirs);
3173 * Returns 1 if a directory can be removed at this point in time.
3174 * We check this by iterating all dir items and checking if the inode behind
3175 * the dir item was already processed.
3177 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen)
3181 struct btrfs_root *root = sctx->parent_root;
3182 struct btrfs_path *path;
3183 struct btrfs_key key;
3184 struct btrfs_key found_key;
3185 struct btrfs_key loc;
3186 struct btrfs_dir_item *di;
3187 struct orphan_dir_info *odi = NULL;
3188 u64 dir_high_seq_ino = 0;
3189 u64 last_dir_index_offset = 0;
3192 * Don't try to rmdir the top/root subvolume dir.
3194 if (dir == BTRFS_FIRST_FREE_OBJECTID)
3197 odi = get_orphan_dir_info(sctx, dir, dir_gen);
3198 if (odi && sctx->cur_ino < odi->dir_high_seq_ino)
3201 path = alloc_path_for_send();
3207 * Find the inode number associated with the last dir index
3208 * entry. This is very likely the inode with the highest number
3209 * of all inodes that have an entry in the directory. We can
3210 * then use it to avoid future calls to can_rmdir(), when
3211 * processing inodes with a lower number, from having to search
3212 * the parent root b+tree for dir index keys.
3215 key.type = BTRFS_DIR_INDEX_KEY;
3216 key.offset = (u64)-1;
3218 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3221 } else if (ret > 0) {
3222 /* Can't happen, the root is never empty. */
3223 ASSERT(path->slots[0] > 0);
3224 if (WARN_ON(path->slots[0] == 0)) {
3231 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3232 if (key.objectid != dir || key.type != BTRFS_DIR_INDEX_KEY) {
3233 /* No index keys, dir can be removed. */
3238 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3239 struct btrfs_dir_item);
3240 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3241 dir_high_seq_ino = loc.objectid;
3242 if (sctx->cur_ino < dir_high_seq_ino) {
3247 btrfs_release_path(path);
3251 key.type = BTRFS_DIR_INDEX_KEY;
3252 key.offset = (odi ? odi->last_dir_index_offset : 0);
3254 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
3255 struct waiting_dir_move *dm;
3257 if (found_key.objectid != key.objectid ||
3258 found_key.type != key.type)
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);
3265 dir_high_seq_ino = max(dir_high_seq_ino, loc.objectid);
3266 last_dir_index_offset = found_key.offset;
3268 dm = get_waiting_dir_move(sctx, loc.objectid);
3270 dm->rmdir_ino = dir;
3271 dm->rmdir_gen = dir_gen;
3276 if (loc.objectid > sctx->cur_ino) {
3285 free_orphan_dir_info(sctx, odi);
3290 btrfs_free_path(path);
3296 odi = add_orphan_dir_info(sctx, dir, dir_gen);
3298 return PTR_ERR(odi);
3303 odi->last_dir_index_offset = last_dir_index_offset;
3304 odi->dir_high_seq_ino = max(odi->dir_high_seq_ino, dir_high_seq_ino);
3309 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3311 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3313 return entry != NULL;
3316 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3318 struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3319 struct rb_node *parent = NULL;
3320 struct waiting_dir_move *entry, *dm;
3322 dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3328 dm->orphanized = orphanized;
3332 entry = rb_entry(parent, struct waiting_dir_move, node);
3333 if (ino < entry->ino) {
3335 } else if (ino > entry->ino) {
3336 p = &(*p)->rb_right;
3343 rb_link_node(&dm->node, parent, p);
3344 rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3348 static struct waiting_dir_move *
3349 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3351 struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3352 struct waiting_dir_move *entry;
3355 entry = rb_entry(n, struct waiting_dir_move, node);
3356 if (ino < entry->ino)
3358 else if (ino > entry->ino)
3366 static void free_waiting_dir_move(struct send_ctx *sctx,
3367 struct waiting_dir_move *dm)
3371 rb_erase(&dm->node, &sctx->waiting_dir_moves);
3375 static int add_pending_dir_move(struct send_ctx *sctx,
3379 struct list_head *new_refs,
3380 struct list_head *deleted_refs,
3381 const bool is_orphan)
3383 struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3384 struct rb_node *parent = NULL;
3385 struct pending_dir_move *entry = NULL, *pm;
3386 struct recorded_ref *cur;
3390 pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3393 pm->parent_ino = parent_ino;
3396 INIT_LIST_HEAD(&pm->list);
3397 INIT_LIST_HEAD(&pm->update_refs);
3398 RB_CLEAR_NODE(&pm->node);
3402 entry = rb_entry(parent, struct pending_dir_move, node);
3403 if (parent_ino < entry->parent_ino) {
3405 } else if (parent_ino > entry->parent_ino) {
3406 p = &(*p)->rb_right;
3413 list_for_each_entry(cur, deleted_refs, list) {
3414 ret = dup_ref(cur, &pm->update_refs);
3418 list_for_each_entry(cur, new_refs, list) {
3419 ret = dup_ref(cur, &pm->update_refs);
3424 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3429 list_add_tail(&pm->list, &entry->list);
3431 rb_link_node(&pm->node, parent, p);
3432 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3437 __free_recorded_refs(&pm->update_refs);
3443 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3446 struct rb_node *n = sctx->pending_dir_moves.rb_node;
3447 struct pending_dir_move *entry;
3450 entry = rb_entry(n, struct pending_dir_move, node);
3451 if (parent_ino < entry->parent_ino)
3453 else if (parent_ino > entry->parent_ino)
3461 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3462 u64 ino, u64 gen, u64 *ancestor_ino)
3465 u64 parent_inode = 0;
3467 u64 start_ino = ino;
3470 while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3471 fs_path_reset(name);
3473 if (is_waiting_for_rm(sctx, ino, gen))
3475 if (is_waiting_for_move(sctx, ino)) {
3476 if (*ancestor_ino == 0)
3477 *ancestor_ino = ino;
3478 ret = get_first_ref(sctx->parent_root, ino,
3479 &parent_inode, &parent_gen, name);
3481 ret = __get_cur_name_and_parent(sctx, ino, gen,
3491 if (parent_inode == start_ino) {
3493 if (*ancestor_ino == 0)
3494 *ancestor_ino = ino;
3503 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3505 struct fs_path *from_path = NULL;
3506 struct fs_path *to_path = NULL;
3507 struct fs_path *name = NULL;
3508 u64 orig_progress = sctx->send_progress;
3509 struct recorded_ref *cur;
3510 u64 parent_ino, parent_gen;
3511 struct waiting_dir_move *dm = NULL;
3518 name = fs_path_alloc();
3519 from_path = fs_path_alloc();
3520 if (!name || !from_path) {
3525 dm = get_waiting_dir_move(sctx, pm->ino);
3527 rmdir_ino = dm->rmdir_ino;
3528 rmdir_gen = dm->rmdir_gen;
3529 is_orphan = dm->orphanized;
3530 free_waiting_dir_move(sctx, dm);
3533 ret = gen_unique_name(sctx, pm->ino,
3534 pm->gen, from_path);
3536 ret = get_first_ref(sctx->parent_root, pm->ino,
3537 &parent_ino, &parent_gen, name);
3540 ret = get_cur_path(sctx, parent_ino, parent_gen,
3544 ret = fs_path_add_path(from_path, name);
3549 sctx->send_progress = sctx->cur_ino + 1;
3550 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3554 LIST_HEAD(deleted_refs);
3555 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3556 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3557 &pm->update_refs, &deleted_refs,
3562 dm = get_waiting_dir_move(sctx, pm->ino);
3564 dm->rmdir_ino = rmdir_ino;
3565 dm->rmdir_gen = rmdir_gen;
3569 fs_path_reset(name);
3572 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3576 ret = send_rename(sctx, from_path, to_path);
3581 struct orphan_dir_info *odi;
3584 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3586 /* already deleted */
3591 ret = can_rmdir(sctx, rmdir_ino, gen);
3597 name = fs_path_alloc();
3602 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3605 ret = send_rmdir(sctx, name);
3611 ret = cache_dir_utimes(sctx, pm->ino, pm->gen);
3616 * After rename/move, need to update the utimes of both new parent(s)
3617 * and old parent(s).
3619 list_for_each_entry(cur, &pm->update_refs, list) {
3621 * The parent inode might have been deleted in the send snapshot
3623 ret = get_inode_info(sctx->send_root, cur->dir, NULL);
3624 if (ret == -ENOENT) {
3631 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
3638 fs_path_free(from_path);
3639 fs_path_free(to_path);
3640 sctx->send_progress = orig_progress;
3645 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3647 if (!list_empty(&m->list))
3649 if (!RB_EMPTY_NODE(&m->node))
3650 rb_erase(&m->node, &sctx->pending_dir_moves);
3651 __free_recorded_refs(&m->update_refs);
3655 static void tail_append_pending_moves(struct send_ctx *sctx,
3656 struct pending_dir_move *moves,
3657 struct list_head *stack)
3659 if (list_empty(&moves->list)) {
3660 list_add_tail(&moves->list, stack);
3663 list_splice_init(&moves->list, &list);
3664 list_add_tail(&moves->list, stack);
3665 list_splice_tail(&list, stack);
3667 if (!RB_EMPTY_NODE(&moves->node)) {
3668 rb_erase(&moves->node, &sctx->pending_dir_moves);
3669 RB_CLEAR_NODE(&moves->node);
3673 static int apply_children_dir_moves(struct send_ctx *sctx)
3675 struct pending_dir_move *pm;
3676 struct list_head stack;
3677 u64 parent_ino = sctx->cur_ino;
3680 pm = get_pending_dir_moves(sctx, parent_ino);
3684 INIT_LIST_HEAD(&stack);
3685 tail_append_pending_moves(sctx, pm, &stack);
3687 while (!list_empty(&stack)) {
3688 pm = list_first_entry(&stack, struct pending_dir_move, list);
3689 parent_ino = pm->ino;
3690 ret = apply_dir_move(sctx, pm);
3691 free_pending_move(sctx, pm);
3694 pm = get_pending_dir_moves(sctx, parent_ino);
3696 tail_append_pending_moves(sctx, pm, &stack);
3701 while (!list_empty(&stack)) {
3702 pm = list_first_entry(&stack, struct pending_dir_move, list);
3703 free_pending_move(sctx, pm);
3709 * We might need to delay a directory rename even when no ancestor directory
3710 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3711 * renamed. This happens when we rename a directory to the old name (the name
3712 * in the parent root) of some other unrelated directory that got its rename
3713 * delayed due to some ancestor with higher number that got renamed.
3719 * |---- a/ (ino 257)
3720 * | |---- file (ino 260)
3722 * |---- b/ (ino 258)
3723 * |---- c/ (ino 259)
3727 * |---- a/ (ino 258)
3728 * |---- x/ (ino 259)
3729 * |---- y/ (ino 257)
3730 * |----- file (ino 260)
3732 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3733 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3734 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3737 * 1 - rename 259 from 'c' to 'x'
3738 * 2 - rename 257 from 'a' to 'x/y'
3739 * 3 - rename 258 from 'b' to 'a'
3741 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3742 * be done right away and < 0 on error.
3744 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3745 struct recorded_ref *parent_ref,
3746 const bool is_orphan)
3748 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3749 struct btrfs_path *path;
3750 struct btrfs_key key;
3751 struct btrfs_key di_key;
3752 struct btrfs_dir_item *di;
3756 struct waiting_dir_move *wdm;
3758 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3761 path = alloc_path_for_send();
3765 key.objectid = parent_ref->dir;
3766 key.type = BTRFS_DIR_ITEM_KEY;
3767 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3769 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3772 } else if (ret > 0) {
3777 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3778 parent_ref->name_len);
3784 * di_key.objectid has the number of the inode that has a dentry in the
3785 * parent directory with the same name that sctx->cur_ino is being
3786 * renamed to. We need to check if that inode is in the send root as
3787 * well and if it is currently marked as an inode with a pending rename,
3788 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3789 * that it happens after that other inode is renamed.
3791 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3792 if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3797 ret = get_inode_gen(sctx->parent_root, di_key.objectid, &left_gen);
3800 ret = get_inode_gen(sctx->send_root, di_key.objectid, &right_gen);
3807 /* Different inode, no need to delay the rename of sctx->cur_ino */
3808 if (right_gen != left_gen) {
3813 wdm = get_waiting_dir_move(sctx, di_key.objectid);
3814 if (wdm && !wdm->orphanized) {
3815 ret = add_pending_dir_move(sctx,
3817 sctx->cur_inode_gen,
3820 &sctx->deleted_refs,
3826 btrfs_free_path(path);
3831 * Check if inode ino2, or any of its ancestors, is inode ino1.
3832 * Return 1 if true, 0 if false and < 0 on error.
3834 static int check_ino_in_path(struct btrfs_root *root,
3839 struct fs_path *fs_path)
3844 return ino1_gen == ino2_gen;
3846 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3851 fs_path_reset(fs_path);
3852 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3856 return parent_gen == ino1_gen;
3863 * Check if inode ino1 is an ancestor of inode ino2 in the given root for any
3864 * possible path (in case ino2 is not a directory and has multiple hard links).
3865 * Return 1 if true, 0 if false and < 0 on error.
3867 static int is_ancestor(struct btrfs_root *root,
3871 struct fs_path *fs_path)
3873 bool free_fs_path = false;
3876 struct btrfs_path *path = NULL;
3877 struct btrfs_key key;
3880 fs_path = fs_path_alloc();
3883 free_fs_path = true;
3886 path = alloc_path_for_send();
3892 key.objectid = ino2;
3893 key.type = BTRFS_INODE_REF_KEY;
3896 btrfs_for_each_slot(root, &key, &key, path, iter_ret) {
3897 struct extent_buffer *leaf = path->nodes[0];
3898 int slot = path->slots[0];
3902 if (key.objectid != ino2)
3904 if (key.type != BTRFS_INODE_REF_KEY &&
3905 key.type != BTRFS_INODE_EXTREF_KEY)
3908 item_size = btrfs_item_size(leaf, slot);
3909 while (cur_offset < item_size) {
3913 if (key.type == BTRFS_INODE_EXTREF_KEY) {
3915 struct btrfs_inode_extref *extref;
3917 ptr = btrfs_item_ptr_offset(leaf, slot);
3918 extref = (struct btrfs_inode_extref *)
3920 parent = btrfs_inode_extref_parent(leaf,
3922 cur_offset += sizeof(*extref);
3923 cur_offset += btrfs_inode_extref_name_len(leaf,
3926 parent = key.offset;
3927 cur_offset = item_size;
3930 ret = get_inode_gen(root, parent, &parent_gen);
3933 ret = check_ino_in_path(root, ino1, ino1_gen,
3934 parent, parent_gen, fs_path);
3944 btrfs_free_path(path);
3946 fs_path_free(fs_path);
3950 static int wait_for_parent_move(struct send_ctx *sctx,
3951 struct recorded_ref *parent_ref,
3952 const bool is_orphan)
3955 u64 ino = parent_ref->dir;
3956 u64 ino_gen = parent_ref->dir_gen;
3957 u64 parent_ino_before, parent_ino_after;
3958 struct fs_path *path_before = NULL;
3959 struct fs_path *path_after = NULL;
3962 path_after = fs_path_alloc();
3963 path_before = fs_path_alloc();
3964 if (!path_after || !path_before) {
3970 * Our current directory inode may not yet be renamed/moved because some
3971 * ancestor (immediate or not) has to be renamed/moved first. So find if
3972 * such ancestor exists and make sure our own rename/move happens after
3973 * that ancestor is processed to avoid path build infinite loops (done
3974 * at get_cur_path()).
