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>
23 #include "btrfs_inode.h"
24 #include "transaction.h"
25 #include "compression.h"
29 * Maximum number of references an extent can have in order for us to attempt to
30 * issue clone operations instead of write operations. This currently exists to
31 * avoid hitting limitations of the backreference walking code (taking a lot of
32 * time and using too much memory for extents with large number of references).
34 #define SEND_MAX_EXTENT_REFS 64
37 * A fs_path is a helper to dynamically build path names with unknown size.
38 * It reallocates the internal buffer on demand.
39 * It allows fast adding of path elements on the right side (normal path) and
40 * fast adding to the left side (reversed path). A reversed path can also be
41 * unreversed if needed.
50 unsigned short buf_len:15;
51 unsigned short reversed:1;
55 * Average path length does not exceed 200 bytes, we'll have
56 * better packing in the slab and higher chance to satisfy
57 * a allocation later during send.
62 #define FS_PATH_INLINE_SIZE \
63 (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
66 /* reused for each extent */
68 struct btrfs_root *root;
75 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128
76 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
79 struct file *send_filp;
85 u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
86 u64 flags; /* 'flags' member of btrfs_ioctl_send_args is u64 */
88 struct btrfs_root *send_root;
89 struct btrfs_root *parent_root;
90 struct clone_root *clone_roots;
93 /* current state of the compare_tree call */
94 struct btrfs_path *left_path;
95 struct btrfs_path *right_path;
96 struct btrfs_key *cmp_key;
99 * infos of the currently processed inode. In case of deleted inodes,
100 * these are the values from the deleted inode.
105 int cur_inode_new_gen;
106 int cur_inode_deleted;
110 u64 cur_inode_last_extent;
111 u64 cur_inode_next_write_offset;
112 bool ignore_cur_inode;
116 struct list_head new_refs;
117 struct list_head deleted_refs;
119 struct radix_tree_root name_cache;
120 struct list_head name_cache_list;
123 struct file_ra_state ra;
126 * We process inodes by their increasing order, so if before an
127 * incremental send we reverse the parent/child relationship of
128 * directories such that a directory with a lower inode number was
129 * the parent of a directory with a higher inode number, and the one
130 * becoming the new parent got renamed too, we can't rename/move the
131 * directory with lower inode number when we finish processing it - we
132 * must process the directory with higher inode number first, then
133 * rename/move it and then rename/move the directory with lower inode
134 * number. Example follows.
136 * Tree state when the first send was performed:
148 * Tree state when the second (incremental) send is performed:
157 * The sequence of steps that lead to the second state was:
159 * mv /a/b/c/d /a/b/c2/d2
160 * mv /a/b/c /a/b/c2/d2/cc
162 * "c" has lower inode number, but we can't move it (2nd mv operation)
163 * before we move "d", which has higher inode number.
165 * So we just memorize which move/rename operations must be performed
166 * later when their respective parent is processed and moved/renamed.
169 /* Indexed by parent directory inode number. */
170 struct rb_root pending_dir_moves;
173 * Reverse index, indexed by the inode number of a directory that
174 * is waiting for the move/rename of its immediate parent before its
175 * own move/rename can be performed.
177 struct rb_root waiting_dir_moves;
180 * A directory that is going to be rm'ed might have a child directory
181 * which is in the pending directory moves index above. In this case,
182 * the directory can only be removed after the move/rename of its child
183 * is performed. Example:
203 * Sequence of steps that lead to the send snapshot:
204 * rm -f /a/b/c/foo.txt
206 * mv /a/b/c/x /a/b/YY
209 * When the child is processed, its move/rename is delayed until its
210 * parent is processed (as explained above), but all other operations
211 * like update utimes, chown, chgrp, etc, are performed and the paths
212 * that it uses for those operations must use the orphanized name of
213 * its parent (the directory we're going to rm later), so we need to
214 * memorize that name.
216 * Indexed by the inode number of the directory to be deleted.
218 struct rb_root orphan_dirs;
221 struct pending_dir_move {
223 struct list_head list;
227 struct list_head update_refs;
230 struct waiting_dir_move {
234 * There might be some directory that could not be removed because it
235 * was waiting for this directory inode to be moved first. Therefore
236 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
243 struct orphan_dir_info {
247 u64 last_dir_index_offset;
250 struct name_cache_entry {
251 struct list_head list;
253 * radix_tree has only 32bit entries but we need to handle 64bit inums.
254 * We use the lower 32bit of the 64bit inum to store it in the tree. If
255 * more then one inum would fall into the same entry, we use radix_list
256 * to store the additional entries. radix_list is also used to store
257 * entries where two entries have the same inum but different
260 struct list_head radix_list;
266 int need_later_update;
272 #define ADVANCE_ONLY_NEXT -1
274 enum btrfs_compare_tree_result {
275 BTRFS_COMPARE_TREE_NEW,
276 BTRFS_COMPARE_TREE_DELETED,
277 BTRFS_COMPARE_TREE_CHANGED,
278 BTRFS_COMPARE_TREE_SAME,
282 static void inconsistent_snapshot_error(struct send_ctx *sctx,
283 enum btrfs_compare_tree_result result,
286 const char *result_string;
289 case BTRFS_COMPARE_TREE_NEW:
290 result_string = "new";
292 case BTRFS_COMPARE_TREE_DELETED:
293 result_string = "deleted";
295 case BTRFS_COMPARE_TREE_CHANGED:
296 result_string = "updated";
298 case BTRFS_COMPARE_TREE_SAME:
300 result_string = "unchanged";
304 result_string = "unexpected";
307 btrfs_err(sctx->send_root->fs_info,
308 "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
309 result_string, what, sctx->cmp_key->objectid,
310 sctx->send_root->root_key.objectid,
312 sctx->parent_root->root_key.objectid : 0));
315 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
317 static struct waiting_dir_move *
318 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
320 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);
322 static int need_send_hole(struct send_ctx *sctx)
324 return (sctx->parent_root && !sctx->cur_inode_new &&
325 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
326 S_ISREG(sctx->cur_inode_mode));
329 static void fs_path_reset(struct fs_path *p)
332 p->start = p->buf + p->buf_len - 1;
342 static struct fs_path *fs_path_alloc(void)
346 p = kmalloc(sizeof(*p), GFP_KERNEL);
350 p->buf = p->inline_buf;
351 p->buf_len = FS_PATH_INLINE_SIZE;
356 static struct fs_path *fs_path_alloc_reversed(void)
368 static void fs_path_free(struct fs_path *p)
372 if (p->buf != p->inline_buf)
377 static int fs_path_len(struct fs_path *p)
379 return p->end - p->start;
382 static int fs_path_ensure_buf(struct fs_path *p, int len)
390 if (p->buf_len >= len)
393 if (len > PATH_MAX) {
398 path_len = p->end - p->start;
399 old_buf_len = p->buf_len;
402 * First time the inline_buf does not suffice
404 if (p->buf == p->inline_buf) {
405 tmp_buf = kmalloc(len, GFP_KERNEL);
407 memcpy(tmp_buf, p->buf, old_buf_len);
409 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
415 * The real size of the buffer is bigger, this will let the fast path
416 * happen most of the time
418 p->buf_len = ksize(p->buf);
421 tmp_buf = p->buf + old_buf_len - path_len - 1;
422 p->end = p->buf + p->buf_len - 1;
423 p->start = p->end - path_len;
424 memmove(p->start, tmp_buf, path_len + 1);
427 p->end = p->start + path_len;
432 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
438 new_len = p->end - p->start + name_len;
439 if (p->start != p->end)
441 ret = fs_path_ensure_buf(p, new_len);
446 if (p->start != p->end)
448 p->start -= name_len;
449 *prepared = p->start;
451 if (p->start != p->end)
462 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
467 ret = fs_path_prepare_for_add(p, name_len, &prepared);
470 memcpy(prepared, name, name_len);
476 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
481 ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
484 memcpy(prepared, p2->start, p2->end - p2->start);
490 static int fs_path_add_from_extent_buffer(struct fs_path *p,
491 struct extent_buffer *eb,
492 unsigned long off, int len)
497 ret = fs_path_prepare_for_add(p, len, &prepared);
501 read_extent_buffer(eb, prepared, off, len);
507 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
511 p->reversed = from->reversed;
514 ret = fs_path_add_path(p, from);
520 static void fs_path_unreverse(struct fs_path *p)
529 len = p->end - p->start;
531 p->end = p->start + len;
532 memmove(p->start, tmp, len + 1);
536 static struct btrfs_path *alloc_path_for_send(void)
538 struct btrfs_path *path;
540 path = btrfs_alloc_path();
543 path->search_commit_root = 1;
544 path->skip_locking = 1;
545 path->need_commit_sem = 1;
549 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
555 ret = kernel_write(filp, buf + pos, len - pos, off);
556 /* TODO handle that correctly */
557 /*if (ret == -ERESTARTSYS) {
571 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
573 struct btrfs_tlv_header *hdr;
574 int total_len = sizeof(*hdr) + len;
575 int left = sctx->send_max_size - sctx->send_size;
577 if (unlikely(left < total_len))
580 hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
581 put_unaligned_le16(attr, &hdr->tlv_type);
582 put_unaligned_le16(len, &hdr->tlv_len);
583 memcpy(hdr + 1, data, len);
584 sctx->send_size += total_len;
589 #define TLV_PUT_DEFINE_INT(bits) \
590 static int tlv_put_u##bits(struct send_ctx *sctx, \
591 u##bits attr, u##bits value) \
593 __le##bits __tmp = cpu_to_le##bits(value); \
594 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp)); \
597 TLV_PUT_DEFINE_INT(64)
599 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
600 const char *str, int len)
604 return tlv_put(sctx, attr, str, len);
607 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
610 return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
613 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
614 struct extent_buffer *eb,
615 struct btrfs_timespec *ts)
617 struct btrfs_timespec bts;
618 read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
619 return tlv_put(sctx, attr, &bts, sizeof(bts));
623 #define TLV_PUT(sctx, attrtype, data, attrlen) \
625 ret = tlv_put(sctx, attrtype, data, attrlen); \
627 goto tlv_put_failure; \
630 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
632 ret = tlv_put_u##bits(sctx, attrtype, value); \
634 goto tlv_put_failure; \
637 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
638 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
639 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
640 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
641 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
643 ret = tlv_put_string(sctx, attrtype, str, len); \
645 goto tlv_put_failure; \
647 #define TLV_PUT_PATH(sctx, attrtype, p) \
649 ret = tlv_put_string(sctx, attrtype, p->start, \
650 p->end - p->start); \
652 goto tlv_put_failure; \
654 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
656 ret = tlv_put_uuid(sctx, attrtype, uuid); \
658 goto tlv_put_failure; \
660 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
662 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
664 goto tlv_put_failure; \
667 static int send_header(struct send_ctx *sctx)
669 struct btrfs_stream_header hdr;
671 strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
672 hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
674 return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
679 * For each command/item we want to send to userspace, we call this function.
681 static int begin_cmd(struct send_ctx *sctx, int cmd)
683 struct btrfs_cmd_header *hdr;
685 if (WARN_ON(!sctx->send_buf))
688 BUG_ON(sctx->send_size);
690 sctx->send_size += sizeof(*hdr);
691 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
692 put_unaligned_le16(cmd, &hdr->cmd);
697 static int send_cmd(struct send_ctx *sctx)
700 struct btrfs_cmd_header *hdr;
703 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
704 put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
705 put_unaligned_le32(0, &hdr->crc);
707 crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
708 put_unaligned_le32(crc, &hdr->crc);
710 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
713 sctx->total_send_size += sctx->send_size;
714 sctx->cmd_send_size[get_unaligned_le16(&hdr->cmd)] += sctx->send_size;
721 * Sends a move instruction to user space
723 static int send_rename(struct send_ctx *sctx,
724 struct fs_path *from, struct fs_path *to)
726 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
729 btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
731 ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
735 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
736 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
738 ret = send_cmd(sctx);
746 * Sends a link instruction to user space
748 static int send_link(struct send_ctx *sctx,
749 struct fs_path *path, struct fs_path *lnk)
751 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
754 btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
756 ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
760 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
761 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
763 ret = send_cmd(sctx);
771 * Sends an unlink instruction to user space
773 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
775 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
778 btrfs_debug(fs_info, "send_unlink %s", path->start);
780 ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
784 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
786 ret = send_cmd(sctx);
794 * Sends a rmdir instruction to user space
796 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
798 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
801 btrfs_debug(fs_info, "send_rmdir %s", path->start);
803 ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
807 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
809 ret = send_cmd(sctx);
817 * Helper function to retrieve some fields from an inode item.
819 static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
820 u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
824 struct btrfs_inode_item *ii;
825 struct btrfs_key key;
828 key.type = BTRFS_INODE_ITEM_KEY;
830 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
837 ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
838 struct btrfs_inode_item);
840 *size = btrfs_inode_size(path->nodes[0], ii);
842 *gen = btrfs_inode_generation(path->nodes[0], ii);
844 *mode = btrfs_inode_mode(path->nodes[0], ii);
846 *uid = btrfs_inode_uid(path->nodes[0], ii);
848 *gid = btrfs_inode_gid(path->nodes[0], ii);
850 *rdev = btrfs_inode_rdev(path->nodes[0], ii);
855 static int get_inode_info(struct btrfs_root *root,
856 u64 ino, u64 *size, u64 *gen,
857 u64 *mode, u64 *uid, u64 *gid,
860 struct btrfs_path *path;
863 path = alloc_path_for_send();
866 ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
868 btrfs_free_path(path);
872 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
877 * Helper function to iterate the entries in ONE btrfs_inode_ref or
878 * btrfs_inode_extref.
879 * The iterate callback may return a non zero value to stop iteration. This can
880 * be a negative value for error codes or 1 to simply stop it.
882 * path must point to the INODE_REF or INODE_EXTREF when called.
884 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
885 struct btrfs_key *found_key, int resolve,
886 iterate_inode_ref_t iterate, void *ctx)
888 struct extent_buffer *eb = path->nodes[0];
889 struct btrfs_item *item;
890 struct btrfs_inode_ref *iref;
891 struct btrfs_inode_extref *extref;
892 struct btrfs_path *tmp_path;
896 int slot = path->slots[0];
903 unsigned long name_off;
904 unsigned long elem_size;
907 p = fs_path_alloc_reversed();
911 tmp_path = alloc_path_for_send();
918 if (found_key->type == BTRFS_INODE_REF_KEY) {
919 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
920 struct btrfs_inode_ref);
921 item = btrfs_item_nr(slot);
922 total = btrfs_item_size(eb, item);
923 elem_size = sizeof(*iref);
925 ptr = btrfs_item_ptr_offset(eb, slot);
926 total = btrfs_item_size_nr(eb, slot);
927 elem_size = sizeof(*extref);
930 while (cur < total) {
933 if (found_key->type == BTRFS_INODE_REF_KEY) {
934 iref = (struct btrfs_inode_ref *)(ptr + cur);
935 name_len = btrfs_inode_ref_name_len(eb, iref);
936 name_off = (unsigned long)(iref + 1);
937 index = btrfs_inode_ref_index(eb, iref);
938 dir = found_key->offset;
940 extref = (struct btrfs_inode_extref *)(ptr + cur);
941 name_len = btrfs_inode_extref_name_len(eb, extref);
942 name_off = (unsigned long)&extref->name;
943 index = btrfs_inode_extref_index(eb, extref);
944 dir = btrfs_inode_extref_parent(eb, extref);
948 start = btrfs_ref_to_path(root, tmp_path, name_len,
952 ret = PTR_ERR(start);
955 if (start < p->buf) {
956 /* overflow , try again with larger buffer */
957 ret = fs_path_ensure_buf(p,
958 p->buf_len + p->buf - start);
961 start = btrfs_ref_to_path(root, tmp_path,
966 ret = PTR_ERR(start);
969 BUG_ON(start < p->buf);
973 ret = fs_path_add_from_extent_buffer(p, eb, name_off,
979 cur += elem_size + name_len;
980 ret = iterate(num, dir, index, p, ctx);
987 btrfs_free_path(tmp_path);
992 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
993 const char *name, int name_len,
994 const char *data, int data_len,
998 * Helper function to iterate the entries in ONE btrfs_dir_item.
999 * The iterate callback may return a non zero value to stop iteration. This can
1000 * be a negative value for error codes or 1 to simply stop it.
1002 * path must point to the dir item when called.
1004 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1005 iterate_dir_item_t iterate, void *ctx)
1008 struct extent_buffer *eb;
1009 struct btrfs_item *item;
1010 struct btrfs_dir_item *di;
1011 struct btrfs_key di_key;
1024 * Start with a small buffer (1 page). If later we end up needing more
1025 * space, which can happen for xattrs on a fs with a leaf size greater
1026 * then the page size, attempt to increase the buffer. Typically xattr
1030 buf = kmalloc(buf_len, GFP_KERNEL);
1036 eb = path->nodes[0];
1037 slot = path->slots[0];
1038 item = btrfs_item_nr(slot);
1039 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1042 total = btrfs_item_size(eb, item);
1045 while (cur < total) {
1046 name_len = btrfs_dir_name_len(eb, di);
1047 data_len = btrfs_dir_data_len(eb, di);
1048 type = btrfs_dir_type(eb, di);
1049 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1051 if (type == BTRFS_FT_XATTR) {
1052 if (name_len > XATTR_NAME_MAX) {
1053 ret = -ENAMETOOLONG;
1056 if (name_len + data_len >
1057 BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1065 if (name_len + data_len > PATH_MAX) {
1066 ret = -ENAMETOOLONG;
1071 if (name_len + data_len > buf_len) {
1072 buf_len = name_len + data_len;
1073 if (is_vmalloc_addr(buf)) {
1077 char *tmp = krealloc(buf, buf_len,
1078 GFP_KERNEL | __GFP_NOWARN);
1085 buf = kvmalloc(buf_len, GFP_KERNEL);
1093 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1094 name_len + data_len);
1096 len = sizeof(*di) + name_len + data_len;
1097 di = (struct btrfs_dir_item *)((char *)di + len);
1100 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1101 data_len, type, ctx);
1117 static int __copy_first_ref(int num, u64 dir, int index,
1118 struct fs_path *p, void *ctx)
1121 struct fs_path *pt = ctx;
1123 ret = fs_path_copy(pt, p);
1127 /* we want the first only */
1132 * Retrieve the first path of an inode. If an inode has more then one
1133 * ref/hardlink, this is ignored.
