1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2012 Alexander Block. All rights reserved.
4 */
5
6 #include <linux/bsearch.h>
7 #include <linux/fs.h>
8 #include <linux/file.h>
9 #include <linux/sort.h>
10 #include <linux/mount.h>
11 #include <linux/xattr.h>
12 #include <linux/posix_acl_xattr.h>
13 #include <linux/radix-tree.h>
14 #include <linux/vmalloc.h>
15 #include <linux/string.h>
16 #include <linux/compat.h>
17 #include <linux/crc32c.h>
18 #include <linux/fsverity.h>
19
20 #include "send.h"
21 #include "ctree.h"
22 #include "backref.h"
23 #include "locking.h"
24 #include "disk-io.h"
25 #include "btrfs_inode.h"
26 #include "transaction.h"
27 #include "compression.h"
28 #include "xattr.h"
29 #include "print-tree.h"
30 #include "accessors.h"
31 #include "dir-item.h"
32 #include "file-item.h"
33 #include "ioctl.h"
34 #include "verity.h"
35 #include "lru_cache.h"
36
37 /*
38 * Maximum number of references an extent can have in order for us to attempt to
39 * issue clone operations instead of write operations. This currently exists to
40 * avoid hitting limitations of the backreference walking code (taking a lot of
41 * time and using too much memory for extents with large number of references).
42 */
43 #define SEND_MAX_EXTENT_REFS 1024
44
45 /*
46 * A fs_path is a helper to dynamically build path names with unknown size.
47 * It reallocates the internal buffer on demand.
48 * It allows fast adding of path elements on the right side (normal path) and
49 * fast adding to the left side (reversed path). A reversed path can also be
50 * unreversed if needed.
51 */
52 struct fs_path {
53 union {
54 struct {
55 char *start;
56 char *end;
57
58 char *buf;
59 unsigned short buf_len:15;
60 unsigned short reversed:1;
61 char inline_buf[];
62 };
63 /*
64 * Average path length does not exceed 200 bytes, we'll have
65 * better packing in the slab and higher chance to satisfy
66 * a allocation later during send.
67 */
68 char pad[256];
69 };
70 };
71 #define FS_PATH_INLINE_SIZE \
72 (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
73
74
75 /* reused for each extent */
76 struct clone_root {
77 struct btrfs_root *root;
78 u64 ino;
79 u64 offset;
80 u64 num_bytes;
81 bool found_ref;
82 };
83
84 #define SEND_MAX_NAME_CACHE_SIZE 256
85
86 /*
87 * Limit the root_ids array of struct backref_cache_entry to 17 elements.
88 * This makes the size of a cache entry to be exactly 192 bytes on x86_64, which
89 * can be satisfied from the kmalloc-192 slab, without wasting any space.
90 * The most common case is to have a single root for cloning, which corresponds
91 * to the send root. Having the user specify more than 16 clone roots is not
92 * common, and in such rare cases we simply don't use caching if the number of
93 * cloning roots that lead down to a leaf is more than 17.
94 */
95 #define SEND_MAX_BACKREF_CACHE_ROOTS 17
96
97 /*
98 * Max number of entries in the cache.
99 * With SEND_MAX_BACKREF_CACHE_ROOTS as 17, the size in bytes, excluding
100 * maple tree's internal nodes, is 24K.
101 */
102 #define SEND_MAX_BACKREF_CACHE_SIZE 128
103
104 /*
105 * A backref cache entry maps a leaf to a list of IDs of roots from which the
106 * leaf is accessible and we can use for clone operations.
107 * With SEND_MAX_BACKREF_CACHE_ROOTS as 12, each cache entry is 128 bytes (on
108 * x86_64).
109 */
110 struct backref_cache_entry {
111 struct btrfs_lru_cache_entry entry;
112 u64 root_ids[SEND_MAX_BACKREF_CACHE_ROOTS];
113 /* Number of valid elements in the root_ids array. */
114 int num_roots;
115 };
116
117 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
118 static_assert(offsetof(struct backref_cache_entry, entry) == 0);
119
120 /*
121 * Max number of entries in the cache that stores directories that were already
122 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
123 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
124 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
125 */
126 #define SEND_MAX_DIR_CREATED_CACHE_SIZE 64
127
128 /*
129 * Max number of entries in the cache that stores directories that were already
130 * created. The cache uses raw struct btrfs_lru_cache_entry entries, so it uses
131 * at most 4096 bytes - sizeof(struct btrfs_lru_cache_entry) is 48 bytes, but
132 * the kmalloc-64 slab is used, so we get 4096 bytes (64 bytes * 64).
133 */
134 #define SEND_MAX_DIR_UTIMES_CACHE_SIZE 64
135
136 struct send_ctx {
137 struct file *send_filp;
138 loff_t send_off;
139 char *send_buf;
140 u32 send_size;
141 u32 send_max_size;
142 /*
143 * Whether BTRFS_SEND_A_DATA attribute was already added to current
144 * command (since protocol v2, data must be the last attribute).
145 */
146 bool put_data;
147 struct page **send_buf_pages;
148 u64 flags; /* 'flags' member of btrfs_ioctl_send_args is u64 */
149 /* Protocol version compatibility requested */
150 u32 proto;
151
152 struct btrfs_root *send_root;
153 struct btrfs_root *parent_root;
154 struct clone_root *clone_roots;
155 int clone_roots_cnt;
156
157 /* current state of the compare_tree call */
158 struct btrfs_path *left_path;
159 struct btrfs_path *right_path;
160 struct btrfs_key *cmp_key;
161
162 /*
163 * Keep track of the generation of the last transaction that was used
164 * for relocating a block group. This is periodically checked in order
165 * to detect if a relocation happened since the last check, so that we
166 * don't operate on stale extent buffers for nodes (level >= 1) or on
167 * stale disk_bytenr values of file extent items.
168 */
169 u64 last_reloc_trans;
170
171 /*
172 * infos of the currently processed inode. In case of deleted inodes,
173 * these are the values from the deleted inode.
174 */
175 u64 cur_ino;
176 u64 cur_inode_gen;
177 u64 cur_inode_size;
178 u64 cur_inode_mode;
179 u64 cur_inode_rdev;
180 u64 cur_inode_last_extent;
181 u64 cur_inode_next_write_offset;
182 bool cur_inode_new;
183 bool cur_inode_new_gen;
184 bool cur_inode_deleted;
185 bool ignore_cur_inode;
186 bool cur_inode_needs_verity;
187 void *verity_descriptor;
188
189 u64 send_progress;
190
191 struct list_head new_refs;
192 struct list_head deleted_refs;
193
194 struct btrfs_lru_cache name_cache;
195
196 /*
197 * The inode we are currently processing. It's not NULL only when we
198 * need to issue write commands for data extents from this inode.
199 */
200 struct inode *cur_inode;
201 struct file_ra_state ra;
202 u64 page_cache_clear_start;
203 bool clean_page_cache;
204
205 /*
206 * We process inodes by their increasing order, so if before an
207 * incremental send we reverse the parent/child relationship of
208 * directories such that a directory with a lower inode number was
209 * the parent of a directory with a higher inode number, and the one
210 * becoming the new parent got renamed too, we can't rename/move the
211 * directory with lower inode number when we finish processing it - we
212 * must process the directory with higher inode number first, then
213 * rename/move it and then rename/move the directory with lower inode
214 * number. Example follows.
215 *
216 * Tree state when the first send was performed:
217 *
218 * .
219 * |-- a (ino 257)
220 * |-- b (ino 258)
221 * |
222 * |
223 * |-- c (ino 259)
224 * | |-- d (ino 260)
225 * |
226 * |-- c2 (ino 261)
227 *
228 * Tree state when the second (incremental) send is performed:
229 *
230 * .
231 * |-- a (ino 257)
232 * |-- b (ino 258)
233 * |-- c2 (ino 261)
234 * |-- d2 (ino 260)
235 * |-- cc (ino 259)
236 *
237 * The sequence of steps that lead to the second state was:
238 *
239 * mv /a/b/c/d /a/b/c2/d2
240 * mv /a/b/c /a/b/c2/d2/cc
241 *
242 * "c" has lower inode number, but we can't move it (2nd mv operation)
243 * before we move "d", which has higher inode number.
244 *
245 * So we just memorize which move/rename operations must be performed
246 * later when their respective parent is processed and moved/renamed.
247 */
248
249 /* Indexed by parent directory inode number. */
250 struct rb_root pending_dir_moves;
251
252 /*
253 * Reverse index, indexed by the inode number of a directory that
254 * is waiting for the move/rename of its immediate parent before its
255 * own move/rename can be performed.
256 */
257 struct rb_root waiting_dir_moves;
258
259 /*
260 * A directory that is going to be rm'ed might have a child directory
261 * which is in the pending directory moves index above. In this case,
262 * the directory can only be removed after the move/rename of its child
263 * is performed. Example:
264 *
265 * Parent snapshot:
266 *
267 * . (ino 256)
268 * |-- a/ (ino 257)
269 * |-- b/ (ino 258)
270 * |-- c/ (ino 259)
271 * | |-- x/ (ino 260)
272 * |
273 * |-- y/ (ino 261)
274 *
275 * Send snapshot:
276 *
277 * . (ino 256)
278 * |-- a/ (ino 257)
279 * |-- b/ (ino 258)
280 * |-- YY/ (ino 261)
281 * |-- x/ (ino 260)
282 *
283 * Sequence of steps that lead to the send snapshot:
284 * rm -f /a/b/c/foo.txt
285 * mv /a/b/y /a/b/YY
286 * mv /a/b/c/x /a/b/YY
287 * rmdir /a/b/c
288 *
289 * When the child is processed, its move/rename is delayed until its
290 * parent is processed (as explained above), but all other operations
291 * like update utimes, chown, chgrp, etc, are performed and the paths
292 * that it uses for those operations must use the orphanized name of
293 * its parent (the directory we're going to rm later), so we need to
294 * memorize that name.
295 *
296 * Indexed by the inode number of the directory to be deleted.
297 */
298 struct rb_root orphan_dirs;
299
300 struct rb_root rbtree_new_refs;
301 struct rb_root rbtree_deleted_refs;
302
303 struct btrfs_lru_cache backref_cache;
304 u64 backref_cache_last_reloc_trans;
305
306 struct btrfs_lru_cache dir_created_cache;
307 struct btrfs_lru_cache dir_utimes_cache;
308 };
309
310 struct pending_dir_move {
311 struct rb_node node;
312 struct list_head list;
313 u64 parent_ino;
314 u64 ino;
315 u64 gen;
316 struct list_head update_refs;
317 };
318
319 struct waiting_dir_move {
320 struct rb_node node;
321 u64 ino;
322 /*
323 * There might be some directory that could not be removed because it
324 * was waiting for this directory inode to be moved first. Therefore
325 * after this directory is moved, we can try to rmdir the ino rmdir_ino.
326 */
327 u64 rmdir_ino;
328 u64 rmdir_gen;
329 bool orphanized;
330 };
331
332 struct orphan_dir_info {
333 struct rb_node node;
334 u64 ino;
335 u64 gen;
336 u64 last_dir_index_offset;
337 u64 dir_high_seq_ino;
338 };
339
340 struct name_cache_entry {
341 /*
342 * The key in the entry is an inode number, and the generation matches
343 * the inode's generation.
344 */
345 struct btrfs_lru_cache_entry entry;
346 u64 parent_ino;
347 u64 parent_gen;
348 int ret;
349 int need_later_update;
350 int name_len;
351 char name[];
352 };
353
354 /* See the comment at lru_cache.h about struct btrfs_lru_cache_entry. */
355 static_assert(offsetof(struct name_cache_entry, entry) == 0);
356
357 #define ADVANCE 1
358 #define ADVANCE_ONLY_NEXT -1
359
360 enum btrfs_compare_tree_result {
361 BTRFS_COMPARE_TREE_NEW,
362 BTRFS_COMPARE_TREE_DELETED,
363 BTRFS_COMPARE_TREE_CHANGED,
364 BTRFS_COMPARE_TREE_SAME,
365 };
366
367 __cold
inconsistent_snapshot_error(struct send_ctx * sctx,enum btrfs_compare_tree_result result,const char * what)368 static void inconsistent_snapshot_error(struct send_ctx *sctx,
369 enum btrfs_compare_tree_result result,
370 const char *what)
371 {
372 const char *result_string;
373
374 switch (result) {
375 case BTRFS_COMPARE_TREE_NEW:
376 result_string = "new";
377 break;
378 case BTRFS_COMPARE_TREE_DELETED:
379 result_string = "deleted";
380 break;
381 case BTRFS_COMPARE_TREE_CHANGED:
382 result_string = "updated";
383 break;
384 case BTRFS_COMPARE_TREE_SAME:
385 ASSERT(0);
386 result_string = "unchanged";
387 break;
388 default:
389 ASSERT(0);
390 result_string = "unexpected";
391 }
392
393 btrfs_err(sctx->send_root->fs_info,
394 "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
395 result_string, what, sctx->cmp_key->objectid,
396 sctx->send_root->root_key.objectid,
397 (sctx->parent_root ?
398 sctx->parent_root->root_key.objectid : 0));
399 }
400
401 __maybe_unused
proto_cmd_ok(const struct send_ctx * sctx,int cmd)402 static bool proto_cmd_ok(const struct send_ctx *sctx, int cmd)
403 {
404 switch (sctx->proto) {
405 case 1: return cmd <= BTRFS_SEND_C_MAX_V1;
406 case 2: return cmd <= BTRFS_SEND_C_MAX_V2;
407 case 3: return cmd <= BTRFS_SEND_C_MAX_V3;
408 default: return false;
409 }
410 }
411
412 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
413
414 static struct waiting_dir_move *
415 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
416
417 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen);
418
need_send_hole(struct send_ctx * sctx)419 static int need_send_hole(struct send_ctx *sctx)
420 {
421 return (sctx->parent_root && !sctx->cur_inode_new &&
422 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
423 S_ISREG(sctx->cur_inode_mode));
424 }
425
fs_path_reset(struct fs_path * p)426 static void fs_path_reset(struct fs_path *p)
427 {
428 if (p->reversed) {
429 p->start = p->buf + p->buf_len - 1;
430 p->end = p->start;
431 *p->start = 0;
432 } else {
433 p->start = p->buf;
434 p->end = p->start;
435 *p->start = 0;
436 }
437 }
438
fs_path_alloc(void)439 static struct fs_path *fs_path_alloc(void)
440 {
441 struct fs_path *p;
442
443 p = kmalloc(sizeof(*p), GFP_KERNEL);
444 if (!p)
445 return NULL;
446 p->reversed = 0;
447 p->buf = p->inline_buf;
448 p->buf_len = FS_PATH_INLINE_SIZE;
449 fs_path_reset(p);
450 return p;
451 }
452
fs_path_alloc_reversed(void)453 static struct fs_path *fs_path_alloc_reversed(void)
454 {
455 struct fs_path *p;
456
457 p = fs_path_alloc();
458 if (!p)
459 return NULL;
460 p->reversed = 1;
461 fs_path_reset(p);
462 return p;
463 }
464
fs_path_free(struct fs_path * p)465 static void fs_path_free(struct fs_path *p)
466 {
467 if (!p)
468 return;
469 if (p->buf != p->inline_buf)
470 kfree(p->buf);
471 kfree(p);
472 }
473
fs_path_len(struct fs_path * p)474 static int fs_path_len(struct fs_path *p)
475 {
476 return p->end - p->start;
477 }
478
fs_path_ensure_buf(struct fs_path * p,int len)479 static int fs_path_ensure_buf(struct fs_path *p, int len)
480 {
481 char *tmp_buf;
482 int path_len;
483 int old_buf_len;
484
485 len++;
486
487 if (p->buf_len >= len)
488 return 0;
489
490 if (len > PATH_MAX) {
491 WARN_ON(1);
492 return -ENOMEM;
493 }
494
495 path_len = p->end - p->start;
496 old_buf_len = p->buf_len;
497
498 /*
499 * Allocate to the next largest kmalloc bucket size, to let
500 * the fast path happen most of the time.
501 */
502 len = kmalloc_size_roundup(len);
503 /*
504 * First time the inline_buf does not suffice
505 */
506 if (p->buf == p->inline_buf) {
507 tmp_buf = kmalloc(len, GFP_KERNEL);
508 if (tmp_buf)
509 memcpy(tmp_buf, p->buf, old_buf_len);
510 } else {
511 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
512 }
513 if (!tmp_buf)
514 return -ENOMEM;
515 p->buf = tmp_buf;
516 p->buf_len = len;
517
518 if (p->reversed) {
519 tmp_buf = p->buf + old_buf_len - path_len - 1;
520 p->end = p->buf + p->buf_len - 1;
521 p->start = p->end - path_len;
522 memmove(p->start, tmp_buf, path_len + 1);
523 } else {
524 p->start = p->buf;
525 p->end = p->start + path_len;
526 }
527 return 0;
528 }
529
fs_path_prepare_for_add(struct fs_path * p,int name_len,char ** prepared)530 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
531 char **prepared)
532 {
533 int ret;
534 int new_len;
535
536 new_len = p->end - p->start + name_len;
537 if (p->start != p->end)
538 new_len++;
539 ret = fs_path_ensure_buf(p, new_len);
540 if (ret < 0)
541 goto out;
542
543 if (p->reversed) {
544 if (p->start != p->end)
545 *--p->start = '/';
546 p->start -= name_len;
547 *prepared = p->start;
548 } else {
549 if (p->start != p->end)
550 *p->end++ = '/';
551 *prepared = p->end;
552 p->end += name_len;
553 *p->end = 0;
554 }
555
556 out:
557 return ret;
558 }
559
fs_path_add(struct fs_path * p,const char * name,int name_len)560 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
561 {
562 int ret;
563 char *prepared;
564
565 ret = fs_path_prepare_for_add(p, name_len, &prepared);
566 if (ret < 0)
567 goto out;
568 memcpy(prepared, name, name_len);
569
570 out:
571 return ret;
572 }
573
fs_path_add_path(struct fs_path * p,struct fs_path * p2)574 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
575 {
576 int ret;
577 char *prepared;
578
579 ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
580 if (ret < 0)
581 goto out;
582 memcpy(prepared, p2->start, p2->end - p2->start);
583
584 out:
585 return ret;
586 }
587
fs_path_add_from_extent_buffer(struct fs_path * p,struct extent_buffer * eb,unsigned long off,int len)588 static int fs_path_add_from_extent_buffer(struct fs_path *p,
589 struct extent_buffer *eb,
590 unsigned long off, int len)
591 {
592 int ret;
593 char *prepared;
594
595 ret = fs_path_prepare_for_add(p, len, &prepared);
596 if (ret < 0)
597 goto out;
598
599 read_extent_buffer(eb, prepared, off, len);
600
601 out:
602 return ret;
603 }
604
fs_path_copy(struct fs_path * p,struct fs_path * from)605 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
606 {
607 p->reversed = from->reversed;
608 fs_path_reset(p);
609
610 return fs_path_add_path(p, from);
611 }
612
fs_path_unreverse(struct fs_path * p)613 static void fs_path_unreverse(struct fs_path *p)
614 {
615 char *tmp;
616 int len;
617
618 if (!p->reversed)
619 return;
620
621 tmp = p->start;
622 len = p->end - p->start;
623 p->start = p->buf;
624 p->end = p->start + len;
625 memmove(p->start, tmp, len + 1);
626 p->reversed = 0;
627 }
628
alloc_path_for_send(void)629 static struct btrfs_path *alloc_path_for_send(void)
630 {
631 struct btrfs_path *path;
632
633 path = btrfs_alloc_path();
634 if (!path)
635 return NULL;
636 path->search_commit_root = 1;
637 path->skip_locking = 1;
638 path->need_commit_sem = 1;
639 return path;
640 }
641
write_buf(struct file * filp,const void * buf,u32 len,loff_t * off)642 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
643 {
644 int ret;
645 u32 pos = 0;
646
647 while (pos < len) {
648 ret = kernel_write(filp, buf + pos, len - pos, off);
649 if (ret < 0)
650 return ret;
651 if (ret == 0)
652 return -EIO;
653 pos += ret;
654 }
655
656 return 0;
657 }
658
tlv_put(struct send_ctx * sctx,u16 attr,const void * data,int len)659 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
660 {
661 struct btrfs_tlv_header *hdr;
662 int total_len = sizeof(*hdr) + len;
663 int left = sctx->send_max_size - sctx->send_size;
664
665 if (WARN_ON_ONCE(sctx->put_data))
666 return -EINVAL;
667
668 if (unlikely(left < total_len))
669 return -EOVERFLOW;
670
671 hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
672 put_unaligned_le16(attr, &hdr->tlv_type);
673 put_unaligned_le16(len, &hdr->tlv_len);
674 memcpy(hdr + 1, data, len);
675 sctx->send_size += total_len;
676
677 return 0;
678 }
679
680 #define TLV_PUT_DEFINE_INT(bits) \
681 static int tlv_put_u##bits(struct send_ctx *sctx, \
682 u##bits attr, u##bits value) \
683 { \
684 __le##bits __tmp = cpu_to_le##bits(value); \
685 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp)); \
686 }
687
688 TLV_PUT_DEFINE_INT(8)
689 TLV_PUT_DEFINE_INT(32)
690 TLV_PUT_DEFINE_INT(64)
691
tlv_put_string(struct send_ctx * sctx,u16 attr,const char * str,int len)692 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
693 const char *str, int len)
694 {
695 if (len == -1)
696 len = strlen(str);
697 return tlv_put(sctx, attr, str, len);
698 }
699
tlv_put_uuid(struct send_ctx * sctx,u16 attr,const u8 * uuid)700 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
701 const u8 *uuid)
702 {
703 return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
704 }
705
tlv_put_btrfs_timespec(struct send_ctx * sctx,u16 attr,struct extent_buffer * eb,struct btrfs_timespec * ts)706 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
707 struct extent_buffer *eb,
708 struct btrfs_timespec *ts)
709 {
710 struct btrfs_timespec bts;
711 read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
712 return tlv_put(sctx, attr, &bts, sizeof(bts));
713 }
714
715
716 #define TLV_PUT(sctx, attrtype, data, attrlen) \
717 do { \
718 ret = tlv_put(sctx, attrtype, data, attrlen); \
719 if (ret < 0) \
720 goto tlv_put_failure; \
721 } while (0)
722
723 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
724 do { \
725 ret = tlv_put_u##bits(sctx, attrtype, value); \
726 if (ret < 0) \
727 goto tlv_put_failure; \
728 } while (0)
729
730 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
731 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
732 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
733 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
734 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
735 do { \
736 ret = tlv_put_string(sctx, attrtype, str, len); \
737 if (ret < 0) \
738 goto tlv_put_failure; \
739 } while (0)
740 #define TLV_PUT_PATH(sctx, attrtype, p) \
741 do { \
742 ret = tlv_put_string(sctx, attrtype, p->start, \
743 p->end - p->start); \
744 if (ret < 0) \
745 goto tlv_put_failure; \
746 } while(0)
747 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
748 do { \
749 ret = tlv_put_uuid(sctx, attrtype, uuid); \
750 if (ret < 0) \
751 goto tlv_put_failure; \
752 } while (0)
753 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
754 do { \
755 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
756 if (ret < 0) \
757 goto tlv_put_failure; \
758 } while (0)
759
send_header(struct send_ctx * sctx)760 static int send_header(struct send_ctx *sctx)
761 {
762 struct btrfs_stream_header hdr;
763
764 strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
765 hdr.version = cpu_to_le32(sctx->proto);
766 return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
767 &sctx->send_off);
768 }
769
770 /*
771 * For each command/item we want to send to userspace, we call this function.
772 */
begin_cmd(struct send_ctx * sctx,int cmd)773 static int begin_cmd(struct send_ctx *sctx, int cmd)
774 {
775 struct btrfs_cmd_header *hdr;
776
777 if (WARN_ON(!sctx->send_buf))
778 return -EINVAL;
779
780 BUG_ON(sctx->send_size);
781
782 sctx->send_size += sizeof(*hdr);
783 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
784 put_unaligned_le16(cmd, &hdr->cmd);
785
786 return 0;
787 }
788
send_cmd(struct send_ctx * sctx)789 static int send_cmd(struct send_ctx *sctx)
790 {
791 int ret;
792 struct btrfs_cmd_header *hdr;
793 u32 crc;
794
795 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
796 put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len);
797 put_unaligned_le32(0, &hdr->crc);
798
799 crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
800 put_unaligned_le32(crc, &hdr->crc);
801
802 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
803 &sctx->send_off);
804
805 sctx->send_size = 0;
806 sctx->put_data = false;
807
808 return ret;
809 }
810
811 /*
812 * Sends a move instruction to user space
813 */
send_rename(struct send_ctx * sctx,struct fs_path * from,struct fs_path * to)814 static int send_rename(struct send_ctx *sctx,
815 struct fs_path *from, struct fs_path *to)
816 {
817 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
818 int ret;
819
820 btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
821
822 ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
823 if (ret < 0)
824 goto out;
825
826 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
827 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
828
829 ret = send_cmd(sctx);
830
831 tlv_put_failure:
832 out:
833 return ret;
834 }
835
836 /*
837 * Sends a link instruction to user space
838 */
send_link(struct send_ctx * sctx,struct fs_path * path,struct fs_path * lnk)839 static int send_link(struct send_ctx *sctx,
840 struct fs_path *path, struct fs_path *lnk)
841 {
842 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
843 int ret;
844
845 btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
846
847 ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
848 if (ret < 0)
849 goto out;
850
851 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
852 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
853
854 ret = send_cmd(sctx);
855
856 tlv_put_failure:
857 out:
858 return ret;
859 }
860
861 /*
862 * Sends an unlink instruction to user space
863 */
send_unlink(struct send_ctx * sctx,struct fs_path * path)864 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
865 {
866 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
867 int ret;
868
869 btrfs_debug(fs_info, "send_unlink %s", path->start);
870
871 ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
872 if (ret < 0)
873 goto out;
874
875 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
876
877 ret = send_cmd(sctx);
878
879 tlv_put_failure:
880 out:
881 return ret;
882 }
883
884 /*
885 * Sends a rmdir instruction to user space
886 */
send_rmdir(struct send_ctx * sctx,struct fs_path * path)887 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
888 {
889 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
890 int ret;
891
892 btrfs_debug(fs_info, "send_rmdir %s", path->start);
893
894 ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
895 if (ret < 0)
896 goto out;
897
898 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
899
900 ret = send_cmd(sctx);
901
902 tlv_put_failure:
903 out:
904 return ret;
905 }
906
907 struct btrfs_inode_info {
908 u64 size;
909 u64 gen;
910 u64 mode;
911 u64 uid;
912 u64 gid;
913 u64 rdev;
914 u64 fileattr;
915 u64 nlink;
916 };
917
918 /*
919 * Helper function to retrieve some fields from an inode item.
920 */
get_inode_info(struct btrfs_root * root,u64 ino,struct btrfs_inode_info * info)921 static int get_inode_info(struct btrfs_root *root, u64 ino,
922 struct btrfs_inode_info *info)
923 {
924 int ret;
925 struct btrfs_path *path;
926 struct btrfs_inode_item *ii;
927 struct btrfs_key key;
928
929 path = alloc_path_for_send();
930 if (!path)
931 return -ENOMEM;
932
933 key.objectid = ino;
934 key.type = BTRFS_INODE_ITEM_KEY;
935 key.offset = 0;
936 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
937 if (ret) {
938 if (ret > 0)
939 ret = -ENOENT;
940 goto out;
941 }
942
943 if (!info)
944 goto out;
945
946 ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
947 struct btrfs_inode_item);
948 info->size = btrfs_inode_size(path->nodes[0], ii);
949 info->gen = btrfs_inode_generation(path->nodes[0], ii);
950 info->mode = btrfs_inode_mode(path->nodes[0], ii);
951 info->uid = btrfs_inode_uid(path->nodes[0], ii);
952 info->gid = btrfs_inode_gid(path->nodes[0], ii);
953 info->rdev = btrfs_inode_rdev(path->nodes[0], ii);
954 info->nlink = btrfs_inode_nlink(path->nodes[0], ii);
955 /*
956 * Transfer the unchanged u64 value of btrfs_inode_item::flags, that's
957 * otherwise logically split to 32/32 parts.
958 */
959 info->fileattr = btrfs_inode_flags(path->nodes[0], ii);
960
961 out:
962 btrfs_free_path(path);
963 return ret;
964 }
965
get_inode_gen(struct btrfs_root * root,u64 ino,u64 * gen)966 static int get_inode_gen(struct btrfs_root *root, u64 ino, u64 *gen)
967 {
968 int ret;
969 struct btrfs_inode_info info = { 0 };
970
971 ASSERT(gen);
972
973 ret = get_inode_info(root, ino, &info);
974 *gen = info.gen;
975 return ret;
976 }
977
978 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
979 struct fs_path *p,
980 void *ctx);
981
982 /*
983 * Helper function to iterate the entries in ONE btrfs_inode_ref or
984 * btrfs_inode_extref.
985 * The iterate callback may return a non zero value to stop iteration. This can
986 * be a negative value for error codes or 1 to simply stop it.
987 *
988 * path must point to the INODE_REF or INODE_EXTREF when called.
989 */
iterate_inode_ref(struct btrfs_root * root,struct btrfs_path * path,struct btrfs_key * found_key,int resolve,iterate_inode_ref_t iterate,void * ctx)990 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
991 struct btrfs_key *found_key, int resolve,
992 iterate_inode_ref_t iterate, void *ctx)
993 {
994 struct extent_buffer *eb = path->nodes[0];
995 struct btrfs_inode_ref *iref;
996 struct btrfs_inode_extref *extref;
997 struct btrfs_path *tmp_path;
998 struct fs_path *p;
999 u32 cur = 0;
1000 u32 total;
1001 int slot = path->slots[0];
1002 u32 name_len;
1003 char *start;
1004 int ret = 0;
1005 int num = 0;
1006 int index;
1007 u64 dir;
1008 unsigned long name_off;
1009 unsigned long elem_size;
1010 unsigned long ptr;
1011
1012 p = fs_path_alloc_reversed();
1013 if (!p)
1014 return -ENOMEM;
1015
1016 tmp_path = alloc_path_for_send();
1017 if (!tmp_path) {
1018 fs_path_free(p);
1019 return -ENOMEM;
1020 }
1021
1022
1023 if (found_key->type == BTRFS_INODE_REF_KEY) {
1024 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
1025 struct btrfs_inode_ref);
1026 total = btrfs_item_size(eb, slot);
1027 elem_size = sizeof(*iref);
1028 } else {
1029 ptr = btrfs_item_ptr_offset(eb, slot);
1030 total = btrfs_item_size(eb, slot);
1031 elem_size = sizeof(*extref);
1032 }
1033
1034 while (cur < total) {
1035 fs_path_reset(p);
1036
1037 if (found_key->type == BTRFS_INODE_REF_KEY) {
1038 iref = (struct btrfs_inode_ref *)(ptr + cur);
1039 name_len = btrfs_inode_ref_name_len(eb, iref);
1040 name_off = (unsigned long)(iref + 1);
1041 index = btrfs_inode_ref_index(eb, iref);
1042 dir = found_key->offset;
1043 } else {
1044 extref = (struct btrfs_inode_extref *)(ptr + cur);
1045 name_len = btrfs_inode_extref_name_len(eb, extref);
1046 name_off = (unsigned long)&extref->name;
1047 index = btrfs_inode_extref_index(eb, extref);
1048 dir = btrfs_inode_extref_parent(eb, extref);
1049 }
1050
1051 if (resolve) {
1052 start = btrfs_ref_to_path(root, tmp_path, name_len,
1053 name_off, eb, dir,
1054 p->buf, p->buf_len);
1055 if (IS_ERR(start)) {
1056 ret = PTR_ERR(start);
1057 goto out;
1058 }
1059 if (start < p->buf) {
1060 /* overflow , try again with larger buffer */
1061 ret = fs_path_ensure_buf(p,
1062 p->buf_len + p->buf - start);
1063 if (ret < 0)
1064 goto out;
1065 start = btrfs_ref_to_path(root, tmp_path,
1066 name_len, name_off,
1067 eb, dir,
1068 p->buf, p->buf_len);
1069 if (IS_ERR(start)) {
1070 ret = PTR_ERR(start);
1071 goto out;
1072 }
1073 if (unlikely(start < p->buf)) {
1074 btrfs_err(root->fs_info,
1075 "send: path ref buffer underflow for key (%llu %u %llu)",
1076 found_key->objectid,
1077 found_key->type,
1078 found_key->offset);
1079 ret = -EINVAL;
1080 goto out;
1081 }
1082 }
1083 p->start = start;
1084 } else {
1085 ret = fs_path_add_from_extent_buffer(p, eb, name_off,
1086 name_len);
1087 if (ret < 0)
1088 goto out;
1089 }
1090
1091 cur += elem_size + name_len;
1092 ret = iterate(num, dir, index, p, ctx);
1093 if (ret)
1094 goto out;
1095 num++;
1096 }
1097
1098 out:
1099 btrfs_free_path(tmp_path);
1100 fs_path_free(p);
1101 return ret;
1102 }
1103
1104 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
1105 const char *name, int name_len,
1106 const char *data, int data_len,
1107 void *ctx);
1108
1109 /*
1110 * Helper function to iterate the entries in ONE btrfs_dir_item.
1111 * The iterate callback may return a non zero value to stop iteration. This can
1112 * be a negative value for error codes or 1 to simply stop it.
1113 *
1114 * path must point to the dir item when called.
1115 */
iterate_dir_item(struct btrfs_root * root,struct btrfs_path * path,iterate_dir_item_t iterate,void * ctx)1116 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1117 iterate_dir_item_t iterate, void *ctx)
1118 {
1119 int ret = 0;
1120 struct extent_buffer *eb;
1121 struct btrfs_dir_item *di;
1122 struct btrfs_key di_key;
1123 char *buf = NULL;
1124 int buf_len;
1125 u32 name_len;
1126 u32 data_len;
1127 u32 cur;
1128 u32 len;
1129 u32 total;
1130 int slot;
1131 int num;
1132
1133 /*
1134 * Start with a small buffer (1 page). If later we end up needing more
1135 * space, which can happen for xattrs on a fs with a leaf size greater
1136 * then the page size, attempt to increase the buffer. Typically xattr
1137 * values are small.
