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

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

Version: ~ [ linux-5.19-rc3 ] ~ [ linux-5.18.5 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.48 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.123 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.199 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.248 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.284 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.319 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ 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 
 30 #include "send.h"
 31 #include "backref.h"
 32 #include "hash.h"
 33 #include "locking.h"
 34 #include "disk-io.h"
 35 #include "btrfs_inode.h"
 36 #include "transaction.h"
 37 #include "compression.h"
 38 
 39 /*
 40  * A fs_path is a helper to dynamically build path names with unknown size.
 41  * It reallocates the internal buffer on demand.
 42  * It allows fast adding of path elements on the right side (normal path) and
 43  * fast adding to the left side (reversed path). A reversed path can also be
 44  * unreversed if needed.
 45  */
 46 struct fs_path {
 47         union {
 48                 struct {
 49                         char *start;
 50                         char *end;
 51 
 52                         char *buf;
 53                         unsigned short buf_len:15;
 54                         unsigned short reversed:1;
 55                         char inline_buf[];
 56                 };
 57                 /*
 58                  * Average path length does not exceed 200 bytes, we'll have
 59                  * better packing in the slab and higher chance to satisfy
 60                  * a allocation later during send.
 61                  */
 62                 char pad[256];
 63         };
 64 };
 65 #define FS_PATH_INLINE_SIZE \
 66         (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
 67 
 68 
 69 /* reused for each extent */
 70 struct clone_root {
 71         struct btrfs_root *root;
 72         u64 ino;
 73         u64 offset;
 74 
 75         u64 found_refs;
 76 };
 77 
 78 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128
 79 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
 80 
 81 struct send_ctx {
 82         struct file *send_filp;
 83         loff_t send_off;
 84         char *send_buf;
 85         u32 send_size;
 86         u32 send_max_size;
 87         u64 total_send_size;
 88         u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
 89         u64 flags;      /* 'flags' member of btrfs_ioctl_send_args is u64 */
 90 
 91         struct btrfs_root *send_root;
 92         struct btrfs_root *parent_root;
 93         struct clone_root *clone_roots;
 94         int clone_roots_cnt;
 95 
 96         /* current state of the compare_tree call */
 97         struct btrfs_path *left_path;
 98         struct btrfs_path *right_path;
 99         struct btrfs_key *cmp_key;
100 
101         /*
102          * infos of the currently processed inode. In case of deleted inodes,
103          * these are the values from the deleted inode.
104          */
105         u64 cur_ino;
106         u64 cur_inode_gen;
107         int cur_inode_new;
108         int cur_inode_new_gen;
109         int cur_inode_deleted;
110         u64 cur_inode_size;
111         u64 cur_inode_mode;
112         u64 cur_inode_rdev;
113         u64 cur_inode_last_extent;
114 
115         u64 send_progress;
116 
117         struct list_head new_refs;
118         struct list_head deleted_refs;
119 
120         struct radix_tree_root name_cache;
121         struct list_head name_cache_list;
122         int name_cache_size;
123 
124         struct file_ra_state ra;
125 
126         char *read_buf;
127 
128         /*
129          * We process inodes by their increasing order, so if before an
130          * incremental send we reverse the parent/child relationship of
131          * directories such that a directory with a lower inode number was
132          * the parent of a directory with a higher inode number, and the one
133          * becoming the new parent got renamed too, we can't rename/move the
134          * directory with lower inode number when we finish processing it - we
135          * must process the directory with higher inode number first, then
136          * rename/move it and then rename/move the directory with lower inode
137          * number. Example follows.
138          *
139          * Tree state when the first send was performed:
140          *
141          * .
142          * |-- a                   (ino 257)
143          *     |-- b               (ino 258)
144          *         |
145          *         |
146          *         |-- c           (ino 259)
147          *         |   |-- d       (ino 260)
148          *         |
149          *         |-- c2          (ino 261)
150          *
151          * Tree state when the second (incremental) send is performed:
152          *
153          * .
154          * |-- a                   (ino 257)
155          *     |-- b               (ino 258)
156          *         |-- c2          (ino 261)
157          *             |-- d2      (ino 260)
158          *                 |-- cc  (ino 259)
159          *
160          * The sequence of steps that lead to the second state was:
161          *
162          * mv /a/b/c/d /a/b/c2/d2
163          * mv /a/b/c /a/b/c2/d2/cc
164          *
165          * "c" has lower inode number, but we can't move it (2nd mv operation)
166          * before we move "d", which has higher inode number.
167          *
168          * So we just memorize which move/rename operations must be performed
169          * later when their respective parent is processed and moved/renamed.
170          */
171 
172         /* Indexed by parent directory inode number. */
173         struct rb_root pending_dir_moves;
174 
175         /*
176          * Reverse index, indexed by the inode number of a directory that
177          * is waiting for the move/rename of its immediate parent before its
178          * own move/rename can be performed.
179          */
180         struct rb_root waiting_dir_moves;
181 
182         /*
183          * A directory that is going to be rm'ed might have a child directory
184          * which is in the pending directory moves index above. In this case,
185          * the directory can only be removed after the move/rename of its child
186          * is performed. Example:
187          *
188          * Parent snapshot:
189          *
190          * .                        (ino 256)
191          * |-- a/                   (ino 257)
192          *     |-- b/               (ino 258)
193          *         |-- c/           (ino 259)
194          *         |   |-- x/       (ino 260)
195          *         |
196          *         |-- y/           (ino 261)
197          *
198          * Send snapshot:
199          *
200          * .                        (ino 256)
201          * |-- a/                   (ino 257)
202          *     |-- b/               (ino 258)
203          *         |-- YY/          (ino 261)
204          *              |-- x/      (ino 260)
205          *
206          * Sequence of steps that lead to the send snapshot:
207          * rm -f /a/b/c/foo.txt
208          * mv /a/b/y /a/b/YY
209          * mv /a/b/c/x /a/b/YY
210          * rmdir /a/b/c
211          *
212          * When the child is processed, its move/rename is delayed until its
213          * parent is processed (as explained above), but all other operations
214          * like update utimes, chown, chgrp, etc, are performed and the paths
215          * that it uses for those operations must use the orphanized name of
216          * its parent (the directory we're going to rm later), so we need to
217          * memorize that name.
218          *
219          * Indexed by the inode number of the directory to be deleted.
220          */
221         struct rb_root orphan_dirs;
222 };
223 
224 struct pending_dir_move {
225         struct rb_node node;
226         struct list_head list;
227         u64 parent_ino;
228         u64 ino;
229         u64 gen;
230         struct list_head update_refs;
231 };
232 
233 struct waiting_dir_move {
234         struct rb_node node;
235         u64 ino;
236         /*
237          * There might be some directory that could not be removed because it
238          * was waiting for this directory inode to be moved first. Therefore
239          * after this directory is moved, we can try to rmdir the ino rmdir_ino.
240          */
241         u64 rmdir_ino;
242         bool orphanized;
243 };
244 
245 struct orphan_dir_info {
246         struct rb_node node;
247         u64 ino;
248         u64 gen;
249 };
250 
251 struct name_cache_entry {
252         struct list_head list;
253         /*
254          * radix_tree has only 32bit entries but we need to handle 64bit inums.
255          * We use the lower 32bit of the 64bit inum to store it in the tree. If
256          * more then one inum would fall into the same entry, we use radix_list
257          * to store the additional entries. radix_list is also used to store
258          * entries where two entries have the same inum but different
259          * generations.
260          */
261         struct list_head radix_list;
262         u64 ino;
263         u64 gen;
264         u64 parent_ino;
265         u64 parent_gen;
266         int ret;
267         int need_later_update;
268         int name_len;
269         char name[];
270 };
271 
272 static void inconsistent_snapshot_error(struct send_ctx *sctx,
273                                         enum btrfs_compare_tree_result result,
274                                         const char *what)
275 {
276         const char *result_string;
277 
278         switch (result) {
279         case BTRFS_COMPARE_TREE_NEW:
280                 result_string = "new";
281                 break;
282         case BTRFS_COMPARE_TREE_DELETED:
283                 result_string = "deleted";
284                 break;
285         case BTRFS_COMPARE_TREE_CHANGED:
286                 result_string = "updated";
287                 break;
288         case BTRFS_COMPARE_TREE_SAME:
289                 ASSERT(0);
290                 result_string = "unchanged";
291                 break;
292         default:
293                 ASSERT(0);
294                 result_string = "unexpected";
295         }
296 
297         btrfs_err(sctx->send_root->fs_info,
298                   "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
299                   result_string, what, sctx->cmp_key->objectid,
300                   sctx->send_root->root_key.objectid,
301                   (sctx->parent_root ?
302                    sctx->parent_root->root_key.objectid : 0));
303 }
304 
305 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
306 
307 static struct waiting_dir_move *
308 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
309 
310 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino);
311 
312 static int need_send_hole(struct send_ctx *sctx)
313 {
314         return (sctx->parent_root && !sctx->cur_inode_new &&
315                 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
316                 S_ISREG(sctx->cur_inode_mode));
317 }
318 
319 static void fs_path_reset(struct fs_path *p)
320 {
321         if (p->reversed) {
322                 p->start = p->buf + p->buf_len - 1;
323                 p->end = p->start;
324                 *p->start = 0;
325         } else {
326                 p->start = p->buf;
327                 p->end = p->start;
328                 *p->start = 0;
329         }
330 }
331 
332 static struct fs_path *fs_path_alloc(void)
333 {
334         struct fs_path *p;
335 
336         p = kmalloc(sizeof(*p), GFP_KERNEL);
337         if (!p)
338                 return NULL;
339         p->reversed = 0;
340         p->buf = p->inline_buf;
341         p->buf_len = FS_PATH_INLINE_SIZE;
342         fs_path_reset(p);
343         return p;
344 }
345 
346 static struct fs_path *fs_path_alloc_reversed(void)
347 {
348         struct fs_path *p;
349 
350         p = fs_path_alloc();
351         if (!p)
352                 return NULL;
353         p->reversed = 1;
354         fs_path_reset(p);
355         return p;
356 }
357 
358 static void fs_path_free(struct fs_path *p)
359 {
360         if (!p)
361                 return;
362         if (p->buf != p->inline_buf)
363                 kfree(p->buf);
364         kfree(p);
365 }
366 
367 static int fs_path_len(struct fs_path *p)
368 {
369         return p->end - p->start;
370 }
371 
372 static int fs_path_ensure_buf(struct fs_path *p, int len)
373 {
374         char *tmp_buf;
375         int path_len;
376         int old_buf_len;
377 
378         len++;
379 
380         if (p->buf_len >= len)
381                 return 0;
382 
383         if (len > PATH_MAX) {
384                 WARN_ON(1);
385                 return -ENOMEM;
386         }
387 
388         path_len = p->end - p->start;
389         old_buf_len = p->buf_len;
390 
391         /*
392          * First time the inline_buf does not suffice
393          */
394         if (p->buf == p->inline_buf) {
395                 tmp_buf = kmalloc(len, GFP_KERNEL);
396                 if (tmp_buf)
397                         memcpy(tmp_buf, p->buf, old_buf_len);
398         } else {
399                 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
400         }
401         if (!tmp_buf)
402                 return -ENOMEM;
403         p->buf = tmp_buf;
404         /*
405          * The real size of the buffer is bigger, this will let the fast path
406          * happen most of the time
407          */
408         p->buf_len = ksize(p->buf);
409 
410         if (p->reversed) {
411                 tmp_buf = p->buf + old_buf_len - path_len - 1;
412                 p->end = p->buf + p->buf_len - 1;
413                 p->start = p->end - path_len;
414                 memmove(p->start, tmp_buf, path_len + 1);
415         } else {
416                 p->start = p->buf;
417                 p->end = p->start + path_len;
418         }
419         return 0;
420 }
421 
422 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
423                                    char **prepared)
424 {
425         int ret;
426         int new_len;
427 
428         new_len = p->end - p->start + name_len;
429         if (p->start != p->end)
430                 new_len++;
431         ret = fs_path_ensure_buf(p, new_len);
432         if (ret < 0)
433                 goto out;
434 
435         if (p->reversed) {
436                 if (p->start != p->end)
437                         *--p->start = '/';
438                 p->start -= name_len;
439                 *prepared = p->start;
440         } else {
441                 if (p->start != p->end)
442                         *p->end++ = '/';
443                 *prepared = p->end;
444                 p->end += name_len;
445                 *p->end = 0;
446         }
447 
448 out:
449         return ret;
450 }
451 
452 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
453 {
454         int ret;
455         char *prepared;
456 
457         ret = fs_path_prepare_for_add(p, name_len, &prepared);
458         if (ret < 0)
459                 goto out;
460         memcpy(prepared, name, name_len);
461 
462 out:
463         return ret;
464 }
465 
466 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
467 {
468         int ret;
469         char *prepared;
470 
471         ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
472         if (ret < 0)
473                 goto out;
474         memcpy(prepared, p2->start, p2->end - p2->start);
475 
476 out:
477         return ret;
478 }
479 
480 static int fs_path_add_from_extent_buffer(struct fs_path *p,
481                                           struct extent_buffer *eb,
482                                           unsigned long off, int len)
483 {
484         int ret;
485         char *prepared;
486 
487         ret = fs_path_prepare_for_add(p, len, &prepared);
488         if (ret < 0)
489                 goto out;
490 
491         read_extent_buffer(eb, prepared, off, len);
492 
493 out:
494         return ret;
495 }
496 
497 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
498 {
499         int ret;
500 
501         p->reversed = from->reversed;
502         fs_path_reset(p);
503 
504         ret = fs_path_add_path(p, from);
505 
506         return ret;
507 }
508 
509 
510 static void fs_path_unreverse(struct fs_path *p)
511 {
512         char *tmp;
513         int len;
514 
515         if (!p->reversed)
516                 return;
517 
518         tmp = p->start;
519         len = p->end - p->start;
520         p->start = p->buf;
521         p->end = p->start + len;
522         memmove(p->start, tmp, len + 1);
523         p->reversed = 0;
524 }
525 
526 static struct btrfs_path *alloc_path_for_send(void)
527 {
528         struct btrfs_path *path;
529 
530         path = btrfs_alloc_path();
531         if (!path)
532                 return NULL;
533         path->search_commit_root = 1;
534         path->skip_locking = 1;
535         path->need_commit_sem = 1;
536         return path;
537 }
538 
539 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
540 {
541         int ret;
542         mm_segment_t old_fs;
543         u32 pos = 0;
544 
545         old_fs = get_fs();
546         set_fs(KERNEL_DS);
547 
548         while (pos < len) {
549                 ret = vfs_write(filp, (__force const char __user *)buf + pos,
550                                 len - pos, off);
551                 /* TODO handle that correctly */
552                 /*if (ret == -ERESTARTSYS) {
553                         continue;
554                 }*/
555                 if (ret < 0)
556                         goto out;
557                 if (ret == 0) {
558                         ret = -EIO;
559                         goto out;
560                 }
561                 pos += ret;
562         }
563 
564         ret = 0;
565 
566 out:
567         set_fs(old_fs);
568         return ret;
569 }
570 
571 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
572 {
573         struct btrfs_tlv_header *hdr;
574         int total_len = sizeof(*hdr) + len;
575         int left = sctx->send_max_size - sctx->send_size;
576 
577         if (unlikely(left < total_len))
578                 return -EOVERFLOW;
579 
580         hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
581         hdr->tlv_type = cpu_to_le16(attr);
582         hdr->tlv_len = cpu_to_le16(len);
583         memcpy(hdr + 1, data, len);
584         sctx->send_size += total_len;
585 
586         return 0;
587 }
588 
589 #define TLV_PUT_DEFINE_INT(bits) \
590         static int tlv_put_u##bits(struct send_ctx *sctx,               \
591                         u##bits attr, u##bits value)                    \
592         {                                                               \
593                 __le##bits __tmp = cpu_to_le##bits(value);              \
594                 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));      \
595         }
596 
597 TLV_PUT_DEFINE_INT(64)
598 
599 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
600                           const char *str, int len)
601 {
602         if (len == -1)
603                 len = strlen(str);
604         return tlv_put(sctx, attr, str, len);
605 }
606 
607 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
608                         const u8 *uuid)
609 {
610         return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
611 }
612 
613 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
614                                   struct extent_buffer *eb,
615                                   struct btrfs_timespec *ts)
616 {
617         struct btrfs_timespec bts;
618         read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
619         return tlv_put(sctx, attr, &bts, sizeof(bts));
620 }
621 
622 
623 #define TLV_PUT(sctx, attrtype, attrlen, data) \
624         do { \
625                 ret = tlv_put(sctx, attrtype, attrlen, data); \
626                 if (ret < 0) \
627                         goto tlv_put_failure; \
628         } while (0)
629 
630 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
631         do { \
632                 ret = tlv_put_u##bits(sctx, attrtype, value); \
633                 if (ret < 0) \
634                         goto tlv_put_failure; \
635         } while (0)
636 
637 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
638 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
639 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
640 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
641 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
642         do { \
643                 ret = tlv_put_string(sctx, attrtype, str, len); \
644                 if (ret < 0) \
645                         goto tlv_put_failure; \
646         } while (0)
647 #define TLV_PUT_PATH(sctx, attrtype, p) \
648         do { \
649                 ret = tlv_put_string(sctx, attrtype, p->start, \
650                         p->end - p->start); \
651                 if (ret < 0) \
652                         goto tlv_put_failure; \
653         } while(0)
654 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
655         do { \
656                 ret = tlv_put_uuid(sctx, attrtype, uuid); \
657                 if (ret < 0) \
658                         goto tlv_put_failure; \
659         } while (0)
660 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
661         do { \
662                 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
663                 if (ret < 0) \
664                         goto tlv_put_failure; \
665         } while (0)
666 
667 static int send_header(struct send_ctx *sctx)
668 {
669         struct btrfs_stream_header hdr;
670 
671         strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
672         hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
673 
674         return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
675                                         &sctx->send_off);
676 }
677 
678 /*
679  * For each command/item we want to send to userspace, we call this function.
680  */
681 static int begin_cmd(struct send_ctx *sctx, int cmd)
682 {
683         struct btrfs_cmd_header *hdr;
684 
685         if (WARN_ON(!sctx->send_buf))
686                 return -EINVAL;
687 
688         BUG_ON(sctx->send_size);
689 
690         sctx->send_size += sizeof(*hdr);
691         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
692         hdr->cmd = cpu_to_le16(cmd);
693 
694         return 0;
695 }
696 
697 static int send_cmd(struct send_ctx *sctx)
698 {
699         int ret;
700         struct btrfs_cmd_header *hdr;
701         u32 crc;
702 
703         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
704         hdr->len = cpu_to_le32(sctx->send_size - sizeof(*hdr));
705         hdr->crc = 0;
706 
707         crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
708         hdr->crc = cpu_to_le32(crc);
709 
710         ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
711                                         &sctx->send_off);
712 
713         sctx->total_send_size += sctx->send_size;
714         sctx->cmd_send_size[le16_to_cpu(hdr->cmd)] += sctx->send_size;
715         sctx->send_size = 0;
716 
717         return ret;
718 }
719 
720 /*
721  * Sends a move instruction to user space
722  */
723 static int send_rename(struct send_ctx *sctx,
724                      struct fs_path *from, struct fs_path *to)
725 {
726         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
727         int ret;
728 
729         btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
730 
731         ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
732         if (ret < 0)
733                 goto out;
734 
735         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
736         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
737 
738         ret = send_cmd(sctx);
739 
740 tlv_put_failure:
741 out:
742         return ret;
743 }
744 
745 /*
746  * Sends a link instruction to user space
747  */
748 static int send_link(struct send_ctx *sctx,
749                      struct fs_path *path, struct fs_path *lnk)
750 {
751         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
752         int ret;
753 
754         btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
755 
756         ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
757         if (ret < 0)
758                 goto out;
759 
760         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
761         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
762 
763         ret = send_cmd(sctx);
764 
765 tlv_put_failure:
766 out:
767         return ret;
768 }
769 
770 /*
771  * Sends an unlink instruction to user space
772  */
773 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
774 {
775         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
776         int ret;
777 
778         btrfs_debug(fs_info, "send_unlink %s", path->start);
779 
780         ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
781         if (ret < 0)
782                 goto out;
783 
784         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
785 
786         ret = send_cmd(sctx);
787 
788 tlv_put_failure:
789 out:
790         return ret;
791 }
792 
793 /*
794  * Sends a rmdir instruction to user space
795  */
796 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
797 {
798         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
799         int ret;
800 
801         btrfs_debug(fs_info, "send_rmdir %s", path->start);
802 
803         ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
804         if (ret < 0)
805                 goto out;
806 
807         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
808 
809         ret = send_cmd(sctx);
810 
811 tlv_put_failure:
812 out:
813         return ret;
814 }
815 
816 /*
817  * Helper function to retrieve some fields from an inode item.
818  */
819 static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
820                           u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
821                           u64 *gid, u64 *rdev)
822 {
823         int ret;
824         struct btrfs_inode_item *ii;
825         struct btrfs_key key;
826 
827         key.objectid = ino;
828         key.type = BTRFS_INODE_ITEM_KEY;
829         key.offset = 0;
830         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
831         if (ret) {
832                 if (ret > 0)
833                         ret = -ENOENT;
834                 return ret;
835         }
836 
837         ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
838                         struct btrfs_inode_item);
839         if (size)
840                 *size = btrfs_inode_size(path->nodes[0], ii);
841         if (gen)
842                 *gen = btrfs_inode_generation(path->nodes[0], ii);
843         if (mode)
844                 *mode = btrfs_inode_mode(path->nodes[0], ii);
845         if (uid)
846                 *uid = btrfs_inode_uid(path->nodes[0], ii);
847         if (gid)
848                 *gid = btrfs_inode_gid(path->nodes[0], ii);
849         if (rdev)
850                 *rdev = btrfs_inode_rdev(path->nodes[0], ii);
851 
852         return ret;
853 }
854 
855 static int get_inode_info(struct btrfs_root *root,
856                           u64 ino, u64 *size, u64 *gen,
857                           u64 *mode, u64 *uid, u64 *gid,
858                           u64 *rdev)
859 {
860         struct btrfs_path *path;
861         int ret;
862 
863         path = alloc_path_for_send();
864         if (!path)
865                 return -ENOMEM;
866         ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
867                                rdev);
868         btrfs_free_path(path);
869         return ret;
870 }
871 
872 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
873                                    struct fs_path *p,
874                                    void *ctx);
875 
876 /*
877  * Helper function to iterate the entries in ONE btrfs_inode_ref or
878  * btrfs_inode_extref.