1135 static int get_inode_path(struct btrfs_root *root,
1136 u64 ino, struct fs_path *path)
1139 struct btrfs_key key, found_key;
1140 struct btrfs_path *p;
1142 p = alloc_path_for_send();
1146 fs_path_reset(path);
1149 key.type = BTRFS_INODE_REF_KEY;
1152 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1159 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1160 if (found_key.objectid != ino ||
1161 (found_key.type != BTRFS_INODE_REF_KEY &&
1162 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1167 ret = iterate_inode_ref(root, p, &found_key, 1,
1168 __copy_first_ref, path);
1178 struct backref_ctx {
1179 struct send_ctx *sctx;
1181 /* number of total found references */
1185 * used for clones found in send_root. clones found behind cur_objectid
1186 * and cur_offset are not considered as allowed clones.
1191 /* may be truncated in case it's the last extent in a file */
1194 /* Just to check for bugs in backref resolving */
1198 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1200 u64 root = (u64)(uintptr_t)key;
1201 const struct clone_root *cr = elt;
1203 if (root < cr->root->root_key.objectid)
1205 if (root > cr->root->root_key.objectid)
1210 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1212 const struct clone_root *cr1 = e1;
1213 const struct clone_root *cr2 = e2;
1215 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1217 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1223 * Called for every backref that is found for the current extent.
1224 * Results are collected in sctx->clone_roots->ino/offset/found_refs
1226 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1228 struct backref_ctx *bctx = ctx_;
1229 struct clone_root *found;
1231 /* First check if the root is in the list of accepted clone sources */
1232 found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1233 bctx->sctx->clone_roots_cnt,
1234 sizeof(struct clone_root),
1235 __clone_root_cmp_bsearch);
1239 if (found->root == bctx->sctx->send_root &&
1240 ino == bctx->cur_objectid &&
1241 offset == bctx->cur_offset) {
1242 bctx->found_itself = 1;
1246 * Make sure we don't consider clones from send_root that are
1247 * behind the current inode/offset.
1249 if (found->root == bctx->sctx->send_root) {
1251 * If the source inode was not yet processed we can't issue a
1252 * clone operation, as the source extent does not exist yet at
1253 * the destination of the stream.
1255 if (ino > bctx->cur_objectid)
1258 * We clone from the inode currently being sent as long as the
1259 * source extent is already processed, otherwise we could try
1260 * to clone from an extent that does not exist yet at the
1261 * destination of the stream.
1263 if (ino == bctx->cur_objectid &&
1264 offset + bctx->extent_len >
1265 bctx->sctx->cur_inode_next_write_offset)
1270 found->found_refs++;
1271 if (ino < found->ino) {
1273 found->offset = offset;
1274 } else if (found->ino == ino) {
1276 * same extent found more then once in the same file.
1278 if (found->offset > offset + bctx->extent_len)
1279 found->offset = offset;
1286 * Given an inode, offset and extent item, it finds a good clone for a clone
1287 * instruction. Returns -ENOENT when none could be found. The function makes
1288 * sure that the returned clone is usable at the point where sending is at the
1289 * moment. This means, that no clones are accepted which lie behind the current
1292 * path must point to the extent item when called.
1294 static int find_extent_clone(struct send_ctx *sctx,
1295 struct btrfs_path *path,
1296 u64 ino, u64 data_offset,
1298 struct clone_root **found)
1300 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1306 u64 extent_item_pos;
1308 struct btrfs_file_extent_item *fi;
1309 struct extent_buffer *eb = path->nodes[0];
1310 struct backref_ctx backref_ctx = {0};
1311 struct clone_root *cur_clone_root;
1312 struct btrfs_key found_key;
1313 struct btrfs_path *tmp_path;
1314 struct btrfs_extent_item *ei;
1318 tmp_path = alloc_path_for_send();
1322 /* We only use this path under the commit sem */
1323 tmp_path->need_commit_sem = 0;
1325 if (data_offset >= ino_size) {
1327 * There may be extents that lie behind the file's size.
1328 * I at least had this in combination with snapshotting while
1329 * writing large files.
1335 fi = btrfs_item_ptr(eb, path->slots[0],
1336 struct btrfs_file_extent_item);
1337 extent_type = btrfs_file_extent_type(eb, fi);
1338 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1342 compressed = btrfs_file_extent_compression(eb, fi);
1344 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1345 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1346 if (disk_byte == 0) {
1350 logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1352 down_read(&fs_info->commit_root_sem);
1353 ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1354 &found_key, &flags);
1355 up_read(&fs_info->commit_root_sem);
1359 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1364 ei = btrfs_item_ptr(tmp_path->nodes[0], tmp_path->slots[0],
1365 struct btrfs_extent_item);
1367 * Backreference walking (iterate_extent_inodes() below) is currently
1368 * too expensive when an extent has a large number of references, both
1369 * in time spent and used memory. So for now just fallback to write
1370 * operations instead of clone operations when an extent has more than
1371 * a certain amount of references.
1373 if (btrfs_extent_refs(tmp_path->nodes[0], ei) > SEND_MAX_EXTENT_REFS) {
1377 btrfs_release_path(tmp_path);
1380 * Setup the clone roots.
1382 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1383 cur_clone_root = sctx->clone_roots + i;
1384 cur_clone_root->ino = (u64)-1;
1385 cur_clone_root->offset = 0;
1386 cur_clone_root->found_refs = 0;
1389 backref_ctx.sctx = sctx;
1390 backref_ctx.found = 0;
1391 backref_ctx.cur_objectid = ino;
1392 backref_ctx.cur_offset = data_offset;
1393 backref_ctx.found_itself = 0;
1394 backref_ctx.extent_len = num_bytes;
1397 * The last extent of a file may be too large due to page alignment.
1398 * We need to adjust extent_len in this case so that the checks in
1399 * __iterate_backrefs work.
1401 if (data_offset + num_bytes >= ino_size)
1402 backref_ctx.extent_len = ino_size - data_offset;
1405 * Now collect all backrefs.
1407 if (compressed == BTRFS_COMPRESS_NONE)
1408 extent_item_pos = logical - found_key.objectid;
1410 extent_item_pos = 0;
1411 ret = iterate_extent_inodes(fs_info, found_key.objectid,
1412 extent_item_pos, 1, __iterate_backrefs,
1413 &backref_ctx, false);
1418 if (!backref_ctx.found_itself) {
1419 /* found a bug in backref code? */
1422 "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1423 ino, data_offset, disk_byte, found_key.objectid);
1427 btrfs_debug(fs_info,
1428 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1429 data_offset, ino, num_bytes, logical);
1431 if (!backref_ctx.found)
1432 btrfs_debug(fs_info, "no clones found");
1434 cur_clone_root = NULL;
1435 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1436 if (sctx->clone_roots[i].found_refs) {
1437 if (!cur_clone_root)
1438 cur_clone_root = sctx->clone_roots + i;
1439 else if (sctx->clone_roots[i].root == sctx->send_root)
1440 /* prefer clones from send_root over others */
1441 cur_clone_root = sctx->clone_roots + i;
1446 if (cur_clone_root) {
1447 *found = cur_clone_root;
1454 btrfs_free_path(tmp_path);
1458 static int read_symlink(struct btrfs_root *root,
1460 struct fs_path *dest)
1463 struct btrfs_path *path;
1464 struct btrfs_key key;
1465 struct btrfs_file_extent_item *ei;
1471 path = alloc_path_for_send();
1476 key.type = BTRFS_EXTENT_DATA_KEY;
1478 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1483 * An empty symlink inode. Can happen in rare error paths when
1484 * creating a symlink (transaction committed before the inode
1485 * eviction handler removed the symlink inode items and a crash
1486 * happened in between or the subvol was snapshoted in between).
1487 * Print an informative message to dmesg/syslog so that the user
1488 * can delete the symlink.
1490 btrfs_err(root->fs_info,
1491 "Found empty symlink inode %llu at root %llu",
1492 ino, root->root_key.objectid);
1497 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1498 struct btrfs_file_extent_item);
1499 type = btrfs_file_extent_type(path->nodes[0], ei);
1500 compression = btrfs_file_extent_compression(path->nodes[0], ei);
1501 BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1502 BUG_ON(compression);
1504 off = btrfs_file_extent_inline_start(ei);
1505 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1507 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1510 btrfs_free_path(path);
1515 * Helper function to generate a file name that is unique in the root of
1516 * send_root and parent_root. This is used to generate names for orphan inodes.
1518 static int gen_unique_name(struct send_ctx *sctx,
1520 struct fs_path *dest)
1523 struct btrfs_path *path;
1524 struct btrfs_dir_item *di;
1529 path = alloc_path_for_send();
1534 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1536 ASSERT(len < sizeof(tmp));
1538 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1539 path, BTRFS_FIRST_FREE_OBJECTID,
1540 tmp, strlen(tmp), 0);
1541 btrfs_release_path(path);
1547 /* not unique, try again */
1552 if (!sctx->parent_root) {
1558 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1559 path, BTRFS_FIRST_FREE_OBJECTID,
1560 tmp, strlen(tmp), 0);
1561 btrfs_release_path(path);
1567 /* not unique, try again */
1575 ret = fs_path_add(dest, tmp, strlen(tmp));
1578 btrfs_free_path(path);
1583 inode_state_no_change,
1584 inode_state_will_create,
1585 inode_state_did_create,
1586 inode_state_will_delete,
1587 inode_state_did_delete,
1590 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1598 ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1600 if (ret < 0 && ret != -ENOENT)
1604 if (!sctx->parent_root) {
1605 right_ret = -ENOENT;
1607 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1608 NULL, NULL, NULL, NULL);
1609 if (ret < 0 && ret != -ENOENT)
1614 if (!left_ret && !right_ret) {
1615 if (left_gen == gen && right_gen == gen) {
1616 ret = inode_state_no_change;
1617 } else if (left_gen == gen) {
1618 if (ino < sctx->send_progress)
1619 ret = inode_state_did_create;
1621 ret = inode_state_will_create;
1622 } else if (right_gen == gen) {
1623 if (ino < sctx->send_progress)
1624 ret = inode_state_did_delete;
1626 ret = inode_state_will_delete;
1630 } else if (!left_ret) {
1631 if (left_gen == gen) {
1632 if (ino < sctx->send_progress)
1633 ret = inode_state_did_create;
1635 ret = inode_state_will_create;
1639 } else if (!right_ret) {
1640 if (right_gen == gen) {
1641 if (ino < sctx->send_progress)
1642 ret = inode_state_did_delete;
1644 ret = inode_state_will_delete;
1656 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1660 if (ino == BTRFS_FIRST_FREE_OBJECTID)
1663 ret = get_cur_inode_state(sctx, ino, gen);
1667 if (ret == inode_state_no_change ||
1668 ret == inode_state_did_create ||
1669 ret == inode_state_will_delete)
1679 * Helper function to lookup a dir item in a dir.
1681 static int lookup_dir_item_inode(struct btrfs_root *root,
1682 u64 dir, const char *name, int name_len,
1687 struct btrfs_dir_item *di;
1688 struct btrfs_key key;
1689 struct btrfs_path *path;
1691 path = alloc_path_for_send();
1695 di = btrfs_lookup_dir_item(NULL, root, path,
1696 dir, name, name_len, 0);
1697 if (IS_ERR_OR_NULL(di)) {
1698 ret = di ? PTR_ERR(di) : -ENOENT;
1701 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1702 if (key.type == BTRFS_ROOT_ITEM_KEY) {
1706 *found_inode = key.objectid;
1707 *found_type = btrfs_dir_type(path->nodes[0], di);
1710 btrfs_free_path(path);
1715 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1716 * generation of the parent dir and the name of the dir entry.
1718 static int get_first_ref(struct btrfs_root *root, u64 ino,
1719 u64 *dir, u64 *dir_gen, struct fs_path *name)
1722 struct btrfs_key key;
1723 struct btrfs_key found_key;
1724 struct btrfs_path *path;
1728 path = alloc_path_for_send();
1733 key.type = BTRFS_INODE_REF_KEY;
1736 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1740 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1742 if (ret || found_key.objectid != ino ||
1743 (found_key.type != BTRFS_INODE_REF_KEY &&
1744 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1749 if (found_key.type == BTRFS_INODE_REF_KEY) {
1750 struct btrfs_inode_ref *iref;
1751 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1752 struct btrfs_inode_ref);
1753 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1754 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1755 (unsigned long)(iref + 1),
1757 parent_dir = found_key.offset;
1759 struct btrfs_inode_extref *extref;
1760 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1761 struct btrfs_inode_extref);
1762 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1763 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1764 (unsigned long)&extref->name, len);
1765 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1769 btrfs_release_path(path);
1772 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1781 btrfs_free_path(path);
1785 static int is_first_ref(struct btrfs_root *root,
1787 const char *name, int name_len)
1790 struct fs_path *tmp_name;
1793 tmp_name = fs_path_alloc();
1797 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1801 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1806 ret = !memcmp(tmp_name->start, name, name_len);
1809 fs_path_free(tmp_name);
1814 * Used by process_recorded_refs to determine if a new ref would overwrite an
1815 * already existing ref. In case it detects an overwrite, it returns the
1816 * inode/gen in who_ino/who_gen.
1817 * When an overwrite is detected, process_recorded_refs does proper orphanizing
1818 * to make sure later references to the overwritten inode are possible.
1819 * Orphanizing is however only required for the first ref of an inode.
1820 * process_recorded_refs does an additional is_first_ref check to see if
1821 * orphanizing is really required.
1823 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1824 const char *name, int name_len,
1825 u64 *who_ino, u64 *who_gen, u64 *who_mode)
1829 u64 other_inode = 0;
1832 if (!sctx->parent_root)
1835 ret = is_inode_existent(sctx, dir, dir_gen);
1840 * If we have a parent root we need to verify that the parent dir was
1841 * not deleted and then re-created, if it was then we have no overwrite
1842 * and we can just unlink this entry.
1844 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1845 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1847 if (ret < 0 && ret != -ENOENT)
1857 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1858 &other_inode, &other_type);
1859 if (ret < 0 && ret != -ENOENT)
1867 * Check if the overwritten ref was already processed. If yes, the ref
1868 * was already unlinked/moved, so we can safely assume that we will not
1869 * overwrite anything at this point in time.
1871 if (other_inode > sctx->send_progress ||
1872 is_waiting_for_move(sctx, other_inode)) {
1873 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1874 who_gen, who_mode, NULL, NULL, NULL);
1879 *who_ino = other_inode;
1889 * Checks if the ref was overwritten by an already processed inode. This is
1890 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1891 * thus the orphan name needs be used.
1892 * process_recorded_refs also uses it to avoid unlinking of refs that were
1895 static int did_overwrite_ref(struct send_ctx *sctx,
1896 u64 dir, u64 dir_gen,
1897 u64 ino, u64 ino_gen,
1898 const char *name, int name_len)
1905 if (!sctx->parent_root)
1908 ret = is_inode_existent(sctx, dir, dir_gen);
1912 if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1913 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1915 if (ret < 0 && ret != -ENOENT)
1925 /* check if the ref was overwritten by another ref */
1926 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1927 &ow_inode, &other_type);
1928 if (ret < 0 && ret != -ENOENT)
1931 /* was never and will never be overwritten */
1936 ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1941 if (ow_inode == ino && gen == ino_gen) {
1947 * We know that it is or will be overwritten. Check this now.
1948 * The current inode being processed might have been the one that caused
1949 * inode 'ino' to be orphanized, therefore check if ow_inode matches
1950 * the current inode being processed.
1952 if ((ow_inode < sctx->send_progress) ||
1953 (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1954 gen == sctx->cur_inode_gen))
1964 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1965 * that got overwritten. This is used by process_recorded_refs to determine
1966 * if it has to use the path as returned by get_cur_path or the orphan name.
1968 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1971 struct fs_path *name = NULL;
1975 if (!sctx->parent_root)
1978 name = fs_path_alloc();
1982 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
1986 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
1987 name->start, fs_path_len(name));
1995 * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
1996 * so we need to do some special handling in case we have clashes. This function
1997 * takes care of this with the help of name_cache_entry::radix_list.
1998 * In case of error, nce is kfreed.
2000 static int name_cache_insert(struct send_ctx *sctx,
2001 struct name_cache_entry *nce)
2004 struct list_head *nce_head;
2006 nce_head = radix_tree_lookup(&sctx->name_cache,
2007 (unsigned long)nce->ino);
2009 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2014 INIT_LIST_HEAD(nce_head);
2016 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2023 list_add_tail(&nce->radix_list, nce_head);
2024 list_add_tail(&nce->list, &sctx->name_cache_list);
2025 sctx->name_cache_size++;
2030 static void name_cache_delete(struct send_ctx *sctx,
2031 struct name_cache_entry *nce)
2033 struct list_head *nce_head;
2035 nce_head = radix_tree_lookup(&sctx->name_cache,
2036 (unsigned long)nce->ino);
2038 btrfs_err(sctx->send_root->fs_info,
2039 "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2040 nce->ino, sctx->name_cache_size);
2043 list_del(&nce->radix_list);
2044 list_del(&nce->list);
2045 sctx->name_cache_size--;
2048 * We may not get to the final release of nce_head if the lookup fails
2050 if (nce_head && list_empty(nce_head)) {
2051 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2056 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2059 struct list_head *nce_head;
2060 struct name_cache_entry *cur;
2062 nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2066 list_for_each_entry(cur, nce_head, radix_list) {
2067 if (cur->ino == ino && cur->gen == gen)
2074 * Remove some entries from the beginning of name_cache_list.
2076 static void name_cache_clean_unused(struct send_ctx *sctx)
2078 struct name_cache_entry *nce;
2080 if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2083 while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2084 nce = list_entry(sctx->name_cache_list.next,
2085 struct name_cache_entry, list);
2086 name_cache_delete(sctx, nce);
2091 static void name_cache_free(struct send_ctx *sctx)
2093 struct name_cache_entry *nce;
2095 while (!list_empty(&sctx->name_cache_list)) {
2096 nce = list_entry(sctx->name_cache_list.next,
2097 struct name_cache_entry, list);
2098 name_cache_delete(sctx, nce);
2104 * Used by get_cur_path for each ref up to the root.
2105 * Returns 0 if it succeeded.
