1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2012 Alexander Block. All rights reserved. 4 */ 5 6 #include <linux/bsearch.h> 7 #include <linux/fs.h> 8 #include <linux/file.h> 9 #include <linux/sort.h> 10 #include <linux/mount.h> 11 #include <linux/xattr.h> 12 #include <linux/posix_acl_xattr.h> 13 #include <linux/radix-tree.h> 14 #include <linux/vmalloc.h> 15 #include <linux/string.h> 16 #include <linux/compat.h> 17 #include <linux/crc32c.h> 18 19 #include "send.h" 20 #include "backref.h" 21 #include "locking.h" 22 #include "disk-io.h" 23 #include "btrfs_inode.h" 24 #include "transaction.h" 25 #include "compression.h" 26 #include "xattr.h" 27 28 /* 29 * Maximum number of references an extent can have in order for us to attempt to 30 * issue clone operations instead of write operations. This currently exists to 31 * avoid hitting limitations of the backreference walking code (taking a lot of 32 * time and using too much memory for extents with large number of references). 33 */ 34 #define SEND_MAX_EXTENT_REFS 64 35 36 /* 37 * A fs_path is a helper to dynamically build path names with unknown size. 38 * It reallocates the internal buffer on demand. 39 * It allows fast adding of path elements on the right side (normal path) and 40 * fast adding to the left side (reversed path). A reversed path can also be 41 * unreversed if needed. 42 */ 43 struct fs_path { 44 union { 45 struct { 46 char *start; 47 char *end; 48 49 char *buf; 50 unsigned short buf_len:15; 51 unsigned short reversed:1; 52 char inline_buf[]; 53 }; 54 /* 55 * Average path length does not exceed 200 bytes, we'll have 56 * better packing in the slab and higher chance to satisfy 57 * a allocation later during send. 58 */ 59 char pad[256]; 60 }; 61 }; 62 #define FS_PATH_INLINE_SIZE \ 63 (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf)) 64 65 66 /* reused for each extent */ 67 struct clone_root { 68 struct btrfs_root *root; 69 u64 ino; 70 u64 offset; 71 72 u64 found_refs; 73 }; 74 75 #define SEND_CTX_MAX_NAME_CACHE_SIZE 128 76 #define SEND_CTX_NAME_CACHE_CLEAN_SIZE (SEND_CTX_MAX_NAME_CACHE_SIZE * 2) 77 78 struct send_ctx { 79 struct file *send_filp; 80 loff_t send_off; 81 char *send_buf; 82 u32 send_size; 83 u32 send_max_size; 84 u64 total_send_size; 85 u64 cmd_send_size[BTRFS_SEND_C_MAX + 1]; 86 u64 flags; /* 'flags' member of btrfs_ioctl_send_args is u64 */ 87 88 struct btrfs_root *send_root; 89 struct btrfs_root *parent_root; 90 struct clone_root *clone_roots; 91 int clone_roots_cnt; 92 93 /* current state of the compare_tree call */ 94 struct btrfs_path *left_path; 95 struct btrfs_path *right_path; 96 struct btrfs_key *cmp_key; 97 98 /* 99 * infos of the currently processed inode. In case of deleted inodes, 100 * these are the values from the deleted inode. 101 */ 102 u64 cur_ino; 103 u64 cur_inode_gen; 104 int cur_inode_new; 105 int cur_inode_new_gen; 106 int cur_inode_deleted; 107 u64 cur_inode_size; 108 u64 cur_inode_mode; 109 u64 cur_inode_rdev; 110 u64 cur_inode_last_extent; 111 u64 cur_inode_next_write_offset; 112 bool ignore_cur_inode; 113 114 u64 send_progress; 115 116 struct list_head new_refs; 117 struct list_head deleted_refs; 118 119 struct radix_tree_root name_cache; 120 struct list_head name_cache_list; 121 int name_cache_size; 122 123 struct file_ra_state ra; 124 125 /* 126 * We process inodes by their increasing order, so if before an 127 * incremental send we reverse the parent/child relationship of 128 * directories such that a directory with a lower inode number was 129 * the parent of a directory with a higher inode number, and the one 130 * becoming the new parent got renamed too, we can't rename/move the 131 * directory with lower inode number when we finish processing it - we 132 * must process the directory with higher inode number first, then 133 * rename/move it and then rename/move the directory with lower inode 134 * number. Example follows. 135 * 136 * Tree state when the first send was performed: 137 * 138 * . 