3976 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3977 u64 parent_ino_after_gen;
3979 if (is_waiting_for_move(sctx, ino)) {
3981 * If the current inode is an ancestor of ino in the
3982 * parent root, we need to delay the rename of the
3983 * current inode, otherwise don't delayed the rename
3984 * because we can end up with a circular dependency
3985 * of renames, resulting in some directories never
3986 * getting the respective rename operations issued in
3987 * the send stream or getting into infinite path build
3990 ret = is_ancestor(sctx->parent_root,
3991 sctx->cur_ino, sctx->cur_inode_gen,
3997 fs_path_reset(path_before);
3998 fs_path_reset(path_after);
4000 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
4001 &parent_ino_after_gen, path_after);
4004 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
4006 if (ret < 0 && ret != -ENOENT) {
4008 } else if (ret == -ENOENT) {
4013 len1 = fs_path_len(path_before);
4014 len2 = fs_path_len(path_after);
4015 if (ino > sctx->cur_ino &&
4016 (parent_ino_before != parent_ino_after || len1 != len2 ||
4017 memcmp(path_before->start, path_after->start, len1))) {
4020 ret = get_inode_gen(sctx->parent_root, ino, &parent_ino_gen);
4023 if (ino_gen == parent_ino_gen) {
4028 ino = parent_ino_after;
4029 ino_gen = parent_ino_after_gen;
4033 fs_path_free(path_before);
4034 fs_path_free(path_after);
4037 ret = add_pending_dir_move(sctx,
4039 sctx->cur_inode_gen,
4042 &sctx->deleted_refs,
4051 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4054 struct fs_path *new_path;
4057 * Our reference's name member points to its full_path member string, so
4058 * we use here a new path.
4060 new_path = fs_path_alloc();
4064 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
4066 fs_path_free(new_path);
4069 ret = fs_path_add(new_path, ref->name, ref->name_len);
4071 fs_path_free(new_path);
4075 fs_path_free(ref->full_path);
4076 set_ref_path(ref, new_path);
4082 * When processing the new references for an inode we may orphanize an existing
4083 * directory inode because its old name conflicts with one of the new references
4084 * of the current inode. Later, when processing another new reference of our
4085 * inode, we might need to orphanize another inode, but the path we have in the
4086 * reference reflects the pre-orphanization name of the directory we previously
4087 * orphanized. For example:
4089 * parent snapshot looks like:
4092 * |----- f1 (ino 257)
4093 * |----- f2 (ino 258)
4094 * |----- d1/ (ino 259)
4095 * |----- d2/ (ino 260)
4097 * send snapshot looks like:
4100 * |----- d1 (ino 258)
4101 * |----- f2/ (ino 259)
4102 * |----- f2_link/ (ino 260)
4103 * | |----- f1 (ino 257)
4105 * |----- d2 (ino 258)
4107 * When processing inode 257 we compute the name for inode 259 as "d1", and we
4108 * cache it in the name cache. Later when we start processing inode 258, when
4109 * collecting all its new references we set a full path of "d1/d2" for its new
4110 * reference with name "d2". When we start processing the new references we
4111 * start by processing the new reference with name "d1", and this results in
4112 * orphanizing inode 259, since its old reference causes a conflict. Then we
4113 * move on the next new reference, with name "d2", and we find out we must
4114 * orphanize inode 260, as its old reference conflicts with ours - but for the
4115 * orphanization we use a source path corresponding to the path we stored in the
4116 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
4117 * receiver fail since the path component "d1/" no longer exists, it was renamed
4118 * to "o259-6-0/" when processing the previous new reference. So in this case we
4119 * must recompute the path in the new reference and use it for the new
4120 * orphanization operation.
4122 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4127 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
4131 fs_path_reset(ref->full_path);
4132 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
4136 ret = fs_path_add(ref->full_path, name, ref->name_len);
4140 /* Update the reference's base name pointer. */
4141 set_ref_path(ref, ref->full_path);
4148 * This does all the move/link/unlink/rmdir magic.
4150 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
4152 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4154 struct recorded_ref *cur;
4155 struct recorded_ref *cur2;
4156 struct list_head check_dirs;
4157 struct fs_path *valid_path = NULL;
4161 int did_overwrite = 0;
4163 u64 last_dir_ino_rm = 0;
4164 bool can_rename = true;
4165 bool orphanized_dir = false;
4166 bool orphanized_ancestor = false;
4168 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
4171 * This should never happen as the root dir always has the same ref
4172 * which is always '..'
4174 BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
4175 INIT_LIST_HEAD(&check_dirs);
4177 valid_path = fs_path_alloc();
4184 * First, check if the first ref of the current inode was overwritten
4185 * before. If yes, we know that the current inode was already orphanized
4186 * and thus use the orphan name. If not, we can use get_cur_path to
4187 * get the path of the first ref as it would like while receiving at
4188 * this point in time.
4189 * New inodes are always orphan at the beginning, so force to use the
4190 * orphan name in this case.
4191 * The first ref is stored in valid_path and will be updated if it
4192 * gets moved around.
4194 if (!sctx->cur_inode_new) {
4195 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
4196 sctx->cur_inode_gen);
4202 if (sctx->cur_inode_new || did_overwrite) {
4203 ret = gen_unique_name(sctx, sctx->cur_ino,
4204 sctx->cur_inode_gen, valid_path);
4209 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4216 * Before doing any rename and link operations, do a first pass on the
4217 * new references to orphanize any unprocessed inodes that may have a
4218 * reference that conflicts with one of the new references of the current
4219 * inode. This needs to happen first because a new reference may conflict
4220 * with the old reference of a parent directory, so we must make sure
4221 * that the path used for link and rename commands don't use an
4222 * orphanized name when an ancestor was not yet orphanized.
4229 * |----- testdir/ (ino 259)
4230 * | |----- a (ino 257)
4232 * |----- b (ino 258)
4237 * |----- testdir_2/ (ino 259)
4238 * | |----- a (ino 260)
4240 * |----- testdir (ino 257)
4241 * |----- b (ino 257)
4242 * |----- b2 (ino 258)
4244 * Processing the new reference for inode 257 with name "b" may happen
4245 * before processing the new reference with name "testdir". If so, we
4246 * must make sure that by the time we send a link command to create the
4247 * hard link "b", inode 259 was already orphanized, since the generated
4248 * path in "valid_path" already contains the orphanized name for 259.
4249 * We are processing inode 257, so only later when processing 259 we do
4250 * the rename operation to change its temporary (orphanized) name to
4253 list_for_each_entry(cur, &sctx->new_refs, list) {
4254 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4257 if (ret == inode_state_will_create)
4261 * Check if this new ref would overwrite the first ref of another
4262 * unprocessed inode. If yes, orphanize the overwritten inode.
4263 * If we find an overwritten ref that is not the first ref,
4266 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4267 cur->name, cur->name_len,
4268 &ow_inode, &ow_gen, &ow_mode);
4272 ret = is_first_ref(sctx->parent_root,
4273 ow_inode, cur->dir, cur->name,
4278 struct name_cache_entry *nce;
4279 struct waiting_dir_move *wdm;
4281 if (orphanized_dir) {
4282 ret = refresh_ref_path(sctx, cur);
4287 ret = orphanize_inode(sctx, ow_inode, ow_gen,
4291 if (S_ISDIR(ow_mode))
4292 orphanized_dir = true;
4295 * If ow_inode has its rename operation delayed
4296 * make sure that its orphanized name is used in
4297 * the source path when performing its rename
4300 wdm = get_waiting_dir_move(sctx, ow_inode);
4302 wdm->orphanized = true;
4305 * Make sure we clear our orphanized inode's
4306 * name from the name cache. This is because the
4307 * inode ow_inode might be an ancestor of some
4308 * other inode that will be orphanized as well
4309 * later and has an inode number greater than
4310 * sctx->send_progress. We need to prevent
4311 * future name lookups from using the old name
4312 * and get instead the orphan name.
4314 nce = name_cache_search(sctx, ow_inode, ow_gen);
4316 btrfs_lru_cache_remove(&sctx->name_cache,
4320 * ow_inode might currently be an ancestor of
4321 * cur_ino, therefore compute valid_path (the
4322 * current path of cur_ino) again because it
4323 * might contain the pre-orphanization name of
4324 * ow_inode, which is no longer valid.
4326 ret = is_ancestor(sctx->parent_root,
4328 sctx->cur_ino, NULL);
4330 orphanized_ancestor = true;
4331 fs_path_reset(valid_path);
4332 ret = get_cur_path(sctx, sctx->cur_ino,
4333 sctx->cur_inode_gen,
4340 * If we previously orphanized a directory that
4341 * collided with a new reference that we already
4342 * processed, recompute the current path because
4343 * that directory may be part of the path.
4345 if (orphanized_dir) {
4346 ret = refresh_ref_path(sctx, cur);
4350 ret = send_unlink(sctx, cur->full_path);
4358 list_for_each_entry(cur, &sctx->new_refs, list) {
4360 * We may have refs where the parent directory does not exist
4361 * yet. This happens if the parent directories inum is higher
4362 * than the current inum. To handle this case, we create the
4363 * parent directory out of order. But we need to check if this
4364 * did already happen before due to other refs in the same dir.
4366 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4369 if (ret == inode_state_will_create) {
4372 * First check if any of the current inodes refs did
4373 * already create the dir.
4375 list_for_each_entry(cur2, &sctx->new_refs, list) {
4378 if (cur2->dir == cur->dir) {
4385 * If that did not happen, check if a previous inode
4386 * did already create the dir.
4389 ret = did_create_dir(sctx, cur->dir);
4393 ret = send_create_inode(sctx, cur->dir);
4396 cache_dir_created(sctx, cur->dir);
4400 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4401 ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4410 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4412 ret = wait_for_parent_move(sctx, cur, is_orphan);
4422 * link/move the ref to the new place. If we have an orphan
4423 * inode, move it and update valid_path. If not, link or move
4424 * it depending on the inode mode.
4426 if (is_orphan && can_rename) {
4427 ret = send_rename(sctx, valid_path, cur->full_path);
4431 ret = fs_path_copy(valid_path, cur->full_path);
4434 } else if (can_rename) {
4435 if (S_ISDIR(sctx->cur_inode_mode)) {
4437 * Dirs can't be linked, so move it. For moved
4438 * dirs, we always have one new and one deleted
4439 * ref. The deleted ref is ignored later.
4441 ret = send_rename(sctx, valid_path,
4444 ret = fs_path_copy(valid_path,
4450 * We might have previously orphanized an inode
4451 * which is an ancestor of our current inode,
4452 * so our reference's full path, which was
4453 * computed before any such orphanizations, must
4456 if (orphanized_dir) {
4457 ret = update_ref_path(sctx, cur);
4461 ret = send_link(sctx, cur->full_path,
4467 ret = dup_ref(cur, &check_dirs);
4472 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4474 * Check if we can already rmdir the directory. If not,
4475 * orphanize it. For every dir item inside that gets deleted
4476 * later, we do this check again and rmdir it then if possible.
4477 * See the use of check_dirs for more details.
4479 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen);
4483 ret = send_rmdir(sctx, valid_path);
4486 } else if (!is_orphan) {
4487 ret = orphanize_inode(sctx, sctx->cur_ino,
4488 sctx->cur_inode_gen, valid_path);
4494 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4495 ret = dup_ref(cur, &check_dirs);
4499 } else if (S_ISDIR(sctx->cur_inode_mode) &&
4500 !list_empty(&sctx->deleted_refs)) {
4502 * We have a moved dir. Add the old parent to check_dirs
4504 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4506 ret = dup_ref(cur, &check_dirs);
4509 } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4511 * We have a non dir inode. Go through all deleted refs and
4512 * unlink them if they were not already overwritten by other
4515 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4516 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4517 sctx->cur_ino, sctx->cur_inode_gen,
4518 cur->name, cur->name_len);
4523 * If we orphanized any ancestor before, we need
4524 * to recompute the full path for deleted names,
4525 * since any such path was computed before we
4526 * processed any references and orphanized any
4529 if (orphanized_ancestor) {
4530 ret = update_ref_path(sctx, cur);
4534 ret = send_unlink(sctx, cur->full_path);
4538 ret = dup_ref(cur, &check_dirs);
4543 * If the inode is still orphan, unlink the orphan. This may
4544 * happen when a previous inode did overwrite the first ref
4545 * of this inode and no new refs were added for the current
4546 * inode. Unlinking does not mean that the inode is deleted in
4547 * all cases. There may still be links to this inode in other
4551 ret = send_unlink(sctx, valid_path);
4558 * We did collect all parent dirs where cur_inode was once located. We
4559 * now go through all these dirs and check if they are pending for
4560 * deletion and if it's finally possible to perform the rmdir now.
4561 * We also update the inode stats of the parent dirs here.
4563 list_for_each_entry(cur, &check_dirs, list) {
4565 * In case we had refs into dirs that were not processed yet,
4566 * we don't need to do the utime and rmdir logic for these dirs.
4567 * The dir will be processed later.