879  * The iterate callback may return a non zero value to stop iteration. This can
880  * be a negative value for error codes or 1 to simply stop it.
881  *
882  * path must point to the INODE_REF or INODE_EXTREF when called.
883  */
884 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
885                              struct btrfs_key *found_key, int resolve,
886                              iterate_inode_ref_t iterate, void *ctx)
887 {
888         struct extent_buffer *eb = path->nodes[0];
889         struct btrfs_item *item;
890         struct btrfs_inode_ref *iref;
891         struct btrfs_inode_extref *extref;
892         struct btrfs_path *tmp_path;
893         struct fs_path *p;
894         u32 cur = 0;
895         u32 total;
896         int slot = path->slots[0];
897         u32 name_len;
898         char *start;
899         int ret = 0;
900         int num = 0;
901         int index;
902         u64 dir;
903         unsigned long name_off;
904         unsigned long elem_size;
905         unsigned long ptr;
906 
907         p = fs_path_alloc_reversed();
908         if (!p)
909                 return -ENOMEM;
910 
911         tmp_path = alloc_path_for_send();
912         if (!tmp_path) {
913                 fs_path_free(p);
914                 return -ENOMEM;
915         }
916 
917 
918         if (found_key->type == BTRFS_INODE_REF_KEY) {
919                 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
920                                                     struct btrfs_inode_ref);
921                 item = btrfs_item_nr(slot);
922                 total = btrfs_item_size(eb, item);
923                 elem_size = sizeof(*iref);
924         } else {
925                 ptr = btrfs_item_ptr_offset(eb, slot);
926                 total = btrfs_item_size_nr(eb, slot);
927                 elem_size = sizeof(*extref);
928         }
929 
930         while (cur < total) {
931                 fs_path_reset(p);
932 
933                 if (found_key->type == BTRFS_INODE_REF_KEY) {
934                         iref = (struct btrfs_inode_ref *)(ptr + cur);
935                         name_len = btrfs_inode_ref_name_len(eb, iref);
936                         name_off = (unsigned long)(iref + 1);
937                         index = btrfs_inode_ref_index(eb, iref);
938                         dir = found_key->offset;
939                 } else {
940                         extref = (struct btrfs_inode_extref *)(ptr + cur);
941                         name_len = btrfs_inode_extref_name_len(eb, extref);
942                         name_off = (unsigned long)&extref->name;
943                         index = btrfs_inode_extref_index(eb, extref);
944                         dir = btrfs_inode_extref_parent(eb, extref);
945                 }
946 
947                 if (resolve) {
948                         start = btrfs_ref_to_path(root, tmp_path, name_len,
949                                                   name_off, eb, dir,
950                                                   p->buf, p->buf_len);
951                         if (IS_ERR(start)) {
952                                 ret = PTR_ERR(start);
953                                 goto out;
954                         }
955                         if (start < p->buf) {
956                                 /* overflow , try again with larger buffer */
957                                 ret = fs_path_ensure_buf(p,
958                                                 p->buf_len + p->buf - start);
959                                 if (ret < 0)
960                                         goto out;
961                                 start = btrfs_ref_to_path(root, tmp_path,
962                                                           name_len, name_off,
963                                                           eb, dir,
964                                                           p->buf, p->buf_len);
965                                 if (IS_ERR(start)) {
966                                         ret = PTR_ERR(start);
967                                         goto out;
968                                 }
969                                 BUG_ON(start < p->buf);
970                         }
971                         p->start = start;
972                 } else {
973                         ret = fs_path_add_from_extent_buffer(p, eb, name_off,
974                                                              name_len);
975                         if (ret < 0)
976                                 goto out;
977                 }
978 
979                 cur += elem_size + name_len;
980                 ret = iterate(num, dir, index, p, ctx);
981                 if (ret)
982                         goto out;
983                 num++;
984         }
985 
986 out:
987         btrfs_free_path(tmp_path);
988         fs_path_free(p);
989         return ret;
990 }
991 
992 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
993                                   const char *name, int name_len,
994                                   const char *data, int data_len,
995                                   u8 type, void *ctx);
996 
997 /*
998  * Helper function to iterate the entries in ONE btrfs_dir_item.
999  * The iterate callback may return a non zero value to stop iteration. This can
1000  * be a negative value for error codes or 1 to simply stop it.
1001  *
1002  * path must point to the dir item when called.
1003  */
1004 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1005                             struct btrfs_key *found_key,
1006                             iterate_dir_item_t iterate, void *ctx)
1007 {
1008         int ret = 0;
1009         struct extent_buffer *eb;
1010         struct btrfs_item *item;
1011         struct btrfs_dir_item *di;
1012         struct btrfs_key di_key;
1013         char *buf = NULL;
1014         int buf_len;
1015         u32 name_len;
1016         u32 data_len;
1017         u32 cur;
1018         u32 len;
1019         u32 total;
1020         int slot;
1021         int num;
1022         u8 type;
1023 
1024         /*
1025          * Start with a small buffer (1 page). If later we end up needing more
1026          * space, which can happen for xattrs on a fs with a leaf size greater
1027          * then the page size, attempt to increase the buffer. Typically xattr
1028          * values are small.
1029          */
1030         buf_len = PATH_MAX;
1031         buf = kmalloc(buf_len, GFP_KERNEL);
1032         if (!buf) {
1033                 ret = -ENOMEM;
1034                 goto out;
1035         }
1036 
1037         eb = path->nodes[0];
1038         slot = path->slots[0];
1039         item = btrfs_item_nr(slot);
1040         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1041         cur = 0;
1042         len = 0;
1043         total = btrfs_item_size(eb, item);
1044 
1045         num = 0;
1046         while (cur < total) {
1047                 name_len = btrfs_dir_name_len(eb, di);
1048                 data_len = btrfs_dir_data_len(eb, di);
1049                 type = btrfs_dir_type(eb, di);
1050                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1051 
1052                 if (type == BTRFS_FT_XATTR) {
1053                         if (name_len > XATTR_NAME_MAX) {
1054                                 ret = -ENAMETOOLONG;
1055                                 goto out;
1056                         }
1057                         if (name_len + data_len >
1058                                         BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1059                                 ret = -E2BIG;
1060                                 goto out;
1061                         }
1062                 } else {
1063                         /*
1064                          * Path too long
1065                          */
1066                         if (name_len + data_len > PATH_MAX) {
1067                                 ret = -ENAMETOOLONG;
1068                                 goto out;
1069                         }
1070                 }
1071 
1072                 if (name_len + data_len > buf_len) {
1073                         buf_len = name_len + data_len;
1074                         if (is_vmalloc_addr(buf)) {
1075                                 vfree(buf);
1076                                 buf = NULL;
1077                         } else {
1078                                 char *tmp = krealloc(buf, buf_len,
1079                                                 GFP_KERNEL | __GFP_NOWARN);
1080 
1081                                 if (!tmp)
1082                                         kfree(buf);
1083                                 buf = tmp;
1084                         }
1085                         if (!buf) {
1086                                 buf = vmalloc(buf_len);
1087                                 if (!buf) {
1088                                         ret = -ENOMEM;
1089                                         goto out;
1090                                 }
1091                         }
1092                 }
1093 
1094                 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1095                                 name_len + data_len);
1096 
1097                 len = sizeof(*di) + name_len + data_len;
1098                 di = (struct btrfs_dir_item *)((char *)di + len);
1099                 cur += len;
1100 
1101                 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1102                                 data_len, type, ctx);
1103                 if (ret < 0)
1104                         goto out;
1105                 if (ret) {
1106                         ret = 0;
1107                         goto out;
1108                 }
1109 
1110                 num++;
1111         }
1112 
1113 out:
1114         kvfree(buf);
1115         return ret;
1116 }
1117 
1118 static int __copy_first_ref(int num, u64 dir, int index,
1119                             struct fs_path *p, void *ctx)
1120 {
1121         int ret;
1122         struct fs_path *pt = ctx;
1123 
1124         ret = fs_path_copy(pt, p);
1125         if (ret < 0)
1126                 return ret;
1127 
1128         /* we want the first only */
1129         return 1;
1130 }
1131 
1132 /*
1133  * Retrieve the first path of an inode. If an inode has more then one
1134  * ref/hardlink, this is ignored.
1135  */
1136 static int get_inode_path(struct btrfs_root *root,
1137                           u64 ino, struct fs_path *path)
1138 {
1139         int ret;
1140         struct btrfs_key key, found_key;
1141         struct btrfs_path *p;
1142 
1143         p = alloc_path_for_send();
1144         if (!p)
1145                 return -ENOMEM;
1146 
1147         fs_path_reset(path);
1148 
1149         key.objectid = ino;
1150         key.type = BTRFS_INODE_REF_KEY;
1151         key.offset = 0;
1152 
1153         ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1154         if (ret < 0)
1155                 goto out;
1156         if (ret) {
1157                 ret = 1;
1158                 goto out;
1159         }
1160         btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1161         if (found_key.objectid != ino ||
1162             (found_key.type != BTRFS_INODE_REF_KEY &&
1163              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1164                 ret = -ENOENT;
1165                 goto out;
1166         }
1167 
1168         ret = iterate_inode_ref(root, p, &found_key, 1,
1169                                 __copy_first_ref, path);
1170         if (ret < 0)
1171                 goto out;
1172         ret = 0;
1173 
1174 out:
1175         btrfs_free_path(p);
1176         return ret;
1177 }
1178 
1179 struct backref_ctx {
1180         struct send_ctx *sctx;
1181 
1182         struct btrfs_path *path;
1183         /* number of total found references */
1184         u64 found;
1185 
1186         /*
1187          * used for clones found in send_root. clones found behind cur_objectid
1188          * and cur_offset are not considered as allowed clones.
1189          */
1190         u64 cur_objectid;
1191         u64 cur_offset;
1192 
1193         /* may be truncated in case it's the last extent in a file */
1194         u64 extent_len;
1195 
1196         /* data offset in the file extent item */
1197         u64 data_offset;
1198 
1199         /* Just to check for bugs in backref resolving */
1200         int found_itself;
1201 };
1202 
1203 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1204 {
1205         u64 root = (u64)(uintptr_t)key;
1206         struct clone_root *cr = (struct clone_root *)elt;
1207 
1208         if (root < cr->root->objectid)
1209                 return -1;
1210         if (root > cr->root->objectid)
1211                 return 1;
1212         return 0;
1213 }
1214 
1215 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1216 {
1217         struct clone_root *cr1 = (struct clone_root *)e1;
1218         struct clone_root *cr2 = (struct clone_root *)e2;
1219 
1220         if (cr1->root->objectid < cr2->root->objectid)
1221                 return -1;
1222         if (cr1->root->objectid > cr2->root->objectid)
1223                 return 1;
1224         return 0;
1225 }
1226 
1227 /*
1228  * Called for every backref that is found for the current extent.
1229  * Results are collected in sctx->clone_roots->ino/offset/found_refs
1230  */
1231 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1232 {
1233         struct backref_ctx *bctx = ctx_;
1234         struct clone_root *found;
1235         int ret;
1236         u64 i_size;
1237 
1238         /* First check if the root is in the list of accepted clone sources */
1239         found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1240                         bctx->sctx->clone_roots_cnt,
1241                         sizeof(struct clone_root),
1242                         __clone_root_cmp_bsearch);
1243         if (!found)
1244                 return 0;
1245 
1246         if (found->root == bctx->sctx->send_root &&
1247             ino == bctx->cur_objectid &&
1248             offset == bctx->cur_offset) {
1249                 bctx->found_itself = 1;
1250         }
1251 
1252         /*
1253          * There are inodes that have extents that lie behind its i_size. Don't
1254          * accept clones from these extents.
1255          */
1256         ret = __get_inode_info(found->root, bctx->path, ino, &i_size, NULL, NULL,
1257                                NULL, NULL, NULL);
1258         btrfs_release_path(bctx->path);
1259         if (ret < 0)
1260                 return ret;
1261 
1262         if (offset + bctx->data_offset + bctx->extent_len > i_size)
1263                 return 0;
1264 
1265         /*
1266          * Make sure we don't consider clones from send_root that are
1267          * behind the current inode/offset.
1268          */
1269         if (found->root == bctx->sctx->send_root) {
1270                 /*
1271                  * TODO for the moment we don't accept clones from the inode
1272                  * that is currently send. We may change this when
1273                  * BTRFS_IOC_CLONE_RANGE supports cloning from and to the same
1274                  * file.
1275                  */
1276                 if (ino >= bctx->cur_objectid)
1277                         return 0;
1278 #if 0
1279                 if (ino > bctx->cur_objectid)
1280                         return 0;
1281                 if (offset + bctx->extent_len > bctx->cur_offset)
1282                         return 0;
1283 #endif
1284         }
1285 
1286         bctx->found++;
1287         found->found_refs++;
1288         if (ino < found->ino) {
1289                 found->ino = ino;
1290                 found->offset = offset;
1291         } else if (found->ino == ino) {
1292                 /*
1293                  * same extent found more then once in the same file.
1294                  */
1295                 if (found->offset > offset + bctx->extent_len)
1296                         found->offset = offset;
1297         }
1298 
1299         return 0;
1300 }
1301 
1302 /*
1303  * Given an inode, offset and extent item, it finds a good clone for a clone
1304  * instruction. Returns -ENOENT when none could be found. The function makes
1305  * sure that the returned clone is usable at the point where sending is at the
1306  * moment. This means, that no clones are accepted which lie behind the current
1307  * inode+offset.
1308  *
1309  * path must point to the extent item when called.
1310  */
1311 static int find_extent_clone(struct send_ctx *sctx,
1312                              struct btrfs_path *path,
1313                              u64 ino, u64 data_offset,
1314                              u64 ino_size,
1315                              struct clone_root **found)
1316 {
1317         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1318         int ret;
1319         int extent_type;
1320         u64 logical;
1321         u64 disk_byte;
1322         u64 num_bytes;
1323         u64 extent_item_pos;
1324         u64 flags = 0;
1325         struct btrfs_file_extent_item *fi;
1326         struct extent_buffer *eb = path->nodes[0];
1327         struct backref_ctx *backref_ctx = NULL;
1328         struct clone_root *cur_clone_root;
1329         struct btrfs_key found_key;
1330         struct btrfs_path *tmp_path;
1331         int compressed;
1332         u32 i;
1333 
1334         tmp_path = alloc_path_for_send();
1335         if (!tmp_path)
1336                 return -ENOMEM;
1337 
1338         /* We only use this path under the commit sem */
1339         tmp_path->need_commit_sem = 0;
1340 
1341         backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1342         if (!backref_ctx) {
1343                 ret = -ENOMEM;
1344                 goto out;
1345         }
1346 
1347         backref_ctx->path = tmp_path;
1348 
1349         if (data_offset >= ino_size) {
1350                 /*
1351                  * There may be extents that lie behind the file's size.
1352                  * I at least had this in combination with snapshotting while
1353                  * writing large files.
1354                  */
1355                 ret = 0;
1356                 goto out;
1357         }
1358 
1359         fi = btrfs_item_ptr(eb, path->slots[0],
1360                         struct btrfs_file_extent_item);
1361         extent_type = btrfs_file_extent_type(eb, fi);
1362         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1363                 ret = -ENOENT;
1364                 goto out;
1365         }
1366         compressed = btrfs_file_extent_compression(eb, fi);
1367 
1368         num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1369         disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1370         if (disk_byte == 0) {
1371                 ret = -ENOENT;
1372                 goto out;
1373         }
1374         logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1375 
1376         down_read(&fs_info->commit_root_sem);
1377         ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1378                                   &found_key, &flags);
1379         up_read(&fs_info->commit_root_sem);
1380         btrfs_release_path(tmp_path);
1381 
1382         if (ret < 0)
1383                 goto out;
1384         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1385                 ret = -EIO;
1386                 goto out;
1387         }
1388 
1389         /*
1390          * Setup the clone roots.
1391          */
1392         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1393                 cur_clone_root = sctx->clone_roots + i;
1394                 cur_clone_root->ino = (u64)-1;
1395                 cur_clone_root->offset = 0;
1396                 cur_clone_root->found_refs = 0;
1397         }
1398 
1399         backref_ctx->sctx = sctx;
1400         backref_ctx->found = 0;
1401         backref_ctx->cur_objectid = ino;
1402         backref_ctx->cur_offset = data_offset;
1403         backref_ctx->found_itself = 0;
1404         backref_ctx->extent_len = num_bytes;
1405         /*
1406          * For non-compressed extents iterate_extent_inodes() gives us extent
1407          * offsets that already take into account the data offset, but not for
1408          * compressed extents, since the offset is logical and not relative to
1409          * the physical extent locations. We must take this into account to
1410          * avoid sending clone offsets that go beyond the source file's size,
1411          * which would result in the clone ioctl failing with -EINVAL on the
1412          * receiving end.
1413          */
1414         if (compressed == BTRFS_COMPRESS_NONE)
1415                 backref_ctx->data_offset = 0;
1416         else
1417                 backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1418 
1419         /*
1420          * The last extent of a file may be too large due to page alignment.
1421          * We need to adjust extent_len in this case so that the checks in
1422          * __iterate_backrefs work.
1423          */
1424         if (data_offset + num_bytes >= ino_size)
1425                 backref_ctx->extent_len = ino_size - data_offset;
1426 
1427         /*
1428          * Now collect all backrefs.
1429          */
1430         if (compressed == BTRFS_COMPRESS_NONE)
1431                 extent_item_pos = logical - found_key.objectid;
1432         else
1433                 extent_item_pos = 0;
1434         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1435                                     extent_item_pos, 1, __iterate_backrefs,
1436                                     backref_ctx);
1437 
1438         if (ret < 0)
1439                 goto out;
1440 
1441         if (!backref_ctx->found_itself) {
1442                 /* found a bug in backref code? */
1443                 ret = -EIO;
1444                 btrfs_err(fs_info,
1445                           "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1446                           ino, data_offset, disk_byte, found_key.objectid);
1447                 goto out;
1448         }
1449 
1450         btrfs_debug(fs_info,
1451                     "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1452                     data_offset, ino, num_bytes, logical);
1453 
1454         if (!backref_ctx->found)
1455                 btrfs_debug(fs_info, "no clones found");
1456 
1457         cur_clone_root = NULL;
1458         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1459                 if (sctx->clone_roots[i].found_refs) {
1460                         if (!cur_clone_root)
1461                                 cur_clone_root = sctx->clone_roots + i;
1462                         else if (sctx->clone_roots[i].root == sctx->send_root)
1463                                 /* prefer clones from send_root over others */
1464                                 cur_clone_root = sctx->clone_roots + i;
1465                 }
1466 
1467         }
1468 
1469         if (cur_clone_root) {
1470                 *found = cur_clone_root;
1471                 ret = 0;
1472         } else {
1473                 ret = -ENOENT;
1474         }
1475 
1476 out:
1477         btrfs_free_path(tmp_path);
1478         kfree(backref_ctx);
1479         return ret;
1480 }
1481 
1482 static int read_symlink(struct btrfs_root *root,
1483                         u64 ino,
1484                         struct fs_path *dest)
1485 {
1486         int ret;
1487         struct btrfs_path *path;
1488         struct btrfs_key key;
1489         struct btrfs_file_extent_item *ei;
1490         u8 type;
1491         u8 compression;
1492         unsigned long off;
1493         int len;
1494 
1495         path = alloc_path_for_send();
1496         if (!path)
1497                 return -ENOMEM;
1498 
1499         key.objectid = ino;
1500         key.type = BTRFS_EXTENT_DATA_KEY;
1501         key.offset = 0;
1502         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1503         if (ret < 0)
1504                 goto out;
1505         if (ret) {
1506                 /*
1507                  * An empty symlink inode. Can happen in rare error paths when
1508                  * creating a symlink (transaction committed before the inode
1509                  * eviction handler removed the symlink inode items and a crash
1510                  * happened in between or the subvol was snapshoted in between).
1511                  * Print an informative message to dmesg/syslog so that the user
1512                  * can delete the symlink.
1513                  */
1514                 btrfs_err(root->fs_info,
1515                           "Found empty symlink inode %llu at root %llu",
1516                           ino, root->root_key.objectid);
1517                 ret = -EIO;
1518                 goto out;
1519         }
1520 
1521         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1522                         struct btrfs_file_extent_item);
1523         type = btrfs_file_extent_type(path->nodes[0], ei);
1524         compression = btrfs_file_extent_compression(path->nodes[0], ei);
1525         BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1526         BUG_ON(compression);
1527 
1528         off = btrfs_file_extent_inline_start(ei);
1529         len = btrfs_file_extent_inline_len(path->nodes[0], path->slots[0], ei);
1530 
1531         ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1532 
1533 out:
1534         btrfs_free_path(path);
1535         return ret;
1536 }
1537 
1538 /*
1539  * Helper function to generate a file name that is unique in the root of
1540  * send_root and parent_root. This is used to generate names for orphan inodes.
