~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/btrfs/send.c

Version: ~ [ linux-5.1-rc2 ] ~ [ linux-5.0.4 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.31 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.108 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.165 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.177 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.137 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.63 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Copyright (C) 2012 Alexander Block.  All rights reserved.
  3  *
  4  * This program is free software; you can redistribute it and/or
  5  * modify it under the terms of the GNU General Public
  6  * License v2 as published by the Free Software Foundation.
  7  *
  8  * This program is distributed in the hope that it will be useful,
  9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 11  * General Public License for more details.
 12  *
 13  * You should have received a copy of the GNU General Public
 14  * License along with this program; if not, write to the
 15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 16  * Boston, MA 021110-1307, USA.
 17  */
 18 
 19 #include <linux/bsearch.h>
 20 #include <linux/fs.h>
 21 #include <linux/file.h>
 22 #include <linux/sort.h>
 23 #include <linux/mount.h>
 24 #include <linux/xattr.h>
 25 #include <linux/posix_acl_xattr.h>
 26 #include <linux/radix-tree.h>
 27 #include <linux/vmalloc.h>
 28 #include <linux/string.h>
 29 #include <linux/compat.h>
 30 
 31 #include "send.h"
 32 #include "backref.h"
 33 #include "hash.h"
 34 #include "locking.h"
 35 #include "disk-io.h"
 36 #include "btrfs_inode.h"
 37 #include "transaction.h"
 38 #include "compression.h"
 39 
 40 /*
 41  * A fs_path is a helper to dynamically build path names with unknown size.
 42  * It reallocates the internal buffer on demand.
 43  * It allows fast adding of path elements on the right side (normal path) and
 44  * fast adding to the left side (reversed path). A reversed path can also be
 45  * unreversed if needed.
 46  */
 47 struct fs_path {
 48         union {
 49                 struct {
 50                         char *start;
 51                         char *end;
 52 
 53                         char *buf;
 54                         unsigned short buf_len:15;
 55                         unsigned short reversed:1;
 56                         char inline_buf[];
 57                 };
 58                 /*
 59                  * Average path length does not exceed 200 bytes, we'll have
 60                  * better packing in the slab and higher chance to satisfy
 61                  * a allocation later during send.
 62                  */
 63                 char pad[256];
 64         };
 65 };
 66 #define FS_PATH_INLINE_SIZE \
 67         (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
 68 
 69 
 70 /* reused for each extent */
 71 struct clone_root {
 72         struct btrfs_root *root;
 73         u64 ino;
 74         u64 offset;
 75 
 76         u64 found_refs;
 77 };
 78 
 79 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128
 80 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
 81 
 82 struct send_ctx {
 83         struct file *send_filp;
 84         loff_t send_off;
 85         char *send_buf;
 86         u32 send_size;
 87         u32 send_max_size;
 88         u64 total_send_size;
 89         u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
 90         u64 flags;      /* 'flags' member of btrfs_ioctl_send_args is u64 */
 91 
 92         struct btrfs_root *send_root;
 93         struct btrfs_root *parent_root;
 94         struct clone_root *clone_roots;
 95         int clone_roots_cnt;
 96 
 97         /* current state of the compare_tree call */
 98         struct btrfs_path *left_path;
 99         struct btrfs_path *right_path;
100         struct btrfs_key *cmp_key;
101 
102         /*
103          * infos of the currently processed inode. In case of deleted inodes,
104          * these are the values from the deleted inode.
105          */
106         u64 cur_ino;
107         u64 cur_inode_gen;
108         int cur_inode_new;
109         int cur_inode_new_gen;
110         int cur_inode_deleted;
111         u64 cur_inode_size;
112         u64 cur_inode_mode;
113         u64 cur_inode_rdev;
114         u64 cur_inode_last_extent;
115 
116         u64 send_progress;
117 
118         struct list_head new_refs;
119         struct list_head deleted_refs;
120 
121         struct radix_tree_root name_cache;
122         struct list_head name_cache_list;
123         int name_cache_size;
124 
125         struct file_ra_state ra;
126 
127         char *read_buf;
128 
129         /*
130          * We process inodes by their increasing order, so if before an
131          * incremental send we reverse the parent/child relationship of
132          * directories such that a directory with a lower inode number was
133          * the parent of a directory with a higher inode number, and the one
134          * becoming the new parent got renamed too, we can't rename/move the
135          * directory with lower inode number when we finish processing it - we
136          * must process the directory with higher inode number first, then
137          * rename/move it and then rename/move the directory with lower inode
138          * number. Example follows.
139          *
140          * Tree state when the first send was performed:
141          *
142          * .
143          * |-- a                   (ino 257)
144          *     |-- b               (ino 258)
145          *         |
146          *         |
147          *         |-- c           (ino 259)
148          *         |   |-- d       (ino 260)
149          *         |
150          *         |-- c2          (ino 261)
151          *
152          * Tree state when the second (incremental) send is performed:
153          *
154          * .
155          * |-- a                   (ino 257)
156          *     |-- b               (ino 258)
157          *         |-- c2          (ino 261)
158          *             |-- d2      (ino 260)
159          *                 |-- cc  (ino 259)
160          *
161          * The sequence of steps that lead to the second state was:
162          *
163          * mv /a/b/c/d /a/b/c2/d2
164          * mv /a/b/c /a/b/c2/d2/cc
165          *
166          * "c" has lower inode number, but we can't move it (2nd mv operation)
167          * before we move "d", which has higher inode number.
168          *
169          * So we just memorize which move/rename operations must be performed
170          * later when their respective parent is processed and moved/renamed.
171          */
172 
173         /* Indexed by parent directory inode number. */
174         struct rb_root pending_dir_moves;
175 
176         /*
177          * Reverse index, indexed by the inode number of a directory that
178          * is waiting for the move/rename of its immediate parent before its
179          * own move/rename can be performed.
180          */
181         struct rb_root waiting_dir_moves;
182 
183         /*
184          * A directory that is going to be rm'ed might have a child directory
185          * which is in the pending directory moves index above. In this case,
186          * the directory can only be removed after the move/rename of its child
187          * is performed. Example:
188          *
189          * Parent snapshot:
190          *
191          * .                        (ino 256)
192          * |-- a/                   (ino 257)
193          *     |-- b/               (ino 258)
194          *         |-- c/           (ino 259)
195          *         |   |-- x/       (ino 260)
196          *         |
197          *         |-- y/           (ino 261)
198          *
199          * Send snapshot:
200          *
201          * .                        (ino 256)
202          * |-- a/                   (ino 257)
203          *     |-- b/               (ino 258)
204          *         |-- YY/          (ino 261)
205          *              |-- x/      (ino 260)
206          *
207          * Sequence of steps that lead to the send snapshot:
208          * rm -f /a/b/c/foo.txt
209          * mv /a/b/y /a/b/YY
210          * mv /a/b/c/x /a/b/YY
211          * rmdir /a/b/c
212          *
213          * When the child is processed, its move/rename is delayed until its
214          * parent is processed (as explained above), but all other operations
215          * like update utimes, chown, chgrp, etc, are performed and the paths
216          * that it uses for those operations must use the orphanized name of
217          * its parent (the directory we're going to rm later), so we need to
218          * memorize that name.
219          *
220          * Indexed by the inode number of the directory to be deleted.
221          */
222         struct rb_root orphan_dirs;
223 };
224 
225 struct pending_dir_move {
226         struct rb_node node;
227         struct list_head list;
228         u64 parent_ino;
229         u64 ino;
230         u64 gen;
231         struct list_head update_refs;
232 };
233 
234 struct waiting_dir_move {
235         struct rb_node node;
236         u64 ino;
237         /*
238          * There might be some directory that could not be removed because it
239          * was waiting for this directory inode to be moved first. Therefore
240          * after this directory is moved, we can try to rmdir the ino rmdir_ino.
241          */
242         u64 rmdir_ino;
243         bool orphanized;
244 };
245 
246 struct orphan_dir_info {
247         struct rb_node node;
248         u64 ino;
249         u64 gen;
250 };
251 
252 struct name_cache_entry {
253         struct list_head list;
254         /*
255          * radix_tree has only 32bit entries but we need to handle 64bit inums.
256          * We use the lower 32bit of the 64bit inum to store it in the tree. If
257          * more then one inum would fall into the same entry, we use radix_list
258          * to store the additional entries. radix_list is also used to store
259          * entries where two entries have the same inum but different
260          * generations.
261          */
262         struct list_head radix_list;
263         u64 ino;
264         u64 gen;
265         u64 parent_ino;
266         u64 parent_gen;
267         int ret;
268         int need_later_update;
269         int name_len;
270         char name[];
271 };
272 
273 static void inconsistent_snapshot_error(struct send_ctx *sctx,
274                                         enum btrfs_compare_tree_result result,
275                                         const char *what)
276 {
277         const char *result_string;
278 
279         switch (result) {
280         case BTRFS_COMPARE_TREE_NEW:
281                 result_string = "new";
282                 break;
283         case BTRFS_COMPARE_TREE_DELETED:
284                 result_string = "deleted";
285                 break;
286         case BTRFS_COMPARE_TREE_CHANGED:
287                 result_string = "updated";
288                 break;
289         case BTRFS_COMPARE_TREE_SAME:
290                 ASSERT(0);
291                 result_string = "unchanged";
292                 break;
293         default:
294                 ASSERT(0);
295                 result_string = "unexpected";
296         }
297 
298         btrfs_err(sctx->send_root->fs_info,
299                   "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
300                   result_string, what, sctx->cmp_key->objectid,
301                   sctx->send_root->root_key.objectid,
302                   (sctx->parent_root ?
303                    sctx->parent_root->root_key.objectid : 0));
304 }
305 
306 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
307 
308 static struct waiting_dir_move *
309 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
310 
311 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino);
312 
313 static int need_send_hole(struct send_ctx *sctx)
314 {
315         return (sctx->parent_root && !sctx->cur_inode_new &&
316                 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
317                 S_ISREG(sctx->cur_inode_mode));
318 }
319 
320 static void fs_path_reset(struct fs_path *p)
321 {
322         if (p->reversed) {
323                 p->start = p->buf + p->buf_len - 1;
324                 p->end = p->start;
325                 *p->start = 0;
326         } else {
327                 p->start = p->buf;
328                 p->end = p->start;
329                 *p->start = 0;
330         }
331 }
332 
333 static struct fs_path *fs_path_alloc(void)
334 {
335         struct fs_path *p;
336 
337         p = kmalloc(sizeof(*p), GFP_KERNEL);
338         if (!p)
339                 return NULL;
340         p->reversed = 0;
341         p->buf = p->inline_buf;
342         p->buf_len = FS_PATH_INLINE_SIZE;
343         fs_path_reset(p);
344         return p;
345 }
346 
347 static struct fs_path *fs_path_alloc_reversed(void)
348 {
349         struct fs_path *p;
350 
351         p = fs_path_alloc();
352         if (!p)
353                 return NULL;
354         p->reversed = 1;
355         fs_path_reset(p);
356         return p;
357 }
358 
359 static void fs_path_free(struct fs_path *p)
360 {
361         if (!p)
362                 return;
363         if (p->buf != p->inline_buf)
364                 kfree(p->buf);
365         kfree(p);
366 }
367 
368 static int fs_path_len(struct fs_path *p)
369 {
370         return p->end - p->start;
371 }
372 
373 static int fs_path_ensure_buf(struct fs_path *p, int len)
374 {
375         char *tmp_buf;
376         int path_len;
377         int old_buf_len;
378 
379         len++;
380 
381         if (p->buf_len >= len)
382                 return 0;
383 
384         if (len > PATH_MAX) {
385                 WARN_ON(1);
386                 return -ENOMEM;
387         }
388 
389         path_len = p->end - p->start;
390         old_buf_len = p->buf_len;
391 
392         /*
393          * First time the inline_buf does not suffice
394          */
395         if (p->buf == p->inline_buf) {
396                 tmp_buf = kmalloc(len, GFP_KERNEL);
397                 if (tmp_buf)
398                         memcpy(tmp_buf, p->buf, old_buf_len);
399         } else {
400                 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
401         }
402         if (!tmp_buf)
403                 return -ENOMEM;
404         p->buf = tmp_buf;
405         /*
406          * The real size of the buffer is bigger, this will let the fast path
407          * happen most of the time
408          */
409         p->buf_len = ksize(p->buf);
410 
411         if (p->reversed) {
412                 tmp_buf = p->buf + old_buf_len - path_len - 1;
413                 p->end = p->buf + p->buf_len - 1;
414                 p->start = p->end - path_len;
415                 memmove(p->start, tmp_buf, path_len + 1);
416         } else {
417                 p->start = p->buf;
418                 p->end = p->start + path_len;
419         }
420         return 0;
421 }
422 
423 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
424                                    char **prepared)
425 {
426         int ret;
427         int new_len;
428 
429         new_len = p->end - p->start + name_len;
430         if (p->start != p->end)
431                 new_len++;
432         ret = fs_path_ensure_buf(p, new_len);
433         if (ret < 0)
434                 goto out;
435 
436         if (p->reversed) {
437                 if (p->start != p->end)
438                         *--p->start = '/';
439                 p->start -= name_len;
440                 *prepared = p->start;
441         } else {
442                 if (p->start != p->end)
443                         *p->end++ = '/';
444                 *prepared = p->end;
445                 p->end += name_len;
446                 *p->end = 0;
447         }
448 
449 out:
450         return ret;
451 }
452 
453 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
454 {
455         int ret;
456         char *prepared;
457 
458         ret = fs_path_prepare_for_add(p, name_len, &prepared);
459         if (ret < 0)
460                 goto out;
461         memcpy(prepared, name, name_len);
462 
463 out:
464         return ret;
465 }
466 
467 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
468 {
469         int ret;
470         char *prepared;
471 
472         ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
473         if (ret < 0)
474                 goto out;
475         memcpy(prepared, p2->start, p2->end - p2->start);
476 
477 out:
478         return ret;
479 }
480 
481 static int fs_path_add_from_extent_buffer(struct fs_path *p,
482                                           struct extent_buffer *eb,
483                                           unsigned long off, int len)
484 {
485         int ret;
486         char *prepared;
487 
488         ret = fs_path_prepare_for_add(p, len, &prepared);
489         if (ret < 0)
490                 goto out;
491 
492         read_extent_buffer(eb, prepared, off, len);
493 
494 out:
495         return ret;
496 }
497 
498 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
499 {
500         int ret;
501 
502         p->reversed = from->reversed;
503         fs_path_reset(p);
504 
505         ret = fs_path_add_path(p, from);
506 
507         return ret;
508 }
509 
510 
511 static void fs_path_unreverse(struct fs_path *p)
512 {
513         char *tmp;
514         int len;
515 
516         if (!p->reversed)
517                 return;
518 
519         tmp = p->start;
520         len = p->end - p->start;
521         p->start = p->buf;
522         p->end = p->start + len;
523         memmove(p->start, tmp, len + 1);
524         p->reversed = 0;
525 }
526 
527 static struct btrfs_path *alloc_path_for_send(void)
528 {
529         struct btrfs_path *path;
530 
531         path = btrfs_alloc_path();
532         if (!path)
533                 return NULL;
534         path->search_commit_root = 1;
535         path->skip_locking = 1;
536         path->need_commit_sem = 1;
537         return path;
538 }
539 
540 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
541 {
542         int ret;
543         u32 pos = 0;
544 
545         while (pos < len) {
546                 ret = kernel_write(filp, buf + pos, len - pos, off);
547                 /* TODO handle that correctly */
548                 /*if (ret == -ERESTARTSYS) {
549                         continue;
550                 }*/
551                 if (ret < 0)
552                         return ret;
553                 if (ret == 0) {
554                         return -EIO;
555                 }
556                 pos += ret;
557         }
558 
559         return 0;
560 }
561 
562 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
563 {
564         struct btrfs_tlv_header *hdr;
565         int total_len = sizeof(*hdr) + len;
566         int left = sctx->send_max_size - sctx->send_size;
567 
568         if (unlikely(left < total_len))
569                 return -EOVERFLOW;
570 
571         hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
572         hdr->tlv_type = cpu_to_le16(attr);
573         hdr->tlv_len = cpu_to_le16(len);
574         memcpy(hdr + 1, data, len);
575         sctx->send_size += total_len;
576 
577         return 0;
578 }
579 
580 #define TLV_PUT_DEFINE_INT(bits) \
581         static int tlv_put_u##bits(struct send_ctx *sctx,               \
582                         u##bits attr, u##bits value)                    \
583         {                                                               \
584                 __le##bits __tmp = cpu_to_le##bits(value);              \
585                 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));      \
586         }
587 
588 TLV_PUT_DEFINE_INT(64)
589 
590 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
591                           const char *str, int len)
592 {
593         if (len == -1)
594                 len = strlen(str);
595         return tlv_put(sctx, attr, str, len);
596 }
597 
598 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
599                         const u8 *uuid)
600 {
601         return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
602 }
603 
604 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
605                                   struct extent_buffer *eb,
606                                   struct btrfs_timespec *ts)
607 {
608         struct btrfs_timespec bts;
609         read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
610         return tlv_put(sctx, attr, &bts, sizeof(bts));
611 }
612 
613 
614 #define TLV_PUT(sctx, attrtype, attrlen, data) \
615         do { \
616                 ret = tlv_put(sctx, attrtype, attrlen, data); \
617                 if (ret < 0) \
618                         goto tlv_put_failure; \
619         } while (0)
620 
621 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
622         do { \
623                 ret = tlv_put_u##bits(sctx, attrtype, value); \
624                 if (ret < 0) \
625                         goto tlv_put_failure; \
626         } while (0)
627 
628 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
629 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
630 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
631 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
632 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
633         do { \
634                 ret = tlv_put_string(sctx, attrtype, str, len); \
635                 if (ret < 0) \
636                         goto tlv_put_failure; \
637         } while (0)
638 #define TLV_PUT_PATH(sctx, attrtype, p) \
639         do { \
640                 ret = tlv_put_string(sctx, attrtype, p->start, \
641                         p->end - p->start); \
642                 if (ret < 0) \
643                         goto tlv_put_failure; \
644         } while(0)
645 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
646         do { \
647                 ret = tlv_put_uuid(sctx, attrtype, uuid); \
648                 if (ret < 0) \
649                         goto tlv_put_failure; \
650         } while (0)
651 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
652         do { \
653                 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
654                 if (ret < 0) \
655                         goto tlv_put_failure; \
656         } while (0)
657 
658 static int send_header(struct send_ctx *sctx)
659 {
660         struct btrfs_stream_header hdr;
661 
662         strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
663         hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
664 
665         return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
666                                         &sctx->send_off);
667 }
668 
669 /*
670  * For each command/item we want to send to userspace, we call this function.
671  */
672 static int begin_cmd(struct send_ctx *sctx, int cmd)
673 {
674         struct btrfs_cmd_header *hdr;
675 
676         if (WARN_ON(!sctx->send_buf))
677                 return -EINVAL;
678 
679         BUG_ON(sctx->send_size);
680 
681         sctx->send_size += sizeof(*hdr);
682         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
683         hdr->cmd = cpu_to_le16(cmd);
684 
685         return 0;
686 }
687 
688 static int send_cmd(struct send_ctx *sctx)
689 {
690         int ret;
691         struct btrfs_cmd_header *hdr;
692         u32 crc;
693 
694         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
695         hdr->len = cpu_to_le32(sctx->send_size - sizeof(*hdr));
696         hdr->crc = 0;
697 
698         crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
699         hdr->crc = cpu_to_le32(crc);
700 
701         ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
702                                         &sctx->send_off);
703 
704         sctx->total_send_size += sctx->send_size;
705         sctx->cmd_send_size[le16_to_cpu(hdr->cmd)] += sctx->send_size;
706         sctx->send_size = 0;
707 
708         return ret;
709 }
710 
711 /*
712  * Sends a move instruction to user space
713  */
714 static int send_rename(struct send_ctx *sctx,
715                      struct fs_path *from, struct fs_path *to)
716 {
717         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
718         int ret;
719 
720         btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
721 
722         ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
723         if (ret < 0)
724                 goto out;
725 
726         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
727         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
728 
729         ret = send_cmd(sctx);
730 
731 tlv_put_failure:
732 out:
733         return ret;
734 }
735 
736 /*
737  * Sends a link instruction to user space
738  */
739 static int send_link(struct send_ctx *sctx,
740                      struct fs_path *path, struct fs_path *lnk)
741 {
742         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
743         int ret;
744 
745         btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
746 
747         ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
748         if (ret < 0)
749                 goto out;
750 
751         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
752         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
753 
754         ret = send_cmd(sctx);
755 
756 tlv_put_failure:
757 out:
758         return ret;
759 }
760 
761 /*
762  * Sends an unlink instruction to user space
763  */
764 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
765 {
766         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
767         int ret;
768 
769         btrfs_debug(fs_info, "send_unlink %s", path->start);
770 
771         ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
772         if (ret < 0)
773                 goto out;
774 
775         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
776 
777         ret = send_cmd(sctx);
778 
779 tlv_put_failure:
780 out:
781         return ret;
782 }
783 
784 /*
785  * Sends a rmdir instruction to user space
786  */
787 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
788 {
789         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
790         int ret;
791 
792         btrfs_debug(fs_info, "send_rmdir %s", path->start);
793 
794         ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
795         if (ret < 0)
796                 goto out;
797 
798         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
799 
800         ret = send_cmd(sctx);
801 
802 tlv_put_failure:
803 out:
804         return ret;
805 }
806 
807 /*
808  * Helper function to retrieve some fields from an inode item.
809  */
810 static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
811                           u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
812                           u64 *gid, u64 *rdev)
813 {
814         int ret;
815         struct btrfs_inode_item *ii;
816         struct btrfs_key key;
817 
818         key.objectid = ino;
819         key.type = BTRFS_INODE_ITEM_KEY;
820         key.offset = 0;
821         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
822         if (ret) {
823                 if (ret > 0)
824                         ret = -ENOENT;
825                 return ret;
826         }
827 
828         ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
829                         struct btrfs_inode_item);
830         if (size)
831                 *size = btrfs_inode_size(path->nodes[0], ii);
832         if (gen)
833                 *gen = btrfs_inode_generation(path->nodes[0], ii);
834         if (mode)
835                 *mode = btrfs_inode_mode(path->nodes[0], ii);
836         if (uid)
837                 *uid = btrfs_inode_uid(path->nodes[0], ii);
838         if (gid)
839                 *gid = btrfs_inode_gid(path->nodes[0], ii);
840         if (rdev)
841                 *rdev = btrfs_inode_rdev(path->nodes[0], ii);
842 
843         return ret;
844 }
845 
846 static int get_inode_info(struct btrfs_root *root,
847                           u64 ino, u64 *size, u64 *gen,
848                           u64 *mode, u64 *uid, u64 *gid,
849                           u64 *rdev)
850 {
851         struct btrfs_path *path;
852         int ret;
853 
854         path = alloc_path_for_send();
855         if (!path)
856                 return -ENOMEM;
857         ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
858                                rdev);
859         btrfs_free_path(path);
860         return ret;
861 }
862 
863 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
864                                    struct fs_path *p,
865                                    void *ctx);
866 
867 /*
868  * Helper function to iterate the entries in ONE btrfs_inode_ref or
869  * btrfs_inode_extref.
870  * The iterate callback may return a non zero value to stop iteration. This can
871  * be a negative value for error codes or 1 to simply stop it.
872  *
873  * path must point to the INODE_REF or INODE_EXTREF when called.