1138 */
1139 buf_len = PATH_MAX;
1140 buf = kmalloc(buf_len, GFP_KERNEL);
1141 if (!buf) {
1142 ret = -ENOMEM;
1143 goto out;
1144 }
1145
1146 eb = path->nodes[0];
1147 slot = path->slots[0];
1148 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1149 cur = 0;
1150 len = 0;
1151 total = btrfs_item_size(eb, slot);
1152
1153 num = 0;
1154 while (cur < total) {
1155 name_len = btrfs_dir_name_len(eb, di);
1156 data_len = btrfs_dir_data_len(eb, di);
1157 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1158
1159 if (btrfs_dir_ftype(eb, di) == BTRFS_FT_XATTR) {
1160 if (name_len > XATTR_NAME_MAX) {
1161 ret = -ENAMETOOLONG;
1162 goto out;
1163 }
1164 if (name_len + data_len >
1165 BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1166 ret = -E2BIG;
1167 goto out;
1168 }
1169 } else {
1170 /*
1171 * Path too long
1172 */
1173 if (name_len + data_len > PATH_MAX) {
1174 ret = -ENAMETOOLONG;
1175 goto out;
1176 }
1177 }
1178
1179 if (name_len + data_len > buf_len) {
1180 buf_len = name_len + data_len;
1181 if (is_vmalloc_addr(buf)) {
1182 vfree(buf);
1183 buf = NULL;
1184 } else {
1185 char *tmp = krealloc(buf, buf_len,
1186 GFP_KERNEL | __GFP_NOWARN);
1187
1188 if (!tmp)
1189 kfree(buf);
1190 buf = tmp;
1191 }
1192 if (!buf) {
1193 buf = kvmalloc(buf_len, GFP_KERNEL);
1194 if (!buf) {
1195 ret = -ENOMEM;
1196 goto out;
1197 }
1198 }
1199 }
1200
1201 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1202 name_len + data_len);
1203
1204 len = sizeof(*di) + name_len + data_len;
1205 di = (struct btrfs_dir_item *)((char *)di + len);
1206 cur += len;
1207
1208 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1209 data_len, ctx);
1210 if (ret < 0)
1211 goto out;
1212 if (ret) {
1213 ret = 0;
1214 goto out;
1215 }
1216
1217 num++;
1218 }
1219
1220 out:
1221 kvfree(buf);
1222 return ret;
1223 }
1224
__copy_first_ref(int num,u64 dir,int index,struct fs_path * p,void * ctx)1225 static int __copy_first_ref(int num, u64 dir, int index,
1226 struct fs_path *p, void *ctx)
1227 {
1228 int ret;
1229 struct fs_path *pt = ctx;
1230
1231 ret = fs_path_copy(pt, p);
1232 if (ret < 0)
1233 return ret;
1234
1235 /* we want the first only */
1236 return 1;
1237 }
1238
1239 /*
1240 * Retrieve the first path of an inode. If an inode has more then one
1241 * ref/hardlink, this is ignored.
1242 */
get_inode_path(struct btrfs_root * root,u64 ino,struct fs_path * path)1243 static int get_inode_path(struct btrfs_root *root,
1244 u64 ino, struct fs_path *path)
1245 {
1246 int ret;
1247 struct btrfs_key key, found_key;
1248 struct btrfs_path *p;
1249
1250 p = alloc_path_for_send();
1251 if (!p)
1252 return -ENOMEM;
1253
1254 fs_path_reset(path);
1255
1256 key.objectid = ino;
1257 key.type = BTRFS_INODE_REF_KEY;
1258 key.offset = 0;
1259
1260 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1261 if (ret < 0)
1262 goto out;
1263 if (ret) {
1264 ret = 1;
1265 goto out;
1266 }
1267 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1268 if (found_key.objectid != ino ||
1269 (found_key.type != BTRFS_INODE_REF_KEY &&
1270 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1271 ret = -ENOENT;
1272 goto out;
1273 }
1274
1275 ret = iterate_inode_ref(root, p, &found_key, 1,
1276 __copy_first_ref, path);
1277 if (ret < 0)
1278 goto out;
1279 ret = 0;
1280
1281 out:
1282 btrfs_free_path(p);
1283 return ret;
1284 }
1285
1286 struct backref_ctx {
1287 struct send_ctx *sctx;
1288
1289 /* number of total found references */
1290 u64 found;
1291
1292 /*
1293 * used for clones found in send_root. clones found behind cur_objectid
1294 * and cur_offset are not considered as allowed clones.
1295 */
1296 u64 cur_objectid;
1297 u64 cur_offset;
1298
1299 /* may be truncated in case it's the last extent in a file */
1300 u64 extent_len;
1301
1302 /* The bytenr the file extent item we are processing refers to. */
1303 u64 bytenr;
1304 /* The owner (root id) of the data backref for the current extent. */
1305 u64 backref_owner;
1306 /* The offset of the data backref for the current extent. */
1307 u64 backref_offset;
1308 };
1309
__clone_root_cmp_bsearch(const void * key,const void * elt)1310 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1311 {
1312 u64 root = (u64)(uintptr_t)key;
1313 const struct clone_root *cr = elt;
1314
1315 if (root < cr->root->root_key.objectid)
1316 return -1;
1317 if (root > cr->root->root_key.objectid)
1318 return 1;
1319 return 0;
1320 }
1321
__clone_root_cmp_sort(const void * e1,const void * e2)1322 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1323 {
1324 const struct clone_root *cr1 = e1;
1325 const struct clone_root *cr2 = e2;
1326
1327 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1328 return -1;
1329 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1330 return 1;
1331 return 0;
1332 }
1333
1334 /*
1335 * Called for every backref that is found for the current extent.
1336 * Results are collected in sctx->clone_roots->ino/offset.
1337 */
iterate_backrefs(u64 ino,u64 offset,u64 num_bytes,u64 root_id,void * ctx_)1338 static int iterate_backrefs(u64 ino, u64 offset, u64 num_bytes, u64 root_id,
1339 void *ctx_)
1340 {
1341 struct backref_ctx *bctx = ctx_;
1342 struct clone_root *clone_root;
1343
1344 /* First check if the root is in the list of accepted clone sources */
1345 clone_root = bsearch((void *)(uintptr_t)root_id, bctx->sctx->clone_roots,
1346 bctx->sctx->clone_roots_cnt,
1347 sizeof(struct clone_root),
1348 __clone_root_cmp_bsearch);
1349 if (!clone_root)
1350 return 0;
1351
1352 /* This is our own reference, bail out as we can't clone from it. */
1353 if (clone_root->root == bctx->sctx->send_root &&
1354 ino == bctx->cur_objectid &&
1355 offset == bctx->cur_offset)
1356 return 0;
1357
1358 /*
1359 * Make sure we don't consider clones from send_root that are
1360 * behind the current inode/offset.
1361 */
1362 if (clone_root->root == bctx->sctx->send_root) {
1363 /*
1364 * If the source inode was not yet processed we can't issue a
1365 * clone operation, as the source extent does not exist yet at
1366 * the destination of the stream.
1367 */
1368 if (ino > bctx->cur_objectid)
1369 return 0;
1370 /*
1371 * We clone from the inode currently being sent as long as the
1372 * source extent is already processed, otherwise we could try
1373 * to clone from an extent that does not exist yet at the
1374 * destination of the stream.
1375 */
1376 if (ino == bctx->cur_objectid &&
1377 offset + bctx->extent_len >
1378 bctx->sctx->cur_inode_next_write_offset)
1379 return 0;
1380 }
1381
1382 bctx->found++;
1383 clone_root->found_ref = true;
1384
1385 /*
1386 * If the given backref refers to a file extent item with a larger
1387 * number of bytes than what we found before, use the new one so that
1388 * we clone more optimally and end up doing less writes and getting
1389 * less exclusive, non-shared extents at the destination.
1390 */
1391 if (num_bytes > clone_root->num_bytes) {
1392 clone_root->ino = ino;
1393 clone_root->offset = offset;
1394 clone_root->num_bytes = num_bytes;
1395
1396 /*
1397 * Found a perfect candidate, so there's no need to continue
1398 * backref walking.
1399 */
1400 if (num_bytes >= bctx->extent_len)
1401 return BTRFS_ITERATE_EXTENT_INODES_STOP;
1402 }
1403
1404 return 0;
1405 }
1406
lookup_backref_cache(u64 leaf_bytenr,void * ctx,const u64 ** root_ids_ret,int * root_count_ret)1407 static bool lookup_backref_cache(u64 leaf_bytenr, void *ctx,
1408 const u64 **root_ids_ret, int *root_count_ret)
1409 {
1410 struct backref_ctx *bctx = ctx;
1411 struct send_ctx *sctx = bctx->sctx;
1412 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1413 const u64 key = leaf_bytenr >> fs_info->sectorsize_bits;
1414 struct btrfs_lru_cache_entry *raw_entry;
1415 struct backref_cache_entry *entry;
1416
1417 if (btrfs_lru_cache_size(&sctx->backref_cache) == 0)
1418 return false;
1419
1420 /*
1421 * If relocation happened since we first filled the cache, then we must
1422 * empty the cache and can not use it, because even though we operate on
1423 * read-only roots, their leaves and nodes may have been reallocated and
1424 * now be used for different nodes/leaves of the same tree or some other
1425 * tree.
1426 *
1427 * We are called from iterate_extent_inodes() while either holding a
1428 * transaction handle or holding fs_info->commit_root_sem, so no need
1429 * to take any lock here.
1430 */
1431 if (fs_info->last_reloc_trans > sctx->backref_cache_last_reloc_trans) {
1432 btrfs_lru_cache_clear(&sctx->backref_cache);
1433 return false;
1434 }
1435
1436 raw_entry = btrfs_lru_cache_lookup(&sctx->backref_cache, key, 0);
1437 if (!raw_entry)
1438 return false;
1439
1440 entry = container_of(raw_entry, struct backref_cache_entry, entry);
1441 *root_ids_ret = entry->root_ids;
1442 *root_count_ret = entry->num_roots;
1443
1444 return true;
1445 }
1446
store_backref_cache(u64 leaf_bytenr,const struct ulist * root_ids,void * ctx)1447 static void store_backref_cache(u64 leaf_bytenr, const struct ulist *root_ids,
1448 void *ctx)
1449 {
1450 struct backref_ctx *bctx = ctx;
1451 struct send_ctx *sctx = bctx->sctx;
1452 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1453 struct backref_cache_entry *new_entry;
1454 struct ulist_iterator uiter;
1455 struct ulist_node *node;
1456 int ret;
1457
1458 /*
1459 * We're called while holding a transaction handle or while holding
1460 * fs_info->commit_root_sem (at iterate_extent_inodes()), so must do a
1461 * NOFS allocation.
1462 */
1463 new_entry = kmalloc(sizeof(struct backref_cache_entry), GFP_NOFS);
1464 /* No worries, cache is optional. */
1465 if (!new_entry)
1466 return;
1467
1468 new_entry->entry.key = leaf_bytenr >> fs_info->sectorsize_bits;
1469 new_entry->entry.gen = 0;
1470 new_entry->num_roots = 0;
1471 ULIST_ITER_INIT(&uiter);
1472 while ((node = ulist_next(root_ids, &uiter)) != NULL) {
1473 const u64 root_id = node->val;
1474 struct clone_root *root;
1475
1476 root = bsearch((void *)(uintptr_t)root_id, sctx->clone_roots,
1477 sctx->clone_roots_cnt, sizeof(struct clone_root),
1478 __clone_root_cmp_bsearch);
1479 if (!root)
1480 continue;
1481
1482 /* Too many roots, just exit, no worries as caching is optional. */
1483 if (new_entry->num_roots >= SEND_MAX_BACKREF_CACHE_ROOTS) {
1484 kfree(new_entry);
1485 return;
1486 }
1487
1488 new_entry->root_ids[new_entry->num_roots] = root_id;
1489 new_entry->num_roots++;
1490 }
1491
1492 /*
1493 * We may have not added any roots to the new cache entry, which means
1494 * none of the roots is part of the list of roots from which we are
1495 * allowed to clone. Cache the new entry as it's still useful to avoid
1496 * backref walking to determine which roots have a path to the leaf.
1497 *
1498 * Also use GFP_NOFS because we're called while holding a transaction
1499 * handle or while holding fs_info->commit_root_sem.
1500 */
1501 ret = btrfs_lru_cache_store(&sctx->backref_cache, &new_entry->entry,
1502 GFP_NOFS);
1503 ASSERT(ret == 0 || ret == -ENOMEM);
1504 if (ret) {
1505 /* Caching is optional, no worries. */
1506 kfree(new_entry);
1507 return;
1508 }
1509
1510 /*
1511 * We are called from iterate_extent_inodes() while either holding a
1512 * transaction handle or holding fs_info->commit_root_sem, so no need
1513 * to take any lock here.
1514 */
1515 if (btrfs_lru_cache_size(&sctx->backref_cache) == 1)
1516 sctx->backref_cache_last_reloc_trans = fs_info->last_reloc_trans;
1517 }
1518
check_extent_item(u64 bytenr,const struct btrfs_extent_item * ei,const struct extent_buffer * leaf,void * ctx)1519 static int check_extent_item(u64 bytenr, const struct btrfs_extent_item *ei,
1520 const struct extent_buffer *leaf, void *ctx)
1521 {
1522 const u64 refs = btrfs_extent_refs(leaf, ei);
1523 const struct backref_ctx *bctx = ctx;
1524 const struct send_ctx *sctx = bctx->sctx;
1525
1526 if (bytenr == bctx->bytenr) {
1527 const u64 flags = btrfs_extent_flags(leaf, ei);
1528
1529 if (WARN_ON(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK))
1530 return -EUCLEAN;
1531
1532 /*
1533 * If we have only one reference and only the send root as a
1534 * clone source - meaning no clone roots were given in the
1535 * struct btrfs_ioctl_send_args passed to the send ioctl - then
1536 * it's our reference and there's no point in doing backref
1537 * walking which is expensive, so exit early.
1538 */
1539 if (refs == 1 && sctx->clone_roots_cnt == 1)
1540 return -ENOENT;
1541 }
1542
1543 /*
1544 * Backreference walking (iterate_extent_inodes() below) is currently
1545 * too expensive when an extent has a large number of references, both
1546 * in time spent and used memory. So for now just fallback to write
1547 * operations instead of clone operations when an extent has more than
1548 * a certain amount of references.
1549 */
1550 if (refs > SEND_MAX_EXTENT_REFS)
1551 return -ENOENT;
1552
1553 return 0;
1554 }
1555
skip_self_data_ref(u64 root,u64 ino,u64 offset,void * ctx)1556 static bool skip_self_data_ref(u64 root, u64 ino, u64 offset, void *ctx)
1557 {
1558 const struct backref_ctx *bctx = ctx;
1559
1560 if (ino == bctx->cur_objectid &&
1561 root == bctx->backref_owner &&
1562 offset == bctx->backref_offset)
1563 return true;
1564
1565 return false;
1566 }
1567
1568 /*
1569 * Given an inode, offset and extent item, it finds a good clone for a clone
1570 * instruction. Returns -ENOENT when none could be found. The function makes
1571 * sure that the returned clone is usable at the point where sending is at the
1572 * moment. This means, that no clones are accepted which lie behind the current
1573 * inode+offset.
1574 *
1575 * path must point to the extent item when called.
1576 */
find_extent_clone(struct send_ctx * sctx,struct btrfs_path * path,u64 ino,u64 data_offset,u64 ino_size,struct clone_root ** found)1577 static int find_extent_clone(struct send_ctx *sctx,
1578 struct btrfs_path *path,
1579 u64 ino, u64 data_offset,
1580 u64 ino_size,
1581 struct clone_root **found)
1582 {
1583 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1584 int ret;
1585 int extent_type;
1586 u64 logical;
1587 u64 disk_byte;
1588 u64 num_bytes;
1589 struct btrfs_file_extent_item *fi;
1590 struct extent_buffer *eb = path->nodes[0];
1591 struct backref_ctx backref_ctx = { 0 };
1592 struct btrfs_backref_walk_ctx backref_walk_ctx = { 0 };
1593 struct clone_root *cur_clone_root;
1594 int compressed;
1595 u32 i;
1596
1597 /*
1598 * With fallocate we can get prealloc extents beyond the inode's i_size,
1599 * so we don't do anything here because clone operations can not clone
1600 * to a range beyond i_size without increasing the i_size of the
1601 * destination inode.
1602 */
1603 if (data_offset >= ino_size)
1604 return 0;
1605
1606 fi = btrfs_item_ptr(eb, path->slots[0], struct btrfs_file_extent_item);
1607 extent_type = btrfs_file_extent_type(eb, fi);
1608 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1609 return -ENOENT;
1610
1611 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1612 if (disk_byte == 0)
1613 return -ENOENT;
1614
1615 compressed = btrfs_file_extent_compression(eb, fi);
1616 num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1617 logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1618
1619 /*
1620 * Setup the clone roots.
1621 */
1622 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1623 cur_clone_root = sctx->clone_roots + i;
1624 cur_clone_root->ino = (u64)-1;
1625 cur_clone_root->offset = 0;
1626 cur_clone_root->num_bytes = 0;
1627 cur_clone_root->found_ref = false;
1628 }
1629
1630 backref_ctx.sctx = sctx;
1631 backref_ctx.cur_objectid = ino;
1632 backref_ctx.cur_offset = data_offset;
1633 backref_ctx.bytenr = disk_byte;
1634 /*
1635 * Use the header owner and not the send root's id, because in case of a
1636 * snapshot we can have shared subtrees.
1637 */
1638 backref_ctx.backref_owner = btrfs_header_owner(eb);
1639 backref_ctx.backref_offset = data_offset - btrfs_file_extent_offset(eb, fi);
1640
1641 /*
1642 * The last extent of a file may be too large due to page alignment.
1643 * We need to adjust extent_len in this case so that the checks in
1644 * iterate_backrefs() work.
1645 */
1646 if (data_offset + num_bytes >= ino_size)
1647 backref_ctx.extent_len = ino_size - data_offset;
1648 else
1649 backref_ctx.extent_len = num_bytes;
1650
1651 /*
1652 * Now collect all backrefs.
1653 */
1654 backref_walk_ctx.bytenr = disk_byte;
1655 if (compressed == BTRFS_COMPRESS_NONE)
1656 backref_walk_ctx.extent_item_pos = btrfs_file_extent_offset(eb, fi);
1657 backref_walk_ctx.fs_info = fs_info;
1658 backref_walk_ctx.cache_lookup = lookup_backref_cache;
1659 backref_walk_ctx.cache_store = store_backref_cache;
1660 backref_walk_ctx.indirect_ref_iterator = iterate_backrefs;
1661 backref_walk_ctx.check_extent_item = check_extent_item;
1662 backref_walk_ctx.user_ctx = &backref_ctx;
1663
1664 /*
1665 * If have a single clone root, then it's the send root and we can tell
1666 * the backref walking code to skip our own backref and not resolve it,
1667 * since we can not use it for cloning - the source and destination
1668 * ranges can't overlap and in case the leaf is shared through a subtree
1669 * due to snapshots, we can't use those other roots since they are not
1670 * in the list of clone roots.
1671 */
1672 if (sctx->clone_roots_cnt == 1)
1673 backref_walk_ctx.skip_data_ref = skip_self_data_ref;
1674
1675 ret = iterate_extent_inodes(&backref_walk_ctx, true, iterate_backrefs,
1676 &backref_ctx);
1677 if (ret < 0)
1678 return ret;
1679
1680 down_read(&fs_info->commit_root_sem);
1681 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
1682 /*
1683 * A transaction commit for a transaction in which block group
1684 * relocation was done just happened.
1685 * The disk_bytenr of the file extent item we processed is
1686 * possibly stale, referring to the extent's location before
1687 * relocation. So act as if we haven't found any clone sources
1688 * and fallback to write commands, which will read the correct
1689 * data from the new extent location. Otherwise we will fail
1690 * below because we haven't found our own back reference or we
1691 * could be getting incorrect sources in case the old extent
1692 * was already reallocated after the relocation.
1693 */
1694 up_read(&fs_info->commit_root_sem);
1695 return -ENOENT;
1696 }
1697 up_read(&fs_info->commit_root_sem);
1698
1699 btrfs_debug(fs_info,
1700 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1701 data_offset, ino, num_bytes, logical);
1702
1703 if (!backref_ctx.found) {
1704 btrfs_debug(fs_info, "no clones found");
1705 return -ENOENT;
1706 }
1707
1708 cur_clone_root = NULL;
1709 for (i = 0; i < sctx->clone_roots_cnt; i++) {
1710 struct clone_root *clone_root = &sctx->clone_roots[i];
1711
1712 if (!clone_root->found_ref)
1713 continue;
1714
1715 /*
1716 * Choose the root from which we can clone more bytes, to
1717 * minimize write operations and therefore have more extent
1718 * sharing at the destination (the same as in the source).
1719 */
1720 if (!cur_clone_root ||
1721 clone_root->num_bytes > cur_clone_root->num_bytes) {
1722 cur_clone_root = clone_root;
1723
1724 /*
1725 * We found an optimal clone candidate (any inode from
1726 * any root is fine), so we're done.
1727 */
1728 if (clone_root->num_bytes >= backref_ctx.extent_len)
1729 break;
1730 }
1731 }
1732
1733 if (cur_clone_root) {
1734 *found = cur_clone_root;
1735 ret = 0;
1736 } else {
1737 ret = -ENOENT;
1738 }
1739
1740 return ret;
1741 }
1742
read_symlink(struct btrfs_root * root,u64 ino,struct fs_path * dest)1743 static int read_symlink(struct btrfs_root *root,
1744 u64 ino,
1745 struct fs_path *dest)
1746 {
1747 int ret;
1748 struct btrfs_path *path;
1749 struct btrfs_key key;
1750 struct btrfs_file_extent_item *ei;
1751 u8 type;
1752 u8 compression;
1753 unsigned long off;
1754 int len;
1755
1756 path = alloc_path_for_send();
1757 if (!path)
1758 return -ENOMEM;
1759
1760 key.objectid = ino;
1761 key.type = BTRFS_EXTENT_DATA_KEY;
1762 key.offset = 0;
1763 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1764 if (ret < 0)
1765 goto out;
1766 if (ret) {
1767 /*
1768 * An empty symlink inode. Can happen in rare error paths when
1769 * creating a symlink (transaction committed before the inode
1770 * eviction handler removed the symlink inode items and a crash
1771 * happened in between or the subvol was snapshoted in between).
1772 * Print an informative message to dmesg/syslog so that the user
1773 * can delete the symlink.
1774 */
1775 btrfs_err(root->fs_info,
1776 "Found empty symlink inode %llu at root %llu",
1777 ino, root->root_key.objectid);
1778 ret = -EIO;
1779 goto out;
1780 }
1781
1782 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1783 struct btrfs_file_extent_item);
1784 type = btrfs_file_extent_type(path->nodes[0], ei);
1785 if (unlikely(type != BTRFS_FILE_EXTENT_INLINE)) {
1786 ret = -EUCLEAN;
1787 btrfs_crit(root->fs_info,
1788 "send: found symlink extent that is not inline, ino %llu root %llu extent type %d",
1789 ino, btrfs_root_id(root), type);
1790 goto out;
1791 }
1792 compression = btrfs_file_extent_compression(path->nodes[0], ei);
1793 if (unlikely(compression != BTRFS_COMPRESS_NONE)) {
1794 ret = -EUCLEAN;
1795 btrfs_crit(root->fs_info,
1796 "send: found symlink extent with compression, ino %llu root %llu compression type %d",
1797 ino, btrfs_root_id(root), compression);
1798 goto out;
1799 }
1800
1801 off = btrfs_file_extent_inline_start(ei);
1802 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1803
1804 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1805
1806 out:
1807 btrfs_free_path(path);
1808 return ret;
1809 }
1810
1811 /*
1812 * Helper function to generate a file name that is unique in the root of
1813 * send_root and parent_root. This is used to generate names for orphan inodes.
1814 */
gen_unique_name(struct send_ctx * sctx,u64 ino,u64 gen,struct fs_path * dest)1815 static int gen_unique_name(struct send_ctx *sctx,
1816 u64 ino, u64 gen,
1817 struct fs_path *dest)
1818 {
1819 int ret = 0;
1820 struct btrfs_path *path;
1821 struct btrfs_dir_item *di;
1822 char tmp[64];
1823 int len;
1824 u64 idx = 0;
1825
1826 path = alloc_path_for_send();
1827 if (!path)
1828 return -ENOMEM;
1829
1830 while (1) {
1831 struct fscrypt_str tmp_name;
1832
1833 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1834 ino, gen, idx);
1835 ASSERT(len < sizeof(tmp));
1836 tmp_name.name = tmp;
1837 tmp_name.len = strlen(tmp);
1838
1839 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1840 path, BTRFS_FIRST_FREE_OBJECTID,
1841 &tmp_name, 0);
1842 btrfs_release_path(path);
1843 if (IS_ERR(di)) {
1844 ret = PTR_ERR(di);
1845 goto out;
1846 }
1847 if (di) {
1848 /* not unique, try again */
1849 idx++;
1850 continue;
1851 }
1852
1853 if (!sctx->parent_root) {
1854 /* unique */
1855 ret = 0;
1856 break;
1857 }
1858
1859 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1860 path, BTRFS_FIRST_FREE_OBJECTID,
1861 &tmp_name, 0);
1862 btrfs_release_path(path);
1863 if (IS_ERR(di)) {
1864 ret = PTR_ERR(di);
1865 goto out;
1866 }
1867 if (di) {
1868 /* not unique, try again */
1869 idx++;
1870 continue;
1871 }
1872 /* unique */
1873 break;
1874 }
1875
1876 ret = fs_path_add(dest, tmp, strlen(tmp));
1877
1878 out:
1879 btrfs_free_path(path);
1880 return ret;
1881 }
1882
1883 enum inode_state {
1884 inode_state_no_change,
1885 inode_state_will_create,
1886 inode_state_did_create,
1887 inode_state_will_delete,
1888 inode_state_did_delete,
1889 };
1890
get_cur_inode_state(struct send_ctx * sctx,u64 ino,u64 gen,u64 * send_gen,u64 * parent_gen)1891 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen,
1892 u64 *send_gen, u64 *parent_gen)
1893 {
1894 int ret;
1895 int left_ret;
1896 int right_ret;
1897 u64 left_gen;
1898 u64 right_gen = 0;
1899 struct btrfs_inode_info info;
1900
1901 ret = get_inode_info(sctx->send_root, ino, &info);
1902 if (ret < 0 && ret != -ENOENT)
1903 goto out;
1904 left_ret = (info.nlink == 0) ? -ENOENT : ret;
1905 left_gen = info.gen;
1906 if (send_gen)
1907 *send_gen = ((left_ret == -ENOENT) ? 0 : info.gen);
1908
1909 if (!sctx->parent_root) {
1910 right_ret = -ENOENT;
1911 } else {
1912 ret = get_inode_info(sctx->parent_root, ino, &info);
1913 if (ret < 0 && ret != -ENOENT)
1914 goto out;
1915 right_ret = (info.nlink == 0) ? -ENOENT : ret;
1916 right_gen = info.gen;
1917 if (parent_gen)
1918 *parent_gen = ((right_ret == -ENOENT) ? 0 : info.gen);
1919 }
1920
1921 if (!left_ret && !right_ret) {
1922 if (left_gen == gen && right_gen == gen) {
1923 ret = inode_state_no_change;
1924 } else if (left_gen == gen) {
1925 if (ino < sctx->send_progress)
1926 ret = inode_state_did_create;
1927 else
1928 ret = inode_state_will_create;
1929 } else if (right_gen == gen) {
1930 if (ino < sctx->send_progress)
1931 ret = inode_state_did_delete;
1932 else
1933 ret = inode_state_will_delete;
1934 } else {
1935 ret = -ENOENT;
1936 }
1937 } else if (!left_ret) {
1938 if (left_gen == gen) {
1939 if (ino < sctx->send_progress)
1940 ret = inode_state_did_create;
1941 else
1942 ret = inode_state_will_create;
1943 } else {
1944 ret = -ENOENT;
1945 }
1946 } else if (!right_ret) {
1947 if (right_gen == gen) {
1948 if (ino < sctx->send_progress)
1949 ret = inode_state_did_delete;
1950 else
1951 ret = inode_state_will_delete;
1952 } else {
1953 ret = -ENOENT;
1954 }
1955 } else {
1956 ret = -ENOENT;
1957 }
1958
1959 out:
1960 return ret;
1961 }
1962
is_inode_existent(struct send_ctx * sctx,u64 ino,u64 gen,u64 * send_gen,u64 * parent_gen)1963 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen,
1964 u64 *send_gen, u64 *parent_gen)
1965 {
1966 int ret;
1967
1968 if (ino == BTRFS_FIRST_FREE_OBJECTID)
1969 return 1;
1970
1971 ret = get_cur_inode_state(sctx, ino, gen, send_gen, parent_gen);
1972 if (ret < 0)
1973 goto out;
1974
1975 if (ret == inode_state_no_change ||
1976 ret == inode_state_did_create ||
1977 ret == inode_state_will_delete)
1978 ret = 1;
1979 else
1980 ret = 0;
1981
1982 out:
1983 return ret;
1984 }
1985
1986 /*
1987 * Helper function to lookup a dir item in a dir.
1988 */
lookup_dir_item_inode(struct btrfs_root * root,u64 dir,const char * name,int name_len,u64 * found_inode)1989 static int lookup_dir_item_inode(struct btrfs_root *root,
1990 u64 dir, const char *name, int name_len,
1991 u64 *found_inode)
1992 {
1993 int ret = 0;
1994 struct btrfs_dir_item *di;
1995 struct btrfs_key key;
1996 struct btrfs_path *path;
1997 struct fscrypt_str name_str = FSTR_INIT((char *)name, name_len);
1998
1999 path = alloc_path_for_send();
2000 if (!path)
2001 return -ENOMEM;
2002
2003 di = btrfs_lookup_dir_item(NULL, root, path, dir, &name_str, 0);
2004 if (IS_ERR_OR_NULL(di)) {
2005 ret = di ? PTR_ERR(di) : -ENOENT;
2006 goto out;
2007 }
2008 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
2009 if (key.type == BTRFS_ROOT_ITEM_KEY) {
2010 ret = -ENOENT;
2011 goto out;
2012 }
2013 *found_inode = key.objectid;
2014
2015 out:
2016 btrfs_free_path(path);
2017 return ret;
2018 }
2019
2020 /*
2021 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
2022 * generation of the parent dir and the name of the dir entry.
2023 */
get_first_ref(struct btrfs_root * root,u64 ino,u64 * dir,u64 * dir_gen,struct fs_path * name)2024 static int get_first_ref(struct btrfs_root *root, u64 ino,
2025 u64 *dir, u64 *dir_gen, struct fs_path *name)
2026 {
2027 int ret;
2028 struct btrfs_key key;
2029 struct btrfs_key found_key;
2030 struct btrfs_path *path;
2031 int len;
2032 u64 parent_dir;
2033
2034 path = alloc_path_for_send();
2035 if (!path)
2036 return -ENOMEM;
2037
2038 key.objectid = ino;
2039 key.type = BTRFS_INODE_REF_KEY;
2040 key.offset = 0;
2041
2042 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
2043 if (ret < 0)
2044 goto out;
2045 if (!ret)
2046 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2047 path->slots[0]);
2048 if (ret || found_key.objectid != ino ||
2049 (found_key.type != BTRFS_INODE_REF_KEY &&
2050 found_key.type != BTRFS_INODE_EXTREF_KEY)) {
2051 ret = -ENOENT;
2052 goto out;
2053 }
2054
2055 if (found_key.type == BTRFS_INODE_REF_KEY) {
2056 struct btrfs_inode_ref *iref;
2057 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2058 struct btrfs_inode_ref);
2059 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
2060 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2061 (unsigned long)(iref + 1),
2062 len);
2063 parent_dir = found_key.offset;
2064 } else {
2065 struct btrfs_inode_extref *extref;
2066 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
2067 struct btrfs_inode_extref);
2068 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
2069 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
2070 (unsigned long)&extref->name, len);
2071 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
2072 }
2073 if (ret < 0)
2074 goto out;
2075 btrfs_release_path(path);
2076
2077 if (dir_gen) {
2078 ret = get_inode_gen(root, parent_dir, dir_gen);
2079 if (ret < 0)
2080 goto out;
2081 }
2082
2083 *dir = parent_dir;
2084
2085 out:
2086 btrfs_free_path(path);
2087 return ret;
2088 }
2089
is_first_ref(struct btrfs_root * root,u64 ino,u64 dir,const char * name,int name_len)2090 static int is_first_ref(struct btrfs_root *root,
2091 u64 ino, u64 dir,
2092 const char *name, int name_len)
2093 {
2094 int ret;
2095 struct fs_path *tmp_name;
2096 u64 tmp_dir;
2097
2098 tmp_name = fs_path_alloc();
2099 if (!tmp_name)
2100 return -ENOMEM;
2101
2102 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
2103 if (ret < 0)
2104 goto out;
2105
2106 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
2107 ret = 0;
2108 goto out;
2109 }
2110
2111 ret = !memcmp(tmp_name->start, name, name_len);
2112
2113 out:
2114 fs_path_free(tmp_name);
2115 return ret;
2116 }
2117
2118 /*
2119 * Used by process_recorded_refs to determine if a new ref would overwrite an
2120 * already existing ref. In case it detects an overwrite, it returns the
2121 * inode/gen in who_ino/who_gen.
2122 * When an overwrite is detected, process_recorded_refs does proper orphanizing
2123 * to make sure later references to the overwritten inode are possible.
2124 * Orphanizing is however only required for the first ref of an inode.
2125 * process_recorded_refs does an additional is_first_ref check to see if
2126 * orphanizing is really required.
2127 */
will_overwrite_ref(struct send_ctx * sctx,u64 dir,u64 dir_gen,const char * name,int name_len,u64 * who_ino,u64 * who_gen,u64 * who_mode)2128 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2129 const char *name, int name_len,
2130 u64 *who_ino, u64 *who_gen, u64 *who_mode)
2131 {
2132 int ret;
2133 u64 parent_root_dir_gen;
2134 u64 other_inode = 0;
2135 struct btrfs_inode_info info;
2136
2137 if (!sctx->parent_root)
2138 return 0;
2139
2140 ret = is_inode_existent(sctx, dir, dir_gen, NULL, &parent_root_dir_gen);
2141 if (ret <= 0)
2142 return 0;
2143
2144 /*
2145 * If we have a parent root we need to verify that the parent dir was
2146 * not deleted and then re-created, if it was then we have no overwrite
2147 * and we can just unlink this entry.
2148 *
2149 * @parent_root_dir_gen was set to 0 if the inode does not exist in the
2150 * parent root.
2151 */
2152 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID &&
2153 parent_root_dir_gen != dir_gen)
2154 return 0;
2155
2156 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
2157 &other_inode);
2158 if (ret == -ENOENT)
2159 return 0;
2160 else if (ret < 0)
2161 return ret;
2162
2163 /*
2164 * Check if the overwritten ref was already processed. If yes, the ref
2165 * was already unlinked/moved, so we can safely assume that we will not
2166 * overwrite anything at this point in time.
2167 */
2168 if (other_inode > sctx->send_progress ||
2169 is_waiting_for_move(sctx, other_inode)) {
2170 ret = get_inode_info(sctx->parent_root, other_inode, &info);
2171 if (ret < 0)
2172 return ret;
2173
2174 *who_ino = other_inode;
2175 *who_gen = info.gen;
2176 *who_mode = info.mode;
2177 return 1;
2178 }
2179
2180 return 0;
2181 }
2182
2183 /*
2184 * Checks if the ref was overwritten by an already processed inode. This is
2185 * used by __get_cur_name_and_parent to find out if the ref was orphanized and
2186 * thus the orphan name needs be used.
