~ [ 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 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * Copyright (C) 2012 Alexander Block.  All rights reserved.
  4  */
  5 
  6 #include <linux/bsearch.h>
  7 #include <linux/fs.h>
  8 #include <linux/file.h>
  9 #include <linux/sort.h>
 10 #include <linux/mount.h>
 11 #include <linux/xattr.h>
 12 #include <linux/posix_acl_xattr.h>
 13 #include <linux/radix-tree.h>
 14 #include <linux/vmalloc.h>
 15 #include <linux/string.h>
 16 #include <linux/compat.h>
 17 #include <linux/crc32c.h>
 18 
 19 #include "send.h"
 20 #include "backref.h"
 21 #include "locking.h"
 22 #include "disk-io.h"
 23 #include "btrfs_inode.h"
 24 #include "transaction.h"
 25 #include "compression.h"
 26 
 27 /*
 28  * A fs_path is a helper to dynamically build path names with unknown size.
 29  * It reallocates the internal buffer on demand.
 30  * It allows fast adding of path elements on the right side (normal path) and
 31  * fast adding to the left side (reversed path). A reversed path can also be
 32  * unreversed if needed.
 33  */
 34 struct fs_path {
 35         union {
 36                 struct {
 37                         char *start;
 38                         char *end;
 39 
 40                         char *buf;
 41                         unsigned short buf_len:15;
 42                         unsigned short reversed:1;
 43                         char inline_buf[];
 44                 };
 45                 /*
 46                  * Average path length does not exceed 200 bytes, we'll have
 47                  * better packing in the slab and higher chance to satisfy
 48                  * a allocation later during send.
 49                  */
 50                 char pad[256];
 51         };
 52 };
 53 #define FS_PATH_INLINE_SIZE \
 54         (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
 55 
 56 
 57 /* reused for each extent */
 58 struct clone_root {
 59         struct btrfs_root *root;
 60         u64 ino;
 61         u64 offset;
 62 
 63         u64 found_refs;
 64 };
 65 
 66 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128
 67 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
 68 
 69 struct send_ctx {
 70         struct file *send_filp;
 71         loff_t send_off;
 72         char *send_buf;
 73         u32 send_size;
 74         u32 send_max_size;
 75         u64 total_send_size;
 76         u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
 77         u64 flags;      /* 'flags' member of btrfs_ioctl_send_args is u64 */
 78 
 79         struct btrfs_root *send_root;
 80         struct btrfs_root *parent_root;
 81         struct clone_root *clone_roots;
 82         int clone_roots_cnt;
 83 
 84         /* current state of the compare_tree call */
 85         struct btrfs_path *left_path;
 86         struct btrfs_path *right_path;
 87         struct btrfs_key *cmp_key;
 88 
 89         /*
 90          * infos of the currently processed inode. In case of deleted inodes,
 91          * these are the values from the deleted inode.
 92          */
 93         u64 cur_ino;
 94         u64 cur_inode_gen;
 95         int cur_inode_new;
 96         int cur_inode_new_gen;
 97         int cur_inode_deleted;
 98         u64 cur_inode_size;
 99         u64 cur_inode_mode;
100         u64 cur_inode_rdev;
101         u64 cur_inode_last_extent;
102         u64 cur_inode_next_write_offset;
103         bool ignore_cur_inode;
104 
105         u64 send_progress;
106 
107         struct list_head new_refs;
108         struct list_head deleted_refs;
109 
110         struct radix_tree_root name_cache;
111         struct list_head name_cache_list;
112         int name_cache_size;
113 
114         struct file_ra_state ra;
115 
116         char *read_buf;
117 
118         /*
119          * We process inodes by their increasing order, so if before an
120          * incremental send we reverse the parent/child relationship of
121          * directories such that a directory with a lower inode number was
122          * the parent of a directory with a higher inode number, and the one
123          * becoming the new parent got renamed too, we can't rename/move the
124          * directory with lower inode number when we finish processing it - we
125          * must process the directory with higher inode number first, then
126          * rename/move it and then rename/move the directory with lower inode
127          * number. Example follows.
128          *
129          * Tree state when the first send was performed:
130          *
131          * .
132          * |-- a                   (ino 257)
133          *     |-- b               (ino 258)
134          *         |
135          *         |
136          *         |-- c           (ino 259)
137          *         |   |-- d       (ino 260)
138          *         |
139          *         |-- c2          (ino 261)
140          *
141          * Tree state when the second (incremental) send is performed:
142          *
143          * .
144          * |-- a                   (ino 257)
145          *     |-- b               (ino 258)
146          *         |-- c2          (ino 261)
147          *             |-- d2      (ino 260)
148          *                 |-- cc  (ino 259)
149          *
150          * The sequence of steps that lead to the second state was:
151          *
152          * mv /a/b/c/d /a/b/c2/d2
153          * mv /a/b/c /a/b/c2/d2/cc
154          *
155          * "c" has lower inode number, but we can't move it (2nd mv operation)
156          * before we move "d", which has higher inode number.
157          *
158          * So we just memorize which move/rename operations must be performed
159          * later when their respective parent is processed and moved/renamed.
160          */
161 
162         /* Indexed by parent directory inode number. */
163         struct rb_root pending_dir_moves;
164 
165         /*
166          * Reverse index, indexed by the inode number of a directory that
167          * is waiting for the move/rename of its immediate parent before its
168          * own move/rename can be performed.
169          */
170         struct rb_root waiting_dir_moves;
171 
172         /*
173          * A directory that is going to be rm'ed might have a child directory
174          * which is in the pending directory moves index above. In this case,
175          * the directory can only be removed after the move/rename of its child
176          * is performed. Example:
177          *
178          * Parent snapshot:
179          *
180          * .                        (ino 256)
181          * |-- a/                   (ino 257)
182          *     |-- b/               (ino 258)
183          *         |-- c/           (ino 259)
184          *         |   |-- x/       (ino 260)
185          *         |
186          *         |-- y/           (ino 261)
187          *
188          * Send snapshot:
189          *
190          * .                        (ino 256)
191          * |-- a/                   (ino 257)
192          *     |-- b/               (ino 258)
193          *         |-- YY/          (ino 261)
194          *              |-- x/      (ino 260)
195          *
196          * Sequence of steps that lead to the send snapshot:
197          * rm -f /a/b/c/foo.txt
198          * mv /a/b/y /a/b/YY
199          * mv /a/b/c/x /a/b/YY
200          * rmdir /a/b/c
201          *
202          * When the child is processed, its move/rename is delayed until its
203          * parent is processed (as explained above), but all other operations
204          * like update utimes, chown, chgrp, etc, are performed and the paths
205          * that it uses for those operations must use the orphanized name of
206          * its parent (the directory we're going to rm later), so we need to
207          * memorize that name.
208          *
209          * Indexed by the inode number of the directory to be deleted.
210          */
211         struct rb_root orphan_dirs;
212 };
213 
214 struct pending_dir_move {
215         struct rb_node node;
216         struct list_head list;
217         u64 parent_ino;
218         u64 ino;
219         u64 gen;
220         struct list_head update_refs;
221 };
222 
223 struct waiting_dir_move {
224         struct rb_node node;
225         u64 ino;
226         /*
227          * There might be some directory that could not be removed because it
228          * was waiting for this directory inode to be moved first. Therefore
229          * after this directory is moved, we can try to rmdir the ino rmdir_ino.
230          */
231         u64 rmdir_ino;
232         bool orphanized;
233 };
234 
235 struct orphan_dir_info {
236         struct rb_node node;
237         u64 ino;
238         u64 gen;
239         u64 last_dir_index_offset;
240 };
241 
242 struct name_cache_entry {
243         struct list_head list;
244         /*
245          * radix_tree has only 32bit entries but we need to handle 64bit inums.
246          * We use the lower 32bit of the 64bit inum to store it in the tree. If
247          * more then one inum would fall into the same entry, we use radix_list
248          * to store the additional entries. radix_list is also used to store
249          * entries where two entries have the same inum but different
250          * generations.
251          */
252         struct list_head radix_list;
253         u64 ino;
254         u64 gen;
255         u64 parent_ino;
256         u64 parent_gen;
257         int ret;
258         int need_later_update;
259         int name_len;
260         char name[];
261 };
262 
263 __cold
264 static void inconsistent_snapshot_error(struct send_ctx *sctx,
265                                         enum btrfs_compare_tree_result result,
266                                         const char *what)
267 {
268         const char *result_string;
269 
270         switch (result) {
271         case BTRFS_COMPARE_TREE_NEW:
272                 result_string = "new";
273                 break;
274         case BTRFS_COMPARE_TREE_DELETED:
275                 result_string = "deleted";
276                 break;
277         case BTRFS_COMPARE_TREE_CHANGED:
278                 result_string = "updated";
279                 break;
280         case BTRFS_COMPARE_TREE_SAME:
281                 ASSERT(0);
282                 result_string = "unchanged";
283                 break;
284         default:
285                 ASSERT(0);
286                 result_string = "unexpected";
287         }
288 
289         btrfs_err(sctx->send_root->fs_info,
290                   "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
291                   result_string, what, sctx->cmp_key->objectid,
292                   sctx->send_root->root_key.objectid,
293                   (sctx->parent_root ?
294                    sctx->parent_root->root_key.objectid : 0));
295 }
296 
297 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
298 
299 static struct waiting_dir_move *
300 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
301 
302 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino);
303 
304 static int need_send_hole(struct send_ctx *sctx)
305 {
306         return (sctx->parent_root && !sctx->cur_inode_new &&
307                 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
308                 S_ISREG(sctx->cur_inode_mode));
309 }
310 
311 static void fs_path_reset(struct fs_path *p)
312 {
313         if (p->reversed) {
314                 p->start = p->buf + p->buf_len - 1;
315                 p->end = p->start;
316                 *p->start = 0;
317         } else {
318                 p->start = p->buf;
319                 p->end = p->start;
320                 *p->start = 0;
321         }
322 }
323 
324 static struct fs_path *fs_path_alloc(void)
325 {
326         struct fs_path *p;
327 
328         p = kmalloc(sizeof(*p), GFP_KERNEL);
329         if (!p)
330                 return NULL;
331         p->reversed = 0;
332         p->buf = p->inline_buf;
333         p->buf_len = FS_PATH_INLINE_SIZE;
334         fs_path_reset(p);
335         return p;
336 }
337 
338 static struct fs_path *fs_path_alloc_reversed(void)
339 {
340         struct fs_path *p;
341 
342         p = fs_path_alloc();
343         if (!p)
344                 return NULL;
345         p->reversed = 1;
346         fs_path_reset(p);
347         return p;
348 }
349 
350 static void fs_path_free(struct fs_path *p)
351 {
352         if (!p)
353                 return;
354         if (p->buf != p->inline_buf)
355                 kfree(p->buf);
356         kfree(p);
357 }
358 
359 static int fs_path_len(struct fs_path *p)
360 {
361         return p->end - p->start;
362 }
363 
364 static int fs_path_ensure_buf(struct fs_path *p, int len)
365 {
366         char *tmp_buf;
367         int path_len;
368         int old_buf_len;
369 
370         len++;
371 
372         if (p->buf_len >= len)
373                 return 0;
374 
375         if (len > PATH_MAX) {
376                 WARN_ON(1);
377                 return -ENOMEM;
378         }
379 
380         path_len = p->end - p->start;
381         old_buf_len = p->buf_len;
382 
383         /*
384          * First time the inline_buf does not suffice
385          */
386         if (p->buf == p->inline_buf) {
387                 tmp_buf = kmalloc(len, GFP_KERNEL);
388                 if (tmp_buf)
389                         memcpy(tmp_buf, p->buf, old_buf_len);
390         } else {
391                 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
392         }
393         if (!tmp_buf)
394                 return -ENOMEM;
395         p->buf = tmp_buf;
396         /*
397          * The real size of the buffer is bigger, this will let the fast path
398          * happen most of the time
399          */
400         p->buf_len = ksize(p->buf);
401 
402         if (p->reversed) {
403                 tmp_buf = p->buf + old_buf_len - path_len - 1;
404                 p->end = p->buf + p->buf_len - 1;
405                 p->start = p->end - path_len;
406                 memmove(p->start, tmp_buf, path_len + 1);
407         } else {
408                 p->start = p->buf;
409                 p->end = p->start + path_len;
410         }
411         return 0;
412 }
413 
414 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
415                                    char **prepared)
416 {
417         int ret;
418         int new_len;
419 
420         new_len = p->end - p->start + name_len;
421         if (p->start != p->end)
422                 new_len++;
423         ret = fs_path_ensure_buf(p, new_len);
424         if (ret < 0)
425                 goto out;
426 
427         if (p->reversed) {
428                 if (p->start != p->end)
429                         *--p->start = '/';
430                 p->start -= name_len;
431                 *prepared = p->start;
432         } else {
433                 if (p->start != p->end)
434                         *p->end++ = '/';
435                 *prepared = p->end;
436                 p->end += name_len;
437                 *p->end = 0;
438         }
439 
440 out:
441         return ret;
442 }
443 
444 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
445 {
446         int ret;
447         char *prepared;
448 
449         ret = fs_path_prepare_for_add(p, name_len, &prepared);
450         if (ret < 0)
451                 goto out;
452         memcpy(prepared, name, name_len);
453 
454 out:
455         return ret;
456 }
457 
458 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
459 {
460         int ret;
461         char *prepared;
462 
463         ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
464         if (ret < 0)
465                 goto out;
466         memcpy(prepared, p2->start, p2->end - p2->start);
467 
468 out:
469         return ret;
470 }
471 
472 static int fs_path_add_from_extent_buffer(struct fs_path *p,
473                                           struct extent_buffer *eb,
474                                           unsigned long off, int len)
475 {
476         int ret;
477         char *prepared;
478 
479         ret = fs_path_prepare_for_add(p, len, &prepared);
480         if (ret < 0)
481                 goto out;
482 
483         read_extent_buffer(eb, prepared, off, len);
484 
485 out:
486         return ret;
487 }
488 
489 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
490 {
491         int ret;
492 
493         p->reversed = from->reversed;
494         fs_path_reset(p);
495 
496         ret = fs_path_add_path(p, from);
497 
498         return ret;
499 }
500 
501 
502 static void fs_path_unreverse(struct fs_path *p)
503 {
504         char *tmp;
505         int len;
506 
507         if (!p->reversed)
508                 return;
509 
510         tmp = p->start;
511         len = p->end - p->start;
512         p->start = p->buf;
513         p->end = p->start + len;
514         memmove(p->start, tmp, len + 1);
515         p->reversed = 0;
516 }
517 
518 static struct btrfs_path *alloc_path_for_send(void)
519 {
520         struct btrfs_path *path;
521 
522         path = btrfs_alloc_path();
523         if (!path)
524                 return NULL;
525         path->search_commit_root = 1;
526         path->skip_locking = 1;
527         path->need_commit_sem = 1;
528         return path;
529 }
530 
531 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
532 {
533         int ret;
534         u32 pos = 0;
535 
536         while (pos < len) {
537                 ret = kernel_write(filp, buf + pos, len - pos, off);
538                 /* TODO handle that correctly */
539                 /*if (ret == -ERESTARTSYS) {
540                         continue;
541                 }*/
542                 if (ret < 0)
543                         return ret;
544                 if (ret == 0) {
545                         return -EIO;
546                 }
547                 pos += ret;
548         }
549 
550         return 0;
551 }
552 
553 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
554 {
555         struct btrfs_tlv_header *hdr;
556         int total_len = sizeof(*hdr) + len;
557         int left = sctx->send_max_size - sctx->send_size;
558 
559         if (unlikely(left < total_len))
560                 return -EOVERFLOW;
561 
562         hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
563         hdr->tlv_type = cpu_to_le16(attr);
564         hdr->tlv_len = cpu_to_le16(len);
565         memcpy(hdr + 1, data, len);
566         sctx->send_size += total_len;
567 
568         return 0;
569 }
570 
571 #define TLV_PUT_DEFINE_INT(bits) \
572         static int tlv_put_u##bits(struct send_ctx *sctx,               \
573                         u##bits attr, u##bits value)                    \
574         {                                                               \
575                 __le##bits __tmp = cpu_to_le##bits(value);              \
576                 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));      \
577         }
578 
579 TLV_PUT_DEFINE_INT(64)
580 
581 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
582                           const char *str, int len)
583 {
584         if (len == -1)
585                 len = strlen(str);
586         return tlv_put(sctx, attr, str, len);
587 }
588 
589 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
590                         const u8 *uuid)
591 {
592         return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
593 }
594 
595 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
596                                   struct extent_buffer *eb,
597                                   struct btrfs_timespec *ts)
598 {
599         struct btrfs_timespec bts;
600         read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
601         return tlv_put(sctx, attr, &bts, sizeof(bts));
602 }
603 
604 
605 #define TLV_PUT(sctx, attrtype, data, attrlen) \
606         do { \
607                 ret = tlv_put(sctx, attrtype, data, attrlen); \
608                 if (ret < 0) \
609                         goto tlv_put_failure; \
610         } while (0)
611 
612 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
613         do { \
614                 ret = tlv_put_u##bits(sctx, attrtype, value); \
615                 if (ret < 0) \
616                         goto tlv_put_failure; \
617         } while (0)
618 
619 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
620 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
621 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
622 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
623 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
624         do { \
625                 ret = tlv_put_string(sctx, attrtype, str, len); \
626                 if (ret < 0) \
627                         goto tlv_put_failure; \
628         } while (0)
629 #define TLV_PUT_PATH(sctx, attrtype, p) \
630         do { \
631                 ret = tlv_put_string(sctx, attrtype, p->start, \
632                         p->end - p->start); \
633                 if (ret < 0) \
634                         goto tlv_put_failure; \
635         } while(0)
636 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
637         do { \
638                 ret = tlv_put_uuid(sctx, attrtype, uuid); \
639                 if (ret < 0) \
640                         goto tlv_put_failure; \
641         } while (0)
642 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
643         do { \
644                 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
645                 if (ret < 0) \
646                         goto tlv_put_failure; \
647         } while (0)
648 
649 static int send_header(struct send_ctx *sctx)
650 {
651         struct btrfs_stream_header hdr;
652 
653         strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
654         hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
655 
656         return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
657                                         &sctx->send_off);
658 }
659 
660 /*
661  * For each command/item we want to send to userspace, we call this function.
662  */
663 static int begin_cmd(struct send_ctx *sctx, int cmd)
664 {
665         struct btrfs_cmd_header *hdr;
666 
667         if (WARN_ON(!sctx->send_buf))
668                 return -EINVAL;
669 
670         BUG_ON(sctx->send_size);
671 
672         sctx->send_size += sizeof(*hdr);
673         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
674         hdr->cmd = cpu_to_le16(cmd);
675 
676         return 0;
677 }
678 
679 static int send_cmd(struct send_ctx *sctx)
680 {
681         int ret;
682         struct btrfs_cmd_header *hdr;
683         u32 crc;
684 
685         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
686         hdr->len = cpu_to_le32(sctx->send_size - sizeof(*hdr));
687         hdr->crc = 0;
688 
689         crc = crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
690         hdr->crc = cpu_to_le32(crc);
691 
692         ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
693                                         &sctx->send_off);
694 
695         sctx->total_send_size += sctx->send_size;
696         sctx->cmd_send_size[le16_to_cpu(hdr->cmd)] += sctx->send_size;
697         sctx->send_size = 0;
698 
699         return ret;
700 }
701 
702 /*
703  * Sends a move instruction to user space
704  */
705 static int send_rename(struct send_ctx *sctx,
706                      struct fs_path *from, struct fs_path *to)
707 {
708         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
709         int ret;
710 
711         btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
712 
713         ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
714         if (ret < 0)
715                 goto out;
716 
717         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
718         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
719 
720         ret = send_cmd(sctx);
721 
722 tlv_put_failure:
723 out:
724         return ret;
725 }
726 
727 /*
728  * Sends a link instruction to user space
729  */
730 static int send_link(struct send_ctx *sctx,
731                      struct fs_path *path, struct fs_path *lnk)
732 {
733         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
734         int ret;
735 
736         btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
737 
738         ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
739         if (ret < 0)
740                 goto out;
741 
742         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
743         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
744 
745         ret = send_cmd(sctx);
746 
747 tlv_put_failure:
748 out:
749         return ret;
750 }
751 
752 /*
753  * Sends an unlink instruction to user space
754  */
755 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
756 {
757         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
758         int ret;
759 
760         btrfs_debug(fs_info, "send_unlink %s", path->start);
761 
762         ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
763         if (ret < 0)
764                 goto out;
765 
766         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
767 
768         ret = send_cmd(sctx);
769 
770 tlv_put_failure:
771 out:
772         return ret;
773 }
774 
775 /*
776  * Sends a rmdir instruction to user space
777  */
778 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
779 {
780         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
781         int ret;
782 
783         btrfs_debug(fs_info, "send_rmdir %s", path->start);
784 
785         ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
786         if (ret < 0)
787                 goto out;
788 
789         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
790 
791         ret = send_cmd(sctx);
792 
793 tlv_put_failure:
794 out:
795         return ret;
796 }
797 
798 /*
799  * Helper function to retrieve some fields from an inode item.
800  */
801 static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
802                           u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
803                           u64 *gid, u64 *rdev)
804 {
805         int ret;
806         struct btrfs_inode_item *ii;
807         struct btrfs_key key;
808 
809         key.objectid = ino;
810         key.type = BTRFS_INODE_ITEM_KEY;
811         key.offset = 0;
812         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
813         if (ret) {
814                 if (ret > 0)
815                         ret = -ENOENT;
816                 return ret;
817         }
818 
819         ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
820                         struct btrfs_inode_item);
821         if (size)
822                 *size = btrfs_inode_size(path->nodes[0], ii);
823         if (gen)
824                 *gen = btrfs_inode_generation(path->nodes[0], ii);
825         if (mode)
826                 *mode = btrfs_inode_mode(path->nodes[0], ii);
827         if (uid)
828                 *uid = btrfs_inode_uid(path->nodes[0], ii);
829         if (gid)
830                 *gid = btrfs_inode_gid(path->nodes[0], ii);
831         if (rdev)
832                 *rdev = btrfs_inode_rdev(path->nodes[0], ii);
833 
834         return ret;
835 }
836 
837 static int get_inode_info(struct btrfs_root *root,
838                           u64 ino, u64 *size, u64 *gen,
839                           u64 *mode, u64 *uid, u64 *gid,
840                           u64 *rdev)
841 {
842         struct btrfs_path *path;
843         int ret;
844 
845         path = alloc_path_for_send();
846         if (!path)
847                 return -ENOMEM;
848         ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
849                                rdev);
850         btrfs_free_path(path);
851         return ret;
852 }
853 
854 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
855                                    struct fs_path *p,
856                                    void *ctx);
857 
858 /*
859  * Helper function to iterate the entries in ONE btrfs_inode_ref or
860  * btrfs_inode_extref.