874  */
875 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
876                              struct btrfs_key *found_key, int resolve,
877                              iterate_inode_ref_t iterate, void *ctx)
878 {
879         struct extent_buffer *eb = path->nodes[0];
880         struct btrfs_item *item;
881         struct btrfs_inode_ref *iref;
882         struct btrfs_inode_extref *extref;
883         struct btrfs_path *tmp_path;
884         struct fs_path *p;
885         u32 cur = 0;
886         u32 total;
887         int slot = path->slots[0];
888         u32 name_len;
889         char *start;
890         int ret = 0;
891         int num = 0;
892         int index;
893         u64 dir;
894         unsigned long name_off;
895         unsigned long elem_size;
896         unsigned long ptr;
897 
898         p = fs_path_alloc_reversed();
899         if (!p)
900                 return -ENOMEM;
901 
902         tmp_path = alloc_path_for_send();
903         if (!tmp_path) {
904                 fs_path_free(p);
905                 return -ENOMEM;
906         }
907 
908 
909         if (found_key->type == BTRFS_INODE_REF_KEY) {
910                 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
911                                                     struct btrfs_inode_ref);
912                 item = btrfs_item_nr(slot);
913                 total = btrfs_item_size(eb, item);
914                 elem_size = sizeof(*iref);
915         } else {
916                 ptr = btrfs_item_ptr_offset(eb, slot);
917                 total = btrfs_item_size_nr(eb, slot);
918                 elem_size = sizeof(*extref);
919         }
920 
921         while (cur < total) {
922                 fs_path_reset(p);
923 
924                 if (found_key->type == BTRFS_INODE_REF_KEY) {
925                         iref = (struct btrfs_inode_ref *)(ptr + cur);
926                         name_len = btrfs_inode_ref_name_len(eb, iref);
927                         name_off = (unsigned long)(iref + 1);
928                         index = btrfs_inode_ref_index(eb, iref);
929                         dir = found_key->offset;
930                 } else {
931                         extref = (struct btrfs_inode_extref *)(ptr + cur);
932                         name_len = btrfs_inode_extref_name_len(eb, extref);
933                         name_off = (unsigned long)&extref->name;
934                         index = btrfs_inode_extref_index(eb, extref);
935                         dir = btrfs_inode_extref_parent(eb, extref);
936                 }
937 
938                 if (resolve) {
939                         start = btrfs_ref_to_path(root, tmp_path, name_len,
940                                                   name_off, eb, dir,
941                                                   p->buf, p->buf_len);
942                         if (IS_ERR(start)) {
943                                 ret = PTR_ERR(start);
944                                 goto out;
945                         }
946                         if (start < p->buf) {
947                                 /* overflow , try again with larger buffer */
948                                 ret = fs_path_ensure_buf(p,
949                                                 p->buf_len + p->buf - start);
950                                 if (ret < 0)
951                                         goto out;
952                                 start = btrfs_ref_to_path(root, tmp_path,
953                                                           name_len, name_off,
954                                                           eb, dir,
955                                                           p->buf, p->buf_len);
956                                 if (IS_ERR(start)) {
957                                         ret = PTR_ERR(start);
958                                         goto out;
959                                 }
960                                 BUG_ON(start < p->buf);
961                         }
962                         p->start = start;
963                 } else {
964                         ret = fs_path_add_from_extent_buffer(p, eb, name_off,
965                                                              name_len);
966                         if (ret < 0)
967                                 goto out;
968                 }
969 
970                 cur += elem_size + name_len;
971                 ret = iterate(num, dir, index, p, ctx);
972                 if (ret)
973                         goto out;
974                 num++;
975         }
976 
977 out:
978         btrfs_free_path(tmp_path);
979         fs_path_free(p);
980         return ret;
981 }
982 
983 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
984                                   const char *name, int name_len,
985                                   const char *data, int data_len,
986                                   u8 type, void *ctx);
987 
988 /*
989  * Helper function to iterate the entries in ONE btrfs_dir_item.
990  * The iterate callback may return a non zero value to stop iteration. This can
991  * be a negative value for error codes or 1 to simply stop it.
992  *
993  * path must point to the dir item when called.
994  */
995 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
996                             iterate_dir_item_t iterate, void *ctx)
997 {
998         int ret = 0;
999         struct extent_buffer *eb;
1000         struct btrfs_item *item;
1001         struct btrfs_dir_item *di;
1002         struct btrfs_key di_key;
1003         char *buf = NULL;
1004         int buf_len;
1005         u32 name_len;
1006         u32 data_len;
1007         u32 cur;
1008         u32 len;
1009         u32 total;
1010         int slot;
1011         int num;
1012         u8 type;
1013 
1014         /*
1015          * Start with a small buffer (1 page). If later we end up needing more
1016          * space, which can happen for xattrs on a fs with a leaf size greater
1017          * then the page size, attempt to increase the buffer. Typically xattr
1018          * values are small.
1019          */
1020         buf_len = PATH_MAX;
1021         buf = kmalloc(buf_len, GFP_KERNEL);
1022         if (!buf) {
1023                 ret = -ENOMEM;
1024                 goto out;
1025         }
1026 
1027         eb = path->nodes[0];
1028         slot = path->slots[0];
1029         item = btrfs_item_nr(slot);
1030         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1031         cur = 0;
1032         len = 0;
1033         total = btrfs_item_size(eb, item);
1034 
1035         num = 0;
1036         while (cur < total) {
1037                 name_len = btrfs_dir_name_len(eb, di);
1038                 data_len = btrfs_dir_data_len(eb, di);
1039                 type = btrfs_dir_type(eb, di);
1040                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1041 
1042                 if (type == BTRFS_FT_XATTR) {
1043                         if (name_len > XATTR_NAME_MAX) {
1044                                 ret = -ENAMETOOLONG;
1045                                 goto out;
1046                         }
1047                         if (name_len + data_len >
1048                                         BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1049                                 ret = -E2BIG;
1050                                 goto out;
1051                         }
1052                 } else {
1053                         /*
1054                          * Path too long
1055                          */
1056                         if (name_len + data_len > PATH_MAX) {
1057                                 ret = -ENAMETOOLONG;
1058                                 goto out;
1059                         }
1060                 }
1061 
1062                 if (name_len + data_len > buf_len) {
1063                         buf_len = name_len + data_len;
1064                         if (is_vmalloc_addr(buf)) {
1065                                 vfree(buf);
1066                                 buf = NULL;
1067                         } else {
1068                                 char *tmp = krealloc(buf, buf_len,
1069                                                 GFP_KERNEL | __GFP_NOWARN);
1070 
1071                                 if (!tmp)
1072                                         kfree(buf);
1073                                 buf = tmp;
1074                         }
1075                         if (!buf) {
1076                                 buf = kvmalloc(buf_len, GFP_KERNEL);
1077                                 if (!buf) {
1078                                         ret = -ENOMEM;
1079                                         goto out;
1080                                 }
1081                         }
1082                 }
1083 
1084                 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1085                                 name_len + data_len);
1086 
1087                 len = sizeof(*di) + name_len + data_len;
1088                 di = (struct btrfs_dir_item *)((char *)di + len);
1089                 cur += len;
1090 
1091                 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1092                                 data_len, type, ctx);
1093                 if (ret < 0)
1094                         goto out;
1095                 if (ret) {
1096                         ret = 0;
1097                         goto out;
1098                 }
1099 
1100                 num++;
1101         }
1102 
1103 out:
1104         kvfree(buf);
1105         return ret;
1106 }
1107 
1108 static int __copy_first_ref(int num, u64 dir, int index,
1109                             struct fs_path *p, void *ctx)
1110 {
1111         int ret;
1112         struct fs_path *pt = ctx;
1113 
1114         ret = fs_path_copy(pt, p);
1115         if (ret < 0)
1116                 return ret;
1117 
1118         /* we want the first only */
1119         return 1;
1120 }
1121 
1122 /*
1123  * Retrieve the first path of an inode. If an inode has more then one
1124  * ref/hardlink, this is ignored.
1125  */
1126 static int get_inode_path(struct btrfs_root *root,
1127                           u64 ino, struct fs_path *path)
1128 {
1129         int ret;
1130         struct btrfs_key key, found_key;
1131         struct btrfs_path *p;
1132 
1133         p = alloc_path_for_send();
1134         if (!p)
1135                 return -ENOMEM;
1136 
1137         fs_path_reset(path);
1138 
1139         key.objectid = ino;
1140         key.type = BTRFS_INODE_REF_KEY;
1141         key.offset = 0;
1142 
1143         ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1144         if (ret < 0)
1145                 goto out;
1146         if (ret) {
1147                 ret = 1;
1148                 goto out;
1149         }
1150         btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1151         if (found_key.objectid != ino ||
1152             (found_key.type != BTRFS_INODE_REF_KEY &&
1153              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1154                 ret = -ENOENT;
1155                 goto out;
1156         }
1157 
1158         ret = iterate_inode_ref(root, p, &found_key, 1,
1159                                 __copy_first_ref, path);
1160         if (ret < 0)
1161                 goto out;
1162         ret = 0;
1163 
1164 out:
1165         btrfs_free_path(p);
1166         return ret;
1167 }
1168 
1169 struct backref_ctx {
1170         struct send_ctx *sctx;
1171 
1172         struct btrfs_path *path;
1173         /* number of total found references */
1174         u64 found;
1175 
1176         /*
1177          * used for clones found in send_root. clones found behind cur_objectid
1178          * and cur_offset are not considered as allowed clones.
1179          */
1180         u64 cur_objectid;
1181         u64 cur_offset;
1182 
1183         /* may be truncated in case it's the last extent in a file */
1184         u64 extent_len;
1185 
1186         /* data offset in the file extent item */
1187         u64 data_offset;
1188 
1189         /* Just to check for bugs in backref resolving */
1190         int found_itself;
1191 };
1192 
1193 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1194 {
1195         u64 root = (u64)(uintptr_t)key;
1196         struct clone_root *cr = (struct clone_root *)elt;
1197 
1198         if (root < cr->root->objectid)
1199                 return -1;
1200         if (root > cr->root->objectid)
1201                 return 1;
1202         return 0;
1203 }
1204 
1205 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1206 {
1207         struct clone_root *cr1 = (struct clone_root *)e1;
1208         struct clone_root *cr2 = (struct clone_root *)e2;
1209 
1210         if (cr1->root->objectid < cr2->root->objectid)
1211                 return -1;
1212         if (cr1->root->objectid > cr2->root->objectid)
1213                 return 1;
1214         return 0;
1215 }
1216 
1217 /*
1218  * Called for every backref that is found for the current extent.
1219  * Results are collected in sctx->clone_roots->ino/offset/found_refs
1220  */
1221 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1222 {
1223         struct backref_ctx *bctx = ctx_;
1224         struct clone_root *found;
1225         int ret;
1226         u64 i_size;
1227 
1228         /* First check if the root is in the list of accepted clone sources */
1229         found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1230                         bctx->sctx->clone_roots_cnt,
1231                         sizeof(struct clone_root),
1232                         __clone_root_cmp_bsearch);
1233         if (!found)
1234                 return 0;
1235 
1236         if (found->root == bctx->sctx->send_root &&
1237             ino == bctx->cur_objectid &&
1238             offset == bctx->cur_offset) {
1239                 bctx->found_itself = 1;
1240         }
1241 
1242         /*
1243          * There are inodes that have extents that lie behind its i_size. Don't
1244          * accept clones from these extents.
1245          */
1246         ret = __get_inode_info(found->root, bctx->path, ino, &i_size, NULL, NULL,
1247                                NULL, NULL, NULL);
1248         btrfs_release_path(bctx->path);
1249         if (ret < 0)
1250                 return ret;
1251 
1252         if (offset + bctx->data_offset + bctx->extent_len > i_size)
1253                 return 0;
1254 
1255         /*
1256          * Make sure we don't consider clones from send_root that are
1257          * behind the current inode/offset.
1258          */
1259         if (found->root == bctx->sctx->send_root) {
1260                 /*
1261                  * TODO for the moment we don't accept clones from the inode
1262                  * that is currently send. We may change this when
1263                  * BTRFS_IOC_CLONE_RANGE supports cloning from and to the same
1264                  * file.
1265                  */
1266                 if (ino >= bctx->cur_objectid)
1267                         return 0;
1268         }
1269 
1270         bctx->found++;
1271         found->found_refs++;
1272         if (ino < found->ino) {
1273                 found->ino = ino;
1274                 found->offset = offset;
1275         } else if (found->ino == ino) {
1276                 /*
1277                  * same extent found more then once in the same file.
1278                  */
1279                 if (found->offset > offset + bctx->extent_len)
1280                         found->offset = offset;
1281         }
1282 
1283         return 0;
1284 }
1285 
1286 /*
1287  * Given an inode, offset and extent item, it finds a good clone for a clone
1288  * instruction. Returns -ENOENT when none could be found. The function makes
1289  * sure that the returned clone is usable at the point where sending is at the
1290  * moment. This means, that no clones are accepted which lie behind the current
1291  * inode+offset.
1292  *
1293  * path must point to the extent item when called.
1294  */
1295 static int find_extent_clone(struct send_ctx *sctx,
1296                              struct btrfs_path *path,
1297                              u64 ino, u64 data_offset,
1298                              u64 ino_size,
1299                              struct clone_root **found)
1300 {
1301         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1302         int ret;
1303         int extent_type;
1304         u64 logical;
1305         u64 disk_byte;
1306         u64 num_bytes;
1307         u64 extent_item_pos;
1308         u64 flags = 0;
1309         struct btrfs_file_extent_item *fi;
1310         struct extent_buffer *eb = path->nodes[0];
1311         struct backref_ctx *backref_ctx = NULL;
1312         struct clone_root *cur_clone_root;
1313         struct btrfs_key found_key;
1314         struct btrfs_path *tmp_path;
1315         int compressed;
1316         u32 i;
1317 
1318         tmp_path = alloc_path_for_send();
1319         if (!tmp_path)
1320                 return -ENOMEM;
1321 
1322         /* We only use this path under the commit sem */
1323         tmp_path->need_commit_sem = 0;
1324 
1325         backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1326         if (!backref_ctx) {
1327                 ret = -ENOMEM;
1328                 goto out;
1329         }
1330 
1331         backref_ctx->path = tmp_path;
1332 
1333         if (data_offset >= ino_size) {
1334                 /*
1335                  * There may be extents that lie behind the file's size.
1336                  * I at least had this in combination with snapshotting while
1337                  * writing large files.
1338                  */
1339                 ret = 0;
1340                 goto out;
1341         }
1342 
1343         fi = btrfs_item_ptr(eb, path->slots[0],
1344                         struct btrfs_file_extent_item);
1345         extent_type = btrfs_file_extent_type(eb, fi);
1346         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1347                 ret = -ENOENT;
1348                 goto out;
1349         }
1350         compressed = btrfs_file_extent_compression(eb, fi);
1351 
1352         num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1353         disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1354         if (disk_byte == 0) {
1355                 ret = -ENOENT;
1356                 goto out;
1357         }
1358         logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1359 
1360         down_read(&fs_info->commit_root_sem);
1361         ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1362                                   &found_key, &flags);
1363         up_read(&fs_info->commit_root_sem);
1364         btrfs_release_path(tmp_path);
1365 
1366         if (ret < 0)
1367                 goto out;
1368         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1369                 ret = -EIO;
1370                 goto out;
1371         }
1372 
1373         /*
1374          * Setup the clone roots.
1375          */
1376         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1377                 cur_clone_root = sctx->clone_roots + i;
1378                 cur_clone_root->ino = (u64)-1;
1379                 cur_clone_root->offset = 0;
1380                 cur_clone_root->found_refs = 0;
1381         }
1382 
1383         backref_ctx->sctx = sctx;
1384         backref_ctx->found = 0;
1385         backref_ctx->cur_objectid = ino;
1386         backref_ctx->cur_offset = data_offset;
1387         backref_ctx->found_itself = 0;
1388         backref_ctx->extent_len = num_bytes;
1389         /*
1390          * For non-compressed extents iterate_extent_inodes() gives us extent
1391          * offsets that already take into account the data offset, but not for
1392          * compressed extents, since the offset is logical and not relative to
1393          * the physical extent locations. We must take this into account to
1394          * avoid sending clone offsets that go beyond the source file's size,
1395          * which would result in the clone ioctl failing with -EINVAL on the
1396          * receiving end.
1397          */
1398         if (compressed == BTRFS_COMPRESS_NONE)
1399                 backref_ctx->data_offset = 0;
1400         else
1401                 backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1402 
1403         /*
1404          * The last extent of a file may be too large due to page alignment.
1405          * We need to adjust extent_len in this case so that the checks in
1406          * __iterate_backrefs work.
1407          */
1408         if (data_offset + num_bytes >= ino_size)
1409                 backref_ctx->extent_len = ino_size - data_offset;
1410 
1411         /*
1412          * Now collect all backrefs.
1413          */
1414         if (compressed == BTRFS_COMPRESS_NONE)
1415                 extent_item_pos = logical - found_key.objectid;
1416         else
1417                 extent_item_pos = 0;
1418         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1419                                     extent_item_pos, 1, __iterate_backrefs,
1420                                     backref_ctx, false);
1421 
1422         if (ret < 0)
1423                 goto out;
1424 
1425         if (!backref_ctx->found_itself) {
1426                 /* found a bug in backref code? */
1427                 ret = -EIO;
1428                 btrfs_err(fs_info,
1429                           "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1430                           ino, data_offset, disk_byte, found_key.objectid);
1431                 goto out;
1432         }
1433 
1434         btrfs_debug(fs_info,
1435                     "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1436                     data_offset, ino, num_bytes, logical);
1437 
1438         if (!backref_ctx->found)
1439                 btrfs_debug(fs_info, "no clones found");
1440 
1441         cur_clone_root = NULL;
1442         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1443                 if (sctx->clone_roots[i].found_refs) {
1444                         if (!cur_clone_root)
1445                                 cur_clone_root = sctx->clone_roots + i;
1446                         else if (sctx->clone_roots[i].root == sctx->send_root)
1447                                 /* prefer clones from send_root over others */
1448                                 cur_clone_root = sctx->clone_roots + i;
1449                 }
1450 
1451         }
1452 
1453         if (cur_clone_root) {
1454                 *found = cur_clone_root;
1455                 ret = 0;
1456         } else {
1457                 ret = -ENOENT;
1458         }
1459 
1460 out:
1461         btrfs_free_path(tmp_path);
1462         kfree(backref_ctx);
1463         return ret;
1464 }
1465 
1466 static int read_symlink(struct btrfs_root *root,
1467                         u64 ino,
1468                         struct fs_path *dest)
1469 {
1470         int ret;
1471         struct btrfs_path *path;
1472         struct btrfs_key key;
1473         struct btrfs_file_extent_item *ei;
1474         u8 type;
1475         u8 compression;
1476         unsigned long off;
1477         int len;
1478 
1479         path = alloc_path_for_send();
1480         if (!path)
1481                 return -ENOMEM;
1482 
1483         key.objectid = ino;
1484         key.type = BTRFS_EXTENT_DATA_KEY;
1485         key.offset = 0;
1486         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1487         if (ret < 0)
1488                 goto out;
1489         if (ret) {
1490                 /*
1491                  * An empty symlink inode. Can happen in rare error paths when
1492                  * creating a symlink (transaction committed before the inode
1493                  * eviction handler removed the symlink inode items and a crash
1494                  * happened in between or the subvol was snapshoted in between).
1495                  * Print an informative message to dmesg/syslog so that the user
1496                  * can delete the symlink.
1497                  */
1498                 btrfs_err(root->fs_info,
1499                           "Found empty symlink inode %llu at root %llu",
1500                           ino, root->root_key.objectid);
1501                 ret = -EIO;
1502                 goto out;
1503         }
1504 
1505         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1506                         struct btrfs_file_extent_item);
1507         type = btrfs_file_extent_type(path->nodes[0], ei);
1508         compression = btrfs_file_extent_compression(path->nodes[0], ei);
1509         BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1510         BUG_ON(compression);
1511 
1512         off = btrfs_file_extent_inline_start(ei);
1513         len = btrfs_file_extent_inline_len(path->nodes[0], path->slots[0], ei);
1514 
1515         ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1516 
1517 out:
1518         btrfs_free_path(path);
1519         return ret;
1520 }
1521 
1522 /*
1523  * Helper function to generate a file name that is unique in the root of
1524  * send_root and parent_root. This is used to generate names for orphan inodes.
1525  */
1526 static int gen_unique_name(struct send_ctx *sctx,
1527                            u64 ino, u64 gen,
1528                            struct fs_path *dest)
1529 {
1530         int ret = 0;
1531         struct btrfs_path *path;
1532         struct btrfs_dir_item *di;
1533         char tmp[64];
1534         int len;
1535         u64 idx = 0;
1536 
1537         path = alloc_path_for_send();
1538         if (!path)
1539                 return -ENOMEM;
1540 
1541         while (1) {
1542                 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1543                                 ino, gen, idx);
1544                 ASSERT(len < sizeof(tmp));
1545 
1546                 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1547                                 path, BTRFS_FIRST_FREE_OBJECTID,
1548                                 tmp, strlen(tmp), 0);
1549                 btrfs_release_path(path);
1550                 if (IS_ERR(di)) {
1551                         ret = PTR_ERR(di);
1552                         goto out;
1553                 }
1554                 if (di) {
1555                         /* not unique, try again */
1556                         idx++;
1557                         continue;
1558                 }
1559 
1560                 if (!sctx->parent_root) {
1561                         /* unique */
1562                         ret = 0;
1563                         break;
1564                 }
1565 
1566                 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1567                                 path, BTRFS_FIRST_FREE_OBJECTID,
1568                                 tmp, strlen(tmp), 0);
1569                 btrfs_release_path(path);
1570                 if (IS_ERR(di)) {
1571                         ret = PTR_ERR(di);
1572                         goto out;
1573                 }
1574                 if (di) {
1575                         /* not unique, try again */
1576                         idx++;
1577                         continue;
1578                 }
1579                 /* unique */
1580                 break;
1581         }
1582 
1583         ret = fs_path_add(dest, tmp, strlen(tmp));
1584 
1585 out:
1586         btrfs_free_path(path);
1587         return ret;
1588 }
1589 
1590 enum inode_state {
1591         inode_state_no_change,
1592         inode_state_will_create,
1593         inode_state_did_create,
1594         inode_state_will_delete,
1595         inode_state_did_delete,
1596 };
1597 
1598 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1599 {
1600         int ret;
1601         int left_ret;
1602         int right_ret;
1603         u64 left_gen;
1604         u64 right_gen;
1605 
1606         ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1607                         NULL, NULL);
1608         if (ret < 0 && ret != -ENOENT)
1609                 goto out;
1610         left_ret = ret;
1611 
1612         if (!sctx->parent_root) {
1613                 right_ret = -ENOENT;
1614         } else {
1615                 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1616                                 NULL, NULL, NULL, NULL);
1617                 if (ret < 0 && ret != -ENOENT)
1618                         goto out;
1619                 right_ret = ret;
1620         }
1621 
1622         if (!left_ret && !right_ret) {
1623                 if (left_gen == gen && right_gen == gen) {
1624                         ret = inode_state_no_change;
1625                 } else if (left_gen == gen) {
1626                         if (ino < sctx->send_progress)
1627                                 ret = inode_state_did_create;
1628                         else
1629                                 ret = inode_state_will_create;
1630                 } else if (right_gen == gen) {
1631                         if (ino < sctx->send_progress)
1632                                 ret = inode_state_did_delete;
1633                         else
1634                                 ret = inode_state_will_delete;
1635                 } else  {
1636                         ret = -ENOENT;
1637                 }
1638         } else if (!left_ret) {
1639                 if (left_gen == gen) {
1640                         if (ino < sctx->send_progress)
1641                                 ret = inode_state_did_create;
1642                         else
1643                                 ret = inode_state_will_create;
1644                 } else {
1645                         ret = -ENOENT;
1646                 }
1647         } else if (!right_ret) {
1648                 if (right_gen == gen) {
1649                         if (ino < sctx->send_progress)
1650                                 ret = inode_state_did_delete;
1651                         else
1652                                 ret = inode_state_will_delete;
1653                 } else {
1654                         ret = -ENOENT;
1655                 }
1656         } else {
1657                 ret = -ENOENT;
1658         }
1659 
1660 out:
1661         return ret;
1662 }
1663 
1664 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1665 {
1666         int ret;
1667 
1668         if (ino == BTRFS_FIRST_FREE_OBJECTID)
1669                 return 1;
1670 
1671         ret = get_cur_inode_state(sctx, ino, gen);
1672         if (ret < 0)
1673                 goto out;
1674 
1675         if (ret == inode_state_no_change ||
1676             ret == inode_state_did_create ||
1677             ret == inode_state_will_delete)
1678                 ret = 1;
1679         else
1680                 ret = 0;
1681 
1682 out:
1683         return ret;
1684 }
1685 
1686 /*
1687  * Helper function to lookup a dir item in a dir.
1688  */
1689 static int lookup_dir_item_inode(struct btrfs_root *root,
1690                                  u64 dir, const char *name, int name_len,
1691                                  u64 *found_inode,
1692                                  u8 *found_type)
1693 {
1694         int ret = 0;
1695         struct btrfs_dir_item *di;
1696         struct btrfs_key key;
1697         struct btrfs_path *path;
1698 
1699         path = alloc_path_for_send();
1700         if (!path)
1701                 return -ENOMEM;
1702 
1703         di = btrfs_lookup_dir_item(NULL, root, path,
1704                         dir, name, name_len, 0);
1705         if (!di) {
1706                 ret = -ENOENT;
1707                 goto out;
1708         }
1709         if (IS_ERR(di)) {
1710                 ret = PTR_ERR(di);
1711                 goto out;
1712         }
1713         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1714         if (key.type == BTRFS_ROOT_ITEM_KEY) {
1715                 ret = -ENOENT;
1716                 goto out;
1717         }
1718         *found_inode = key.objectid;
1719         *found_type = btrfs_dir_type(path->nodes[0], di);
1720 
1721 out:
1722         btrfs_free_path(path);
1723         return ret;
1724 }
1725 
1726 /*
1727  * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1728  * generation of the parent dir and the name of the dir entry.