2187 * process_recorded_refs also uses it to avoid unlinking of refs that were
2188 * overwritten.
2189 */
did_overwrite_ref(struct send_ctx * sctx,u64 dir,u64 dir_gen,u64 ino,u64 ino_gen,const char * name,int name_len)2190 static int did_overwrite_ref(struct send_ctx *sctx,
2191 u64 dir, u64 dir_gen,
2192 u64 ino, u64 ino_gen,
2193 const char *name, int name_len)
2194 {
2195 int ret;
2196 u64 ow_inode;
2197 u64 ow_gen = 0;
2198 u64 send_root_dir_gen;
2199
2200 if (!sctx->parent_root)
2201 return 0;
2202
2203 ret = is_inode_existent(sctx, dir, dir_gen, &send_root_dir_gen, NULL);
2204 if (ret <= 0)
2205 return ret;
2206
2207 /*
2208 * @send_root_dir_gen was set to 0 if the inode does not exist in the
2209 * send root.
2210 */
2211 if (dir != BTRFS_FIRST_FREE_OBJECTID && send_root_dir_gen != dir_gen)
2212 return 0;
2213
2214 /* check if the ref was overwritten by another ref */
2215 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
2216 &ow_inode);
2217 if (ret == -ENOENT) {
2218 /* was never and will never be overwritten */
2219 return 0;
2220 } else if (ret < 0) {
2221 return ret;
2222 }
2223
2224 if (ow_inode == ino) {
2225 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2226 if (ret < 0)
2227 return ret;
2228
2229 /* It's the same inode, so no overwrite happened. */
2230 if (ow_gen == ino_gen)
2231 return 0;
2232 }
2233
2234 /*
2235 * We know that it is or will be overwritten. Check this now.
2236 * The current inode being processed might have been the one that caused
2237 * inode 'ino' to be orphanized, therefore check if ow_inode matches
2238 * the current inode being processed.
2239 */
2240 if (ow_inode < sctx->send_progress)
2241 return 1;
2242
2243 if (ino != sctx->cur_ino && ow_inode == sctx->cur_ino) {
2244 if (ow_gen == 0) {
2245 ret = get_inode_gen(sctx->send_root, ow_inode, &ow_gen);
2246 if (ret < 0)
2247 return ret;
2248 }
2249 if (ow_gen == sctx->cur_inode_gen)
2250 return 1;
2251 }
2252
2253 return 0;
2254 }
2255
2256 /*
2257 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
2258 * that got overwritten. This is used by process_recorded_refs to determine
2259 * if it has to use the path as returned by get_cur_path or the orphan name.
2260 */
did_overwrite_first_ref(struct send_ctx * sctx,u64 ino,u64 gen)2261 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
2262 {
2263 int ret = 0;
2264 struct fs_path *name = NULL;
2265 u64 dir;
2266 u64 dir_gen;
2267
2268 if (!sctx->parent_root)
2269 goto out;
2270
2271 name = fs_path_alloc();
2272 if (!name)
2273 return -ENOMEM;
2274
2275 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2276 if (ret < 0)
2277 goto out;
2278
2279 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2280 name->start, fs_path_len(name));
2281
2282 out:
2283 fs_path_free(name);
2284 return ret;
2285 }
2286
name_cache_search(struct send_ctx * sctx,u64 ino,u64 gen)2287 static inline struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2288 u64 ino, u64 gen)
2289 {
2290 struct btrfs_lru_cache_entry *entry;
2291
2292 entry = btrfs_lru_cache_lookup(&sctx->name_cache, ino, gen);
2293 if (!entry)
2294 return NULL;
2295
2296 return container_of(entry, struct name_cache_entry, entry);
2297 }
2298
2299 /*
2300 * Used by get_cur_path for each ref up to the root.
2301 * Returns 0 if it succeeded.
2302 * Returns 1 if the inode is not existent or got overwritten. In that case, the
2303 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2304 * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2305 * Returns <0 in case of error.
2306 */
__get_cur_name_and_parent(struct send_ctx * sctx,u64 ino,u64 gen,u64 * parent_ino,u64 * parent_gen,struct fs_path * dest)2307 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2308 u64 ino, u64 gen,
2309 u64 *parent_ino,
2310 u64 *parent_gen,
2311 struct fs_path *dest)
2312 {
2313 int ret;
2314 int nce_ret;
2315 struct name_cache_entry *nce;
2316
2317 /*
2318 * First check if we already did a call to this function with the same
2319 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2320 * return the cached result.
2321 */
2322 nce = name_cache_search(sctx, ino, gen);
2323 if (nce) {
2324 if (ino < sctx->send_progress && nce->need_later_update) {
2325 btrfs_lru_cache_remove(&sctx->name_cache, &nce->entry);
2326 nce = NULL;
2327 } else {
2328 *parent_ino = nce->parent_ino;
2329 *parent_gen = nce->parent_gen;
2330 ret = fs_path_add(dest, nce->name, nce->name_len);
2331 if (ret < 0)
2332 goto out;
2333 ret = nce->ret;
2334 goto out;
2335 }
2336 }
2337
2338 /*
2339 * If the inode is not existent yet, add the orphan name and return 1.
2340 * This should only happen for the parent dir that we determine in
2341 * record_new_ref_if_needed().
2342 */
2343 ret = is_inode_existent(sctx, ino, gen, NULL, NULL);
2344 if (ret < 0)
2345 goto out;
2346
2347 if (!ret) {
2348 ret = gen_unique_name(sctx, ino, gen, dest);
2349 if (ret < 0)
2350 goto out;
2351 ret = 1;
2352 goto out_cache;
2353 }
2354
2355 /*
2356 * Depending on whether the inode was already processed or not, use
2357 * send_root or parent_root for ref lookup.
2358 */
2359 if (ino < sctx->send_progress)
2360 ret = get_first_ref(sctx->send_root, ino,
2361 parent_ino, parent_gen, dest);
2362 else
2363 ret = get_first_ref(sctx->parent_root, ino,
2364 parent_ino, parent_gen, dest);
2365 if (ret < 0)
2366 goto out;
2367
2368 /*
2369 * Check if the ref was overwritten by an inode's ref that was processed
2370 * earlier. If yes, treat as orphan and return 1.
2371 */
2372 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2373 dest->start, dest->end - dest->start);
2374 if (ret < 0)
2375 goto out;
2376 if (ret) {
2377 fs_path_reset(dest);
2378 ret = gen_unique_name(sctx, ino, gen, dest);
2379 if (ret < 0)
2380 goto out;
2381 ret = 1;
2382 }
2383
2384 out_cache:
2385 /*
2386 * Store the result of the lookup in the name cache.
2387 */
2388 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2389 if (!nce) {
2390 ret = -ENOMEM;
2391 goto out;
2392 }
2393
2394 nce->entry.key = ino;
2395 nce->entry.gen = gen;
2396 nce->parent_ino = *parent_ino;
2397 nce->parent_gen = *parent_gen;
2398 nce->name_len = fs_path_len(dest);
2399 nce->ret = ret;
2400 strcpy(nce->name, dest->start);
2401
2402 if (ino < sctx->send_progress)
2403 nce->need_later_update = 0;
2404 else
2405 nce->need_later_update = 1;
2406
2407 nce_ret = btrfs_lru_cache_store(&sctx->name_cache, &nce->entry, GFP_KERNEL);
2408 if (nce_ret < 0) {
2409 kfree(nce);
2410 ret = nce_ret;
2411 }
2412
2413 out:
2414 return ret;
2415 }
2416
2417 /*
2418 * Magic happens here. This function returns the first ref to an inode as it
2419 * would look like while receiving the stream at this point in time.
2420 * We walk the path up to the root. For every inode in between, we check if it
2421 * was already processed/sent. If yes, we continue with the parent as found
2422 * in send_root. If not, we continue with the parent as found in parent_root.
2423 * If we encounter an inode that was deleted at this point in time, we use the
2424 * inodes "orphan" name instead of the real name and stop. Same with new inodes
2425 * that were not created yet and overwritten inodes/refs.
2426 *
2427 * When do we have orphan inodes:
2428 * 1. When an inode is freshly created and thus no valid refs are available yet
2429 * 2. When a directory lost all it's refs (deleted) but still has dir items
2430 * inside which were not processed yet (pending for move/delete). If anyone
2431 * tried to get the path to the dir items, it would get a path inside that
2432 * orphan directory.
2433 * 3. When an inode is moved around or gets new links, it may overwrite the ref
2434 * of an unprocessed inode. If in that case the first ref would be
2435 * overwritten, the overwritten inode gets "orphanized". Later when we
2436 * process this overwritten inode, it is restored at a new place by moving
2437 * the orphan inode.
2438 *
2439 * sctx->send_progress tells this function at which point in time receiving
2440 * would be.
2441 */
get_cur_path(struct send_ctx * sctx,u64 ino,u64 gen,struct fs_path * dest)2442 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2443 struct fs_path *dest)
2444 {
2445 int ret = 0;
2446 struct fs_path *name = NULL;
2447 u64 parent_inode = 0;
2448 u64 parent_gen = 0;
2449 int stop = 0;
2450
2451 name = fs_path_alloc();
2452 if (!name) {
2453 ret = -ENOMEM;
2454 goto out;
2455 }
2456
2457 dest->reversed = 1;
2458 fs_path_reset(dest);
2459
2460 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2461 struct waiting_dir_move *wdm;
2462
2463 fs_path_reset(name);
2464
2465 if (is_waiting_for_rm(sctx, ino, gen)) {
2466 ret = gen_unique_name(sctx, ino, gen, name);
2467 if (ret < 0)
2468 goto out;
2469 ret = fs_path_add_path(dest, name);
2470 break;
2471 }
2472
2473 wdm = get_waiting_dir_move(sctx, ino);
2474 if (wdm && wdm->orphanized) {
2475 ret = gen_unique_name(sctx, ino, gen, name);
2476 stop = 1;
2477 } else if (wdm) {
2478 ret = get_first_ref(sctx->parent_root, ino,
2479 &parent_inode, &parent_gen, name);
2480 } else {
2481 ret = __get_cur_name_and_parent(sctx, ino, gen,
2482 &parent_inode,
2483 &parent_gen, name);
2484 if (ret)
2485 stop = 1;
2486 }
2487
2488 if (ret < 0)
2489 goto out;
2490
2491 ret = fs_path_add_path(dest, name);
2492 if (ret < 0)
2493 goto out;
2494
2495 ino = parent_inode;
2496 gen = parent_gen;
2497 }
2498
2499 out:
2500 fs_path_free(name);
2501 if (!ret)
2502 fs_path_unreverse(dest);
2503 return ret;
2504 }
2505
2506 /*
2507 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2508 */
send_subvol_begin(struct send_ctx * sctx)2509 static int send_subvol_begin(struct send_ctx *sctx)
2510 {
2511 int ret;
2512 struct btrfs_root *send_root = sctx->send_root;
2513 struct btrfs_root *parent_root = sctx->parent_root;
2514 struct btrfs_path *path;
2515 struct btrfs_key key;
2516 struct btrfs_root_ref *ref;
2517 struct extent_buffer *leaf;
2518 char *name = NULL;
2519 int namelen;
2520
2521 path = btrfs_alloc_path();
2522 if (!path)
2523 return -ENOMEM;
2524
2525 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2526 if (!name) {
2527 btrfs_free_path(path);
2528 return -ENOMEM;
2529 }
2530
2531 key.objectid = send_root->root_key.objectid;
2532 key.type = BTRFS_ROOT_BACKREF_KEY;
2533 key.offset = 0;
2534
2535 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2536 &key, path, 1, 0);
2537 if (ret < 0)
2538 goto out;
2539 if (ret) {
2540 ret = -ENOENT;
2541 goto out;
2542 }
2543
2544 leaf = path->nodes[0];
2545 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2546 if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2547 key.objectid != send_root->root_key.objectid) {
2548 ret = -ENOENT;
2549 goto out;
2550 }
2551 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2552 namelen = btrfs_root_ref_name_len(leaf, ref);
2553 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2554 btrfs_release_path(path);
2555
2556 if (parent_root) {
2557 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2558 if (ret < 0)
2559 goto out;
2560 } else {
2561 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2562 if (ret < 0)
2563 goto out;
2564 }
2565
2566 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2567
2568 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2569 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2570 sctx->send_root->root_item.received_uuid);
2571 else
2572 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2573 sctx->send_root->root_item.uuid);
2574
2575 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2576 btrfs_root_ctransid(&sctx->send_root->root_item));
2577 if (parent_root) {
2578 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2579 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2580 parent_root->root_item.received_uuid);
2581 else
2582 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2583 parent_root->root_item.uuid);
2584 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2585 btrfs_root_ctransid(&sctx->parent_root->root_item));
2586 }
2587
2588 ret = send_cmd(sctx);
2589
2590 tlv_put_failure:
2591 out:
2592 btrfs_free_path(path);
2593 kfree(name);
2594 return ret;
2595 }
2596
send_truncate(struct send_ctx * sctx,u64 ino,u64 gen,u64 size)2597 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2598 {
2599 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2600 int ret = 0;
2601 struct fs_path *p;
2602
2603 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2604
2605 p = fs_path_alloc();
2606 if (!p)
2607 return -ENOMEM;
2608
2609 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2610 if (ret < 0)
2611 goto out;
2612
2613 ret = get_cur_path(sctx, ino, gen, p);
2614 if (ret < 0)
2615 goto out;
2616 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2617 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2618
2619 ret = send_cmd(sctx);
2620
2621 tlv_put_failure:
2622 out:
2623 fs_path_free(p);
2624 return ret;
2625 }
2626
send_chmod(struct send_ctx * sctx,u64 ino,u64 gen,u64 mode)2627 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2628 {
2629 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2630 int ret = 0;
2631 struct fs_path *p;
2632
2633 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2634
2635 p = fs_path_alloc();
2636 if (!p)
2637 return -ENOMEM;
2638
2639 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2640 if (ret < 0)
2641 goto out;
2642
2643 ret = get_cur_path(sctx, ino, gen, p);
2644 if (ret < 0)
2645 goto out;
2646 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2647 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2648
2649 ret = send_cmd(sctx);
2650
2651 tlv_put_failure:
2652 out:
2653 fs_path_free(p);
2654 return ret;
2655 }
2656
send_fileattr(struct send_ctx * sctx,u64 ino,u64 gen,u64 fileattr)2657 static int send_fileattr(struct send_ctx *sctx, u64 ino, u64 gen, u64 fileattr)
2658 {
2659 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2660 int ret = 0;
2661 struct fs_path *p;
2662
2663 if (sctx->proto < 2)
2664 return 0;
2665
2666 btrfs_debug(fs_info, "send_fileattr %llu fileattr=%llu", ino, fileattr);
2667
2668 p = fs_path_alloc();
2669 if (!p)
2670 return -ENOMEM;
2671
2672 ret = begin_cmd(sctx, BTRFS_SEND_C_FILEATTR);
2673 if (ret < 0)
2674 goto out;
2675
2676 ret = get_cur_path(sctx, ino, gen, p);
2677 if (ret < 0)
2678 goto out;
2679 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2680 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILEATTR, fileattr);
2681
2682 ret = send_cmd(sctx);
2683
2684 tlv_put_failure:
2685 out:
2686 fs_path_free(p);
2687 return ret;
2688 }
2689
send_chown(struct send_ctx * sctx,u64 ino,u64 gen,u64 uid,u64 gid)2690 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2691 {
2692 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2693 int ret = 0;
2694 struct fs_path *p;
2695
2696 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2697 ino, uid, gid);
2698
2699 p = fs_path_alloc();
2700 if (!p)
2701 return -ENOMEM;
2702
2703 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2704 if (ret < 0)
2705 goto out;
2706
2707 ret = get_cur_path(sctx, ino, gen, p);
2708 if (ret < 0)
2709 goto out;
2710 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2711 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2712 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2713
2714 ret = send_cmd(sctx);
2715
2716 tlv_put_failure:
2717 out:
2718 fs_path_free(p);
2719 return ret;
2720 }
2721
send_utimes(struct send_ctx * sctx,u64 ino,u64 gen)2722 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2723 {
2724 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2725 int ret = 0;
2726 struct fs_path *p = NULL;
2727 struct btrfs_inode_item *ii;
2728 struct btrfs_path *path = NULL;
2729 struct extent_buffer *eb;
2730 struct btrfs_key key;
2731 int slot;
2732
2733 btrfs_debug(fs_info, "send_utimes %llu", ino);
2734
2735 p = fs_path_alloc();
2736 if (!p)
2737 return -ENOMEM;
2738
2739 path = alloc_path_for_send();
2740 if (!path) {
2741 ret = -ENOMEM;
2742 goto out;
2743 }
2744
2745 key.objectid = ino;
2746 key.type = BTRFS_INODE_ITEM_KEY;
2747 key.offset = 0;
2748 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2749 if (ret > 0)
2750 ret = -ENOENT;
2751 if (ret < 0)
2752 goto out;
2753
2754 eb = path->nodes[0];
2755 slot = path->slots[0];
2756 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2757
2758 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2759 if (ret < 0)
2760 goto out;
2761
2762 ret = get_cur_path(sctx, ino, gen, p);
2763 if (ret < 0)
2764 goto out;
2765 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2766 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2767 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2768 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2769 if (sctx->proto >= 2)
2770 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_OTIME, eb, &ii->otime);
2771
2772 ret = send_cmd(sctx);
2773
2774 tlv_put_failure:
2775 out:
2776 fs_path_free(p);
2777 btrfs_free_path(path);
2778 return ret;
2779 }
2780
2781 /*
2782 * If the cache is full, we can't remove entries from it and do a call to
2783 * send_utimes() for each respective inode, because we might be finishing
2784 * processing an inode that is a directory and it just got renamed, and existing
2785 * entries in the cache may refer to inodes that have the directory in their
2786 * full path - in which case we would generate outdated paths (pre-rename)
2787 * for the inodes that the cache entries point to. Instead of prunning the
2788 * cache when inserting, do it after we finish processing each inode at
2789 * finish_inode_if_needed().
2790 */
cache_dir_utimes(struct send_ctx * sctx,u64 dir,u64 gen)2791 static int cache_dir_utimes(struct send_ctx *sctx, u64 dir, u64 gen)
2792 {
2793 struct btrfs_lru_cache_entry *entry;
2794 int ret;
2795
2796 entry = btrfs_lru_cache_lookup(&sctx->dir_utimes_cache, dir, gen);
2797 if (entry != NULL)
2798 return 0;
2799
2800 /* Caching is optional, don't fail if we can't allocate memory. */
2801 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2802 if (!entry)
2803 return send_utimes(sctx, dir, gen);
2804
2805 entry->key = dir;
2806 entry->gen = gen;
2807
2808 ret = btrfs_lru_cache_store(&sctx->dir_utimes_cache, entry, GFP_KERNEL);
2809 ASSERT(ret != -EEXIST);
2810 if (ret) {
2811 kfree(entry);
2812 return send_utimes(sctx, dir, gen);
2813 }
2814
2815 return 0;
2816 }
2817
trim_dir_utimes_cache(struct send_ctx * sctx)2818 static int trim_dir_utimes_cache(struct send_ctx *sctx)
2819 {
2820 while (btrfs_lru_cache_size(&sctx->dir_utimes_cache) >
2821 SEND_MAX_DIR_UTIMES_CACHE_SIZE) {
2822 struct btrfs_lru_cache_entry *lru;
2823 int ret;
2824
2825 lru = btrfs_lru_cache_lru_entry(&sctx->dir_utimes_cache);
2826 ASSERT(lru != NULL);
2827
2828 ret = send_utimes(sctx, lru->key, lru->gen);
2829 if (ret)
2830 return ret;
2831
2832 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, lru);
2833 }
2834
2835 return 0;
2836 }
2837
2838 /*
2839 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2840 * a valid path yet because we did not process the refs yet. So, the inode
2841 * is created as orphan.
2842 */
send_create_inode(struct send_ctx * sctx,u64 ino)2843 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2844 {
2845 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2846 int ret = 0;
2847 struct fs_path *p;
2848 int cmd;
2849 struct btrfs_inode_info info;
2850 u64 gen;
2851 u64 mode;
2852 u64 rdev;
2853
2854 btrfs_debug(fs_info, "send_create_inode %llu", ino);
2855
2856 p = fs_path_alloc();
2857 if (!p)
2858 return -ENOMEM;
2859
2860 if (ino != sctx->cur_ino) {
2861 ret = get_inode_info(sctx->send_root, ino, &info);
2862 if (ret < 0)
2863 goto out;
2864 gen = info.gen;
2865 mode = info.mode;
2866 rdev = info.rdev;
2867 } else {
2868 gen = sctx->cur_inode_gen;
2869 mode = sctx->cur_inode_mode;
2870 rdev = sctx->cur_inode_rdev;
2871 }
2872
2873 if (S_ISREG(mode)) {
2874 cmd = BTRFS_SEND_C_MKFILE;
2875 } else if (S_ISDIR(mode)) {
2876 cmd = BTRFS_SEND_C_MKDIR;
2877 } else if (S_ISLNK(mode)) {
2878 cmd = BTRFS_SEND_C_SYMLINK;
2879 } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2880 cmd = BTRFS_SEND_C_MKNOD;
2881 } else if (S_ISFIFO(mode)) {
2882 cmd = BTRFS_SEND_C_MKFIFO;
2883 } else if (S_ISSOCK(mode)) {
2884 cmd = BTRFS_SEND_C_MKSOCK;
2885 } else {
2886 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2887 (int)(mode & S_IFMT));
2888 ret = -EOPNOTSUPP;
2889 goto out;
2890 }
2891
2892 ret = begin_cmd(sctx, cmd);
2893 if (ret < 0)
2894 goto out;
2895
2896 ret = gen_unique_name(sctx, ino, gen, p);
2897 if (ret < 0)
2898 goto out;
2899
2900 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2901 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2902
2903 if (S_ISLNK(mode)) {
2904 fs_path_reset(p);
2905 ret = read_symlink(sctx->send_root, ino, p);
2906 if (ret < 0)
2907 goto out;
2908 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2909 } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2910 S_ISFIFO(mode) || S_ISSOCK(mode)) {
2911 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2912 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2913 }
2914
2915 ret = send_cmd(sctx);
2916 if (ret < 0)
2917 goto out;
2918
2919
2920 tlv_put_failure:
2921 out:
2922 fs_path_free(p);
2923 return ret;
2924 }
2925
cache_dir_created(struct send_ctx * sctx,u64 dir)2926 static void cache_dir_created(struct send_ctx *sctx, u64 dir)
2927 {
2928 struct btrfs_lru_cache_entry *entry;
2929 int ret;
2930
2931 /* Caching is optional, ignore any failures. */
2932 entry = kmalloc(sizeof(*entry), GFP_KERNEL);
2933 if (!entry)
2934 return;
2935
2936 entry->key = dir;
2937 entry->gen = 0;
2938 ret = btrfs_lru_cache_store(&sctx->dir_created_cache, entry, GFP_KERNEL);
2939 if (ret < 0)
2940 kfree(entry);
2941 }
2942
2943 /*
2944 * We need some special handling for inodes that get processed before the parent
2945 * directory got created. See process_recorded_refs for details.
2946 * This function does the check if we already created the dir out of order.
2947 */
did_create_dir(struct send_ctx * sctx,u64 dir)2948 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2949 {
2950 int ret = 0;
2951 int iter_ret = 0;
2952 struct btrfs_path *path = NULL;
2953 struct btrfs_key key;
2954 struct btrfs_key found_key;
2955 struct btrfs_key di_key;
2956 struct btrfs_dir_item *di;
2957
2958 if (btrfs_lru_cache_lookup(&sctx->dir_created_cache, dir, 0))
2959 return 1;
2960
2961 path = alloc_path_for_send();
2962 if (!path)
2963 return -ENOMEM;
2964
2965 key.objectid = dir;
2966 key.type = BTRFS_DIR_INDEX_KEY;
2967 key.offset = 0;
2968
2969 btrfs_for_each_slot(sctx->send_root, &key, &found_key, path, iter_ret) {
2970 struct extent_buffer *eb = path->nodes[0];
2971
2972 if (found_key.objectid != key.objectid ||
2973 found_key.type != key.type) {
2974 ret = 0;
2975 break;
2976 }
2977
2978 di = btrfs_item_ptr(eb, path->slots[0], struct btrfs_dir_item);
2979 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2980
2981 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2982 di_key.objectid < sctx->send_progress) {
2983 ret = 1;
2984 cache_dir_created(sctx, dir);
2985 break;
2986 }
2987 }
2988 /* Catch error found during iteration */
2989 if (iter_ret < 0)
2990 ret = iter_ret;
2991
2992 btrfs_free_path(path);
2993 return ret;
2994 }
2995
2996 /*
2997 * Only creates the inode if it is:
2998 * 1. Not a directory
2999 * 2. Or a directory which was not created already due to out of order
3000 * directories. See did_create_dir and process_recorded_refs for details.
3001 */
send_create_inode_if_needed(struct send_ctx * sctx)3002 static int send_create_inode_if_needed(struct send_ctx *sctx)
3003 {
3004 int ret;
3005
3006 if (S_ISDIR(sctx->cur_inode_mode)) {
3007 ret = did_create_dir(sctx, sctx->cur_ino);
3008 if (ret < 0)
3009 return ret;
3010 else if (ret > 0)
3011 return 0;
3012 }
3013
3014 ret = send_create_inode(sctx, sctx->cur_ino);
3015
3016 if (ret == 0 && S_ISDIR(sctx->cur_inode_mode))
3017 cache_dir_created(sctx, sctx->cur_ino);
3018
3019 return ret;
3020 }
3021
3022 struct recorded_ref {
3023 struct list_head list;
3024 char *name;
3025 struct fs_path *full_path;
3026 u64 dir;
3027 u64 dir_gen;
3028 int name_len;
3029 struct rb_node node;
3030 struct rb_root *root;
3031 };
3032
recorded_ref_alloc(void)3033 static struct recorded_ref *recorded_ref_alloc(void)
3034 {
3035 struct recorded_ref *ref;
3036
3037 ref = kzalloc(sizeof(*ref), GFP_KERNEL);
3038 if (!ref)
3039 return NULL;
3040 RB_CLEAR_NODE(&ref->node);
3041 INIT_LIST_HEAD(&ref->list);
3042 return ref;
3043 }
3044
recorded_ref_free(struct recorded_ref * ref)3045 static void recorded_ref_free(struct recorded_ref *ref)
3046 {
3047 if (!ref)
3048 return;
3049 if (!RB_EMPTY_NODE(&ref->node))
3050 rb_erase(&ref->node, ref->root);
3051 list_del(&ref->list);
3052 fs_path_free(ref->full_path);
3053 kfree(ref);
3054 }
3055
set_ref_path(struct recorded_ref * ref,struct fs_path * path)3056 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
3057 {
3058 ref->full_path = path;
3059 ref->name = (char *)kbasename(ref->full_path->start);
3060 ref->name_len = ref->full_path->end - ref->name;
3061 }
3062
dup_ref(struct recorded_ref * ref,struct list_head * list)3063 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
3064 {
3065 struct recorded_ref *new;
3066
3067 new = recorded_ref_alloc();
3068 if (!new)
3069 return -ENOMEM;
3070
3071 new->dir = ref->dir;
3072 new->dir_gen = ref->dir_gen;
3073 list_add_tail(&new->list, list);
3074 return 0;
3075 }
3076
__free_recorded_refs(struct list_head * head)3077 static void __free_recorded_refs(struct list_head *head)
3078 {
3079 struct recorded_ref *cur;
3080
3081 while (!list_empty(head)) {
3082 cur = list_entry(head->next, struct recorded_ref, list);
3083 recorded_ref_free(cur);
3084 }
3085 }
3086
free_recorded_refs(struct send_ctx * sctx)3087 static void free_recorded_refs(struct send_ctx *sctx)
3088 {
3089 __free_recorded_refs(&sctx->new_refs);
3090 __free_recorded_refs(&sctx->deleted_refs);
3091 }
3092
3093 /*
3094 * Renames/moves a file/dir to its orphan name. Used when the first
3095 * ref of an unprocessed inode gets overwritten and for all non empty
3096 * directories.
3097 */
orphanize_inode(struct send_ctx * sctx,u64 ino,u64 gen,struct fs_path * path)3098 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
3099 struct fs_path *path)
3100 {
3101 int ret;
3102 struct fs_path *orphan;
3103
3104 orphan = fs_path_alloc();
3105 if (!orphan)
3106 return -ENOMEM;
3107
3108 ret = gen_unique_name(sctx, ino, gen, orphan);
3109 if (ret < 0)
3110 goto out;
3111
3112 ret = send_rename(sctx, path, orphan);
3113
3114 out:
3115 fs_path_free(orphan);
3116 return ret;
3117 }
3118
add_orphan_dir_info(struct send_ctx * sctx,u64 dir_ino,u64 dir_gen)3119 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx,
3120 u64 dir_ino, u64 dir_gen)
3121 {
3122 struct rb_node **p = &sctx->orphan_dirs.rb_node;
3123 struct rb_node *parent = NULL;
3124 struct orphan_dir_info *entry, *odi;
3125
3126 while (*p) {
3127 parent = *p;
3128 entry = rb_entry(parent, struct orphan_dir_info, node);
3129 if (dir_ino < entry->ino)
3130 p = &(*p)->rb_left;
3131 else if (dir_ino > entry->ino)
3132 p = &(*p)->rb_right;
3133 else if (dir_gen < entry->gen)
3134 p = &(*p)->rb_left;
3135 else if (dir_gen > entry->gen)
3136 p = &(*p)->rb_right;
3137 else
3138 return entry;
3139 }
3140
3141 odi = kmalloc(sizeof(*odi), GFP_KERNEL);
3142 if (!odi)
3143 return ERR_PTR(-ENOMEM);
3144 odi->ino = dir_ino;
3145 odi->gen = dir_gen;
3146 odi->last_dir_index_offset = 0;
3147 odi->dir_high_seq_ino = 0;
3148
3149 rb_link_node(&odi->node, parent, p);
3150 rb_insert_color(&odi->node, &sctx->orphan_dirs);
3151 return odi;
3152 }
3153
get_orphan_dir_info(struct send_ctx * sctx,u64 dir_ino,u64 gen)3154 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx,
3155 u64 dir_ino, u64 gen)
3156 {
3157 struct rb_node *n = sctx->orphan_dirs.rb_node;
3158 struct orphan_dir_info *entry;
3159
3160 while (n) {
3161 entry = rb_entry(n, struct orphan_dir_info, node);
3162 if (dir_ino < entry->ino)
3163 n = n->rb_left;
3164 else if (dir_ino > entry->ino)
3165 n = n->rb_right;
3166 else if (gen < entry->gen)
3167 n = n->rb_left;
3168 else if (gen > entry->gen)
3169 n = n->rb_right;
3170 else
3171 return entry;
3172 }
3173 return NULL;
3174 }
3175
is_waiting_for_rm(struct send_ctx * sctx,u64 dir_ino,u64 gen)3176 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen)
3177 {
3178 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen);
3179
3180 return odi != NULL;
3181 }
3182
free_orphan_dir_info(struct send_ctx * sctx,struct orphan_dir_info * odi)3183 static void free_orphan_dir_info(struct send_ctx *sctx,
3184 struct orphan_dir_info *odi)
3185 {
3186 if (!odi)
3187 return;
3188 rb_erase(&odi->node, &sctx->orphan_dirs);
3189 kfree(odi);
3190 }
3191
3192 /*
3193 * Returns 1 if a directory can be removed at this point in time.
3194 * We check this by iterating all dir items and checking if the inode behind
3195 * the dir item was already processed.
3196 */
can_rmdir(struct send_ctx * sctx,u64 dir,u64 dir_gen)3197 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen)
3198 {
3199 int ret = 0;
3200 int iter_ret = 0;
3201 struct btrfs_root *root = sctx->parent_root;
3202 struct btrfs_path *path;
3203 struct btrfs_key key;
3204 struct btrfs_key found_key;
3205 struct btrfs_key loc;
3206 struct btrfs_dir_item *di;
3207 struct orphan_dir_info *odi = NULL;
3208 u64 dir_high_seq_ino = 0;
3209 u64 last_dir_index_offset = 0;
3210
3211 /*
3212 * Don't try to rmdir the top/root subvolume dir.
3213 */
3214 if (dir == BTRFS_FIRST_FREE_OBJECTID)
3215 return 0;
3216
3217 odi = get_orphan_dir_info(sctx, dir, dir_gen);
3218 if (odi && sctx->cur_ino < odi->dir_high_seq_ino)
3219 return 0;
3220
3221 path = alloc_path_for_send();
3222 if (!path)
3223 return -ENOMEM;
3224
3225 if (!odi) {
3226 /*
3227 * Find the inode number associated with the last dir index
3228 * entry. This is very likely the inode with the highest number
3229 * of all inodes that have an entry in the directory. We can
3230 * then use it to avoid future calls to can_rmdir(), when
3231 * processing inodes with a lower number, from having to search
3232 * the parent root b+tree for dir index keys.