1541  */
1542 static int gen_unique_name(struct send_ctx *sctx,
1543                            u64 ino, u64 gen,
1544                            struct fs_path *dest)
1545 {
1546         int ret = 0;
1547         struct btrfs_path *path;
1548         struct btrfs_dir_item *di;
1549         char tmp[64];
1550         int len;
1551         u64 idx = 0;
1552 
1553         path = alloc_path_for_send();
1554         if (!path)
1555                 return -ENOMEM;
1556 
1557         while (1) {
1558                 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1559                                 ino, gen, idx);
1560                 ASSERT(len < sizeof(tmp));
1561 
1562                 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1563                                 path, BTRFS_FIRST_FREE_OBJECTID,
1564                                 tmp, strlen(tmp), 0);
1565                 btrfs_release_path(path);
1566                 if (IS_ERR(di)) {
1567                         ret = PTR_ERR(di);
1568                         goto out;
1569                 }
1570                 if (di) {
1571                         /* not unique, try again */
1572                         idx++;
1573                         continue;
1574                 }
1575 
1576                 if (!sctx->parent_root) {
1577                         /* unique */
1578                         ret = 0;
1579                         break;
1580                 }
1581 
1582                 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1583                                 path, BTRFS_FIRST_FREE_OBJECTID,
1584                                 tmp, strlen(tmp), 0);
1585                 btrfs_release_path(path);
1586                 if (IS_ERR(di)) {
1587                         ret = PTR_ERR(di);
1588                         goto out;
1589                 }
1590                 if (di) {
1591                         /* not unique, try again */
1592                         idx++;
1593                         continue;
1594                 }
1595                 /* unique */
1596                 break;
1597         }
1598 
1599         ret = fs_path_add(dest, tmp, strlen(tmp));
1600 
1601 out:
1602         btrfs_free_path(path);
1603         return ret;
1604 }
1605 
1606 enum inode_state {
1607         inode_state_no_change,
1608         inode_state_will_create,
1609         inode_state_did_create,
1610         inode_state_will_delete,
1611         inode_state_did_delete,
1612 };
1613 
1614 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1615 {
1616         int ret;
1617         int left_ret;
1618         int right_ret;
1619         u64 left_gen;
1620         u64 right_gen;
1621 
1622         ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1623                         NULL, NULL);
1624         if (ret < 0 && ret != -ENOENT)
1625                 goto out;
1626         left_ret = ret;
1627 
1628         if (!sctx->parent_root) {
1629                 right_ret = -ENOENT;
1630         } else {
1631                 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1632                                 NULL, NULL, NULL, NULL);
1633                 if (ret < 0 && ret != -ENOENT)
1634                         goto out;
1635                 right_ret = ret;
1636         }
1637 
1638         if (!left_ret && !right_ret) {
1639                 if (left_gen == gen && right_gen == gen) {
1640                         ret = inode_state_no_change;
1641                 } else if (left_gen == gen) {
1642                         if (ino < sctx->send_progress)
1643                                 ret = inode_state_did_create;
1644                         else
1645                                 ret = inode_state_will_create;
1646                 } else if (right_gen == gen) {
1647                         if (ino < sctx->send_progress)
1648                                 ret = inode_state_did_delete;
1649                         else
1650                                 ret = inode_state_will_delete;
1651                 } else  {
1652                         ret = -ENOENT;
1653                 }
1654         } else if (!left_ret) {
1655                 if (left_gen == gen) {
1656                         if (ino < sctx->send_progress)
1657                                 ret = inode_state_did_create;
1658                         else
1659                                 ret = inode_state_will_create;
1660                 } else {
1661                         ret = -ENOENT;
1662                 }
1663         } else if (!right_ret) {
1664                 if (right_gen == gen) {
1665                         if (ino < sctx->send_progress)
1666                                 ret = inode_state_did_delete;
1667                         else
1668                                 ret = inode_state_will_delete;
1669                 } else {
1670                         ret = -ENOENT;
1671                 }
1672         } else {
1673                 ret = -ENOENT;
1674         }
1675 
1676 out:
1677         return ret;
1678 }
1679 
1680 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1681 {
1682         int ret;
1683 
1684         if (ino == BTRFS_FIRST_FREE_OBJECTID)
1685                 return 1;
1686 
1687         ret = get_cur_inode_state(sctx, ino, gen);
1688         if (ret < 0)
1689                 goto out;
1690 
1691         if (ret == inode_state_no_change ||
1692             ret == inode_state_did_create ||
1693             ret == inode_state_will_delete)
1694                 ret = 1;
1695         else
1696                 ret = 0;
1697 
1698 out:
1699         return ret;
1700 }
1701 
1702 /*
1703  * Helper function to lookup a dir item in a dir.
1704  */
1705 static int lookup_dir_item_inode(struct btrfs_root *root,
1706                                  u64 dir, const char *name, int name_len,
1707                                  u64 *found_inode,
1708                                  u8 *found_type)
1709 {
1710         int ret = 0;
1711         struct btrfs_dir_item *di;
1712         struct btrfs_key key;
1713         struct btrfs_path *path;
1714 
1715         path = alloc_path_for_send();
1716         if (!path)
1717                 return -ENOMEM;
1718 
1719         di = btrfs_lookup_dir_item(NULL, root, path,
1720                         dir, name, name_len, 0);
1721         if (!di) {
1722                 ret = -ENOENT;
1723                 goto out;
1724         }
1725         if (IS_ERR(di)) {
1726                 ret = PTR_ERR(di);
1727                 goto out;
1728         }
1729         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1730         if (key.type == BTRFS_ROOT_ITEM_KEY) {
1731                 ret = -ENOENT;
1732                 goto out;
1733         }
1734         *found_inode = key.objectid;
1735         *found_type = btrfs_dir_type(path->nodes[0], di);
1736 
1737 out:
1738         btrfs_free_path(path);
1739         return ret;
1740 }
1741 
1742 /*
1743  * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1744  * generation of the parent dir and the name of the dir entry.
1745  */
1746 static int get_first_ref(struct btrfs_root *root, u64 ino,
1747                          u64 *dir, u64 *dir_gen, struct fs_path *name)
1748 {
1749         int ret;
1750         struct btrfs_key key;
1751         struct btrfs_key found_key;
1752         struct btrfs_path *path;
1753         int len;
1754         u64 parent_dir;
1755 
1756         path = alloc_path_for_send();
1757         if (!path)
1758                 return -ENOMEM;
1759 
1760         key.objectid = ino;
1761         key.type = BTRFS_INODE_REF_KEY;
1762         key.offset = 0;
1763 
1764         ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1765         if (ret < 0)
1766                 goto out;
1767         if (!ret)
1768                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1769                                 path->slots[0]);
1770         if (ret || found_key.objectid != ino ||
1771             (found_key.type != BTRFS_INODE_REF_KEY &&
1772              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1773                 ret = -ENOENT;
1774                 goto out;
1775         }
1776 
1777         if (found_key.type == BTRFS_INODE_REF_KEY) {
1778                 struct btrfs_inode_ref *iref;
1779                 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1780                                       struct btrfs_inode_ref);
1781                 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1782                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1783                                                      (unsigned long)(iref + 1),
1784                                                      len);
1785                 parent_dir = found_key.offset;
1786         } else {
1787                 struct btrfs_inode_extref *extref;
1788                 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1789                                         struct btrfs_inode_extref);
1790                 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1791                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1792                                         (unsigned long)&extref->name, len);
1793                 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1794         }
1795         if (ret < 0)
1796                 goto out;
1797         btrfs_release_path(path);
1798 
1799         if (dir_gen) {
1800                 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1801                                      NULL, NULL, NULL);
1802                 if (ret < 0)
1803                         goto out;
1804         }
1805 
1806         *dir = parent_dir;
1807 
1808 out:
1809         btrfs_free_path(path);
1810         return ret;
1811 }
1812 
1813 static int is_first_ref(struct btrfs_root *root,
1814                         u64 ino, u64 dir,
1815                         const char *name, int name_len)
1816 {
1817         int ret;
1818         struct fs_path *tmp_name;
1819         u64 tmp_dir;
1820 
1821         tmp_name = fs_path_alloc();
1822         if (!tmp_name)
1823                 return -ENOMEM;
1824 
1825         ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1826         if (ret < 0)
1827                 goto out;
1828 
1829         if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1830                 ret = 0;
1831                 goto out;
1832         }
1833 
1834         ret = !memcmp(tmp_name->start, name, name_len);
1835 
1836 out:
1837         fs_path_free(tmp_name);
1838         return ret;
1839 }
1840 
1841 /*
1842  * Used by process_recorded_refs to determine if a new ref would overwrite an
1843  * already existing ref. In case it detects an overwrite, it returns the
1844  * inode/gen in who_ino/who_gen.
1845  * When an overwrite is detected, process_recorded_refs does proper orphanizing
1846  * to make sure later references to the overwritten inode are possible.
1847  * Orphanizing is however only required for the first ref of an inode.
1848  * process_recorded_refs does an additional is_first_ref check to see if
1849  * orphanizing is really required.
1850  */
1851 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1852                               const char *name, int name_len,
1853                               u64 *who_ino, u64 *who_gen)
1854 {
1855         int ret = 0;
1856         u64 gen;
1857         u64 other_inode = 0;
1858         u8 other_type = 0;
1859 
1860         if (!sctx->parent_root)
1861                 goto out;
1862 
1863         ret = is_inode_existent(sctx, dir, dir_gen);
1864         if (ret <= 0)
1865                 goto out;
1866 
1867         /*
1868          * If we have a parent root we need to verify that the parent dir was
1869          * not deleted and then re-created, if it was then we have no overwrite
1870          * and we can just unlink this entry.
1871          */
1872         if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1873                 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1874                                      NULL, NULL, NULL);
1875                 if (ret < 0 && ret != -ENOENT)
1876                         goto out;
1877                 if (ret) {
1878                         ret = 0;
1879                         goto out;
1880                 }
1881                 if (gen != dir_gen)
1882                         goto out;
1883         }
1884 
1885         ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1886                         &other_inode, &other_type);
1887         if (ret < 0 && ret != -ENOENT)
1888                 goto out;
1889         if (ret) {
1890                 ret = 0;
1891                 goto out;
1892         }
1893 
1894         /*
1895          * Check if the overwritten ref was already processed. If yes, the ref
1896          * was already unlinked/moved, so we can safely assume that we will not
1897          * overwrite anything at this point in time.
1898          */
1899         if (other_inode > sctx->send_progress ||
1900             is_waiting_for_move(sctx, other_inode)) {
1901                 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1902                                 who_gen, NULL, NULL, NULL, NULL);
1903                 if (ret < 0)
1904                         goto out;
1905 
1906                 ret = 1;
1907                 *who_ino = other_inode;
1908         } else {
1909                 ret = 0;
1910         }
1911 
1912 out:
1913         return ret;
1914 }
1915 
1916 /*
1917  * Checks if the ref was overwritten by an already processed inode. This is
1918  * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1919  * thus the orphan name needs be used.
1920  * process_recorded_refs also uses it to avoid unlinking of refs that were
1921  * overwritten.
1922  */
1923 static int did_overwrite_ref(struct send_ctx *sctx,
1924                             u64 dir, u64 dir_gen,
1925                             u64 ino, u64 ino_gen,
1926                             const char *name, int name_len)
1927 {
1928         int ret = 0;
1929         u64 gen;
1930         u64 ow_inode;
1931         u8 other_type;
1932 
1933         if (!sctx->parent_root)
1934                 goto out;
1935 
1936         ret = is_inode_existent(sctx, dir, dir_gen);
1937         if (ret <= 0)
1938                 goto out;
1939 
1940         if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1941                 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1942                                      NULL, NULL, NULL);
1943                 if (ret < 0 && ret != -ENOENT)
1944                         goto out;
1945                 if (ret) {
1946                         ret = 0;
1947                         goto out;
1948                 }
1949                 if (gen != dir_gen)
1950                         goto out;
1951         }
1952 
1953         /* check if the ref was overwritten by another ref */
1954         ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1955                         &ow_inode, &other_type);
1956         if (ret < 0 && ret != -ENOENT)
1957                 goto out;
1958         if (ret) {
1959                 /* was never and will never be overwritten */
1960                 ret = 0;
1961                 goto out;
1962         }
1963 
1964         ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1965                         NULL, NULL);
1966         if (ret < 0)
1967                 goto out;
1968 
1969         if (ow_inode == ino && gen == ino_gen) {
1970                 ret = 0;
1971                 goto out;
1972         }
1973 
1974         /*
1975          * We know that it is or will be overwritten. Check this now.
1976          * The current inode being processed might have been the one that caused
1977          * inode 'ino' to be orphanized, therefore check if ow_inode matches
1978          * the current inode being processed.
1979          */
1980         if ((ow_inode < sctx->send_progress) ||
1981             (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1982              gen == sctx->cur_inode_gen))
1983                 ret = 1;
1984         else
1985                 ret = 0;
1986 
1987 out:
1988         return ret;
1989 }
1990 
1991 /*
1992  * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1993  * that got overwritten. This is used by process_recorded_refs to determine
1994  * if it has to use the path as returned by get_cur_path or the orphan name.
1995  */
1996 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1997 {
1998         int ret = 0;
1999         struct fs_path *name = NULL;
2000         u64 dir;
2001         u64 dir_gen;
2002 
2003         if (!sctx->parent_root)
2004                 goto out;
2005 
2006         name = fs_path_alloc();
2007         if (!name)
2008                 return -ENOMEM;
2009 
2010         ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2011         if (ret < 0)
2012                 goto out;
2013 
2014         ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2015                         name->start, fs_path_len(name));
2016 
2017 out:
2018         fs_path_free(name);
2019         return ret;
2020 }
2021 
2022 /*
2023  * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
2024  * so we need to do some special handling in case we have clashes. This function
2025  * takes care of this with the help of name_cache_entry::radix_list.
2026  * In case of error, nce is kfreed.
2027  */
2028 static int name_cache_insert(struct send_ctx *sctx,
2029                              struct name_cache_entry *nce)
2030 {
2031         int ret = 0;
2032         struct list_head *nce_head;
2033 
2034         nce_head = radix_tree_lookup(&sctx->name_cache,
2035                         (unsigned long)nce->ino);
2036         if (!nce_head) {
2037                 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2038                 if (!nce_head) {
2039                         kfree(nce);
2040                         return -ENOMEM;
2041                 }
2042                 INIT_LIST_HEAD(nce_head);
2043 
2044                 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2045                 if (ret < 0) {
2046                         kfree(nce_head);
2047                         kfree(nce);
2048                         return ret;
2049                 }
2050         }
2051         list_add_tail(&nce->radix_list, nce_head);
2052         list_add_tail(&nce->list, &sctx->name_cache_list);
2053         sctx->name_cache_size++;
2054 
2055         return ret;
2056 }
2057 
2058 static void name_cache_delete(struct send_ctx *sctx,
2059                               struct name_cache_entry *nce)
2060 {
2061         struct list_head *nce_head;
2062 
2063         nce_head = radix_tree_lookup(&sctx->name_cache,
2064                         (unsigned long)nce->ino);
2065         if (!nce_head) {
2066                 btrfs_err(sctx->send_root->fs_info,
2067               "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2068                         nce->ino, sctx->name_cache_size);
2069         }
2070 
2071         list_del(&nce->radix_list);
2072         list_del(&nce->list);
2073         sctx->name_cache_size--;
2074 
2075         /*
2076          * We may not get to the final release of nce_head if the lookup fails
2077          */
2078         if (nce_head && list_empty(nce_head)) {
2079                 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2080                 kfree(nce_head);
2081         }
2082 }
2083 
2084 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2085                                                     u64 ino, u64 gen)
2086 {
2087         struct list_head *nce_head;
2088         struct name_cache_entry *cur;
2089 
2090         nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2091         if (!nce_head)
2092                 return NULL;
2093 
2094         list_for_each_entry(cur, nce_head, radix_list) {
2095                 if (cur->ino == ino && cur->gen == gen)
2096                         return cur;
2097         }
2098         return NULL;
2099 }
2100 
2101 /*
2102  * Removes the entry from the list and adds it back to the end. This marks the
2103  * entry as recently used so that name_cache_clean_unused does not remove it.
2104  */
2105 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2106 {
2107         list_del(&nce->list);
2108         list_add_tail(&nce->list, &sctx->name_cache_list);
2109 }
2110 
2111 /*
2112  * Remove some entries from the beginning of name_cache_list.
2113  */
2114 static void name_cache_clean_unused(struct send_ctx *sctx)
2115 {
2116         struct name_cache_entry *nce;
2117 
2118         if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2119                 return;
2120 
2121         while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2122                 nce = list_entry(sctx->name_cache_list.next,
2123                                 struct name_cache_entry, list);
2124                 name_cache_delete(sctx, nce);
2125                 kfree(nce);
2126         }
2127 }
2128 
2129 static void name_cache_free(struct send_ctx *sctx)
2130 {
2131         struct name_cache_entry *nce;
2132 
2133         while (!list_empty(&sctx->name_cache_list)) {
2134                 nce = list_entry(sctx->name_cache_list.next,
2135                                 struct name_cache_entry, list);
2136                 name_cache_delete(sctx, nce);
2137                 kfree(nce);
2138         }
2139 }
2140 
2141 /*
2142  * Used by get_cur_path for each ref up to the root.
2143  * Returns 0 if it succeeded.
2144  * Returns 1 if the inode is not existent or got overwritten. In that case, the
2145  * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2146  * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2147  * Returns <0 in case of error.
2148  */
2149 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2150                                      u64 ino, u64 gen,
2151                                      u64 *parent_ino,
2152                                      u64 *parent_gen,
2153                                      struct fs_path *dest)
2154 {
2155         int ret;
2156         int nce_ret;
2157         struct name_cache_entry *nce = NULL;
2158 
2159         /*
2160          * First check if we already did a call to this function with the same
2161          * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2162          * return the cached result.
2163          */
2164         nce = name_cache_search(sctx, ino, gen);
2165         if (nce) {
2166                 if (ino < sctx->send_progress && nce->need_later_update) {
2167                         name_cache_delete(sctx, nce);
2168                         kfree(nce);
2169                         nce = NULL;
2170                 } else {
2171                         name_cache_used(sctx, nce);
2172                         *parent_ino = nce->parent_ino;
2173                         *parent_gen = nce->parent_gen;
2174                         ret = fs_path_add(dest, nce->name, nce->name_len);
2175                         if (ret < 0)
2176                                 goto out;
2177                         ret = nce->ret;
2178                         goto out;
2179                 }
2180         }
2181 
2182         /*
2183          * If the inode is not existent yet, add the orphan name and return 1.
2184          * This should only happen for the parent dir that we determine in
2185          * __record_new_ref
2186          */
2187         ret = is_inode_existent(sctx, ino, gen);
2188         if (ret < 0)
2189                 goto out;
2190 
2191         if (!ret) {
2192                 ret = gen_unique_name(sctx, ino, gen, dest);
2193                 if (ret < 0)
2194                         goto out;
2195                 ret = 1;
2196                 goto out_cache;
2197         }
2198 
2199         /*
2200          * Depending on whether the inode was already processed or not, use
2201          * send_root or parent_root for ref lookup.
2202          */
2203         if (ino < sctx->send_progress)
2204                 ret = get_first_ref(sctx->send_root, ino,
2205                                     parent_ino, parent_gen, dest);
2206         else
2207                 ret = get_first_ref(sctx->parent_root, ino,
2208                                     parent_ino, parent_gen, dest);
2209         if (ret < 0)
2210                 goto out;
2211 
2212         /*
2213          * Check if the ref was overwritten by an inode's ref that was processed
2214          * earlier. If yes, treat as orphan and return 1.
2215          */
2216         ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2217                         dest->start, dest->end - dest->start);
2218         if (ret < 0)
2219                 goto out;
2220         if (ret) {
2221                 fs_path_reset(dest);
2222                 ret = gen_unique_name(sctx, ino, gen, dest);
2223                 if (ret < 0)
2224                         goto out;
2225                 ret = 1;
2226         }
2227 
2228 out_cache:
2229         /*
2230          * Store the result of the lookup in the name cache.
2231          */
2232         nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2233         if (!nce) {
2234                 ret = -ENOMEM;
2235                 goto out;
2236         }
2237 
2238         nce->ino = ino;
2239         nce->gen = gen;
2240         nce->parent_ino = *parent_ino;
2241         nce->parent_gen = *parent_gen;
2242         nce->name_len = fs_path_len(dest);
2243         nce->ret = ret;
2244         strcpy(nce->name, dest->start);
2245 
2246         if (ino < sctx->send_progress)
2247                 nce->need_later_update = 0;
2248         else
2249                 nce->need_later_update = 1;
2250 
2251         nce_ret = name_cache_insert(sctx, nce);
2252         if (nce_ret < 0)
2253                 ret = nce_ret;
2254         name_cache_clean_unused(sctx);
2255 
2256 out:
2257         return ret;
2258 }
2259 
2260 /*
2261  * Magic happens here. This function returns the first ref to an inode as it
2262  * would look like while receiving the stream at this point in time.
2263  * We walk the path up to the root. For every inode in between, we check if it
2264  * was already processed/sent. If yes, we continue with the parent as found
2265  * in send_root. If not, we continue with the parent as found in parent_root.
2266  * If we encounter an inode that was deleted at this point in time, we use the
2267  * inodes "orphan" name instead of the real name and stop. Same with new inodes
2268  * that were not created yet and overwritten inodes/refs.
2269  *
2270  * When do we have have orphan inodes:
2271  * 1. When an inode is freshly created and thus no valid refs are available yet
2272  * 2. When a directory lost all it's refs (deleted) but still has dir items
2273  *    inside which were not processed yet (pending for move/delete). If anyone
2274  *    tried to get the path to the dir items, it would get a path inside that
2275  *    orphan directory.
2276  * 3. When an inode is moved around or gets new links, it may overwrite the ref
2277  *    of an unprocessed inode. If in that case the first ref would be
2278  *    overwritten, the overwritten inode gets "orphanized". Later when we
2279  *    process this overwritten inode, it is restored at a new place by moving
2280  *    the orphan inode.
2281  *
2282  * sctx->send_progress tells this function at which point in time receiving
2283  * would be.