2106 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2107 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2108 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2109 * Returns <0 in case of error.
2111 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2115 struct fs_path *dest)
2119 struct name_cache_entry *nce = NULL;
2122 * First check if we already did a call to this function with the same
2123 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2124 * return the cached result.
2126 nce = name_cache_search(sctx, ino, gen);
2128 if (ino < sctx->send_progress && nce->need_later_update) {
2129 name_cache_delete(sctx, nce);
2134 * Removes the entry from the list and adds it back to
2135 * the end. This marks the entry as recently used so
2136 * that name_cache_clean_unused does not remove it.
2138 list_move_tail(&nce->list, &sctx->name_cache_list);
2140 *parent_ino = nce->parent_ino;
2141 *parent_gen = nce->parent_gen;
2142 ret = fs_path_add(dest, nce->name, nce->name_len);
2151 * If the inode is not existent yet, add the orphan name and return 1.
2152 * This should only happen for the parent dir that we determine in
2155 ret = is_inode_existent(sctx, ino, gen);
2160 ret = gen_unique_name(sctx, ino, gen, dest);
2168 * Depending on whether the inode was already processed or not, use
2169 * send_root or parent_root for ref lookup.
2171 if (ino < sctx->send_progress)
2172 ret = get_first_ref(sctx->send_root, ino,
2173 parent_ino, parent_gen, dest);
2175 ret = get_first_ref(sctx->parent_root, ino,
2176 parent_ino, parent_gen, dest);
2181 * Check if the ref was overwritten by an inode's ref that was processed
2182 * earlier. If yes, treat as orphan and return 1.
2184 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2185 dest->start, dest->end - dest->start);
2189 fs_path_reset(dest);
2190 ret = gen_unique_name(sctx, ino, gen, dest);
2198 * Store the result of the lookup in the name cache.
2200 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2208 nce->parent_ino = *parent_ino;
2209 nce->parent_gen = *parent_gen;
2210 nce->name_len = fs_path_len(dest);
2212 strcpy(nce->name, dest->start);
2214 if (ino < sctx->send_progress)
2215 nce->need_later_update = 0;
2217 nce->need_later_update = 1;
2219 nce_ret = name_cache_insert(sctx, nce);
2222 name_cache_clean_unused(sctx);
2229 * Magic happens here. This function returns the first ref to an inode as it
2230 * would look like while receiving the stream at this point in time.
2231 * We walk the path up to the root. For every inode in between, we check if it
2232 * was already processed/sent. If yes, we continue with the parent as found
2233 * in send_root. If not, we continue with the parent as found in parent_root.
2234 * If we encounter an inode that was deleted at this point in time, we use the
2235 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2236 * that were not created yet and overwritten inodes/refs.
2238 * When do we have orphan inodes:
2239 * 1. When an inode is freshly created and thus no valid refs are available yet
2240 * 2. When a directory lost all it's refs (deleted) but still has dir items
2241 * inside which were not processed yet (pending for move/delete). If anyone
2242 * tried to get the path to the dir items, it would get a path inside that
2244 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2245 * of an unprocessed inode. If in that case the first ref would be
2246 * overwritten, the overwritten inode gets "orphanized". Later when we
2247 * process this overwritten inode, it is restored at a new place by moving
2250 * sctx->send_progress tells this function at which point in time receiving
2253 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2254 struct fs_path *dest)
2257 struct fs_path *name = NULL;
2258 u64 parent_inode = 0;
2262 name = fs_path_alloc();
2269 fs_path_reset(dest);
2271 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2272 struct waiting_dir_move *wdm;
2274 fs_path_reset(name);
2276 if (is_waiting_for_rm(sctx, ino, gen)) {
2277 ret = gen_unique_name(sctx, ino, gen, name);
2280 ret = fs_path_add_path(dest, name);
2284 wdm = get_waiting_dir_move(sctx, ino);
2285 if (wdm && wdm->orphanized) {
2286 ret = gen_unique_name(sctx, ino, gen, name);
2289 ret = get_first_ref(sctx->parent_root, ino,
2290 &parent_inode, &parent_gen, name);
2292 ret = __get_cur_name_and_parent(sctx, ino, gen,
2302 ret = fs_path_add_path(dest, name);
2313 fs_path_unreverse(dest);
2318 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2320 static int send_subvol_begin(struct send_ctx *sctx)
2323 struct btrfs_root *send_root = sctx->send_root;
2324 struct btrfs_root *parent_root = sctx->parent_root;
2325 struct btrfs_path *path;
2326 struct btrfs_key key;
2327 struct btrfs_root_ref *ref;
2328 struct extent_buffer *leaf;
2332 path = btrfs_alloc_path();
2336 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2338 btrfs_free_path(path);
2342 key.objectid = send_root->root_key.objectid;
2343 key.type = BTRFS_ROOT_BACKREF_KEY;
2346 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2355 leaf = path->nodes[0];
2356 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2357 if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2358 key.objectid != send_root->root_key.objectid) {
2362 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2363 namelen = btrfs_root_ref_name_len(leaf, ref);
2364 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2365 btrfs_release_path(path);
2368 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2372 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2377 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2379 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2380 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2381 sctx->send_root->root_item.received_uuid);
2383 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2384 sctx->send_root->root_item.uuid);
2386 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2387 btrfs_root_ctransid(&sctx->send_root->root_item));
2389 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2390 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2391 parent_root->root_item.received_uuid);
2393 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2394 parent_root->root_item.uuid);
2395 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2396 btrfs_root_ctransid(&sctx->parent_root->root_item));
2399 ret = send_cmd(sctx);
2403 btrfs_free_path(path);
2408 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2410 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2414 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2416 p = fs_path_alloc();
2420 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2424 ret = get_cur_path(sctx, ino, gen, p);
2427 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2428 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2430 ret = send_cmd(sctx);
2438 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2440 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2444 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2446 p = fs_path_alloc();
2450 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2454 ret = get_cur_path(sctx, ino, gen, p);
2457 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2458 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2460 ret = send_cmd(sctx);
2468 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2470 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2474 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2477 p = fs_path_alloc();
2481 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2485 ret = get_cur_path(sctx, ino, gen, p);
2488 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2489 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2490 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2492 ret = send_cmd(sctx);
2500 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2502 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2504 struct fs_path *p = NULL;
2505 struct btrfs_inode_item *ii;
2506 struct btrfs_path *path = NULL;
2507 struct extent_buffer *eb;
2508 struct btrfs_key key;
2511 btrfs_debug(fs_info, "send_utimes %llu", ino);
2513 p = fs_path_alloc();
2517 path = alloc_path_for_send();
2524 key.type = BTRFS_INODE_ITEM_KEY;
2526 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2532 eb = path->nodes[0];
2533 slot = path->slots[0];
2534 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2536 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2540 ret = get_cur_path(sctx, ino, gen, p);
2543 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2544 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2545 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2546 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2547 /* TODO Add otime support when the otime patches get into upstream */
2549 ret = send_cmd(sctx);
2554 btrfs_free_path(path);
2559 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2560 * a valid path yet because we did not process the refs yet. So, the inode
2561 * is created as orphan.
2563 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2565 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2573 btrfs_debug(fs_info, "send_create_inode %llu", ino);
2575 p = fs_path_alloc();
2579 if (ino != sctx->cur_ino) {
2580 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2585 gen = sctx->cur_inode_gen;
2586 mode = sctx->cur_inode_mode;
2587 rdev = sctx->cur_inode_rdev;
2590 if (S_ISREG(mode)) {
2591 cmd = BTRFS_SEND_C_MKFILE;
2592 } else if (S_ISDIR(mode)) {
2593 cmd = BTRFS_SEND_C_MKDIR;
2594 } else if (S_ISLNK(mode)) {
2595 cmd = BTRFS_SEND_C_SYMLINK;
2596 } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2597 cmd = BTRFS_SEND_C_MKNOD;
2598 } else if (S_ISFIFO(mode)) {
2599 cmd = BTRFS_SEND_C_MKFIFO;
2600 } else if (S_ISSOCK(mode)) {
2601 cmd = BTRFS_SEND_C_MKSOCK;
2603 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2604 (int)(mode & S_IFMT));
2609 ret = begin_cmd(sctx, cmd);
2613 ret = gen_unique_name(sctx, ino, gen, p);
2617 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2618 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2620 if (S_ISLNK(mode)) {
2622 ret = read_symlink(sctx->send_root, ino, p);
2625 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2626 } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2627 S_ISFIFO(mode) || S_ISSOCK(mode)) {
2628 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2629 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2632 ret = send_cmd(sctx);
2644 * We need some special handling for inodes that get processed before the parent
2645 * directory got created. See process_recorded_refs for details.
2646 * This function does the check if we already created the dir out of order.
2648 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2651 struct btrfs_path *path = NULL;
2652 struct btrfs_key key;
2653 struct btrfs_key found_key;
2654 struct btrfs_key di_key;
2655 struct extent_buffer *eb;
2656 struct btrfs_dir_item *di;
2659 path = alloc_path_for_send();
2666 key.type = BTRFS_DIR_INDEX_KEY;
2668 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2673 eb = path->nodes[0];
2674 slot = path->slots[0];
2675 if (slot >= btrfs_header_nritems(eb)) {
2676 ret = btrfs_next_leaf(sctx->send_root, path);
2679 } else if (ret > 0) {
2686 btrfs_item_key_to_cpu(eb, &found_key, slot);
2687 if (found_key.objectid != key.objectid ||
2688 found_key.type != key.type) {
2693 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2694 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2696 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2697 di_key.objectid < sctx->send_progress) {
2706 btrfs_free_path(path);
2711 * Only creates the inode if it is:
2712 * 1. Not a directory
2713 * 2. Or a directory which was not created already due to out of order
2714 * directories. See did_create_dir and process_recorded_refs for details.
2716 static int send_create_inode_if_needed(struct send_ctx *sctx)
2720 if (S_ISDIR(sctx->cur_inode_mode)) {
2721 ret = did_create_dir(sctx, sctx->cur_ino);
2730 ret = send_create_inode(sctx, sctx->cur_ino);
2738 struct recorded_ref {
2739 struct list_head list;
2741 struct fs_path *full_path;
2747 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
2749 ref->full_path = path;
2750 ref->name = (char *)kbasename(ref->full_path->start);
2751 ref->name_len = ref->full_path->end - ref->name;
2755 * We need to process new refs before deleted refs, but compare_tree gives us
2756 * everything mixed. So we first record all refs and later process them.
2757 * This function is a helper to record one ref.
2759 static int __record_ref(struct list_head *head, u64 dir,
2760 u64 dir_gen, struct fs_path *path)
2762 struct recorded_ref *ref;
2764 ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2769 ref->dir_gen = dir_gen;
2770 set_ref_path(ref, path);
2771 list_add_tail(&ref->list, head);
2775 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2777 struct recorded_ref *new;
2779 new = kmalloc(sizeof(*ref), GFP_KERNEL);
2783 new->dir = ref->dir;
2784 new->dir_gen = ref->dir_gen;
2785 new->full_path = NULL;
2786 INIT_LIST_HEAD(&new->list);
2787 list_add_tail(&new->list, list);
2791 static void __free_recorded_refs(struct list_head *head)
2793 struct recorded_ref *cur;
2795 while (!list_empty(head)) {
2796 cur = list_entry(head->next, struct recorded_ref, list);
2797 fs_path_free(cur->full_path);
2798 list_del(&cur->list);
2803 static void free_recorded_refs(struct send_ctx *sctx)
2805 __free_recorded_refs(&sctx->new_refs);
2806 __free_recorded_refs(&sctx->deleted_refs);
2810 * Renames/moves a file/dir to its orphan name. Used when the first
2811 * ref of an unprocessed inode gets overwritten and for all non empty
2814 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2815 struct fs_path *path)
2818 struct fs_path *orphan;
2820 orphan = fs_path_alloc();
2824 ret = gen_unique_name(sctx, ino, gen, orphan);
2828 ret = send_rename(sctx, path, orphan);
2831 fs_path_free(orphan);
2835 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
2836 u64 dir_ino, u64 dir_gen)
2838 struct rb_node **p = &sctx->orphan_dirs.rb_node;
2839 struct rb_node *parent = NULL;
2840 struct orphan_dir_info *entry, *odi;
2844 entry = rb_entry(parent, struct orphan_dir_info, node);
2845 if (dir_ino < entry->ino)
2847 else if (dir_ino > entry->ino)
2848 p = &(*p)->rb_right;
2849 else if (dir_gen < entry->gen)
2851 else if (dir_gen > entry->gen)
2852 p = &(*p)->rb_right;
2857 odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2859 return ERR_PTR(-ENOMEM);
2862 odi->last_dir_index_offset = 0;
2864 rb_link_node(&odi->node, parent, p);
2865 rb_insert_color(&odi->node, &sctx->orphan_dirs);
2869 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
2870 u64 dir_ino, u64 gen)
2872 struct rb_node *n = sctx->orphan_dirs.rb_node;
2873 struct orphan_dir_info *entry;
2876 entry = rb_entry(n, struct orphan_dir_info, node);
2877 if (dir_ino < entry->ino)
2879 else if (dir_ino > entry->ino)
2881 else if (gen < entry->gen)
2883 else if (gen > entry->gen)
2891 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
2893 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
2898 static void free_orphan_dir_info(struct send_ctx *sctx,
2899 struct orphan_dir_info *odi)
2903 rb_erase(&odi->node, &sctx->orphan_dirs);
2908 * Returns 1 if a directory can be removed at this point in time.
2909 * We check this by iterating all dir items and checking if the inode behind
2910 * the dir item was already processed.
2912 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2916 struct btrfs_root *root = sctx->parent_root;
2917 struct btrfs_path *path;
2918 struct btrfs_key key;
2919 struct btrfs_key found_key;
2920 struct btrfs_key loc;
2921 struct btrfs_dir_item *di;
2922 struct orphan_dir_info *odi = NULL;
2925 * Don't try to rmdir the top/root subvolume dir.
2927 if (dir == BTRFS_FIRST_FREE_OBJECTID)
2930 path = alloc_path_for_send();
2935 key.type = BTRFS_DIR_INDEX_KEY;
2938 odi = get_orphan_dir_info(sctx, dir, dir_gen);
2940 key.offset = odi->last_dir_index_offset;
2942 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2947 struct waiting_dir_move *dm;
2949 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2950 ret = btrfs_next_leaf(root, path);
2957 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2959 if (found_key.objectid != key.objectid ||
2960 found_key.type != key.type)
2963 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2964 struct btrfs_dir_item);
2965 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2967 dm = get_waiting_dir_move(sctx, loc.objectid);
2969 odi = add_orphan_dir_info(sctx, dir, dir_gen);
2975 odi->last_dir_index_offset = found_key.offset;
2976 dm->rmdir_ino = dir;
2977 dm->rmdir_gen = dir_gen;
2982 if (loc.objectid > send_progress) {
2983 odi = add_orphan_dir_info(sctx, dir, dir_gen);
2989 odi->last_dir_index_offset = found_key.offset;
2996 free_orphan_dir_info(sctx, odi);
3001 btrfs_free_path(path);
3005 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3007 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3009 return entry != NULL;
3012 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3014 struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3015 struct rb_node *parent = NULL;
3016 struct waiting_dir_move *entry, *dm;
3018 dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3024 dm->orphanized = orphanized;
3028 entry = rb_entry(parent, struct waiting_dir_move, node);
3029 if (ino < entry->ino) {
3031 } else if (ino > entry->ino) {
3032 p = &(*p)->rb_right;
3039 rb_link_node(&dm->node, parent, p);
3040 rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3044 static struct waiting_dir_move *
3045 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3047 struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3048 struct waiting_dir_move *entry;
3051 entry = rb_entry(n, struct waiting_dir_move, node);
3052 if (ino < entry->ino)
3054 else if (ino > entry->ino)
3062 static void free_waiting_dir_move(struct send_ctx *sctx,
3063 struct waiting_dir_move *dm)
3067 rb_erase(&dm->node, &sctx->waiting_dir_moves);
3071 static int add_pending_dir_move(struct send_ctx *sctx,
3075 struct list_head *new_refs,
3076 struct list_head *deleted_refs,
3077 const bool is_orphan)
3079 struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3080 struct rb_node *parent = NULL;
3081 struct pending_dir_move *entry = NULL, *pm;
3082 struct recorded_ref *cur;
3086 pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3089 pm->parent_ino = parent_ino;
3092 INIT_LIST_HEAD(&pm->list);
3093 INIT_LIST_HEAD(&pm->update_refs);
3094 RB_CLEAR_NODE(&pm->node);
3098 entry = rb_entry(parent, struct pending_dir_move, node);
3099 if (parent_ino < entry->parent_ino) {
3101 } else if (parent_ino > entry->parent_ino) {
3102 p = &(*p)->rb_right;
3109 list_for_each_entry(cur, deleted_refs, list) {
3110 ret = dup_ref(cur, &pm->update_refs);
3114 list_for_each_entry(cur, new_refs, list) {
3115 ret = dup_ref(cur, &pm->update_refs);
3120 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3125 list_add_tail(&pm->list, &entry->list);
3127 rb_link_node(&pm->node, parent, p);
3128 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3133 __free_recorded_refs(&pm->update_refs);
3139 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3142 struct rb_node *n = sctx->pending_dir_moves.rb_node;
3143 struct pending_dir_move *entry;
3146 entry = rb_entry(n, struct pending_dir_move, node);
3147 if (parent_ino < entry->parent_ino)
3149 else if (parent_ino > entry->parent_ino)
3157 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3158 u64 ino, u64 gen, u64 *ancestor_ino)
3161 u64 parent_inode = 0;
3163 u64 start_ino = ino;
3166 while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3167 fs_path_reset(name);
3169 if (is_waiting_for_rm(sctx, ino, gen))
3171 if (is_waiting_for_move(sctx, ino)) {
3172 if (*ancestor_ino == 0)
3173 *ancestor_ino = ino;
3174 ret = get_first_ref(sctx->parent_root, ino,
3175 &parent_inode, &parent_gen, name);
3177 ret = __get_cur_name_and_parent(sctx, ino, gen,
3187 if (parent_inode == start_ino) {
3189 if (*ancestor_ino == 0)
3190 *ancestor_ino = ino;
3199 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3201 struct fs_path *from_path = NULL;
3202 struct fs_path *to_path = NULL;
3203 struct fs_path *name = NULL;
3204 u64 orig_progress = sctx->send_progress;
3205 struct recorded_ref *cur;
3206 u64 parent_ino, parent_gen;
3207 struct waiting_dir_move *dm = NULL;
3214 name = fs_path_alloc();
3215 from_path = fs_path_alloc();
3216 if (!name || !from_path) {
3221 dm = get_waiting_dir_move(sctx, pm->ino);
3223 rmdir_ino = dm->rmdir_ino;
3224 rmdir_gen = dm->rmdir_gen;
3225 is_orphan = dm->orphanized;
3226 free_waiting_dir_move(sctx, dm);
3229 ret = gen_unique_name(sctx, pm->ino,
3230 pm->gen, from_path);
3232 ret = get_first_ref(sctx->parent_root, pm->ino,
3233 &parent_ino, &parent_gen, name);
3236 ret = get_cur_path(sctx, parent_ino, parent_gen,
3240 ret = fs_path_add_path(from_path, name);
3245 sctx->send_progress = sctx->cur_ino + 1;
3246 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3250 LIST_HEAD(deleted_refs);
3251 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3252 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3253 &pm->update_refs, &deleted_refs,
3258 dm = get_waiting_dir_move(sctx, pm->ino);
3260 dm->rmdir_ino = rmdir_ino;
3261 dm->rmdir_gen = rmdir_gen;
3265 fs_path_reset(name);
3268 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3272 ret = send_rename(sctx, from_path, to_path);
3277 struct orphan_dir_info *odi;
3280 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3282 /* already deleted */
3287 ret = can_rmdir(sctx, rmdir_ino, gen, sctx->cur_ino);
3293 name = fs_path_alloc();
3298 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3301 ret = send_rmdir(sctx, name);
3307 ret = send_utimes(sctx, pm->ino, pm->gen);
3312 * After rename/move, need to update the utimes of both new parent(s)
3313 * and old parent(s).