3233 */
3234 key.objectid = dir;
3235 key.type = BTRFS_DIR_INDEX_KEY;
3236 key.offset = (u64)-1;
3237
3238 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3239 if (ret < 0) {
3240 goto out;
3241 } else if (ret > 0) {
3242 /* Can't happen, the root is never empty. */
3243 ASSERT(path->slots[0] > 0);
3244 if (WARN_ON(path->slots[0] == 0)) {
3245 ret = -EUCLEAN;
3246 goto out;
3247 }
3248 path->slots[0]--;
3249 }
3250
3251 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3252 if (key.objectid != dir || key.type != BTRFS_DIR_INDEX_KEY) {
3253 /* No index keys, dir can be removed. */
3254 ret = 1;
3255 goto out;
3256 }
3257
3258 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3259 struct btrfs_dir_item);
3260 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3261 dir_high_seq_ino = loc.objectid;
3262 if (sctx->cur_ino < dir_high_seq_ino) {
3263 ret = 0;
3264 goto out;
3265 }
3266
3267 btrfs_release_path(path);
3268 }
3269
3270 key.objectid = dir;
3271 key.type = BTRFS_DIR_INDEX_KEY;
3272 key.offset = (odi ? odi->last_dir_index_offset : 0);
3273
3274 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
3275 struct waiting_dir_move *dm;
3276
3277 if (found_key.objectid != key.objectid ||
3278 found_key.type != key.type)
3279 break;
3280
3281 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
3282 struct btrfs_dir_item);
3283 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
3284
3285 dir_high_seq_ino = max(dir_high_seq_ino, loc.objectid);
3286 last_dir_index_offset = found_key.offset;
3287
3288 dm = get_waiting_dir_move(sctx, loc.objectid);
3289 if (dm) {
3290 dm->rmdir_ino = dir;
3291 dm->rmdir_gen = dir_gen;
3292 ret = 0;
3293 goto out;
3294 }
3295
3296 if (loc.objectid > sctx->cur_ino) {
3297 ret = 0;
3298 goto out;
3299 }
3300 }
3301 if (iter_ret < 0) {
3302 ret = iter_ret;
3303 goto out;
3304 }
3305 free_orphan_dir_info(sctx, odi);
3306
3307 ret = 1;
3308
3309 out:
3310 btrfs_free_path(path);
3311
3312 if (ret)
3313 return ret;
3314
3315 if (!odi) {
3316 odi = add_orphan_dir_info(sctx, dir, dir_gen);
3317 if (IS_ERR(odi))
3318 return PTR_ERR(odi);
3319
3320 odi->gen = dir_gen;
3321 }
3322
3323 odi->last_dir_index_offset = last_dir_index_offset;
3324 odi->dir_high_seq_ino = max(odi->dir_high_seq_ino, dir_high_seq_ino);
3325
3326 return 0;
3327 }
3328
is_waiting_for_move(struct send_ctx * sctx,u64 ino)3329 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3330 {
3331 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3332
3333 return entry != NULL;
3334 }
3335
add_waiting_dir_move(struct send_ctx * sctx,u64 ino,bool orphanized)3336 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3337 {
3338 struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3339 struct rb_node *parent = NULL;
3340 struct waiting_dir_move *entry, *dm;
3341
3342 dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3343 if (!dm)
3344 return -ENOMEM;
3345 dm->ino = ino;
3346 dm->rmdir_ino = 0;
3347 dm->rmdir_gen = 0;
3348 dm->orphanized = orphanized;
3349
3350 while (*p) {
3351 parent = *p;
3352 entry = rb_entry(parent, struct waiting_dir_move, node);
3353 if (ino < entry->ino) {
3354 p = &(*p)->rb_left;
3355 } else if (ino > entry->ino) {
3356 p = &(*p)->rb_right;
3357 } else {
3358 kfree(dm);
3359 return -EEXIST;
3360 }
3361 }
3362
3363 rb_link_node(&dm->node, parent, p);
3364 rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3365 return 0;
3366 }
3367
3368 static struct waiting_dir_move *
get_waiting_dir_move(struct send_ctx * sctx,u64 ino)3369 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3370 {
3371 struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3372 struct waiting_dir_move *entry;
3373
3374 while (n) {
3375 entry = rb_entry(n, struct waiting_dir_move, node);
3376 if (ino < entry->ino)
3377 n = n->rb_left;
3378 else if (ino > entry->ino)
3379 n = n->rb_right;
3380 else
3381 return entry;
3382 }
3383 return NULL;
3384 }
3385
free_waiting_dir_move(struct send_ctx * sctx,struct waiting_dir_move * dm)3386 static void free_waiting_dir_move(struct send_ctx *sctx,
3387 struct waiting_dir_move *dm)
3388 {
3389 if (!dm)
3390 return;
3391 rb_erase(&dm->node, &sctx->waiting_dir_moves);
3392 kfree(dm);
3393 }
3394
add_pending_dir_move(struct send_ctx * sctx,u64 ino,u64 ino_gen,u64 parent_ino,struct list_head * new_refs,struct list_head * deleted_refs,const bool is_orphan)3395 static int add_pending_dir_move(struct send_ctx *sctx,
3396 u64 ino,
3397 u64 ino_gen,
3398 u64 parent_ino,
3399 struct list_head *new_refs,
3400 struct list_head *deleted_refs,
3401 const bool is_orphan)
3402 {
3403 struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3404 struct rb_node *parent = NULL;
3405 struct pending_dir_move *entry = NULL, *pm;
3406 struct recorded_ref *cur;
3407 int exists = 0;
3408 int ret;
3409
3410 pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3411 if (!pm)
3412 return -ENOMEM;
3413 pm->parent_ino = parent_ino;
3414 pm->ino = ino;
3415 pm->gen = ino_gen;
3416 INIT_LIST_HEAD(&pm->list);
3417 INIT_LIST_HEAD(&pm->update_refs);
3418 RB_CLEAR_NODE(&pm->node);
3419
3420 while (*p) {
3421 parent = *p;
3422 entry = rb_entry(parent, struct pending_dir_move, node);
3423 if (parent_ino < entry->parent_ino) {
3424 p = &(*p)->rb_left;
3425 } else if (parent_ino > entry->parent_ino) {
3426 p = &(*p)->rb_right;
3427 } else {
3428 exists = 1;
3429 break;
3430 }
3431 }
3432
3433 list_for_each_entry(cur, deleted_refs, list) {
3434 ret = dup_ref(cur, &pm->update_refs);
3435 if (ret < 0)
3436 goto out;
3437 }
3438 list_for_each_entry(cur, new_refs, list) {
3439 ret = dup_ref(cur, &pm->update_refs);
3440 if (ret < 0)
3441 goto out;
3442 }
3443
3444 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3445 if (ret)
3446 goto out;
3447
3448 if (exists) {
3449 list_add_tail(&pm->list, &entry->list);
3450 } else {
3451 rb_link_node(&pm->node, parent, p);
3452 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3453 }
3454 ret = 0;
3455 out:
3456 if (ret) {
3457 __free_recorded_refs(&pm->update_refs);
3458 kfree(pm);
3459 }
3460 return ret;
3461 }
3462
get_pending_dir_moves(struct send_ctx * sctx,u64 parent_ino)3463 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3464 u64 parent_ino)
3465 {
3466 struct rb_node *n = sctx->pending_dir_moves.rb_node;
3467 struct pending_dir_move *entry;
3468
3469 while (n) {
3470 entry = rb_entry(n, struct pending_dir_move, node);
3471 if (parent_ino < entry->parent_ino)
3472 n = n->rb_left;
3473 else if (parent_ino > entry->parent_ino)
3474 n = n->rb_right;
3475 else
3476 return entry;
3477 }
3478 return NULL;
3479 }
3480
path_loop(struct send_ctx * sctx,struct fs_path * name,u64 ino,u64 gen,u64 * ancestor_ino)3481 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3482 u64 ino, u64 gen, u64 *ancestor_ino)
3483 {
3484 int ret = 0;
3485 u64 parent_inode = 0;
3486 u64 parent_gen = 0;
3487 u64 start_ino = ino;
3488
3489 *ancestor_ino = 0;
3490 while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3491 fs_path_reset(name);
3492
3493 if (is_waiting_for_rm(sctx, ino, gen))
3494 break;
3495 if (is_waiting_for_move(sctx, ino)) {
3496 if (*ancestor_ino == 0)
3497 *ancestor_ino = ino;
3498 ret = get_first_ref(sctx->parent_root, ino,
3499 &parent_inode, &parent_gen, name);
3500 } else {
3501 ret = __get_cur_name_and_parent(sctx, ino, gen,
3502 &parent_inode,
3503 &parent_gen, name);
3504 if (ret > 0) {
3505 ret = 0;
3506 break;
3507 }
3508 }
3509 if (ret < 0)
3510 break;
3511 if (parent_inode == start_ino) {
3512 ret = 1;
3513 if (*ancestor_ino == 0)
3514 *ancestor_ino = ino;
3515 break;
3516 }
3517 ino = parent_inode;
3518 gen = parent_gen;
3519 }
3520 return ret;
3521 }
3522
apply_dir_move(struct send_ctx * sctx,struct pending_dir_move * pm)3523 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3524 {
3525 struct fs_path *from_path = NULL;
3526 struct fs_path *to_path = NULL;
3527 struct fs_path *name = NULL;
3528 u64 orig_progress = sctx->send_progress;
3529 struct recorded_ref *cur;
3530 u64 parent_ino, parent_gen;
3531 struct waiting_dir_move *dm = NULL;
3532 u64 rmdir_ino = 0;
3533 u64 rmdir_gen;
3534 u64 ancestor;
3535 bool is_orphan;
3536 int ret;
3537
3538 name = fs_path_alloc();
3539 from_path = fs_path_alloc();
3540 if (!name || !from_path) {
3541 ret = -ENOMEM;
3542 goto out;
3543 }
3544
3545 dm = get_waiting_dir_move(sctx, pm->ino);
3546 ASSERT(dm);
3547 rmdir_ino = dm->rmdir_ino;
3548 rmdir_gen = dm->rmdir_gen;
3549 is_orphan = dm->orphanized;
3550 free_waiting_dir_move(sctx, dm);
3551
3552 if (is_orphan) {
3553 ret = gen_unique_name(sctx, pm->ino,
3554 pm->gen, from_path);
3555 } else {
3556 ret = get_first_ref(sctx->parent_root, pm->ino,
3557 &parent_ino, &parent_gen, name);
3558 if (ret < 0)
3559 goto out;
3560 ret = get_cur_path(sctx, parent_ino, parent_gen,
3561 from_path);
3562 if (ret < 0)
3563 goto out;
3564 ret = fs_path_add_path(from_path, name);
3565 }
3566 if (ret < 0)
3567 goto out;
3568
3569 sctx->send_progress = sctx->cur_ino + 1;
3570 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3571 if (ret < 0)
3572 goto out;
3573 if (ret) {
3574 LIST_HEAD(deleted_refs);
3575 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3576 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3577 &pm->update_refs, &deleted_refs,
3578 is_orphan);
3579 if (ret < 0)
3580 goto out;
3581 if (rmdir_ino) {
3582 dm = get_waiting_dir_move(sctx, pm->ino);
3583 ASSERT(dm);
3584 dm->rmdir_ino = rmdir_ino;
3585 dm->rmdir_gen = rmdir_gen;
3586 }
3587 goto out;
3588 }
3589 fs_path_reset(name);
3590 to_path = name;
3591 name = NULL;
3592 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3593 if (ret < 0)
3594 goto out;
3595
3596 ret = send_rename(sctx, from_path, to_path);
3597 if (ret < 0)
3598 goto out;
3599
3600 if (rmdir_ino) {
3601 struct orphan_dir_info *odi;
3602 u64 gen;
3603
3604 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen);
3605 if (!odi) {
3606 /* already deleted */
3607 goto finish;
3608 }
3609 gen = odi->gen;
3610
3611 ret = can_rmdir(sctx, rmdir_ino, gen);
3612 if (ret < 0)
3613 goto out;
3614 if (!ret)
3615 goto finish;
3616
3617 name = fs_path_alloc();
3618 if (!name) {
3619 ret = -ENOMEM;
3620 goto out;
3621 }
3622 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3623 if (ret < 0)
3624 goto out;
3625 ret = send_rmdir(sctx, name);
3626 if (ret < 0)
3627 goto out;
3628 }
3629
3630 finish:
3631 ret = cache_dir_utimes(sctx, pm->ino, pm->gen);
3632 if (ret < 0)
3633 goto out;
3634
3635 /*
3636 * After rename/move, need to update the utimes of both new parent(s)
3637 * and old parent(s).
3638 */
3639 list_for_each_entry(cur, &pm->update_refs, list) {
3640 /*
3641 * The parent inode might have been deleted in the send snapshot
3642 */
3643 ret = get_inode_info(sctx->send_root, cur->dir, NULL);
3644 if (ret == -ENOENT) {
3645 ret = 0;
3646 continue;
3647 }
3648 if (ret < 0)
3649 goto out;
3650
3651 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
3652 if (ret < 0)
3653 goto out;
3654 }
3655
3656 out:
3657 fs_path_free(name);
3658 fs_path_free(from_path);
3659 fs_path_free(to_path);
3660 sctx->send_progress = orig_progress;
3661
3662 return ret;
3663 }
3664
free_pending_move(struct send_ctx * sctx,struct pending_dir_move * m)3665 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3666 {
3667 if (!list_empty(&m->list))
3668 list_del(&m->list);
3669 if (!RB_EMPTY_NODE(&m->node))
3670 rb_erase(&m->node, &sctx->pending_dir_moves);
3671 __free_recorded_refs(&m->update_refs);
3672 kfree(m);
3673 }
3674
tail_append_pending_moves(struct send_ctx * sctx,struct pending_dir_move * moves,struct list_head * stack)3675 static void tail_append_pending_moves(struct send_ctx *sctx,
3676 struct pending_dir_move *moves,
3677 struct list_head *stack)
3678 {
3679 if (list_empty(&moves->list)) {
3680 list_add_tail(&moves->list, stack);
3681 } else {
3682 LIST_HEAD(list);
3683 list_splice_init(&moves->list, &list);
3684 list_add_tail(&moves->list, stack);
3685 list_splice_tail(&list, stack);
3686 }
3687 if (!RB_EMPTY_NODE(&moves->node)) {
3688 rb_erase(&moves->node, &sctx->pending_dir_moves);
3689 RB_CLEAR_NODE(&moves->node);
3690 }
3691 }
3692
apply_children_dir_moves(struct send_ctx * sctx)3693 static int apply_children_dir_moves(struct send_ctx *sctx)
3694 {
3695 struct pending_dir_move *pm;
3696 LIST_HEAD(stack);
3697 u64 parent_ino = sctx->cur_ino;
3698 int ret = 0;
3699
3700 pm = get_pending_dir_moves(sctx, parent_ino);
3701 if (!pm)
3702 return 0;
3703
3704 tail_append_pending_moves(sctx, pm, &stack);
3705
3706 while (!list_empty(&stack)) {
3707 pm = list_first_entry(&stack, struct pending_dir_move, list);
3708 parent_ino = pm->ino;
3709 ret = apply_dir_move(sctx, pm);
3710 free_pending_move(sctx, pm);
3711 if (ret)
3712 goto out;
3713 pm = get_pending_dir_moves(sctx, parent_ino);
3714 if (pm)
3715 tail_append_pending_moves(sctx, pm, &stack);
3716 }
3717 return 0;
3718
3719 out:
3720 while (!list_empty(&stack)) {
3721 pm = list_first_entry(&stack, struct pending_dir_move, list);
3722 free_pending_move(sctx, pm);
3723 }
3724 return ret;
3725 }
3726
3727 /*
3728 * We might need to delay a directory rename even when no ancestor directory
3729 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3730 * renamed. This happens when we rename a directory to the old name (the name
3731 * in the parent root) of some other unrelated directory that got its rename
3732 * delayed due to some ancestor with higher number that got renamed.
3733 *
3734 * Example:
3735 *
3736 * Parent snapshot:
3737 * . (ino 256)
3738 * |---- a/ (ino 257)
3739 * | |---- file (ino 260)
3740 * |
3741 * |---- b/ (ino 258)
3742 * |---- c/ (ino 259)
3743 *
3744 * Send snapshot:
3745 * . (ino 256)
3746 * |---- a/ (ino 258)
3747 * |---- x/ (ino 259)
3748 * |---- y/ (ino 257)
3749 * |----- file (ino 260)
3750 *
3751 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3752 * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3753 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3754 * must issue is:
3755 *
3756 * 1 - rename 259 from 'c' to 'x'
3757 * 2 - rename 257 from 'a' to 'x/y'
3758 * 3 - rename 258 from 'b' to 'a'
3759 *
3760 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3761 * be done right away and < 0 on error.
3762 */
wait_for_dest_dir_move(struct send_ctx * sctx,struct recorded_ref * parent_ref,const bool is_orphan)3763 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3764 struct recorded_ref *parent_ref,
3765 const bool is_orphan)
3766 {
3767 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3768 struct btrfs_path *path;
3769 struct btrfs_key key;
3770 struct btrfs_key di_key;
3771 struct btrfs_dir_item *di;
3772 u64 left_gen;
3773 u64 right_gen;
3774 int ret = 0;
3775 struct waiting_dir_move *wdm;
3776
3777 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3778 return 0;
3779
3780 path = alloc_path_for_send();
3781 if (!path)
3782 return -ENOMEM;
3783
3784 key.objectid = parent_ref->dir;
3785 key.type = BTRFS_DIR_ITEM_KEY;
3786 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3787
3788 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3789 if (ret < 0) {
3790 goto out;
3791 } else if (ret > 0) {
3792 ret = 0;
3793 goto out;
3794 }
3795
3796 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3797 parent_ref->name_len);
3798 if (!di) {
3799 ret = 0;
3800 goto out;
3801 }
3802 /*
3803 * di_key.objectid has the number of the inode that has a dentry in the
3804 * parent directory with the same name that sctx->cur_ino is being
3805 * renamed to. We need to check if that inode is in the send root as
3806 * well and if it is currently marked as an inode with a pending rename,
3807 * if it is, we need to delay the rename of sctx->cur_ino as well, so
3808 * that it happens after that other inode is renamed.
3809 */
3810 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3811 if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3812 ret = 0;
3813 goto out;
3814 }
3815
3816 ret = get_inode_gen(sctx->parent_root, di_key.objectid, &left_gen);
3817 if (ret < 0)
3818 goto out;
3819 ret = get_inode_gen(sctx->send_root, di_key.objectid, &right_gen);
3820 if (ret < 0) {
3821 if (ret == -ENOENT)
3822 ret = 0;
3823 goto out;
3824 }
3825
3826 /* Different inode, no need to delay the rename of sctx->cur_ino */
3827 if (right_gen != left_gen) {
3828 ret = 0;
3829 goto out;
3830 }
3831
3832 wdm = get_waiting_dir_move(sctx, di_key.objectid);
3833 if (wdm && !wdm->orphanized) {
3834 ret = add_pending_dir_move(sctx,
3835 sctx->cur_ino,
3836 sctx->cur_inode_gen,
3837 di_key.objectid,
3838 &sctx->new_refs,
3839 &sctx->deleted_refs,
3840 is_orphan);
3841 if (!ret)
3842 ret = 1;
3843 }
3844 out:
3845 btrfs_free_path(path);
3846 return ret;
3847 }
3848
3849 /*
3850 * Check if inode ino2, or any of its ancestors, is inode ino1.
3851 * Return 1 if true, 0 if false and < 0 on error.
3852 */
check_ino_in_path(struct btrfs_root * root,const u64 ino1,const u64 ino1_gen,const u64 ino2,const u64 ino2_gen,struct fs_path * fs_path)3853 static int check_ino_in_path(struct btrfs_root *root,
3854 const u64 ino1,
3855 const u64 ino1_gen,
3856 const u64 ino2,
3857 const u64 ino2_gen,
3858 struct fs_path *fs_path)
3859 {
3860 u64 ino = ino2;
3861
3862 if (ino1 == ino2)
3863 return ino1_gen == ino2_gen;
3864
3865 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3866 u64 parent;
3867 u64 parent_gen;
3868 int ret;
3869
3870 fs_path_reset(fs_path);
3871 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3872 if (ret < 0)
3873 return ret;
3874 if (parent == ino1)
3875 return parent_gen == ino1_gen;
3876 ino = parent;
3877 }
3878 return 0;
3879 }
3880
3881 /*
3882 * Check if inode ino1 is an ancestor of inode ino2 in the given root for any
3883 * possible path (in case ino2 is not a directory and has multiple hard links).
3884 * Return 1 if true, 0 if false and < 0 on error.
3885 */
is_ancestor(struct btrfs_root * root,const u64 ino1,const u64 ino1_gen,const u64 ino2,struct fs_path * fs_path)3886 static int is_ancestor(struct btrfs_root *root,
3887 const u64 ino1,
3888 const u64 ino1_gen,
3889 const u64 ino2,
3890 struct fs_path *fs_path)
3891 {
3892 bool free_fs_path = false;
3893 int ret = 0;
3894 int iter_ret = 0;
3895 struct btrfs_path *path = NULL;
3896 struct btrfs_key key;
3897
3898 if (!fs_path) {
3899 fs_path = fs_path_alloc();
3900 if (!fs_path)
3901 return -ENOMEM;
3902 free_fs_path = true;
3903 }
3904
3905 path = alloc_path_for_send();
3906 if (!path) {
3907 ret = -ENOMEM;
3908 goto out;
3909 }
3910
3911 key.objectid = ino2;
3912 key.type = BTRFS_INODE_REF_KEY;
3913 key.offset = 0;
3914
3915 btrfs_for_each_slot(root, &key, &key, path, iter_ret) {
3916 struct extent_buffer *leaf = path->nodes[0];
3917 int slot = path->slots[0];
3918 u32 cur_offset = 0;
3919 u32 item_size;
3920
3921 if (key.objectid != ino2)
3922 break;
3923 if (key.type != BTRFS_INODE_REF_KEY &&
3924 key.type != BTRFS_INODE_EXTREF_KEY)
3925 break;
3926
3927 item_size = btrfs_item_size(leaf, slot);
3928 while (cur_offset < item_size) {
3929 u64 parent;
3930 u64 parent_gen;
3931
3932 if (key.type == BTRFS_INODE_EXTREF_KEY) {
3933 unsigned long ptr;
3934 struct btrfs_inode_extref *extref;
3935
3936 ptr = btrfs_item_ptr_offset(leaf, slot);
3937 extref = (struct btrfs_inode_extref *)
3938 (ptr + cur_offset);
3939 parent = btrfs_inode_extref_parent(leaf,
3940 extref);
3941 cur_offset += sizeof(*extref);
3942 cur_offset += btrfs_inode_extref_name_len(leaf,
3943 extref);
3944 } else {
3945 parent = key.offset;
3946 cur_offset = item_size;
3947 }
3948
3949 ret = get_inode_gen(root, parent, &parent_gen);
3950 if (ret < 0)
3951 goto out;
3952 ret = check_ino_in_path(root, ino1, ino1_gen,
3953 parent, parent_gen, fs_path);
3954 if (ret)
3955 goto out;
3956 }
3957 }
3958 ret = 0;
3959 if (iter_ret < 0)
3960 ret = iter_ret;
3961
3962 out:
3963 btrfs_free_path(path);
3964 if (free_fs_path)
3965 fs_path_free(fs_path);
3966 return ret;
3967 }
3968
wait_for_parent_move(struct send_ctx * sctx,struct recorded_ref * parent_ref,const bool is_orphan)3969 static int wait_for_parent_move(struct send_ctx *sctx,
3970 struct recorded_ref *parent_ref,
3971 const bool is_orphan)
3972 {
3973 int ret = 0;
3974 u64 ino = parent_ref->dir;
3975 u64 ino_gen = parent_ref->dir_gen;
3976 u64 parent_ino_before, parent_ino_after;
3977 struct fs_path *path_before = NULL;
3978 struct fs_path *path_after = NULL;
3979 int len1, len2;
3980
3981 path_after = fs_path_alloc();
3982 path_before = fs_path_alloc();
3983 if (!path_after || !path_before) {
3984 ret = -ENOMEM;
3985 goto out;
3986 }
3987
3988 /*
3989 * Our current directory inode may not yet be renamed/moved because some
3990 * ancestor (immediate or not) has to be renamed/moved first. So find if
3991 * such ancestor exists and make sure our own rename/move happens after
3992 * that ancestor is processed to avoid path build infinite loops (done
3993 * at get_cur_path()).
3994 */
3995 while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3996 u64 parent_ino_after_gen;
3997
3998 if (is_waiting_for_move(sctx, ino)) {
3999 /*
4000 * If the current inode is an ancestor of ino in the
4001 * parent root, we need to delay the rename of the
4002 * current inode, otherwise don't delayed the rename
4003 * because we can end up with a circular dependency
4004 * of renames, resulting in some directories never
4005 * getting the respective rename operations issued in
4006 * the send stream or getting into infinite path build
4007 * loops.
4008 */
4009 ret = is_ancestor(sctx->parent_root,
4010 sctx->cur_ino, sctx->cur_inode_gen,
4011 ino, path_before);
4012 if (ret)
4013 break;
4014 }
4015
4016 fs_path_reset(path_before);
4017 fs_path_reset(path_after);
4018
4019 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
4020 &parent_ino_after_gen, path_after);
4021 if (ret < 0)
4022 goto out;
4023 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
4024 NULL, path_before);
4025 if (ret < 0 && ret != -ENOENT) {
4026 goto out;
4027 } else if (ret == -ENOENT) {
4028 ret = 0;
4029 break;
4030 }
4031
4032 len1 = fs_path_len(path_before);
4033 len2 = fs_path_len(path_after);
4034 if (ino > sctx->cur_ino &&
4035 (parent_ino_before != parent_ino_after || len1 != len2 ||
4036 memcmp(path_before->start, path_after->start, len1))) {
4037 u64 parent_ino_gen;
4038
4039 ret = get_inode_gen(sctx->parent_root, ino, &parent_ino_gen);
4040 if (ret < 0)
4041 goto out;
4042 if (ino_gen == parent_ino_gen) {
4043 ret = 1;
4044 break;
4045 }
4046 }
4047 ino = parent_ino_after;
4048 ino_gen = parent_ino_after_gen;
4049 }
4050
4051 out:
4052 fs_path_free(path_before);
4053 fs_path_free(path_after);
4054
4055 if (ret == 1) {
4056 ret = add_pending_dir_move(sctx,
4057 sctx->cur_ino,
4058 sctx->cur_inode_gen,
4059 ino,
4060 &sctx->new_refs,
4061 &sctx->deleted_refs,
4062 is_orphan);
4063 if (!ret)
4064 ret = 1;
4065 }
4066
4067 return ret;
4068 }
4069
update_ref_path(struct send_ctx * sctx,struct recorded_ref * ref)4070 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4071 {
4072 int ret;
4073 struct fs_path *new_path;
4074
4075 /*
4076 * Our reference's name member points to its full_path member string, so
4077 * we use here a new path.
4078 */
4079 new_path = fs_path_alloc();
4080 if (!new_path)
4081 return -ENOMEM;
4082
4083 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
4084 if (ret < 0) {
4085 fs_path_free(new_path);
4086 return ret;
4087 }
4088 ret = fs_path_add(new_path, ref->name, ref->name_len);
4089 if (ret < 0) {
4090 fs_path_free(new_path);
4091 return ret;
4092 }
4093
4094 fs_path_free(ref->full_path);
4095 set_ref_path(ref, new_path);
4096
4097 return 0;
4098 }
4099
4100 /*
4101 * When processing the new references for an inode we may orphanize an existing
4102 * directory inode because its old name conflicts with one of the new references
4103 * of the current inode. Later, when processing another new reference of our
4104 * inode, we might need to orphanize another inode, but the path we have in the
4105 * reference reflects the pre-orphanization name of the directory we previously
4106 * orphanized. For example:
4107 *
4108 * parent snapshot looks like:
4109 *
4110 * . (ino 256)
4111 * |----- f1 (ino 257)
4112 * |----- f2 (ino 258)
4113 * |----- d1/ (ino 259)
4114 * |----- d2/ (ino 260)
4115 *
4116 * send snapshot looks like:
4117 *
4118 * . (ino 256)
4119 * |----- d1 (ino 258)
4120 * |----- f2/ (ino 259)
4121 * |----- f2_link/ (ino 260)
4122 * | |----- f1 (ino 257)
4123 * |
4124 * |----- d2 (ino 258)
4125 *
4126 * When processing inode 257 we compute the name for inode 259 as "d1", and we
4127 * cache it in the name cache. Later when we start processing inode 258, when
4128 * collecting all its new references we set a full path of "d1/d2" for its new
4129 * reference with name "d2". When we start processing the new references we
4130 * start by processing the new reference with name "d1", and this results in
4131 * orphanizing inode 259, since its old reference causes a conflict. Then we
4132 * move on the next new reference, with name "d2", and we find out we must
4133 * orphanize inode 260, as its old reference conflicts with ours - but for the
4134 * orphanization we use a source path corresponding to the path we stored in the
4135 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the
4136 * receiver fail since the path component "d1/" no longer exists, it was renamed
4137 * to "o259-6-0/" when processing the previous new reference. So in this case we
4138 * must recompute the path in the new reference and use it for the new
4139 * orphanization operation.
4140 */
refresh_ref_path(struct send_ctx * sctx,struct recorded_ref * ref)4141 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
4142 {
4143 char *name;
4144 int ret;
4145
4146 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL);
4147 if (!name)
4148 return -ENOMEM;
4149
4150 fs_path_reset(ref->full_path);
4151 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path);
4152 if (ret < 0)
4153 goto out;
4154
4155 ret = fs_path_add(ref->full_path, name, ref->name_len);
4156 if (ret < 0)
4157 goto out;
4158
4159 /* Update the reference's base name pointer. */
4160 set_ref_path(ref, ref->full_path);
4161 out:
4162 kfree(name);
4163 return ret;
4164 }
4165
4166 /*
4167 * This does all the move/link/unlink/rmdir magic.
4168 */
process_recorded_refs(struct send_ctx * sctx,int * pending_move)4169 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
4170 {
4171 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4172 int ret = 0;
4173 struct recorded_ref *cur;
4174 struct recorded_ref *cur2;
4175 LIST_HEAD(check_dirs);
4176 struct fs_path *valid_path = NULL;
4177 u64 ow_inode = 0;
4178 u64 ow_gen;
4179 u64 ow_mode;
4180 int did_overwrite = 0;
4181 int is_orphan = 0;
4182 u64 last_dir_ino_rm = 0;
4183 bool can_rename = true;
4184 bool orphanized_dir = false;
4185 bool orphanized_ancestor = false;
4186
4187 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
4188
4189 /*
4190 * This should never happen as the root dir always has the same ref
4191 * which is always '..'
4192 */
4193 BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
4194
4195 valid_path = fs_path_alloc();
4196 if (!valid_path) {
4197 ret = -ENOMEM;
4198 goto out;
4199 }
4200
4201 /*
4202 * First, check if the first ref of the current inode was overwritten
4203 * before. If yes, we know that the current inode was already orphanized
4204 * and thus use the orphan name. If not, we can use get_cur_path to
4205 * get the path of the first ref as it would like while receiving at
4206 * this point in time.
4207 * New inodes are always orphan at the beginning, so force to use the
4208 * orphan name in this case.
4209 * The first ref is stored in valid_path and will be updated if it
4210 * gets moved around.
4211 */
4212 if (!sctx->cur_inode_new) {
4213 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
4214 sctx->cur_inode_gen);
4215 if (ret < 0)
4216 goto out;
4217 if (ret)
4218 did_overwrite = 1;
4219 }
4220 if (sctx->cur_inode_new || did_overwrite) {
4221 ret = gen_unique_name(sctx, sctx->cur_ino,
4222 sctx->cur_inode_gen, valid_path);
4223 if (ret < 0)
4224 goto out;
4225 is_orphan = 1;
4226 } else {
4227 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4228 valid_path);
4229 if (ret < 0)
4230 goto out;
4231 }
4232
4233 /*
4234 * Before doing any rename and link operations, do a first pass on the
4235 * new references to orphanize any unprocessed inodes that may have a
4236 * reference that conflicts with one of the new references of the current
4237 * inode. This needs to happen first because a new reference may conflict
4238 * with the old reference of a parent directory, so we must make sure
4239 * that the path used for link and rename commands don't use an
4240 * orphanized name when an ancestor was not yet orphanized.
4241 *
4242 * Example:
4243 *
4244 * Parent snapshot:
4245 *
4246 * . (ino 256)
4247 * |----- testdir/ (ino 259)
4248 * | |----- a (ino 257)
4249 * |
4250 * |----- b (ino 258)
4251 *
4252 * Send snapshot:
4253 *
4254 * . (ino 256)
4255 * |----- testdir_2/ (ino 259)
4256 * | |----- a (ino 260)
4257 * |
4258 * |----- testdir (ino 257)
4259 * |----- b (ino 257)
4260 * |----- b2 (ino 258)
4261 *
4262 * Processing the new reference for inode 257 with name "b" may happen
4263 * before processing the new reference with name "testdir". If so, we
4264 * must make sure that by the time we send a link command to create the
4265 * hard link "b", inode 259 was already orphanized, since the generated
4266 * path in "valid_path" already contains the orphanized name for 259.
4267 * We are processing inode 257, so only later when processing 259 we do
4268 * the rename operation to change its temporary (orphanized) name to
4269 * "testdir_2".
4270 */
4271 list_for_each_entry(cur, &sctx->new_refs, list) {
4272 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4273 if (ret < 0)
4274 goto out;
4275 if (ret == inode_state_will_create)
4276 continue;
4277
4278 /*
4279 * Check if this new ref would overwrite the first ref of another
4280 * unprocessed inode. If yes, orphanize the overwritten inode.
4281 * If we find an overwritten ref that is not the first ref,
4282 * simply unlink it.
4283 */
4284 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4285 cur->name, cur->name_len,
4286 &ow_inode, &ow_gen, &ow_mode);
4287 if (ret < 0)
4288 goto out;
4289 if (ret) {
4290 ret = is_first_ref(sctx->parent_root,
4291 ow_inode, cur->dir, cur->name,
4292 cur->name_len);
4293 if (ret < 0)
4294 goto out;
4295 if (ret) {
4296 struct name_cache_entry *nce;
4297 struct waiting_dir_move *wdm;
4298
4299 if (orphanized_dir) {
4300 ret = refresh_ref_path(sctx, cur);
4301 if (ret < 0)
4302 goto out;
4303 }
4304
4305 ret = orphanize_inode(sctx, ow_inode, ow_gen,
4306 cur->full_path);
4307 if (ret < 0)
4308 goto out;
4309 if (S_ISDIR(ow_mode))
4310 orphanized_dir = true;
4311
4312 /*
4313 * If ow_inode has its rename operation delayed
4314 * make sure that its orphanized name is used in
4315 * the source path when performing its rename
4316 * operation.
4317 */
4318 wdm = get_waiting_dir_move(sctx, ow_inode);
4319 if (wdm)
4320 wdm->orphanized = true;
4321
4322 /*
4323 * Make sure we clear our orphanized inode's
4324 * name from the name cache. This is because the
4325 * inode ow_inode might be an ancestor of some
4326 * other inode that will be orphanized as well
4327 * later and has an inode number greater than
4328 * sctx->send_progress. We need to prevent
4329 * future name lookups from using the old name
4330 * and get instead the orphan name.
4331 */
4332 nce = name_cache_search(sctx, ow_inode, ow_gen);
4333 if (nce)
4334 btrfs_lru_cache_remove(&sctx->name_cache,
4335 &nce->entry);
4336
4337 /*
4338 * ow_inode might currently be an ancestor of
4339 * cur_ino, therefore compute valid_path (the
4340 * current path of cur_ino) again because it
4341 * might contain the pre-orphanization name of
4342 * ow_inode, which is no longer valid.
4343 */
4344 ret = is_ancestor(sctx->parent_root,
4345 ow_inode, ow_gen,
4346 sctx->cur_ino, NULL);
4347 if (ret > 0) {
4348 orphanized_ancestor = true;
4349 fs_path_reset(valid_path);
4350 ret = get_cur_path(sctx, sctx->cur_ino,
4351 sctx->cur_inode_gen,
4352 valid_path);
4353 }
4354 if (ret < 0)
4355 goto out;
4356 } else {
4357 /*
4358 * If we previously orphanized a directory that
4359 * collided with a new reference that we already
4360 * processed, recompute the current path because
4361 * that directory may be part of the path.