2284  */
2285 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2286                         struct fs_path *dest)
2287 {
2288         int ret = 0;
2289         struct fs_path *name = NULL;
2290         u64 parent_inode = 0;
2291         u64 parent_gen = 0;
2292         int stop = 0;
2293 
2294         name = fs_path_alloc();
2295         if (!name) {
2296                 ret = -ENOMEM;
2297                 goto out;
2298         }
2299 
2300         dest->reversed = 1;
2301         fs_path_reset(dest);
2302 
2303         while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2304                 struct waiting_dir_move *wdm;
2305 
2306                 fs_path_reset(name);
2307 
2308                 if (is_waiting_for_rm(sctx, ino)) {
2309                         ret = gen_unique_name(sctx, ino, gen, name);
2310                         if (ret < 0)
2311                                 goto out;
2312                         ret = fs_path_add_path(dest, name);
2313                         break;
2314                 }
2315 
2316                 wdm = get_waiting_dir_move(sctx, ino);
2317                 if (wdm && wdm->orphanized) {
2318                         ret = gen_unique_name(sctx, ino, gen, name);
2319                         stop = 1;
2320                 } else if (wdm) {
2321                         ret = get_first_ref(sctx->parent_root, ino,
2322                                             &parent_inode, &parent_gen, name);
2323                 } else {
2324                         ret = __get_cur_name_and_parent(sctx, ino, gen,
2325                                                         &parent_inode,
2326                                                         &parent_gen, name);
2327                         if (ret)
2328                                 stop = 1;
2329                 }
2330 
2331                 if (ret < 0)
2332                         goto out;
2333 
2334                 ret = fs_path_add_path(dest, name);
2335                 if (ret < 0)
2336                         goto out;
2337 
2338                 ino = parent_inode;
2339                 gen = parent_gen;
2340         }
2341 
2342 out:
2343         fs_path_free(name);
2344         if (!ret)
2345                 fs_path_unreverse(dest);
2346         return ret;
2347 }
2348 
2349 /*
2350  * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2351  */
2352 static int send_subvol_begin(struct send_ctx *sctx)
2353 {
2354         int ret;
2355         struct btrfs_root *send_root = sctx->send_root;
2356         struct btrfs_root *parent_root = sctx->parent_root;
2357         struct btrfs_path *path;
2358         struct btrfs_key key;
2359         struct btrfs_root_ref *ref;
2360         struct extent_buffer *leaf;
2361         char *name = NULL;
2362         int namelen;
2363 
2364         path = btrfs_alloc_path();
2365         if (!path)
2366                 return -ENOMEM;
2367 
2368         name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2369         if (!name) {
2370                 btrfs_free_path(path);
2371                 return -ENOMEM;
2372         }
2373 
2374         key.objectid = send_root->objectid;
2375         key.type = BTRFS_ROOT_BACKREF_KEY;
2376         key.offset = 0;
2377 
2378         ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2379                                 &key, path, 1, 0);
2380         if (ret < 0)
2381                 goto out;
2382         if (ret) {
2383                 ret = -ENOENT;
2384                 goto out;
2385         }
2386 
2387         leaf = path->nodes[0];
2388         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2389         if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2390             key.objectid != send_root->objectid) {
2391                 ret = -ENOENT;
2392                 goto out;
2393         }
2394         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2395         namelen = btrfs_root_ref_name_len(leaf, ref);
2396         read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2397         btrfs_release_path(path);
2398 
2399         if (parent_root) {
2400                 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2401                 if (ret < 0)
2402                         goto out;
2403         } else {
2404                 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2405                 if (ret < 0)
2406                         goto out;
2407         }
2408 
2409         TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2410 
2411         if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2412                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2413                             sctx->send_root->root_item.received_uuid);
2414         else
2415                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2416                             sctx->send_root->root_item.uuid);
2417 
2418         TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2419                     le64_to_cpu(sctx->send_root->root_item.ctransid));
2420         if (parent_root) {
2421                 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2422                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2423                                      parent_root->root_item.received_uuid);
2424                 else
2425                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2426                                      parent_root->root_item.uuid);
2427                 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2428                             le64_to_cpu(sctx->parent_root->root_item.ctransid));
2429         }
2430 
2431         ret = send_cmd(sctx);
2432 
2433 tlv_put_failure:
2434 out:
2435         btrfs_free_path(path);
2436         kfree(name);
2437         return ret;
2438 }
2439 
2440 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2441 {
2442         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2443         int ret = 0;
2444         struct fs_path *p;
2445 
2446         btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2447 
2448         p = fs_path_alloc();
2449         if (!p)
2450                 return -ENOMEM;
2451 
2452         ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2453         if (ret < 0)
2454                 goto out;
2455 
2456         ret = get_cur_path(sctx, ino, gen, p);
2457         if (ret < 0)
2458                 goto out;
2459         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2460         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2461 
2462         ret = send_cmd(sctx);
2463 
2464 tlv_put_failure:
2465 out:
2466         fs_path_free(p);
2467         return ret;
2468 }
2469 
2470 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2471 {
2472         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2473         int ret = 0;
2474         struct fs_path *p;
2475 
2476         btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2477 
2478         p = fs_path_alloc();
2479         if (!p)
2480                 return -ENOMEM;
2481 
2482         ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2483         if (ret < 0)
2484                 goto out;
2485 
2486         ret = get_cur_path(sctx, ino, gen, p);
2487         if (ret < 0)
2488                 goto out;
2489         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2490         TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2491 
2492         ret = send_cmd(sctx);
2493 
2494 tlv_put_failure:
2495 out:
2496         fs_path_free(p);
2497         return ret;
2498 }
2499 
2500 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2501 {
2502         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2503         int ret = 0;
2504         struct fs_path *p;
2505 
2506         btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2507                     ino, uid, gid);
2508 
2509         p = fs_path_alloc();
2510         if (!p)
2511                 return -ENOMEM;
2512 
2513         ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2514         if (ret < 0)
2515                 goto out;
2516 
2517         ret = get_cur_path(sctx, ino, gen, p);
2518         if (ret < 0)
2519                 goto out;
2520         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2521         TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2522         TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2523 
2524         ret = send_cmd(sctx);
2525 
2526 tlv_put_failure:
2527 out:
2528         fs_path_free(p);
2529         return ret;
2530 }
2531 
2532 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2533 {
2534         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2535         int ret = 0;
2536         struct fs_path *p = NULL;
2537         struct btrfs_inode_item *ii;
2538         struct btrfs_path *path = NULL;
2539         struct extent_buffer *eb;
2540         struct btrfs_key key;
2541         int slot;
2542 
2543         btrfs_debug(fs_info, "send_utimes %llu", ino);
2544 
2545         p = fs_path_alloc();
2546         if (!p)
2547                 return -ENOMEM;
2548 
2549         path = alloc_path_for_send();
2550         if (!path) {
2551                 ret = -ENOMEM;
2552                 goto out;
2553         }
2554 
2555         key.objectid = ino;
2556         key.type = BTRFS_INODE_ITEM_KEY;
2557         key.offset = 0;
2558         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2559         if (ret > 0)
2560                 ret = -ENOENT;
2561         if (ret < 0)
2562                 goto out;
2563 
2564         eb = path->nodes[0];
2565         slot = path->slots[0];
2566         ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2567 
2568         ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2569         if (ret < 0)
2570                 goto out;
2571 
2572         ret = get_cur_path(sctx, ino, gen, p);
2573         if (ret < 0)
2574                 goto out;
2575         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2576         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2577         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2578         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2579         /* TODO Add otime support when the otime patches get into upstream */
2580 
2581         ret = send_cmd(sctx);
2582 
2583 tlv_put_failure:
2584 out:
2585         fs_path_free(p);
2586         btrfs_free_path(path);
2587         return ret;
2588 }
2589 
2590 /*
2591  * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2592  * a valid path yet because we did not process the refs yet. So, the inode
2593  * is created as orphan.
2594  */
2595 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2596 {
2597         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2598         int ret = 0;
2599         struct fs_path *p;
2600         int cmd;
2601         u64 gen;
2602         u64 mode;
2603         u64 rdev;
2604 
2605         btrfs_debug(fs_info, "send_create_inode %llu", ino);
2606 
2607         p = fs_path_alloc();
2608         if (!p)
2609                 return -ENOMEM;
2610 
2611         if (ino != sctx->cur_ino) {
2612                 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2613                                      NULL, NULL, &rdev);
2614                 if (ret < 0)
2615                         goto out;
2616         } else {
2617                 gen = sctx->cur_inode_gen;
2618                 mode = sctx->cur_inode_mode;
2619                 rdev = sctx->cur_inode_rdev;
2620         }
2621 
2622         if (S_ISREG(mode)) {
2623                 cmd = BTRFS_SEND_C_MKFILE;
2624         } else if (S_ISDIR(mode)) {
2625                 cmd = BTRFS_SEND_C_MKDIR;
2626         } else if (S_ISLNK(mode)) {
2627                 cmd = BTRFS_SEND_C_SYMLINK;
2628         } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2629                 cmd = BTRFS_SEND_C_MKNOD;
2630         } else if (S_ISFIFO(mode)) {
2631                 cmd = BTRFS_SEND_C_MKFIFO;
2632         } else if (S_ISSOCK(mode)) {
2633                 cmd = BTRFS_SEND_C_MKSOCK;
2634         } else {
2635                 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2636                                 (int)(mode & S_IFMT));
2637                 ret = -ENOTSUPP;
2638                 goto out;
2639         }
2640 
2641         ret = begin_cmd(sctx, cmd);
2642         if (ret < 0)
2643                 goto out;
2644 
2645         ret = gen_unique_name(sctx, ino, gen, p);
2646         if (ret < 0)
2647                 goto out;
2648 
2649         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2650         TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2651 
2652         if (S_ISLNK(mode)) {
2653                 fs_path_reset(p);
2654                 ret = read_symlink(sctx->send_root, ino, p);
2655                 if (ret < 0)
2656                         goto out;
2657                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2658         } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2659                    S_ISFIFO(mode) || S_ISSOCK(mode)) {
2660                 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2661                 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2662         }
2663 
2664         ret = send_cmd(sctx);
2665         if (ret < 0)
2666                 goto out;
2667 
2668 
2669 tlv_put_failure:
2670 out:
2671         fs_path_free(p);
2672         return ret;
2673 }
2674 
2675 /*
2676  * We need some special handling for inodes that get processed before the parent
2677  * directory got created. See process_recorded_refs for details.
2678  * This function does the check if we already created the dir out of order.
2679  */
2680 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2681 {
2682         int ret = 0;
2683         struct btrfs_path *path = NULL;
2684         struct btrfs_key key;
2685         struct btrfs_key found_key;
2686         struct btrfs_key di_key;
2687         struct extent_buffer *eb;
2688         struct btrfs_dir_item *di;
2689         int slot;
2690 
2691         path = alloc_path_for_send();
2692         if (!path) {
2693                 ret = -ENOMEM;
2694                 goto out;
2695         }
2696 
2697         key.objectid = dir;
2698         key.type = BTRFS_DIR_INDEX_KEY;
2699         key.offset = 0;
2700         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2701         if (ret < 0)
2702                 goto out;
2703 
2704         while (1) {
2705                 eb = path->nodes[0];
2706                 slot = path->slots[0];
2707                 if (slot >= btrfs_header_nritems(eb)) {
2708                         ret = btrfs_next_leaf(sctx->send_root, path);
2709                         if (ret < 0) {
2710                                 goto out;
2711                         } else if (ret > 0) {
2712                                 ret = 0;
2713                                 break;
2714                         }
2715                         continue;
2716                 }
2717 
2718                 btrfs_item_key_to_cpu(eb, &found_key, slot);
2719                 if (found_key.objectid != key.objectid ||
2720                     found_key.type != key.type) {
2721                         ret = 0;
2722                         goto out;
2723                 }
2724 
2725                 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2726                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2727 
2728                 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2729                     di_key.objectid < sctx->send_progress) {
2730                         ret = 1;
2731                         goto out;
2732                 }
2733 
2734                 path->slots[0]++;
2735         }
2736 
2737 out:
2738         btrfs_free_path(path);
2739         return ret;
2740 }
2741 
2742 /*
2743  * Only creates the inode if it is:
2744  * 1. Not a directory
2745  * 2. Or a directory which was not created already due to out of order
2746  *    directories. See did_create_dir and process_recorded_refs for details.
2747  */
2748 static int send_create_inode_if_needed(struct send_ctx *sctx)
2749 {
2750         int ret;
2751 
2752         if (S_ISDIR(sctx->cur_inode_mode)) {
2753                 ret = did_create_dir(sctx, sctx->cur_ino);
2754                 if (ret < 0)
2755                         goto out;
2756                 if (ret) {
2757                         ret = 0;
2758                         goto out;
2759                 }
2760         }
2761 
2762         ret = send_create_inode(sctx, sctx->cur_ino);
2763         if (ret < 0)
2764                 goto out;
2765 
2766 out:
2767         return ret;
2768 }
2769 
2770 struct recorded_ref {
2771         struct list_head list;
2772         char *dir_path;
2773         char *name;
2774         struct fs_path *full_path;
2775         u64 dir;
2776         u64 dir_gen;
2777         int dir_path_len;
2778         int name_len;
2779 };
2780 
2781 /*
2782  * We need to process new refs before deleted refs, but compare_tree gives us
2783  * everything mixed. So we first record all refs and later process them.
2784  * This function is a helper to record one ref.
2785  */
2786 static int __record_ref(struct list_head *head, u64 dir,
2787                       u64 dir_gen, struct fs_path *path)
2788 {
2789         struct recorded_ref *ref;
2790 
2791         ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2792         if (!ref)
2793                 return -ENOMEM;
2794 
2795         ref->dir = dir;
2796         ref->dir_gen = dir_gen;
2797         ref->full_path = path;
2798 
2799         ref->name = (char *)kbasename(ref->full_path->start);
2800         ref->name_len = ref->full_path->end - ref->name;
2801         ref->dir_path = ref->full_path->start;
2802         if (ref->name == ref->full_path->start)
2803                 ref->dir_path_len = 0;
2804         else
2805                 ref->dir_path_len = ref->full_path->end -
2806                                 ref->full_path->start - 1 - ref->name_len;
2807 
2808         list_add_tail(&ref->list, head);
2809         return 0;
2810 }
2811 
2812 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2813 {
2814         struct recorded_ref *new;
2815 
2816         new = kmalloc(sizeof(*ref), GFP_KERNEL);
2817         if (!new)
2818                 return -ENOMEM;
2819 
2820         new->dir = ref->dir;
2821         new->dir_gen = ref->dir_gen;
2822         new->full_path = NULL;
2823         INIT_LIST_HEAD(&new->list);
2824         list_add_tail(&new->list, list);
2825         return 0;
2826 }
2827 
2828 static void __free_recorded_refs(struct list_head *head)
2829 {
2830         struct recorded_ref *cur;
2831 
2832         while (!list_empty(head)) {
2833                 cur = list_entry(head->next, struct recorded_ref, list);
2834                 fs_path_free(cur->full_path);
2835                 list_del(&cur->list);
2836                 kfree(cur);
2837         }
2838 }
2839 
2840 static void free_recorded_refs(struct send_ctx *sctx)
2841 {
2842         __free_recorded_refs(&sctx->new_refs);
2843         __free_recorded_refs(&sctx->deleted_refs);
2844 }
2845 
2846 /*
2847  * Renames/moves a file/dir to its orphan name. Used when the first
2848  * ref of an unprocessed inode gets overwritten and for all non empty
2849  * directories.
2850  */
2851 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2852                           struct fs_path *path)
2853 {
2854         int ret;
2855         struct fs_path *orphan;
2856 
2857         orphan = fs_path_alloc();
2858         if (!orphan)
2859                 return -ENOMEM;
2860 
2861         ret = gen_unique_name(sctx, ino, gen, orphan);
2862         if (ret < 0)
2863                 goto out;
2864 
2865         ret = send_rename(sctx, path, orphan);
2866 
2867 out:
2868         fs_path_free(orphan);
2869         return ret;
2870 }
2871 
2872 static struct orphan_dir_info *
2873 add_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2874 {
2875         struct rb_node **p = &sctx->orphan_dirs.rb_node;
2876         struct rb_node *parent = NULL;
2877         struct orphan_dir_info *entry, *odi;
2878 
2879         odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2880         if (!odi)
2881                 return ERR_PTR(-ENOMEM);
2882         odi->ino = dir_ino;
2883         odi->gen = 0;
2884 
2885         while (*p) {
2886                 parent = *p;
2887                 entry = rb_entry(parent, struct orphan_dir_info, node);
2888                 if (dir_ino < entry->ino) {
2889                         p = &(*p)->rb_left;
2890                 } else if (dir_ino > entry->ino) {
2891                         p = &(*p)->rb_right;
2892                 } else {
2893                         kfree(odi);
2894                         return entry;
2895                 }
2896         }
2897 
2898         rb_link_node(&odi->node, parent, p);
2899         rb_insert_color(&odi->node, &sctx->orphan_dirs);
2900         return odi;
2901 }
2902 
2903 static struct orphan_dir_info *
2904 get_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2905 {
2906         struct rb_node *n = sctx->orphan_dirs.rb_node;
2907         struct orphan_dir_info *entry;
2908 
2909         while (n) {
2910                 entry = rb_entry(n, struct orphan_dir_info, node);
2911                 if (dir_ino < entry->ino)
2912                         n = n->rb_left;
2913                 else if (dir_ino > entry->ino)
2914                         n = n->rb_right;
2915                 else
2916                         return entry;
2917         }
2918         return NULL;
2919 }
2920 
2921 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino)
2922 {
2923         struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino);
2924 
2925         return odi != NULL;
2926 }
2927 
2928 static void free_orphan_dir_info(struct send_ctx *sctx,
2929                                  struct orphan_dir_info *odi)
2930 {
2931         if (!odi)
2932                 return;
2933         rb_erase(&odi->node, &sctx->orphan_dirs);
2934         kfree(odi);
2935 }
2936 
2937 /*
2938  * Returns 1 if a directory can be removed at this point in time.
2939  * We check this by iterating all dir items and checking if the inode behind
2940  * the dir item was already processed.
2941  */
2942 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2943                      u64 send_progress)
2944 {
2945         int ret = 0;
2946         struct btrfs_root *root = sctx->parent_root;
2947         struct btrfs_path *path;
2948         struct btrfs_key key;
2949         struct btrfs_key found_key;
2950         struct btrfs_key loc;
2951         struct btrfs_dir_item *di;
2952 
2953         /*
2954          * Don't try to rmdir the top/root subvolume dir.
2955          */
2956         if (dir == BTRFS_FIRST_FREE_OBJECTID)
2957                 return 0;
2958 
2959         path = alloc_path_for_send();
2960         if (!path)
2961                 return -ENOMEM;
2962 
2963         key.objectid = dir;
2964         key.type = BTRFS_DIR_INDEX_KEY;
2965         key.offset = 0;
2966         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2967         if (ret < 0)
2968                 goto out;
2969 
2970         while (1) {
2971                 struct waiting_dir_move *dm;
2972 
2973                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2974                         ret = btrfs_next_leaf(root, path);
2975                         if (ret < 0)
2976                                 goto out;
2977                         else if (ret > 0)
2978                                 break;
2979                         continue;
2980                 }
2981                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2982                                       path->slots[0]);
2983                 if (found_key.objectid != key.objectid ||
2984                     found_key.type != key.type)
2985                         break;
2986 
2987                 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2988                                 struct btrfs_dir_item);
2989                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2990 
2991                 dm = get_waiting_dir_move(sctx, loc.objectid);
2992                 if (dm) {
2993                         struct orphan_dir_info *odi;
2994 
2995                         odi = add_orphan_dir_info(sctx, dir);
2996                         if (IS_ERR(odi)) {
2997                                 ret = PTR_ERR(odi);
2998                                 goto out;
2999                         }
3000                         odi->gen = dir_gen;
3001                         dm->rmdir_ino = dir;
3002                         ret = 0;
3003                         goto out;
3004                 }
3005 
3006                 if (loc.objectid > send_progress) {
3007                         struct orphan_dir_info *odi;
3008 
3009                         odi = get_orphan_dir_info(sctx, dir);
3010                         free_orphan_dir_info(sctx, odi);
3011                         ret = 0;
3012                         goto out;
3013                 }
3014 
3015                 path->slots[0]++;
3016         }
3017 
3018         ret = 1;
3019 
3020 out:
3021         btrfs_free_path(path);
3022         return ret;
3023 }
3024 
3025 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3026 {
3027         struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3028 
3029         return entry != NULL;
3030 }
3031 
3032 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3033 {
3034         struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3035         struct rb_node *parent = NULL;
3036         struct waiting_dir_move *entry, *dm;
3037 
3038         dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3039         if (!dm)
3040                 return -ENOMEM;
3041         dm->ino = ino;
3042         dm->rmdir_ino = 0;
3043         dm->orphanized = orphanized;
3044 
3045         while (*p) {
3046                 parent = *p;
3047                 entry = rb_entry(parent, struct waiting_dir_move, node);
3048                 if (ino < entry->ino) {
3049                         p = &(*p)->rb_left;
3050                 } else if (ino > entry->ino) {
3051                         p = &(*p)->rb_right;
3052                 } else {
3053                         kfree(dm);
3054                         return -EEXIST;
3055                 }
3056         }
3057 
3058         rb_link_node(&dm->node, parent, p);
3059         rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3060         return 0;
3061 }
3062 
3063 static struct waiting_dir_move *
3064 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3065 {
3066         struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3067         struct waiting_dir_move *entry;
3068 
3069         while (n) {
3070                 entry = rb_entry(n, struct waiting_dir_move, node);
3071                 if (ino < entry->ino)
3072                         n = n->rb_left;
3073                 else if (ino > entry->ino)
3074                         n = n->rb_right;
3075                 else
3076                         return entry;
3077         }
3078         return NULL;
3079 }
3080 
3081 static void free_waiting_dir_move(struct send_ctx *sctx,
3082                                   struct waiting_dir_move *dm)
3083 {
3084         if (!dm)
3085                 return;
3086         rb_erase(&dm->node, &sctx->waiting_dir_moves);
3087         kfree(dm);
3088 }
3089 
3090 static int add_pending_dir_move(struct send_ctx *sctx,
3091                                 u64 ino,
3092                                 u64 ino_gen,
3093                                 u64 parent_ino,
3094                                 struct list_head *new_refs,
3095                                 struct list_head *deleted_refs,
3096                                 const bool is_orphan)
3097 {
3098         struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3099         struct rb_node *parent = NULL;
3100         struct pending_dir_move *entry = NULL, *pm;
3101         struct recorded_ref *cur;
3102         int exists = 0;
3103         int ret;
3104 
3105         pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3106         if (!pm)
3107                 return -ENOMEM;
3108         pm->parent_ino = parent_ino;
3109         pm->ino = ino;
3110         pm->gen = ino_gen;
3111         INIT_LIST_HEAD(&pm->list);
3112         INIT_LIST_HEAD(&pm->update_refs);
3113         RB_CLEAR_NODE(&pm->node);
3114 
3115         while (*p) {
3116                 parent = *p;
3117                 entry = rb_entry(parent, struct pending_dir_move, node);
3118                 if (parent_ino < entry->parent_ino) {
3119                         p = &(*p)->rb_left;
3120                 } else if (parent_ino > entry->parent_ino) {
3121                         p = &(*p)->rb_right;
3122                 } else {
3123                         exists = 1;
3124                         break;
3125                 }
3126         }
3127 
3128         list_for_each_entry(cur, deleted_refs, list) {
3129                 ret = dup_ref(cur, &pm->update_refs);
3130                 if (ret < 0)
3131                         goto out;
3132         }
3133         list_for_each_entry(cur, new_refs, list) {
3134                 ret = dup_ref(cur, &pm->update_refs);
3135                 if (ret < 0)
3136                         goto out;
3137         }
3138 
3139         ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3140         if (ret)
3141                 goto out;
3142 
3143         if (exists) {
3144                 list_add_tail(&pm->list, &entry->list);
3145         } else {
3146                 rb_link_node(&pm->node, parent, p);
3147                 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3148         }
3149         ret = 0;
3150 out:
3151         if (ret) {
3152                 __free_recorded_refs(&pm->update_refs);
3153                 kfree(pm);
3154         }
3155         return ret;
3156 }
3157 
3158 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3159                                                       u64 parent_ino)
3160 {
3161         struct rb_node *n = sctx->pending_dir_moves.rb_node;
3162         struct pending_dir_move *entry;
3163 
3164         while (n) {
3165                 entry = rb_entry(n, struct pending_dir_move, node);
3166                 if (parent_ino < entry->parent_ino)
3167                         n = n->rb_left;
3168                 else if (parent_ino > entry->parent_ino)
3169                         n = n->rb_right;
3170                 else
3171                         return entry;
3172         }
3173         return NULL;
3174 }
3175 
3176 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3177                      u64 ino, u64 gen, u64 *ancestor_ino)
3178 {
3179         int ret = 0;
3180         u64 parent_inode = 0;
3181         u64 parent_gen = 0;
3182         u64 start_ino = ino;
3183 
3184         *ancestor_ino = 0;
3185         while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3186                 fs_path_reset(name);
3187 
3188                 if (is_waiting_for_rm(sctx, ino))
3189                         break;
3190                 if (is_waiting_for_move(sctx, ino)) {
3191                         if (*ancestor_ino == 0)
3192                                 *ancestor_ino = ino;
3193                         ret = get_first_ref(sctx->parent_root, ino,
3194                                             &parent_inode, &parent_gen, name);
3195                 } else {
3196                         ret = __get_cur_name_and_parent(sctx, ino, gen,
3197                                                         &parent_inode,
3198                                                         &parent_gen, name);
3199                         if (ret > 0) {
3200                                 ret = 0;
3201                                 break;
3202                         }
3203                 }
3204                 if (ret < 0)
3205                         break;
3206                 if (parent_inode == start_ino) {
3207                         ret = 1;
3208                         if (*ancestor_ino == 0)
3209                                 *ancestor_ino = ino;
3210                         break;
3211                 }
3212                 ino = parent_inode;
3213                 gen = parent_gen;
3214         }
3215         return ret;
3216 }
3217 
3218 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3219 {
3220         struct fs_path *from_path = NULL;
3221         struct fs_path *to_path = NULL;
3222         struct fs_path *name = NULL;
3223         u64 orig_progress = sctx->send_progress;
3224         struct recorded_ref *cur;
3225         u64 parent_ino, parent_gen;
3226         struct waiting_dir_move *dm = NULL;
3227         u64 rmdir_ino = 0;
3228         u64 ancestor;
3229         bool is_orphan;
3230         int ret;
3231 
3232         name = fs_path_alloc();
3233         from_path = fs_path_alloc();
3234         if (!name || !from_path) {
3235                 ret = -ENOMEM;
3236                 goto out;
3237         }
3238 
3239         dm = get_waiting_dir_move(sctx, pm->ino);
3240         ASSERT(dm);
3241         rmdir_ino = dm->rmdir_ino;
3242         is_orphan = dm->orphanized;
3243         free_waiting_dir_move(sctx, dm);
3244 
3245         if (is_orphan) {
3246                 ret = gen_unique_name(sctx, pm->ino,
3247                                       pm->gen, from_path);
3248         } else {
3249                 ret = get_first_ref(sctx->parent_root, pm->ino,
3250                                     &parent_ino, &parent_gen, name);
3251                 if (ret < 0)
3252                         goto out;
3253                 ret = get_cur_path(sctx, parent_ino, parent_gen,
3254                                    from_path);
3255                 if (ret < 0)
3256                         goto out;
3257                 ret = fs_path_add_path(from_path, name);
3258         }
3259         if (ret < 0)
3260                 goto out;
3261 
3262         sctx->send_progress = sctx->cur_ino + 1;
3263         ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3264         if (ret < 0)
3265                 goto out;
3266         if (ret) {
3267                 LIST_HEAD(deleted_refs);
3268                 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3269                 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3270                                            &pm->update_refs, &deleted_refs,
3271                                            is_orphan);
3272                 if (ret < 0)
3273                         goto out;
3274                 if (rmdir_ino) {
3275                         dm = get_waiting_dir_move(sctx, pm->ino);
3276                         ASSERT(dm);
3277                         dm->rmdir_ino = rmdir_ino;
3278                 }
3279                 goto out;
3280         }
3281         fs_path_reset(name);
3282         to_path = name;
3283         name = NULL;
3284         ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3285         if (ret < 0)
3286                 goto out;
3287 
3288         ret = send_rename(sctx, from_path, to_path);
3289         if (ret < 0)
3290                 goto out;
3291 
3292         if (rmdir_ino) {
3293                 struct orphan_dir_info *odi;
3294 
3295                 odi = get_orphan_dir_info(sctx, rmdir_ino);
3296                 if (!odi) {
3297                         /* already deleted */
3298                         goto finish;
3299                 }
3300                 ret = can_rmdir(sctx, rmdir_ino, odi->gen, sctx->cur_ino);
3301                 if (ret < 0)
3302                         goto out;
3303                 if (!ret)
3304                         goto finish;
3305 
3306                 name = fs_path_alloc();
3307                 if (!name) {
3308                         ret = -ENOMEM;
3309                         goto out;
3310                 }
3311                 ret = get_cur_path(sctx, rmdir_ino, odi->gen, name);
3312                 if (ret < 0)
3313                         goto out;
3314                 ret = send_rmdir(sctx, name);
3315                 if (ret < 0)
3316                         goto out;
3317                 free_orphan_dir_info(sctx, odi);
3318         }
3319 
3320 finish:
3321         ret = send_utimes(sctx, pm->ino, pm->gen);
3322         if (ret < 0)
3323                 goto out;
3324 
3325         /*
3326          * After rename/move, need to update the utimes of both new parent(s)
3327          * and old parent(s).
