~ [ 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 #include "xattr.h"
 27 
 28 /*
 29  * Maximum number of references an extent can have in order for us to attempt to
 30  * issue clone operations instead of write operations. This currently exists to
 31  * avoid hitting limitations of the backreference walking code (taking a lot of
 32  * time and using too much memory for extents with large number of references).
 33  */
 34 #define SEND_MAX_EXTENT_REFS    64
 35 
 36 /*
 37  * A fs_path is a helper to dynamically build path names with unknown size.
 38  * It reallocates the internal buffer on demand.
 39  * It allows fast adding of path elements on the right side (normal path) and
 40  * fast adding to the left side (reversed path). A reversed path can also be
 41  * unreversed if needed.
 42  */
 43 struct fs_path {
 44         union {
 45                 struct {
 46                         char *start;
 47                         char *end;
 48 
 49                         char *buf;
 50                         unsigned short buf_len:15;
 51                         unsigned short reversed:1;
 52                         char inline_buf[];
 53                 };
 54                 /*
 55                  * Average path length does not exceed 200 bytes, we'll have
 56                  * better packing in the slab and higher chance to satisfy
 57                  * a allocation later during send.
 58                  */
 59                 char pad[256];
 60         };
 61 };
 62 #define FS_PATH_INLINE_SIZE \
 63         (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf))
 64 
 65 
 66 /* reused for each extent */
 67 struct clone_root {
 68         struct btrfs_root *root;
 69         u64 ino;
 70         u64 offset;
 71 
 72         u64 found_refs;
 73 };
 74 
 75 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128
 76 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2)
 77 
 78 struct send_ctx {
 79         struct file *send_filp;
 80         loff_t send_off;
 81         char *send_buf;
 82         u32 send_size;
 83         u32 send_max_size;
 84         u64 total_send_size;
 85         u64 cmd_send_size[BTRFS_SEND_C_MAX + 1];
 86         u64 flags;      /* 'flags' member of btrfs_ioctl_send_args is u64 */
 87 
 88         struct btrfs_root *send_root;
 89         struct btrfs_root *parent_root;
 90         struct clone_root *clone_roots;
 91         int clone_roots_cnt;
 92 
 93         /* current state of the compare_tree call */
 94         struct btrfs_path *left_path;
 95         struct btrfs_path *right_path;
 96         struct btrfs_key *cmp_key;
 97 
 98         /*
 99          * infos of the currently processed inode. In case of deleted inodes,
100          * these are the values from the deleted inode.
101          */
102         u64 cur_ino;
103         u64 cur_inode_gen;
104         int cur_inode_new;
105         int cur_inode_new_gen;
106         int cur_inode_deleted;
107         u64 cur_inode_size;
108         u64 cur_inode_mode;
109         u64 cur_inode_rdev;
110         u64 cur_inode_last_extent;
111         u64 cur_inode_next_write_offset;
112         bool ignore_cur_inode;
113 
114         u64 send_progress;
115 
116         struct list_head new_refs;
117         struct list_head deleted_refs;
118 
119         struct radix_tree_root name_cache;
120         struct list_head name_cache_list;
121         int name_cache_size;
122 
123         struct file_ra_state ra;
124 
125         char *read_buf;
126 
127         /*
128          * We process inodes by their increasing order, so if before an
129          * incremental send we reverse the parent/child relationship of
130          * directories such that a directory with a lower inode number was
131          * the parent of a directory with a higher inode number, and the one
132          * becoming the new parent got renamed too, we can't rename/move the
133          * directory with lower inode number when we finish processing it - we
134          * must process the directory with higher inode number first, then
135          * rename/move it and then rename/move the directory with lower inode
136          * number. Example follows.
137          *
138          * Tree state when the first send was performed:
139          *
140          * .
141          * |-- a                   (ino 257)
142          *     |-- b               (ino 258)
143          *         |
144          *         |
145          *         |-- c           (ino 259)
146          *         |   |-- d       (ino 260)
147          *         |
148          *         |-- c2          (ino 261)
149          *
150          * Tree state when the second (incremental) send is performed:
151          *
152          * .
153          * |-- a                   (ino 257)
154          *     |-- b               (ino 258)
155          *         |-- c2          (ino 261)
156          *             |-- d2      (ino 260)
157          *                 |-- cc  (ino 259)
158          *
159          * The sequence of steps that lead to the second state was:
160          *
161          * mv /a/b/c/d /a/b/c2/d2
162          * mv /a/b/c /a/b/c2/d2/cc
163          *
164          * "c" has lower inode number, but we can't move it (2nd mv operation)
165          * before we move "d", which has higher inode number.
166          *
167          * So we just memorize which move/rename operations must be performed
168          * later when their respective parent is processed and moved/renamed.
169          */
170 
171         /* Indexed by parent directory inode number. */
172         struct rb_root pending_dir_moves;
173 
174         /*
175          * Reverse index, indexed by the inode number of a directory that
176          * is waiting for the move/rename of its immediate parent before its
177          * own move/rename can be performed.
178          */
179         struct rb_root waiting_dir_moves;
180 
181         /*
182          * A directory that is going to be rm'ed might have a child directory
183          * which is in the pending directory moves index above. In this case,
184          * the directory can only be removed after the move/rename of its child
185          * is performed. Example:
186          *
187          * Parent snapshot:
188          *
189          * .                        (ino 256)
190          * |-- a/                   (ino 257)
191          *     |-- b/               (ino 258)
192          *         |-- c/           (ino 259)
193          *         |   |-- x/       (ino 260)
194          *         |
195          *         |-- y/           (ino 261)
196          *
197          * Send snapshot:
198          *
199          * .                        (ino 256)
200          * |-- a/                   (ino 257)
201          *     |-- b/               (ino 258)
202          *         |-- YY/          (ino 261)
203          *              |-- x/      (ino 260)
204          *
205          * Sequence of steps that lead to the send snapshot:
206          * rm -f /a/b/c/foo.txt
207          * mv /a/b/y /a/b/YY
208          * mv /a/b/c/x /a/b/YY
209          * rmdir /a/b/c
210          *
211          * When the child is processed, its move/rename is delayed until its
212          * parent is processed (as explained above), but all other operations
213          * like update utimes, chown, chgrp, etc, are performed and the paths
214          * that it uses for those operations must use the orphanized name of
215          * its parent (the directory we're going to rm later), so we need to
216          * memorize that name.
217          *
218          * Indexed by the inode number of the directory to be deleted.
219          */
220         struct rb_root orphan_dirs;
221 };
222 
223 struct pending_dir_move {
224         struct rb_node node;
225         struct list_head list;
226         u64 parent_ino;
227         u64 ino;
228         u64 gen;
229         struct list_head update_refs;
230 };
231 
232 struct waiting_dir_move {
233         struct rb_node node;
234         u64 ino;
235         /*
236          * There might be some directory that could not be removed because it
237          * was waiting for this directory inode to be moved first. Therefore
238          * after this directory is moved, we can try to rmdir the ino rmdir_ino.
239          */
240         u64 rmdir_ino;
241         bool orphanized;
242 };
243 
244 struct orphan_dir_info {
245         struct rb_node node;
246         u64 ino;
247         u64 gen;
248         u64 last_dir_index_offset;
249 };
250 
251 struct name_cache_entry {
252         struct list_head list;
253         /*
254          * radix_tree has only 32bit entries but we need to handle 64bit inums.
255          * We use the lower 32bit of the 64bit inum to store it in the tree. If
256          * more then one inum would fall into the same entry, we use radix_list
257          * to store the additional entries. radix_list is also used to store
258          * entries where two entries have the same inum but different
259          * generations.
260          */
261         struct list_head radix_list;
262         u64 ino;
263         u64 gen;
264         u64 parent_ino;
265         u64 parent_gen;
266         int ret;
267         int need_later_update;
268         int name_len;
269         char name[];
270 };
271 
272 #define ADVANCE                                                 1
273 #define ADVANCE_ONLY_NEXT                                       -1
274 
275 enum btrfs_compare_tree_result {
276         BTRFS_COMPARE_TREE_NEW,
277         BTRFS_COMPARE_TREE_DELETED,
278         BTRFS_COMPARE_TREE_CHANGED,
279         BTRFS_COMPARE_TREE_SAME,
280 };
281 typedef int (*btrfs_changed_cb_t)(struct btrfs_path *left_path,
282                                   struct btrfs_path *right_path,
283                                   struct btrfs_key *key,
284                                   enum btrfs_compare_tree_result result,
285                                   void *ctx);
286 
287 __cold
288 static void inconsistent_snapshot_error(struct send_ctx *sctx,
289                                         enum btrfs_compare_tree_result result,
290                                         const char *what)
291 {
292         const char *result_string;
293 
294         switch (result) {
295         case BTRFS_COMPARE_TREE_NEW:
296                 result_string = "new";
297                 break;
298         case BTRFS_COMPARE_TREE_DELETED:
299                 result_string = "deleted";
300                 break;
301         case BTRFS_COMPARE_TREE_CHANGED:
302                 result_string = "updated";
303                 break;
304         case BTRFS_COMPARE_TREE_SAME:
305                 ASSERT(0);
306                 result_string = "unchanged";
307                 break;
308         default:
309                 ASSERT(0);
310                 result_string = "unexpected";
311         }
312 
313         btrfs_err(sctx->send_root->fs_info,
314                   "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu",
315                   result_string, what, sctx->cmp_key->objectid,
316                   sctx->send_root->root_key.objectid,
317                   (sctx->parent_root ?
318                    sctx->parent_root->root_key.objectid : 0));
319 }
320 
321 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino);
322 
323 static struct waiting_dir_move *
324 get_waiting_dir_move(struct send_ctx *sctx, u64 ino);
325 
326 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino);
327 
328 static int need_send_hole(struct send_ctx *sctx)
329 {
330         return (sctx->parent_root && !sctx->cur_inode_new &&
331                 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted &&
332                 S_ISREG(sctx->cur_inode_mode));
333 }
334 
335 static void fs_path_reset(struct fs_path *p)
336 {
337         if (p->reversed) {
338                 p->start = p->buf + p->buf_len - 1;
339                 p->end = p->start;
340                 *p->start = 0;
341         } else {
342                 p->start = p->buf;
343                 p->end = p->start;
344                 *p->start = 0;
345         }
346 }
347 
348 static struct fs_path *fs_path_alloc(void)
349 {
350         struct fs_path *p;
351 
352         p = kmalloc(sizeof(*p), GFP_KERNEL);
353         if (!p)
354                 return NULL;
355         p->reversed = 0;
356         p->buf = p->inline_buf;
357         p->buf_len = FS_PATH_INLINE_SIZE;
358         fs_path_reset(p);
359         return p;
360 }
361 
362 static struct fs_path *fs_path_alloc_reversed(void)
363 {
364         struct fs_path *p;
365 
366         p = fs_path_alloc();
367         if (!p)
368                 return NULL;
369         p->reversed = 1;
370         fs_path_reset(p);
371         return p;
372 }
373 
374 static void fs_path_free(struct fs_path *p)
375 {
376         if (!p)
377                 return;
378         if (p->buf != p->inline_buf)
379                 kfree(p->buf);
380         kfree(p);
381 }
382 
383 static int fs_path_len(struct fs_path *p)
384 {
385         return p->end - p->start;
386 }
387 
388 static int fs_path_ensure_buf(struct fs_path *p, int len)
389 {
390         char *tmp_buf;
391         int path_len;
392         int old_buf_len;
393 
394         len++;
395 
396         if (p->buf_len >= len)
397                 return 0;
398 
399         if (len > PATH_MAX) {
400                 WARN_ON(1);
401                 return -ENOMEM;
402         }
403 
404         path_len = p->end - p->start;
405         old_buf_len = p->buf_len;
406 
407         /*
408          * First time the inline_buf does not suffice
409          */
410         if (p->buf == p->inline_buf) {
411                 tmp_buf = kmalloc(len, GFP_KERNEL);
412                 if (tmp_buf)
413                         memcpy(tmp_buf, p->buf, old_buf_len);
414         } else {
415                 tmp_buf = krealloc(p->buf, len, GFP_KERNEL);
416         }
417         if (!tmp_buf)
418                 return -ENOMEM;
419         p->buf = tmp_buf;
420         /*
421          * The real size of the buffer is bigger, this will let the fast path
422          * happen most of the time
423          */
424         p->buf_len = ksize(p->buf);
425 
426         if (p->reversed) {
427                 tmp_buf = p->buf + old_buf_len - path_len - 1;
428                 p->end = p->buf + p->buf_len - 1;
429                 p->start = p->end - path_len;
430                 memmove(p->start, tmp_buf, path_len + 1);
431         } else {
432                 p->start = p->buf;
433                 p->end = p->start + path_len;
434         }
435         return 0;
436 }
437 
438 static int fs_path_prepare_for_add(struct fs_path *p, int name_len,
439                                    char **prepared)
440 {
441         int ret;
442         int new_len;
443 
444         new_len = p->end - p->start + name_len;
445         if (p->start != p->end)
446                 new_len++;
447         ret = fs_path_ensure_buf(p, new_len);
448         if (ret < 0)
449                 goto out;
450 
451         if (p->reversed) {
452                 if (p->start != p->end)
453                         *--p->start = '/';
454                 p->start -= name_len;
455                 *prepared = p->start;
456         } else {
457                 if (p->start != p->end)
458                         *p->end++ = '/';
459                 *prepared = p->end;
460                 p->end += name_len;
461                 *p->end = 0;
462         }
463 
464 out:
465         return ret;
466 }
467 
468 static int fs_path_add(struct fs_path *p, const char *name, int name_len)
469 {
470         int ret;
471         char *prepared;
472 
473         ret = fs_path_prepare_for_add(p, name_len, &prepared);
474         if (ret < 0)
475                 goto out;
476         memcpy(prepared, name, name_len);
477 
478 out:
479         return ret;
480 }
481 
482 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2)
483 {
484         int ret;
485         char *prepared;
486 
487         ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared);
488         if (ret < 0)
489                 goto out;
490         memcpy(prepared, p2->start, p2->end - p2->start);
491 
492 out:
493         return ret;
494 }
495 
496 static int fs_path_add_from_extent_buffer(struct fs_path *p,
497                                           struct extent_buffer *eb,
498                                           unsigned long off, int len)
499 {
500         int ret;
501         char *prepared;
502 
503         ret = fs_path_prepare_for_add(p, len, &prepared);
504         if (ret < 0)
505                 goto out;
506 
507         read_extent_buffer(eb, prepared, off, len);
508 
509 out:
510         return ret;
511 }
512 
513 static int fs_path_copy(struct fs_path *p, struct fs_path *from)
514 {
515         int ret;
516 
517         p->reversed = from->reversed;
518         fs_path_reset(p);
519 
520         ret = fs_path_add_path(p, from);
521 
522         return ret;
523 }
524 
525 
526 static void fs_path_unreverse(struct fs_path *p)
527 {
528         char *tmp;
529         int len;
530 
531         if (!p->reversed)
532                 return;
533 
534         tmp = p->start;
535         len = p->end - p->start;
536         p->start = p->buf;
537         p->end = p->start + len;
538         memmove(p->start, tmp, len + 1);
539         p->reversed = 0;
540 }
541 
542 static struct btrfs_path *alloc_path_for_send(void)
543 {
544         struct btrfs_path *path;
545 
546         path = btrfs_alloc_path();
547         if (!path)
548                 return NULL;
549         path->search_commit_root = 1;
550         path->skip_locking = 1;
551         path->need_commit_sem = 1;
552         return path;
553 }
554 
555 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off)
556 {
557         int ret;
558         u32 pos = 0;
559 
560         while (pos < len) {
561                 ret = kernel_write(filp, buf + pos, len - pos, off);
562                 /* TODO handle that correctly */
563                 /*if (ret == -ERESTARTSYS) {
564                         continue;
565                 }*/
566                 if (ret < 0)
567                         return ret;
568                 if (ret == 0) {
569                         return -EIO;
570                 }
571                 pos += ret;
572         }
573 
574         return 0;
575 }
576 
577 static int tlv_put(struct send_ctx *sctx, u16 attr, const void *data, int len)
578 {
579         struct btrfs_tlv_header *hdr;
580         int total_len = sizeof(*hdr) + len;
581         int left = sctx->send_max_size - sctx->send_size;
582 
583         if (unlikely(left < total_len))
584                 return -EOVERFLOW;
585 
586         hdr = (struct btrfs_tlv_header *) (sctx->send_buf + sctx->send_size);
587         hdr->tlv_type = cpu_to_le16(attr);
588         hdr->tlv_len = cpu_to_le16(len);
589         memcpy(hdr + 1, data, len);
590         sctx->send_size += total_len;
591 
592         return 0;
593 }
594 
595 #define TLV_PUT_DEFINE_INT(bits) \
596         static int tlv_put_u##bits(struct send_ctx *sctx,               \
597                         u##bits attr, u##bits value)                    \
598         {                                                               \
599                 __le##bits __tmp = cpu_to_le##bits(value);              \
600                 return tlv_put(sctx, attr, &__tmp, sizeof(__tmp));      \
601         }
602 
603 TLV_PUT_DEFINE_INT(64)
604 
605 static int tlv_put_string(struct send_ctx *sctx, u16 attr,
606                           const char *str, int len)
607 {
608         if (len == -1)
609                 len = strlen(str);
610         return tlv_put(sctx, attr, str, len);
611 }
612 
613 static int tlv_put_uuid(struct send_ctx *sctx, u16 attr,
614                         const u8 *uuid)
615 {
616         return tlv_put(sctx, attr, uuid, BTRFS_UUID_SIZE);
617 }
618 
619 static int tlv_put_btrfs_timespec(struct send_ctx *sctx, u16 attr,
620                                   struct extent_buffer *eb,
621                                   struct btrfs_timespec *ts)
622 {
623         struct btrfs_timespec bts;
624         read_extent_buffer(eb, &bts, (unsigned long)ts, sizeof(bts));
625         return tlv_put(sctx, attr, &bts, sizeof(bts));
626 }
627 
628 
629 #define TLV_PUT(sctx, attrtype, data, attrlen) \
630         do { \
631                 ret = tlv_put(sctx, attrtype, data, attrlen); \
632                 if (ret < 0) \
633                         goto tlv_put_failure; \
634         } while (0)
635 
636 #define TLV_PUT_INT(sctx, attrtype, bits, value) \
637         do { \
638                 ret = tlv_put_u##bits(sctx, attrtype, value); \
639                 if (ret < 0) \
640                         goto tlv_put_failure; \
641         } while (0)
642 
643 #define TLV_PUT_U8(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 8, data)
644 #define TLV_PUT_U16(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 16, data)
645 #define TLV_PUT_U32(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 32, data)
646 #define TLV_PUT_U64(sctx, attrtype, data) TLV_PUT_INT(sctx, attrtype, 64, data)
647 #define TLV_PUT_STRING(sctx, attrtype, str, len) \
648         do { \
649                 ret = tlv_put_string(sctx, attrtype, str, len); \
650                 if (ret < 0) \
651                         goto tlv_put_failure; \
652         } while (0)
653 #define TLV_PUT_PATH(sctx, attrtype, p) \
654         do { \
655                 ret = tlv_put_string(sctx, attrtype, p->start, \
656                         p->end - p->start); \
657                 if (ret < 0) \
658                         goto tlv_put_failure; \
659         } while(0)
660 #define TLV_PUT_UUID(sctx, attrtype, uuid) \
661         do { \
662                 ret = tlv_put_uuid(sctx, attrtype, uuid); \
663                 if (ret < 0) \
664                         goto tlv_put_failure; \
665         } while (0)
666 #define TLV_PUT_BTRFS_TIMESPEC(sctx, attrtype, eb, ts) \
667         do { \
668                 ret = tlv_put_btrfs_timespec(sctx, attrtype, eb, ts); \
669                 if (ret < 0) \
670                         goto tlv_put_failure; \
671         } while (0)
672 
673 static int send_header(struct send_ctx *sctx)
674 {
675         struct btrfs_stream_header hdr;
676 
677         strcpy(hdr.magic, BTRFS_SEND_STREAM_MAGIC);
678         hdr.version = cpu_to_le32(BTRFS_SEND_STREAM_VERSION);
679 
680         return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
681                                         &sctx->send_off);
682 }
683 
684 /*
685  * For each command/item we want to send to userspace, we call this function.
686  */
687 static int begin_cmd(struct send_ctx *sctx, int cmd)
688 {
689         struct btrfs_cmd_header *hdr;
690 
691         if (WARN_ON(!sctx->send_buf))
692                 return -EINVAL;
693 
694         BUG_ON(sctx->send_size);
695 
696         sctx->send_size += sizeof(*hdr);
697         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
698         hdr->cmd = cpu_to_le16(cmd);
699 
700         return 0;
701 }
702 
703 static int send_cmd(struct send_ctx *sctx)
704 {
705         int ret;
706         struct btrfs_cmd_header *hdr;
707         u32 crc;
708 
709         hdr = (struct btrfs_cmd_header *)sctx->send_buf;
710         hdr->len = cpu_to_le32(sctx->send_size - sizeof(*hdr));
711         hdr->crc = 0;
712 
713         crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size);
714         hdr->crc = cpu_to_le32(crc);
715 
716         ret = write_buf(sctx->send_filp, sctx->send_buf, sctx->send_size,
717                                         &sctx->send_off);
718 
719         sctx->total_send_size += sctx->send_size;
720         sctx->cmd_send_size[le16_to_cpu(hdr->cmd)] += sctx->send_size;
721         sctx->send_size = 0;
722 
723         return ret;
724 }
725 
726 /*
727  * Sends a move instruction to user space
728  */
729 static int send_rename(struct send_ctx *sctx,
730                      struct fs_path *from, struct fs_path *to)
731 {
732         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
733         int ret;
734 
735         btrfs_debug(fs_info, "send_rename %s -> %s", from->start, to->start);
736 
737         ret = begin_cmd(sctx, BTRFS_SEND_C_RENAME);
738         if (ret < 0)
739                 goto out;
740 
741         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, from);
742         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_TO, to);
743 
744         ret = send_cmd(sctx);
745 
746 tlv_put_failure:
747 out:
748         return ret;
749 }
750 
751 /*
752  * Sends a link instruction to user space
753  */
754 static int send_link(struct send_ctx *sctx,
755                      struct fs_path *path, struct fs_path *lnk)
756 {
757         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
758         int ret;
759 
760         btrfs_debug(fs_info, "send_link %s -> %s", path->start, lnk->start);
761 
762         ret = begin_cmd(sctx, BTRFS_SEND_C_LINK);
763         if (ret < 0)
764                 goto out;
765 
766         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
767         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, lnk);
768 
769         ret = send_cmd(sctx);
770 
771 tlv_put_failure:
772 out:
773         return ret;
774 }
775 
776 /*
777  * Sends an unlink instruction to user space
778  */
779 static int send_unlink(struct send_ctx *sctx, struct fs_path *path)
780 {
781         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
782         int ret;
783 
784         btrfs_debug(fs_info, "send_unlink %s", path->start);
785 
786         ret = begin_cmd(sctx, BTRFS_SEND_C_UNLINK);
787         if (ret < 0)
788                 goto out;
789 
790         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
791 
792         ret = send_cmd(sctx);
793 
794 tlv_put_failure:
795 out:
796         return ret;
797 }
798 
799 /*
800  * Sends a rmdir instruction to user space
801  */
802 static int send_rmdir(struct send_ctx *sctx, struct fs_path *path)
803 {
804         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
805         int ret;
806 
807         btrfs_debug(fs_info, "send_rmdir %s", path->start);
808 
809         ret = begin_cmd(sctx, BTRFS_SEND_C_RMDIR);
810         if (ret < 0)
811                 goto out;
812 
813         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
814 
815         ret = send_cmd(sctx);
816 
817 tlv_put_failure:
818 out:
819         return ret;
820 }
821 
822 /*
823  * Helper function to retrieve some fields from an inode item.