4569 if (cur->dir > sctx->cur_ino)
4572 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4576 if (ret == inode_state_did_create ||
4577 ret == inode_state_no_change) {
4578 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
4581 } else if (ret == inode_state_did_delete &&
4582 cur->dir != last_dir_ino_rm) {
4583 ret = can_rmdir(sctx, cur->dir, cur->dir_gen);
4587 ret = get_cur_path(sctx, cur->dir,
4588 cur->dir_gen, valid_path);
4591 ret = send_rmdir(sctx, valid_path);
4594 last_dir_ino_rm = cur->dir;
4602 __free_recorded_refs(&check_dirs);
4603 free_recorded_refs(sctx);
4604 fs_path_free(valid_path);
4608 static int rbtree_ref_comp(const void *k, const struct rb_node *node)
4610 const struct recorded_ref *data = k;
4611 const struct recorded_ref *ref = rb_entry(node, struct recorded_ref, node);
4614 if (data->dir > ref->dir)
4616 if (data->dir < ref->dir)
4618 if (data->dir_gen > ref->dir_gen)
4620 if (data->dir_gen < ref->dir_gen)
4622 if (data->name_len > ref->name_len)
4624 if (data->name_len < ref->name_len)
4626 result = strcmp(data->name, ref->name);
4634 static bool rbtree_ref_less(struct rb_node *node, const struct rb_node *parent)
4636 const struct recorded_ref *entry = rb_entry(node, struct recorded_ref, node);
4638 return rbtree_ref_comp(entry, parent) < 0;
4641 static int record_ref_in_tree(struct rb_root *root, struct list_head *refs,
4642 struct fs_path *name, u64 dir, u64 dir_gen,
4643 struct send_ctx *sctx)
4646 struct fs_path *path = NULL;
4647 struct recorded_ref *ref = NULL;
4649 path = fs_path_alloc();
4655 ref = recorded_ref_alloc();
4661 ret = get_cur_path(sctx, dir, dir_gen, path);
4664 ret = fs_path_add_path(path, name);
4669 ref->dir_gen = dir_gen;
4670 set_ref_path(ref, path);
4671 list_add_tail(&ref->list, refs);
4672 rb_add(&ref->node, root, rbtree_ref_less);
4676 if (path && (!ref || !ref->full_path))
4678 recorded_ref_free(ref);
4683 static int record_new_ref_if_needed(int num, u64 dir, int index,
4684 struct fs_path *name, void *ctx)
4687 struct send_ctx *sctx = ctx;
4688 struct rb_node *node = NULL;
4689 struct recorded_ref data;
4690 struct recorded_ref *ref;
4693 ret = get_inode_gen(sctx->send_root, dir, &dir_gen);
4698 data.dir_gen = dir_gen;
4699 set_ref_path(&data, name);
4700 node = rb_find(&data, &sctx->rbtree_deleted_refs, rbtree_ref_comp);
4702 ref = rb_entry(node, struct recorded_ref, node);
4703 recorded_ref_free(ref);
4705 ret = record_ref_in_tree(&sctx->rbtree_new_refs,
4706 &sctx->new_refs, name, dir, dir_gen,
4713 static int record_deleted_ref_if_needed(int num, u64 dir, int index,
4714 struct fs_path *name, void *ctx)
4717 struct send_ctx *sctx = ctx;
4718 struct rb_node *node = NULL;
4719 struct recorded_ref data;
4720 struct recorded_ref *ref;
4723 ret = get_inode_gen(sctx->parent_root, dir, &dir_gen);
4728 data.dir_gen = dir_gen;
4729 set_ref_path(&data, name);
4730 node = rb_find(&data, &sctx->rbtree_new_refs, rbtree_ref_comp);
4732 ref = rb_entry(node, struct recorded_ref, node);
4733 recorded_ref_free(ref);
4735 ret = record_ref_in_tree(&sctx->rbtree_deleted_refs,
4736 &sctx->deleted_refs, name, dir,
4743 static int record_new_ref(struct send_ctx *sctx)
4747 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4748 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4757 static int record_deleted_ref(struct send_ctx *sctx)
4761 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4762 sctx->cmp_key, 0, record_deleted_ref_if_needed,
4772 static int record_changed_ref(struct send_ctx *sctx)
4776 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4777 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4780 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4781 sctx->cmp_key, 0, record_deleted_ref_if_needed, sctx);
4791 * Record and process all refs at once. Needed when an inode changes the
4792 * generation number, which means that it was deleted and recreated.
4794 static int process_all_refs(struct send_ctx *sctx,
4795 enum btrfs_compare_tree_result cmd)
4799 struct btrfs_root *root;
4800 struct btrfs_path *path;
4801 struct btrfs_key key;
4802 struct btrfs_key found_key;
4803 iterate_inode_ref_t cb;
4804 int pending_move = 0;
4806 path = alloc_path_for_send();
4810 if (cmd == BTRFS_COMPARE_TREE_NEW) {
4811 root = sctx->send_root;
4812 cb = record_new_ref_if_needed;
4813 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4814 root = sctx->parent_root;
4815 cb = record_deleted_ref_if_needed;
4817 btrfs_err(sctx->send_root->fs_info,
4818 "Wrong command %d in process_all_refs", cmd);
4823 key.objectid = sctx->cmp_key->objectid;
4824 key.type = BTRFS_INODE_REF_KEY;
4826 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
4827 if (found_key.objectid != key.objectid ||
4828 (found_key.type != BTRFS_INODE_REF_KEY &&
4829 found_key.type != BTRFS_INODE_EXTREF_KEY))
4832 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4836 /* Catch error found during iteration */
4841 btrfs_release_path(path);
4844 * We don't actually care about pending_move as we are simply
4845 * re-creating this inode and will be rename'ing it into place once we
4846 * rename the parent directory.
4848 ret = process_recorded_refs(sctx, &pending_move);
4850 btrfs_free_path(path);
4854 static int send_set_xattr(struct send_ctx *sctx,
4855 struct fs_path *path,
4856 const char *name, int name_len,
4857 const char *data, int data_len)
4861 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4865 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4866 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4867 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4869 ret = send_cmd(sctx);
4876 static int send_remove_xattr(struct send_ctx *sctx,
4877 struct fs_path *path,
4878 const char *name, int name_len)
4882 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4886 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4887 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4889 ret = send_cmd(sctx);
4896 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4897 const char *name, int name_len, const char *data,
4898 int data_len, void *ctx)
4901 struct send_ctx *sctx = ctx;
4903 struct posix_acl_xattr_header dummy_acl;
4905 /* Capabilities are emitted by finish_inode_if_needed */
4906 if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4909 p = fs_path_alloc();
4914 * This hack is needed because empty acls are stored as zero byte
4915 * data in xattrs. Problem with that is, that receiving these zero byte
4916 * acls will fail later. To fix this, we send a dummy acl list that
4917 * only contains the version number and no entries.
4919 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4920 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4921 if (data_len == 0) {
4922 dummy_acl.a_version =
4923 cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4924 data = (char *)&dummy_acl;
4925 data_len = sizeof(dummy_acl);
4929 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4933 ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4940 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4941 const char *name, int name_len,
4942 const char *data, int data_len, void *ctx)
4945 struct send_ctx *sctx = ctx;
4948 p = fs_path_alloc();
4952 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4956 ret = send_remove_xattr(sctx, p, name, name_len);
4963 static int process_new_xattr(struct send_ctx *sctx)
4967 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4968 __process_new_xattr, sctx);
4973 static int process_deleted_xattr(struct send_ctx *sctx)
4975 return iterate_dir_item(sctx->parent_root, sctx->right_path,
4976 __process_deleted_xattr, sctx);
4979 struct find_xattr_ctx {
4987 static int __find_xattr(int num, struct btrfs_key *di_key, const char *name,
4988 int name_len, const char *data, int data_len, void *vctx)
4990 struct find_xattr_ctx *ctx = vctx;
4992 if (name_len == ctx->name_len &&
4993 strncmp(name, ctx->name, name_len) == 0) {
4994 ctx->found_idx = num;
4995 ctx->found_data_len = data_len;
4996 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4997 if (!ctx->found_data)
5004 static int find_xattr(struct btrfs_root *root,
5005 struct btrfs_path *path,
5006 struct btrfs_key *key,
5007 const char *name, int name_len,
5008 char **data, int *data_len)
5011 struct find_xattr_ctx ctx;
5014 ctx.name_len = name_len;
5016 ctx.found_data = NULL;
5017 ctx.found_data_len = 0;
5019 ret = iterate_dir_item(root, path, __find_xattr, &ctx);
5023 if (ctx.found_idx == -1)
5026 *data = ctx.found_data;
5027 *data_len = ctx.found_data_len;
5029 kfree(ctx.found_data);
5031 return ctx.found_idx;
5035 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
5036 const char *name, int name_len,
5037 const char *data, int data_len,
5041 struct send_ctx *sctx = ctx;
5042 char *found_data = NULL;
5043 int found_data_len = 0;
5045 ret = find_xattr(sctx->parent_root, sctx->right_path,
5046 sctx->cmp_key, name, name_len, &found_data,
5048 if (ret == -ENOENT) {
5049 ret = __process_new_xattr(num, di_key, name, name_len, data,
5051 } else if (ret >= 0) {
5052 if (data_len != found_data_len ||
5053 memcmp(data, found_data, data_len)) {
5054 ret = __process_new_xattr(num, di_key, name, name_len,
5055 data, data_len, ctx);
5065 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
5066 const char *name, int name_len,
5067 const char *data, int data_len,
5071 struct send_ctx *sctx = ctx;
5073 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
5074 name, name_len, NULL, NULL);
5076 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
5084 static int process_changed_xattr(struct send_ctx *sctx)
5088 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
5089 __process_changed_new_xattr, sctx);
5092 ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
5093 __process_changed_deleted_xattr, sctx);
5099 static int process_all_new_xattrs(struct send_ctx *sctx)
5103 struct btrfs_root *root;
5104 struct btrfs_path *path;
5105 struct btrfs_key key;
5106 struct btrfs_key found_key;
5108 path = alloc_path_for_send();
5112 root = sctx->send_root;
5114 key.objectid = sctx->cmp_key->objectid;
5115 key.type = BTRFS_XATTR_ITEM_KEY;
5117 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
5118 if (found_key.objectid != key.objectid ||
5119 found_key.type != key.type) {
5124 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
5128 /* Catch error found during iteration */
5132 btrfs_free_path(path);
5136 static int send_verity(struct send_ctx *sctx, struct fs_path *path,
5137 struct fsverity_descriptor *desc)
5141 ret = begin_cmd(sctx, BTRFS_SEND_C_ENABLE_VERITY);
5145 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
5146 TLV_PUT_U8(sctx, BTRFS_SEND_A_VERITY_ALGORITHM,
5147 le8_to_cpu(desc->hash_algorithm));
5148 TLV_PUT_U32(sctx, BTRFS_SEND_A_VERITY_BLOCK_SIZE,
5149 1U << le8_to_cpu(desc->log_blocksize));
5150 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SALT_DATA, desc->salt,
5151 le8_to_cpu(desc->salt_size));
5152 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SIG_DATA, desc->signature,
5153 le32_to_cpu(desc->sig_size));
5155 ret = send_cmd(sctx);
5162 static int process_verity(struct send_ctx *sctx)
5165 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5166 struct inode *inode;
5169 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, sctx->send_root);
5171 return PTR_ERR(inode);
5173 ret = btrfs_get_verity_descriptor(inode, NULL, 0);
5177 if (ret > FS_VERITY_MAX_DESCRIPTOR_SIZE) {
5181 if (!sctx->verity_descriptor) {
5182 sctx->verity_descriptor = kvmalloc(FS_VERITY_MAX_DESCRIPTOR_SIZE,
5184 if (!sctx->verity_descriptor) {
5190 ret = btrfs_get_verity_descriptor(inode, sctx->verity_descriptor, ret);
5194 p = fs_path_alloc();
5199 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5203 ret = send_verity(sctx, p, sctx->verity_descriptor);
5214 static inline u64 max_send_read_size(const struct send_ctx *sctx)
5216 return sctx->send_max_size - SZ_16K;
5219 static int put_data_header(struct send_ctx *sctx, u32 len)
5221 if (WARN_ON_ONCE(sctx->put_data))
5223 sctx->put_data = true;
5224 if (sctx->proto >= 2) {
5226 * Since v2, the data attribute header doesn't include a length,
5227 * it is implicitly to the end of the command.
5229 if (sctx->send_max_size - sctx->send_size < sizeof(__le16) + len)
5231 put_unaligned_le16(BTRFS_SEND_A_DATA, sctx->send_buf + sctx->send_size);
5232 sctx->send_size += sizeof(__le16);
5234 struct btrfs_tlv_header *hdr;
5236 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
5238 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
5239 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
5240 put_unaligned_le16(len, &hdr->tlv_len);
5241 sctx->send_size += sizeof(*hdr);
5246 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
5248 struct btrfs_root *root = sctx->send_root;
5249 struct btrfs_fs_info *fs_info = root->fs_info;
5251 pgoff_t index = offset >> PAGE_SHIFT;
5253 unsigned pg_offset = offset_in_page(offset);
5256 ret = put_data_header(sctx, len);
5260 last_index = (offset + len - 1) >> PAGE_SHIFT;
5262 while (index <= last_index) {
5263 unsigned cur_len = min_t(unsigned, len,
5264 PAGE_SIZE - pg_offset);
5266 page = find_lock_page(sctx->cur_inode->i_mapping, index);
5268 page_cache_sync_readahead(sctx->cur_inode->i_mapping,
5269 &sctx->ra, NULL, index,
5270 last_index + 1 - index);
5272 page = find_or_create_page(sctx->cur_inode->i_mapping,
5280 if (PageReadahead(page))
5281 page_cache_async_readahead(sctx->cur_inode->i_mapping,
5282 &sctx->ra, NULL, page_folio(page),
5283 index, last_index + 1 - index);
5285 if (!PageUptodate(page)) {
5286 btrfs_read_folio(NULL, page_folio(page));
5288 if (!PageUptodate(page)) {
5291 "send: IO error at offset %llu for inode %llu root %llu",
5292 page_offset(page), sctx->cur_ino,
5293 sctx->send_root->root_key.objectid);
5300 memcpy_from_page(sctx->send_buf + sctx->send_size, page,
5301 pg_offset, cur_len);
5307 sctx->send_size += cur_len;
5314 * Read some bytes from the current inode/file and send a write command to
5317 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5319 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5323 p = fs_path_alloc();
5327 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5329 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5333 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5337 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5338 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5339 ret = put_file_data(sctx, offset, len);
5343 ret = send_cmd(sctx);
5352 * Send a clone command to user space.
5354 static int send_clone(struct send_ctx *sctx,
5355 u64 offset, u32 len,
5356 struct clone_root *clone_root)
5362 btrfs_debug(sctx->send_root->fs_info,
5363 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5364 offset, len, clone_root->root->root_key.objectid,
5365 clone_root->ino, clone_root->offset);
5367 p = fs_path_alloc();
5371 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5375 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5379 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5380 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5381 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5383 if (clone_root->root == sctx->send_root) {
5384 ret = get_inode_gen(sctx->send_root, clone_root->ino, &gen);
5387 ret = get_cur_path(sctx, clone_root->ino, gen, p);
5389 ret = get_inode_path(clone_root->root, clone_root->ino, p);
5395 * If the parent we're using has a received_uuid set then use that as
5396 * our clone source as that is what we will look for when doing a
5399 * This covers the case that we create a snapshot off of a received
5400 * subvolume and then use that as the parent and try to receive on a
5403 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5404 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5405 clone_root->root->root_item.received_uuid);
5407 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5408 clone_root->root->root_item.uuid);
5409 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5410 btrfs_root_ctransid(&clone_root->root->root_item));
5411 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5412 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5413 clone_root->offset);
5415 ret = send_cmd(sctx);
5424 * Send an update extent command to user space.
5426 static int send_update_extent(struct send_ctx *sctx,
5427 u64 offset, u32 len)
5432 p = fs_path_alloc();
5436 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5440 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5444 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5445 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5446 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5448 ret = send_cmd(sctx);
5456 static int send_hole(struct send_ctx *sctx, u64 end)
5458 struct fs_path *p = NULL;
5459 u64 read_size = max_send_read_size(sctx);
5460 u64 offset = sctx->cur_inode_last_extent;
5464 * A hole that starts at EOF or beyond it. Since we do not yet support
5465 * fallocate (for extent preallocation and hole punching), sending a
5466 * write of zeroes starting at EOF or beyond would later require issuing
5467 * a truncate operation which would undo the write and achieve nothing.