1729  */
1730 static int get_first_ref(struct btrfs_root *root, u64 ino,
1731                          u64 *dir, u64 *dir_gen, struct fs_path *name)
1732 {
1733         int ret;
1734         struct btrfs_key key;
1735         struct btrfs_key found_key;
1736         struct btrfs_path *path;
1737         int len;
1738         u64 parent_dir;
1739 
1740         path = alloc_path_for_send();
1741         if (!path)
1742                 return -ENOMEM;
1743 
1744         key.objectid = ino;
1745         key.type = BTRFS_INODE_REF_KEY;
1746         key.offset = 0;
1747 
1748         ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1749         if (ret < 0)
1750                 goto out;
1751         if (!ret)
1752                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1753                                 path->slots[0]);
1754         if (ret || found_key.objectid != ino ||
1755             (found_key.type != BTRFS_INODE_REF_KEY &&
1756              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1757                 ret = -ENOENT;
1758                 goto out;
1759         }
1760 
1761         if (found_key.type == BTRFS_INODE_REF_KEY) {
1762                 struct btrfs_inode_ref *iref;
1763                 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1764                                       struct btrfs_inode_ref);
1765                 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1766                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1767                                                      (unsigned long)(iref + 1),
1768                                                      len);
1769                 parent_dir = found_key.offset;
1770         } else {
1771                 struct btrfs_inode_extref *extref;
1772                 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1773                                         struct btrfs_inode_extref);
1774                 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1775                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1776                                         (unsigned long)&extref->name, len);
1777                 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1778         }
1779         if (ret < 0)
1780                 goto out;
1781         btrfs_release_path(path);
1782 
1783         if (dir_gen) {
1784                 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1785                                      NULL, NULL, NULL);
1786                 if (ret < 0)
1787                         goto out;
1788         }
1789 
1790         *dir = parent_dir;
1791 
1792 out:
1793         btrfs_free_path(path);
1794         return ret;
1795 }
1796 
1797 static int is_first_ref(struct btrfs_root *root,
1798                         u64 ino, u64 dir,
1799                         const char *name, int name_len)
1800 {
1801         int ret;
1802         struct fs_path *tmp_name;
1803         u64 tmp_dir;
1804 
1805         tmp_name = fs_path_alloc();
1806         if (!tmp_name)
1807                 return -ENOMEM;
1808 
1809         ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1810         if (ret < 0)
1811                 goto out;
1812 
1813         if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1814                 ret = 0;
1815                 goto out;
1816         }
1817 
1818         ret = !memcmp(tmp_name->start, name, name_len);
1819 
1820 out:
1821         fs_path_free(tmp_name);
1822         return ret;
1823 }
1824 
1825 /*
1826  * Used by process_recorded_refs to determine if a new ref would overwrite an
1827  * already existing ref. In case it detects an overwrite, it returns the
1828  * inode/gen in who_ino/who_gen.
1829  * When an overwrite is detected, process_recorded_refs does proper orphanizing
1830  * to make sure later references to the overwritten inode are possible.
1831  * Orphanizing is however only required for the first ref of an inode.
1832  * process_recorded_refs does an additional is_first_ref check to see if
1833  * orphanizing is really required.
1834  */
1835 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1836                               const char *name, int name_len,
1837                               u64 *who_ino, u64 *who_gen, u64 *who_mode)
1838 {
1839         int ret = 0;
1840         u64 gen;
1841         u64 other_inode = 0;
1842         u8 other_type = 0;
1843 
1844         if (!sctx->parent_root)
1845                 goto out;
1846 
1847         ret = is_inode_existent(sctx, dir, dir_gen);
1848         if (ret <= 0)
1849                 goto out;
1850 
1851         /*
1852          * If we have a parent root we need to verify that the parent dir was
1853          * not deleted and then re-created, if it was then we have no overwrite
1854          * and we can just unlink this entry.
1855          */
1856         if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1857                 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1858                                      NULL, NULL, NULL);
1859                 if (ret < 0 && ret != -ENOENT)
1860                         goto out;
1861                 if (ret) {
1862                         ret = 0;
1863                         goto out;
1864                 }
1865                 if (gen != dir_gen)
1866                         goto out;
1867         }
1868 
1869         ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1870                         &other_inode, &other_type);
1871         if (ret < 0 && ret != -ENOENT)
1872                 goto out;
1873         if (ret) {
1874                 ret = 0;
1875                 goto out;
1876         }
1877 
1878         /*
1879          * Check if the overwritten ref was already processed. If yes, the ref
1880          * was already unlinked/moved, so we can safely assume that we will not
1881          * overwrite anything at this point in time.
1882          */
1883         if (other_inode > sctx->send_progress ||
1884             is_waiting_for_move(sctx, other_inode)) {
1885                 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1886                                 who_gen, who_mode, NULL, NULL, NULL);
1887                 if (ret < 0)
1888                         goto out;
1889 
1890                 ret = 1;
1891                 *who_ino = other_inode;
1892         } else {
1893                 ret = 0;
1894         }
1895 
1896 out:
1897         return ret;
1898 }
1899 
1900 /*
1901  * Checks if the ref was overwritten by an already processed inode. This is
1902  * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1903  * thus the orphan name needs be used.
1904  * process_recorded_refs also uses it to avoid unlinking of refs that were
1905  * overwritten.
1906  */
1907 static int did_overwrite_ref(struct send_ctx *sctx,
1908                             u64 dir, u64 dir_gen,
1909                             u64 ino, u64 ino_gen,
1910                             const char *name, int name_len)
1911 {
1912         int ret = 0;
1913         u64 gen;
1914         u64 ow_inode;
1915         u8 other_type;
1916 
1917         if (!sctx->parent_root)
1918                 goto out;
1919 
1920         ret = is_inode_existent(sctx, dir, dir_gen);
1921         if (ret <= 0)
1922                 goto out;
1923 
1924         if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1925                 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1926                                      NULL, NULL, NULL);
1927                 if (ret < 0 && ret != -ENOENT)
1928                         goto out;
1929                 if (ret) {
1930                         ret = 0;
1931                         goto out;
1932                 }
1933                 if (gen != dir_gen)
1934                         goto out;
1935         }
1936 
1937         /* check if the ref was overwritten by another ref */
1938         ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1939                         &ow_inode, &other_type);
1940         if (ret < 0 && ret != -ENOENT)
1941                 goto out;
1942         if (ret) {
1943                 /* was never and will never be overwritten */
1944                 ret = 0;
1945                 goto out;
1946         }
1947 
1948         ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1949                         NULL, NULL);
1950         if (ret < 0)
1951                 goto out;
1952 
1953         if (ow_inode == ino && gen == ino_gen) {
1954                 ret = 0;
1955                 goto out;
1956         }
1957 
1958         /*
1959          * We know that it is or will be overwritten. Check this now.
1960          * The current inode being processed might have been the one that caused
1961          * inode 'ino' to be orphanized, therefore check if ow_inode matches
1962          * the current inode being processed.
1963          */
1964         if ((ow_inode < sctx->send_progress) ||
1965             (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1966              gen == sctx->cur_inode_gen))
1967                 ret = 1;
1968         else
1969                 ret = 0;
1970 
1971 out:
1972         return ret;
1973 }
1974 
1975 /*
1976  * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1977  * that got overwritten. This is used by process_recorded_refs to determine
1978  * if it has to use the path as returned by get_cur_path or the orphan name.
1979  */
1980 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1981 {
1982         int ret = 0;
1983         struct fs_path *name = NULL;
1984         u64 dir;
1985         u64 dir_gen;
1986 
1987         if (!sctx->parent_root)
1988                 goto out;
1989 
1990         name = fs_path_alloc();
1991         if (!name)
1992                 return -ENOMEM;
1993 
1994         ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
1995         if (ret < 0)
1996                 goto out;
1997 
1998         ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
1999                         name->start, fs_path_len(name));
2000 
2001 out:
2002         fs_path_free(name);
2003         return ret;
2004 }
2005 
2006 /*
2007  * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
2008  * so we need to do some special handling in case we have clashes. This function
2009  * takes care of this with the help of name_cache_entry::radix_list.
2010  * In case of error, nce is kfreed.
2011  */
2012 static int name_cache_insert(struct send_ctx *sctx,
2013                              struct name_cache_entry *nce)
2014 {
2015         int ret = 0;
2016         struct list_head *nce_head;
2017 
2018         nce_head = radix_tree_lookup(&sctx->name_cache,
2019                         (unsigned long)nce->ino);
2020         if (!nce_head) {
2021                 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2022                 if (!nce_head) {
2023                         kfree(nce);
2024                         return -ENOMEM;
2025                 }
2026                 INIT_LIST_HEAD(nce_head);
2027 
2028                 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2029                 if (ret < 0) {
2030                         kfree(nce_head);
2031                         kfree(nce);
2032                         return ret;
2033                 }
2034         }
2035         list_add_tail(&nce->radix_list, nce_head);
2036         list_add_tail(&nce->list, &sctx->name_cache_list);
2037         sctx->name_cache_size++;
2038 
2039         return ret;
2040 }
2041 
2042 static void name_cache_delete(struct send_ctx *sctx,
2043                               struct name_cache_entry *nce)
2044 {
2045         struct list_head *nce_head;
2046 
2047         nce_head = radix_tree_lookup(&sctx->name_cache,
2048                         (unsigned long)nce->ino);
2049         if (!nce_head) {
2050                 btrfs_err(sctx->send_root->fs_info,
2051               "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2052                         nce->ino, sctx->name_cache_size);
2053         }
2054 
2055         list_del(&nce->radix_list);
2056         list_del(&nce->list);
2057         sctx->name_cache_size--;
2058 
2059         /*
2060          * We may not get to the final release of nce_head if the lookup fails
2061          */
2062         if (nce_head && list_empty(nce_head)) {
2063                 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2064                 kfree(nce_head);
2065         }
2066 }
2067 
2068 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2069                                                     u64 ino, u64 gen)
2070 {
2071         struct list_head *nce_head;
2072         struct name_cache_entry *cur;
2073 
2074         nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2075         if (!nce_head)
2076                 return NULL;
2077 
2078         list_for_each_entry(cur, nce_head, radix_list) {
2079                 if (cur->ino == ino && cur->gen == gen)
2080                         return cur;
2081         }
2082         return NULL;
2083 }
2084 
2085 /*
2086  * Removes the entry from the list and adds it back to the end. This marks the
2087  * entry as recently used so that name_cache_clean_unused does not remove it.
2088  */
2089 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2090 {
2091         list_del(&nce->list);
2092         list_add_tail(&nce->list, &sctx->name_cache_list);
2093 }
2094 
2095 /*
2096  * Remove some entries from the beginning of name_cache_list.
2097  */
2098 static void name_cache_clean_unused(struct send_ctx *sctx)
2099 {
2100         struct name_cache_entry *nce;
2101 
2102         if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2103                 return;
2104 
2105         while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2106                 nce = list_entry(sctx->name_cache_list.next,
2107                                 struct name_cache_entry, list);
2108                 name_cache_delete(sctx, nce);
2109                 kfree(nce);
2110         }
2111 }
2112 
2113 static void name_cache_free(struct send_ctx *sctx)
2114 {
2115         struct name_cache_entry *nce;
2116 
2117         while (!list_empty(&sctx->name_cache_list)) {
2118                 nce = list_entry(sctx->name_cache_list.next,
2119                                 struct name_cache_entry, list);
2120                 name_cache_delete(sctx, nce);
2121                 kfree(nce);
2122         }
2123 }
2124 
2125 /*
2126  * Used by get_cur_path for each ref up to the root.
2127  * Returns 0 if it succeeded.
2128  * Returns 1 if the inode is not existent or got overwritten. In that case, the
2129  * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2130  * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2131  * Returns <0 in case of error.
2132  */
2133 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2134                                      u64 ino, u64 gen,
2135                                      u64 *parent_ino,
2136                                      u64 *parent_gen,
2137                                      struct fs_path *dest)
2138 {
2139         int ret;
2140         int nce_ret;
2141         struct name_cache_entry *nce = NULL;
2142 
2143         /*
2144          * First check if we already did a call to this function with the same
2145          * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2146          * return the cached result.
2147          */
2148         nce = name_cache_search(sctx, ino, gen);
2149         if (nce) {
2150                 if (ino < sctx->send_progress && nce->need_later_update) {
2151                         name_cache_delete(sctx, nce);
2152                         kfree(nce);
2153                         nce = NULL;
2154                 } else {
2155                         name_cache_used(sctx, nce);
2156                         *parent_ino = nce->parent_ino;
2157                         *parent_gen = nce->parent_gen;
2158                         ret = fs_path_add(dest, nce->name, nce->name_len);
2159                         if (ret < 0)
2160                                 goto out;
2161                         ret = nce->ret;
2162                         goto out;
2163                 }
2164         }
2165 
2166         /*
2167          * If the inode is not existent yet, add the orphan name and return 1.
2168          * This should only happen for the parent dir that we determine in
2169          * __record_new_ref
2170          */
2171         ret = is_inode_existent(sctx, ino, gen);
2172         if (ret < 0)
2173                 goto out;
2174 
2175         if (!ret) {
2176                 ret = gen_unique_name(sctx, ino, gen, dest);
2177                 if (ret < 0)
2178                         goto out;
2179                 ret = 1;
2180                 goto out_cache;
2181         }
2182 
2183         /*
2184          * Depending on whether the inode was already processed or not, use
2185          * send_root or parent_root for ref lookup.
2186          */
2187         if (ino < sctx->send_progress)
2188                 ret = get_first_ref(sctx->send_root, ino,
2189                                     parent_ino, parent_gen, dest);
2190         else
2191                 ret = get_first_ref(sctx->parent_root, ino,
2192                                     parent_ino, parent_gen, dest);
2193         if (ret < 0)
2194                 goto out;
2195 
2196         /*
2197          * Check if the ref was overwritten by an inode's ref that was processed
2198          * earlier. If yes, treat as orphan and return 1.
2199          */
2200         ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2201                         dest->start, dest->end - dest->start);
2202         if (ret < 0)
2203                 goto out;
2204         if (ret) {
2205                 fs_path_reset(dest);
2206                 ret = gen_unique_name(sctx, ino, gen, dest);
2207                 if (ret < 0)
2208                         goto out;
2209                 ret = 1;
2210         }
2211 
2212 out_cache:
2213         /*
2214          * Store the result of the lookup in the name cache.
2215          */
2216         nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2217         if (!nce) {
2218                 ret = -ENOMEM;
2219                 goto out;
2220         }
2221 
2222         nce->ino = ino;
2223         nce->gen = gen;
2224         nce->parent_ino = *parent_ino;
2225         nce->parent_gen = *parent_gen;
2226         nce->name_len = fs_path_len(dest);
2227         nce->ret = ret;
2228         strcpy(nce->name, dest->start);
2229 
2230         if (ino < sctx->send_progress)
2231                 nce->need_later_update = 0;
2232         else
2233                 nce->need_later_update = 1;
2234 
2235         nce_ret = name_cache_insert(sctx, nce);
2236         if (nce_ret < 0)
2237                 ret = nce_ret;
2238         name_cache_clean_unused(sctx);
2239 
2240 out:
2241         return ret;
2242 }
2243 
2244 /*
2245  * Magic happens here. This function returns the first ref to an inode as it
2246  * would look like while receiving the stream at this point in time.
2247  * We walk the path up to the root. For every inode in between, we check if it
2248  * was already processed/sent. If yes, we continue with the parent as found
2249  * in send_root. If not, we continue with the parent as found in parent_root.
2250  * If we encounter an inode that was deleted at this point in time, we use the
2251  * inodes "orphan" name instead of the real name and stop. Same with new inodes
2252  * that were not created yet and overwritten inodes/refs.
2253  *
2254  * When do we have have orphan inodes:
2255  * 1. When an inode is freshly created and thus no valid refs are available yet
2256  * 2. When a directory lost all it's refs (deleted) but still has dir items
2257  *    inside which were not processed yet (pending for move/delete). If anyone
2258  *    tried to get the path to the dir items, it would get a path inside that
2259  *    orphan directory.
2260  * 3. When an inode is moved around or gets new links, it may overwrite the ref
2261  *    of an unprocessed inode. If in that case the first ref would be
2262  *    overwritten, the overwritten inode gets "orphanized". Later when we
2263  *    process this overwritten inode, it is restored at a new place by moving
2264  *    the orphan inode.
2265  *
2266  * sctx->send_progress tells this function at which point in time receiving
2267  * would be.
2268  */
2269 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2270                         struct fs_path *dest)
2271 {
2272         int ret = 0;
2273         struct fs_path *name = NULL;
2274         u64 parent_inode = 0;
2275         u64 parent_gen = 0;
2276         int stop = 0;
2277 
2278         name = fs_path_alloc();
2279         if (!name) {
2280                 ret = -ENOMEM;
2281                 goto out;
2282         }
2283 
2284         dest->reversed = 1;
2285         fs_path_reset(dest);
2286 
2287         while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2288                 struct waiting_dir_move *wdm;
2289 
2290                 fs_path_reset(name);
2291 
2292                 if (is_waiting_for_rm(sctx, ino)) {
2293                         ret = gen_unique_name(sctx, ino, gen, name);
2294                         if (ret < 0)
2295                                 goto out;
2296                         ret = fs_path_add_path(dest, name);
2297                         break;
2298                 }
2299 
2300                 wdm = get_waiting_dir_move(sctx, ino);
2301                 if (wdm && wdm->orphanized) {
2302                         ret = gen_unique_name(sctx, ino, gen, name);
2303                         stop = 1;
2304                 } else if (wdm) {
2305                         ret = get_first_ref(sctx->parent_root, ino,
2306                                             &parent_inode, &parent_gen, name);
2307                 } else {
2308                         ret = __get_cur_name_and_parent(sctx, ino, gen,
2309                                                         &parent_inode,
2310                                                         &parent_gen, name);
2311                         if (ret)
2312                                 stop = 1;
2313                 }
2314 
2315                 if (ret < 0)
2316                         goto out;
2317 
2318                 ret = fs_path_add_path(dest, name);
2319                 if (ret < 0)
2320                         goto out;
2321 
2322                 ino = parent_inode;
2323                 gen = parent_gen;
2324         }
2325 
2326 out:
2327         fs_path_free(name);
2328         if (!ret)
2329                 fs_path_unreverse(dest);
2330         return ret;
2331 }
2332 
2333 /*
2334  * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2335  */
2336 static int send_subvol_begin(struct send_ctx *sctx)
2337 {
2338         int ret;
2339         struct btrfs_root *send_root = sctx->send_root;
2340         struct btrfs_root *parent_root = sctx->parent_root;
2341         struct btrfs_path *path;
2342         struct btrfs_key key;
2343         struct btrfs_root_ref *ref;
2344         struct extent_buffer *leaf;
2345         char *name = NULL;
2346         int namelen;
2347 
2348         path = btrfs_alloc_path();
2349         if (!path)
2350                 return -ENOMEM;
2351 
2352         name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2353         if (!name) {
2354                 btrfs_free_path(path);
2355                 return -ENOMEM;
2356         }
2357 
2358         key.objectid = send_root->objectid;
2359         key.type = BTRFS_ROOT_BACKREF_KEY;
2360         key.offset = 0;
2361 
2362         ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2363                                 &key, path, 1, 0);
2364         if (ret < 0)
2365                 goto out;
2366         if (ret) {
2367                 ret = -ENOENT;
2368                 goto out;
2369         }
2370 
2371         leaf = path->nodes[0];
2372         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2373         if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2374             key.objectid != send_root->objectid) {
2375                 ret = -ENOENT;
2376                 goto out;
2377         }
2378         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2379         namelen = btrfs_root_ref_name_len(leaf, ref);
2380         read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2381         btrfs_release_path(path);
2382 
2383         if (parent_root) {
2384                 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2385                 if (ret < 0)
2386                         goto out;
2387         } else {
2388                 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2389                 if (ret < 0)
2390                         goto out;
2391         }
2392 
2393         TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2394 
2395         if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2396                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2397                             sctx->send_root->root_item.received_uuid);
2398         else
2399                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2400                             sctx->send_root->root_item.uuid);
2401 
2402         TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2403                     le64_to_cpu(sctx->send_root->root_item.ctransid));
2404         if (parent_root) {
2405                 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2406                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2407                                      parent_root->root_item.received_uuid);
2408                 else
2409                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2410                                      parent_root->root_item.uuid);
2411                 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2412                             le64_to_cpu(sctx->parent_root->root_item.ctransid));
2413         }
2414 
2415         ret = send_cmd(sctx);
2416 
2417 tlv_put_failure:
2418 out:
2419         btrfs_free_path(path);
2420         kfree(name);
2421         return ret;
2422 }
2423 
2424 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2425 {
2426         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2427         int ret = 0;
2428         struct fs_path *p;
2429 
2430         btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2431 
2432         p = fs_path_alloc();
2433         if (!p)
2434                 return -ENOMEM;
2435 
2436         ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2437         if (ret < 0)
2438                 goto out;
2439 
2440         ret = get_cur_path(sctx, ino, gen, p);
2441         if (ret < 0)
2442                 goto out;
2443         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2444         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2445 
2446         ret = send_cmd(sctx);
2447 
2448 tlv_put_failure:
2449 out:
2450         fs_path_free(p);
2451         return ret;
2452 }
2453 
2454 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2455 {
2456         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2457         int ret = 0;
2458         struct fs_path *p;
2459 
2460         btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2461 
2462         p = fs_path_alloc();
2463         if (!p)
2464                 return -ENOMEM;
2465 
2466         ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2467         if (ret < 0)
2468                 goto out;
2469 
2470         ret = get_cur_path(sctx, ino, gen, p);
2471         if (ret < 0)
2472                 goto out;
2473         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2474         TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2475 
2476         ret = send_cmd(sctx);
2477 
2478 tlv_put_failure:
2479 out:
2480         fs_path_free(p);
2481         return ret;
2482 }
2483 
2484 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2485 {
2486         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2487         int ret = 0;
2488         struct fs_path *p;
2489 
2490         btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2491                     ino, uid, gid);
2492 
2493         p = fs_path_alloc();
2494         if (!p)
2495                 return -ENOMEM;
2496 
2497         ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2498         if (ret < 0)
2499                 goto out;
2500 
2501         ret = get_cur_path(sctx, ino, gen, p);
2502         if (ret < 0)
2503                 goto out;
2504         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2505         TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2506         TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2507 
2508         ret = send_cmd(sctx);
2509 
2510 tlv_put_failure:
2511 out:
2512         fs_path_free(p);
2513         return ret;
2514 }
2515 
2516 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2517 {
2518         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2519         int ret = 0;
2520         struct fs_path *p = NULL;
2521         struct btrfs_inode_item *ii;
2522         struct btrfs_path *path = NULL;
2523         struct extent_buffer *eb;
2524         struct btrfs_key key;
2525         int slot;
2526 
2527         btrfs_debug(fs_info, "send_utimes %llu", ino);
2528 
2529         p = fs_path_alloc();
2530         if (!p)
2531                 return -ENOMEM;
2532 
2533         path = alloc_path_for_send();
2534         if (!path) {
2535                 ret = -ENOMEM;
2536                 goto out;
2537         }
2538 
2539         key.objectid = ino;
2540         key.type = BTRFS_INODE_ITEM_KEY;
2541         key.offset = 0;
2542         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2543         if (ret > 0)
2544                 ret = -ENOENT;
2545         if (ret < 0)
2546                 goto out;
2547 
2548         eb = path->nodes[0];
2549         slot = path->slots[0];
2550         ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2551 
2552         ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2553         if (ret < 0)
2554                 goto out;
2555 
2556         ret = get_cur_path(sctx, ino, gen, p);
2557         if (ret < 0)
2558                 goto out;
2559         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2560         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2561         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2562         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2563         /* TODO Add otime support when the otime patches get into upstream */
2564 
2565         ret = send_cmd(sctx);
2566 
2567 tlv_put_failure:
2568 out:
2569         fs_path_free(p);
2570         btrfs_free_path(path);
2571         return ret;
2572 }
2573 
2574 /*
2575  * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2576  * a valid path yet because we did not process the refs yet. So, the inode
2577  * is created as orphan.
2578  */
2579 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2580 {
2581         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2582         int ret = 0;
2583         struct fs_path *p;
2584         int cmd;
2585         u64 gen;
2586         u64 mode;
2587         u64 rdev;
2588 
2589         btrfs_debug(fs_info, "send_create_inode %llu", ino);
2590 
2591         p = fs_path_alloc();
2592         if (!p)
2593                 return -ENOMEM;
2594 
2595         if (ino != sctx->cur_ino) {
2596                 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2597                                      NULL, NULL, &rdev);
2598                 if (ret < 0)
2599                         goto out;
2600         } else {
2601                 gen = sctx->cur_inode_gen;
2602                 mode = sctx->cur_inode_mode;
2603                 rdev = sctx->cur_inode_rdev;
2604         }
2605 
2606         if (S_ISREG(mode)) {
2607                 cmd = BTRFS_SEND_C_MKFILE;
2608         } else if (S_ISDIR(mode)) {
2609                 cmd = BTRFS_SEND_C_MKDIR;
2610         } else if (S_ISLNK(mode)) {
2611                 cmd = BTRFS_SEND_C_SYMLINK;
2612         } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2613                 cmd = BTRFS_SEND_C_MKNOD;
2614         } else if (S_ISFIFO(mode)) {
2615                 cmd = BTRFS_SEND_C_MKFIFO;
2616         } else if (S_ISSOCK(mode)) {
2617                 cmd = BTRFS_SEND_C_MKSOCK;
2618         } else {
2619                 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2620                                 (int)(mode & S_IFMT));
2621                 ret = -EOPNOTSUPP;
2622                 goto out;
2623         }
2624 
2625         ret = begin_cmd(sctx, cmd);
2626         if (ret < 0)
2627                 goto out;
2628 
2629         ret = gen_unique_name(sctx, ino, gen, p);
2630         if (ret < 0)
2631                 goto out;
2632 
2633         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2634         TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2635 
2636         if (S_ISLNK(mode)) {
2637                 fs_path_reset(p);
2638                 ret = read_symlink(sctx->send_root, ino, p);
2639                 if (ret < 0)
2640                         goto out;
2641                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2642         } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2643                    S_ISFIFO(mode) || S_ISSOCK(mode)) {
2644                 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2645                 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2646         }
2647 
2648         ret = send_cmd(sctx);
2649         if (ret < 0)
2650                 goto out;
2651 
2652 
2653 tlv_put_failure:
2654 out:
2655         fs_path_free(p);
2656         return ret;
2657 }
2658 
2659 /*
2660  * We need some special handling for inodes that get processed before the parent
2661  * directory got created. See process_recorded_refs for details.