3315 list_for_each_entry(cur, &pm->update_refs, list) {
3317 * The parent inode might have been deleted in the send snapshot
3319 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3320 NULL, NULL, NULL, NULL, NULL);
3321 if (ret == -ENOENT) {
3328 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3335 fs_path_free(from_path);
3336 fs_path_free(to_path);
3337 sctx->send_progress = orig_progress;
3342 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3344 if (!list_empty(&m->list))
3346 if (!RB_EMPTY_NODE(&m->node))
3347 rb_erase(&m->node, &sctx->pending_dir_moves);
3348 __free_recorded_refs(&m->update_refs);
3352 static void tail_append_pending_moves(struct send_ctx *sctx,
3353 struct pending_dir_move *moves,
3354 struct list_head *stack)
3356 if (list_empty(&moves->list)) {
3357 list_add_tail(&moves->list, stack);
3360 list_splice_init(&moves->list, &list);
3361 list_add_tail(&moves->list, stack);
3362 list_splice_tail(&list, stack);
3364 if (!RB_EMPTY_NODE(&moves->node)) {
3365 rb_erase(&moves->node, &sctx->pending_dir_moves);
3366 RB_CLEAR_NODE(&moves->node);
3370 static int apply_children_dir_moves(struct send_ctx *sctx)
3372 struct pending_dir_move *pm;
3373 struct list_head stack;
3374 u64 parent_ino = sctx->cur_ino;
3377 pm = get_pending_dir_moves(sctx, parent_ino);
3381 INIT_LIST_HEAD(&stack);
3382 tail_append_pending_moves(sctx, pm, &stack);
3384 while (!list_empty(&stack)) {
3385 pm = list_first_entry(&stack, struct pending_dir_move, list);
3386 parent_ino = pm->ino;
3387 ret = apply_dir_move(sctx, pm);
3388 free_pending_move(sctx, pm);
3391 pm = get_pending_dir_moves(sctx, parent_ino);
3393 tail_append_pending_moves(sctx, pm, &stack);
3398 while (!list_empty(&stack)) {
3399 pm = list_first_entry(&stack, struct pending_dir_move, list);
3400 free_pending_move(sctx, pm);
3406 * We might need to delay a directory rename even when no ancestor directory
3407 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3408 * renamed. This happens when we rename a directory to the old name (the name
3409 * in the parent root) of some other unrelated directory that got its rename
3410 * delayed due to some ancestor with higher number that got renamed.
3416 * |---- a/ (ino 257)
3417 * | |---- file (ino 260)
3419 * |---- b/ (ino 258)
3420 * |---- c/ (ino 259)
3424 * |---- a/ (ino 258)
3425 * |---- x/ (ino 259)
3426 * |---- y/ (ino 257)
3427 * |----- file (ino 260)
3429 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3430 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3431 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3434 * 1 - rename 259 from 'c' to 'x'
3435 * 2 - rename 257 from 'a' to 'x/y'
3436 * 3 - rename 258 from 'b' to 'a'
3438 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3439 * be done right away and < 0 on error.
3441 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3442 struct recorded_ref *parent_ref,
3443 const bool is_orphan)
3445 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3446 struct btrfs_path *path;
3447 struct btrfs_key key;
3448 struct btrfs_key di_key;
3449 struct btrfs_dir_item *di;
3453 struct waiting_dir_move *wdm;
3455 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3458 path = alloc_path_for_send();
3462 key.objectid = parent_ref->dir;
3463 key.type = BTRFS_DIR_ITEM_KEY;
3464 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3466 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3469 } else if (ret > 0) {
3474 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3475 parent_ref->name_len);
3481 * di_key.objectid has the number of the inode that has a dentry in the
3482 * parent directory with the same name that sctx->cur_ino is being
3483 * renamed to. We need to check if that inode is in the send root as
3484 * well and if it is currently marked as an inode with a pending rename,
3485 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3486 * that it happens after that other inode is renamed.
3488 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3489 if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3494 ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3495 &left_gen, NULL, NULL, NULL, NULL);
3498 ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3499 &right_gen, NULL, NULL, NULL, NULL);
3506 /* Different inode, no need to delay the rename of sctx->cur_ino */
3507 if (right_gen != left_gen) {
3512 wdm = get_waiting_dir_move(sctx, di_key.objectid);
3513 if (wdm && !wdm->orphanized) {
3514 ret = add_pending_dir_move(sctx,
3516 sctx->cur_inode_gen,
3519 &sctx->deleted_refs,
3525 btrfs_free_path(path);
3530 * Check if inode ino2, or any of its ancestors, is inode ino1.
3531 * Return 1 if true, 0 if false and < 0 on error.
3533 static int check_ino_in_path(struct btrfs_root *root,
3538 struct fs_path *fs_path)
3543 return ino1_gen == ino2_gen;
3545 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3550 fs_path_reset(fs_path);
3551 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3555 return parent_gen == ino1_gen;
3562 * Check if ino ino1 is an ancestor of inode ino2 in the given root for any
3563 * possible path (in case ino2 is not a directory and has multiple hard links).
3564 * Return 1 if true, 0 if false and < 0 on error.
3566 static int is_ancestor(struct btrfs_root *root,
3570 struct fs_path *fs_path)
3572 bool free_fs_path = false;
3574 struct btrfs_path *path = NULL;
3575 struct btrfs_key key;
3578 fs_path = fs_path_alloc();
3581 free_fs_path = true;
3584 path = alloc_path_for_send();
3590 key.objectid = ino2;
3591 key.type = BTRFS_INODE_REF_KEY;
3594 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3599 struct extent_buffer *leaf = path->nodes[0];
3600 int slot = path->slots[0];
3604 if (slot >= btrfs_header_nritems(leaf)) {
3605 ret = btrfs_next_leaf(root, path);
3613 btrfs_item_key_to_cpu(leaf, &key, slot);
3614 if (key.objectid != ino2)
3616 if (key.type != BTRFS_INODE_REF_KEY &&
3617 key.type != BTRFS_INODE_EXTREF_KEY)
3620 item_size = btrfs_item_size_nr(leaf, slot);
3621 while (cur_offset < item_size) {
3625 if (key.type == BTRFS_INODE_EXTREF_KEY) {
3627 struct btrfs_inode_extref *extref;
3629 ptr = btrfs_item_ptr_offset(leaf, slot);
3630 extref = (struct btrfs_inode_extref *)
3632 parent = btrfs_inode_extref_parent(leaf,
3634 cur_offset += sizeof(*extref);
3635 cur_offset += btrfs_inode_extref_name_len(leaf,
3638 parent = key.offset;
3639 cur_offset = item_size;
3642 ret = get_inode_info(root, parent, NULL, &parent_gen,
3643 NULL, NULL, NULL, NULL);
3646 ret = check_ino_in_path(root, ino1, ino1_gen,
3647 parent, parent_gen, fs_path);
3655 btrfs_free_path(path);
3657 fs_path_free(fs_path);
3661 static int wait_for_parent_move(struct send_ctx *sctx,
3662 struct recorded_ref *parent_ref,
3663 const bool is_orphan)
3666 u64 ino = parent_ref->dir;
3667 u64 ino_gen = parent_ref->dir_gen;
3668 u64 parent_ino_before, parent_ino_after;
3669 struct fs_path *path_before = NULL;
3670 struct fs_path *path_after = NULL;
3673 path_after = fs_path_alloc();
3674 path_before = fs_path_alloc();
3675 if (!path_after || !path_before) {
3681 * Our current directory inode may not yet be renamed/moved because some
3682 * ancestor (immediate or not) has to be renamed/moved first. So find if
3683 * such ancestor exists and make sure our own rename/move happens after
3684 * that ancestor is processed to avoid path build infinite loops (done
3685 * at get_cur_path()).
3687 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3688 u64 parent_ino_after_gen;
3690 if (is_waiting_for_move(sctx, ino)) {
3692 * If the current inode is an ancestor of ino in the
3693 * parent root, we need to delay the rename of the
3694 * current inode, otherwise don't delayed the rename
3695 * because we can end up with a circular dependency
3696 * of renames, resulting in some directories never
3697 * getting the respective rename operations issued in
3698 * the send stream or getting into infinite path build
3701 ret = is_ancestor(sctx->parent_root,
3702 sctx->cur_ino, sctx->cur_inode_gen,
3708 fs_path_reset(path_before);
3709 fs_path_reset(path_after);
3711 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3712 &parent_ino_after_gen, path_after);
3715 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3717 if (ret < 0 && ret != -ENOENT) {
3719 } else if (ret == -ENOENT) {
3724 len1 = fs_path_len(path_before);
3725 len2 = fs_path_len(path_after);
3726 if (ino > sctx->cur_ino &&
3727 (parent_ino_before != parent_ino_after || len1 != len2 ||
3728 memcmp(path_before->start, path_after->start, len1))) {
3731 ret = get_inode_info(sctx->parent_root, ino, NULL,
3732 &parent_ino_gen, NULL, NULL, NULL,
3736 if (ino_gen == parent_ino_gen) {
3741 ino = parent_ino_after;
3742 ino_gen = parent_ino_after_gen;
3746 fs_path_free(path_before);
3747 fs_path_free(path_after);
3750 ret = add_pending_dir_move(sctx,
3752 sctx->cur_inode_gen,
3755 &sctx->deleted_refs,
3764 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3767 struct fs_path *new_path;
3770 * Our reference's name member points to its full_path member string, so
3771 * we use here a new path.
3773 new_path = fs_path_alloc();
3777 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
3779 fs_path_free(new_path);
3782 ret = fs_path_add(new_path, ref->name, ref->name_len);
3784 fs_path_free(new_path);
3788 fs_path_free(ref->full_path);
3789 set_ref_path(ref, new_path);
3795 * When processing the new references for an inode we may orphanize an existing
3796 * directory inode because its old name conflicts with one of the new references
3797 * of the current inode. Later, when processing another new reference of our
3798 * inode, we might need to orphanize another inode, but the path we have in the
3799 * reference reflects the pre-orphanization name of the directory we previously
3800 * orphanized. For example:
3802 * parent snapshot looks like:
3805 * |----- f1 (ino 257)
3806 * |----- f2 (ino 258)
3807 * |----- d1/ (ino 259)
3808 * |----- d2/ (ino 260)
3810 * send snapshot looks like:
3813 * |----- d1 (ino 258)
3814 * |----- f2/ (ino 259)
3815 * |----- f2_link/ (ino 260)
3816 * | |----- f1 (ino 257)
3818 * |----- d2 (ino 258)
3820 * When processing inode 257 we compute the name for inode 259 as "d1", and we
3821 * cache it in the name cache. Later when we start processing inode 258, when
3822 * collecting all its new references we set a full path of "d1/d2" for its new
3823 * reference with name "d2". When we start processing the new references we
3824 * start by processing the new reference with name "d1", and this results in
3825 * orphanizing inode 259, since its old reference causes a conflict. Then we
3826 * move on the next new reference, with name "d2", and we find out we must
3827 * orphanize inode 260, as its old reference conflicts with ours - but for the
3828 * orphanization we use a source path corresponding to the path we stored in the
3829 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
3830 * receiver fail since the path component "d1/" no longer exists, it was renamed
3831 * to "o259-6-0/" when processing the previous new reference. So in this case we
3832 * must recompute the path in the new reference and use it for the new
3833 * orphanization operation.
3835 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3840 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
3844 fs_path_reset(ref->full_path);
3845 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
3849 ret = fs_path_add(ref->full_path, name, ref->name_len);
3853 /* Update the reference's base name pointer. */
3854 set_ref_path(ref, ref->full_path);
3861 * This does all the move/link/unlink/rmdir magic.
3863 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3865 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3867 struct recorded_ref *cur;
3868 struct recorded_ref *cur2;
3869 struct list_head check_dirs;
3870 struct fs_path *valid_path = NULL;
3874 int did_overwrite = 0;
3876 u64 last_dir_ino_rm = 0;
3877 bool can_rename = true;
3878 bool orphanized_dir = false;
3879 bool orphanized_ancestor = false;
3881 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3884 * This should never happen as the root dir always has the same ref
3885 * which is always '..'
3887 BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3888 INIT_LIST_HEAD(&check_dirs);
3890 valid_path = fs_path_alloc();
3897 * First, check if the first ref of the current inode was overwritten
3898 * before. If yes, we know that the current inode was already orphanized
3899 * and thus use the orphan name. If not, we can use get_cur_path to
3900 * get the path of the first ref as it would like while receiving at
3901 * this point in time.
3902 * New inodes are always orphan at the beginning, so force to use the
3903 * orphan name in this case.
3904 * The first ref is stored in valid_path and will be updated if it
3905 * gets moved around.
3907 if (!sctx->cur_inode_new) {
3908 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3909 sctx->cur_inode_gen);
3915 if (sctx->cur_inode_new || did_overwrite) {
3916 ret = gen_unique_name(sctx, sctx->cur_ino,
3917 sctx->cur_inode_gen, valid_path);
3922 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3929 * Before doing any rename and link operations, do a first pass on the
3930 * new references to orphanize any unprocessed inodes that may have a
3931 * reference that conflicts with one of the new references of the current
3932 * inode. This needs to happen first because a new reference may conflict
3933 * with the old reference of a parent directory, so we must make sure
3934 * that the path used for link and rename commands don't use an
3935 * orphanized name when an ancestor was not yet orphanized.
3942 * |----- testdir/ (ino 259)
3943 * | |----- a (ino 257)
3945 * |----- b (ino 258)
3950 * |----- testdir_2/ (ino 259)
3951 * | |----- a (ino 260)
3953 * |----- testdir (ino 257)
3954 * |----- b (ino 257)
3955 * |----- b2 (ino 258)
3957 * Processing the new reference for inode 257 with name "b" may happen
3958 * before processing the new reference with name "testdir". If so, we
3959 * must make sure that by the time we send a link command to create the
3960 * hard link "b", inode 259 was already orphanized, since the generated
3961 * path in "valid_path" already contains the orphanized name for 259.
3962 * We are processing inode 257, so only later when processing 259 we do
3963 * the rename operation to change its temporary (orphanized) name to
3966 list_for_each_entry(cur, &sctx->new_refs, list) {
3967 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3970 if (ret == inode_state_will_create)
3974 * Check if this new ref would overwrite the first ref of another
3975 * unprocessed inode. If yes, orphanize the overwritten inode.
3976 * If we find an overwritten ref that is not the first ref,
3979 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3980 cur->name, cur->name_len,
3981 &ow_inode, &ow_gen, &ow_mode);
3985 ret = is_first_ref(sctx->parent_root,
3986 ow_inode, cur->dir, cur->name,
3991 struct name_cache_entry *nce;
3992 struct waiting_dir_move *wdm;
3994 if (orphanized_dir) {
3995 ret = refresh_ref_path(sctx, cur);
4000 ret = orphanize_inode(sctx, ow_inode, ow_gen,
4004 if (S_ISDIR(ow_mode))
4005 orphanized_dir = true;
4008 * If ow_inode has its rename operation delayed
4009 * make sure that its orphanized name is used in
4010 * the source path when performing its rename
4013 if (is_waiting_for_move(sctx, ow_inode)) {
4014 wdm = get_waiting_dir_move(sctx,
4017 wdm->orphanized = true;
4021 * Make sure we clear our orphanized inode's
4022 * name from the name cache. This is because the
4023 * inode ow_inode might be an ancestor of some
4024 * other inode that will be orphanized as well
4025 * later and has an inode number greater than
4026 * sctx->send_progress. We need to prevent
4027 * future name lookups from using the old name
4028 * and get instead the orphan name.
4030 nce = name_cache_search(sctx, ow_inode, ow_gen);
4032 name_cache_delete(sctx, nce);
4037 * ow_inode might currently be an ancestor of
4038 * cur_ino, therefore compute valid_path (the
4039 * current path of cur_ino) again because it
4040 * might contain the pre-orphanization name of
4041 * ow_inode, which is no longer valid.
4043 ret = is_ancestor(sctx->parent_root,
4045 sctx->cur_ino, NULL);
4047 orphanized_ancestor = true;
4048 fs_path_reset(valid_path);
4049 ret = get_cur_path(sctx, sctx->cur_ino,
4050 sctx->cur_inode_gen,
4057 * If we previously orphanized a directory that
4058 * collided with a new reference that we already
4059 * processed, recompute the current path because
4060 * that directory may be part of the path.