861  * The iterate callback may return a non zero value to stop iteration. This can
862  * be a negative value for error codes or 1 to simply stop it.
863  *
864  * path must point to the INODE_REF or INODE_EXTREF when called.
865  */
866 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
867                              struct btrfs_key *found_key, int resolve,
868                              iterate_inode_ref_t iterate, void *ctx)
869 {
870         struct extent_buffer *eb = path->nodes[0];
871         struct btrfs_item *item;
872         struct btrfs_inode_ref *iref;
873         struct btrfs_inode_extref *extref;
874         struct btrfs_path *tmp_path;
875         struct fs_path *p;
876         u32 cur = 0;
877         u32 total;
878         int slot = path->slots[0];
879         u32 name_len;
880         char *start;
881         int ret = 0;
882         int num = 0;
883         int index;
884         u64 dir;
885         unsigned long name_off;
886         unsigned long elem_size;
887         unsigned long ptr;
888 
889         p = fs_path_alloc_reversed();
890         if (!p)
891                 return -ENOMEM;
892 
893         tmp_path = alloc_path_for_send();
894         if (!tmp_path) {
895                 fs_path_free(p);
896                 return -ENOMEM;
897         }
898 
899 
900         if (found_key->type == BTRFS_INODE_REF_KEY) {
901                 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
902                                                     struct btrfs_inode_ref);
903                 item = btrfs_item_nr(slot);
904                 total = btrfs_item_size(eb, item);
905                 elem_size = sizeof(*iref);
906         } else {
907                 ptr = btrfs_item_ptr_offset(eb, slot);
908                 total = btrfs_item_size_nr(eb, slot);
909                 elem_size = sizeof(*extref);
910         }
911 
912         while (cur < total) {
913                 fs_path_reset(p);
914 
915                 if (found_key->type == BTRFS_INODE_REF_KEY) {
916                         iref = (struct btrfs_inode_ref *)(ptr + cur);
917                         name_len = btrfs_inode_ref_name_len(eb, iref);
918                         name_off = (unsigned long)(iref + 1);
919                         index = btrfs_inode_ref_index(eb, iref);
920                         dir = found_key->offset;
921                 } else {
922                         extref = (struct btrfs_inode_extref *)(ptr + cur);
923                         name_len = btrfs_inode_extref_name_len(eb, extref);
924                         name_off = (unsigned long)&extref->name;
925                         index = btrfs_inode_extref_index(eb, extref);
926                         dir = btrfs_inode_extref_parent(eb, extref);
927                 }
928 
929                 if (resolve) {
930                         start = btrfs_ref_to_path(root, tmp_path, name_len,
931                                                   name_off, eb, dir,
932                                                   p->buf, p->buf_len);
933                         if (IS_ERR(start)) {
934                                 ret = PTR_ERR(start);
935                                 goto out;
936                         }
937                         if (start < p->buf) {
938                                 /* overflow , try again with larger buffer */
939                                 ret = fs_path_ensure_buf(p,
940                                                 p->buf_len + p->buf - start);
941                                 if (ret < 0)
942                                         goto out;
943                                 start = btrfs_ref_to_path(root, tmp_path,
944                                                           name_len, name_off,
945                                                           eb, dir,
946                                                           p->buf, p->buf_len);
947                                 if (IS_ERR(start)) {
948                                         ret = PTR_ERR(start);
949                                         goto out;
950                                 }
951                                 BUG_ON(start < p->buf);
952                         }
953                         p->start = start;
954                 } else {
955                         ret = fs_path_add_from_extent_buffer(p, eb, name_off,
956                                                              name_len);
957                         if (ret < 0)
958                                 goto out;
959                 }
960 
961                 cur += elem_size + name_len;
962                 ret = iterate(num, dir, index, p, ctx);
963                 if (ret)
964                         goto out;
965                 num++;
966         }
967 
968 out:
969         btrfs_free_path(tmp_path);
970         fs_path_free(p);
971         return ret;
972 }
973 
974 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
975                                   const char *name, int name_len,
976                                   const char *data, int data_len,
977                                   u8 type, void *ctx);
978 
979 /*
980  * Helper function to iterate the entries in ONE btrfs_dir_item.
981  * The iterate callback may return a non zero value to stop iteration. This can
982  * be a negative value for error codes or 1 to simply stop it.
983  *
984  * path must point to the dir item when called.
985  */
986 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
987                             iterate_dir_item_t iterate, void *ctx)
988 {
989         int ret = 0;
990         struct extent_buffer *eb;
991         struct btrfs_item *item;
992         struct btrfs_dir_item *di;
993         struct btrfs_key di_key;
994         char *buf = NULL;
995         int buf_len;
996         u32 name_len;
997         u32 data_len;
998         u32 cur;
999         u32 len;
1000         u32 total;
1001         int slot;
1002         int num;
1003         u8 type;
1004 
1005         /*
1006          * Start with a small buffer (1 page). If later we end up needing more
1007          * space, which can happen for xattrs on a fs with a leaf size greater
1008          * then the page size, attempt to increase the buffer. Typically xattr
1009          * values are small.
1010          */
1011         buf_len = PATH_MAX;
1012         buf = kmalloc(buf_len, GFP_KERNEL);
1013         if (!buf) {
1014                 ret = -ENOMEM;
1015                 goto out;
1016         }
1017 
1018         eb = path->nodes[0];
1019         slot = path->slots[0];
1020         item = btrfs_item_nr(slot);
1021         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1022         cur = 0;
1023         len = 0;
1024         total = btrfs_item_size(eb, item);
1025 
1026         num = 0;
1027         while (cur < total) {
1028                 name_len = btrfs_dir_name_len(eb, di);
1029                 data_len = btrfs_dir_data_len(eb, di);
1030                 type = btrfs_dir_type(eb, di);
1031                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1032 
1033                 if (type == BTRFS_FT_XATTR) {
1034                         if (name_len > XATTR_NAME_MAX) {
1035                                 ret = -ENAMETOOLONG;
1036                                 goto out;
1037                         }
1038                         if (name_len + data_len >
1039                                         BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1040                                 ret = -E2BIG;
1041                                 goto out;
1042                         }
1043                 } else {
1044                         /*
1045                          * Path too long
1046                          */
1047                         if (name_len + data_len > PATH_MAX) {
1048                                 ret = -ENAMETOOLONG;
1049                                 goto out;
1050                         }
1051                 }
1052 
1053                 if (name_len + data_len > buf_len) {
1054                         buf_len = name_len + data_len;
1055                         if (is_vmalloc_addr(buf)) {
1056                                 vfree(buf);
1057                                 buf = NULL;
1058                         } else {
1059                                 char *tmp = krealloc(buf, buf_len,
1060                                                 GFP_KERNEL | __GFP_NOWARN);
1061 
1062                                 if (!tmp)
1063                                         kfree(buf);
1064                                 buf = tmp;
1065                         }
1066                         if (!buf) {
1067                                 buf = kvmalloc(buf_len, GFP_KERNEL);
1068                                 if (!buf) {
1069                                         ret = -ENOMEM;
1070                                         goto out;
1071                                 }
1072                         }
1073                 }
1074 
1075                 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1076                                 name_len + data_len);
1077 
1078                 len = sizeof(*di) + name_len + data_len;
1079                 di = (struct btrfs_dir_item *)((char *)di + len);
1080                 cur += len;
1081 
1082                 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1083                                 data_len, type, ctx);
1084                 if (ret < 0)
1085                         goto out;
1086                 if (ret) {
1087                         ret = 0;
1088                         goto out;
1089                 }
1090 
1091                 num++;
1092         }
1093 
1094 out:
1095         kvfree(buf);
1096         return ret;
1097 }
1098 
1099 static int __copy_first_ref(int num, u64 dir, int index,
1100                             struct fs_path *p, void *ctx)
1101 {
1102         int ret;
1103         struct fs_path *pt = ctx;
1104 
1105         ret = fs_path_copy(pt, p);
1106         if (ret < 0)
1107                 return ret;
1108 
1109         /* we want the first only */
1110         return 1;
1111 }
1112 
1113 /*
1114  * Retrieve the first path of an inode. If an inode has more then one
1115  * ref/hardlink, this is ignored.
1116  */
1117 static int get_inode_path(struct btrfs_root *root,
1118                           u64 ino, struct fs_path *path)
1119 {
1120         int ret;
1121         struct btrfs_key key, found_key;
1122         struct btrfs_path *p;
1123 
1124         p = alloc_path_for_send();
1125         if (!p)
1126                 return -ENOMEM;
1127 
1128         fs_path_reset(path);
1129 
1130         key.objectid = ino;
1131         key.type = BTRFS_INODE_REF_KEY;
1132         key.offset = 0;
1133 
1134         ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1135         if (ret < 0)
1136                 goto out;
1137         if (ret) {
1138                 ret = 1;
1139                 goto out;
1140         }
1141         btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1142         if (found_key.objectid != ino ||
1143             (found_key.type != BTRFS_INODE_REF_KEY &&
1144              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1145                 ret = -ENOENT;
1146                 goto out;
1147         }
1148 
1149         ret = iterate_inode_ref(root, p, &found_key, 1,
1150                                 __copy_first_ref, path);
1151         if (ret < 0)
1152                 goto out;
1153         ret = 0;
1154 
1155 out:
1156         btrfs_free_path(p);
1157         return ret;
1158 }
1159 
1160 struct backref_ctx {
1161         struct send_ctx *sctx;
1162 
1163         /* number of total found references */
1164         u64 found;
1165 
1166         /*
1167          * used for clones found in send_root. clones found behind cur_objectid
1168          * and cur_offset are not considered as allowed clones.
1169          */
1170         u64 cur_objectid;
1171         u64 cur_offset;
1172 
1173         /* may be truncated in case it's the last extent in a file */
1174         u64 extent_len;
1175 
1176         /* data offset in the file extent item */
1177         u64 data_offset;
1178 
1179         /* Just to check for bugs in backref resolving */
1180         int found_itself;
1181 };
1182 
1183 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1184 {
1185         u64 root = (u64)(uintptr_t)key;
1186         struct clone_root *cr = (struct clone_root *)elt;
1187 
1188         if (root < cr->root->root_key.objectid)
1189                 return -1;
1190         if (root > cr->root->root_key.objectid)
1191                 return 1;
1192         return 0;
1193 }
1194 
1195 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1196 {
1197         struct clone_root *cr1 = (struct clone_root *)e1;
1198         struct clone_root *cr2 = (struct clone_root *)e2;
1199 
1200         if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1201                 return -1;
1202         if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1203                 return 1;
1204         return 0;
1205 }
1206 
1207 /*
1208  * Called for every backref that is found for the current extent.
1209  * Results are collected in sctx->clone_roots->ino/offset/found_refs
1210  */
1211 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1212 {
1213         struct backref_ctx *bctx = ctx_;
1214         struct clone_root *found;
1215 
1216         /* First check if the root is in the list of accepted clone sources */
1217         found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1218                         bctx->sctx->clone_roots_cnt,
1219                         sizeof(struct clone_root),
1220                         __clone_root_cmp_bsearch);
1221         if (!found)
1222                 return 0;
1223 
1224         if (found->root == bctx->sctx->send_root &&
1225             ino == bctx->cur_objectid &&
1226             offset == bctx->cur_offset) {
1227                 bctx->found_itself = 1;
1228         }
1229 
1230         /*
1231          * Make sure we don't consider clones from send_root that are
1232          * behind the current inode/offset.
1233          */
1234         if (found->root == bctx->sctx->send_root) {
1235                 /*
1236                  * TODO for the moment we don't accept clones from the inode
1237                  * that is currently send. We may change this when
1238                  * BTRFS_IOC_CLONE_RANGE supports cloning from and to the same
1239                  * file.
1240                  */
1241                 if (ino >= bctx->cur_objectid)
1242                         return 0;
1243         }
1244 
1245         bctx->found++;
1246         found->found_refs++;
1247         if (ino < found->ino) {
1248                 found->ino = ino;
1249                 found->offset = offset;
1250         } else if (found->ino == ino) {
1251                 /*
1252                  * same extent found more then once in the same file.
1253                  */
1254                 if (found->offset > offset + bctx->extent_len)
1255                         found->offset = offset;
1256         }
1257 
1258         return 0;
1259 }
1260 
1261 /*
1262  * Given an inode, offset and extent item, it finds a good clone for a clone
1263  * instruction. Returns -ENOENT when none could be found. The function makes
1264  * sure that the returned clone is usable at the point where sending is at the
1265  * moment. This means, that no clones are accepted which lie behind the current
1266  * inode+offset.
1267  *
1268  * path must point to the extent item when called.
1269  */
1270 static int find_extent_clone(struct send_ctx *sctx,
1271                              struct btrfs_path *path,
1272                              u64 ino, u64 data_offset,
1273                              u64 ino_size,
1274                              struct clone_root **found)
1275 {
1276         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1277         int ret;
1278         int extent_type;
1279         u64 logical;
1280         u64 disk_byte;
1281         u64 num_bytes;
1282         u64 extent_item_pos;
1283         u64 flags = 0;
1284         struct btrfs_file_extent_item *fi;
1285         struct extent_buffer *eb = path->nodes[0];
1286         struct backref_ctx *backref_ctx = NULL;
1287         struct clone_root *cur_clone_root;
1288         struct btrfs_key found_key;
1289         struct btrfs_path *tmp_path;
1290         int compressed;
1291         u32 i;
1292 
1293         tmp_path = alloc_path_for_send();
1294         if (!tmp_path)
1295                 return -ENOMEM;
1296 
1297         /* We only use this path under the commit sem */
1298         tmp_path->need_commit_sem = 0;
1299 
1300         backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1301         if (!backref_ctx) {
1302                 ret = -ENOMEM;
1303                 goto out;
1304         }
1305 
1306         if (data_offset >= ino_size) {
1307                 /*
1308                  * There may be extents that lie behind the file's size.
1309                  * I at least had this in combination with snapshotting while
1310                  * writing large files.
1311                  */
1312                 ret = 0;
1313                 goto out;
1314         }
1315 
1316         fi = btrfs_item_ptr(eb, path->slots[0],
1317                         struct btrfs_file_extent_item);
1318         extent_type = btrfs_file_extent_type(eb, fi);
1319         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1320                 ret = -ENOENT;
1321                 goto out;
1322         }
1323         compressed = btrfs_file_extent_compression(eb, fi);
1324 
1325         num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1326         disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1327         if (disk_byte == 0) {
1328                 ret = -ENOENT;
1329                 goto out;
1330         }
1331         logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1332 
1333         down_read(&fs_info->commit_root_sem);
1334         ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1335                                   &found_key, &flags);
1336         up_read(&fs_info->commit_root_sem);
1337         btrfs_release_path(tmp_path);
1338 
1339         if (ret < 0)
1340                 goto out;
1341         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1342                 ret = -EIO;
1343                 goto out;
1344         }
1345 
1346         /*
1347          * Setup the clone roots.
1348          */
1349         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1350                 cur_clone_root = sctx->clone_roots + i;
1351                 cur_clone_root->ino = (u64)-1;
1352                 cur_clone_root->offset = 0;
1353                 cur_clone_root->found_refs = 0;
1354         }
1355 
1356         backref_ctx->sctx = sctx;
1357         backref_ctx->found = 0;
1358         backref_ctx->cur_objectid = ino;
1359         backref_ctx->cur_offset = data_offset;
1360         backref_ctx->found_itself = 0;
1361         backref_ctx->extent_len = num_bytes;
1362         /*
1363          * For non-compressed extents iterate_extent_inodes() gives us extent
1364          * offsets that already take into account the data offset, but not for
1365          * compressed extents, since the offset is logical and not relative to
1366          * the physical extent locations. We must take this into account to
1367          * avoid sending clone offsets that go beyond the source file's size,
1368          * which would result in the clone ioctl failing with -EINVAL on the
1369          * receiving end.
1370          */
1371         if (compressed == BTRFS_COMPRESS_NONE)
1372                 backref_ctx->data_offset = 0;
1373         else
1374                 backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1375 
1376         /*
1377          * The last extent of a file may be too large due to page alignment.
1378          * We need to adjust extent_len in this case so that the checks in
1379          * __iterate_backrefs work.
1380          */
1381         if (data_offset + num_bytes >= ino_size)
1382                 backref_ctx->extent_len = ino_size - data_offset;
1383 
1384         /*
1385          * Now collect all backrefs.
1386          */
1387         if (compressed == BTRFS_COMPRESS_NONE)
1388                 extent_item_pos = logical - found_key.objectid;
1389         else
1390                 extent_item_pos = 0;
1391         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1392                                     extent_item_pos, 1, __iterate_backrefs,
1393                                     backref_ctx, false);
1394 
1395         if (ret < 0)
1396                 goto out;
1397 
1398         if (!backref_ctx->found_itself) {
1399                 /* found a bug in backref code? */
1400                 ret = -EIO;
1401                 btrfs_err(fs_info,
1402                           "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1403                           ino, data_offset, disk_byte, found_key.objectid);
1404                 goto out;
1405         }
1406 
1407         btrfs_debug(fs_info,
1408                     "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1409                     data_offset, ino, num_bytes, logical);
1410 
1411         if (!backref_ctx->found)
1412                 btrfs_debug(fs_info, "no clones found");
1413 
1414         cur_clone_root = NULL;
1415         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1416                 if (sctx->clone_roots[i].found_refs) {
1417                         if (!cur_clone_root)
1418                                 cur_clone_root = sctx->clone_roots + i;
1419                         else if (sctx->clone_roots[i].root == sctx->send_root)
1420                                 /* prefer clones from send_root over others */
1421                                 cur_clone_root = sctx->clone_roots + i;
1422                 }
1423 
1424         }
1425 
1426         if (cur_clone_root) {
1427                 *found = cur_clone_root;
1428                 ret = 0;
1429         } else {
1430                 ret = -ENOENT;
1431         }
1432 
1433 out:
1434         btrfs_free_path(tmp_path);
1435         kfree(backref_ctx);
1436         return ret;
1437 }
1438 
1439 static int read_symlink(struct btrfs_root *root,
1440                         u64 ino,
1441                         struct fs_path *dest)
1442 {
1443         int ret;
1444         struct btrfs_path *path;
1445         struct btrfs_key key;
1446         struct btrfs_file_extent_item *ei;
1447         u8 type;
1448         u8 compression;
1449         unsigned long off;
1450         int len;
1451 
1452         path = alloc_path_for_send();
1453         if (!path)
1454                 return -ENOMEM;
1455 
1456         key.objectid = ino;
1457         key.type = BTRFS_EXTENT_DATA_KEY;
1458         key.offset = 0;
1459         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1460         if (ret < 0)
1461                 goto out;
1462         if (ret) {
1463                 /*
1464                  * An empty symlink inode. Can happen in rare error paths when
1465                  * creating a symlink (transaction committed before the inode
1466                  * eviction handler removed the symlink inode items and a crash
1467                  * happened in between or the subvol was snapshoted in between).
1468                  * Print an informative message to dmesg/syslog so that the user
1469                  * can delete the symlink.
1470                  */
1471                 btrfs_err(root->fs_info,
1472                           "Found empty symlink inode %llu at root %llu",
1473                           ino, root->root_key.objectid);
1474                 ret = -EIO;
1475                 goto out;
1476         }
1477 
1478         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1479                         struct btrfs_file_extent_item);
1480         type = btrfs_file_extent_type(path->nodes[0], ei);
1481         compression = btrfs_file_extent_compression(path->nodes[0], ei);
1482         BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1483         BUG_ON(compression);
1484 
1485         off = btrfs_file_extent_inline_start(ei);
1486         len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1487 
1488         ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1489 
1490 out:
1491         btrfs_free_path(path);
1492         return ret;
1493 }
1494 
1495 /*
1496  * Helper function to generate a file name that is unique in the root of
1497  * send_root and parent_root. This is used to generate names for orphan inodes.