2662  * This function does the check if we already created the dir out of order.
2663  */
2664 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2665 {
2666         int ret = 0;
2667         struct btrfs_path *path = NULL;
2668         struct btrfs_key key;
2669         struct btrfs_key found_key;
2670         struct btrfs_key di_key;
2671         struct extent_buffer *eb;
2672         struct btrfs_dir_item *di;
2673         int slot;
2674 
2675         path = alloc_path_for_send();
2676         if (!path) {
2677                 ret = -ENOMEM;
2678                 goto out;
2679         }
2680 
2681         key.objectid = dir;
2682         key.type = BTRFS_DIR_INDEX_KEY;
2683         key.offset = 0;
2684         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2685         if (ret < 0)
2686                 goto out;
2687 
2688         while (1) {
2689                 eb = path->nodes[0];
2690                 slot = path->slots[0];
2691                 if (slot >= btrfs_header_nritems(eb)) {
2692                         ret = btrfs_next_leaf(sctx->send_root, path);
2693                         if (ret < 0) {
2694                                 goto out;
2695                         } else if (ret > 0) {
2696                                 ret = 0;
2697                                 break;
2698                         }
2699                         continue;
2700                 }
2701 
2702                 btrfs_item_key_to_cpu(eb, &found_key, slot);
2703                 if (found_key.objectid != key.objectid ||
2704                     found_key.type != key.type) {
2705                         ret = 0;
2706                         goto out;
2707                 }
2708 
2709                 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2710                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2711 
2712                 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2713                     di_key.objectid < sctx->send_progress) {
2714                         ret = 1;
2715                         goto out;
2716                 }
2717 
2718                 path->slots[0]++;
2719         }
2720 
2721 out:
2722         btrfs_free_path(path);
2723         return ret;
2724 }
2725 
2726 /*
2727  * Only creates the inode if it is:
2728  * 1. Not a directory
2729  * 2. Or a directory which was not created already due to out of order
2730  *    directories. See did_create_dir and process_recorded_refs for details.
2731  */
2732 static int send_create_inode_if_needed(struct send_ctx *sctx)
2733 {
2734         int ret;
2735 
2736         if (S_ISDIR(sctx->cur_inode_mode)) {
2737                 ret = did_create_dir(sctx, sctx->cur_ino);
2738                 if (ret < 0)
2739                         goto out;
2740                 if (ret) {
2741                         ret = 0;
2742                         goto out;
2743                 }
2744         }
2745 
2746         ret = send_create_inode(sctx, sctx->cur_ino);
2747         if (ret < 0)
2748                 goto out;
2749 
2750 out:
2751         return ret;
2752 }
2753 
2754 struct recorded_ref {
2755         struct list_head list;
2756         char *name;
2757         struct fs_path *full_path;
2758         u64 dir;
2759         u64 dir_gen;
2760         int name_len;
2761 };
2762 
2763 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
2764 {
2765         ref->full_path = path;
2766         ref->name = (char *)kbasename(ref->full_path->start);
2767         ref->name_len = ref->full_path->end - ref->name;
2768 }
2769 
2770 /*
2771  * We need to process new refs before deleted refs, but compare_tree gives us
2772  * everything mixed. So we first record all refs and later process them.
2773  * This function is a helper to record one ref.
2774  */
2775 static int __record_ref(struct list_head *head, u64 dir,
2776                       u64 dir_gen, struct fs_path *path)
2777 {
2778         struct recorded_ref *ref;
2779 
2780         ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2781         if (!ref)
2782                 return -ENOMEM;
2783 
2784         ref->dir = dir;
2785         ref->dir_gen = dir_gen;
2786         set_ref_path(ref, path);
2787         list_add_tail(&ref->list, head);
2788         return 0;
2789 }
2790 
2791 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2792 {
2793         struct recorded_ref *new;
2794 
2795         new = kmalloc(sizeof(*ref), GFP_KERNEL);
2796         if (!new)
2797                 return -ENOMEM;
2798 
2799         new->dir = ref->dir;
2800         new->dir_gen = ref->dir_gen;
2801         new->full_path = NULL;
2802         INIT_LIST_HEAD(&new->list);
2803         list_add_tail(&new->list, list);
2804         return 0;
2805 }
2806 
2807 static void __free_recorded_refs(struct list_head *head)
2808 {
2809         struct recorded_ref *cur;
2810 
2811         while (!list_empty(head)) {
2812                 cur = list_entry(head->next, struct recorded_ref, list);
2813                 fs_path_free(cur->full_path);
2814                 list_del(&cur->list);
2815                 kfree(cur);
2816         }
2817 }
2818 
2819 static void free_recorded_refs(struct send_ctx *sctx)
2820 {
2821         __free_recorded_refs(&sctx->new_refs);
2822         __free_recorded_refs(&sctx->deleted_refs);
2823 }
2824 
2825 /*
2826  * Renames/moves a file/dir to its orphan name. Used when the first
2827  * ref of an unprocessed inode gets overwritten and for all non empty
2828  * directories.
2829  */
2830 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2831                           struct fs_path *path)
2832 {
2833         int ret;
2834         struct fs_path *orphan;
2835 
2836         orphan = fs_path_alloc();
2837         if (!orphan)
2838                 return -ENOMEM;
2839 
2840         ret = gen_unique_name(sctx, ino, gen, orphan);
2841         if (ret < 0)
2842                 goto out;
2843 
2844         ret = send_rename(sctx, path, orphan);
2845 
2846 out:
2847         fs_path_free(orphan);
2848         return ret;
2849 }
2850 
2851 static struct orphan_dir_info *
2852 add_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2853 {
2854         struct rb_node **p = &sctx->orphan_dirs.rb_node;
2855         struct rb_node *parent = NULL;
2856         struct orphan_dir_info *entry, *odi;
2857 
2858         odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2859         if (!odi)
2860                 return ERR_PTR(-ENOMEM);
2861         odi->ino = dir_ino;
2862         odi->gen = 0;
2863 
2864         while (*p) {
2865                 parent = *p;
2866                 entry = rb_entry(parent, struct orphan_dir_info, node);
2867                 if (dir_ino < entry->ino) {
2868                         p = &(*p)->rb_left;
2869                 } else if (dir_ino > entry->ino) {
2870                         p = &(*p)->rb_right;
2871                 } else {
2872                         kfree(odi);
2873                         return entry;
2874                 }
2875         }
2876 
2877         rb_link_node(&odi->node, parent, p);
2878         rb_insert_color(&odi->node, &sctx->orphan_dirs);
2879         return odi;
2880 }
2881 
2882 static struct orphan_dir_info *
2883 get_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2884 {
2885         struct rb_node *n = sctx->orphan_dirs.rb_node;
2886         struct orphan_dir_info *entry;
2887 
2888         while (n) {
2889                 entry = rb_entry(n, struct orphan_dir_info, node);
2890                 if (dir_ino < entry->ino)
2891                         n = n->rb_left;
2892                 else if (dir_ino > entry->ino)
2893                         n = n->rb_right;
2894                 else
2895                         return entry;
2896         }
2897         return NULL;
2898 }
2899 
2900 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino)
2901 {
2902         struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino);
2903 
2904         return odi != NULL;
2905 }
2906 
2907 static void free_orphan_dir_info(struct send_ctx *sctx,
2908                                  struct orphan_dir_info *odi)
2909 {
2910         if (!odi)
2911                 return;
2912         rb_erase(&odi->node, &sctx->orphan_dirs);
2913         kfree(odi);
2914 }
2915 
2916 /*
2917  * Returns 1 if a directory can be removed at this point in time.
2918  * We check this by iterating all dir items and checking if the inode behind
2919  * the dir item was already processed.
2920  */
2921 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2922                      u64 send_progress)
2923 {
2924         int ret = 0;
2925         struct btrfs_root *root = sctx->parent_root;
2926         struct btrfs_path *path;
2927         struct btrfs_key key;
2928         struct btrfs_key found_key;
2929         struct btrfs_key loc;
2930         struct btrfs_dir_item *di;
2931 
2932         /*
2933          * Don't try to rmdir the top/root subvolume dir.
2934          */
2935         if (dir == BTRFS_FIRST_FREE_OBJECTID)
2936                 return 0;
2937 
2938         path = alloc_path_for_send();
2939         if (!path)
2940                 return -ENOMEM;
2941 
2942         key.objectid = dir;
2943         key.type = BTRFS_DIR_INDEX_KEY;
2944         key.offset = 0;
2945         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2946         if (ret < 0)
2947                 goto out;
2948 
2949         while (1) {
2950                 struct waiting_dir_move *dm;
2951 
2952                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2953                         ret = btrfs_next_leaf(root, path);
2954                         if (ret < 0)
2955                                 goto out;
2956                         else if (ret > 0)
2957                                 break;
2958                         continue;
2959                 }
2960                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2961                                       path->slots[0]);
2962                 if (found_key.objectid != key.objectid ||
2963                     found_key.type != key.type)
2964                         break;
2965 
2966                 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2967                                 struct btrfs_dir_item);
2968                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2969 
2970                 dm = get_waiting_dir_move(sctx, loc.objectid);
2971                 if (dm) {
2972                         struct orphan_dir_info *odi;
2973 
2974                         odi = add_orphan_dir_info(sctx, dir);
2975                         if (IS_ERR(odi)) {
2976                                 ret = PTR_ERR(odi);
2977                                 goto out;
2978                         }
2979                         odi->gen = dir_gen;
2980                         dm->rmdir_ino = dir;
2981                         ret = 0;
2982                         goto out;
2983                 }
2984 
2985                 if (loc.objectid > send_progress) {
2986                         struct orphan_dir_info *odi;
2987 
2988                         odi = get_orphan_dir_info(sctx, dir);
2989                         free_orphan_dir_info(sctx, odi);
2990                         ret = 0;
2991                         goto out;
2992                 }
2993 
2994                 path->slots[0]++;
2995         }
2996 
2997         ret = 1;
2998 
2999 out:
3000         btrfs_free_path(path);
3001         return ret;
3002 }
3003 
3004 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3005 {
3006         struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3007 
3008         return entry != NULL;
3009 }
3010 
3011 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3012 {
3013         struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3014         struct rb_node *parent = NULL;
3015         struct waiting_dir_move *entry, *dm;
3016 
3017         dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3018         if (!dm)
3019                 return -ENOMEM;
3020         dm->ino = ino;
3021         dm->rmdir_ino = 0;
3022         dm->orphanized = orphanized;
3023 
3024         while (*p) {
3025                 parent = *p;
3026                 entry = rb_entry(parent, struct waiting_dir_move, node);
3027                 if (ino < entry->ino) {
3028                         p = &(*p)->rb_left;
3029                 } else if (ino > entry->ino) {
3030                         p = &(*p)->rb_right;
3031                 } else {
3032                         kfree(dm);
3033                         return -EEXIST;
3034                 }
3035         }
3036 
3037         rb_link_node(&dm->node, parent, p);
3038         rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3039         return 0;
3040 }
3041 
3042 static struct waiting_dir_move *
3043 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3044 {
3045         struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3046         struct waiting_dir_move *entry;
3047 
3048         while (n) {
3049                 entry = rb_entry(n, struct waiting_dir_move, node);
3050                 if (ino < entry->ino)
3051                         n = n->rb_left;
3052                 else if (ino > entry->ino)
3053                         n = n->rb_right;
3054                 else
3055                         return entry;
3056         }
3057         return NULL;
3058 }
3059 
3060 static void free_waiting_dir_move(struct send_ctx *sctx,
3061                                   struct waiting_dir_move *dm)
3062 {
3063         if (!dm)
3064                 return;
3065         rb_erase(&dm->node, &sctx->waiting_dir_moves);
3066         kfree(dm);
3067 }
3068 
3069 static int add_pending_dir_move(struct send_ctx *sctx,
3070                                 u64 ino,
3071                                 u64 ino_gen,
3072                                 u64 parent_ino,
3073                                 struct list_head *new_refs,
3074                                 struct list_head *deleted_refs,
3075                                 const bool is_orphan)
3076 {
3077         struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3078         struct rb_node *parent = NULL;
3079         struct pending_dir_move *entry = NULL, *pm;
3080         struct recorded_ref *cur;
3081         int exists = 0;
3082         int ret;
3083 
3084         pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3085         if (!pm)
3086                 return -ENOMEM;
3087         pm->parent_ino = parent_ino;
3088         pm->ino = ino;
3089         pm->gen = ino_gen;
3090         INIT_LIST_HEAD(&pm->list);
3091         INIT_LIST_HEAD(&pm->update_refs);
3092         RB_CLEAR_NODE(&pm->node);
3093 
3094         while (*p) {
3095                 parent = *p;
3096                 entry = rb_entry(parent, struct pending_dir_move, node);
3097                 if (parent_ino < entry->parent_ino) {
3098                         p = &(*p)->rb_left;
3099                 } else if (parent_ino > entry->parent_ino) {
3100                         p = &(*p)->rb_right;
3101                 } else {
3102                         exists = 1;
3103                         break;
3104                 }
3105         }
3106 
3107         list_for_each_entry(cur, deleted_refs, list) {
3108                 ret = dup_ref(cur, &pm->update_refs);
3109                 if (ret < 0)
3110                         goto out;
3111         }
3112         list_for_each_entry(cur, new_refs, list) {
3113                 ret = dup_ref(cur, &pm->update_refs);
3114                 if (ret < 0)
3115                         goto out;
3116         }
3117 
3118         ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3119         if (ret)
3120                 goto out;
3121 
3122         if (exists) {
3123                 list_add_tail(&pm->list, &entry->list);
3124         } else {
3125                 rb_link_node(&pm->node, parent, p);
3126                 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3127         }
3128         ret = 0;
3129 out:
3130         if (ret) {
3131                 __free_recorded_refs(&pm->update_refs);
3132                 kfree(pm);
3133         }
3134         return ret;
3135 }
3136 
3137 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3138                                                       u64 parent_ino)
3139 {
3140         struct rb_node *n = sctx->pending_dir_moves.rb_node;
3141         struct pending_dir_move *entry;
3142 
3143         while (n) {
3144                 entry = rb_entry(n, struct pending_dir_move, node);
3145                 if (parent_ino < entry->parent_ino)
3146                         n = n->rb_left;
3147                 else if (parent_ino > entry->parent_ino)
3148                         n = n->rb_right;
3149                 else
3150                         return entry;
3151         }
3152         return NULL;
3153 }
3154 
3155 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3156                      u64 ino, u64 gen, u64 *ancestor_ino)
3157 {
3158         int ret = 0;
3159         u64 parent_inode = 0;
3160         u64 parent_gen = 0;
3161         u64 start_ino = ino;
3162 
3163         *ancestor_ino = 0;
3164         while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3165                 fs_path_reset(name);
3166 
3167                 if (is_waiting_for_rm(sctx, ino))
3168                         break;
3169                 if (is_waiting_for_move(sctx, ino)) {
3170                         if (*ancestor_ino == 0)
3171                                 *ancestor_ino = ino;
3172                         ret = get_first_ref(sctx->parent_root, ino,
3173                                             &parent_inode, &parent_gen, name);
3174                 } else {
3175                         ret = __get_cur_name_and_parent(sctx, ino, gen,
3176                                                         &parent_inode,
3177                                                         &parent_gen, name);
3178                         if (ret > 0) {
3179                                 ret = 0;
3180                                 break;
3181                         }
3182                 }
3183                 if (ret < 0)
3184                         break;
3185                 if (parent_inode == start_ino) {
3186                         ret = 1;
3187                         if (*ancestor_ino == 0)
3188                                 *ancestor_ino = ino;
3189                         break;
3190                 }
3191                 ino = parent_inode;
3192                 gen = parent_gen;
3193         }
3194         return ret;
3195 }
3196 
3197 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3198 {
3199         struct fs_path *from_path = NULL;
3200         struct fs_path *to_path = NULL;
3201         struct fs_path *name = NULL;
3202         u64 orig_progress = sctx->send_progress;
3203         struct recorded_ref *cur;
3204         u64 parent_ino, parent_gen;
3205         struct waiting_dir_move *dm = NULL;
3206         u64 rmdir_ino = 0;
3207         u64 ancestor;
3208         bool is_orphan;
3209         int ret;
3210 
3211         name = fs_path_alloc();
3212         from_path = fs_path_alloc();
3213         if (!name || !from_path) {
3214                 ret = -ENOMEM;
3215                 goto out;
3216         }
3217 
3218         dm = get_waiting_dir_move(sctx, pm->ino);
3219         ASSERT(dm);
3220         rmdir_ino = dm->rmdir_ino;
3221         is_orphan = dm->orphanized;
3222         free_waiting_dir_move(sctx, dm);
3223 
3224         if (is_orphan) {
3225                 ret = gen_unique_name(sctx, pm->ino,
3226                                       pm->gen, from_path);
3227         } else {
3228                 ret = get_first_ref(sctx->parent_root, pm->ino,
3229                                     &parent_ino, &parent_gen, name);
3230                 if (ret < 0)
3231                         goto out;
3232                 ret = get_cur_path(sctx, parent_ino, parent_gen,
3233                                    from_path);
3234                 if (ret < 0)
3235                         goto out;
3236                 ret = fs_path_add_path(from_path, name);
3237         }
3238         if (ret < 0)
3239                 goto out;
3240 
3241         sctx->send_progress = sctx->cur_ino + 1;
3242         ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3243         if (ret < 0)
3244                 goto out;
3245         if (ret) {
3246                 LIST_HEAD(deleted_refs);
3247                 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3248                 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3249                                            &pm->update_refs, &deleted_refs,
3250                                            is_orphan);
3251                 if (ret < 0)
3252                         goto out;
3253                 if (rmdir_ino) {
3254                         dm = get_waiting_dir_move(sctx, pm->ino);
3255                         ASSERT(dm);
3256                         dm->rmdir_ino = rmdir_ino;
3257                 }
3258                 goto out;
3259         }
3260         fs_path_reset(name);
3261         to_path = name;
3262         name = NULL;
3263         ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3264         if (ret < 0)
3265                 goto out;
3266 
3267         ret = send_rename(sctx, from_path, to_path);
3268         if (ret < 0)
3269                 goto out;
3270 
3271         if (rmdir_ino) {
3272                 struct orphan_dir_info *odi;
3273 
3274                 odi = get_orphan_dir_info(sctx, rmdir_ino);
3275                 if (!odi) {
3276                         /* already deleted */
3277                         goto finish;
3278                 }
3279                 ret = can_rmdir(sctx, rmdir_ino, odi->gen, sctx->cur_ino);
3280                 if (ret < 0)
3281                         goto out;
3282                 if (!ret)
3283                         goto finish;
3284 
3285                 name = fs_path_alloc();
3286                 if (!name) {
3287                         ret = -ENOMEM;
3288                         goto out;
3289                 }
3290                 ret = get_cur_path(sctx, rmdir_ino, odi->gen, name);
3291                 if (ret < 0)
3292                         goto out;
3293                 ret = send_rmdir(sctx, name);
3294                 if (ret < 0)
3295                         goto out;
3296                 free_orphan_dir_info(sctx, odi);
3297         }
3298 
3299 finish:
3300         ret = send_utimes(sctx, pm->ino, pm->gen);
3301         if (ret < 0)
3302                 goto out;
3303 
3304         /*
3305          * After rename/move, need to update the utimes of both new parent(s)
3306          * and old parent(s).
3307          */
3308         list_for_each_entry(cur, &pm->update_refs, list) {
3309                 /*
3310                  * The parent inode might have been deleted in the send snapshot
3311                  */
3312                 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3313                                      NULL, NULL, NULL, NULL, NULL);
3314                 if (ret == -ENOENT) {
3315                         ret = 0;
3316                         continue;
3317                 }
3318                 if (ret < 0)
3319                         goto out;
3320 
3321                 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3322                 if (ret < 0)
3323                         goto out;
3324         }
3325 
3326 out:
3327         fs_path_free(name);
3328         fs_path_free(from_path);
3329         fs_path_free(to_path);
3330         sctx->send_progress = orig_progress;
3331 
3332         return ret;
3333 }
3334 
3335 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3336 {
3337         if (!list_empty(&m->list))
3338                 list_del(&m->list);
3339         if (!RB_EMPTY_NODE(&m->node))
3340                 rb_erase(&m->node, &sctx->pending_dir_moves);
3341         __free_recorded_refs(&m->update_refs);
3342         kfree(m);
3343 }
3344 
3345 static void tail_append_pending_moves(struct pending_dir_move *moves,
3346                                       struct list_head *stack)
3347 {
3348         if (list_empty(&moves->list)) {
3349                 list_add_tail(&moves->list, stack);
3350         } else {
3351                 LIST_HEAD(list);
3352                 list_splice_init(&moves->list, &list);
3353                 list_add_tail(&moves->list, stack);
3354                 list_splice_tail(&list, stack);
3355         }
3356 }
3357 
3358 static int apply_children_dir_moves(struct send_ctx *sctx)
3359 {
3360         struct pending_dir_move *pm;
3361         struct list_head stack;
3362         u64 parent_ino = sctx->cur_ino;
3363         int ret = 0;
3364 
3365         pm = get_pending_dir_moves(sctx, parent_ino);
3366         if (!pm)
3367                 return 0;
3368 
3369         INIT_LIST_HEAD(&stack);
3370         tail_append_pending_moves(pm, &stack);
3371 
3372         while (!list_empty(&stack)) {
3373                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3374                 parent_ino = pm->ino;
3375                 ret = apply_dir_move(sctx, pm);
3376                 free_pending_move(sctx, pm);
3377                 if (ret)
3378                         goto out;
3379                 pm = get_pending_dir_moves(sctx, parent_ino);
3380                 if (pm)
3381                         tail_append_pending_moves(pm, &stack);
3382         }
3383         return 0;
3384 
3385 out:
3386         while (!list_empty(&stack)) {
3387                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3388                 free_pending_move(sctx, pm);
3389         }
3390         return ret;
3391 }
3392 
3393 /*
3394  * We might need to delay a directory rename even when no ancestor directory
3395  * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3396  * renamed. This happens when we rename a directory to the old name (the name
3397  * in the parent root) of some other unrelated directory that got its rename
3398  * delayed due to some ancestor with higher number that got renamed.
3399  *
3400  * Example:
3401  *
3402  * Parent snapshot:
3403  * .                                       (ino 256)
3404  * |---- a/                                (ino 257)
3405  * |     |---- file                        (ino 260)
3406  * |
3407  * |---- b/                                (ino 258)
3408  * |---- c/                                (ino 259)
3409  *
3410  * Send snapshot:
3411  * .                                       (ino 256)
3412  * |---- a/                                (ino 258)
3413  * |---- x/                                (ino 259)
3414  *       |---- y/                          (ino 257)
3415  *             |----- file                 (ino 260)
3416  *
3417  * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3418  * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3419  * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3420  * must issue is:
3421  *
3422  * 1 - rename 259 from 'c' to 'x'
3423  * 2 - rename 257 from 'a' to 'x/y'
3424  * 3 - rename 258 from 'b' to 'a'
3425  *
3426  * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3427  * be done right away and < 0 on error.
3428  */
3429 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3430                                   struct recorded_ref *parent_ref,
3431                                   const bool is_orphan)
3432 {
3433         struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3434         struct btrfs_path *path;
3435         struct btrfs_key key;
3436         struct btrfs_key di_key;
3437         struct btrfs_dir_item *di;
3438         u64 left_gen;
3439         u64 right_gen;
3440         int ret = 0;
3441         struct waiting_dir_move *wdm;
3442 
3443         if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3444                 return 0;
3445 
3446         path = alloc_path_for_send();
3447         if (!path)
3448                 return -ENOMEM;
3449 
3450         key.objectid = parent_ref->dir;
3451         key.type = BTRFS_DIR_ITEM_KEY;
3452         key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3453 
3454         ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3455         if (ret < 0) {
3456                 goto out;
3457         } else if (ret > 0) {
3458                 ret = 0;
3459                 goto out;
3460         }
3461 
3462         di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3463                                        parent_ref->name_len);
3464         if (!di) {
3465                 ret = 0;
3466                 goto out;
3467         }
3468         /*
3469          * di_key.objectid has the number of the inode that has a dentry in the
3470          * parent directory with the same name that sctx->cur_ino is being
3471          * renamed to. We need to check if that inode is in the send root as
3472          * well and if it is currently marked as an inode with a pending rename,
3473          * if it is, we need to delay the rename of sctx->cur_ino as well, so
3474          * that it happens after that other inode is renamed.
