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 if (unlikely(start < p->buf)) {
970 btrfs_err(root->fs_info,
971 "send: path ref buffer underflow for key (%llu %u %llu)",
981 ret = fs_path_add_from_extent_buffer(p, eb, name_off,
987 cur += elem_size + name_len;
988 ret = iterate(num, dir, index, p, ctx);
995 btrfs_free_path(tmp_path);
1000 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
1001 const char *name, int name_len,
1002 const char *data, int data_len,
1003 u8 type, void *ctx);
1006 * Helper function to iterate the entries in ONE btrfs_dir_item.
1007 * The iterate callback may return a non zero value to stop iteration. This can
1008 * be a negative value for error codes or 1 to simply stop it.
1010 * path must point to the dir item when called.
1012 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1013 iterate_dir_item_t iterate, void *ctx)
1016 struct extent_buffer *eb;
1017 struct btrfs_item *item;
1018 struct btrfs_dir_item *di;
1019 struct btrfs_key di_key;
1032 * Start with a small buffer (1 page). If later we end up needing more
1033 * space, which can happen for xattrs on a fs with a leaf size greater
1034 * then the page size, attempt to increase the buffer. Typically xattr
1038 buf = kmalloc(buf_len, GFP_KERNEL);
1044 eb = path->nodes[0];
1045 slot = path->slots[0];
1046 item = btrfs_item_nr(slot);
1047 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1050 total = btrfs_item_size(eb, item);
1053 while (cur < total) {
1054 name_len = btrfs_dir_name_len(eb, di);
1055 data_len = btrfs_dir_data_len(eb, di);
1056 type = btrfs_dir_type(eb, di);
1057 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1059 if (type == BTRFS_FT_XATTR) {
1060 if (name_len > XATTR_NAME_MAX) {
1061 ret = -ENAMETOOLONG;
1064 if (name_len + data_len >
1065 BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1073 if (name_len + data_len > PATH_MAX) {
1074 ret = -ENAMETOOLONG;
1079 if (name_len + data_len > buf_len) {
1080 buf_len = name_len + data_len;
1081 if (is_vmalloc_addr(buf)) {
1085 char *tmp = krealloc(buf, buf_len,
1086 GFP_KERNEL | __GFP_NOWARN);
1093 buf = kvmalloc(buf_len, GFP_KERNEL);
1101 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1102 name_len + data_len);
1104 len = sizeof(*di) + name_len + data_len;
1105 di = (struct btrfs_dir_item *)((char *)di + len);
1108 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1109 data_len, type, ctx);
1125 static int __copy_first_ref(int num, u64 dir, int index,
1126 struct fs_path *p, void *ctx)
1129 struct fs_path *pt = ctx;
1131 ret = fs_path_copy(pt, p);
1135 /* we want the first only */
1140 * Retrieve the first path of an inode. If an inode has more then one
1141 * ref/hardlink, this is ignored.
1143 static int get_inode_path(struct btrfs_root *root,
1144 u64 ino, struct fs_path *path)
1147 struct btrfs_key key, found_key;
1148 struct btrfs_path *p;
1150 p = alloc_path_for_send();
1154 fs_path_reset(path);
1157 key.type = BTRFS_INODE_REF_KEY;
1160 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1167 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1168 if (found_key.objectid != ino ||
1169 (found_key.type != BTRFS_INODE_REF_KEY &&
1170 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1175 ret = iterate_inode_ref(root, p, &found_key, 1,
1176 __copy_first_ref, path);
1186 struct backref_ctx {
1187 struct send_ctx *sctx;
1189 /* number of total found references */
1193 * used for clones found in send_root. clones found behind cur_objectid
1194 * and cur_offset are not considered as allowed clones.
1199 /* may be truncated in case it's the last extent in a file */
1202 /* data offset in the file extent item */
1205 /* Just to check for bugs in backref resolving */
1209 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1211 u64 root = (u64)(uintptr_t)key;
1212 struct clone_root *cr = (struct clone_root *)elt;
1214 if (root < cr->root->root_key.objectid)
1216 if (root > cr->root->root_key.objectid)
1221 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1223 struct clone_root *cr1 = (struct clone_root *)e1;
1224 struct clone_root *cr2 = (struct clone_root *)e2;
1226 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1228 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1234 * Called for every backref that is found for the current extent.
1235 * Results are collected in sctx->clone_roots->ino/offset/found_refs
1237 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1239 struct backref_ctx *bctx = ctx_;
1240 struct clone_root *found;
1242 /* First check if the root is in the list of accepted clone sources */
1243 found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1244 bctx->sctx->clone_roots_cnt,
1245 sizeof(struct clone_root),
1246 __clone_root_cmp_bsearch);
1250 if (found->root == bctx->sctx->send_root &&
1251 ino == bctx->cur_objectid &&
1252 offset == bctx->cur_offset) {
1253 bctx->found_itself = 1;
1257 * Make sure we don't consider clones from send_root that are
1258 * behind the current inode/offset.
1260 if (found->root == bctx->sctx->send_root) {
1262 * If the source inode was not yet processed we can't issue a
1263 * clone operation, as the source extent does not exist yet at
1264 * the destination of the stream.
1266 if (ino > bctx->cur_objectid)
1269 * We clone from the inode currently being sent as long as the
1270 * source extent is already processed, otherwise we could try
1271 * to clone from an extent that does not exist yet at the
1272 * destination of the stream.
1274 if (ino == bctx->cur_objectid &&
1275 offset + bctx->extent_len >
1276 bctx->sctx->cur_inode_next_write_offset)
1281 found->found_refs++;
1282 if (ino < found->ino) {
1284 found->offset = offset;
1285 } else if (found->ino == ino) {
1287 * same extent found more then once in the same file.
1289 if (found->offset > offset + bctx->extent_len)
1290 found->offset = offset;
1297 * Given an inode, offset and extent item, it finds a good clone for a clone
1298 * instruction. Returns -ENOENT when none could be found. The function makes
1299 * sure that the returned clone is usable at the point where sending is at the
1300 * moment. This means, that no clones are accepted which lie behind the current
1303 * path must point to the extent item when called.
1305 static int find_extent_clone(struct send_ctx *sctx,
1306 struct btrfs_path *path,
1307 u64 ino, u64 data_offset,
1309 struct clone_root **found)
1311 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1317 u64 extent_item_pos;
1319 struct btrfs_file_extent_item *fi;
1320 struct extent_buffer *eb = path->nodes[0];
1321 struct backref_ctx *backref_ctx = NULL;
1322 struct clone_root *cur_clone_root;
1323 struct btrfs_key found_key;
1324 struct btrfs_path *tmp_path;
1325 struct btrfs_extent_item *ei;
1329 tmp_path = alloc_path_for_send();
1333 /* We only use this path under the commit sem */
1334 tmp_path->need_commit_sem = 0;
1336 backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1342 if (data_offset >= ino_size) {
1344 * There may be extents that lie behind the file's size.
1345 * I at least had this in combination with snapshotting while
1346 * writing large files.
1352 fi = btrfs_item_ptr(eb, path->slots[0],
1353 struct btrfs_file_extent_item);
1354 extent_type = btrfs_file_extent_type(eb, fi);
1355 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1359 compressed = btrfs_file_extent_compression(eb, fi);
1361 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1362 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1363 if (disk_byte == 0) {
1367 logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1369 down_read(&fs_info->commit_root_sem);
1370 ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1371 &found_key, &flags);
1372 up_read(&fs_info->commit_root_sem);
1376 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1381 ei = btrfs_item_ptr(tmp_path->nodes[0], tmp_path->slots[0],
1382 struct btrfs_extent_item);
1384 * Backreference walking (iterate_extent_inodes() below) is currently
1385 * too expensive when an extent has a large number of references, both
1386 * in time spent and used memory. So for now just fallback to write
1387 * operations instead of clone operations when an extent has more than
1388 * a certain amount of references.
1390 if (btrfs_extent_refs(tmp_path->nodes[0], ei) > SEND_MAX_EXTENT_REFS) {
1394 btrfs_release_path(tmp_path);
1397 * Setup the clone roots.
1399 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1400 cur_clone_root = sctx->clone_roots + i;
1401 cur_clone_root->ino = (u64)-1;
1402 cur_clone_root->offset = 0;
1403 cur_clone_root->found_refs = 0;
1406 backref_ctx->sctx = sctx;
1407 backref_ctx->found = 0;
1408 backref_ctx->cur_objectid = ino;
1409 backref_ctx->cur_offset = data_offset;
1410 backref_ctx->found_itself = 0;
1411 backref_ctx->extent_len = num_bytes;
1413 * For non-compressed extents iterate_extent_inodes() gives us extent
1414 * offsets that already take into account the data offset, but not for
1415 * compressed extents, since the offset is logical and not relative to
1416 * the physical extent locations. We must take this into account to
1417 * avoid sending clone offsets that go beyond the source file's size,
1418 * which would result in the clone ioctl failing with -EINVAL on the
1421 if (compressed == BTRFS_COMPRESS_NONE)
1422 backref_ctx->data_offset = 0;
1424 backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1427 * The last extent of a file may be too large due to page alignment.
1428 * We need to adjust extent_len in this case so that the checks in
1429 * __iterate_backrefs work.
1431 if (data_offset + num_bytes >= ino_size)
1432 backref_ctx->extent_len = ino_size - data_offset;
1435 * Now collect all backrefs.
1437 if (compressed == BTRFS_COMPRESS_NONE)
1438 extent_item_pos = logical - found_key.objectid;
1440 extent_item_pos = 0;
1441 ret = iterate_extent_inodes(fs_info, found_key.objectid,
1442 extent_item_pos, 1, __iterate_backrefs,
1443 backref_ctx, false);
1448 if (!backref_ctx->found_itself) {
1449 /* found a bug in backref code? */
1452 "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1453 ino, data_offset, disk_byte, found_key.objectid);
1457 btrfs_debug(fs_info,
1458 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1459 data_offset, ino, num_bytes, logical);
1461 if (!backref_ctx->found)
1462 btrfs_debug(fs_info, "no clones found");
1464 cur_clone_root = NULL;
1465 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1466 if (sctx->clone_roots[i].found_refs) {
1467 if (!cur_clone_root)
1468 cur_clone_root = sctx->clone_roots + i;
1469 else if (sctx->clone_roots[i].root == sctx->send_root)
1470 /* prefer clones from send_root over others */
1471 cur_clone_root = sctx->clone_roots + i;
1476 if (cur_clone_root) {
1477 *found = cur_clone_root;
1484 btrfs_free_path(tmp_path);
1489 static int read_symlink(struct btrfs_root *root,
1491 struct fs_path *dest)
1494 struct btrfs_path *path;
1495 struct btrfs_key key;
1496 struct btrfs_file_extent_item *ei;
1502 path = alloc_path_for_send();
1507 key.type = BTRFS_EXTENT_DATA_KEY;
1509 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1514 * An empty symlink inode. Can happen in rare error paths when
1515 * creating a symlink (transaction committed before the inode
1516 * eviction handler removed the symlink inode items and a crash
1517 * happened in between or the subvol was snapshoted in between).
1518 * Print an informative message to dmesg/syslog so that the user
1519 * can delete the symlink.
1521 btrfs_err(root->fs_info,
1522 "Found empty symlink inode %llu at root %llu",
1523 ino, root->root_key.objectid);
1528 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1529 struct btrfs_file_extent_item);
1530 type = btrfs_file_extent_type(path->nodes[0], ei);
1531 compression = btrfs_file_extent_compression(path->nodes[0], ei);
1532 BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1533 BUG_ON(compression);
1535 off = btrfs_file_extent_inline_start(ei);
1536 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1538 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1541 btrfs_free_path(path);
1546 * Helper function to generate a file name that is unique in the root of
1547 * send_root and parent_root. This is used to generate names for orphan inodes.
1549 static int gen_unique_name(struct send_ctx *sctx,
1551 struct fs_path *dest)
1554 struct btrfs_path *path;
1555 struct btrfs_dir_item *di;
1560 path = alloc_path_for_send();
1565 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1567 ASSERT(len < sizeof(tmp));
1569 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1570 path, BTRFS_FIRST_FREE_OBJECTID,
1571 tmp, strlen(tmp), 0);
1572 btrfs_release_path(path);
1578 /* not unique, try again */
1583 if (!sctx->parent_root) {
1589 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1590 path, BTRFS_FIRST_FREE_OBJECTID,
1591 tmp, strlen(tmp), 0);
1592 btrfs_release_path(path);
1598 /* not unique, try again */
1606 ret = fs_path_add(dest, tmp, strlen(tmp));
1609 btrfs_free_path(path);
1614 inode_state_no_change,
1615 inode_state_will_create,
1616 inode_state_did_create,
1617 inode_state_will_delete,
1618 inode_state_did_delete,
1621 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1629 ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1631 if (ret < 0 && ret != -ENOENT)
1635 if (!sctx->parent_root) {
1636 right_ret = -ENOENT;
1638 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1639 NULL, NULL, NULL, NULL);
1640 if (ret < 0 && ret != -ENOENT)
1645 if (!left_ret && !right_ret) {
1646 if (left_gen == gen && right_gen == gen) {
1647 ret = inode_state_no_change;
1648 } else if (left_gen == gen) {
1649 if (ino < sctx->send_progress)
1650 ret = inode_state_did_create;
1652 ret = inode_state_will_create;
1653 } else if (right_gen == gen) {
1654 if (ino < sctx->send_progress)
1655 ret = inode_state_did_delete;
1657 ret = inode_state_will_delete;
1661 } else if (!left_ret) {
1662 if (left_gen == gen) {
1663 if (ino < sctx->send_progress)
1664 ret = inode_state_did_create;
1666 ret = inode_state_will_create;
1670 } else if (!right_ret) {
1671 if (right_gen == gen) {
1672 if (ino < sctx->send_progress)
1673 ret = inode_state_did_delete;
1675 ret = inode_state_will_delete;
1687 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1691 if (ino == BTRFS_FIRST_FREE_OBJECTID)
1694 ret = get_cur_inode_state(sctx, ino, gen);
1698 if (ret == inode_state_no_change ||
1699 ret == inode_state_did_create ||
1700 ret == inode_state_will_delete)
1710 * Helper function to lookup a dir item in a dir.
1712 static int lookup_dir_item_inode(struct btrfs_root *root,
1713 u64 dir, const char *name, int name_len,
1718 struct btrfs_dir_item *di;
1719 struct btrfs_key key;
1720 struct btrfs_path *path;
1722 path = alloc_path_for_send();
1726 di = btrfs_lookup_dir_item(NULL, root, path,
1727 dir, name, name_len, 0);
1728 if (IS_ERR_OR_NULL(di)) {
1729 ret = di ? PTR_ERR(di) : -ENOENT;
1732 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1733 if (key.type == BTRFS_ROOT_ITEM_KEY) {
1737 *found_inode = key.objectid;
1738 *found_type = btrfs_dir_type(path->nodes[0], di);
1741 btrfs_free_path(path);
1746 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1747 * generation of the parent dir and the name of the dir entry.
1749 static int get_first_ref(struct btrfs_root *root, u64 ino,
1750 u64 *dir, u64 *dir_gen, struct fs_path *name)
1753 struct btrfs_key key;
1754 struct btrfs_key found_key;
1755 struct btrfs_path *path;
1759 path = alloc_path_for_send();
1764 key.type = BTRFS_INODE_REF_KEY;
1767 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1771 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1773 if (ret || found_key.objectid != ino ||
1774 (found_key.type != BTRFS_INODE_REF_KEY &&
1775 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1780 if (found_key.type == BTRFS_INODE_REF_KEY) {
1781 struct btrfs_inode_ref *iref;
1782 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1783 struct btrfs_inode_ref);
1784 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1785 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1786 (unsigned long)(iref + 1),
1788 parent_dir = found_key.offset;
1790 struct btrfs_inode_extref *extref;
1791 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1792 struct btrfs_inode_extref);
1793 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1794 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1795 (unsigned long)&extref->name, len);
1796 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1800 btrfs_release_path(path);
1803 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1812 btrfs_free_path(path);
1816 static int is_first_ref(struct btrfs_root *root,
1818 const char *name, int name_len)
1821 struct fs_path *tmp_name;
1824 tmp_name = fs_path_alloc();
1828 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1832 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1837 ret = !memcmp(tmp_name->start, name, name_len);
1840 fs_path_free(tmp_name);
1845 * Used by process_recorded_refs to determine if a new ref would overwrite an
1846 * already existing ref. In case it detects an overwrite, it returns the
1847 * inode/gen in who_ino/who_gen.