1498  */
1499 static int gen_unique_name(struct send_ctx *sctx,
1500                            u64 ino, u64 gen,
1501                            struct fs_path *dest)
1502 {
1503         int ret = 0;
1504         struct btrfs_path *path;
1505         struct btrfs_dir_item *di;
1506         char tmp[64];
1507         int len;
1508         u64 idx = 0;
1509 
1510         path = alloc_path_for_send();
1511         if (!path)
1512                 return -ENOMEM;
1513 
1514         while (1) {
1515                 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1516                                 ino, gen, idx);
1517                 ASSERT(len < sizeof(tmp));
1518 
1519                 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1520                                 path, BTRFS_FIRST_FREE_OBJECTID,
1521                                 tmp, strlen(tmp), 0);
1522                 btrfs_release_path(path);
1523                 if (IS_ERR(di)) {
1524                         ret = PTR_ERR(di);
1525                         goto out;
1526                 }
1527                 if (di) {
1528                         /* not unique, try again */
1529                         idx++;
1530                         continue;
1531                 }
1532 
1533                 if (!sctx->parent_root) {
1534                         /* unique */
1535                         ret = 0;
1536                         break;
1537                 }
1538 
1539                 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1540                                 path, BTRFS_FIRST_FREE_OBJECTID,
1541                                 tmp, strlen(tmp), 0);
1542                 btrfs_release_path(path);
1543                 if (IS_ERR(di)) {
1544                         ret = PTR_ERR(di);
1545                         goto out;
1546                 }
1547                 if (di) {
1548                         /* not unique, try again */
1549                         idx++;
1550                         continue;
1551                 }
1552                 /* unique */
1553                 break;
1554         }
1555 
1556         ret = fs_path_add(dest, tmp, strlen(tmp));
1557 
1558 out:
1559         btrfs_free_path(path);
1560         return ret;
1561 }
1562 
1563 enum inode_state {
1564         inode_state_no_change,
1565         inode_state_will_create,
1566         inode_state_did_create,
1567         inode_state_will_delete,
1568         inode_state_did_delete,
1569 };
1570 
1571 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1572 {
1573         int ret;
1574         int left_ret;
1575         int right_ret;
1576         u64 left_gen;
1577         u64 right_gen;
1578 
1579         ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1580                         NULL, NULL);
1581         if (ret < 0 && ret != -ENOENT)
1582                 goto out;
1583         left_ret = ret;
1584 
1585         if (!sctx->parent_root) {
1586                 right_ret = -ENOENT;
1587         } else {
1588                 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1589                                 NULL, NULL, NULL, NULL);
1590                 if (ret < 0 && ret != -ENOENT)
1591                         goto out;
1592                 right_ret = ret;
1593         }
1594 
1595         if (!left_ret && !right_ret) {
1596                 if (left_gen == gen && right_gen == gen) {
1597                         ret = inode_state_no_change;
1598                 } else if (left_gen == gen) {
1599                         if (ino < sctx->send_progress)
1600                                 ret = inode_state_did_create;
1601                         else
1602                                 ret = inode_state_will_create;
1603                 } else if (right_gen == gen) {
1604                         if (ino < sctx->send_progress)
1605                                 ret = inode_state_did_delete;
1606                         else
1607                                 ret = inode_state_will_delete;
1608                 } else  {
1609                         ret = -ENOENT;
1610                 }
1611         } else if (!left_ret) {
1612                 if (left_gen == gen) {
1613                         if (ino < sctx->send_progress)
1614                                 ret = inode_state_did_create;
1615                         else
1616                                 ret = inode_state_will_create;
1617                 } else {
1618                         ret = -ENOENT;
1619                 }
1620         } else if (!right_ret) {
1621                 if (right_gen == gen) {
1622                         if (ino < sctx->send_progress)
1623                                 ret = inode_state_did_delete;
1624                         else
1625                                 ret = inode_state_will_delete;
1626                 } else {
1627                         ret = -ENOENT;
1628                 }
1629         } else {
1630                 ret = -ENOENT;
1631         }
1632 
1633 out:
1634         return ret;
1635 }
1636 
1637 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1638 {
1639         int ret;
1640 
1641         if (ino == BTRFS_FIRST_FREE_OBJECTID)
1642                 return 1;
1643 
1644         ret = get_cur_inode_state(sctx, ino, gen);
1645         if (ret < 0)
1646                 goto out;
1647 
1648         if (ret == inode_state_no_change ||
1649             ret == inode_state_did_create ||
1650             ret == inode_state_will_delete)
1651                 ret = 1;
1652         else
1653                 ret = 0;
1654 
1655 out:
1656         return ret;
1657 }
1658 
1659 /*
1660  * Helper function to lookup a dir item in a dir.
1661  */
1662 static int lookup_dir_item_inode(struct btrfs_root *root,
1663                                  u64 dir, const char *name, int name_len,
1664                                  u64 *found_inode,
1665                                  u8 *found_type)
1666 {
1667         int ret = 0;
1668         struct btrfs_dir_item *di;
1669         struct btrfs_key key;
1670         struct btrfs_path *path;
1671 
1672         path = alloc_path_for_send();
1673         if (!path)
1674                 return -ENOMEM;
1675 
1676         di = btrfs_lookup_dir_item(NULL, root, path,
1677                         dir, name, name_len, 0);
1678         if (IS_ERR_OR_NULL(di)) {
1679                 ret = di ? PTR_ERR(di) : -ENOENT;
1680                 goto out;
1681         }
1682         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1683         if (key.type == BTRFS_ROOT_ITEM_KEY) {
1684                 ret = -ENOENT;
1685                 goto out;
1686         }
1687         *found_inode = key.objectid;
1688         *found_type = btrfs_dir_type(path->nodes[0], di);
1689 
1690 out:
1691         btrfs_free_path(path);
1692         return ret;
1693 }
1694 
1695 /*
1696  * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1697  * generation of the parent dir and the name of the dir entry.
1698  */
1699 static int get_first_ref(struct btrfs_root *root, u64 ino,
1700                          u64 *dir, u64 *dir_gen, struct fs_path *name)
1701 {
1702         int ret;
1703         struct btrfs_key key;
1704         struct btrfs_key found_key;
1705         struct btrfs_path *path;
1706         int len;
1707         u64 parent_dir;
1708 
1709         path = alloc_path_for_send();
1710         if (!path)
1711                 return -ENOMEM;
1712 
1713         key.objectid = ino;
1714         key.type = BTRFS_INODE_REF_KEY;
1715         key.offset = 0;
1716 
1717         ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1718         if (ret < 0)
1719                 goto out;
1720         if (!ret)
1721                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1722                                 path->slots[0]);
1723         if (ret || found_key.objectid != ino ||
1724             (found_key.type != BTRFS_INODE_REF_KEY &&
1725              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1726                 ret = -ENOENT;
1727                 goto out;
1728         }
1729 
1730         if (found_key.type == BTRFS_INODE_REF_KEY) {
1731                 struct btrfs_inode_ref *iref;
1732                 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1733                                       struct btrfs_inode_ref);
1734                 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1735                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1736                                                      (unsigned long)(iref + 1),
1737                                                      len);
1738                 parent_dir = found_key.offset;
1739         } else {
1740                 struct btrfs_inode_extref *extref;
1741                 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1742                                         struct btrfs_inode_extref);
1743                 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1744                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1745                                         (unsigned long)&extref->name, len);
1746                 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1747         }
1748         if (ret < 0)
1749                 goto out;
1750         btrfs_release_path(path);
1751 
1752         if (dir_gen) {
1753                 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1754                                      NULL, NULL, NULL);
1755                 if (ret < 0)
1756                         goto out;
1757         }
1758 
1759         *dir = parent_dir;
1760 
1761 out:
1762         btrfs_free_path(path);
1763         return ret;
1764 }
1765 
1766 static int is_first_ref(struct btrfs_root *root,
1767                         u64 ino, u64 dir,
1768                         const char *name, int name_len)
1769 {
1770         int ret;
1771         struct fs_path *tmp_name;
1772         u64 tmp_dir;
1773 
1774         tmp_name = fs_path_alloc();
1775         if (!tmp_name)
1776                 return -ENOMEM;
1777 
1778         ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1779         if (ret < 0)
1780                 goto out;
1781 
1782         if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1783                 ret = 0;
1784                 goto out;
1785         }
1786 
1787         ret = !memcmp(tmp_name->start, name, name_len);
1788 
1789 out:
1790         fs_path_free(tmp_name);
1791         return ret;
1792 }
1793 
1794 /*
1795  * Used by process_recorded_refs to determine if a new ref would overwrite an
1796  * already existing ref. In case it detects an overwrite, it returns the
1797  * inode/gen in who_ino/who_gen.
1798  * When an overwrite is detected, process_recorded_refs does proper orphanizing
1799  * to make sure later references to the overwritten inode are possible.
1800  * Orphanizing is however only required for the first ref of an inode.
1801  * process_recorded_refs does an additional is_first_ref check to see if
1802  * orphanizing is really required.
1803  */
1804 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1805                               const char *name, int name_len,
1806                               u64 *who_ino, u64 *who_gen, u64 *who_mode)
1807 {
1808         int ret = 0;
1809         u64 gen;
1810         u64 other_inode = 0;
1811         u8 other_type = 0;
1812 
1813         if (!sctx->parent_root)
1814                 goto out;
1815 
1816         ret = is_inode_existent(sctx, dir, dir_gen);
1817         if (ret <= 0)
1818                 goto out;
1819 
1820         /*
1821          * If we have a parent root we need to verify that the parent dir was
1822          * not deleted and then re-created, if it was then we have no overwrite
1823          * and we can just unlink this entry.
1824          */
1825         if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1826                 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1827                                      NULL, NULL, NULL);
1828                 if (ret < 0 && ret != -ENOENT)
1829                         goto out;
1830                 if (ret) {
1831                         ret = 0;
1832                         goto out;
1833                 }
1834                 if (gen != dir_gen)
1835                         goto out;
1836         }
1837 
1838         ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1839                         &other_inode, &other_type);
1840         if (ret < 0 && ret != -ENOENT)
1841                 goto out;
1842         if (ret) {
1843                 ret = 0;
1844                 goto out;
1845         }
1846 
1847         /*
1848          * Check if the overwritten ref was already processed. If yes, the ref
1849          * was already unlinked/moved, so we can safely assume that we will not
1850          * overwrite anything at this point in time.
1851          */
1852         if (other_inode > sctx->send_progress ||
1853             is_waiting_for_move(sctx, other_inode)) {
1854                 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1855                                 who_gen, who_mode, NULL, NULL, NULL);
1856                 if (ret < 0)
1857                         goto out;
1858 
1859                 ret = 1;
1860                 *who_ino = other_inode;
1861         } else {
1862                 ret = 0;
1863         }
1864 
1865 out:
1866         return ret;
1867 }
1868 
1869 /*
1870  * Checks if the ref was overwritten by an already processed inode. This is
1871  * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1872  * thus the orphan name needs be used.
1873  * process_recorded_refs also uses it to avoid unlinking of refs that were
1874  * overwritten.
1875  */
1876 static int did_overwrite_ref(struct send_ctx *sctx,
1877                             u64 dir, u64 dir_gen,
1878                             u64 ino, u64 ino_gen,
1879                             const char *name, int name_len)
1880 {
1881         int ret = 0;
1882         u64 gen;
1883         u64 ow_inode;
1884         u8 other_type;
1885 
1886         if (!sctx->parent_root)
1887                 goto out;
1888 
1889         ret = is_inode_existent(sctx, dir, dir_gen);
1890         if (ret <= 0)
1891                 goto out;
1892 
1893         if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1894                 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1895                                      NULL, NULL, NULL);
1896                 if (ret < 0 && ret != -ENOENT)
1897                         goto out;
1898                 if (ret) {
1899                         ret = 0;
1900                         goto out;
1901                 }
1902                 if (gen != dir_gen)
1903                         goto out;
1904         }
1905 
1906         /* check if the ref was overwritten by another ref */
1907         ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1908                         &ow_inode, &other_type);
1909         if (ret < 0 && ret != -ENOENT)
1910                 goto out;
1911         if (ret) {
1912                 /* was never and will never be overwritten */
1913                 ret = 0;
1914                 goto out;
1915         }
1916 
1917         ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1918                         NULL, NULL);
1919         if (ret < 0)
1920                 goto out;
1921 
1922         if (ow_inode == ino && gen == ino_gen) {
1923                 ret = 0;
1924                 goto out;
1925         }
1926 
1927         /*
1928          * We know that it is or will be overwritten. Check this now.
1929          * The current inode being processed might have been the one that caused
1930          * inode 'ino' to be orphanized, therefore check if ow_inode matches
1931          * the current inode being processed.
1932          */
1933         if ((ow_inode < sctx->send_progress) ||
1934             (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1935              gen == sctx->cur_inode_gen))
1936                 ret = 1;
1937         else
1938                 ret = 0;
1939 
1940 out:
1941         return ret;
1942 }
1943 
1944 /*
1945  * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1946  * that got overwritten. This is used by process_recorded_refs to determine
1947  * if it has to use the path as returned by get_cur_path or the orphan name.
1948  */
1949 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1950 {
1951         int ret = 0;
1952         struct fs_path *name = NULL;
1953         u64 dir;
1954         u64 dir_gen;
1955 
1956         if (!sctx->parent_root)
1957                 goto out;
1958 
1959         name = fs_path_alloc();
1960         if (!name)
1961                 return -ENOMEM;
1962 
1963         ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
1964         if (ret < 0)
1965                 goto out;
1966 
1967         ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
1968                         name->start, fs_path_len(name));
1969 
1970 out:
1971         fs_path_free(name);
1972         return ret;
1973 }
1974 
1975 /*
1976  * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
1977  * so we need to do some special handling in case we have clashes. This function
1978  * takes care of this with the help of name_cache_entry::radix_list.
1979  * In case of error, nce is kfreed.
1980  */
1981 static int name_cache_insert(struct send_ctx *sctx,
1982                              struct name_cache_entry *nce)
1983 {
1984         int ret = 0;
1985         struct list_head *nce_head;
1986 
1987         nce_head = radix_tree_lookup(&sctx->name_cache,
1988                         (unsigned long)nce->ino);
1989         if (!nce_head) {
1990                 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
1991                 if (!nce_head) {
1992                         kfree(nce);
1993                         return -ENOMEM;
1994                 }
1995                 INIT_LIST_HEAD(nce_head);
1996 
1997                 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
1998                 if (ret < 0) {
1999                         kfree(nce_head);
2000                         kfree(nce);
2001                         return ret;
2002                 }
2003         }
2004         list_add_tail(&nce->radix_list, nce_head);
2005         list_add_tail(&nce->list, &sctx->name_cache_list);
2006         sctx->name_cache_size++;
2007 
2008         return ret;
2009 }
2010 
2011 static void name_cache_delete(struct send_ctx *sctx,
2012                               struct name_cache_entry *nce)
2013 {
2014         struct list_head *nce_head;
2015 
2016         nce_head = radix_tree_lookup(&sctx->name_cache,
2017                         (unsigned long)nce->ino);
2018         if (!nce_head) {
2019                 btrfs_err(sctx->send_root->fs_info,
2020               "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2021                         nce->ino, sctx->name_cache_size);
2022         }
2023 
2024         list_del(&nce->radix_list);
2025         list_del(&nce->list);
2026         sctx->name_cache_size--;
2027 
2028         /*
2029          * We may not get to the final release of nce_head if the lookup fails
2030          */
2031         if (nce_head && list_empty(nce_head)) {
2032                 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2033                 kfree(nce_head);
2034         }
2035 }
2036 
2037 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2038                                                     u64 ino, u64 gen)
2039 {
2040         struct list_head *nce_head;
2041         struct name_cache_entry *cur;
2042 
2043         nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2044         if (!nce_head)
2045                 return NULL;
2046 
2047         list_for_each_entry(cur, nce_head, radix_list) {
2048                 if (cur->ino == ino && cur->gen == gen)
2049                         return cur;
2050         }
2051         return NULL;
2052 }
2053 
2054 /*
2055  * Removes the entry from the list and adds it back to the end. This marks the
2056  * entry as recently used so that name_cache_clean_unused does not remove it.
2057  */
2058 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2059 {
2060         list_del(&nce->list);
2061         list_add_tail(&nce->list, &sctx->name_cache_list);
2062 }
2063 
2064 /*
2065  * Remove some entries from the beginning of name_cache_list.
2066  */
2067 static void name_cache_clean_unused(struct send_ctx *sctx)
2068 {
2069         struct name_cache_entry *nce;
2070 
2071         if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2072                 return;
2073 
2074         while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2075                 nce = list_entry(sctx->name_cache_list.next,
2076                                 struct name_cache_entry, list);
2077                 name_cache_delete(sctx, nce);
2078                 kfree(nce);
2079         }
2080 }
2081 
2082 static void name_cache_free(struct send_ctx *sctx)
2083 {
2084         struct name_cache_entry *nce;
2085 
2086         while (!list_empty(&sctx->name_cache_list)) {
2087                 nce = list_entry(sctx->name_cache_list.next,
2088                                 struct name_cache_entry, list);
2089                 name_cache_delete(sctx, nce);
2090                 kfree(nce);
2091         }
2092 }
2093 
2094 /*
2095  * Used by get_cur_path for each ref up to the root.
2096  * Returns 0 if it succeeded.
2097  * Returns 1 if the inode is not existent or got overwritten. In that case, the
2098  * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2099  * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2100  * Returns <0 in case of error.
2101  */
2102 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2103                                      u64 ino, u64 gen,
2104                                      u64 *parent_ino,
2105                                      u64 *parent_gen,
2106                                      struct fs_path *dest)
2107 {
2108         int ret;
2109         int nce_ret;
2110         struct name_cache_entry *nce = NULL;
2111 
2112         /*
2113          * First check if we already did a call to this function with the same
2114          * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2115          * return the cached result.
2116          */
2117         nce = name_cache_search(sctx, ino, gen);
2118         if (nce) {
2119                 if (ino < sctx->send_progress && nce->need_later_update) {
2120                         name_cache_delete(sctx, nce);
2121                         kfree(nce);
2122                         nce = NULL;
2123                 } else {
2124                         name_cache_used(sctx, nce);
2125                         *parent_ino = nce->parent_ino;
2126                         *parent_gen = nce->parent_gen;
2127                         ret = fs_path_add(dest, nce->name, nce->name_len);
2128                         if (ret < 0)
2129                                 goto out;
2130                         ret = nce->ret;
2131                         goto out;
2132                 }
2133         }
2134 
2135         /*
2136          * If the inode is not existent yet, add the orphan name and return 1.
2137          * This should only happen for the parent dir that we determine in
2138          * __record_new_ref
2139          */
2140         ret = is_inode_existent(sctx, ino, gen);
2141         if (ret < 0)
2142                 goto out;
2143 
2144         if (!ret) {
2145                 ret = gen_unique_name(sctx, ino, gen, dest);
2146                 if (ret < 0)
2147                         goto out;
2148                 ret = 1;
2149                 goto out_cache;
2150         }
2151 
2152         /*
2153          * Depending on whether the inode was already processed or not, use
2154          * send_root or parent_root for ref lookup.
2155          */
2156         if (ino < sctx->send_progress)
2157                 ret = get_first_ref(sctx->send_root, ino,
2158                                     parent_ino, parent_gen, dest);
2159         else
2160                 ret = get_first_ref(sctx->parent_root, ino,
2161                                     parent_ino, parent_gen, dest);
2162         if (ret < 0)
2163                 goto out;
2164 
2165         /*
2166          * Check if the ref was overwritten by an inode's ref that was processed
2167          * earlier. If yes, treat as orphan and return 1.
2168          */
2169         ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2170                         dest->start, dest->end - dest->start);
2171         if (ret < 0)
2172                 goto out;
2173         if (ret) {
2174                 fs_path_reset(dest);
2175                 ret = gen_unique_name(sctx, ino, gen, dest);
2176                 if (ret < 0)
2177                         goto out;
2178                 ret = 1;
2179         }
2180 
2181 out_cache:
2182         /*
2183          * Store the result of the lookup in the name cache.
2184          */
2185         nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2186         if (!nce) {
2187                 ret = -ENOMEM;
2188                 goto out;
2189         }
2190 
2191         nce->ino = ino;
2192         nce->gen = gen;
2193         nce->parent_ino = *parent_ino;
2194         nce->parent_gen = *parent_gen;
2195         nce->name_len = fs_path_len(dest);
2196         nce->ret = ret;
2197         strcpy(nce->name, dest->start);
2198 
2199         if (ino < sctx->send_progress)
2200                 nce->need_later_update = 0;
2201         else
2202                 nce->need_later_update = 1;
2203 
2204         nce_ret = name_cache_insert(sctx, nce);
2205         if (nce_ret < 0)
2206                 ret = nce_ret;
2207         name_cache_clean_unused(sctx);
2208 
2209 out:
2210         return ret;
2211 }
2212 
2213 /*
2214  * Magic happens here. This function returns the first ref to an inode as it
2215  * would look like while receiving the stream at this point in time.
2216  * We walk the path up to the root. For every inode in between, we check if it
2217  * was already processed/sent. If yes, we continue with the parent as found
2218  * in send_root. If not, we continue with the parent as found in parent_root.
2219  * If we encounter an inode that was deleted at this point in time, we use the
2220  * inodes "orphan" name instead of the real name and stop. Same with new inodes
2221  * that were not created yet and overwritten inodes/refs.
2222  *
2223  * When do we have orphan inodes:
2224  * 1. When an inode is freshly created and thus no valid refs are available yet
2225  * 2. When a directory lost all it's refs (deleted) but still has dir items
2226  *    inside which were not processed yet (pending for move/delete). If anyone
2227  *    tried to get the path to the dir items, it would get a path inside that
2228  *    orphan directory.
2229  * 3. When an inode is moved around or gets new links, it may overwrite the ref
2230  *    of an unprocessed inode. If in that case the first ref would be
2231  *    overwritten, the overwritten inode gets "orphanized". Later when we
2232  *    process this overwritten inode, it is restored at a new place by moving
2233  *    the orphan inode.
2234  *
2235  * sctx->send_progress tells this function at which point in time receiving
2236  * would be.