4362 */
4363 if (orphanized_dir) {
4364 ret = refresh_ref_path(sctx, cur);
4365 if (ret < 0)
4366 goto out;
4367 }
4368 ret = send_unlink(sctx, cur->full_path);
4369 if (ret < 0)
4370 goto out;
4371 }
4372 }
4373
4374 }
4375
4376 list_for_each_entry(cur, &sctx->new_refs, list) {
4377 /*
4378 * We may have refs where the parent directory does not exist
4379 * yet. This happens if the parent directories inum is higher
4380 * than the current inum. To handle this case, we create the
4381 * parent directory out of order. But we need to check if this
4382 * did already happen before due to other refs in the same dir.
4383 */
4384 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4385 if (ret < 0)
4386 goto out;
4387 if (ret == inode_state_will_create) {
4388 ret = 0;
4389 /*
4390 * First check if any of the current inodes refs did
4391 * already create the dir.
4392 */
4393 list_for_each_entry(cur2, &sctx->new_refs, list) {
4394 if (cur == cur2)
4395 break;
4396 if (cur2->dir == cur->dir) {
4397 ret = 1;
4398 break;
4399 }
4400 }
4401
4402 /*
4403 * If that did not happen, check if a previous inode
4404 * did already create the dir.
4405 */
4406 if (!ret)
4407 ret = did_create_dir(sctx, cur->dir);
4408 if (ret < 0)
4409 goto out;
4410 if (!ret) {
4411 ret = send_create_inode(sctx, cur->dir);
4412 if (ret < 0)
4413 goto out;
4414 cache_dir_created(sctx, cur->dir);
4415 }
4416 }
4417
4418 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4419 ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4420 if (ret < 0)
4421 goto out;
4422 if (ret == 1) {
4423 can_rename = false;
4424 *pending_move = 1;
4425 }
4426 }
4427
4428 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4429 can_rename) {
4430 ret = wait_for_parent_move(sctx, cur, is_orphan);
4431 if (ret < 0)
4432 goto out;
4433 if (ret == 1) {
4434 can_rename = false;
4435 *pending_move = 1;
4436 }
4437 }
4438
4439 /*
4440 * link/move the ref to the new place. If we have an orphan
4441 * inode, move it and update valid_path. If not, link or move
4442 * it depending on the inode mode.
4443 */
4444 if (is_orphan && can_rename) {
4445 ret = send_rename(sctx, valid_path, cur->full_path);
4446 if (ret < 0)
4447 goto out;
4448 is_orphan = 0;
4449 ret = fs_path_copy(valid_path, cur->full_path);
4450 if (ret < 0)
4451 goto out;
4452 } else if (can_rename) {
4453 if (S_ISDIR(sctx->cur_inode_mode)) {
4454 /*
4455 * Dirs can't be linked, so move it. For moved
4456 * dirs, we always have one new and one deleted
4457 * ref. The deleted ref is ignored later.
4458 */
4459 ret = send_rename(sctx, valid_path,
4460 cur->full_path);
4461 if (!ret)
4462 ret = fs_path_copy(valid_path,
4463 cur->full_path);
4464 if (ret < 0)
4465 goto out;
4466 } else {
4467 /*
4468 * We might have previously orphanized an inode
4469 * which is an ancestor of our current inode,
4470 * so our reference's full path, which was
4471 * computed before any such orphanizations, must
4472 * be updated.
4473 */
4474 if (orphanized_dir) {
4475 ret = update_ref_path(sctx, cur);
4476 if (ret < 0)
4477 goto out;
4478 }
4479 ret = send_link(sctx, cur->full_path,
4480 valid_path);
4481 if (ret < 0)
4482 goto out;
4483 }
4484 }
4485 ret = dup_ref(cur, &check_dirs);
4486 if (ret < 0)
4487 goto out;
4488 }
4489
4490 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4491 /*
4492 * Check if we can already rmdir the directory. If not,
4493 * orphanize it. For every dir item inside that gets deleted
4494 * later, we do this check again and rmdir it then if possible.
4495 * See the use of check_dirs for more details.
4496 */
4497 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen);
4498 if (ret < 0)
4499 goto out;
4500 if (ret) {
4501 ret = send_rmdir(sctx, valid_path);
4502 if (ret < 0)
4503 goto out;
4504 } else if (!is_orphan) {
4505 ret = orphanize_inode(sctx, sctx->cur_ino,
4506 sctx->cur_inode_gen, valid_path);
4507 if (ret < 0)
4508 goto out;
4509 is_orphan = 1;
4510 }
4511
4512 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4513 ret = dup_ref(cur, &check_dirs);
4514 if (ret < 0)
4515 goto out;
4516 }
4517 } else if (S_ISDIR(sctx->cur_inode_mode) &&
4518 !list_empty(&sctx->deleted_refs)) {
4519 /*
4520 * We have a moved dir. Add the old parent to check_dirs
4521 */
4522 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4523 list);
4524 ret = dup_ref(cur, &check_dirs);
4525 if (ret < 0)
4526 goto out;
4527 } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4528 /*
4529 * We have a non dir inode. Go through all deleted refs and
4530 * unlink them if they were not already overwritten by other
4531 * inodes.
4532 */
4533 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4534 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4535 sctx->cur_ino, sctx->cur_inode_gen,
4536 cur->name, cur->name_len);
4537 if (ret < 0)
4538 goto out;
4539 if (!ret) {
4540 /*
4541 * If we orphanized any ancestor before, we need
4542 * to recompute the full path for deleted names,
4543 * since any such path was computed before we
4544 * processed any references and orphanized any
4545 * ancestor inode.
4546 */
4547 if (orphanized_ancestor) {
4548 ret = update_ref_path(sctx, cur);
4549 if (ret < 0)
4550 goto out;
4551 }
4552 ret = send_unlink(sctx, cur->full_path);
4553 if (ret < 0)
4554 goto out;
4555 }
4556 ret = dup_ref(cur, &check_dirs);
4557 if (ret < 0)
4558 goto out;
4559 }
4560 /*
4561 * If the inode is still orphan, unlink the orphan. This may
4562 * happen when a previous inode did overwrite the first ref
4563 * of this inode and no new refs were added for the current
4564 * inode. Unlinking does not mean that the inode is deleted in
4565 * all cases. There may still be links to this inode in other
4566 * places.
4567 */
4568 if (is_orphan) {
4569 ret = send_unlink(sctx, valid_path);
4570 if (ret < 0)
4571 goto out;
4572 }
4573 }
4574
4575 /*
4576 * We did collect all parent dirs where cur_inode was once located. We
4577 * now go through all these dirs and check if they are pending for
4578 * deletion and if it's finally possible to perform the rmdir now.
4579 * We also update the inode stats of the parent dirs here.
4580 */
4581 list_for_each_entry(cur, &check_dirs, list) {
4582 /*
4583 * In case we had refs into dirs that were not processed yet,
4584 * we don't need to do the utime and rmdir logic for these dirs.
4585 * The dir will be processed later.
4586 */
4587 if (cur->dir > sctx->cur_ino)
4588 continue;
4589
4590 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen, NULL, NULL);
4591 if (ret < 0)
4592 goto out;
4593
4594 if (ret == inode_state_did_create ||
4595 ret == inode_state_no_change) {
4596 ret = cache_dir_utimes(sctx, cur->dir, cur->dir_gen);
4597 if (ret < 0)
4598 goto out;
4599 } else if (ret == inode_state_did_delete &&
4600 cur->dir != last_dir_ino_rm) {
4601 ret = can_rmdir(sctx, cur->dir, cur->dir_gen);
4602 if (ret < 0)
4603 goto out;
4604 if (ret) {
4605 ret = get_cur_path(sctx, cur->dir,
4606 cur->dir_gen, valid_path);
4607 if (ret < 0)
4608 goto out;
4609 ret = send_rmdir(sctx, valid_path);
4610 if (ret < 0)
4611 goto out;
4612 last_dir_ino_rm = cur->dir;
4613 }
4614 }
4615 }
4616
4617 ret = 0;
4618
4619 out:
4620 __free_recorded_refs(&check_dirs);
4621 free_recorded_refs(sctx);
4622 fs_path_free(valid_path);
4623 return ret;
4624 }
4625
rbtree_ref_comp(const void * k,const struct rb_node * node)4626 static int rbtree_ref_comp(const void *k, const struct rb_node *node)
4627 {
4628 const struct recorded_ref *data = k;
4629 const struct recorded_ref *ref = rb_entry(node, struct recorded_ref, node);
4630 int result;
4631
4632 if (data->dir > ref->dir)
4633 return 1;
4634 if (data->dir < ref->dir)
4635 return -1;
4636 if (data->dir_gen > ref->dir_gen)
4637 return 1;
4638 if (data->dir_gen < ref->dir_gen)
4639 return -1;
4640 if (data->name_len > ref->name_len)
4641 return 1;
4642 if (data->name_len < ref->name_len)
4643 return -1;
4644 result = strcmp(data->name, ref->name);
4645 if (result > 0)
4646 return 1;
4647 if (result < 0)
4648 return -1;
4649 return 0;
4650 }
4651
rbtree_ref_less(struct rb_node * node,const struct rb_node * parent)4652 static bool rbtree_ref_less(struct rb_node *node, const struct rb_node *parent)
4653 {
4654 const struct recorded_ref *entry = rb_entry(node, struct recorded_ref, node);
4655
4656 return rbtree_ref_comp(entry, parent) < 0;
4657 }
4658
record_ref_in_tree(struct rb_root * root,struct list_head * refs,struct fs_path * name,u64 dir,u64 dir_gen,struct send_ctx * sctx)4659 static int record_ref_in_tree(struct rb_root *root, struct list_head *refs,
4660 struct fs_path *name, u64 dir, u64 dir_gen,
4661 struct send_ctx *sctx)
4662 {
4663 int ret = 0;
4664 struct fs_path *path = NULL;
4665 struct recorded_ref *ref = NULL;
4666
4667 path = fs_path_alloc();
4668 if (!path) {
4669 ret = -ENOMEM;
4670 goto out;
4671 }
4672
4673 ref = recorded_ref_alloc();
4674 if (!ref) {
4675 ret = -ENOMEM;
4676 goto out;
4677 }
4678
4679 ret = get_cur_path(sctx, dir, dir_gen, path);
4680 if (ret < 0)
4681 goto out;
4682 ret = fs_path_add_path(path, name);
4683 if (ret < 0)
4684 goto out;
4685
4686 ref->dir = dir;
4687 ref->dir_gen = dir_gen;
4688 set_ref_path(ref, path);
4689 list_add_tail(&ref->list, refs);
4690 rb_add(&ref->node, root, rbtree_ref_less);
4691 ref->root = root;
4692 out:
4693 if (ret) {
4694 if (path && (!ref || !ref->full_path))
4695 fs_path_free(path);
4696 recorded_ref_free(ref);
4697 }
4698 return ret;
4699 }
4700
record_new_ref_if_needed(int num,u64 dir,int index,struct fs_path * name,void * ctx)4701 static int record_new_ref_if_needed(int num, u64 dir, int index,
4702 struct fs_path *name, void *ctx)
4703 {
4704 int ret = 0;
4705 struct send_ctx *sctx = ctx;
4706 struct rb_node *node = NULL;
4707 struct recorded_ref data;
4708 struct recorded_ref *ref;
4709 u64 dir_gen;
4710
4711 ret = get_inode_gen(sctx->send_root, dir, &dir_gen);
4712 if (ret < 0)
4713 goto out;
4714
4715 data.dir = dir;
4716 data.dir_gen = dir_gen;
4717 set_ref_path(&data, name);
4718 node = rb_find(&data, &sctx->rbtree_deleted_refs, rbtree_ref_comp);
4719 if (node) {
4720 ref = rb_entry(node, struct recorded_ref, node);
4721 recorded_ref_free(ref);
4722 } else {
4723 ret = record_ref_in_tree(&sctx->rbtree_new_refs,
4724 &sctx->new_refs, name, dir, dir_gen,
4725 sctx);
4726 }
4727 out:
4728 return ret;
4729 }
4730
record_deleted_ref_if_needed(int num,u64 dir,int index,struct fs_path * name,void * ctx)4731 static int record_deleted_ref_if_needed(int num, u64 dir, int index,
4732 struct fs_path *name, void *ctx)
4733 {
4734 int ret = 0;
4735 struct send_ctx *sctx = ctx;
4736 struct rb_node *node = NULL;
4737 struct recorded_ref data;
4738 struct recorded_ref *ref;
4739 u64 dir_gen;
4740
4741 ret = get_inode_gen(sctx->parent_root, dir, &dir_gen);
4742 if (ret < 0)
4743 goto out;
4744
4745 data.dir = dir;
4746 data.dir_gen = dir_gen;
4747 set_ref_path(&data, name);
4748 node = rb_find(&data, &sctx->rbtree_new_refs, rbtree_ref_comp);
4749 if (node) {
4750 ref = rb_entry(node, struct recorded_ref, node);
4751 recorded_ref_free(ref);
4752 } else {
4753 ret = record_ref_in_tree(&sctx->rbtree_deleted_refs,
4754 &sctx->deleted_refs, name, dir,
4755 dir_gen, sctx);
4756 }
4757 out:
4758 return ret;
4759 }
4760
record_new_ref(struct send_ctx * sctx)4761 static int record_new_ref(struct send_ctx *sctx)
4762 {
4763 int ret;
4764
4765 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4766 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4767 if (ret < 0)
4768 goto out;
4769 ret = 0;
4770
4771 out:
4772 return ret;
4773 }
4774
record_deleted_ref(struct send_ctx * sctx)4775 static int record_deleted_ref(struct send_ctx *sctx)
4776 {
4777 int ret;
4778
4779 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4780 sctx->cmp_key, 0, record_deleted_ref_if_needed,
4781 sctx);
4782 if (ret < 0)
4783 goto out;
4784 ret = 0;
4785
4786 out:
4787 return ret;
4788 }
4789
record_changed_ref(struct send_ctx * sctx)4790 static int record_changed_ref(struct send_ctx *sctx)
4791 {
4792 int ret = 0;
4793
4794 ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4795 sctx->cmp_key, 0, record_new_ref_if_needed, sctx);
4796 if (ret < 0)
4797 goto out;
4798 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4799 sctx->cmp_key, 0, record_deleted_ref_if_needed, sctx);
4800 if (ret < 0)
4801 goto out;
4802 ret = 0;
4803
4804 out:
4805 return ret;
4806 }
4807
4808 /*
4809 * Record and process all refs at once. Needed when an inode changes the
4810 * generation number, which means that it was deleted and recreated.
4811 */
process_all_refs(struct send_ctx * sctx,enum btrfs_compare_tree_result cmd)4812 static int process_all_refs(struct send_ctx *sctx,
4813 enum btrfs_compare_tree_result cmd)
4814 {
4815 int ret = 0;
4816 int iter_ret = 0;
4817 struct btrfs_root *root;
4818 struct btrfs_path *path;
4819 struct btrfs_key key;
4820 struct btrfs_key found_key;
4821 iterate_inode_ref_t cb;
4822 int pending_move = 0;
4823
4824 path = alloc_path_for_send();
4825 if (!path)
4826 return -ENOMEM;
4827
4828 if (cmd == BTRFS_COMPARE_TREE_NEW) {
4829 root = sctx->send_root;
4830 cb = record_new_ref_if_needed;
4831 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4832 root = sctx->parent_root;
4833 cb = record_deleted_ref_if_needed;
4834 } else {
4835 btrfs_err(sctx->send_root->fs_info,
4836 "Wrong command %d in process_all_refs", cmd);
4837 ret = -EINVAL;
4838 goto out;
4839 }
4840
4841 key.objectid = sctx->cmp_key->objectid;
4842 key.type = BTRFS_INODE_REF_KEY;
4843 key.offset = 0;
4844 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
4845 if (found_key.objectid != key.objectid ||
4846 (found_key.type != BTRFS_INODE_REF_KEY &&
4847 found_key.type != BTRFS_INODE_EXTREF_KEY))
4848 break;
4849
4850 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4851 if (ret < 0)
4852 goto out;
4853 }
4854 /* Catch error found during iteration */
4855 if (iter_ret < 0) {
4856 ret = iter_ret;
4857 goto out;
4858 }
4859 btrfs_release_path(path);
4860
4861 /*
4862 * We don't actually care about pending_move as we are simply
4863 * re-creating this inode and will be rename'ing it into place once we
4864 * rename the parent directory.
4865 */
4866 ret = process_recorded_refs(sctx, &pending_move);
4867 out:
4868 btrfs_free_path(path);
4869 return ret;
4870 }
4871
send_set_xattr(struct send_ctx * sctx,struct fs_path * path,const char * name,int name_len,const char * data,int data_len)4872 static int send_set_xattr(struct send_ctx *sctx,
4873 struct fs_path *path,
4874 const char *name, int name_len,
4875 const char *data, int data_len)
4876 {
4877 int ret = 0;
4878
4879 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4880 if (ret < 0)
4881 goto out;
4882
4883 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4884 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4885 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4886
4887 ret = send_cmd(sctx);
4888
4889 tlv_put_failure:
4890 out:
4891 return ret;
4892 }
4893
send_remove_xattr(struct send_ctx * sctx,struct fs_path * path,const char * name,int name_len)4894 static int send_remove_xattr(struct send_ctx *sctx,
4895 struct fs_path *path,
4896 const char *name, int name_len)
4897 {
4898 int ret = 0;
4899
4900 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4901 if (ret < 0)
4902 goto out;
4903
4904 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4905 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4906
4907 ret = send_cmd(sctx);
4908
4909 tlv_put_failure:
4910 out:
4911 return ret;
4912 }
4913
__process_new_xattr(int num,struct btrfs_key * di_key,const char * name,int name_len,const char * data,int data_len,void * ctx)4914 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4915 const char *name, int name_len, const char *data,
4916 int data_len, void *ctx)
4917 {
4918 int ret;
4919 struct send_ctx *sctx = ctx;
4920 struct fs_path *p;
4921 struct posix_acl_xattr_header dummy_acl;
4922
4923 /* Capabilities are emitted by finish_inode_if_needed */
4924 if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4925 return 0;
4926
4927 p = fs_path_alloc();
4928 if (!p)
4929 return -ENOMEM;
4930
4931 /*
4932 * This hack is needed because empty acls are stored as zero byte
4933 * data in xattrs. Problem with that is, that receiving these zero byte
4934 * acls will fail later. To fix this, we send a dummy acl list that
4935 * only contains the version number and no entries.
4936 */
4937 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4938 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4939 if (data_len == 0) {
4940 dummy_acl.a_version =
4941 cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4942 data = (char *)&dummy_acl;
4943 data_len = sizeof(dummy_acl);
4944 }
4945 }
4946
4947 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4948 if (ret < 0)
4949 goto out;
4950
4951 ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4952
4953 out:
4954 fs_path_free(p);
4955 return ret;
4956 }
4957
__process_deleted_xattr(int num,struct btrfs_key * di_key,const char * name,int name_len,const char * data,int data_len,void * ctx)4958 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4959 const char *name, int name_len,
4960 const char *data, int data_len, void *ctx)
4961 {
4962 int ret;
4963 struct send_ctx *sctx = ctx;
4964 struct fs_path *p;
4965
4966 p = fs_path_alloc();
4967 if (!p)
4968 return -ENOMEM;
4969
4970 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4971 if (ret < 0)
4972 goto out;
4973
4974 ret = send_remove_xattr(sctx, p, name, name_len);
4975
4976 out:
4977 fs_path_free(p);
4978 return ret;
4979 }
4980
process_new_xattr(struct send_ctx * sctx)4981 static int process_new_xattr(struct send_ctx *sctx)
4982 {
4983 int ret = 0;
4984
4985 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4986 __process_new_xattr, sctx);
4987
4988 return ret;
4989 }
4990
process_deleted_xattr(struct send_ctx * sctx)4991 static int process_deleted_xattr(struct send_ctx *sctx)
4992 {
4993 return iterate_dir_item(sctx->parent_root, sctx->right_path,
4994 __process_deleted_xattr, sctx);
4995 }
4996
4997 struct find_xattr_ctx {
4998 const char *name;
4999 int name_len;
5000 int found_idx;
5001 char *found_data;
5002 int found_data_len;
5003 };
5004
__find_xattr(int num,struct btrfs_key * di_key,const char * name,int name_len,const char * data,int data_len,void * vctx)5005 static int __find_xattr(int num, struct btrfs_key *di_key, const char *name,
5006 int name_len, const char *data, int data_len, void *vctx)
5007 {
5008 struct find_xattr_ctx *ctx = vctx;
5009
5010 if (name_len == ctx->name_len &&
5011 strncmp(name, ctx->name, name_len) == 0) {
5012 ctx->found_idx = num;
5013 ctx->found_data_len = data_len;
5014 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
5015 if (!ctx->found_data)
5016 return -ENOMEM;
5017 return 1;
5018 }
5019 return 0;
5020 }
5021
find_xattr(struct btrfs_root * root,struct btrfs_path * path,struct btrfs_key * key,const char * name,int name_len,char ** data,int * data_len)5022 static int find_xattr(struct btrfs_root *root,
5023 struct btrfs_path *path,
5024 struct btrfs_key *key,
5025 const char *name, int name_len,
5026 char **data, int *data_len)
5027 {
5028 int ret;
5029 struct find_xattr_ctx ctx;
5030
5031 ctx.name = name;
5032 ctx.name_len = name_len;
5033 ctx.found_idx = -1;
5034 ctx.found_data = NULL;
5035 ctx.found_data_len = 0;
5036
5037 ret = iterate_dir_item(root, path, __find_xattr, &ctx);
5038 if (ret < 0)
5039 return ret;
5040
5041 if (ctx.found_idx == -1)
5042 return -ENOENT;
5043 if (data) {
5044 *data = ctx.found_data;
5045 *data_len = ctx.found_data_len;
5046 } else {
5047 kfree(ctx.found_data);
5048 }
5049 return ctx.found_idx;
5050 }
5051
5052
__process_changed_new_xattr(int num,struct btrfs_key * di_key,const char * name,int name_len,const char * data,int data_len,void * ctx)5053 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
5054 const char *name, int name_len,
5055 const char *data, int data_len,
5056 void *ctx)
5057 {
5058 int ret;
5059 struct send_ctx *sctx = ctx;
5060 char *found_data = NULL;
5061 int found_data_len = 0;
5062
5063 ret = find_xattr(sctx->parent_root, sctx->right_path,
5064 sctx->cmp_key, name, name_len, &found_data,
5065 &found_data_len);
5066 if (ret == -ENOENT) {
5067 ret = __process_new_xattr(num, di_key, name, name_len, data,
5068 data_len, ctx);
5069 } else if (ret >= 0) {
5070 if (data_len != found_data_len ||
5071 memcmp(data, found_data, data_len)) {
5072 ret = __process_new_xattr(num, di_key, name, name_len,
5073 data, data_len, ctx);
5074 } else {
5075 ret = 0;
5076 }
5077 }
5078
5079 kfree(found_data);
5080 return ret;
5081 }
5082
__process_changed_deleted_xattr(int num,struct btrfs_key * di_key,const char * name,int name_len,const char * data,int data_len,void * ctx)5083 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
5084 const char *name, int name_len,
5085 const char *data, int data_len,
5086 void *ctx)
5087 {
5088 int ret;
5089 struct send_ctx *sctx = ctx;
5090
5091 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
5092 name, name_len, NULL, NULL);
5093 if (ret == -ENOENT)
5094 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
5095 data_len, ctx);
5096 else if (ret >= 0)
5097 ret = 0;
5098
5099 return ret;
5100 }
5101
process_changed_xattr(struct send_ctx * sctx)5102 static int process_changed_xattr(struct send_ctx *sctx)
5103 {
5104 int ret = 0;
5105
5106 ret = iterate_dir_item(sctx->send_root, sctx->left_path,
5107 __process_changed_new_xattr, sctx);
5108 if (ret < 0)
5109 goto out;
5110 ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
5111 __process_changed_deleted_xattr, sctx);
5112
5113 out:
5114 return ret;
5115 }
5116
process_all_new_xattrs(struct send_ctx * sctx)5117 static int process_all_new_xattrs(struct send_ctx *sctx)
5118 {
5119 int ret = 0;
5120 int iter_ret = 0;
5121 struct btrfs_root *root;
5122 struct btrfs_path *path;
5123 struct btrfs_key key;
5124 struct btrfs_key found_key;
5125
5126 path = alloc_path_for_send();
5127 if (!path)
5128 return -ENOMEM;
5129
5130 root = sctx->send_root;
5131
5132 key.objectid = sctx->cmp_key->objectid;
5133 key.type = BTRFS_XATTR_ITEM_KEY;
5134 key.offset = 0;
5135 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
5136 if (found_key.objectid != key.objectid ||
5137 found_key.type != key.type) {
5138 ret = 0;
5139 break;
5140 }
5141
5142 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
5143 if (ret < 0)
5144 break;
5145 }
5146 /* Catch error found during iteration */
5147 if (iter_ret < 0)
5148 ret = iter_ret;
5149
5150 btrfs_free_path(path);
5151 return ret;
5152 }
5153
send_verity(struct send_ctx * sctx,struct fs_path * path,struct fsverity_descriptor * desc)5154 static int send_verity(struct send_ctx *sctx, struct fs_path *path,
5155 struct fsverity_descriptor *desc)
5156 {
5157 int ret;
5158
5159 ret = begin_cmd(sctx, BTRFS_SEND_C_ENABLE_VERITY);
5160 if (ret < 0)
5161 goto out;
5162
5163 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
5164 TLV_PUT_U8(sctx, BTRFS_SEND_A_VERITY_ALGORITHM,
5165 le8_to_cpu(desc->hash_algorithm));
5166 TLV_PUT_U32(sctx, BTRFS_SEND_A_VERITY_BLOCK_SIZE,
5167 1U << le8_to_cpu(desc->log_blocksize));
5168 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SALT_DATA, desc->salt,
5169 le8_to_cpu(desc->salt_size));
5170 TLV_PUT(sctx, BTRFS_SEND_A_VERITY_SIG_DATA, desc->signature,
5171 le32_to_cpu(desc->sig_size));
5172
5173 ret = send_cmd(sctx);
5174
5175 tlv_put_failure:
5176 out:
5177 return ret;
5178 }
5179
process_verity(struct send_ctx * sctx)5180 static int process_verity(struct send_ctx *sctx)
5181 {
5182 int ret = 0;
5183 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5184 struct inode *inode;
5185 struct fs_path *p;
5186
5187 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, sctx->send_root);
5188 if (IS_ERR(inode))
5189 return PTR_ERR(inode);
5190
5191 ret = btrfs_get_verity_descriptor(inode, NULL, 0);
5192 if (ret < 0)
5193 goto iput;
5194
5195 if (ret > FS_VERITY_MAX_DESCRIPTOR_SIZE) {
5196 ret = -EMSGSIZE;
5197 goto iput;
5198 }
5199 if (!sctx->verity_descriptor) {
5200 sctx->verity_descriptor = kvmalloc(FS_VERITY_MAX_DESCRIPTOR_SIZE,
5201 GFP_KERNEL);
5202 if (!sctx->verity_descriptor) {
5203 ret = -ENOMEM;
5204 goto iput;
5205 }
5206 }
5207
5208 ret = btrfs_get_verity_descriptor(inode, sctx->verity_descriptor, ret);
5209 if (ret < 0)
5210 goto iput;
5211
5212 p = fs_path_alloc();
5213 if (!p) {
5214 ret = -ENOMEM;
5215 goto iput;
5216 }
5217 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5218 if (ret < 0)
5219 goto free_path;
5220
5221 ret = send_verity(sctx, p, sctx->verity_descriptor);
5222 if (ret < 0)
5223 goto free_path;
5224
5225 free_path:
5226 fs_path_free(p);
5227 iput:
5228 iput(inode);
5229 return ret;
5230 }
5231
max_send_read_size(const struct send_ctx * sctx)5232 static inline u64 max_send_read_size(const struct send_ctx *sctx)
5233 {
5234 return sctx->send_max_size - SZ_16K;
5235 }
5236
put_data_header(struct send_ctx * sctx,u32 len)5237 static int put_data_header(struct send_ctx *sctx, u32 len)
5238 {
5239 if (WARN_ON_ONCE(sctx->put_data))
5240 return -EINVAL;
5241 sctx->put_data = true;
5242 if (sctx->proto >= 2) {
5243 /*
5244 * Since v2, the data attribute header doesn't include a length,
5245 * it is implicitly to the end of the command.
5246 */
5247 if (sctx->send_max_size - sctx->send_size < sizeof(__le16) + len)
5248 return -EOVERFLOW;
5249 put_unaligned_le16(BTRFS_SEND_A_DATA, sctx->send_buf + sctx->send_size);
5250 sctx->send_size += sizeof(__le16);
5251 } else {
5252 struct btrfs_tlv_header *hdr;
5253
5254 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len)
5255 return -EOVERFLOW;
5256 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size);
5257 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type);
5258 put_unaligned_le16(len, &hdr->tlv_len);
5259 sctx->send_size += sizeof(*hdr);
5260 }
5261 return 0;
5262 }
5263
put_file_data(struct send_ctx * sctx,u64 offset,u32 len)5264 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len)
5265 {
5266 struct btrfs_root *root = sctx->send_root;
5267 struct btrfs_fs_info *fs_info = root->fs_info;
5268 struct page *page;
5269 pgoff_t index = offset >> PAGE_SHIFT;
5270 pgoff_t last_index;
5271 unsigned pg_offset = offset_in_page(offset);
5272 int ret;
5273
5274 ret = put_data_header(sctx, len);
5275 if (ret)
5276 return ret;
5277
5278 last_index = (offset + len - 1) >> PAGE_SHIFT;
5279
5280 while (index <= last_index) {
5281 unsigned cur_len = min_t(unsigned, len,
5282 PAGE_SIZE - pg_offset);
5283
5284 page = find_lock_page(sctx->cur_inode->i_mapping, index);
5285 if (!page) {
5286 page_cache_sync_readahead(sctx->cur_inode->i_mapping,
5287 &sctx->ra, NULL, index,
5288 last_index + 1 - index);
5289
5290 page = find_or_create_page(sctx->cur_inode->i_mapping,
5291 index, GFP_KERNEL);
5292 if (!page) {
5293 ret = -ENOMEM;
5294 break;
5295 }
5296 }
5297
5298 if (PageReadahead(page))
5299 page_cache_async_readahead(sctx->cur_inode->i_mapping,
5300 &sctx->ra, NULL, page_folio(page),
5301 index, last_index + 1 - index);
5302
5303 if (!PageUptodate(page)) {
5304 btrfs_read_folio(NULL, page_folio(page));
5305 lock_page(page);
5306 if (!PageUptodate(page)) {
5307 unlock_page(page);
5308 btrfs_err(fs_info,
5309 "send: IO error at offset %llu for inode %llu root %llu",
5310 page_offset(page), sctx->cur_ino,
5311 sctx->send_root->root_key.objectid);
5312 put_page(page);
5313 ret = -EIO;
5314 break;
5315 }
5316 }
5317
5318 memcpy_from_page(sctx->send_buf + sctx->send_size, page,
5319 pg_offset, cur_len);
5320 unlock_page(page);
5321 put_page(page);
5322 index++;
5323 pg_offset = 0;
5324 len -= cur_len;
5325 sctx->send_size += cur_len;
5326 }
5327
5328 return ret;
5329 }
5330
5331 /*
5332 * Read some bytes from the current inode/file and send a write command to
5333 * user space.
5334 */
send_write(struct send_ctx * sctx,u64 offset,u32 len)5335 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
5336 {
5337 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
5338 int ret = 0;
5339 struct fs_path *p;
5340
5341 p = fs_path_alloc();
5342 if (!p)
5343 return -ENOMEM;
5344
5345 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
5346
5347 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5348 if (ret < 0)
5349 goto out;
5350
5351 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5352 if (ret < 0)
5353 goto out;
5354
5355 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5356 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5357 ret = put_file_data(sctx, offset, len);
5358 if (ret < 0)
5359 goto out;
5360
5361 ret = send_cmd(sctx);
5362
5363 tlv_put_failure:
5364 out:
5365 fs_path_free(p);
5366 return ret;
5367 }
5368
5369 /*
5370 * Send a clone command to user space.
5371 */
send_clone(struct send_ctx * sctx,u64 offset,u32 len,struct clone_root * clone_root)5372 static int send_clone(struct send_ctx *sctx,
5373 u64 offset, u32 len,
5374 struct clone_root *clone_root)
5375 {
5376 int ret = 0;
5377 struct fs_path *p;
5378 u64 gen;
5379
5380 btrfs_debug(sctx->send_root->fs_info,
5381 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
5382 offset, len, clone_root->root->root_key.objectid,
5383 clone_root->ino, clone_root->offset);
5384
5385 p = fs_path_alloc();
5386 if (!p)
5387 return -ENOMEM;
5388
5389 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
5390 if (ret < 0)
5391 goto out;
5392
5393 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5394 if (ret < 0)
5395 goto out;
5396
5397 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5398 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
5399 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5400
5401 if (clone_root->root == sctx->send_root) {
5402 ret = get_inode_gen(sctx->send_root, clone_root->ino, &gen);
5403 if (ret < 0)
5404 goto out;
5405 ret = get_cur_path(sctx, clone_root->ino, gen, p);
5406 } else {
5407 ret = get_inode_path(clone_root->root, clone_root->ino, p);
5408 }
5409 if (ret < 0)
5410 goto out;
5411
5412 /*
5413 * If the parent we're using has a received_uuid set then use that as
5414 * our clone source as that is what we will look for when doing a
5415 * receive.
5416 *
5417 * This covers the case that we create a snapshot off of a received
5418 * subvolume and then use that as the parent and try to receive on a
5419 * different host.
5420 */
5421 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
5422 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5423 clone_root->root->root_item.received_uuid);
5424 else
5425 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
5426 clone_root->root->root_item.uuid);
5427 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
5428 btrfs_root_ctransid(&clone_root->root->root_item));
5429 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
5430 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
5431 clone_root->offset);
5432
5433 ret = send_cmd(sctx);
5434
5435 tlv_put_failure:
5436 out:
5437 fs_path_free(p);
5438 return ret;
5439 }
5440
5441 /*
5442 * Send an update extent command to user space.
5443 */
send_update_extent(struct send_ctx * sctx,u64 offset,u32 len)5444 static int send_update_extent(struct send_ctx *sctx,
5445 u64 offset, u32 len)
5446 {
5447 int ret = 0;
5448 struct fs_path *p;
5449
5450 p = fs_path_alloc();
5451 if (!p)
5452 return -ENOMEM;
5453
5454 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5455 if (ret < 0)
5456 goto out;
5457
5458 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5459 if (ret < 0)
5460 goto out;
5461
5462 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5463 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5464 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5465
5466 ret = send_cmd(sctx);
5467
5468 tlv_put_failure:
5469 out:
5470 fs_path_free(p);
5471 return ret;
5472 }
5473
send_hole(struct send_ctx * sctx,u64 end)5474 static int send_hole(struct send_ctx *sctx, u64 end)
5475 {
5476 struct fs_path *p = NULL;
5477 u64 read_size = max_send_read_size(sctx);
5478 u64 offset = sctx->cur_inode_last_extent;
5479 int ret = 0;
5480
5481 /*
5482 * A hole that starts at EOF or beyond it. Since we do not yet support
5483 * fallocate (for extent preallocation and hole punching), sending a
5484 * write of zeroes starting at EOF or beyond would later require issuing
5485 * a truncate operation which would undo the write and achieve nothing.