1848 * When an overwrite is detected, process_recorded_refs does proper orphanizing
1849 * to make sure later references to the overwritten inode are possible.
1850 * Orphanizing is however only required for the first ref of an inode.
1851 * process_recorded_refs does an additional is_first_ref check to see if
1852 * orphanizing is really required.
1854 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1855 const char *name, int name_len,
1856 u64 *who_ino, u64 *who_gen, u64 *who_mode)
1860 u64 other_inode = 0;
1863 if (!sctx->parent_root)
1866 ret = is_inode_existent(sctx, dir, dir_gen);
1871 * If we have a parent root we need to verify that the parent dir was
1872 * not deleted and then re-created, if it was then we have no overwrite
1873 * and we can just unlink this entry.
1875 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1876 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1878 if (ret < 0 && ret != -ENOENT)
1888 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1889 &other_inode, &other_type);
1890 if (ret < 0 && ret != -ENOENT)
1898 * Check if the overwritten ref was already processed. If yes, the ref
1899 * was already unlinked/moved, so we can safely assume that we will not
1900 * overwrite anything at this point in time.
1902 if (other_inode > sctx->send_progress ||
1903 is_waiting_for_move(sctx, other_inode)) {
1904 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1905 who_gen, who_mode, NULL, NULL, NULL);
1910 *who_ino = other_inode;
1920 * Checks if the ref was overwritten by an already processed inode. This is
1921 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1922 * thus the orphan name needs be used.
1923 * process_recorded_refs also uses it to avoid unlinking of refs that were
1926 static int did_overwrite_ref(struct send_ctx *sctx,
1927 u64 dir, u64 dir_gen,
1928 u64 ino, u64 ino_gen,
1929 const char *name, int name_len)
1936 if (!sctx->parent_root)
1939 ret = is_inode_existent(sctx, dir, dir_gen);
1943 if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1944 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1946 if (ret < 0 && ret != -ENOENT)
1956 /* check if the ref was overwritten by another ref */
1957 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1958 &ow_inode, &other_type);
1959 if (ret < 0 && ret != -ENOENT)
1962 /* was never and will never be overwritten */
1967 ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1972 if (ow_inode == ino && gen == ino_gen) {
1978 * We know that it is or will be overwritten. Check this now.
1979 * The current inode being processed might have been the one that caused
1980 * inode 'ino' to be orphanized, therefore check if ow_inode matches
1981 * the current inode being processed.
1983 if ((ow_inode < sctx->send_progress) ||
1984 (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1985 gen == sctx->cur_inode_gen))
1995 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1996 * that got overwritten. This is used by process_recorded_refs to determine
1997 * if it has to use the path as returned by get_cur_path or the orphan name.
1999 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
2002 struct fs_path *name = NULL;
2006 if (!sctx->parent_root)
2009 name = fs_path_alloc();
2013 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2017 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2018 name->start, fs_path_len(name));
2026 * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
2027 * so we need to do some special handling in case we have clashes. This function
2028 * takes care of this with the help of name_cache_entry::radix_list.
2029 * In case of error, nce is kfreed.
2031 static int name_cache_insert(struct send_ctx *sctx,
2032 struct name_cache_entry *nce)
2035 struct list_head *nce_head;
2037 nce_head = radix_tree_lookup(&sctx->name_cache,
2038 (unsigned long)nce->ino);
2040 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2045 INIT_LIST_HEAD(nce_head);
2047 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2054 list_add_tail(&nce->radix_list, nce_head);
2055 list_add_tail(&nce->list, &sctx->name_cache_list);
2056 sctx->name_cache_size++;
2061 static void name_cache_delete(struct send_ctx *sctx,
2062 struct name_cache_entry *nce)
2064 struct list_head *nce_head;
2066 nce_head = radix_tree_lookup(&sctx->name_cache,
2067 (unsigned long)nce->ino);
2069 btrfs_err(sctx->send_root->fs_info,
2070 "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2071 nce->ino, sctx->name_cache_size);
2074 list_del(&nce->radix_list);
2075 list_del(&nce->list);
2076 sctx->name_cache_size--;
2079 * We may not get to the final release of nce_head if the lookup fails
2081 if (nce_head && list_empty(nce_head)) {
2082 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2087 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2090 struct list_head *nce_head;
2091 struct name_cache_entry *cur;
2093 nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2097 list_for_each_entry(cur, nce_head, radix_list) {
2098 if (cur->ino == ino && cur->gen == gen)
2105 * Removes the entry from the list and adds it back to the end. This marks the
2106 * entry as recently used so that name_cache_clean_unused does not remove it.
2108 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2110 list_del(&nce->list);
2111 list_add_tail(&nce->list, &sctx->name_cache_list);
2115 * Remove some entries from the beginning of name_cache_list.
2117 static void name_cache_clean_unused(struct send_ctx *sctx)
2119 struct name_cache_entry *nce;
2121 if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2124 while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2125 nce = list_entry(sctx->name_cache_list.next,
2126 struct name_cache_entry, list);
2127 name_cache_delete(sctx, nce);
2132 static void name_cache_free(struct send_ctx *sctx)
2134 struct name_cache_entry *nce;
2136 while (!list_empty(&sctx->name_cache_list)) {
2137 nce = list_entry(sctx->name_cache_list.next,
2138 struct name_cache_entry, list);
2139 name_cache_delete(sctx, nce);
2145 * Used by get_cur_path for each ref up to the root.
2146 * Returns 0 if it succeeded.
2147 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2148 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2149 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2150 * Returns <0 in case of error.
2152 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2156 struct fs_path *dest)
2160 struct name_cache_entry *nce = NULL;
2163 * First check if we already did a call to this function with the same
2164 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2165 * return the cached result.
2167 nce = name_cache_search(sctx, ino, gen);
2169 if (ino < sctx->send_progress && nce->need_later_update) {
2170 name_cache_delete(sctx, nce);
2174 name_cache_used(sctx, nce);
2175 *parent_ino = nce->parent_ino;
2176 *parent_gen = nce->parent_gen;
2177 ret = fs_path_add(dest, nce->name, nce->name_len);
2186 * If the inode is not existent yet, add the orphan name and return 1.
2187 * This should only happen for the parent dir that we determine in
2190 ret = is_inode_existent(sctx, ino, gen);
2195 ret = gen_unique_name(sctx, ino, gen, dest);
2203 * Depending on whether the inode was already processed or not, use
2204 * send_root or parent_root for ref lookup.
2206 if (ino < sctx->send_progress)
2207 ret = get_first_ref(sctx->send_root, ino,
2208 parent_ino, parent_gen, dest);
2210 ret = get_first_ref(sctx->parent_root, ino,
2211 parent_ino, parent_gen, dest);
2216 * Check if the ref was overwritten by an inode's ref that was processed
2217 * earlier. If yes, treat as orphan and return 1.
2219 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2220 dest->start, dest->end - dest->start);
2224 fs_path_reset(dest);
2225 ret = gen_unique_name(sctx, ino, gen, dest);
2233 * Store the result of the lookup in the name cache.
2235 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2243 nce->parent_ino = *parent_ino;
2244 nce->parent_gen = *parent_gen;
2245 nce->name_len = fs_path_len(dest);
2247 strcpy(nce->name, dest->start);
2249 if (ino < sctx->send_progress)
2250 nce->need_later_update = 0;
2252 nce->need_later_update = 1;
2254 nce_ret = name_cache_insert(sctx, nce);
2257 name_cache_clean_unused(sctx);
2264 * Magic happens here. This function returns the first ref to an inode as it
2265 * would look like while receiving the stream at this point in time.
2266 * We walk the path up to the root. For every inode in between, we check if it
2267 * was already processed/sent. If yes, we continue with the parent as found
2268 * in send_root. If not, we continue with the parent as found in parent_root.
2269 * If we encounter an inode that was deleted at this point in time, we use the
2270 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2271 * that were not created yet and overwritten inodes/refs.
2273 * When do we have orphan inodes:
2274 * 1. When an inode is freshly created and thus no valid refs are available yet
2275 * 2. When a directory lost all it's refs (deleted) but still has dir items
2276 * inside which were not processed yet (pending for move/delete). If anyone
2277 * tried to get the path to the dir items, it would get a path inside that
2279 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2280 * of an unprocessed inode. If in that case the first ref would be
2281 * overwritten, the overwritten inode gets "orphanized". Later when we
2282 * process this overwritten inode, it is restored at a new place by moving
2285 * sctx->send_progress tells this function at which point in time receiving
2288 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2289 struct fs_path *dest)
2292 struct fs_path *name = NULL;
2293 u64 parent_inode = 0;
2297 name = fs_path_alloc();
2304 fs_path_reset(dest);
2306 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2307 struct waiting_dir_move *wdm;
2309 fs_path_reset(name);
2311 if (is_waiting_for_rm(sctx, ino, gen)) {
2312 ret = gen_unique_name(sctx, ino, gen, name);
2315 ret = fs_path_add_path(dest, name);
2319 wdm = get_waiting_dir_move(sctx, ino);
2320 if (wdm && wdm->orphanized) {
2321 ret = gen_unique_name(sctx, ino, gen, name);
2324 ret = get_first_ref(sctx->parent_root, ino,
2325 &parent_inode, &parent_gen, name);
2327 ret = __get_cur_name_and_parent(sctx, ino, gen,
2337 ret = fs_path_add_path(dest, name);
2348 fs_path_unreverse(dest);
2353 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2355 static int send_subvol_begin(struct send_ctx *sctx)
2358 struct btrfs_root *send_root = sctx->send_root;
2359 struct btrfs_root *parent_root = sctx->parent_root;
2360 struct btrfs_path *path;
2361 struct btrfs_key key;
2362 struct btrfs_root_ref *ref;
2363 struct extent_buffer *leaf;
2367 path = btrfs_alloc_path();
2371 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2373 btrfs_free_path(path);
2377 key.objectid = send_root->root_key.objectid;
2378 key.type = BTRFS_ROOT_BACKREF_KEY;
2381 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2390 leaf = path->nodes[0];
2391 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2392 if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2393 key.objectid != send_root->root_key.objectid) {
2397 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2398 namelen = btrfs_root_ref_name_len(leaf, ref);
2399 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2400 btrfs_release_path(path);
2403 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2407 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2412 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2414 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2415 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2416 sctx->send_root->root_item.received_uuid);
2418 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2419 sctx->send_root->root_item.uuid);
2421 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2422 le64_to_cpu(sctx->send_root->root_item.ctransid));
2424 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2425 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2426 parent_root->root_item.received_uuid);
2428 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2429 parent_root->root_item.uuid);
2430 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2431 le64_to_cpu(sctx->parent_root->root_item.ctransid));
2434 ret = send_cmd(sctx);
2438 btrfs_free_path(path);
2443 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2445 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2449 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2451 p = fs_path_alloc();
2455 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2459 ret = get_cur_path(sctx, ino, gen, p);
2462 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2463 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2465 ret = send_cmd(sctx);
2473 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2475 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2479 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2481 p = fs_path_alloc();
2485 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2489 ret = get_cur_path(sctx, ino, gen, p);
2492 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2493 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2495 ret = send_cmd(sctx);
2503 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2505 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2509 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2512 p = fs_path_alloc();
2516 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2520 ret = get_cur_path(sctx, ino, gen, p);
2523 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2524 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2525 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2527 ret = send_cmd(sctx);
2535 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2537 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2539 struct fs_path *p = NULL;
2540 struct btrfs_inode_item *ii;
2541 struct btrfs_path *path = NULL;
2542 struct extent_buffer *eb;
2543 struct btrfs_key key;
2546 btrfs_debug(fs_info, "send_utimes %llu", ino);
2548 p = fs_path_alloc();
2552 path = alloc_path_for_send();
2559 key.type = BTRFS_INODE_ITEM_KEY;
2561 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2567 eb = path->nodes[0];
2568 slot = path->slots[0];
2569 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2571 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2575 ret = get_cur_path(sctx, ino, gen, p);
2578 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2579 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2580 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2581 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2582 /* TODO Add otime support when the otime patches get into upstream */
2584 ret = send_cmd(sctx);
2589 btrfs_free_path(path);
2594 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2595 * a valid path yet because we did not process the refs yet. So, the inode
2596 * is created as orphan.
2598 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2600 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2608 btrfs_debug(fs_info, "send_create_inode %llu", ino);
2610 p = fs_path_alloc();
2614 if (ino != sctx->cur_ino) {
2615 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2620 gen = sctx->cur_inode_gen;
2621 mode = sctx->cur_inode_mode;
2622 rdev = sctx->cur_inode_rdev;
2625 if (S_ISREG(mode)) {
2626 cmd = BTRFS_SEND_C_MKFILE;
2627 } else if (S_ISDIR(mode)) {
2628 cmd = BTRFS_SEND_C_MKDIR;
2629 } else if (S_ISLNK(mode)) {
2630 cmd = BTRFS_SEND_C_SYMLINK;
2631 } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2632 cmd = BTRFS_SEND_C_MKNOD;
2633 } else if (S_ISFIFO(mode)) {
2634 cmd = BTRFS_SEND_C_MKFIFO;
2635 } else if (S_ISSOCK(mode)) {
2636 cmd = BTRFS_SEND_C_MKSOCK;
2638 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2639 (int)(mode & S_IFMT));
2644 ret = begin_cmd(sctx, cmd);
2648 ret = gen_unique_name(sctx, ino, gen, p);
2652 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2653 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2655 if (S_ISLNK(mode)) {
2657 ret = read_symlink(sctx->send_root, ino, p);
2660 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2661 } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2662 S_ISFIFO(mode) || S_ISSOCK(mode)) {
2663 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2664 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2667 ret = send_cmd(sctx);
2679 * We need some special handling for inodes that get processed before the parent
2680 * directory got created. See process_recorded_refs for details.
2681 * This function does the check if we already created the dir out of order.
2683 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2686 struct btrfs_path *path = NULL;
2687 struct btrfs_key key;
2688 struct btrfs_key found_key;
2689 struct btrfs_key di_key;
2690 struct extent_buffer *eb;
2691 struct btrfs_dir_item *di;
2694 path = alloc_path_for_send();
2701 key.type = BTRFS_DIR_INDEX_KEY;
2703 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2708 eb = path->nodes[0];
2709 slot = path->slots[0];
2710 if (slot >= btrfs_header_nritems(eb)) {
2711 ret = btrfs_next_leaf(sctx->send_root, path);
2714 } else if (ret > 0) {
2721 btrfs_item_key_to_cpu(eb, &found_key, slot);
2722 if (found_key.objectid != key.objectid ||
2723 found_key.type != key.type) {
2728 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2729 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2731 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2732 di_key.objectid < sctx->send_progress) {
2741 btrfs_free_path(path);
2746 * Only creates the inode if it is:
2747 * 1. Not a directory
2748 * 2. Or a directory which was not created already due to out of order
2749 * directories. See did_create_dir and process_recorded_refs for details.
2751 static int send_create_inode_if_needed(struct send_ctx *sctx)
2755 if (S_ISDIR(sctx->cur_inode_mode)) {
2756 ret = did_create_dir(sctx, sctx->cur_ino);
2765 ret = send_create_inode(sctx, sctx->cur_ino);
2773 struct recorded_ref {
2774 struct list_head list;
2776 struct fs_path *full_path;
2782 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
2784 ref->full_path = path;
2785 ref->name = (char *)kbasename(ref->full_path->start);
2786 ref->name_len = ref->full_path->end - ref->name;
2790 * We need to process new refs before deleted refs, but compare_tree gives us
2791 * everything mixed. So we first record all refs and later process them.
2792 * This function is a helper to record one ref.