2237  */
2238 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2239                         struct fs_path *dest)
2240 {
2241         int ret = 0;
2242         struct fs_path *name = NULL;
2243         u64 parent_inode = 0;
2244         u64 parent_gen = 0;
2245         int stop = 0;
2246 
2247         name = fs_path_alloc();
2248         if (!name) {
2249                 ret = -ENOMEM;
2250                 goto out;
2251         }
2252 
2253         dest->reversed = 1;
2254         fs_path_reset(dest);
2255 
2256         while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2257                 struct waiting_dir_move *wdm;
2258 
2259                 fs_path_reset(name);
2260 
2261                 if (is_waiting_for_rm(sctx, ino)) {
2262                         ret = gen_unique_name(sctx, ino, gen, name);
2263                         if (ret < 0)
2264                                 goto out;
2265                         ret = fs_path_add_path(dest, name);
2266                         break;
2267                 }
2268 
2269                 wdm = get_waiting_dir_move(sctx, ino);
2270                 if (wdm && wdm->orphanized) {
2271                         ret = gen_unique_name(sctx, ino, gen, name);
2272                         stop = 1;
2273                 } else if (wdm) {
2274                         ret = get_first_ref(sctx->parent_root, ino,
2275                                             &parent_inode, &parent_gen, name);
2276                 } else {
2277                         ret = __get_cur_name_and_parent(sctx, ino, gen,
2278                                                         &parent_inode,
2279                                                         &parent_gen, name);
2280                         if (ret)
2281                                 stop = 1;
2282                 }
2283 
2284                 if (ret < 0)
2285                         goto out;
2286 
2287                 ret = fs_path_add_path(dest, name);
2288                 if (ret < 0)
2289                         goto out;
2290 
2291                 ino = parent_inode;
2292                 gen = parent_gen;
2293         }
2294 
2295 out:
2296         fs_path_free(name);
2297         if (!ret)
2298                 fs_path_unreverse(dest);
2299         return ret;
2300 }
2301 
2302 /*
2303  * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2304  */
2305 static int send_subvol_begin(struct send_ctx *sctx)
2306 {
2307         int ret;
2308         struct btrfs_root *send_root = sctx->send_root;
2309         struct btrfs_root *parent_root = sctx->parent_root;
2310         struct btrfs_path *path;
2311         struct btrfs_key key;
2312         struct btrfs_root_ref *ref;
2313         struct extent_buffer *leaf;
2314         char *name = NULL;
2315         int namelen;
2316 
2317         path = btrfs_alloc_path();
2318         if (!path)
2319                 return -ENOMEM;
2320 
2321         name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2322         if (!name) {
2323                 btrfs_free_path(path);
2324                 return -ENOMEM;
2325         }
2326 
2327         key.objectid = send_root->root_key.objectid;
2328         key.type = BTRFS_ROOT_BACKREF_KEY;
2329         key.offset = 0;
2330 
2331         ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2332                                 &key, path, 1, 0);
2333         if (ret < 0)
2334                 goto out;
2335         if (ret) {
2336                 ret = -ENOENT;
2337                 goto out;
2338         }
2339 
2340         leaf = path->nodes[0];
2341         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2342         if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2343             key.objectid != send_root->root_key.objectid) {
2344                 ret = -ENOENT;
2345                 goto out;
2346         }
2347         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2348         namelen = btrfs_root_ref_name_len(leaf, ref);
2349         read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2350         btrfs_release_path(path);
2351 
2352         if (parent_root) {
2353                 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2354                 if (ret < 0)
2355                         goto out;
2356         } else {
2357                 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2358                 if (ret < 0)
2359                         goto out;
2360         }
2361 
2362         TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2363 
2364         if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2365                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2366                             sctx->send_root->root_item.received_uuid);
2367         else
2368                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2369                             sctx->send_root->root_item.uuid);
2370 
2371         TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2372                     le64_to_cpu(sctx->send_root->root_item.ctransid));
2373         if (parent_root) {
2374                 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2375                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2376                                      parent_root->root_item.received_uuid);
2377                 else
2378                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2379                                      parent_root->root_item.uuid);
2380                 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2381                             le64_to_cpu(sctx->parent_root->root_item.ctransid));
2382         }
2383 
2384         ret = send_cmd(sctx);
2385 
2386 tlv_put_failure:
2387 out:
2388         btrfs_free_path(path);
2389         kfree(name);
2390         return ret;
2391 }
2392 
2393 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2394 {
2395         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2396         int ret = 0;
2397         struct fs_path *p;
2398 
2399         btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2400 
2401         p = fs_path_alloc();
2402         if (!p)
2403                 return -ENOMEM;
2404 
2405         ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2406         if (ret < 0)
2407                 goto out;
2408 
2409         ret = get_cur_path(sctx, ino, gen, p);
2410         if (ret < 0)
2411                 goto out;
2412         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2413         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2414 
2415         ret = send_cmd(sctx);
2416 
2417 tlv_put_failure:
2418 out:
2419         fs_path_free(p);
2420         return ret;
2421 }
2422 
2423 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2424 {
2425         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2426         int ret = 0;
2427         struct fs_path *p;
2428 
2429         btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2430 
2431         p = fs_path_alloc();
2432         if (!p)
2433                 return -ENOMEM;
2434 
2435         ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2436         if (ret < 0)
2437                 goto out;
2438 
2439         ret = get_cur_path(sctx, ino, gen, p);
2440         if (ret < 0)
2441                 goto out;
2442         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2443         TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2444 
2445         ret = send_cmd(sctx);
2446 
2447 tlv_put_failure:
2448 out:
2449         fs_path_free(p);
2450         return ret;
2451 }
2452 
2453 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2454 {
2455         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2456         int ret = 0;
2457         struct fs_path *p;
2458 
2459         btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2460                     ino, uid, gid);
2461 
2462         p = fs_path_alloc();
2463         if (!p)
2464                 return -ENOMEM;
2465 
2466         ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2467         if (ret < 0)
2468                 goto out;
2469 
2470         ret = get_cur_path(sctx, ino, gen, p);
2471         if (ret < 0)
2472                 goto out;
2473         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2474         TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2475         TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2476 
2477         ret = send_cmd(sctx);
2478 
2479 tlv_put_failure:
2480 out:
2481         fs_path_free(p);
2482         return ret;
2483 }
2484 
2485 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2486 {
2487         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2488         int ret = 0;
2489         struct fs_path *p = NULL;
2490         struct btrfs_inode_item *ii;
2491         struct btrfs_path *path = NULL;
2492         struct extent_buffer *eb;
2493         struct btrfs_key key;
2494         int slot;
2495 
2496         btrfs_debug(fs_info, "send_utimes %llu", ino);
2497 
2498         p = fs_path_alloc();
2499         if (!p)
2500                 return -ENOMEM;
2501 
2502         path = alloc_path_for_send();
2503         if (!path) {
2504                 ret = -ENOMEM;
2505                 goto out;
2506         }
2507 
2508         key.objectid = ino;
2509         key.type = BTRFS_INODE_ITEM_KEY;
2510         key.offset = 0;
2511         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2512         if (ret > 0)
2513                 ret = -ENOENT;
2514         if (ret < 0)
2515                 goto out;
2516 
2517         eb = path->nodes[0];
2518         slot = path->slots[0];
2519         ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2520 
2521         ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2522         if (ret < 0)
2523                 goto out;
2524 
2525         ret = get_cur_path(sctx, ino, gen, p);
2526         if (ret < 0)
2527                 goto out;
2528         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2529         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2530         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2531         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2532         /* TODO Add otime support when the otime patches get into upstream */
2533 
2534         ret = send_cmd(sctx);
2535 
2536 tlv_put_failure:
2537 out:
2538         fs_path_free(p);
2539         btrfs_free_path(path);
2540         return ret;
2541 }
2542 
2543 /*
2544  * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2545  * a valid path yet because we did not process the refs yet. So, the inode
2546  * is created as orphan.
2547  */
2548 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2549 {
2550         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2551         int ret = 0;
2552         struct fs_path *p;
2553         int cmd;
2554         u64 gen;
2555         u64 mode;
2556         u64 rdev;
2557 
2558         btrfs_debug(fs_info, "send_create_inode %llu", ino);
2559 
2560         p = fs_path_alloc();
2561         if (!p)
2562                 return -ENOMEM;
2563 
2564         if (ino != sctx->cur_ino) {
2565                 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2566                                      NULL, NULL, &rdev);
2567                 if (ret < 0)
2568                         goto out;
2569         } else {
2570                 gen = sctx->cur_inode_gen;
2571                 mode = sctx->cur_inode_mode;
2572                 rdev = sctx->cur_inode_rdev;
2573         }
2574 
2575         if (S_ISREG(mode)) {
2576                 cmd = BTRFS_SEND_C_MKFILE;
2577         } else if (S_ISDIR(mode)) {
2578                 cmd = BTRFS_SEND_C_MKDIR;
2579         } else if (S_ISLNK(mode)) {
2580                 cmd = BTRFS_SEND_C_SYMLINK;
2581         } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2582                 cmd = BTRFS_SEND_C_MKNOD;
2583         } else if (S_ISFIFO(mode)) {
2584                 cmd = BTRFS_SEND_C_MKFIFO;
2585         } else if (S_ISSOCK(mode)) {
2586                 cmd = BTRFS_SEND_C_MKSOCK;
2587         } else {
2588                 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2589                                 (int)(mode & S_IFMT));
2590                 ret = -EOPNOTSUPP;
2591                 goto out;
2592         }
2593 
2594         ret = begin_cmd(sctx, cmd);
2595         if (ret < 0)
2596                 goto out;
2597 
2598         ret = gen_unique_name(sctx, ino, gen, p);
2599         if (ret < 0)
2600                 goto out;
2601 
2602         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2603         TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2604 
2605         if (S_ISLNK(mode)) {
2606                 fs_path_reset(p);
2607                 ret = read_symlink(sctx->send_root, ino, p);
2608                 if (ret < 0)
2609                         goto out;
2610                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2611         } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2612                    S_ISFIFO(mode) || S_ISSOCK(mode)) {
2613                 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2614                 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2615         }
2616 
2617         ret = send_cmd(sctx);
2618         if (ret < 0)
2619                 goto out;
2620 
2621 
2622 tlv_put_failure:
2623 out:
2624         fs_path_free(p);
2625         return ret;
2626 }
2627 
2628 /*
2629  * We need some special handling for inodes that get processed before the parent
2630  * directory got created. See process_recorded_refs for details.
2631  * This function does the check if we already created the dir out of order.
2632  */
2633 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2634 {
2635         int ret = 0;
2636         struct btrfs_path *path = NULL;
2637         struct btrfs_key key;
2638         struct btrfs_key found_key;
2639         struct btrfs_key di_key;
2640         struct extent_buffer *eb;
2641         struct btrfs_dir_item *di;
2642         int slot;
2643 
2644         path = alloc_path_for_send();
2645         if (!path) {
2646                 ret = -ENOMEM;
2647                 goto out;
2648         }
2649 
2650         key.objectid = dir;
2651         key.type = BTRFS_DIR_INDEX_KEY;
2652         key.offset = 0;
2653         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2654         if (ret < 0)
2655                 goto out;
2656 
2657         while (1) {
2658                 eb = path->nodes[0];
2659                 slot = path->slots[0];
2660                 if (slot >= btrfs_header_nritems(eb)) {
2661                         ret = btrfs_next_leaf(sctx->send_root, path);
2662                         if (ret < 0) {
2663                                 goto out;
2664                         } else if (ret > 0) {
2665                                 ret = 0;
2666                                 break;
2667                         }
2668                         continue;
2669                 }
2670 
2671                 btrfs_item_key_to_cpu(eb, &found_key, slot);
2672                 if (found_key.objectid != key.objectid ||
2673                     found_key.type != key.type) {
2674                         ret = 0;
2675                         goto out;
2676                 }
2677 
2678                 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2679                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2680 
2681                 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2682                     di_key.objectid < sctx->send_progress) {
2683                         ret = 1;
2684                         goto out;
2685                 }
2686 
2687                 path->slots[0]++;
2688         }
2689 
2690 out:
2691         btrfs_free_path(path);
2692         return ret;
2693 }
2694 
2695 /*
2696  * Only creates the inode if it is:
2697  * 1. Not a directory
2698  * 2. Or a directory which was not created already due to out of order
2699  *    directories. See did_create_dir and process_recorded_refs for details.
2700  */
2701 static int send_create_inode_if_needed(struct send_ctx *sctx)
2702 {
2703         int ret;
2704 
2705         if (S_ISDIR(sctx->cur_inode_mode)) {
2706                 ret = did_create_dir(sctx, sctx->cur_ino);
2707                 if (ret < 0)
2708                         goto out;
2709                 if (ret) {
2710                         ret = 0;
2711                         goto out;
2712                 }
2713         }
2714 
2715         ret = send_create_inode(sctx, sctx->cur_ino);
2716         if (ret < 0)
2717                 goto out;
2718 
2719 out:
2720         return ret;
2721 }
2722 
2723 struct recorded_ref {
2724         struct list_head list;
2725         char *name;
2726         struct fs_path *full_path;
2727         u64 dir;
2728         u64 dir_gen;
2729         int name_len;
2730 };
2731 
2732 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
2733 {
2734         ref->full_path = path;
2735         ref->name = (char *)kbasename(ref->full_path->start);
2736         ref->name_len = ref->full_path->end - ref->name;
2737 }
2738 
2739 /*
2740  * We need to process new refs before deleted refs, but compare_tree gives us
2741  * everything mixed. So we first record all refs and later process them.
2742  * This function is a helper to record one ref.
2743  */
2744 static int __record_ref(struct list_head *head, u64 dir,
2745                       u64 dir_gen, struct fs_path *path)
2746 {
2747         struct recorded_ref *ref;
2748 
2749         ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2750         if (!ref)
2751                 return -ENOMEM;
2752 
2753         ref->dir = dir;
2754         ref->dir_gen = dir_gen;
2755         set_ref_path(ref, path);
2756         list_add_tail(&ref->list, head);
2757         return 0;
2758 }
2759 
2760 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2761 {
2762         struct recorded_ref *new;
2763 
2764         new = kmalloc(sizeof(*ref), GFP_KERNEL);
2765         if (!new)
2766                 return -ENOMEM;
2767 
2768         new->dir = ref->dir;
2769         new->dir_gen = ref->dir_gen;
2770         new->full_path = NULL;
2771         INIT_LIST_HEAD(&new->list);
2772         list_add_tail(&new->list, list);
2773         return 0;
2774 }
2775 
2776 static void __free_recorded_refs(struct list_head *head)
2777 {
2778         struct recorded_ref *cur;
2779 
2780         while (!list_empty(head)) {
2781                 cur = list_entry(head->next, struct recorded_ref, list);
2782                 fs_path_free(cur->full_path);
2783                 list_del(&cur->list);
2784                 kfree(cur);
2785         }
2786 }
2787 
2788 static void free_recorded_refs(struct send_ctx *sctx)
2789 {
2790         __free_recorded_refs(&sctx->new_refs);
2791         __free_recorded_refs(&sctx->deleted_refs);
2792 }
2793 
2794 /*
2795  * Renames/moves a file/dir to its orphan name. Used when the first
2796  * ref of an unprocessed inode gets overwritten and for all non empty
2797  * directories.
2798  */
2799 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2800                           struct fs_path *path)
2801 {
2802         int ret;
2803         struct fs_path *orphan;
2804 
2805         orphan = fs_path_alloc();
2806         if (!orphan)
2807                 return -ENOMEM;
2808 
2809         ret = gen_unique_name(sctx, ino, gen, orphan);
2810         if (ret < 0)
2811                 goto out;
2812 
2813         ret = send_rename(sctx, path, orphan);
2814 
2815 out:
2816         fs_path_free(orphan);
2817         return ret;
2818 }
2819 
2820 static struct orphan_dir_info *
2821 add_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2822 {
2823         struct rb_node **p = &sctx->orphan_dirs.rb_node;
2824         struct rb_node *parent = NULL;
2825         struct orphan_dir_info *entry, *odi;
2826 
2827         while (*p) {
2828                 parent = *p;
2829                 entry = rb_entry(parent, struct orphan_dir_info, node);
2830                 if (dir_ino < entry->ino) {
2831                         p = &(*p)->rb_left;
2832                 } else if (dir_ino > entry->ino) {
2833                         p = &(*p)->rb_right;
2834                 } else {
2835                         return entry;
2836                 }
2837         }
2838 
2839         odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2840         if (!odi)
2841                 return ERR_PTR(-ENOMEM);
2842         odi->ino = dir_ino;
2843         odi->gen = 0;
2844         odi->last_dir_index_offset = 0;
2845 
2846         rb_link_node(&odi->node, parent, p);
2847         rb_insert_color(&odi->node, &sctx->orphan_dirs);
2848         return odi;
2849 }
2850 
2851 static struct orphan_dir_info *
2852 get_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2853 {
2854         struct rb_node *n = sctx->orphan_dirs.rb_node;
2855         struct orphan_dir_info *entry;
2856 
2857         while (n) {
2858                 entry = rb_entry(n, struct orphan_dir_info, node);
2859                 if (dir_ino < entry->ino)
2860                         n = n->rb_left;
2861                 else if (dir_ino > entry->ino)
2862                         n = n->rb_right;
2863                 else
2864                         return entry;
2865         }
2866         return NULL;
2867 }
2868 
2869 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino)
2870 {
2871         struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino);
2872 
2873         return odi != NULL;
2874 }
2875 
2876 static void free_orphan_dir_info(struct send_ctx *sctx,
2877                                  struct orphan_dir_info *odi)
2878 {
2879         if (!odi)
2880                 return;
2881         rb_erase(&odi->node, &sctx->orphan_dirs);
2882         kfree(odi);
2883 }
2884 
2885 /*
2886  * Returns 1 if a directory can be removed at this point in time.
2887  * We check this by iterating all dir items and checking if the inode behind
2888  * the dir item was already processed.
2889  */
2890 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2891                      u64 send_progress)
2892 {
2893         int ret = 0;
2894         struct btrfs_root *root = sctx->parent_root;
2895         struct btrfs_path *path;
2896         struct btrfs_key key;
2897         struct btrfs_key found_key;
2898         struct btrfs_key loc;
2899         struct btrfs_dir_item *di;
2900         struct orphan_dir_info *odi = NULL;
2901 
2902         /*
2903          * Don't try to rmdir the top/root subvolume dir.
2904          */
2905         if (dir == BTRFS_FIRST_FREE_OBJECTID)
2906                 return 0;
2907 
2908         path = alloc_path_for_send();
2909         if (!path)
2910                 return -ENOMEM;
2911 
2912         key.objectid = dir;
2913         key.type = BTRFS_DIR_INDEX_KEY;
2914         key.offset = 0;
2915 
2916         odi = get_orphan_dir_info(sctx, dir);
2917         if (odi)
2918                 key.offset = odi->last_dir_index_offset;
2919 
2920         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2921         if (ret < 0)
2922                 goto out;
2923 
2924         while (1) {
2925                 struct waiting_dir_move *dm;
2926 
2927                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2928                         ret = btrfs_next_leaf(root, path);
2929                         if (ret < 0)
2930                                 goto out;
2931                         else if (ret > 0)
2932                                 break;
2933                         continue;
2934                 }
2935                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2936                                       path->slots[0]);
2937                 if (found_key.objectid != key.objectid ||
2938                     found_key.type != key.type)
2939                         break;
2940 
2941                 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2942                                 struct btrfs_dir_item);
2943                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2944 
2945                 dm = get_waiting_dir_move(sctx, loc.objectid);
2946                 if (dm) {
2947                         odi = add_orphan_dir_info(sctx, dir);
2948                         if (IS_ERR(odi)) {
2949                                 ret = PTR_ERR(odi);
2950                                 goto out;
2951                         }
2952                         odi->gen = dir_gen;
2953                         odi->last_dir_index_offset = found_key.offset;
2954                         dm->rmdir_ino = dir;
2955                         ret = 0;
2956                         goto out;
2957                 }
2958 
2959                 if (loc.objectid > send_progress) {
2960                         odi = add_orphan_dir_info(sctx, dir);
2961                         if (IS_ERR(odi)) {
2962                                 ret = PTR_ERR(odi);
2963                                 goto out;
2964                         }
2965                         odi->gen = dir_gen;
2966                         odi->last_dir_index_offset = found_key.offset;
2967                         ret = 0;
2968                         goto out;
2969                 }
2970 
2971                 path->slots[0]++;
2972         }
2973         free_orphan_dir_info(sctx, odi);
2974 
2975         ret = 1;
2976 
2977 out:
2978         btrfs_free_path(path);
2979         return ret;
2980 }
2981 
2982 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
2983 {
2984         struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
2985 
2986         return entry != NULL;
2987 }
2988 
2989 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
2990 {
2991         struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
2992         struct rb_node *parent = NULL;
2993         struct waiting_dir_move *entry, *dm;
2994 
2995         dm = kmalloc(sizeof(*dm), GFP_KERNEL);
2996         if (!dm)
2997                 return -ENOMEM;
2998         dm->ino = ino;
2999         dm->rmdir_ino = 0;
3000         dm->orphanized = orphanized;
3001 
3002         while (*p) {
3003                 parent = *p;
3004                 entry = rb_entry(parent, struct waiting_dir_move, node);
3005                 if (ino < entry->ino) {
3006                         p = &(*p)->rb_left;
3007                 } else if (ino > entry->ino) {
3008                         p = &(*p)->rb_right;
3009                 } else {
3010                         kfree(dm);
3011                         return -EEXIST;
3012                 }
3013         }
3014 
3015         rb_link_node(&dm->node, parent, p);
3016         rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3017         return 0;
3018 }
3019 
3020 static struct waiting_dir_move *
3021 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3022 {
3023         struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3024         struct waiting_dir_move *entry;
3025 
3026         while (n) {
3027                 entry = rb_entry(n, struct waiting_dir_move, node);
3028                 if (ino < entry->ino)
3029                         n = n->rb_left;
3030                 else if (ino > entry->ino)
3031                         n = n->rb_right;
3032                 else
3033                         return entry;
3034         }
3035         return NULL;
3036 }
3037 
3038 static void free_waiting_dir_move(struct send_ctx *sctx,
3039                                   struct waiting_dir_move *dm)
3040 {
3041         if (!dm)
3042                 return;
3043         rb_erase(&dm->node, &sctx->waiting_dir_moves);
3044         kfree(dm);
3045 }
3046 
3047 static int add_pending_dir_move(struct send_ctx *sctx,
3048                                 u64 ino,
3049                                 u64 ino_gen,
3050                                 u64 parent_ino,
3051                                 struct list_head *new_refs,
3052                                 struct list_head *deleted_refs,
3053                                 const bool is_orphan)
3054 {
3055         struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3056         struct rb_node *parent = NULL;
3057         struct pending_dir_move *entry = NULL, *pm;
3058         struct recorded_ref *cur;
3059         int exists = 0;
3060         int ret;
3061 
3062         pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3063         if (!pm)
3064                 return -ENOMEM;
3065         pm->parent_ino = parent_ino;
3066         pm->ino = ino;
3067         pm->gen = ino_gen;
3068         INIT_LIST_HEAD(&pm->list);
3069         INIT_LIST_HEAD(&pm->update_refs);
3070         RB_CLEAR_NODE(&pm->node);
3071 
3072         while (*p) {
3073                 parent = *p;
3074                 entry = rb_entry(parent, struct pending_dir_move, node);
3075                 if (parent_ino < entry->parent_ino) {
3076                         p = &(*p)->rb_left;
3077                 } else if (parent_ino > entry->parent_ino) {
3078                         p = &(*p)->rb_right;
3079                 } else {
3080                         exists = 1;
3081                         break;
3082                 }
3083         }
3084 
3085         list_for_each_entry(cur, deleted_refs, list) {
3086                 ret = dup_ref(cur, &pm->update_refs);
3087                 if (ret < 0)
3088                         goto out;
3089         }
3090         list_for_each_entry(cur, new_refs, list) {
3091                 ret = dup_ref(cur, &pm->update_refs);
3092                 if (ret < 0)
3093                         goto out;
3094         }
3095 
3096         ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3097         if (ret)
3098                 goto out;
3099 
3100         if (exists) {
3101                 list_add_tail(&pm->list, &entry->list);
3102         } else {
3103                 rb_link_node(&pm->node, parent, p);
3104                 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3105         }
3106         ret = 0;
3107 out:
3108         if (ret) {
3109                 __free_recorded_refs(&pm->update_refs);
3110                 kfree(pm);
3111         }
3112         return ret;
3113 }
3114 
3115 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3116                                                       u64 parent_ino)
3117 {
3118         struct rb_node *n = sctx->pending_dir_moves.rb_node;
3119         struct pending_dir_move *entry;
3120 
3121         while (n) {
3122                 entry = rb_entry(n, struct pending_dir_move, node);
3123                 if (parent_ino < entry->parent_ino)
3124                         n = n->rb_left;
3125                 else if (parent_ino > entry->parent_ino)
3126                         n = n->rb_right;
3127                 else
3128                         return entry;
3129         }
3130         return NULL;
3131 }
3132 
3133 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3134                      u64 ino, u64 gen, u64 *ancestor_ino)
3135 {
3136         int ret = 0;
3137         u64 parent_inode = 0;
3138         u64 parent_gen = 0;
3139         u64 start_ino = ino;
3140 
3141         *ancestor_ino = 0;
3142         while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3143                 fs_path_reset(name);
3144 
3145                 if (is_waiting_for_rm(sctx, ino))
3146                         break;
3147                 if (is_waiting_for_move(sctx, ino)) {
3148                         if (*ancestor_ino == 0)
3149                                 *ancestor_ino = ino;
3150                         ret = get_first_ref(sctx->parent_root, ino,
3151                                             &parent_inode, &parent_gen, name);
3152                 } else {
3153                         ret = __get_cur_name_and_parent(sctx, ino, gen,
3154                                                         &parent_inode,
3155                                                         &parent_gen, name);
3156                         if (ret > 0) {
3157                                 ret = 0;
3158                                 break;
3159                         }
3160                 }
3161                 if (ret < 0)
3162                         break;
3163                 if (parent_inode == start_ino) {
3164                         ret = 1;
3165                         if (*ancestor_ino == 0)
3166                                 *ancestor_ino = ino;
3167                         break;
3168                 }
3169                 ino = parent_inode;
3170                 gen = parent_gen;
3171         }
3172         return ret;
3173 }
3174 
3175 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3176 {
3177         struct fs_path *from_path = NULL;
3178         struct fs_path *to_path = NULL;
3179         struct fs_path *name = NULL;
3180         u64 orig_progress = sctx->send_progress;
3181         struct recorded_ref *cur;
3182         u64 parent_ino, parent_gen;
3183         struct waiting_dir_move *dm = NULL;
3184         u64 rmdir_ino = 0;
3185         u64 ancestor;
3186         bool is_orphan;
3187         int ret;
3188 
3189         name = fs_path_alloc();
3190         from_path = fs_path_alloc();
3191         if (!name || !from_path) {
3192                 ret = -ENOMEM;
3193                 goto out;
3194         }
3195 
3196         dm = get_waiting_dir_move(sctx, pm->ino);
3197         ASSERT(dm);
3198         rmdir_ino = dm->rmdir_ino;
3199         is_orphan = dm->orphanized;
3200         free_waiting_dir_move(sctx, dm);
3201 
3202         if (is_orphan) {
3203                 ret = gen_unique_name(sctx, pm->ino,
3204                                       pm->gen, from_path);
3205         } else {
3206                 ret = get_first_ref(sctx->parent_root, pm->ino,
3207                                     &parent_ino, &parent_gen, name);
3208                 if (ret < 0)
3209                         goto out;
3210                 ret = get_cur_path(sctx, parent_ino, parent_gen,
3211                                    from_path);
3212                 if (ret < 0)
3213                         goto out;
3214                 ret = fs_path_add_path(from_path, name);
3215         }
3216         if (ret < 0)
3217                 goto out;
3218 
3219         sctx->send_progress = sctx->cur_ino + 1;
3220         ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3221         if (ret < 0)
3222                 goto out;
3223         if (ret) {
3224                 LIST_HEAD(deleted_refs);
3225                 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3226                 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3227                                            &pm->update_refs, &deleted_refs,
3228                                            is_orphan);
3229                 if (ret < 0)
3230                         goto out;
3231                 if (rmdir_ino) {
3232                         dm = get_waiting_dir_move(sctx, pm->ino);
3233                         ASSERT(dm);
3234                         dm->rmdir_ino = rmdir_ino;
3235                 }
3236                 goto out;
3237         }
3238         fs_path_reset(name);
3239         to_path = name;
3240         name = NULL;
3241         ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3242         if (ret < 0)
3243                 goto out;
3244 
3245         ret = send_rename(sctx, from_path, to_path);
3246         if (ret < 0)
3247                 goto out;
3248 
3249         if (rmdir_ino) {
3250                 struct orphan_dir_info *odi;
3251                 u64 gen;
3252 
3253                 odi = get_orphan_dir_info(sctx, rmdir_ino);
3254                 if (!odi) {
3255                         /* already deleted */
3256                         goto finish;
3257                 }
3258                 gen = odi->gen;
3259 
3260                 ret = can_rmdir(sctx, rmdir_ino, gen, sctx->cur_ino);
3261                 if (ret < 0)
3262                         goto out;
3263                 if (!ret)
3264                         goto finish;
3265 
3266                 name = fs_path_alloc();
3267                 if (!name) {
3268                         ret = -ENOMEM;
3269                         goto out;
3270                 }
3271                 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3272                 if (ret < 0)
3273                         goto out;
3274                 ret = send_rmdir(sctx, name);
3275                 if (ret < 0)
3276                         goto out;
3277         }
3278 
3279 finish:
3280         ret = send_utimes(sctx, pm->ino, pm->gen);
3281         if (ret < 0)
3282                 goto out;
3283 
3284         /*
3285          * After rename/move, need to update the utimes of both new parent(s)
3286          * and old parent(s).