139 * |-- a (ino 257) 140 * |-- b (ino 258) 141 * | 142 * | 143 * |-- c (ino 259) 144 * | |-- d (ino 260) 145 * | 146 * |-- c2 (ino 261) 147 * 148 * Tree state when the second (incremental) send is performed: 149 * 150 * . 151 * |-- a (ino 257) 152 * |-- b (ino 258) 153 * |-- c2 (ino 261) 154 * |-- d2 (ino 260) 155 * |-- cc (ino 259) 156 * 157 * The sequence of steps that lead to the second state was: 158 * 159 * mv /a/b/c/d /a/b/c2/d2 160 * mv /a/b/c /a/b/c2/d2/cc 161 * 162 * "c" has lower inode number, but we can't move it (2nd mv operation) 163 * before we move "d", which has higher inode number. 164 * 165 * So we just memorize which move/rename operations must be performed 166 * later when their respective parent is processed and moved/renamed. 167 */ 168 169 /* Indexed by parent directory inode number. */ 170 struct rb_root pending_dir_moves; 171 172 /* 173 * Reverse index, indexed by the inode number of a directory that 174 * is waiting for the move/rename of its immediate parent before its 175 * own move/rename can be performed. 176 */ 177 struct rb_root waiting_dir_moves; 178 179 /* 180 * A directory that is going to be rm'ed might have a child directory 181 * which is in the pending directory moves index above. In this case, 182 * the directory can only be removed after the move/rename of its child 183 * is performed. Example: 184 * 185 * Parent snapshot: 186 * 187 * . (ino 256) 188 * |-- a/ (ino 257) 189 * |-- b/ (ino 258) 190 * |-- c/ (ino 259) 191 * | |-- x/ (ino 260) 192 * | 193 * |-- y/ (ino 261) 194 * 195 * Send snapshot: 196 * 197 * . (ino 256) 198 * |-- a/ (ino 257) 199 * |-- b/ (ino 258) 200 * |-- YY/ (ino 261) 201 * |-- x/ (ino 260) 202 * 203 * Sequence of steps that lead to the send snapshot: 204 * rm -f /a/b/c/foo.txt 205 * mv /a/b/y /a/b/YY 206 * mv /a/b/c/x /a/b/YY 207 * rmdir /a/b/c 208 * 209 * When the child is processed, its move/rename is delayed until its 210 * parent is processed (as explained above), but all other operations 211 * like update utimes, chown, chgrp, etc, are performed and the paths 212 * that it uses for those operations must use the orphanized name of 213 * its parent (the directory we're going to rm later), so we need to 214 * memorize that name. 215 * 216 * Indexed by the inode number of the directory to be deleted. 217 */ 218 struct rb_root orphan_dirs; 219 }; 220 221 struct pending_dir_move { 222 struct rb_node node; 223 struct list_head list; 224 u64 parent_ino; 225 u64 ino; 226 u64 gen; 227 struct list_head update_refs; 228 }; 229 230 struct waiting_dir_move { 231 struct rb_node node; 232 u64 ino; 233 /* 234 * There might be some directory that could not be removed because it 235 * was waiting for this directory inode to be moved first. Therefore 236 * after this directory is moved, we can try to rmdir the ino rmdir_ino. 237 */ 238 u64 rmdir_ino; 239 u64 rmdir_gen; 240 bool orphanized; 241 }; 242 243 struct orphan_dir_info { 244 struct rb_node node; 245 u64 ino; 246 u64 gen; 247 u64 last_dir_index_offset; 248 }; 249 250 struct name_cache_entry { 251 struct list_head list; 252 /* 253 * radix_tree has only 32bit entries but we need to handle 64bit inums. 254 * We use the lower 32bit of the 64bit inum to store it in the tree. If 255 * more then one inum would fall into the same entry, we use radix_list 256 * to store the additional entries. radix_list is also used to store 257 * entries where two entries have the same inum but different 258 * generations. 259 */ 260 struct list_head radix_list; 261 u64 ino; 262 u64 gen; 263 u64 parent_ino; 264 u64 parent_gen; 265 int ret; 266 int need_later_update; 267 int name_len; 268 char name[]; 269 }; 270 271 #define ADVANCE 1 272 #define ADVANCE_ONLY_NEXT -1 273 274 enum btrfs_compare_tree_result { 275 BTRFS_COMPARE_TREE_NEW, 276 BTRFS_COMPARE_TREE_DELETED, 277 BTRFS_COMPARE_TREE_CHANGED, 278 BTRFS_COMPARE_TREE_SAME, 279 }; 280 281 __cold 282 static void inconsistent_snapshot_error(struct send_ctx *sctx, 283 enum btrfs_compare_tree_result result, 284 const char *what) 285 { 286 const char *result_string; 287 288 switch (result) { 