824  */
825 static int __get_inode_info(struct btrfs_root *root, struct btrfs_path *path,
826                           u64 ino, u64 *size, u64 *gen, u64 *mode, u64 *uid,
827                           u64 *gid, u64 *rdev)
828 {
829         int ret;
830         struct btrfs_inode_item *ii;
831         struct btrfs_key key;
832 
833         key.objectid = ino;
834         key.type = BTRFS_INODE_ITEM_KEY;
835         key.offset = 0;
836         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
837         if (ret) {
838                 if (ret > 0)
839                         ret = -ENOENT;
840                 return ret;
841         }
842 
843         ii = btrfs_item_ptr(path->nodes[0], path->slots[0],
844                         struct btrfs_inode_item);
845         if (size)
846                 *size = btrfs_inode_size(path->nodes[0], ii);
847         if (gen)
848                 *gen = btrfs_inode_generation(path->nodes[0], ii);
849         if (mode)
850                 *mode = btrfs_inode_mode(path->nodes[0], ii);
851         if (uid)
852                 *uid = btrfs_inode_uid(path->nodes[0], ii);
853         if (gid)
854                 *gid = btrfs_inode_gid(path->nodes[0], ii);
855         if (rdev)
856                 *rdev = btrfs_inode_rdev(path->nodes[0], ii);
857 
858         return ret;
859 }
860 
861 static int get_inode_info(struct btrfs_root *root,
862                           u64 ino, u64 *size, u64 *gen,
863                           u64 *mode, u64 *uid, u64 *gid,
864                           u64 *rdev)
865 {
866         struct btrfs_path *path;
867         int ret;
868 
869         path = alloc_path_for_send();
870         if (!path)
871                 return -ENOMEM;
872         ret = __get_inode_info(root, path, ino, size, gen, mode, uid, gid,
873                                rdev);
874         btrfs_free_path(path);
875         return ret;
876 }
877 
878 typedef int (*iterate_inode_ref_t)(int num, u64 dir, int index,
879                                    struct fs_path *p,
880                                    void *ctx);
881 
882 /*
883  * Helper function to iterate the entries in ONE btrfs_inode_ref or
884  * btrfs_inode_extref.
885  * The iterate callback may return a non zero value to stop iteration. This can
886  * be a negative value for error codes or 1 to simply stop it.
887  *
888  * path must point to the INODE_REF or INODE_EXTREF when called.
889  */
890 static int iterate_inode_ref(struct btrfs_root *root, struct btrfs_path *path,
891                              struct btrfs_key *found_key, int resolve,
892                              iterate_inode_ref_t iterate, void *ctx)
893 {
894         struct extent_buffer *eb = path->nodes[0];
895         struct btrfs_item *item;
896         struct btrfs_inode_ref *iref;
897         struct btrfs_inode_extref *extref;
898         struct btrfs_path *tmp_path;
899         struct fs_path *p;
900         u32 cur = 0;
901         u32 total;
902         int slot = path->slots[0];
903         u32 name_len;
904         char *start;
905         int ret = 0;
906         int num = 0;
907         int index;
908         u64 dir;
909         unsigned long name_off;
910         unsigned long elem_size;
911         unsigned long ptr;
912 
913         p = fs_path_alloc_reversed();
914         if (!p)
915                 return -ENOMEM;
916 
917         tmp_path = alloc_path_for_send();
918         if (!tmp_path) {
919                 fs_path_free(p);
920                 return -ENOMEM;
921         }
922 
923 
924         if (found_key->type == BTRFS_INODE_REF_KEY) {
925                 ptr = (unsigned long)btrfs_item_ptr(eb, slot,
926                                                     struct btrfs_inode_ref);
927                 item = btrfs_item_nr(slot);
928                 total = btrfs_item_size(eb, item);
929                 elem_size = sizeof(*iref);
930         } else {
931                 ptr = btrfs_item_ptr_offset(eb, slot);
932                 total = btrfs_item_size_nr(eb, slot);
933                 elem_size = sizeof(*extref);
934         }
935 
936         while (cur < total) {
937                 fs_path_reset(p);
938 
939                 if (found_key->type == BTRFS_INODE_REF_KEY) {
940                         iref = (struct btrfs_inode_ref *)(ptr + cur);
941                         name_len = btrfs_inode_ref_name_len(eb, iref);
942                         name_off = (unsigned long)(iref + 1);
943                         index = btrfs_inode_ref_index(eb, iref);
944                         dir = found_key->offset;
945                 } else {
946                         extref = (struct btrfs_inode_extref *)(ptr + cur);
947                         name_len = btrfs_inode_extref_name_len(eb, extref);
948                         name_off = (unsigned long)&extref->name;
949                         index = btrfs_inode_extref_index(eb, extref);
950                         dir = btrfs_inode_extref_parent(eb, extref);
951                 }
952 
953                 if (resolve) {
954                         start = btrfs_ref_to_path(root, tmp_path, name_len,
955                                                   name_off, eb, dir,
956                                                   p->buf, p->buf_len);
957                         if (IS_ERR(start)) {
958                                 ret = PTR_ERR(start);
959                                 goto out;
960                         }
961                         if (start < p->buf) {
962                                 /* overflow , try again with larger buffer */
963                                 ret = fs_path_ensure_buf(p,
964                                                 p->buf_len + p->buf - start);
965                                 if (ret < 0)
966                                         goto out;
967                                 start = btrfs_ref_to_path(root, tmp_path,
968                                                           name_len, name_off,
969                                                           eb, dir,
970                                                           p->buf, p->buf_len);
971                                 if (IS_ERR(start)) {
972                                         ret = PTR_ERR(start);
973                                         goto out;
974                                 }
975                                 BUG_ON(start < p->buf);
976                         }
977                         p->start = start;
978                 } else {
979                         ret = fs_path_add_from_extent_buffer(p, eb, name_off,
980                                                              name_len);
981                         if (ret < 0)
982                                 goto out;
983                 }
984 
985                 cur += elem_size + name_len;
986                 ret = iterate(num, dir, index, p, ctx);
987                 if (ret)
988                         goto out;
989                 num++;
990         }
991 
992 out:
993         btrfs_free_path(tmp_path);
994         fs_path_free(p);
995         return ret;
996 }
997 
998 typedef int (*iterate_dir_item_t)(int num, struct btrfs_key *di_key,
999                                   const char *name, int name_len,
1000                                   const char *data, int data_len,
1001                                   u8 type, void *ctx);
1002 
1003 /*
1004  * Helper function to iterate the entries in ONE btrfs_dir_item.
1005  * The iterate callback may return a non zero value to stop iteration. This can
1006  * be a negative value for error codes or 1 to simply stop it.
1007  *
1008  * path must point to the dir item when called.
1009  */
1010 static int iterate_dir_item(struct btrfs_root *root, struct btrfs_path *path,
1011                             iterate_dir_item_t iterate, void *ctx)
1012 {
1013         int ret = 0;
1014         struct extent_buffer *eb;
1015         struct btrfs_item *item;
1016         struct btrfs_dir_item *di;
1017         struct btrfs_key di_key;
1018         char *buf = NULL;
1019         int buf_len;
1020         u32 name_len;
1021         u32 data_len;
1022         u32 cur;
1023         u32 len;
1024         u32 total;
1025         int slot;
1026         int num;
1027         u8 type;
1028 
1029         /*
1030          * Start with a small buffer (1 page). If later we end up needing more
1031          * space, which can happen for xattrs on a fs with a leaf size greater
1032          * then the page size, attempt to increase the buffer. Typically xattr
1033          * values are small.
1034          */
1035         buf_len = PATH_MAX;
1036         buf = kmalloc(buf_len, GFP_KERNEL);
1037         if (!buf) {
1038                 ret = -ENOMEM;
1039                 goto out;
1040         }
1041 
1042         eb = path->nodes[0];
1043         slot = path->slots[0];
1044         item = btrfs_item_nr(slot);
1045         di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
1046         cur = 0;
1047         len = 0;
1048         total = btrfs_item_size(eb, item);
1049 
1050         num = 0;
1051         while (cur < total) {
1052                 name_len = btrfs_dir_name_len(eb, di);
1053                 data_len = btrfs_dir_data_len(eb, di);
1054                 type = btrfs_dir_type(eb, di);
1055                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
1056 
1057                 if (type == BTRFS_FT_XATTR) {
1058                         if (name_len > XATTR_NAME_MAX) {
1059                                 ret = -ENAMETOOLONG;
1060                                 goto out;
1061                         }
1062                         if (name_len + data_len >
1063                                         BTRFS_MAX_XATTR_SIZE(root->fs_info)) {
1064                                 ret = -E2BIG;
1065                                 goto out;
1066                         }
1067                 } else {
1068                         /*
1069                          * Path too long
1070                          */
1071                         if (name_len + data_len > PATH_MAX) {
1072                                 ret = -ENAMETOOLONG;
1073                                 goto out;
1074                         }
1075                 }
1076 
1077                 if (name_len + data_len > buf_len) {
1078                         buf_len = name_len + data_len;
1079                         if (is_vmalloc_addr(buf)) {
1080                                 vfree(buf);
1081                                 buf = NULL;
1082                         } else {
1083                                 char *tmp = krealloc(buf, buf_len,
1084                                                 GFP_KERNEL | __GFP_NOWARN);
1085 
1086                                 if (!tmp)
1087                                         kfree(buf);
1088                                 buf = tmp;
1089                         }
1090                         if (!buf) {
1091                                 buf = kvmalloc(buf_len, GFP_KERNEL);
1092                                 if (!buf) {
1093                                         ret = -ENOMEM;
1094                                         goto out;
1095                                 }
1096                         }
1097                 }
1098 
1099                 read_extent_buffer(eb, buf, (unsigned long)(di + 1),
1100                                 name_len + data_len);
1101 
1102                 len = sizeof(*di) + name_len + data_len;
1103                 di = (struct btrfs_dir_item *)((char *)di + len);
1104                 cur += len;
1105 
1106                 ret = iterate(num, &di_key, buf, name_len, buf + name_len,
1107                                 data_len, type, ctx);
1108                 if (ret < 0)
1109                         goto out;
1110                 if (ret) {
1111                         ret = 0;
1112                         goto out;
1113                 }
1114 
1115                 num++;
1116         }
1117 
1118 out:
1119         kvfree(buf);
1120         return ret;
1121 }
1122 
1123 static int __copy_first_ref(int num, u64 dir, int index,
1124                             struct fs_path *p, void *ctx)
1125 {
1126         int ret;
1127         struct fs_path *pt = ctx;
1128 
1129         ret = fs_path_copy(pt, p);
1130         if (ret < 0)
1131                 return ret;
1132 
1133         /* we want the first only */
1134         return 1;
1135 }
1136 
1137 /*
1138  * Retrieve the first path of an inode. If an inode has more then one
1139  * ref/hardlink, this is ignored.
1140  */
1141 static int get_inode_path(struct btrfs_root *root,
1142                           u64 ino, struct fs_path *path)
1143 {
1144         int ret;
1145         struct btrfs_key key, found_key;
1146         struct btrfs_path *p;
1147 
1148         p = alloc_path_for_send();
1149         if (!p)
1150                 return -ENOMEM;
1151 
1152         fs_path_reset(path);
1153 
1154         key.objectid = ino;
1155         key.type = BTRFS_INODE_REF_KEY;
1156         key.offset = 0;
1157 
1158         ret = btrfs_search_slot_for_read(root, &key, p, 1, 0);
1159         if (ret < 0)
1160                 goto out;
1161         if (ret) {
1162                 ret = 1;
1163                 goto out;
1164         }
1165         btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]);
1166         if (found_key.objectid != ino ||
1167             (found_key.type != BTRFS_INODE_REF_KEY &&
1168              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1169                 ret = -ENOENT;
1170                 goto out;
1171         }
1172 
1173         ret = iterate_inode_ref(root, p, &found_key, 1,
1174                                 __copy_first_ref, path);
1175         if (ret < 0)
1176                 goto out;
1177         ret = 0;
1178 
1179 out:
1180         btrfs_free_path(p);
1181         return ret;
1182 }
1183 
1184 struct backref_ctx {
1185         struct send_ctx *sctx;
1186 
1187         /* number of total found references */
1188         u64 found;
1189 
1190         /*
1191          * used for clones found in send_root. clones found behind cur_objectid
1192          * and cur_offset are not considered as allowed clones.
1193          */
1194         u64 cur_objectid;
1195         u64 cur_offset;
1196 
1197         /* may be truncated in case it's the last extent in a file */
1198         u64 extent_len;
1199 
1200         /* data offset in the file extent item */
1201         u64 data_offset;
1202 
1203         /* Just to check for bugs in backref resolving */
1204         int found_itself;
1205 };
1206 
1207 static int __clone_root_cmp_bsearch(const void *key, const void *elt)
1208 {
1209         u64 root = (u64)(uintptr_t)key;
1210         struct clone_root *cr = (struct clone_root *)elt;
1211 
1212         if (root < cr->root->root_key.objectid)
1213                 return -1;
1214         if (root > cr->root->root_key.objectid)
1215                 return 1;
1216         return 0;
1217 }
1218 
1219 static int __clone_root_cmp_sort(const void *e1, const void *e2)
1220 {
1221         struct clone_root *cr1 = (struct clone_root *)e1;
1222         struct clone_root *cr2 = (struct clone_root *)e2;
1223 
1224         if (cr1->root->root_key.objectid < cr2->root->root_key.objectid)
1225                 return -1;
1226         if (cr1->root->root_key.objectid > cr2->root->root_key.objectid)
1227                 return 1;
1228         return 0;
1229 }
1230 
1231 /*
1232  * Called for every backref that is found for the current extent.
1233  * Results are collected in sctx->clone_roots->ino/offset/found_refs
1234  */
1235 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_)
1236 {
1237         struct backref_ctx *bctx = ctx_;
1238         struct clone_root *found;
1239 
1240         /* First check if the root is in the list of accepted clone sources */
1241         found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots,
1242                         bctx->sctx->clone_roots_cnt,
1243                         sizeof(struct clone_root),
1244                         __clone_root_cmp_bsearch);
1245         if (!found)
1246                 return 0;
1247 
1248         if (found->root == bctx->sctx->send_root &&
1249             ino == bctx->cur_objectid &&
1250             offset == bctx->cur_offset) {
1251                 bctx->found_itself = 1;
1252         }
1253 
1254         /*
1255          * Make sure we don't consider clones from send_root that are
1256          * behind the current inode/offset.
1257          */
1258         if (found->root == bctx->sctx->send_root) {
1259                 /*
1260                  * If the source inode was not yet processed we can't issue a
1261                  * clone operation, as the source extent does not exist yet at
1262                  * the destination of the stream.
1263                  */
1264                 if (ino > bctx->cur_objectid)
1265                         return 0;
1266                 /*
1267                  * We clone from the inode currently being sent as long as the
1268                  * source extent is already processed, otherwise we could try
1269                  * to clone from an extent that does not exist yet at the
1270                  * destination of the stream.
1271                  */
1272                 if (ino == bctx->cur_objectid &&
1273                     offset + bctx->extent_len >
1274                     bctx->sctx->cur_inode_next_write_offset)
1275                         return 0;
1276         }
1277 
1278         bctx->found++;
1279         found->found_refs++;
1280         if (ino < found->ino) {
1281                 found->ino = ino;
1282                 found->offset = offset;
1283         } else if (found->ino == ino) {
1284                 /*
1285                  * same extent found more then once in the same file.
1286                  */
1287                 if (found->offset > offset + bctx->extent_len)
1288                         found->offset = offset;
1289         }
1290 
1291         return 0;
1292 }
1293 
1294 /*
1295  * Given an inode, offset and extent item, it finds a good clone for a clone
1296  * instruction. Returns -ENOENT when none could be found. The function makes
1297  * sure that the returned clone is usable at the point where sending is at the
1298  * moment. This means, that no clones are accepted which lie behind the current
1299  * inode+offset.
1300  *
1301  * path must point to the extent item when called.
1302  */
1303 static int find_extent_clone(struct send_ctx *sctx,
1304                              struct btrfs_path *path,
1305                              u64 ino, u64 data_offset,
1306                              u64 ino_size,
1307                              struct clone_root **found)
1308 {
1309         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
1310         int ret;
1311         int extent_type;
1312         u64 logical;
1313         u64 disk_byte;
1314         u64 num_bytes;
1315         u64 extent_item_pos;
1316         u64 flags = 0;
1317         struct btrfs_file_extent_item *fi;
1318         struct extent_buffer *eb = path->nodes[0];
1319         struct backref_ctx *backref_ctx = NULL;
1320         struct clone_root *cur_clone_root;
1321         struct btrfs_key found_key;
1322         struct btrfs_path *tmp_path;
1323         struct btrfs_extent_item *ei;
1324         int compressed;
1325         u32 i;
1326 
1327         tmp_path = alloc_path_for_send();
1328         if (!tmp_path)
1329                 return -ENOMEM;
1330 
1331         /* We only use this path under the commit sem */
1332         tmp_path->need_commit_sem = 0;
1333 
1334         backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL);
1335         if (!backref_ctx) {
1336                 ret = -ENOMEM;
1337                 goto out;
1338         }
1339 
1340         if (data_offset >= ino_size) {
1341                 /*
1342                  * There may be extents that lie behind the file's size.
1343                  * I at least had this in combination with snapshotting while
1344                  * writing large files.
1345                  */
1346                 ret = 0;
1347                 goto out;
1348         }
1349 
1350         fi = btrfs_item_ptr(eb, path->slots[0],
1351                         struct btrfs_file_extent_item);
1352         extent_type = btrfs_file_extent_type(eb, fi);
1353         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1354                 ret = -ENOENT;
1355                 goto out;
1356         }
1357         compressed = btrfs_file_extent_compression(eb, fi);
1358 
1359         num_bytes = btrfs_file_extent_num_bytes(eb, fi);
1360         disk_byte = btrfs_file_extent_disk_bytenr(eb, fi);
1361         if (disk_byte == 0) {
1362                 ret = -ENOENT;
1363                 goto out;
1364         }
1365         logical = disk_byte + btrfs_file_extent_offset(eb, fi);
1366 
1367         down_read(&fs_info->commit_root_sem);
1368         ret = extent_from_logical(fs_info, disk_byte, tmp_path,
1369                                   &found_key, &flags);
1370         up_read(&fs_info->commit_root_sem);
1371 
1372         if (ret < 0)
1373                 goto out;
1374         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1375                 ret = -EIO;
1376                 goto out;
1377         }
1378 
1379         ei = btrfs_item_ptr(tmp_path->nodes[0], tmp_path->slots[0],
1380                             struct btrfs_extent_item);
1381         /*
1382          * Backreference walking (iterate_extent_inodes() below) is currently
1383          * too expensive when an extent has a large number of references, both
1384          * in time spent and used memory. So for now just fallback to write
1385          * operations instead of clone operations when an extent has more than
1386          * a certain amount of references.
1387          */
1388         if (btrfs_extent_refs(tmp_path->nodes[0], ei) > SEND_MAX_EXTENT_REFS) {
1389                 ret = -ENOENT;
1390                 goto out;
1391         }
1392         btrfs_release_path(tmp_path);
1393 
1394         /*
1395          * Setup the clone roots.
1396          */
1397         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1398                 cur_clone_root = sctx->clone_roots + i;
1399                 cur_clone_root->ino = (u64)-1;
1400                 cur_clone_root->offset = 0;
1401                 cur_clone_root->found_refs = 0;
1402         }
1403 
1404         backref_ctx->sctx = sctx;
1405         backref_ctx->found = 0;
1406         backref_ctx->cur_objectid = ino;
1407         backref_ctx->cur_offset = data_offset;
1408         backref_ctx->found_itself = 0;
1409         backref_ctx->extent_len = num_bytes;
1410         /*
1411          * For non-compressed extents iterate_extent_inodes() gives us extent
1412          * offsets that already take into account the data offset, but not for
1413          * compressed extents, since the offset is logical and not relative to
1414          * the physical extent locations. We must take this into account to
1415          * avoid sending clone offsets that go beyond the source file's size,
1416          * which would result in the clone ioctl failing with -EINVAL on the
1417          * receiving end.
1418          */
1419         if (compressed == BTRFS_COMPRESS_NONE)
1420                 backref_ctx->data_offset = 0;
1421         else
1422                 backref_ctx->data_offset = btrfs_file_extent_offset(eb, fi);
1423 
1424         /*
1425          * The last extent of a file may be too large due to page alignment.
1426          * We need to adjust extent_len in this case so that the checks in
1427          * __iterate_backrefs work.
1428          */
1429         if (data_offset + num_bytes >= ino_size)
1430                 backref_ctx->extent_len = ino_size - data_offset;
1431 
1432         /*
1433          * Now collect all backrefs.
1434          */
1435         if (compressed == BTRFS_COMPRESS_NONE)
1436                 extent_item_pos = logical - found_key.objectid;
1437         else
1438                 extent_item_pos = 0;
1439         ret = iterate_extent_inodes(fs_info, found_key.objectid,
1440                                     extent_item_pos, 1, __iterate_backrefs,
1441                                     backref_ctx, false);
1442 
1443         if (ret < 0)
1444                 goto out;
1445 
1446         if (!backref_ctx->found_itself) {
1447                 /* found a bug in backref code? */
1448                 ret = -EIO;
1449                 btrfs_err(fs_info,
1450                           "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu",
1451                           ino, data_offset, disk_byte, found_key.objectid);
1452                 goto out;
1453         }
1454 
1455         btrfs_debug(fs_info,
1456                     "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu",
1457                     data_offset, ino, num_bytes, logical);
1458 
1459         if (!backref_ctx->found)
1460                 btrfs_debug(fs_info, "no clones found");
1461 
1462         cur_clone_root = NULL;
1463         for (i = 0; i < sctx->clone_roots_cnt; i++) {
1464                 if (sctx->clone_roots[i].found_refs) {
1465                         if (!cur_clone_root)
1466                                 cur_clone_root = sctx->clone_roots + i;
1467                         else if (sctx->clone_roots[i].root == sctx->send_root)
1468                                 /* prefer clones from send_root over others */
1469                                 cur_clone_root = sctx->clone_roots + i;
1470                 }
1471 
1472         }
1473 
1474         if (cur_clone_root) {
1475                 *found = cur_clone_root;
1476                 ret = 0;
1477         } else {
1478                 ret = -ENOENT;
1479         }
1480 
1481 out:
1482         btrfs_free_path(tmp_path);
1483         kfree(backref_ctx);
1484         return ret;
1485 }
1486 
1487 static int read_symlink(struct btrfs_root *root,
1488                         u64 ino,
1489                         struct fs_path *dest)
1490 {
1491         int ret;
1492         struct btrfs_path *path;
1493         struct btrfs_key key;
1494         struct btrfs_file_extent_item *ei;
1495         u8 type;
1496         u8 compression;
1497         unsigned long off;
1498         int len;
1499 
1500         path = alloc_path_for_send();
1501         if (!path)
1502                 return -ENOMEM;
1503 
1504         key.objectid = ino;
1505         key.type = BTRFS_EXTENT_DATA_KEY;
1506         key.offset = 0;
1507         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1508         if (ret < 0)
1509                 goto out;
1510         if (ret) {
1511                 /*
1512                  * An empty symlink inode. Can happen in rare error paths when
1513                  * creating a symlink (transaction committed before the inode
1514                  * eviction handler removed the symlink inode items and a crash
1515                  * happened in between or the subvol was snapshoted in between).
1516                  * Print an informative message to dmesg/syslog so that the user
1517                  * can delete the symlink.
1518                  */
1519                 btrfs_err(root->fs_info,
1520                           "Found empty symlink inode %llu at root %llu",
1521                           ino, root->root_key.objectid);
1522                 ret = -EIO;
1523                 goto out;
1524         }
1525 
1526         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
1527                         struct btrfs_file_extent_item);
1528         type = btrfs_file_extent_type(path->nodes[0], ei);
1529         compression = btrfs_file_extent_compression(path->nodes[0], ei);
1530         BUG_ON(type != BTRFS_FILE_EXTENT_INLINE);
1531         BUG_ON(compression);
1532 
1533         off = btrfs_file_extent_inline_start(ei);
1534         len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
1535 
1536         ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len);
1537 
1538 out:
1539         btrfs_free_path(path);
1540         return ret;
1541 }
1542 
1543 /*
1544  * Helper function to generate a file name that is unique in the root of
1545  * send_root and parent_root. This is used to generate names for orphan inodes.