3287          */
3288         list_for_each_entry(cur, &pm->update_refs, list) {
3289                 /*
3290                  * The parent inode might have been deleted in the send snapshot
3291                  */
3292                 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3293                                      NULL, NULL, NULL, NULL, NULL);
3294                 if (ret == -ENOENT) {
3295                         ret = 0;
3296                         continue;
3297                 }
3298                 if (ret < 0)
3299                         goto out;
3300 
3301                 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3302                 if (ret < 0)
3303                         goto out;
3304         }
3305 
3306 out:
3307         fs_path_free(name);
3308         fs_path_free(from_path);
3309         fs_path_free(to_path);
3310         sctx->send_progress = orig_progress;
3311 
3312         return ret;
3313 }
3314 
3315 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3316 {
3317         if (!list_empty(&m->list))
3318                 list_del(&m->list);
3319         if (!RB_EMPTY_NODE(&m->node))
3320                 rb_erase(&m->node, &sctx->pending_dir_moves);
3321         __free_recorded_refs(&m->update_refs);
3322         kfree(m);
3323 }
3324 
3325 static void tail_append_pending_moves(struct send_ctx *sctx,
3326                                       struct pending_dir_move *moves,
3327                                       struct list_head *stack)
3328 {
3329         if (list_empty(&moves->list)) {
3330                 list_add_tail(&moves->list, stack);
3331         } else {
3332                 LIST_HEAD(list);
3333                 list_splice_init(&moves->list, &list);
3334                 list_add_tail(&moves->list, stack);
3335                 list_splice_tail(&list, stack);
3336         }
3337         if (!RB_EMPTY_NODE(&moves->node)) {
3338                 rb_erase(&moves->node, &sctx->pending_dir_moves);
3339                 RB_CLEAR_NODE(&moves->node);
3340         }
3341 }
3342 
3343 static int apply_children_dir_moves(struct send_ctx *sctx)
3344 {
3345         struct pending_dir_move *pm;
3346         struct list_head stack;
3347         u64 parent_ino = sctx->cur_ino;
3348         int ret = 0;
3349 
3350         pm = get_pending_dir_moves(sctx, parent_ino);
3351         if (!pm)
3352                 return 0;
3353 
3354         INIT_LIST_HEAD(&stack);
3355         tail_append_pending_moves(sctx, pm, &stack);
3356 
3357         while (!list_empty(&stack)) {
3358                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3359                 parent_ino = pm->ino;
3360                 ret = apply_dir_move(sctx, pm);
3361                 free_pending_move(sctx, pm);
3362                 if (ret)
3363                         goto out;
3364                 pm = get_pending_dir_moves(sctx, parent_ino);
3365                 if (pm)
3366                         tail_append_pending_moves(sctx, pm, &stack);
3367         }
3368         return 0;
3369 
3370 out:
3371         while (!list_empty(&stack)) {
3372                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3373                 free_pending_move(sctx, pm);
3374         }
3375         return ret;
3376 }
3377 
3378 /*
3379  * We might need to delay a directory rename even when no ancestor directory
3380  * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3381  * renamed. This happens when we rename a directory to the old name (the name
3382  * in the parent root) of some other unrelated directory that got its rename
3383  * delayed due to some ancestor with higher number that got renamed.
3384  *
3385  * Example:
3386  *
3387  * Parent snapshot:
3388  * .                                       (ino 256)
3389  * |---- a/                                (ino 257)
3390  * |     |---- file                        (ino 260)
3391  * |
3392  * |---- b/                                (ino 258)
3393  * |---- c/                                (ino 259)
3394  *
3395  * Send snapshot:
3396  * .                                       (ino 256)
3397  * |---- a/                                (ino 258)
3398  * |---- x/                                (ino 259)
3399  *       |---- y/                          (ino 257)
3400  *             |----- file                 (ino 260)
3401  *
3402  * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3403  * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3404  * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3405  * must issue is:
3406  *
3407  * 1 - rename 259 from 'c' to 'x'
3408  * 2 - rename 257 from 'a' to 'x/y'
3409  * 3 - rename 258 from 'b' to 'a'
3410  *
3411  * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3412  * be done right away and < 0 on error.
3413  */
3414 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3415                                   struct recorded_ref *parent_ref,
3416                                   const bool is_orphan)
3417 {
3418         struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3419         struct btrfs_path *path;
3420         struct btrfs_key key;
3421         struct btrfs_key di_key;
3422         struct btrfs_dir_item *di;
3423         u64 left_gen;
3424         u64 right_gen;
3425         int ret = 0;
3426         struct waiting_dir_move *wdm;
3427 
3428         if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3429                 return 0;
3430 
3431         path = alloc_path_for_send();
3432         if (!path)
3433                 return -ENOMEM;
3434 
3435         key.objectid = parent_ref->dir;
3436         key.type = BTRFS_DIR_ITEM_KEY;
3437         key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3438 
3439         ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3440         if (ret < 0) {
3441                 goto out;
3442         } else if (ret > 0) {
3443                 ret = 0;
3444                 goto out;
3445         }
3446 
3447         di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3448                                        parent_ref->name_len);
3449         if (!di) {
3450                 ret = 0;
3451                 goto out;
3452         }
3453         /*
3454          * di_key.objectid has the number of the inode that has a dentry in the
3455          * parent directory with the same name that sctx->cur_ino is being
3456          * renamed to. We need to check if that inode is in the send root as
3457          * well and if it is currently marked as an inode with a pending rename,
3458          * if it is, we need to delay the rename of sctx->cur_ino as well, so
3459          * that it happens after that other inode is renamed.
3460          */
3461         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3462         if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3463                 ret = 0;
3464                 goto out;
3465         }
3466 
3467         ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3468                              &left_gen, NULL, NULL, NULL, NULL);
3469         if (ret < 0)
3470                 goto out;
3471         ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3472                              &right_gen, NULL, NULL, NULL, NULL);
3473         if (ret < 0) {
3474                 if (ret == -ENOENT)
3475                         ret = 0;
3476                 goto out;
3477         }
3478 
3479         /* Different inode, no need to delay the rename of sctx->cur_ino */
3480         if (right_gen != left_gen) {
3481                 ret = 0;
3482                 goto out;
3483         }
3484 
3485         wdm = get_waiting_dir_move(sctx, di_key.objectid);
3486         if (wdm && !wdm->orphanized) {
3487                 ret = add_pending_dir_move(sctx,
3488                                            sctx->cur_ino,
3489                                            sctx->cur_inode_gen,
3490                                            di_key.objectid,
3491                                            &sctx->new_refs,
3492                                            &sctx->deleted_refs,
3493                                            is_orphan);
3494                 if (!ret)
3495                         ret = 1;
3496         }
3497 out:
3498         btrfs_free_path(path);
3499         return ret;
3500 }
3501 
3502 /*
3503  * Check if inode ino2, or any of its ancestors, is inode ino1.
3504  * Return 1 if true, 0 if false and < 0 on error.
3505  */
3506 static int check_ino_in_path(struct btrfs_root *root,
3507                              const u64 ino1,
3508                              const u64 ino1_gen,
3509                              const u64 ino2,
3510                              const u64 ino2_gen,
3511                              struct fs_path *fs_path)
3512 {
3513         u64 ino = ino2;
3514 
3515         if (ino1 == ino2)
3516                 return ino1_gen == ino2_gen;
3517 
3518         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3519                 u64 parent;
3520                 u64 parent_gen;
3521                 int ret;
3522 
3523                 fs_path_reset(fs_path);
3524                 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3525                 if (ret < 0)
3526                         return ret;
3527                 if (parent == ino1)
3528                         return parent_gen == ino1_gen;
3529                 ino = parent;
3530         }
3531         return 0;
3532 }
3533 
3534 /*
3535  * Check if ino ino1 is an ancestor of inode ino2 in the given root for any
3536  * possible path (in case ino2 is not a directory and has multiple hard links).
3537  * Return 1 if true, 0 if false and < 0 on error.
3538  */
3539 static int is_ancestor(struct btrfs_root *root,
3540                        const u64 ino1,
3541                        const u64 ino1_gen,
3542                        const u64 ino2,
3543                        struct fs_path *fs_path)
3544 {
3545         bool free_fs_path = false;
3546         int ret = 0;
3547         struct btrfs_path *path = NULL;
3548         struct btrfs_key key;
3549 
3550         if (!fs_path) {
3551                 fs_path = fs_path_alloc();
3552                 if (!fs_path)
3553                         return -ENOMEM;
3554                 free_fs_path = true;
3555         }
3556 
3557         path = alloc_path_for_send();
3558         if (!path) {
3559                 ret = -ENOMEM;
3560                 goto out;
3561         }
3562 
3563         key.objectid = ino2;
3564         key.type = BTRFS_INODE_REF_KEY;
3565         key.offset = 0;
3566 
3567         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3568         if (ret < 0)
3569                 goto out;
3570 
3571         while (true) {
3572                 struct extent_buffer *leaf = path->nodes[0];
3573                 int slot = path->slots[0];
3574                 u32 cur_offset = 0;
3575                 u32 item_size;
3576 
3577                 if (slot >= btrfs_header_nritems(leaf)) {
3578                         ret = btrfs_next_leaf(root, path);
3579                         if (ret < 0)
3580                                 goto out;
3581                         if (ret > 0)
3582                                 break;
3583                         continue;
3584                 }
3585 
3586                 btrfs_item_key_to_cpu(leaf, &key, slot);
3587                 if (key.objectid != ino2)
3588                         break;
3589                 if (key.type != BTRFS_INODE_REF_KEY &&
3590                     key.type != BTRFS_INODE_EXTREF_KEY)
3591                         break;
3592 
3593                 item_size = btrfs_item_size_nr(leaf, slot);
3594                 while (cur_offset < item_size) {
3595                         u64 parent;
3596                         u64 parent_gen;
3597 
3598                         if (key.type == BTRFS_INODE_EXTREF_KEY) {
3599                                 unsigned long ptr;
3600                                 struct btrfs_inode_extref *extref;
3601 
3602                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3603                                 extref = (struct btrfs_inode_extref *)
3604                                         (ptr + cur_offset);
3605                                 parent = btrfs_inode_extref_parent(leaf,
3606                                                                    extref);
3607                                 cur_offset += sizeof(*extref);
3608                                 cur_offset += btrfs_inode_extref_name_len(leaf,
3609                                                                   extref);
3610                         } else {
3611                                 parent = key.offset;
3612                                 cur_offset = item_size;
3613                         }
3614 
3615                         ret = get_inode_info(root, parent, NULL, &parent_gen,
3616                                              NULL, NULL, NULL, NULL);
3617                         if (ret < 0)
3618                                 goto out;
3619                         ret = check_ino_in_path(root, ino1, ino1_gen,
3620                                                 parent, parent_gen, fs_path);
3621                         if (ret)
3622                                 goto out;
3623                 }
3624                 path->slots[0]++;
3625         }
3626         ret = 0;
3627  out:
3628         btrfs_free_path(path);
3629         if (free_fs_path)
3630                 fs_path_free(fs_path);
3631         return ret;
3632 }
3633 
3634 static int wait_for_parent_move(struct send_ctx *sctx,
3635                                 struct recorded_ref *parent_ref,
3636                                 const bool is_orphan)
3637 {
3638         int ret = 0;
3639         u64 ino = parent_ref->dir;
3640         u64 ino_gen = parent_ref->dir_gen;
3641         u64 parent_ino_before, parent_ino_after;
3642         struct fs_path *path_before = NULL;
3643         struct fs_path *path_after = NULL;
3644         int len1, len2;
3645 
3646         path_after = fs_path_alloc();
3647         path_before = fs_path_alloc();
3648         if (!path_after || !path_before) {
3649                 ret = -ENOMEM;
3650                 goto out;
3651         }
3652 
3653         /*
3654          * Our current directory inode may not yet be renamed/moved because some
3655          * ancestor (immediate or not) has to be renamed/moved first. So find if
3656          * such ancestor exists and make sure our own rename/move happens after
3657          * that ancestor is processed to avoid path build infinite loops (done
3658          * at get_cur_path()).
3659          */
3660         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3661                 u64 parent_ino_after_gen;
3662 
3663                 if (is_waiting_for_move(sctx, ino)) {
3664                         /*
3665                          * If the current inode is an ancestor of ino in the
3666                          * parent root, we need to delay the rename of the
3667                          * current inode, otherwise don't delayed the rename
3668                          * because we can end up with a circular dependency
3669                          * of renames, resulting in some directories never
3670                          * getting the respective rename operations issued in
3671                          * the send stream or getting into infinite path build
3672                          * loops.
3673                          */
3674                         ret = is_ancestor(sctx->parent_root,
3675                                           sctx->cur_ino, sctx->cur_inode_gen,
3676                                           ino, path_before);
3677                         if (ret)
3678                                 break;
3679                 }
3680 
3681                 fs_path_reset(path_before);
3682                 fs_path_reset(path_after);
3683 
3684                 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3685                                     &parent_ino_after_gen, path_after);
3686                 if (ret < 0)
3687                         goto out;
3688                 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3689                                     NULL, path_before);
3690                 if (ret < 0 && ret != -ENOENT) {
3691                         goto out;
3692                 } else if (ret == -ENOENT) {
3693                         ret = 0;
3694                         break;
3695                 }
3696 
3697                 len1 = fs_path_len(path_before);
3698                 len2 = fs_path_len(path_after);
3699                 if (ino > sctx->cur_ino &&
3700                     (parent_ino_before != parent_ino_after || len1 != len2 ||
3701                      memcmp(path_before->start, path_after->start, len1))) {
3702                         u64 parent_ino_gen;
3703 
3704                         ret = get_inode_info(sctx->parent_root, ino, NULL,
3705                                              &parent_ino_gen, NULL, NULL, NULL,
3706                                              NULL);
3707                         if (ret < 0)
3708                                 goto out;
3709                         if (ino_gen == parent_ino_gen) {
3710                                 ret = 1;
3711                                 break;
3712                         }
3713                 }
3714                 ino = parent_ino_after;
3715                 ino_gen = parent_ino_after_gen;
3716         }
3717 
3718 out:
3719         fs_path_free(path_before);
3720         fs_path_free(path_after);
3721 
3722         if (ret == 1) {
3723                 ret = add_pending_dir_move(sctx,
3724                                            sctx->cur_ino,
3725                                            sctx->cur_inode_gen,
3726                                            ino,
3727                                            &sctx->new_refs,
3728                                            &sctx->deleted_refs,
3729                                            is_orphan);
3730                 if (!ret)
3731                         ret = 1;
3732         }
3733 
3734         return ret;
3735 }
3736 
3737 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3738 {
3739         int ret;
3740         struct fs_path *new_path;
3741 
3742         /*
3743          * Our reference's name member points to its full_path member string, so
3744          * we use here a new path.
3745          */
3746         new_path = fs_path_alloc();
3747         if (!new_path)
3748                 return -ENOMEM;
3749 
3750         ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
3751         if (ret < 0) {
3752                 fs_path_free(new_path);
3753                 return ret;
3754         }
3755         ret = fs_path_add(new_path, ref->name, ref->name_len);
3756         if (ret < 0) {
3757                 fs_path_free(new_path);
3758                 return ret;
3759         }
3760 
3761         fs_path_free(ref->full_path);
3762         set_ref_path(ref, new_path);
3763 
3764         return 0;
3765 }
3766 
3767 /*
3768  * This does all the move/link/unlink/rmdir magic.
3769  */
3770 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3771 {
3772         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3773         int ret = 0;
3774         struct recorded_ref *cur;
3775         struct recorded_ref *cur2;
3776         struct list_head check_dirs;
3777         struct fs_path *valid_path = NULL;
3778         u64 ow_inode = 0;
3779         u64 ow_gen;
3780         u64 ow_mode;
3781         int did_overwrite = 0;
3782         int is_orphan = 0;
3783         u64 last_dir_ino_rm = 0;
3784         bool can_rename = true;
3785         bool orphanized_dir = false;
3786         bool orphanized_ancestor = false;
3787 
3788         btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3789 
3790         /*
3791          * This should never happen as the root dir always has the same ref
3792          * which is always '..'
3793          */
3794         BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3795         INIT_LIST_HEAD(&check_dirs);
3796 
3797         valid_path = fs_path_alloc();
3798         if (!valid_path) {
3799                 ret = -ENOMEM;
3800                 goto out;
3801         }
3802 
3803         /*
3804          * First, check if the first ref of the current inode was overwritten
3805          * before. If yes, we know that the current inode was already orphanized
3806          * and thus use the orphan name. If not, we can use get_cur_path to
3807          * get the path of the first ref as it would like while receiving at
3808          * this point in time.
3809          * New inodes are always orphan at the beginning, so force to use the
3810          * orphan name in this case.
3811          * The first ref is stored in valid_path and will be updated if it
3812          * gets moved around.
3813          */
3814         if (!sctx->cur_inode_new) {
3815                 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3816                                 sctx->cur_inode_gen);
3817                 if (ret < 0)
3818                         goto out;
3819                 if (ret)
3820                         did_overwrite = 1;
3821         }
3822         if (sctx->cur_inode_new || did_overwrite) {
3823                 ret = gen_unique_name(sctx, sctx->cur_ino,
3824                                 sctx->cur_inode_gen, valid_path);
3825                 if (ret < 0)
3826                         goto out;
3827                 is_orphan = 1;
3828         } else {
3829                 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3830                                 valid_path);
3831                 if (ret < 0)
3832                         goto out;
3833         }
3834 
3835         list_for_each_entry(cur, &sctx->new_refs, list) {
3836                 /*
3837                  * We may have refs where the parent directory does not exist
3838                  * yet. This happens if the parent directories inum is higher
3839                  * than the current inum. To handle this case, we create the
3840                  * parent directory out of order. But we need to check if this
3841                  * did already happen before due to other refs in the same dir.