2794 static int __record_ref(struct list_head *head, u64 dir,
2795 u64 dir_gen, struct fs_path *path)
2797 struct recorded_ref *ref;
2799 ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2804 ref->dir_gen = dir_gen;
2805 set_ref_path(ref, path);
2806 list_add_tail(&ref->list, head);
2810 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2812 struct recorded_ref *new;
2814 new = kmalloc(sizeof(*ref), GFP_KERNEL);
2818 new->dir = ref->dir;
2819 new->dir_gen = ref->dir_gen;
2820 new->full_path = NULL;
2821 INIT_LIST_HEAD(&new->list);
2822 list_add_tail(&new->list, list);
2826 static void __free_recorded_refs(struct list_head *head)
2828 struct recorded_ref *cur;
2830 while (!list_empty(head)) {
2831 cur = list_entry(head->next, struct recorded_ref, list);
2832 fs_path_free(cur->full_path);
2833 list_del(&cur->list);
2838 static void free_recorded_refs(struct send_ctx *sctx)
2840 __free_recorded_refs(&sctx->new_refs);
2841 __free_recorded_refs(&sctx->deleted_refs);
2845 * Renames/moves a file/dir to its orphan name. Used when the first
2846 * ref of an unprocessed inode gets overwritten and for all non empty
2849 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2850 struct fs_path *path)
2853 struct fs_path *orphan;
2855 orphan = fs_path_alloc();
2859 ret = gen_unique_name(sctx, ino, gen, orphan);
2863 ret = send_rename(sctx, path, orphan);
2866 fs_path_free(orphan);
2870 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
2871 u64 dir_ino, u64 dir_gen)
2873 struct rb_node **p = &sctx->orphan_dirs.rb_node;
2874 struct rb_node *parent = NULL;
2875 struct orphan_dir_info *entry, *odi;
2879 entry = rb_entry(parent, struct orphan_dir_info, node);
2880 if (dir_ino < entry->ino)
2882 else if (dir_ino > entry->ino)
2883 p = &(*p)->rb_right;
2884 else if (dir_gen < entry->gen)
2886 else if (dir_gen > entry->gen)
2887 p = &(*p)->rb_right;
2892 odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2894 return ERR_PTR(-ENOMEM);
2897 odi->last_dir_index_offset = 0;
2899 rb_link_node(&odi->node, parent, p);
2900 rb_insert_color(&odi->node, &sctx->orphan_dirs);
2904 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
2905 u64 dir_ino, u64 gen)
2907 struct rb_node *n = sctx->orphan_dirs.rb_node;
2908 struct orphan_dir_info *entry;
2911 entry = rb_entry(n, struct orphan_dir_info, node);
2912 if (dir_ino < entry->ino)
2914 else if (dir_ino > entry->ino)
2916 else if (gen < entry->gen)
2918 else if (gen > entry->gen)
2926 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
2928 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
2933 static void free_orphan_dir_info(struct send_ctx *sctx,
2934 struct orphan_dir_info *odi)
2938 rb_erase(&odi->node, &sctx->orphan_dirs);
2943 * Returns 1 if a directory can be removed at this point in time.
2944 * We check this by iterating all dir items and checking if the inode behind
2945 * the dir item was already processed.
2947 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2951 struct btrfs_root *root = sctx->parent_root;
2952 struct btrfs_path *path;
2953 struct btrfs_key key;
2954 struct btrfs_key found_key;
2955 struct btrfs_key loc;
2956 struct btrfs_dir_item *di;
2957 struct orphan_dir_info *odi = NULL;
2960 * Don't try to rmdir the top/root subvolume dir.
2962 if (dir == BTRFS_FIRST_FREE_OBJECTID)
2965 path = alloc_path_for_send();
2970 key.type = BTRFS_DIR_INDEX_KEY;
2973 odi = get_orphan_dir_info(sctx, dir, dir_gen);
2975 key.offset = odi->last_dir_index_offset;
2977 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2982 struct waiting_dir_move *dm;
2984 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2985 ret = btrfs_next_leaf(root, path);
2992 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2994 if (found_key.objectid != key.objectid ||
2995 found_key.type != key.type)
2998 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2999 struct btrfs_dir_item);
3000 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3002 dm = get_waiting_dir_move(sctx, loc.objectid);
3004 odi = add_orphan_dir_info(sctx, dir, dir_gen);
3010 odi->last_dir_index_offset = found_key.offset;
3011 dm->rmdir_ino = dir;
3012 dm->rmdir_gen = dir_gen;
3017 if (loc.objectid > send_progress) {
3018 odi = add_orphan_dir_info(sctx, dir, dir_gen);
3024 odi->last_dir_index_offset = found_key.offset;
3031 free_orphan_dir_info(sctx, odi);
3036 btrfs_free_path(path);
3040 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3042 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3044 return entry != NULL;
3047 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3049 struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3050 struct rb_node *parent = NULL;
3051 struct waiting_dir_move *entry, *dm;
3053 dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3059 dm->orphanized = orphanized;
3063 entry = rb_entry(parent, struct waiting_dir_move, node);
3064 if (ino < entry->ino) {
3066 } else if (ino > entry->ino) {
3067 p = &(*p)->rb_right;
3074 rb_link_node(&dm->node, parent, p);
3075 rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3079 static struct waiting_dir_move *
3080 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3082 struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3083 struct waiting_dir_move *entry;
3086 entry = rb_entry(n, struct waiting_dir_move, node);
3087 if (ino < entry->ino)
3089 else if (ino > entry->ino)
3097 static void free_waiting_dir_move(struct send_ctx *sctx,
3098 struct waiting_dir_move *dm)
3102 rb_erase(&dm->node, &sctx->waiting_dir_moves);
3106 static int add_pending_dir_move(struct send_ctx *sctx,
3110 struct list_head *new_refs,
3111 struct list_head *deleted_refs,
3112 const bool is_orphan)
3114 struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3115 struct rb_node *parent = NULL;
3116 struct pending_dir_move *entry = NULL, *pm;
3117 struct recorded_ref *cur;
3121 pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3124 pm->parent_ino = parent_ino;
3127 INIT_LIST_HEAD(&pm->list);
3128 INIT_LIST_HEAD(&pm->update_refs);
3129 RB_CLEAR_NODE(&pm->node);
3133 entry = rb_entry(parent, struct pending_dir_move, node);
3134 if (parent_ino < entry->parent_ino) {
3136 } else if (parent_ino > entry->parent_ino) {
3137 p = &(*p)->rb_right;
3144 list_for_each_entry(cur, deleted_refs, list) {
3145 ret = dup_ref(cur, &pm->update_refs);
3149 list_for_each_entry(cur, new_refs, list) {
3150 ret = dup_ref(cur, &pm->update_refs);
3155 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3160 list_add_tail(&pm->list, &entry->list);
3162 rb_link_node(&pm->node, parent, p);
3163 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3168 __free_recorded_refs(&pm->update_refs);
3174 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3177 struct rb_node *n = sctx->pending_dir_moves.rb_node;
3178 struct pending_dir_move *entry;
3181 entry = rb_entry(n, struct pending_dir_move, node);
3182 if (parent_ino < entry->parent_ino)
3184 else if (parent_ino > entry->parent_ino)
3192 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3193 u64 ino, u64 gen, u64 *ancestor_ino)
3196 u64 parent_inode = 0;
3198 u64 start_ino = ino;
3201 while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3202 fs_path_reset(name);
3204 if (is_waiting_for_rm(sctx, ino, gen))
3206 if (is_waiting_for_move(sctx, ino)) {
3207 if (*ancestor_ino == 0)
3208 *ancestor_ino = ino;
3209 ret = get_first_ref(sctx->parent_root, ino,
3210 &parent_inode, &parent_gen, name);
3212 ret = __get_cur_name_and_parent(sctx, ino, gen,
3222 if (parent_inode == start_ino) {
3224 if (*ancestor_ino == 0)
3225 *ancestor_ino = ino;
3234 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3236 struct fs_path *from_path = NULL;
3237 struct fs_path *to_path = NULL;
3238 struct fs_path *name = NULL;
3239 u64 orig_progress = sctx->send_progress;
3240 struct recorded_ref *cur;
3241 u64 parent_ino, parent_gen;
3242 struct waiting_dir_move *dm = NULL;
3249 name = fs_path_alloc();
3250 from_path = fs_path_alloc();
3251 if (!name || !from_path) {
3256 dm = get_waiting_dir_move(sctx, pm->ino);
3258 rmdir_ino = dm->rmdir_ino;
3259 rmdir_gen = dm->rmdir_gen;
3260 is_orphan = dm->orphanized;
3261 free_waiting_dir_move(sctx, dm);
3264 ret = gen_unique_name(sctx, pm->ino,
3265 pm->gen, from_path);
3267 ret = get_first_ref(sctx->parent_root, pm->ino,
3268 &parent_ino, &parent_gen, name);
3271 ret = get_cur_path(sctx, parent_ino, parent_gen,
3275 ret = fs_path_add_path(from_path, name);
3280 sctx->send_progress = sctx->cur_ino + 1;
3281 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3285 LIST_HEAD(deleted_refs);
3286 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3287 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3288 &pm->update_refs, &deleted_refs,
3293 dm = get_waiting_dir_move(sctx, pm->ino);
3295 dm->rmdir_ino = rmdir_ino;
3296 dm->rmdir_gen = rmdir_gen;
3300 fs_path_reset(name);
3303 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3307 ret = send_rename(sctx, from_path, to_path);
3312 struct orphan_dir_info *odi;
3315 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3317 /* already deleted */
3322 ret = can_rmdir(sctx, rmdir_ino, gen, sctx->cur_ino);
3328 name = fs_path_alloc();
3333 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3336 ret = send_rmdir(sctx, name);
3342 ret = send_utimes(sctx, pm->ino, pm->gen);
3347 * After rename/move, need to update the utimes of both new parent(s)
3348 * and old parent(s).
3350 list_for_each_entry(cur, &pm->update_refs, list) {
3352 * The parent inode might have been deleted in the send snapshot
3354 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3355 NULL, NULL, NULL, NULL, NULL);
3356 if (ret == -ENOENT) {
3363 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3370 fs_path_free(from_path);
3371 fs_path_free(to_path);
3372 sctx->send_progress = orig_progress;
3377 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3379 if (!list_empty(&m->list))
3381 if (!RB_EMPTY_NODE(&m->node))
3382 rb_erase(&m->node, &sctx->pending_dir_moves);
3383 __free_recorded_refs(&m->update_refs);
3387 static void tail_append_pending_moves(struct send_ctx *sctx,
3388 struct pending_dir_move *moves,
3389 struct list_head *stack)
3391 if (list_empty(&moves->list)) {
3392 list_add_tail(&moves->list, stack);
3395 list_splice_init(&moves->list, &list);
3396 list_add_tail(&moves->list, stack);
3397 list_splice_tail(&list, stack);
3399 if (!RB_EMPTY_NODE(&moves->node)) {
3400 rb_erase(&moves->node, &sctx->pending_dir_moves);
3401 RB_CLEAR_NODE(&moves->node);
3405 static int apply_children_dir_moves(struct send_ctx *sctx)
3407 struct pending_dir_move *pm;
3408 struct list_head stack;
3409 u64 parent_ino = sctx->cur_ino;
3412 pm = get_pending_dir_moves(sctx, parent_ino);
3416 INIT_LIST_HEAD(&stack);
3417 tail_append_pending_moves(sctx, pm, &stack);
3419 while (!list_empty(&stack)) {
3420 pm = list_first_entry(&stack, struct pending_dir_move, list);
3421 parent_ino = pm->ino;
3422 ret = apply_dir_move(sctx, pm);
3423 free_pending_move(sctx, pm);
3426 pm = get_pending_dir_moves(sctx, parent_ino);
3428 tail_append_pending_moves(sctx, pm, &stack);
3433 while (!list_empty(&stack)) {
3434 pm = list_first_entry(&stack, struct pending_dir_move, list);
3435 free_pending_move(sctx, pm);
3441 * We might need to delay a directory rename even when no ancestor directory
3442 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3443 * renamed. This happens when we rename a directory to the old name (the name
3444 * in the parent root) of some other unrelated directory that got its rename
3445 * delayed due to some ancestor with higher number that got renamed.
3451 * |---- a/ (ino 257)
3452 * | |---- file (ino 260)
3454 * |---- b/ (ino 258)
3455 * |---- c/ (ino 259)
3459 * |---- a/ (ino 258)
3460 * |---- x/ (ino 259)
3461 * |---- y/ (ino 257)
3462 * |----- file (ino 260)
3464 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3465 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3466 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3469 * 1 - rename 259 from 'c' to 'x'
3470 * 2 - rename 257 from 'a' to 'x/y'
3471 * 3 - rename 258 from 'b' to 'a'
3473 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3474 * be done right away and < 0 on error.
3476 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3477 struct recorded_ref *parent_ref,
3478 const bool is_orphan)
3480 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3481 struct btrfs_path *path;
3482 struct btrfs_key key;
3483 struct btrfs_key di_key;
3484 struct btrfs_dir_item *di;
3488 struct waiting_dir_move *wdm;
3490 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3493 path = alloc_path_for_send();
3497 key.objectid = parent_ref->dir;
3498 key.type = BTRFS_DIR_ITEM_KEY;
3499 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3501 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3504 } else if (ret > 0) {
3509 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3510 parent_ref->name_len);
3516 * di_key.objectid has the number of the inode that has a dentry in the
3517 * parent directory with the same name that sctx->cur_ino is being
3518 * renamed to. We need to check if that inode is in the send root as
3519 * well and if it is currently marked as an inode with a pending rename,
3520 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3521 * that it happens after that other inode is renamed.
3523 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3524 if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3529 ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3530 &left_gen, NULL, NULL, NULL, NULL);
3533 ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3534 &right_gen, NULL, NULL, NULL, NULL);
3541 /* Different inode, no need to delay the rename of sctx->cur_ino */
3542 if (right_gen != left_gen) {
3547 wdm = get_waiting_dir_move(sctx, di_key.objectid);
3548 if (wdm && !wdm->orphanized) {
3549 ret = add_pending_dir_move(sctx,
3551 sctx->cur_inode_gen,
3554 &sctx->deleted_refs,
3560 btrfs_free_path(path);
3565 * Check if inode ino2, or any of its ancestors, is inode ino1.
3566 * Return 1 if true, 0 if false and < 0 on error.
3568 static int check_ino_in_path(struct btrfs_root *root,
3573 struct fs_path *fs_path)
3578 return ino1_gen == ino2_gen;
3580 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3585 fs_path_reset(fs_path);
3586 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3590 return parent_gen == ino1_gen;
3597 * Check if ino ino1 is an ancestor of inode ino2 in the given root for any
3598 * possible path (in case ino2 is not a directory and has multiple hard links).
3599 * Return 1 if true, 0 if false and < 0 on error.
3601 static int is_ancestor(struct btrfs_root *root,
3605 struct fs_path *fs_path)
3607 bool free_fs_path = false;
3609 struct btrfs_path *path = NULL;
3610 struct btrfs_key key;
3613 fs_path = fs_path_alloc();
3616 free_fs_path = true;
3619 path = alloc_path_for_send();
3625 key.objectid = ino2;
3626 key.type = BTRFS_INODE_REF_KEY;
3629 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3634 struct extent_buffer *leaf = path->nodes[0];
3635 int slot = path->slots[0];
3639 if (slot >= btrfs_header_nritems(leaf)) {
3640 ret = btrfs_next_leaf(root, path);
3648 btrfs_item_key_to_cpu(leaf, &key, slot);
3649 if (key.objectid != ino2)
3651 if (key.type != BTRFS_INODE_REF_KEY &&
3652 key.type != BTRFS_INODE_EXTREF_KEY)
3655 item_size = btrfs_item_size_nr(leaf, slot);
3656 while (cur_offset < item_size) {
3660 if (key.type == BTRFS_INODE_EXTREF_KEY) {
3662 struct btrfs_inode_extref *extref;
3664 ptr = btrfs_item_ptr_offset(leaf, slot);
3665 extref = (struct btrfs_inode_extref *)
3667 parent = btrfs_inode_extref_parent(leaf,
3669 cur_offset += sizeof(*extref);
3670 cur_offset += btrfs_inode_extref_name_len(leaf,
3673 parent = key.offset;
3674 cur_offset = item_size;
3677 ret = get_inode_info(root, parent, NULL, &parent_gen,
3678 NULL, NULL, NULL, NULL);
3681 ret = check_ino_in_path(root, ino1, ino1_gen,
3682 parent, parent_gen, fs_path);
3690 btrfs_free_path(path);
3692 fs_path_free(fs_path);
3696 static int wait_for_parent_move(struct send_ctx *sctx,
3697 struct recorded_ref *parent_ref,
3698 const bool is_orphan)
3701 u64 ino = parent_ref->dir;
3702 u64 ino_gen = parent_ref->dir_gen;
3703 u64 parent_ino_before, parent_ino_after;
3704 struct fs_path *path_before = NULL;
3705 struct fs_path *path_after = NULL;
3708 path_after = fs_path_alloc();
3709 path_before = fs_path_alloc();
3710 if (!path_after || !path_before) {
3716 * Our current directory inode may not yet be renamed/moved because some
3717 * ancestor (immediate or not) has to be renamed/moved first. So find if
3718 * such ancestor exists and make sure our own rename/move happens after
3719 * that ancestor is processed to avoid path build infinite loops (done
3720 * at get_cur_path()).