1546  */
1547 static int gen_unique_name(struct send_ctx *sctx,
1548                            u64 ino, u64 gen,
1549                            struct fs_path *dest)
1550 {
1551         int ret = 0;
1552         struct btrfs_path *path;
1553         struct btrfs_dir_item *di;
1554         char tmp[64];
1555         int len;
1556         u64 idx = 0;
1557 
1558         path = alloc_path_for_send();
1559         if (!path)
1560                 return -ENOMEM;
1561 
1562         while (1) {
1563                 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu",
1564                                 ino, gen, idx);
1565                 ASSERT(len < sizeof(tmp));
1566 
1567                 di = btrfs_lookup_dir_item(NULL, sctx->send_root,
1568                                 path, BTRFS_FIRST_FREE_OBJECTID,
1569                                 tmp, strlen(tmp), 0);
1570                 btrfs_release_path(path);
1571                 if (IS_ERR(di)) {
1572                         ret = PTR_ERR(di);
1573                         goto out;
1574                 }
1575                 if (di) {
1576                         /* not unique, try again */
1577                         idx++;
1578                         continue;
1579                 }
1580 
1581                 if (!sctx->parent_root) {
1582                         /* unique */
1583                         ret = 0;
1584                         break;
1585                 }
1586 
1587                 di = btrfs_lookup_dir_item(NULL, sctx->parent_root,
1588                                 path, BTRFS_FIRST_FREE_OBJECTID,
1589                                 tmp, strlen(tmp), 0);
1590                 btrfs_release_path(path);
1591                 if (IS_ERR(di)) {
1592                         ret = PTR_ERR(di);
1593                         goto out;
1594                 }
1595                 if (di) {
1596                         /* not unique, try again */
1597                         idx++;
1598                         continue;
1599                 }
1600                 /* unique */
1601                 break;
1602         }
1603 
1604         ret = fs_path_add(dest, tmp, strlen(tmp));
1605 
1606 out:
1607         btrfs_free_path(path);
1608         return ret;
1609 }
1610 
1611 enum inode_state {
1612         inode_state_no_change,
1613         inode_state_will_create,
1614         inode_state_did_create,
1615         inode_state_will_delete,
1616         inode_state_did_delete,
1617 };
1618 
1619 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen)
1620 {
1621         int ret;
1622         int left_ret;
1623         int right_ret;
1624         u64 left_gen;
1625         u64 right_gen;
1626 
1627         ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL,
1628                         NULL, NULL);
1629         if (ret < 0 && ret != -ENOENT)
1630                 goto out;
1631         left_ret = ret;
1632 
1633         if (!sctx->parent_root) {
1634                 right_ret = -ENOENT;
1635         } else {
1636                 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen,
1637                                 NULL, NULL, NULL, NULL);
1638                 if (ret < 0 && ret != -ENOENT)
1639                         goto out;
1640                 right_ret = ret;
1641         }
1642 
1643         if (!left_ret && !right_ret) {
1644                 if (left_gen == gen && right_gen == gen) {
1645                         ret = inode_state_no_change;
1646                 } else if (left_gen == gen) {
1647                         if (ino < sctx->send_progress)
1648                                 ret = inode_state_did_create;
1649                         else
1650                                 ret = inode_state_will_create;
1651                 } else if (right_gen == gen) {
1652                         if (ino < sctx->send_progress)
1653                                 ret = inode_state_did_delete;
1654                         else
1655                                 ret = inode_state_will_delete;
1656                 } else  {
1657                         ret = -ENOENT;
1658                 }
1659         } else if (!left_ret) {
1660                 if (left_gen == gen) {
1661                         if (ino < sctx->send_progress)
1662                                 ret = inode_state_did_create;
1663                         else
1664                                 ret = inode_state_will_create;
1665                 } else {
1666                         ret = -ENOENT;
1667                 }
1668         } else if (!right_ret) {
1669                 if (right_gen == gen) {
1670                         if (ino < sctx->send_progress)
1671                                 ret = inode_state_did_delete;
1672                         else
1673                                 ret = inode_state_will_delete;
1674                 } else {
1675                         ret = -ENOENT;
1676                 }
1677         } else {
1678                 ret = -ENOENT;
1679         }
1680 
1681 out:
1682         return ret;
1683 }
1684 
1685 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen)
1686 {
1687         int ret;
1688 
1689         if (ino == BTRFS_FIRST_FREE_OBJECTID)
1690                 return 1;
1691 
1692         ret = get_cur_inode_state(sctx, ino, gen);
1693         if (ret < 0)
1694                 goto out;
1695 
1696         if (ret == inode_state_no_change ||
1697             ret == inode_state_did_create ||
1698             ret == inode_state_will_delete)
1699                 ret = 1;
1700         else
1701                 ret = 0;
1702 
1703 out:
1704         return ret;
1705 }
1706 
1707 /*
1708  * Helper function to lookup a dir item in a dir.
1709  */
1710 static int lookup_dir_item_inode(struct btrfs_root *root,
1711                                  u64 dir, const char *name, int name_len,
1712                                  u64 *found_inode,
1713                                  u8 *found_type)
1714 {
1715         int ret = 0;
1716         struct btrfs_dir_item *di;
1717         struct btrfs_key key;
1718         struct btrfs_path *path;
1719 
1720         path = alloc_path_for_send();
1721         if (!path)
1722                 return -ENOMEM;
1723 
1724         di = btrfs_lookup_dir_item(NULL, root, path,
1725                         dir, name, name_len, 0);
1726         if (IS_ERR_OR_NULL(di)) {
1727                 ret = di ? PTR_ERR(di) : -ENOENT;
1728                 goto out;
1729         }
1730         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
1731         if (key.type == BTRFS_ROOT_ITEM_KEY) {
1732                 ret = -ENOENT;
1733                 goto out;
1734         }
1735         *found_inode = key.objectid;
1736         *found_type = btrfs_dir_type(path->nodes[0], di);
1737 
1738 out:
1739         btrfs_free_path(path);
1740         return ret;
1741 }
1742 
1743 /*
1744  * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir,
1745  * generation of the parent dir and the name of the dir entry.
1746  */
1747 static int get_first_ref(struct btrfs_root *root, u64 ino,
1748                          u64 *dir, u64 *dir_gen, struct fs_path *name)
1749 {
1750         int ret;
1751         struct btrfs_key key;
1752         struct btrfs_key found_key;
1753         struct btrfs_path *path;
1754         int len;
1755         u64 parent_dir;
1756 
1757         path = alloc_path_for_send();
1758         if (!path)
1759                 return -ENOMEM;
1760 
1761         key.objectid = ino;
1762         key.type = BTRFS_INODE_REF_KEY;
1763         key.offset = 0;
1764 
1765         ret = btrfs_search_slot_for_read(root, &key, path, 1, 0);
1766         if (ret < 0)
1767                 goto out;
1768         if (!ret)
1769                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1770                                 path->slots[0]);
1771         if (ret || found_key.objectid != ino ||
1772             (found_key.type != BTRFS_INODE_REF_KEY &&
1773              found_key.type != BTRFS_INODE_EXTREF_KEY)) {
1774                 ret = -ENOENT;
1775                 goto out;
1776         }
1777 
1778         if (found_key.type == BTRFS_INODE_REF_KEY) {
1779                 struct btrfs_inode_ref *iref;
1780                 iref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1781                                       struct btrfs_inode_ref);
1782                 len = btrfs_inode_ref_name_len(path->nodes[0], iref);
1783                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1784                                                      (unsigned long)(iref + 1),
1785                                                      len);
1786                 parent_dir = found_key.offset;
1787         } else {
1788                 struct btrfs_inode_extref *extref;
1789                 extref = btrfs_item_ptr(path->nodes[0], path->slots[0],
1790                                         struct btrfs_inode_extref);
1791                 len = btrfs_inode_extref_name_len(path->nodes[0], extref);
1792                 ret = fs_path_add_from_extent_buffer(name, path->nodes[0],
1793                                         (unsigned long)&extref->name, len);
1794                 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref);
1795         }
1796         if (ret < 0)
1797                 goto out;
1798         btrfs_release_path(path);
1799 
1800         if (dir_gen) {
1801                 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL,
1802                                      NULL, NULL, NULL);
1803                 if (ret < 0)
1804                         goto out;
1805         }
1806 
1807         *dir = parent_dir;
1808 
1809 out:
1810         btrfs_free_path(path);
1811         return ret;
1812 }
1813 
1814 static int is_first_ref(struct btrfs_root *root,
1815                         u64 ino, u64 dir,
1816                         const char *name, int name_len)
1817 {
1818         int ret;
1819         struct fs_path *tmp_name;
1820         u64 tmp_dir;
1821 
1822         tmp_name = fs_path_alloc();
1823         if (!tmp_name)
1824                 return -ENOMEM;
1825 
1826         ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name);
1827         if (ret < 0)
1828                 goto out;
1829 
1830         if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) {
1831                 ret = 0;
1832                 goto out;
1833         }
1834 
1835         ret = !memcmp(tmp_name->start, name, name_len);
1836 
1837 out:
1838         fs_path_free(tmp_name);
1839         return ret;
1840 }
1841 
1842 /*
1843  * Used by process_recorded_refs to determine if a new ref would overwrite an
1844  * already existing ref. In case it detects an overwrite, it returns the
1845  * inode/gen in who_ino/who_gen.
1846  * When an overwrite is detected, process_recorded_refs does proper orphanizing
1847  * to make sure later references to the overwritten inode are possible.
1848  * Orphanizing is however only required for the first ref of an inode.
1849  * process_recorded_refs does an additional is_first_ref check to see if
1850  * orphanizing is really required.
1851  */
1852 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen,
1853                               const char *name, int name_len,
1854                               u64 *who_ino, u64 *who_gen, u64 *who_mode)
1855 {
1856         int ret = 0;
1857         u64 gen;
1858         u64 other_inode = 0;
1859         u8 other_type = 0;
1860 
1861         if (!sctx->parent_root)
1862                 goto out;
1863 
1864         ret = is_inode_existent(sctx, dir, dir_gen);
1865         if (ret <= 0)
1866                 goto out;
1867 
1868         /*
1869          * If we have a parent root we need to verify that the parent dir was
1870          * not deleted and then re-created, if it was then we have no overwrite
1871          * and we can just unlink this entry.
1872          */
1873         if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) {
1874                 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL,
1875                                      NULL, NULL, NULL);
1876                 if (ret < 0 && ret != -ENOENT)
1877                         goto out;
1878                 if (ret) {
1879                         ret = 0;
1880                         goto out;
1881                 }
1882                 if (gen != dir_gen)
1883                         goto out;
1884         }
1885 
1886         ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len,
1887                         &other_inode, &other_type);
1888         if (ret < 0 && ret != -ENOENT)
1889                 goto out;
1890         if (ret) {
1891                 ret = 0;
1892                 goto out;
1893         }
1894 
1895         /*
1896          * Check if the overwritten ref was already processed. If yes, the ref
1897          * was already unlinked/moved, so we can safely assume that we will not
1898          * overwrite anything at this point in time.
1899          */
1900         if (other_inode > sctx->send_progress ||
1901             is_waiting_for_move(sctx, other_inode)) {
1902                 ret = get_inode_info(sctx->parent_root, other_inode, NULL,
1903                                 who_gen, who_mode, NULL, NULL, NULL);
1904                 if (ret < 0)
1905                         goto out;
1906 
1907                 ret = 1;
1908                 *who_ino = other_inode;
1909         } else {
1910                 ret = 0;
1911         }
1912 
1913 out:
1914         return ret;
1915 }
1916 
1917 /*
1918  * Checks if the ref was overwritten by an already processed inode. This is
1919  * used by __get_cur_name_and_parent to find out if the ref was orphanized and
1920  * thus the orphan name needs be used.
1921  * process_recorded_refs also uses it to avoid unlinking of refs that were
1922  * overwritten.
1923  */
1924 static int did_overwrite_ref(struct send_ctx *sctx,
1925                             u64 dir, u64 dir_gen,
1926                             u64 ino, u64 ino_gen,
1927                             const char *name, int name_len)
1928 {
1929         int ret = 0;
1930         u64 gen;
1931         u64 ow_inode;
1932         u8 other_type;
1933 
1934         if (!sctx->parent_root)
1935                 goto out;
1936 
1937         ret = is_inode_existent(sctx, dir, dir_gen);
1938         if (ret <= 0)
1939                 goto out;
1940 
1941         if (dir != BTRFS_FIRST_FREE_OBJECTID) {
1942                 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL,
1943                                      NULL, NULL, NULL);
1944                 if (ret < 0 && ret != -ENOENT)
1945                         goto out;
1946                 if (ret) {
1947                         ret = 0;
1948                         goto out;
1949                 }
1950                 if (gen != dir_gen)
1951                         goto out;
1952         }
1953 
1954         /* check if the ref was overwritten by another ref */
1955         ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len,
1956                         &ow_inode, &other_type);
1957         if (ret < 0 && ret != -ENOENT)
1958                 goto out;
1959         if (ret) {
1960                 /* was never and will never be overwritten */
1961                 ret = 0;
1962                 goto out;
1963         }
1964 
1965         ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL,
1966                         NULL, NULL);
1967         if (ret < 0)
1968                 goto out;
1969 
1970         if (ow_inode == ino && gen == ino_gen) {
1971                 ret = 0;
1972                 goto out;
1973         }
1974 
1975         /*
1976          * We know that it is or will be overwritten. Check this now.
1977          * The current inode being processed might have been the one that caused
1978          * inode 'ino' to be orphanized, therefore check if ow_inode matches
1979          * the current inode being processed.
1980          */
1981         if ((ow_inode < sctx->send_progress) ||
1982             (ino != sctx->cur_ino && ow_inode == sctx->cur_ino &&
1983              gen == sctx->cur_inode_gen))
1984                 ret = 1;
1985         else
1986                 ret = 0;
1987 
1988 out:
1989         return ret;
1990 }
1991 
1992 /*
1993  * Same as did_overwrite_ref, but also checks if it is the first ref of an inode
1994  * that got overwritten. This is used by process_recorded_refs to determine
1995  * if it has to use the path as returned by get_cur_path or the orphan name.
1996  */
1997 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen)
1998 {
1999         int ret = 0;
2000         struct fs_path *name = NULL;
2001         u64 dir;
2002         u64 dir_gen;
2003 
2004         if (!sctx->parent_root)
2005                 goto out;
2006 
2007         name = fs_path_alloc();
2008         if (!name)
2009                 return -ENOMEM;
2010 
2011         ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name);
2012         if (ret < 0)
2013                 goto out;
2014 
2015         ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen,
2016                         name->start, fs_path_len(name));
2017 
2018 out:
2019         fs_path_free(name);
2020         return ret;
2021 }
2022 
2023 /*
2024  * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit,
2025  * so we need to do some special handling in case we have clashes. This function
2026  * takes care of this with the help of name_cache_entry::radix_list.
2027  * In case of error, nce is kfreed.
2028  */
2029 static int name_cache_insert(struct send_ctx *sctx,
2030                              struct name_cache_entry *nce)
2031 {
2032         int ret = 0;
2033         struct list_head *nce_head;
2034 
2035         nce_head = radix_tree_lookup(&sctx->name_cache,
2036                         (unsigned long)nce->ino);
2037         if (!nce_head) {
2038                 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL);
2039                 if (!nce_head) {
2040                         kfree(nce);
2041                         return -ENOMEM;
2042                 }
2043                 INIT_LIST_HEAD(nce_head);
2044 
2045                 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head);
2046                 if (ret < 0) {
2047                         kfree(nce_head);
2048                         kfree(nce);
2049                         return ret;
2050                 }
2051         }
2052         list_add_tail(&nce->radix_list, nce_head);
2053         list_add_tail(&nce->list, &sctx->name_cache_list);
2054         sctx->name_cache_size++;
2055 
2056         return ret;
2057 }
2058 
2059 static void name_cache_delete(struct send_ctx *sctx,
2060                               struct name_cache_entry *nce)
2061 {
2062         struct list_head *nce_head;
2063 
2064         nce_head = radix_tree_lookup(&sctx->name_cache,
2065                         (unsigned long)nce->ino);
2066         if (!nce_head) {
2067                 btrfs_err(sctx->send_root->fs_info,
2068               "name_cache_delete lookup failed ino %llu cache size %d, leaking memory",
2069                         nce->ino, sctx->name_cache_size);
2070         }
2071 
2072         list_del(&nce->radix_list);
2073         list_del(&nce->list);
2074         sctx->name_cache_size--;
2075 
2076         /*
2077          * We may not get to the final release of nce_head if the lookup fails
2078          */
2079         if (nce_head && list_empty(nce_head)) {
2080                 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino);
2081                 kfree(nce_head);
2082         }
2083 }
2084 
2085 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx,
2086                                                     u64 ino, u64 gen)
2087 {
2088         struct list_head *nce_head;
2089         struct name_cache_entry *cur;
2090 
2091         nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino);
2092         if (!nce_head)
2093                 return NULL;
2094 
2095         list_for_each_entry(cur, nce_head, radix_list) {
2096                 if (cur->ino == ino && cur->gen == gen)
2097                         return cur;
2098         }
2099         return NULL;
2100 }
2101 
2102 /*
2103  * Removes the entry from the list and adds it back to the end. This marks the
2104  * entry as recently used so that name_cache_clean_unused does not remove it.
2105  */
2106 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce)
2107 {
2108         list_del(&nce->list);
2109         list_add_tail(&nce->list, &sctx->name_cache_list);
2110 }
2111 
2112 /*
2113  * Remove some entries from the beginning of name_cache_list.
2114  */
2115 static void name_cache_clean_unused(struct send_ctx *sctx)
2116 {
2117         struct name_cache_entry *nce;
2118 
2119         if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE)
2120                 return;
2121 
2122         while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) {
2123                 nce = list_entry(sctx->name_cache_list.next,
2124                                 struct name_cache_entry, list);
2125                 name_cache_delete(sctx, nce);
2126                 kfree(nce);
2127         }
2128 }
2129 
2130 static void name_cache_free(struct send_ctx *sctx)
2131 {
2132         struct name_cache_entry *nce;
2133 
2134         while (!list_empty(&sctx->name_cache_list)) {
2135                 nce = list_entry(sctx->name_cache_list.next,
2136                                 struct name_cache_entry, list);
2137                 name_cache_delete(sctx, nce);
2138                 kfree(nce);
2139         }
2140 }
2141 
2142 /*
2143  * Used by get_cur_path for each ref up to the root.
2144  * Returns 0 if it succeeded.
2145  * Returns 1 if the inode is not existent or got overwritten. In that case, the
2146  * name is an orphan name. This instructs get_cur_path to stop iterating. If 1
2147  * is returned, parent_ino/parent_gen are not guaranteed to be valid.
2148  * Returns <0 in case of error.
2149  */
2150 static int __get_cur_name_and_parent(struct send_ctx *sctx,
2151                                      u64 ino, u64 gen,
2152                                      u64 *parent_ino,
2153                                      u64 *parent_gen,
2154                                      struct fs_path *dest)
2155 {
2156         int ret;
2157         int nce_ret;
2158         struct name_cache_entry *nce = NULL;
2159 
2160         /*
2161          * First check if we already did a call to this function with the same
2162          * ino/gen. If yes, check if the cache entry is still up-to-date. If yes
2163          * return the cached result.
2164          */
2165         nce = name_cache_search(sctx, ino, gen);
2166         if (nce) {
2167                 if (ino < sctx->send_progress && nce->need_later_update) {
2168                         name_cache_delete(sctx, nce);
2169                         kfree(nce);
2170                         nce = NULL;
2171                 } else {
2172                         name_cache_used(sctx, nce);
2173                         *parent_ino = nce->parent_ino;
2174                         *parent_gen = nce->parent_gen;
2175                         ret = fs_path_add(dest, nce->name, nce->name_len);
2176                         if (ret < 0)
2177                                 goto out;
2178                         ret = nce->ret;
2179                         goto out;
2180                 }
2181         }
2182 
2183         /*
2184          * If the inode is not existent yet, add the orphan name and return 1.
2185          * This should only happen for the parent dir that we determine in
2186          * __record_new_ref
2187          */
2188         ret = is_inode_existent(sctx, ino, gen);
2189         if (ret < 0)
2190                 goto out;
2191 
2192         if (!ret) {
2193                 ret = gen_unique_name(sctx, ino, gen, dest);
2194                 if (ret < 0)
2195                         goto out;
2196                 ret = 1;
2197                 goto out_cache;
2198         }
2199 
2200         /*
2201          * Depending on whether the inode was already processed or not, use
2202          * send_root or parent_root for ref lookup.
2203          */
2204         if (ino < sctx->send_progress)
2205                 ret = get_first_ref(sctx->send_root, ino,
2206                                     parent_ino, parent_gen, dest);
2207         else
2208                 ret = get_first_ref(sctx->parent_root, ino,
2209                                     parent_ino, parent_gen, dest);
2210         if (ret < 0)
2211                 goto out;
2212 
2213         /*
2214          * Check if the ref was overwritten by an inode's ref that was processed
2215          * earlier. If yes, treat as orphan and return 1.
2216          */
2217         ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen,
2218                         dest->start, dest->end - dest->start);
2219         if (ret < 0)
2220                 goto out;
2221         if (ret) {
2222                 fs_path_reset(dest);
2223                 ret = gen_unique_name(sctx, ino, gen, dest);
2224                 if (ret < 0)
2225                         goto out;
2226                 ret = 1;
2227         }
2228 
2229 out_cache:
2230         /*
2231          * Store the result of the lookup in the name cache.
2232          */
2233         nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL);
2234         if (!nce) {
2235                 ret = -ENOMEM;
2236                 goto out;
2237         }
2238 
2239         nce->ino = ino;
2240         nce->gen = gen;
2241         nce->parent_ino = *parent_ino;
2242         nce->parent_gen = *parent_gen;
2243         nce->name_len = fs_path_len(dest);
2244         nce->ret = ret;
2245         strcpy(nce->name, dest->start);
2246 
2247         if (ino < sctx->send_progress)
2248                 nce->need_later_update = 0;
2249         else
2250                 nce->need_later_update = 1;
2251 
2252         nce_ret = name_cache_insert(sctx, nce);
2253         if (nce_ret < 0)
2254                 ret = nce_ret;
2255         name_cache_clean_unused(sctx);
2256 
2257 out:
2258         return ret;
2259 }
2260 
2261 /*
2262  * Magic happens here. This function returns the first ref to an inode as it
2263  * would look like while receiving the stream at this point in time.
2264  * We walk the path up to the root. For every inode in between, we check if it
2265  * was already processed/sent. If yes, we continue with the parent as found
2266  * in send_root. If not, we continue with the parent as found in parent_root.
2267  * If we encounter an inode that was deleted at this point in time, we use the
2268  * inodes "orphan" name instead of the real name and stop. Same with new inodes
2269  * that were not created yet and overwritten inodes/refs.
2270  *
2271  * When do we have orphan inodes:
2272  * 1. When an inode is freshly created and thus no valid refs are available yet
2273  * 2. When a directory lost all it's refs (deleted) but still has dir items
2274  *    inside which were not processed yet (pending for move/delete). If anyone
2275  *    tried to get the path to the dir items, it would get a path inside that
2276  *    orphan directory.
2277  * 3. When an inode is moved around or gets new links, it may overwrite the ref
2278  *    of an unprocessed inode. If in that case the first ref would be
2279  *    overwritten, the overwritten inode gets "orphanized". Later when we
2280  *    process this overwritten inode, it is restored at a new place by moving
2281  *    the orphan inode.