5469 if (offset >= sctx->cur_inode_size)
5473 * Don't go beyond the inode's i_size due to prealloc extents that start
5476 end = min_t(u64, end, sctx->cur_inode_size);
5478 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5479 return send_update_extent(sctx, offset, end - offset);
5481 p = fs_path_alloc();
5484 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5486 goto tlv_put_failure;
5487 while (offset < end) {
5488 u64 len = min(end - offset, read_size);
5490 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5493 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5494 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5495 ret = put_data_header(sctx, len);
5498 memset(sctx->send_buf + sctx->send_size, 0, len);
5499 sctx->send_size += len;
5500 ret = send_cmd(sctx);
5505 sctx->cur_inode_next_write_offset = offset;
5511 static int send_encoded_inline_extent(struct send_ctx *sctx,
5512 struct btrfs_path *path, u64 offset,
5515 struct btrfs_root *root = sctx->send_root;
5516 struct btrfs_fs_info *fs_info = root->fs_info;
5517 struct inode *inode;
5518 struct fs_path *fspath;
5519 struct extent_buffer *leaf = path->nodes[0];
5520 struct btrfs_key key;
5521 struct btrfs_file_extent_item *ei;
5526 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5528 return PTR_ERR(inode);
5530 fspath = fs_path_alloc();
5536 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5540 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5544 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5545 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5546 ram_bytes = btrfs_file_extent_ram_bytes(leaf, ei);
5547 inline_size = btrfs_file_extent_inline_item_len(leaf, path->slots[0]);
5549 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5550 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5551 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5552 min(key.offset + ram_bytes - offset, len));
5553 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN, ram_bytes);
5554 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET, offset - key.offset);
5555 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5556 btrfs_file_extent_compression(leaf, ei));
5559 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5561 ret = put_data_header(sctx, inline_size);
5564 read_extent_buffer(leaf, sctx->send_buf + sctx->send_size,
5565 btrfs_file_extent_inline_start(ei), inline_size);
5566 sctx->send_size += inline_size;
5568 ret = send_cmd(sctx);
5572 fs_path_free(fspath);
5577 static int send_encoded_extent(struct send_ctx *sctx, struct btrfs_path *path,
5578 u64 offset, u64 len)
5580 struct btrfs_root *root = sctx->send_root;
5581 struct btrfs_fs_info *fs_info = root->fs_info;
5582 struct inode *inode;
5583 struct fs_path *fspath;
5584 struct extent_buffer *leaf = path->nodes[0];
5585 struct btrfs_key key;
5586 struct btrfs_file_extent_item *ei;
5587 u64 disk_bytenr, disk_num_bytes;
5589 struct btrfs_cmd_header *hdr;
5593 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5595 return PTR_ERR(inode);
5597 fspath = fs_path_alloc();
5603 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5607 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5611 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5612 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5613 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
5614 disk_num_bytes = btrfs_file_extent_disk_num_bytes(leaf, ei);
5616 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5617 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5618 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5619 min(key.offset + btrfs_file_extent_num_bytes(leaf, ei) - offset,
5621 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN,
5622 btrfs_file_extent_ram_bytes(leaf, ei));
5623 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET,
5624 offset - key.offset + btrfs_file_extent_offset(leaf, ei));
5625 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5626 btrfs_file_extent_compression(leaf, ei));
5629 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5630 TLV_PUT_U32(sctx, BTRFS_SEND_A_ENCRYPTION, 0);
5632 ret = put_data_header(sctx, disk_num_bytes);
5637 * We want to do I/O directly into the send buffer, so get the next page
5638 * boundary in the send buffer. This means that there may be a gap
5639 * between the beginning of the command and the file data.
5641 data_offset = PAGE_ALIGN(sctx->send_size);
5642 if (data_offset > sctx->send_max_size ||
5643 sctx->send_max_size - data_offset < disk_num_bytes) {
5649 * Note that send_buf is a mapping of send_buf_pages, so this is really
5650 * reading into send_buf.
5652 ret = btrfs_encoded_read_regular_fill_pages(BTRFS_I(inode), offset,
5653 disk_bytenr, disk_num_bytes,
5654 sctx->send_buf_pages +
5655 (data_offset >> PAGE_SHIFT));
5659 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
5660 hdr->len = cpu_to_le32(sctx->send_size + disk_num_bytes - sizeof(*hdr));
5662 crc = btrfs_crc32c(0, sctx->send_buf, sctx->send_size);
5663 crc = btrfs_crc32c(crc, sctx->send_buf + data_offset, disk_num_bytes);
5664 hdr->crc = cpu_to_le32(crc);
5666 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
5669 ret = write_buf(sctx->send_filp, sctx->send_buf + data_offset,
5670 disk_num_bytes, &sctx->send_off);
5672 sctx->send_size = 0;
5673 sctx->put_data = false;
5677 fs_path_free(fspath);
5682 static int send_extent_data(struct send_ctx *sctx, struct btrfs_path *path,
5683 const u64 offset, const u64 len)
5685 const u64 end = offset + len;
5686 struct extent_buffer *leaf = path->nodes[0];
5687 struct btrfs_file_extent_item *ei;
5688 u64 read_size = max_send_read_size(sctx);
5691 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5692 return send_update_extent(sctx, offset, len);
5694 ei = btrfs_item_ptr(leaf, path->slots[0],
5695 struct btrfs_file_extent_item);
5696 if ((sctx->flags & BTRFS_SEND_FLAG_COMPRESSED) &&
5697 btrfs_file_extent_compression(leaf, ei) != BTRFS_COMPRESS_NONE) {
5698 bool is_inline = (btrfs_file_extent_type(leaf, ei) ==
5699 BTRFS_FILE_EXTENT_INLINE);
5702 * Send the compressed extent unless the compressed data is
5703 * larger than the decompressed data. This can happen if we're
5704 * not sending the entire extent, either because it has been
5705 * partially overwritten/truncated or because this is a part of
5706 * the extent that we couldn't clone in clone_range().
5709 btrfs_file_extent_inline_item_len(leaf,
5710 path->slots[0]) <= len) {
5711 return send_encoded_inline_extent(sctx, path, offset,
5713 } else if (!is_inline &&
5714 btrfs_file_extent_disk_num_bytes(leaf, ei) <= len) {
5715 return send_encoded_extent(sctx, path, offset, len);
5719 if (sctx->cur_inode == NULL) {
5720 struct btrfs_root *root = sctx->send_root;
5722 sctx->cur_inode = btrfs_iget(root->fs_info->sb, sctx->cur_ino, root);
5723 if (IS_ERR(sctx->cur_inode)) {
5724 int err = PTR_ERR(sctx->cur_inode);
5726 sctx->cur_inode = NULL;
5729 memset(&sctx->ra, 0, sizeof(struct file_ra_state));
5730 file_ra_state_init(&sctx->ra, sctx->cur_inode->i_mapping);
5733 * It's very likely there are no pages from this inode in the page
5734 * cache, so after reading extents and sending their data, we clean
5735 * the page cache to avoid trashing the page cache (adding pressure
5736 * to the page cache and forcing eviction of other data more useful
5737 * for applications).
5739 * We decide if we should clean the page cache simply by checking
5740 * if the inode's mapping nrpages is 0 when we first open it, and
5741 * not by using something like filemap_range_has_page() before
5742 * reading an extent because when we ask the readahead code to
5743 * read a given file range, it may (and almost always does) read
5744 * pages from beyond that range (see the documentation for
5745 * page_cache_sync_readahead()), so it would not be reliable,
5746 * because after reading the first extent future calls to
5747 * filemap_range_has_page() would return true because the readahead
5748 * on the previous extent resulted in reading pages of the current
5751 sctx->clean_page_cache = (sctx->cur_inode->i_mapping->nrpages == 0);
5752 sctx->page_cache_clear_start = round_down(offset, PAGE_SIZE);
5755 while (sent < len) {
5756 u64 size = min(len - sent, read_size);
5759 ret = send_write(sctx, offset + sent, size);
5765 if (sctx->clean_page_cache && PAGE_ALIGNED(end)) {
5767 * Always operate only on ranges that are a multiple of the page
5768 * size. This is not only to prevent zeroing parts of a page in
5769 * the case of subpage sector size, but also to guarantee we evict
5770 * pages, as passing a range that is smaller than page size does
5771 * not evict the respective page (only zeroes part of its content).
5773 * Always start from the end offset of the last range cleared.
5774 * This is because the readahead code may (and very often does)
5775 * reads pages beyond the range we request for readahead. So if
5776 * we have an extent layout like this:
5778 * [ extent A ] [ extent B ] [ extent C ]
5780 * When we ask page_cache_sync_readahead() to read extent A, it
5781 * may also trigger reads for pages of extent B. If we are doing
5782 * an incremental send and extent B has not changed between the
5783 * parent and send snapshots, some or all of its pages may end
5784 * up being read and placed in the page cache. So when truncating
5785 * the page cache we always start from the end offset of the
5786 * previously processed extent up to the end of the current
5789 truncate_inode_pages_range(&sctx->cur_inode->i_data,
5790 sctx->page_cache_clear_start,
5792 sctx->page_cache_clear_start = end;
5799 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5800 * found, call send_set_xattr function to emit it.
5802 * Return 0 if there isn't a capability, or when the capability was emitted
5803 * successfully, or < 0 if an error occurred.
5805 static int send_capabilities(struct send_ctx *sctx)
5807 struct fs_path *fspath = NULL;
5808 struct btrfs_path *path;
5809 struct btrfs_dir_item *di;
5810 struct extent_buffer *leaf;
5811 unsigned long data_ptr;
5816 path = alloc_path_for_send();
5820 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5821 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5823 /* There is no xattr for this inode */
5825 } else if (IS_ERR(di)) {
5830 leaf = path->nodes[0];
5831 buf_len = btrfs_dir_data_len(leaf, di);
5833 fspath = fs_path_alloc();
5834 buf = kmalloc(buf_len, GFP_KERNEL);
5835 if (!fspath || !buf) {
5840 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5844 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5845 read_extent_buffer(leaf, buf, data_ptr, buf_len);
5847 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5848 strlen(XATTR_NAME_CAPS), buf, buf_len);
5851 fs_path_free(fspath);
5852 btrfs_free_path(path);
5856 static int clone_range(struct send_ctx *sctx, struct btrfs_path *dst_path,
5857 struct clone_root *clone_root, const u64 disk_byte,
5858 u64 data_offset, u64 offset, u64 len)
5860 struct btrfs_path *path;
5861 struct btrfs_key key;
5863 struct btrfs_inode_info info;
5864 u64 clone_src_i_size = 0;
5867 * Prevent cloning from a zero offset with a length matching the sector
5868 * size because in some scenarios this will make the receiver fail.
5870 * For example, if in the source filesystem the extent at offset 0
5871 * has a length of sectorsize and it was written using direct IO, then
5872 * it can never be an inline extent (even if compression is enabled).
5873 * Then this extent can be cloned in the original filesystem to a non
5874 * zero file offset, but it may not be possible to clone in the
5875 * destination filesystem because it can be inlined due to compression
5876 * on the destination filesystem (as the receiver's write operations are
5877 * always done using buffered IO). The same happens when the original
5878 * filesystem does not have compression enabled but the destination
5881 if (clone_root->offset == 0 &&
5882 len == sctx->send_root->fs_info->sectorsize)
5883 return send_extent_data(sctx, dst_path, offset, len);
5885 path = alloc_path_for_send();
5890 * There are inodes that have extents that lie behind its i_size. Don't
5891 * accept clones from these extents.
5893 ret = get_inode_info(clone_root->root, clone_root->ino, &info);
5894 btrfs_release_path(path);
5897 clone_src_i_size = info.size;
5900 * We can't send a clone operation for the entire range if we find
5901 * extent items in the respective range in the source file that
5902 * refer to different extents or if we find holes.
5903 * So check for that and do a mix of clone and regular write/copy
5904 * operations if needed.
5908 * mkfs.btrfs -f /dev/sda
5909 * mount /dev/sda /mnt
5910 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5911 * cp --reflink=always /mnt/foo /mnt/bar
5912 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5913 * btrfs subvolume snapshot -r /mnt /mnt/snap
5915 * If when we send the snapshot and we are processing file bar (which
5916 * has a higher inode number than foo) we blindly send a clone operation
5917 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5918 * a file bar that matches the content of file foo - iow, doesn't match
5919 * the content from bar in the original filesystem.
5921 key.objectid = clone_root->ino;
5922 key.type = BTRFS_EXTENT_DATA_KEY;
5923 key.offset = clone_root->offset;
5924 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5927 if (ret > 0 && path->slots[0] > 0) {
5928 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5929 if (key.objectid == clone_root->ino &&
5930 key.type == BTRFS_EXTENT_DATA_KEY)
5935 struct extent_buffer *leaf = path->nodes[0];
5936 int slot = path->slots[0];
5937 struct btrfs_file_extent_item *ei;
5941 u64 clone_data_offset;
5942 bool crossed_src_i_size = false;
5944 if (slot >= btrfs_header_nritems(leaf)) {
5945 ret = btrfs_next_leaf(clone_root->root, path);
5953 btrfs_item_key_to_cpu(leaf, &key, slot);
5956 * We might have an implicit trailing hole (NO_HOLES feature
5957 * enabled). We deal with it after leaving this loop.
5959 if (key.objectid != clone_root->ino ||
5960 key.type != BTRFS_EXTENT_DATA_KEY)
5963 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5964 type = btrfs_file_extent_type(leaf, ei);
5965 if (type == BTRFS_FILE_EXTENT_INLINE) {
5966 ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5967 ext_len = PAGE_ALIGN(ext_len);
5969 ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5972 if (key.offset + ext_len <= clone_root->offset)
5975 if (key.offset > clone_root->offset) {
5976 /* Implicit hole, NO_HOLES feature enabled. */
5977 u64 hole_len = key.offset - clone_root->offset;
5981 ret = send_extent_data(sctx, dst_path, offset,
5990 clone_root->offset += hole_len;
5991 data_offset += hole_len;
5994 if (key.offset >= clone_root->offset + len)
5997 if (key.offset >= clone_src_i_size)
6000 if (key.offset + ext_len > clone_src_i_size) {
6001 ext_len = clone_src_i_size - key.offset;
6002 crossed_src_i_size = true;
6005 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
6006 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
6007 clone_root->offset = key.offset;
6008 if (clone_data_offset < data_offset &&
6009 clone_data_offset + ext_len > data_offset) {
6012 extent_offset = data_offset - clone_data_offset;
6013 ext_len -= extent_offset;
6014 clone_data_offset += extent_offset;
6015 clone_root->offset += extent_offset;
6019 clone_len = min_t(u64, ext_len, len);
6021 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
6022 clone_data_offset == data_offset) {
6023 const u64 src_end = clone_root->offset + clone_len;
6024 const u64 sectorsize = SZ_64K;
6027 * We can't clone the last block, when its size is not
6028 * sector size aligned, into the middle of a file. If we
6029 * do so, the receiver will get a failure (-EINVAL) when
6030 * trying to clone or will silently corrupt the data in
6031 * the destination file if it's on a kernel without the
6032 * fix introduced by commit ac765f83f1397646
6033 * ("Btrfs: fix data corruption due to cloning of eof
6036 * So issue a clone of the aligned down range plus a
6037 * regular write for the eof block, if we hit that case.