4062 if (orphanized_dir) {
4063 ret = refresh_ref_path(sctx, cur);
4067 ret = send_unlink(sctx, cur->full_path);
4075 list_for_each_entry(cur, &sctx->new_refs, list) {
4077 * We may have refs where the parent directory does not exist
4078 * yet. This happens if the parent directories inum is higher
4079 * than the current inum. To handle this case, we create the
4080 * parent directory out of order. But we need to check if this
4081 * did already happen before due to other refs in the same dir.
4083 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4086 if (ret == inode_state_will_create) {
4089 * First check if any of the current inodes refs did
4090 * already create the dir.
4092 list_for_each_entry(cur2, &sctx->new_refs, list) {
4095 if (cur2->dir == cur->dir) {
4102 * If that did not happen, check if a previous inode
4103 * did already create the dir.
4106 ret = did_create_dir(sctx, cur->dir);
4110 ret = send_create_inode(sctx, cur->dir);
4116 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4117 ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4126 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4128 ret = wait_for_parent_move(sctx, cur, is_orphan);
4138 * link/move the ref to the new place. If we have an orphan
4139 * inode, move it and update valid_path. If not, link or move
4140 * it depending on the inode mode.
4142 if (is_orphan && can_rename) {
4143 ret = send_rename(sctx, valid_path, cur->full_path);
4147 ret = fs_path_copy(valid_path, cur->full_path);
4150 } else if (can_rename) {
4151 if (S_ISDIR(sctx->cur_inode_mode)) {
4153 * Dirs can't be linked, so move it. For moved
4154 * dirs, we always have one new and one deleted
4155 * ref. The deleted ref is ignored later.
4157 ret = send_rename(sctx, valid_path,
4160 ret = fs_path_copy(valid_path,
4166 * We might have previously orphanized an inode
4167 * which is an ancestor of our current inode,
4168 * so our reference's full path, which was
4169 * computed before any such orphanizations, must
4172 if (orphanized_dir) {
4173 ret = update_ref_path(sctx, cur);
4177 ret = send_link(sctx, cur->full_path,
4183 ret = dup_ref(cur, &check_dirs);
4188 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4190 * Check if we can already rmdir the directory. If not,
4191 * orphanize it. For every dir item inside that gets deleted
4192 * later, we do this check again and rmdir it then if possible.
4193 * See the use of check_dirs for more details.
4195 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4200 ret = send_rmdir(sctx, valid_path);
4203 } else if (!is_orphan) {
4204 ret = orphanize_inode(sctx, sctx->cur_ino,
4205 sctx->cur_inode_gen, valid_path);
4211 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4212 ret = dup_ref(cur, &check_dirs);
4216 } else if (S_ISDIR(sctx->cur_inode_mode) &&
4217 !list_empty(&sctx->deleted_refs)) {
4219 * We have a moved dir. Add the old parent to check_dirs
4221 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4223 ret = dup_ref(cur, &check_dirs);
4226 } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4228 * We have a non dir inode. Go through all deleted refs and
4229 * unlink them if they were not already overwritten by other
4232 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4233 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4234 sctx->cur_ino, sctx->cur_inode_gen,
4235 cur->name, cur->name_len);
4240 * If we orphanized any ancestor before, we need
4241 * to recompute the full path for deleted names,
4242 * since any such path was computed before we
4243 * processed any references and orphanized any
4246 if (orphanized_ancestor) {
4247 ret = update_ref_path(sctx, cur);
4251 ret = send_unlink(sctx, cur->full_path);
4255 ret = dup_ref(cur, &check_dirs);
4260 * If the inode is still orphan, unlink the orphan. This may
4261 * happen when a previous inode did overwrite the first ref
4262 * of this inode and no new refs were added for the current
4263 * inode. Unlinking does not mean that the inode is deleted in
4264 * all cases. There may still be links to this inode in other
4268 ret = send_unlink(sctx, valid_path);
4275 * We did collect all parent dirs where cur_inode was once located. We
4276 * now go through all these dirs and check if they are pending for
4277 * deletion and if it's finally possible to perform the rmdir now.
4278 * We also update the inode stats of the parent dirs here.
4280 list_for_each_entry(cur, &check_dirs, list) {
4282 * In case we had refs into dirs that were not processed yet,
4283 * we don't need to do the utime and rmdir logic for these dirs.
4284 * The dir will be processed later.
4286 if (cur->dir > sctx->cur_ino)
4289 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4293 if (ret == inode_state_did_create ||
4294 ret == inode_state_no_change) {
4295 /* TODO delayed utimes */
4296 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4299 } else if (ret == inode_state_did_delete &&
4300 cur->dir != last_dir_ino_rm) {
4301 ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4306 ret = get_cur_path(sctx, cur->dir,
4307 cur->dir_gen, valid_path);
4310 ret = send_rmdir(sctx, valid_path);
4313 last_dir_ino_rm = cur->dir;
4321 __free_recorded_refs(&check_dirs);
4322 free_recorded_refs(sctx);
4323 fs_path_free(valid_path);
4327 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name,
4328 void *ctx, struct list_head *refs)
4331 struct send_ctx *sctx = ctx;
4335 p = fs_path_alloc();
4339 ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4344 ret = get_cur_path(sctx, dir, gen, p);
4347 ret = fs_path_add_path(p, name);
4351 ret = __record_ref(refs, dir, gen, p);
4359 static int __record_new_ref(int num, u64 dir, int index,
4360 struct fs_path *name,
4363 struct send_ctx *sctx = ctx;
4364 return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs);
4368 static int __record_deleted_ref(int num, u64 dir, int index,
4369 struct fs_path *name,
4372 struct send_ctx *sctx = ctx;
4373 return record_ref(sctx->parent_root, dir, name, ctx,
4374 &sctx->deleted_refs);
4377 static int record_new_ref(struct send_ctx *sctx)
4381 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4382 sctx->cmp_key, 0, __record_new_ref, sctx);
4391 static int record_deleted_ref(struct send_ctx *sctx)
4395 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4396 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4405 struct find_ref_ctx {
4408 struct btrfs_root *root;
4409 struct fs_path *name;
4413 static int __find_iref(int num, u64 dir, int index,
4414 struct fs_path *name,
4417 struct find_ref_ctx *ctx = ctx_;
4421 if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4422 strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4424 * To avoid doing extra lookups we'll only do this if everything
4427 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4431 if (dir_gen != ctx->dir_gen)
4433 ctx->found_idx = num;
4439 static int find_iref(struct btrfs_root *root,
4440 struct btrfs_path *path,
4441 struct btrfs_key *key,
4442 u64 dir, u64 dir_gen, struct fs_path *name)
4445 struct find_ref_ctx ctx;
4449 ctx.dir_gen = dir_gen;
4453 ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4457 if (ctx.found_idx == -1)
4460 return ctx.found_idx;
4463 static int __record_changed_new_ref(int num, u64 dir, int index,
4464 struct fs_path *name,
4469 struct send_ctx *sctx = ctx;
4471 ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4476 ret = find_iref(sctx->parent_root, sctx->right_path,
4477 sctx->cmp_key, dir, dir_gen, name);
4479 ret = __record_new_ref(num, dir, index, name, sctx);
4486 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4487 struct fs_path *name,
4492 struct send_ctx *sctx = ctx;
4494 ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4499 ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4500 dir, dir_gen, name);
4502 ret = __record_deleted_ref(num, dir, index, name, sctx);
4509 static int record_changed_ref(struct send_ctx *sctx)
4513 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4514 sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4517 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4518 sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4528 * Record and process all refs at once. Needed when an inode changes the
4529 * generation number, which means that it was deleted and recreated.
4531 static int process_all_refs(struct send_ctx *sctx,
4532 enum btrfs_compare_tree_result cmd)
4535 struct btrfs_root *root;
4536 struct btrfs_path *path;
4537 struct btrfs_key key;
4538 struct btrfs_key found_key;
4539 struct extent_buffer *eb;
4541 iterate_inode_ref_t cb;
4542 int pending_move = 0;
4544 path = alloc_path_for_send();
4548 if (cmd == BTRFS_COMPARE_TREE_NEW) {
4549 root = sctx->send_root;
4550 cb = __record_new_ref;
4551 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4552 root = sctx->parent_root;
4553 cb = __record_deleted_ref;
4555 btrfs_err(sctx->send_root->fs_info,
4556 "Wrong command %d in process_all_refs", cmd);
4561 key.objectid = sctx->cmp_key->objectid;
4562 key.type = BTRFS_INODE_REF_KEY;
4564 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4569 eb = path->nodes[0];
4570 slot = path->slots[0];
4571 if (slot >= btrfs_header_nritems(eb)) {
4572 ret = btrfs_next_leaf(root, path);
4580 btrfs_item_key_to_cpu(eb, &found_key, slot);
4582 if (found_key.objectid != key.objectid ||
4583 (found_key.type != BTRFS_INODE_REF_KEY &&
4584 found_key.type != BTRFS_INODE_EXTREF_KEY))
4587 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4593 btrfs_release_path(path);
4596 * We don't actually care about pending_move as we are simply
4597 * re-creating this inode and will be rename'ing it into place once we
4598 * rename the parent directory.
4600 ret = process_recorded_refs(sctx, &pending_move);
4602 btrfs_free_path(path);
4606 static int send_set_xattr(struct send_ctx *sctx,
4607 struct fs_path *path,
4608 const char *name, int name_len,
4609 const char *data, int data_len)
4613 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4617 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4618 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4619 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4621 ret = send_cmd(sctx);
4628 static int send_remove_xattr(struct send_ctx *sctx,
4629 struct fs_path *path,
4630 const char *name, int name_len)
4634 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4638 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4639 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4641 ret = send_cmd(sctx);
4648 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4649 const char *name, int name_len,
4650 const char *data, int data_len,
4654 struct send_ctx *sctx = ctx;
4656 struct posix_acl_xattr_header dummy_acl;
4658 /* Capabilities are emitted by finish_inode_if_needed */
4659 if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4662 p = fs_path_alloc();
4667 * This hack is needed because empty acls are stored as zero byte
4668 * data in xattrs. Problem with that is, that receiving these zero byte
4669 * acls will fail later. To fix this, we send a dummy acl list that
4670 * only contains the version number and no entries.
4672 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4673 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4674 if (data_len == 0) {
4675 dummy_acl.a_version =
4676 cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4677 data = (char *)&dummy_acl;
4678 data_len = sizeof(dummy_acl);
4682 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4686 ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4693 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4694 const char *name, int name_len,
4695 const char *data, int data_len,
4699 struct send_ctx *sctx = ctx;
4702 p = fs_path_alloc();
4706 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4710 ret = send_remove_xattr(sctx, p, name, name_len);
4717 static int process_new_xattr(struct send_ctx *sctx)
4721 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4722 __process_new_xattr, sctx);
4727 static int process_deleted_xattr(struct send_ctx *sctx)
4729 return iterate_dir_item(sctx->parent_root, sctx->right_path,
4730 __process_deleted_xattr, sctx);
4733 struct find_xattr_ctx {
4741 static int __find_xattr(int num, struct btrfs_key *di_key,
4742 const char *name, int name_len,
4743 const char *data, int data_len,
4744 u8 type, void *vctx)
4746 struct find_xattr_ctx *ctx = vctx;
4748 if (name_len == ctx->name_len &&
4749 strncmp(name, ctx->name, name_len) == 0) {
4750 ctx->found_idx = num;
4751 ctx->found_data_len = data_len;
4752 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4753 if (!ctx->found_data)
4760 static int find_xattr(struct btrfs_root *root,
4761 struct btrfs_path *path,
4762 struct btrfs_key *key,
4763 const char *name, int name_len,
4764 char **data, int *data_len)
4767 struct find_xattr_ctx ctx;
4770 ctx.name_len = name_len;
4772 ctx.found_data = NULL;
4773 ctx.found_data_len = 0;
4775 ret = iterate_dir_item(root, path, __find_xattr, &ctx);
4779 if (ctx.found_idx == -1)
4782 *data = ctx.found_data;
4783 *data_len = ctx.found_data_len;
4785 kfree(ctx.found_data);
4787 return ctx.found_idx;
4791 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4792 const char *name, int name_len,
4793 const char *data, int data_len,
4797 struct send_ctx *sctx = ctx;
4798 char *found_data = NULL;
4799 int found_data_len = 0;
4801 ret = find_xattr(sctx->parent_root, sctx->right_path,
4802 sctx->cmp_key, name, name_len, &found_data,
4804 if (ret == -ENOENT) {
4805 ret = __process_new_xattr(num, di_key, name, name_len, data,
4806 data_len, type, ctx);
4807 } else if (ret >= 0) {
4808 if (data_len != found_data_len ||
4809 memcmp(data, found_data, data_len)) {
4810 ret = __process_new_xattr(num, di_key, name, name_len,
4811 data, data_len, type, ctx);
4821 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4822 const char *name, int name_len,
4823 const char *data, int data_len,
4827 struct send_ctx *sctx = ctx;
4829 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4830 name, name_len, NULL, NULL);
4832 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4833 data_len, type, ctx);
4840 static int process_changed_xattr(struct send_ctx *sctx)
4844 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4845 __process_changed_new_xattr, sctx);
4848 ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4849 __process_changed_deleted_xattr, sctx);
4855 static int process_all_new_xattrs(struct send_ctx *sctx)
4858 struct btrfs_root *root;
4859 struct btrfs_path *path;
4860 struct btrfs_key key;
4861 struct btrfs_key found_key;
4862 struct extent_buffer *eb;
4865 path = alloc_path_for_send();
4869 root = sctx->send_root;
4871 key.objectid = sctx->cmp_key->objectid;
4872 key.type = BTRFS_XATTR_ITEM_KEY;
4874 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4879 eb = path->nodes[0];
4880 slot = path->slots[0];
4881 if (slot >= btrfs_header_nritems(eb)) {
4882 ret = btrfs_next_leaf(root, path);
4885 } else if (ret > 0) {
4892 btrfs_item_key_to_cpu(eb, &found_key, slot);
4893 if (found_key.objectid != key.objectid ||
4894 found_key.type != key.type) {
4899 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
4907 btrfs_free_path(path);
4911 static inline u64 max_send_read_size(const struct send_ctx *sctx)
4913 return sctx->send_max_size - SZ_16K;
4916 static int put_data_header(struct send_ctx *sctx, u32 len)
4918 struct btrfs_tlv_header *hdr;
4920 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
4922 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
4923 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
4924 put_unaligned_le16(len, &hdr->tlv_len);
4925 sctx->send_size += sizeof(*hdr);
4929 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
4931 struct btrfs_root *root = sctx->send_root;
4932 struct btrfs_fs_info *fs_info = root->fs_info;
4933 struct inode *inode;
4935 pgoff_t index = offset >> PAGE_SHIFT;
4937 unsigned pg_offset = offset_in_page(offset);
4940 ret = put_data_header(sctx, len);
4944 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
4946 return PTR_ERR(inode);
4948 last_index = (offset + len - 1) >> PAGE_SHIFT;
4950 /* initial readahead */
4951 memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4952 file_ra_state_init(&sctx->ra, inode->i_mapping);
4954 while (index <= last_index) {
4955 unsigned cur_len = min_t(unsigned, len,
4956 PAGE_SIZE - pg_offset);
4958 page = find_lock_page(inode->i_mapping, index);
4960 page_cache_sync_readahead(inode->i_mapping, &sctx->ra,
4961 NULL, index, last_index + 1 - index);
4963 page = find_or_create_page(inode->i_mapping, index,
4971 if (PageReadahead(page)) {
4972 page_cache_async_readahead(inode->i_mapping, &sctx->ra,
4973 NULL, page, index, last_index + 1 - index);
4976 if (!PageUptodate(page)) {
4977 btrfs_readpage(NULL, page);
4979 if (!PageUptodate(page)) {
4987 memcpy_from_page(sctx->send_buf + sctx->send_size, page,
4988 pg_offset, cur_len);
4994 sctx->send_size += cur_len;
5001 * Read some bytes from the current inode/file and send a write command to
5004 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5006 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5010 p = fs_path_alloc();
5014 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5016 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5020 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5024 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5025 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5026 ret = put_file_data(sctx, offset, len);
5030 ret = send_cmd(sctx);
5039 * Send a clone command to user space.
5041 static int send_clone(struct send_ctx *sctx,
5042 u64 offset, u32 len,
5043 struct clone_root *clone_root)
5049 btrfs_debug(sctx->send_root->fs_info,
5050 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5051 offset, len, clone_root->root->root_key.objectid,
5052 clone_root->ino, clone_root->offset);
5054 p = fs_path_alloc();
5058 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5062 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5066 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5067 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5068 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5070 if (clone_root->root == sctx->send_root) {
5071 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
5072 &gen, NULL, NULL, NULL, NULL);
5075 ret = get_cur_path(sctx, clone_root->ino, gen, p);
5077 ret = get_inode_path(clone_root->root, clone_root->ino, p);
5083 * If the parent we're using has a received_uuid set then use that as
5084 * our clone source as that is what we will look for when doing a
5087 * This covers the case that we create a snapshot off of a received
5088 * subvolume and then use that as the parent and try to receive on a
5091 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5092 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5093 clone_root->root->root_item.received_uuid);
5095 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5096 clone_root->root->root_item.uuid);
5097 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5098 btrfs_root_ctransid(&clone_root->root->root_item));
5099 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5100 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5101 clone_root->offset);
5103 ret = send_cmd(sctx);
5112 * Send an update extent command to user space.
5114 static int send_update_extent(struct send_ctx *sctx,
5115 u64 offset, u32 len)
5120 p = fs_path_alloc();
5124 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5128 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5132 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5133 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5134 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5136 ret = send_cmd(sctx);
5144 static int send_hole(struct send_ctx *sctx, u64 end)
5146 struct fs_path *p = NULL;
5147 u64 read_size = max_send_read_size(sctx);
5148 u64 offset = sctx->cur_inode_last_extent;
5152 * A hole that starts at EOF or beyond it. Since we do not yet support
5153 * fallocate (for extent preallocation and hole punching), sending a
5154 * write of zeroes starting at EOF or beyond would later require issuing
5155 * a truncate operation which would undo the write and achieve nothing.