3328          */
3329         list_for_each_entry(cur, &pm->update_refs, list) {
3330                 /*
3331                  * The parent inode might have been deleted in the send snapshot
3332                  */
3333                 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3334                                      NULL, NULL, NULL, NULL, NULL);
3335                 if (ret == -ENOENT) {
3336                         ret = 0;
3337                         continue;
3338                 }
3339                 if (ret < 0)
3340                         goto out;
3341 
3342                 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3343                 if (ret < 0)
3344                         goto out;
3345         }
3346 
3347 out:
3348         fs_path_free(name);
3349         fs_path_free(from_path);
3350         fs_path_free(to_path);
3351         sctx->send_progress = orig_progress;
3352 
3353         return ret;
3354 }
3355 
3356 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3357 {
3358         if (!list_empty(&m->list))
3359                 list_del(&m->list);
3360         if (!RB_EMPTY_NODE(&m->node))
3361                 rb_erase(&m->node, &sctx->pending_dir_moves);
3362         __free_recorded_refs(&m->update_refs);
3363         kfree(m);
3364 }
3365 
3366 static void tail_append_pending_moves(struct pending_dir_move *moves,
3367                                       struct list_head *stack)
3368 {
3369         if (list_empty(&moves->list)) {
3370                 list_add_tail(&moves->list, stack);
3371         } else {
3372                 LIST_HEAD(list);
3373                 list_splice_init(&moves->list, &list);
3374                 list_add_tail(&moves->list, stack);
3375                 list_splice_tail(&list, stack);
3376         }
3377 }
3378 
3379 static int apply_children_dir_moves(struct send_ctx *sctx)
3380 {
3381         struct pending_dir_move *pm;
3382         struct list_head stack;
3383         u64 parent_ino = sctx->cur_ino;
3384         int ret = 0;
3385 
3386         pm = get_pending_dir_moves(sctx, parent_ino);
3387         if (!pm)
3388                 return 0;
3389 
3390         INIT_LIST_HEAD(&stack);
3391         tail_append_pending_moves(pm, &stack);
3392 
3393         while (!list_empty(&stack)) {
3394                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3395                 parent_ino = pm->ino;
3396                 ret = apply_dir_move(sctx, pm);
3397                 free_pending_move(sctx, pm);
3398                 if (ret)
3399                         goto out;
3400                 pm = get_pending_dir_moves(sctx, parent_ino);
3401                 if (pm)
3402                         tail_append_pending_moves(pm, &stack);
3403         }
3404         return 0;
3405 
3406 out:
3407         while (!list_empty(&stack)) {
3408                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3409                 free_pending_move(sctx, pm);
3410         }
3411         return ret;
3412 }
3413 
3414 /*
3415  * We might need to delay a directory rename even when no ancestor directory
3416  * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3417  * renamed. This happens when we rename a directory to the old name (the name
3418  * in the parent root) of some other unrelated directory that got its rename
3419  * delayed due to some ancestor with higher number that got renamed.
3420  *
3421  * Example:
3422  *
3423  * Parent snapshot:
3424  * .                                       (ino 256)
3425  * |---- a/                                (ino 257)
3426  * |     |---- file                        (ino 260)
3427  * |
3428  * |---- b/                                (ino 258)
3429  * |---- c/                                (ino 259)
3430  *
3431  * Send snapshot:
3432  * .                                       (ino 256)
3433  * |---- a/                                (ino 258)
3434  * |---- x/                                (ino 259)
3435  *       |---- y/                          (ino 257)
3436  *             |----- file                 (ino 260)
3437  *
3438  * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3439  * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3440  * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3441  * must issue is:
3442  *
3443  * 1 - rename 259 from 'c' to 'x'
3444  * 2 - rename 257 from 'a' to 'x/y'
3445  * 3 - rename 258 from 'b' to 'a'
3446  *
3447  * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3448  * be done right away and < 0 on error.
3449  */
3450 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3451                                   struct recorded_ref *parent_ref,
3452                                   const bool is_orphan)
3453 {
3454         struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3455         struct btrfs_path *path;
3456         struct btrfs_key key;
3457         struct btrfs_key di_key;
3458         struct btrfs_dir_item *di;
3459         u64 left_gen;
3460         u64 right_gen;
3461         int ret = 0;
3462         struct waiting_dir_move *wdm;
3463 
3464         if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3465                 return 0;
3466 
3467         path = alloc_path_for_send();
3468         if (!path)
3469                 return -ENOMEM;
3470 
3471         key.objectid = parent_ref->dir;
3472         key.type = BTRFS_DIR_ITEM_KEY;
3473         key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3474 
3475         ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3476         if (ret < 0) {
3477                 goto out;
3478         } else if (ret > 0) {
3479                 ret = 0;
3480                 goto out;
3481         }
3482 
3483         di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3484                                        parent_ref->name_len);
3485         if (!di) {
3486                 ret = 0;
3487                 goto out;
3488         }
3489         /*
3490          * di_key.objectid has the number of the inode that has a dentry in the
3491          * parent directory with the same name that sctx->cur_ino is being
3492          * renamed to. We need to check if that inode is in the send root as
3493          * well and if it is currently marked as an inode with a pending rename,
3494          * if it is, we need to delay the rename of sctx->cur_ino as well, so
3495          * that it happens after that other inode is renamed.
3496          */
3497         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3498         if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3499                 ret = 0;
3500                 goto out;
3501         }
3502 
3503         ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3504                              &left_gen, NULL, NULL, NULL, NULL);
3505         if (ret < 0)
3506                 goto out;
3507         ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3508                              &right_gen, NULL, NULL, NULL, NULL);
3509         if (ret < 0) {
3510                 if (ret == -ENOENT)
3511                         ret = 0;
3512                 goto out;
3513         }
3514 
3515         /* Different inode, no need to delay the rename of sctx->cur_ino */
3516         if (right_gen != left_gen) {
3517                 ret = 0;
3518                 goto out;
3519         }
3520 
3521         wdm = get_waiting_dir_move(sctx, di_key.objectid);
3522         if (wdm && !wdm->orphanized) {
3523                 ret = add_pending_dir_move(sctx,
3524                                            sctx->cur_ino,
3525                                            sctx->cur_inode_gen,
3526                                            di_key.objectid,
3527                                            &sctx->new_refs,
3528                                            &sctx->deleted_refs,
3529                                            is_orphan);
3530                 if (!ret)
3531                         ret = 1;
3532         }
3533 out:
3534         btrfs_free_path(path);
3535         return ret;
3536 }
3537 
3538 /*
3539  * Check if ino ino1 is an ancestor of inode ino2 in the given root.
3540  * Return 1 if true, 0 if false and < 0 on error.
3541  */
3542 static int is_ancestor(struct btrfs_root *root,
3543                        const u64 ino1,
3544                        const u64 ino1_gen,
3545                        const u64 ino2,
3546                        struct fs_path *fs_path)
3547 {
3548         u64 ino = ino2;
3549 
3550         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3551                 int ret;
3552                 u64 parent;
3553                 u64 parent_gen;
3554 
3555                 fs_path_reset(fs_path);
3556                 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3557                 if (ret < 0) {
3558                         if (ret == -ENOENT && ino == ino2)
3559                                 ret = 0;
3560                         return ret;
3561                 }
3562                 if (parent == ino1)
3563                         return parent_gen == ino1_gen ? 1 : 0;
3564                 ino = parent;
3565         }
3566         return 0;
3567 }
3568 
3569 static int wait_for_parent_move(struct send_ctx *sctx,
3570                                 struct recorded_ref *parent_ref,
3571                                 const bool is_orphan)
3572 {
3573         int ret = 0;
3574         u64 ino = parent_ref->dir;
3575         u64 ino_gen = parent_ref->dir_gen;
3576         u64 parent_ino_before, parent_ino_after;
3577         struct fs_path *path_before = NULL;
3578         struct fs_path *path_after = NULL;
3579         int len1, len2;
3580 
3581         path_after = fs_path_alloc();
3582         path_before = fs_path_alloc();
3583         if (!path_after || !path_before) {
3584                 ret = -ENOMEM;
3585                 goto out;
3586         }
3587 
3588         /*
3589          * Our current directory inode may not yet be renamed/moved because some
3590          * ancestor (immediate or not) has to be renamed/moved first. So find if
3591          * such ancestor exists and make sure our own rename/move happens after
3592          * that ancestor is processed to avoid path build infinite loops (done
3593          * at get_cur_path()).
3594          */
3595         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3596                 u64 parent_ino_after_gen;
3597 
3598                 if (is_waiting_for_move(sctx, ino)) {
3599                         /*
3600                          * If the current inode is an ancestor of ino in the
3601                          * parent root, we need to delay the rename of the
3602                          * current inode, otherwise don't delayed the rename
3603                          * because we can end up with a circular dependency
3604                          * of renames, resulting in some directories never
3605                          * getting the respective rename operations issued in
3606                          * the send stream or getting into infinite path build
3607                          * loops.
3608                          */
3609                         ret = is_ancestor(sctx->parent_root,
3610                                           sctx->cur_ino, sctx->cur_inode_gen,
3611                                           ino, path_before);
3612                         if (ret)
3613                                 break;
3614                 }
3615 
3616                 fs_path_reset(path_before);
3617                 fs_path_reset(path_after);
3618 
3619                 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3620                                     &parent_ino_after_gen, path_after);
3621                 if (ret < 0)
3622                         goto out;
3623                 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3624                                     NULL, path_before);
3625                 if (ret < 0 && ret != -ENOENT) {
3626                         goto out;
3627                 } else if (ret == -ENOENT) {
3628                         ret = 0;
3629                         break;
3630                 }
3631 
3632                 len1 = fs_path_len(path_before);
3633                 len2 = fs_path_len(path_after);
3634                 if (ino > sctx->cur_ino &&
3635                     (parent_ino_before != parent_ino_after || len1 != len2 ||
3636                      memcmp(path_before->start, path_after->start, len1))) {
3637                         u64 parent_ino_gen;
3638 
3639                         ret = get_inode_info(sctx->parent_root, ino, NULL,
3640                                              &parent_ino_gen, NULL, NULL, NULL,
3641                                              NULL);
3642                         if (ret < 0)
3643                                 goto out;
3644                         if (ino_gen == parent_ino_gen) {
3645                                 ret = 1;
3646                                 break;
3647                         }
3648                 }
3649                 ino = parent_ino_after;
3650                 ino_gen = parent_ino_after_gen;
3651         }
3652 
3653 out:
3654         fs_path_free(path_before);
3655         fs_path_free(path_after);
3656 
3657         if (ret == 1) {
3658                 ret = add_pending_dir_move(sctx,
3659                                            sctx->cur_ino,
3660                                            sctx->cur_inode_gen,
3661                                            ino,
3662                                            &sctx->new_refs,
3663                                            &sctx->deleted_refs,
3664                                            is_orphan);
3665                 if (!ret)
3666                         ret = 1;
3667         }
3668 
3669         return ret;
3670 }
3671 
3672 /*
3673  * This does all the move/link/unlink/rmdir magic.
3674  */
3675 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3676 {
3677         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3678         int ret = 0;
3679         struct recorded_ref *cur;
3680         struct recorded_ref *cur2;
3681         struct list_head check_dirs;
3682         struct fs_path *valid_path = NULL;
3683         u64 ow_inode = 0;
3684         u64 ow_gen;
3685         int did_overwrite = 0;
3686         int is_orphan = 0;
3687         u64 last_dir_ino_rm = 0;
3688         bool can_rename = true;
3689 
3690         btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3691 
3692         /*
3693          * This should never happen as the root dir always has the same ref
3694          * which is always '..'
3695          */
3696         BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3697         INIT_LIST_HEAD(&check_dirs);
3698 
3699         valid_path = fs_path_alloc();
3700         if (!valid_path) {
3701                 ret = -ENOMEM;
3702                 goto out;
3703         }
3704 
3705         /*
3706          * First, check if the first ref of the current inode was overwritten
3707          * before. If yes, we know that the current inode was already orphanized
3708          * and thus use the orphan name. If not, we can use get_cur_path to
3709          * get the path of the first ref as it would like while receiving at
3710          * this point in time.
3711          * New inodes are always orphan at the beginning, so force to use the
3712          * orphan name in this case.
3713          * The first ref is stored in valid_path and will be updated if it
3714          * gets moved around.
3715          */
3716         if (!sctx->cur_inode_new) {
3717                 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3718                                 sctx->cur_inode_gen);
3719                 if (ret < 0)
3720                         goto out;
3721                 if (ret)
3722                         did_overwrite = 1;
3723         }
3724         if (sctx->cur_inode_new || did_overwrite) {
3725                 ret = gen_unique_name(sctx, sctx->cur_ino,
3726                                 sctx->cur_inode_gen, valid_path);
3727                 if (ret < 0)
3728                         goto out;
3729                 is_orphan = 1;
3730         } else {
3731                 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3732                                 valid_path);
3733                 if (ret < 0)
3734                         goto out;
3735         }
3736 
3737         list_for_each_entry(cur, &sctx->new_refs, list) {
3738                 /*
3739                  * We may have refs where the parent directory does not exist
3740                  * yet. This happens if the parent directories inum is higher
3741                  * the the current inum. To handle this case, we create the
3742                  * parent directory out of order. But we need to check if this
3743                  * did already happen before due to other refs in the same dir.
3744                  */
3745                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3746                 if (ret < 0)
3747                         goto out;
3748                 if (ret == inode_state_will_create) {
3749                         ret = 0;
3750                         /*
3751                          * First check if any of the current inodes refs did
3752                          * already create the dir.
3753                          */
3754                         list_for_each_entry(cur2, &sctx->new_refs, list) {
3755                                 if (cur == cur2)
3756                                         break;
3757                                 if (cur2->dir == cur->dir) {
3758                                         ret = 1;
3759                                         break;
3760                                 }
3761                         }
3762 
3763                         /*
3764                          * If that did not happen, check if a previous inode
3765                          * did already create the dir.
3766                          */
3767                         if (!ret)
3768                                 ret = did_create_dir(sctx, cur->dir);
3769                         if (ret < 0)
3770                                 goto out;
3771                         if (!ret) {
3772                                 ret = send_create_inode(sctx, cur->dir);
3773                                 if (ret < 0)
3774                                         goto out;
3775                         }
3776                 }
3777 
3778                 /*
3779                  * Check if this new ref would overwrite the first ref of
3780                  * another unprocessed inode. If yes, orphanize the
3781                  * overwritten inode. If we find an overwritten ref that is
3782                  * not the first ref, simply unlink it.
3783                  */
3784                 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3785                                 cur->name, cur->name_len,
3786                                 &ow_inode, &ow_gen);
3787                 if (ret < 0)
3788                         goto out;
3789                 if (ret) {
3790                         ret = is_first_ref(sctx->parent_root,
3791                                            ow_inode, cur->dir, cur->name,
3792                                            cur->name_len);
3793                         if (ret < 0)
3794                                 goto out;
3795                         if (ret) {
3796                                 struct name_cache_entry *nce;
3797                                 struct waiting_dir_move *wdm;
3798 
3799                                 ret = orphanize_inode(sctx, ow_inode, ow_gen,
3800                                                 cur->full_path);
3801                                 if (ret < 0)
3802                                         goto out;
3803 
3804                                 /*
3805                                  * If ow_inode has its rename operation delayed
3806                                  * make sure that its orphanized name is used in
3807                                  * the source path when performing its rename
3808                                  * operation.
3809                                  */
3810                                 if (is_waiting_for_move(sctx, ow_inode)) {
3811                                         wdm = get_waiting_dir_move(sctx,
3812                                                                    ow_inode);
3813                                         ASSERT(wdm);
3814                                         wdm->orphanized = true;
3815                                 }
3816 
3817                                 /*
3818                                  * Make sure we clear our orphanized inode's
3819                                  * name from the name cache. This is because the
3820                                  * inode ow_inode might be an ancestor of some
3821                                  * other inode that will be orphanized as well
3822                                  * later and has an inode number greater than
3823                                  * sctx->send_progress. We need to prevent
3824                                  * future name lookups from using the old name
3825                                  * and get instead the orphan name.
3826                                  */
3827                                 nce = name_cache_search(sctx, ow_inode, ow_gen);
3828                                 if (nce) {
3829                                         name_cache_delete(sctx, nce);
3830                                         kfree(nce);
3831                                 }
3832 
3833                                 /*
3834                                  * ow_inode might currently be an ancestor of
3835                                  * cur_ino, therefore compute valid_path (the
3836                                  * current path of cur_ino) again because it
3837                                  * might contain the pre-orphanization name of
3838                                  * ow_inode, which is no longer valid.
3839                                  */
3840                                 fs_path_reset(valid_path);
3841                                 ret = get_cur_path(sctx, sctx->cur_ino,
3842                                            sctx->cur_inode_gen, valid_path);
3843                                 if (ret < 0)
3844                                         goto out;
3845                         } else {
3846                                 ret = send_unlink(sctx, cur->full_path);
3847                                 if (ret < 0)
3848                                         goto out;
3849                         }
3850                 }
3851 
3852                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
3853                         ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
3854                         if (ret < 0)
3855                                 goto out;
3856                         if (ret == 1) {
3857                                 can_rename = false;
3858                                 *pending_move = 1;
3859                         }
3860                 }
3861 
3862                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
3863                     can_rename) {
3864                         ret = wait_for_parent_move(sctx, cur, is_orphan);
3865                         if (ret < 0)
3866                                 goto out;
3867                         if (ret == 1) {
3868                                 can_rename = false;
3869                                 *pending_move = 1;
3870                         }
3871                 }
3872 
3873                 /*
3874                  * link/move the ref to the new place. If we have an orphan
3875                  * inode, move it and update valid_path. If not, link or move
3876                  * it depending on the inode mode.