6039 * Also, we use the maximum possible sector size, 64K,
6040 * because we don't know what's the sector size of the
6041 * filesystem that receives the stream, so we have to
6042 * assume the largest possible sector size.
6044 if (src_end == clone_src_i_size &&
6045 !IS_ALIGNED(src_end, sectorsize) &&
6046 offset + clone_len < sctx->cur_inode_size) {
6049 slen = ALIGN_DOWN(src_end - clone_root->offset,
6052 ret = send_clone(sctx, offset, slen,
6057 ret = send_extent_data(sctx, dst_path,
6061 ret = send_clone(sctx, offset, clone_len,
6064 } else if (crossed_src_i_size && clone_len < len) {
6066 * If we are at i_size of the clone source inode and we
6067 * can not clone from it, terminate the loop. This is
6068 * to avoid sending two write operations, one with a
6069 * length matching clone_len and the final one after
6070 * this loop with a length of len - clone_len.
6072 * When using encoded writes (BTRFS_SEND_FLAG_COMPRESSED
6073 * was passed to the send ioctl), this helps avoid
6074 * sending an encoded write for an offset that is not
6075 * sector size aligned, in case the i_size of the source
6076 * inode is not sector size aligned. That will make the
6077 * receiver fallback to decompression of the data and
6078 * writing it using regular buffered IO, therefore while
6079 * not incorrect, it's not optimal due decompression and
6080 * possible re-compression at the receiver.
6084 ret = send_extent_data(sctx, dst_path, offset,
6094 offset += clone_len;
6095 clone_root->offset += clone_len;
6098 * If we are cloning from the file we are currently processing,
6099 * and using the send root as the clone root, we must stop once
6100 * the current clone offset reaches the current eof of the file
6101 * at the receiver, otherwise we would issue an invalid clone
6102 * operation (source range going beyond eof) and cause the
6103 * receiver to fail. So if we reach the current eof, bail out
6104 * and fallback to a regular write.
6106 if (clone_root->root == sctx->send_root &&
6107 clone_root->ino == sctx->cur_ino &&
6108 clone_root->offset >= sctx->cur_inode_next_write_offset)
6111 data_offset += clone_len;
6117 ret = send_extent_data(sctx, dst_path, offset, len);
6121 btrfs_free_path(path);
6125 static int send_write_or_clone(struct send_ctx *sctx,
6126 struct btrfs_path *path,
6127 struct btrfs_key *key,
6128 struct clone_root *clone_root)
6131 u64 offset = key->offset;
6133 u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
6135 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
6139 if (clone_root && IS_ALIGNED(end, bs)) {
6140 struct btrfs_file_extent_item *ei;
6144 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6145 struct btrfs_file_extent_item);
6146 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
6147 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
6148 ret = clone_range(sctx, path, clone_root, disk_byte,
6149 data_offset, offset, end - offset);
6151 ret = send_extent_data(sctx, path, offset, end - offset);
6153 sctx->cur_inode_next_write_offset = end;
6157 static int is_extent_unchanged(struct send_ctx *sctx,
6158 struct btrfs_path *left_path,
6159 struct btrfs_key *ekey)
6162 struct btrfs_key key;
6163 struct btrfs_path *path = NULL;
6164 struct extent_buffer *eb;
6166 struct btrfs_key found_key;
6167 struct btrfs_file_extent_item *ei;
6172 u64 left_offset_fixed;
6180 path = alloc_path_for_send();
6184 eb = left_path->nodes[0];
6185 slot = left_path->slots[0];
6186 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6187 left_type = btrfs_file_extent_type(eb, ei);
6189 if (left_type != BTRFS_FILE_EXTENT_REG) {
6193 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6194 left_len = btrfs_file_extent_num_bytes(eb, ei);
6195 left_offset = btrfs_file_extent_offset(eb, ei);
6196 left_gen = btrfs_file_extent_generation(eb, ei);
6199 * Following comments will refer to these graphics. L is the left
6200 * extents which we are checking at the moment. 1-8 are the right
6201 * extents that we iterate.
6204 * |-1-|-2a-|-3-|-4-|-5-|-6-|
6207 * |--1--|-2b-|...(same as above)
6209 * Alternative situation. Happens on files where extents got split.
6211 * |-----------7-----------|-6-|
6213 * Alternative situation. Happens on files which got larger.
6216 * Nothing follows after 8.
6219 key.objectid = ekey->objectid;
6220 key.type = BTRFS_EXTENT_DATA_KEY;
6221 key.offset = ekey->offset;
6222 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
6231 * Handle special case where the right side has no extents at all.
6233 eb = path->nodes[0];
6234 slot = path->slots[0];
6235 btrfs_item_key_to_cpu(eb, &found_key, slot);
6236 if (found_key.objectid != key.objectid ||
6237 found_key.type != key.type) {
6238 /* If we're a hole then just pretend nothing changed */
6239 ret = (left_disknr) ? 0 : 1;
6244 * We're now on 2a, 2b or 7.
6247 while (key.offset < ekey->offset + left_len) {
6248 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6249 right_type = btrfs_file_extent_type(eb, ei);
6250 if (right_type != BTRFS_FILE_EXTENT_REG &&
6251 right_type != BTRFS_FILE_EXTENT_INLINE) {
6256 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6257 right_len = btrfs_file_extent_ram_bytes(eb, ei);
6258 right_len = PAGE_ALIGN(right_len);
6260 right_len = btrfs_file_extent_num_bytes(eb, ei);
6264 * Are we at extent 8? If yes, we know the extent is changed.
6265 * This may only happen on the first iteration.
6267 if (found_key.offset + right_len <= ekey->offset) {
6268 /* If we're a hole just pretend nothing changed */
6269 ret = (left_disknr) ? 0 : 1;
6274 * We just wanted to see if when we have an inline extent, what
6275 * follows it is a regular extent (wanted to check the above
6276 * condition for inline extents too). This should normally not
6277 * happen but it's possible for example when we have an inline
6278 * compressed extent representing data with a size matching
6279 * the page size (currently the same as sector size).
6281 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6286 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6287 right_offset = btrfs_file_extent_offset(eb, ei);
6288 right_gen = btrfs_file_extent_generation(eb, ei);
6290 left_offset_fixed = left_offset;
6291 if (key.offset < ekey->offset) {
6292 /* Fix the right offset for 2a and 7. */
6293 right_offset += ekey->offset - key.offset;
6295 /* Fix the left offset for all behind 2a and 2b */
6296 left_offset_fixed += key.offset - ekey->offset;
6300 * Check if we have the same extent.
6302 if (left_disknr != right_disknr ||
6303 left_offset_fixed != right_offset ||
6304 left_gen != right_gen) {
6310 * Go to the next extent.
6312 ret = btrfs_next_item(sctx->parent_root, path);
6316 eb = path->nodes[0];
6317 slot = path->slots[0];
6318 btrfs_item_key_to_cpu(eb, &found_key, slot);
6320 if (ret || found_key.objectid != key.objectid ||
6321 found_key.type != key.type) {
6322 key.offset += right_len;
6325 if (found_key.offset != key.offset + right_len) {
6333 * We're now behind the left extent (treat as unchanged) or at the end
6334 * of the right side (treat as changed).
6336 if (key.offset >= ekey->offset + left_len)
6343 btrfs_free_path(path);
6347 static int get_last_extent(struct send_ctx *sctx, u64 offset)
6349 struct btrfs_path *path;
6350 struct btrfs_root *root = sctx->send_root;
6351 struct btrfs_key key;
6354 path = alloc_path_for_send();
6358 sctx->cur_inode_last_extent = 0;
6360 key.objectid = sctx->cur_ino;
6361 key.type = BTRFS_EXTENT_DATA_KEY;
6362 key.offset = offset;
6363 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
6367 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
6368 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
6371 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6373 btrfs_free_path(path);
6377 static int range_is_hole_in_parent(struct send_ctx *sctx,
6381 struct btrfs_path *path;
6382 struct btrfs_key key;
6383 struct btrfs_root *root = sctx->parent_root;
6384 u64 search_start = start;
6387 path = alloc_path_for_send();
6391 key.objectid = sctx->cur_ino;
6392 key.type = BTRFS_EXTENT_DATA_KEY;
6393 key.offset = search_start;
6394 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6397 if (ret > 0 && path->slots[0] > 0)
6400 while (search_start < end) {
6401 struct extent_buffer *leaf = path->nodes[0];
6402 int slot = path->slots[0];
6403 struct btrfs_file_extent_item *fi;
6406 if (slot >= btrfs_header_nritems(leaf)) {
6407 ret = btrfs_next_leaf(root, path);
6415 btrfs_item_key_to_cpu(leaf, &key, slot);
6416 if (key.objectid < sctx->cur_ino ||
6417 key.type < BTRFS_EXTENT_DATA_KEY)
6419 if (key.objectid > sctx->cur_ino ||
6420 key.type > BTRFS_EXTENT_DATA_KEY ||
6424 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6425 extent_end = btrfs_file_extent_end(path);
6426 if (extent_end <= start)
6428 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
6429 search_start = extent_end;
6439 btrfs_free_path(path);
6443 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
6444 struct btrfs_key *key)
6448 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
6451 if (sctx->cur_inode_last_extent == (u64)-1) {
6452 ret = get_last_extent(sctx, key->offset - 1);
6457 if (path->slots[0] == 0 &&
6458 sctx->cur_inode_last_extent < key->offset) {
6460 * We might have skipped entire leafs that contained only
6461 * file extent items for our current inode. These leafs have
6462 * a generation number smaller (older) than the one in the
6463 * current leaf and the leaf our last extent came from, and
6464 * are located between these 2 leafs.
6466 ret = get_last_extent(sctx, key->offset - 1);
6471 if (sctx->cur_inode_last_extent < key->offset) {
6472 ret = range_is_hole_in_parent(sctx,
6473 sctx->cur_inode_last_extent,
6478 ret = send_hole(sctx, key->offset);
6482 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6486 static int process_extent(struct send_ctx *sctx,
6487 struct btrfs_path *path,
6488 struct btrfs_key *key)
6490 struct clone_root *found_clone = NULL;
6493 if (S_ISLNK(sctx->cur_inode_mode))
6496 if (sctx->parent_root && !sctx->cur_inode_new) {
6497 ret = is_extent_unchanged(sctx, path, key);
6505 struct btrfs_file_extent_item *ei;
6508 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6509 struct btrfs_file_extent_item);
6510 type = btrfs_file_extent_type(path->nodes[0], ei);
6511 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
6512 type == BTRFS_FILE_EXTENT_REG) {
6514 * The send spec does not have a prealloc command yet,
6515 * so just leave a hole for prealloc'ed extents until
6516 * we have enough commands queued up to justify rev'ing
6519 if (type == BTRFS_FILE_EXTENT_PREALLOC) {
6524 /* Have a hole, just skip it. */
6525 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
6532 ret = find_extent_clone(sctx, path, key->objectid, key->offset,
6533 sctx->cur_inode_size, &found_clone);
6534 if (ret != -ENOENT && ret < 0)
6537 ret = send_write_or_clone(sctx, path, key, found_clone);
6541 ret = maybe_send_hole(sctx, path, key);
6546 static int process_all_extents(struct send_ctx *sctx)
6550 struct btrfs_root *root;
6551 struct btrfs_path *path;
6552 struct btrfs_key key;
6553 struct btrfs_key found_key;
6555 root = sctx->send_root;
6556 path = alloc_path_for_send();
6560 key.objectid = sctx->cmp_key->objectid;
6561 key.type = BTRFS_EXTENT_DATA_KEY;
6563 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
6564 if (found_key.objectid != key.objectid ||
6565 found_key.type != key.type) {
6570 ret = process_extent(sctx, path, &found_key);
6574 /* Catch error found during iteration */
6578 btrfs_free_path(path);
6582 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6584 int *refs_processed)
6588 if (sctx->cur_ino == 0)
6590 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6591 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6593 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6596 ret = process_recorded_refs(sctx, pending_move);
6600 *refs_processed = 1;
6605 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6608 struct btrfs_inode_info info;
6619 bool need_fileattr = false;
6620 int need_truncate = 1;
6621 int pending_move = 0;
6622 int refs_processed = 0;
6624 if (sctx->ignore_cur_inode)
6627 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6633 * We have processed the refs and thus need to advance send_progress.
6634 * Now, calls to get_cur_xxx will take the updated refs of the current
6635 * inode into account.
6637 * On the other hand, if our current inode is a directory and couldn't
6638 * be moved/renamed because its parent was renamed/moved too and it has
6639 * a higher inode number, we can only move/rename our current inode
6640 * after we moved/renamed its parent. Therefore in this case operate on
6641 * the old path (pre move/rename) of our current inode, and the
6642 * move/rename will be performed later.
6644 if (refs_processed && !pending_move)
6645 sctx->send_progress = sctx->cur_ino + 1;
6647 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6649 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6651 ret = get_inode_info(sctx->send_root, sctx->cur_ino, &info);
6654 left_mode = info.mode;
6655 left_uid = info.uid;
6656 left_gid = info.gid;
6657 left_fileattr = info.fileattr;
6659 if (!sctx->parent_root || sctx->cur_inode_new) {
6661 if (!S_ISLNK(sctx->cur_inode_mode))
6663 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6668 ret = get_inode_info(sctx->parent_root, sctx->cur_ino, &info);
6671 old_size = info.size;
6672 right_mode = info.mode;
6673 right_uid = info.uid;
6674 right_gid = info.gid;
6675 right_fileattr = info.fileattr;
6677 if (left_uid != right_uid || left_gid != right_gid)
6679 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6681 if (!S_ISLNK(sctx->cur_inode_mode) && left_fileattr != right_fileattr)
6682 need_fileattr = true;
6683 if ((old_size == sctx->cur_inode_size) ||
6684 (sctx->cur_inode_size > old_size &&
6685 sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6689 if (S_ISREG(sctx->cur_inode_mode)) {
6690 if (need_send_hole(sctx)) {
6691 if (sctx->cur_inode_last_extent == (u64)-1 ||
6692 sctx->cur_inode_last_extent <
6693 sctx->cur_inode_size) {
6694 ret = get_last_extent(sctx, (u64)-1);
6698 if (sctx->cur_inode_last_extent <
6699 sctx->cur_inode_size) {
6700 ret = send_hole(sctx, sctx->cur_inode_size);
6705 if (need_truncate) {
6706 ret = send_truncate(sctx, sctx->cur_ino,
6707 sctx->cur_inode_gen,
6708 sctx->cur_inode_size);
6715 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6716 left_uid, left_gid);
6721 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6726 if (need_fileattr) {
6727 ret = send_fileattr(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6733 if (proto_cmd_ok(sctx, BTRFS_SEND_C_ENABLE_VERITY)
6734 && sctx->cur_inode_needs_verity) {
6735 ret = process_verity(sctx);
6740 ret = send_capabilities(sctx);
6745 * If other directory inodes depended on our current directory
6746 * inode's move/rename, now do their move/rename operations.