2282  *
2283  * sctx->send_progress tells this function at which point in time receiving
2284  * would be.
2285  */
2286 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen,
2287                         struct fs_path *dest)
2288 {
2289         int ret = 0;
2290         struct fs_path *name = NULL;
2291         u64 parent_inode = 0;
2292         u64 parent_gen = 0;
2293         int stop = 0;
2294 
2295         name = fs_path_alloc();
2296         if (!name) {
2297                 ret = -ENOMEM;
2298                 goto out;
2299         }
2300 
2301         dest->reversed = 1;
2302         fs_path_reset(dest);
2303 
2304         while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) {
2305                 struct waiting_dir_move *wdm;
2306 
2307                 fs_path_reset(name);
2308 
2309                 if (is_waiting_for_rm(sctx, ino)) {
2310                         ret = gen_unique_name(sctx, ino, gen, name);
2311                         if (ret < 0)
2312                                 goto out;
2313                         ret = fs_path_add_path(dest, name);
2314                         break;
2315                 }
2316 
2317                 wdm = get_waiting_dir_move(sctx, ino);
2318                 if (wdm && wdm->orphanized) {
2319                         ret = gen_unique_name(sctx, ino, gen, name);
2320                         stop = 1;
2321                 } else if (wdm) {
2322                         ret = get_first_ref(sctx->parent_root, ino,
2323                                             &parent_inode, &parent_gen, name);
2324                 } else {
2325                         ret = __get_cur_name_and_parent(sctx, ino, gen,
2326                                                         &parent_inode,
2327                                                         &parent_gen, name);
2328                         if (ret)
2329                                 stop = 1;
2330                 }
2331 
2332                 if (ret < 0)
2333                         goto out;
2334 
2335                 ret = fs_path_add_path(dest, name);
2336                 if (ret < 0)
2337                         goto out;
2338 
2339                 ino = parent_inode;
2340                 gen = parent_gen;
2341         }
2342 
2343 out:
2344         fs_path_free(name);
2345         if (!ret)
2346                 fs_path_unreverse(dest);
2347         return ret;
2348 }
2349 
2350 /*
2351  * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace
2352  */
2353 static int send_subvol_begin(struct send_ctx *sctx)
2354 {
2355         int ret;
2356         struct btrfs_root *send_root = sctx->send_root;
2357         struct btrfs_root *parent_root = sctx->parent_root;
2358         struct btrfs_path *path;
2359         struct btrfs_key key;
2360         struct btrfs_root_ref *ref;
2361         struct extent_buffer *leaf;
2362         char *name = NULL;
2363         int namelen;
2364 
2365         path = btrfs_alloc_path();
2366         if (!path)
2367                 return -ENOMEM;
2368 
2369         name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL);
2370         if (!name) {
2371                 btrfs_free_path(path);
2372                 return -ENOMEM;
2373         }
2374 
2375         key.objectid = send_root->root_key.objectid;
2376         key.type = BTRFS_ROOT_BACKREF_KEY;
2377         key.offset = 0;
2378 
2379         ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root,
2380                                 &key, path, 1, 0);
2381         if (ret < 0)
2382                 goto out;
2383         if (ret) {
2384                 ret = -ENOENT;
2385                 goto out;
2386         }
2387 
2388         leaf = path->nodes[0];
2389         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2390         if (key.type != BTRFS_ROOT_BACKREF_KEY ||
2391             key.objectid != send_root->root_key.objectid) {
2392                 ret = -ENOENT;
2393                 goto out;
2394         }
2395         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
2396         namelen = btrfs_root_ref_name_len(leaf, ref);
2397         read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen);
2398         btrfs_release_path(path);
2399 
2400         if (parent_root) {
2401                 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT);
2402                 if (ret < 0)
2403                         goto out;
2404         } else {
2405                 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL);
2406                 if (ret < 0)
2407                         goto out;
2408         }
2409 
2410         TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen);
2411 
2412         if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid))
2413                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2414                             sctx->send_root->root_item.received_uuid);
2415         else
2416                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID,
2417                             sctx->send_root->root_item.uuid);
2418 
2419         TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID,
2420                     le64_to_cpu(sctx->send_root->root_item.ctransid));
2421         if (parent_root) {
2422                 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid))
2423                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2424                                      parent_root->root_item.received_uuid);
2425                 else
2426                         TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
2427                                      parent_root->root_item.uuid);
2428                 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
2429                             le64_to_cpu(sctx->parent_root->root_item.ctransid));
2430         }
2431 
2432         ret = send_cmd(sctx);
2433 
2434 tlv_put_failure:
2435 out:
2436         btrfs_free_path(path);
2437         kfree(name);
2438         return ret;
2439 }
2440 
2441 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size)
2442 {
2443         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2444         int ret = 0;
2445         struct fs_path *p;
2446 
2447         btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size);
2448 
2449         p = fs_path_alloc();
2450         if (!p)
2451                 return -ENOMEM;
2452 
2453         ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE);
2454         if (ret < 0)
2455                 goto out;
2456 
2457         ret = get_cur_path(sctx, ino, gen, p);
2458         if (ret < 0)
2459                 goto out;
2460         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2461         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size);
2462 
2463         ret = send_cmd(sctx);
2464 
2465 tlv_put_failure:
2466 out:
2467         fs_path_free(p);
2468         return ret;
2469 }
2470 
2471 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode)
2472 {
2473         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2474         int ret = 0;
2475         struct fs_path *p;
2476 
2477         btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode);
2478 
2479         p = fs_path_alloc();
2480         if (!p)
2481                 return -ENOMEM;
2482 
2483         ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD);
2484         if (ret < 0)
2485                 goto out;
2486 
2487         ret = get_cur_path(sctx, ino, gen, p);
2488         if (ret < 0)
2489                 goto out;
2490         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2491         TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777);
2492 
2493         ret = send_cmd(sctx);
2494 
2495 tlv_put_failure:
2496 out:
2497         fs_path_free(p);
2498         return ret;
2499 }
2500 
2501 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid)
2502 {
2503         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2504         int ret = 0;
2505         struct fs_path *p;
2506 
2507         btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu",
2508                     ino, uid, gid);
2509 
2510         p = fs_path_alloc();
2511         if (!p)
2512                 return -ENOMEM;
2513 
2514         ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN);
2515         if (ret < 0)
2516                 goto out;
2517 
2518         ret = get_cur_path(sctx, ino, gen, p);
2519         if (ret < 0)
2520                 goto out;
2521         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2522         TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid);
2523         TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid);
2524 
2525         ret = send_cmd(sctx);
2526 
2527 tlv_put_failure:
2528 out:
2529         fs_path_free(p);
2530         return ret;
2531 }
2532 
2533 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen)
2534 {
2535         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2536         int ret = 0;
2537         struct fs_path *p = NULL;
2538         struct btrfs_inode_item *ii;
2539         struct btrfs_path *path = NULL;
2540         struct extent_buffer *eb;
2541         struct btrfs_key key;
2542         int slot;
2543 
2544         btrfs_debug(fs_info, "send_utimes %llu", ino);
2545 
2546         p = fs_path_alloc();
2547         if (!p)
2548                 return -ENOMEM;
2549 
2550         path = alloc_path_for_send();
2551         if (!path) {
2552                 ret = -ENOMEM;
2553                 goto out;
2554         }
2555 
2556         key.objectid = ino;
2557         key.type = BTRFS_INODE_ITEM_KEY;
2558         key.offset = 0;
2559         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2560         if (ret > 0)
2561                 ret = -ENOENT;
2562         if (ret < 0)
2563                 goto out;
2564 
2565         eb = path->nodes[0];
2566         slot = path->slots[0];
2567         ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item);
2568 
2569         ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES);
2570         if (ret < 0)
2571                 goto out;
2572 
2573         ret = get_cur_path(sctx, ino, gen, p);
2574         if (ret < 0)
2575                 goto out;
2576         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2577         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime);
2578         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime);
2579         TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime);
2580         /* TODO Add otime support when the otime patches get into upstream */
2581 
2582         ret = send_cmd(sctx);
2583 
2584 tlv_put_failure:
2585 out:
2586         fs_path_free(p);
2587         btrfs_free_path(path);
2588         return ret;
2589 }
2590 
2591 /*
2592  * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have
2593  * a valid path yet because we did not process the refs yet. So, the inode
2594  * is created as orphan.
2595  */
2596 static int send_create_inode(struct send_ctx *sctx, u64 ino)
2597 {
2598         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
2599         int ret = 0;
2600         struct fs_path *p;
2601         int cmd;
2602         u64 gen;
2603         u64 mode;
2604         u64 rdev;
2605 
2606         btrfs_debug(fs_info, "send_create_inode %llu", ino);
2607 
2608         p = fs_path_alloc();
2609         if (!p)
2610                 return -ENOMEM;
2611 
2612         if (ino != sctx->cur_ino) {
2613                 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode,
2614                                      NULL, NULL, &rdev);
2615                 if (ret < 0)
2616                         goto out;
2617         } else {
2618                 gen = sctx->cur_inode_gen;
2619                 mode = sctx->cur_inode_mode;
2620                 rdev = sctx->cur_inode_rdev;
2621         }
2622 
2623         if (S_ISREG(mode)) {
2624                 cmd = BTRFS_SEND_C_MKFILE;
2625         } else if (S_ISDIR(mode)) {
2626                 cmd = BTRFS_SEND_C_MKDIR;
2627         } else if (S_ISLNK(mode)) {
2628                 cmd = BTRFS_SEND_C_SYMLINK;
2629         } else if (S_ISCHR(mode) || S_ISBLK(mode)) {
2630                 cmd = BTRFS_SEND_C_MKNOD;
2631         } else if (S_ISFIFO(mode)) {
2632                 cmd = BTRFS_SEND_C_MKFIFO;
2633         } else if (S_ISSOCK(mode)) {
2634                 cmd = BTRFS_SEND_C_MKSOCK;
2635         } else {
2636                 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o",
2637                                 (int)(mode & S_IFMT));
2638                 ret = -EOPNOTSUPP;
2639                 goto out;
2640         }
2641 
2642         ret = begin_cmd(sctx, cmd);
2643         if (ret < 0)
2644                 goto out;
2645 
2646         ret = gen_unique_name(sctx, ino, gen, p);
2647         if (ret < 0)
2648                 goto out;
2649 
2650         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
2651         TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino);
2652 
2653         if (S_ISLNK(mode)) {
2654                 fs_path_reset(p);
2655                 ret = read_symlink(sctx->send_root, ino, p);
2656                 if (ret < 0)
2657                         goto out;
2658                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p);
2659         } else if (S_ISCHR(mode) || S_ISBLK(mode) ||
2660                    S_ISFIFO(mode) || S_ISSOCK(mode)) {
2661                 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev));
2662                 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode);
2663         }
2664 
2665         ret = send_cmd(sctx);
2666         if (ret < 0)
2667                 goto out;
2668 
2669 
2670 tlv_put_failure:
2671 out:
2672         fs_path_free(p);
2673         return ret;
2674 }
2675 
2676 /*
2677  * We need some special handling for inodes that get processed before the parent
2678  * directory got created. See process_recorded_refs for details.
2679  * This function does the check if we already created the dir out of order.
2680  */
2681 static int did_create_dir(struct send_ctx *sctx, u64 dir)
2682 {
2683         int ret = 0;
2684         struct btrfs_path *path = NULL;
2685         struct btrfs_key key;
2686         struct btrfs_key found_key;
2687         struct btrfs_key di_key;
2688         struct extent_buffer *eb;
2689         struct btrfs_dir_item *di;
2690         int slot;
2691 
2692         path = alloc_path_for_send();
2693         if (!path) {
2694                 ret = -ENOMEM;
2695                 goto out;
2696         }
2697 
2698         key.objectid = dir;
2699         key.type = BTRFS_DIR_INDEX_KEY;
2700         key.offset = 0;
2701         ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0);
2702         if (ret < 0)
2703                 goto out;
2704 
2705         while (1) {
2706                 eb = path->nodes[0];
2707                 slot = path->slots[0];
2708                 if (slot >= btrfs_header_nritems(eb)) {
2709                         ret = btrfs_next_leaf(sctx->send_root, path);
2710                         if (ret < 0) {
2711                                 goto out;
2712                         } else if (ret > 0) {
2713                                 ret = 0;
2714                                 break;
2715                         }
2716                         continue;
2717                 }
2718 
2719                 btrfs_item_key_to_cpu(eb, &found_key, slot);
2720                 if (found_key.objectid != key.objectid ||
2721                     found_key.type != key.type) {
2722                         ret = 0;
2723                         goto out;
2724                 }
2725 
2726                 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item);
2727                 btrfs_dir_item_key_to_cpu(eb, di, &di_key);
2728 
2729                 if (di_key.type != BTRFS_ROOT_ITEM_KEY &&
2730                     di_key.objectid < sctx->send_progress) {
2731                         ret = 1;
2732                         goto out;
2733                 }
2734 
2735                 path->slots[0]++;
2736         }
2737 
2738 out:
2739         btrfs_free_path(path);
2740         return ret;
2741 }
2742 
2743 /*
2744  * Only creates the inode if it is:
2745  * 1. Not a directory
2746  * 2. Or a directory which was not created already due to out of order
2747  *    directories. See did_create_dir and process_recorded_refs for details.
2748  */
2749 static int send_create_inode_if_needed(struct send_ctx *sctx)
2750 {
2751         int ret;
2752 
2753         if (S_ISDIR(sctx->cur_inode_mode)) {
2754                 ret = did_create_dir(sctx, sctx->cur_ino);
2755                 if (ret < 0)
2756                         goto out;
2757                 if (ret) {
2758                         ret = 0;
2759                         goto out;
2760                 }
2761         }
2762 
2763         ret = send_create_inode(sctx, sctx->cur_ino);
2764         if (ret < 0)
2765                 goto out;
2766 
2767 out:
2768         return ret;
2769 }
2770 
2771 struct recorded_ref {
2772         struct list_head list;
2773         char *name;
2774         struct fs_path *full_path;
2775         u64 dir;
2776         u64 dir_gen;
2777         int name_len;
2778 };
2779 
2780 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path)
2781 {
2782         ref->full_path = path;
2783         ref->name = (char *)kbasename(ref->full_path->start);
2784         ref->name_len = ref->full_path->end - ref->name;
2785 }
2786 
2787 /*
2788  * We need to process new refs before deleted refs, but compare_tree gives us
2789  * everything mixed. So we first record all refs and later process them.
2790  * This function is a helper to record one ref.
2791  */
2792 static int __record_ref(struct list_head *head, u64 dir,
2793                       u64 dir_gen, struct fs_path *path)
2794 {
2795         struct recorded_ref *ref;
2796 
2797         ref = kmalloc(sizeof(*ref), GFP_KERNEL);
2798         if (!ref)
2799                 return -ENOMEM;
2800 
2801         ref->dir = dir;
2802         ref->dir_gen = dir_gen;
2803         set_ref_path(ref, path);
2804         list_add_tail(&ref->list, head);
2805         return 0;
2806 }
2807 
2808 static int dup_ref(struct recorded_ref *ref, struct list_head *list)
2809 {
2810         struct recorded_ref *new;
2811 
2812         new = kmalloc(sizeof(*ref), GFP_KERNEL);
2813         if (!new)
2814                 return -ENOMEM;
2815 
2816         new->dir = ref->dir;
2817         new->dir_gen = ref->dir_gen;
2818         new->full_path = NULL;
2819         INIT_LIST_HEAD(&new->list);
2820         list_add_tail(&new->list, list);
2821         return 0;
2822 }
2823 
2824 static void __free_recorded_refs(struct list_head *head)
2825 {
2826         struct recorded_ref *cur;
2827 
2828         while (!list_empty(head)) {
2829                 cur = list_entry(head->next, struct recorded_ref, list);
2830                 fs_path_free(cur->full_path);
2831                 list_del(&cur->list);
2832                 kfree(cur);
2833         }
2834 }
2835 
2836 static void free_recorded_refs(struct send_ctx *sctx)
2837 {
2838         __free_recorded_refs(&sctx->new_refs);
2839         __free_recorded_refs(&sctx->deleted_refs);
2840 }
2841 
2842 /*
2843  * Renames/moves a file/dir to its orphan name. Used when the first
2844  * ref of an unprocessed inode gets overwritten and for all non empty
2845  * directories.
2846  */
2847 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen,
2848                           struct fs_path *path)
2849 {
2850         int ret;
2851         struct fs_path *orphan;
2852 
2853         orphan = fs_path_alloc();
2854         if (!orphan)
2855                 return -ENOMEM;
2856 
2857         ret = gen_unique_name(sctx, ino, gen, orphan);
2858         if (ret < 0)
2859                 goto out;
2860 
2861         ret = send_rename(sctx, path, orphan);
2862 
2863 out:
2864         fs_path_free(orphan);
2865         return ret;
2866 }
2867 
2868 static struct orphan_dir_info *
2869 add_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2870 {
2871         struct rb_node **p = &sctx->orphan_dirs.rb_node;
2872         struct rb_node *parent = NULL;
2873         struct orphan_dir_info *entry, *odi;
2874 
2875         while (*p) {
2876                 parent = *p;
2877                 entry = rb_entry(parent, struct orphan_dir_info, node);
2878                 if (dir_ino < entry->ino) {
2879                         p = &(*p)->rb_left;
2880                 } else if (dir_ino > entry->ino) {
2881                         p = &(*p)->rb_right;
2882                 } else {
2883                         return entry;
2884                 }
2885         }
2886 
2887         odi = kmalloc(sizeof(*odi), GFP_KERNEL);
2888         if (!odi)
2889                 return ERR_PTR(-ENOMEM);
2890         odi->ino = dir_ino;
2891         odi->gen = 0;
2892         odi->last_dir_index_offset = 0;
2893 
2894         rb_link_node(&odi->node, parent, p);
2895         rb_insert_color(&odi->node, &sctx->orphan_dirs);
2896         return odi;
2897 }
2898 
2899 static struct orphan_dir_info *
2900 get_orphan_dir_info(struct send_ctx *sctx, u64 dir_ino)
2901 {
2902         struct rb_node *n = sctx->orphan_dirs.rb_node;
2903         struct orphan_dir_info *entry;
2904 
2905         while (n) {
2906                 entry = rb_entry(n, struct orphan_dir_info, node);
2907                 if (dir_ino < entry->ino)
2908                         n = n->rb_left;
2909                 else if (dir_ino > entry->ino)
2910                         n = n->rb_right;
2911                 else
2912                         return entry;
2913         }
2914         return NULL;
2915 }
2916 
2917 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino)
2918 {
2919         struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino);
2920 
2921         return odi != NULL;
2922 }
2923 
2924 static void free_orphan_dir_info(struct send_ctx *sctx,
2925                                  struct orphan_dir_info *odi)
2926 {
2927         if (!odi)
2928                 return;
2929         rb_erase(&odi->node, &sctx->orphan_dirs);
2930         kfree(odi);
2931 }
2932 
2933 /*
2934  * Returns 1 if a directory can be removed at this point in time.
2935  * We check this by iterating all dir items and checking if the inode behind
2936  * the dir item was already processed.
2937  */
2938 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen,
2939                      u64 send_progress)
2940 {
2941         int ret = 0;
2942         struct btrfs_root *root = sctx->parent_root;
2943         struct btrfs_path *path;
2944         struct btrfs_key key;
2945         struct btrfs_key found_key;
2946         struct btrfs_key loc;
2947         struct btrfs_dir_item *di;
2948         struct orphan_dir_info *odi = NULL;
2949 
2950         /*
2951          * Don't try to rmdir the top/root subvolume dir.