3877                  */
3878                 if (is_orphan && can_rename) {
3879                         ret = send_rename(sctx, valid_path, cur->full_path);
3880                         if (ret < 0)
3881                                 goto out;
3882                         is_orphan = 0;
3883                         ret = fs_path_copy(valid_path, cur->full_path);
3884                         if (ret < 0)
3885                                 goto out;
3886                 } else if (can_rename) {
3887                         if (S_ISDIR(sctx->cur_inode_mode)) {
3888                                 /*
3889                                  * Dirs can't be linked, so move it. For moved
3890                                  * dirs, we always have one new and one deleted
3891                                  * ref. The deleted ref is ignored later.
3892                                  */
3893                                 ret = send_rename(sctx, valid_path,
3894                                                   cur->full_path);
3895                                 if (!ret)
3896                                         ret = fs_path_copy(valid_path,
3897                                                            cur->full_path);
3898                                 if (ret < 0)
3899                                         goto out;
3900                         } else {
3901                                 ret = send_link(sctx, cur->full_path,
3902                                                 valid_path);
3903                                 if (ret < 0)
3904                                         goto out;
3905                         }
3906                 }
3907                 ret = dup_ref(cur, &check_dirs);
3908                 if (ret < 0)
3909                         goto out;
3910         }
3911 
3912         if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
3913                 /*
3914                  * Check if we can already rmdir the directory. If not,
3915                  * orphanize it. For every dir item inside that gets deleted
3916                  * later, we do this check again and rmdir it then if possible.
3917                  * See the use of check_dirs for more details.
3918                  */
3919                 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3920                                 sctx->cur_ino);
3921                 if (ret < 0)
3922                         goto out;
3923                 if (ret) {
3924                         ret = send_rmdir(sctx, valid_path);
3925                         if (ret < 0)
3926                                 goto out;
3927                 } else if (!is_orphan) {
3928                         ret = orphanize_inode(sctx, sctx->cur_ino,
3929                                         sctx->cur_inode_gen, valid_path);
3930                         if (ret < 0)
3931                                 goto out;
3932                         is_orphan = 1;
3933                 }
3934 
3935                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
3936                         ret = dup_ref(cur, &check_dirs);
3937                         if (ret < 0)
3938                                 goto out;
3939                 }
3940         } else if (S_ISDIR(sctx->cur_inode_mode) &&
3941                    !list_empty(&sctx->deleted_refs)) {
3942                 /*
3943                  * We have a moved dir. Add the old parent to check_dirs
3944                  */
3945                 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
3946                                 list);
3947                 ret = dup_ref(cur, &check_dirs);
3948                 if (ret < 0)
3949                         goto out;
3950         } else if (!S_ISDIR(sctx->cur_inode_mode)) {
3951                 /*
3952                  * We have a non dir inode. Go through all deleted refs and
3953                  * unlink them if they were not already overwritten by other
3954                  * inodes.
3955                  */
3956                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
3957                         ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3958                                         sctx->cur_ino, sctx->cur_inode_gen,
3959                                         cur->name, cur->name_len);
3960                         if (ret < 0)
3961                                 goto out;
3962                         if (!ret) {
3963                                 ret = send_unlink(sctx, cur->full_path);
3964                                 if (ret < 0)
3965                                         goto out;
3966                         }
3967                         ret = dup_ref(cur, &check_dirs);
3968                         if (ret < 0)
3969                                 goto out;
3970                 }
3971                 /*
3972                  * If the inode is still orphan, unlink the orphan. This may
3973                  * happen when a previous inode did overwrite the first ref
3974                  * of this inode and no new refs were added for the current
3975                  * inode. Unlinking does not mean that the inode is deleted in
3976                  * all cases. There may still be links to this inode in other
3977                  * places.
3978                  */
3979                 if (is_orphan) {
3980                         ret = send_unlink(sctx, valid_path);
3981                         if (ret < 0)
3982                                 goto out;
3983                 }
3984         }
3985 
3986         /*
3987          * We did collect all parent dirs where cur_inode was once located. We
3988          * now go through all these dirs and check if they are pending for
3989          * deletion and if it's finally possible to perform the rmdir now.
3990          * We also update the inode stats of the parent dirs here.
3991          */
3992         list_for_each_entry(cur, &check_dirs, list) {
3993                 /*
3994                  * In case we had refs into dirs that were not processed yet,
3995                  * we don't need to do the utime and rmdir logic for these dirs.
3996                  * The dir will be processed later.
3997                  */
3998                 if (cur->dir > sctx->cur_ino)
3999                         continue;
4000 
4001                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4002                 if (ret < 0)
4003                         goto out;
4004 
4005                 if (ret == inode_state_did_create ||
4006                     ret == inode_state_no_change) {
4007                         /* TODO delayed utimes */
4008                         ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4009                         if (ret < 0)
4010                                 goto out;
4011                 } else if (ret == inode_state_did_delete &&
4012                            cur->dir != last_dir_ino_rm) {
4013                         ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4014                                         sctx->cur_ino);
4015                         if (ret < 0)
4016                                 goto out;
4017                         if (ret) {
4018                                 ret = get_cur_path(sctx, cur->dir,
4019                                                    cur->dir_gen, valid_path);
4020                                 if (ret < 0)
4021                                         goto out;
4022                                 ret = send_rmdir(sctx, valid_path);
4023                                 if (ret < 0)
4024                                         goto out;
4025                                 last_dir_ino_rm = cur->dir;
4026                         }
4027                 }
4028         }
4029 
4030         ret = 0;
4031 
4032 out:
4033         __free_recorded_refs(&check_dirs);
4034         free_recorded_refs(sctx);
4035         fs_path_free(valid_path);
4036         return ret;
4037 }
4038 
4039 static int record_ref(struct btrfs_root *root, int num, u64 dir, int index,
4040                       struct fs_path *name, void *ctx, struct list_head *refs)
4041 {
4042         int ret = 0;
4043         struct send_ctx *sctx = ctx;
4044         struct fs_path *p;
4045         u64 gen;
4046 
4047         p = fs_path_alloc();
4048         if (!p)
4049                 return -ENOMEM;
4050 
4051         ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4052                         NULL, NULL);
4053         if (ret < 0)
4054                 goto out;
4055 
4056         ret = get_cur_path(sctx, dir, gen, p);
4057         if (ret < 0)
4058                 goto out;
4059         ret = fs_path_add_path(p, name);
4060         if (ret < 0)
4061                 goto out;
4062 
4063         ret = __record_ref(refs, dir, gen, p);
4064 
4065 out:
4066         if (ret)
4067                 fs_path_free(p);
4068         return ret;
4069 }
4070 
4071 static int __record_new_ref(int num, u64 dir, int index,
4072                             struct fs_path *name,
4073                             void *ctx)
4074 {
4075         struct send_ctx *sctx = ctx;
4076         return record_ref(sctx->send_root, num, dir, index, name,
4077                           ctx, &sctx->new_refs);
4078 }
4079 
4080 
4081 static int __record_deleted_ref(int num, u64 dir, int index,
4082                                 struct fs_path *name,
4083                                 void *ctx)
4084 {
4085         struct send_ctx *sctx = ctx;
4086         return record_ref(sctx->parent_root, num, dir, index, name,
4087                           ctx, &sctx->deleted_refs);
4088 }
4089 
4090 static int record_new_ref(struct send_ctx *sctx)
4091 {
4092         int ret;
4093 
4094         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4095                                 sctx->cmp_key, 0, __record_new_ref, sctx);
4096         if (ret < 0)
4097                 goto out;
4098         ret = 0;
4099 
4100 out:
4101         return ret;
4102 }
4103 
4104 static int record_deleted_ref(struct send_ctx *sctx)
4105 {
4106         int ret;
4107 
4108         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4109                                 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4110         if (ret < 0)
4111                 goto out;
4112         ret = 0;
4113 
4114 out:
4115         return ret;
4116 }
4117 
4118 struct find_ref_ctx {
4119         u64 dir;
4120         u64 dir_gen;
4121         struct btrfs_root *root;
4122         struct fs_path *name;
4123         int found_idx;
4124 };
4125 
4126 static int __find_iref(int num, u64 dir, int index,
4127                        struct fs_path *name,
4128                        void *ctx_)
4129 {
4130         struct find_ref_ctx *ctx = ctx_;
4131         u64 dir_gen;
4132         int ret;
4133 
4134         if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4135             strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4136                 /*
4137                  * To avoid doing extra lookups we'll only do this if everything
4138                  * else matches.
4139                  */
4140                 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4141                                      NULL, NULL, NULL);
4142                 if (ret)
4143                         return ret;
4144                 if (dir_gen != ctx->dir_gen)
4145                         return 0;
4146                 ctx->found_idx = num;
4147                 return 1;
4148         }
4149         return 0;
4150 }
4151 
4152 static int find_iref(struct btrfs_root *root,
4153                      struct btrfs_path *path,
4154                      struct btrfs_key *key,
4155                      u64 dir, u64 dir_gen, struct fs_path *name)
4156 {
4157         int ret;
4158         struct find_ref_ctx ctx;
4159 
4160         ctx.dir = dir;
4161         ctx.name = name;
4162         ctx.dir_gen = dir_gen;
4163         ctx.found_idx = -1;
4164         ctx.root = root;
4165 
4166         ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4167         if (ret < 0)
4168                 return ret;
4169 
4170         if (ctx.found_idx == -1)
4171                 return -ENOENT;
4172 
4173         return ctx.found_idx;
4174 }
4175 
4176 static int __record_changed_new_ref(int num, u64 dir, int index,
4177                                     struct fs_path *name,
4178                                     void *ctx)
4179 {
4180         u64 dir_gen;
4181         int ret;
4182         struct send_ctx *sctx = ctx;
4183 
4184         ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4185                              NULL, NULL, NULL);
4186         if (ret)
4187                 return ret;
4188 
4189         ret = find_iref(sctx->parent_root, sctx->right_path,
4190                         sctx->cmp_key, dir, dir_gen, name);
4191         if (ret == -ENOENT)
4192                 ret = __record_new_ref(num, dir, index, name, sctx);
4193         else if (ret > 0)
4194                 ret = 0;
4195 
4196         return ret;
4197 }
4198 
4199 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4200                                         struct fs_path *name,
4201                                         void *ctx)
4202 {
4203         u64 dir_gen;
4204         int ret;
4205         struct send_ctx *sctx = ctx;
4206 
4207         ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4208                              NULL, NULL, NULL);
4209         if (ret)
4210                 return ret;
4211 
4212         ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4213                         dir, dir_gen, name);
4214         if (ret == -ENOENT)
4215                 ret = __record_deleted_ref(num, dir, index, name, sctx);
4216         else if (ret > 0)
4217                 ret = 0;
4218 
4219         return ret;
4220 }
4221 
4222 static int record_changed_ref(struct send_ctx *sctx)
4223 {
4224         int ret = 0;
4225 
4226         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4227                         sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4228         if (ret < 0)
4229                 goto out;
4230         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4231                         sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4232         if (ret < 0)
4233                 goto out;
4234         ret = 0;
4235 
4236 out:
4237         return ret;
4238 }
4239 
4240 /*
4241  * Record and process all refs at once. Needed when an inode changes the
4242  * generation number, which means that it was deleted and recreated.
4243  */
4244 static int process_all_refs(struct send_ctx *sctx,
4245                             enum btrfs_compare_tree_result cmd)
4246 {
4247         int ret;
4248         struct btrfs_root *root;
4249         struct btrfs_path *path;
4250         struct btrfs_key key;
4251         struct btrfs_key found_key;
4252         struct extent_buffer *eb;
4253         int slot;
4254         iterate_inode_ref_t cb;
4255         int pending_move = 0;
4256 
4257         path = alloc_path_for_send();
4258         if (!path)
4259                 return -ENOMEM;
4260 
4261         if (cmd == BTRFS_COMPARE_TREE_NEW) {
4262                 root = sctx->send_root;
4263                 cb = __record_new_ref;
4264         } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4265                 root = sctx->parent_root;
4266                 cb = __record_deleted_ref;
4267         } else {
4268                 btrfs_err(sctx->send_root->fs_info,
4269                                 "Wrong command %d in process_all_refs", cmd);
4270                 ret = -EINVAL;
4271                 goto out;
4272         }
4273 
4274         key.objectid = sctx->cmp_key->objectid;
4275         key.type = BTRFS_INODE_REF_KEY;
4276         key.offset = 0;
4277         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4278         if (ret < 0)
4279                 goto out;
4280 
4281         while (1) {
4282                 eb = path->nodes[0];
4283                 slot = path->slots[0];
4284                 if (slot >= btrfs_header_nritems(eb)) {
4285                         ret = btrfs_next_leaf(root, path);
4286                         if (ret < 0)
4287                                 goto out;
4288                         else if (ret > 0)
4289                                 break;
4290                         continue;
4291                 }
4292 
4293                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4294 
4295                 if (found_key.objectid != key.objectid ||
4296                     (found_key.type != BTRFS_INODE_REF_KEY &&
4297                      found_key.type != BTRFS_INODE_EXTREF_KEY))
4298                         break;
4299 
4300                 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4301                 if (ret < 0)
4302                         goto out;
4303 
4304                 path->slots[0]++;
4305         }
4306         btrfs_release_path(path);
4307 
4308         /*
4309          * We don't actually care about pending_move as we are simply
4310          * re-creating this inode and will be rename'ing it into place once we
4311          * rename the parent directory.
4312          */
4313         ret = process_recorded_refs(sctx, &pending_move);
4314 out:
4315         btrfs_free_path(path);
4316         return ret;
4317 }
4318 
4319 static int send_set_xattr(struct send_ctx *sctx,
4320                           struct fs_path *path,
4321                           const char *name, int name_len,
4322                           const char *data, int data_len)
4323 {
4324         int ret = 0;
4325 
4326         ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4327         if (ret < 0)
4328                 goto out;
4329 
4330         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4331         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4332         TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4333 
4334         ret = send_cmd(sctx);
4335 
4336 tlv_put_failure:
4337 out:
4338         return ret;
4339 }
4340 
4341 static int send_remove_xattr(struct send_ctx *sctx,
4342                           struct fs_path *path,
4343                           const char *name, int name_len)
4344 {
4345         int ret = 0;
4346 
4347         ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4348         if (ret < 0)
4349                 goto out;
4350 
4351         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4352         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4353 
4354         ret = send_cmd(sctx);
4355 
4356 tlv_put_failure:
4357 out:
4358         return ret;
4359 }
4360 
4361 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4362                                const char *name, int name_len,
4363                                const char *data, int data_len,
4364                                u8 type, void *ctx)
4365 {
4366         int ret;
4367         struct send_ctx *sctx = ctx;
4368         struct fs_path *p;
4369         struct posix_acl_xattr_header dummy_acl;
4370 
4371         p = fs_path_alloc();
4372         if (!p)
4373                 return -ENOMEM;
4374 
4375         /*
4376          * This hack is needed because empty acls are stored as zero byte
4377          * data in xattrs. Problem with that is, that receiving these zero byte
4378          * acls will fail later. To fix this, we send a dummy acl list that
4379          * only contains the version number and no entries.
4380          */
4381         if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4382             !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4383                 if (data_len == 0) {
4384                         dummy_acl.a_version =
4385                                         cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4386                         data = (char *)&dummy_acl;
4387                         data_len = sizeof(dummy_acl);
4388                 }
4389         }
4390 
4391         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4392         if (ret < 0)
4393                 goto out;
4394 
4395         ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4396 
4397 out:
4398         fs_path_free(p);
4399         return ret;
4400 }
4401 
4402 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4403                                    const char *name, int name_len,
4404                                    const char *data, int data_len,
4405                                    u8 type, void *ctx)
4406 {
4407         int ret;
4408         struct send_ctx *sctx = ctx;
4409         struct fs_path *p;
4410 
4411         p = fs_path_alloc();
4412         if (!p)
4413                 return -ENOMEM;
4414 
4415         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4416         if (ret < 0)
4417                 goto out;
4418 
4419         ret = send_remove_xattr(sctx, p, name, name_len);
4420 
4421 out:
4422         fs_path_free(p);
4423         return ret;
4424 }
4425 
4426 static int process_new_xattr(struct send_ctx *sctx)
4427 {
4428         int ret = 0;
4429 
4430         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4431                                sctx->cmp_key, __process_new_xattr, sctx);
4432 
4433         return ret;
4434 }
4435 
4436 static int process_deleted_xattr(struct send_ctx *sctx)
4437 {
4438         return iterate_dir_item(sctx->parent_root, sctx->right_path,
4439                                 sctx->cmp_key, __process_deleted_xattr, sctx);
4440 }
4441 
4442 struct find_xattr_ctx {
4443         const char *name;
4444         int name_len;
4445         int found_idx;
4446         char *found_data;
4447         int found_data_len;
4448 };
4449 
4450 static int __find_xattr(int num, struct btrfs_key *di_key,
4451                         const char *name, int name_len,
4452                         const char *data, int data_len,
4453                         u8 type, void *vctx)
4454 {
4455         struct find_xattr_ctx *ctx = vctx;
4456 
4457         if (name_len == ctx->name_len &&
4458             strncmp(name, ctx->name, name_len) == 0) {
4459                 ctx->found_idx = num;
4460                 ctx->found_data_len = data_len;
4461                 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4462                 if (!ctx->found_data)
4463                         return -ENOMEM;
4464                 return 1;
4465         }
4466         return 0;
4467 }
4468 
4469 static int find_xattr(struct btrfs_root *root,
4470                       struct btrfs_path *path,
4471                       struct btrfs_key *key,
4472                       const char *name, int name_len,
4473                       char **data, int *data_len)
4474 {
4475         int ret;
4476         struct find_xattr_ctx ctx;
4477 
4478         ctx.name = name;
4479         ctx.name_len = name_len;
4480         ctx.found_idx = -1;
4481         ctx.found_data = NULL;
4482         ctx.found_data_len = 0;
4483 
4484         ret = iterate_dir_item(root, path, key, __find_xattr, &ctx);
4485         if (ret < 0)
4486                 return ret;
4487 
4488         if (ctx.found_idx == -1)
4489                 return -ENOENT;
4490         if (data) {
4491                 *data = ctx.found_data;
4492                 *data_len = ctx.found_data_len;
4493         } else {
4494                 kfree(ctx.found_data);
4495         }
4496         return ctx.found_idx;
4497 }
4498 
4499 
4500 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4501                                        const char *name, int name_len,
4502                                        const char *data, int data_len,
4503                                        u8 type, void *ctx)
4504 {
4505         int ret;
4506         struct send_ctx *sctx = ctx;
4507         char *found_data = NULL;
4508         int found_data_len  = 0;
4509 
4510         ret = find_xattr(sctx->parent_root, sctx->right_path,
4511                          sctx->cmp_key, name, name_len, &found_data,
4512                          &found_data_len);
4513         if (ret == -ENOENT) {
4514                 ret = __process_new_xattr(num, di_key, name, name_len, data,
4515                                 data_len, type, ctx);
4516         } else if (ret >= 0) {
4517                 if (data_len != found_data_len ||
4518                     memcmp(data, found_data, data_len)) {
4519                         ret = __process_new_xattr(num, di_key, name, name_len,
4520                                         data, data_len, type, ctx);
4521                 } else {
4522                         ret = 0;
4523                 }
4524         }
4525 
4526         kfree(found_data);
4527         return ret;
4528 }
4529 
4530 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4531                                            const char *name, int name_len,
4532                                            const char *data, int data_len,
4533                                            u8 type, void *ctx)
4534 {
4535         int ret;
4536         struct send_ctx *sctx = ctx;
4537 
4538         ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4539                          name, name_len, NULL, NULL);
4540         if (ret == -ENOENT)
4541                 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4542                                 data_len, type, ctx);
4543         else if (ret >= 0)
4544                 ret = 0;
4545 
4546         return ret;
4547 }
4548 
4549 static int process_changed_xattr(struct send_ctx *sctx)
4550 {
4551         int ret = 0;
4552 
4553         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4554                         sctx->cmp_key, __process_changed_new_xattr, sctx);
4555         if (ret < 0)
4556                 goto out;
4557         ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4558                         sctx->cmp_key, __process_changed_deleted_xattr, sctx);
4559 
4560 out:
4561         return ret;
4562 }
4563 
4564 static int process_all_new_xattrs(struct send_ctx *sctx)
4565 {
4566         int ret;
4567         struct btrfs_root *root;
4568         struct btrfs_path *path;
4569         struct btrfs_key key;
4570         struct btrfs_key found_key;
4571         struct extent_buffer *eb;
4572         int slot;
4573 
4574         path = alloc_path_for_send();
4575         if (!path)
4576                 return -ENOMEM;
4577 
4578         root = sctx->send_root;
4579 
4580         key.objectid = sctx->cmp_key->objectid;
4581         key.type = BTRFS_XATTR_ITEM_KEY;
4582         key.offset = 0;
4583         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4584         if (ret < 0)
4585                 goto out;
4586 
4587         while (1) {
4588                 eb = path->nodes[0];
4589                 slot = path->slots[0];
4590                 if (slot >= btrfs_header_nritems(eb)) {
4591                         ret = btrfs_next_leaf(root, path);
4592                         if (ret < 0) {
4593                                 goto out;
4594                         } else if (ret > 0) {
4595                                 ret = 0;
4596                                 break;
4597                         }
4598                         continue;
4599                 }
4600 
4601                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4602                 if (found_key.objectid != key.objectid ||
4603                     found_key.type != key.type) {
4604                         ret = 0;
4605                         goto out;
4606                 }
4607 
4608                 ret = iterate_dir_item(root, path, &found_key,
4609                                        __process_new_xattr, sctx);
4610                 if (ret < 0)
4611                         goto out;
4612 
4613                 path->slots[0]++;
4614         }
4615 
4616 out:
4617         btrfs_free_path(path);
4618         return ret;
4619 }
4620 
4621 static ssize_t fill_read_buf(struct send_ctx *sctx, u64 offset, u32 len)
4622 {
4623         struct btrfs_root *root = sctx->send_root;
4624         struct btrfs_fs_info *fs_info = root->fs_info;
4625         struct inode *inode;
4626         struct page *page;
4627         char *addr;
4628         struct btrfs_key key;
4629         pgoff_t index = offset >> PAGE_SHIFT;
4630         pgoff_t last_index;
4631         unsigned pg_offset = offset & ~PAGE_MASK;
4632         ssize_t ret = 0;
4633 
4634         key.objectid = sctx->cur_ino;
4635         key.type = BTRFS_INODE_ITEM_KEY;
4636         key.offset = 0;
4637 
4638         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4639         if (IS_ERR(inode))
4640                 return PTR_ERR(inode);
4641 
4642         if (offset + len > i_size_read(inode)) {
4643                 if (offset > i_size_read(inode))
4644                         len = 0;
4645                 else
4646                         len = offset - i_size_read(inode);
4647         }
4648         if (len == 0)
4649                 goto out;
4650 
4651         last_index = (offset + len - 1) >> PAGE_SHIFT;
4652 
4653         /* initial readahead */
4654         memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4655         file_ra_state_init(&sctx->ra, inode->i_mapping);
4656         btrfs_force_ra(inode->i_mapping, &sctx->ra, NULL, index,
4657                        last_index - index + 1);
4658 
4659         while (index <= last_index) {
4660                 unsigned cur_len = min_t(unsigned, len,
4661                                          PAGE_SIZE - pg_offset);
4662                 page = find_or_create_page(inode->i_mapping, index, GFP_KERNEL);
4663                 if (!page) {
4664                         ret = -ENOMEM;
4665                         break;
4666                 }
4667 
4668                 if (!PageUptodate(page)) {
4669                         btrfs_readpage(NULL, page);
4670                         lock_page(page);
4671                         if (!PageUptodate(page)) {
4672                                 unlock_page(page);
4673                                 put_page(page);
4674                                 ret = -EIO;
4675                                 break;
4676                         }
4677                 }
4678 
4679                 addr = kmap(page);
4680                 memcpy(sctx->read_buf + ret, addr + pg_offset, cur_len);
4681                 kunmap(page);
4682                 unlock_page(page);
4683                 put_page(page);
4684                 index++;
4685                 pg_offset = 0;
4686                 len -= cur_len;
4687                 ret += cur_len;
4688         }
4689 out:
4690         iput(inode);
4691         return ret;
4692 }
4693 
4694 /*
4695  * Read some bytes from the current inode/file and send a write command to
4696  * user space.