5157 if (offset >= sctx->cur_inode_size)
5161 * Don't go beyond the inode's i_size due to prealloc extents that start
5164 end = min_t(u64, end, sctx->cur_inode_size);
5166 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5167 return send_update_extent(sctx, offset, end - offset);
5169 p = fs_path_alloc();
5172 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5174 goto tlv_put_failure;
5175 while (offset < end) {
5176 u64 len = min(end - offset, read_size);
5178 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5181 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5182 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5183 ret = put_data_header(sctx, len);
5186 memset(sctx->send_buf + sctx->send_size, 0, len);
5187 sctx->send_size += len;
5188 ret = send_cmd(sctx);
5193 sctx->cur_inode_next_write_offset = offset;
5199 static int send_extent_data(struct send_ctx *sctx,
5203 u64 read_size = max_send_read_size(sctx);
5206 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5207 return send_update_extent(sctx, offset, len);
5209 while (sent < len) {
5210 u64 size = min(len - sent, read_size);
5213 ret = send_write(sctx, offset + sent, size);
5222 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5223 * found, call send_set_xattr function to emit it.
5225 * Return 0 if there isn't a capability, or when the capability was emitted
5226 * successfully, or < 0 if an error occurred.
5228 static int send_capabilities(struct send_ctx *sctx)
5230 struct fs_path *fspath = NULL;
5231 struct btrfs_path *path;
5232 struct btrfs_dir_item *di;
5233 struct extent_buffer *leaf;
5234 unsigned long data_ptr;
5239 path = alloc_path_for_send();
5243 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5244 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5246 /* There is no xattr for this inode */
5248 } else if (IS_ERR(di)) {
5253 leaf = path->nodes[0];
5254 buf_len = btrfs_dir_data_len(leaf, di);
5256 fspath = fs_path_alloc();
5257 buf = kmalloc(buf_len, GFP_KERNEL);
5258 if (!fspath || !buf) {
5263 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5267 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5268 read_extent_buffer(leaf, buf, data_ptr, buf_len);
5270 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5271 strlen(XATTR_NAME_CAPS), buf, buf_len);
5274 fs_path_free(fspath);
5275 btrfs_free_path(path);
5279 static int clone_range(struct send_ctx *sctx,
5280 struct clone_root *clone_root,
5281 const u64 disk_byte,
5286 struct btrfs_path *path;
5287 struct btrfs_key key;
5289 u64 clone_src_i_size = 0;
5292 * Prevent cloning from a zero offset with a length matching the sector
5293 * size because in some scenarios this will make the receiver fail.
5295 * For example, if in the source filesystem the extent at offset 0
5296 * has a length of sectorsize and it was written using direct IO, then
5297 * it can never be an inline extent (even if compression is enabled).
5298 * Then this extent can be cloned in the original filesystem to a non
5299 * zero file offset, but it may not be possible to clone in the
5300 * destination filesystem because it can be inlined due to compression
5301 * on the destination filesystem (as the receiver's write operations are
5302 * always done using buffered IO). The same happens when the original
5303 * filesystem does not have compression enabled but the destination
5306 if (clone_root->offset == 0 &&
5307 len == sctx->send_root->fs_info->sectorsize)
5308 return send_extent_data(sctx, offset, len);
5310 path = alloc_path_for_send();
5315 * There are inodes that have extents that lie behind its i_size. Don't
5316 * accept clones from these extents.
5318 ret = __get_inode_info(clone_root->root, path, clone_root->ino,
5319 &clone_src_i_size, NULL, NULL, NULL, NULL, NULL);
5320 btrfs_release_path(path);
5325 * We can't send a clone operation for the entire range if we find
5326 * extent items in the respective range in the source file that
5327 * refer to different extents or if we find holes.
5328 * So check for that and do a mix of clone and regular write/copy
5329 * operations if needed.
5333 * mkfs.btrfs -f /dev/sda
5334 * mount /dev/sda /mnt
5335 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5336 * cp --reflink=always /mnt/foo /mnt/bar
5337 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5338 * btrfs subvolume snapshot -r /mnt /mnt/snap
5340 * If when we send the snapshot and we are processing file bar (which
5341 * has a higher inode number than foo) we blindly send a clone operation
5342 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5343 * a file bar that matches the content of file foo - iow, doesn't match
5344 * the content from bar in the original filesystem.
5346 key.objectid = clone_root->ino;
5347 key.type = BTRFS_EXTENT_DATA_KEY;
5348 key.offset = clone_root->offset;
5349 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5352 if (ret > 0 && path->slots[0] > 0) {
5353 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5354 if (key.objectid == clone_root->ino &&
5355 key.type == BTRFS_EXTENT_DATA_KEY)
5360 struct extent_buffer *leaf = path->nodes[0];
5361 int slot = path->slots[0];
5362 struct btrfs_file_extent_item *ei;
5366 u64 clone_data_offset;
5368 if (slot >= btrfs_header_nritems(leaf)) {
5369 ret = btrfs_next_leaf(clone_root->root, path);
5377 btrfs_item_key_to_cpu(leaf, &key, slot);
5380 * We might have an implicit trailing hole (NO_HOLES feature
5381 * enabled). We deal with it after leaving this loop.
5383 if (key.objectid != clone_root->ino ||
5384 key.type != BTRFS_EXTENT_DATA_KEY)
5387 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5388 type = btrfs_file_extent_type(leaf, ei);
5389 if (type == BTRFS_FILE_EXTENT_INLINE) {
5390 ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5391 ext_len = PAGE_ALIGN(ext_len);
5393 ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5396 if (key.offset + ext_len <= clone_root->offset)
5399 if (key.offset > clone_root->offset) {
5400 /* Implicit hole, NO_HOLES feature enabled. */
5401 u64 hole_len = key.offset - clone_root->offset;
5405 ret = send_extent_data(sctx, offset, hole_len);
5413 clone_root->offset += hole_len;
5414 data_offset += hole_len;
5417 if (key.offset >= clone_root->offset + len)
5420 if (key.offset >= clone_src_i_size)
5423 if (key.offset + ext_len > clone_src_i_size)
5424 ext_len = clone_src_i_size - key.offset;
5426 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
5427 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
5428 clone_root->offset = key.offset;
5429 if (clone_data_offset < data_offset &&
5430 clone_data_offset + ext_len > data_offset) {
5433 extent_offset = data_offset - clone_data_offset;
5434 ext_len -= extent_offset;
5435 clone_data_offset += extent_offset;
5436 clone_root->offset += extent_offset;
5440 clone_len = min_t(u64, ext_len, len);
5442 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5443 clone_data_offset == data_offset) {
5444 const u64 src_end = clone_root->offset + clone_len;
5445 const u64 sectorsize = SZ_64K;
5448 * We can't clone the last block, when its size is not
5449 * sector size aligned, into the middle of a file. If we
5450 * do so, the receiver will get a failure (-EINVAL) when
5451 * trying to clone or will silently corrupt the data in
5452 * the destination file if it's on a kernel without the
5453 * fix introduced by commit ac765f83f1397646
5454 * ("Btrfs: fix data corruption due to cloning of eof
5457 * So issue a clone of the aligned down range plus a
5458 * regular write for the eof block, if we hit that case.
5460 * Also, we use the maximum possible sector size, 64K,
5461 * because we don't know what's the sector size of the
5462 * filesystem that receives the stream, so we have to
5463 * assume the largest possible sector size.
5465 if (src_end == clone_src_i_size &&
5466 !IS_ALIGNED(src_end, sectorsize) &&
5467 offset + clone_len < sctx->cur_inode_size) {
5470 slen = ALIGN_DOWN(src_end - clone_root->offset,
5473 ret = send_clone(sctx, offset, slen,
5478 ret = send_extent_data(sctx, offset + slen,
5481 ret = send_clone(sctx, offset, clone_len,
5485 ret = send_extent_data(sctx, offset, clone_len);
5494 offset += clone_len;
5495 clone_root->offset += clone_len;
5498 * If we are cloning from the file we are currently processing,
5499 * and using the send root as the clone root, we must stop once
5500 * the current clone offset reaches the current eof of the file
5501 * at the receiver, otherwise we would issue an invalid clone
5502 * operation (source range going beyond eof) and cause the
5503 * receiver to fail. So if we reach the current eof, bail out
5504 * and fallback to a regular write.
5506 if (clone_root->root == sctx->send_root &&
5507 clone_root->ino == sctx->cur_ino &&
5508 clone_root->offset >= sctx->cur_inode_next_write_offset)
5511 data_offset += clone_len;
5517 ret = send_extent_data(sctx, offset, len);
5521 btrfs_free_path(path);
5525 static int send_write_or_clone(struct send_ctx *sctx,
5526 struct btrfs_path *path,
5527 struct btrfs_key *key,
5528 struct clone_root *clone_root)
5531 u64 offset = key->offset;
5533 u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5535 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
5539 if (clone_root && IS_ALIGNED(end, bs)) {
5540 struct btrfs_file_extent_item *ei;
5544 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5545 struct btrfs_file_extent_item);
5546 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5547 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5548 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5549 offset, end - offset);
5551 ret = send_extent_data(sctx, offset, end - offset);
5553 sctx->cur_inode_next_write_offset = end;
5557 static int is_extent_unchanged(struct send_ctx *sctx,
5558 struct btrfs_path *left_path,
5559 struct btrfs_key *ekey)
5562 struct btrfs_key key;
5563 struct btrfs_path *path = NULL;
5564 struct extent_buffer *eb;
5566 struct btrfs_key found_key;
5567 struct btrfs_file_extent_item *ei;
5572 u64 left_offset_fixed;
5580 path = alloc_path_for_send();
5584 eb = left_path->nodes[0];
5585 slot = left_path->slots[0];
5586 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5587 left_type = btrfs_file_extent_type(eb, ei);
5589 if (left_type != BTRFS_FILE_EXTENT_REG) {
5593 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5594 left_len = btrfs_file_extent_num_bytes(eb, ei);
5595 left_offset = btrfs_file_extent_offset(eb, ei);
5596 left_gen = btrfs_file_extent_generation(eb, ei);
5599 * Following comments will refer to these graphics. L is the left
5600 * extents which we are checking at the moment. 1-8 are the right
5601 * extents that we iterate.
5604 * |-1-|-2a-|-3-|-4-|-5-|-6-|
5607 * |--1--|-2b-|...(same as above)
5609 * Alternative situation. Happens on files where extents got split.
5611 * |-----------7-----------|-6-|
5613 * Alternative situation. Happens on files which got larger.
5616 * Nothing follows after 8.
5619 key.objectid = ekey->objectid;
5620 key.type = BTRFS_EXTENT_DATA_KEY;
5621 key.offset = ekey->offset;
5622 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5631 * Handle special case where the right side has no extents at all.
5633 eb = path->nodes[0];
5634 slot = path->slots[0];
5635 btrfs_item_key_to_cpu(eb, &found_key, slot);
5636 if (found_key.objectid != key.objectid ||
5637 found_key.type != key.type) {
5638 /* If we're a hole then just pretend nothing changed */
5639 ret = (left_disknr) ? 0 : 1;
5644 * We're now on 2a, 2b or 7.
5647 while (key.offset < ekey->offset + left_len) {
5648 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5649 right_type = btrfs_file_extent_type(eb, ei);
5650 if (right_type != BTRFS_FILE_EXTENT_REG &&
5651 right_type != BTRFS_FILE_EXTENT_INLINE) {
5656 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5657 right_len = btrfs_file_extent_ram_bytes(eb, ei);
5658 right_len = PAGE_ALIGN(right_len);
5660 right_len = btrfs_file_extent_num_bytes(eb, ei);
5664 * Are we at extent 8? If yes, we know the extent is changed.
5665 * This may only happen on the first iteration.
5667 if (found_key.offset + right_len <= ekey->offset) {
5668 /* If we're a hole just pretend nothing changed */
5669 ret = (left_disknr) ? 0 : 1;
5674 * We just wanted to see if when we have an inline extent, what
5675 * follows it is a regular extent (wanted to check the above
5676 * condition for inline extents too). This should normally not
5677 * happen but it's possible for example when we have an inline
5678 * compressed extent representing data with a size matching
5679 * the page size (currently the same as sector size).
5681 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5686 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5687 right_offset = btrfs_file_extent_offset(eb, ei);
5688 right_gen = btrfs_file_extent_generation(eb, ei);
5690 left_offset_fixed = left_offset;
5691 if (key.offset < ekey->offset) {
5692 /* Fix the right offset for 2a and 7. */
5693 right_offset += ekey->offset - key.offset;
5695 /* Fix the left offset for all behind 2a and 2b */
5696 left_offset_fixed += key.offset - ekey->offset;
5700 * Check if we have the same extent.
5702 if (left_disknr != right_disknr ||
5703 left_offset_fixed != right_offset ||
5704 left_gen != right_gen) {
5710 * Go to the next extent.
5712 ret = btrfs_next_item(sctx->parent_root, path);
5716 eb = path->nodes[0];
5717 slot = path->slots[0];
5718 btrfs_item_key_to_cpu(eb, &found_key, slot);
5720 if (ret || found_key.objectid != key.objectid ||
5721 found_key.type != key.type) {
5722 key.offset += right_len;
5725 if (found_key.offset != key.offset + right_len) {
5733 * We're now behind the left extent (treat as unchanged) or at the end
5734 * of the right side (treat as changed).
5736 if (key.offset >= ekey->offset + left_len)
5743 btrfs_free_path(path);
5747 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5749 struct btrfs_path *path;
5750 struct btrfs_root *root = sctx->send_root;
5751 struct btrfs_key key;
5754 path = alloc_path_for_send();
5758 sctx->cur_inode_last_extent = 0;
5760 key.objectid = sctx->cur_ino;
5761 key.type = BTRFS_EXTENT_DATA_KEY;
5762 key.offset = offset;
5763 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5767 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5768 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5771 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5773 btrfs_free_path(path);
5777 static int range_is_hole_in_parent(struct send_ctx *sctx,
5781 struct btrfs_path *path;
5782 struct btrfs_key key;
5783 struct btrfs_root *root = sctx->parent_root;
5784 u64 search_start = start;
5787 path = alloc_path_for_send();
5791 key.objectid = sctx->cur_ino;
5792 key.type = BTRFS_EXTENT_DATA_KEY;
5793 key.offset = search_start;
5794 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5797 if (ret > 0 && path->slots[0] > 0)
5800 while (search_start < end) {
5801 struct extent_buffer *leaf = path->nodes[0];
5802 int slot = path->slots[0];
5803 struct btrfs_file_extent_item *fi;
5806 if (slot >= btrfs_header_nritems(leaf)) {
5807 ret = btrfs_next_leaf(root, path);
5815 btrfs_item_key_to_cpu(leaf, &key, slot);
5816 if (key.objectid < sctx->cur_ino ||
5817 key.type < BTRFS_EXTENT_DATA_KEY)
5819 if (key.objectid > sctx->cur_ino ||
5820 key.type > BTRFS_EXTENT_DATA_KEY ||
5824 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5825 extent_end = btrfs_file_extent_end(path);
5826 if (extent_end <= start)
5828 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5829 search_start = extent_end;
5839 btrfs_free_path(path);
5843 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5844 struct btrfs_key *key)
5848 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5851 if (sctx->cur_inode_last_extent == (u64)-1) {
5852 ret = get_last_extent(sctx, key->offset - 1);
5857 if (path->slots[0] == 0 &&
5858 sctx->cur_inode_last_extent < key->offset) {
5860 * We might have skipped entire leafs that contained only
5861 * file extent items for our current inode. These leafs have
5862 * a generation number smaller (older) than the one in the
5863 * current leaf and the leaf our last extent came from, and
5864 * are located between these 2 leafs.
5866 ret = get_last_extent(sctx, key->offset - 1);
5871 if (sctx->cur_inode_last_extent < key->offset) {
5872 ret = range_is_hole_in_parent(sctx,
5873 sctx->cur_inode_last_extent,
5878 ret = send_hole(sctx, key->offset);
5882 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5886 static int process_extent(struct send_ctx *sctx,
5887 struct btrfs_path *path,
5888 struct btrfs_key *key)
5890 struct clone_root *found_clone = NULL;
5893 if (S_ISLNK(sctx->cur_inode_mode))
5896 if (sctx->parent_root && !sctx->cur_inode_new) {
5897 ret = is_extent_unchanged(sctx, path, key);
5905 struct btrfs_file_extent_item *ei;
5908 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5909 struct btrfs_file_extent_item);
5910 type = btrfs_file_extent_type(path->nodes[0], ei);
5911 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5912 type == BTRFS_FILE_EXTENT_REG) {
5914 * The send spec does not have a prealloc command yet,
5915 * so just leave a hole for prealloc'ed extents until
5916 * we have enough commands queued up to justify rev'ing
5919 if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5924 /* Have a hole, just skip it. */
5925 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5932 ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5933 sctx->cur_inode_size, &found_clone);
5934 if (ret != -ENOENT && ret < 0)
5937 ret = send_write_or_clone(sctx, path, key, found_clone);
5941 ret = maybe_send_hole(sctx, path, key);
5946 static int process_all_extents(struct send_ctx *sctx)
5949 struct btrfs_root *root;
5950 struct btrfs_path *path;
5951 struct btrfs_key key;
5952 struct btrfs_key found_key;
5953 struct extent_buffer *eb;
5956 root = sctx->send_root;
5957 path = alloc_path_for_send();
5961 key.objectid = sctx->cmp_key->objectid;
5962 key.type = BTRFS_EXTENT_DATA_KEY;
5964 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5969 eb = path->nodes[0];
5970 slot = path->slots[0];
5972 if (slot >= btrfs_header_nritems(eb)) {
5973 ret = btrfs_next_leaf(root, path);
5976 } else if (ret > 0) {
5983 btrfs_item_key_to_cpu(eb, &found_key, slot);
5985 if (found_key.objectid != key.objectid ||
5986 found_key.type != key.type) {
5991 ret = process_extent(sctx, path, &found_key);
5999 btrfs_free_path(path);
6003 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6005 int *refs_processed)
6009 if (sctx->cur_ino == 0)
6011 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6012 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6014 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6017 ret = process_recorded_refs(sctx, pending_move);
6021 *refs_processed = 1;
6026 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6037 int need_truncate = 1;
6038 int pending_move = 0;
6039 int refs_processed = 0;
6041 if (sctx->ignore_cur_inode)
6044 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6050 * We have processed the refs and thus need to advance send_progress.
6051 * Now, calls to get_cur_xxx will take the updated refs of the current
6052 * inode into account.