3842                  */
3843                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3844                 if (ret < 0)
3845                         goto out;
3846                 if (ret == inode_state_will_create) {
3847                         ret = 0;
3848                         /*
3849                          * First check if any of the current inodes refs did
3850                          * already create the dir.
3851                          */
3852                         list_for_each_entry(cur2, &sctx->new_refs, list) {
3853                                 if (cur == cur2)
3854                                         break;
3855                                 if (cur2->dir == cur->dir) {
3856                                         ret = 1;
3857                                         break;
3858                                 }
3859                         }
3860 
3861                         /*
3862                          * If that did not happen, check if a previous inode
3863                          * did already create the dir.
3864                          */
3865                         if (!ret)
3866                                 ret = did_create_dir(sctx, cur->dir);
3867                         if (ret < 0)
3868                                 goto out;
3869                         if (!ret) {
3870                                 ret = send_create_inode(sctx, cur->dir);
3871                                 if (ret < 0)
3872                                         goto out;
3873                         }
3874                 }
3875 
3876                 /*
3877                  * Check if this new ref would overwrite the first ref of
3878                  * another unprocessed inode. If yes, orphanize the
3879                  * overwritten inode. If we find an overwritten ref that is
3880                  * not the first ref, simply unlink it.
3881                  */
3882                 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3883                                 cur->name, cur->name_len,
3884                                 &ow_inode, &ow_gen, &ow_mode);
3885                 if (ret < 0)
3886                         goto out;
3887                 if (ret) {
3888                         ret = is_first_ref(sctx->parent_root,
3889                                            ow_inode, cur->dir, cur->name,
3890                                            cur->name_len);
3891                         if (ret < 0)
3892                                 goto out;
3893                         if (ret) {
3894                                 struct name_cache_entry *nce;
3895                                 struct waiting_dir_move *wdm;
3896 
3897                                 ret = orphanize_inode(sctx, ow_inode, ow_gen,
3898                                                 cur->full_path);
3899                                 if (ret < 0)
3900                                         goto out;
3901                                 if (S_ISDIR(ow_mode))
3902                                         orphanized_dir = true;
3903 
3904                                 /*
3905                                  * If ow_inode has its rename operation delayed
3906                                  * make sure that its orphanized name is used in
3907                                  * the source path when performing its rename
3908                                  * operation.
3909                                  */
3910                                 if (is_waiting_for_move(sctx, ow_inode)) {
3911                                         wdm = get_waiting_dir_move(sctx,
3912                                                                    ow_inode);
3913                                         ASSERT(wdm);
3914                                         wdm->orphanized = true;
3915                                 }
3916 
3917                                 /*
3918                                  * Make sure we clear our orphanized inode's
3919                                  * name from the name cache. This is because the
3920                                  * inode ow_inode might be an ancestor of some
3921                                  * other inode that will be orphanized as well
3922                                  * later and has an inode number greater than
3923                                  * sctx->send_progress. We need to prevent
3924                                  * future name lookups from using the old name
3925                                  * and get instead the orphan name.
3926                                  */
3927                                 nce = name_cache_search(sctx, ow_inode, ow_gen);
3928                                 if (nce) {
3929                                         name_cache_delete(sctx, nce);
3930                                         kfree(nce);
3931                                 }
3932 
3933                                 /*
3934                                  * ow_inode might currently be an ancestor of
3935                                  * cur_ino, therefore compute valid_path (the
3936                                  * current path of cur_ino) again because it
3937                                  * might contain the pre-orphanization name of
3938                                  * ow_inode, which is no longer valid.
3939                                  */
3940                                 ret = is_ancestor(sctx->parent_root,
3941                                                   ow_inode, ow_gen,
3942                                                   sctx->cur_ino, NULL);
3943                                 if (ret > 0) {
3944                                         orphanized_ancestor = true;
3945                                         fs_path_reset(valid_path);
3946                                         ret = get_cur_path(sctx, sctx->cur_ino,
3947                                                            sctx->cur_inode_gen,
3948                                                            valid_path);
3949                                 }
3950                                 if (ret < 0)
3951                                         goto out;
3952                         } else {
3953                                 ret = send_unlink(sctx, cur->full_path);
3954                                 if (ret < 0)
3955                                         goto out;
3956                         }
3957                 }
3958 
3959                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
3960                         ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
3961                         if (ret < 0)
3962                                 goto out;
3963                         if (ret == 1) {
3964                                 can_rename = false;
3965                                 *pending_move = 1;
3966                         }
3967                 }
3968 
3969                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
3970                     can_rename) {
3971                         ret = wait_for_parent_move(sctx, cur, is_orphan);
3972                         if (ret < 0)
3973                                 goto out;
3974                         if (ret == 1) {
3975                                 can_rename = false;
3976                                 *pending_move = 1;
3977                         }
3978                 }
3979 
3980                 /*
3981                  * link/move the ref to the new place. If we have an orphan
3982                  * inode, move it and update valid_path. If not, link or move
3983                  * it depending on the inode mode.
3984                  */
3985                 if (is_orphan && can_rename) {
3986                         ret = send_rename(sctx, valid_path, cur->full_path);
3987                         if (ret < 0)
3988                                 goto out;
3989                         is_orphan = 0;
3990                         ret = fs_path_copy(valid_path, cur->full_path);
3991                         if (ret < 0)
3992                                 goto out;
3993                 } else if (can_rename) {
3994                         if (S_ISDIR(sctx->cur_inode_mode)) {
3995                                 /*
3996                                  * Dirs can't be linked, so move it. For moved
3997                                  * dirs, we always have one new and one deleted
3998                                  * ref. The deleted ref is ignored later.
3999                                  */
4000                                 ret = send_rename(sctx, valid_path,
4001                                                   cur->full_path);
4002                                 if (!ret)
4003                                         ret = fs_path_copy(valid_path,
4004                                                            cur->full_path);
4005                                 if (ret < 0)
4006                                         goto out;
4007                         } else {
4008                                 /*
4009                                  * We might have previously orphanized an inode
4010                                  * which is an ancestor of our current inode,
4011                                  * so our reference's full path, which was
4012                                  * computed before any such orphanizations, must
4013                                  * be updated.
4014                                  */
4015                                 if (orphanized_dir) {
4016                                         ret = update_ref_path(sctx, cur);
4017                                         if (ret < 0)
4018                                                 goto out;
4019                                 }
4020                                 ret = send_link(sctx, cur->full_path,
4021                                                 valid_path);
4022                                 if (ret < 0)
4023                                         goto out;
4024                         }
4025                 }
4026                 ret = dup_ref(cur, &check_dirs);
4027                 if (ret < 0)
4028                         goto out;
4029         }
4030 
4031         if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4032                 /*
4033                  * Check if we can already rmdir the directory. If not,
4034                  * orphanize it. For every dir item inside that gets deleted
4035                  * later, we do this check again and rmdir it then if possible.
4036                  * See the use of check_dirs for more details.
4037                  */
4038                 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4039                                 sctx->cur_ino);
4040                 if (ret < 0)
4041                         goto out;
4042                 if (ret) {
4043                         ret = send_rmdir(sctx, valid_path);
4044                         if (ret < 0)
4045                                 goto out;
4046                 } else if (!is_orphan) {
4047                         ret = orphanize_inode(sctx, sctx->cur_ino,
4048                                         sctx->cur_inode_gen, valid_path);
4049                         if (ret < 0)
4050                                 goto out;
4051                         is_orphan = 1;
4052                 }
4053 
4054                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4055                         ret = dup_ref(cur, &check_dirs);
4056                         if (ret < 0)
4057                                 goto out;
4058                 }
4059         } else if (S_ISDIR(sctx->cur_inode_mode) &&
4060                    !list_empty(&sctx->deleted_refs)) {
4061                 /*
4062                  * We have a moved dir. Add the old parent to check_dirs
4063                  */
4064                 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4065                                 list);
4066                 ret = dup_ref(cur, &check_dirs);
4067                 if (ret < 0)
4068                         goto out;
4069         } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4070                 /*
4071                  * We have a non dir inode. Go through all deleted refs and
4072                  * unlink them if they were not already overwritten by other
4073                  * inodes.
4074                  */
4075                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4076                         ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4077                                         sctx->cur_ino, sctx->cur_inode_gen,
4078                                         cur->name, cur->name_len);
4079                         if (ret < 0)
4080                                 goto out;
4081                         if (!ret) {
4082                                 /*
4083                                  * If we orphanized any ancestor before, we need
4084                                  * to recompute the full path for deleted names,
4085                                  * since any such path was computed before we
4086                                  * processed any references and orphanized any
4087                                  * ancestor inode.
4088                                  */
4089                                 if (orphanized_ancestor) {
4090                                         ret = update_ref_path(sctx, cur);
4091                                         if (ret < 0)
4092                                                 goto out;
4093                                 }
4094                                 ret = send_unlink(sctx, cur->full_path);
4095                                 if (ret < 0)
4096                                         goto out;
4097                         }
4098                         ret = dup_ref(cur, &check_dirs);
4099                         if (ret < 0)
4100                                 goto out;
4101                 }
4102                 /*
4103                  * If the inode is still orphan, unlink the orphan. This may
4104                  * happen when a previous inode did overwrite the first ref
4105                  * of this inode and no new refs were added for the current
4106                  * inode. Unlinking does not mean that the inode is deleted in
4107                  * all cases. There may still be links to this inode in other
4108                  * places.
4109                  */
4110                 if (is_orphan) {
4111                         ret = send_unlink(sctx, valid_path);
4112                         if (ret < 0)
4113                                 goto out;
4114                 }
4115         }
4116 
4117         /*
4118          * We did collect all parent dirs where cur_inode was once located. We
4119          * now go through all these dirs and check if they are pending for
4120          * deletion and if it's finally possible to perform the rmdir now.
4121          * We also update the inode stats of the parent dirs here.
4122          */
4123         list_for_each_entry(cur, &check_dirs, list) {
4124                 /*
4125                  * In case we had refs into dirs that were not processed yet,
4126                  * we don't need to do the utime and rmdir logic for these dirs.
4127                  * The dir will be processed later.
4128                  */
4129                 if (cur->dir > sctx->cur_ino)
4130                         continue;
4131 
4132                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4133                 if (ret < 0)
4134                         goto out;
4135 
4136                 if (ret == inode_state_did_create ||
4137                     ret == inode_state_no_change) {
4138                         /* TODO delayed utimes */
4139                         ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4140                         if (ret < 0)
4141                                 goto out;
4142                 } else if (ret == inode_state_did_delete &&
4143                            cur->dir != last_dir_ino_rm) {
4144                         ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4145                                         sctx->cur_ino);
4146                         if (ret < 0)
4147                                 goto out;
4148                         if (ret) {
4149                                 ret = get_cur_path(sctx, cur->dir,
4150                                                    cur->dir_gen, valid_path);
4151                                 if (ret < 0)
4152                                         goto out;
4153                                 ret = send_rmdir(sctx, valid_path);
4154                                 if (ret < 0)
4155                                         goto out;
4156                                 last_dir_ino_rm = cur->dir;
4157                         }
4158                 }
4159         }
4160 
4161         ret = 0;
4162 
4163 out:
4164         __free_recorded_refs(&check_dirs);
4165         free_recorded_refs(sctx);
4166         fs_path_free(valid_path);
4167         return ret;
4168 }
4169 
4170 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name,
4171                       void *ctx, struct list_head *refs)
4172 {
4173         int ret = 0;
4174         struct send_ctx *sctx = ctx;
4175         struct fs_path *p;
4176         u64 gen;
4177 
4178         p = fs_path_alloc();
4179         if (!p)
4180                 return -ENOMEM;
4181 
4182         ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4183                         NULL, NULL);
4184         if (ret < 0)
4185                 goto out;
4186 
4187         ret = get_cur_path(sctx, dir, gen, p);
4188         if (ret < 0)
4189                 goto out;
4190         ret = fs_path_add_path(p, name);
4191         if (ret < 0)
4192                 goto out;
4193 
4194         ret = __record_ref(refs, dir, gen, p);
4195 
4196 out:
4197         if (ret)
4198                 fs_path_free(p);
4199         return ret;
4200 }
4201 
4202 static int __record_new_ref(int num, u64 dir, int index,
4203                             struct fs_path *name,
4204                             void *ctx)
4205 {
4206         struct send_ctx *sctx = ctx;
4207         return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs);
4208 }
4209 
4210 
4211 static int __record_deleted_ref(int num, u64 dir, int index,
4212                                 struct fs_path *name,
4213                                 void *ctx)
4214 {
4215         struct send_ctx *sctx = ctx;
4216         return record_ref(sctx->parent_root, dir, name, ctx,
4217                           &sctx->deleted_refs);
4218 }
4219 
4220 static int record_new_ref(struct send_ctx *sctx)
4221 {
4222         int ret;
4223 
4224         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4225                                 sctx->cmp_key, 0, __record_new_ref, sctx);
4226         if (ret < 0)
4227                 goto out;
4228         ret = 0;
4229 
4230 out:
4231         return ret;
4232 }
4233 
4234 static int record_deleted_ref(struct send_ctx *sctx)
4235 {
4236         int ret;
4237 
4238         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4239                                 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4240         if (ret < 0)
4241                 goto out;
4242         ret = 0;
4243 
4244 out:
4245         return ret;
4246 }
4247 
4248 struct find_ref_ctx {
4249         u64 dir;
4250         u64 dir_gen;
4251         struct btrfs_root *root;
4252         struct fs_path *name;
4253         int found_idx;
4254 };
4255 
4256 static int __find_iref(int num, u64 dir, int index,
4257                        struct fs_path *name,
4258                        void *ctx_)
4259 {
4260         struct find_ref_ctx *ctx = ctx_;
4261         u64 dir_gen;
4262         int ret;
4263 
4264         if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4265             strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4266                 /*
4267                  * To avoid doing extra lookups we'll only do this if everything
4268                  * else matches.
4269                  */
4270                 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4271                                      NULL, NULL, NULL);
4272                 if (ret)
4273                         return ret;
4274                 if (dir_gen != ctx->dir_gen)
4275                         return 0;
4276                 ctx->found_idx = num;
4277                 return 1;
4278         }
4279         return 0;
4280 }
4281 
4282 static int find_iref(struct btrfs_root *root,
4283                      struct btrfs_path *path,
4284                      struct btrfs_key *key,
4285                      u64 dir, u64 dir_gen, struct fs_path *name)
4286 {
4287         int ret;
4288         struct find_ref_ctx ctx;
4289 
4290         ctx.dir = dir;
4291         ctx.name = name;
4292         ctx.dir_gen = dir_gen;
4293         ctx.found_idx = -1;
4294         ctx.root = root;
4295 
4296         ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4297         if (ret < 0)
4298                 return ret;
4299 
4300         if (ctx.found_idx == -1)
4301                 return -ENOENT;
4302 
4303         return ctx.found_idx;
4304 }
4305 
4306 static int __record_changed_new_ref(int num, u64 dir, int index,
4307                                     struct fs_path *name,
4308                                     void *ctx)
4309 {
4310         u64 dir_gen;
4311         int ret;
4312         struct send_ctx *sctx = ctx;
4313 
4314         ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4315                              NULL, NULL, NULL);
4316         if (ret)
4317                 return ret;
4318 
4319         ret = find_iref(sctx->parent_root, sctx->right_path,
4320                         sctx->cmp_key, dir, dir_gen, name);
4321         if (ret == -ENOENT)
4322                 ret = __record_new_ref(num, dir, index, name, sctx);
4323         else if (ret > 0)
4324                 ret = 0;
4325 
4326         return ret;
4327 }
4328 
4329 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4330                                         struct fs_path *name,
4331                                         void *ctx)
4332 {
4333         u64 dir_gen;
4334         int ret;
4335         struct send_ctx *sctx = ctx;
4336 
4337         ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4338                              NULL, NULL, NULL);
4339         if (ret)
4340                 return ret;
4341 
4342         ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4343                         dir, dir_gen, name);
4344         if (ret == -ENOENT)
4345                 ret = __record_deleted_ref(num, dir, index, name, sctx);
4346         else if (ret > 0)
4347                 ret = 0;
4348 
4349         return ret;
4350 }
4351 
4352 static int record_changed_ref(struct send_ctx *sctx)
4353 {
4354         int ret = 0;
4355 
4356         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4357                         sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4358         if (ret < 0)
4359                 goto out;
4360         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4361                         sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4362         if (ret < 0)
4363                 goto out;
4364         ret = 0;
4365 
4366 out:
4367         return ret;
4368 }
4369 
4370 /*
4371  * Record and process all refs at once. Needed when an inode changes the
4372  * generation number, which means that it was deleted and recreated.
4373  */
4374 static int process_all_refs(struct send_ctx *sctx,
4375                             enum btrfs_compare_tree_result cmd)
4376 {
4377         int ret;
4378         struct btrfs_root *root;
4379         struct btrfs_path *path;
4380         struct btrfs_key key;
4381         struct btrfs_key found_key;
4382         struct extent_buffer *eb;
4383         int slot;
4384         iterate_inode_ref_t cb;
4385         int pending_move = 0;
4386 
4387         path = alloc_path_for_send();
4388         if (!path)
4389                 return -ENOMEM;
4390 
4391         if (cmd == BTRFS_COMPARE_TREE_NEW) {
4392                 root = sctx->send_root;
4393                 cb = __record_new_ref;
4394         } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4395                 root = sctx->parent_root;
4396                 cb = __record_deleted_ref;
4397         } else {
4398                 btrfs_err(sctx->send_root->fs_info,
4399                                 "Wrong command %d in process_all_refs", cmd);
4400                 ret = -EINVAL;
4401                 goto out;
4402         }
4403 
4404         key.objectid = sctx->cmp_key->objectid;
4405         key.type = BTRFS_INODE_REF_KEY;
4406         key.offset = 0;
4407         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4408         if (ret < 0)
4409                 goto out;
4410 
4411         while (1) {
4412                 eb = path->nodes[0];
4413                 slot = path->slots[0];
4414                 if (slot >= btrfs_header_nritems(eb)) {
4415                         ret = btrfs_next_leaf(root, path);
4416                         if (ret < 0)
4417                                 goto out;
4418                         else if (ret > 0)
4419                                 break;
4420                         continue;
4421                 }
4422 
4423                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4424 
4425                 if (found_key.objectid != key.objectid ||
4426                     (found_key.type != BTRFS_INODE_REF_KEY &&
4427                      found_key.type != BTRFS_INODE_EXTREF_KEY))
4428                         break;
4429 
4430                 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4431                 if (ret < 0)
4432                         goto out;
4433 
4434                 path->slots[0]++;
4435         }
4436         btrfs_release_path(path);
4437 
4438         /*
4439          * We don't actually care about pending_move as we are simply
4440          * re-creating this inode and will be rename'ing it into place once we
4441          * rename the parent directory.
4442          */
4443         ret = process_recorded_refs(sctx, &pending_move);
4444 out:
4445         btrfs_free_path(path);
4446         return ret;
4447 }
4448 
4449 static int send_set_xattr(struct send_ctx *sctx,
4450                           struct fs_path *path,
4451                           const char *name, int name_len,
4452                           const char *data, int data_len)
4453 {
4454         int ret = 0;
4455 
4456         ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4457         if (ret < 0)
4458                 goto out;
4459 
4460         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4461         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4462         TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4463 
4464         ret = send_cmd(sctx);
4465 
4466 tlv_put_failure:
4467 out:
4468         return ret;
4469 }
4470 
4471 static int send_remove_xattr(struct send_ctx *sctx,
4472                           struct fs_path *path,
4473                           const char *name, int name_len)
4474 {
4475         int ret = 0;
4476 
4477         ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4478         if (ret < 0)
4479                 goto out;
4480 
4481         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4482         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4483 
4484         ret = send_cmd(sctx);
4485 
4486 tlv_put_failure:
4487 out:
4488         return ret;
4489 }
4490 
4491 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4492                                const char *name, int name_len,
4493                                const char *data, int data_len,
4494                                u8 type, void *ctx)
4495 {
4496         int ret;
4497         struct send_ctx *sctx = ctx;
4498         struct fs_path *p;
4499         struct posix_acl_xattr_header dummy_acl;
4500 
4501         p = fs_path_alloc();
4502         if (!p)
4503                 return -ENOMEM;
4504 
4505         /*
4506          * This hack is needed because empty acls are stored as zero byte
4507          * data in xattrs. Problem with that is, that receiving these zero byte
4508          * acls will fail later. To fix this, we send a dummy acl list that
4509          * only contains the version number and no entries.