3722 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3723 u64 parent_ino_after_gen;
3725 if (is_waiting_for_move(sctx, ino)) {
3727 * If the current inode is an ancestor of ino in the
3728 * parent root, we need to delay the rename of the
3729 * current inode, otherwise don't delayed the rename
3730 * because we can end up with a circular dependency
3731 * of renames, resulting in some directories never
3732 * getting the respective rename operations issued in
3733 * the send stream or getting into infinite path build
3736 ret = is_ancestor(sctx->parent_root,
3737 sctx->cur_ino, sctx->cur_inode_gen,
3743 fs_path_reset(path_before);
3744 fs_path_reset(path_after);
3746 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3747 &parent_ino_after_gen, path_after);
3750 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3752 if (ret < 0 && ret != -ENOENT) {
3754 } else if (ret == -ENOENT) {
3759 len1 = fs_path_len(path_before);
3760 len2 = fs_path_len(path_after);
3761 if (ino > sctx->cur_ino &&
3762 (parent_ino_before != parent_ino_after || len1 != len2 ||
3763 memcmp(path_before->start, path_after->start, len1))) {
3766 ret = get_inode_info(sctx->parent_root, ino, NULL,
3767 &parent_ino_gen, NULL, NULL, NULL,
3771 if (ino_gen == parent_ino_gen) {
3776 ino = parent_ino_after;
3777 ino_gen = parent_ino_after_gen;
3781 fs_path_free(path_before);
3782 fs_path_free(path_after);
3785 ret = add_pending_dir_move(sctx,
3787 sctx->cur_inode_gen,
3790 &sctx->deleted_refs,
3799 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3802 struct fs_path *new_path;
3805 * Our reference's name member points to its full_path member string, so
3806 * we use here a new path.
3808 new_path = fs_path_alloc();
3812 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
3814 fs_path_free(new_path);
3817 ret = fs_path_add(new_path, ref->name, ref->name_len);
3819 fs_path_free(new_path);
3823 fs_path_free(ref->full_path);
3824 set_ref_path(ref, new_path);
3830 * When processing the new references for an inode we may orphanize an existing
3831 * directory inode because its old name conflicts with one of the new references
3832 * of the current inode. Later, when processing another new reference of our
3833 * inode, we might need to orphanize another inode, but the path we have in the
3834 * reference reflects the pre-orphanization name of the directory we previously
3835 * orphanized. For example:
3837 * parent snapshot looks like:
3840 * |----- f1 (ino 257)
3841 * |----- f2 (ino 258)
3842 * |----- d1/ (ino 259)
3843 * |----- d2/ (ino 260)
3845 * send snapshot looks like:
3848 * |----- d1 (ino 258)
3849 * |----- f2/ (ino 259)
3850 * |----- f2_link/ (ino 260)
3851 * | |----- f1 (ino 257)
3853 * |----- d2 (ino 258)
3855 * When processing inode 257 we compute the name for inode 259 as "d1", and we
3856 * cache it in the name cache. Later when we start processing inode 258, when
3857 * collecting all its new references we set a full path of "d1/d2" for its new
3858 * reference with name "d2". When we start processing the new references we
3859 * start by processing the new reference with name "d1", and this results in
3860 * orphanizing inode 259, since its old reference causes a conflict. Then we
3861 * move on the next new reference, with name "d2", and we find out we must
3862 * orphanize inode 260, as its old reference conflicts with ours - but for the
3863 * orphanization we use a source path corresponding to the path we stored in the
3864 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
3865 * receiver fail since the path component "d1/" no longer exists, it was renamed
3866 * to "o259-6-0/" when processing the previous new reference. So in this case we
3867 * must recompute the path in the new reference and use it for the new
3868 * orphanization operation.
3870 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3875 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
3879 fs_path_reset(ref->full_path);
3880 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
3884 ret = fs_path_add(ref->full_path, name, ref->name_len);
3888 /* Update the reference's base name pointer. */
3889 set_ref_path(ref, ref->full_path);
3896 * This does all the move/link/unlink/rmdir magic.
3898 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3900 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3902 struct recorded_ref *cur;
3903 struct recorded_ref *cur2;
3904 struct list_head check_dirs;
3905 struct fs_path *valid_path = NULL;
3909 int did_overwrite = 0;
3911 u64 last_dir_ino_rm = 0;
3912 bool can_rename = true;
3913 bool orphanized_dir = false;
3914 bool orphanized_ancestor = false;
3916 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3919 * This should never happen as the root dir always has the same ref
3920 * which is always '..'
3922 BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3923 INIT_LIST_HEAD(&check_dirs);
3925 valid_path = fs_path_alloc();
3932 * First, check if the first ref of the current inode was overwritten
3933 * before. If yes, we know that the current inode was already orphanized
3934 * and thus use the orphan name. If not, we can use get_cur_path to
3935 * get the path of the first ref as it would like while receiving at
3936 * this point in time.
3937 * New inodes are always orphan at the beginning, so force to use the
3938 * orphan name in this case.
3939 * The first ref is stored in valid_path and will be updated if it
3940 * gets moved around.
3942 if (!sctx->cur_inode_new) {
3943 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3944 sctx->cur_inode_gen);
3950 if (sctx->cur_inode_new || did_overwrite) {
3951 ret = gen_unique_name(sctx, sctx->cur_ino,
3952 sctx->cur_inode_gen, valid_path);
3957 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3964 * Before doing any rename and link operations, do a first pass on the
3965 * new references to orphanize any unprocessed inodes that may have a
3966 * reference that conflicts with one of the new references of the current
3967 * inode. This needs to happen first because a new reference may conflict
3968 * with the old reference of a parent directory, so we must make sure
3969 * that the path used for link and rename commands don't use an
3970 * orphanized name when an ancestor was not yet orphanized.
3977 * |----- testdir/ (ino 259)
3978 * | |----- a (ino 257)
3980 * |----- b (ino 258)
3985 * |----- testdir_2/ (ino 259)
3986 * | |----- a (ino 260)
3988 * |----- testdir (ino 257)
3989 * |----- b (ino 257)
3990 * |----- b2 (ino 258)
3992 * Processing the new reference for inode 257 with name "b" may happen
3993 * before processing the new reference with name "testdir". If so, we
3994 * must make sure that by the time we send a link command to create the
3995 * hard link "b", inode 259 was already orphanized, since the generated
3996 * path in "valid_path" already contains the orphanized name for 259.
3997 * We are processing inode 257, so only later when processing 259 we do
3998 * the rename operation to change its temporary (orphanized) name to
4001 list_for_each_entry(cur, &sctx->new_refs, list) {
4002 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4005 if (ret == inode_state_will_create)
4009 * Check if this new ref would overwrite the first ref of another
4010 * unprocessed inode. If yes, orphanize the overwritten inode.
4011 * If we find an overwritten ref that is not the first ref,
4014 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4015 cur->name, cur->name_len,
4016 &ow_inode, &ow_gen, &ow_mode);
4020 ret = is_first_ref(sctx->parent_root,
4021 ow_inode, cur->dir, cur->name,
4026 struct name_cache_entry *nce;
4027 struct waiting_dir_move *wdm;
4029 if (orphanized_dir) {
4030 ret = refresh_ref_path(sctx, cur);
4035 ret = orphanize_inode(sctx, ow_inode, ow_gen,
4039 if (S_ISDIR(ow_mode))
4040 orphanized_dir = true;
4043 * If ow_inode has its rename operation delayed
4044 * make sure that its orphanized name is used in
4045 * the source path when performing its rename
4048 if (is_waiting_for_move(sctx, ow_inode)) {
4049 wdm = get_waiting_dir_move(sctx,
4052 wdm->orphanized = true;
4056 * Make sure we clear our orphanized inode's
4057 * name from the name cache. This is because the
4058 * inode ow_inode might be an ancestor of some
4059 * other inode that will be orphanized as well
4060 * later and has an inode number greater than
4061 * sctx->send_progress. We need to prevent
4062 * future name lookups from using the old name
4063 * and get instead the orphan name.
4065 nce = name_cache_search(sctx, ow_inode, ow_gen);
4067 name_cache_delete(sctx, nce);
4072 * ow_inode might currently be an ancestor of
4073 * cur_ino, therefore compute valid_path (the
4074 * current path of cur_ino) again because it
4075 * might contain the pre-orphanization name of
4076 * ow_inode, which is no longer valid.
4078 ret = is_ancestor(sctx->parent_root,
4080 sctx->cur_ino, NULL);
4082 orphanized_ancestor = true;
4083 fs_path_reset(valid_path);
4084 ret = get_cur_path(sctx, sctx->cur_ino,
4085 sctx->cur_inode_gen,
4092 * If we previously orphanized a directory that
4093 * collided with a new reference that we already
4094 * processed, recompute the current path because
4095 * that directory may be part of the path.
4097 if (orphanized_dir) {
4098 ret = refresh_ref_path(sctx, cur);
4102 ret = send_unlink(sctx, cur->full_path);
4110 list_for_each_entry(cur, &sctx->new_refs, list) {
4112 * We may have refs where the parent directory does not exist
4113 * yet. This happens if the parent directories inum is higher
4114 * than the current inum. To handle this case, we create the
4115 * parent directory out of order. But we need to check if this
4116 * did already happen before due to other refs in the same dir.
4118 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4121 if (ret == inode_state_will_create) {
4124 * First check if any of the current inodes refs did
4125 * already create the dir.
4127 list_for_each_entry(cur2, &sctx->new_refs, list) {
4130 if (cur2->dir == cur->dir) {
4137 * If that did not happen, check if a previous inode
4138 * did already create the dir.
4141 ret = did_create_dir(sctx, cur->dir);
4145 ret = send_create_inode(sctx, cur->dir);
4151 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4152 ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4161 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4163 ret = wait_for_parent_move(sctx, cur, is_orphan);
4173 * link/move the ref to the new place. If we have an orphan
4174 * inode, move it and update valid_path. If not, link or move
4175 * it depending on the inode mode.
4177 if (is_orphan && can_rename) {
4178 ret = send_rename(sctx, valid_path, cur->full_path);
4182 ret = fs_path_copy(valid_path, cur->full_path);
4185 } else if (can_rename) {
4186 if (S_ISDIR(sctx->cur_inode_mode)) {
4188 * Dirs can't be linked, so move it. For moved
4189 * dirs, we always have one new and one deleted
4190 * ref. The deleted ref is ignored later.
4192 ret = send_rename(sctx, valid_path,
4195 ret = fs_path_copy(valid_path,
4201 * We might have previously orphanized an inode
4202 * which is an ancestor of our current inode,
4203 * so our reference's full path, which was
4204 * computed before any such orphanizations, must
4207 if (orphanized_dir) {
4208 ret = update_ref_path(sctx, cur);
4212 ret = send_link(sctx, cur->full_path,
4218 ret = dup_ref(cur, &check_dirs);
4223 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4225 * Check if we can already rmdir the directory. If not,
4226 * orphanize it. For every dir item inside that gets deleted
4227 * later, we do this check again and rmdir it then if possible.
4228 * See the use of check_dirs for more details.
4230 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4235 ret = send_rmdir(sctx, valid_path);
4238 } else if (!is_orphan) {
4239 ret = orphanize_inode(sctx, sctx->cur_ino,
4240 sctx->cur_inode_gen, valid_path);
4246 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4247 ret = dup_ref(cur, &check_dirs);
4251 } else if (S_ISDIR(sctx->cur_inode_mode) &&
4252 !list_empty(&sctx->deleted_refs)) {
4254 * We have a moved dir. Add the old parent to check_dirs
4256 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4258 ret = dup_ref(cur, &check_dirs);
4261 } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4263 * We have a non dir inode. Go through all deleted refs and
4264 * unlink them if they were not already overwritten by other
4267 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4268 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4269 sctx->cur_ino, sctx->cur_inode_gen,
4270 cur->name, cur->name_len);
4275 * If we orphanized any ancestor before, we need
4276 * to recompute the full path for deleted names,
4277 * since any such path was computed before we
4278 * processed any references and orphanized any
4281 if (orphanized_ancestor) {
4282 ret = update_ref_path(sctx, cur);
4286 ret = send_unlink(sctx, cur->full_path);
4290 ret = dup_ref(cur, &check_dirs);
4295 * If the inode is still orphan, unlink the orphan. This may
4296 * happen when a previous inode did overwrite the first ref
4297 * of this inode and no new refs were added for the current
4298 * inode. Unlinking does not mean that the inode is deleted in
4299 * all cases. There may still be links to this inode in other
4303 ret = send_unlink(sctx, valid_path);
4310 * We did collect all parent dirs where cur_inode was once located. We
4311 * now go through all these dirs and check if they are pending for
4312 * deletion and if it's finally possible to perform the rmdir now.
4313 * We also update the inode stats of the parent dirs here.
4315 list_for_each_entry(cur, &check_dirs, list) {
4317 * In case we had refs into dirs that were not processed yet,
4318 * we don't need to do the utime and rmdir logic for these dirs.
4319 * The dir will be processed later.
4321 if (cur->dir > sctx->cur_ino)
4324 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4328 if (ret == inode_state_did_create ||
4329 ret == inode_state_no_change) {
4330 /* TODO delayed utimes */
4331 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4334 } else if (ret == inode_state_did_delete &&
4335 cur->dir != last_dir_ino_rm) {
4336 ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4341 ret = get_cur_path(sctx, cur->dir,
4342 cur->dir_gen, valid_path);
4345 ret = send_rmdir(sctx, valid_path);
4348 last_dir_ino_rm = cur->dir;
4356 __free_recorded_refs(&check_dirs);
4357 free_recorded_refs(sctx);
4358 fs_path_free(valid_path);
4362 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name,
4363 void *ctx, struct list_head *refs)
4366 struct send_ctx *sctx = ctx;
4370 p = fs_path_alloc();
4374 ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4379 ret = get_cur_path(sctx, dir, gen, p);
4382 ret = fs_path_add_path(p, name);
4386 ret = __record_ref(refs, dir, gen, p);
4394 static int __record_new_ref(int num, u64 dir, int index,
4395 struct fs_path *name,
4398 struct send_ctx *sctx = ctx;
4399 return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs);
4403 static int __record_deleted_ref(int num, u64 dir, int index,
4404 struct fs_path *name,
4407 struct send_ctx *sctx = ctx;
4408 return record_ref(sctx->parent_root, dir, name, ctx,
4409 &sctx->deleted_refs);
4412 static int record_new_ref(struct send_ctx *sctx)
4416 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4417 sctx->cmp_key, 0, __record_new_ref, sctx);
4426 static int record_deleted_ref(struct send_ctx *sctx)
4430 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4431 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4440 struct find_ref_ctx {
4443 struct btrfs_root *root;
4444 struct fs_path *name;
4448 static int __find_iref(int num, u64 dir, int index,
4449 struct fs_path *name,
4452 struct find_ref_ctx *ctx = ctx_;
4456 if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4457 strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4459 * To avoid doing extra lookups we'll only do this if everything
4462 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4466 if (dir_gen != ctx->dir_gen)
4468 ctx->found_idx = num;
4474 static int find_iref(struct btrfs_root *root,
4475 struct btrfs_path *path,
4476 struct btrfs_key *key,
4477 u64 dir, u64 dir_gen, struct fs_path *name)
4480 struct find_ref_ctx ctx;
4484 ctx.dir_gen = dir_gen;
4488 ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4492 if (ctx.found_idx == -1)
4495 return ctx.found_idx;
4498 static int __record_changed_new_ref(int num, u64 dir, int index,
4499 struct fs_path *name,
4504 struct send_ctx *sctx = ctx;
4506 ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4511 ret = find_iref(sctx->parent_root, sctx->right_path,
4512 sctx->cmp_key, dir, dir_gen, name);
4514 ret = __record_new_ref(num, dir, index, name, sctx);
4521 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4522 struct fs_path *name,
4527 struct send_ctx *sctx = ctx;
4529 ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4534 ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4535 dir, dir_gen, name);
4537 ret = __record_deleted_ref(num, dir, index, name, sctx);
4544 static int record_changed_ref(struct send_ctx *sctx)
4548 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4549 sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4552 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4553 sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4563 * Record and process all refs at once. Needed when an inode changes the
4564 * generation number, which means that it was deleted and recreated.