2952          */
2953         if (dir == BTRFS_FIRST_FREE_OBJECTID)
2954                 return 0;
2955 
2956         path = alloc_path_for_send();
2957         if (!path)
2958                 return -ENOMEM;
2959 
2960         key.objectid = dir;
2961         key.type = BTRFS_DIR_INDEX_KEY;
2962         key.offset = 0;
2963 
2964         odi = get_orphan_dir_info(sctx, dir);
2965         if (odi)
2966                 key.offset = odi->last_dir_index_offset;
2967 
2968         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2969         if (ret < 0)
2970                 goto out;
2971 
2972         while (1) {
2973                 struct waiting_dir_move *dm;
2974 
2975                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2976                         ret = btrfs_next_leaf(root, path);
2977                         if (ret < 0)
2978                                 goto out;
2979                         else if (ret > 0)
2980                                 break;
2981                         continue;
2982                 }
2983                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2984                                       path->slots[0]);
2985                 if (found_key.objectid != key.objectid ||
2986                     found_key.type != key.type)
2987                         break;
2988 
2989                 di = btrfs_item_ptr(path->nodes[0], path->slots[0],
2990                                 struct btrfs_dir_item);
2991                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc);
2992 
2993                 dm = get_waiting_dir_move(sctx, loc.objectid);
2994                 if (dm) {
2995                         odi = add_orphan_dir_info(sctx, dir);
2996                         if (IS_ERR(odi)) {
2997                                 ret = PTR_ERR(odi);
2998                                 goto out;
2999                         }
3000                         odi->gen = dir_gen;
3001                         odi->last_dir_index_offset = found_key.offset;
3002                         dm->rmdir_ino = dir;
3003                         ret = 0;
3004                         goto out;
3005                 }
3006 
3007                 if (loc.objectid > send_progress) {
3008                         odi = add_orphan_dir_info(sctx, dir);
3009                         if (IS_ERR(odi)) {
3010                                 ret = PTR_ERR(odi);
3011                                 goto out;
3012                         }
3013                         odi->gen = dir_gen;
3014                         odi->last_dir_index_offset = found_key.offset;
3015                         ret = 0;
3016                         goto out;
3017                 }
3018 
3019                 path->slots[0]++;
3020         }
3021         free_orphan_dir_info(sctx, odi);
3022 
3023         ret = 1;
3024 
3025 out:
3026         btrfs_free_path(path);
3027         return ret;
3028 }
3029 
3030 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino)
3031 {
3032         struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino);
3033 
3034         return entry != NULL;
3035 }
3036 
3037 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized)
3038 {
3039         struct rb_node **p = &sctx->waiting_dir_moves.rb_node;
3040         struct rb_node *parent = NULL;
3041         struct waiting_dir_move *entry, *dm;
3042 
3043         dm = kmalloc(sizeof(*dm), GFP_KERNEL);
3044         if (!dm)
3045                 return -ENOMEM;
3046         dm->ino = ino;
3047         dm->rmdir_ino = 0;
3048         dm->orphanized = orphanized;
3049 
3050         while (*p) {
3051                 parent = *p;
3052                 entry = rb_entry(parent, struct waiting_dir_move, node);
3053                 if (ino < entry->ino) {
3054                         p = &(*p)->rb_left;
3055                 } else if (ino > entry->ino) {
3056                         p = &(*p)->rb_right;
3057                 } else {
3058                         kfree(dm);
3059                         return -EEXIST;
3060                 }
3061         }
3062 
3063         rb_link_node(&dm->node, parent, p);
3064         rb_insert_color(&dm->node, &sctx->waiting_dir_moves);
3065         return 0;
3066 }
3067 
3068 static struct waiting_dir_move *
3069 get_waiting_dir_move(struct send_ctx *sctx, u64 ino)
3070 {
3071         struct rb_node *n = sctx->waiting_dir_moves.rb_node;
3072         struct waiting_dir_move *entry;
3073 
3074         while (n) {
3075                 entry = rb_entry(n, struct waiting_dir_move, node);
3076                 if (ino < entry->ino)
3077                         n = n->rb_left;
3078                 else if (ino > entry->ino)
3079                         n = n->rb_right;
3080                 else
3081                         return entry;
3082         }
3083         return NULL;
3084 }
3085 
3086 static void free_waiting_dir_move(struct send_ctx *sctx,
3087                                   struct waiting_dir_move *dm)
3088 {
3089         if (!dm)
3090                 return;
3091         rb_erase(&dm->node, &sctx->waiting_dir_moves);
3092         kfree(dm);
3093 }
3094 
3095 static int add_pending_dir_move(struct send_ctx *sctx,
3096                                 u64 ino,
3097                                 u64 ino_gen,
3098                                 u64 parent_ino,
3099                                 struct list_head *new_refs,
3100                                 struct list_head *deleted_refs,
3101                                 const bool is_orphan)
3102 {
3103         struct rb_node **p = &sctx->pending_dir_moves.rb_node;
3104         struct rb_node *parent = NULL;
3105         struct pending_dir_move *entry = NULL, *pm;
3106         struct recorded_ref *cur;
3107         int exists = 0;
3108         int ret;
3109 
3110         pm = kmalloc(sizeof(*pm), GFP_KERNEL);
3111         if (!pm)
3112                 return -ENOMEM;
3113         pm->parent_ino = parent_ino;
3114         pm->ino = ino;
3115         pm->gen = ino_gen;
3116         INIT_LIST_HEAD(&pm->list);
3117         INIT_LIST_HEAD(&pm->update_refs);
3118         RB_CLEAR_NODE(&pm->node);
3119 
3120         while (*p) {
3121                 parent = *p;
3122                 entry = rb_entry(parent, struct pending_dir_move, node);
3123                 if (parent_ino < entry->parent_ino) {
3124                         p = &(*p)->rb_left;
3125                 } else if (parent_ino > entry->parent_ino) {
3126                         p = &(*p)->rb_right;
3127                 } else {
3128                         exists = 1;
3129                         break;
3130                 }
3131         }
3132 
3133         list_for_each_entry(cur, deleted_refs, list) {
3134                 ret = dup_ref(cur, &pm->update_refs);
3135                 if (ret < 0)
3136                         goto out;
3137         }
3138         list_for_each_entry(cur, new_refs, list) {
3139                 ret = dup_ref(cur, &pm->update_refs);
3140                 if (ret < 0)
3141                         goto out;
3142         }
3143 
3144         ret = add_waiting_dir_move(sctx, pm->ino, is_orphan);
3145         if (ret)
3146                 goto out;
3147 
3148         if (exists) {
3149                 list_add_tail(&pm->list, &entry->list);
3150         } else {
3151                 rb_link_node(&pm->node, parent, p);
3152                 rb_insert_color(&pm->node, &sctx->pending_dir_moves);
3153         }
3154         ret = 0;
3155 out:
3156         if (ret) {
3157                 __free_recorded_refs(&pm->update_refs);
3158                 kfree(pm);
3159         }
3160         return ret;
3161 }
3162 
3163 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx,
3164                                                       u64 parent_ino)
3165 {
3166         struct rb_node *n = sctx->pending_dir_moves.rb_node;
3167         struct pending_dir_move *entry;
3168 
3169         while (n) {
3170                 entry = rb_entry(n, struct pending_dir_move, node);
3171                 if (parent_ino < entry->parent_ino)
3172                         n = n->rb_left;
3173                 else if (parent_ino > entry->parent_ino)
3174                         n = n->rb_right;
3175                 else
3176                         return entry;
3177         }
3178         return NULL;
3179 }
3180 
3181 static int path_loop(struct send_ctx *sctx, struct fs_path *name,
3182                      u64 ino, u64 gen, u64 *ancestor_ino)
3183 {
3184         int ret = 0;
3185         u64 parent_inode = 0;
3186         u64 parent_gen = 0;
3187         u64 start_ino = ino;
3188 
3189         *ancestor_ino = 0;
3190         while (ino != BTRFS_FIRST_FREE_OBJECTID) {
3191                 fs_path_reset(name);
3192 
3193                 if (is_waiting_for_rm(sctx, ino))
3194                         break;
3195                 if (is_waiting_for_move(sctx, ino)) {
3196                         if (*ancestor_ino == 0)
3197                                 *ancestor_ino = ino;
3198                         ret = get_first_ref(sctx->parent_root, ino,
3199                                             &parent_inode, &parent_gen, name);
3200                 } else {
3201                         ret = __get_cur_name_and_parent(sctx, ino, gen,
3202                                                         &parent_inode,
3203                                                         &parent_gen, name);
3204                         if (ret > 0) {
3205                                 ret = 0;
3206                                 break;
3207                         }
3208                 }
3209                 if (ret < 0)
3210                         break;
3211                 if (parent_inode == start_ino) {
3212                         ret = 1;
3213                         if (*ancestor_ino == 0)
3214                                 *ancestor_ino = ino;
3215                         break;
3216                 }
3217                 ino = parent_inode;
3218                 gen = parent_gen;
3219         }
3220         return ret;
3221 }
3222 
3223 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm)
3224 {
3225         struct fs_path *from_path = NULL;
3226         struct fs_path *to_path = NULL;
3227         struct fs_path *name = NULL;
3228         u64 orig_progress = sctx->send_progress;
3229         struct recorded_ref *cur;
3230         u64 parent_ino, parent_gen;
3231         struct waiting_dir_move *dm = NULL;
3232         u64 rmdir_ino = 0;
3233         u64 ancestor;
3234         bool is_orphan;
3235         int ret;
3236 
3237         name = fs_path_alloc();
3238         from_path = fs_path_alloc();
3239         if (!name || !from_path) {
3240                 ret = -ENOMEM;
3241                 goto out;
3242         }
3243 
3244         dm = get_waiting_dir_move(sctx, pm->ino);
3245         ASSERT(dm);
3246         rmdir_ino = dm->rmdir_ino;
3247         is_orphan = dm->orphanized;
3248         free_waiting_dir_move(sctx, dm);
3249 
3250         if (is_orphan) {
3251                 ret = gen_unique_name(sctx, pm->ino,
3252                                       pm->gen, from_path);
3253         } else {
3254                 ret = get_first_ref(sctx->parent_root, pm->ino,
3255                                     &parent_ino, &parent_gen, name);
3256                 if (ret < 0)
3257                         goto out;
3258                 ret = get_cur_path(sctx, parent_ino, parent_gen,
3259                                    from_path);
3260                 if (ret < 0)
3261                         goto out;
3262                 ret = fs_path_add_path(from_path, name);
3263         }
3264         if (ret < 0)
3265                 goto out;
3266 
3267         sctx->send_progress = sctx->cur_ino + 1;
3268         ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor);
3269         if (ret < 0)
3270                 goto out;
3271         if (ret) {
3272                 LIST_HEAD(deleted_refs);
3273                 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID);
3274                 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor,
3275                                            &pm->update_refs, &deleted_refs,
3276                                            is_orphan);
3277                 if (ret < 0)
3278                         goto out;
3279                 if (rmdir_ino) {
3280                         dm = get_waiting_dir_move(sctx, pm->ino);
3281                         ASSERT(dm);
3282                         dm->rmdir_ino = rmdir_ino;
3283                 }
3284                 goto out;
3285         }
3286         fs_path_reset(name);
3287         to_path = name;
3288         name = NULL;
3289         ret = get_cur_path(sctx, pm->ino, pm->gen, to_path);
3290         if (ret < 0)
3291                 goto out;
3292 
3293         ret = send_rename(sctx, from_path, to_path);
3294         if (ret < 0)
3295                 goto out;
3296 
3297         if (rmdir_ino) {
3298                 struct orphan_dir_info *odi;
3299                 u64 gen;
3300 
3301                 odi = get_orphan_dir_info(sctx, rmdir_ino);
3302                 if (!odi) {
3303                         /* already deleted */
3304                         goto finish;
3305                 }
3306                 gen = odi->gen;
3307 
3308                 ret = can_rmdir(sctx, rmdir_ino, gen, sctx->cur_ino);
3309                 if (ret < 0)
3310                         goto out;
3311                 if (!ret)
3312                         goto finish;
3313 
3314                 name = fs_path_alloc();
3315                 if (!name) {
3316                         ret = -ENOMEM;
3317                         goto out;
3318                 }
3319                 ret = get_cur_path(sctx, rmdir_ino, gen, name);
3320                 if (ret < 0)
3321                         goto out;
3322                 ret = send_rmdir(sctx, name);
3323                 if (ret < 0)
3324                         goto out;
3325         }
3326 
3327 finish:
3328         ret = send_utimes(sctx, pm->ino, pm->gen);
3329         if (ret < 0)
3330                 goto out;
3331 
3332         /*
3333          * After rename/move, need to update the utimes of both new parent(s)
3334          * and old parent(s).
3335          */
3336         list_for_each_entry(cur, &pm->update_refs, list) {
3337                 /*
3338                  * The parent inode might have been deleted in the send snapshot
3339                  */
3340                 ret = get_inode_info(sctx->send_root, cur->dir, NULL,
3341                                      NULL, NULL, NULL, NULL, NULL);
3342                 if (ret == -ENOENT) {
3343                         ret = 0;
3344                         continue;
3345                 }
3346                 if (ret < 0)
3347                         goto out;
3348 
3349                 ret = send_utimes(sctx, cur->dir, cur->dir_gen);
3350                 if (ret < 0)
3351                         goto out;
3352         }
3353 
3354 out:
3355         fs_path_free(name);
3356         fs_path_free(from_path);
3357         fs_path_free(to_path);
3358         sctx->send_progress = orig_progress;
3359 
3360         return ret;
3361 }
3362 
3363 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m)
3364 {
3365         if (!list_empty(&m->list))
3366                 list_del(&m->list);
3367         if (!RB_EMPTY_NODE(&m->node))
3368                 rb_erase(&m->node, &sctx->pending_dir_moves);
3369         __free_recorded_refs(&m->update_refs);
3370         kfree(m);
3371 }
3372 
3373 static void tail_append_pending_moves(struct send_ctx *sctx,
3374                                       struct pending_dir_move *moves,
3375                                       struct list_head *stack)
3376 {
3377         if (list_empty(&moves->list)) {
3378                 list_add_tail(&moves->list, stack);
3379         } else {
3380                 LIST_HEAD(list);
3381                 list_splice_init(&moves->list, &list);
3382                 list_add_tail(&moves->list, stack);
3383                 list_splice_tail(&list, stack);
3384         }
3385         if (!RB_EMPTY_NODE(&moves->node)) {
3386                 rb_erase(&moves->node, &sctx->pending_dir_moves);
3387                 RB_CLEAR_NODE(&moves->node);
3388         }
3389 }
3390 
3391 static int apply_children_dir_moves(struct send_ctx *sctx)
3392 {
3393         struct pending_dir_move *pm;
3394         struct list_head stack;
3395         u64 parent_ino = sctx->cur_ino;
3396         int ret = 0;
3397 
3398         pm = get_pending_dir_moves(sctx, parent_ino);
3399         if (!pm)
3400                 return 0;
3401 
3402         INIT_LIST_HEAD(&stack);
3403         tail_append_pending_moves(sctx, pm, &stack);
3404 
3405         while (!list_empty(&stack)) {
3406                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3407                 parent_ino = pm->ino;
3408                 ret = apply_dir_move(sctx, pm);
3409                 free_pending_move(sctx, pm);
3410                 if (ret)
3411                         goto out;
3412                 pm = get_pending_dir_moves(sctx, parent_ino);
3413                 if (pm)
3414                         tail_append_pending_moves(sctx, pm, &stack);
3415         }
3416         return 0;
3417 
3418 out:
3419         while (!list_empty(&stack)) {
3420                 pm = list_first_entry(&stack, struct pending_dir_move, list);
3421                 free_pending_move(sctx, pm);
3422         }
3423         return ret;
3424 }
3425 
3426 /*
3427  * We might need to delay a directory rename even when no ancestor directory
3428  * (in the send root) with a higher inode number than ours (sctx->cur_ino) was
3429  * renamed. This happens when we rename a directory to the old name (the name
3430  * in the parent root) of some other unrelated directory that got its rename
3431  * delayed due to some ancestor with higher number that got renamed.
3432  *
3433  * Example:
3434  *
3435  * Parent snapshot:
3436  * .                                       (ino 256)
3437  * |---- a/                                (ino 257)
3438  * |     |---- file                        (ino 260)
3439  * |
3440  * |---- b/                                (ino 258)
3441  * |---- c/                                (ino 259)
3442  *
3443  * Send snapshot:
3444  * .                                       (ino 256)
3445  * |---- a/                                (ino 258)
3446  * |---- x/                                (ino 259)
3447  *       |---- y/                          (ino 257)
3448  *             |----- file                 (ino 260)
3449  *
3450  * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257
3451  * from 'a' to 'x/y' happening first, which in turn depends on the rename of
3452  * inode 259 from 'c' to 'x'. So the order of rename commands the send stream
3453  * must issue is:
3454  *
3455  * 1 - rename 259 from 'c' to 'x'
3456  * 2 - rename 257 from 'a' to 'x/y'
3457  * 3 - rename 258 from 'b' to 'a'
3458  *
3459  * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can
3460  * be done right away and < 0 on error.
3461  */
3462 static int wait_for_dest_dir_move(struct send_ctx *sctx,
3463                                   struct recorded_ref *parent_ref,
3464                                   const bool is_orphan)
3465 {
3466         struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info;
3467         struct btrfs_path *path;
3468         struct btrfs_key key;
3469         struct btrfs_key di_key;
3470         struct btrfs_dir_item *di;
3471         u64 left_gen;
3472         u64 right_gen;
3473         int ret = 0;
3474         struct waiting_dir_move *wdm;
3475 
3476         if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves))
3477                 return 0;
3478 
3479         path = alloc_path_for_send();
3480         if (!path)
3481                 return -ENOMEM;
3482 
3483         key.objectid = parent_ref->dir;
3484         key.type = BTRFS_DIR_ITEM_KEY;
3485         key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len);
3486 
3487         ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0);
3488         if (ret < 0) {
3489                 goto out;
3490         } else if (ret > 0) {
3491                 ret = 0;
3492                 goto out;
3493         }
3494 
3495         di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name,
3496                                        parent_ref->name_len);
3497         if (!di) {
3498                 ret = 0;
3499                 goto out;
3500         }
3501         /*
3502          * di_key.objectid has the number of the inode that has a dentry in the
3503          * parent directory with the same name that sctx->cur_ino is being
3504          * renamed to. We need to check if that inode is in the send root as
3505          * well and if it is currently marked as an inode with a pending rename,
3506          * if it is, we need to delay the rename of sctx->cur_ino as well, so
3507          * that it happens after that other inode is renamed.
3508          */
3509         btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key);
3510         if (di_key.type != BTRFS_INODE_ITEM_KEY) {
3511                 ret = 0;
3512                 goto out;
3513         }
3514 
3515         ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL,
3516                              &left_gen, NULL, NULL, NULL, NULL);
3517         if (ret < 0)
3518                 goto out;
3519         ret = get_inode_info(sctx->send_root, di_key.objectid, NULL,
3520                              &right_gen, NULL, NULL, NULL, NULL);
3521         if (ret < 0) {
3522                 if (ret == -ENOENT)
3523                         ret = 0;
3524                 goto out;
3525         }
3526 
3527         /* Different inode, no need to delay the rename of sctx->cur_ino */
3528         if (right_gen != left_gen) {
3529                 ret = 0;
3530                 goto out;
3531         }
3532 
3533         wdm = get_waiting_dir_move(sctx, di_key.objectid);
3534         if (wdm && !wdm->orphanized) {
3535                 ret = add_pending_dir_move(sctx,
3536                                            sctx->cur_ino,
3537                                            sctx->cur_inode_gen,
3538                                            di_key.objectid,
3539                                            &sctx->new_refs,
3540                                            &sctx->deleted_refs,
3541                                            is_orphan);
3542                 if (!ret)
3543                         ret = 1;
3544         }
3545 out:
3546         btrfs_free_path(path);
3547         return ret;
3548 }
3549 
3550 /*
3551  * Check if inode ino2, or any of its ancestors, is inode ino1.
3552  * Return 1 if true, 0 if false and < 0 on error.
3553  */
3554 static int check_ino_in_path(struct btrfs_root *root,
3555                              const u64 ino1,
3556                              const u64 ino1_gen,
3557                              const u64 ino2,
3558                              const u64 ino2_gen,
3559                              struct fs_path *fs_path)
3560 {
3561         u64 ino = ino2;
3562 
3563         if (ino1 == ino2)
3564                 return ino1_gen == ino2_gen;
3565 
3566         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3567                 u64 parent;
3568                 u64 parent_gen;
3569                 int ret;
3570 
3571                 fs_path_reset(fs_path);
3572                 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path);
3573                 if (ret < 0)
3574                         return ret;
3575                 if (parent == ino1)
3576                         return parent_gen == ino1_gen;
3577                 ino = parent;
3578         }
3579         return 0;
3580 }
3581 
3582 /*
3583  * Check if ino ino1 is an ancestor of inode ino2 in the given root for any
3584  * possible path (in case ino2 is not a directory and has multiple hard links).
3585  * Return 1 if true, 0 if false and < 0 on error.
3586  */
3587 static int is_ancestor(struct btrfs_root *root,
3588                        const u64 ino1,
3589                        const u64 ino1_gen,
3590                        const u64 ino2,
3591                        struct fs_path *fs_path)
3592 {
3593         bool free_fs_path = false;
3594         int ret = 0;
3595         struct btrfs_path *path = NULL;
3596         struct btrfs_key key;
3597 
3598         if (!fs_path) {
3599                 fs_path = fs_path_alloc();
3600                 if (!fs_path)
3601                         return -ENOMEM;
3602                 free_fs_path = true;
3603         }
3604 
3605         path = alloc_path_for_send();
3606         if (!path) {
3607                 ret = -ENOMEM;
3608                 goto out;
3609         }
3610 
3611         key.objectid = ino2;
3612         key.type = BTRFS_INODE_REF_KEY;
3613         key.offset = 0;
3614 
3615         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3616         if (ret < 0)
3617                 goto out;
3618 
3619         while (true) {
3620                 struct extent_buffer *leaf = path->nodes[0];
3621                 int slot = path->slots[0];
3622                 u32 cur_offset = 0;
3623                 u32 item_size;
3624 
3625                 if (slot >= btrfs_header_nritems(leaf)) {
3626                         ret = btrfs_next_leaf(root, path);
3627                         if (ret < 0)
3628                                 goto out;
3629                         if (ret > 0)
3630                                 break;
3631                         continue;
3632                 }
3633 
3634                 btrfs_item_key_to_cpu(leaf, &key, slot);
3635                 if (key.objectid != ino2)
3636                         break;
3637                 if (key.type != BTRFS_INODE_REF_KEY &&
3638                     key.type != BTRFS_INODE_EXTREF_KEY)
3639                         break;
3640 
3641                 item_size = btrfs_item_size_nr(leaf, slot);
3642                 while (cur_offset < item_size) {
3643                         u64 parent;
3644                         u64 parent_gen;
3645 
3646                         if (key.type == BTRFS_INODE_EXTREF_KEY) {
3647                                 unsigned long ptr;
3648                                 struct btrfs_inode_extref *extref;
3649 
3650                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3651                                 extref = (struct btrfs_inode_extref *)
3652                                         (ptr + cur_offset);
3653                                 parent = btrfs_inode_extref_parent(leaf,
3654                                                                    extref);
3655                                 cur_offset += sizeof(*extref);
3656                                 cur_offset += btrfs_inode_extref_name_len(leaf,
3657                                                                   extref);
3658                         } else {
3659                                 parent = key.offset;
3660                                 cur_offset = item_size;
3661                         }
3662 
3663                         ret = get_inode_info(root, parent, NULL, &parent_gen,
3664                                              NULL, NULL, NULL, NULL);
3665                         if (ret < 0)
3666                                 goto out;
3667                         ret = check_ino_in_path(root, ino1, ino1_gen,
3668                                                 parent, parent_gen, fs_path);
3669                         if (ret)
3670                                 goto out;
3671                 }
3672                 path->slots[0]++;
3673         }
3674         ret = 0;
3675  out:
3676         btrfs_free_path(path);
3677         if (free_fs_path)
3678                 fs_path_free(fs_path);
3679         return ret;
3680 }
3681 
3682 static int wait_for_parent_move(struct send_ctx *sctx,
3683                                 struct recorded_ref *parent_ref,
3684                                 const bool is_orphan)
3685 {
3686         int ret = 0;
3687         u64 ino = parent_ref->dir;
3688         u64 ino_gen = parent_ref->dir_gen;
3689         u64 parent_ino_before, parent_ino_after;
3690         struct fs_path *path_before = NULL;
3691         struct fs_path *path_after = NULL;
3692         int len1, len2;
3693 
3694         path_after = fs_path_alloc();
3695         path_before = fs_path_alloc();
3696         if (!path_after || !path_before) {
3697                 ret = -ENOMEM;
3698                 goto out;
3699         }
3700 
3701         /*
3702          * Our current directory inode may not yet be renamed/moved because some
3703          * ancestor (immediate or not) has to be renamed/moved first. So find if
3704          * such ancestor exists and make sure our own rename/move happens after
3705          * that ancestor is processed to avoid path build infinite loops (done
3706          * at get_cur_path()).
3707          */
3708         while (ino > BTRFS_FIRST_FREE_OBJECTID) {
3709                 u64 parent_ino_after_gen;
3710 
3711                 if (is_waiting_for_move(sctx, ino)) {
3712                         /*
3713                          * If the current inode is an ancestor of ino in the
3714                          * parent root, we need to delay the rename of the
3715                          * current inode, otherwise don't delayed the rename
3716                          * because we can end up with a circular dependency
3717                          * of renames, resulting in some directories never
3718                          * getting the respective rename operations issued in
3719                          * the send stream or getting into infinite path build
3720                          * loops.
3721                          */
3722                         ret = is_ancestor(sctx->parent_root,
3723                                           sctx->cur_ino, sctx->cur_inode_gen,
3724                                           ino, path_before);
3725                         if (ret)
3726                                 break;
3727                 }
3728 
3729                 fs_path_reset(path_before);
3730                 fs_path_reset(path_after);
3731 
3732                 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after,
3733                                     &parent_ino_after_gen, path_after);
3734                 if (ret < 0)
3735                         goto out;
3736                 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before,
3737                                     NULL, path_before);
3738                 if (ret < 0 && ret != -ENOENT) {
3739                         goto out;
3740                 } else if (ret == -ENOENT) {
3741                         ret = 0;
3742                         break;
3743                 }
3744 
3745                 len1 = fs_path_len(path_before);
3746                 len2 = fs_path_len(path_after);
3747                 if (ino > sctx->cur_ino &&
3748                     (parent_ino_before != parent_ino_after || len1 != len2 ||
3749                      memcmp(path_before->start, path_after->start, len1))) {
3750                         u64 parent_ino_gen;
3751 
3752                         ret = get_inode_info(sctx->parent_root, ino, NULL,
3753                                              &parent_ino_gen, NULL, NULL, NULL,
3754                                              NULL);
3755                         if (ret < 0)
3756                                 goto out;
3757                         if (ino_gen == parent_ino_gen) {
3758                                 ret = 1;
3759                                 break;
3760                         }
3761                 }
3762                 ino = parent_ino_after;
3763                 ino_gen = parent_ino_after_gen;
3764         }
3765 
3766 out:
3767         fs_path_free(path_before);
3768         fs_path_free(path_after);
3769 
3770         if (ret == 1) {
3771                 ret = add_pending_dir_move(sctx,
3772                                            sctx->cur_ino,
3773                                            sctx->cur_inode_gen,
3774                                            ino,
3775                                            &sctx->new_refs,
3776                                            &sctx->deleted_refs,
3777                                            is_orphan);
3778                 if (!ret)
3779                         ret = 1;
3780         }
3781 
3782         return ret;
3783 }
3784 
3785 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref)
3786 {
3787         int ret;
3788         struct fs_path *new_path;
3789 
3790         /*
3791          * Our reference's name member points to its full_path member string, so
3792          * we use here a new path.