6054 * On the other hand, if our current inode is a directory and couldn't
6055 * be moved/renamed because its parent was renamed/moved too and it has
6056 * a higher inode number, we can only move/rename our current inode
6057 * after we moved/renamed its parent. Therefore in this case operate on
6058 * the old path (pre move/rename) of our current inode, and the
6059 * move/rename will be performed later.
6061 if (refs_processed && !pending_move)
6062 sctx->send_progress = sctx->cur_ino + 1;
6064 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6066 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6069 ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
6070 &left_mode, &left_uid, &left_gid, NULL);
6074 if (!sctx->parent_root || sctx->cur_inode_new) {
6076 if (!S_ISLNK(sctx->cur_inode_mode))
6078 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6083 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
6084 &old_size, NULL, &right_mode, &right_uid,
6089 if (left_uid != right_uid || left_gid != right_gid)
6091 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6093 if ((old_size == sctx->cur_inode_size) ||
6094 (sctx->cur_inode_size > old_size &&
6095 sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6099 if (S_ISREG(sctx->cur_inode_mode)) {
6100 if (need_send_hole(sctx)) {
6101 if (sctx->cur_inode_last_extent == (u64)-1 ||
6102 sctx->cur_inode_last_extent <
6103 sctx->cur_inode_size) {
6104 ret = get_last_extent(sctx, (u64)-1);
6108 if (sctx->cur_inode_last_extent <
6109 sctx->cur_inode_size) {
6110 ret = send_hole(sctx, sctx->cur_inode_size);
6115 if (need_truncate) {
6116 ret = send_truncate(sctx, sctx->cur_ino,
6117 sctx->cur_inode_gen,
6118 sctx->cur_inode_size);
6125 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6126 left_uid, left_gid);
6131 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6137 ret = send_capabilities(sctx);
6142 * If other directory inodes depended on our current directory
6143 * inode's move/rename, now do their move/rename operations.
6145 if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6146 ret = apply_children_dir_moves(sctx);
6150 * Need to send that every time, no matter if it actually
6151 * changed between the two trees as we have done changes to
6152 * the inode before. If our inode is a directory and it's
6153 * waiting to be moved/renamed, we will send its utimes when
6154 * it's moved/renamed, therefore we don't need to do it here.
6156 sctx->send_progress = sctx->cur_ino + 1;
6157 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6166 struct parent_paths_ctx {
6167 struct list_head *refs;
6168 struct send_ctx *sctx;
6171 static int record_parent_ref(int num, u64 dir, int index, struct fs_path *name,
6174 struct parent_paths_ctx *ppctx = ctx;
6176 return record_ref(ppctx->sctx->parent_root, dir, name, ppctx->sctx,
6181 * Issue unlink operations for all paths of the current inode found in the
6184 static int btrfs_unlink_all_paths(struct send_ctx *sctx)
6186 LIST_HEAD(deleted_refs);
6187 struct btrfs_path *path;
6188 struct btrfs_key key;
6189 struct parent_paths_ctx ctx;
6192 path = alloc_path_for_send();
6196 key.objectid = sctx->cur_ino;
6197 key.type = BTRFS_INODE_REF_KEY;
6199 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
6203 ctx.refs = &deleted_refs;
6207 struct extent_buffer *eb = path->nodes[0];
6208 int slot = path->slots[0];
6210 if (slot >= btrfs_header_nritems(eb)) {
6211 ret = btrfs_next_leaf(sctx->parent_root, path);
6219 btrfs_item_key_to_cpu(eb, &key, slot);
6220 if (key.objectid != sctx->cur_ino)
6222 if (key.type != BTRFS_INODE_REF_KEY &&
6223 key.type != BTRFS_INODE_EXTREF_KEY)
6226 ret = iterate_inode_ref(sctx->parent_root, path, &key, 1,
6227 record_parent_ref, &ctx);
6234 while (!list_empty(&deleted_refs)) {
6235 struct recorded_ref *ref;
6237 ref = list_first_entry(&deleted_refs, struct recorded_ref, list);
6238 ret = send_unlink(sctx, ref->full_path);
6241 fs_path_free(ref->full_path);
6242 list_del(&ref->list);
6247 btrfs_free_path(path);
6249 __free_recorded_refs(&deleted_refs);
6253 static int changed_inode(struct send_ctx *sctx,
6254 enum btrfs_compare_tree_result result)
6257 struct btrfs_key *key = sctx->cmp_key;
6258 struct btrfs_inode_item *left_ii = NULL;
6259 struct btrfs_inode_item *right_ii = NULL;
6263 sctx->cur_ino = key->objectid;
6264 sctx->cur_inode_new_gen = 0;
6265 sctx->cur_inode_last_extent = (u64)-1;
6266 sctx->cur_inode_next_write_offset = 0;
6267 sctx->ignore_cur_inode = false;
6270 * Set send_progress to current inode. This will tell all get_cur_xxx
6271 * functions that the current inode's refs are not updated yet. Later,
6272 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6274 sctx->send_progress = sctx->cur_ino;
6276 if (result == BTRFS_COMPARE_TREE_NEW ||
6277 result == BTRFS_COMPARE_TREE_CHANGED) {
6278 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6279 sctx->left_path->slots[0],
6280 struct btrfs_inode_item);
6281 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6284 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6285 sctx->right_path->slots[0],
6286 struct btrfs_inode_item);
6287 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6290 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6291 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6292 sctx->right_path->slots[0],
6293 struct btrfs_inode_item);
6295 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6299 * The cur_ino = root dir case is special here. We can't treat
6300 * the inode as deleted+reused because it would generate a
6301 * stream that tries to delete/mkdir the root dir.
6303 if (left_gen != right_gen &&
6304 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6305 sctx->cur_inode_new_gen = 1;
6309 * Normally we do not find inodes with a link count of zero (orphans)
6310 * because the most common case is to create a snapshot and use it
6311 * for a send operation. However other less common use cases involve
6312 * using a subvolume and send it after turning it to RO mode just
6313 * after deleting all hard links of a file while holding an open
6314 * file descriptor against it or turning a RO snapshot into RW mode,
6315 * keep an open file descriptor against a file, delete it and then
6316 * turn the snapshot back to RO mode before using it for a send
6317 * operation. So if we find such cases, ignore the inode and all its
6318 * items completely if it's a new inode, or if it's a changed inode
6319 * make sure all its previous paths (from the parent snapshot) are all
6320 * unlinked and all other the inode items are ignored.
6322 if (result == BTRFS_COMPARE_TREE_NEW ||
6323 result == BTRFS_COMPARE_TREE_CHANGED) {
6326 nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6328 sctx->ignore_cur_inode = true;
6329 if (result == BTRFS_COMPARE_TREE_CHANGED)
6330 ret = btrfs_unlink_all_paths(sctx);
6335 if (result == BTRFS_COMPARE_TREE_NEW) {
6336 sctx->cur_inode_gen = left_gen;
6337 sctx->cur_inode_new = 1;
6338 sctx->cur_inode_deleted = 0;
6339 sctx->cur_inode_size = btrfs_inode_size(
6340 sctx->left_path->nodes[0], left_ii);
6341 sctx->cur_inode_mode = btrfs_inode_mode(
6342 sctx->left_path->nodes[0], left_ii);
6343 sctx->cur_inode_rdev = btrfs_inode_rdev(
6344 sctx->left_path->nodes[0], left_ii);
6345 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6346 ret = send_create_inode_if_needed(sctx);
6347 } else if (result == BTRFS_COMPARE_TREE_DELETED) {
6348 sctx->cur_inode_gen = right_gen;
6349 sctx->cur_inode_new = 0;
6350 sctx->cur_inode_deleted = 1;
6351 sctx->cur_inode_size = btrfs_inode_size(
6352 sctx->right_path->nodes[0], right_ii);
6353 sctx->cur_inode_mode = btrfs_inode_mode(
6354 sctx->right_path->nodes[0], right_ii);
6355 } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6357 * We need to do some special handling in case the inode was
6358 * reported as changed with a changed generation number. This
6359 * means that the original inode was deleted and new inode
6360 * reused the same inum. So we have to treat the old inode as
6361 * deleted and the new one as new.
6363 if (sctx->cur_inode_new_gen) {
6365 * First, process the inode as if it was deleted.
6367 sctx->cur_inode_gen = right_gen;
6368 sctx->cur_inode_new = 0;
6369 sctx->cur_inode_deleted = 1;
6370 sctx->cur_inode_size = btrfs_inode_size(
6371 sctx->right_path->nodes[0], right_ii);
6372 sctx->cur_inode_mode = btrfs_inode_mode(
6373 sctx->right_path->nodes[0], right_ii);
6374 ret = process_all_refs(sctx,
6375 BTRFS_COMPARE_TREE_DELETED);
6380 * Now process the inode as if it was new.
6382 sctx->cur_inode_gen = left_gen;
6383 sctx->cur_inode_new = 1;
6384 sctx->cur_inode_deleted = 0;
6385 sctx->cur_inode_size = btrfs_inode_size(
6386 sctx->left_path->nodes[0], left_ii);
6387 sctx->cur_inode_mode = btrfs_inode_mode(
6388 sctx->left_path->nodes[0], left_ii);
6389 sctx->cur_inode_rdev = btrfs_inode_rdev(
6390 sctx->left_path->nodes[0], left_ii);
6391 ret = send_create_inode_if_needed(sctx);
6395 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
6399 * Advance send_progress now as we did not get into
6400 * process_recorded_refs_if_needed in the new_gen case.
6402 sctx->send_progress = sctx->cur_ino + 1;
6405 * Now process all extents and xattrs of the inode as if
6406 * they were all new.
6408 ret = process_all_extents(sctx);
6411 ret = process_all_new_xattrs(sctx);
6415 sctx->cur_inode_gen = left_gen;
6416 sctx->cur_inode_new = 0;
6417 sctx->cur_inode_new_gen = 0;
6418 sctx->cur_inode_deleted = 0;
6419 sctx->cur_inode_size = btrfs_inode_size(
6420 sctx->left_path->nodes[0], left_ii);
6421 sctx->cur_inode_mode = btrfs_inode_mode(
6422 sctx->left_path->nodes[0], left_ii);
6431 * We have to process new refs before deleted refs, but compare_trees gives us
6432 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
6433 * first and later process them in process_recorded_refs.
6434 * For the cur_inode_new_gen case, we skip recording completely because
6435 * changed_inode did already initiate processing of refs. The reason for this is
6436 * that in this case, compare_tree actually compares the refs of 2 different
6437 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
6438 * refs of the right tree as deleted and all refs of the left tree as new.
6440 static int changed_ref(struct send_ctx *sctx,
6441 enum btrfs_compare_tree_result result)
6445 if (sctx->cur_ino != sctx->cmp_key->objectid) {
6446 inconsistent_snapshot_error(sctx, result, "reference");
6450 if (!sctx->cur_inode_new_gen &&
6451 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
6452 if (result == BTRFS_COMPARE_TREE_NEW)
6453 ret = record_new_ref(sctx);
6454 else if (result == BTRFS_COMPARE_TREE_DELETED)
6455 ret = record_deleted_ref(sctx);
6456 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6457 ret = record_changed_ref(sctx);
6464 * Process new/deleted/changed xattrs. We skip processing in the
6465 * cur_inode_new_gen case because changed_inode did already initiate processing
6466 * of xattrs. The reason is the same as in changed_ref
6468 static int changed_xattr(struct send_ctx *sctx,
6469 enum btrfs_compare_tree_result result)
6473 if (sctx->cur_ino != sctx->cmp_key->objectid) {
6474 inconsistent_snapshot_error(sctx, result, "xattr");
6478 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6479 if (result == BTRFS_COMPARE_TREE_NEW)
6480 ret = process_new_xattr(sctx);
6481 else if (result == BTRFS_COMPARE_TREE_DELETED)
6482 ret = process_deleted_xattr(sctx);
6483 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6484 ret = process_changed_xattr(sctx);
6491 * Process new/deleted/changed extents. We skip processing in the
6492 * cur_inode_new_gen case because changed_inode did already initiate processing
6493 * of extents. The reason is the same as in changed_ref
6495 static int changed_extent(struct send_ctx *sctx,
6496 enum btrfs_compare_tree_result result)
6501 * We have found an extent item that changed without the inode item
6502 * having changed. This can happen either after relocation (where the
6503 * disk_bytenr of an extent item is replaced at
6504 * relocation.c:replace_file_extents()) or after deduplication into a
6505 * file in both the parent and send snapshots (where an extent item can
6506 * get modified or replaced with a new one). Note that deduplication
6507 * updates the inode item, but it only changes the iversion (sequence
6508 * field in the inode item) of the inode, so if a file is deduplicated
6509 * the same amount of times in both the parent and send snapshots, its
6510 * iversion becomes the same in both snapshots, whence the inode item is
6511 * the same on both snapshots.
6513 if (sctx->cur_ino != sctx->cmp_key->objectid)
6516 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6517 if (result != BTRFS_COMPARE_TREE_DELETED)
6518 ret = process_extent(sctx, sctx->left_path,
6525 static int dir_changed(struct send_ctx *sctx, u64 dir)
6527 u64 orig_gen, new_gen;
6530 ret = get_inode_info(sctx->send_root, dir, NULL, &new_gen, NULL, NULL,
6535 ret = get_inode_info(sctx->parent_root, dir, NULL, &orig_gen, NULL,
6540 return (orig_gen != new_gen) ? 1 : 0;
6543 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
6544 struct btrfs_key *key)
6546 struct btrfs_inode_extref *extref;
6547 struct extent_buffer *leaf;
6548 u64 dirid = 0, last_dirid = 0;
6555 /* Easy case, just check this one dirid */
6556 if (key->type == BTRFS_INODE_REF_KEY) {
6557 dirid = key->offset;
6559 ret = dir_changed(sctx, dirid);
6563 leaf = path->nodes[0];
6564 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
6565 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
6566 while (cur_offset < item_size) {
6567 extref = (struct btrfs_inode_extref *)(ptr +
6569 dirid = btrfs_inode_extref_parent(leaf, extref);
6570 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
6571 cur_offset += ref_name_len + sizeof(*extref);
6572 if (dirid == last_dirid)
6574 ret = dir_changed(sctx, dirid);
6584 * Updates compare related fields in sctx and simply forwards to the actual
6585 * changed_xxx functions.
6587 static int changed_cb(struct btrfs_path *left_path,
6588 struct btrfs_path *right_path,
6589 struct btrfs_key *key,
6590 enum btrfs_compare_tree_result result,
6591 struct send_ctx *sctx)
6595 if (result == BTRFS_COMPARE_TREE_SAME) {
6596 if (key->type == BTRFS_INODE_REF_KEY ||
6597 key->type == BTRFS_INODE_EXTREF_KEY) {
6598 ret = compare_refs(sctx, left_path, key);
6603 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
6604 return maybe_send_hole(sctx, left_path, key);
6608 result = BTRFS_COMPARE_TREE_CHANGED;
6612 sctx->left_path = left_path;
6613 sctx->right_path = right_path;
6614 sctx->cmp_key = key;
6616 ret = finish_inode_if_needed(sctx, 0);
6620 /* Ignore non-FS objects */
6621 if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
6622 key->objectid == BTRFS_FREE_SPACE_OBJECTID)
6625 if (key->type == BTRFS_INODE_ITEM_KEY) {
6626 ret = changed_inode(sctx, result);
6627 } else if (!sctx->ignore_cur_inode) {
6628 if (key->type == BTRFS_INODE_REF_KEY ||
6629 key->type == BTRFS_INODE_EXTREF_KEY)
6630 ret = changed_ref(sctx, result);
6631 else if (key->type == BTRFS_XATTR_ITEM_KEY)
6632 ret = changed_xattr(sctx, result);
6633 else if (key->type == BTRFS_EXTENT_DATA_KEY)
6634 ret = changed_extent(sctx, result);
6641 static int full_send_tree(struct send_ctx *sctx)
6644 struct btrfs_root *send_root = sctx->send_root;
6645 struct btrfs_key key;
6646 struct btrfs_path *path;
6647 struct extent_buffer *eb;
6650 path = alloc_path_for_send();
6653 path->reada = READA_FORWARD_ALWAYS;
6655 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
6656 key.type = BTRFS_INODE_ITEM_KEY;
6659 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
6666 eb = path->nodes[0];
6667 slot = path->slots[0];
6668 btrfs_item_key_to_cpu(eb, &key, slot);
6670 ret = changed_cb(path, NULL, &key,
6671 BTRFS_COMPARE_TREE_NEW, sctx);
6675 ret = btrfs_next_item(send_root, path);
6685 ret = finish_inode_if_needed(sctx, 1);
6688 btrfs_free_path(path);
6692 static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
6694 struct extent_buffer *eb;
6695 struct extent_buffer *parent = path->nodes[*level];
6696 int slot = path->slots[*level];
6697 const int nritems = btrfs_header_nritems(parent);
6701 BUG_ON(*level == 0);
6702 eb = btrfs_read_node_slot(parent, slot);
6707 * Trigger readahead for the next leaves we will process, so that it is
6708 * very likely that when we need them they are already in memory and we
6709 * will not block on disk IO. For nodes we only do readahead for one,
6710 * since the time window between processing nodes is typically larger.
6712 reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
6714 for (slot++; slot < nritems && reada_done < reada_max; slot++) {
6715 if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
6716 btrfs_readahead_node_child(parent, slot);
6717 reada_done += eb->fs_info->nodesize;
6721 path->nodes[*level - 1] = eb;
6722 path->slots[*level - 1] = 0;
6727 static int tree_move_next_or_upnext(struct btrfs_path *path,
6728 int *level, int root_level)
6732 nritems = btrfs_header_nritems(path->nodes[*level]);
6734 path->slots[*level]++;
6736 while (path->slots[*level] >= nritems) {
6737 if (*level == root_level)
6741 path->slots[*level] = 0;
6742 free_extent_buffer(path->nodes[*level]);
6743 path->nodes[*level] = NULL;
6745 path->slots[*level]++;
6747 nritems = btrfs_header_nritems(path->nodes[*level]);
6754 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
6757 static int tree_advance(struct btrfs_path *path,
6758 int *level, int root_level,
6760 struct btrfs_key *key,
6765 if (*level == 0 || !allow_down) {
6766 ret = tree_move_next_or_upnext(path, level, root_level);
6768 ret = tree_move_down(path, level, reada_min_gen);
6772 btrfs_item_key_to_cpu(path->nodes[*level], key,
6773 path->slots[*level]);
6775 btrfs_node_key_to_cpu(path->nodes[*level], key,
6776 path->slots[*level]);
6781 static int tree_compare_item(struct btrfs_path *left_path,
6782 struct btrfs_path *right_path,
6787 unsigned long off1, off2;
6789 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
6790 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
6794 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
6795 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
6796 right_path->slots[0]);
6798 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
6800 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
6807 * This function compares two trees and calls the provided callback for
6808 * every changed/new/deleted item it finds.