6748 if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6749 ret = apply_children_dir_moves(sctx);
6753 * Need to send that every time, no matter if it actually
6754 * changed between the two trees as we have done changes to
6755 * the inode before. If our inode is a directory and it's
6756 * waiting to be moved/renamed, we will send its utimes when
6757 * it's moved/renamed, therefore we don't need to do it here.
6759 sctx->send_progress = sctx->cur_ino + 1;
6762 * If the current inode is a non-empty directory, delay issuing
6763 * the utimes command for it, as it's very likely we have inodes
6764 * with an higher number inside it. We want to issue the utimes
6765 * command only after adding all dentries to it.
6767 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_size > 0)
6768 ret = cache_dir_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6770 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6778 ret = trim_dir_utimes_cache(sctx);
6783 static void close_current_inode(struct send_ctx *sctx)
6787 if (sctx->cur_inode == NULL)
6790 i_size = i_size_read(sctx->cur_inode);
6793 * If we are doing an incremental send, we may have extents between the
6794 * last processed extent and the i_size that have not been processed
6795 * because they haven't changed but we may have read some of their pages
6796 * through readahead, see the comments at send_extent_data().
6798 if (sctx->clean_page_cache && sctx->page_cache_clear_start < i_size)
6799 truncate_inode_pages_range(&sctx->cur_inode->i_data,
6800 sctx->page_cache_clear_start,
6801 round_up(i_size, PAGE_SIZE) - 1);
6803 iput(sctx->cur_inode);
6804 sctx->cur_inode = NULL;
6807 static int changed_inode(struct send_ctx *sctx,
6808 enum btrfs_compare_tree_result result)
6811 struct btrfs_key *key = sctx->cmp_key;
6812 struct btrfs_inode_item *left_ii = NULL;
6813 struct btrfs_inode_item *right_ii = NULL;
6817 close_current_inode(sctx);
6819 sctx->cur_ino = key->objectid;
6820 sctx->cur_inode_new_gen = false;
6821 sctx->cur_inode_last_extent = (u64)-1;
6822 sctx->cur_inode_next_write_offset = 0;
6823 sctx->ignore_cur_inode = false;
6826 * Set send_progress to current inode. This will tell all get_cur_xxx
6827 * functions that the current inode's refs are not updated yet. Later,
6828 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6830 sctx->send_progress = sctx->cur_ino;
6832 if (result == BTRFS_COMPARE_TREE_NEW ||
6833 result == BTRFS_COMPARE_TREE_CHANGED) {
6834 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6835 sctx->left_path->slots[0],
6836 struct btrfs_inode_item);
6837 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6840 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6841 sctx->right_path->slots[0],
6842 struct btrfs_inode_item);
6843 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6846 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6847 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6848 sctx->right_path->slots[0],
6849 struct btrfs_inode_item);
6851 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6855 * The cur_ino = root dir case is special here. We can't treat
6856 * the inode as deleted+reused because it would generate a
6857 * stream that tries to delete/mkdir the root dir.
6859 if (left_gen != right_gen &&
6860 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6861 sctx->cur_inode_new_gen = true;
6865 * Normally we do not find inodes with a link count of zero (orphans)
6866 * because the most common case is to create a snapshot and use it
6867 * for a send operation. However other less common use cases involve
6868 * using a subvolume and send it after turning it to RO mode just
6869 * after deleting all hard links of a file while holding an open
6870 * file descriptor against it or turning a RO snapshot into RW mode,
6871 * keep an open file descriptor against a file, delete it and then
6872 * turn the snapshot back to RO mode before using it for a send
6873 * operation. The former is what the receiver operation does.
6874 * Therefore, if we want to send these snapshots soon after they're
6875 * received, we need to handle orphan inodes as well. Moreover, orphans
6876 * can appear not only in the send snapshot but also in the parent
6877 * snapshot. Here are several cases:
6879 * Case 1: BTRFS_COMPARE_TREE_NEW
6880 * | send snapshot | action
6881 * --------------------------------
6882 * nlink | 0 | ignore
6884 * Case 2: BTRFS_COMPARE_TREE_DELETED
6885 * | parent snapshot | action
6886 * ----------------------------------
6887 * nlink | 0 | as usual
6888 * Note: No unlinks will be sent because there're no paths for it.
6890 * Case 3: BTRFS_COMPARE_TREE_CHANGED
6891 * | | parent snapshot | send snapshot | action
6892 * -----------------------------------------------------------------------
6893 * subcase 1 | nlink | 0 | 0 | ignore
6894 * subcase 2 | nlink | >0 | 0 | new_gen(deletion)
6895 * subcase 3 | nlink | 0 | >0 | new_gen(creation)
6898 if (result == BTRFS_COMPARE_TREE_NEW) {
6899 if (btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii) == 0) {
6900 sctx->ignore_cur_inode = true;
6903 sctx->cur_inode_gen = left_gen;
6904 sctx->cur_inode_new = true;
6905 sctx->cur_inode_deleted = false;
6906 sctx->cur_inode_size = btrfs_inode_size(
6907 sctx->left_path->nodes[0], left_ii);
6908 sctx->cur_inode_mode = btrfs_inode_mode(
6909 sctx->left_path->nodes[0], left_ii);
6910 sctx->cur_inode_rdev = btrfs_inode_rdev(
6911 sctx->left_path->nodes[0], left_ii);
6912 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6913 ret = send_create_inode_if_needed(sctx);
6914 } else if (result == BTRFS_COMPARE_TREE_DELETED) {
6915 sctx->cur_inode_gen = right_gen;
6916 sctx->cur_inode_new = false;
6917 sctx->cur_inode_deleted = true;
6918 sctx->cur_inode_size = btrfs_inode_size(
6919 sctx->right_path->nodes[0], right_ii);
6920 sctx->cur_inode_mode = btrfs_inode_mode(
6921 sctx->right_path->nodes[0], right_ii);
6922 } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6923 u32 new_nlinks, old_nlinks;
6925 new_nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6926 old_nlinks = btrfs_inode_nlink(sctx->right_path->nodes[0], right_ii);
6927 if (new_nlinks == 0 && old_nlinks == 0) {
6928 sctx->ignore_cur_inode = true;
6930 } else if (new_nlinks == 0 || old_nlinks == 0) {
6931 sctx->cur_inode_new_gen = 1;
6934 * We need to do some special handling in case the inode was
6935 * reported as changed with a changed generation number. This
6936 * means that the original inode was deleted and new inode
6937 * reused the same inum. So we have to treat the old inode as
6938 * deleted and the new one as new.
6940 if (sctx->cur_inode_new_gen) {
6942 * First, process the inode as if it was deleted.
6944 if (old_nlinks > 0) {
6945 sctx->cur_inode_gen = right_gen;
6946 sctx->cur_inode_new = false;
6947 sctx->cur_inode_deleted = true;
6948 sctx->cur_inode_size = btrfs_inode_size(
6949 sctx->right_path->nodes[0], right_ii);
6950 sctx->cur_inode_mode = btrfs_inode_mode(
6951 sctx->right_path->nodes[0], right_ii);
6952 ret = process_all_refs(sctx,
6953 BTRFS_COMPARE_TREE_DELETED);
6959 * Now process the inode as if it was new.
6961 if (new_nlinks > 0) {
6962 sctx->cur_inode_gen = left_gen;
6963 sctx->cur_inode_new = true;
6964 sctx->cur_inode_deleted = false;
6965 sctx->cur_inode_size = btrfs_inode_size(
6966 sctx->left_path->nodes[0],
6968 sctx->cur_inode_mode = btrfs_inode_mode(
6969 sctx->left_path->nodes[0],
6971 sctx->cur_inode_rdev = btrfs_inode_rdev(
6972 sctx->left_path->nodes[0],
6974 ret = send_create_inode_if_needed(sctx);
6978 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
6982 * Advance send_progress now as we did not get
6983 * into process_recorded_refs_if_needed in the
6986 sctx->send_progress = sctx->cur_ino + 1;
6989 * Now process all extents and xattrs of the
6990 * inode as if they were all new.
6992 ret = process_all_extents(sctx);
6995 ret = process_all_new_xattrs(sctx);
7000 sctx->cur_inode_gen = left_gen;
7001 sctx->cur_inode_new = false;
7002 sctx->cur_inode_new_gen = false;
7003 sctx->cur_inode_deleted = false;
7004 sctx->cur_inode_size = btrfs_inode_size(
7005 sctx->left_path->nodes[0], left_ii);
7006 sctx->cur_inode_mode = btrfs_inode_mode(
7007 sctx->left_path->nodes[0], left_ii);
7016 * We have to process new refs before deleted refs, but compare_trees gives us
7017 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
7018 * first and later process them in process_recorded_refs.
7019 * For the cur_inode_new_gen case, we skip recording completely because
7020 * changed_inode did already initiate processing of refs. The reason for this is
7021 * that in this case, compare_tree actually compares the refs of 2 different
7022 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
7023 * refs of the right tree as deleted and all refs of the left tree as new.
7025 static int changed_ref(struct send_ctx *sctx,
7026 enum btrfs_compare_tree_result result)
7030 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7031 inconsistent_snapshot_error(sctx, result, "reference");
7035 if (!sctx->cur_inode_new_gen &&
7036 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
7037 if (result == BTRFS_COMPARE_TREE_NEW)
7038 ret = record_new_ref(sctx);
7039 else if (result == BTRFS_COMPARE_TREE_DELETED)
7040 ret = record_deleted_ref(sctx);
7041 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7042 ret = record_changed_ref(sctx);
7049 * Process new/deleted/changed xattrs. We skip processing in the
7050 * cur_inode_new_gen case because changed_inode did already initiate processing
7051 * of xattrs. The reason is the same as in changed_ref
7053 static int changed_xattr(struct send_ctx *sctx,
7054 enum btrfs_compare_tree_result result)
7058 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7059 inconsistent_snapshot_error(sctx, result, "xattr");
7063 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7064 if (result == BTRFS_COMPARE_TREE_NEW)
7065 ret = process_new_xattr(sctx);
7066 else if (result == BTRFS_COMPARE_TREE_DELETED)
7067 ret = process_deleted_xattr(sctx);
7068 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7069 ret = process_changed_xattr(sctx);
7076 * Process new/deleted/changed extents. We skip processing in the
7077 * cur_inode_new_gen case because changed_inode did already initiate processing
7078 * of extents. The reason is the same as in changed_ref
7080 static int changed_extent(struct send_ctx *sctx,
7081 enum btrfs_compare_tree_result result)
7086 * We have found an extent item that changed without the inode item
7087 * having changed. This can happen either after relocation (where the
7088 * disk_bytenr of an extent item is replaced at
7089 * relocation.c:replace_file_extents()) or after deduplication into a
7090 * file in both the parent and send snapshots (where an extent item can
7091 * get modified or replaced with a new one). Note that deduplication
7092 * updates the inode item, but it only changes the iversion (sequence
7093 * field in the inode item) of the inode, so if a file is deduplicated
7094 * the same amount of times in both the parent and send snapshots, its
7095 * iversion becomes the same in both snapshots, whence the inode item is
7096 * the same on both snapshots.
7098 if (sctx->cur_ino != sctx->cmp_key->objectid)
7101 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7102 if (result != BTRFS_COMPARE_TREE_DELETED)
7103 ret = process_extent(sctx, sctx->left_path,
7110 static int changed_verity(struct send_ctx *sctx, enum btrfs_compare_tree_result result)
7114 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7115 if (result == BTRFS_COMPARE_TREE_NEW)
7116 sctx->cur_inode_needs_verity = true;
7121 static int dir_changed(struct send_ctx *sctx, u64 dir)
7123 u64 orig_gen, new_gen;
7126 ret = get_inode_gen(sctx->send_root, dir, &new_gen);
7130 ret = get_inode_gen(sctx->parent_root, dir, &orig_gen);
7134 return (orig_gen != new_gen) ? 1 : 0;
7137 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
7138 struct btrfs_key *key)
7140 struct btrfs_inode_extref *extref;
7141 struct extent_buffer *leaf;
7142 u64 dirid = 0, last_dirid = 0;
7149 /* Easy case, just check this one dirid */
7150 if (key->type == BTRFS_INODE_REF_KEY) {
7151 dirid = key->offset;
7153 ret = dir_changed(sctx, dirid);
7157 leaf = path->nodes[0];
7158 item_size = btrfs_item_size(leaf, path->slots[0]);
7159 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
7160 while (cur_offset < item_size) {
7161 extref = (struct btrfs_inode_extref *)(ptr +
7163 dirid = btrfs_inode_extref_parent(leaf, extref);
7164 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
7165 cur_offset += ref_name_len + sizeof(*extref);
7166 if (dirid == last_dirid)
7168 ret = dir_changed(sctx, dirid);
7178 * Updates compare related fields in sctx and simply forwards to the actual
7179 * changed_xxx functions.
7181 static int changed_cb(struct btrfs_path *left_path,
7182 struct btrfs_path *right_path,
7183 struct btrfs_key *key,
7184 enum btrfs_compare_tree_result result,
7185 struct send_ctx *sctx)
7190 * We can not hold the commit root semaphore here. This is because in
7191 * the case of sending and receiving to the same filesystem, using a
7192 * pipe, could result in a deadlock:
7194 * 1) The task running send blocks on the pipe because it's full;
7196 * 2) The task running receive, which is the only consumer of the pipe,
7197 * is waiting for a transaction commit (for example due to a space
7198 * reservation when doing a write or triggering a transaction commit
7199 * when creating a subvolume);
7201 * 3) The transaction is waiting to write lock the commit root semaphore,
7202 * but can not acquire it since it's being held at 1).
7204 * Down this call chain we write to the pipe through kernel_write().
7205 * The same type of problem can also happen when sending to a file that
7206 * is stored in the same filesystem - when reserving space for a write
7207 * into the file, we can trigger a transaction commit.
7209 * Our caller has supplied us with clones of leaves from the send and
7210 * parent roots, so we're safe here from a concurrent relocation and
7211 * further reallocation of metadata extents while we are here. Below we
7212 * also assert that the leaves are clones.
7214 lockdep_assert_not_held(&sctx->send_root->fs_info->commit_root_sem);
7217 * We always have a send root, so left_path is never NULL. We will not
7218 * have a leaf when we have reached the end of the send root but have
7219 * not yet reached the end of the parent root.
7221 if (left_path->nodes[0])
7222 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7223 &left_path->nodes[0]->bflags));
7225 * When doing a full send we don't have a parent root, so right_path is
7226 * NULL. When doing an incremental send, we may have reached the end of
7227 * the parent root already, so we don't have a leaf at right_path.