4510          */
4511         if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4512             !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4513                 if (data_len == 0) {
4514                         dummy_acl.a_version =
4515                                         cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4516                         data = (char *)&dummy_acl;
4517                         data_len = sizeof(dummy_acl);
4518                 }
4519         }
4520 
4521         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4522         if (ret < 0)
4523                 goto out;
4524 
4525         ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4526 
4527 out:
4528         fs_path_free(p);
4529         return ret;
4530 }
4531 
4532 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4533                                    const char *name, int name_len,
4534                                    const char *data, int data_len,
4535                                    u8 type, void *ctx)
4536 {
4537         int ret;
4538         struct send_ctx *sctx = ctx;
4539         struct fs_path *p;
4540 
4541         p = fs_path_alloc();
4542         if (!p)
4543                 return -ENOMEM;
4544 
4545         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4546         if (ret < 0)
4547                 goto out;
4548 
4549         ret = send_remove_xattr(sctx, p, name, name_len);
4550 
4551 out:
4552         fs_path_free(p);
4553         return ret;
4554 }
4555 
4556 static int process_new_xattr(struct send_ctx *sctx)
4557 {
4558         int ret = 0;
4559 
4560         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4561                                __process_new_xattr, sctx);
4562 
4563         return ret;
4564 }
4565 
4566 static int process_deleted_xattr(struct send_ctx *sctx)
4567 {
4568         return iterate_dir_item(sctx->parent_root, sctx->right_path,
4569                                 __process_deleted_xattr, sctx);
4570 }
4571 
4572 struct find_xattr_ctx {
4573         const char *name;
4574         int name_len;
4575         int found_idx;
4576         char *found_data;
4577         int found_data_len;
4578 };
4579 
4580 static int __find_xattr(int num, struct btrfs_key *di_key,
4581                         const char *name, int name_len,
4582                         const char *data, int data_len,
4583                         u8 type, void *vctx)
4584 {
4585         struct find_xattr_ctx *ctx = vctx;
4586 
4587         if (name_len == ctx->name_len &&
4588             strncmp(name, ctx->name, name_len) == 0) {
4589                 ctx->found_idx = num;
4590                 ctx->found_data_len = data_len;
4591                 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4592                 if (!ctx->found_data)
4593                         return -ENOMEM;
4594                 return 1;
4595         }
4596         return 0;
4597 }
4598 
4599 static int find_xattr(struct btrfs_root *root,
4600                       struct btrfs_path *path,
4601                       struct btrfs_key *key,
4602                       const char *name, int name_len,
4603                       char **data, int *data_len)
4604 {
4605         int ret;
4606         struct find_xattr_ctx ctx;
4607 
4608         ctx.name = name;
4609         ctx.name_len = name_len;
4610         ctx.found_idx = -1;
4611         ctx.found_data = NULL;
4612         ctx.found_data_len = 0;
4613 
4614         ret = iterate_dir_item(root, path, __find_xattr, &ctx);
4615         if (ret < 0)
4616                 return ret;
4617 
4618         if (ctx.found_idx == -1)
4619                 return -ENOENT;
4620         if (data) {
4621                 *data = ctx.found_data;
4622                 *data_len = ctx.found_data_len;
4623         } else {
4624                 kfree(ctx.found_data);
4625         }
4626         return ctx.found_idx;
4627 }
4628 
4629 
4630 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4631                                        const char *name, int name_len,
4632                                        const char *data, int data_len,
4633                                        u8 type, void *ctx)
4634 {
4635         int ret;
4636         struct send_ctx *sctx = ctx;
4637         char *found_data = NULL;
4638         int found_data_len  = 0;
4639 
4640         ret = find_xattr(sctx->parent_root, sctx->right_path,
4641                          sctx->cmp_key, name, name_len, &found_data,
4642                          &found_data_len);
4643         if (ret == -ENOENT) {
4644                 ret = __process_new_xattr(num, di_key, name, name_len, data,
4645                                 data_len, type, ctx);
4646         } else if (ret >= 0) {
4647                 if (data_len != found_data_len ||
4648                     memcmp(data, found_data, data_len)) {
4649                         ret = __process_new_xattr(num, di_key, name, name_len,
4650                                         data, data_len, type, ctx);
4651                 } else {
4652                         ret = 0;
4653                 }
4654         }
4655 
4656         kfree(found_data);
4657         return ret;
4658 }
4659 
4660 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4661                                            const char *name, int name_len,
4662                                            const char *data, int data_len,
4663                                            u8 type, void *ctx)
4664 {
4665         int ret;
4666         struct send_ctx *sctx = ctx;
4667 
4668         ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4669                          name, name_len, NULL, NULL);
4670         if (ret == -ENOENT)
4671                 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4672                                 data_len, type, ctx);
4673         else if (ret >= 0)
4674                 ret = 0;
4675 
4676         return ret;
4677 }
4678 
4679 static int process_changed_xattr(struct send_ctx *sctx)
4680 {
4681         int ret = 0;
4682 
4683         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4684                         __process_changed_new_xattr, sctx);
4685         if (ret < 0)
4686                 goto out;
4687         ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4688                         __process_changed_deleted_xattr, sctx);
4689 
4690 out:
4691         return ret;
4692 }
4693 
4694 static int process_all_new_xattrs(struct send_ctx *sctx)
4695 {
4696         int ret;
4697         struct btrfs_root *root;
4698         struct btrfs_path *path;
4699         struct btrfs_key key;
4700         struct btrfs_key found_key;
4701         struct extent_buffer *eb;
4702         int slot;
4703 
4704         path = alloc_path_for_send();
4705         if (!path)
4706                 return -ENOMEM;
4707 
4708         root = sctx->send_root;
4709 
4710         key.objectid = sctx->cmp_key->objectid;
4711         key.type = BTRFS_XATTR_ITEM_KEY;
4712         key.offset = 0;
4713         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4714         if (ret < 0)
4715                 goto out;
4716 
4717         while (1) {
4718                 eb = path->nodes[0];
4719                 slot = path->slots[0];
4720                 if (slot >= btrfs_header_nritems(eb)) {
4721                         ret = btrfs_next_leaf(root, path);
4722                         if (ret < 0) {
4723                                 goto out;
4724                         } else if (ret > 0) {
4725                                 ret = 0;
4726                                 break;
4727                         }
4728                         continue;
4729                 }
4730 
4731                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4732                 if (found_key.objectid != key.objectid ||
4733                     found_key.type != key.type) {
4734                         ret = 0;
4735                         goto out;
4736                 }
4737 
4738                 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
4739                 if (ret < 0)
4740                         goto out;
4741 
4742                 path->slots[0]++;
4743         }
4744 
4745 out:
4746         btrfs_free_path(path);
4747         return ret;
4748 }
4749 
4750 static ssize_t fill_read_buf(struct send_ctx *sctx, u64 offset, u32 len)
4751 {
4752         struct btrfs_root *root = sctx->send_root;
4753         struct btrfs_fs_info *fs_info = root->fs_info;
4754         struct inode *inode;
4755         struct page *page;
4756         char *addr;
4757         struct btrfs_key key;
4758         pgoff_t index = offset >> PAGE_SHIFT;
4759         pgoff_t last_index;
4760         unsigned pg_offset = offset_in_page(offset);
4761         ssize_t ret = 0;
4762 
4763         key.objectid = sctx->cur_ino;
4764         key.type = BTRFS_INODE_ITEM_KEY;
4765         key.offset = 0;
4766 
4767         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
4768         if (IS_ERR(inode))
4769                 return PTR_ERR(inode);
4770 
4771         if (offset + len > i_size_read(inode)) {
4772                 if (offset > i_size_read(inode))
4773                         len = 0;
4774                 else
4775                         len = offset - i_size_read(inode);
4776         }
4777         if (len == 0)
4778                 goto out;
4779 
4780         last_index = (offset + len - 1) >> PAGE_SHIFT;
4781 
4782         /* initial readahead */
4783         memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4784         file_ra_state_init(&sctx->ra, inode->i_mapping);
4785 
4786         while (index <= last_index) {
4787                 unsigned cur_len = min_t(unsigned, len,
4788                                          PAGE_SIZE - pg_offset);
4789 
4790                 page = find_lock_page(inode->i_mapping, index);
4791                 if (!page) {
4792                         page_cache_sync_readahead(inode->i_mapping, &sctx->ra,
4793                                 NULL, index, last_index + 1 - index);
4794 
4795                         page = find_or_create_page(inode->i_mapping, index,
4796                                         GFP_KERNEL);
4797                         if (!page) {
4798                                 ret = -ENOMEM;
4799                                 break;
4800                         }
4801                 }
4802 
4803                 if (PageReadahead(page)) {
4804                         page_cache_async_readahead(inode->i_mapping, &sctx->ra,
4805                                 NULL, page, index, last_index + 1 - index);
4806                 }
4807 
4808                 if (!PageUptodate(page)) {
4809                         btrfs_readpage(NULL, page);
4810                         lock_page(page);
4811                         if (!PageUptodate(page)) {
4812                                 unlock_page(page);
4813                                 put_page(page);
4814                                 ret = -EIO;
4815                                 break;
4816                         }
4817                 }
4818 
4819                 addr = kmap(page);
4820                 memcpy(sctx->read_buf + ret, addr + pg_offset, cur_len);
4821                 kunmap(page);
4822                 unlock_page(page);
4823                 put_page(page);
4824                 index++;
4825                 pg_offset = 0;
4826                 len -= cur_len;
4827                 ret += cur_len;
4828         }
4829 out:
4830         iput(inode);
4831         return ret;
4832 }
4833 
4834 /*
4835  * Read some bytes from the current inode/file and send a write command to
4836  * user space.
4837  */
4838 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
4839 {
4840         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4841         int ret = 0;
4842         struct fs_path *p;
4843         ssize_t num_read = 0;
4844 
4845         p = fs_path_alloc();
4846         if (!p)
4847                 return -ENOMEM;
4848 
4849         btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
4850 
4851         num_read = fill_read_buf(sctx, offset, len);
4852         if (num_read <= 0) {
4853                 if (num_read < 0)
4854                         ret = num_read;
4855                 goto out;
4856         }
4857 
4858         ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4859         if (ret < 0)
4860                 goto out;
4861 
4862         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4863         if (ret < 0)
4864                 goto out;
4865 
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, num_read);
4869 
4870         ret = send_cmd(sctx);
4871 
4872 tlv_put_failure:
4873 out:
4874         fs_path_free(p);
4875         if (ret < 0)
4876                 return ret;
4877         return num_read;
4878 }
4879 
4880 /*
4881  * Send a clone command to user space.
4882  */
4883 static int send_clone(struct send_ctx *sctx,
4884                       u64 offset, u32 len,
4885                       struct clone_root *clone_root)
4886 {
4887         int ret = 0;
4888         struct fs_path *p;
4889         u64 gen;
4890 
4891         btrfs_debug(sctx->send_root->fs_info,
4892                     "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
4893                     offset, len, clone_root->root->root_key.objectid,
4894                     clone_root->ino, clone_root->offset);
4895 
4896         p = fs_path_alloc();
4897         if (!p)
4898                 return -ENOMEM;
4899 
4900         ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
4901         if (ret < 0)
4902                 goto out;
4903 
4904         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4905         if (ret < 0)
4906                 goto out;
4907 
4908         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4909         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
4910         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4911 
4912         if (clone_root->root == sctx->send_root) {
4913                 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
4914                                 &gen, NULL, NULL, NULL, NULL);
4915                 if (ret < 0)
4916                         goto out;
4917                 ret = get_cur_path(sctx, clone_root->ino, gen, p);
4918         } else {
4919                 ret = get_inode_path(clone_root->root, clone_root->ino, p);
4920         }
4921         if (ret < 0)
4922                 goto out;
4923 
4924         /*
4925          * If the parent we're using has a received_uuid set then use that as
4926          * our clone source as that is what we will look for when doing a
4927          * receive.
4928          *
4929          * This covers the case that we create a snapshot off of a received
4930          * subvolume and then use that as the parent and try to receive on a
4931          * different host.
4932          */
4933         if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
4934                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4935                              clone_root->root->root_item.received_uuid);
4936         else
4937                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4938                              clone_root->root->root_item.uuid);
4939         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
4940                     le64_to_cpu(clone_root->root->root_item.ctransid));
4941         TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
4942         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
4943                         clone_root->offset);
4944 
4945         ret = send_cmd(sctx);
4946 
4947 tlv_put_failure:
4948 out:
4949         fs_path_free(p);
4950         return ret;
4951 }
4952 
4953 /*
4954  * Send an update extent command to user space.
4955  */
4956 static int send_update_extent(struct send_ctx *sctx,
4957                               u64 offset, u32 len)
4958 {
4959         int ret = 0;
4960         struct fs_path *p;
4961 
4962         p = fs_path_alloc();
4963         if (!p)
4964                 return -ENOMEM;
4965 
4966         ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
4967         if (ret < 0)
4968                 goto out;
4969 
4970         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4971         if (ret < 0)
4972                 goto out;
4973 
4974         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4975         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4976         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
4977 
4978         ret = send_cmd(sctx);
4979 
4980 tlv_put_failure:
4981 out:
4982         fs_path_free(p);
4983         return ret;
4984 }
4985 
4986 static int send_hole(struct send_ctx *sctx, u64 end)
4987 {
4988         struct fs_path *p = NULL;
4989         u64 offset = sctx->cur_inode_last_extent;
4990         u64 len;
4991         int ret = 0;
4992 
4993         /*
4994          * A hole that starts at EOF or beyond it. Since we do not yet support
4995          * fallocate (for extent preallocation and hole punching), sending a
4996          * write of zeroes starting at EOF or beyond would later require issuing
4997          * a truncate operation which would undo the write and achieve nothing.
4998          */
4999         if (offset >= sctx->cur_inode_size)
5000                 return 0;
5001 
5002         /*
5003          * Don't go beyond the inode's i_size due to prealloc extents that start
5004          * after the i_size.
5005          */
5006         end = min_t(u64, end, sctx->cur_inode_size);
5007 
5008         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5009                 return send_update_extent(sctx, offset, end - offset);
5010 
5011         p = fs_path_alloc();
5012         if (!p)
5013                 return -ENOMEM;
5014         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5015         if (ret < 0)
5016                 goto tlv_put_failure;
5017         memset(sctx->read_buf, 0, BTRFS_SEND_READ_SIZE);
5018         while (offset < end) {
5019                 len = min_t(u64, end - offset, BTRFS_SEND_READ_SIZE);
5020 
5021                 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5022                 if (ret < 0)
5023                         break;
5024                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5025                 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5026                 TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, len);
5027                 ret = send_cmd(sctx);
5028                 if (ret < 0)
5029                         break;
5030                 offset += len;
5031         }
5032         sctx->cur_inode_next_write_offset = offset;
5033 tlv_put_failure:
5034         fs_path_free(p);
5035         return ret;
5036 }
5037 
5038 static int send_extent_data(struct send_ctx *sctx,
5039                             const u64 offset,
5040                             const u64 len)
5041 {
5042         u64 sent = 0;
5043 
5044         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5045                 return send_update_extent(sctx, offset, len);
5046 
5047         while (sent < len) {
5048                 u64 size = len - sent;
5049                 int ret;
5050 
5051                 if (size > BTRFS_SEND_READ_SIZE)
5052                         size = BTRFS_SEND_READ_SIZE;
5053                 ret = send_write(sctx, offset + sent, size);
5054                 if (ret < 0)
5055                         return ret;
5056                 if (!ret)
5057                         break;
5058                 sent += ret;
5059         }
5060         return 0;
5061 }
5062 
5063 static int clone_range(struct send_ctx *sctx,
5064                        struct clone_root *clone_root,
5065                        const u64 disk_byte,
5066                        u64 data_offset,
5067                        u64 offset,
5068                        u64 len)
5069 {
5070         struct btrfs_path *path;
5071         struct btrfs_key key;
5072         int ret;
5073         u64 clone_src_i_size;
5074 
5075         /*
5076          * Prevent cloning from a zero offset with a length matching the sector
5077          * size because in some scenarios this will make the receiver fail.
5078          *
5079          * For example, if in the source filesystem the extent at offset 0
5080          * has a length of sectorsize and it was written using direct IO, then
5081          * it can never be an inline extent (even if compression is enabled).
5082          * Then this extent can be cloned in the original filesystem to a non
5083          * zero file offset, but it may not be possible to clone in the
5084          * destination filesystem because it can be inlined due to compression
5085          * on the destination filesystem (as the receiver's write operations are
5086          * always done using buffered IO). The same happens when the original
5087          * filesystem does not have compression enabled but the destination
5088          * filesystem has.
5089          */
5090         if (clone_root->offset == 0 &&
5091             len == sctx->send_root->fs_info->sectorsize)
5092                 return send_extent_data(sctx, offset, len);
5093 
5094         path = alloc_path_for_send();
5095         if (!path)
5096                 return -ENOMEM;
5097 
5098         /*
5099          * There are inodes that have extents that lie behind its i_size. Don't
5100          * accept clones from these extents.
5101          */
5102         ret = __get_inode_info(clone_root->root, path, clone_root->ino,
5103                                &clone_src_i_size, NULL, NULL, NULL, NULL, NULL);
5104         btrfs_release_path(path);
5105         if (ret < 0)
5106                 goto out;
5107 
5108         /*
5109          * We can't send a clone operation for the entire range if we find
5110          * extent items in the respective range in the source file that
5111          * refer to different extents or if we find holes.
5112          * So check for that and do a mix of clone and regular write/copy
5113          * operations if needed.
5114          *
5115          * Example:
5116          *
5117          * mkfs.btrfs -f /dev/sda
5118          * mount /dev/sda /mnt
5119          * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5120          * cp --reflink=always /mnt/foo /mnt/bar
5121          * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5122          * btrfs subvolume snapshot -r /mnt /mnt/snap
5123          *
5124          * If when we send the snapshot and we are processing file bar (which
5125          * has a higher inode number than foo) we blindly send a clone operation
5126          * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5127          * a file bar that matches the content of file foo - iow, doesn't match
5128          * the content from bar in the original filesystem.
5129          */
5130         key.objectid = clone_root->ino;
5131         key.type = BTRFS_EXTENT_DATA_KEY;
5132         key.offset = clone_root->offset;
5133         ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5134         if (ret < 0)
5135                 goto out;
5136         if (ret > 0 && path->slots[0] > 0) {
5137                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5138                 if (key.objectid == clone_root->ino &&
5139                     key.type == BTRFS_EXTENT_DATA_KEY)
5140                         path->slots[0]--;
5141         }
5142 
5143         while (true) {
5144                 struct extent_buffer *leaf = path->nodes[0];
5145                 int slot = path->slots[0];
5146                 struct btrfs_file_extent_item *ei;
5147                 u8 type;
5148                 u64 ext_len;
5149                 u64 clone_len;
5150                 u64 clone_data_offset;
5151 
5152                 if (slot >= btrfs_header_nritems(leaf)) {
5153                         ret = btrfs_next_leaf(clone_root->root, path);
5154                         if (ret < 0)
5155                                 goto out;
5156                         else if (ret > 0)
5157                                 break;
5158                         continue;
5159                 }
5160 
5161                 btrfs_item_key_to_cpu(leaf, &key, slot);
5162 
5163                 /*
5164                  * We might have an implicit trailing hole (NO_HOLES feature
5165                  * enabled). We deal with it after leaving this loop.
5166                  */
5167                 if (key.objectid != clone_root->ino ||
5168                     key.type != BTRFS_EXTENT_DATA_KEY)
5169                         break;
5170 
5171                 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5172                 type = btrfs_file_extent_type(leaf, ei);
5173                 if (type == BTRFS_FILE_EXTENT_INLINE) {
5174                         ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5175                         ext_len = PAGE_ALIGN(ext_len);
5176                 } else {
5177                         ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5178                 }
5179 
5180                 if (key.offset + ext_len <= clone_root->offset)
5181                         goto next;
5182 
5183                 if (key.offset > clone_root->offset) {
5184                         /* Implicit hole, NO_HOLES feature enabled. */
5185                         u64 hole_len = key.offset - clone_root->offset;
5186 
5187                         if (hole_len > len)
5188                                 hole_len = len;
5189                         ret = send_extent_data(sctx, offset, hole_len);
5190                         if (ret < 0)
5191                                 goto out;
5192 
5193                         len -= hole_len;
5194                         if (len == 0)
5195                                 break;
5196                         offset += hole_len;
5197                         clone_root->offset += hole_len;
5198                         data_offset += hole_len;
5199                 }
5200 
5201                 if (key.offset >= clone_root->offset + len)
5202                         break;
5203 
5204                 if (key.offset >= clone_src_i_size)
5205                         break;
5206 
5207                 if (key.offset + ext_len > clone_src_i_size)
5208                         ext_len = clone_src_i_size - key.offset;
5209 
5210                 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
5211                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
5212                         clone_root->offset = key.offset;
5213                         if (clone_data_offset < data_offset &&
5214                                 clone_data_offset + ext_len > data_offset) {
5215                                 u64 extent_offset;
5216 
5217                                 extent_offset = data_offset - clone_data_offset;
5218                                 ext_len -= extent_offset;
5219                                 clone_data_offset += extent_offset;
5220                                 clone_root->offset += extent_offset;
5221                         }
5222                 }
5223 
5224                 clone_len = min_t(u64, ext_len, len);
5225 
5226                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5227                     clone_data_offset == data_offset) {
5228                         const u64 src_end = clone_root->offset + clone_len;
5229                         const u64 sectorsize = SZ_64K;
5230 
5231                         /*
5232                          * We can't clone the last block, when its size is not
5233                          * sector size aligned, into the middle of a file. If we
5234                          * do so, the receiver will get a failure (-EINVAL) when
5235                          * trying to clone or will silently corrupt the data in
5236                          * the destination file if it's on a kernel without the
5237                          * fix introduced by commit ac765f83f1397646
5238                          * ("Btrfs: fix data corruption due to cloning of eof
5239                          * block).
5240                          *
5241                          * So issue a clone of the aligned down range plus a
5242                          * regular write for the eof block, if we hit that case.
5243                          *
5244                          * Also, we use the maximum possible sector size, 64K,
5245                          * because we don't know what's the sector size of the
5246                          * filesystem that receives the stream, so we have to
5247                          * assume the largest possible sector size.