6809 * If shared tree blocks are encountered, whole subtrees are skipped, making
6810 * the compare pretty fast on snapshotted subvolumes.
6812 * This currently works on commit roots only. As commit roots are read only,
6813 * we don't do any locking. The commit roots are protected with transactions.
6814 * Transactions are ended and rejoined when a commit is tried in between.
6816 * This function checks for modifications done to the trees while comparing.
6817 * If it detects a change, it aborts immediately.
6819 static int btrfs_compare_trees(struct btrfs_root *left_root,
6820 struct btrfs_root *right_root, struct send_ctx *sctx)
6822 struct btrfs_fs_info *fs_info = left_root->fs_info;
6825 struct btrfs_path *left_path = NULL;
6826 struct btrfs_path *right_path = NULL;
6827 struct btrfs_key left_key;
6828 struct btrfs_key right_key;
6829 char *tmp_buf = NULL;
6830 int left_root_level;
6831 int right_root_level;
6834 int left_end_reached;
6835 int right_end_reached;
6844 left_path = btrfs_alloc_path();
6849 right_path = btrfs_alloc_path();
6855 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
6861 left_path->search_commit_root = 1;
6862 left_path->skip_locking = 1;
6863 right_path->search_commit_root = 1;
6864 right_path->skip_locking = 1;
6867 * Strategy: Go to the first items of both trees. Then do
6869 * If both trees are at level 0
6870 * Compare keys of current items
6871 * If left < right treat left item as new, advance left tree
6873 * If left > right treat right item as deleted, advance right tree
6875 * If left == right do deep compare of items, treat as changed if
6876 * needed, advance both trees and repeat
6877 * If both trees are at the same level but not at level 0
6878 * Compare keys of current nodes/leafs
6879 * If left < right advance left tree and repeat
6880 * If left > right advance right tree and repeat
6881 * If left == right compare blockptrs of the next nodes/leafs
6882 * If they match advance both trees but stay at the same level
6884 * If they don't match advance both trees while allowing to go
6886 * If tree levels are different
6887 * Advance the tree that needs it and repeat
6889 * Advancing a tree means:
6890 * If we are at level 0, try to go to the next slot. If that's not
6891 * possible, go one level up and repeat. Stop when we found a level
6892 * where we could go to the next slot. We may at this point be on a
6895 * If we are not at level 0 and not on shared tree blocks, go one
6898 * If we are not at level 0 and on shared tree blocks, go one slot to
6899 * the right if possible or go up and right.
6902 down_read(&fs_info->commit_root_sem);
6903 left_level = btrfs_header_level(left_root->commit_root);
6904 left_root_level = left_level;
6905 left_path->nodes[left_level] =
6906 btrfs_clone_extent_buffer(left_root->commit_root);
6907 if (!left_path->nodes[left_level]) {
6908 up_read(&fs_info->commit_root_sem);
6913 right_level = btrfs_header_level(right_root->commit_root);
6914 right_root_level = right_level;
6915 right_path->nodes[right_level] =
6916 btrfs_clone_extent_buffer(right_root->commit_root);
6917 if (!right_path->nodes[right_level]) {
6918 up_read(&fs_info->commit_root_sem);
6923 * Our right root is the parent root, while the left root is the "send"
6924 * root. We know that all new nodes/leaves in the left root must have
6925 * a generation greater than the right root's generation, so we trigger
6926 * readahead for those nodes and leaves of the left root, as we know we
6927 * will need to read them at some point.
6929 reada_min_gen = btrfs_header_generation(right_root->commit_root);
6930 up_read(&fs_info->commit_root_sem);
6932 if (left_level == 0)
6933 btrfs_item_key_to_cpu(left_path->nodes[left_level],
6934 &left_key, left_path->slots[left_level]);
6936 btrfs_node_key_to_cpu(left_path->nodes[left_level],
6937 &left_key, left_path->slots[left_level]);
6938 if (right_level == 0)
6939 btrfs_item_key_to_cpu(right_path->nodes[right_level],
6940 &right_key, right_path->slots[right_level]);
6942 btrfs_node_key_to_cpu(right_path->nodes[right_level],
6943 &right_key, right_path->slots[right_level]);
6945 left_end_reached = right_end_reached = 0;
6946 advance_left = advance_right = 0;
6950 if (advance_left && !left_end_reached) {
6951 ret = tree_advance(left_path, &left_level,
6953 advance_left != ADVANCE_ONLY_NEXT,
6954 &left_key, reada_min_gen);
6956 left_end_reached = ADVANCE;
6961 if (advance_right && !right_end_reached) {
6962 ret = tree_advance(right_path, &right_level,
6964 advance_right != ADVANCE_ONLY_NEXT,
6965 &right_key, reada_min_gen);
6967 right_end_reached = ADVANCE;
6973 if (left_end_reached && right_end_reached) {
6976 } else if (left_end_reached) {
6977 if (right_level == 0) {
6978 ret = changed_cb(left_path, right_path,
6980 BTRFS_COMPARE_TREE_DELETED,
6985 advance_right = ADVANCE;
6987 } else if (right_end_reached) {
6988 if (left_level == 0) {
6989 ret = changed_cb(left_path, right_path,
6991 BTRFS_COMPARE_TREE_NEW,
6996 advance_left = ADVANCE;
7000 if (left_level == 0 && right_level == 0) {
7001 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7003 ret = changed_cb(left_path, right_path,
7005 BTRFS_COMPARE_TREE_NEW,
7009 advance_left = ADVANCE;
7010 } else if (cmp > 0) {
7011 ret = changed_cb(left_path, right_path,
7013 BTRFS_COMPARE_TREE_DELETED,
7017 advance_right = ADVANCE;
7019 enum btrfs_compare_tree_result result;
7021 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7022 ret = tree_compare_item(left_path, right_path,
7025 result = BTRFS_COMPARE_TREE_CHANGED;
7027 result = BTRFS_COMPARE_TREE_SAME;
7028 ret = changed_cb(left_path, right_path,
7029 &left_key, result, sctx);
7032 advance_left = ADVANCE;
7033 advance_right = ADVANCE;
7035 } else if (left_level == right_level) {
7036 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7038 advance_left = ADVANCE;
7039 } else if (cmp > 0) {
7040 advance_right = ADVANCE;
7042 left_blockptr = btrfs_node_blockptr(
7043 left_path->nodes[left_level],
7044 left_path->slots[left_level]);
7045 right_blockptr = btrfs_node_blockptr(
7046 right_path->nodes[right_level],
7047 right_path->slots[right_level]);
7048 left_gen = btrfs_node_ptr_generation(
7049 left_path->nodes[left_level],
7050 left_path->slots[left_level]);
7051 right_gen = btrfs_node_ptr_generation(
7052 right_path->nodes[right_level],
7053 right_path->slots[right_level]);
7054 if (left_blockptr == right_blockptr &&
7055 left_gen == right_gen) {
7057 * As we're on a shared block, don't
7058 * allow to go deeper.
7060 advance_left = ADVANCE_ONLY_NEXT;
7061 advance_right = ADVANCE_ONLY_NEXT;
7063 advance_left = ADVANCE;
7064 advance_right = ADVANCE;
7067 } else if (left_level < right_level) {
7068 advance_right = ADVANCE;
7070 advance_left = ADVANCE;
7075 btrfs_free_path(left_path);
7076 btrfs_free_path(right_path);
7081 static int send_subvol(struct send_ctx *sctx)
7085 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7086 ret = send_header(sctx);
7091 ret = send_subvol_begin(sctx);
7095 if (sctx->parent_root) {
7096 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7099 ret = finish_inode_if_needed(sctx, 1);
7103 ret = full_send_tree(sctx);
7109 free_recorded_refs(sctx);
7114 * If orphan cleanup did remove any orphans from a root, it means the tree
7115 * was modified and therefore the commit root is not the same as the current
7116 * root anymore. This is a problem, because send uses the commit root and
7117 * therefore can see inode items that don't exist in the current root anymore,
7118 * and for example make calls to btrfs_iget, which will do tree lookups based
7119 * on the current root and not on the commit root. Those lookups will fail,
7120 * returning a -ESTALE error, and making send fail with that error. So make
7121 * sure a send does not see any orphans we have just removed, and that it will
7122 * see the same inodes regardless of whether a transaction commit happened
7123 * before it started (meaning that the commit root will be the same as the
7124 * current root) or not.
7126 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
7129 struct btrfs_trans_handle *trans = NULL;
7132 if (sctx->parent_root &&
7133 sctx->parent_root->node != sctx->parent_root->commit_root)
7136 for (i = 0; i < sctx->clone_roots_cnt; i++)
7137 if (sctx->clone_roots[i].root->node !=
7138 sctx->clone_roots[i].root->commit_root)
7142 return btrfs_end_transaction(trans);
7147 /* Use any root, all fs roots will get their commit roots updated. */
7149 trans = btrfs_join_transaction(sctx->send_root);
7151 return PTR_ERR(trans);
7155 return btrfs_commit_transaction(trans);
7159 * Make sure any existing dellaloc is flushed for any root used by a send
7160 * operation so that we do not miss any data and we do not race with writeback
7161 * finishing and changing a tree while send is using the tree. This could
7162 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
7163 * a send operation then uses the subvolume.
7164 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
7166 static int flush_delalloc_roots(struct send_ctx *sctx)
7168 struct btrfs_root *root = sctx->parent_root;
7173 ret = btrfs_start_delalloc_snapshot(root, false);
7176 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
7179 for (i = 0; i < sctx->clone_roots_cnt; i++) {
7180 root = sctx->clone_roots[i].root;
7181 ret = btrfs_start_delalloc_snapshot(root, false);
7184 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
7190 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
7192 spin_lock(&root->root_item_lock);
7193 root->send_in_progress--;
7195 * Not much left to do, we don't know why it's unbalanced and
7196 * can't blindly reset it to 0.
7198 if (root->send_in_progress < 0)
7199 btrfs_err(root->fs_info,
7200 "send_in_progress unbalanced %d root %llu",
7201 root->send_in_progress, root->root_key.objectid);
7202 spin_unlock(&root->root_item_lock);
7205 static void dedupe_in_progress_warn(const struct btrfs_root *root)
7207 btrfs_warn_rl(root->fs_info,
7208 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
7209 root->root_key.objectid, root->dedupe_in_progress);
7212 long btrfs_ioctl_send(struct file *mnt_file, struct btrfs_ioctl_send_args *arg)
7215 struct btrfs_root *send_root = BTRFS_I(file_inode(mnt_file))->root;
7216 struct btrfs_fs_info *fs_info = send_root->fs_info;
7217 struct btrfs_root *clone_root;
7218 struct send_ctx *sctx = NULL;
7220 u64 *clone_sources_tmp = NULL;
7221 int clone_sources_to_rollback = 0;
7223 int sort_clone_roots = 0;
7225 if (!capable(CAP_SYS_ADMIN))
7229 * The subvolume must remain read-only during send, protect against
7230 * making it RW. This also protects against deletion.
7232 spin_lock(&send_root->root_item_lock);
7233 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
7234 dedupe_in_progress_warn(send_root);
7235 spin_unlock(&send_root->root_item_lock);
7238 send_root->send_in_progress++;
7239 spin_unlock(&send_root->root_item_lock);
7242 * Userspace tools do the checks and warn the user if it's
7245 if (!btrfs_root_readonly(send_root)) {
7251 * Check that we don't overflow at later allocations, we request
7252 * clone_sources_count + 1 items, and compare to unsigned long inside
7255 if (arg->clone_sources_count >
7256 ULONG_MAX / sizeof(struct clone_root) - 1) {
7261 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
7266 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
7272 INIT_LIST_HEAD(&sctx->new_refs);
7273 INIT_LIST_HEAD(&sctx->deleted_refs);
7274 INIT_RADIX_TREE(&sctx->name_cache, GFP_KERNEL);
7275 INIT_LIST_HEAD(&sctx->name_cache_list);
7277 sctx->flags = arg->flags;
7279 sctx->send_filp = fget(arg->send_fd);
7280 if (!sctx->send_filp) {
7285 sctx->send_root = send_root;
7287 * Unlikely but possible, if the subvolume is marked for deletion but
7288 * is slow to remove the directory entry, send can still be started
7290 if (btrfs_root_dead(sctx->send_root)) {
7295 sctx->clone_roots_cnt = arg->clone_sources_count;
7297 sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
7298 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
7299 if (!sctx->send_buf) {
7304 sctx->pending_dir_moves = RB_ROOT;
7305 sctx->waiting_dir_moves = RB_ROOT;
7306 sctx->orphan_dirs = RB_ROOT;
7308 sctx->clone_roots = kvcalloc(sizeof(*sctx->clone_roots),
7309 arg->clone_sources_count + 1,
7311 if (!sctx->clone_roots) {
7316 alloc_size = array_size(sizeof(*arg->clone_sources),
7317 arg->clone_sources_count);
7319 if (arg->clone_sources_count) {
7320 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
7321 if (!clone_sources_tmp) {
7326 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
7333 for (i = 0; i < arg->clone_sources_count; i++) {
7334 clone_root = btrfs_get_fs_root(fs_info,
7335 clone_sources_tmp[i], true);
7336 if (IS_ERR(clone_root)) {
7337 ret = PTR_ERR(clone_root);
7340 spin_lock(&clone_root->root_item_lock);
7341 if (!btrfs_root_readonly(clone_root) ||
7342 btrfs_root_dead(clone_root)) {
7343 spin_unlock(&clone_root->root_item_lock);
7344 btrfs_put_root(clone_root);
7348 if (clone_root->dedupe_in_progress) {
7349 dedupe_in_progress_warn(clone_root);
7350 spin_unlock(&clone_root->root_item_lock);
7351 btrfs_put_root(clone_root);
7355 clone_root->send_in_progress++;
7356 spin_unlock(&clone_root->root_item_lock);
7358 sctx->clone_roots[i].root = clone_root;
7359 clone_sources_to_rollback = i + 1;
7361 kvfree(clone_sources_tmp);
7362 clone_sources_tmp = NULL;
7365 if (arg->parent_root) {
7366 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
7368 if (IS_ERR(sctx->parent_root)) {
7369 ret = PTR_ERR(sctx->parent_root);
7373 spin_lock(&sctx->parent_root->root_item_lock);
7374 sctx->parent_root->send_in_progress++;
7375 if (!btrfs_root_readonly(sctx->parent_root) ||
7376 btrfs_root_dead(sctx->parent_root)) {
7377 spin_unlock(&sctx->parent_root->root_item_lock);
7381 if (sctx->parent_root->dedupe_in_progress) {
7382 dedupe_in_progress_warn(sctx->parent_root);
7383 spin_unlock(&sctx->parent_root->root_item_lock);
7387 spin_unlock(&sctx->parent_root->root_item_lock);
7391 * Clones from send_root are allowed, but only if the clone source
7392 * is behind the current send position. This is checked while searching
7393 * for possible clone sources.
7395 sctx->clone_roots[sctx->clone_roots_cnt++].root =
7396 btrfs_grab_root(sctx->send_root);
7398 /* We do a bsearch later */
7399 sort(sctx->clone_roots, sctx->clone_roots_cnt,
7400 sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
7402 sort_clone_roots = 1;
7404 ret = flush_delalloc_roots(sctx);
7408 ret = ensure_commit_roots_uptodate(sctx);
7412 spin_lock(&fs_info->send_reloc_lock);
7413 if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags)) {
7414 spin_unlock(&fs_info->send_reloc_lock);
7415 btrfs_warn_rl(fs_info,
7416 "cannot run send because a relocation operation is in progress");
7420 fs_info->send_in_progress++;
7421 spin_unlock(&fs_info->send_reloc_lock);
7423 ret = send_subvol(sctx);
7424 spin_lock(&fs_info->send_reloc_lock);
7425 fs_info->send_in_progress--;
7426 spin_unlock(&fs_info->send_reloc_lock);
7430 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
7431 ret = begin_cmd(sctx, BTRFS_SEND_C_END);
7434 ret = send_cmd(sctx);
7440 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
7441 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
7443 struct pending_dir_move *pm;
7445 n = rb_first(&sctx->pending_dir_moves);
7446 pm = rb_entry(n, struct pending_dir_move, node);
7447 while (!list_empty(&pm->list)) {
7448 struct pending_dir_move *pm2;
7450 pm2 = list_first_entry(&pm->list,
7451 struct pending_dir_move, list);
7452 free_pending_move(sctx, pm2);
7454 free_pending_move(sctx, pm);
7457 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
7458 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
7460 struct waiting_dir_move *dm;
7462 n = rb_first(&sctx->waiting_dir_moves);
7463 dm = rb_entry(n, struct waiting_dir_move, node);
7464 rb_erase(&dm->node, &sctx->waiting_dir_moves);
7468 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
7469 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
7471 struct orphan_dir_info *odi;
7473 n = rb_first(&sctx->orphan_dirs);
7474 odi = rb_entry(n, struct orphan_dir_info, node);
7475 free_orphan_dir_info(sctx, odi);
7478 if (sort_clone_roots) {
7479 for (i = 0; i < sctx->clone_roots_cnt; i++) {
7480 btrfs_root_dec_send_in_progress(
7481 sctx->clone_roots[i].root);
7482 btrfs_put_root(sctx->clone_roots[i].root);
7485 for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
7486 btrfs_root_dec_send_in_progress(
7487 sctx->clone_roots[i].root);
7488 btrfs_put_root(sctx->clone_roots[i].root);
7491 btrfs_root_dec_send_in_progress(send_root);
7493 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
7494 btrfs_root_dec_send_in_progress(sctx->parent_root);
7495 btrfs_put_root(sctx->parent_root);
7498 kvfree(clone_sources_tmp);
7501 if (sctx->send_filp)
7502 fput(sctx->send_filp);
7504 kvfree(sctx->clone_roots);
7505 kvfree(sctx->send_buf);
7507 name_cache_free(sctx);