7229 if (right_path && right_path->nodes[0])
7230 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7231 &right_path->nodes[0]->bflags));
7233 if (result == BTRFS_COMPARE_TREE_SAME) {
7234 if (key->type == BTRFS_INODE_REF_KEY ||
7235 key->type == BTRFS_INODE_EXTREF_KEY) {
7236 ret = compare_refs(sctx, left_path, key);
7241 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
7242 return maybe_send_hole(sctx, left_path, key);
7246 result = BTRFS_COMPARE_TREE_CHANGED;
7250 sctx->left_path = left_path;
7251 sctx->right_path = right_path;
7252 sctx->cmp_key = key;
7254 ret = finish_inode_if_needed(sctx, 0);
7258 /* Ignore non-FS objects */
7259 if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
7260 key->objectid == BTRFS_FREE_SPACE_OBJECTID)
7263 if (key->type == BTRFS_INODE_ITEM_KEY) {
7264 ret = changed_inode(sctx, result);
7265 } else if (!sctx->ignore_cur_inode) {
7266 if (key->type == BTRFS_INODE_REF_KEY ||
7267 key->type == BTRFS_INODE_EXTREF_KEY)
7268 ret = changed_ref(sctx, result);
7269 else if (key->type == BTRFS_XATTR_ITEM_KEY)
7270 ret = changed_xattr(sctx, result);
7271 else if (key->type == BTRFS_EXTENT_DATA_KEY)
7272 ret = changed_extent(sctx, result);
7273 else if (key->type == BTRFS_VERITY_DESC_ITEM_KEY &&
7275 ret = changed_verity(sctx, result);
7282 static int search_key_again(const struct send_ctx *sctx,
7283 struct btrfs_root *root,
7284 struct btrfs_path *path,
7285 const struct btrfs_key *key)
7289 if (!path->need_commit_sem)
7290 lockdep_assert_held_read(&root->fs_info->commit_root_sem);
7293 * Roots used for send operations are readonly and no one can add,
7294 * update or remove keys from them, so we should be able to find our
7295 * key again. The only exception is deduplication, which can operate on
7296 * readonly roots and add, update or remove keys to/from them - but at
7297 * the moment we don't allow it to run in parallel with send.
7299 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7302 btrfs_print_tree(path->nodes[path->lowest_level], false);
7303 btrfs_err(root->fs_info,
7304 "send: key (%llu %u %llu) not found in %s root %llu, lowest_level %d, slot %d",
7305 key->objectid, key->type, key->offset,
7306 (root == sctx->parent_root ? "parent" : "send"),
7307 root->root_key.objectid, path->lowest_level,
7308 path->slots[path->lowest_level]);
7315 static int full_send_tree(struct send_ctx *sctx)
7318 struct btrfs_root *send_root = sctx->send_root;
7319 struct btrfs_key key;
7320 struct btrfs_fs_info *fs_info = send_root->fs_info;
7321 struct btrfs_path *path;
7323 path = alloc_path_for_send();
7326 path->reada = READA_FORWARD_ALWAYS;
7328 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
7329 key.type = BTRFS_INODE_ITEM_KEY;
7332 down_read(&fs_info->commit_root_sem);
7333 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7334 up_read(&fs_info->commit_root_sem);
7336 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
7343 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
7345 ret = changed_cb(path, NULL, &key,
7346 BTRFS_COMPARE_TREE_NEW, sctx);
7350 down_read(&fs_info->commit_root_sem);
7351 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7352 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7353 up_read(&fs_info->commit_root_sem);
7355 * A transaction used for relocating a block group was
7356 * committed or is about to finish its commit. Release
7357 * our path (leaf) and restart the search, so that we
7358 * avoid operating on any file extent items that are
7359 * stale, with a disk_bytenr that reflects a pre
7360 * relocation value. This way we avoid as much as
7361 * possible to fallback to regular writes when checking
7362 * if we can clone file ranges.
7364 btrfs_release_path(path);
7365 ret = search_key_again(sctx, send_root, path, &key);
7369 up_read(&fs_info->commit_root_sem);
7372 ret = btrfs_next_item(send_root, path);
7382 ret = finish_inode_if_needed(sctx, 1);
7385 btrfs_free_path(path);
7389 static int replace_node_with_clone(struct btrfs_path *path, int level)
7391 struct extent_buffer *clone;
7393 clone = btrfs_clone_extent_buffer(path->nodes[level]);
7397 free_extent_buffer(path->nodes[level]);
7398 path->nodes[level] = clone;
7403 static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
7405 struct extent_buffer *eb;
7406 struct extent_buffer *parent = path->nodes[*level];
7407 int slot = path->slots[*level];
7408 const int nritems = btrfs_header_nritems(parent);
7412 lockdep_assert_held_read(&parent->fs_info->commit_root_sem);
7414 BUG_ON(*level == 0);
7415 eb = btrfs_read_node_slot(parent, slot);
7420 * Trigger readahead for the next leaves we will process, so that it is
7421 * very likely that when we need them they are already in memory and we
7422 * will not block on disk IO. For nodes we only do readahead for one,
7423 * since the time window between processing nodes is typically larger.
7425 reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
7427 for (slot++; slot < nritems && reada_done < reada_max; slot++) {
7428 if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
7429 btrfs_readahead_node_child(parent, slot);
7430 reada_done += eb->fs_info->nodesize;
7434 path->nodes[*level - 1] = eb;
7435 path->slots[*level - 1] = 0;
7439 return replace_node_with_clone(path, 0);
7444 static int tree_move_next_or_upnext(struct btrfs_path *path,
7445 int *level, int root_level)
7449 nritems = btrfs_header_nritems(path->nodes[*level]);
7451 path->slots[*level]++;
7453 while (path->slots[*level] >= nritems) {
7454 if (*level == root_level) {
7455 path->slots[*level] = nritems - 1;
7460 path->slots[*level] = 0;
7461 free_extent_buffer(path->nodes[*level]);
7462 path->nodes[*level] = NULL;
7464 path->slots[*level]++;
7466 nritems = btrfs_header_nritems(path->nodes[*level]);
7473 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
7476 static int tree_advance(struct btrfs_path *path,
7477 int *level, int root_level,
7479 struct btrfs_key *key,
7484 if (*level == 0 || !allow_down) {
7485 ret = tree_move_next_or_upnext(path, level, root_level);
7487 ret = tree_move_down(path, level, reada_min_gen);
7491 * Even if we have reached the end of a tree, ret is -1, update the key
7492 * anyway, so that in case we need to restart due to a block group
7493 * relocation, we can assert that the last key of the root node still
7494 * exists in the tree.
7497 btrfs_item_key_to_cpu(path->nodes[*level], key,
7498 path->slots[*level]);
7500 btrfs_node_key_to_cpu(path->nodes[*level], key,
7501 path->slots[*level]);
7506 static int tree_compare_item(struct btrfs_path *left_path,
7507 struct btrfs_path *right_path,
7512 unsigned long off1, off2;
7514 len1 = btrfs_item_size(left_path->nodes[0], left_path->slots[0]);
7515 len2 = btrfs_item_size(right_path->nodes[0], right_path->slots[0]);
7519 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
7520 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
7521 right_path->slots[0]);
7523 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
7525 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
7532 * A transaction used for relocating a block group was committed or is about to
7533 * finish its commit. Release our paths and restart the search, so that we are
7534 * not using stale extent buffers:
7536 * 1) For levels > 0, we are only holding references of extent buffers, without
7537 * any locks on them, which does not prevent them from having been relocated
7538 * and reallocated after the last time we released the commit root semaphore.
7539 * The exception are the root nodes, for which we always have a clone, see
7540 * the comment at btrfs_compare_trees();
7542 * 2) For leaves, level 0, we are holding copies (clones) of extent buffers, so
7543 * we are safe from the concurrent relocation and reallocation. However they
7544 * can have file extent items with a pre relocation disk_bytenr value, so we
7545 * restart the start from the current commit roots and clone the new leaves so
7546 * that we get the post relocation disk_bytenr values. Not doing so, could
7547 * make us clone the wrong data in case there are new extents using the old
7548 * disk_bytenr that happen to be shared.
7550 static int restart_after_relocation(struct btrfs_path *left_path,
7551 struct btrfs_path *right_path,
7552 const struct btrfs_key *left_key,
7553 const struct btrfs_key *right_key,
7556 const struct send_ctx *sctx)
7561 lockdep_assert_held_read(&sctx->send_root->fs_info->commit_root_sem);
7563 btrfs_release_path(left_path);
7564 btrfs_release_path(right_path);
7567 * Since keys can not be added or removed to/from our roots because they
7568 * are readonly and we do not allow deduplication to run in parallel
7569 * (which can add, remove or change keys), the layout of the trees should
7572 left_path->lowest_level = left_level;
7573 ret = search_key_again(sctx, sctx->send_root, left_path, left_key);
7577 right_path->lowest_level = right_level;
7578 ret = search_key_again(sctx, sctx->parent_root, right_path, right_key);
7583 * If the lowest level nodes are leaves, clone them so that they can be
7584 * safely used by changed_cb() while not under the protection of the
7585 * commit root semaphore, even if relocation and reallocation happens in
7588 if (left_level == 0) {
7589 ret = replace_node_with_clone(left_path, 0);
7594 if (right_level == 0) {
7595 ret = replace_node_with_clone(right_path, 0);
7601 * Now clone the root nodes (unless they happen to be the leaves we have
7602 * already cloned). This is to protect against concurrent snapshotting of
7603 * the send and parent roots (see the comment at btrfs_compare_trees()).
7605 root_level = btrfs_header_level(sctx->send_root->commit_root);
7606 if (root_level > 0) {
7607 ret = replace_node_with_clone(left_path, root_level);
7612 root_level = btrfs_header_level(sctx->parent_root->commit_root);
7613 if (root_level > 0) {
7614 ret = replace_node_with_clone(right_path, root_level);
7623 * This function compares two trees and calls the provided callback for
7624 * every changed/new/deleted item it finds.
7625 * If shared tree blocks are encountered, whole subtrees are skipped, making
7626 * the compare pretty fast on snapshotted subvolumes.
7628 * This currently works on commit roots only. As commit roots are read only,
7629 * we don't do any locking. The commit roots are protected with transactions.
7630 * Transactions are ended and rejoined when a commit is tried in between.
7632 * This function checks for modifications done to the trees while comparing.
7633 * If it detects a change, it aborts immediately.
7635 static int btrfs_compare_trees(struct btrfs_root *left_root,
7636 struct btrfs_root *right_root, struct send_ctx *sctx)
7638 struct btrfs_fs_info *fs_info = left_root->fs_info;
7641 struct btrfs_path *left_path = NULL;
7642 struct btrfs_path *right_path = NULL;
7643 struct btrfs_key left_key;
7644 struct btrfs_key right_key;
7645 char *tmp_buf = NULL;
7646 int left_root_level;
7647 int right_root_level;
7650 int left_end_reached = 0;
7651 int right_end_reached = 0;
7652 int advance_left = 0;
7653 int advance_right = 0;
7660 left_path = btrfs_alloc_path();
7665 right_path = btrfs_alloc_path();
7671 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
7677 left_path->search_commit_root = 1;
7678 left_path->skip_locking = 1;
7679 right_path->search_commit_root = 1;
7680 right_path->skip_locking = 1;
7683 * Strategy: Go to the first items of both trees. Then do
7685 * If both trees are at level 0
7686 * Compare keys of current items
7687 * If left < right treat left item as new, advance left tree
7689 * If left > right treat right item as deleted, advance right tree
7691 * If left == right do deep compare of items, treat as changed if
7692 * needed, advance both trees and repeat
7693 * If both trees are at the same level but not at level 0
7694 * Compare keys of current nodes/leafs
7695 * If left < right advance left tree and repeat
7696 * If left > right advance right tree and repeat
7697 * If left == right compare blockptrs of the next nodes/leafs
7698 * If they match advance both trees but stay at the same level
7700 * If they don't match advance both trees while allowing to go
7702 * If tree levels are different
7703 * Advance the tree that needs it and repeat
7705 * Advancing a tree means:
7706 * If we are at level 0, try to go to the next slot. If that's not
7707 * possible, go one level up and repeat. Stop when we found a level
7708 * where we could go to the next slot. We may at this point be on a
7711 * If we are not at level 0 and not on shared tree blocks, go one
7714 * If we are not at level 0 and on shared tree blocks, go one slot to
7715 * the right if possible or go up and right.
7718 down_read(&fs_info->commit_root_sem);
7719 left_level = btrfs_header_level(left_root->commit_root);
7720 left_root_level = left_level;
7722 * We clone the root node of the send and parent roots to prevent races
7723 * with snapshot creation of these roots. Snapshot creation COWs the
7724 * root node of a tree, so after the transaction is committed the old
7725 * extent can be reallocated while this send operation is still ongoing.
7726 * So we clone them, under the commit root semaphore, to be race free.
7728 left_path->nodes[left_level] =
7729 btrfs_clone_extent_buffer(left_root->commit_root);
7730 if (!left_path->nodes[left_level]) {
7735 right_level = btrfs_header_level(right_root->commit_root);
7736 right_root_level = right_level;
7737 right_path->nodes[right_level] =
7738 btrfs_clone_extent_buffer(right_root->commit_root);
7739 if (!right_path->nodes[right_level]) {
7744 * Our right root is the parent root, while the left root is the "send"
7745 * root. We know that all new nodes/leaves in the left root must have
7746 * a generation greater than the right root's generation, so we trigger
7747 * readahead for those nodes and leaves of the left root, as we know we
7748 * will need to read them at some point.