5486 */
5487 if (offset >= sctx->cur_inode_size)
5488 return 0;
5489
5490 /*
5491 * Don't go beyond the inode's i_size due to prealloc extents that start
5492 * after the i_size.
5493 */
5494 end = min_t(u64, end, sctx->cur_inode_size);
5495
5496 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5497 return send_update_extent(sctx, offset, end - offset);
5498
5499 p = fs_path_alloc();
5500 if (!p)
5501 return -ENOMEM;
5502 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5503 if (ret < 0)
5504 goto tlv_put_failure;
5505 while (offset < end) {
5506 u64 len = min(end - offset, read_size);
5507
5508 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5509 if (ret < 0)
5510 break;
5511 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5512 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5513 ret = put_data_header(sctx, len);
5514 if (ret < 0)
5515 break;
5516 memset(sctx->send_buf + sctx->send_size, 0, len);
5517 sctx->send_size += len;
5518 ret = send_cmd(sctx);
5519 if (ret < 0)
5520 break;
5521 offset += len;
5522 }
5523 sctx->cur_inode_next_write_offset = offset;
5524 tlv_put_failure:
5525 fs_path_free(p);
5526 return ret;
5527 }
5528
send_encoded_inline_extent(struct send_ctx * sctx,struct btrfs_path * path,u64 offset,u64 len)5529 static int send_encoded_inline_extent(struct send_ctx *sctx,
5530 struct btrfs_path *path, u64 offset,
5531 u64 len)
5532 {
5533 struct btrfs_root *root = sctx->send_root;
5534 struct btrfs_fs_info *fs_info = root->fs_info;
5535 struct inode *inode;
5536 struct fs_path *fspath;
5537 struct extent_buffer *leaf = path->nodes[0];
5538 struct btrfs_key key;
5539 struct btrfs_file_extent_item *ei;
5540 u64 ram_bytes;
5541 size_t inline_size;
5542 int ret;
5543
5544 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5545 if (IS_ERR(inode))
5546 return PTR_ERR(inode);
5547
5548 fspath = fs_path_alloc();
5549 if (!fspath) {
5550 ret = -ENOMEM;
5551 goto out;
5552 }
5553
5554 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5555 if (ret < 0)
5556 goto out;
5557
5558 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5559 if (ret < 0)
5560 goto out;
5561
5562 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5563 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5564 ram_bytes = btrfs_file_extent_ram_bytes(leaf, ei);
5565 inline_size = btrfs_file_extent_inline_item_len(leaf, path->slots[0]);
5566
5567 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5568 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5569 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5570 min(key.offset + ram_bytes - offset, len));
5571 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN, ram_bytes);
5572 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET, offset - key.offset);
5573 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5574 btrfs_file_extent_compression(leaf, ei));
5575 if (ret < 0)
5576 goto out;
5577 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5578
5579 ret = put_data_header(sctx, inline_size);
5580 if (ret < 0)
5581 goto out;
5582 read_extent_buffer(leaf, sctx->send_buf + sctx->send_size,
5583 btrfs_file_extent_inline_start(ei), inline_size);
5584 sctx->send_size += inline_size;
5585
5586 ret = send_cmd(sctx);
5587
5588 tlv_put_failure:
5589 out:
5590 fs_path_free(fspath);
5591 iput(inode);
5592 return ret;
5593 }
5594
send_encoded_extent(struct send_ctx * sctx,struct btrfs_path * path,u64 offset,u64 len)5595 static int send_encoded_extent(struct send_ctx *sctx, struct btrfs_path *path,
5596 u64 offset, u64 len)
5597 {
5598 struct btrfs_root *root = sctx->send_root;
5599 struct btrfs_fs_info *fs_info = root->fs_info;
5600 struct inode *inode;
5601 struct fs_path *fspath;
5602 struct extent_buffer *leaf = path->nodes[0];
5603 struct btrfs_key key;
5604 struct btrfs_file_extent_item *ei;
5605 u64 disk_bytenr, disk_num_bytes;
5606 u32 data_offset;
5607 struct btrfs_cmd_header *hdr;
5608 u32 crc;
5609 int ret;
5610
5611 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root);
5612 if (IS_ERR(inode))
5613 return PTR_ERR(inode);
5614
5615 fspath = fs_path_alloc();
5616 if (!fspath) {
5617 ret = -ENOMEM;
5618 goto out;
5619 }
5620
5621 ret = begin_cmd(sctx, BTRFS_SEND_C_ENCODED_WRITE);
5622 if (ret < 0)
5623 goto out;
5624
5625 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5626 if (ret < 0)
5627 goto out;
5628
5629 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
5630 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
5631 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
5632 disk_num_bytes = btrfs_file_extent_disk_num_bytes(leaf, ei);
5633
5634 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, fspath);
5635 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5636 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_FILE_LEN,
5637 min(key.offset + btrfs_file_extent_num_bytes(leaf, ei) - offset,
5638 len));
5639 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_LEN,
5640 btrfs_file_extent_ram_bytes(leaf, ei));
5641 TLV_PUT_U64(sctx, BTRFS_SEND_A_UNENCODED_OFFSET,
5642 offset - key.offset + btrfs_file_extent_offset(leaf, ei));
5643 ret = btrfs_encoded_io_compression_from_extent(fs_info,
5644 btrfs_file_extent_compression(leaf, ei));
5645 if (ret < 0)
5646 goto out;
5647 TLV_PUT_U32(sctx, BTRFS_SEND_A_COMPRESSION, ret);
5648 TLV_PUT_U32(sctx, BTRFS_SEND_A_ENCRYPTION, 0);
5649
5650 ret = put_data_header(sctx, disk_num_bytes);
5651 if (ret < 0)
5652 goto out;
5653
5654 /*
5655 * We want to do I/O directly into the send buffer, so get the next page
5656 * boundary in the send buffer. This means that there may be a gap
5657 * between the beginning of the command and the file data.
5658 */
5659 data_offset = PAGE_ALIGN(sctx->send_size);
5660 if (data_offset > sctx->send_max_size ||
5661 sctx->send_max_size - data_offset < disk_num_bytes) {
5662 ret = -EOVERFLOW;
5663 goto out;
5664 }
5665
5666 /*
5667 * Note that send_buf is a mapping of send_buf_pages, so this is really
5668 * reading into send_buf.
5669 */
5670 ret = btrfs_encoded_read_regular_fill_pages(BTRFS_I(inode), offset,
5671 disk_bytenr, disk_num_bytes,
5672 sctx->send_buf_pages +
5673 (data_offset >> PAGE_SHIFT));
5674 if (ret)
5675 goto out;
5676
5677 hdr = (struct btrfs_cmd_header *)sctx->send_buf;
5678 hdr->len = cpu_to_le32(sctx->send_size + disk_num_bytes - sizeof(*hdr));
5679 hdr->crc = 0;
5680 crc = btrfs_crc32c(0, sctx->send_buf, sctx->send_size);
5681 crc = btrfs_crc32c(crc, sctx->send_buf + data_offset, disk_num_bytes);
5682 hdr->crc = cpu_to_le32(crc);
5683
5684 ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
5685 &sctx->send_off);
5686 if (!ret) {
5687 ret = write_buf(sctx->send_filp, sctx->send_buf + data_offset,
5688 disk_num_bytes, &sctx->send_off);
5689 }
5690 sctx->send_size = 0;
5691 sctx->put_data = false;
5692
5693 tlv_put_failure:
5694 out:
5695 fs_path_free(fspath);
5696 iput(inode);
5697 return ret;
5698 }
5699
send_extent_data(struct send_ctx * sctx,struct btrfs_path * path,const u64 offset,const u64 len)5700 static int send_extent_data(struct send_ctx *sctx, struct btrfs_path *path,
5701 const u64 offset, const u64 len)
5702 {
5703 const u64 end = offset + len;
5704 struct extent_buffer *leaf = path->nodes[0];
5705 struct btrfs_file_extent_item *ei;
5706 u64 read_size = max_send_read_size(sctx);
5707 u64 sent = 0;
5708
5709 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5710 return send_update_extent(sctx, offset, len);
5711
5712 ei = btrfs_item_ptr(leaf, path->slots[0],
5713 struct btrfs_file_extent_item);
5714 if ((sctx->flags & BTRFS_SEND_FLAG_COMPRESSED) &&
5715 btrfs_file_extent_compression(leaf, ei) != BTRFS_COMPRESS_NONE) {
5716 bool is_inline = (btrfs_file_extent_type(leaf, ei) ==
5717 BTRFS_FILE_EXTENT_INLINE);
5718
5719 /*
5720 * Send the compressed extent unless the compressed data is
5721 * larger than the decompressed data. This can happen if we're
5722 * not sending the entire extent, either because it has been
5723 * partially overwritten/truncated or because this is a part of
5724 * the extent that we couldn't clone in clone_range().
5725 */
5726 if (is_inline &&
5727 btrfs_file_extent_inline_item_len(leaf,
5728 path->slots[0]) <= len) {
5729 return send_encoded_inline_extent(sctx, path, offset,
5730 len);
5731 } else if (!is_inline &&
5732 btrfs_file_extent_disk_num_bytes(leaf, ei) <= len) {
5733 return send_encoded_extent(sctx, path, offset, len);
5734 }
5735 }
5736
5737 if (sctx->cur_inode == NULL) {
5738 struct btrfs_root *root = sctx->send_root;
5739
5740 sctx->cur_inode = btrfs_iget(root->fs_info->sb, sctx->cur_ino, root);
5741 if (IS_ERR(sctx->cur_inode)) {
5742 int err = PTR_ERR(sctx->cur_inode);
5743
5744 sctx->cur_inode = NULL;
5745 return err;
5746 }
5747 memset(&sctx->ra, 0, sizeof(struct file_ra_state));
5748 file_ra_state_init(&sctx->ra, sctx->cur_inode->i_mapping);
5749
5750 /*
5751 * It's very likely there are no pages from this inode in the page
5752 * cache, so after reading extents and sending their data, we clean
5753 * the page cache to avoid trashing the page cache (adding pressure
5754 * to the page cache and forcing eviction of other data more useful
5755 * for applications).
5756 *
5757 * We decide if we should clean the page cache simply by checking
5758 * if the inode's mapping nrpages is 0 when we first open it, and
5759 * not by using something like filemap_range_has_page() before
5760 * reading an extent because when we ask the readahead code to
5761 * read a given file range, it may (and almost always does) read
5762 * pages from beyond that range (see the documentation for
5763 * page_cache_sync_readahead()), so it would not be reliable,
5764 * because after reading the first extent future calls to
5765 * filemap_range_has_page() would return true because the readahead
5766 * on the previous extent resulted in reading pages of the current
5767 * extent as well.
5768 */
5769 sctx->clean_page_cache = (sctx->cur_inode->i_mapping->nrpages == 0);
5770 sctx->page_cache_clear_start = round_down(offset, PAGE_SIZE);
5771 }
5772
5773 while (sent < len) {
5774 u64 size = min(len - sent, read_size);
5775 int ret;
5776
5777 ret = send_write(sctx, offset + sent, size);
5778 if (ret < 0)
5779 return ret;
5780 sent += size;
5781 }
5782
5783 if (sctx->clean_page_cache && PAGE_ALIGNED(end)) {
5784 /*
5785 * Always operate only on ranges that are a multiple of the page
5786 * size. This is not only to prevent zeroing parts of a page in
5787 * the case of subpage sector size, but also to guarantee we evict
5788 * pages, as passing a range that is smaller than page size does
5789 * not evict the respective page (only zeroes part of its content).
5790 *
5791 * Always start from the end offset of the last range cleared.
5792 * This is because the readahead code may (and very often does)
5793 * reads pages beyond the range we request for readahead. So if
5794 * we have an extent layout like this:
5795 *
5796 * [ extent A ] [ extent B ] [ extent C ]
5797 *
5798 * When we ask page_cache_sync_readahead() to read extent A, it
5799 * may also trigger reads for pages of extent B. If we are doing
5800 * an incremental send and extent B has not changed between the
5801 * parent and send snapshots, some or all of its pages may end
5802 * up being read and placed in the page cache. So when truncating
5803 * the page cache we always start from the end offset of the
5804 * previously processed extent up to the end of the current
5805 * extent.
5806 */
5807 truncate_inode_pages_range(&sctx->cur_inode->i_data,
5808 sctx->page_cache_clear_start,
5809 end - 1);
5810 sctx->page_cache_clear_start = end;
5811 }
5812
5813 return 0;
5814 }
5815
5816 /*
5817 * Search for a capability xattr related to sctx->cur_ino. If the capability is
5818 * found, call send_set_xattr function to emit it.
5819 *
5820 * Return 0 if there isn't a capability, or when the capability was emitted
5821 * successfully, or < 0 if an error occurred.
5822 */
send_capabilities(struct send_ctx * sctx)5823 static int send_capabilities(struct send_ctx *sctx)
5824 {
5825 struct fs_path *fspath = NULL;
5826 struct btrfs_path *path;
5827 struct btrfs_dir_item *di;
5828 struct extent_buffer *leaf;
5829 unsigned long data_ptr;
5830 char *buf = NULL;
5831 int buf_len;
5832 int ret = 0;
5833
5834 path = alloc_path_for_send();
5835 if (!path)
5836 return -ENOMEM;
5837
5838 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5839 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5840 if (!di) {
5841 /* There is no xattr for this inode */
5842 goto out;
5843 } else if (IS_ERR(di)) {
5844 ret = PTR_ERR(di);
5845 goto out;
5846 }
5847
5848 leaf = path->nodes[0];
5849 buf_len = btrfs_dir_data_len(leaf, di);
5850
5851 fspath = fs_path_alloc();
5852 buf = kmalloc(buf_len, GFP_KERNEL);
5853 if (!fspath || !buf) {
5854 ret = -ENOMEM;
5855 goto out;
5856 }
5857
5858 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5859 if (ret < 0)
5860 goto out;
5861
5862 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5863 read_extent_buffer(leaf, buf, data_ptr, buf_len);
5864
5865 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5866 strlen(XATTR_NAME_CAPS), buf, buf_len);
5867 out:
5868 kfree(buf);
5869 fs_path_free(fspath);
5870 btrfs_free_path(path);
5871 return ret;
5872 }
5873
clone_range(struct send_ctx * sctx,struct btrfs_path * dst_path,struct clone_root * clone_root,const u64 disk_byte,u64 data_offset,u64 offset,u64 len)5874 static int clone_range(struct send_ctx *sctx, struct btrfs_path *dst_path,
5875 struct clone_root *clone_root, const u64 disk_byte,
5876 u64 data_offset, u64 offset, u64 len)
5877 {
5878 struct btrfs_path *path;
5879 struct btrfs_key key;
5880 int ret;
5881 struct btrfs_inode_info info;
5882 u64 clone_src_i_size = 0;
5883
5884 /*
5885 * Prevent cloning from a zero offset with a length matching the sector
5886 * size because in some scenarios this will make the receiver fail.
5887 *
5888 * For example, if in the source filesystem the extent at offset 0
5889 * has a length of sectorsize and it was written using direct IO, then
5890 * it can never be an inline extent (even if compression is enabled).
5891 * Then this extent can be cloned in the original filesystem to a non
5892 * zero file offset, but it may not be possible to clone in the
5893 * destination filesystem because it can be inlined due to compression
5894 * on the destination filesystem (as the receiver's write operations are
5895 * always done using buffered IO). The same happens when the original
5896 * filesystem does not have compression enabled but the destination
5897 * filesystem has.
5898 */
5899 if (clone_root->offset == 0 &&
5900 len == sctx->send_root->fs_info->sectorsize)
5901 return send_extent_data(sctx, dst_path, offset, len);
5902
5903 path = alloc_path_for_send();
5904 if (!path)
5905 return -ENOMEM;
5906
5907 /*
5908 * There are inodes that have extents that lie behind its i_size. Don't
5909 * accept clones from these extents.
5910 */
5911 ret = get_inode_info(clone_root->root, clone_root->ino, &info);
5912 btrfs_release_path(path);
5913 if (ret < 0)
5914 goto out;
5915 clone_src_i_size = info.size;
5916
5917 /*
5918 * We can't send a clone operation for the entire range if we find
5919 * extent items in the respective range in the source file that
5920 * refer to different extents or if we find holes.
5921 * So check for that and do a mix of clone and regular write/copy
5922 * operations if needed.
5923 *
5924 * Example:
5925 *
5926 * mkfs.btrfs -f /dev/sda
5927 * mount /dev/sda /mnt
5928 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5929 * cp --reflink=always /mnt/foo /mnt/bar
5930 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5931 * btrfs subvolume snapshot -r /mnt /mnt/snap
5932 *
5933 * If when we send the snapshot and we are processing file bar (which
5934 * has a higher inode number than foo) we blindly send a clone operation
5935 * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5936 * a file bar that matches the content of file foo - iow, doesn't match
5937 * the content from bar in the original filesystem.
5938 */
5939 key.objectid = clone_root->ino;
5940 key.type = BTRFS_EXTENT_DATA_KEY;
5941 key.offset = clone_root->offset;
5942 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5943 if (ret < 0)
5944 goto out;
5945 if (ret > 0 && path->slots[0] > 0) {
5946 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5947 if (key.objectid == clone_root->ino &&
5948 key.type == BTRFS_EXTENT_DATA_KEY)
5949 path->slots[0]--;
5950 }
5951
5952 while (true) {
5953 struct extent_buffer *leaf = path->nodes[0];
5954 int slot = path->slots[0];
5955 struct btrfs_file_extent_item *ei;
5956 u8 type;
5957 u64 ext_len;
5958 u64 clone_len;
5959 u64 clone_data_offset;
5960 bool crossed_src_i_size = false;
5961
5962 if (slot >= btrfs_header_nritems(leaf)) {
5963 ret = btrfs_next_leaf(clone_root->root, path);
5964 if (ret < 0)
5965 goto out;
5966 else if (ret > 0)
5967 break;
5968 continue;
5969 }
5970
5971 btrfs_item_key_to_cpu(leaf, &key, slot);
5972
5973 /*
5974 * We might have an implicit trailing hole (NO_HOLES feature
5975 * enabled). We deal with it after leaving this loop.
5976 */
5977 if (key.objectid != clone_root->ino ||
5978 key.type != BTRFS_EXTENT_DATA_KEY)
5979 break;
5980
5981 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5982 type = btrfs_file_extent_type(leaf, ei);
5983 if (type == BTRFS_FILE_EXTENT_INLINE) {
5984 ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5985 ext_len = PAGE_ALIGN(ext_len);
5986 } else {
5987 ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5988 }
5989
5990 if (key.offset + ext_len <= clone_root->offset)
5991 goto next;
5992
5993 if (key.offset > clone_root->offset) {
5994 /* Implicit hole, NO_HOLES feature enabled. */
5995 u64 hole_len = key.offset - clone_root->offset;
5996
5997 if (hole_len > len)
5998 hole_len = len;
5999 ret = send_extent_data(sctx, dst_path, offset,
6000 hole_len);
6001 if (ret < 0)
6002 goto out;
6003
6004 len -= hole_len;
6005 if (len == 0)
6006 break;
6007 offset += hole_len;
6008 clone_root->offset += hole_len;
6009 data_offset += hole_len;
6010 }
6011
6012 if (key.offset >= clone_root->offset + len)
6013 break;
6014
6015 if (key.offset >= clone_src_i_size)
6016 break;
6017
6018 if (key.offset + ext_len > clone_src_i_size) {
6019 ext_len = clone_src_i_size - key.offset;
6020 crossed_src_i_size = true;
6021 }
6022
6023 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
6024 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
6025 clone_root->offset = key.offset;
6026 if (clone_data_offset < data_offset &&
6027 clone_data_offset + ext_len > data_offset) {
6028 u64 extent_offset;
6029
6030 extent_offset = data_offset - clone_data_offset;
6031 ext_len -= extent_offset;
6032 clone_data_offset += extent_offset;
6033 clone_root->offset += extent_offset;
6034 }
6035 }
6036
6037 clone_len = min_t(u64, ext_len, len);
6038
6039 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
6040 clone_data_offset == data_offset) {
6041 const u64 src_end = clone_root->offset + clone_len;
6042 const u64 sectorsize = SZ_64K;
6043
6044 /*
6045 * We can't clone the last block, when its size is not
6046 * sector size aligned, into the middle of a file. If we
6047 * do so, the receiver will get a failure (-EINVAL) when
6048 * trying to clone or will silently corrupt the data in
6049 * the destination file if it's on a kernel without the
6050 * fix introduced by commit ac765f83f1397646
6051 * ("Btrfs: fix data corruption due to cloning of eof
6052 * block).
6053 *
6054 * So issue a clone of the aligned down range plus a
6055 * regular write for the eof block, if we hit that case.
6056 *
6057 * Also, we use the maximum possible sector size, 64K,
6058 * because we don't know what's the sector size of the
6059 * filesystem that receives the stream, so we have to
6060 * assume the largest possible sector size.
6061 */
6062 if (src_end == clone_src_i_size &&
6063 !IS_ALIGNED(src_end, sectorsize) &&
6064 offset + clone_len < sctx->cur_inode_size) {
6065 u64 slen;
6066
6067 slen = ALIGN_DOWN(src_end - clone_root->offset,
6068 sectorsize);
6069 if (slen > 0) {
6070 ret = send_clone(sctx, offset, slen,
6071 clone_root);
6072 if (ret < 0)
6073 goto out;
6074 }
6075 ret = send_extent_data(sctx, dst_path,
6076 offset + slen,
6077 clone_len - slen);
6078 } else {
6079 ret = send_clone(sctx, offset, clone_len,
6080 clone_root);
6081 }
6082 } else if (crossed_src_i_size && clone_len < len) {
6083 /*
6084 * If we are at i_size of the clone source inode and we
6085 * can not clone from it, terminate the loop. This is
6086 * to avoid sending two write operations, one with a
6087 * length matching clone_len and the final one after
6088 * this loop with a length of len - clone_len.
6089 *
6090 * When using encoded writes (BTRFS_SEND_FLAG_COMPRESSED
6091 * was passed to the send ioctl), this helps avoid
6092 * sending an encoded write for an offset that is not
6093 * sector size aligned, in case the i_size of the source
6094 * inode is not sector size aligned. That will make the
6095 * receiver fallback to decompression of the data and
6096 * writing it using regular buffered IO, therefore while
6097 * not incorrect, it's not optimal due decompression and
6098 * possible re-compression at the receiver.
6099 */
6100 break;
6101 } else {
6102 ret = send_extent_data(sctx, dst_path, offset,
6103 clone_len);
6104 }
6105
6106 if (ret < 0)
6107 goto out;
6108
6109 len -= clone_len;
6110 if (len == 0)
6111 break;
6112 offset += clone_len;
6113 clone_root->offset += clone_len;
6114
6115 /*
6116 * If we are cloning from the file we are currently processing,
6117 * and using the send root as the clone root, we must stop once
6118 * the current clone offset reaches the current eof of the file
6119 * at the receiver, otherwise we would issue an invalid clone
6120 * operation (source range going beyond eof) and cause the
6121 * receiver to fail. So if we reach the current eof, bail out
6122 * and fallback to a regular write.
6123 */
6124 if (clone_root->root == sctx->send_root &&
6125 clone_root->ino == sctx->cur_ino &&
6126 clone_root->offset >= sctx->cur_inode_next_write_offset)
6127 break;
6128
6129 data_offset += clone_len;
6130 next:
6131 path->slots[0]++;
6132 }
6133
6134 if (len > 0)
6135 ret = send_extent_data(sctx, dst_path, offset, len);
6136 else
6137 ret = 0;
6138 out:
6139 btrfs_free_path(path);
6140 return ret;
6141 }
6142
send_write_or_clone(struct send_ctx * sctx,struct btrfs_path * path,struct btrfs_key * key,struct clone_root * clone_root)6143 static int send_write_or_clone(struct send_ctx *sctx,
6144 struct btrfs_path *path,
6145 struct btrfs_key *key,
6146 struct clone_root *clone_root)
6147 {
6148 int ret = 0;
6149 u64 offset = key->offset;
6150 u64 end;
6151 u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
6152
6153 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size);
6154 if (offset >= end)
6155 return 0;
6156
6157 if (clone_root && IS_ALIGNED(end, bs)) {
6158 struct btrfs_file_extent_item *ei;
6159 u64 disk_byte;
6160 u64 data_offset;
6161
6162 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6163 struct btrfs_file_extent_item);
6164 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
6165 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
6166 ret = clone_range(sctx, path, clone_root, disk_byte,
6167 data_offset, offset, end - offset);
6168 } else {
6169 ret = send_extent_data(sctx, path, offset, end - offset);
6170 }
6171 sctx->cur_inode_next_write_offset = end;
6172 return ret;
6173 }
6174
is_extent_unchanged(struct send_ctx * sctx,struct btrfs_path * left_path,struct btrfs_key * ekey)6175 static int is_extent_unchanged(struct send_ctx *sctx,
6176 struct btrfs_path *left_path,
6177 struct btrfs_key *ekey)
6178 {
6179 int ret = 0;
6180 struct btrfs_key key;
6181 struct btrfs_path *path = NULL;
6182 struct extent_buffer *eb;
6183 int slot;
6184 struct btrfs_key found_key;
6185 struct btrfs_file_extent_item *ei;
6186 u64 left_disknr;
6187 u64 right_disknr;
6188 u64 left_offset;
6189 u64 right_offset;
6190 u64 left_offset_fixed;
6191 u64 left_len;
6192 u64 right_len;
6193 u64 left_gen;
6194 u64 right_gen;
6195 u8 left_type;
6196 u8 right_type;
6197
6198 path = alloc_path_for_send();
6199 if (!path)
6200 return -ENOMEM;
6201
6202 eb = left_path->nodes[0];
6203 slot = left_path->slots[0];
6204 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6205 left_type = btrfs_file_extent_type(eb, ei);
6206
6207 if (left_type != BTRFS_FILE_EXTENT_REG) {
6208 ret = 0;
6209 goto out;
6210 }
6211 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6212 left_len = btrfs_file_extent_num_bytes(eb, ei);
6213 left_offset = btrfs_file_extent_offset(eb, ei);
6214 left_gen = btrfs_file_extent_generation(eb, ei);
6215
6216 /*
6217 * Following comments will refer to these graphics. L is the left
6218 * extents which we are checking at the moment. 1-8 are the right
6219 * extents that we iterate.
6220 *
6221 * |-----L-----|
6222 * |-1-|-2a-|-3-|-4-|-5-|-6-|
6223 *
6224 * |-----L-----|
6225 * |--1--|-2b-|...(same as above)
6226 *
6227 * Alternative situation. Happens on files where extents got split.
6228 * |-----L-----|
6229 * |-----------7-----------|-6-|
6230 *
6231 * Alternative situation. Happens on files which got larger.
6232 * |-----L-----|
6233 * |-8-|
6234 * Nothing follows after 8.
6235 */
6236
6237 key.objectid = ekey->objectid;
6238 key.type = BTRFS_EXTENT_DATA_KEY;
6239 key.offset = ekey->offset;
6240 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
6241 if (ret < 0)
6242 goto out;
6243 if (ret) {
6244 ret = 0;
6245 goto out;
6246 }
6247
6248 /*
6249 * Handle special case where the right side has no extents at all.
6250 */
6251 eb = path->nodes[0];
6252 slot = path->slots[0];
6253 btrfs_item_key_to_cpu(eb, &found_key, slot);
6254 if (found_key.objectid != key.objectid ||
6255 found_key.type != key.type) {
6256 /* If we're a hole then just pretend nothing changed */
6257 ret = (left_disknr) ? 0 : 1;
6258 goto out;
6259 }
6260
6261 /*
6262 * We're now on 2a, 2b or 7.
6263 */
6264 key = found_key;
6265 while (key.offset < ekey->offset + left_len) {
6266 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
6267 right_type = btrfs_file_extent_type(eb, ei);
6268 if (right_type != BTRFS_FILE_EXTENT_REG &&
6269 right_type != BTRFS_FILE_EXTENT_INLINE) {
6270 ret = 0;
6271 goto out;
6272 }
6273
6274 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6275 right_len = btrfs_file_extent_ram_bytes(eb, ei);
6276 right_len = PAGE_ALIGN(right_len);
6277 } else {
6278 right_len = btrfs_file_extent_num_bytes(eb, ei);
6279 }
6280
6281 /*
6282 * Are we at extent 8? If yes, we know the extent is changed.
6283 * This may only happen on the first iteration.
6284 */
6285 if (found_key.offset + right_len <= ekey->offset) {
6286 /* If we're a hole just pretend nothing changed */
6287 ret = (left_disknr) ? 0 : 1;
6288 goto out;
6289 }
6290
6291 /*
6292 * We just wanted to see if when we have an inline extent, what
6293 * follows it is a regular extent (wanted to check the above
6294 * condition for inline extents too). This should normally not
6295 * happen but it's possible for example when we have an inline
6296 * compressed extent representing data with a size matching
6297 * the page size (currently the same as sector size).
6298 */
6299 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
6300 ret = 0;
6301 goto out;
6302 }
6303
6304 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
6305 right_offset = btrfs_file_extent_offset(eb, ei);
6306 right_gen = btrfs_file_extent_generation(eb, ei);
6307
6308 left_offset_fixed = left_offset;
6309 if (key.offset < ekey->offset) {
6310 /* Fix the right offset for 2a and 7. */
6311 right_offset += ekey->offset - key.offset;
6312 } else {
6313 /* Fix the left offset for all behind 2a and 2b */
6314 left_offset_fixed += key.offset - ekey->offset;
6315 }
6316
6317 /*
6318 * Check if we have the same extent.
6319 */
6320 if (left_disknr != right_disknr ||
6321 left_offset_fixed != right_offset ||
6322 left_gen != right_gen) {
6323 ret = 0;
6324 goto out;
6325 }
6326
6327 /*
6328 * Go to the next extent.
6329 */
6330 ret = btrfs_next_item(sctx->parent_root, path);
6331 if (ret < 0)
6332 goto out;
6333 if (!ret) {
6334 eb = path->nodes[0];
6335 slot = path->slots[0];
6336 btrfs_item_key_to_cpu(eb, &found_key, slot);
6337 }
6338 if (ret || found_key.objectid != key.objectid ||
6339 found_key.type != key.type) {
6340 key.offset += right_len;
6341 break;
6342 }
6343 if (found_key.offset != key.offset + right_len) {
6344 ret = 0;
6345 goto out;
6346 }
6347 key = found_key;
6348 }
6349
6350 /*
6351 * We're now behind the left extent (treat as unchanged) or at the end
6352 * of the right side (treat as changed).
6353 */
6354 if (key.offset >= ekey->offset + left_len)
6355 ret = 1;
6356 else
6357 ret = 0;
6358
6359
6360 out:
6361 btrfs_free_path(path);
6362 return ret;
6363 }
6364
get_last_extent(struct send_ctx * sctx,u64 offset)6365 static int get_last_extent(struct send_ctx *sctx, u64 offset)
6366 {
6367 struct btrfs_path *path;
6368 struct btrfs_root *root = sctx->send_root;
6369 struct btrfs_key key;
6370 int ret;
6371
6372 path = alloc_path_for_send();
6373 if (!path)
6374 return -ENOMEM;
6375
6376 sctx->cur_inode_last_extent = 0;
6377
6378 key.objectid = sctx->cur_ino;
6379 key.type = BTRFS_EXTENT_DATA_KEY;
6380 key.offset = offset;
6381 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
6382 if (ret < 0)
6383 goto out;
6384 ret = 0;
6385 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
6386 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
6387 goto out;
6388
6389 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6390 out:
6391 btrfs_free_path(path);
6392 return ret;
6393 }
6394
range_is_hole_in_parent(struct send_ctx * sctx,const u64 start,const u64 end)6395 static int range_is_hole_in_parent(struct send_ctx *sctx,
6396 const u64 start,
6397 const u64 end)
6398 {
6399 struct btrfs_path *path;
6400 struct btrfs_key key;
6401 struct btrfs_root *root = sctx->parent_root;
6402 u64 search_start = start;
6403 int ret;
6404
6405 path = alloc_path_for_send();
6406 if (!path)
6407 return -ENOMEM;
6408
6409 key.objectid = sctx->cur_ino;
6410 key.type = BTRFS_EXTENT_DATA_KEY;
6411 key.offset = search_start;
6412 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6413 if (ret < 0)
6414 goto out;
6415 if (ret > 0 && path->slots[0] > 0)
6416 path->slots[0]--;
6417
6418 while (search_start < end) {
6419 struct extent_buffer *leaf = path->nodes[0];
6420 int slot = path->slots[0];
6421 struct btrfs_file_extent_item *fi;
6422 u64 extent_end;
6423
6424 if (slot >= btrfs_header_nritems(leaf)) {
6425 ret = btrfs_next_leaf(root, path);
6426 if (ret < 0)
6427 goto out;
6428 else if (ret > 0)
6429 break;
6430 continue;
6431 }
6432
6433 btrfs_item_key_to_cpu(leaf, &key, slot);
6434 if (key.objectid < sctx->cur_ino ||
6435 key.type < BTRFS_EXTENT_DATA_KEY)
6436 goto next;
6437 if (key.objectid > sctx->cur_ino ||
6438 key.type > BTRFS_EXTENT_DATA_KEY ||
6439 key.offset >= end)
6440 break;
6441
6442 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6443 extent_end = btrfs_file_extent_end(path);
6444 if (extent_end <= start)
6445 goto next;
6446 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
6447 search_start = extent_end;
6448 goto next;
6449 }
6450 ret = 0;
6451 goto out;
6452 next:
6453 path->slots[0]++;
6454 }
6455 ret = 1;
6456 out:
6457 btrfs_free_path(path);
6458 return ret;
6459 }
6460
maybe_send_hole(struct send_ctx * sctx,struct btrfs_path * path,struct btrfs_key * key)6461 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
6462 struct btrfs_key *key)
6463 {
6464 int ret = 0;
6465
6466 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
6467 return 0;
6468
6469 if (sctx->cur_inode_last_extent == (u64)-1) {
6470 ret = get_last_extent(sctx, key->offset - 1);
6471 if (ret)
6472 return ret;
6473 }
6474
6475 if (path->slots[0] == 0 &&
6476 sctx->cur_inode_last_extent < key->offset) {
6477 /*
6478 * We might have skipped entire leafs that contained only
6479 * file extent items for our current inode. These leafs have
6480 * a generation number smaller (older) than the one in the
6481 * current leaf and the leaf our last extent came from, and
6482 * are located between these 2 leafs.