3475          */
3476         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3477         if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3478                 ret = 0;
3479                 goto out;
3480         }
3481 
3482         ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3483                              &left_gen, NULL, NULL, NULL, NULL);
3484         if (ret < 0)
3485                 goto out;
3486         ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3487                              &right_gen, NULL, NULL, NULL, NULL);
3488         if (ret < 0) {
3489                 if (ret == -ENOENT)
3490                         ret = 0;
3491                 goto out;
3492         }
3493 
3494         /* Different inode, no need to delay the rename of sctx->cur_ino */
3495         if (right_gen != left_gen) {
3496                 ret = 0;
3497                 goto out;
3498         }
3499 
3500         wdm = get_waiting_dir_move(sctx, di_key.objectid);
3501         if (wdm && !wdm->orphanized) {
3502                 ret = add_pending_dir_move(sctx,
3503                                            sctx->cur_ino,
3504                                            sctx->cur_inode_gen,
3505                                            di_key.objectid,
3506                                            &sctx->new_refs,
3507                                            &sctx->deleted_refs,
3508                                            is_orphan);
3509                 if (!ret)
3510                         ret = 1;
3511         }
3512 out:
3513         btrfs_free_path(path);
3514         return ret;
3515 }
3516 
3517 /*
3518  * Check if inode ino2, or any of its ancestors, is inode ino1.
3519  * Return 1 if true, 0 if false and < 0 on error.
3520  */
3521 static int check_ino_in_path(struct btrfs_root *root,
3522                              const u64 ino1,
3523                              const u64 ino1_gen,
3524                              const u64 ino2,
3525                              const u64 ino2_gen,
3526                              struct fs_path *fs_path)
3527 {
3528         u64 ino = ino2;
3529 
3530         if (ino1 == ino2)
3531                 return ino1_gen == ino2_gen;
3532 
3533         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3534                 u64 parent;
3535                 u64 parent_gen;
3536                 int ret;
3537 
3538                 fs_path_reset(fs_path);
3539                 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3540                 if (ret < 0)
3541                         return ret;
3542                 if (parent == ino1)
3543                         return parent_gen == ino1_gen;
3544                 ino = parent;
3545         }
3546         return 0;
3547 }
3548 
3549 /*
3550  * Check if ino ino1 is an ancestor of inode ino2 in the given root for any
3551  * possible path (in case ino2 is not a directory and has multiple hard links).
3552  * Return 1 if true, 0 if false and < 0 on error.
3553  */
3554 static int is_ancestor(struct btrfs_root *root,
3555                        const u64 ino1,
3556                        const u64 ino1_gen,
3557                        const u64 ino2,
3558                        struct fs_path *fs_path)
3559 {
3560         bool free_fs_path = false;
3561         int ret = 0;
3562         struct btrfs_path *path = NULL;
3563         struct btrfs_key key;
3564 
3565         if (!fs_path) {
3566                 fs_path = fs_path_alloc();
3567                 if (!fs_path)
3568                         return -ENOMEM;
3569                 free_fs_path = true;
3570         }
3571 
3572         path = alloc_path_for_send();
3573         if (!path) {
3574                 ret = -ENOMEM;
3575                 goto out;
3576         }
3577 
3578         key.objectid = ino2;
3579         key.type = BTRFS_INODE_REF_KEY;
3580         key.offset = 0;
3581 
3582         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3583         if (ret < 0)
3584                 goto out;
3585 
3586         while (true) {
3587                 struct extent_buffer *leaf = path->nodes[0];
3588                 int slot = path->slots[0];
3589                 u32 cur_offset = 0;
3590                 u32 item_size;
3591 
3592                 if (slot >= btrfs_header_nritems(leaf)) {
3593                         ret = btrfs_next_leaf(root, path);
3594                         if (ret < 0)
3595                                 goto out;
3596                         if (ret > 0)
3597                                 break;
3598                         continue;
3599                 }
3600 
3601                 btrfs_item_key_to_cpu(leaf, &key, slot);
3602                 if (key.objectid != ino2)
3603                         break;
3604                 if (key.type != BTRFS_INODE_REF_KEY &&
3605                     key.type != BTRFS_INODE_EXTREF_KEY)
3606                         break;
3607 
3608                 item_size = btrfs_item_size_nr(leaf, slot);
3609                 while (cur_offset < item_size) {
3610                         u64 parent;
3611                         u64 parent_gen;
3612 
3613                         if (key.type == BTRFS_INODE_EXTREF_KEY) {
3614                                 unsigned long ptr;
3615                                 struct btrfs_inode_extref *extref;
3616 
3617                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3618                                 extref = (struct btrfs_inode_extref *)
3619                                         (ptr + cur_offset);
3620                                 parent = btrfs_inode_extref_parent(leaf,
3621                                                                    extref);
3622                                 cur_offset += sizeof(*extref);
3623                                 cur_offset += btrfs_inode_extref_name_len(leaf,
3624                                                                   extref);
3625                         } else {
3626                                 parent = key.offset;
3627                                 cur_offset = item_size;
3628                         }
3629 
3630                         ret = get_inode_info(root, parent, NULL, &parent_gen,
3631                                              NULL, NULL, NULL, NULL);
3632                         if (ret < 0)
3633                                 goto out;
3634                         ret = check_ino_in_path(root, ino1, ino1_gen,
3635                                                 parent, parent_gen, fs_path);
3636                         if (ret)
3637                                 goto out;
3638                 }
3639                 path->slots[0]++;
3640         }
3641         ret = 0;
3642  out:
3643         btrfs_free_path(path);
3644         if (free_fs_path)
3645                 fs_path_free(fs_path);
3646         return ret;
3647 }
3648 
3649 static int wait_for_parent_move(struct send_ctx *sctx,
3650                                 struct recorded_ref *parent_ref,
3651                                 const bool is_orphan)
3652 {
3653         int ret = 0;
3654         u64 ino = parent_ref->dir;
3655         u64 ino_gen = parent_ref->dir_gen;
3656         u64 parent_ino_before, parent_ino_after;
3657         struct fs_path *path_before = NULL;
3658         struct fs_path *path_after = NULL;
3659         int len1, len2;
3660 
3661         path_after = fs_path_alloc();
3662         path_before = fs_path_alloc();
3663         if (!path_after || !path_before) {
3664                 ret = -ENOMEM;
3665                 goto out;
3666         }
3667 
3668         /*
3669          * Our current directory inode may not yet be renamed/moved because some
3670          * ancestor (immediate or not) has to be renamed/moved first. So find if
3671          * such ancestor exists and make sure our own rename/move happens after
3672          * that ancestor is processed to avoid path build infinite loops (done
3673          * at get_cur_path()).
3674          */
3675         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3676                 u64 parent_ino_after_gen;
3677 
3678                 if (is_waiting_for_move(sctx, ino)) {
3679                         /*
3680                          * If the current inode is an ancestor of ino in the
3681                          * parent root, we need to delay the rename of the
3682                          * current inode, otherwise don't delayed the rename
3683                          * because we can end up with a circular dependency
3684                          * of renames, resulting in some directories never
3685                          * getting the respective rename operations issued in
3686                          * the send stream or getting into infinite path build
3687                          * loops.
3688                          */
3689                         ret = is_ancestor(sctx->parent_root,
3690                                           sctx->cur_ino, sctx->cur_inode_gen,
3691                                           ino, path_before);
3692                         if (ret)
3693                                 break;
3694                 }
3695 
3696                 fs_path_reset(path_before);
3697                 fs_path_reset(path_after);
3698 
3699                 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3700                                     &parent_ino_after_gen, path_after);
3701                 if (ret < 0)
3702                         goto out;
3703                 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3704                                     NULL, path_before);
3705                 if (ret < 0 && ret != -ENOENT) {
3706                         goto out;
3707                 } else if (ret == -ENOENT) {
3708                         ret = 0;
3709                         break;
3710                 }
3711 
3712                 len1 = fs_path_len(path_before);
3713                 len2 = fs_path_len(path_after);
3714                 if (ino > sctx->cur_ino &&
3715                     (parent_ino_before != parent_ino_after || len1 != len2 ||
3716                      memcmp(path_before->start, path_after->start, len1))) {
3717                         u64 parent_ino_gen;
3718 
3719                         ret = get_inode_info(sctx->parent_root, ino, NULL,
3720                                              &parent_ino_gen, NULL, NULL, NULL,
3721                                              NULL);
3722                         if (ret < 0)
3723                                 goto out;
3724                         if (ino_gen == parent_ino_gen) {
3725                                 ret = 1;
3726                                 break;
3727                         }
3728                 }
3729                 ino = parent_ino_after;
3730                 ino_gen = parent_ino_after_gen;
3731         }
3732 
3733 out:
3734         fs_path_free(path_before);
3735         fs_path_free(path_after);
3736 
3737         if (ret == 1) {
3738                 ret = add_pending_dir_move(sctx,
3739                                            sctx->cur_ino,
3740                                            sctx->cur_inode_gen,
3741                                            ino,
3742                                            &sctx->new_refs,
3743                                            &sctx->deleted_refs,
3744                                            is_orphan);
3745                 if (!ret)
3746                         ret = 1;
3747         }
3748 
3749         return ret;
3750 }
3751 
3752 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3753 {
3754         int ret;
3755         struct fs_path *new_path;
3756 
3757         /*
3758          * Our reference's name member points to its full_path member string, so
3759          * we use here a new path.
3760          */
3761         new_path = fs_path_alloc();
3762         if (!new_path)
3763                 return -ENOMEM;
3764 
3765         ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
3766         if (ret < 0) {
3767                 fs_path_free(new_path);
3768                 return ret;
3769         }
3770         ret = fs_path_add(new_path, ref->name, ref->name_len);
3771         if (ret < 0) {
3772                 fs_path_free(new_path);
3773                 return ret;
3774         }
3775 
3776         fs_path_free(ref->full_path);
3777         set_ref_path(ref, new_path);
3778 
3779         return 0;
3780 }
3781 
3782 /*
3783  * This does all the move/link/unlink/rmdir magic.
3784  */
3785 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3786 {
3787         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3788         int ret = 0;
3789         struct recorded_ref *cur;
3790         struct recorded_ref *cur2;
3791         struct list_head check_dirs;
3792         struct fs_path *valid_path = NULL;
3793         u64 ow_inode = 0;
3794         u64 ow_gen;
3795         u64 ow_mode;
3796         int did_overwrite = 0;
3797         int is_orphan = 0;
3798         u64 last_dir_ino_rm = 0;
3799         bool can_rename = true;
3800         bool orphanized_dir = false;
3801         bool orphanized_ancestor = false;
3802 
3803         btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3804 
3805         /*
3806          * This should never happen as the root dir always has the same ref
3807          * which is always '..'
3808          */
3809         BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3810         INIT_LIST_HEAD(&check_dirs);
3811 
3812         valid_path = fs_path_alloc();
3813         if (!valid_path) {
3814                 ret = -ENOMEM;
3815                 goto out;
3816         }
3817 
3818         /*
3819          * First, check if the first ref of the current inode was overwritten
3820          * before. If yes, we know that the current inode was already orphanized
3821          * and thus use the orphan name. If not, we can use get_cur_path to
3822          * get the path of the first ref as it would like while receiving at
3823          * this point in time.
3824          * New inodes are always orphan at the beginning, so force to use the
3825          * orphan name in this case.
3826          * The first ref is stored in valid_path and will be updated if it
3827          * gets moved around.
3828          */
3829         if (!sctx->cur_inode_new) {
3830                 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3831                                 sctx->cur_inode_gen);
3832                 if (ret < 0)
3833                         goto out;
3834                 if (ret)
3835                         did_overwrite = 1;
3836         }
3837         if (sctx->cur_inode_new || did_overwrite) {
3838                 ret = gen_unique_name(sctx, sctx->cur_ino,
3839                                 sctx->cur_inode_gen, valid_path);
3840                 if (ret < 0)
3841                         goto out;
3842                 is_orphan = 1;
3843         } else {
3844                 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3845                                 valid_path);
3846                 if (ret < 0)
3847                         goto out;
3848         }
3849 
3850         list_for_each_entry(cur, &sctx->new_refs, list) {
3851                 /*
3852                  * We may have refs where the parent directory does not exist
3853                  * yet. This happens if the parent directories inum is higher
3854                  * the the current inum. To handle this case, we create the
3855                  * parent directory out of order. But we need to check if this
3856                  * did already happen before due to other refs in the same dir.
3857                  */
3858                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3859                 if (ret < 0)
3860                         goto out;
3861                 if (ret == inode_state_will_create) {
3862                         ret = 0;
3863                         /*
3864                          * First check if any of the current inodes refs did
3865                          * already create the dir.
3866                          */
3867                         list_for_each_entry(cur2, &sctx->new_refs, list) {
3868                                 if (cur == cur2)
3869                                         break;
3870                                 if (cur2->dir == cur->dir) {
3871                                         ret = 1;
3872                                         break;
3873                                 }
3874                         }
3875 
3876                         /*
3877                          * If that did not happen, check if a previous inode
3878                          * did already create the dir.
3879                          */
3880                         if (!ret)
3881                                 ret = did_create_dir(sctx, cur->dir);
3882                         if (ret < 0)
3883                                 goto out;
3884                         if (!ret) {
3885                                 ret = send_create_inode(sctx, cur->dir);
3886                                 if (ret < 0)
3887                                         goto out;
3888                         }
3889                 }
3890 
3891                 /*
3892                  * Check if this new ref would overwrite the first ref of
3893                  * another unprocessed inode. If yes, orphanize the
3894                  * overwritten inode. If we find an overwritten ref that is
3895                  * not the first ref, simply unlink it.
3896                  */
3897                 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3898                                 cur->name, cur->name_len,
3899                                 &ow_inode, &ow_gen, &ow_mode);
3900                 if (ret < 0)
3901                         goto out;
3902                 if (ret) {
3903                         ret = is_first_ref(sctx->parent_root,
3904                                            ow_inode, cur->dir, cur->name,
3905                                            cur->name_len);
3906                         if (ret < 0)
3907                                 goto out;
3908                         if (ret) {
3909                                 struct name_cache_entry *nce;
3910                                 struct waiting_dir_move *wdm;
3911 
3912                                 ret = orphanize_inode(sctx, ow_inode, ow_gen,
3913                                                 cur->full_path);
3914                                 if (ret < 0)
3915                                         goto out;
3916                                 if (S_ISDIR(ow_mode))
3917                                         orphanized_dir = true;
3918 
3919                                 /*
3920                                  * If ow_inode has its rename operation delayed
3921                                  * make sure that its orphanized name is used in
3922                                  * the source path when performing its rename
3923                                  * operation.
3924                                  */
3925                                 if (is_waiting_for_move(sctx, ow_inode)) {
3926                                         wdm = get_waiting_dir_move(sctx,
3927                                                                    ow_inode);
3928                                         ASSERT(wdm);
3929                                         wdm->orphanized = true;
3930                                 }
3931 
3932                                 /*
3933                                  * Make sure we clear our orphanized inode's
3934                                  * name from the name cache. This is because the
3935                                  * inode ow_inode might be an ancestor of some
3936                                  * other inode that will be orphanized as well
3937                                  * later and has an inode number greater than
3938                                  * sctx->send_progress. We need to prevent
3939                                  * future name lookups from using the old name
3940                                  * and get instead the orphan name.
3941                                  */
3942                                 nce = name_cache_search(sctx, ow_inode, ow_gen);
3943                                 if (nce) {
3944                                         name_cache_delete(sctx, nce);
3945                                         kfree(nce);
3946                                 }
3947 
3948                                 /*
3949                                  * ow_inode might currently be an ancestor of
3950                                  * cur_ino, therefore compute valid_path (the
3951                                  * current path of cur_ino) again because it
3952                                  * might contain the pre-orphanization name of
3953                                  * ow_inode, which is no longer valid.
3954                                  */
3955                                 ret = is_ancestor(sctx->parent_root,
3956                                                   ow_inode, ow_gen,
3957                                                   sctx->cur_ino, NULL);
3958                                 if (ret > 0) {
3959                                         orphanized_ancestor = true;
3960                                         fs_path_reset(valid_path);
3961                                         ret = get_cur_path(sctx, sctx->cur_ino,
3962                                                            sctx->cur_inode_gen,
3963                                                            valid_path);
3964                                 }
3965                                 if (ret < 0)
3966                                         goto out;
3967                         } else {
3968                                 ret = send_unlink(sctx, cur->full_path);
3969                                 if (ret < 0)
3970                                         goto out;
3971                         }
3972                 }
3973 
3974                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
3975                         ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
3976                         if (ret < 0)
3977                                 goto out;
3978                         if (ret == 1) {
3979                                 can_rename = false;
3980                                 *pending_move = 1;
3981                         }
3982                 }
3983 
3984                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
3985                     can_rename) {
3986                         ret = wait_for_parent_move(sctx, cur, is_orphan);
3987                         if (ret < 0)
3988                                 goto out;
3989                         if (ret == 1) {
3990                                 can_rename = false;
3991                                 *pending_move = 1;
3992                         }
3993                 }
3994 
3995                 /*
3996                  * link/move the ref to the new place. If we have an orphan
3997                  * inode, move it and update valid_path. If not, link or move
3998                  * it depending on the inode mode.
3999                  */
4000                 if (is_orphan && can_rename) {
4001                         ret = send_rename(sctx, valid_path, cur->full_path);
4002                         if (ret < 0)
4003                                 goto out;
4004                         is_orphan = 0;
4005                         ret = fs_path_copy(valid_path, cur->full_path);
4006                         if (ret < 0)
4007                                 goto out;
4008                 } else if (can_rename) {
4009                         if (S_ISDIR(sctx->cur_inode_mode)) {
4010                                 /*
4011                                  * Dirs can't be linked, so move it. For moved
4012                                  * dirs, we always have one new and one deleted
4013                                  * ref. The deleted ref is ignored later.
4014                                  */
4015                                 ret = send_rename(sctx, valid_path,
4016                                                   cur->full_path);
4017                                 if (!ret)
4018                                         ret = fs_path_copy(valid_path,
4019                                                            cur->full_path);
4020                                 if (ret < 0)
4021                                         goto out;
4022                         } else {
4023                                 /*
4024                                  * We might have previously orphanized an inode
4025                                  * which is an ancestor of our current inode,
4026                                  * so our reference's full path, which was
4027                                  * computed before any such orphanizations, must
4028                                  * be updated.
4029                                  */
4030                                 if (orphanized_dir) {
4031                                         ret = update_ref_path(sctx, cur);
4032                                         if (ret < 0)
4033                                                 goto out;
4034                                 }
4035                                 ret = send_link(sctx, cur->full_path,
4036                                                 valid_path);
4037                                 if (ret < 0)
4038                                         goto out;
4039                         }
4040                 }
4041                 ret = dup_ref(cur, &check_dirs);
4042                 if (ret < 0)
4043                         goto out;
4044         }
4045 
4046         if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4047                 /*
4048                  * Check if we can already rmdir the directory. If not,
4049                  * orphanize it. For every dir item inside that gets deleted
4050                  * later, we do this check again and rmdir it then if possible.
4051                  * See the use of check_dirs for more details.
4052                  */
4053                 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4054                                 sctx->cur_ino);
4055                 if (ret < 0)
4056                         goto out;
4057                 if (ret) {
4058                         ret = send_rmdir(sctx, valid_path);
4059                         if (ret < 0)
4060                                 goto out;
4061                 } else if (!is_orphan) {
4062                         ret = orphanize_inode(sctx, sctx->cur_ino,
4063                                         sctx->cur_inode_gen, valid_path);
4064                         if (ret < 0)
4065                                 goto out;
4066                         is_orphan = 1;
4067                 }
4068 
4069                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4070                         ret = dup_ref(cur, &check_dirs);
4071                         if (ret < 0)
4072                                 goto out;
4073                 }
4074         } else if (S_ISDIR(sctx->cur_inode_mode) &&
4075                    !list_empty(&sctx->deleted_refs)) {
4076                 /*
4077                  * We have a moved dir. Add the old parent to check_dirs
4078                  */
4079                 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4080                                 list);
4081                 ret = dup_ref(cur, &check_dirs);
4082                 if (ret < 0)
4083                         goto out;
4084         } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4085                 /*
4086                  * We have a non dir inode. Go through all deleted refs and
4087                  * unlink them if they were not already overwritten by other
4088                  * inodes.
4089                  */
4090                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4091                         ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4092                                         sctx->cur_ino, sctx->cur_inode_gen,
4093                                         cur->name, cur->name_len);
4094                         if (ret < 0)
4095                                 goto out;
4096                         if (!ret) {
4097                                 /*
4098                                  * If we orphanized any ancestor before, we need
4099                                  * to recompute the full path for deleted names,
4100                                  * since any such path was computed before we
4101                                  * processed any references and orphanized any
4102                                  * ancestor inode.
4103                                  */
4104                                 if (orphanized_ancestor) {
4105                                         ret = update_ref_path(sctx, cur);
4106                                         if (ret < 0)
4107                                                 goto out;
4108                                 }
4109                                 ret = send_unlink(sctx, cur->full_path);
4110                                 if (ret < 0)
4111                                         goto out;
4112                         }
4113                         ret = dup_ref(cur, &check_dirs);
4114                         if (ret < 0)
4115                                 goto out;
4116                 }
4117                 /*
4118                  * If the inode is still orphan, unlink the orphan. This may
4119                  * happen when a previous inode did overwrite the first ref
4120                  * of this inode and no new refs were added for the current
4121                  * inode. Unlinking does not mean that the inode is deleted in
4122                  * all cases. There may still be links to this inode in other
4123                  * places.
4124                  */
4125                 if (is_orphan) {
4126                         ret = send_unlink(sctx, valid_path);
4127                         if (ret < 0)
4128                                 goto out;
4129                 }
4130         }
4131 
4132         /*
4133          * We did collect all parent dirs where cur_inode was once located. We
4134          * now go through all these dirs and check if they are pending for
4135          * deletion and if it's finally possible to perform the rmdir now.
4136          * We also update the inode stats of the parent dirs here.
4137          */
4138         list_for_each_entry(cur, &check_dirs, list) {
4139                 /*
4140                  * In case we had refs into dirs that were not processed yet,
4141                  * we don't need to do the utime and rmdir logic for these dirs.
4142                  * The dir will be processed later.
4143                  */
4144                 if (cur->dir > sctx->cur_ino)
4145                         continue;
4146 
4147                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4148                 if (ret < 0)
4149                         goto out;
4150 
4151                 if (ret == inode_state_did_create ||
4152                     ret == inode_state_no_change) {
4153                         /* TODO delayed utimes */
4154                         ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4155                         if (ret < 0)
4156                                 goto out;
4157                 } else if (ret == inode_state_did_delete &&
4158                            cur->dir != last_dir_ino_rm) {
4159                         ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4160                                         sctx->cur_ino);
4161                         if (ret < 0)
4162                                 goto out;
4163                         if (ret) {
4164                                 ret = get_cur_path(sctx, cur->dir,
4165                                                    cur->dir_gen, valid_path);
4166                                 if (ret < 0)
4167                                         goto out;
4168                                 ret = send_rmdir(sctx, valid_path);
4169                                 if (ret < 0)
4170                                         goto out;
4171                                 last_dir_ino_rm = cur->dir;
4172                         }
4173                 }
4174         }
4175 
4176         ret = 0;
4177 
4178 out:
4179         __free_recorded_refs(&check_dirs);
4180         free_recorded_refs(sctx);
4181         fs_path_free(valid_path);
4182         return ret;
4183 }
4184 
4185 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name,
4186                       void *ctx, struct list_head *refs)
4187 {
4188         int ret = 0;
4189         struct send_ctx *sctx = ctx;
4190         struct fs_path *p;
4191         u64 gen;
4192 
4193         p = fs_path_alloc();
4194         if (!p)
4195                 return -ENOMEM;
4196 
4197         ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4198                         NULL, NULL);
4199         if (ret < 0)
4200                 goto out;
4201 
4202         ret = get_cur_path(sctx, dir, gen, p);
4203         if (ret < 0)
4204                 goto out;
4205         ret = fs_path_add_path(p, name);
4206         if (ret < 0)
4207                 goto out;
4208 
4209         ret = __record_ref(refs, dir, gen, p);
4210 
4211 out:
4212         if (ret)
4213                 fs_path_free(p);
4214         return ret;
4215 }
4216 
4217 static int __record_new_ref(int num, u64 dir, int index,
4218                             struct fs_path *name,
4219                             void *ctx)
4220 {
4221         struct send_ctx *sctx = ctx;
4222         return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs);
4223 }
4224 
4225 
4226 static int __record_deleted_ref(int num, u64 dir, int index,
4227                                 struct fs_path *name,
4228                                 void *ctx)
4229 {
4230         struct send_ctx *sctx = ctx;
4231         return record_ref(sctx->parent_root, dir, name, ctx,
4232                           &sctx->deleted_refs);
4233 }
4234 
4235 static int record_new_ref(struct send_ctx *sctx)
4236 {
4237         int ret;
4238 
4239         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4240                                 sctx->cmp_key, 0, __record_new_ref, sctx);
4241         if (ret < 0)
4242                 goto out;
4243         ret = 0;
4244 
4245 out:
4246         return ret;
4247 }
4248 
4249 static int record_deleted_ref(struct send_ctx *sctx)
4250 {
4251         int ret;
4252 
4253         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4254                                 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4255         if (ret < 0)
4256                 goto out;
4257         ret = 0;
4258 
4259 out:
4260         return ret;
4261 }
4262 
4263 struct find_ref_ctx {
4264         u64 dir;
4265         u64 dir_gen;
4266         struct btrfs_root *root;
4267         struct fs_path *name;
4268         int found_idx;
4269 };
4270 
4271 static int __find_iref(int num, u64 dir, int index,
4272                        struct fs_path *name,
4273                        void *ctx_)
4274 {
4275         struct find_ref_ctx *ctx = ctx_;
4276         u64 dir_gen;
4277         int ret;
4278 
4279         if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4280             strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4281                 /*
4282                  * To avoid doing extra lookups we'll only do this if everything
4283                  * else matches.