5248                          */
5249                         if (src_end == clone_src_i_size &&
5250                             !IS_ALIGNED(src_end, sectorsize) &&
5251                             offset + clone_len < sctx->cur_inode_size) {
5252                                 u64 slen;
5253 
5254                                 slen = ALIGN_DOWN(src_end - clone_root->offset,
5255                                                   sectorsize);
5256                                 if (slen > 0) {
5257                                         ret = send_clone(sctx, offset, slen,
5258                                                          clone_root);
5259                                         if (ret < 0)
5260                                                 goto out;
5261                                 }
5262                                 ret = send_extent_data(sctx, offset + slen,
5263                                                        clone_len - slen);
5264                         } else {
5265                                 ret = send_clone(sctx, offset, clone_len,
5266                                                  clone_root);
5267                         }
5268                 } else {
5269                         ret = send_extent_data(sctx, offset, clone_len);
5270                 }
5271 
5272                 if (ret < 0)
5273                         goto out;
5274 
5275                 len -= clone_len;
5276                 if (len == 0)
5277                         break;
5278                 offset += clone_len;
5279                 clone_root->offset += clone_len;
5280                 data_offset += clone_len;
5281 next:
5282                 path->slots[0]++;
5283         }
5284 
5285         if (len > 0)
5286                 ret = send_extent_data(sctx, offset, len);
5287         else
5288                 ret = 0;
5289 out:
5290         btrfs_free_path(path);
5291         return ret;
5292 }
5293 
5294 static int send_write_or_clone(struct send_ctx *sctx,
5295                                struct btrfs_path *path,
5296                                struct btrfs_key *key,
5297                                struct clone_root *clone_root)
5298 {
5299         int ret = 0;
5300         struct btrfs_file_extent_item *ei;
5301         u64 offset = key->offset;
5302         u64 len;
5303         u8 type;
5304         u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5305 
5306         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5307                         struct btrfs_file_extent_item);
5308         type = btrfs_file_extent_type(path->nodes[0], ei);
5309         if (type == BTRFS_FILE_EXTENT_INLINE) {
5310                 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
5311                 /*
5312                  * it is possible the inline item won't cover the whole page,
5313                  * but there may be items after this page.  Make
5314                  * sure to send the whole thing
5315                  */
5316                 len = PAGE_ALIGN(len);
5317         } else {
5318                 len = btrfs_file_extent_num_bytes(path->nodes[0], ei);
5319         }
5320 
5321         if (offset >= sctx->cur_inode_size) {
5322                 ret = 0;
5323                 goto out;
5324         }
5325         if (offset + len > sctx->cur_inode_size)
5326                 len = sctx->cur_inode_size - offset;
5327         if (len == 0) {
5328                 ret = 0;
5329                 goto out;
5330         }
5331 
5332         if (clone_root && IS_ALIGNED(offset + len, bs)) {
5333                 u64 disk_byte;
5334                 u64 data_offset;
5335 
5336                 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5337                 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5338                 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5339                                   offset, len);
5340         } else {
5341                 ret = send_extent_data(sctx, offset, len);
5342         }
5343         sctx->cur_inode_next_write_offset = offset + len;
5344 out:
5345         return ret;
5346 }
5347 
5348 static int is_extent_unchanged(struct send_ctx *sctx,
5349                                struct btrfs_path *left_path,
5350                                struct btrfs_key *ekey)
5351 {
5352         int ret = 0;
5353         struct btrfs_key key;
5354         struct btrfs_path *path = NULL;
5355         struct extent_buffer *eb;
5356         int slot;
5357         struct btrfs_key found_key;
5358         struct btrfs_file_extent_item *ei;
5359         u64 left_disknr;
5360         u64 right_disknr;
5361         u64 left_offset;
5362         u64 right_offset;
5363         u64 left_offset_fixed;
5364         u64 left_len;
5365         u64 right_len;
5366         u64 left_gen;
5367         u64 right_gen;
5368         u8 left_type;
5369         u8 right_type;
5370 
5371         path = alloc_path_for_send();
5372         if (!path)
5373                 return -ENOMEM;
5374 
5375         eb = left_path->nodes[0];
5376         slot = left_path->slots[0];
5377         ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5378         left_type = btrfs_file_extent_type(eb, ei);
5379 
5380         if (left_type != BTRFS_FILE_EXTENT_REG) {
5381                 ret = 0;
5382                 goto out;
5383         }
5384         left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5385         left_len = btrfs_file_extent_num_bytes(eb, ei);
5386         left_offset = btrfs_file_extent_offset(eb, ei);
5387         left_gen = btrfs_file_extent_generation(eb, ei);
5388 
5389         /*
5390          * Following comments will refer to these graphics. L is the left
5391          * extents which we are checking at the moment. 1-8 are the right
5392          * extents that we iterate.
5393          *
5394          *       |-----L-----|
5395          * |-1-|-2a-|-3-|-4-|-5-|-6-|
5396          *
5397          *       |-----L-----|
5398          * |--1--|-2b-|...(same as above)
5399          *
5400          * Alternative situation. Happens on files where extents got split.
5401          *       |-----L-----|
5402          * |-----------7-----------|-6-|
5403          *
5404          * Alternative situation. Happens on files which got larger.
5405          *       |-----L-----|
5406          * |-8-|
5407          * Nothing follows after 8.
5408          */
5409 
5410         key.objectid = ekey->objectid;
5411         key.type = BTRFS_EXTENT_DATA_KEY;
5412         key.offset = ekey->offset;
5413         ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5414         if (ret < 0)
5415                 goto out;
5416         if (ret) {
5417                 ret = 0;
5418                 goto out;
5419         }
5420 
5421         /*
5422          * Handle special case where the right side has no extents at all.
5423          */
5424         eb = path->nodes[0];
5425         slot = path->slots[0];
5426         btrfs_item_key_to_cpu(eb, &found_key, slot);
5427         if (found_key.objectid != key.objectid ||
5428             found_key.type != key.type) {
5429                 /* If we're a hole then just pretend nothing changed */
5430                 ret = (left_disknr) ? 0 : 1;
5431                 goto out;
5432         }
5433 
5434         /*
5435          * We're now on 2a, 2b or 7.
5436          */
5437         key = found_key;
5438         while (key.offset < ekey->offset + left_len) {
5439                 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5440                 right_type = btrfs_file_extent_type(eb, ei);
5441                 if (right_type != BTRFS_FILE_EXTENT_REG &&
5442                     right_type != BTRFS_FILE_EXTENT_INLINE) {
5443                         ret = 0;
5444                         goto out;
5445                 }
5446 
5447                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5448                         right_len = btrfs_file_extent_ram_bytes(eb, ei);
5449                         right_len = PAGE_ALIGN(right_len);
5450                 } else {
5451                         right_len = btrfs_file_extent_num_bytes(eb, ei);
5452                 }
5453 
5454                 /*
5455                  * Are we at extent 8? If yes, we know the extent is changed.
5456                  * This may only happen on the first iteration.
5457                  */
5458                 if (found_key.offset + right_len <= ekey->offset) {
5459                         /* If we're a hole just pretend nothing changed */
5460                         ret = (left_disknr) ? 0 : 1;
5461                         goto out;
5462                 }
5463 
5464                 /*
5465                  * We just wanted to see if when we have an inline extent, what
5466                  * follows it is a regular extent (wanted to check the above
5467                  * condition for inline extents too). This should normally not
5468                  * happen but it's possible for example when we have an inline
5469                  * compressed extent representing data with a size matching
5470                  * the page size (currently the same as sector size).
5471                  */
5472                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5473                         ret = 0;
5474                         goto out;
5475                 }
5476 
5477                 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5478                 right_offset = btrfs_file_extent_offset(eb, ei);
5479                 right_gen = btrfs_file_extent_generation(eb, ei);
5480 
5481                 left_offset_fixed = left_offset;
5482                 if (key.offset < ekey->offset) {
5483                         /* Fix the right offset for 2a and 7. */
5484                         right_offset += ekey->offset - key.offset;
5485                 } else {
5486                         /* Fix the left offset for all behind 2a and 2b */
5487                         left_offset_fixed += key.offset - ekey->offset;
5488                 }
5489 
5490                 /*
5491                  * Check if we have the same extent.
5492                  */
5493                 if (left_disknr != right_disknr ||
5494                     left_offset_fixed != right_offset ||
5495                     left_gen != right_gen) {
5496                         ret = 0;
5497                         goto out;
5498                 }
5499 
5500                 /*
5501                  * Go to the next extent.
5502                  */
5503                 ret = btrfs_next_item(sctx->parent_root, path);
5504                 if (ret < 0)
5505                         goto out;
5506                 if (!ret) {
5507                         eb = path->nodes[0];
5508                         slot = path->slots[0];
5509                         btrfs_item_key_to_cpu(eb, &found_key, slot);
5510                 }
5511                 if (ret || found_key.objectid != key.objectid ||
5512                     found_key.type != key.type) {
5513                         key.offset += right_len;
5514                         break;
5515                 }
5516                 if (found_key.offset != key.offset + right_len) {
5517                         ret = 0;
5518                         goto out;
5519                 }
5520                 key = found_key;
5521         }
5522 
5523         /*
5524          * We're now behind the left extent (treat as unchanged) or at the end
5525          * of the right side (treat as changed).
5526          */
5527         if (key.offset >= ekey->offset + left_len)
5528                 ret = 1;
5529         else
5530                 ret = 0;
5531 
5532 
5533 out:
5534         btrfs_free_path(path);
5535         return ret;
5536 }
5537 
5538 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5539 {
5540         struct btrfs_path *path;
5541         struct btrfs_root *root = sctx->send_root;
5542         struct btrfs_file_extent_item *fi;
5543         struct btrfs_key key;
5544         u64 extent_end;
5545         u8 type;
5546         int ret;
5547 
5548         path = alloc_path_for_send();
5549         if (!path)
5550                 return -ENOMEM;
5551 
5552         sctx->cur_inode_last_extent = 0;
5553 
5554         key.objectid = sctx->cur_ino;
5555         key.type = BTRFS_EXTENT_DATA_KEY;
5556         key.offset = offset;
5557         ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5558         if (ret < 0)
5559                 goto out;
5560         ret = 0;
5561         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5562         if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5563                 goto out;
5564 
5565         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5566                             struct btrfs_file_extent_item);
5567         type = btrfs_file_extent_type(path->nodes[0], fi);
5568         if (type == BTRFS_FILE_EXTENT_INLINE) {
5569                 u64 size = btrfs_file_extent_ram_bytes(path->nodes[0], fi);
5570                 extent_end = ALIGN(key.offset + size,
5571                                    sctx->send_root->fs_info->sectorsize);
5572         } else {
5573                 extent_end = key.offset +
5574                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5575         }
5576         sctx->cur_inode_last_extent = extent_end;
5577 out:
5578         btrfs_free_path(path);
5579         return ret;
5580 }
5581 
5582 static int range_is_hole_in_parent(struct send_ctx *sctx,
5583                                    const u64 start,
5584                                    const u64 end)
5585 {
5586         struct btrfs_path *path;
5587         struct btrfs_key key;
5588         struct btrfs_root *root = sctx->parent_root;
5589         u64 search_start = start;
5590         int ret;
5591 
5592         path = alloc_path_for_send();
5593         if (!path)
5594                 return -ENOMEM;
5595 
5596         key.objectid = sctx->cur_ino;
5597         key.type = BTRFS_EXTENT_DATA_KEY;
5598         key.offset = search_start;
5599         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5600         if (ret < 0)
5601                 goto out;
5602         if (ret > 0 && path->slots[0] > 0)
5603                 path->slots[0]--;
5604 
5605         while (search_start < end) {
5606                 struct extent_buffer *leaf = path->nodes[0];
5607                 int slot = path->slots[0];
5608                 struct btrfs_file_extent_item *fi;
5609                 u64 extent_end;
5610 
5611                 if (slot >= btrfs_header_nritems(leaf)) {
5612                         ret = btrfs_next_leaf(root, path);
5613                         if (ret < 0)
5614                                 goto out;
5615                         else if (ret > 0)
5616                                 break;
5617                         continue;
5618                 }
5619 
5620                 btrfs_item_key_to_cpu(leaf, &key, slot);
5621                 if (key.objectid < sctx->cur_ino ||
5622                     key.type < BTRFS_EXTENT_DATA_KEY)
5623                         goto next;
5624                 if (key.objectid > sctx->cur_ino ||
5625                     key.type > BTRFS_EXTENT_DATA_KEY ||
5626                     key.offset >= end)
5627                         break;
5628 
5629                 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5630                 if (btrfs_file_extent_type(leaf, fi) ==
5631                     BTRFS_FILE_EXTENT_INLINE) {
5632                         u64 size = btrfs_file_extent_ram_bytes(leaf, fi);
5633 
5634                         extent_end = ALIGN(key.offset + size,
5635                                            root->fs_info->sectorsize);
5636                 } else {
5637                         extent_end = key.offset +
5638                                 btrfs_file_extent_num_bytes(leaf, fi);
5639                 }
5640                 if (extent_end <= start)
5641                         goto next;
5642                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5643                         search_start = extent_end;
5644                         goto next;
5645                 }
5646                 ret = 0;
5647                 goto out;
5648 next:
5649                 path->slots[0]++;
5650         }
5651         ret = 1;
5652 out:
5653         btrfs_free_path(path);
5654         return ret;
5655 }
5656 
5657 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5658                            struct btrfs_key *key)
5659 {
5660         struct btrfs_file_extent_item *fi;
5661         u64 extent_end;
5662         u8 type;
5663         int ret = 0;
5664 
5665         if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5666                 return 0;
5667 
5668         if (sctx->cur_inode_last_extent == (u64)-1) {
5669                 ret = get_last_extent(sctx, key->offset - 1);
5670                 if (ret)
5671                         return ret;
5672         }
5673 
5674         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
5675                             struct btrfs_file_extent_item);
5676         type = btrfs_file_extent_type(path->nodes[0], fi);
5677         if (type == BTRFS_FILE_EXTENT_INLINE) {
5678                 u64 size = btrfs_file_extent_ram_bytes(path->nodes[0], fi);
5679                 extent_end = ALIGN(key->offset + size,
5680                                    sctx->send_root->fs_info->sectorsize);
5681         } else {
5682                 extent_end = key->offset +
5683                         btrfs_file_extent_num_bytes(path->nodes[0], fi);
5684         }
5685 
5686         if (path->slots[0] == 0 &&
5687             sctx->cur_inode_last_extent < key->offset) {
5688                 /*
5689                  * We might have skipped entire leafs that contained only
5690                  * file extent items for our current inode. These leafs have
5691                  * a generation number smaller (older) than the one in the
5692                  * current leaf and the leaf our last extent came from, and
5693                  * are located between these 2 leafs.
5694                  */
5695                 ret = get_last_extent(sctx, key->offset - 1);
5696                 if (ret)
5697                         return ret;
5698         }
5699 
5700         if (sctx->cur_inode_last_extent < key->offset) {
5701                 ret = range_is_hole_in_parent(sctx,
5702                                               sctx->cur_inode_last_extent,
5703                                               key->offset);
5704                 if (ret < 0)
5705                         return ret;
5706                 else if (ret == 0)
5707                         ret = send_hole(sctx, key->offset);
5708                 else
5709                         ret = 0;
5710         }
5711         sctx->cur_inode_last_extent = extent_end;
5712         return ret;
5713 }
5714 
5715 static int process_extent(struct send_ctx *sctx,
5716                           struct btrfs_path *path,
5717                           struct btrfs_key *key)
5718 {
5719         struct clone_root *found_clone = NULL;
5720         int ret = 0;
5721 
5722         if (S_ISLNK(sctx->cur_inode_mode))
5723                 return 0;
5724 
5725         if (sctx->parent_root && !sctx->cur_inode_new) {
5726                 ret = is_extent_unchanged(sctx, path, key);
5727                 if (ret < 0)
5728                         goto out;
5729                 if (ret) {
5730                         ret = 0;
5731                         goto out_hole;
5732                 }
5733         } else {
5734                 struct btrfs_file_extent_item *ei;
5735                 u8 type;
5736 
5737                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5738                                     struct btrfs_file_extent_item);
5739                 type = btrfs_file_extent_type(path->nodes[0], ei);
5740                 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5741                     type == BTRFS_FILE_EXTENT_REG) {
5742                         /*
5743                          * The send spec does not have a prealloc command yet,
5744                          * so just leave a hole for prealloc'ed extents until
5745                          * we have enough commands queued up to justify rev'ing
5746                          * the send spec.
5747                          */
5748                         if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5749                                 ret = 0;
5750                                 goto out;
5751                         }
5752 
5753                         /* Have a hole, just skip it. */
5754                         if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5755                                 ret = 0;
5756                                 goto out;
5757                         }
5758                 }
5759         }
5760 
5761         ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5762                         sctx->cur_inode_size, &found_clone);
5763         if (ret != -ENOENT && ret < 0)
5764                 goto out;
5765 
5766         ret = send_write_or_clone(sctx, path, key, found_clone);
5767         if (ret)
5768                 goto out;
5769 out_hole:
5770         ret = maybe_send_hole(sctx, path, key);
5771 out:
5772         return ret;
5773 }
5774 
5775 static int process_all_extents(struct send_ctx *sctx)
5776 {
5777         int ret;
5778         struct btrfs_root *root;
5779         struct btrfs_path *path;
5780         struct btrfs_key key;
5781         struct btrfs_key found_key;
5782         struct extent_buffer *eb;
5783         int slot;
5784 
5785         root = sctx->send_root;
5786         path = alloc_path_for_send();
5787         if (!path)
5788                 return -ENOMEM;
5789 
5790         key.objectid = sctx->cmp_key->objectid;
5791         key.type = BTRFS_EXTENT_DATA_KEY;
5792         key.offset = 0;
5793         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5794         if (ret < 0)
5795                 goto out;
5796 
5797         while (1) {
5798                 eb = path->nodes[0];
5799                 slot = path->slots[0];
5800 
5801                 if (slot >= btrfs_header_nritems(eb)) {
5802                         ret = btrfs_next_leaf(root, path);
5803                         if (ret < 0) {
5804                                 goto out;
5805                         } else if (ret > 0) {
5806                                 ret = 0;
5807                                 break;
5808                         }
5809                         continue;
5810                 }
5811 
5812                 btrfs_item_key_to_cpu(eb, &found_key, slot);
5813 
5814                 if (found_key.objectid != key.objectid ||
5815                     found_key.type != key.type) {
5816                         ret = 0;
5817                         goto out;
5818                 }
5819 
5820                 ret = process_extent(sctx, path, &found_key);
5821                 if (ret < 0)
5822                         goto out;
5823 
5824                 path->slots[0]++;
5825         }
5826 
5827 out:
5828         btrfs_free_path(path);
5829         return ret;
5830 }
5831 
5832 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
5833                                            int *pending_move,
5834                                            int *refs_processed)
5835 {
5836         int ret = 0;
5837 
5838         if (sctx->cur_ino == 0)
5839                 goto out;
5840         if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
5841             sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
5842                 goto out;
5843         if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
5844                 goto out;
5845 
5846         ret = process_recorded_refs(sctx, pending_move);
5847         if (ret < 0)
5848                 goto out;
5849 
5850         *refs_processed = 1;
5851 out:
5852         return ret;
5853 }
5854 
5855 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
5856 {
5857         int ret = 0;
5858         u64 left_mode;
5859         u64 left_uid;
5860         u64 left_gid;
5861         u64 right_mode;
5862         u64 right_uid;
5863         u64 right_gid;
5864         int need_chmod = 0;
5865         int need_chown = 0;
5866         int need_truncate = 1;
5867         int pending_move = 0;
5868         int refs_processed = 0;
5869 
5870         if (sctx->ignore_cur_inode)
5871                 return 0;
5872 
5873         ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
5874                                               &refs_processed);
5875         if (ret < 0)
5876                 goto out;
5877 
5878         /*
5879          * We have processed the refs and thus need to advance send_progress.
5880          * Now, calls to get_cur_xxx will take the updated refs of the current
5881          * inode into account.
5882          *
5883          * On the other hand, if our current inode is a directory and couldn't
5884          * be moved/renamed because its parent was renamed/moved too and it has
5885          * a higher inode number, we can only move/rename our current inode
5886          * after we moved/renamed its parent. Therefore in this case operate on
5887          * the old path (pre move/rename) of our current inode, and the
5888          * move/rename will be performed later.
5889          */
5890         if (refs_processed && !pending_move)
5891                 sctx->send_progress = sctx->cur_ino + 1;
5892 
5893         if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
5894                 goto out;
5895         if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
5896                 goto out;
5897 
5898         ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
5899                         &left_mode, &left_uid, &left_gid, NULL);
5900         if (ret < 0)
5901                 goto out;
5902 
5903         if (!sctx->parent_root || sctx->cur_inode_new) {
5904                 need_chown = 1;
5905                 if (!S_ISLNK(sctx->cur_inode_mode))
5906                         need_chmod = 1;
5907                 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
5908                         need_truncate = 0;
5909         } else {
5910                 u64 old_size;
5911 
5912                 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
5913                                 &old_size, NULL, &right_mode, &right_uid,
5914                                 &right_gid, NULL);
5915                 if (ret < 0)
5916                         goto out;
5917 
5918                 if (left_uid != right_uid || left_gid != right_gid)
5919                         need_chown = 1;
5920                 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
5921                         need_chmod = 1;
5922                 if ((old_size == sctx->cur_inode_size) ||
5923                     (sctx->cur_inode_size > old_size &&
5924                      sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
5925                         need_truncate = 0;
5926         }
5927 
5928         if (S_ISREG(sctx->cur_inode_mode)) {
5929                 if (need_send_hole(sctx)) {
5930                         if (sctx->cur_inode_last_extent == (u64)-1 ||
5931                             sctx->cur_inode_last_extent <
5932                             sctx->cur_inode_size) {
5933                                 ret = get_last_extent(sctx, (u64)-1);
5934                                 if (ret)
5935                                         goto out;
5936                         }
5937                         if (sctx->cur_inode_last_extent <
5938                             sctx->cur_inode_size) {
5939                                 ret = send_hole(sctx, sctx->cur_inode_size);
5940                                 if (ret)
5941                                         goto out;
5942                         }
5943                 }
5944                 if (need_truncate) {
5945                         ret = send_truncate(sctx, sctx->cur_ino,
5946                                             sctx->cur_inode_gen,
5947                                             sctx->cur_inode_size);
5948                         if (ret < 0)
5949                                 goto out;
5950                 }
5951         }
5952 
5953         if (need_chown) {
5954                 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5955                                 left_uid, left_gid);
5956                 if (ret < 0)
5957                         goto out;
5958         }
5959         if (need_chmod) {
5960                 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen,
5961                                 left_mode);
5962                 if (ret < 0)
5963                         goto out;
5964         }
5965 
5966         /*
5967          * If other directory inodes depended on our current directory
5968          * inode's move/rename, now do their move/rename operations.
5969          */
5970         if (!is_waiting_for_move(sctx, sctx->cur_ino)) {
5971                 ret = apply_children_dir_moves(sctx);
5972                 if (ret)
5973                         goto out;
5974                 /*
5975                  * Need to send that every time, no matter if it actually
5976                  * changed between the two trees as we have done changes to
5977                  * the inode before. If our inode is a directory and it's
5978                  * waiting to be moved/renamed, we will send its utimes when
5979                  * it's moved/renamed, therefore we don't need to do it here.
5980                  */
5981                 sctx->send_progress = sctx->cur_ino + 1;
5982                 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen);
5983                 if (ret < 0)
5984                         goto out;
5985         }
5986 
5987 out:
5988         return ret;
5989 }
5990 
5991 struct parent_paths_ctx {
5992         struct list_head *refs;
5993         struct send_ctx *sctx;
5994 };
5995 
5996 static int record_parent_ref(int num, u64 dir, int index, struct fs_path *name,
5997                              void *ctx)
5998 {
5999         struct parent_paths_ctx *ppctx = ctx;
6000 
6001         return record_ref(ppctx->sctx->parent_root, dir, name, ppctx->sctx,
6002                           ppctx->refs);
6003 }
6004 
6005 /*
6006  * Issue unlink operations for all paths of the current inode found in the
6007  * parent snapshot.
6008  */
6009 static int btrfs_unlink_all_paths(struct send_ctx *sctx)
6010 {
6011         LIST_HEAD(deleted_refs);
6012         struct btrfs_path *path;