3793          */
3794         new_path = fs_path_alloc();
3795         if (!new_path)
3796                 return -ENOMEM;
3797 
3798         ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path);
3799         if (ret < 0) {
3800                 fs_path_free(new_path);
3801                 return ret;
3802         }
3803         ret = fs_path_add(new_path, ref->name, ref->name_len);
3804         if (ret < 0) {
3805                 fs_path_free(new_path);
3806                 return ret;
3807         }
3808 
3809         fs_path_free(ref->full_path);
3810         set_ref_path(ref, new_path);
3811 
3812         return 0;
3813 }
3814 
3815 /*
3816  * This does all the move/link/unlink/rmdir magic.
3817  */
3818 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move)
3819 {
3820         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
3821         int ret = 0;
3822         struct recorded_ref *cur;
3823         struct recorded_ref *cur2;
3824         struct list_head check_dirs;
3825         struct fs_path *valid_path = NULL;
3826         u64 ow_inode = 0;
3827         u64 ow_gen;
3828         u64 ow_mode;
3829         int did_overwrite = 0;
3830         int is_orphan = 0;
3831         u64 last_dir_ino_rm = 0;
3832         bool can_rename = true;
3833         bool orphanized_dir = false;
3834         bool orphanized_ancestor = false;
3835 
3836         btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino);
3837 
3838         /*
3839          * This should never happen as the root dir always has the same ref
3840          * which is always '..'
3841          */
3842         BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID);
3843         INIT_LIST_HEAD(&check_dirs);
3844 
3845         valid_path = fs_path_alloc();
3846         if (!valid_path) {
3847                 ret = -ENOMEM;
3848                 goto out;
3849         }
3850 
3851         /*
3852          * First, check if the first ref of the current inode was overwritten
3853          * before. If yes, we know that the current inode was already orphanized
3854          * and thus use the orphan name. If not, we can use get_cur_path to
3855          * get the path of the first ref as it would like while receiving at
3856          * this point in time.
3857          * New inodes are always orphan at the beginning, so force to use the
3858          * orphan name in this case.
3859          * The first ref is stored in valid_path and will be updated if it
3860          * gets moved around.
3861          */
3862         if (!sctx->cur_inode_new) {
3863                 ret = did_overwrite_first_ref(sctx, sctx->cur_ino,
3864                                 sctx->cur_inode_gen);
3865                 if (ret < 0)
3866                         goto out;
3867                 if (ret)
3868                         did_overwrite = 1;
3869         }
3870         if (sctx->cur_inode_new || did_overwrite) {
3871                 ret = gen_unique_name(sctx, sctx->cur_ino,
3872                                 sctx->cur_inode_gen, valid_path);
3873                 if (ret < 0)
3874                         goto out;
3875                 is_orphan = 1;
3876         } else {
3877                 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen,
3878                                 valid_path);
3879                 if (ret < 0)
3880                         goto out;
3881         }
3882 
3883         list_for_each_entry(cur, &sctx->new_refs, list) {
3884                 /*
3885                  * We may have refs where the parent directory does not exist
3886                  * yet. This happens if the parent directories inum is higher
3887                  * than the current inum. To handle this case, we create the
3888                  * parent directory out of order. But we need to check if this
3889                  * did already happen before due to other refs in the same dir.
3890                  */
3891                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
3892                 if (ret < 0)
3893                         goto out;
3894                 if (ret == inode_state_will_create) {
3895                         ret = 0;
3896                         /*
3897                          * First check if any of the current inodes refs did
3898                          * already create the dir.
3899                          */
3900                         list_for_each_entry(cur2, &sctx->new_refs, list) {
3901                                 if (cur == cur2)
3902                                         break;
3903                                 if (cur2->dir == cur->dir) {
3904                                         ret = 1;
3905                                         break;
3906                                 }
3907                         }
3908 
3909                         /*
3910                          * If that did not happen, check if a previous inode
3911                          * did already create the dir.
3912                          */
3913                         if (!ret)
3914                                 ret = did_create_dir(sctx, cur->dir);
3915                         if (ret < 0)
3916                                 goto out;
3917                         if (!ret) {
3918                                 ret = send_create_inode(sctx, cur->dir);
3919                                 if (ret < 0)
3920                                         goto out;
3921                         }
3922                 }
3923 
3924                 /*
3925                  * Check if this new ref would overwrite the first ref of
3926                  * another unprocessed inode. If yes, orphanize the
3927                  * overwritten inode. If we find an overwritten ref that is
3928                  * not the first ref, simply unlink it.
3929                  */
3930                 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen,
3931                                 cur->name, cur->name_len,
3932                                 &ow_inode, &ow_gen, &ow_mode);
3933                 if (ret < 0)
3934                         goto out;
3935                 if (ret) {
3936                         ret = is_first_ref(sctx->parent_root,
3937                                            ow_inode, cur->dir, cur->name,
3938                                            cur->name_len);
3939                         if (ret < 0)
3940                                 goto out;
3941                         if (ret) {
3942                                 struct name_cache_entry *nce;
3943                                 struct waiting_dir_move *wdm;
3944 
3945                                 ret = orphanize_inode(sctx, ow_inode, ow_gen,
3946                                                 cur->full_path);
3947                                 if (ret < 0)
3948                                         goto out;
3949                                 if (S_ISDIR(ow_mode))
3950                                         orphanized_dir = true;
3951 
3952                                 /*
3953                                  * If ow_inode has its rename operation delayed
3954                                  * make sure that its orphanized name is used in
3955                                  * the source path when performing its rename
3956                                  * operation.
3957                                  */
3958                                 if (is_waiting_for_move(sctx, ow_inode)) {
3959                                         wdm = get_waiting_dir_move(sctx,
3960                                                                    ow_inode);
3961                                         ASSERT(wdm);
3962                                         wdm->orphanized = true;
3963                                 }
3964 
3965                                 /*
3966                                  * Make sure we clear our orphanized inode's
3967                                  * name from the name cache. This is because the
3968                                  * inode ow_inode might be an ancestor of some
3969                                  * other inode that will be orphanized as well
3970                                  * later and has an inode number greater than
3971                                  * sctx->send_progress. We need to prevent
3972                                  * future name lookups from using the old name
3973                                  * and get instead the orphan name.
3974                                  */
3975                                 nce = name_cache_search(sctx, ow_inode, ow_gen);
3976                                 if (nce) {
3977                                         name_cache_delete(sctx, nce);
3978                                         kfree(nce);
3979                                 }
3980 
3981                                 /*
3982                                  * ow_inode might currently be an ancestor of
3983                                  * cur_ino, therefore compute valid_path (the
3984                                  * current path of cur_ino) again because it
3985                                  * might contain the pre-orphanization name of
3986                                  * ow_inode, which is no longer valid.
3987                                  */
3988                                 ret = is_ancestor(sctx->parent_root,
3989                                                   ow_inode, ow_gen,
3990                                                   sctx->cur_ino, NULL);
3991                                 if (ret > 0) {
3992                                         orphanized_ancestor = true;
3993                                         fs_path_reset(valid_path);
3994                                         ret = get_cur_path(sctx, sctx->cur_ino,
3995                                                            sctx->cur_inode_gen,
3996                                                            valid_path);
3997                                 }
3998                                 if (ret < 0)
3999                                         goto out;
4000                         } else {
4001                                 ret = send_unlink(sctx, cur->full_path);
4002                                 if (ret < 0)
4003                                         goto out;
4004                         }
4005                 }
4006 
4007                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) {
4008                         ret = wait_for_dest_dir_move(sctx, cur, is_orphan);
4009                         if (ret < 0)
4010                                 goto out;
4011                         if (ret == 1) {
4012                                 can_rename = false;
4013                                 *pending_move = 1;
4014                         }
4015                 }
4016 
4017                 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root &&
4018                     can_rename) {
4019                         ret = wait_for_parent_move(sctx, cur, is_orphan);
4020                         if (ret < 0)
4021                                 goto out;
4022                         if (ret == 1) {
4023                                 can_rename = false;
4024                                 *pending_move = 1;
4025                         }
4026                 }
4027 
4028                 /*
4029                  * link/move the ref to the new place. If we have an orphan
4030                  * inode, move it and update valid_path. If not, link or move
4031                  * it depending on the inode mode.
4032                  */
4033                 if (is_orphan && can_rename) {
4034                         ret = send_rename(sctx, valid_path, cur->full_path);
4035                         if (ret < 0)
4036                                 goto out;
4037                         is_orphan = 0;
4038                         ret = fs_path_copy(valid_path, cur->full_path);
4039                         if (ret < 0)
4040                                 goto out;
4041                 } else if (can_rename) {
4042                         if (S_ISDIR(sctx->cur_inode_mode)) {
4043                                 /*
4044                                  * Dirs can't be linked, so move it. For moved
4045                                  * dirs, we always have one new and one deleted
4046                                  * ref. The deleted ref is ignored later.
4047                                  */
4048                                 ret = send_rename(sctx, valid_path,
4049                                                   cur->full_path);
4050                                 if (!ret)
4051                                         ret = fs_path_copy(valid_path,
4052                                                            cur->full_path);
4053                                 if (ret < 0)
4054                                         goto out;
4055                         } else {
4056                                 /*
4057                                  * We might have previously orphanized an inode
4058                                  * which is an ancestor of our current inode,
4059                                  * so our reference's full path, which was
4060                                  * computed before any such orphanizations, must
4061                                  * be updated.
4062                                  */
4063                                 if (orphanized_dir) {
4064                                         ret = update_ref_path(sctx, cur);
4065                                         if (ret < 0)
4066                                                 goto out;
4067                                 }
4068                                 ret = send_link(sctx, cur->full_path,
4069                                                 valid_path);
4070                                 if (ret < 0)
4071                                         goto out;
4072                         }
4073                 }
4074                 ret = dup_ref(cur, &check_dirs);
4075                 if (ret < 0)
4076                         goto out;
4077         }
4078 
4079         if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) {
4080                 /*
4081                  * Check if we can already rmdir the directory. If not,
4082                  * orphanize it. For every dir item inside that gets deleted
4083                  * later, we do this check again and rmdir it then if possible.
4084                  * See the use of check_dirs for more details.
4085                  */
4086                 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen,
4087                                 sctx->cur_ino);
4088                 if (ret < 0)
4089                         goto out;
4090                 if (ret) {
4091                         ret = send_rmdir(sctx, valid_path);
4092                         if (ret < 0)
4093                                 goto out;
4094                 } else if (!is_orphan) {
4095                         ret = orphanize_inode(sctx, sctx->cur_ino,
4096                                         sctx->cur_inode_gen, valid_path);
4097                         if (ret < 0)
4098                                 goto out;
4099                         is_orphan = 1;
4100                 }
4101 
4102                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4103                         ret = dup_ref(cur, &check_dirs);
4104                         if (ret < 0)
4105                                 goto out;
4106                 }
4107         } else if (S_ISDIR(sctx->cur_inode_mode) &&
4108                    !list_empty(&sctx->deleted_refs)) {
4109                 /*
4110                  * We have a moved dir. Add the old parent to check_dirs
4111                  */
4112                 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref,
4113                                 list);
4114                 ret = dup_ref(cur, &check_dirs);
4115                 if (ret < 0)
4116                         goto out;
4117         } else if (!S_ISDIR(sctx->cur_inode_mode)) {
4118                 /*
4119                  * We have a non dir inode. Go through all deleted refs and
4120                  * unlink them if they were not already overwritten by other
4121                  * inodes.
4122                  */
4123                 list_for_each_entry(cur, &sctx->deleted_refs, list) {
4124                         ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen,
4125                                         sctx->cur_ino, sctx->cur_inode_gen,
4126                                         cur->name, cur->name_len);
4127                         if (ret < 0)
4128                                 goto out;
4129                         if (!ret) {
4130                                 /*
4131                                  * If we orphanized any ancestor before, we need
4132                                  * to recompute the full path for deleted names,
4133                                  * since any such path was computed before we
4134                                  * processed any references and orphanized any
4135                                  * ancestor inode.
4136                                  */
4137                                 if (orphanized_ancestor) {
4138                                         ret = update_ref_path(sctx, cur);
4139                                         if (ret < 0)
4140                                                 goto out;
4141                                 }
4142                                 ret = send_unlink(sctx, cur->full_path);
4143                                 if (ret < 0)
4144                                         goto out;
4145                         }
4146                         ret = dup_ref(cur, &check_dirs);
4147                         if (ret < 0)
4148                                 goto out;
4149                 }
4150                 /*
4151                  * If the inode is still orphan, unlink the orphan. This may
4152                  * happen when a previous inode did overwrite the first ref
4153                  * of this inode and no new refs were added for the current
4154                  * inode. Unlinking does not mean that the inode is deleted in
4155                  * all cases. There may still be links to this inode in other
4156                  * places.
4157                  */
4158                 if (is_orphan) {
4159                         ret = send_unlink(sctx, valid_path);
4160                         if (ret < 0)
4161                                 goto out;
4162                 }
4163         }
4164 
4165         /*
4166          * We did collect all parent dirs where cur_inode was once located. We
4167          * now go through all these dirs and check if they are pending for
4168          * deletion and if it's finally possible to perform the rmdir now.
4169          * We also update the inode stats of the parent dirs here.
4170          */
4171         list_for_each_entry(cur, &check_dirs, list) {
4172                 /*
4173                  * In case we had refs into dirs that were not processed yet,
4174                  * we don't need to do the utime and rmdir logic for these dirs.
4175                  * The dir will be processed later.
4176                  */
4177                 if (cur->dir > sctx->cur_ino)
4178                         continue;
4179 
4180                 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen);
4181                 if (ret < 0)
4182                         goto out;
4183 
4184                 if (ret == inode_state_did_create ||
4185                     ret == inode_state_no_change) {
4186                         /* TODO delayed utimes */
4187                         ret = send_utimes(sctx, cur->dir, cur->dir_gen);
4188                         if (ret < 0)
4189                                 goto out;
4190                 } else if (ret == inode_state_did_delete &&
4191                            cur->dir != last_dir_ino_rm) {
4192                         ret = can_rmdir(sctx, cur->dir, cur->dir_gen,
4193                                         sctx->cur_ino);
4194                         if (ret < 0)
4195                                 goto out;
4196                         if (ret) {
4197                                 ret = get_cur_path(sctx, cur->dir,
4198                                                    cur->dir_gen, valid_path);
4199                                 if (ret < 0)
4200                                         goto out;
4201                                 ret = send_rmdir(sctx, valid_path);
4202                                 if (ret < 0)
4203                                         goto out;
4204                                 last_dir_ino_rm = cur->dir;
4205                         }
4206                 }
4207         }
4208 
4209         ret = 0;
4210 
4211 out:
4212         __free_recorded_refs(&check_dirs);
4213         free_recorded_refs(sctx);
4214         fs_path_free(valid_path);
4215         return ret;
4216 }
4217 
4218 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name,
4219                       void *ctx, struct list_head *refs)
4220 {
4221         int ret = 0;
4222         struct send_ctx *sctx = ctx;
4223         struct fs_path *p;
4224         u64 gen;
4225 
4226         p = fs_path_alloc();
4227         if (!p)
4228                 return -ENOMEM;
4229 
4230         ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL,
4231                         NULL, NULL);
4232         if (ret < 0)
4233                 goto out;
4234 
4235         ret = get_cur_path(sctx, dir, gen, p);
4236         if (ret < 0)
4237                 goto out;
4238         ret = fs_path_add_path(p, name);
4239         if (ret < 0)
4240                 goto out;
4241 
4242         ret = __record_ref(refs, dir, gen, p);
4243 
4244 out:
4245         if (ret)
4246                 fs_path_free(p);
4247         return ret;
4248 }
4249 
4250 static int __record_new_ref(int num, u64 dir, int index,
4251                             struct fs_path *name,
4252                             void *ctx)
4253 {
4254         struct send_ctx *sctx = ctx;
4255         return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs);
4256 }
4257 
4258 
4259 static int __record_deleted_ref(int num, u64 dir, int index,
4260                                 struct fs_path *name,
4261                                 void *ctx)
4262 {
4263         struct send_ctx *sctx = ctx;
4264         return record_ref(sctx->parent_root, dir, name, ctx,
4265                           &sctx->deleted_refs);
4266 }
4267 
4268 static int record_new_ref(struct send_ctx *sctx)
4269 {
4270         int ret;
4271 
4272         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4273                                 sctx->cmp_key, 0, __record_new_ref, sctx);
4274         if (ret < 0)
4275                 goto out;
4276         ret = 0;
4277 
4278 out:
4279         return ret;
4280 }
4281 
4282 static int record_deleted_ref(struct send_ctx *sctx)
4283 {
4284         int ret;
4285 
4286         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4287                                 sctx->cmp_key, 0, __record_deleted_ref, sctx);
4288         if (ret < 0)
4289                 goto out;
4290         ret = 0;
4291 
4292 out:
4293         return ret;
4294 }
4295 
4296 struct find_ref_ctx {
4297         u64 dir;
4298         u64 dir_gen;
4299         struct btrfs_root *root;
4300         struct fs_path *name;
4301         int found_idx;
4302 };
4303 
4304 static int __find_iref(int num, u64 dir, int index,
4305                        struct fs_path *name,
4306                        void *ctx_)
4307 {
4308         struct find_ref_ctx *ctx = ctx_;
4309         u64 dir_gen;
4310         int ret;
4311 
4312         if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) &&
4313             strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) {
4314                 /*
4315                  * To avoid doing extra lookups we'll only do this if everything
4316                  * else matches.
4317                  */
4318                 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL,
4319                                      NULL, NULL, NULL);
4320                 if (ret)
4321                         return ret;
4322                 if (dir_gen != ctx->dir_gen)
4323                         return 0;
4324                 ctx->found_idx = num;
4325                 return 1;
4326         }
4327         return 0;
4328 }
4329 
4330 static int find_iref(struct btrfs_root *root,
4331                      struct btrfs_path *path,
4332                      struct btrfs_key *key,
4333                      u64 dir, u64 dir_gen, struct fs_path *name)
4334 {
4335         int ret;
4336         struct find_ref_ctx ctx;
4337 
4338         ctx.dir = dir;
4339         ctx.name = name;
4340         ctx.dir_gen = dir_gen;
4341         ctx.found_idx = -1;
4342         ctx.root = root;
4343 
4344         ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx);
4345         if (ret < 0)
4346                 return ret;
4347 
4348         if (ctx.found_idx == -1)
4349                 return -ENOENT;
4350 
4351         return ctx.found_idx;
4352 }
4353 
4354 static int __record_changed_new_ref(int num, u64 dir, int index,
4355                                     struct fs_path *name,
4356                                     void *ctx)
4357 {
4358         u64 dir_gen;
4359         int ret;
4360         struct send_ctx *sctx = ctx;
4361 
4362         ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL,
4363                              NULL, NULL, NULL);
4364         if (ret)
4365                 return ret;
4366 
4367         ret = find_iref(sctx->parent_root, sctx->right_path,
4368                         sctx->cmp_key, dir, dir_gen, name);
4369         if (ret == -ENOENT)
4370                 ret = __record_new_ref(num, dir, index, name, sctx);
4371         else if (ret > 0)
4372                 ret = 0;
4373 
4374         return ret;
4375 }
4376 
4377 static int __record_changed_deleted_ref(int num, u64 dir, int index,
4378                                         struct fs_path *name,
4379                                         void *ctx)
4380 {
4381         u64 dir_gen;
4382         int ret;
4383         struct send_ctx *sctx = ctx;
4384 
4385         ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL,
4386                              NULL, NULL, NULL);
4387         if (ret)
4388                 return ret;
4389 
4390         ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key,
4391                         dir, dir_gen, name);
4392         if (ret == -ENOENT)
4393                 ret = __record_deleted_ref(num, dir, index, name, sctx);
4394         else if (ret > 0)
4395                 ret = 0;
4396 
4397         return ret;
4398 }
4399 
4400 static int record_changed_ref(struct send_ctx *sctx)
4401 {
4402         int ret = 0;
4403 
4404         ret = iterate_inode_ref(sctx->send_root, sctx->left_path,
4405                         sctx->cmp_key, 0, __record_changed_new_ref, sctx);
4406         if (ret < 0)
4407                 goto out;
4408         ret = iterate_inode_ref(sctx->parent_root, sctx->right_path,
4409                         sctx->cmp_key, 0, __record_changed_deleted_ref, sctx);
4410         if (ret < 0)
4411                 goto out;
4412         ret = 0;
4413 
4414 out:
4415         return ret;
4416 }
4417 
4418 /*
4419  * Record and process all refs at once. Needed when an inode changes the
4420  * generation number, which means that it was deleted and recreated.
4421  */
4422 static int process_all_refs(struct send_ctx *sctx,
4423                             enum btrfs_compare_tree_result cmd)
4424 {
4425         int ret;
4426         struct btrfs_root *root;
4427         struct btrfs_path *path;
4428         struct btrfs_key key;
4429         struct btrfs_key found_key;
4430         struct extent_buffer *eb;
4431         int slot;
4432         iterate_inode_ref_t cb;
4433         int pending_move = 0;
4434 
4435         path = alloc_path_for_send();
4436         if (!path)
4437                 return -ENOMEM;
4438 
4439         if (cmd == BTRFS_COMPARE_TREE_NEW) {
4440                 root = sctx->send_root;
4441                 cb = __record_new_ref;
4442         } else if (cmd == BTRFS_COMPARE_TREE_DELETED) {
4443                 root = sctx->parent_root;
4444                 cb = __record_deleted_ref;
4445         } else {
4446                 btrfs_err(sctx->send_root->fs_info,
4447                                 "Wrong command %d in process_all_refs", cmd);
4448                 ret = -EINVAL;
4449                 goto out;
4450         }
4451 
4452         key.objectid = sctx->cmp_key->objectid;
4453         key.type = BTRFS_INODE_REF_KEY;
4454         key.offset = 0;
4455         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4456         if (ret < 0)
4457                 goto out;
4458 
4459         while (1) {
4460                 eb = path->nodes[0];
4461                 slot = path->slots[0];
4462                 if (slot >= btrfs_header_nritems(eb)) {
4463                         ret = btrfs_next_leaf(root, path);
4464                         if (ret < 0)
4465                                 goto out;
4466                         else if (ret > 0)
4467                                 break;
4468                         continue;
4469                 }
4470 
4471                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4472 
4473                 if (found_key.objectid != key.objectid ||
4474                     (found_key.type != BTRFS_INODE_REF_KEY &&
4475                      found_key.type != BTRFS_INODE_EXTREF_KEY))
4476                         break;
4477 
4478                 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx);
4479                 if (ret < 0)
4480                         goto out;
4481 
4482                 path->slots[0]++;
4483         }
4484         btrfs_release_path(path);
4485 
4486         /*
4487          * We don't actually care about pending_move as we are simply
4488          * re-creating this inode and will be rename'ing it into place once we
4489          * rename the parent directory.
4490          */
4491         ret = process_recorded_refs(sctx, &pending_move);
4492 out:
4493         btrfs_free_path(path);
4494         return ret;
4495 }
4496 
4497 static int send_set_xattr(struct send_ctx *sctx,
4498                           struct fs_path *path,
4499                           const char *name, int name_len,
4500                           const char *data, int data_len)
4501 {
4502         int ret = 0;
4503 
4504         ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR);
4505         if (ret < 0)
4506                 goto out;
4507 
4508         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4509         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4510         TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len);
4511 
4512         ret = send_cmd(sctx);
4513 
4514 tlv_put_failure:
4515 out:
4516         return ret;
4517 }
4518 
4519 static int send_remove_xattr(struct send_ctx *sctx,
4520                           struct fs_path *path,
4521                           const char *name, int name_len)
4522 {
4523         int ret = 0;
4524 
4525         ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR);
4526         if (ret < 0)
4527                 goto out;
4528 
4529         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path);
4530         TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len);
4531 
4532         ret = send_cmd(sctx);
4533 
4534 tlv_put_failure:
4535 out:
4536         return ret;
4537 }
4538 
4539 static int __process_new_xattr(int num, struct btrfs_key *di_key,
4540                                const char *name, int name_len,
4541                                const char *data, int data_len,
4542                                u8 type, void *ctx)
4543 {
4544         int ret;
4545         struct send_ctx *sctx = ctx;
4546         struct fs_path *p;
4547         struct posix_acl_xattr_header dummy_acl;
4548 
4549         /* Capabilities are emitted by finish_inode_if_needed */
4550         if (!strncmp(name, XATTR_NAME_CAPS, name_len))
4551                 return 0;
4552 
4553         p = fs_path_alloc();
4554         if (!p)
4555                 return -ENOMEM;
4556 
4557         /*
4558          * This hack is needed because empty acls are stored as zero byte
4559          * data in xattrs. Problem with that is, that receiving these zero byte
4560          * acls will fail later. To fix this, we send a dummy acl list that
4561          * only contains the version number and no entries.