4284                  */
4285                 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4286                                      NULL, NULL, NULL);
4287                 if (ret)
4288                         return ret;
4289                 if (dir_gen != ctx->dir_gen)
4290                         return 0;
4291                 ctx->found_idx = num;
4292                 return 1;
4293         }
4294         return 0;
4295 }
4296 
4297 static int find_iref(struct btrfs_root *root,
4298                      struct btrfs_path *path,
4299                      struct btrfs_key *key,
4300                      u64 dir, u64 dir_gen, struct fs_path *name)
4301 {
4302         int ret;
4303         struct find_ref_ctx ctx;
4304 
4305         ctx.dir = dir;
4306         ctx.name = name;
4307         ctx.dir_gen = dir_gen;
4308         ctx.found_idx = -1;
4309         ctx.root = root;
4310 
4311         ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4312         if (ret < 0)
4313                 return ret;
4314 
4315         if (ctx.found_idx == -1)
4316                 return -ENOENT;
4317 
4318         return ctx.found_idx;
4319 }
4320 
4321 static int __record_changed_new_ref(int num, u64 dir, int index,
4322                                     struct fs_path *name,
4323                                     void *ctx)
4324 {
4325         u64 dir_gen;
4326         int ret;
4327         struct send_ctx *sctx = ctx;
4328 
4329         ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4330                              NULL, NULL, NULL);
4331         if (ret)
4332                 return ret;
4333 
4334         ret = find_iref(sctx->parent_root, sctx->right_path,
4335                         sctx->cmp_key, dir, dir_gen, name);
4336         if (ret == -ENOENT)
4337                 ret = __record_new_ref(num, dir, index, name, sctx);
4338         else if (ret > 0)
4339                 ret = 0;
4340 
4341         return ret;
4342 }
4343 
4344 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4345                                         struct fs_path *name,
4346                                         void *ctx)
4347 {
4348         u64 dir_gen;
4349         int ret;
4350         struct send_ctx *sctx = ctx;
4351 
4352         ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4353                              NULL, NULL, NULL);
4354         if (ret)
4355                 return ret;
4356 
4357         ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4358                         dir, dir_gen, name);
4359         if (ret == -ENOENT)
4360                 ret = __record_deleted_ref(num, dir, index, name, sctx);
4361         else if (ret > 0)
4362                 ret = 0;
4363 
4364         return ret;
4365 }
4366 
4367 static int record_changed_ref(struct send_ctx *sctx)
4368 {
4369         int ret = 0;
4370 
4371         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4372                         sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4373         if (ret < 0)
4374                 goto out;
4375         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4376                         sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4377         if (ret < 0)
4378                 goto out;
4379         ret = 0;
4380 
4381 out:
4382         return ret;
4383 }
4384 
4385 /*
4386  * Record and process all refs at once. Needed when an inode changes the
4387  * generation number, which means that it was deleted and recreated.
4388  */
4389 static int process_all_refs(struct send_ctx *sctx,
4390                             enum btrfs_compare_tree_result cmd)
4391 {
4392         int ret;
4393         struct btrfs_root *root;
4394         struct btrfs_path *path;
4395         struct btrfs_key key;
4396         struct btrfs_key found_key;
4397         struct extent_buffer *eb;
4398         int slot;
4399         iterate_inode_ref_t cb;
4400         int pending_move = 0;
4401 
4402         path = alloc_path_for_send();
4403         if (!path)
4404                 return -ENOMEM;
4405 
4406         if (cmd == BTRFS_COMPARE_TREE_NEW) {
4407                 root = sctx->send_root;
4408                 cb = __record_new_ref;
4409         } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4410                 root = sctx->parent_root;
4411                 cb = __record_deleted_ref;
4412         } else {
4413                 btrfs_err(sctx->send_root->fs_info,
4414                                 "Wrong command %d in process_all_refs", cmd);
4415                 ret = -EINVAL;
4416                 goto out;
4417         }
4418 
4419         key.objectid = sctx->cmp_key->objectid;
4420         key.type = BTRFS_INODE_REF_KEY;
4421         key.offset = 0;
4422         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4423         if (ret < 0)
4424                 goto out;
4425 
4426         while (1) {
4427                 eb = path->nodes[0];
4428                 slot = path->slots[0];
4429                 if (slot >= btrfs_header_nritems(eb)) {
4430                         ret = btrfs_next_leaf(root, path);
4431                         if (ret < 0)
4432                                 goto out;
4433                         else if (ret > 0)
4434                                 break;
4435                         continue;
4436                 }
4437 
4438                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4439 
4440                 if (found_key.objectid != key.objectid ||
4441                     (found_key.type != BTRFS_INODE_REF_KEY &&
4442                      found_key.type != BTRFS_INODE_EXTREF_KEY))
4443                         break;
4444 
4445                 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4446                 if (ret < 0)
4447                         goto out;
4448 
4449                 path->slots[0]++;
4450         }
4451         btrfs_release_path(path);
4452 
4453         /*
4454          * We don't actually care about pending_move as we are simply
4455          * re-creating this inode and will be rename'ing it into place once we
4456          * rename the parent directory.
4457          */
4458         ret = process_recorded_refs(sctx, &pending_move);
4459 out:
4460         btrfs_free_path(path);
4461         return ret;
4462 }
4463 
4464 static int send_set_xattr(struct send_ctx *sctx,
4465                           struct fs_path *path,
4466                           const char *name, int name_len,
4467                           const char *data, int data_len)
4468 {
4469         int ret = 0;
4470 
4471         ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4472         if (ret < 0)
4473                 goto out;
4474 
4475         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4476         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4477         TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4478 
4479         ret = send_cmd(sctx);
4480 
4481 tlv_put_failure:
4482 out:
4483         return ret;
4484 }
4485 
4486 static int send_remove_xattr(struct send_ctx *sctx,
4487                           struct fs_path *path,
4488                           const char *name, int name_len)
4489 {
4490         int ret = 0;
4491 
4492         ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4493         if (ret < 0)
4494                 goto out;
4495 
4496         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4497         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4498 
4499         ret = send_cmd(sctx);
4500 
4501 tlv_put_failure:
4502 out:
4503         return ret;
4504 }
4505 
4506 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4507                                const char *name, int name_len,
4508                                const char *data, int data_len,
4509                                u8 type, void *ctx)
4510 {
4511         int ret;
4512         struct send_ctx *sctx = ctx;
4513         struct fs_path *p;
4514         struct posix_acl_xattr_header dummy_acl;
4515 
4516         p = fs_path_alloc();
4517         if (!p)
4518                 return -ENOMEM;
4519 
4520         /*
4521          * This hack is needed because empty acls are stored as zero byte
4522          * data in xattrs. Problem with that is, that receiving these zero byte
4523          * acls will fail later. To fix this, we send a dummy acl list that
4524          * only contains the version number and no entries.
4525          */
4526         if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4527             !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4528                 if (data_len == 0) {
4529                         dummy_acl.a_version =
4530                                         cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4531                         data = (char *)&dummy_acl;
4532                         data_len = sizeof(dummy_acl);
4533                 }
4534         }
4535 
4536         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4537         if (ret < 0)
4538                 goto out;
4539 
4540         ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4541 
4542 out:
4543         fs_path_free(p);
4544         return ret;
4545 }
4546 
4547 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4548                                    const char *name, int name_len,
4549                                    const char *data, int data_len,
4550                                    u8 type, void *ctx)
4551 {
4552         int ret;
4553         struct send_ctx *sctx = ctx;
4554         struct fs_path *p;
4555 
4556         p = fs_path_alloc();
4557         if (!p)
4558                 return -ENOMEM;
4559 
4560         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4561         if (ret < 0)
4562                 goto out;
4563 
4564         ret = send_remove_xattr(sctx, p, name, name_len);
4565 
4566 out:
4567         fs_path_free(p);
4568         return ret;
4569 }
4570 
4571 static int process_new_xattr(struct send_ctx *sctx)
4572 {
4573         int ret = 0;
4574 
4575         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4576                                __process_new_xattr, sctx);
4577 
4578         return ret;
4579 }
4580 
4581 static int process_deleted_xattr(struct send_ctx *sctx)
4582 {
4583         return iterate_dir_item(sctx->parent_root, sctx->right_path,
4584                                 __process_deleted_xattr, sctx);
4585 }
4586 
4587 struct find_xattr_ctx {
4588         const char *name;
4589         int name_len;
4590         int found_idx;
4591         char *found_data;
4592         int found_data_len;
4593 };
4594 
4595 static int __find_xattr(int num, struct btrfs_key *di_key,
4596                         const char *name, int name_len,
4597                         const char *data, int data_len,
4598                         u8 type, void *vctx)
4599 {
4600         struct find_xattr_ctx *ctx = vctx;
4601 
4602         if (name_len == ctx->name_len &&
4603             strncmp(name, ctx->name, name_len) == 0) {
4604                 ctx->found_idx = num;
4605                 ctx->found_data_len = data_len;
4606                 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4607                 if (!ctx->found_data)
4608                         return -ENOMEM;
4609                 return 1;
4610         }
4611         return 0;
4612 }
4613 
4614 static int find_xattr(struct btrfs_root *root,
4615                       struct btrfs_path *path,
4616                       struct btrfs_key *key,
4617                       const char *name, int name_len,
4618                       char **data, int *data_len)
4619 {
4620         int ret;
4621         struct find_xattr_ctx ctx;
4622 
4623         ctx.name = name;
4624         ctx.name_len = name_len;
4625         ctx.found_idx = -1;
4626         ctx.found_data = NULL;
4627         ctx.found_data_len = 0;
4628 
4629         ret = iterate_dir_item(root, path, __find_xattr, &ctx);
4630         if (ret < 0)
4631                 return ret;
4632 
4633         if (ctx.found_idx == -1)
4634                 return -ENOENT;
4635         if (data) {
4636                 *data = ctx.found_data;
4637                 *data_len = ctx.found_data_len;
4638         } else {
4639                 kfree(ctx.found_data);
4640         }
4641         return ctx.found_idx;
4642 }
4643 
4644 
4645 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4646                                        const char *name, int name_len,
4647                                        const char *data, int data_len,
4648                                        u8 type, void *ctx)
4649 {
4650         int ret;
4651         struct send_ctx *sctx = ctx;
4652         char *found_data = NULL;
4653         int found_data_len  = 0;
4654 
4655         ret = find_xattr(sctx->parent_root, sctx->right_path,
4656                          sctx->cmp_key, name, name_len, &found_data,
4657                          &found_data_len);
4658         if (ret == -ENOENT) {
4659                 ret = __process_new_xattr(num, di_key, name, name_len, data,
4660                                 data_len, type, ctx);
4661         } else if (ret >= 0) {
4662                 if (data_len != found_data_len ||
4663                     memcmp(data, found_data, data_len)) {
4664                         ret = __process_new_xattr(num, di_key, name, name_len,
4665                                         data, data_len, type, ctx);
4666                 } else {
4667                         ret = 0;
4668                 }
4669         }
4670 
4671         kfree(found_data);
4672         return ret;
4673 }
4674 
4675 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4676                                            const char *name, int name_len,
4677                                            const char *data, int data_len,
4678                                            u8 type, void *ctx)
4679 {
4680         int ret;
4681         struct send_ctx *sctx = ctx;
4682 
4683         ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4684                          name, name_len, NULL, NULL);
4685         if (ret == -ENOENT)
4686                 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4687                                 data_len, type, ctx);
4688         else if (ret >= 0)
4689                 ret = 0;
4690 
4691         return ret;
4692 }
4693 
4694 static int process_changed_xattr(struct send_ctx *sctx)
4695 {
4696         int ret = 0;
4697 
4698         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4699                         __process_changed_new_xattr, sctx);
4700         if (ret < 0)
4701                 goto out;
4702         ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4703                         __process_changed_deleted_xattr, sctx);
4704 
4705 out:
4706         return ret;
4707 }
4708 
4709 static int process_all_new_xattrs(struct send_ctx *sctx)
4710 {
4711         int ret;
4712         struct btrfs_root *root;
4713         struct btrfs_path *path;
4714         struct btrfs_key key;
4715         struct btrfs_key found_key;
4716         struct extent_buffer *eb;
4717         int slot;
4718 
4719         path = alloc_path_for_send();
4720         if (!path)
4721                 return -ENOMEM;
4722 
4723         root = sctx->send_root;
4724 
4725         key.objectid = sctx->cmp_key->objectid;
4726         key.type = BTRFS_XATTR_ITEM_KEY;
4727         key.offset = 0;
4728         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4729         if (ret < 0)
4730                 goto out;
4731 
4732         while (1) {
4733                 eb = path->nodes[0];
4734                 slot = path->slots[0];
4735                 if (slot >= btrfs_header_nritems(eb)) {
4736                         ret = btrfs_next_leaf(root, path);
4737                         if (ret < 0) {
4738                                 goto out;
4739                         } else if (ret > 0) {
4740                                 ret = 0;
4741                                 break;
4742                         }
4743                         continue;
4744                 }
4745 
4746                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4747                 if (found_key.objectid != key.objectid ||
4748                     found_key.type != key.type) {
4749                         ret = 0;
4750                         goto out;
4751                 }
4752 
4753                 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
4754                 if (ret < 0)
4755                         goto out;
4756 
4757                 path->slots[0]++;
4758         }
4759 
4760 out:
4761         btrfs_free_path(path);
4762         return ret;
4763 }
4764 
4765 static ssize_t fill_read_buf(struct send_ctx *sctx, u64 offset, u32 len)
4766 {
4767         struct btrfs_root *root = sctx->send_root;
4768         struct btrfs_fs_info *fs_info = root->fs_info;
4769         struct inode *inode;
4770         struct page *page;
4771         char *addr;
4772         struct btrfs_key key;
4773         pgoff_t index = offset >> PAGE_SHIFT;
4774         pgoff_t last_index;
4775         unsigned pg_offset = offset & ~PAGE_MASK;
4776         ssize_t ret = 0;
4777 
4778         key.objectid = sctx->cur_ino;
4779         key.type = BTRFS_INODE_ITEM_KEY;
4780         key.offset = 0;
4781 
4782         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4783         if (IS_ERR(inode))
4784                 return PTR_ERR(inode);
4785 
4786         if (offset + len > i_size_read(inode)) {
4787                 if (offset > i_size_read(inode))
4788                         len = 0;
4789                 else
4790                         len = offset - i_size_read(inode);
4791         }
4792         if (len == 0)
4793                 goto out;
4794 
4795         last_index = (offset + len - 1) >> PAGE_SHIFT;
4796 
4797         /* initial readahead */
4798         memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4799         file_ra_state_init(&sctx->ra, inode->i_mapping);
4800 
4801         while (index <= last_index) {
4802                 unsigned cur_len = min_t(unsigned, len,
4803                                          PAGE_SIZE - pg_offset);
4804 
4805                 page = find_lock_page(inode->i_mapping, index);
4806                 if (!page) {
4807                         page_cache_sync_readahead(inode->i_mapping, &sctx->ra,
4808                                 NULL, index, last_index + 1 - index);
4809 
4810                         page = find_or_create_page(inode->i_mapping, index,
4811                                         GFP_KERNEL);
4812                         if (!page) {
4813                                 ret = -ENOMEM;
4814                                 break;
4815                         }
4816                 }
4817 
4818                 if (PageReadahead(page)) {
4819                         page_cache_async_readahead(inode->i_mapping, &sctx->ra,
4820                                 NULL, page, index, last_index + 1 - index);
4821                 }
4822 
4823                 if (!PageUptodate(page)) {
4824                         btrfs_readpage(NULL, page);
4825                         lock_page(page);
4826                         if (!PageUptodate(page)) {
4827                                 unlock_page(page);
4828                                 put_page(page);
4829                                 ret = -EIO;
4830                                 break;
4831                         }
4832                 }
4833 
4834                 addr = kmap(page);
4835                 memcpy(sctx->read_buf + ret, addr + pg_offset, cur_len);
4836                 kunmap(page);
4837                 unlock_page(page);
4838                 put_page(page);
4839                 index++;
4840                 pg_offset = 0;
4841                 len -= cur_len;
4842                 ret += cur_len;
4843         }
4844 out:
4845         iput(inode);
4846         return ret;
4847 }
4848 
4849 /*
4850  * Read some bytes from the current inode/file and send a write command to
4851  * user space.
4852  */
4853 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
4854 {
4855         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4856         int ret = 0;
4857         struct fs_path *p;
4858         ssize_t num_read = 0;
4859 
4860         p = fs_path_alloc();
4861         if (!p)
4862                 return -ENOMEM;
4863 
4864         btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
4865 
4866         num_read = fill_read_buf(sctx, offset, len);
4867         if (num_read <= 0) {
4868                 if (num_read < 0)
4869                         ret = num_read;
4870                 goto out;
4871         }
4872 
4873         ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4874         if (ret < 0)
4875                 goto out;
4876 
4877         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4878         if (ret < 0)
4879                 goto out;
4880 
4881         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4882         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4883         TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, num_read);
4884 
4885         ret = send_cmd(sctx);
4886 
4887 tlv_put_failure:
4888 out:
4889         fs_path_free(p);
4890         if (ret < 0)
4891                 return ret;
4892         return num_read;
4893 }
4894 
4895 /*
4896  * Send a clone command to user space.
4897  */
4898 static int send_clone(struct send_ctx *sctx,
4899                       u64 offset, u32 len,
4900                       struct clone_root *clone_root)
4901 {
4902         int ret = 0;
4903         struct fs_path *p;
4904         u64 gen;
4905 
4906         btrfs_debug(sctx->send_root->fs_info,
4907                     "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
4908                     offset, len, clone_root->root->objectid, clone_root->ino,
4909                     clone_root->offset);
4910 
4911         p = fs_path_alloc();
4912         if (!p)
4913                 return -ENOMEM;
4914 
4915         ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
4916         if (ret < 0)
4917                 goto out;
4918 
4919         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4920         if (ret < 0)
4921                 goto out;
4922 
4923         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4924         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
4925         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4926 
4927         if (clone_root->root == sctx->send_root) {
4928                 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
4929                                 &gen, NULL, NULL, NULL, NULL);
4930                 if (ret < 0)
4931                         goto out;
4932                 ret = get_cur_path(sctx, clone_root->ino, gen, p);
4933         } else {
4934                 ret = get_inode_path(clone_root->root, clone_root->ino, p);
4935         }
4936         if (ret < 0)
4937                 goto out;
4938 
4939         /*
4940          * If the parent we're using has a received_uuid set then use that as
4941          * our clone source as that is what we will look for when doing a
4942          * receive.
4943          *
4944          * This covers the case that we create a snapshot off of a received
4945          * subvolume and then use that as the parent and try to receive on a
4946          * different host.
4947          */
4948         if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
4949                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4950                              clone_root->root->root_item.received_uuid);
4951         else
4952                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4953                              clone_root->root->root_item.uuid);
4954         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
4955                     le64_to_cpu(clone_root->root->root_item.ctransid));
4956         TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
4957         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
4958                         clone_root->offset);
4959 
4960         ret = send_cmd(sctx);
4961 
4962 tlv_put_failure:
4963 out:
4964         fs_path_free(p);
4965         return ret;
4966 }
4967 
4968 /*
4969  * Send an update extent command to user space.
4970  */
4971 static int send_update_extent(struct send_ctx *sctx,
4972                               u64 offset, u32 len)
4973 {
4974         int ret = 0;
4975         struct fs_path *p;
4976 
4977         p = fs_path_alloc();
4978         if (!p)
4979                 return -ENOMEM;
4980 
4981         ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
4982         if (ret < 0)
4983                 goto out;
4984 
4985         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4986         if (ret < 0)
4987                 goto out;
4988 
4989         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4990         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4991         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
4992 
4993         ret = send_cmd(sctx);
4994 
4995 tlv_put_failure:
4996 out:
4997         fs_path_free(p);
4998         return ret;
4999 }
5000 
5001 static int send_hole(struct send_ctx *sctx, u64 end)
5002 {
5003         struct fs_path *p = NULL;
5004         u64 offset = sctx->cur_inode_last_extent;
5005         u64 len;
5006         int ret = 0;
5007 
5008         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5009                 return send_update_extent(sctx, offset, end - offset);
5010 
5011         p = fs_path_alloc();
5012         if (!p)
5013                 return -ENOMEM;
5014         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5015         if (ret < 0)
5016                 goto tlv_put_failure;
5017         memset(sctx->read_buf, 0, BTRFS_SEND_READ_SIZE);
5018         while (offset < end) {
5019                 len = min_t(u64, end - offset, BTRFS_SEND_READ_SIZE);
5020 
5021                 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5022                 if (ret < 0)
5023                         break;
5024                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5025                 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5026                 TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, len);
5027                 ret = send_cmd(sctx);
5028                 if (ret < 0)
5029                         break;
5030                 offset += len;
5031         }
5032 tlv_put_failure:
5033         fs_path_free(p);
5034         return ret;
5035 }
5036 
5037 static int send_extent_data(struct send_ctx *sctx,
5038                             const u64 offset,
5039                             const u64 len)
5040 {
5041         u64 sent = 0;
5042 
5043         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5044                 return send_update_extent(sctx, offset, len);
5045 
5046         while (sent < len) {
5047                 u64 size = len - sent;
5048                 int ret;
5049 
5050                 if (size > BTRFS_SEND_READ_SIZE)
5051                         size = BTRFS_SEND_READ_SIZE;
5052                 ret = send_write(sctx, offset + sent, size);
5053                 if (ret < 0)
5054                         return ret;
5055                 if (!ret)
5056                         break;
5057                 sent += ret;
5058         }
5059         return 0;
5060 }
5061 
5062 static int clone_range(struct send_ctx *sctx,
5063                        struct clone_root *clone_root,
5064                        const u64 disk_byte,
5065                        u64 data_offset,
5066                        u64 offset,
5067                        u64 len)
5068 {
5069         struct btrfs_path *path;
5070         struct btrfs_key key;
5071         int ret;
5072 
5073         /*
5074          * Prevent cloning from a zero offset with a length matching the sector
5075          * size because in some scenarios this will make the receiver fail.
5076          *
5077          * For example, if in the source filesystem the extent at offset 0
5078          * has a length of sectorsize and it was written using direct IO, then
5079          * it can never be an inline extent (even if compression is enabled).
5080          * Then this extent can be cloned in the original filesystem to a non
5081          * zero file offset, but it may not be possible to clone in the
5082          * destination filesystem because it can be inlined due to compression
5083          * on the destination filesystem (as the receiver's write operations are
5084          * always done using buffered IO). The same happens when the original
5085          * filesystem does not have compression enabled but the destination
5086          * filesystem has.
5087          */
5088         if (clone_root->offset == 0 &&
5089             len == sctx->send_root->fs_info->sectorsize)
5090                 return send_extent_data(sctx, offset, len);
5091 
5092         path = alloc_path_for_send();
5093         if (!path)
5094                 return -ENOMEM;
5095 
5096         /*
5097          * We can't send a clone operation for the entire range if we find
5098          * extent items in the respective range in the source file that
5099          * refer to different extents or if we find holes.
5100          * So check for that and do a mix of clone and regular write/copy
5101          * operations if needed.
5102          *
5103          * Example:
5104          *
5105          * mkfs.btrfs -f /dev/sda
5106          * mount /dev/sda /mnt
5107          * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5108          * cp --reflink=always /mnt/foo /mnt/bar
5109          * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5110          * btrfs subvolume snapshot -r /mnt /mnt/snap
5111          *
5112          * If when we send the snapshot and we are processing file bar (which
5113          * has a higher inode number than foo) we blindly send a clone operation
5114          * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5115          * a file bar that matches the content of file foo - iow, doesn't match
5116          * the content from bar in the original filesystem.
5117          */
5118         key.objectid = clone_root->ino;
5119         key.type = BTRFS_EXTENT_DATA_KEY;
5120         key.offset = clone_root->offset;
5121         ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5122         if (ret < 0)
5123                 goto out;
5124         if (ret > 0 && path->slots[0] > 0) {
5125                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5126                 if (key.objectid == clone_root->ino &&
5127                     key.type == BTRFS_EXTENT_DATA_KEY)
5128                         path->slots[0]--;
5129         }
5130 
5131         while (true) {
5132                 struct extent_buffer *leaf = path->nodes[0];
5133                 int slot = path->slots[0];
5134                 struct btrfs_file_extent_item *ei;
5135                 u8 type;
5136                 u64 ext_len;
5137                 u64 clone_len;
5138 
5139                 if (slot >= btrfs_header_nritems(leaf)) {
5140                         ret = btrfs_next_leaf(clone_root->root, path);
5141                         if (ret < 0)
5142                                 goto out;
5143                         else if (ret > 0)
5144                                 break;
5145                         continue;
5146                 }
5147 
5148                 btrfs_item_key_to_cpu(leaf, &key, slot);
5149 
5150                 /*
5151                  * We might have an implicit trailing hole (NO_HOLES feature
5152                  * enabled). We deal with it after leaving this loop.