4697  */
4698 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
4699 {
4700         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4701         int ret = 0;
4702         struct fs_path *p;
4703         ssize_t num_read = 0;
4704 
4705         p = fs_path_alloc();
4706         if (!p)
4707                 return -ENOMEM;
4708 
4709         btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
4710 
4711         num_read = fill_read_buf(sctx, offset, len);
4712         if (num_read <= 0) {
4713                 if (num_read < 0)
4714                         ret = num_read;
4715                 goto out;
4716         }
4717 
4718         ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4719         if (ret < 0)
4720                 goto out;
4721 
4722         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4723         if (ret < 0)
4724                 goto out;
4725 
4726         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4727         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4728         TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, num_read);
4729 
4730         ret = send_cmd(sctx);
4731 
4732 tlv_put_failure:
4733 out:
4734         fs_path_free(p);
4735         if (ret < 0)
4736                 return ret;
4737         return num_read;
4738 }
4739 
4740 /*
4741  * Send a clone command to user space.
4742  */
4743 static int send_clone(struct send_ctx *sctx,
4744                       u64 offset, u32 len,
4745                       struct clone_root *clone_root)
4746 {
4747         int ret = 0;
4748         struct fs_path *p;
4749         u64 gen;
4750 
4751         btrfs_debug(sctx->send_root->fs_info,
4752                     "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
4753                     offset, len, clone_root->root->objectid, clone_root->ino,
4754                     clone_root->offset);
4755 
4756         p = fs_path_alloc();
4757         if (!p)
4758                 return -ENOMEM;
4759 
4760         ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
4761         if (ret < 0)
4762                 goto out;
4763 
4764         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4765         if (ret < 0)
4766                 goto out;
4767 
4768         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4769         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
4770         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4771 
4772         if (clone_root->root == sctx->send_root) {
4773                 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
4774                                 &gen, NULL, NULL, NULL, NULL);
4775                 if (ret < 0)
4776                         goto out;
4777                 ret = get_cur_path(sctx, clone_root->ino, gen, p);
4778         } else {
4779                 ret = get_inode_path(clone_root->root, clone_root->ino, p);
4780         }
4781         if (ret < 0)
4782                 goto out;
4783 
4784         /*
4785          * If the parent we're using has a received_uuid set then use that as
4786          * our clone source as that is what we will look for when doing a
4787          * receive.
4788          *
4789          * This covers the case that we create a snapshot off of a received
4790          * subvolume and then use that as the parent and try to receive on a
4791          * different host.
4792          */
4793         if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
4794                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4795                              clone_root->root->root_item.received_uuid);
4796         else
4797                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4798                              clone_root->root->root_item.uuid);
4799         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
4800                     le64_to_cpu(clone_root->root->root_item.ctransid));
4801         TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
4802         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
4803                         clone_root->offset);
4804 
4805         ret = send_cmd(sctx);
4806 
4807 tlv_put_failure:
4808 out:
4809         fs_path_free(p);
4810         return ret;
4811 }
4812 
4813 /*
4814  * Send an update extent command to user space.
4815  */
4816 static int send_update_extent(struct send_ctx *sctx,
4817                               u64 offset, u32 len)
4818 {
4819         int ret = 0;
4820         struct fs_path *p;
4821 
4822         p = fs_path_alloc();
4823         if (!p)
4824                 return -ENOMEM;
4825 
4826         ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
4827         if (ret < 0)
4828                 goto out;
4829 
4830         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4831         if (ret < 0)
4832                 goto out;
4833 
4834         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4835         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4836         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
4837 
4838         ret = send_cmd(sctx);
4839 
4840 tlv_put_failure:
4841 out:
4842         fs_path_free(p);
4843         return ret;
4844 }
4845 
4846 static int send_hole(struct send_ctx *sctx, u64 end)
4847 {
4848         struct fs_path *p = NULL;
4849         u64 offset = sctx->cur_inode_last_extent;
4850         u64 len;
4851         int ret = 0;
4852 
4853         p = fs_path_alloc();
4854         if (!p)
4855                 return -ENOMEM;
4856         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4857         if (ret < 0)
4858                 goto tlv_put_failure;
4859         memset(sctx->read_buf, 0, BTRFS_SEND_READ_SIZE);
4860         while (offset < end) {
4861                 len = min_t(u64, end - offset, BTRFS_SEND_READ_SIZE);
4862 
4863                 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4864                 if (ret < 0)
4865                         break;
4866                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4867                 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4868                 TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, len);
4869                 ret = send_cmd(sctx);
4870                 if (ret < 0)
4871                         break;
4872                 offset += len;
4873         }
4874 tlv_put_failure:
4875         fs_path_free(p);
4876         return ret;
4877 }
4878 
4879 static int send_extent_data(struct send_ctx *sctx,
4880                             const u64 offset,
4881                             const u64 len)
4882 {
4883         u64 sent = 0;
4884 
4885         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
4886                 return send_update_extent(sctx, offset, len);
4887 
4888         while (sent < len) {
4889                 u64 size = len - sent;
4890                 int ret;
4891 
4892                 if (size > BTRFS_SEND_READ_SIZE)
4893                         size = BTRFS_SEND_READ_SIZE;
4894                 ret = send_write(sctx, offset + sent, size);
4895                 if (ret < 0)
4896                         return ret;
4897                 if (!ret)
4898                         break;
4899                 sent += ret;
4900         }
4901         return 0;
4902 }
4903 
4904 static int clone_range(struct send_ctx *sctx,
4905                        struct clone_root *clone_root,
4906                        const u64 disk_byte,
4907                        u64 data_offset,
4908                        u64 offset,
4909                        u64 len)
4910 {
4911         struct btrfs_path *path;
4912         struct btrfs_key key;
4913         int ret;
4914 
4915         path = alloc_path_for_send();
4916         if (!path)
4917                 return -ENOMEM;
4918 
4919         /*
4920          * We can't send a clone operation for the entire range if we find
4921          * extent items in the respective range in the source file that
4922          * refer to different extents or if we find holes.
4923          * So check for that and do a mix of clone and regular write/copy
4924          * operations if needed.
4925          *
4926          * Example:
4927          *
4928          * mkfs.btrfs -f /dev/sda
4929          * mount /dev/sda /mnt
4930          * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
4931          * cp --reflink=always /mnt/foo /mnt/bar
4932          * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
4933          * btrfs subvolume snapshot -r /mnt /mnt/snap
4934          *
4935          * If when we send the snapshot and we are processing file bar (which
4936          * has a higher inode number than foo) we blindly send a clone operation
4937          * for the [0, 100K[ range from foo to bar, the receiver ends up getting
4938          * a file bar that matches the content of file foo - iow, doesn't match
4939          * the content from bar in the original filesystem.
4940          */
4941         key.objectid = clone_root->ino;
4942         key.type = BTRFS_EXTENT_DATA_KEY;
4943         key.offset = clone_root->offset;
4944         ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
4945         if (ret < 0)
4946                 goto out;
4947         if (ret > 0 && path->slots[0] > 0) {
4948                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
4949                 if (key.objectid == clone_root->ino &&
4950                     key.type == BTRFS_EXTENT_DATA_KEY)
4951                         path->slots[0]--;
4952         }
4953 
4954         while (true) {
4955                 struct extent_buffer *leaf = path->nodes[0];
4956                 int slot = path->slots[0];
4957                 struct btrfs_file_extent_item *ei;
4958                 u8 type;
4959                 u64 ext_len;
4960                 u64 clone_len;
4961 
4962                 if (slot >= btrfs_header_nritems(leaf)) {
4963                         ret = btrfs_next_leaf(clone_root->root, path);
4964                         if (ret < 0)
4965                                 goto out;
4966                         else if (ret > 0)
4967                                 break;
4968                         continue;
4969                 }
4970 
4971                 btrfs_item_key_to_cpu(leaf, &key, slot);
4972 
4973                 /*
4974                  * We might have an implicit trailing hole (NO_HOLES feature
4975                  * enabled). We deal with it after leaving this loop.
4976                  */
4977                 if (key.objectid != clone_root->ino ||
4978                     key.type != BTRFS_EXTENT_DATA_KEY)
4979                         break;
4980 
4981                 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4982                 type = btrfs_file_extent_type(leaf, ei);
4983                 if (type == BTRFS_FILE_EXTENT_INLINE) {
4984                         ext_len = btrfs_file_extent_inline_len(leaf, slot, ei);
4985                         ext_len = PAGE_ALIGN(ext_len);
4986                 } else {
4987                         ext_len = btrfs_file_extent_num_bytes(leaf, ei);
4988                 }
4989 
4990                 if (key.offset + ext_len <= clone_root->offset)
4991                         goto next;
4992 
4993                 if (key.offset > clone_root->offset) {
4994                         /* Implicit hole, NO_HOLES feature enabled. */
4995                         u64 hole_len = key.offset - clone_root->offset;
4996 
4997                         if (hole_len > len)
4998                                 hole_len = len;
4999                         ret = send_extent_data(sctx, offset, hole_len);
5000                         if (ret < 0)
5001                                 goto out;
5002 
5003                         len -= hole_len;
5004                         if (len == 0)
5005                                 break;
5006                         offset += hole_len;
5007                         clone_root->offset += hole_len;
5008                         data_offset += hole_len;
5009                 }
5010 
5011                 if (key.offset >= clone_root->offset + len)
5012                         break;
5013 
5014                 clone_len = min_t(u64, ext_len, len);
5015 
5016                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5017                     btrfs_file_extent_offset(leaf, ei) == data_offset)
5018                         ret = send_clone(sctx, offset, clone_len, clone_root);
5019                 else
5020                         ret = send_extent_data(sctx, offset, clone_len);
5021 
5022                 if (ret < 0)
5023                         goto out;
5024 
5025                 len -= clone_len;
5026                 if (len == 0)
5027                         break;
5028                 offset += clone_len;
5029                 clone_root->offset += clone_len;
5030                 data_offset += clone_len;
5031 next:
5032                 path->slots[0]++;
5033         }
5034 
5035         if (len > 0)
5036                 ret = send_extent_data(sctx, offset, len);
5037         else
5038                 ret = 0;
5039 out:
5040         btrfs_free_path(path);
5041         return ret;
5042 }
5043 
5044 static int send_write_or_clone(struct send_ctx *sctx,
5045                                struct btrfs_path *path,
5046                                struct btrfs_key *key,
5047                                struct clone_root *clone_root)
5048 {
5049         int ret = 0;
5050         struct btrfs_file_extent_item *ei;
5051         u64 offset = key->offset;
5052         u64 len;
5053         u8 type;
5054         u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5055 
5056         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5057                         struct btrfs_file_extent_item);
5058         type = btrfs_file_extent_type(path->nodes[0], ei);
5059         if (type == BTRFS_FILE_EXTENT_INLINE) {
5060                 len = btrfs_file_extent_inline_len(path->nodes[0],
5061                                                    path->slots[0], ei);
5062                 /*
5063                  * it is possible the inline item won't cover the whole page,
5064                  * but there may be items after this page.  Make
5065                  * sure to send the whole thing
5066                  */
5067                 len = PAGE_ALIGN(len);
5068         } else {
5069                 len = btrfs_file_extent_num_bytes(path->nodes[0], ei);
5070         }
5071 
5072         if (offset + len > sctx->cur_inode_size)
5073                 len = sctx->cur_inode_size - offset;
5074         if (len == 0) {
5075                 ret = 0;
5076                 goto out;
5077         }
5078 
5079         if (clone_root && IS_ALIGNED(offset + len, bs)) {
5080                 u64 disk_byte;
5081                 u64 data_offset;
5082 
5083                 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5084                 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5085                 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5086                                   offset, len);
5087         } else {
5088                 ret = send_extent_data(sctx, offset, len);
5089         }
5090 out:
5091         return ret;
5092 }
5093 
5094 static int is_extent_unchanged(struct send_ctx *sctx,
5095                                struct btrfs_path *left_path,
5096                                struct btrfs_key *ekey)
5097 {
5098         int ret = 0;
5099         struct btrfs_key key;
5100         struct btrfs_path *path = NULL;
5101         struct extent_buffer *eb;
5102         int slot;
5103         struct btrfs_key found_key;
5104         struct btrfs_file_extent_item *ei;
5105         u64 left_disknr;
5106         u64 right_disknr;
5107         u64 left_offset;
5108         u64 right_offset;
5109         u64 left_offset_fixed;
5110         u64 left_len;
5111         u64 right_len;
5112         u64 left_gen;
5113         u64 right_gen;
5114         u8 left_type;
5115         u8 right_type;
5116 
5117         path = alloc_path_for_send();
5118         if (!path)
5119                 return -ENOMEM;
5120 
5121         eb = left_path->nodes[0];
5122         slot = left_path->slots[0];
5123         ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5124         left_type = btrfs_file_extent_type(eb, ei);
5125 
5126         if (left_type != BTRFS_FILE_EXTENT_REG) {
5127                 ret = 0;
5128                 goto out;
5129         }
5130         left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5131         left_len = btrfs_file_extent_num_bytes(eb, ei);
5132         left_offset = btrfs_file_extent_offset(eb, ei);
5133         left_gen = btrfs_file_extent_generation(eb, ei);
5134 
5135         /*
5136          * Following comments will refer to these graphics. L is the left
5137          * extents which we are checking at the moment. 1-8 are the right
5138          * extents that we iterate.
5139          *
5140          *       |-----L-----|
5141          * |-1-|-2a-|-3-|-4-|-5-|-6-|
5142          *
5143          *       |-----L-----|
5144          * |--1--|-2b-|...(same as above)
5145          *
5146          * Alternative situation. Happens on files where extents got split.
5147          *       |-----L-----|
5148          * |-----------7-----------|-6-|
5149          *
5150          * Alternative situation. Happens on files which got larger.
5151          *       |-----L-----|
5152          * |-8-|
5153          * Nothing follows after 8.
5154          */
5155 
5156         key.objectid = ekey->objectid;
5157         key.type = BTRFS_EXTENT_DATA_KEY;
5158         key.offset = ekey->offset;
5159         ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5160         if (ret < 0)
5161                 goto out;
5162         if (ret) {
5163                 ret = 0;
5164                 goto out;
5165         }
5166 
5167         /*
5168          * Handle special case where the right side has no extents at all.
5169          */
5170         eb = path->nodes[0];
5171         slot = path->slots[0];
5172         btrfs_item_key_to_cpu(eb, &found_key, slot);
5173         if (found_key.objectid != key.objectid ||
5174             found_key.type != key.type) {
5175                 /* If we're a hole then just pretend nothing changed */
5176                 ret = (left_disknr) ? 0 : 1;
5177                 goto out;
5178         }
5179 
5180         /*
5181          * We're now on 2a, 2b or 7.
5182          */
5183         key = found_key;
5184         while (key.offset < ekey->offset + left_len) {
5185                 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5186                 right_type = btrfs_file_extent_type(eb, ei);
5187                 if (right_type != BTRFS_FILE_EXTENT_REG &&
5188                     right_type != BTRFS_FILE_EXTENT_INLINE) {
5189                         ret = 0;
5190                         goto out;
5191                 }
5192 
5193                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5194                         right_len = btrfs_file_extent_inline_len(eb, slot, ei);
5195                         right_len = PAGE_ALIGN(right_len);
5196                 } else {
5197                         right_len = btrfs_file_extent_num_bytes(eb, ei);
5198                 }
5199 
5200                 /*
5201                  * Are we at extent 8? If yes, we know the extent is changed.
5202                  * This may only happen on the first iteration.
5203                  */
5204                 if (found_key.offset + right_len <= ekey->offset) {
5205                         /* If we're a hole just pretend nothing changed */
5206                         ret = (left_disknr) ? 0 : 1;
5207                         goto out;
5208                 }
5209 
5210                 /*
5211                  * We just wanted to see if when we have an inline extent, what
5212                  * follows it is a regular extent (wanted to check the above
5213                  * condition for inline extents too). This should normally not
5214                  * happen but it's possible for example when we have an inline
5215                  * compressed extent representing data with a size matching
5216                  * the page size (currently the same as sector size).
5217                  */
5218                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5219                         ret = 0;
5220                         goto out;
5221                 }
5222 
5223                 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5224                 right_offset = btrfs_file_extent_offset(eb, ei);
5225                 right_gen = btrfs_file_extent_generation(eb, ei);
5226 
5227                 left_offset_fixed = left_offset;
5228                 if (key.offset < ekey->offset) {
5229                         /* Fix the right offset for 2a and 7. */
5230                         right_offset += ekey->offset - key.offset;
5231                 } else {
5232                         /* Fix the left offset for all behind 2a and 2b */
5233                         left_offset_fixed += key.offset - ekey->offset;
5234                 }
5235 
5236                 /*
5237                  * Check if we have the same extent.
5238                  */
5239                 if (left_disknr != right_disknr ||
5240                     left_offset_fixed != right_offset ||
5241                     left_gen != right_gen) {
5242                         ret = 0;
5243                         goto out;
5244                 }
5245 
5246                 /*
5247                  * Go to the next extent.
5248                  */
5249                 ret = btrfs_next_item(sctx->parent_root, path);
5250                 if (ret < 0)
5251                         goto out;
5252                 if (!ret) {
5253                         eb = path->nodes[0];
5254                         slot = path->slots[0];
5255                         btrfs_item_key_to_cpu(eb, &found_key, slot);
5256                 }
5257                 if (ret || found_key.objectid != key.objectid ||
5258                     found_key.type != key.type) {
5259                         key.offset += right_len;
5260                         break;
5261                 }
5262                 if (found_key.offset != key.offset + right_len) {
5263                         ret = 0;
5264                         goto out;
5265                 }
5266                 key = found_key;
5267         }
5268 
5269         /*
5270          * We're now behind the left extent (treat as unchanged) or at the end
5271          * of the right side (treat as changed).
5272          */
5273         if (key.offset >= ekey->offset + left_len)
5274                 ret = 1;
5275         else
5276                 ret = 0;
5277 
5278 
5279 out:
5280         btrfs_free_path(path);
5281         return ret;
5282 }
5283 
5284 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5285 {
5286         struct btrfs_path *path;
5287         struct btrfs_root *root = sctx->send_root;
5288         struct btrfs_file_extent_item *fi;
5289         struct btrfs_key key;
5290         u64 extent_end;
5291         u8 type;
5292         int ret;
5293 
5294         path = alloc_path_for_send();
5295         if (!path)
5296                 return -ENOMEM;
5297 
5298         sctx->cur_inode_last_extent = 0;
5299 
5300         key.objectid = sctx->cur_ino;
5301         key.type = BTRFS_EXTENT_DATA_KEY;
5302         key.offset = offset;
5303         ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5304         if (ret < 0)
5305                 goto out;
5306         ret = 0;
5307         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5308         if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5309                 goto out;
5310 
5311         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5312                             struct btrfs_file_extent_item);
5313         type = btrfs_file_extent_type(path->nodes[0], fi);
5314         if (type == BTRFS_FILE_EXTENT_INLINE) {
5315                 u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5316                                                         path->slots[0], fi);
5317                 extent_end = ALIGN(key.offset + size,
5318                                    sctx->send_root->fs_info->sectorsize);
5319         } else {
5320                 extent_end = key.offset +
5321                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5322         }
5323         sctx->cur_inode_last_extent = extent_end;
5324 out:
5325         btrfs_free_path(path);
5326         return ret;
5327 }
5328 
5329 static int range_is_hole_in_parent(struct send_ctx *sctx,
5330                                    const u64 start,
5331                                    const u64 end)
5332 {
5333         struct btrfs_path *path;
5334         struct btrfs_key key;
5335         struct btrfs_root *root = sctx->parent_root;
5336         u64 search_start = start;
5337         int ret;
5338 
5339         path = alloc_path_for_send();
5340         if (!path)
5341                 return -ENOMEM;
5342 
5343         key.objectid = sctx->cur_ino;
5344         key.type = BTRFS_EXTENT_DATA_KEY;
5345         key.offset = search_start;
5346         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5347         if (ret < 0)
5348                 goto out;
5349         if (ret > 0 && path->slots[0] > 0)
5350                 path->slots[0]--;
5351 
5352         while (search_start < end) {
5353                 struct extent_buffer *leaf = path->nodes[0];
5354                 int slot = path->slots[0];
5355                 struct btrfs_file_extent_item *fi;
5356                 u64 extent_end;
5357 
5358                 if (slot >= btrfs_header_nritems(leaf)) {
5359                         ret = btrfs_next_leaf(root, path);
5360                         if (ret < 0)
5361                                 goto out;
5362                         else if (ret > 0)
5363                                 break;
5364                         continue;
5365                 }
5366 
5367                 btrfs_item_key_to_cpu(leaf, &key, slot);
5368                 if (key.objectid < sctx->cur_ino ||
5369                     key.type < BTRFS_EXTENT_DATA_KEY)
5370                         goto next;
5371                 if (key.objectid > sctx->cur_ino ||
5372                     key.type > BTRFS_EXTENT_DATA_KEY ||
5373                     key.offset >= end)
5374                         break;
5375 
5376                 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5377                 if (btrfs_file_extent_type(leaf, fi) ==
5378                     BTRFS_FILE_EXTENT_INLINE) {
5379                         u64 size = btrfs_file_extent_inline_len(leaf, slot, fi);
5380 
5381                         extent_end = ALIGN(key.offset + size,
5382                                            root->fs_info->sectorsize);
5383                 } else {
5384                         extent_end = key.offset +
5385                                 btrfs_file_extent_num_bytes(leaf, fi);
5386                 }
5387                 if (extent_end <= start)
5388                         goto next;
5389                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5390                         search_start = extent_end;
5391                         goto next;
5392                 }
5393                 ret = 0;
5394                 goto out;
5395 next:
5396                 path->slots[0]++;
5397         }
5398         ret = 1;
5399 out:
5400         btrfs_free_path(path);
5401         return ret;
5402 }
5403 
5404 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5405                            struct btrfs_key *key)
5406 {
5407         struct btrfs_file_extent_item *fi;
5408         u64 extent_end;
5409         u8 type;
5410         int ret = 0;
5411 
5412         if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5413                 return 0;
5414 
5415         if (sctx->cur_inode_last_extent == (u64)-1) {
5416                 ret = get_last_extent(sctx, key->offset - 1);
5417                 if (ret)
5418                         return ret;
5419         }
5420 
5421         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5422                             struct btrfs_file_extent_item);
5423         type = btrfs_file_extent_type(path->nodes[0], fi);
5424         if (type == BTRFS_FILE_EXTENT_INLINE) {
5425                 u64 size = btrfs_file_extent_inline_len(path->nodes[0],
5426                                                         path->slots[0], fi);
5427                 extent_end = ALIGN(key->offset + size,
5428                                    sctx->send_root->fs_info->sectorsize);
5429         } else {
5430                 extent_end = key->offset +
5431                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5432         }
5433 
5434         if (path->slots[0] == 0 &&
5435             sctx->cur_inode_last_extent < key->offset) {
5436                 /*
5437                  * We might have skipped entire leafs that contained only
5438                  * file extent items for our current inode. These leafs have
5439                  * a generation number smaller (older) than the one in the
5440                  * current leaf and the leaf our last extent came from, and
5441                  * are located between these 2 leafs.