4566 static int process_all_refs(struct send_ctx *sctx,
4567 enum btrfs_compare_tree_result cmd)
4570 struct btrfs_root *root;
4571 struct btrfs_path *path;
4572 struct btrfs_key key;
4573 struct btrfs_key found_key;
4574 struct extent_buffer *eb;
4576 iterate_inode_ref_t cb;
4577 int pending_move = 0;
4579 path = alloc_path_for_send();
4583 if (cmd == BTRFS_COMPARE_TREE_NEW) {
4584 root = sctx->send_root;
4585 cb = __record_new_ref;
4586 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4587 root = sctx->parent_root;
4588 cb = __record_deleted_ref;
4590 btrfs_err(sctx->send_root->fs_info,
4591 "Wrong command %d in process_all_refs", cmd);
4596 key.objectid = sctx->cmp_key->objectid;
4597 key.type = BTRFS_INODE_REF_KEY;
4599 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4604 eb = path->nodes[0];
4605 slot = path->slots[0];
4606 if (slot >= btrfs_header_nritems(eb)) {
4607 ret = btrfs_next_leaf(root, path);
4615 btrfs_item_key_to_cpu(eb, &found_key, slot);
4617 if (found_key.objectid != key.objectid ||
4618 (found_key.type != BTRFS_INODE_REF_KEY &&
4619 found_key.type != BTRFS_INODE_EXTREF_KEY))
4622 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4628 btrfs_release_path(path);
4631 * We don't actually care about pending_move as we are simply
4632 * re-creating this inode and will be rename'ing it into place once we
4633 * rename the parent directory.
4635 ret = process_recorded_refs(sctx, &pending_move);
4637 btrfs_free_path(path);
4641 static int send_set_xattr(struct send_ctx *sctx,
4642 struct fs_path *path,
4643 const char *name, int name_len,
4644 const char *data, int data_len)
4648 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4652 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4653 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4654 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4656 ret = send_cmd(sctx);
4663 static int send_remove_xattr(struct send_ctx *sctx,
4664 struct fs_path *path,
4665 const char *name, int name_len)
4669 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4673 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4674 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4676 ret = send_cmd(sctx);
4683 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4684 const char *name, int name_len,
4685 const char *data, int data_len,
4689 struct send_ctx *sctx = ctx;
4691 struct posix_acl_xattr_header dummy_acl;
4693 /* Capabilities are emitted by finish_inode_if_needed */
4694 if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4697 p = fs_path_alloc();
4702 * This hack is needed because empty acls are stored as zero byte
4703 * data in xattrs. Problem with that is, that receiving these zero byte
4704 * acls will fail later. To fix this, we send a dummy acl list that
4705 * only contains the version number and no entries.
4707 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4708 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4709 if (data_len == 0) {
4710 dummy_acl.a_version =
4711 cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4712 data = (char *)&dummy_acl;
4713 data_len = sizeof(dummy_acl);
4717 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4721 ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4728 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4729 const char *name, int name_len,
4730 const char *data, int data_len,
4734 struct send_ctx *sctx = ctx;
4737 p = fs_path_alloc();
4741 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4745 ret = send_remove_xattr(sctx, p, name, name_len);
4752 static int process_new_xattr(struct send_ctx *sctx)
4756 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4757 __process_new_xattr, sctx);
4762 static int process_deleted_xattr(struct send_ctx *sctx)
4764 return iterate_dir_item(sctx->parent_root, sctx->right_path,
4765 __process_deleted_xattr, sctx);
4768 struct find_xattr_ctx {
4776 static int __find_xattr(int num, struct btrfs_key *di_key,
4777 const char *name, int name_len,
4778 const char *data, int data_len,
4779 u8 type, void *vctx)
4781 struct find_xattr_ctx *ctx = vctx;
4783 if (name_len == ctx->name_len &&
4784 strncmp(name, ctx->name, name_len) == 0) {
4785 ctx->found_idx = num;
4786 ctx->found_data_len = data_len;
4787 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4788 if (!ctx->found_data)
4795 static int find_xattr(struct btrfs_root *root,
4796 struct btrfs_path *path,
4797 struct btrfs_key *key,
4798 const char *name, int name_len,
4799 char **data, int *data_len)
4802 struct find_xattr_ctx ctx;
4805 ctx.name_len = name_len;
4807 ctx.found_data = NULL;
4808 ctx.found_data_len = 0;
4810 ret = iterate_dir_item(root, path, __find_xattr, &ctx);
4814 if (ctx.found_idx == -1)
4817 *data = ctx.found_data;
4818 *data_len = ctx.found_data_len;
4820 kfree(ctx.found_data);
4822 return ctx.found_idx;
4826 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4827 const char *name, int name_len,
4828 const char *data, int data_len,
4832 struct send_ctx *sctx = ctx;
4833 char *found_data = NULL;
4834 int found_data_len = 0;
4836 ret = find_xattr(sctx->parent_root, sctx->right_path,
4837 sctx->cmp_key, name, name_len, &found_data,
4839 if (ret == -ENOENT) {
4840 ret = __process_new_xattr(num, di_key, name, name_len, data,
4841 data_len, type, ctx);
4842 } else if (ret >= 0) {
4843 if (data_len != found_data_len ||
4844 memcmp(data, found_data, data_len)) {
4845 ret = __process_new_xattr(num, di_key, name, name_len,
4846 data, data_len, type, ctx);
4856 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4857 const char *name, int name_len,
4858 const char *data, int data_len,
4862 struct send_ctx *sctx = ctx;
4864 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4865 name, name_len, NULL, NULL);
4867 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4868 data_len, type, ctx);
4875 static int process_changed_xattr(struct send_ctx *sctx)
4879 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4880 __process_changed_new_xattr, sctx);
4883 ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4884 __process_changed_deleted_xattr, sctx);
4890 static int process_all_new_xattrs(struct send_ctx *sctx)
4893 struct btrfs_root *root;
4894 struct btrfs_path *path;
4895 struct btrfs_key key;
4896 struct btrfs_key found_key;
4897 struct extent_buffer *eb;
4900 path = alloc_path_for_send();
4904 root = sctx->send_root;
4906 key.objectid = sctx->cmp_key->objectid;
4907 key.type = BTRFS_XATTR_ITEM_KEY;
4909 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4914 eb = path->nodes[0];
4915 slot = path->slots[0];
4916 if (slot >= btrfs_header_nritems(eb)) {
4917 ret = btrfs_next_leaf(root, path);
4920 } else if (ret > 0) {
4927 btrfs_item_key_to_cpu(eb, &found_key, slot);
4928 if (found_key.objectid != key.objectid ||
4929 found_key.type != key.type) {
4934 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
4942 btrfs_free_path(path);
4946 static inline u64 max_send_read_size(const struct send_ctx *sctx)
4948 return sctx->send_max_size - SZ_16K;
4951 static int put_data_header(struct send_ctx *sctx, u32 len)
4953 struct btrfs_tlv_header *hdr;
4955 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
4957 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
4958 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
4959 put_unaligned_le16(len, &hdr->tlv_len);
4960 sctx->send_size += sizeof(*hdr);
4964 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
4966 struct btrfs_root *root = sctx->send_root;
4967 struct btrfs_fs_info *fs_info = root->fs_info;
4968 struct inode *inode;
4971 pgoff_t index = offset >> PAGE_SHIFT;
4973 unsigned pg_offset = offset_in_page(offset);
4976 ret = put_data_header(sctx, len);
4980 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
4982 return PTR_ERR(inode);
4984 last_index = (offset + len - 1) >> PAGE_SHIFT;
4986 /* initial readahead */
4987 memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4988 file_ra_state_init(&sctx->ra, inode->i_mapping);
4990 while (index <= last_index) {
4991 unsigned cur_len = min_t(unsigned, len,
4992 PAGE_SIZE - pg_offset);
4994 page = find_lock_page(inode->i_mapping, index);
4996 page_cache_sync_readahead(inode->i_mapping, &sctx->ra,
4997 NULL, index, last_index + 1 - index);
4999 page = find_or_create_page(inode->i_mapping, index,
5007 if (PageReadahead(page)) {
5008 page_cache_async_readahead(inode->i_mapping, &sctx->ra,
5009 NULL, page, index, last_index + 1 - index);
5012 if (!PageUptodate(page)) {
5013 btrfs_readpage(NULL, page);
5015 if (!PageUptodate(page)) {
5018 "send: IO error at offset %llu for inode %llu root %llu",
5019 page_offset(page), sctx->cur_ino,
5020 sctx->send_root->root_key.objectid);
5028 memcpy(sctx->send_buf + sctx->send_size, addr + pg_offset,
5036 sctx->send_size += cur_len;
5043 * Read some bytes from the current inode/file and send a write command to
5046 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5048 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5052 p = fs_path_alloc();
5056 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5058 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5062 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5066 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5067 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5068 ret = put_file_data(sctx, offset, len);
5072 ret = send_cmd(sctx);
5081 * Send a clone command to user space.
5083 static int send_clone(struct send_ctx *sctx,
5084 u64 offset, u32 len,
5085 struct clone_root *clone_root)
5091 btrfs_debug(sctx->send_root->fs_info,
5092 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5093 offset, len, clone_root->root->root_key.objectid,
5094 clone_root->ino, clone_root->offset);
5096 p = fs_path_alloc();
5100 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5104 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5108 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5109 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5110 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5112 if (clone_root->root == sctx->send_root) {
5113 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
5114 &gen, NULL, NULL, NULL, NULL);
5117 ret = get_cur_path(sctx, clone_root->ino, gen, p);
5119 ret = get_inode_path(clone_root->root, clone_root->ino, p);
5125 * If the parent we're using has a received_uuid set then use that as
5126 * our clone source as that is what we will look for when doing a
5129 * This covers the case that we create a snapshot off of a received
5130 * subvolume and then use that as the parent and try to receive on a
5133 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5134 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5135 clone_root->root->root_item.received_uuid);
5137 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5138 clone_root->root->root_item.uuid);
5139 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5140 le64_to_cpu(clone_root->root->root_item.ctransid));
5141 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5142 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5143 clone_root->offset);
5145 ret = send_cmd(sctx);
5154 * Send an update extent command to user space.
5156 static int send_update_extent(struct send_ctx *sctx,
5157 u64 offset, u32 len)
5162 p = fs_path_alloc();
5166 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5170 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5174 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5175 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5176 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5178 ret = send_cmd(sctx);
5186 static int send_hole(struct send_ctx *sctx, u64 end)
5188 struct fs_path *p = NULL;
5189 u64 read_size = max_send_read_size(sctx);
5190 u64 offset = sctx->cur_inode_last_extent;
5194 * A hole that starts at EOF or beyond it. Since we do not yet support
5195 * fallocate (for extent preallocation and hole punching), sending a
5196 * write of zeroes starting at EOF or beyond would later require issuing
5197 * a truncate operation which would undo the write and achieve nothing.
5199 if (offset >= sctx->cur_inode_size)
5203 * Don't go beyond the inode's i_size due to prealloc extents that start
5206 end = min_t(u64, end, sctx->cur_inode_size);
5208 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5209 return send_update_extent(sctx, offset, end - offset);
5211 p = fs_path_alloc();
5214 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5216 goto tlv_put_failure;
5217 while (offset < end) {
5218 u64 len = min(end - offset, read_size);
5220 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5223 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5224 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5225 ret = put_data_header(sctx, len);
5228 memset(sctx->send_buf + sctx->send_size, 0, len);
5229 sctx->send_size += len;
5230 ret = send_cmd(sctx);
5235 sctx->cur_inode_next_write_offset = offset;
5241 static int send_extent_data(struct send_ctx *sctx,
5245 u64 read_size = max_send_read_size(sctx);
5248 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5249 return send_update_extent(sctx, offset, len);
5251 while (sent < len) {
5252 u64 size = min(len - sent, read_size);
5255 ret = send_write(sctx, offset + sent, size);
5264 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5265 * found, call send_set_xattr function to emit it.
5267 * Return 0 if there isn't a capability, or when the capability was emitted
5268 * successfully, or < 0 if an error occurred.
5270 static int send_capabilities(struct send_ctx *sctx)
5272 struct fs_path *fspath = NULL;
5273 struct btrfs_path *path;
5274 struct btrfs_dir_item *di;
5275 struct extent_buffer *leaf;
5276 unsigned long data_ptr;
5281 path = alloc_path_for_send();
5285 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5286 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5288 /* There is no xattr for this inode */
5290 } else if (IS_ERR(di)) {
5295 leaf = path->nodes[0];
5296 buf_len = btrfs_dir_data_len(leaf, di);
5298 fspath = fs_path_alloc();
5299 buf = kmalloc(buf_len, GFP_KERNEL);
5300 if (!fspath || !buf) {
5305 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5309 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5310 read_extent_buffer(leaf, buf, data_ptr, buf_len);
5312 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5313 strlen(XATTR_NAME_CAPS), buf, buf_len);
5316 fs_path_free(fspath);
5317 btrfs_free_path(path);
5321 static int clone_range(struct send_ctx *sctx,
5322 struct clone_root *clone_root,
5323 const u64 disk_byte,
5328 struct btrfs_path *path;
5329 struct btrfs_key key;
5331 u64 clone_src_i_size = 0;
5334 * Prevent cloning from a zero offset with a length matching the sector
5335 * size because in some scenarios this will make the receiver fail.
5337 * For example, if in the source filesystem the extent at offset 0
5338 * has a length of sectorsize and it was written using direct IO, then
5339 * it can never be an inline extent (even if compression is enabled).
5340 * Then this extent can be cloned in the original filesystem to a non
5341 * zero file offset, but it may not be possible to clone in the
5342 * destination filesystem because it can be inlined due to compression
5343 * on the destination filesystem (as the receiver's write operations are
5344 * always done using buffered IO). The same happens when the original
5345 * filesystem does not have compression enabled but the destination
5348 if (clone_root->offset == 0 &&
5349 len == sctx->send_root->fs_info->sectorsize)
5350 return send_extent_data(sctx, offset, len);
5352 path = alloc_path_for_send();
5357 * There are inodes that have extents that lie behind its i_size. Don't
5358 * accept clones from these extents.
5360 ret = __get_inode_info(clone_root->root, path, clone_root->ino,
5361 &clone_src_i_size, NULL, NULL, NULL, NULL, NULL);
5362 btrfs_release_path(path);
5367 * We can't send a clone operation for the entire range if we find
5368 * extent items in the respective range in the source file that
5369 * refer to different extents or if we find holes.
5370 * So check for that and do a mix of clone and regular write/copy
5371 * operations if needed.
5375 * mkfs.btrfs -f /dev/sda
5376 * mount /dev/sda /mnt
5377 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5378 * cp --reflink=always /mnt/foo /mnt/bar
5379 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5380 * btrfs subvolume snapshot -r /mnt /mnt/snap
5382 * If when we send the snapshot and we are processing file bar (which
5383 * has a higher inode number than foo) we blindly send a clone operation
5384 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5385 * a file bar that matches the content of file foo - iow, doesn't match
5386 * the content from bar in the original filesystem.
5388 key.objectid = clone_root->ino;
5389 key.type = BTRFS_EXTENT_DATA_KEY;
5390 key.offset = clone_root->offset;
5391 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5394 if (ret > 0 && path->slots[0] > 0) {
5395 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5396 if (key.objectid == clone_root->ino &&
5397 key.type == BTRFS_EXTENT_DATA_KEY)
5402 struct extent_buffer *leaf = path->nodes[0];
5403 int slot = path->slots[0];
5404 struct btrfs_file_extent_item *ei;
5408 u64 clone_data_offset;
5409 bool crossed_src_i_size = false;
5411 if (slot >= btrfs_header_nritems(leaf)) {
5412 ret = btrfs_next_leaf(clone_root->root, path);
5420 btrfs_item_key_to_cpu(leaf, &key, slot);
5423 * We might have an implicit trailing hole (NO_HOLES feature
5424 * enabled). We deal with it after leaving this loop.