5153                  */
5154                 if (key.objectid != clone_root->ino ||
5155                     key.type != BTRFS_EXTENT_DATA_KEY)
5156                         break;
5157 
5158                 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5159                 type = btrfs_file_extent_type(leaf, ei);
5160                 if (type == BTRFS_FILE_EXTENT_INLINE) {
5161                         ext_len = btrfs_file_extent_inline_len(leaf, slot, ei);
5162                         ext_len = PAGE_ALIGN(ext_len);
5163                 } else {
5164                         ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5165                 }
5166 
5167                 if (key.offset + ext_len <= clone_root->offset)
5168                         goto next;
5169 
5170                 if (key.offset > clone_root->offset) {
5171                         /* Implicit hole, NO_HOLES feature enabled. */
5172                         u64 hole_len = key.offset - clone_root->offset;
5173 
5174                         if (hole_len > len)
5175                                 hole_len = len;
5176                         ret = send_extent_data(sctx, offset, hole_len);
5177                         if (ret < 0)
5178                                 goto out;
5179 
5180                         len -= hole_len;
5181                         if (len == 0)
5182                                 break;
5183                         offset += hole_len;
5184                         clone_root->offset += hole_len;
5185                         data_offset += hole_len;
5186                 }
5187 
5188                 if (key.offset >= clone_root->offset + len)
5189                         break;
5190 
5191                 clone_len = min_t(u64, ext_len, len);
5192 
5193                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5194                     btrfs_file_extent_offset(leaf, ei) == data_offset)
5195                         ret = send_clone(sctx, offset, clone_len, clone_root);
5196                 else
5197                         ret = send_extent_data(sctx, offset, clone_len);
5198 
5199                 if (ret < 0)
5200                         goto out;
5201 
5202                 len -= clone_len;
5203                 if (len == 0)
5204                         break;
5205                 offset += clone_len;
5206                 clone_root->offset += clone_len;
5207                 data_offset += clone_len;
5208 next:
5209                 path->slots[0]++;
5210         }
5211 
5212         if (len > 0)
5213                 ret = send_extent_data(sctx, offset, len);
5214         else
5215                 ret = 0;
5216 out:
5217         btrfs_free_path(path);
5218         return ret;
5219 }
5220 
5221 static int send_write_or_clone(struct send_ctx *sctx,
5222                                struct btrfs_path *path,
5223                                struct btrfs_key *key,
5224                                struct clone_root *clone_root)
5225 {
5226         int ret = 0;
5227         struct btrfs_file_extent_item *ei;
5228         u64 offset = key->offset;
5229         u64 len;
5230         u8 type;
5231         u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5232 
5233         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5234                         struct btrfs_file_extent_item);
5235         type = btrfs_file_extent_type(path->nodes[0], ei);
5236         if (type == BTRFS_FILE_EXTENT_INLINE) {
5237                 len = btrfs_file_extent_inline_len(path->nodes[0],
5238                                                    path->slots[0], ei);
5239                 /*
5240                  * it is possible the inline item won't cover the whole page,
5241                  * but there may be items after this page.  Make
5242                  * sure to send the whole thing
5243                  */
5244                 len = PAGE_ALIGN(len);
5245         } else {
5246                 len = btrfs_file_extent_num_bytes(path->nodes[0], ei);
5247         }
5248 
5249         if (offset + len > sctx->cur_inode_size)
5250                 len = sctx->cur_inode_size - offset;
5251         if (len == 0) {
5252                 ret = 0;
5253                 goto out;
5254         }
5255 
5256         if (clone_root && IS_ALIGNED(offset + len, bs)) {
5257                 u64 disk_byte;
5258                 u64 data_offset;
5259 
5260                 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5261                 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5262                 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5263                                   offset, len);
5264         } else {
5265                 ret = send_extent_data(sctx, offset, len);
5266         }
5267 out:
5268         return ret;
5269 }
5270 
5271 static int is_extent_unchanged(struct send_ctx *sctx,
5272                                struct btrfs_path *left_path,
5273                                struct btrfs_key *ekey)
5274 {
5275         int ret = 0;
5276         struct btrfs_key key;
5277         struct btrfs_path *path = NULL;
5278         struct extent_buffer *eb;
5279         int slot;
5280         struct btrfs_key found_key;
5281         struct btrfs_file_extent_item *ei;
5282         u64 left_disknr;
5283         u64 right_disknr;
5284         u64 left_offset;
5285         u64 right_offset;
5286         u64 left_offset_fixed;
5287         u64 left_len;
5288         u64 right_len;
5289         u64 left_gen;
5290         u64 right_gen;
5291         u8 left_type;
5292         u8 right_type;
5293 
5294         path = alloc_path_for_send();
5295         if (!path)
5296                 return -ENOMEM;
5297 
5298         eb = left_path->nodes[0];
5299         slot = left_path->slots[0];
5300         ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5301         left_type = btrfs_file_extent_type(eb, ei);
5302 
5303         if (left_type != BTRFS_FILE_EXTENT_REG) {
5304                 ret = 0;
5305                 goto out;
5306         }
5307         left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5308         left_len = btrfs_file_extent_num_bytes(eb, ei);
5309         left_offset = btrfs_file_extent_offset(eb, ei);
5310         left_gen = btrfs_file_extent_generation(eb, ei);
5311 
5312         /*
5313          * Following comments will refer to these graphics. L is the left
5314          * extents which we are checking at the moment. 1-8 are the right
5315          * extents that we iterate.
5316          *
5317          *       |-----L-----|
5318          * |-1-|-2a-|-3-|-4-|-5-|-6-|
5319          *
5320          *       |-----L-----|
5321          * |--1--|-2b-|...(same as above)
5322          *
5323          * Alternative situation. Happens on files where extents got split.
5324          *       |-----L-----|
5325          * |-----------7-----------|-6-|
5326          *
5327          * Alternative situation. Happens on files which got larger.
5328          *       |-----L-----|
5329          * |-8-|
5330          * Nothing follows after 8.
5331          */
5332 
5333         key.objectid = ekey->objectid;
5334         key.type = BTRFS_EXTENT_DATA_KEY;
5335         key.offset = ekey->offset;
5336         ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5337         if (ret < 0)
5338                 goto out;
5339         if (ret) {
5340                 ret = 0;
5341                 goto out;
5342         }
5343 
5344         /*
5345          * Handle special case where the right side has no extents at all.
5346          */
5347         eb = path->nodes[0];
5348         slot = path->slots[0];
5349         btrfs_item_key_to_cpu(eb, &found_key, slot);
5350         if (found_key.objectid != key.objectid ||
5351             found_key.type != key.type) {
5352                 /* If we're a hole then just pretend nothing changed */
5353                 ret = (left_disknr) ? 0 : 1;
5354                 goto out;
5355         }
5356 
5357         /*
5358          * We're now on 2a, 2b or 7.
5359          */
5360         key = found_key;
5361         while (key.offset < ekey->offset + left_len) {
5362                 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5363                 right_type = btrfs_file_extent_type(eb, ei);
5364                 if (right_type != BTRFS_FILE_EXTENT_REG &&
5365                     right_type != BTRFS_FILE_EXTENT_INLINE) {
5366                         ret = 0;
5367                         goto out;
5368                 }
5369 
5370                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5371                         right_len = btrfs_file_extent_inline_len(eb, slot, ei);
5372                         right_len = PAGE_ALIGN(right_len);
5373                 } else {
5374                         right_len = btrfs_file_extent_num_bytes(eb, ei);
5375                 }
5376 
5377                 /*
5378                  * Are we at extent 8? If yes, we know the extent is changed.
5379                  * This may only happen on the first iteration.
5380                  */
5381                 if (found_key.offset + right_len <= ekey->offset) {
5382                         /* If we're a hole just pretend nothing changed */
5383                         ret = (left_disknr) ? 0 : 1;
5384                         goto out;
5385                 }
5386 
5387                 /*
5388                  * We just wanted to see if when we have an inline extent, what
5389                  * follows it is a regular extent (wanted to check the above
5390                  * condition for inline extents too). This should normally not
5391                  * happen but it's possible for example when we have an inline
5392                  * compressed extent representing data with a size matching
5393                  * the page size (currently the same as sector size).
5394                  */
5395                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5396                         ret = 0;
5397                         goto out;
5398                 }
5399 
5400                 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5401                 right_offset = btrfs_file_extent_offset(eb, ei);
5402                 right_gen = btrfs_file_extent_generation(eb, ei);
5403 
5404                 left_offset_fixed = left_offset;
5405                 if (key.offset < ekey->offset) {
5406                         /* Fix the right offset for 2a and 7. */
5407                         right_offset += ekey->offset - key.offset;
5408                 } else {
5409                         /* Fix the left offset for all behind 2a and 2b */
5410                         left_offset_fixed += key.offset - ekey->offset;
5411                 }
5412 
5413                 /*
5414                  * Check if we have the same extent.
5415                  */
5416                 if (left_disknr != right_disknr ||
5417                     left_offset_fixed != right_offset ||
5418                     left_gen != right_gen) {
5419                         ret = 0;
5420                         goto out;
5421                 }
5422 
5423                 /*
5424                  * Go to the next extent.
5425                  */
5426                 ret = btrfs_next_item(sctx->parent_root, path);
5427                 if (ret < 0)
5428                         goto out;
5429                 if (!ret) {
5430                         eb = path->nodes[0];
5431                         slot = path->slots[0];
5432                         btrfs_item_key_to_cpu(eb, &found_key, slot);
5433                 }
5434                 if (ret || found_key.objectid != key.objectid ||
5435                     found_key.type != key.type) {
5436                         key.offset += right_len;
5437                         break;
5438                 }
5439                 if (found_key.offset != key.offset + right_len) {
5440                         ret = 0;
5441                         goto out;
5442                 }
5443                 key = found_key;
5444         }
5445 
5446         /*
5447          * We're now behind the left extent (treat as unchanged) or at the end
5448          * of the right side (treat as changed).
5449          */
5450         if (key.offset >= ekey->offset + left_len)
5451                 ret = 1;
5452         else
5453                 ret = 0;
5454 
5455 
5456 out:
5457         btrfs_free_path(path);
5458         return ret;
5459 }
5460 
5461 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5462 {
5463         struct btrfs_path *path;
5464         struct btrfs_root *root = sctx->send_root;
5465         struct btrfs_file_extent_item *fi;
5466         struct btrfs_key key;
5467         u64 extent_end;
5468         u8 type;
5469         int ret;
5470 
5471         path = alloc_path_for_send();
5472         if (!path)
5473                 return -ENOMEM;
5474 
5475         sctx->cur_inode_last_extent = 0;
5476 
5477         key.objectid = sctx->cur_ino;
5478         key.type = BTRFS_EXTENT_DATA_KEY;
5479         key.offset = offset;
5480         ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5481         if (ret < 0)
5482                 goto out;
5483         ret = 0;
5484         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5485         if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5486                 goto out;
5487 
5488         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5489                             struct btrfs_file_extent_item);
5490         type = btrfs_file_extent_type(path->nodes[0], fi);
5491         if (type == BTRFS_FILE_EXTENT_INLINE) {
5492                 u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5493                                                         path->slots[0], fi);
5494                 extent_end = ALIGN(key.offset + size,
5495                                    sctx->send_root->fs_info->sectorsize);
5496         } else {
5497                 extent_end = key.offset +
5498                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5499         }
5500         sctx->cur_inode_last_extent = extent_end;
5501 out:
5502         btrfs_free_path(path);
5503         return ret;
5504 }
5505 
5506 static int range_is_hole_in_parent(struct send_ctx *sctx,
5507                                    const u64 start,
5508                                    const u64 end)
5509 {
5510         struct btrfs_path *path;
5511         struct btrfs_key key;
5512         struct btrfs_root *root = sctx->parent_root;
5513         u64 search_start = start;
5514         int ret;
5515 
5516         path = alloc_path_for_send();
5517         if (!path)
5518                 return -ENOMEM;
5519 
5520         key.objectid = sctx->cur_ino;
5521         key.type = BTRFS_EXTENT_DATA_KEY;
5522         key.offset = search_start;
5523         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5524         if (ret < 0)
5525                 goto out;
5526         if (ret > 0 && path->slots[0] > 0)
5527                 path->slots[0]--;
5528 
5529         while (search_start < end) {
5530                 struct extent_buffer *leaf = path->nodes[0];
5531                 int slot = path->slots[0];
5532                 struct btrfs_file_extent_item *fi;
5533                 u64 extent_end;
5534 
5535                 if (slot >= btrfs_header_nritems(leaf)) {
5536                         ret = btrfs_next_leaf(root, path);
5537                         if (ret < 0)
5538                                 goto out;
5539                         else if (ret > 0)
5540                                 break;
5541                         continue;
5542                 }
5543 
5544                 btrfs_item_key_to_cpu(leaf, &key, slot);
5545                 if (key.objectid < sctx->cur_ino ||
5546                     key.type < BTRFS_EXTENT_DATA_KEY)
5547                         goto next;
5548                 if (key.objectid > sctx->cur_ino ||
5549                     key.type > BTRFS_EXTENT_DATA_KEY ||
5550                     key.offset >= end)
5551                         break;
5552 
5553                 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5554                 if (btrfs_file_extent_type(leaf, fi) ==
5555                     BTRFS_FILE_EXTENT_INLINE) {
5556                         u64 size = btrfs_file_extent_inline_len(leaf, slot, fi);
5557 
5558                         extent_end = ALIGN(key.offset + size,
5559                                            root->fs_info->sectorsize);
5560                 } else {
5561                         extent_end = key.offset +
5562                                 btrfs_file_extent_num_bytes(leaf, fi);
5563                 }
5564                 if (extent_end <= start)
5565                         goto next;
5566                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5567                         search_start = extent_end;
5568                         goto next;
5569                 }
5570                 ret = 0;
5571                 goto out;
5572 next:
5573                 path->slots[0]++;
5574         }
5575         ret = 1;
5576 out:
5577         btrfs_free_path(path);
5578         return ret;
5579 }
5580 
5581 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5582                            struct btrfs_key *key)
5583 {
5584         struct btrfs_file_extent_item *fi;
5585         u64 extent_end;
5586         u8 type;
5587         int ret = 0;
5588 
5589         if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5590                 return 0;
5591 
5592         if (sctx->cur_inode_last_extent == (u64)-1) {
5593                 ret = get_last_extent(sctx, key->offset - 1);
5594                 if (ret)
5595                         return ret;
5596         }
5597 
5598         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5599                             struct btrfs_file_extent_item);
5600         type = btrfs_file_extent_type(path->nodes[0], fi);
5601         if (type == BTRFS_FILE_EXTENT_INLINE) {
5602                 u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5603                                                         path->slots[0], fi);
5604                 extent_end = ALIGN(key->offset + size,
5605                                    sctx->send_root->fs_info->sectorsize);
5606         } else {
5607                 extent_end = key->offset +
5608                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5609         }
5610 
5611         if (path->slots[0] == 0 &&
5612             sctx->cur_inode_last_extent < key->offset) {
5613                 /*
5614                  * We might have skipped entire leafs that contained only
5615                  * file extent items for our current inode. These leafs have
5616                  * a generation number smaller (older) than the one in the
5617                  * current leaf and the leaf our last extent came from, and
5618                  * are located between these 2 leafs.
5619                  */
5620                 ret = get_last_extent(sctx, key->offset - 1);
5621                 if (ret)
5622                         return ret;
5623         }
5624 
5625         if (sctx->cur_inode_last_extent < key->offset) {
5626                 ret = range_is_hole_in_parent(sctx,
5627                                               sctx->cur_inode_last_extent,
5628                                               key->offset);
5629                 if (ret < 0)
5630                         return ret;
5631                 else if (ret == 0)
5632                         ret = send_hole(sctx, key->offset);
5633                 else
5634                         ret = 0;
5635         }
5636         sctx->cur_inode_last_extent = extent_end;
5637         return ret;
5638 }
5639 
5640 static int process_extent(struct send_ctx *sctx,
5641                           struct btrfs_path *path,
5642                           struct btrfs_key *key)
5643 {
5644         struct clone_root *found_clone = NULL;
5645         int ret = 0;
5646 
5647         if (S_ISLNK(sctx->cur_inode_mode))
5648                 return 0;
5649 
5650         if (sctx->parent_root && !sctx->cur_inode_new) {
5651                 ret = is_extent_unchanged(sctx, path, key);
5652                 if (ret < 0)
5653                         goto out;
5654                 if (ret) {
5655                         ret = 0;
5656                         goto out_hole;
5657                 }
5658         } else {
5659                 struct btrfs_file_extent_item *ei;
5660                 u8 type;
5661 
5662                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5663                                     struct btrfs_file_extent_item);
5664                 type = btrfs_file_extent_type(path->nodes[0], ei);
5665                 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5666                     type == BTRFS_FILE_EXTENT_REG) {
5667                         /*
5668                          * The send spec does not have a prealloc command yet,
5669                          * so just leave a hole for prealloc'ed extents until
5670                          * we have enough commands queued up to justify rev'ing
5671                          * the send spec.
5672                          */
5673                         if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5674                                 ret = 0;
5675                                 goto out;
5676                         }
5677 
5678                         /* Have a hole, just skip it. */
5679                         if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5680                                 ret = 0;
5681                                 goto out;
5682                         }
5683                 }
5684         }
5685 
5686         ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5687                         sctx->cur_inode_size, &found_clone);
5688         if (ret != -ENOENT && ret < 0)
5689                 goto out;
5690 
5691         ret = send_write_or_clone(sctx, path, key, found_clone);
5692         if (ret)
5693                 goto out;
5694 out_hole:
5695         ret = maybe_send_hole(sctx, path, key);
5696 out:
5697         return ret;
5698 }
5699 
5700 static int process_all_extents(struct send_ctx *sctx)
5701 {
5702         int ret;
5703         struct btrfs_root *root;
5704         struct btrfs_path *path;
5705         struct btrfs_key key;
5706         struct btrfs_key found_key;
5707         struct extent_buffer *eb;
5708         int slot;
5709 
5710         root = sctx->send_root;
5711         path = alloc_path_for_send();
5712         if (!path)
5713                 return -ENOMEM;
5714 
5715         key.objectid = sctx->cmp_key->objectid;
5716         key.type = BTRFS_EXTENT_DATA_KEY;
5717         key.offset = 0;
5718         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5719         if (ret < 0)
5720                 goto out;
5721 
5722         while (1) {
5723                 eb = path->nodes[0];
5724                 slot = path->slots[0];
5725 
5726                 if (slot >= btrfs_header_nritems(eb)) {
5727                         ret = btrfs_next_leaf(root, path);
5728                         if (ret < 0) {
5729                                 goto out;
5730                         } else if (ret > 0) {
5731                                 ret = 0;
5732                                 break;
5733                         }
5734                         continue;
5735                 }
5736 
5737                 btrfs_item_key_to_cpu(eb, &found_key, slot);
5738 
5739                 if (found_key.objectid != key.objectid ||
5740                     found_key.type != key.type) {
5741                         ret = 0;
5742                         goto out;
5743                 }
5744 
5745                 ret = process_extent(sctx, path, &found_key);
5746                 if (ret < 0)
5747                         goto out;
5748 
5749                 path->slots[0]++;
5750         }
5751 
5752 out:
5753         btrfs_free_path(path);
5754         return ret;
5755 }
5756 
5757 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
5758                                            int *pending_move,
5759                                            int *refs_processed)
5760 {
5761         int ret = 0;
5762 
5763         if (sctx->cur_ino == 0)
5764                 goto out;
5765         if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
5766             sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
5767                 goto out;
5768         if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
5769                 goto out;
5770 
5771         ret = process_recorded_refs(sctx, pending_move);
5772         if (ret < 0)
5773                 goto out;
5774 
5775         *refs_processed = 1;
5776 out:
5777         return ret;
5778 }
5779 
5780 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
5781 {
5782         int ret = 0;
5783         u64 left_mode;
5784         u64 left_uid;
5785         u64 left_gid;
5786         u64 right_mode;
5787         u64 right_uid;
5788         u64 right_gid;
5789         int need_chmod = 0;
5790         int need_chown = 0;
5791         int pending_move = 0;
5792         int refs_processed = 0;
5793 
5794         ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
5795                                               &refs_processed);
5796         if (ret < 0)
5797                 goto out;
5798 
5799         /*
5800          * We have processed the refs and thus need to advance send_progress.
5801          * Now, calls to get_cur_xxx will take the updated refs of the current
5802          * inode into account.
5803          *
5804          * On the other hand, if our current inode is a directory and couldn't
5805          * be moved/renamed because its parent was renamed/moved too and it has
5806          * a higher inode number, we can only move/rename our current inode
5807          * after we moved/renamed its parent. Therefore in this case operate on
5808          * the old path (pre move/rename) of our current inode, and the
5809          * move/rename will be performed later.
5810          */
5811         if (refs_processed && !pending_move)
5812                 sctx->send_progress = sctx->cur_ino + 1;
5813 
5814         if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
5815                 goto out;
5816         if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
5817                 goto out;
5818 
5819         ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
5820                         &left_mode, &left_uid, &left_gid, NULL);
5821         if (ret < 0)
5822                 goto out;
5823 
5824         if (!sctx->parent_root || sctx->cur_inode_new) {
5825                 need_chown = 1;
5826                 if (!S_ISLNK(sctx->cur_inode_mode))
5827                         need_chmod = 1;
5828         } else {
5829                 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
5830                                 NULL, NULL, &right_mode, &right_uid,
5831                                 &right_gid, NULL);
5832                 if (ret < 0)
5833                         goto out;
5834 
5835                 if (left_uid != right_uid || left_gid != right_gid)
5836                         need_chown = 1;
5837                 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
5838                         need_chmod = 1;
5839         }
5840 
5841         if (S_ISREG(sctx->cur_inode_mode)) {
5842                 if (need_send_hole(sctx)) {
5843                         if (sctx->cur_inode_last_extent == (u64)-1 ||
5844                             sctx->cur_inode_last_extent <
5845                             sctx->cur_inode_size) {
5846                                 ret = get_last_extent(sctx, (u64)-1);
5847                                 if (ret)
5848                                         goto out;
5849                         }
5850                         if (sctx->cur_inode_last_extent <
5851                             sctx->cur_inode_size) {
5852                                 ret = send_hole(sctx, sctx->cur_inode_size);
5853                                 if (ret)
5854                                         goto out;
5855                         }
5856                 }
5857                 ret = send_truncate(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5858                                 sctx->cur_inode_size);
5859                 if (ret < 0)
5860                         goto out;
5861         }
5862 
5863         if (need_chown) {
5864                 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5865                                 left_uid, left_gid);
5866                 if (ret < 0)
5867                         goto out;
5868         }
5869         if (need_chmod) {
5870                 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5871                                 left_mode);
5872                 if (ret < 0)
5873                         goto out;
5874         }
5875 
5876         /*
5877          * If other directory inodes depended on our current directory
5878          * inode's move/rename, now do their move/rename operations.
5879          */
5880         if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
5881                 ret = apply_children_dir_moves(sctx);
5882                 if (ret)
5883                         goto out;
5884                 /*
5885                  * Need to send that every time, no matter if it actually
5886                  * changed between the two trees as we have done changes to
5887                  * the inode before. If our inode is a directory and it's
5888                  * waiting to be moved/renamed, we will send its utimes when
5889                  * it's moved/renamed, therefore we don't need to do it here.
5890                  */
5891                 sctx->send_progress = sctx->cur_ino + 1;
5892                 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
5893                 if (ret < 0)
5894                         goto out;
5895         }
5896 
5897 out:
5898         return ret;
5899 }
5900 
5901 static int changed_inode(struct send_ctx *sctx,
5902                          enum btrfs_compare_tree_result result)
5903 {
5904         int ret = 0;
5905         struct btrfs_key *key = sctx->cmp_key;
5906         struct btrfs_inode_item *left_ii = NULL;
5907         struct btrfs_inode_item *right_ii = NULL;
5908         u64 left_gen = 0;
5909         u64 right_gen = 0;
5910 
5911         sctx->cur_ino = key->objectid;
5912         sctx->cur_inode_new_gen = 0;
5913         sctx->cur_inode_last_extent = (u64)-1;
5914 
5915         /*
5916          * Set send_progress to current inode. This will tell all get_cur_xxx
5917          * functions that the current inode's refs are not updated yet. Later,
5918          * when process_recorded_refs is finished, it is set to cur_ino + 1.
