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