5426 if (key.objectid != clone_root->ino ||
5427 key.type != BTRFS_EXTENT_DATA_KEY)
5430 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5431 type = btrfs_file_extent_type(leaf, ei);
5432 if (type == BTRFS_FILE_EXTENT_INLINE) {
5433 ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5434 ext_len = PAGE_ALIGN(ext_len);
5436 ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5439 if (key.offset + ext_len <= clone_root->offset)
5442 if (key.offset > clone_root->offset) {
5443 /* Implicit hole, NO_HOLES feature enabled. */
5444 u64 hole_len = key.offset - clone_root->offset;
5448 ret = send_extent_data(sctx, offset, hole_len);
5456 clone_root->offset += hole_len;
5457 data_offset += hole_len;
5460 if (key.offset >= clone_root->offset + len)
5463 if (key.offset >= clone_src_i_size)
5466 if (key.offset + ext_len > clone_src_i_size) {
5467 ext_len = clone_src_i_size - key.offset;
5468 crossed_src_i_size = true;
5471 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
5472 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
5473 clone_root->offset = key.offset;
5474 if (clone_data_offset < data_offset &&
5475 clone_data_offset + ext_len > data_offset) {
5478 extent_offset = data_offset - clone_data_offset;
5479 ext_len -= extent_offset;
5480 clone_data_offset += extent_offset;
5481 clone_root->offset += extent_offset;
5485 clone_len = min_t(u64, ext_len, len);
5487 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5488 clone_data_offset == data_offset) {
5489 const u64 src_end = clone_root->offset + clone_len;
5490 const u64 sectorsize = SZ_64K;
5493 * We can't clone the last block, when its size is not
5494 * sector size aligned, into the middle of a file. If we
5495 * do so, the receiver will get a failure (-EINVAL) when
5496 * trying to clone or will silently corrupt the data in
5497 * the destination file if it's on a kernel without the
5498 * fix introduced by commit ac765f83f1397646
5499 * ("Btrfs: fix data corruption due to cloning of eof
5502 * So issue a clone of the aligned down range plus a
5503 * regular write for the eof block, if we hit that case.
5505 * Also, we use the maximum possible sector size, 64K,
5506 * because we don't know what's the sector size of the
5507 * filesystem that receives the stream, so we have to
5508 * assume the largest possible sector size.
5510 if (src_end == clone_src_i_size &&
5511 !IS_ALIGNED(src_end, sectorsize) &&
5512 offset + clone_len < sctx->cur_inode_size) {
5515 slen = ALIGN_DOWN(src_end - clone_root->offset,
5518 ret = send_clone(sctx, offset, slen,
5523 ret = send_extent_data(sctx, offset + slen,
5526 ret = send_clone(sctx, offset, clone_len,
5529 } else if (crossed_src_i_size && clone_len < len) {
5531 * If we are at i_size of the clone source inode and we
5532 * can not clone from it, terminate the loop. This is
5533 * to avoid sending two write operations, one with a
5534 * length matching clone_len and the final one after
5535 * this loop with a length of len - clone_len.
5537 * When using encoded writes (BTRFS_SEND_FLAG_COMPRESSED
5538 * was passed to the send ioctl), this helps avoid
5539 * sending an encoded write for an offset that is not
5540 * sector size aligned, in case the i_size of the source
5541 * inode is not sector size aligned. That will make the
5542 * receiver fallback to decompression of the data and
5543 * writing it using regular buffered IO, therefore while
5544 * not incorrect, it's not optimal due decompression and
5545 * possible re-compression at the receiver.
5549 ret = send_extent_data(sctx, offset, clone_len);
5558 offset += clone_len;
5559 clone_root->offset += clone_len;
5562 * If we are cloning from the file we are currently processing,
5563 * and using the send root as the clone root, we must stop once
5564 * the current clone offset reaches the current eof of the file
5565 * at the receiver, otherwise we would issue an invalid clone
5566 * operation (source range going beyond eof) and cause the
5567 * receiver to fail. So if we reach the current eof, bail out
5568 * and fallback to a regular write.
5570 if (clone_root->root == sctx->send_root &&
5571 clone_root->ino == sctx->cur_ino &&
5572 clone_root->offset >= sctx->cur_inode_next_write_offset)
5575 data_offset += clone_len;
5581 ret = send_extent_data(sctx, offset, len);
5585 btrfs_free_path(path);
5589 static int send_write_or_clone(struct send_ctx *sctx,
5590 struct btrfs_path *path,
5591 struct btrfs_key *key,
5592 struct clone_root *clone_root)
5595 u64 offset = key->offset;
5597 u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5599 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
5603 if (clone_root && IS_ALIGNED(end, bs)) {
5604 struct btrfs_file_extent_item *ei;
5608 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5609 struct btrfs_file_extent_item);
5610 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5611 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5612 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5613 offset, end - offset);
5615 ret = send_extent_data(sctx, offset, end - offset);
5617 sctx->cur_inode_next_write_offset = end;
5621 static int is_extent_unchanged(struct send_ctx *sctx,
5622 struct btrfs_path *left_path,
5623 struct btrfs_key *ekey)
5626 struct btrfs_key key;
5627 struct btrfs_path *path = NULL;
5628 struct extent_buffer *eb;
5630 struct btrfs_key found_key;
5631 struct btrfs_file_extent_item *ei;
5636 u64 left_offset_fixed;
5644 path = alloc_path_for_send();
5648 eb = left_path->nodes[0];
5649 slot = left_path->slots[0];
5650 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5651 left_type = btrfs_file_extent_type(eb, ei);
5653 if (left_type != BTRFS_FILE_EXTENT_REG) {
5657 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5658 left_len = btrfs_file_extent_num_bytes(eb, ei);
5659 left_offset = btrfs_file_extent_offset(eb, ei);
5660 left_gen = btrfs_file_extent_generation(eb, ei);
5663 * Following comments will refer to these graphics. L is the left
5664 * extents which we are checking at the moment. 1-8 are the right
5665 * extents that we iterate.
5668 * |-1-|-2a-|-3-|-4-|-5-|-6-|
5671 * |--1--|-2b-|...(same as above)
5673 * Alternative situation. Happens on files where extents got split.
5675 * |-----------7-----------|-6-|
5677 * Alternative situation. Happens on files which got larger.
5680 * Nothing follows after 8.
5683 key.objectid = ekey->objectid;
5684 key.type = BTRFS_EXTENT_DATA_KEY;
5685 key.offset = ekey->offset;
5686 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5695 * Handle special case where the right side has no extents at all.
5697 eb = path->nodes[0];
5698 slot = path->slots[0];
5699 btrfs_item_key_to_cpu(eb, &found_key, slot);
5700 if (found_key.objectid != key.objectid ||
5701 found_key.type != key.type) {
5702 /* If we're a hole then just pretend nothing changed */
5703 ret = (left_disknr) ? 0 : 1;
5708 * We're now on 2a, 2b or 7.
5711 while (key.offset < ekey->offset + left_len) {
5712 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5713 right_type = btrfs_file_extent_type(eb, ei);
5714 if (right_type != BTRFS_FILE_EXTENT_REG &&
5715 right_type != BTRFS_FILE_EXTENT_INLINE) {
5720 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5721 right_len = btrfs_file_extent_ram_bytes(eb, ei);
5722 right_len = PAGE_ALIGN(right_len);
5724 right_len = btrfs_file_extent_num_bytes(eb, ei);
5728 * Are we at extent 8? If yes, we know the extent is changed.
5729 * This may only happen on the first iteration.
5731 if (found_key.offset + right_len <= ekey->offset) {
5732 /* If we're a hole just pretend nothing changed */
5733 ret = (left_disknr) ? 0 : 1;
5738 * We just wanted to see if when we have an inline extent, what
5739 * follows it is a regular extent (wanted to check the above
5740 * condition for inline extents too). This should normally not
5741 * happen but it's possible for example when we have an inline
5742 * compressed extent representing data with a size matching
5743 * the page size (currently the same as sector size).
5745 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5750 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5751 right_offset = btrfs_file_extent_offset(eb, ei);
5752 right_gen = btrfs_file_extent_generation(eb, ei);
5754 left_offset_fixed = left_offset;
5755 if (key.offset < ekey->offset) {
5756 /* Fix the right offset for 2a and 7. */
5757 right_offset += ekey->offset - key.offset;
5759 /* Fix the left offset for all behind 2a and 2b */
5760 left_offset_fixed += key.offset - ekey->offset;
5764 * Check if we have the same extent.
5766 if (left_disknr != right_disknr ||
5767 left_offset_fixed != right_offset ||
5768 left_gen != right_gen) {
5774 * Go to the next extent.
5776 ret = btrfs_next_item(sctx->parent_root, path);
5780 eb = path->nodes[0];
5781 slot = path->slots[0];
5782 btrfs_item_key_to_cpu(eb, &found_key, slot);
5784 if (ret || found_key.objectid != key.objectid ||
5785 found_key.type != key.type) {
5786 key.offset += right_len;
5789 if (found_key.offset != key.offset + right_len) {
5797 * We're now behind the left extent (treat as unchanged) or at the end
5798 * of the right side (treat as changed).
5800 if (key.offset >= ekey->offset + left_len)
5807 btrfs_free_path(path);
5811 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5813 struct btrfs_path *path;
5814 struct btrfs_root *root = sctx->send_root;
5815 struct btrfs_key key;
5818 path = alloc_path_for_send();
5822 sctx->cur_inode_last_extent = 0;
5824 key.objectid = sctx->cur_ino;
5825 key.type = BTRFS_EXTENT_DATA_KEY;
5826 key.offset = offset;
5827 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5831 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5832 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5835 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5837 btrfs_free_path(path);
5841 static int range_is_hole_in_parent(struct send_ctx *sctx,
5845 struct btrfs_path *path;
5846 struct btrfs_key key;
5847 struct btrfs_root *root = sctx->parent_root;
5848 u64 search_start = start;
5851 path = alloc_path_for_send();
5855 key.objectid = sctx->cur_ino;
5856 key.type = BTRFS_EXTENT_DATA_KEY;
5857 key.offset = search_start;
5858 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5861 if (ret > 0 && path->slots[0] > 0)
5864 while (search_start < end) {
5865 struct extent_buffer *leaf = path->nodes[0];
5866 int slot = path->slots[0];
5867 struct btrfs_file_extent_item *fi;
5870 if (slot >= btrfs_header_nritems(leaf)) {
5871 ret = btrfs_next_leaf(root, path);
5879 btrfs_item_key_to_cpu(leaf, &key, slot);
5880 if (key.objectid < sctx->cur_ino ||
5881 key.type < BTRFS_EXTENT_DATA_KEY)
5883 if (key.objectid > sctx->cur_ino ||
5884 key.type > BTRFS_EXTENT_DATA_KEY ||
5888 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5889 extent_end = btrfs_file_extent_end(path);
5890 if (extent_end <= start)
5892 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5893 search_start = extent_end;
5903 btrfs_free_path(path);
5907 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5908 struct btrfs_key *key)
5912 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5915 if (sctx->cur_inode_last_extent == (u64)-1) {
5916 ret = get_last_extent(sctx, key->offset - 1);
5921 if (path->slots[0] == 0 &&
5922 sctx->cur_inode_last_extent < key->offset) {
5924 * We might have skipped entire leafs that contained only
5925 * file extent items for our current inode. These leafs have
5926 * a generation number smaller (older) than the one in the
5927 * current leaf and the leaf our last extent came from, and
5928 * are located between these 2 leafs.
5930 ret = get_last_extent(sctx, key->offset - 1);
5935 if (sctx->cur_inode_last_extent < key->offset) {
5936 ret = range_is_hole_in_parent(sctx,
5937 sctx->cur_inode_last_extent,
5942 ret = send_hole(sctx, key->offset);
5946 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5950 static int process_extent(struct send_ctx *sctx,
5951 struct btrfs_path *path,
5952 struct btrfs_key *key)
5954 struct clone_root *found_clone = NULL;
5957 if (S_ISLNK(sctx->cur_inode_mode))
5960 if (sctx->parent_root && !sctx->cur_inode_new) {
5961 ret = is_extent_unchanged(sctx, path, key);
5969 struct btrfs_file_extent_item *ei;
5972 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5973 struct btrfs_file_extent_item);
5974 type = btrfs_file_extent_type(path->nodes[0], ei);
5975 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5976 type == BTRFS_FILE_EXTENT_REG) {
5978 * The send spec does not have a prealloc command yet,
5979 * so just leave a hole for prealloc'ed extents until
5980 * we have enough commands queued up to justify rev'ing
5983 if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5988 /* Have a hole, just skip it. */
5989 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5996 ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5997 sctx->cur_inode_size, &found_clone);
5998 if (ret != -ENOENT && ret < 0)
6001 ret = send_write_or_clone(sctx, path, key, found_clone);
6005 ret = maybe_send_hole(sctx, path, key);
6010 static int process_all_extents(struct send_ctx *sctx)
6013 struct btrfs_root *root;
6014 struct btrfs_path *path;
6015 struct btrfs_key key;
6016 struct btrfs_key found_key;
6017 struct extent_buffer *eb;
6020 root = sctx->send_root;
6021 path = alloc_path_for_send();
6025 key.objectid = sctx->cmp_key->objectid;
6026 key.type = BTRFS_EXTENT_DATA_KEY;
6028 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6033 eb = path->nodes[0];
6034 slot = path->slots[0];
6036 if (slot >= btrfs_header_nritems(eb)) {
6037 ret = btrfs_next_leaf(root, path);
6040 } else if (ret > 0) {
6047 btrfs_item_key_to_cpu(eb, &found_key, slot);
6049 if (found_key.objectid != key.objectid ||
6050 found_key.type != key.type) {
6055 ret = process_extent(sctx, path, &found_key);
6063 btrfs_free_path(path);
6067 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6069 int *refs_processed)
6073 if (sctx->cur_ino == 0)
6075 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6076 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6078 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6081 ret = process_recorded_refs(sctx, pending_move);
6085 *refs_processed = 1;
6090 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6101 int need_truncate = 1;
6102 int pending_move = 0;
6103 int refs_processed = 0;
6105 if (sctx->ignore_cur_inode)
6108 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6114 * We have processed the refs and thus need to advance send_progress.
6115 * Now, calls to get_cur_xxx will take the updated refs of the current
6116 * inode into account.
6118 * On the other hand, if our current inode is a directory and couldn't
6119 * be moved/renamed because its parent was renamed/moved too and it has
6120 * a higher inode number, we can only move/rename our current inode
6121 * after we moved/renamed its parent. Therefore in this case operate on
6122 * the old path (pre move/rename) of our current inode, and the
6123 * move/rename will be performed later.
6125 if (refs_processed && !pending_move)
6126 sctx->send_progress = sctx->cur_ino + 1;
6128 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6130 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6133 ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
6134 &left_mode, &left_uid, &left_gid, NULL);
6138 if (!sctx->parent_root || sctx->cur_inode_new) {
6140 if (!S_ISLNK(sctx->cur_inode_mode))
6142 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6147 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
6148 &old_size, NULL, &right_mode, &right_uid,
6153 if (left_uid != right_uid || left_gid != right_gid)
6155 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6157 if ((old_size == sctx->cur_inode_size) ||
6158 (sctx->cur_inode_size > old_size &&
6159 sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6163 if (S_ISREG(sctx->cur_inode_mode)) {
6164 if (need_send_hole(sctx)) {
6165 if (sctx->cur_inode_last_extent == (u64)-1 ||
6166 sctx->cur_inode_last_extent <
6167 sctx->cur_inode_size) {
6168 ret = get_last_extent(sctx, (u64)-1);
6172 if (sctx->cur_inode_last_extent <
6173 sctx->cur_inode_size) {
6174 ret = send_hole(sctx, sctx->cur_inode_size);
6179 if (need_truncate) {
6180 ret = send_truncate(sctx, sctx->cur_ino,
6181 sctx->cur_inode_gen,
6182 sctx->cur_inode_size);
6189 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6190 left_uid, left_gid);
6195 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6201 ret = send_capabilities(sctx);
6206 * If other directory inodes depended on our current directory
6207 * inode's move/rename, now do their move/rename operations.
6209 if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6210 ret = apply_children_dir_moves(sctx);
6214 * Need to send that every time, no matter if it actually
6215 * changed between the two trees as we have done changes to
6216 * the inode before. If our inode is a directory and it's
6217 * waiting to be moved/renamed, we will send its utimes when
6218 * it's moved/renamed, therefore we don't need to do it here.
