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

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

Version: ~ [ linux-5.19-rc8 ] ~ [ linux-5.18.14 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.57 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.133 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.207 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.253 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.289 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.324 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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