7750 reada_min_gen = btrfs_header_generation(right_root->commit_root);
7752 if (left_level == 0)
7753 btrfs_item_key_to_cpu(left_path->nodes[left_level],
7754 &left_key, left_path->slots[left_level]);
7756 btrfs_node_key_to_cpu(left_path->nodes[left_level],
7757 &left_key, left_path->slots[left_level]);
7758 if (right_level == 0)
7759 btrfs_item_key_to_cpu(right_path->nodes[right_level],
7760 &right_key, right_path->slots[right_level]);
7762 btrfs_node_key_to_cpu(right_path->nodes[right_level],
7763 &right_key, right_path->slots[right_level]);
7765 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7768 if (need_resched() ||
7769 rwsem_is_contended(&fs_info->commit_root_sem)) {
7770 up_read(&fs_info->commit_root_sem);
7772 down_read(&fs_info->commit_root_sem);
7775 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7776 ret = restart_after_relocation(left_path, right_path,
7777 &left_key, &right_key,
7778 left_level, right_level,
7782 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7785 if (advance_left && !left_end_reached) {
7786 ret = tree_advance(left_path, &left_level,
7788 advance_left != ADVANCE_ONLY_NEXT,
7789 &left_key, reada_min_gen);
7791 left_end_reached = ADVANCE;
7796 if (advance_right && !right_end_reached) {
7797 ret = tree_advance(right_path, &right_level,
7799 advance_right != ADVANCE_ONLY_NEXT,
7800 &right_key, reada_min_gen);
7802 right_end_reached = ADVANCE;
7808 if (left_end_reached && right_end_reached) {
7811 } else if (left_end_reached) {
7812 if (right_level == 0) {
7813 up_read(&fs_info->commit_root_sem);
7814 ret = changed_cb(left_path, right_path,
7816 BTRFS_COMPARE_TREE_DELETED,
7820 down_read(&fs_info->commit_root_sem);
7822 advance_right = ADVANCE;
7824 } else if (right_end_reached) {
7825 if (left_level == 0) {
7826 up_read(&fs_info->commit_root_sem);
7827 ret = changed_cb(left_path, right_path,
7829 BTRFS_COMPARE_TREE_NEW,
7833 down_read(&fs_info->commit_root_sem);
7835 advance_left = ADVANCE;
7839 if (left_level == 0 && right_level == 0) {
7840 up_read(&fs_info->commit_root_sem);
7841 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7843 ret = changed_cb(left_path, right_path,
7845 BTRFS_COMPARE_TREE_NEW,
7847 advance_left = ADVANCE;
7848 } else if (cmp > 0) {
7849 ret = changed_cb(left_path, right_path,
7851 BTRFS_COMPARE_TREE_DELETED,
7853 advance_right = ADVANCE;
7855 enum btrfs_compare_tree_result result;
7857 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7858 ret = tree_compare_item(left_path, right_path,
7861 result = BTRFS_COMPARE_TREE_CHANGED;
7863 result = BTRFS_COMPARE_TREE_SAME;
7864 ret = changed_cb(left_path, right_path,
7865 &left_key, result, sctx);
7866 advance_left = ADVANCE;
7867 advance_right = ADVANCE;
7872 down_read(&fs_info->commit_root_sem);
7873 } else if (left_level == right_level) {
7874 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7876 advance_left = ADVANCE;
7877 } else if (cmp > 0) {
7878 advance_right = ADVANCE;
7880 left_blockptr = btrfs_node_blockptr(
7881 left_path->nodes[left_level],
7882 left_path->slots[left_level]);
7883 right_blockptr = btrfs_node_blockptr(
7884 right_path->nodes[right_level],
7885 right_path->slots[right_level]);
7886 left_gen = btrfs_node_ptr_generation(
7887 left_path->nodes[left_level],
7888 left_path->slots[left_level]);
7889 right_gen = btrfs_node_ptr_generation(
7890 right_path->nodes[right_level],
7891 right_path->slots[right_level]);
7892 if (left_blockptr == right_blockptr &&
7893 left_gen == right_gen) {
7895 * As we're on a shared block, don't
7896 * allow to go deeper.
7898 advance_left = ADVANCE_ONLY_NEXT;
7899 advance_right = ADVANCE_ONLY_NEXT;
7901 advance_left = ADVANCE;
7902 advance_right = ADVANCE;
7905 } else if (left_level < right_level) {
7906 advance_right = ADVANCE;
7908 advance_left = ADVANCE;
7913 up_read(&fs_info->commit_root_sem);
7915 btrfs_free_path(left_path);
7916 btrfs_free_path(right_path);
7921 static int send_subvol(struct send_ctx *sctx)
7925 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7926 ret = send_header(sctx);
7931 ret = send_subvol_begin(sctx);
7935 if (sctx->parent_root) {
7936 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7939 ret = finish_inode_if_needed(sctx, 1);
7943 ret = full_send_tree(sctx);
7949 free_recorded_refs(sctx);
7954 * If orphan cleanup did remove any orphans from a root, it means the tree
7955 * was modified and therefore the commit root is not the same as the current
7956 * root anymore. This is a problem, because send uses the commit root and
7957 * therefore can see inode items that don't exist in the current root anymore,
7958 * and for example make calls to btrfs_iget, which will do tree lookups based
7959 * on the current root and not on the commit root. Those lookups will fail,
7960 * returning a -ESTALE error, and making send fail with that error. So make
7961 * sure a send does not see any orphans we have just removed, and that it will
7962 * see the same inodes regardless of whether a transaction commit happened
7963 * before it started (meaning that the commit root will be the same as the
7964 * current root) or not.
7966 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
7969 struct btrfs_trans_handle *trans = NULL;
7972 if (sctx->parent_root &&
7973 sctx->parent_root->node != sctx->parent_root->commit_root)
7976 for (i = 0; i < sctx->clone_roots_cnt; i++)
7977 if (sctx->clone_roots[i].root->node !=
7978 sctx->clone_roots[i].root->commit_root)
7982 return btrfs_end_transaction(trans);
7987 /* Use any root, all fs roots will get their commit roots updated. */
7989 trans = btrfs_join_transaction(sctx->send_root);
7991 return PTR_ERR(trans);
7995 return btrfs_commit_transaction(trans);
7999 * Make sure any existing dellaloc is flushed for any root used by a send
8000 * operation so that we do not miss any data and we do not race with writeback
8001 * finishing and changing a tree while send is using the tree. This could
8002 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
8003 * a send operation then uses the subvolume.
8004 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
8006 static int flush_delalloc_roots(struct send_ctx *sctx)
8008 struct btrfs_root *root = sctx->parent_root;
8013 ret = btrfs_start_delalloc_snapshot(root, false);
8016 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8019 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8020 root = sctx->clone_roots[i].root;
8021 ret = btrfs_start_delalloc_snapshot(root, false);
8024 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8030 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
8032 spin_lock(&root->root_item_lock);
8033 root->send_in_progress--;
8035 * Not much left to do, we don't know why it's unbalanced and
8036 * can't blindly reset it to 0.
8038 if (root->send_in_progress < 0)
8039 btrfs_err(root->fs_info,
8040 "send_in_progress unbalanced %d root %llu",
8041 root->send_in_progress, root->root_key.objectid);
8042 spin_unlock(&root->root_item_lock);
8045 static void dedupe_in_progress_warn(const struct btrfs_root *root)
8047 btrfs_warn_rl(root->fs_info,
8048 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
8049 root->root_key.objectid, root->dedupe_in_progress);
8052 long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
8055 struct btrfs_root *send_root = BTRFS_I(inode)->root;
8056 struct btrfs_fs_info *fs_info = send_root->fs_info;
8057 struct btrfs_root *clone_root;
8058 struct send_ctx *sctx = NULL;
8060 u64 *clone_sources_tmp = NULL;
8061 int clone_sources_to_rollback = 0;
8063 int sort_clone_roots = 0;
8064 struct btrfs_lru_cache_entry *entry;
8065 struct btrfs_lru_cache_entry *tmp;
8067 if (!capable(CAP_SYS_ADMIN))
8071 * The subvolume must remain read-only during send, protect against
8072 * making it RW. This also protects against deletion.
8074 spin_lock(&send_root->root_item_lock);
8075 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
8076 dedupe_in_progress_warn(send_root);
8077 spin_unlock(&send_root->root_item_lock);
8080 send_root->send_in_progress++;
8081 spin_unlock(&send_root->root_item_lock);
8084 * Userspace tools do the checks and warn the user if it's
8087 if (!btrfs_root_readonly(send_root)) {
8093 * Check that we don't overflow at later allocations, we request
8094 * clone_sources_count + 1 items, and compare to unsigned long inside
8095 * access_ok. Also set an upper limit for allocation size so this can't
8096 * easily exhaust memory. Max number of clone sources is about 200K.
8098 if (arg->clone_sources_count > SZ_8M / sizeof(struct clone_root)) {
8103 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
8108 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
8114 INIT_LIST_HEAD(&sctx->new_refs);
8115 INIT_LIST_HEAD(&sctx->deleted_refs);
8117 btrfs_lru_cache_init(&sctx->name_cache, SEND_MAX_NAME_CACHE_SIZE);
8118 btrfs_lru_cache_init(&sctx->backref_cache, SEND_MAX_BACKREF_CACHE_SIZE);
8119 btrfs_lru_cache_init(&sctx->dir_created_cache,
8120 SEND_MAX_DIR_CREATED_CACHE_SIZE);
8122 * This cache is periodically trimmed to a fixed size elsewhere, see
8123 * cache_dir_utimes() and trim_dir_utimes_cache().
8125 btrfs_lru_cache_init(&sctx->dir_utimes_cache, 0);
8127 sctx->pending_dir_moves = RB_ROOT;
8128 sctx->waiting_dir_moves = RB_ROOT;
8129 sctx->orphan_dirs = RB_ROOT;
8130 sctx->rbtree_new_refs = RB_ROOT;
8131 sctx->rbtree_deleted_refs = RB_ROOT;
8133 sctx->flags = arg->flags;
8135 if (arg->flags & BTRFS_SEND_FLAG_VERSION) {
8136 if (arg->version > BTRFS_SEND_STREAM_VERSION) {
8140 /* Zero means "use the highest version" */
8141 sctx->proto = arg->version ?: BTRFS_SEND_STREAM_VERSION;
8145 if ((arg->flags & BTRFS_SEND_FLAG_COMPRESSED) && sctx->proto < 2) {
8150 sctx->send_filp = fget(arg->send_fd);
8151 if (!sctx->send_filp) {
8156 sctx->send_root = send_root;
8158 * Unlikely but possible, if the subvolume is marked for deletion but
8159 * is slow to remove the directory entry, send can still be started
8161 if (btrfs_root_dead(sctx->send_root)) {
8166 sctx->clone_roots_cnt = arg->clone_sources_count;
8168 if (sctx->proto >= 2) {
8169 u32 send_buf_num_pages;
8171 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V2;
8172 sctx->send_buf = vmalloc(sctx->send_max_size);
8173 if (!sctx->send_buf) {
8177 send_buf_num_pages = sctx->send_max_size >> PAGE_SHIFT;
8178 sctx->send_buf_pages = kcalloc(send_buf_num_pages,
8179 sizeof(*sctx->send_buf_pages),
8181 if (!sctx->send_buf_pages) {
8185 for (i = 0; i < send_buf_num_pages; i++) {
8186 sctx->send_buf_pages[i] =
8187 vmalloc_to_page(sctx->send_buf + (i << PAGE_SHIFT));
8190 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
8191 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
8193 if (!sctx->send_buf) {
8198 sctx->clone_roots = kvcalloc(sizeof(*sctx->clone_roots),
8199 arg->clone_sources_count + 1,
8201 if (!sctx->clone_roots) {
8206 alloc_size = array_size(sizeof(*arg->clone_sources),
8207 arg->clone_sources_count);
8209 if (arg->clone_sources_count) {
8210 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
8211 if (!clone_sources_tmp) {
8216 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
8223 for (i = 0; i < arg->clone_sources_count; i++) {
8224 clone_root = btrfs_get_fs_root(fs_info,
8225 clone_sources_tmp[i], true);
8226 if (IS_ERR(clone_root)) {
8227 ret = PTR_ERR(clone_root);
8230 spin_lock(&clone_root->root_item_lock);
8231 if (!btrfs_root_readonly(clone_root) ||
8232 btrfs_root_dead(clone_root)) {
8233 spin_unlock(&clone_root->root_item_lock);
8234 btrfs_put_root(clone_root);
8238 if (clone_root->dedupe_in_progress) {
8239 dedupe_in_progress_warn(clone_root);
8240 spin_unlock(&clone_root->root_item_lock);
8241 btrfs_put_root(clone_root);
8245 clone_root->send_in_progress++;
8246 spin_unlock(&clone_root->root_item_lock);
8248 sctx->clone_roots[i].root = clone_root;
8249 clone_sources_to_rollback = i + 1;
8251 kvfree(clone_sources_tmp);
8252 clone_sources_tmp = NULL;
8255 if (arg->parent_root) {
8256 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
8258 if (IS_ERR(sctx->parent_root)) {
8259 ret = PTR_ERR(sctx->parent_root);
8263 spin_lock(&sctx->parent_root->root_item_lock);
8264 sctx->parent_root->send_in_progress++;
8265 if (!btrfs_root_readonly(sctx->parent_root) ||
8266 btrfs_root_dead(sctx->parent_root)) {
8267 spin_unlock(&sctx->parent_root->root_item_lock);
8271 if (sctx->parent_root->dedupe_in_progress) {
8272 dedupe_in_progress_warn(sctx->parent_root);
8273 spin_unlock(&sctx->parent_root->root_item_lock);
8277 spin_unlock(&sctx->parent_root->root_item_lock);
8281 * Clones from send_root are allowed, but only if the clone source
8282 * is behind the current send position. This is checked while searching
8283 * for possible clone sources.
8285 sctx->clone_roots[sctx->clone_roots_cnt++].root =
8286 btrfs_grab_root(sctx->send_root);
8288 /* We do a bsearch later */
8289 sort(sctx->clone_roots, sctx->clone_roots_cnt,
8290 sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
8292 sort_clone_roots = 1;
8294 ret = flush_delalloc_roots(sctx);
8298 ret = ensure_commit_roots_uptodate(sctx);
8302 ret = send_subvol(sctx);
8306 btrfs_lru_cache_for_each_entry_safe(&sctx->dir_utimes_cache, entry, tmp) {
8307 ret = send_utimes(sctx, entry->key, entry->gen);
8310 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, entry);
8313 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
8314 ret = begin_cmd(sctx, BTRFS_SEND_C_END);
8317 ret = send_cmd(sctx);
8323 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
8324 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
8326 struct pending_dir_move *pm;
8328 n = rb_first(&sctx->pending_dir_moves);
8329 pm = rb_entry(n, struct pending_dir_move, node);
8330 while (!list_empty(&pm->list)) {
8331 struct pending_dir_move *pm2;
8333 pm2 = list_first_entry(&pm->list,
8334 struct pending_dir_move, list);
8335 free_pending_move(sctx, pm2);
8337 free_pending_move(sctx, pm);
8340 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
8341 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
8343 struct waiting_dir_move *dm;
8345 n = rb_first(&sctx->waiting_dir_moves);
8346 dm = rb_entry(n, struct waiting_dir_move, node);
8347 rb_erase(&dm->node, &sctx->waiting_dir_moves);
8351 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
8352 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
8354 struct orphan_dir_info *odi;
8356 n = rb_first(&sctx->orphan_dirs);
8357 odi = rb_entry(n, struct orphan_dir_info, node);
8358 free_orphan_dir_info(sctx, odi);
8361 if (sort_clone_roots) {
8362 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8363 btrfs_root_dec_send_in_progress(
8364 sctx->clone_roots[i].root);
8365 btrfs_put_root(sctx->clone_roots[i].root);
8368 for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
8369 btrfs_root_dec_send_in_progress(
8370 sctx->clone_roots[i].root);
8371 btrfs_put_root(sctx->clone_roots[i].root);
8374 btrfs_root_dec_send_in_progress(send_root);
8376 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
8377 btrfs_root_dec_send_in_progress(sctx->parent_root);
8378 btrfs_put_root(sctx->parent_root);
8381 kvfree(clone_sources_tmp);
8384 if (sctx->send_filp)
8385 fput(sctx->send_filp);
8387 kvfree(sctx->clone_roots);
8388 kfree(sctx->send_buf_pages);
8389 kvfree(sctx->send_buf);
8390 kvfree(sctx->verity_descriptor);
8392 close_current_inode(sctx);
8394 btrfs_lru_cache_clear(&sctx->name_cache);
8395 btrfs_lru_cache_clear(&sctx->backref_cache);
8396 btrfs_lru_cache_clear(&sctx->dir_created_cache);
8397 btrfs_lru_cache_clear(&sctx->dir_utimes_cache);