5442                  */
5443                 ret = get_last_extent(sctx, key->offset - 1);
5444                 if (ret)
5445                         return ret;
5446         }
5447 
5448         if (sctx->cur_inode_last_extent < key->offset) {
5449                 ret = range_is_hole_in_parent(sctx,
5450                                               sctx->cur_inode_last_extent,
5451                                               key->offset);
5452                 if (ret < 0)
5453                         return ret;
5454                 else if (ret == 0)
5455                         ret = send_hole(sctx, key->offset);
5456                 else
5457                         ret = 0;
5458         }
5459         sctx->cur_inode_last_extent = extent_end;
5460         return ret;
5461 }
5462 
5463 static int process_extent(struct send_ctx *sctx,
5464                           struct btrfs_path *path,
5465                           struct btrfs_key *key)
5466 {
5467         struct clone_root *found_clone = NULL;
5468         int ret = 0;
5469 
5470         if (S_ISLNK(sctx->cur_inode_mode))
5471                 return 0;
5472 
5473         if (sctx->parent_root && !sctx->cur_inode_new) {
5474                 ret = is_extent_unchanged(sctx, path, key);
5475                 if (ret < 0)
5476                         goto out;
5477                 if (ret) {
5478                         ret = 0;
5479                         goto out_hole;
5480                 }
5481         } else {
5482                 struct btrfs_file_extent_item *ei;
5483                 u8 type;
5484 
5485                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5486                                     struct btrfs_file_extent_item);
5487                 type = btrfs_file_extent_type(path->nodes[0], ei);
5488                 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5489                     type == BTRFS_FILE_EXTENT_REG) {
5490                         /*
5491                          * The send spec does not have a prealloc command yet,
5492                          * so just leave a hole for prealloc'ed extents until
5493                          * we have enough commands queued up to justify rev'ing
5494                          * the send spec.
5495                          */
5496                         if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5497                                 ret = 0;
5498                                 goto out;
5499                         }
5500 
5501                         /* Have a hole, just skip it. */
5502                         if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5503                                 ret = 0;
5504                                 goto out;
5505                         }
5506                 }
5507         }
5508 
5509         ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5510                         sctx->cur_inode_size, &found_clone);
5511         if (ret != -ENOENT && ret < 0)
5512                 goto out;
5513 
5514         ret = send_write_or_clone(sctx, path, key, found_clone);
5515         if (ret)
5516                 goto out;
5517 out_hole:
5518         ret = maybe_send_hole(sctx, path, key);
5519 out:
5520         return ret;
5521 }
5522 
5523 static int process_all_extents(struct send_ctx *sctx)
5524 {
5525         int ret;
5526         struct btrfs_root *root;
5527         struct btrfs_path *path;
5528         struct btrfs_key key;
5529         struct btrfs_key found_key;
5530         struct extent_buffer *eb;
5531         int slot;
5532 
5533         root = sctx->send_root;
5534         path = alloc_path_for_send();
5535         if (!path)
5536                 return -ENOMEM;
5537 
5538         key.objectid = sctx->cmp_key->objectid;
5539         key.type = BTRFS_EXTENT_DATA_KEY;
5540         key.offset = 0;
5541         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5542         if (ret < 0)
5543                 goto out;
5544 
5545         while (1) {
5546                 eb = path->nodes[0];
5547                 slot = path->slots[0];
5548 
5549                 if (slot >= btrfs_header_nritems(eb)) {
5550                         ret = btrfs_next_leaf(root, path);
5551                         if (ret < 0) {
5552                                 goto out;
5553                         } else if (ret > 0) {
5554                                 ret = 0;
5555                                 break;
5556                         }
5557                         continue;
5558                 }
5559 
5560                 btrfs_item_key_to_cpu(eb, &found_key, slot);
5561 
5562                 if (found_key.objectid != key.objectid ||
5563                     found_key.type != key.type) {
5564                         ret = 0;
5565                         goto out;
5566                 }
5567 
5568                 ret = process_extent(sctx, path, &found_key);
5569                 if (ret < 0)
5570                         goto out;
5571 
5572                 path->slots[0]++;
5573         }
5574 
5575 out:
5576         btrfs_free_path(path);
5577         return ret;
5578 }
5579 
5580 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
5581                                            int *pending_move,
5582                                            int *refs_processed)
5583 {
5584         int ret = 0;
5585 
5586         if (sctx->cur_ino == 0)
5587                 goto out;
5588         if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
5589             sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
5590                 goto out;
5591         if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
5592                 goto out;
5593 
5594         ret = process_recorded_refs(sctx, pending_move);
5595         if (ret < 0)
5596                 goto out;
5597 
5598         *refs_processed = 1;
5599 out:
5600         return ret;
5601 }
5602 
5603 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
5604 {
5605         int ret = 0;
5606         u64 left_mode;
5607         u64 left_uid;
5608         u64 left_gid;
5609         u64 right_mode;
5610         u64 right_uid;
5611         u64 right_gid;
5612         int need_chmod = 0;
5613         int need_chown = 0;
5614         int pending_move = 0;
5615         int refs_processed = 0;
5616 
5617         ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
5618                                               &refs_processed);
5619         if (ret < 0)
5620                 goto out;
5621 
5622         /*
5623          * We have processed the refs and thus need to advance send_progress.
5624          * Now, calls to get_cur_xxx will take the updated refs of the current
5625          * inode into account.
5626          *
5627          * On the other hand, if our current inode is a directory and couldn't
5628          * be moved/renamed because its parent was renamed/moved too and it has
5629          * a higher inode number, we can only move/rename our current inode
5630          * after we moved/renamed its parent. Therefore in this case operate on
5631          * the old path (pre move/rename) of our current inode, and the
5632          * move/rename will be performed later.
5633          */
5634         if (refs_processed && !pending_move)
5635                 sctx->send_progress = sctx->cur_ino + 1;
5636 
5637         if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
5638                 goto out;
5639         if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
5640                 goto out;
5641 
5642         ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
5643                         &left_mode, &left_uid, &left_gid, NULL);
5644         if (ret < 0)
5645                 goto out;
5646 
5647         if (!sctx->parent_root || sctx->cur_inode_new) {
5648                 need_chown = 1;
5649                 if (!S_ISLNK(sctx->cur_inode_mode))
5650                         need_chmod = 1;
5651         } else {
5652                 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
5653                                 NULL, NULL, &right_mode, &right_uid,
5654                                 &right_gid, NULL);
5655                 if (ret < 0)
5656                         goto out;
5657 
5658                 if (left_uid != right_uid || left_gid != right_gid)
5659                         need_chown = 1;
5660                 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
5661                         need_chmod = 1;
5662         }
5663 
5664         if (S_ISREG(sctx->cur_inode_mode)) {
5665                 if (need_send_hole(sctx)) {
5666                         if (sctx->cur_inode_last_extent == (u64)-1 ||
5667                             sctx->cur_inode_last_extent <
5668                             sctx->cur_inode_size) {
5669                                 ret = get_last_extent(sctx, (u64)-1);
5670                                 if (ret)
5671                                         goto out;
5672                         }
5673                         if (sctx->cur_inode_last_extent <
5674                             sctx->cur_inode_size) {
5675                                 ret = send_hole(sctx, sctx->cur_inode_size);
5676                                 if (ret)
5677                                         goto out;
5678                         }
5679                 }
5680                 ret = send_truncate(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5681                                 sctx->cur_inode_size);
5682                 if (ret < 0)
5683                         goto out;
5684         }
5685 
5686         if (need_chown) {
5687                 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5688                                 left_uid, left_gid);
5689                 if (ret < 0)
5690                         goto out;
5691         }
5692         if (need_chmod) {
5693                 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5694                                 left_mode);
5695                 if (ret < 0)
5696                         goto out;
5697         }
5698 
5699         /*
5700          * If other directory inodes depended on our current directory
5701          * inode's move/rename, now do their move/rename operations.
5702          */
5703         if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
5704                 ret = apply_children_dir_moves(sctx);
5705                 if (ret)
5706                         goto out;
5707                 /*
5708                  * Need to send that every time, no matter if it actually
5709                  * changed between the two trees as we have done changes to
5710                  * the inode before. If our inode is a directory and it's
5711                  * waiting to be moved/renamed, we will send its utimes when
5712                  * it's moved/renamed, therefore we don't need to do it here.
5713                  */
5714                 sctx->send_progress = sctx->cur_ino + 1;
5715                 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
5716                 if (ret < 0)
5717                         goto out;
5718         }
5719 
5720 out:
5721         return ret;
5722 }
5723 
5724 static int changed_inode(struct send_ctx *sctx,
5725                          enum btrfs_compare_tree_result result)
5726 {
5727         int ret = 0;
5728         struct btrfs_key *key = sctx->cmp_key;
5729         struct btrfs_inode_item *left_ii = NULL;
5730         struct btrfs_inode_item *right_ii = NULL;
5731         u64 left_gen = 0;
5732         u64 right_gen = 0;
5733 
5734         sctx->cur_ino = key->objectid;
5735         sctx->cur_inode_new_gen = 0;
5736         sctx->cur_inode_last_extent = (u64)-1;
5737 
5738         /*
5739          * Set send_progress to current inode. This will tell all get_cur_xxx
5740          * functions that the current inode's refs are not updated yet. Later,
5741          * when process_recorded_refs is finished, it is set to cur_ino + 1.
5742          */
5743         sctx->send_progress = sctx->cur_ino;
5744 
5745         if (result == BTRFS_COMPARE_TREE_NEW ||
5746             result == BTRFS_COMPARE_TREE_CHANGED) {
5747                 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0],
5748                                 sctx->left_path->slots[0],
5749                                 struct btrfs_inode_item);
5750                 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0],
5751                                 left_ii);
5752         } else {
5753                 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5754                                 sctx->right_path->slots[0],
5755                                 struct btrfs_inode_item);
5756                 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5757                                 right_ii);
5758         }
5759         if (result == BTRFS_COMPARE_TREE_CHANGED) {
5760                 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0],
5761                                 sctx->right_path->slots[0],
5762                                 struct btrfs_inode_item);
5763 
5764                 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0],
5765                                 right_ii);
5766 
5767                 /*
5768                  * The cur_ino = root dir case is special here. We can't treat
5769                  * the inode as deleted+reused because it would generate a
5770                  * stream that tries to delete/mkdir the root dir.
5771                  */
5772                 if (left_gen != right_gen &&
5773                     sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5774                         sctx->cur_inode_new_gen = 1;
5775         }
5776 
5777         if (result == BTRFS_COMPARE_TREE_NEW) {
5778                 sctx->cur_inode_gen = left_gen;
5779                 sctx->cur_inode_new = 1;
5780                 sctx->cur_inode_deleted = 0;
5781                 sctx->cur_inode_size = btrfs_inode_size(
5782                                 sctx->left_path->nodes[0], left_ii);
5783                 sctx->cur_inode_mode = btrfs_inode_mode(
5784                                 sctx->left_path->nodes[0], left_ii);
5785                 sctx->cur_inode_rdev = btrfs_inode_rdev(
5786                                 sctx->left_path->nodes[0], left_ii);
5787                 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID)
5788                         ret = send_create_inode_if_needed(sctx);
5789         } else if (result == BTRFS_COMPARE_TREE_DELETED) {
5790                 sctx->cur_inode_gen = right_gen;
5791                 sctx->cur_inode_new = 0;
5792                 sctx->cur_inode_deleted = 1;
5793                 sctx->cur_inode_size = btrfs_inode_size(
5794                                 sctx->right_path->nodes[0], right_ii);
5795                 sctx->cur_inode_mode = btrfs_inode_mode(
5796                                 sctx->right_path->nodes[0], right_ii);
5797         } else if (result == BTRFS_COMPARE_TREE_CHANGED) {
5798                 /*
5799                  * We need to do some special handling in case the inode was
5800                  * reported as changed with a changed generation number. This
5801                  * means that the original inode was deleted and new inode
5802                  * reused the same inum. So we have to treat the old inode as
5803                  * deleted and the new one as new.
5804                  */
5805                 if (sctx->cur_inode_new_gen) {
5806                         /*
5807                          * First, process the inode as if it was deleted.
5808                          */
5809                         sctx->cur_inode_gen = right_gen;
5810                         sctx->cur_inode_new = 0;
5811                         sctx->cur_inode_deleted = 1;
5812                         sctx->cur_inode_size = btrfs_inode_size(
5813                                         sctx->right_path->nodes[0], right_ii);
5814                         sctx->cur_inode_mode = btrfs_inode_mode(
5815                                         sctx->right_path->nodes[0], right_ii);
5816                         ret = process_all_refs(sctx,
5817                                         BTRFS_COMPARE_TREE_DELETED);
5818                         if (ret < 0)
5819                                 goto out;
5820 
5821                         /*
5822                          * Now process the inode as if it was new.
5823                          */
5824                         sctx->cur_inode_gen = left_gen;
5825                         sctx->cur_inode_new = 1;
5826                         sctx->cur_inode_deleted = 0;
5827                         sctx->cur_inode_size = btrfs_inode_size(
5828                                         sctx->left_path->nodes[0], left_ii);
5829                         sctx->cur_inode_mode = btrfs_inode_mode(
5830                                         sctx->left_path->nodes[0], left_ii);
5831                         sctx->cur_inode_rdev = btrfs_inode_rdev(
5832                                         sctx->left_path->nodes[0], left_ii);
5833                         ret = send_create_inode_if_needed(sctx);
5834                         if (ret < 0)
5835                                 goto out;
5836 
5837                         ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW);
5838                         if (ret < 0)
5839                                 goto out;
5840                         /*
5841                          * Advance send_progress now as we did not get into
5842                          * process_recorded_refs_if_needed in the new_gen case.
5843                          */
5844                         sctx->send_progress = sctx->cur_ino + 1;
5845 
5846                         /*
5847                          * Now process all extents and xattrs of the inode as if
5848                          * they were all new.
5849                          */
5850                         ret = process_all_extents(sctx);
5851                         if (ret < 0)
5852                                 goto out;
5853                         ret = process_all_new_xattrs(sctx);
5854                         if (ret < 0)
5855                                 goto out;
5856                 } else {
5857                         sctx->cur_inode_gen = left_gen;
5858                         sctx->cur_inode_new = 0;
5859                         sctx->cur_inode_new_gen = 0;
5860                         sctx->cur_inode_deleted = 0;
5861                         sctx->cur_inode_size = btrfs_inode_size(
5862                                         sctx->left_path->nodes[0], left_ii);
5863                         sctx->cur_inode_mode = btrfs_inode_mode(
5864                                         sctx->left_path->nodes[0], left_ii);
5865                 }
5866         }
5867 
5868 out:
5869         return ret;
5870 }
5871 
5872 /*
5873  * We have to process new refs before deleted refs, but compare_trees gives us
5874  * the new and deleted refs mixed. To fix this, we record the new/deleted refs
5875  * first and later process them in process_recorded_refs.
5876  * For the cur_inode_new_gen case, we skip recording completely because
5877  * changed_inode did already initiate processing of refs. The reason for this is
5878  * that in this case, compare_tree actually compares the refs of 2 different
5879  * inodes. To fix this, process_all_refs is used in changed_inode to handle all
5880  * refs of the right tree as deleted and all refs of the left tree as new.
5881  */
5882 static int changed_ref(struct send_ctx *sctx,
5883                        enum btrfs_compare_tree_result result)
5884 {
5885         int ret = 0;
5886 
5887         if (sctx->cur_ino != sctx->cmp_key->objectid) {
5888                 inconsistent_snapshot_error(sctx, result, "reference");
5889                 return -EIO;
5890         }
5891 
5892         if (!sctx->cur_inode_new_gen &&
5893             sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) {
5894                 if (result == BTRFS_COMPARE_TREE_NEW)
5895                         ret = record_new_ref(sctx);
5896                 else if (result == BTRFS_COMPARE_TREE_DELETED)
5897                         ret = record_deleted_ref(sctx);
5898                 else if (result == BTRFS_COMPARE_TREE_CHANGED)
5899                         ret = record_changed_ref(sctx);
5900         }
5901 
5902         return ret;
5903 }
5904 
5905 /*
5906  * Process new/deleted/changed xattrs. We skip processing in the
5907  * cur_inode_new_gen case because changed_inode did already initiate processing
5908  * of xattrs. The reason is the same as in changed_ref
5909  */
5910 static int changed_xattr(struct send_ctx *sctx,
5911                          enum btrfs_compare_tree_result result)
5912 {
5913         int ret = 0;
5914 
5915         if (sctx->cur_ino != sctx->cmp_key->objectid) {
5916                 inconsistent_snapshot_error(sctx, result, "xattr");
5917                 return -EIO;
5918         }
5919 
5920         if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) {
5921                 if (result == BTRFS_COMPARE_TREE_NEW)
5922                         ret = process_new_xattr(sctx);
5923                 else if (result == BTRFS_COMPARE_TREE_DELETED)
5924                         ret = process_deleted_xattr(sctx);
5925                 else if (result == BTRFS_COMPARE_TREE_CHANGED)
5926                         ret = process_changed_xattr(sctx);
5927         }
5928 
5929         return ret;
5930 }
5931 
5932 /*
5933  * Process new/deleted/changed extents. We skip processing in the
5934  * cur_inode_new_gen case because changed_inode did already initiate processing
5935  * of extents. The reason is the same as in changed_ref
5936  */
5937 static int changed_extent(struct send_ctx *sctx,
5938                           enum btrfs_compare_tree_result result)
5939 {
5940         int ret = 0;
5941 
5942         if (sctx->cur_ino != sctx->cmp_key->objectid) {
5943 
5944                 if (result == BTRFS_COMPARE_TREE_CHANGED) {
5945                         struct extent_buffer *leaf_l;
5946                         struct extent_buffer *leaf_r;
5947                         struct btrfs_file_extent_item *ei_l;
5948                         struct btrfs_file_extent_item *ei_r;
5949 
5950                         leaf_l = sctx->left_path->nodes[0];
5951                         leaf_r = sctx->right_path->nodes[0];
5952                         ei_l = btrfs_item_ptr(leaf_l,
5953                                               sctx->left_path->slots[0],
5954                                               struct btrfs_file_extent_item);
5955                         ei_r = btrfs_item_ptr(leaf_r,
5956                                               sctx->right_path->slots[0],
5957                                               struct btrfs_file_extent_item);
5958 
5959                         /*
5960                          * We may have found an extent item that has changed
5961                          * only its disk_bytenr field and the corresponding
5962                          * inode item was not updated. This case happens due to
5963                          * very specific timings during relocation when a leaf
5964                          * that contains file extent items is COWed while
5965                          * relocation is ongoing and its in the stage where it
5966                          * updates data pointers. So when this happens we can
5967                          * safely ignore it since we know it's the same extent,
5968                          * but just at different logical and physical locations
5969                          * (when an extent is fully replaced with a new one, we
5970                          * know the generation number must have changed too,
5971                          * since snapshot creation implies committing the current
5972                          * transaction, and the inode item must have been updated
5973                          * as well).
5974                          * This replacement of the disk_bytenr happens at
5975                          * relocation.c:replace_file_extents() through
5976                          * relocation.c:btrfs_reloc_cow_block().
5977                          */
5978                         if (btrfs_file_extent_generation(leaf_l, ei_l) ==
5979                             btrfs_file_extent_generation(leaf_r, ei_r) &&
5980                             btrfs_file_extent_ram_bytes(leaf_l, ei_l) ==
5981                             btrfs_file_extent_ram_bytes(leaf_r, ei_r) &&
5982                             btrfs_file_extent_compression(leaf_l, ei_l) ==
5983                             btrfs_file_extent_compression(leaf_r, ei_r) &&
5984                             btrfs_file_extent_encryption(leaf_l, ei_l) ==
5985                             btrfs_file_extent_encryption(leaf_r, ei_r) &&
5986                             btrfs_file_extent_other_encoding(leaf_l, ei_l) ==
5987                             btrfs_file_extent_other_encoding(leaf_r, ei_r) &&
5988                             btrfs_file_extent_type(leaf_l, ei_l) ==
5989