5919          */
5920         sctx->send_progress = sctx->cur_ino;
5921 
5922         if (result == BTRFS_COMPARE_TREE_NEW ||
5923             result == BTRFS_COMPARE_TREE_CHANGED) {
5924                 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
5925                                 sctx->left_path->slots[0],
5926                                 struct btrfs_inode_item);
5927                 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
5928                                 left_ii);
5929         } else {
5930                 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5931                                 sctx->right_path->slots[0],
5932                                 struct btrfs_inode_item);
5933                 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5934                                 right_ii);
5935         }
5936         if (result == BTRFS_COMPARE_TREE_CHANGED) {
5937                 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5938                                 sctx->right_path->slots[0],
5939                                 struct btrfs_inode_item);
5940 
5941                 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5942                                 right_ii);
5943 
5944                 /*
5945                  * The cur_ino = root dir case is special here. We can't treat
5946                  * the inode as deleted+reused because it would generate a
5947                  * stream that tries to delete/mkdir the root dir.
5948                  */
5949                 if (left_gen != right_gen &&
5950                     sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5951                         sctx->cur_inode_new_gen = 1;
5952         }
5953 
5954         if (result == BTRFS_COMPARE_TREE_NEW) {
5955                 sctx->cur_inode_gen = left_gen;
5956                 sctx->cur_inode_new = 1;
5957                 sctx->cur_inode_deleted = 0;
5958                 sctx->cur_inode_size = btrfs_inode_size(
5959                                 sctx->left_path->nodes[0], left_ii);
5960                 sctx->cur_inode_mode = btrfs_inode_mode(
5961                                 sctx->left_path->nodes[0], left_ii);
5962                 sctx->cur_inode_rdev = btrfs_inode_rdev(
5963                                 sctx->left_path->nodes[0], left_ii);
5964                 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5965                         ret = send_create_inode_if_needed(sctx);
5966         } else if (result == BTRFS_COMPARE_TREE_DELETED) {
5967                 sctx->cur_inode_gen = right_gen;
5968                 sctx->cur_inode_new = 0;
5969                 sctx->cur_inode_deleted = 1;
5970                 sctx->cur_inode_size = btrfs_inode_size(
5971                                 sctx->right_path->nodes[0], right_ii);
5972                 sctx->cur_inode_mode = btrfs_inode_mode(
5973                                 sctx->right_path->nodes[0], right_ii);
5974         } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
5975                 /*
5976                  * We need to do some special handling in case the inode was
5977                  * reported as changed with a changed generation number. This
5978                  * means that the original inode was deleted and new inode
5979                  * reused the same inum. So we have to treat the old inode as
5980                  * deleted and the new one as new.
5981                  */
5982                 if (sctx->cur_inode_new_gen) {
5983                         /*
5984                          * First, process the inode as if it was deleted.
5985                          */
5986                         sctx->cur_inode_gen = right_gen;
5987                         sctx->cur_inode_new = 0;
5988                         sctx->cur_inode_deleted = 1;
5989                         sctx->cur_inode_size = btrfs_inode_size(
5990                                         sctx->right_path->nodes[0], right_ii);
5991                         sctx->cur_inode_mode = btrfs_inode_mode(
5992                                         sctx->right_path->nodes[0], right_ii);
5993                         ret = process_all_refs(sctx,
5994                                         BTRFS_COMPARE_TREE_DELETED);
5995                         if (ret < 0)
5996                                 goto out;
5997 
5998                         /*
5999                          * Now process the inode as if it was new.
6000                          */
6001                         sctx->cur_inode_gen = left_gen;
6002                         sctx->cur_inode_new = 1;
6003                         sctx->cur_inode_deleted = 0;
6004                         sctx->cur_inode_size = btrfs_inode_size(
6005                                         sctx->left_path->nodes[0], left_ii);
6006                         sctx->cur_inode_mode = btrfs_inode_mode(
6007                                         sctx->left_path->nodes[0], left_ii);
6008                         sctx->cur_inode_rdev = btrfs_inode_rdev(
6009                                         sctx->left_path->nodes[0], left_ii);
6010                         ret = send_create_inode_if_needed(sctx);
6011                         if (ret < 0)
6012                                 goto out;
6013 
6014                         ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
6015                         if (ret < 0)
6016                                 goto out;
6017                         /*
6018                          * Advance send_progress now as we did not get into
6019                          * process_recorded_refs_if_needed in the new_gen case.
6020                          */
6021                         sctx->send_progress = sctx->cur_ino + 1;
6022 
6023                         /*
6024                          * Now process all extents and xattrs of the inode as if
6025                          * they were all new.
6026                          */
6027                         ret = process_all_extents(sctx);
6028                         if (ret < 0)
6029                                 goto out;
6030                         ret = process_all_new_xattrs(sctx);
6031                         if (ret < 0)
6032                                 goto out;
6033                 } else {
6034                         sctx->cur_inode_gen = left_gen;
6035                         sctx->cur_inode_new = 0;
6036                         sctx->cur_inode_new_gen = 0;
6037                         sctx->cur_inode_deleted = 0;
6038                         sctx->cur_inode_size = btrfs_inode_size(
6039                                         sctx->left_path->nodes[0], left_ii);
6040                         sctx->cur_inode_mode = btrfs_inode_mode(
6041                                         sctx->left_path->nodes[0], left_ii);
6042                 }
6043         }
6044 
6045 out:
6046         return ret;
6047 }
6048 
6049 /*
6050  * We have to process new refs before deleted refs, but compare_trees gives us
6051  * the new and deleted refs mixed. To fix this, we record the new/deleted refs
6052  * first and later process them in process_recorded_refs.
6053  * For the cur_inode_new_gen case, we skip recording completely because
6054  * changed_inode did already initiate processing of refs. The reason for this is
6055  * that in this case, compare_tree actually compares the refs of 2 different
6056  * inodes. To fix this, process_all_refs is used in changed_inode to handle all
6057  * refs of the right tree as deleted and all refs of the left tree as new.
6058  */
6059 static int changed_ref(struct send_ctx *sctx,
6060                        enum btrfs_compare_tree_result result)
6061 {
6062         int ret = 0;
6063 
6064         if (sctx->cur_ino != sctx->cmp_key->objectid) {
6065                 inconsistent_snapshot_error(sctx, result, "reference");
6066                 return -EIO;
6067         }
6068 
6069         if (!sctx->cur_inode_new_gen &&
6070             sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
6071                 if (result == BTRFS_COMPARE_TREE_NEW)
6072                         ret = record_new_ref(sctx);
6073                 else if (result == BTRFS_COMPARE_TREE_DELETED)
6074                         ret = record_deleted_ref(sctx);
6075                 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6076                         ret = record_changed_ref(sctx);
6077         }
6078 
6079         return ret;
6080 }
6081 
6082 /*
6083  * Process new/deleted/changed xattrs. We skip processing in the
6084  * cur_inode_new_gen case because changed_inode did already initiate processing
6085  * of xattrs. The reason is the same as in changed_ref
6086  */
6087 static int changed_xattr(struct send_ctx *sctx,
6088                          enum btrfs_compare_tree_result result)
6089 {
6090         int ret = 0;
6091 
6092         if (sctx->cur_ino != sctx->cmp_key->objectid) {
6093                 inconsistent_snapshot_error(sctx, result, "xattr");
6094                 return -EIO;
6095         }
6096 
6097         if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6098                 if (result == BTRFS_COMPARE_TREE_NEW)
6099                         ret = process_new_xattr(sctx);
6100                 else if (result == BTRFS_COMPARE_TREE_DELETED)
6101                         ret = process_deleted_xattr(sctx);
6102                 else if (result == BTRFS_COMPARE_TREE_CHANGED)
6103                         ret = process_changed_xattr(sctx);
6104         }
6105 
6106         return ret;
6107 }
6108 
6109 /*
6110  * Process new/deleted/changed extents. We skip processing in the
6111  * cur_inode_new_gen case because changed_inode did already initiate processing
6112  * of extents. The reason is the same as in changed_ref
6113  */
6114 static int changed_extent(struct send_ctx *sctx,
6115                           enum btrfs_compare_tree_result result)
6116 {
6117         int ret = 0;
6118 
6119         if (sctx->cur_ino != sctx->cmp_key->objectid) {
6120 
6121                 if (result == BTRFS_COMPARE_TREE_CHANGED) {
6122                         struct extent_buffer *leaf_l;
6123                         struct extent_buffer *leaf_r;
6124                         struct btrfs_file_extent_item *ei_l;
6125                         struct btrfs_file_extent_item *ei_r;
6126 
6127                         leaf_l = sctx->left_path->nodes[0];
6128                         leaf_r = sctx->right_path->nodes[0];
6129                         ei_l = btrfs_item_ptr(leaf_l,
6130                                               sctx->left_path->slots[0],
6131                                               struct btrfs_file_extent_item);
6132                         ei_r = btrfs_item_ptr(leaf_r,
6133                                               sctx->right_path->slots[0],
6134                                               struct btrfs_file_extent_item);
6135 
6136                         /*
6137                          * We may have found an extent item that has changed
6138                          * only its disk_bytenr field and the corresponding
6139                          * inode item was not updated. This case happens due to
6140                          * very specific timings during relocation when a leaf
6141                          * that contains file extent items is COWed while
6142                          * relocation is ongoing and its in the stage where it
6143                          * updates data pointers. So when this happens we can
6144                          * safely ignore it since we know it's the same extent,
6145                          * but just at different logical and physical locations
6146                          * (when an extent is fully replaced with a new one, we
6147                          * know the generation number must have changed too,
6148                          * since snapshot creation implies committing the current
6149                          * transaction, and the inode item must have been updated
6150                          * as well).
6151                          * This replacement of the disk_bytenr happens at
6152                          * relocation.c:replace_file_extents() through
6153                          * relocation.c:btrfs_reloc_cow_block().
6154                          */
6155                         if (btrfs_file_extent_generation(leaf_l, ei_l) ==
6156                             btrfs_file_extent_generation(leaf_r, ei_r) &&
6157                             btrfs_file_extent_ram_bytes(leaf_l, ei_l) ==
6158                             btrfs_file_extent_ram_bytes(leaf_r, ei_r) &&
6159                             btrfs_file_extent_compression(leaf_l, ei_l) ==
6160                             btrfs_file_extent_compression(leaf_r, ei_r) &&
6161                             btrfs_file_extent_encryption(leaf_l, ei_l) ==
6162                             btrfs_file_extent_encryption(leaf_r, ei_r) &&
6163                             btrfs_file_extent_other_encoding(leaf_l, ei_l) ==
6164                             btrfs_file_extent_other_encoding(leaf_r, ei_r) &&
6165                             btrfs_file_extent_type(leaf_l, ei_l) ==
6166                             btrfs_file_extent_type(leaf_r, ei_r) &&
6167                             btrfs_file_extent_disk_bytenr(leaf_l, ei_l) !=
6168                             btrfs_file_extent_disk_bytenr(leaf_r, ei_r) &&
6169                             btrfs_file_extent_disk_num_bytes(leaf_l, ei_l) ==
6170                             btrfs_file_extent_disk_num_bytes(leaf_r, ei_r) &&
6171                             btrfs_file_extent_offset(leaf_l, ei_l) ==
6172                             btrfs_file_extent_offset(leaf_r, ei_r) &&
6173                             btrfs_file_extent_num_bytes(leaf_l, ei_l) ==
6174                             btrfs_file_extent_num_bytes(leaf_r, ei_r))
6175                                 return 0;
6176                 }
6177 
6178                 inconsistent_snapshot_error(sctx, result, "extent");
6179                 return -EIO;
6180         }
6181 
6182         if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
6183                 if (result != BTRFS_COMPARE_TREE_DELETED)
6184                         ret = process_extent(sctx, sctx->left_path,
6185                                         sctx->cmp_key);
6186         }
6187 
6188         return ret;
6189 }
6190 
6191 static int dir_changed(struct send_ctx *sctx, u64 dir)
6192 {
6193         u64 orig_gen, new_gen;
6194         int ret;
6195 
6196         ret = get_inode_info(sctx->send_root, dir, NULL, &new_gen, NULL, NULL,
6197                              NULL, NULL);
6198         if (ret)
6199                 return ret;
6200 
6201         ret = get_inode_info(sctx->parent_root, dir, NULL, &orig_gen, NULL,
6202                              NULL, NULL, NULL);
6203         if (ret)
6204                 return ret;
6205 
6206         return (orig_gen != new_gen) ? 1 : 0;
6207 }
6208 
6209 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path,
6210                         struct btrfs_key *key)
6211 {
6212         struct btrfs_inode_extref *extref;
6213         struct extent_buffer *leaf;
6214         u64 dirid = 0, last_dirid = 0;
6215         unsigned long ptr;
6216         u32 item_size;
6217         u32 cur_offset = 0;
6218         int ref_name_len;
6219         int ret = 0;
6220 
6221         /* Easy case, just check this one dirid */
6222         if (key->type == BTRFS_INODE_REF_KEY) {
6223                 dirid = key->offset;
6224 
6225                 ret = dir_changed(sctx, dirid);
6226                 goto out;
6227         }
6228 
6229         leaf = path->nodes[0];
6230         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
6231         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
6232         while (cur_offset < item_size) {
6233                 extref = (struct btrfs_inode_extref *)(ptr +
6234                                                        cur_offset);
6235                 dirid = btrfs_inode_extref_parent(leaf, extref);
6236                 ref_name_len = btrfs_inode_extref_name_len(leaf, extref);
6237                 cur_offset += ref_name_len + sizeof(*extref);
6238                 if (dirid == last_dirid)
6239                         continue;
6240                 ret = dir_changed(sctx, dirid);
6241                 if (ret)
6242                         break;
6243                 last_dirid = dirid;
6244         }
6245 out:
6246         return ret;
6247 }
6248 
6249 /*
6250  * Updates compare related fields in sctx and simply forwards to the actual
6251  * changed_xxx functions.
6252  */
6253 static int changed_cb(struct btrfs_path *left_path,
6254                       struct btrfs_path *right_path,
6255                       struct btrfs_key *key,
6256                       enum btrfs_compare_tree_result result,
6257                       void *ctx)
6258 {
6259         int ret = 0;
6260         struct send_ctx *sctx = ctx;
6261 
6262         if (result == BTRFS_COMPARE_TREE_SAME) {
6263                 if (key->type == BTRFS_INODE_REF_KEY ||
6264                     key->type == BTRFS_INODE_EXTREF_KEY) {
6265                         ret = compare_refs(sctx, left_path, key);
6266                         if (!ret)
6267                                 return 0;
6268                         if (ret < 0)
6269                                 return ret;
6270                 } else if (key->type == BTRFS_EXTENT_DATA_KEY) {
6271                         return maybe_send_hole(sctx, left_path, key);
6272                 } else {
6273                         return 0;
6274                 }
6275                 result = BTRFS_COMPARE_TREE_CHANGED;
6276                 ret = 0;
6277         }
6278 
6279         sctx->left_path = left_path;
6280         sctx->right_path = right_path;
6281         sctx->cmp_key = key;
6282 
6283         ret = finish_inode_if_needed(sctx, 0);
6284         if (ret < 0)
6285                 goto out;
6286 
6287         /* Ignore non-FS objects */
6288         if (key->objectid == BTRFS_FREE_INO_OBJECTID ||
6289             key->objectid == BTRFS_FREE_SPACE_OBJECTID)
6290                 goto out;
6291 
6292         if (key->type == BTRFS_INODE_ITEM_KEY)
6293                 ret = changed_inode(sctx, result);
6294         else if (key->type == BTRFS_INODE_REF_KEY ||
6295                  key->type == BTRFS_INODE_EXTREF_KEY)
6296                 ret = changed_ref(sctx, result);
6297         else if (key->type == BTRFS_XATTR_ITEM_KEY)
6298                 ret = changed_xattr(sctx, result);
6299         else if (key->type == BTRFS_EXTENT_DATA_KEY)
6300                 ret = changed_extent(sctx, result);
6301 
6302 out:
6303         return ret;
6304 }
6305 
6306 static int full_send_tree(struct send_ctx *sctx)
6307 {
6308         int ret;
6309         struct btrfs_root *send_root = sctx->send_root;
6310         struct btrfs_key key;
6311         struct btrfs_key found_key;
6312         struct btrfs_path *path;
6313         struct extent_buffer *eb;
6314         int slot;
6315 
6316         path = alloc_path_for_send();
6317         if (!path)
6318                 return -ENOMEM;
6319 
6320         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
6321         key.type = BTRFS_INODE_ITEM_KEY;
6322         key.offset = 0;
6323 
6324         ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0);
6325         if (ret < 0)
6326                 goto out;
6327         if (ret)
6328                 goto out_finish;
6329 
6330         while (1) {
6331                 eb = path->nodes[0];
6332                 slot = path->slots[0];
6333                 btrfs_item_key_to_cpu(eb, &found_key, slot);
6334 
6335                 ret = changed_cb(path, NULL, &found_key,
6336                                  BTRFS_COMPARE_TREE_NEW, sctx);
6337                 if (ret < 0)
6338                         goto out;
6339 
6340                 key.objectid = found_key.objectid;
6341                 key.type = found_key.type;
6342                 key.offset = found_key.offset + 1;
6343 
6344                 ret = btrfs_next_item(send_root, path);
6345                 if (ret < 0)
6346                         goto out;
6347                 if (ret) {
6348                         ret  = 0;
6349                         break;
6350                 }
6351         }
6352 
6353 out_finish:
6354         ret = finish_inode_if_needed(sctx, 1);
6355 
6356 out:
6357         btrfs_free_path(path);
6358         return ret;
6359 }
6360 
6361 static int send_subvol(struct send_ctx *sctx)
6362 {
6363         int ret;
6364 
6365         if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) {
6366                 ret = send_header(sctx);
6367                 if (ret < 0)
6368                         goto out;
6369         }
6370 
6371         ret = send_subvol_begin(sctx);
6372         if (ret < 0)
6373                 goto out;
6374 
6375         if (sctx->parent_root) {
6376                 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root,
6377                                 changed_cb, sctx);
6378                 if (ret < 0)
6379                         goto out;
6380                 ret = finish_inode_if_needed(sctx, 1);
6381                 if (ret < 0)
6382                         goto out;
6383         } else {
6384                 ret = full_send_tree(sctx);
6385                 if (ret < 0)
6386                         goto out;
6387         }
6388 
6389 out:
6390         free_recorded_refs(sctx);
6391         return ret;
6392 }
6393 
6394 /*
6395  * If orphan cleanup did remove any orphans from a root, it means the tree
6396  * was modified and therefore the commit root is not the same as the current
6397  * root anymore. This is a problem, because send uses the commit root and
6398  * therefore can see inode items that don't exist in the current root anymore,
6399  * and for example make calls to btrfs_iget, which will do tree lookups based
6400  * on the current root and not on the commit root. Those lookups will fail,
6401  * returning a -ESTALE error, and making send fail with that error. So make
6402  * sure a send does not see any orphans we have just removed, and that it will
6403  * see the same inodes regardless of whether a transaction commit happened
6404  * before it started (meaning that the commit root will be the same as the
6405  * current root) or not.
6406  */
6407 static int ensure_commit_roots_uptodate(struct send_ctx *sctx)
6408 {
6409         int i;
6410         struct btrfs_trans_handle *trans = NULL;
6411 
6412 again:
6413         if (sctx->parent_root &&
6414             sctx->parent_root->node != sctx->parent_root->commit_root)
6415                 goto commit_trans;
6416 
6417         for (i = 0; i < sctx->clone_roots_cnt; i++)
6418                 if (sctx->clone_roots[i].root->node !=
6419                     sctx->clone_roots[i].root->commit_root)
6420                         goto commit_trans;
6421 
6422         if (trans)
6423                 return btrfs_end_transaction(trans);
6424 
6425         return 0;
6426 
6427 commit_trans:
6428         /* Use any root, all fs roots will get their commit roots updated. */
6429         if (!trans) {
6430                 trans = btrfs_join_transaction(sctx->send_root);
6431                 if (IS_ERR(trans))
6432                         return PTR_ERR(trans);
6433                 goto again;
6434         }
6435 
6436         return btrfs_commit_transaction(trans);
6437 }
6438 
6439 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root)
6440 {
6441         spin_lock(&root->root_item_lock);
6442         root->send_in_progress--;
6443         /*
6444          * Not much left to do, we don't know why it's unbalanced and
6445          * can't blindly reset it to 0.
6446          */
6447         if (root->send_in_progress < 0)
6448                 btrfs_err(root->fs_info,
6449                           "send_in_progres unbalanced %d root %llu",
6450                           root->send_in_progress, root->root_key.objectid);
6451         spin_unlock(&root->root_item_lock);
6452 }
6453 
6454 long btrfs_ioctl_send(struct file *mnt_file, struct btrfs_ioctl_send_args *arg)
6455 {
6456         int ret = 0;
6457         struct btrfs_root *send_root = BTRFS_I(file_inode(mnt_file))->root;
6458         struct btrfs_fs_info *fs_info = send_root->fs_info;
6459         struct btrfs_root *clone_root;
6460         struct btrfs_key key;
6461         struct send_ctx *sctx = NULL;
6462         u32 i;
6463         u64 *clone_sources_tmp = NULL;
6464         int clone_sources_to_rollback = 0;
6465         unsigned alloc_size;
6466         int sort_clone_roots = 0;
6467         int index;
6468 
6469         if (!capable(CAP_SYS_ADMIN))
6470                 return -EPERM;
6471 
6472         /*
6473          * The subvolume must remain read-only during send, protect against
6474          * making it RW. This also protects against deletion.
6475          */
6476         spin_lock(&send_root->root_item_lock);
6477         send_root->send_in_progress++;
6478         spin_unlock(&send_root->root_item_lock);
6479 
6480         /*
6481          * This is done when we lookup the root, it should already be complete
6482          * by the time we get here.
6483          */
6484         WARN_ON(send_root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE);
6485 
6486         /*
6487          * Userspace tools do the checks and warn the user if it's
6488          * not RO.
6489          */
6490         if (!btrfs_root_readonly(send_root)) {
6491                 ret = -EPERM;
6492                 goto out;
6493         }
6494 
6495         /*
6496          * Check that we don't overflow at later allocations, we request
6497          * clone_sources_count + 1 items, and compare to unsigned long inside
6498          * access_ok.
6499          */
6500         if (arg->clone_sources_count >
6501             ULONG_MAX / sizeof(struct clone_root) - 1) {
6502                 ret = -EINVAL;
6503                 goto out;
6504         }
6505 
6506         if (!access_ok(VERIFY_READ, arg->clone_sources,
6507                         sizeof(*arg->clone_sources) *
6508                         arg->clone_sources_count)) {
6509                 ret = -EFAULT;
6510                 goto out;
6511         }
6512 
6513         if (arg->flags & ~BTRFS_SEND_FLAG_MASK) {
6514                 ret = -EINVAL;
6515                 goto out;
6516         }
6517 
6518         sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL);
6519         if (!sctx) {
6520                 ret = -ENOMEM;
6521                 goto out;
6522         }
6523 
6524         INIT_LIST_HEAD(&sctx->new_refs);
6525         INIT_LIST_HEAD(&sctx->deleted_refs);
6526         INIT_RADIX_TREE(&sctx->name_cache, GFP_KERNEL);
6527         INIT_LIST_HEAD(&sctx->name_cache_list);
6528 
6529         sctx->flags = arg->flags;
6530 
6531         sctx->send_filp = fget(arg->send_fd);
6532         if (!sctx->send_filp) {
6533                 ret = -EBADF;
6534                 goto out;
6535         }
6536 
6537         sctx->send_root = send_root;
6538         /*
6539          * Unlikely but possible, if the subvolume is marked for deletion but
6540          * is slow to remove the directory entry, send can still be started
6541          */
6542         if (btrfs_root_dead(sctx->send_root)) {
6543                 ret = -EPERM;
6544                 goto out;
6545         }
6546 
6547         sctx->clone_roots_cnt = arg->clone_sources_count;
6548 
6549         sctx->send_max_size = BTRFS_SEND_BUF_SIZE;
6550         sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL);
6551         if (!sctx->send_buf) {
6552                 ret = -ENOMEM;
6553                 goto out;
6554         }
6555 
6556         sctx->read_buf = kvmalloc(BTRFS_SEND_READ_SIZE, GFP_KERNEL);
6557         if (!sctx->read_buf) {
6558                 ret = -ENOMEM;
6559                 goto out;
6560         }
6561 
6562         sctx->pending_dir_moves = RB_ROOT;
6563         sctx->waiting_dir_moves = RB_ROOT;
6564         sctx->orphan_dirs = RB_ROOT;
6565 
6566         alloc_size = sizeof(struct clone_root) * (arg->clone_sources_count + 1);
6567 
6568         sctx->clone_roots = kzalloc(alloc_size, GFP_KERNEL);
6569         if (!sctx->clone_roots) {
6570                 ret = -ENOMEM;
6571                 goto out;
6572         }
6573 
6574         alloc_size = arg->clone_sources_count * sizeof(*arg->clone_sources);
6575 
6576         if (arg->clone_sources_count) {
6577                 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL);
6578                 if (!clone_sources_tmp) {
6579                         ret = -ENOMEM;
6580                         goto out;
6581                 }
6582 
6583                 ret = copy_from_user(clone_sources_tmp, arg->clone_sources,
6584                                 alloc_size);
6585                 if (ret) {
6586                         ret = -EFAULT;