4562          */
4563         if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) ||
4564             !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) {
4565                 if (data_len == 0) {
4566                         dummy_acl.a_version =
4567                                         cpu_to_le32(POSIX_ACL_XATTR_VERSION);
4568                         data = (char *)&dummy_acl;
4569                         data_len = sizeof(dummy_acl);
4570                 }
4571         }
4572 
4573         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4574         if (ret < 0)
4575                 goto out;
4576 
4577         ret = send_set_xattr(sctx, p, name, name_len, data, data_len);
4578 
4579 out:
4580         fs_path_free(p);
4581         return ret;
4582 }
4583 
4584 static int __process_deleted_xattr(int num, struct btrfs_key *di_key,
4585                                    const char *name, int name_len,
4586                                    const char *data, int data_len,
4587                                    u8 type, void *ctx)
4588 {
4589         int ret;
4590         struct send_ctx *sctx = ctx;
4591         struct fs_path *p;
4592 
4593         p = fs_path_alloc();
4594         if (!p)
4595                 return -ENOMEM;
4596 
4597         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4598         if (ret < 0)
4599                 goto out;
4600 
4601         ret = send_remove_xattr(sctx, p, name, name_len);
4602 
4603 out:
4604         fs_path_free(p);
4605         return ret;
4606 }
4607 
4608 static int process_new_xattr(struct send_ctx *sctx)
4609 {
4610         int ret = 0;
4611 
4612         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4613                                __process_new_xattr, sctx);
4614 
4615         return ret;
4616 }
4617 
4618 static int process_deleted_xattr(struct send_ctx *sctx)
4619 {
4620         return iterate_dir_item(sctx->parent_root, sctx->right_path,
4621                                 __process_deleted_xattr, sctx);
4622 }
4623 
4624 struct find_xattr_ctx {
4625         const char *name;
4626         int name_len;
4627         int found_idx;
4628         char *found_data;
4629         int found_data_len;
4630 };
4631 
4632 static int __find_xattr(int num, struct btrfs_key *di_key,
4633                         const char *name, int name_len,
4634                         const char *data, int data_len,
4635                         u8 type, void *vctx)
4636 {
4637         struct find_xattr_ctx *ctx = vctx;
4638 
4639         if (name_len == ctx->name_len &&
4640             strncmp(name, ctx->name, name_len) == 0) {
4641                 ctx->found_idx = num;
4642                 ctx->found_data_len = data_len;
4643                 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL);
4644                 if (!ctx->found_data)
4645                         return -ENOMEM;
4646                 return 1;
4647         }
4648         return 0;
4649 }
4650 
4651 static int find_xattr(struct btrfs_root *root,
4652                       struct btrfs_path *path,
4653                       struct btrfs_key *key,
4654                       const char *name, int name_len,
4655                       char **data, int *data_len)
4656 {
4657         int ret;
4658         struct find_xattr_ctx ctx;
4659 
4660         ctx.name = name;
4661         ctx.name_len = name_len;
4662         ctx.found_idx = -1;
4663         ctx.found_data = NULL;
4664         ctx.found_data_len = 0;
4665 
4666         ret = iterate_dir_item(root, path, __find_xattr, &ctx);
4667         if (ret < 0)
4668                 return ret;
4669 
4670         if (ctx.found_idx == -1)
4671                 return -ENOENT;
4672         if (data) {
4673                 *data = ctx.found_data;
4674                 *data_len = ctx.found_data_len;
4675         } else {
4676                 kfree(ctx.found_data);
4677         }
4678         return ctx.found_idx;
4679 }
4680 
4681 
4682 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key,
4683                                        const char *name, int name_len,
4684                                        const char *data, int data_len,
4685                                        u8 type, void *ctx)
4686 {
4687         int ret;
4688         struct send_ctx *sctx = ctx;
4689         char *found_data = NULL;
4690         int found_data_len  = 0;
4691 
4692         ret = find_xattr(sctx->parent_root, sctx->right_path,
4693                          sctx->cmp_key, name, name_len, &found_data,
4694                          &found_data_len);
4695         if (ret == -ENOENT) {
4696                 ret = __process_new_xattr(num, di_key, name, name_len, data,
4697                                 data_len, type, ctx);
4698         } else if (ret >= 0) {
4699                 if (data_len != found_data_len ||
4700                     memcmp(data, found_data, data_len)) {
4701                         ret = __process_new_xattr(num, di_key, name, name_len,
4702                                         data, data_len, type, ctx);
4703                 } else {
4704                         ret = 0;
4705                 }
4706         }
4707 
4708         kfree(found_data);
4709         return ret;
4710 }
4711 
4712 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key,
4713                                            const char *name, int name_len,
4714                                            const char *data, int data_len,
4715                                            u8 type, void *ctx)
4716 {
4717         int ret;
4718         struct send_ctx *sctx = ctx;
4719 
4720         ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key,
4721                          name, name_len, NULL, NULL);
4722         if (ret == -ENOENT)
4723                 ret = __process_deleted_xattr(num, di_key, name, name_len, data,
4724                                 data_len, type, ctx);
4725         else if (ret >= 0)
4726                 ret = 0;
4727 
4728         return ret;
4729 }
4730 
4731 static int process_changed_xattr(struct send_ctx *sctx)
4732 {
4733         int ret = 0;
4734 
4735         ret = iterate_dir_item(sctx->send_root, sctx->left_path,
4736                         __process_changed_new_xattr, sctx);
4737         if (ret < 0)
4738                 goto out;
4739         ret = iterate_dir_item(sctx->parent_root, sctx->right_path,
4740                         __process_changed_deleted_xattr, sctx);
4741 
4742 out:
4743         return ret;
4744 }
4745 
4746 static int process_all_new_xattrs(struct send_ctx *sctx)
4747 {
4748         int ret;
4749         struct btrfs_root *root;
4750         struct btrfs_path *path;
4751         struct btrfs_key key;
4752         struct btrfs_key found_key;
4753         struct extent_buffer *eb;
4754         int slot;
4755 
4756         path = alloc_path_for_send();
4757         if (!path)
4758                 return -ENOMEM;
4759 
4760         root = sctx->send_root;
4761 
4762         key.objectid = sctx->cmp_key->objectid;
4763         key.type = BTRFS_XATTR_ITEM_KEY;
4764         key.offset = 0;
4765         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4766         if (ret < 0)
4767                 goto out;
4768 
4769         while (1) {
4770                 eb = path->nodes[0];
4771                 slot = path->slots[0];
4772                 if (slot >= btrfs_header_nritems(eb)) {
4773                         ret = btrfs_next_leaf(root, path);
4774                         if (ret < 0) {
4775                                 goto out;
4776                         } else if (ret > 0) {
4777                                 ret = 0;
4778                                 break;
4779                         }
4780                         continue;
4781                 }
4782 
4783                 btrfs_item_key_to_cpu(eb, &found_key, slot);
4784                 if (found_key.objectid != key.objectid ||
4785                     found_key.type != key.type) {
4786                         ret = 0;
4787                         goto out;
4788                 }
4789 
4790                 ret = iterate_dir_item(root, path, __process_new_xattr, sctx);
4791                 if (ret < 0)
4792                         goto out;
4793 
4794                 path->slots[0]++;
4795         }
4796 
4797 out:
4798         btrfs_free_path(path);
4799         return ret;
4800 }
4801 
4802 static ssize_t fill_read_buf(struct send_ctx *sctx, u64 offset, u32 len)
4803 {
4804         struct btrfs_root *root = sctx->send_root;
4805         struct btrfs_fs_info *fs_info = root->fs_info;
4806         struct inode *inode;
4807         struct page *page;
4808         char *addr;
4809         struct btrfs_key key;
4810         pgoff_t index = offset >> PAGE_SHIFT;
4811         pgoff_t last_index;
4812         unsigned pg_offset = offset_in_page(offset);
4813         ssize_t ret = 0;
4814 
4815         key.objectid = sctx->cur_ino;
4816         key.type = BTRFS_INODE_ITEM_KEY;
4817         key.offset = 0;
4818 
4819         inode = btrfs_iget(fs_info->sb, &key, root);
4820         if (IS_ERR(inode))
4821                 return PTR_ERR(inode);
4822 
4823         if (offset + len > i_size_read(inode)) {
4824                 if (offset > i_size_read(inode))
4825                         len = 0;
4826                 else
4827                         len = offset - i_size_read(inode);
4828         }
4829         if (len == 0)
4830                 goto out;
4831 
4832         last_index = (offset + len - 1) >> PAGE_SHIFT;
4833 
4834         /* initial readahead */
4835         memset(&sctx->ra, 0, sizeof(struct file_ra_state));
4836         file_ra_state_init(&sctx->ra, inode->i_mapping);
4837 
4838         while (index <= last_index) {
4839                 unsigned cur_len = min_t(unsigned, len,
4840                                          PAGE_SIZE - pg_offset);
4841 
4842                 page = find_lock_page(inode->i_mapping, index);
4843                 if (!page) {
4844                         page_cache_sync_readahead(inode->i_mapping, &sctx->ra,
4845                                 NULL, index, last_index + 1 - index);
4846 
4847                         page = find_or_create_page(inode->i_mapping, index,
4848                                         GFP_KERNEL);
4849                         if (!page) {
4850                                 ret = -ENOMEM;
4851                                 break;
4852                         }
4853                 }
4854 
4855                 if (PageReadahead(page)) {
4856                         page_cache_async_readahead(inode->i_mapping, &sctx->ra,
4857                                 NULL, page, index, last_index + 1 - index);
4858                 }
4859 
4860                 if (!PageUptodate(page)) {
4861                         btrfs_readpage(NULL, page);
4862                         lock_page(page);
4863                         if (!PageUptodate(page)) {
4864                                 unlock_page(page);
4865                                 put_page(page);
4866                                 ret = -EIO;
4867                                 break;
4868                         }
4869                 }
4870 
4871                 addr = kmap(page);
4872                 memcpy(sctx->read_buf + ret, addr + pg_offset, cur_len);
4873                 kunmap(page);
4874                 unlock_page(page);
4875                 put_page(page);
4876                 index++;
4877                 pg_offset = 0;
4878                 len -= cur_len;
4879                 ret += cur_len;
4880         }
4881 out:
4882         iput(inode);
4883         return ret;
4884 }
4885 
4886 /*
4887  * Read some bytes from the current inode/file and send a write command to
4888  * user space.
4889  */
4890 static int send_write(struct send_ctx *sctx, u64 offset, u32 len)
4891 {
4892         struct btrfs_fs_info *fs_info = sctx->send_root->fs_info;
4893         int ret = 0;
4894         struct fs_path *p;
4895         ssize_t num_read = 0;
4896 
4897         p = fs_path_alloc();
4898         if (!p)
4899                 return -ENOMEM;
4900 
4901         btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len);
4902 
4903         num_read = fill_read_buf(sctx, offset, len);
4904         if (num_read <= 0) {
4905                 if (num_read < 0)
4906                         ret = num_read;
4907                 goto out;
4908         }
4909 
4910         ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
4911         if (ret < 0)
4912                 goto out;
4913 
4914         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4915         if (ret < 0)
4916                 goto out;
4917 
4918         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4919         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4920         TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, num_read);
4921 
4922         ret = send_cmd(sctx);
4923 
4924 tlv_put_failure:
4925 out:
4926         fs_path_free(p);
4927         if (ret < 0)
4928                 return ret;
4929         return num_read;
4930 }
4931 
4932 /*
4933  * Send a clone command to user space.
4934  */
4935 static int send_clone(struct send_ctx *sctx,
4936                       u64 offset, u32 len,
4937                       struct clone_root *clone_root)
4938 {
4939         int ret = 0;
4940         struct fs_path *p;
4941         u64 gen;
4942 
4943         btrfs_debug(sctx->send_root->fs_info,
4944                     "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu",
4945                     offset, len, clone_root->root->root_key.objectid,
4946                     clone_root->ino, clone_root->offset);
4947 
4948         p = fs_path_alloc();
4949         if (!p)
4950                 return -ENOMEM;
4951 
4952         ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE);
4953         if (ret < 0)
4954                 goto out;
4955 
4956         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
4957         if (ret < 0)
4958                 goto out;
4959 
4960         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
4961         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len);
4962         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
4963 
4964         if (clone_root->root == sctx->send_root) {
4965                 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL,
4966                                 &gen, NULL, NULL, NULL, NULL);
4967                 if (ret < 0)
4968                         goto out;
4969                 ret = get_cur_path(sctx, clone_root->ino, gen, p);
4970         } else {
4971                 ret = get_inode_path(clone_root->root, clone_root->ino, p);
4972         }
4973         if (ret < 0)
4974                 goto out;
4975 
4976         /*
4977          * If the parent we're using has a received_uuid set then use that as
4978          * our clone source as that is what we will look for when doing a
4979          * receive.
4980          *
4981          * This covers the case that we create a snapshot off of a received
4982          * subvolume and then use that as the parent and try to receive on a
4983          * different host.
4984          */
4985         if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid))
4986                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4987                              clone_root->root->root_item.received_uuid);
4988         else
4989                 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID,
4990                              clone_root->root->root_item.uuid);
4991         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID,
4992                     le64_to_cpu(clone_root->root->root_item.ctransid));
4993         TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p);
4994         TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET,
4995                         clone_root->offset);
4996 
4997         ret = send_cmd(sctx);
4998 
4999 tlv_put_failure:
5000 out:
5001         fs_path_free(p);
5002         return ret;
5003 }
5004 
5005 /*
5006  * Send an update extent command to user space.
5007  */
5008 static int send_update_extent(struct send_ctx *sctx,
5009                               u64 offset, u32 len)
5010 {
5011         int ret = 0;
5012         struct fs_path *p;
5013 
5014         p = fs_path_alloc();
5015         if (!p)
5016                 return -ENOMEM;
5017 
5018         ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT);
5019         if (ret < 0)
5020                 goto out;
5021 
5022         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5023         if (ret < 0)
5024                 goto out;
5025 
5026         TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5027         TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5028         TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len);
5029 
5030         ret = send_cmd(sctx);
5031 
5032 tlv_put_failure:
5033 out:
5034         fs_path_free(p);
5035         return ret;
5036 }
5037 
5038 static int send_hole(struct send_ctx *sctx, u64 end)
5039 {
5040         struct fs_path *p = NULL;
5041         u64 offset = sctx->cur_inode_last_extent;
5042         u64 len;
5043         int ret = 0;
5044 
5045         /*
5046          * A hole that starts at EOF or beyond it. Since we do not yet support
5047          * fallocate (for extent preallocation and hole punching), sending a
5048          * write of zeroes starting at EOF or beyond would later require issuing
5049          * a truncate operation which would undo the write and achieve nothing.
5050          */
5051         if (offset >= sctx->cur_inode_size)
5052                 return 0;
5053 
5054         /*
5055          * Don't go beyond the inode's i_size due to prealloc extents that start
5056          * after the i_size.
5057          */
5058         end = min_t(u64, end, sctx->cur_inode_size);
5059 
5060         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5061                 return send_update_extent(sctx, offset, end - offset);
5062 
5063         p = fs_path_alloc();
5064         if (!p)
5065                 return -ENOMEM;
5066         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p);
5067         if (ret < 0)
5068                 goto tlv_put_failure;
5069         memset(sctx->read_buf, 0, BTRFS_SEND_READ_SIZE);
5070         while (offset < end) {
5071                 len = min_t(u64, end - offset, BTRFS_SEND_READ_SIZE);
5072 
5073                 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE);
5074                 if (ret < 0)
5075                         break;
5076                 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p);
5077                 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset);
5078                 TLV_PUT(sctx, BTRFS_SEND_A_DATA, sctx->read_buf, len);
5079                 ret = send_cmd(sctx);
5080                 if (ret < 0)
5081                         break;
5082                 offset += len;
5083         }
5084         sctx->cur_inode_next_write_offset = offset;
5085 tlv_put_failure:
5086         fs_path_free(p);
5087         return ret;
5088 }
5089 
5090 static int send_extent_data(struct send_ctx *sctx,
5091                             const u64 offset,
5092                             const u64 len)
5093 {
5094         u64 sent = 0;
5095 
5096         if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA)
5097                 return send_update_extent(sctx, offset, len);
5098 
5099         while (sent < len) {
5100                 u64 size = len - sent;
5101                 int ret;
5102 
5103                 if (size > BTRFS_SEND_READ_SIZE)
5104                         size = BTRFS_SEND_READ_SIZE;
5105                 ret = send_write(sctx, offset + sent, size);
5106                 if (ret < 0)
5107                         return ret;
5108                 if (!ret)
5109                         break;
5110                 sent += ret;
5111         }
5112         return 0;
5113 }
5114 
5115 /*
5116  * Search for a capability xattr related to sctx->cur_ino. If the capability is
5117  * found, call send_set_xattr function to emit it.
5118  *
5119  * Return 0 if there isn't a capability, or when the capability was emitted
5120  * successfully, or < 0 if an error occurred.
5121  */
5122 static int send_capabilities(struct send_ctx *sctx)
5123 {
5124         struct fs_path *fspath = NULL;
5125         struct btrfs_path *path;
5126         struct btrfs_dir_item *di;
5127         struct extent_buffer *leaf;
5128         unsigned long data_ptr;
5129         char *buf = NULL;
5130         int buf_len;
5131         int ret = 0;
5132 
5133         path = alloc_path_for_send();
5134         if (!path)
5135                 return -ENOMEM;
5136 
5137         di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino,
5138                                 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0);
5139         if (!di) {
5140                 /* There is no xattr for this inode */
5141                 goto out;
5142         } else if (IS_ERR(di)) {
5143                 ret = PTR_ERR(di);
5144                 goto out;
5145         }
5146 
5147         leaf = path->nodes[0];
5148         buf_len = btrfs_dir_data_len(leaf, di);
5149 
5150         fspath = fs_path_alloc();
5151         buf = kmalloc(buf_len, GFP_KERNEL);
5152         if (!fspath || !buf) {
5153                 ret = -ENOMEM;
5154                 goto out;
5155         }
5156 
5157         ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath);
5158         if (ret < 0)
5159                 goto out;
5160 
5161         data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di);
5162         read_extent_buffer(leaf, buf, data_ptr, buf_len);
5163 
5164         ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS,
5165                         strlen(XATTR_NAME_CAPS), buf, buf_len);
5166 out:
5167         kfree(buf);
5168         fs_path_free(fspath);
5169         btrfs_free_path(path);
5170         return ret;
5171 }
5172 
5173 static int clone_range(struct send_ctx *sctx,
5174                        struct clone_root *clone_root,
5175                        const u64 disk_byte,
5176                        u64 data_offset,
5177                        u64 offset,
5178                        u64 len)
5179 {
5180         struct btrfs_path *path;
5181         struct btrfs_key key;
5182         int ret;
5183         u64 clone_src_i_size = 0;
5184 
5185         /*
5186          * Prevent cloning from a zero offset with a length matching the sector
5187          * size because in some scenarios this will make the receiver fail.
5188          *
5189          * For example, if in the source filesystem the extent at offset 0
5190          * has a length of sectorsize and it was written using direct IO, then
5191          * it can never be an inline extent (even if compression is enabled).
5192          * Then this extent can be cloned in the original filesystem to a non
5193          * zero file offset, but it may not be possible to clone in the
5194          * destination filesystem because it can be inlined due to compression
5195          * on the destination filesystem (as the receiver's write operations are
5196          * always done using buffered IO). The same happens when the original
5197          * filesystem does not have compression enabled but the destination
5198          * filesystem has.
5199          */
5200         if (clone_root->offset == 0 &&
5201             len == sctx->send_root->fs_info->sectorsize)
5202                 return send_extent_data(sctx, offset, len);
5203 
5204         path = alloc_path_for_send();
5205         if (!path)
5206                 return -ENOMEM;
5207 
5208         /*
5209          * There are inodes that have extents that lie behind its i_size. Don't
5210          * accept clones from these extents.
5211          */
5212         ret = __get_inode_info(clone_root->root, path, clone_root->ino,
5213                                &clone_src_i_size, NULL, NULL, NULL, NULL, NULL);
5214         btrfs_release_path(path);
5215         if (ret < 0)
5216                 goto out;
5217 
5218         /*
5219          * We can't send a clone operation for the entire range if we find
5220          * extent items in the respective range in the source file that
5221          * refer to different extents or if we find holes.
5222          * So check for that and do a mix of clone and regular write/copy
5223          * operations if needed.
5224          *
5225          * Example:
5226          *
5227          * mkfs.btrfs -f /dev/sda
5228          * mount /dev/sda /mnt
5229          * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo
5230          * cp --reflink=always /mnt/foo /mnt/bar
5231          * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo
5232          * btrfs subvolume snapshot -r /mnt /mnt/snap
5233          *
5234          * If when we send the snapshot and we are processing file bar (which
5235          * has a higher inode number than foo) we blindly send a clone operation
5236          * for the [0, 100K[ range from foo to bar, the receiver ends up getting
5237          * a file bar that matches the content of file foo - iow, doesn't match
5238          * the content from bar in the original filesystem.
5239          */
5240         key.objectid = clone_root->ino;
5241         key.type = BTRFS_EXTENT_DATA_KEY;
5242         key.offset = clone_root->offset;
5243         ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0);
5244         if (ret < 0)
5245                 goto out;
5246         if (ret > 0 && path->slots[0] > 0) {
5247                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
5248                 if (key.objectid == clone_root->ino &&
5249                     key.type == BTRFS_EXTENT_DATA_KEY)
5250                         path->slots[0]--;
5251         }
5252 
5253         while (true) {
5254                 struct extent_buffer *leaf = path->nodes[0];
5255                 int slot = path->slots[0];
5256                 struct btrfs_file_extent_item *ei;
5257                 u8 type;
5258                 u64 ext_len;
5259                 u64 clone_len;
5260                 u64 clone_data_offset;
5261 
5262                 if (slot >= btrfs_header_nritems(leaf)) {
5263                         ret = btrfs_next_leaf(clone_root->root, path);
5264                         if (ret < 0)
5265                                 goto out;
5266                         else if (ret > 0)
5267                                 break;
5268                         continue;
5269                 }
5270 
5271                 btrfs_item_key_to_cpu(leaf, &key, slot);
5272 
5273                 /*
5274                  * We might have an implicit trailing hole (NO_HOLES feature
5275                  * enabled). We deal with it after leaving this loop.