6483 */
6484 ret = get_last_extent(sctx, key->offset - 1);
6485 if (ret)
6486 return ret;
6487 }
6488
6489 if (sctx->cur_inode_last_extent < key->offset) {
6490 ret = range_is_hole_in_parent(sctx,
6491 sctx->cur_inode_last_extent,
6492 key->offset);
6493 if (ret < 0)
6494 return ret;
6495 else if (ret == 0)
6496 ret = send_hole(sctx, key->offset);
6497 else
6498 ret = 0;
6499 }
6500 sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
6501 return ret;
6502 }
6503
process_extent(struct send_ctx * sctx,struct btrfs_path * path,struct btrfs_key * key)6504 static int process_extent(struct send_ctx *sctx,
6505 struct btrfs_path *path,
6506 struct btrfs_key *key)
6507 {
6508 struct clone_root *found_clone = NULL;
6509 int ret = 0;
6510
6511 if (S_ISLNK(sctx->cur_inode_mode))
6512 return 0;
6513
6514 if (sctx->parent_root && !sctx->cur_inode_new) {
6515 ret = is_extent_unchanged(sctx, path, key);
6516 if (ret < 0)
6517 goto out;
6518 if (ret) {
6519 ret = 0;
6520 goto out_hole;
6521 }
6522 } else {
6523 struct btrfs_file_extent_item *ei;
6524 u8 type;
6525
6526 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
6527 struct btrfs_file_extent_item);
6528 type = btrfs_file_extent_type(path->nodes[0], ei);
6529 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
6530 type == BTRFS_FILE_EXTENT_REG) {
6531 /*
6532 * The send spec does not have a prealloc command yet,
6533 * so just leave a hole for prealloc'ed extents until
6534 * we have enough commands queued up to justify rev'ing
6535 * the send spec.
6536 */
6537 if (type == BTRFS_FILE_EXTENT_PREALLOC) {
6538 ret = 0;
6539 goto out;
6540 }
6541
6542 /* Have a hole, just skip it. */
6543 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
6544 ret = 0;
6545 goto out;
6546 }
6547 }
6548 }
6549
6550 ret = find_extent_clone(sctx, path, key->objectid, key->offset,
6551 sctx->cur_inode_size, &found_clone);
6552 if (ret != -ENOENT && ret < 0)
6553 goto out;
6554
6555 ret = send_write_or_clone(sctx, path, key, found_clone);
6556 if (ret)
6557 goto out;
6558 out_hole:
6559 ret = maybe_send_hole(sctx, path, key);
6560 out:
6561 return ret;
6562 }
6563
process_all_extents(struct send_ctx * sctx)6564 static int process_all_extents(struct send_ctx *sctx)
6565 {
6566 int ret = 0;
6567 int iter_ret = 0;
6568 struct btrfs_root *root;
6569 struct btrfs_path *path;
6570 struct btrfs_key key;
6571 struct btrfs_key found_key;
6572
6573 root = sctx->send_root;
6574 path = alloc_path_for_send();
6575 if (!path)
6576 return -ENOMEM;
6577
6578 key.objectid = sctx->cmp_key->objectid;
6579 key.type = BTRFS_EXTENT_DATA_KEY;
6580 key.offset = 0;
6581 btrfs_for_each_slot(root, &key, &found_key, path, iter_ret) {
6582 if (found_key.objectid != key.objectid ||
6583 found_key.type != key.type) {
6584 ret = 0;
6585 break;
6586 }
6587
6588 ret = process_extent(sctx, path, &found_key);
6589 if (ret < 0)
6590 break;
6591 }
6592 /* Catch error found during iteration */
6593 if (iter_ret < 0)
6594 ret = iter_ret;
6595
6596 btrfs_free_path(path);
6597 return ret;
6598 }
6599
process_recorded_refs_if_needed(struct send_ctx * sctx,int at_end,int * pending_move,int * refs_processed)6600 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
6601 int *pending_move,
6602 int *refs_processed)
6603 {
6604 int ret = 0;
6605
6606 if (sctx->cur_ino == 0)
6607 goto out;
6608 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
6609 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
6610 goto out;
6611 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
6612 goto out;
6613
6614 ret = process_recorded_refs(sctx, pending_move);
6615 if (ret < 0)
6616 goto out;
6617
6618 *refs_processed = 1;
6619 out:
6620 return ret;
6621 }
6622
finish_inode_if_needed(struct send_ctx * sctx,int at_end)6623 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
6624 {
6625 int ret = 0;
6626 struct btrfs_inode_info info;
6627 u64 left_mode;
6628 u64 left_uid;
6629 u64 left_gid;
6630 u64 left_fileattr;
6631 u64 right_mode;
6632 u64 right_uid;
6633 u64 right_gid;
6634 u64 right_fileattr;
6635 int need_chmod = 0;
6636 int need_chown = 0;
6637 bool need_fileattr = false;
6638 int need_truncate = 1;
6639 int pending_move = 0;
6640 int refs_processed = 0;
6641
6642 if (sctx->ignore_cur_inode)
6643 return 0;
6644
6645 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
6646 &refs_processed);
6647 if (ret < 0)
6648 goto out;
6649
6650 /*
6651 * We have processed the refs and thus need to advance send_progress.
6652 * Now, calls to get_cur_xxx will take the updated refs of the current
6653 * inode into account.
6654 *
6655 * On the other hand, if our current inode is a directory and couldn't
6656 * be moved/renamed because its parent was renamed/moved too and it has
6657 * a higher inode number, we can only move/rename our current inode
6658 * after we moved/renamed its parent. Therefore in this case operate on
6659 * the old path (pre move/rename) of our current inode, and the
6660 * move/rename will be performed later.
6661 */
6662 if (refs_processed && !pending_move)
6663 sctx->send_progress = sctx->cur_ino + 1;
6664
6665 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
6666 goto out;
6667 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
6668 goto out;
6669 ret = get_inode_info(sctx->send_root, sctx->cur_ino, &info);
6670 if (ret < 0)
6671 goto out;
6672 left_mode = info.mode;
6673 left_uid = info.uid;
6674 left_gid = info.gid;
6675 left_fileattr = info.fileattr;
6676
6677 if (!sctx->parent_root || sctx->cur_inode_new) {
6678 need_chown = 1;
6679 if (!S_ISLNK(sctx->cur_inode_mode))
6680 need_chmod = 1;
6681 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
6682 need_truncate = 0;
6683 } else {
6684 u64 old_size;
6685
6686 ret = get_inode_info(sctx->parent_root, sctx->cur_ino, &info);
6687 if (ret < 0)
6688 goto out;
6689 old_size = info.size;
6690 right_mode = info.mode;
6691 right_uid = info.uid;
6692 right_gid = info.gid;
6693 right_fileattr = info.fileattr;
6694
6695 if (left_uid != right_uid || left_gid != right_gid)
6696 need_chown = 1;
6697 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
6698 need_chmod = 1;
6699 if (!S_ISLNK(sctx->cur_inode_mode) && left_fileattr != right_fileattr)
6700 need_fileattr = true;
6701 if ((old_size == sctx->cur_inode_size) ||
6702 (sctx->cur_inode_size > old_size &&
6703 sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
6704 need_truncate = 0;
6705 }
6706
6707 if (S_ISREG(sctx->cur_inode_mode)) {
6708 if (need_send_hole(sctx)) {
6709 if (sctx->cur_inode_last_extent == (u64)-1 ||
6710 sctx->cur_inode_last_extent <
6711 sctx->cur_inode_size) {
6712 ret = get_last_extent(sctx, (u64)-1);
6713 if (ret)
6714 goto out;
6715 }
6716 if (sctx->cur_inode_last_extent < sctx->cur_inode_size) {
6717 ret = range_is_hole_in_parent(sctx,
6718 sctx->cur_inode_last_extent,
6719 sctx->cur_inode_size);
6720 if (ret < 0) {
6721 goto out;
6722 } else if (ret == 0) {
6723 ret = send_hole(sctx, sctx->cur_inode_size);
6724 if (ret < 0)
6725 goto out;
6726 } else {
6727 /* Range is already a hole, skip. */
6728 ret = 0;
6729 }
6730 }
6731 }
6732 if (need_truncate) {
6733 ret = send_truncate(sctx, sctx->cur_ino,
6734 sctx->cur_inode_gen,
6735 sctx->cur_inode_size);
6736 if (ret < 0)
6737 goto out;
6738 }
6739 }
6740
6741 if (need_chown) {
6742 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6743 left_uid, left_gid);
6744 if (ret < 0)
6745 goto out;
6746 }
6747 if (need_chmod) {
6748 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6749 left_mode);
6750 if (ret < 0)
6751 goto out;
6752 }
6753 if (need_fileattr) {
6754 ret = send_fileattr(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6755 left_fileattr);
6756 if (ret < 0)
6757 goto out;
6758 }
6759
6760 if (proto_cmd_ok(sctx, BTRFS_SEND_C_ENABLE_VERITY)
6761 && sctx->cur_inode_needs_verity) {
6762 ret = process_verity(sctx);
6763 if (ret < 0)
6764 goto out;
6765 }
6766
6767 ret = send_capabilities(sctx);
6768 if (ret < 0)
6769 goto out;
6770
6771 /*
6772 * If other directory inodes depended on our current directory
6773 * inode's move/rename, now do their move/rename operations.
6774 */
6775 if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
6776 ret = apply_children_dir_moves(sctx);
6777 if (ret)
6778 goto out;
6779 /*
6780 * Need to send that every time, no matter if it actually
6781 * changed between the two trees as we have done changes to
6782 * the inode before. If our inode is a directory and it's
6783 * waiting to be moved/renamed, we will send its utimes when
6784 * it's moved/renamed, therefore we don't need to do it here.
6785 */
6786 sctx->send_progress = sctx->cur_ino + 1;
6787
6788 /*
6789 * If the current inode is a non-empty directory, delay issuing
6790 * the utimes command for it, as it's very likely we have inodes
6791 * with an higher number inside it. We want to issue the utimes
6792 * command only after adding all dentries to it.
6793 */
6794 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_size > 0)
6795 ret = cache_dir_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6796 else
6797 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
6798
6799 if (ret < 0)
6800 goto out;
6801 }
6802
6803 out:
6804 if (!ret)
6805 ret = trim_dir_utimes_cache(sctx);
6806
6807 return ret;
6808 }
6809
close_current_inode(struct send_ctx * sctx)6810 static void close_current_inode(struct send_ctx *sctx)
6811 {
6812 u64 i_size;
6813
6814 if (sctx->cur_inode == NULL)
6815 return;
6816
6817 i_size = i_size_read(sctx->cur_inode);
6818
6819 /*
6820 * If we are doing an incremental send, we may have extents between the
6821 * last processed extent and the i_size that have not been processed
6822 * because they haven't changed but we may have read some of their pages
6823 * through readahead, see the comments at send_extent_data().
6824 */
6825 if (sctx->clean_page_cache && sctx->page_cache_clear_start < i_size)
6826 truncate_inode_pages_range(&sctx->cur_inode->i_data,
6827 sctx->page_cache_clear_start,
6828 round_up(i_size, PAGE_SIZE) - 1);
6829
6830 iput(sctx->cur_inode);
6831 sctx->cur_inode = NULL;
6832 }
6833
changed_inode(struct send_ctx * sctx,enum btrfs_compare_tree_result result)6834 static int changed_inode(struct send_ctx *sctx,
6835 enum btrfs_compare_tree_result result)
6836 {
6837 int ret = 0;
6838 struct btrfs_key *key = sctx->cmp_key;
6839 struct btrfs_inode_item *left_ii = NULL;
6840 struct btrfs_inode_item *right_ii = NULL;
6841 u64 left_gen = 0;
6842 u64 right_gen = 0;
6843
6844 close_current_inode(sctx);
6845
6846 sctx->cur_ino = key->objectid;
6847 sctx->cur_inode_new_gen = false;
6848 sctx->cur_inode_last_extent = (u64)-1;
6849 sctx->cur_inode_next_write_offset = 0;
6850 sctx->ignore_cur_inode = false;
6851
6852 /*
6853 * Set send_progress to current inode. This will tell all get_cur_xxx
6854 * functions that the current inode's refs are not updated yet. Later,
6855 * when process_recorded_refs is finished, it is set to cur_ino + 1.
6856 */
6857 sctx->send_progress = sctx->cur_ino;
6858
6859 if (result == BTRFS_COMPARE_TREE_NEW ||
6860 result == BTRFS_COMPARE_TREE_CHANGED) {
6861 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
6862 sctx->left_path->slots[0],
6863 struct btrfs_inode_item);
6864 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
6865 left_ii);
6866 } else {
6867 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6868 sctx->right_path->slots[0],
6869 struct btrfs_inode_item);
6870 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6871 right_ii);
6872 }
6873 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6874 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
6875 sctx->right_path->slots[0],
6876 struct btrfs_inode_item);
6877
6878 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
6879 right_ii);
6880
6881 /*
6882 * The cur_ino = root dir case is special here. We can't treat
6883 * the inode as deleted+reused because it would generate a
6884 * stream that tries to delete/mkdir the root dir.
6885 */
6886 if (left_gen != right_gen &&
6887 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6888 sctx->cur_inode_new_gen = true;
6889 }
6890
6891 /*
6892 * Normally we do not find inodes with a link count of zero (orphans)
6893 * because the most common case is to create a snapshot and use it
6894 * for a send operation. However other less common use cases involve
6895 * using a subvolume and send it after turning it to RO mode just
6896 * after deleting all hard links of a file while holding an open
6897 * file descriptor against it or turning a RO snapshot into RW mode,
6898 * keep an open file descriptor against a file, delete it and then
6899 * turn the snapshot back to RO mode before using it for a send
6900 * operation. The former is what the receiver operation does.
6901 * Therefore, if we want to send these snapshots soon after they're
6902 * received, we need to handle orphan inodes as well. Moreover, orphans
6903 * can appear not only in the send snapshot but also in the parent
6904 * snapshot. Here are several cases:
6905 *
6906 * Case 1: BTRFS_COMPARE_TREE_NEW
6907 * | send snapshot | action
6908 * --------------------------------
6909 * nlink | 0 | ignore
6910 *
6911 * Case 2: BTRFS_COMPARE_TREE_DELETED
6912 * | parent snapshot | action
6913 * ----------------------------------
6914 * nlink | 0 | as usual
6915 * Note: No unlinks will be sent because there're no paths for it.
6916 *
6917 * Case 3: BTRFS_COMPARE_TREE_CHANGED
6918 * | | parent snapshot | send snapshot | action
6919 * -----------------------------------------------------------------------
6920 * subcase 1 | nlink | 0 | 0 | ignore
6921 * subcase 2 | nlink | >0 | 0 | new_gen(deletion)
6922 * subcase 3 | nlink | 0 | >0 | new_gen(creation)
6923 *
6924 */
6925 if (result == BTRFS_COMPARE_TREE_NEW) {
6926 if (btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii) == 0) {
6927 sctx->ignore_cur_inode = true;
6928 goto out;
6929 }
6930 sctx->cur_inode_gen = left_gen;
6931 sctx->cur_inode_new = true;
6932 sctx->cur_inode_deleted = false;
6933 sctx->cur_inode_size = btrfs_inode_size(
6934 sctx->left_path->nodes[0], left_ii);
6935 sctx->cur_inode_mode = btrfs_inode_mode(
6936 sctx->left_path->nodes[0], left_ii);
6937 sctx->cur_inode_rdev = btrfs_inode_rdev(
6938 sctx->left_path->nodes[0], left_ii);
6939 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
6940 ret = send_create_inode_if_needed(sctx);
6941 } else if (result == BTRFS_COMPARE_TREE_DELETED) {
6942 sctx->cur_inode_gen = right_gen;
6943 sctx->cur_inode_new = false;
6944 sctx->cur_inode_deleted = true;
6945 sctx->cur_inode_size = btrfs_inode_size(
6946 sctx->right_path->nodes[0], right_ii);
6947 sctx->cur_inode_mode = btrfs_inode_mode(
6948 sctx->right_path->nodes[0], right_ii);
6949 } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
6950 u32 new_nlinks, old_nlinks;
6951
6952 new_nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii);
6953 old_nlinks = btrfs_inode_nlink(sctx->right_path->nodes[0], right_ii);
6954 if (new_nlinks == 0 && old_nlinks == 0) {
6955 sctx->ignore_cur_inode = true;
6956 goto out;
6957 } else if (new_nlinks == 0 || old_nlinks == 0) {
6958 sctx->cur_inode_new_gen = 1;
6959 }
6960 /*
6961 * We need to do some special handling in case the inode was
6962 * reported as changed with a changed generation number. This
6963 * means that the original inode was deleted and new inode
6964 * reused the same inum. So we have to treat the old inode as
6965 * deleted and the new one as new.
6966 */
6967 if (sctx->cur_inode_new_gen) {
6968 /*
6969 * First, process the inode as if it was deleted.
6970 */
6971 if (old_nlinks > 0) {
6972 sctx->cur_inode_gen = right_gen;
6973 sctx->cur_inode_new = false;
6974 sctx->cur_inode_deleted = true;
6975 sctx->cur_inode_size = btrfs_inode_size(
6976 sctx->right_path->nodes[0], right_ii);
6977 sctx->cur_inode_mode = btrfs_inode_mode(
6978 sctx->right_path->nodes[0], right_ii);
6979 ret = process_all_refs(sctx,
6980 BTRFS_COMPARE_TREE_DELETED);
6981 if (ret < 0)
6982 goto out;
6983 }
6984
6985 /*
6986 * Now process the inode as if it was new.
6987 */
6988 if (new_nlinks > 0) {
6989 sctx->cur_inode_gen = left_gen;
6990 sctx->cur_inode_new = true;
6991 sctx->cur_inode_deleted = false;
6992 sctx->cur_inode_size = btrfs_inode_size(
6993 sctx->left_path->nodes[0],
6994 left_ii);
6995 sctx->cur_inode_mode = btrfs_inode_mode(
6996 sctx->left_path->nodes[0],
6997 left_ii);
6998 sctx->cur_inode_rdev = btrfs_inode_rdev(
6999 sctx->left_path->nodes[0],
7000 left_ii);
7001 ret = send_create_inode_if_needed(sctx);
7002 if (ret < 0)
7003 goto out;
7004
7005 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
7006 if (ret < 0)
7007 goto out;
7008 /*
7009 * Advance send_progress now as we did not get
7010 * into process_recorded_refs_if_needed in the
7011 * new_gen case.
7012 */
7013 sctx->send_progress = sctx->cur_ino + 1;
7014
7015 /*
7016 * Now process all extents and xattrs of the
7017 * inode as if they were all new.
7018 */
7019 ret = process_all_extents(sctx);
7020 if (ret < 0)
7021 goto out;
7022 ret = process_all_new_xattrs(sctx);
7023 if (ret < 0)
7024 goto out;
7025 }
7026 } else {
7027 sctx->cur_inode_gen = left_gen;
7028 sctx->cur_inode_new = false;
7029 sctx->cur_inode_new_gen = false;
7030 sctx->cur_inode_deleted = false;
7031 sctx->cur_inode_size = btrfs_inode_size(
7032 sctx->left_path->nodes[0], left_ii);
7033 sctx->cur_inode_mode = btrfs_inode_mode(
7034 sctx->left_path->nodes[0], left_ii);
7035 }
7036 }
7037
7038 out:
7039 return ret;
7040 }
7041
7042 /*
7043 * We have to process new refs before deleted refs, but compare_trees gives us
7044 * the new and deleted refs mixed. To fix this, we record the new/deleted refs
7045 * first and later process them in process_recorded_refs.
7046 * For the cur_inode_new_gen case, we skip recording completely because
7047 * changed_inode did already initiate processing of refs. The reason for this is
7048 * that in this case, compare_tree actually compares the refs of 2 different
7049 * inodes. To fix this, process_all_refs is used in changed_inode to handle all
7050 * refs of the right tree as deleted and all refs of the left tree as new.
7051 */
changed_ref(struct send_ctx * sctx,enum btrfs_compare_tree_result result)7052 static int changed_ref(struct send_ctx *sctx,
7053 enum btrfs_compare_tree_result result)
7054 {
7055 int ret = 0;
7056
7057 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7058 inconsistent_snapshot_error(sctx, result, "reference");
7059 return -EIO;
7060 }
7061
7062 if (!sctx->cur_inode_new_gen &&
7063 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
7064 if (result == BTRFS_COMPARE_TREE_NEW)
7065 ret = record_new_ref(sctx);
7066 else if (result == BTRFS_COMPARE_TREE_DELETED)
7067 ret = record_deleted_ref(sctx);
7068 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7069 ret = record_changed_ref(sctx);
7070 }
7071
7072 return ret;
7073 }
7074
7075 /*
7076 * Process new/deleted/changed xattrs. We skip processing in the
7077 * cur_inode_new_gen case because changed_inode did already initiate processing
7078 * of xattrs. The reason is the same as in changed_ref
7079 */
changed_xattr(struct send_ctx * sctx,enum btrfs_compare_tree_result result)7080 static int changed_xattr(struct send_ctx *sctx,
7081 enum btrfs_compare_tree_result result)
7082 {
7083 int ret = 0;
7084
7085 if (sctx->cur_ino != sctx->cmp_key->objectid) {
7086 inconsistent_snapshot_error(sctx, result, "xattr");
7087 return -EIO;
7088 }
7089
7090 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7091 if (result == BTRFS_COMPARE_TREE_NEW)
7092 ret = process_new_xattr(sctx);
7093 else if (result == BTRFS_COMPARE_TREE_DELETED)
7094 ret = process_deleted_xattr(sctx);
7095 else if (result == BTRFS_COMPARE_TREE_CHANGED)
7096 ret = process_changed_xattr(sctx);
7097 }
7098
7099 return ret;
7100 }
7101
7102 /*
7103 * Process new/deleted/changed extents. We skip processing in the
7104 * cur_inode_new_gen case because changed_inode did already initiate processing
7105 * of extents. The reason is the same as in changed_ref
7106 */
changed_extent(struct send_ctx * sctx,enum btrfs_compare_tree_result result)7107 static int changed_extent(struct send_ctx *sctx,
7108 enum btrfs_compare_tree_result result)
7109 {
7110 int ret = 0;
7111
7112 /*
7113 * We have found an extent item that changed without the inode item
7114 * having changed. This can happen either after relocation (where the
7115 * disk_bytenr of an extent item is replaced at
7116 * relocation.c:replace_file_extents()) or after deduplication into a
7117 * file in both the parent and send snapshots (where an extent item can
7118 * get modified or replaced with a new one). Note that deduplication
7119 * updates the inode item, but it only changes the iversion (sequence
7120 * field in the inode item) of the inode, so if a file is deduplicated
7121 * the same amount of times in both the parent and send snapshots, its
7122 * iversion becomes the same in both snapshots, whence the inode item is
7123 * the same on both snapshots.
7124 */
7125 if (sctx->cur_ino != sctx->cmp_key->objectid)
7126 return 0;
7127
7128 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7129 if (result != BTRFS_COMPARE_TREE_DELETED)
7130 ret = process_extent(sctx, sctx->left_path,
7131 sctx->cmp_key);
7132 }
7133
7134 return ret;
7135 }
7136
changed_verity(struct send_ctx * sctx,enum btrfs_compare_tree_result result)7137 static int changed_verity(struct send_ctx *sctx, enum btrfs_compare_tree_result result)
7138 {
7139 int ret = 0;
7140
7141 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
7142 if (result == BTRFS_COMPARE_TREE_NEW)
7143 sctx->cur_inode_needs_verity = true;
7144 }
7145 return ret;
7146 }
7147
dir_changed(struct send_ctx * sctx,u64 dir)7148 static int dir_changed(struct send_ctx *sctx, u64 dir)
7149 {
7150 u64 orig_gen, new_gen;
7151 int ret;
7152
7153 ret = get_inode_gen(sctx->send_root, dir, &new_gen);
7154 if (ret)
7155 return ret;
7156
7157 ret = get_inode_gen(sctx->parent_root, dir, &orig_gen);
7158 if (ret)
7159 return ret;
7160
7161 return (orig_gen != new_gen) ? 1 : 0;
7162 }
7163
compare_refs(struct send_ctx * sctx,struct btrfs_path * path,struct btrfs_key * key)7164 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
7165 struct btrfs_key *key)
7166 {
7167 struct btrfs_inode_extref *extref;
7168 struct extent_buffer *leaf;
7169 u64 dirid = 0, last_dirid = 0;
7170 unsigned long ptr;
7171 u32 item_size;
7172 u32 cur_offset = 0;
7173 int ref_name_len;
7174 int ret = 0;
7175
7176 /* Easy case, just check this one dirid */
7177 if (key->type == BTRFS_INODE_REF_KEY) {
7178 dirid = key->offset;
7179
7180 ret = dir_changed(sctx, dirid);
7181 goto out;
7182 }
7183
7184 leaf = path->nodes[0];
7185 item_size = btrfs_item_size(leaf, path->slots[0]);
7186 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
7187 while (cur_offset < item_size) {
7188 extref = (struct btrfs_inode_extref *)(ptr +
7189 cur_offset);
7190 dirid = btrfs_inode_extref_parent(leaf, extref);
7191 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
7192 cur_offset += ref_name_len + sizeof(*extref);
7193 if (dirid == last_dirid)
7194 continue;
7195 ret = dir_changed(sctx, dirid);
7196 if (ret)
7197 break;
7198 last_dirid = dirid;
7199 }
7200 out:
7201 return ret;
7202 }
7203
7204 /*
7205 * Updates compare related fields in sctx and simply forwards to the actual
7206 * changed_xxx functions.
7207 */
changed_cb(struct btrfs_path * left_path,struct btrfs_path * right_path,struct btrfs_key * key,enum btrfs_compare_tree_result result,struct send_ctx * sctx)7208 static int changed_cb(struct btrfs_path *left_path,
7209 struct btrfs_path *right_path,
7210 struct btrfs_key *key,
7211 enum btrfs_compare_tree_result result,
7212 struct send_ctx *sctx)
7213 {
7214 int ret = 0;
7215
7216 /*
7217 * We can not hold the commit root semaphore here. This is because in
7218 * the case of sending and receiving to the same filesystem, using a
7219 * pipe, could result in a deadlock:
7220 *
7221 * 1) The task running send blocks on the pipe because it's full;
7222 *
7223 * 2) The task running receive, which is the only consumer of the pipe,
7224 * is waiting for a transaction commit (for example due to a space
7225 * reservation when doing a write or triggering a transaction commit
7226 * when creating a subvolume);
7227 *
7228 * 3) The transaction is waiting to write lock the commit root semaphore,
7229 * but can not acquire it since it's being held at 1).
7230 *
7231 * Down this call chain we write to the pipe through kernel_write().
7232 * The same type of problem can also happen when sending to a file that
7233 * is stored in the same filesystem - when reserving space for a write
7234 * into the file, we can trigger a transaction commit.
7235 *
7236 * Our caller has supplied us with clones of leaves from the send and
7237 * parent roots, so we're safe here from a concurrent relocation and
7238 * further reallocation of metadata extents while we are here. Below we
7239 * also assert that the leaves are clones.
7240 */
7241 lockdep_assert_not_held(&sctx->send_root->fs_info->commit_root_sem);
7242
7243 /*
7244 * We always have a send root, so left_path is never NULL. We will not
7245 * have a leaf when we have reached the end of the send root but have
7246 * not yet reached the end of the parent root.
7247 */
7248 if (left_path->nodes[0])
7249 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7250 &left_path->nodes[0]->bflags));
7251 /*
7252 * When doing a full send we don't have a parent root, so right_path is
7253 * NULL. When doing an incremental send, we may have reached the end of
7254 * the parent root already, so we don't have a leaf at right_path.
7255 */
7256 if (right_path && right_path->nodes[0])
7257 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED,
7258 &right_path->nodes[0]->bflags));
7259
7260 if (result == BTRFS_COMPARE_TREE_SAME) {
7261 if (key->type == BTRFS_INODE_REF_KEY ||
7262 key->type == BTRFS_INODE_EXTREF_KEY) {
7263 ret = compare_refs(sctx, left_path, key);
7264 if (!ret)
7265 return 0;
7266 if (ret < 0)
7267 return ret;
7268 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
7269 return maybe_send_hole(sctx, left_path, key);
7270 } else {
7271 return 0;
7272 }
7273 result = BTRFS_COMPARE_TREE_CHANGED;
7274 ret = 0;
7275 }
7276
7277 sctx->left_path = left_path;
7278 sctx->right_path = right_path;
7279 sctx->cmp_key = key;
7280
7281 ret = finish_inode_if_needed(sctx, 0);
7282 if (ret < 0)
7283 goto out;
7284
7285 /* Ignore non-FS objects */
7286 if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
7287 key->objectid == BTRFS_FREE_SPACE_OBJECTID)
7288 goto out;
7289
7290 if (key->type == BTRFS_INODE_ITEM_KEY) {
7291 ret = changed_inode(sctx, result);
7292 } else if (!sctx->ignore_cur_inode) {
7293 if (key->type == BTRFS_INODE_REF_KEY ||
7294 key->type == BTRFS_INODE_EXTREF_KEY)
7295 ret = changed_ref(sctx, result);
7296 else if (key->type == BTRFS_XATTR_ITEM_KEY)
7297 ret = changed_xattr(sctx, result);
7298 else if (key->type == BTRFS_EXTENT_DATA_KEY)
7299 ret = changed_extent(sctx, result);
7300 else if (key->type == BTRFS_VERITY_DESC_ITEM_KEY &&
7301 key->offset == 0)
7302 ret = changed_verity(sctx, result);
7303 }
7304
7305 out:
7306 return ret;
7307 }
7308
search_key_again(const struct send_ctx * sctx,struct btrfs_root * root,struct btrfs_path * path,const struct btrfs_key * key)7309 static int search_key_again(const struct send_ctx *sctx,
7310 struct btrfs_root *root,
7311 struct btrfs_path *path,
7312 const struct btrfs_key *key)
7313 {
7314 int ret;
7315
7316 if (!path->need_commit_sem)
7317 lockdep_assert_held_read(&root->fs_info->commit_root_sem);
7318
7319 /*
7320 * Roots used for send operations are readonly and no one can add,
7321 * update or remove keys from them, so we should be able to find our
7322 * key again. The only exception is deduplication, which can operate on
7323 * readonly roots and add, update or remove keys to/from them - but at
7324 * the moment we don't allow it to run in parallel with send.
7325 */
7326 ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7327 ASSERT(ret <= 0);
7328 if (ret > 0) {
7329 btrfs_print_tree(path->nodes[path->lowest_level], false);
7330 btrfs_err(root->fs_info,
7331 "send: key (%llu %u %llu) not found in %s root %llu, lowest_level %d, slot %d",
7332 key->objectid, key->type, key->offset,
7333 (root == sctx->parent_root ? "parent" : "send"),
7334 root->root_key.objectid, path->lowest_level,
7335 path->slots[path->lowest_level]);
7336 return -EUCLEAN;
7337 }
7338
7339 return ret;
7340 }
7341
full_send_tree(struct send_ctx * sctx)7342 static int full_send_tree(struct send_ctx *sctx)
7343 {
7344 int ret;
7345 struct btrfs_root *send_root = sctx->send_root;
7346 struct btrfs_key key;
7347 struct btrfs_fs_info *fs_info = send_root->fs_info;
7348 struct btrfs_path *path;
7349
7350 path = alloc_path_for_send();
7351 if (!path)
7352 return -ENOMEM;
7353 path->reada = READA_FORWARD_ALWAYS;
7354
7355 key.objectid = BTRFS_FIRST_FREE_OBJECTID;
7356 key.type = BTRFS_INODE_ITEM_KEY;
7357 key.offset = 0;
7358
7359 down_read(&fs_info->commit_root_sem);
7360 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7361 up_read(&fs_info->commit_root_sem);
7362
7363 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
7364 if (ret < 0)
7365 goto out;
7366 if (ret)
7367 goto out_finish;
7368
7369 while (1) {
7370 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
7371
7372 ret = changed_cb(path, NULL, &key,
7373 BTRFS_COMPARE_TREE_NEW, sctx);
7374 if (ret < 0)
7375 goto out;
7376
7377 down_read(&fs_info->commit_root_sem);
7378 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7379 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7380 up_read(&fs_info->commit_root_sem);
7381 /*
7382 * A transaction used for relocating a block group was
7383 * committed or is about to finish its commit. Release
7384 * our path (leaf) and restart the search, so that we
7385 * avoid operating on any file extent items that are
7386 * stale, with a disk_bytenr that reflects a pre
7387 * relocation value. This way we avoid as much as
7388 * possible to fallback to regular writes when checking
7389 * if we can clone file ranges.
7390 */
7391 btrfs_release_path(path);
7392 ret = search_key_again(sctx, send_root, path, &key);
7393 if (ret < 0)
7394 goto out;
7395 } else {
7396 up_read(&fs_info->commit_root_sem);
7397 }
7398
7399 ret = btrfs_next_item(send_root, path);
7400 if (ret < 0)
7401 goto out;
7402 if (ret) {
7403 ret = 0;
7404 break;
7405 }
7406 }
7407
7408 out_finish:
7409 ret = finish_inode_if_needed(sctx, 1);
7410
7411 out:
7412 btrfs_free_path(path);
7413 return ret;
7414 }
7415
replace_node_with_clone(struct btrfs_path * path,int level)7416 static int replace_node_with_clone(struct btrfs_path *path, int level)
7417 {
7418 struct extent_buffer *clone;
7419
7420 clone = btrfs_clone_extent_buffer(path->nodes[level]);
7421 if (!clone)
7422 return -ENOMEM;
7423
7424 free_extent_buffer(path->nodes[level]);
7425 path->nodes[level] = clone;
7426
7427 return 0;
7428 }
7429
tree_move_down(struct btrfs_path * path,int * level,u64 reada_min_gen)7430 static int tree_move_down(struct btrfs_path *path, int *level, u64 reada_min_gen)
7431 {
7432 struct extent_buffer *eb;
7433 struct extent_buffer *parent = path->nodes[*level];
7434 int slot = path->slots[*level];
7435 const int nritems = btrfs_header_nritems(parent);
7436 u64 reada_max;
7437 u64 reada_done = 0;
7438
7439 lockdep_assert_held_read(&parent->fs_info->commit_root_sem);
7440
7441 BUG_ON(*level == 0);
7442 eb = btrfs_read_node_slot(parent, slot);
7443 if (IS_ERR(eb))
7444 return PTR_ERR(eb);
7445
7446 /*
7447 * Trigger readahead for the next leaves we will process, so that it is
7448 * very likely that when we need them they are already in memory and we
7449 * will not block on disk IO. For nodes we only do readahead for one,
7450 * since the time window between processing nodes is typically larger.