6220 sctx->send_progress = sctx->cur_ino + 1;
6221 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6230 struct parent_paths_ctx {
6231 struct list_head *refs;
6232 struct send_ctx *sctx;
6235 static int record_parent_ref(int num, u64 dir, int index, struct fs_path *name,
6238 struct parent_paths_ctx *ppctx = ctx;
6240 return record_ref(ppctx->sctx->parent_root, dir, name, ppctx->sctx,
6245 * Issue unlink operations for all paths of the current inode found in the
6248 static int btrfs_unlink_all_paths(struct send_ctx *sctx)
6250 LIST_HEAD(deleted_refs);
6251 struct btrfs_path *path;
6252 struct btrfs_key key;
6253 struct parent_paths_ctx ctx;
6256 path = alloc_path_for_send();
6260 key.objectid = sctx->cur_ino;
6261 key.type = BTRFS_INODE_REF_KEY;
6263 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
6267 ctx.refs = &deleted_refs;
6271 struct extent_buffer *eb = path->nodes[0];
6272 int slot = path->slots[0];
6274 if (slot >= btrfs_header_nritems(eb)) {
6275 ret = btrfs_next_leaf(sctx->parent_root, path);
6283 btrfs_item_key_to_cpu(eb, &key, slot);
6284 if (key.objectid != sctx->cur_ino)
6286 if (key.type != BTRFS_INODE_REF_KEY &&
6287 key.type != BTRFS_INODE_EXTREF_KEY)
6290 ret = iterate_inode_ref(sctx->parent_root, path, &key, 1,
6291 record_parent_ref, &ctx);
6298 while (!list_empty(&deleted_refs)) {
6299 struct recorded_ref *ref;
6301 ref = list_first_entry(&deleted_refs, struct recorded_ref, list);
6302 ret = send_unlink(sctx, ref->full_path);
6305 fs_path_free(ref->full_path);
6306 list_del(&ref->list);
6311 btrfs_free_path(path);
6313 __free_recorded_refs(&deleted_refs);
6317 static int changed_inode(struct send_ctx *sctx,
6318 enum btrfs_compare_tree_result result)
6321 struct btrfs_key *key = sctx->cmp_key;
6322 struct btrfs_inode_item *left_ii = NULL;
6323 struct btrfs_inode_item *right_ii = NULL;
6327 sctx->cur_ino = key->objectid;
6328 sctx->cur_inode_new_gen = 0;
6329 sctx->cur_inode_last_extent = (u64)-1;
6330 sctx->cur_inode_next_write_offset = 0;
6331 sctx->ignore_cur_inode = false;
6334 * Set send_progress to current inode. This will tell all get_cur_xxx
6335 * functions that the current inode's refs are not updated yet. Later,
6336 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6338 sctx->send_progress = sctx->cur_ino;
6340 if (result == BTRFS_COMPARE_TREE_NEW ||
6341 result == BTRFS_COMPARE_TREE_CHANGED) {
6342 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6343 sctx->left_path->slots[0],
6344 struct btrfs_inode_item);
6345 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6348 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6349 sctx->right_path->slots[0],
6350 struct btrfs_inode_item);
6351 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6354 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6355 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6356 sctx->right_path->slots[0],
6357 struct btrfs_inode_item);
6359 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6363 * The cur_ino = root dir case is special here. We can't treat
6364 * the inode as deleted+reused because it would generate a
6365 * stream that tries to delete/mkdir the root dir.
6367 if (left_gen != right_gen &&
6368 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6369 sctx->cur_inode_new_gen = 1;
6373 * Normally we do not find inodes with a link count of zero (orphans)
6374 * because the most common case is to create a snapshot and use it
6375 * for a send operation. However other less common use cases involve
6376 * using a subvolume and send it after turning it to RO mode just
6377 * after deleting all hard links of a file while holding an open
6378 * file descriptor against it or turning a RO snapshot into RW mode,
6379 * keep an open file descriptor against a file, delete it and then
6380 * turn the snapshot back to RO mode before using it for a send
6381 * operation. So if we find such cases, ignore the inode and all its
6382 * items completely if it's a new inode, or if it's a changed inode
6383 * make sure all its previous paths (from the parent snapshot) are all
6384 * unlinked and all other the inode items are ignored.
6386 if (result == BTRFS_COMPARE_TREE_NEW ||
6387 result == BTRFS_COMPARE_TREE_CHANGED) {
6390 nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6392 sctx->ignore_cur_inode = true;
6393 if (result == BTRFS_COMPARE_TREE_CHANGED)
6394 ret = btrfs_unlink_all_paths(sctx);
6399 if (result == BTRFS_COMPARE_TREE_NEW) {
6400 sctx->cur_inode_gen = left_gen;
6401 sctx->cur_inode_new = 1;
6402 sctx->cur_inode_deleted = 0;
6403 sctx->cur_inode_size = btrfs_inode_size(
6404 sctx->left_path->nodes[0], left_ii);
6405 sctx->cur_inode_mode = btrfs_inode_mode(
6406 sctx->left_path->nodes[0], left_ii);
6407 sctx->cur_inode_rdev = btrfs_inode_rdev(
6408 sctx->left_path->nodes[0], left_ii);
6409 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6410 ret = send_create_inode_if_needed(sctx);
6411 } else if (result == BTRFS_COMPARE_TREE_DELETED) {
6412 sctx->cur_inode_gen = right_gen;
6413 sctx->cur_inode_new = 0;
6414 sctx->cur_inode_deleted = 1;
6415 sctx->cur_inode_size = btrfs_inode_size(
6416 sctx->right_path->nodes[0], right_ii);
6417 sctx->cur_inode_mode = btrfs_inode_mode(
6418 sctx->right_path->nodes[0], right_ii);
6419 } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6421 * We need to do some special handling in case the inode was
6422 * reported as changed with a changed generation number. This
6423 * means that the original inode was deleted and new inode
6424 * reused the same inum. So we have to treat the old inode as
6425 * deleted and the new one as new.
6427 if (sctx->cur_inode_new_gen) {
6429 * First, process the inode as if it was deleted.
6431 sctx->cur_inode_gen = right_gen;
6432 sctx->cur_inode_new = 0;
6433 sctx->cur_inode_deleted = 1;
6434 sctx->cur_inode_size = btrfs_inode_size(
6435 sctx->right_path->nodes[0], right_ii);
6436 sctx->cur_inode_mode = btrfs_inode_mode(
6437 sctx->right_path->nodes[0], right_ii);
6438 ret = process_all_refs(sctx,
6439 BTRFS_COMPARE_TREE_DELETED);
6444 * Now process the inode as if it was new.
6446 sctx->cur_inode_gen = left_gen;
6447 sctx->cur_inode_new = 1;
6448 sctx->cur_inode_deleted = 0;
6449 sctx->cur_inode_size = btrfs_inode_size(
6450 sctx->left_path->nodes[0], left_ii);
6451 sctx->cur_inode_mode = btrfs_inode_mode(
6452 sctx->left_path->nodes[0], left_ii);
6453 sctx->cur_inode_rdev = btrfs_inode_rdev(
6454 sctx->left_path->nodes[0], left_ii);
6455 ret = send_create_inode_if_needed(sctx);
6459 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
6463 * Advance send_progress now as we did not get into
6464 * process_recorded_refs_if_needed in the new_gen case.
6466 sctx->send_progress = sctx->cur_ino + 1;
6469 * Now process all extents and xattrs of the inode as if
6470 * they were all new.
6472 ret = process_all_extents(sctx);
6475 ret = process_all_new_xattrs(sctx);
6479 sctx->cur_inode_gen = left_gen;
6480 sctx->cur_inode_new = 0;
6481 sctx->cur_inode_new_gen = 0;
6482 sctx->cur_inode_deleted = 0;
6483 sctx->cur_inode_size = btrfs_inode_size(
6484 sctx->left_path->nodes[0], left_ii);
6485 sctx->cur_inode_mode = btrfs_inode_mode(
6486 sctx->left_path->nodes[0], left_ii);
6495 * We have to process new refs before deleted refs, but compare_trees gives us
6496 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
6497 * first and later process them in process_recorded_refs.
6498 * For the cur_inode_new_gen case, we skip recording completely because
6499 * changed_inode did already initiate processing of refs. The reason for this is
6500 * that in this case, compare_tree actually compares the refs of 2 different
6501 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
6502 * refs of the right tree as deleted and all refs of the left tree as new.
6504 static int changed_ref(struct send_ctx *sctx,
6505 enum btrfs_compare_tree_result result)
6509 if (sctx->cur_ino != sctx->cmp_key->objectid) {
6510 inconsistent_snapshot_error(sctx, result, "reference");
6514 if (!sctx->cur_inode_new_gen &&
6515 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
6516 if (result == BTRFS_COMPARE_TREE_NEW)
6517 ret = record_new_ref(sctx);
6518 else if (result == BTRFS_COMPARE_TREE_DELETED)
6519 ret = record_deleted_ref(sctx);
6520 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6521 ret = record_changed_ref(sctx);
6528 * Process new/deleted/changed xattrs. We skip processing in the
6529 * cur_inode_new_gen case because changed_inode did already initiate processing
6530 * of xattrs. The reason is the same as in changed_ref
6532 static int changed_xattr(struct send_ctx *sctx,
6533 enum btrfs_compare_tree_result result)
6537 if (sctx->cur_ino != sctx->cmp_key->objectid) {
6538 inconsistent_snapshot_error(sctx, result, "xattr");
6542 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6543 if (result == BTRFS_COMPARE_TREE_NEW)
6544 ret = process_new_xattr(sctx);
6545 else if (result == BTRFS_COMPARE_TREE_DELETED)
6546 ret = process_deleted_xattr(sctx);
6547 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6548 ret = process_changed_xattr(sctx);
6555 * Process new/deleted/changed extents. We skip processing in the
6556 * cur_inode_new_gen case because changed_inode did already initiate processing
6557 * of extents. The reason is the same as in changed_ref
6559 static int changed_extent(struct send_ctx *sctx,
6560 enum btrfs_compare_tree_result result)
6565 * We have found an extent item that changed without the inode item
6566 * having changed. This can happen either after relocation (where the
6567 * disk_bytenr of an extent item is replaced at
6568 * relocation.c:replace_file_extents()) or after deduplication into a
6569 * file in both the parent and send snapshots (where an extent item can
6570 * get modified or replaced with a new one). Note that deduplication
6571 * updates the inode item, but it only changes the iversion (sequence
6572 * field in the inode item) of the inode, so if a file is deduplicated
6573 * the same amount of times in both the parent and send snapshots, its
6574 * iversion becames the same in both snapshots, whence the inode item is
6575 * the same on both snapshots.
6577 if (sctx->cur_ino != sctx->cmp_key->objectid)
6580 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6581 if (result != BTRFS_COMPARE_TREE_DELETED)
6582 ret = process_extent(sctx, sctx->left_path,
6589 static int dir_changed(struct send_ctx *sctx, u64 dir)
6591 u64 orig_gen, new_gen;
6594 ret = get_inode_info(sctx->send_root, dir, NULL, &new_gen, NULL, NULL,
6599 ret = get_inode_info(sctx->parent_root, dir, NULL, &orig_gen, NULL,
6604 return (orig_gen != new_gen) ? 1 : 0;
6607 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
6608 struct btrfs_key *key)
6610 struct btrfs_inode_extref *extref;
6611 struct extent_buffer *leaf;
6612 u64 dirid = 0, last_dirid = 0;
6619 /* Easy case, just check this one dirid */
6620 if (key->type == BTRFS_INODE_REF_KEY) {
6621 dirid = key->offset;
6623 ret = dir_changed(sctx, dirid);
6627 leaf = path->nodes[0];
6628 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
6629 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
6630 while (cur_offset < item_size) {
6631 extref = (struct btrfs_inode_extref *)(ptr +
6633 dirid = btrfs_inode_extref_parent(leaf, extref);
6634 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
6635 cur_offset += ref_name_len + sizeof(*extref);
6636 if (dirid == last_dirid)
6638 ret = dir_changed(sctx, dirid);
6648 * Updates compare related fields in sctx and simply forwards to the actual
6649 * changed_xxx functions.
6651 static int changed_cb(struct btrfs_path *left_path,
6652 struct btrfs_path *right_path,
6653 struct btrfs_key *key,
6654 enum btrfs_compare_tree_result result,
6658 struct send_ctx *sctx = ctx;
6660 if (result == BTRFS_COMPARE_TREE_SAME) {
6661 if (key->type == BTRFS_INODE_REF_KEY ||
6662 key->type == BTRFS_INODE_EXTREF_KEY) {
6663 ret = compare_refs(sctx, left_path, key);
6668 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
6669 return maybe_send_hole(sctx, left_path, key);
6673 result = BTRFS_COMPARE_TREE_CHANGED;
6677 sctx->left_path = left_path;
6678 sctx->right_path = right_path;
6679 sctx->cmp_key = key;
6681 ret = finish_inode_if_needed(sctx, 0);
6685 /* Ignore non-FS objects */
6686 if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
6687 key->objectid == BTRFS_FREE_SPACE_OBJECTID)
6690 if (key->type == BTRFS_INODE_ITEM_KEY) {
6691 ret = changed_inode(sctx, result);
6692 } else if (!sctx->ignore_cur_inode) {
6693 if (key->type == BTRFS_INODE_REF_KEY ||
6694 key->type == BTRFS_INODE_EXTREF_KEY)
6695 ret = changed_ref(sctx, result);
6696 else if (key->type == BTRFS_XATTR_ITEM_KEY)
6697 ret = changed_xattr(sctx, result);
6698 else if (key->type == BTRFS_EXTENT_DATA_KEY)
6699 ret = changed_extent(sctx, result);
6706 static int full_send_tree(struct send_ctx *sctx)
6709 struct btrfs_root *send_root = sctx->send_root;
6710 struct btrfs_key key;
6711 struct btrfs_path *path;
6712 struct extent_buffer *eb;
6715 path = alloc_path_for_send();
6719 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
6720 key.type = BTRFS_INODE_ITEM_KEY;
6723 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
6730 eb = path->nodes[0];
6731 slot = path->slots[0];
6732 btrfs_item_key_to_cpu(eb, &key, slot);
6734 ret = changed_cb(path, NULL, &key,
6735 BTRFS_COMPARE_TREE_NEW, sctx);
6739 ret = btrfs_next_item(send_root, path);
6749 ret = finish_inode_if_needed(sctx, 1);
6752 btrfs_free_path(path);
6756 static int tree_move_down(struct btrfs_path *path, int *level)
6758 struct extent_buffer *eb;
6760 BUG_ON(*level == 0);
6761 eb = btrfs_read_node_slot(path->nodes[*level], path->slots[*level]);
6765 path->nodes[*level - 1] = eb;
6766 path->slots[*level - 1] = 0;
6771 static int tree_move_next_or_upnext(struct btrfs_path *path,
6772 int *level, int root_level)
6776 nritems = btrfs_header_nritems(path->nodes[*level]);
6778 path->slots[*level]++;
6780 while (path->slots[*level] >= nritems) {
6781 if (*level == root_level)
6785 path->slots[*level] = 0;
6786 free_extent_buffer(path->nodes[*level]);
6787 path->nodes[*level] = NULL;
6789 path->slots[*level]++;
6791 nritems = btrfs_header_nritems(path->nodes[*level]);
6798 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
6801 static int tree_advance(struct btrfs_path *path,
6802 int *level, int root_level,
6804 struct btrfs_key *key)
6808 if (*level == 0 || !allow_down) {
6809 ret = tree_move_next_or_upnext(path, level, root_level);
6811 ret = tree_move_down(path, level);
6815 btrfs_item_key_to_cpu(path->nodes[*level], key,
6816 path->slots[*level]);
6818 btrfs_node_key_to_cpu(path->nodes[*level], key,
6819 path->slots[*level]);
6824 static int tree_compare_item(struct btrfs_path *left_path,
6825 struct btrfs_path *right_path,
6830 unsigned long off1, off2;
6832 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]);
6833 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]);
6837 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
6838 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
6839 right_path->slots[0]);
6841 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
6843 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
6850 * This function compares two trees and calls the provided callback for
6851 * every changed/new/deleted item it finds.
6852 * If shared tree blocks are encountered, whole subtrees are skipped, making
6853 * the compare pretty fast on snapshotted subvolumes.
6855 * This currently works on commit roots only. As commit roots are read only,
6856 * we don't do any locking. The commit roots are protected with transactions.
6857 * Transactions are ended and rejoined when a commit is tried in between.
6859 * This function checks for modifications done to the trees while comparing.
6860 * If it detects a change, it aborts immediately.