5276                  */
5277                 if (key.objectid != clone_root->ino ||
5278                     key.type != BTRFS_EXTENT_DATA_KEY)
5279                         break;
5280 
5281                 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5282                 type = btrfs_file_extent_type(leaf, ei);
5283                 if (type == BTRFS_FILE_EXTENT_INLINE) {
5284                         ext_len = btrfs_file_extent_ram_bytes(leaf, ei);
5285                         ext_len = PAGE_ALIGN(ext_len);
5286                 } else {
5287                         ext_len = btrfs_file_extent_num_bytes(leaf, ei);
5288                 }
5289 
5290                 if (key.offset + ext_len <= clone_root->offset)
5291                         goto next;
5292 
5293                 if (key.offset > clone_root->offset) {
5294                         /* Implicit hole, NO_HOLES feature enabled. */
5295                         u64 hole_len = key.offset - clone_root->offset;
5296 
5297                         if (hole_len > len)
5298                                 hole_len = len;
5299                         ret = send_extent_data(sctx, offset, hole_len);
5300                         if (ret < 0)
5301                                 goto out;
5302 
5303                         len -= hole_len;
5304                         if (len == 0)
5305                                 break;
5306                         offset += hole_len;
5307                         clone_root->offset += hole_len;
5308                         data_offset += hole_len;
5309                 }
5310 
5311                 if (key.offset >= clone_root->offset + len)
5312                         break;
5313 
5314                 if (key.offset >= clone_src_i_size)
5315                         break;
5316 
5317                 if (key.offset + ext_len > clone_src_i_size)
5318                         ext_len = clone_src_i_size - key.offset;
5319 
5320                 clone_data_offset = btrfs_file_extent_offset(leaf, ei);
5321                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) {
5322                         clone_root->offset = key.offset;
5323                         if (clone_data_offset < data_offset &&
5324                                 clone_data_offset + ext_len > data_offset) {
5325                                 u64 extent_offset;
5326 
5327                                 extent_offset = data_offset - clone_data_offset;
5328                                 ext_len -= extent_offset;
5329                                 clone_data_offset += extent_offset;
5330                                 clone_root->offset += extent_offset;
5331                         }
5332                 }
5333 
5334                 clone_len = min_t(u64, ext_len, len);
5335 
5336                 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte &&
5337                     clone_data_offset == data_offset) {
5338                         const u64 src_end = clone_root->offset + clone_len;
5339                         const u64 sectorsize = SZ_64K;
5340 
5341                         /*
5342                          * We can't clone the last block, when its size is not
5343                          * sector size aligned, into the middle of a file. If we
5344                          * do so, the receiver will get a failure (-EINVAL) when
5345                          * trying to clone or will silently corrupt the data in
5346                          * the destination file if it's on a kernel without the
5347                          * fix introduced by commit ac765f83f1397646
5348                          * ("Btrfs: fix data corruption due to cloning of eof
5349                          * block).
5350                          *
5351                          * So issue a clone of the aligned down range plus a
5352                          * regular write for the eof block, if we hit that case.
5353                          *
5354                          * Also, we use the maximum possible sector size, 64K,
5355                          * because we don't know what's the sector size of the
5356                          * filesystem that receives the stream, so we have to
5357                          * assume the largest possible sector size.
5358                          */
5359                         if (src_end == clone_src_i_size &&
5360                             !IS_ALIGNED(src_end, sectorsize) &&
5361                             offset + clone_len < sctx->cur_inode_size) {
5362                                 u64 slen;
5363 
5364                                 slen = ALIGN_DOWN(src_end - clone_root->offset,
5365                                                   sectorsize);
5366                                 if (slen > 0) {
5367                                         ret = send_clone(sctx, offset, slen,
5368                                                          clone_root);
5369                                         if (ret < 0)
5370                                                 goto out;
5371                                 }
5372                                 ret = send_extent_data(sctx, offset + slen,
5373                                                        clone_len - slen);
5374                         } else {
5375                                 ret = send_clone(sctx, offset, clone_len,
5376                                                  clone_root);
5377                         }
5378                 } else {
5379                         ret = send_extent_data(sctx, offset, clone_len);
5380                 }
5381 
5382                 if (ret < 0)
5383                         goto out;
5384 
5385                 len -= clone_len;
5386                 if (len == 0)
5387                         break;
5388                 offset += clone_len;
5389                 clone_root->offset += clone_len;
5390                 data_offset += clone_len;
5391 next:
5392                 path->slots[0]++;
5393         }
5394 
5395         if (len > 0)
5396                 ret = send_extent_data(sctx, offset, len);
5397         else
5398                 ret = 0;
5399 out:
5400         btrfs_free_path(path);
5401         return ret;
5402 }
5403 
5404 static int send_write_or_clone(struct send_ctx *sctx,
5405                                struct btrfs_path *path,
5406                                struct btrfs_key *key,
5407                                struct clone_root *clone_root)
5408 {
5409         int ret = 0;
5410         struct btrfs_file_extent_item *ei;
5411         u64 offset = key->offset;
5412         u64 len;
5413         u8 type;
5414         u64 bs = sctx->send_root->fs_info->sb->s_blocksize;
5415 
5416         ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5417                         struct btrfs_file_extent_item);
5418         type = btrfs_file_extent_type(path->nodes[0], ei);
5419         if (type == BTRFS_FILE_EXTENT_INLINE) {
5420                 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei);
5421                 /*
5422                  * it is possible the inline item won't cover the whole page,
5423                  * but there may be items after this page.  Make
5424                  * sure to send the whole thing
5425                  */
5426                 len = PAGE_ALIGN(len);
5427         } else {
5428                 len = btrfs_file_extent_num_bytes(path->nodes[0], ei);
5429         }
5430 
5431         if (offset >= sctx->cur_inode_size) {
5432                 ret = 0;
5433                 goto out;
5434         }
5435         if (offset + len > sctx->cur_inode_size)
5436                 len = sctx->cur_inode_size - offset;
5437         if (len == 0) {
5438                 ret = 0;
5439                 goto out;
5440         }
5441 
5442         if (clone_root && IS_ALIGNED(offset + len, bs)) {
5443                 u64 disk_byte;
5444                 u64 data_offset;
5445 
5446                 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei);
5447                 data_offset = btrfs_file_extent_offset(path->nodes[0], ei);
5448                 ret = clone_range(sctx, clone_root, disk_byte, data_offset,
5449                                   offset, len);
5450         } else {
5451                 ret = send_extent_data(sctx, offset, len);
5452         }
5453         sctx->cur_inode_next_write_offset = offset + len;
5454 out:
5455         return ret;
5456 }
5457 
5458 static int is_extent_unchanged(struct send_ctx *sctx,
5459                                struct btrfs_path *left_path,
5460                                struct btrfs_key *ekey)
5461 {
5462         int ret = 0;
5463         struct btrfs_key key;
5464         struct btrfs_path *path = NULL;
5465         struct extent_buffer *eb;
5466         int slot;
5467         struct btrfs_key found_key;
5468         struct btrfs_file_extent_item *ei;
5469         u64 left_disknr;
5470         u64 right_disknr;
5471         u64 left_offset;
5472         u64 right_offset;
5473         u64 left_offset_fixed;
5474         u64 left_len;
5475         u64 right_len;
5476         u64 left_gen;
5477         u64 right_gen;
5478         u8 left_type;
5479         u8 right_type;
5480 
5481         path = alloc_path_for_send();
5482         if (!path)
5483                 return -ENOMEM;
5484 
5485         eb = left_path->nodes[0];
5486         slot = left_path->slots[0];
5487         ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5488         left_type = btrfs_file_extent_type(eb, ei);
5489 
5490         if (left_type != BTRFS_FILE_EXTENT_REG) {
5491                 ret = 0;
5492                 goto out;
5493         }
5494         left_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5495         left_len = btrfs_file_extent_num_bytes(eb, ei);
5496         left_offset = btrfs_file_extent_offset(eb, ei);
5497         left_gen = btrfs_file_extent_generation(eb, ei);
5498 
5499         /*
5500          * Following comments will refer to these graphics. L is the left
5501          * extents which we are checking at the moment. 1-8 are the right
5502          * extents that we iterate.
5503          *
5504          *       |-----L-----|
5505          * |-1-|-2a-|-3-|-4-|-5-|-6-|
5506          *
5507          *       |-----L-----|
5508          * |--1--|-2b-|...(same as above)
5509          *
5510          * Alternative situation. Happens on files where extents got split.
5511          *       |-----L-----|
5512          * |-----------7-----------|-6-|
5513          *
5514          * Alternative situation. Happens on files which got larger.
5515          *       |-----L-----|
5516          * |-8-|
5517          * Nothing follows after 8.
5518          */
5519 
5520         key.objectid = ekey->objectid;
5521         key.type = BTRFS_EXTENT_DATA_KEY;
5522         key.offset = ekey->offset;
5523         ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0);
5524         if (ret < 0)
5525                 goto out;
5526         if (ret) {
5527                 ret = 0;
5528                 goto out;
5529         }
5530 
5531         /*
5532          * Handle special case where the right side has no extents at all.
5533          */
5534         eb = path->nodes[0];
5535         slot = path->slots[0];
5536         btrfs_item_key_to_cpu(eb, &found_key, slot);
5537         if (found_key.objectid != key.objectid ||
5538             found_key.type != key.type) {
5539                 /* If we're a hole then just pretend nothing changed */
5540                 ret = (left_disknr) ? 0 : 1;
5541                 goto out;
5542         }
5543 
5544         /*
5545          * We're now on 2a, 2b or 7.
5546          */
5547         key = found_key;
5548         while (key.offset < ekey->offset + left_len) {
5549                 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
5550                 right_type = btrfs_file_extent_type(eb, ei);
5551                 if (right_type != BTRFS_FILE_EXTENT_REG &&
5552                     right_type != BTRFS_FILE_EXTENT_INLINE) {
5553                         ret = 0;
5554                         goto out;
5555                 }
5556 
5557                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5558                         right_len = btrfs_file_extent_ram_bytes(eb, ei);
5559                         right_len = PAGE_ALIGN(right_len);
5560                 } else {
5561                         right_len = btrfs_file_extent_num_bytes(eb, ei);
5562                 }
5563 
5564                 /*
5565                  * Are we at extent 8? If yes, we know the extent is changed.
5566                  * This may only happen on the first iteration.
5567                  */
5568                 if (found_key.offset + right_len <= ekey->offset) {
5569                         /* If we're a hole just pretend nothing changed */
5570                         ret = (left_disknr) ? 0 : 1;
5571                         goto out;
5572                 }
5573 
5574                 /*
5575                  * We just wanted to see if when we have an inline extent, what
5576                  * follows it is a regular extent (wanted to check the above
5577                  * condition for inline extents too). This should normally not
5578                  * happen but it's possible for example when we have an inline
5579                  * compressed extent representing data with a size matching
5580                  * the page size (currently the same as sector size).
5581                  */
5582                 if (right_type == BTRFS_FILE_EXTENT_INLINE) {
5583                         ret = 0;
5584                         goto out;
5585                 }
5586 
5587                 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei);
5588                 right_offset = btrfs_file_extent_offset(eb, ei);
5589                 right_gen = btrfs_file_extent_generation(eb, ei);
5590 
5591                 left_offset_fixed = left_offset;
5592                 if (key.offset < ekey->offset) {
5593                         /* Fix the right offset for 2a and 7. */
5594                         right_offset += ekey->offset - key.offset;
5595                 } else {
5596                         /* Fix the left offset for all behind 2a and 2b */
5597                         left_offset_fixed += key.offset - ekey->offset;
5598                 }
5599 
5600                 /*
5601                  * Check if we have the same extent.
5602                  */
5603                 if (left_disknr != right_disknr ||
5604                     left_offset_fixed != right_offset ||
5605                     left_gen != right_gen) {
5606                         ret = 0;
5607                         goto out;
5608                 }
5609 
5610                 /*
5611                  * Go to the next extent.
5612                  */
5613                 ret = btrfs_next_item(sctx->parent_root, path);
5614                 if (ret < 0)
5615                         goto out;
5616                 if (!ret) {
5617                         eb = path->nodes[0];
5618                         slot = path->slots[0];
5619                         btrfs_item_key_to_cpu(eb, &found_key, slot);
5620                 }
5621                 if (ret || found_key.objectid != key.objectid ||
5622                     found_key.type != key.type) {
5623                         key.offset += right_len;
5624                         break;
5625                 }
5626                 if (found_key.offset != key.offset + right_len) {
5627                         ret = 0;
5628                         goto out;
5629                 }
5630                 key = found_key;
5631         }
5632 
5633         /*
5634          * We're now behind the left extent (treat as unchanged) or at the end
5635          * of the right side (treat as changed).
5636          */
5637         if (key.offset >= ekey->offset + left_len)
5638                 ret = 1;
5639         else
5640                 ret = 0;
5641 
5642 
5643 out:
5644         btrfs_free_path(path);
5645         return ret;
5646 }
5647 
5648 static int get_last_extent(struct send_ctx *sctx, u64 offset)
5649 {
5650         struct btrfs_path *path;
5651         struct btrfs_root *root = sctx->send_root;
5652         struct btrfs_key key;
5653         int ret;
5654 
5655         path = alloc_path_for_send();
5656         if (!path)
5657                 return -ENOMEM;
5658 
5659         sctx->cur_inode_last_extent = 0;
5660 
5661         key.objectid = sctx->cur_ino;
5662         key.type = BTRFS_EXTENT_DATA_KEY;
5663         key.offset = offset;
5664         ret = btrfs_search_slot_for_read(root, &key, path, 0, 1);
5665         if (ret < 0)
5666                 goto out;
5667         ret = 0;
5668         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
5669         if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY)
5670                 goto out;
5671 
5672         sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5673 out:
5674         btrfs_free_path(path);
5675         return ret;
5676 }
5677 
5678 static int range_is_hole_in_parent(struct send_ctx *sctx,
5679                                    const u64 start,
5680                                    const u64 end)
5681 {
5682         struct btrfs_path *path;
5683         struct btrfs_key key;
5684         struct btrfs_root *root = sctx->parent_root;
5685         u64 search_start = start;
5686         int ret;
5687 
5688         path = alloc_path_for_send();
5689         if (!path)
5690                 return -ENOMEM;
5691 
5692         key.objectid = sctx->cur_ino;
5693         key.type = BTRFS_EXTENT_DATA_KEY;
5694         key.offset = search_start;
5695         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5696         if (ret < 0)
5697                 goto out;
5698         if (ret > 0 && path->slots[0] > 0)
5699                 path->slots[0]--;
5700 
5701         while (search_start < end) {
5702                 struct extent_buffer *leaf = path->nodes[0];
5703                 int slot = path->slots[0];
5704                 struct btrfs_file_extent_item *fi;
5705                 u64 extent_end;
5706 
5707                 if (slot >= btrfs_header_nritems(leaf)) {
5708                         ret = btrfs_next_leaf(root, path);
5709                         if (ret < 0)
5710                                 goto out;
5711                         else if (ret > 0)
5712                                 break;
5713                         continue;
5714                 }
5715 
5716                 btrfs_item_key_to_cpu(leaf, &key, slot);
5717                 if (key.objectid < sctx->cur_ino ||
5718                     key.type < BTRFS_EXTENT_DATA_KEY)
5719                         goto next;
5720                 if (key.objectid > sctx->cur_ino ||
5721                     key.type > BTRFS_EXTENT_DATA_KEY ||
5722                     key.offset >= end)
5723                         break;
5724 
5725                 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5726                 extent_end = btrfs_file_extent_end(path);
5727                 if (extent_end <= start)
5728                         goto next;
5729                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) {
5730                         search_start = extent_end;
5731                         goto next;
5732                 }
5733                 ret = 0;
5734                 goto out;
5735 next:
5736                 path->slots[0]++;
5737         }
5738         ret = 1;
5739 out:
5740         btrfs_free_path(path);
5741         return ret;
5742 }
5743 
5744 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path,
5745                            struct btrfs_key *key)
5746 {
5747         int ret = 0;
5748 
5749         if (sctx->cur_ino != key->objectid || !need_send_hole(sctx))
5750                 return 0;
5751 
5752         if (sctx->cur_inode_last_extent == (u64)-1) {
5753                 ret = get_last_extent(sctx, key->offset - 1);
5754                 if (ret)
5755                         return ret;
5756         }
5757 
5758         if (path->slots[0] == 0 &&
5759             sctx->cur_inode_last_extent < key->offset) {
5760                 /*
5761                  * We might have skipped entire leafs that contained only
5762                  * file extent items for our current inode. These leafs have
5763                  * a generation number smaller (older) than the one in the
5764                  * current leaf and the leaf our last extent came from, and
5765                  * are located between these 2 leafs.
5766                  */
5767                 ret = get_last_extent(sctx, key->offset - 1);
5768                 if (ret)
5769                         return ret;
5770         }
5771 
5772         if (sctx->cur_inode_last_extent < key->offset) {
5773                 ret = range_is_hole_in_parent(sctx,
5774                                               sctx->cur_inode_last_extent,
5775                                               key->offset);
5776                 if (ret < 0)
5777                         return ret;
5778                 else if (ret == 0)
5779                         ret = send_hole(sctx, key->offset);
5780                 else
5781                         ret = 0;
5782         }
5783         sctx->cur_inode_last_extent = btrfs_file_extent_end(path);
5784         return ret;
5785 }
5786 
5787 static int process_extent(struct send_ctx *sctx,
5788                           struct btrfs_path *path,
5789                           struct btrfs_key *key)
5790 {
5791         struct clone_root *found_clone = NULL;
5792         int ret = 0;
5793 
5794         if (S_ISLNK(sctx->cur_inode_mode))
5795                 return 0;
5796 
5797         if (sctx->parent_root && !sctx->cur_inode_new) {
5798                 ret = is_extent_unchanged(sctx, path, key);
5799                 if (ret < 0)
5800                         goto out;
5801                 if (ret) {
5802                         ret = 0;
5803                         goto out_hole;
5804                 }
5805         } else {
5806                 struct btrfs_file_extent_item *ei;
5807                 u8 type;
5808 
5809                 ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
5810                                     struct btrfs_file_extent_item);
5811                 type = btrfs_file_extent_type(path->nodes[0], ei);
5812                 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
5813                     type == BTRFS_FILE_EXTENT_REG) {
5814                         /*
5815                          * The send spec does not have a prealloc command yet,
5816                          * so just leave a hole for prealloc'ed extents until
5817                          * we have enough commands queued up to justify rev'ing
5818                          * the send spec.
5819                          */
5820                         if (type == BTRFS_FILE_EXTENT_PREALLOC) {
5821                                 ret = 0;
5822                                 goto out;
5823                         }
5824 
5825                         /* Have a hole, just skip it. */
5826                         if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) {
5827                                 ret = 0;
5828                                 goto out;
5829                         }
5830                 }
5831         }
5832 
5833         ret = find_extent_clone(sctx, path, key->objectid, key->offset,
5834                         sctx->cur_inode_size, &found_clone);
5835         if (ret != -ENOENT && ret < 0)
5836                 goto out;
5837 
5838         ret = send_write_or_clone(sctx, path, key, found_clone);
5839         if (ret)
5840                 goto out;
5841 out_hole:
5842         ret = maybe_send_hole(sctx, path, key);
5843 out:
5844         return ret;
5845 }
5846 
5847 static int process_all_extents(struct send_ctx *sctx)
5848 {
5849         int ret;
5850         struct btrfs_root *root;
5851         struct btrfs_path *path;
5852         struct btrfs_key key;
5853         struct btrfs_key found_key;
5854         struct extent_buffer *eb;
5855         int slot;
5856 
5857         root = sctx->send_root;
5858         path = alloc_path_for_send();
5859         if (!path)
5860                 return -ENOMEM;
5861 
5862         key.objectid = sctx->cmp_key->objectid;
5863         key.type = BTRFS_EXTENT_DATA_KEY;
5864         key.offset = 0;
5865         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5866         if (ret < 0)
5867                 goto out;
5868 
5869         while (1) {
5870                 eb = path->nodes[0];
5871                 slot = path->slots[0];
5872 
5873                 if (slot >= btrfs_header_nritems(eb)) {
5874                         ret = btrfs_next_leaf(root, path);
5875                         if (ret < 0) {
5876                                 goto out;
5877                         } else if (ret > 0) {
5878                                 ret = 0;
5879                                 break;
5880                         }
5881                         continue;
5882                 }
5883 
5884                 btrfs_item_key_to_cpu(eb, &found_key, slot);
5885 
5886                 if (found_key.objectid != key.objectid ||
5887                     found_key.type != key.type) {
5888                         ret = 0;
5889                         goto out;
5890                 }
5891 
5892                 ret = process_extent(sctx, path, &found_key);
5893                 if (ret < 0)
5894                         goto out;
5895 
5896                 path->slots[0]++;
5897         }
5898 
5899 out:
5900         btrfs_free_path(path);
5901         return ret;
5902 }
5903 
5904 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end,
5905                                            int *pending_move,
5906                                            int *refs_processed)
5907 {
5908         int ret = 0;
5909 
5910         if (sctx->cur_ino == 0)
5911                 goto out;
5912         if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
5913             sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY)
5914                 goto out;
5915         if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
5916                 goto out;
5917 
5918         ret = process_recorded_refs(sctx, pending_move);
5919         if (ret < 0)
5920                 goto out;
5921 
5922         *refs_processed = 1;
5923 out:
5924         return ret;
5925 }
5926 
5927 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end)
5928 {
5929         int ret = 0;
5930         u64 left_mode;
5931         u64 left_uid;
5932         u64 left_gid;
5933         u64 right_mode;
5934         u64 right_uid;
5935         u64 right_gid;
5936         int need_chmod = 0;
5937         int need_chown = 0;
5938         int need_truncate = 1;
5939         int pending_move = 0;
5940         int refs_processed = 0;
5941 
5942         if (sctx->ignore_cur_inode)
5943                 return 0;
5944 
5945         ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move,
5946                                               &refs_processed);
5947         if (ret < 0)
5948                 goto out;
5949 
5950         /*
5951          * We have processed the refs and thus need to advance send_progress.
5952          * Now, calls to get_cur_xxx will take the updated refs of the current
5953          * inode into account.
5954          *
5955          * On the other hand, if our current inode is a directory and couldn't
5956          * be moved/renamed because its parent was renamed/moved too and it has
5957          * a higher inode number, we can only move/rename our current inode
5958          * after we moved/renamed its parent. Therefore in this case operate on
5959          * the old path (pre move/rename) of our current inode, and the
5960          * move/rename will be performed later.
5961          */
5962         if (refs_processed && !pending_move)
5963                 sctx->send_progress = sctx->cur_ino + 1;
5964 
5965         if (sctx->cur_ino == 0 || sctx->cur_inode_deleted)
5966                 goto out;
5967         if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino)
5968                 goto out;
5969 
5970         ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL,
5971                         &left_mode, &left_uid, &left_gid, NULL);
5972         if (ret < 0)
5973                 goto out;
5974 
5975         if (!sctx->parent_root || sctx->cur_inode_new) {
5976                 need_chown = 1;
5977                 if (!S_ISLNK(sctx->cur_inode_mode))
5978                         need_chmod = 1;
5979                 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size)
5980                         need_truncate = 0;
5981         } else {
5982                 u64 old_size;
5983 
5984                 ret = get_inode_info(sctx->parent_root, sctx->cur_ino,
5985                                 &old_size, NULL, &right_mode, &right_uid,
5986                                 &right_gid, NULL);
5987                 if (ret < 0)
5988                         goto out;
5989 
5990                 if (left_uid != right_uid || left_gid != right_gid)
5991                         need_chown = 1;
5992                 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode)
5993                         need_chmod = 1;
5994                 if ((old_size == sctx->cur_inode_size) ||
5995                     (sctx->cur_inode_size > old_size &&
5996                      sctx->cur_inode_next_write_offset == sctx->cur_inode_size))
5997                         need_truncate = 0;
5998         }
5999 
6000         if (S_ISREG(sctx->cur_inode_mode)) {
6001                 if (need_send_hole(sctx)) {
6002                         if (sctx->cur_inode_last_extent == (u64)-1 ||
6003                             sctx->cur_inode_last_extent <
6004                             sctx->cur_inode_size) {
6005                                 ret = get_last_extent(sctx, (u64)-1);
6006                                 if (ret)
6007                                         goto out;
6008                         }
6009                         if (sctx->cur_inode_last_extent <
6010                             sctx->cur_inode_size) {
6011                                 ret = send_hole(sctx, sctx->cur_inode_size);
6012                                 if (ret)
6013                                         goto out;
6014                         }
6015                 }
6016                 if (need_truncate) {
6017                         ret = send_truncate(sctx, sctx->cur_ino,
6018                                             sctx->cur_inode_gen,
6019                                             sctx->cur_inode_size);
6020                         if (ret < 0)
6021                                 goto out;
6022                 }
6023         }
6024 
6025         if (need_chown) {
6026                 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen,
6027                                 left_uid, left_gid);
6028                 if (ret < 0)
6029                         goto out;
6030         }
6031         if (need_c