7451 */
7452 reada_max = (*level == 1 ? SZ_128K : eb->fs_info->nodesize);
7453
7454 for (slot++; slot < nritems && reada_done < reada_max; slot++) {
7455 if (btrfs_node_ptr_generation(parent, slot) > reada_min_gen) {
7456 btrfs_readahead_node_child(parent, slot);
7457 reada_done += eb->fs_info->nodesize;
7458 }
7459 }
7460
7461 path->nodes[*level - 1] = eb;
7462 path->slots[*level - 1] = 0;
7463 (*level)--;
7464
7465 if (*level == 0)
7466 return replace_node_with_clone(path, 0);
7467
7468 return 0;
7469 }
7470
tree_move_next_or_upnext(struct btrfs_path * path,int * level,int root_level)7471 static int tree_move_next_or_upnext(struct btrfs_path *path,
7472 int *level, int root_level)
7473 {
7474 int ret = 0;
7475 int nritems;
7476 nritems = btrfs_header_nritems(path->nodes[*level]);
7477
7478 path->slots[*level]++;
7479
7480 while (path->slots[*level] >= nritems) {
7481 if (*level == root_level) {
7482 path->slots[*level] = nritems - 1;
7483 return -1;
7484 }
7485
7486 /* move upnext */
7487 path->slots[*level] = 0;
7488 free_extent_buffer(path->nodes[*level]);
7489 path->nodes[*level] = NULL;
7490 (*level)++;
7491 path->slots[*level]++;
7492
7493 nritems = btrfs_header_nritems(path->nodes[*level]);
7494 ret = 1;
7495 }
7496 return ret;
7497 }
7498
7499 /*
7500 * Returns 1 if it had to move up and next. 0 is returned if it moved only next
7501 * or down.
7502 */
tree_advance(struct btrfs_path * path,int * level,int root_level,int allow_down,struct btrfs_key * key,u64 reada_min_gen)7503 static int tree_advance(struct btrfs_path *path,
7504 int *level, int root_level,
7505 int allow_down,
7506 struct btrfs_key *key,
7507 u64 reada_min_gen)
7508 {
7509 int ret;
7510
7511 if (*level == 0 || !allow_down) {
7512 ret = tree_move_next_or_upnext(path, level, root_level);
7513 } else {
7514 ret = tree_move_down(path, level, reada_min_gen);
7515 }
7516
7517 /*
7518 * Even if we have reached the end of a tree, ret is -1, update the key
7519 * anyway, so that in case we need to restart due to a block group
7520 * relocation, we can assert that the last key of the root node still
7521 * exists in the tree.
7522 */
7523 if (*level == 0)
7524 btrfs_item_key_to_cpu(path->nodes[*level], key,
7525 path->slots[*level]);
7526 else
7527 btrfs_node_key_to_cpu(path->nodes[*level], key,
7528 path->slots[*level]);
7529
7530 return ret;
7531 }
7532
tree_compare_item(struct btrfs_path * left_path,struct btrfs_path * right_path,char * tmp_buf)7533 static int tree_compare_item(struct btrfs_path *left_path,
7534 struct btrfs_path *right_path,
7535 char *tmp_buf)
7536 {
7537 int cmp;
7538 int len1, len2;
7539 unsigned long off1, off2;
7540
7541 len1 = btrfs_item_size(left_path->nodes[0], left_path->slots[0]);
7542 len2 = btrfs_item_size(right_path->nodes[0], right_path->slots[0]);
7543 if (len1 != len2)
7544 return 1;
7545
7546 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]);
7547 off2 = btrfs_item_ptr_offset(right_path->nodes[0],
7548 right_path->slots[0]);
7549
7550 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1);
7551
7552 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1);
7553 if (cmp)
7554 return 1;
7555 return 0;
7556 }
7557
7558 /*
7559 * A transaction used for relocating a block group was committed or is about to
7560 * finish its commit. Release our paths and restart the search, so that we are
7561 * not using stale extent buffers:
7562 *
7563 * 1) For levels > 0, we are only holding references of extent buffers, without
7564 * any locks on them, which does not prevent them from having been relocated
7565 * and reallocated after the last time we released the commit root semaphore.
7566 * The exception are the root nodes, for which we always have a clone, see
7567 * the comment at btrfs_compare_trees();
7568 *
7569 * 2) For leaves, level 0, we are holding copies (clones) of extent buffers, so
7570 * we are safe from the concurrent relocation and reallocation. However they
7571 * can have file extent items with a pre relocation disk_bytenr value, so we
7572 * restart the start from the current commit roots and clone the new leaves so
7573 * that we get the post relocation disk_bytenr values. Not doing so, could
7574 * make us clone the wrong data in case there are new extents using the old
7575 * disk_bytenr that happen to be shared.
7576 */
restart_after_relocation(struct btrfs_path * left_path,struct btrfs_path * right_path,const struct btrfs_key * left_key,const struct btrfs_key * right_key,int left_level,int right_level,const struct send_ctx * sctx)7577 static int restart_after_relocation(struct btrfs_path *left_path,
7578 struct btrfs_path *right_path,
7579 const struct btrfs_key *left_key,
7580 const struct btrfs_key *right_key,
7581 int left_level,
7582 int right_level,
7583 const struct send_ctx *sctx)
7584 {
7585 int root_level;
7586 int ret;
7587
7588 lockdep_assert_held_read(&sctx->send_root->fs_info->commit_root_sem);
7589
7590 btrfs_release_path(left_path);
7591 btrfs_release_path(right_path);
7592
7593 /*
7594 * Since keys can not be added or removed to/from our roots because they
7595 * are readonly and we do not allow deduplication to run in parallel
7596 * (which can add, remove or change keys), the layout of the trees should
7597 * not change.
7598 */
7599 left_path->lowest_level = left_level;
7600 ret = search_key_again(sctx, sctx->send_root, left_path, left_key);
7601 if (ret < 0)
7602 return ret;
7603
7604 right_path->lowest_level = right_level;
7605 ret = search_key_again(sctx, sctx->parent_root, right_path, right_key);
7606 if (ret < 0)
7607 return ret;
7608
7609 /*
7610 * If the lowest level nodes are leaves, clone them so that they can be
7611 * safely used by changed_cb() while not under the protection of the
7612 * commit root semaphore, even if relocation and reallocation happens in
7613 * parallel.
7614 */
7615 if (left_level == 0) {
7616 ret = replace_node_with_clone(left_path, 0);
7617 if (ret < 0)
7618 return ret;
7619 }
7620
7621 if (right_level == 0) {
7622 ret = replace_node_with_clone(right_path, 0);
7623 if (ret < 0)
7624 return ret;
7625 }
7626
7627 /*
7628 * Now clone the root nodes (unless they happen to be the leaves we have
7629 * already cloned). This is to protect against concurrent snapshotting of
7630 * the send and parent roots (see the comment at btrfs_compare_trees()).
7631 */
7632 root_level = btrfs_header_level(sctx->send_root->commit_root);
7633 if (root_level > 0) {
7634 ret = replace_node_with_clone(left_path, root_level);
7635 if (ret < 0)
7636 return ret;
7637 }
7638
7639 root_level = btrfs_header_level(sctx->parent_root->commit_root);
7640 if (root_level > 0) {
7641 ret = replace_node_with_clone(right_path, root_level);
7642 if (ret < 0)
7643 return ret;
7644 }
7645
7646 return 0;
7647 }
7648
7649 /*
7650 * This function compares two trees and calls the provided callback for
7651 * every changed/new/deleted item it finds.
7652 * If shared tree blocks are encountered, whole subtrees are skipped, making
7653 * the compare pretty fast on snapshotted subvolumes.
7654 *
7655 * This currently works on commit roots only. As commit roots are read only,
7656 * we don't do any locking. The commit roots are protected with transactions.
7657 * Transactions are ended and rejoined when a commit is tried in between.
7658 *
7659 * This function checks for modifications done to the trees while comparing.
7660 * If it detects a change, it aborts immediately.
7661 */
btrfs_compare_trees(struct btrfs_root * left_root,struct btrfs_root * right_root,struct send_ctx * sctx)7662 static int btrfs_compare_trees(struct btrfs_root *left_root,
7663 struct btrfs_root *right_root, struct send_ctx *sctx)
7664 {
7665 struct btrfs_fs_info *fs_info = left_root->fs_info;
7666 int ret;
7667 int cmp;
7668 struct btrfs_path *left_path = NULL;
7669 struct btrfs_path *right_path = NULL;
7670 struct btrfs_key left_key;
7671 struct btrfs_key right_key;
7672 char *tmp_buf = NULL;
7673 int left_root_level;
7674 int right_root_level;
7675 int left_level;
7676 int right_level;
7677 int left_end_reached = 0;
7678 int right_end_reached = 0;
7679 int advance_left = 0;
7680 int advance_right = 0;
7681 u64 left_blockptr;
7682 u64 right_blockptr;
7683 u64 left_gen;
7684 u64 right_gen;
7685 u64 reada_min_gen;
7686
7687 left_path = btrfs_alloc_path();
7688 if (!left_path) {
7689 ret = -ENOMEM;
7690 goto out;
7691 }
7692 right_path = btrfs_alloc_path();
7693 if (!right_path) {
7694 ret = -ENOMEM;
7695 goto out;
7696 }
7697
7698 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
7699 if (!tmp_buf) {
7700 ret = -ENOMEM;
7701 goto out;
7702 }
7703
7704 left_path->search_commit_root = 1;
7705 left_path->skip_locking = 1;
7706 right_path->search_commit_root = 1;
7707 right_path->skip_locking = 1;
7708
7709 /*
7710 * Strategy: Go to the first items of both trees. Then do
7711 *
7712 * If both trees are at level 0
7713 * Compare keys of current items
7714 * If left < right treat left item as new, advance left tree
7715 * and repeat
7716 * If left > right treat right item as deleted, advance right tree
7717 * and repeat
7718 * If left == right do deep compare of items, treat as changed if
7719 * needed, advance both trees and repeat
7720 * If both trees are at the same level but not at level 0
7721 * Compare keys of current nodes/leafs
7722 * If left < right advance left tree and repeat
7723 * If left > right advance right tree and repeat
7724 * If left == right compare blockptrs of the next nodes/leafs
7725 * If they match advance both trees but stay at the same level
7726 * and repeat
7727 * If they don't match advance both trees while allowing to go
7728 * deeper and repeat
7729 * If tree levels are different
7730 * Advance the tree that needs it and repeat
7731 *
7732 * Advancing a tree means:
7733 * If we are at level 0, try to go to the next slot. If that's not
7734 * possible, go one level up and repeat. Stop when we found a level
7735 * where we could go to the next slot. We may at this point be on a
7736 * node or a leaf.
7737 *
7738 * If we are not at level 0 and not on shared tree blocks, go one
7739 * level deeper.
7740 *
7741 * If we are not at level 0 and on shared tree blocks, go one slot to
7742 * the right if possible or go up and right.
7743 */
7744
7745 down_read(&fs_info->commit_root_sem);
7746 left_level = btrfs_header_level(left_root->commit_root);
7747 left_root_level = left_level;
7748 /*
7749 * We clone the root node of the send and parent roots to prevent races
7750 * with snapshot creation of these roots. Snapshot creation COWs the
7751 * root node of a tree, so after the transaction is committed the old
7752 * extent can be reallocated while this send operation is still ongoing.
7753 * So we clone them, under the commit root semaphore, to be race free.
7754 */
7755 left_path->nodes[left_level] =
7756 btrfs_clone_extent_buffer(left_root->commit_root);
7757 if (!left_path->nodes[left_level]) {
7758 ret = -ENOMEM;
7759 goto out_unlock;
7760 }
7761
7762 right_level = btrfs_header_level(right_root->commit_root);
7763 right_root_level = right_level;
7764 right_path->nodes[right_level] =
7765 btrfs_clone_extent_buffer(right_root->commit_root);
7766 if (!right_path->nodes[right_level]) {
7767 ret = -ENOMEM;
7768 goto out_unlock;
7769 }
7770 /*
7771 * Our right root is the parent root, while the left root is the "send"
7772 * root. We know that all new nodes/leaves in the left root must have
7773 * a generation greater than the right root's generation, so we trigger
7774 * readahead for those nodes and leaves of the left root, as we know we
7775 * will need to read them at some point.
7776 */
7777 reada_min_gen = btrfs_header_generation(right_root->commit_root);
7778
7779 if (left_level == 0)
7780 btrfs_item_key_to_cpu(left_path->nodes[left_level],
7781 &left_key, left_path->slots[left_level]);
7782 else
7783 btrfs_node_key_to_cpu(left_path->nodes[left_level],
7784 &left_key, left_path->slots[left_level]);
7785 if (right_level == 0)
7786 btrfs_item_key_to_cpu(right_path->nodes[right_level],
7787 &right_key, right_path->slots[right_level]);
7788 else
7789 btrfs_node_key_to_cpu(right_path->nodes[right_level],
7790 &right_key, right_path->slots[right_level]);
7791
7792 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7793
7794 while (1) {
7795 if (need_resched() ||
7796 rwsem_is_contended(&fs_info->commit_root_sem)) {
7797 up_read(&fs_info->commit_root_sem);
7798 cond_resched();
7799 down_read(&fs_info->commit_root_sem);
7800 }
7801
7802 if (fs_info->last_reloc_trans > sctx->last_reloc_trans) {
7803 ret = restart_after_relocation(left_path, right_path,
7804 &left_key, &right_key,
7805 left_level, right_level,
7806 sctx);
7807 if (ret < 0)
7808 goto out_unlock;
7809 sctx->last_reloc_trans = fs_info->last_reloc_trans;
7810 }
7811
7812 if (advance_left && !left_end_reached) {
7813 ret = tree_advance(left_path, &left_level,
7814 left_root_level,
7815 advance_left != ADVANCE_ONLY_NEXT,
7816 &left_key, reada_min_gen);
7817 if (ret == -1)
7818 left_end_reached = ADVANCE;
7819 else if (ret < 0)
7820 goto out_unlock;
7821 advance_left = 0;
7822 }
7823 if (advance_right && !right_end_reached) {
7824 ret = tree_advance(right_path, &right_level,
7825 right_root_level,
7826 advance_right != ADVANCE_ONLY_NEXT,
7827 &right_key, reada_min_gen);
7828 if (ret == -1)
7829 right_end_reached = ADVANCE;
7830 else if (ret < 0)
7831 goto out_unlock;
7832 advance_right = 0;
7833 }
7834
7835 if (left_end_reached && right_end_reached) {
7836 ret = 0;
7837 goto out_unlock;
7838 } else if (left_end_reached) {
7839 if (right_level == 0) {
7840 up_read(&fs_info->commit_root_sem);
7841 ret = changed_cb(left_path, right_path,
7842 &right_key,
7843 BTRFS_COMPARE_TREE_DELETED,
7844 sctx);
7845 if (ret < 0)
7846 goto out;
7847 down_read(&fs_info->commit_root_sem);
7848 }
7849 advance_right = ADVANCE;
7850 continue;
7851 } else if (right_end_reached) {
7852 if (left_level == 0) {
7853 up_read(&fs_info->commit_root_sem);
7854 ret = changed_cb(left_path, right_path,
7855 &left_key,
7856 BTRFS_COMPARE_TREE_NEW,
7857 sctx);
7858 if (ret < 0)
7859 goto out;
7860 down_read(&fs_info->commit_root_sem);
7861 }
7862 advance_left = ADVANCE;
7863 continue;
7864 }
7865
7866 if (left_level == 0 && right_level == 0) {
7867 up_read(&fs_info->commit_root_sem);
7868 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7869 if (cmp < 0) {
7870 ret = changed_cb(left_path, right_path,
7871 &left_key,
7872 BTRFS_COMPARE_TREE_NEW,
7873 sctx);
7874 advance_left = ADVANCE;
7875 } else if (cmp > 0) {
7876 ret = changed_cb(left_path, right_path,
7877 &right_key,
7878 BTRFS_COMPARE_TREE_DELETED,
7879 sctx);
7880 advance_right = ADVANCE;
7881 } else {
7882 enum btrfs_compare_tree_result result;
7883
7884 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0]));
7885 ret = tree_compare_item(left_path, right_path,
7886 tmp_buf);
7887 if (ret)
7888 result = BTRFS_COMPARE_TREE_CHANGED;
7889 else
7890 result = BTRFS_COMPARE_TREE_SAME;
7891 ret = changed_cb(left_path, right_path,
7892 &left_key, result, sctx);
7893 advance_left = ADVANCE;
7894 advance_right = ADVANCE;
7895 }
7896
7897 if (ret < 0)
7898 goto out;
7899 down_read(&fs_info->commit_root_sem);
7900 } else if (left_level == right_level) {
7901 cmp = btrfs_comp_cpu_keys(&left_key, &right_key);
7902 if (cmp < 0) {
7903 advance_left = ADVANCE;
7904 } else if (cmp > 0) {
7905 advance_right = ADVANCE;
7906 } else {
7907 left_blockptr = btrfs_node_blockptr(
7908 left_path->nodes[left_level],
7909 left_path->slots[left_level]);
7910 right_blockptr = btrfs_node_blockptr(
7911 right_path->nodes[right_level],
7912 right_path->slots[right_level]);
7913 left_gen = btrfs_node_ptr_generation(
7914 left_path->nodes[left_level],
7915 left_path->slots[left_level]);
7916 right_gen = btrfs_node_ptr_generation(
7917 right_path->nodes[right_level],
7918 right_path->slots[right_level]);
7919 if (left_blockptr == right_blockptr &&
7920 left_gen == right_gen) {
7921 /*
7922 * As we're on a shared block, don't
7923 * allow to go deeper.
7924 */
7925 advance_left = ADVANCE_ONLY_NEXT;
7926 advance_right = ADVANCE_ONLY_NEXT;
7927 } else {
7928 advance_left = ADVANCE;
7929 advance_right = ADVANCE;
7930 }
7931 }
7932 } else if (left_level < right_level) {
7933 advance_right = ADVANCE;
7934 } else {
7935 advance_left = ADVANCE;
7936 }
7937 }
7938
7939 out_unlock:
7940 up_read(&fs_info->commit_root_sem);
7941 out:
7942 btrfs_free_path(left_path);
7943 btrfs_free_path(right_path);
7944 kvfree(tmp_buf);
7945 return ret;
7946 }
7947
send_subvol(struct send_ctx * sctx)7948 static int send_subvol(struct send_ctx *sctx)
7949 {
7950 int ret;
7951
7952 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
7953 ret = send_header(sctx);
7954 if (ret < 0)
7955 goto out;
7956 }
7957
7958 ret = send_subvol_begin(sctx);
7959 if (ret < 0)
7960 goto out;
7961
7962 if (sctx->parent_root) {
7963 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx);
7964 if (ret < 0)
7965 goto out;
7966 ret = finish_inode_if_needed(sctx, 1);
7967 if (ret < 0)
7968 goto out;
7969 } else {
7970 ret = full_send_tree(sctx);
7971 if (ret < 0)
7972 goto out;
7973 }
7974
7975 out:
7976 free_recorded_refs(sctx);
7977 return ret;
7978 }
7979
7980 /*
7981 * If orphan cleanup did remove any orphans from a root, it means the tree
7982 * was modified and therefore the commit root is not the same as the current
7983 * root anymore. This is a problem, because send uses the commit root and
7984 * therefore can see inode items that don't exist in the current root anymore,
7985 * and for example make calls to btrfs_iget, which will do tree lookups based
7986 * on the current root and not on the commit root. Those lookups will fail,
7987 * returning a -ESTALE error, and making send fail with that error. So make
7988 * sure a send does not see any orphans we have just removed, and that it will
7989 * see the same inodes regardless of whether a transaction commit happened
7990 * before it started (meaning that the commit root will be the same as the
7991 * current root) or not.
7992 */
ensure_commit_roots_uptodate(struct send_ctx * sctx)7993 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
7994 {
7995 int i;
7996 struct btrfs_trans_handle *trans = NULL;
7997
7998 again:
7999 if (sctx->parent_root &&
8000 sctx->parent_root->node != sctx->parent_root->commit_root)
8001 goto commit_trans;
8002
8003 for (i = 0; i < sctx->clone_roots_cnt; i++)
8004 if (sctx->clone_roots[i].root->node !=
8005 sctx->clone_roots[i].root->commit_root)
8006 goto commit_trans;
8007
8008 if (trans)
8009 return btrfs_end_transaction(trans);
8010
8011 return 0;
8012
8013 commit_trans:
8014 /* Use any root, all fs roots will get their commit roots updated. */
8015 if (!trans) {
8016 trans = btrfs_join_transaction(sctx->send_root);
8017 if (IS_ERR(trans))
8018 return PTR_ERR(trans);
8019 goto again;
8020 }
8021
8022 return btrfs_commit_transaction(trans);
8023 }
8024
8025 /*
8026 * Make sure any existing dellaloc is flushed for any root used by a send
8027 * operation so that we do not miss any data and we do not race with writeback
8028 * finishing and changing a tree while send is using the tree. This could
8029 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and
8030 * a send operation then uses the subvolume.
8031 * After flushing delalloc ensure_commit_roots_uptodate() must be called.
8032 */
flush_delalloc_roots(struct send_ctx * sctx)8033 static int flush_delalloc_roots(struct send_ctx *sctx)
8034 {
8035 struct btrfs_root *root = sctx->parent_root;
8036 int ret;
8037 int i;
8038
8039 if (root) {
8040 ret = btrfs_start_delalloc_snapshot(root, false);
8041 if (ret)
8042 return ret;
8043 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8044 }
8045
8046 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8047 root = sctx->clone_roots[i].root;
8048 ret = btrfs_start_delalloc_snapshot(root, false);
8049 if (ret)
8050 return ret;
8051 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX);
8052 }
8053
8054 return 0;
8055 }
8056
btrfs_root_dec_send_in_progress(struct btrfs_root * root)8057 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
8058 {
8059 spin_lock(&root->root_item_lock);
8060 root->send_in_progress--;
8061 /*
8062 * Not much left to do, we don't know why it's unbalanced and
8063 * can't blindly reset it to 0.
8064 */
8065 if (root->send_in_progress < 0)
8066 btrfs_err(root->fs_info,
8067 "send_in_progress unbalanced %d root %llu",
8068 root->send_in_progress, root->root_key.objectid);
8069 spin_unlock(&root->root_item_lock);
8070 }
8071
dedupe_in_progress_warn(const struct btrfs_root * root)8072 static void dedupe_in_progress_warn(const struct btrfs_root *root)
8073 {
8074 btrfs_warn_rl(root->fs_info,
8075 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)",
8076 root->root_key.objectid, root->dedupe_in_progress);
8077 }
8078
btrfs_ioctl_send(struct inode * inode,struct btrfs_ioctl_send_args * arg)8079 long btrfs_ioctl_send(struct inode *inode, struct btrfs_ioctl_send_args *arg)
8080 {
8081 int ret = 0;
8082 struct btrfs_root *send_root = BTRFS_I(inode)->root;
8083 struct btrfs_fs_info *fs_info = send_root->fs_info;
8084 struct btrfs_root *clone_root;
8085 struct send_ctx *sctx = NULL;
8086 u32 i;
8087 u64 *clone_sources_tmp = NULL;
8088 int clone_sources_to_rollback = 0;
8089 size_t alloc_size;
8090 int sort_clone_roots = 0;
8091 struct btrfs_lru_cache_entry *entry;
8092 struct btrfs_lru_cache_entry *tmp;
8093
8094 if (!capable(CAP_SYS_ADMIN))
8095 return -EPERM;
8096
8097 /*
8098 * The subvolume must remain read-only during send, protect against
8099 * making it RW. This also protects against deletion.
8100 */
8101 spin_lock(&send_root->root_item_lock);
8102 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) {
8103 dedupe_in_progress_warn(send_root);
8104 spin_unlock(&send_root->root_item_lock);
8105 return -EAGAIN;
8106 }
8107 send_root->send_in_progress++;
8108 spin_unlock(&send_root->root_item_lock);
8109
8110 /*
8111 * Userspace tools do the checks and warn the user if it's
8112 * not RO.
8113 */
8114 if (!btrfs_root_readonly(send_root)) {
8115 ret = -EPERM;
8116 goto out;
8117 }
8118
8119 /*
8120 * Check that we don't overflow at later allocations, we request
8121 * clone_sources_count + 1 items, and compare to unsigned long inside
8122 * access_ok. Also set an upper limit for allocation size so this can't
8123 * easily exhaust memory. Max number of clone sources is about 200K.
8124 */
8125 if (arg->clone_sources_count > SZ_8M / sizeof(struct clone_root)) {
8126 ret = -EINVAL;
8127 goto out;
8128 }
8129
8130 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
8131 ret = -EOPNOTSUPP;
8132 goto out;
8133 }
8134
8135 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
8136 if (!sctx) {
8137 ret = -ENOMEM;
8138 goto out;
8139 }
8140
8141 INIT_LIST_HEAD(&sctx->new_refs);
8142 INIT_LIST_HEAD(&sctx->deleted_refs);
8143
8144 btrfs_lru_cache_init(&sctx->name_cache, SEND_MAX_NAME_CACHE_SIZE);
8145 btrfs_lru_cache_init(&sctx->backref_cache, SEND_MAX_BACKREF_CACHE_SIZE);
8146 btrfs_lru_cache_init(&sctx->dir_created_cache,
8147 SEND_MAX_DIR_CREATED_CACHE_SIZE);
8148 /*
8149 * This cache is periodically trimmed to a fixed size elsewhere, see
8150 * cache_dir_utimes() and trim_dir_utimes_cache().
8151 */
8152 btrfs_lru_cache_init(&sctx->dir_utimes_cache, 0);
8153
8154 sctx->pending_dir_moves = RB_ROOT;
8155 sctx->waiting_dir_moves = RB_ROOT;
8156 sctx->orphan_dirs = RB_ROOT;
8157 sctx->rbtree_new_refs = RB_ROOT;
8158 sctx->rbtree_deleted_refs = RB_ROOT;
8159
8160 sctx->flags = arg->flags;
8161
8162 if (arg->flags & BTRFS_SEND_FLAG_VERSION) {
8163 if (arg->version > BTRFS_SEND_STREAM_VERSION) {
8164 ret = -EPROTO;
8165 goto out;
8166 }
8167 /* Zero means "use the highest version" */
8168 sctx->proto = arg->version ?: BTRFS_SEND_STREAM_VERSION;
8169 } else {
8170 sctx->proto = 1;
8171 }
8172 if ((arg->flags & BTRFS_SEND_FLAG_COMPRESSED) && sctx->proto < 2) {
8173 ret = -EINVAL;
8174 goto out;
8175 }
8176
8177 sctx->send_filp = fget(arg->send_fd);
8178 if (!sctx->send_filp || !(sctx->send_filp->f_mode & FMODE_WRITE)) {
8179 ret = -EBADF;
8180 goto out;
8181 }
8182
8183 sctx->send_root = send_root;
8184 /*
8185 * Unlikely but possible, if the subvolume is marked for deletion but
8186 * is slow to remove the directory entry, send can still be started
8187 */
8188 if (btrfs_root_dead(sctx->send_root)) {
8189 ret = -EPERM;
8190 goto out;
8191 }
8192
8193 sctx->clone_roots_cnt = arg->clone_sources_count;
8194
8195 if (sctx->proto >= 2) {
8196 u32 send_buf_num_pages;
8197
8198 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V2;
8199 sctx->send_buf = vmalloc(sctx->send_max_size);
8200 if (!sctx->send_buf) {
8201 ret = -ENOMEM;
8202 goto out;
8203 }
8204 send_buf_num_pages = sctx->send_max_size >> PAGE_SHIFT;
8205 sctx->send_buf_pages = kcalloc(send_buf_num_pages,
8206 sizeof(*sctx->send_buf_pages),
8207 GFP_KERNEL);
8208 if (!sctx->send_buf_pages) {
8209 ret = -ENOMEM;
8210 goto out;
8211 }
8212 for (i = 0; i < send_buf_num_pages; i++) {
8213 sctx->send_buf_pages[i] =
8214 vmalloc_to_page(sctx->send_buf + (i << PAGE_SHIFT));
8215 }
8216 } else {
8217 sctx->send_max_size = BTRFS_SEND_BUF_SIZE_V1;
8218 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
8219 }
8220 if (!sctx->send_buf) {
8221 ret = -ENOMEM;
8222 goto out;
8223 }
8224
8225 sctx->clone_roots = kvcalloc(arg->clone_sources_count + 1,
8226 sizeof(*sctx->clone_roots),
8227 GFP_KERNEL);
8228 if (!sctx->clone_roots) {
8229 ret = -ENOMEM;
8230 goto out;
8231 }
8232
8233 alloc_size = array_size(sizeof(*arg->clone_sources),
8234 arg->clone_sources_count);
8235
8236 if (arg->clone_sources_count) {
8237 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
8238 if (!clone_sources_tmp) {
8239 ret = -ENOMEM;
8240 goto out;
8241 }
8242
8243 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
8244 alloc_size);
8245 if (ret) {
8246 ret = -EFAULT;
8247 goto out;
8248 }
8249
8250 for (i = 0; i < arg->clone_sources_count; i++) {
8251 clone_root = btrfs_get_fs_root(fs_info,
8252 clone_sources_tmp[i], true);
8253 if (IS_ERR(clone_root)) {
8254 ret = PTR_ERR(clone_root);
8255 goto out;
8256 }
8257 spin_lock(&clone_root->root_item_lock);
8258 if (!btrfs_root_readonly(clone_root) ||
8259 btrfs_root_dead(clone_root)) {
8260 spin_unlock(&clone_root->root_item_lock);
8261 btrfs_put_root(clone_root);
8262 ret = -EPERM;
8263 goto out;
8264 }
8265 if (clone_root->dedupe_in_progress) {
8266 dedupe_in_progress_warn(clone_root);
8267 spin_unlock(&clone_root->root_item_lock);
8268 btrfs_put_root(clone_root);
8269 ret = -EAGAIN;
8270 goto out;
8271 }
8272 clone_root->send_in_progress++;
8273 spin_unlock(&clone_root->root_item_lock);
8274
8275 sctx->clone_roots[i].root = clone_root;
8276 clone_sources_to_rollback = i + 1;
8277 }
8278 kvfree(clone_sources_tmp);
8279 clone_sources_tmp = NULL;
8280 }
8281
8282 if (arg->parent_root) {
8283 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root,
8284 true);
8285 if (IS_ERR(sctx->parent_root)) {
8286 ret = PTR_ERR(sctx->parent_root);
8287 goto out;
8288 }
8289
8290 spin_lock(&sctx->parent_root->root_item_lock);
8291 sctx->parent_root->send_in_progress++;
8292 if (!btrfs_root_readonly(sctx->parent_root) ||
8293 btrfs_root_dead(sctx->parent_root)) {
8294 spin_unlock(&sctx->parent_root->root_item_lock);
8295 ret = -EPERM;
8296 goto out;
8297 }
8298 if (sctx->parent_root->dedupe_in_progress) {
8299 dedupe_in_progress_warn(sctx->parent_root);
8300 spin_unlock(&sctx->parent_root->root_item_lock);
8301 ret = -EAGAIN;
8302 goto out;
8303 }
8304 spin_unlock(&sctx->parent_root->root_item_lock);
8305 }
8306
8307 /*
8308 * Clones from send_root are allowed, but only if the clone source
8309 * is behind the current send position. This is checked while searching
8310 * for possible clone sources.
8311 */
8312 sctx->clone_roots[sctx->clone_roots_cnt++].root =
8313 btrfs_grab_root(sctx->send_root);
8314
8315 /* We do a bsearch later */
8316 sort(sctx->clone_roots, sctx->clone_roots_cnt,
8317 sizeof(*sctx->clone_roots), __clone_root_cmp_sort,
8318 NULL);
8319 sort_clone_roots = 1;
8320
8321 ret = flush_delalloc_roots(sctx);
8322 if (ret)
8323 goto out;
8324
8325 ret = ensure_commit_roots_uptodate(sctx);
8326 if (ret)
8327 goto out;
8328
8329 ret = send_subvol(sctx);
8330 if (ret < 0)
8331 goto out;
8332
8333 btrfs_lru_cache_for_each_entry_safe(&sctx->dir_utimes_cache, entry, tmp) {
8334 ret = send_utimes(sctx, entry->key, entry->gen);
8335 if (ret < 0)
8336 goto out;
8337 btrfs_lru_cache_remove(&sctx->dir_utimes_cache, entry);
8338 }
8339
8340 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) {
8341 ret = begin_cmd(sctx, BTRFS_SEND_C_END);
8342 if (ret < 0)
8343 goto out;
8344 ret = send_cmd(sctx);
8345 if (ret < 0)
8346 goto out;
8347 }
8348
8349 out:
8350 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves));
8351 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) {
8352 struct rb_node *n;
8353 struct pending_dir_move *pm;
8354
8355 n = rb_first(&sctx->pending_dir_moves);
8356 pm = rb_entry(n, struct pending_dir_move, node);
8357 while (!list_empty(&pm->list)) {
8358 struct pending_dir_move *pm2;
8359
8360 pm2 = list_first_entry(&pm->list,
8361 struct pending_dir_move, list);
8362 free_pending_move(sctx, pm2);
8363 }
8364 free_pending_move(sctx, pm);
8365 }
8366
8367 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves));
8368 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) {
8369 struct rb_node *n;
8370 struct waiting_dir_move *dm;
8371
8372 n = rb_first(&sctx->waiting_dir_moves);
8373 dm = rb_entry(n, struct waiting_dir_move, node);
8374 rb_erase(&dm->node, &sctx->waiting_dir_moves);
8375 kfree(dm);
8376 }
8377
8378 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs));
8379 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) {
8380 struct rb_node *n;
8381 struct orphan_dir_info *odi;
8382
8383 n = rb_first(&sctx->orphan_dirs);
8384 odi = rb_entry(n, struct orphan_dir_info, node);
8385 free_orphan_dir_info(sctx, odi);
8386 }
8387
8388 if (sort_clone_roots) {
8389 for (i = 0; i < sctx->clone_roots_cnt; i++) {
8390 btrfs_root_dec_send_in_progress(
8391 sctx->clone_roots[i].root);
8392 btrfs_put_root(sctx->clone_roots[i].root);
8393 }
8394 } else {
8395 for (i = 0; sctx && i < clone_sources_to_rollback; i++) {
8396 btrfs_root_dec_send_in_progress(
8397 sctx->clone_roots[i].root);
8398 btrfs_put_root(sctx->clone_roots[i].root);
8399 }
8400
8401 btrfs_root_dec_send_in_progress(send_root);
8402 }
8403 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) {
8404 btrfs_root_dec_send_in_progress(sctx->parent_root);
8405 btrfs_put_root(sctx->parent_root);
8406 }
8407
8408 kvfree(clone_sources_tmp);
8409
8410 if (sctx) {
8411 if (sctx->send_filp)
8412 fput(sctx->send_filp);
8413
8414 kvfree(sctx->clone_roots);
8415 kfree(sctx->send_buf_pages);
8416 kvfree(sctx->send_buf);
8417 kvfree(sctx->verity_descriptor);
8418
8419 close_current_inode(sctx);
8420
8421 btrfs_lru_cache_clear(&sctx->name_cache);
8422 btrfs_lru_cache_clear(&sctx->backref_cache);
8423 btrfs_lru_cache_clear(&sctx->dir_created_cache);
8424 btrfs_lru_cache_clear(&sctx->dir_utimes_cache);
8425
8426 kfree(sctx);
8427 }
8428
8429 return ret;
8430 }
8431