6862 static int btrfs_compare_trees(struct btrfs_root *left_root,
6863 struct btrfs_root *right_root, void *ctx)
6865 struct btrfs_fs_info *fs_info = left_root->fs_info;
6868 struct btrfs_path *left_path = NULL;
6869 struct btrfs_path *right_path = NULL;
6870 struct btrfs_key left_key;
6871 struct btrfs_key right_key;
6872 char *tmp_buf = NULL;
6873 int left_root_level;
6874 int right_root_level;
6877 int left_end_reached;
6878 int right_end_reached;
6886 left_path = btrfs_alloc_path();
6891 right_path = btrfs_alloc_path();
6897 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
6903 left_path->search_commit_root = 1;
6904 left_path->skip_locking = 1;
6905 right_path->search_commit_root = 1;
6906 right_path->skip_locking = 1;
6909 * Strategy: Go to the first items of both trees. Then do
6911 * If both trees are at level 0
6912 * Compare keys of current items
6913 * If left < right treat left item as new, advance left tree
6915 * If left > right treat right item as deleted, advance right tree
6917 * If left == right do deep compare of items, treat as changed if
6918 * needed, advance both trees and repeat
6919 * If both trees are at the same level but not at level 0
6920 * Compare keys of current nodes/leafs
6921 * If left < right advance left tree and repeat
6922 * If left > right advance right tree and repeat
6923 * If left == right compare blockptrs of the next nodes/leafs
6924 * If they match advance both trees but stay at the same level
6926 * If they don't match advance both trees while allowing to go
6928 * If tree levels are different
6929 * Advance the tree that needs it and repeat
6931 * Advancing a tree means:
6932 * If we are at level 0, try to go to the next slot. If that's not
6933 * possible, go one level up and repeat. Stop when we found a level
6934 * where we could go to the next slot. We may at this point be on a
6937 * If we are not at level 0 and not on shared tree blocks, go one
6940 * If we are not at level 0 and on shared tree blocks, go one slot to
6941 * the right if possible or go up and right.
6944 down_read(&fs_info->commit_root_sem);
6945 left_level = btrfs_header_level(left_root->commit_root);
6946 left_root_level = left_level;
6947 left_path->nodes[left_level] =
6948 btrfs_clone_extent_buffer(left_root->commit_root);
6949 if (!left_path->nodes[left_level]) {
6950 up_read(&fs_info->commit_root_sem);
6955 right_level = btrfs_header_level(right_root->commit_root);
6956 right_root_level = right_level;
6957 right_path->nodes[right_level] =
6958 btrfs_clone_extent_buffer(right_root->commit_root);
6959 if (!right_path->nodes[right_level]) {
6960 up_read(&fs_info->commit_root_sem);
6964 up_read(&fs_info->commit_root_sem);
6966 if (left_level == 0)
6967 btrfs_item_key_to_cpu(left_path->nodes[left_level],
6968 &left_key, left_path->slots[left_level]);
6970 btrfs_node_key_to_cpu(left_path->nodes[left_level],
6971 &left_key, left_path->slots[left_level]);
6972 if (right_level == 0)
6973 btrfs_item_key_to_cpu(right_path->nodes[right_level],
6974 &right_key, right_path->slots[right_level]);
6976 btrfs_node_key_to_cpu(right_path->nodes[right_level],
6977 &right_key, right_path->slots[right_level]);
6979 left_end_reached = right_end_reached = 0;
6980 advance_left = advance_right = 0;
6984 if (advance_left && !left_end_reached) {
6985 ret = tree_advance(left_path, &left_level,
6987 advance_left != ADVANCE_ONLY_NEXT,
6990 left_end_reached = ADVANCE;
6995 if (advance_right && !right_end_reached) {
6996 ret = tree_advance(right_path, &right_level,
6998 advance_right != ADVANCE_ONLY_NEXT,
7001 right_end_reached = ADVANCE;
7007 if (left_end_reached && right_end_reached) {
7010 } else if (left_end_reached) {
7011 if (right_level == 0) {
7012 ret = changed_cb(left_path, right_path,
7014 BTRFS_COMPARE_TREE_DELETED,
7019 advance_right = ADVANCE;
7021 } else if (right_end_reached) {
7022 if (left_level == 0) {
7023 ret = changed_cb(left_path, right_path,
7025 BTRFS_COMPARE_TREE_NEW,
7030 advance_left = ADVANCE;
7034 if (left_level == 0 && right_level == 0) {
7035 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7037 ret = changed_cb(left_path, right_path,
7039 BTRFS_COMPARE_TREE_NEW,
7043 advance_left = ADVANCE;
7044 } else if (cmp > 0) {
7045 ret = changed_cb(left_path, right_path,
7047 BTRFS_COMPARE_TREE_DELETED,
7051 advance_right = ADVANCE;
7053 enum btrfs_compare_tree_result result;
7055 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7056 ret = tree_compare_item(left_path, right_path,
7059 result = BTRFS_COMPARE_TREE_CHANGED;
7061 result = BTRFS_COMPARE_TREE_SAME;
7062 ret = changed_cb(left_path, right_path,
7063 &left_key, result, ctx);
7066 advance_left = ADVANCE;
7067 advance_right = ADVANCE;
7069 } else if (left_level == right_level) {
7070 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7072 advance_left = ADVANCE;
7073 } else if (cmp > 0) {
7074 advance_right = ADVANCE;
7076 left_blockptr = btrfs_node_blockptr(
7077 left_path->nodes[left_level],
7078 left_path->slots[left_level]);
7079 right_blockptr = btrfs_node_blockptr(
7080 right_path->nodes[right_level],
7081 right_path->slots[right_level]);
7082 left_gen = btrfs_node_ptr_generation(
7083 left_path->nodes[left_level],
7084 left_path->slots[left_level]);
7085 right_gen = btrfs_node_ptr_generation(
7086 right_path->nodes[right_level],
7087 right_path->slots[right_level]);
7088 if (left_blockptr == right_blockptr &&
7089 left_gen == right_gen) {
7091 * As we're on a shared block, don't
7092 * allow to go deeper.
7094 advance_left = ADVANCE_ONLY_NEXT;
7095 advance_right = ADVANCE_ONLY_NEXT;
7097 advance_left = ADVANCE;
7098 advance_right = ADVANCE;
7101 } else if (left_level < right_level) {
7102 advance_right = ADVANCE;
7104 advance_left = ADVANCE;
7109 btrfs_free_path(left_path);
7110 btrfs_free_path(right_path);
7115 static int send_subvol(struct send_ctx *sctx)
7119 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7120 ret = send_header(sctx);
7125 ret = send_subvol_begin(sctx);
7129 if (sctx->parent_root) {
7130 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7133 ret = finish_inode_if_needed(sctx, 1);
7137 ret = full_send_tree(sctx);
7143 free_recorded_refs(sctx);
7148 * If orphan cleanup did remove any orphans from a root, it means the tree
7149 * was modified and therefore the commit root is not the same as the current
7150 * root anymore. This is a problem, because send uses the commit root and
7151 * therefore can see inode items that don't exist in the current root anymore,
7152 * and for example make calls to btrfs_iget, which will do tree lookups based
7153 * on the current root and not on the commit root. Those lookups will fail,
7154 * returning a -ESTALE error, and making send fail with that error. So make
7155 * sure a send does not see any orphans we have just removed, and that it will
7156 * see the same inodes regardless of whether a transaction commit happened
7157 * before it started (meaning that the commit root will be the same as the
7158 * current root) or not.
7160 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
7163 struct btrfs_trans_handle *trans = NULL;
7166 if (sctx->parent_root &&
7167 sctx->parent_root->node != sctx->parent_root->commit_root)
7170 for (i = 0; i < sctx->clone_roots_cnt; i++)
7171 if (sctx->clone_roots[i].root->node !=
7172 sctx->clone_roots[i].root->commit_root)
7176 return btrfs_end_transaction(trans);
7181 /* Use any root, all fs roots will get their commit roots updated. */
7183 trans = btrfs_join_transaction(sctx->send_root);
7185 return PTR_ERR(trans);
7189 return btrfs_commit_transaction(trans);
7193 * Make sure any existing dellaloc is flushed for any root used by a send
7194 * operation so that we do not miss any data and we do not race with writeback
7195 * finishing and changing a tree while send is using the tree. This could
7196 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
7197 * a send operation then uses the subvolume.
7198 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
7200 static int flush_delalloc_roots(struct send_ctx *sctx)
7202 struct btrfs_root *root = sctx->parent_root;
7207 ret = btrfs_start_delalloc_snapshot(root);
7210 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
7213 for (i = 0; i < sctx->clone_roots_cnt; i++) {
7214 root = sctx->clone_roots[i].root;
7215 ret = btrfs_start_delalloc_snapshot(root);
7218 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
7224 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
7226 spin_lock(&root->root_item_lock);
7227 root->send_in_progress--;
7229 * Not much left to do, we don't know why it's unbalanced and
7230 * can't blindly reset it to 0.
7232 if (root->send_in_progress < 0)
7233 btrfs_err(root->fs_info,
7234 "send_in_progress unbalanced %d root %llu",
7235 root->send_in_progress, root->root_key.objectid);
7236 spin_unlock(&root->root_item_lock);
7239 static void dedupe_in_progress_warn(const struct btrfs_root *root)
7241 btrfs_warn_rl(root->fs_info,
7242 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
7243 root->root_key.objectid, root->dedupe_in_progress);
7246 long btrfs_ioctl_send(struct file *mnt_file, struct btrfs_ioctl_send_args *arg)
7249 struct btrfs_root *send_root = BTRFS_I(file_inode(mnt_file))->root;
7250 struct btrfs_fs_info *fs_info = send_root->fs_info;
7251 struct btrfs_root *clone_root;
7252 struct send_ctx *sctx = NULL;
7254 u64 *clone_sources_tmp = NULL;
7255 int clone_sources_to_rollback = 0;
7257 int sort_clone_roots = 0;
7259 if (!capable(CAP_SYS_ADMIN))
7263 * The subvolume must remain read-only during send, protect against
7264 * making it RW. This also protects against deletion.
7266 spin_lock(&send_root->root_item_lock);
7267 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
7268 dedupe_in_progress_warn(send_root);
7269 spin_unlock(&send_root->root_item_lock);
7272 send_root->send_in_progress++;
7273 spin_unlock(&send_root->root_item_lock);
7276 * Userspace tools do the checks and warn the user if it's
7279 if (!btrfs_root_readonly(send_root)) {
7285 * Check that we don't overflow at later allocations, we request
7286 * clone_sources_count + 1 items, and compare to unsigned long inside
7287 * access_ok. Also set an upper limit for allocation size so this can't
7288 * easily exhaust memory. Max number of clone sources is about 200K.
7290 if (arg->clone_sources_count > SZ_8M / sizeof(struct clone_root)) {
7295 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
7300 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
7306 INIT_LIST_HEAD(&sctx->new_refs);
7307 INIT_LIST_HEAD(&sctx->deleted_refs);
7308 INIT_RADIX_TREE(&sctx->name_cache, GFP_KERNEL);
7309 INIT_LIST_HEAD(&sctx->name_cache_list);
7311 sctx->flags = arg->flags;
7313 sctx->send_filp = fget(arg->send_fd);
7314 if (!sctx->send_filp || !(sctx->send_filp->f_mode & FMODE_WRITE)) {
7319 sctx->send_root = send_root;
7321 * Unlikely but possible, if the subvolume is marked for deletion but
7322 * is slow to remove the directory entry, send can still be started
7324 if (btrfs_root_dead(sctx->send_root)) {
7329 sctx->clone_roots_cnt = arg->clone_sources_count;
7331 sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
7332 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
7333 if (!sctx->send_buf) {
7338 sctx->pending_dir_moves = RB_ROOT;
7339 sctx->waiting_dir_moves = RB_ROOT;
7340 sctx->orphan_dirs = RB_ROOT;
7342 sctx->clone_roots = kvcalloc(sizeof(*sctx->clone_roots),
7343 arg->clone_sources_count + 1,
7345 if (!sctx->clone_roots) {
7350 alloc_size = array_size(sizeof(*arg->clone_sources),
7351 arg->clone_sources_count);
7353 if (arg->clone_sources_count) {
7354 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
7355 if (!clone_sources_tmp) {
7360 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
7367 for (i = 0; i < arg->clone_sources_count; i++) {
7368 clone_root = btrfs_get_fs_root(fs_info,
7369 clone_sources_tmp[i], true);
7370 if (IS_ERR(clone_root)) {
7371 ret = PTR_ERR(clone_root);
7374 spin_lock(&clone_root->root_item_lock);
7375 if (!btrfs_root_readonly(clone_root) ||
7376 btrfs_root_dead(clone_root)) {
7377 spin_unlock(&clone_root->root_item_lock);
7378 btrfs_put_root(clone_root);
7382 if (clone_root->dedupe_in_progress) {
7383 dedupe_in_progress_warn(clone_root);
7384 spin_unlock(&clone_root->root_item_lock);
7385 btrfs_put_root(clone_root);
7389 clone_root->send_in_progress++;
7390 spin_unlock(&clone_root->root_item_lock);
7392 sctx->clone_roots[i].root = clone_root;
7393 clone_sources_to_rollback = i + 1;
7395 kvfree(clone_sources_tmp);
7396 clone_sources_tmp = NULL;
7399 if (arg->parent_root) {
7400 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
7402 if (IS_ERR(sctx->parent_root)) {
7403 ret = PTR_ERR(sctx->parent_root);
7407 spin_lock(&sctx->parent_root->root_item_lock);
7408 sctx->parent_root->send_in_progress++;
7409 if (!btrfs_root_readonly(sctx->parent_root) ||
7410 btrfs_root_dead(sctx->parent_root)) {
7411 spin_unlock(&sctx->parent_root->root_item_lock);
7415 if (sctx->parent_root->dedupe_in_progress) {
7416 dedupe_in_progress_warn(sctx->parent_root);
7417 spin_unlock(&sctx->parent_root->root_item_lock);
7421 spin_unlock(&sctx->parent_root->root_item_lock);
7425 * Clones from send_root are allowed, but only if the clone source
7426 * is behind the current send position. This is checked while searching
7427 * for possible clone sources.
7429 sctx->clone_roots[sctx->clone_roots_cnt++].root =
7430 btrfs_grab_root(sctx->send_root);
7432 /* We do a bsearch later */
7433 sort(sctx->clone_roots, sctx->clone_roots_cnt,
7434 sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
7436 sort_clone_roots = 1;
7438 ret = flush_delalloc_roots(sctx);
7442 ret = ensure_commit_roots_uptodate(sctx);
7446 mutex_lock(&fs_info->balance_mutex);
7447 if (test_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags)) {
7448 mutex_unlock(&fs_info->balance_mutex);
7449 btrfs_warn_rl(fs_info,
7450 "cannot run send because a balance operation is in progress");
7454 fs_info->send_in_progress++;
7455 mutex_unlock(&fs_info->balance_mutex);
7457 current->journal_info = BTRFS_SEND_TRANS_STUB;
7458 ret = send_subvol(sctx);
7459 current->journal_info = NULL;
7460 mutex_lock(&fs_info->balance_mutex);
7461 fs_info->send_in_progress--;
7462 mutex_unlock(&fs_info->balance_mutex);
7466 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
7467 ret = begin_cmd(sctx, BTRFS_SEND_C_END);
7470 ret = send_cmd(sctx);
7476 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
7477 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
7479 struct pending_dir_move *pm;
7481 n = rb_first(&sctx->pending_dir_moves);
7482 pm = rb_entry(n, struct pending_dir_move, node);
7483 while (!list_empty(&pm->list)) {
7484 struct pending_dir_move *pm2;
7486 pm2 = list_first_entry(&pm->list,
7487 struct pending_dir_move, list);
7488 free_pending_move(sctx, pm2);
7490 free_pending_move(sctx, pm);
7493 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
7494 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
7496 struct waiting_dir_move *dm;
7498 n = rb_first(&sctx->waiting_dir_moves);
7499 dm = rb_entry(n, struct waiting_dir_move, node);
7500 rb_erase(&dm->node, &sctx->waiting_dir_moves);
7504 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
7505 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
7507 struct orphan_dir_info *odi;
7509 n = rb_first(&sctx->orphan_dirs);
7510 odi = rb_entry(n, struct orphan_dir_info, node);
7511 free_orphan_dir_info(sctx, odi);
7514 if (sort_clone_roots) {
7515 for (i = 0; i < sctx->clone_roots_cnt; i++) {
7516 btrfs_root_dec_send_in_progress(
7517 sctx->clone_roots[i].root);
7518 btrfs_put_root(sctx->clone_roots[i].root);
7521 for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
7522 btrfs_root_dec_send_in_progress(
7523 sctx->clone_roots[i].root);
7524 btrfs_put_root(sctx->clone_roots[i].root);
7527 btrfs_root_dec_send_in_progress(send_root);
7529 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
7530 btrfs_root_dec_send_in_progress(sctx->parent_root);
7531 btrfs_put_root(sctx->parent_root);
7534 kvfree(clone_sources_tmp);
7537 if (sctx->send_filp)
7538 fput(sctx->send_filp);
7540 kvfree(sctx->clone_roots);
7541 kvfree(sctx->send_buf);
7543 name_cache_free(sctx);