289 case BTRFS_COMPARE_TREE_NEW: 290 result_string = "new"; 291 break; 292 case BTRFS_COMPARE_TREE_DELETED: 293 result_string = "deleted"; 294 break; 295 case BTRFS_COMPARE_TREE_CHANGED: 296 result_string = "updated"; 297 break; 298 case BTRFS_COMPARE_TREE_SAME: 299 ASSERT(0); 300 result_string = "unchanged"; 301 break; 302 default: 303 ASSERT(0); 304 result_string = "unexpected"; 305 } 306 307 btrfs_err(sctx->send_root->fs_info, 308 "Send: inconsistent snapshot, found %s %s for inode %llu without updated inode item, send root is %llu, parent root is %llu", 309 result_string, what, sctx->cmp_key->objectid, 310 sctx->send_root->root_key.objectid, 311 (sctx->parent_root ? 312 sctx->parent_root->root_key.objectid : 0)); 313 } 314 315 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino); 316 317 static struct waiting_dir_move * 318 get_waiting_dir_move(struct send_ctx *sctx, u64 ino); 319 320 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen); 321 322 static int need_send_hole(struct send_ctx *sctx) 323 { 324 return (sctx->parent_root && !sctx->cur_inode_new && 325 !sctx->cur_inode_new_gen && !sctx->cur_inode_deleted && 326 S_ISREG(sctx->cur_inode_mode)); 327 } 328 329 static void fs_path_reset(struct fs_path *p) 330 { 331 if (p->reversed) { 332 p->start = p->buf + p->buf_len - 1; 333 p->end = p->start; 334 *p->start = 0; 335 } else { 336 p->start = p->buf; 337 p->end = p->start; 338 *p->start = 0; 339 } 340 } 341 342 static struct fs_path *fs_path_alloc(void) 343 { 344 struct fs_path *p; 345 346 p = kmalloc(sizeof(*p), GFP_KERNEL); 347 if (!p) 348 return NULL; 349 p->reversed = 0; 350 p->buf = p->inline_buf; 351 p->buf_len = FS_PATH_INLINE_SIZE; 352 fs_path_reset(p); 353 return p; 354 } 355 356 static struct fs_path *fs_path_alloc_reversed(void) 357 { 358 struct fs_path *p; 359 360 p = fs_path_alloc(); 361 if (!p) 362 return NULL; 363 p->reversed = 1; 364 fs_path_reset(p); 365 return p; 366 } 367 368 static void fs_path_free(struct fs_path *p) 369 { 370 if (!p) 371 return; 372 if (p->buf != p->inline_buf) 373 kfree(p->buf); 374 kfree(p); 375 } 376 377 static int fs_path_len(struct fs_path *p) 378 { 379 return p->end - p->start; 380 } 381 382 static int fs_path_ensure_buf(struct fs_path *p, int len) 383 { 384 char *tmp_buf; 385 int path_len; 386 int old_buf_len; 387 388 len++; 389 390 if (p->buf_len >= len) 391 return 0; 392 393 if (len > PATH_MAX) { 394 WARN_ON(1); 395 return -ENOMEM; 396 } 397 398 path_len = p->end - p->start; 399 old_buf_len = p->buf_len; 400 401 /* 402 * First time the inline_buf does not suffice 403 */ 404 if (p->buf == p->inline_buf) { 405 tmp_buf = kmalloc(len, GFP_KERNEL); 406 if (tmp_buf) 407 memcpy(tmp_buf, p->buf, old_buf_len); 408 } else { 409 tmp_buf = krealloc(p->buf, len, GFP_KERNEL); 410 } 411 if (!tmp_buf) 412 return -ENOMEM; 413 p->buf = tmp_buf; 414 /* 415 * The real size of the buffer is bigger, this will let the fast path 416 * happen most of the time 417 */ 418 p->buf_len = ksize(p->buf); 419 420 if (p->reversed) { 421 tmp_buf = p->buf + old_buf_len - path_len - 1; 422 p->end = p->buf + p->buf_len - 1; 423 p->start = p->end - path_len; 424 memmove(p->start, tmp_buf, path_len + 1); 425 } else { 426 p->start = p->buf; 427 p->end = p->start + path_len; 428 } 429 return 0; 430 } 431 432 static int fs_path_prepare_for_add(struct fs_path *p, int name_len, 433 char **prepared) 434 { 435 int ret; 436 int new_len; 437 438 new_len = p->end - p->start + name_len; 439 if (p->start != p->end) 440 new_len++; 441 ret = fs_path_ensure_buf(p, new_len); 442 if (ret < 0) 443 goto out; 444 445 if (p->reversed) { 446 if (p->start != p->end) 447 *--p->start = '/'; 448 p->start -= name_len; 449 *prepared = p->start; 450 } else { 451 if (p->start != p->end) 452 *p->end++ = '/'; 453 *prepared = p->end; 454 p->end += name_len; 455 *p->end = 0; 456 } 457 458 out: 459 return ret; 460 } 461 462 static int fs_path_add(struct fs_path *p, const char *name, int name_len) 463 { 464 int ret; 465 char *prepared; 466 467 ret = fs_path_prepare_for_add(p, name_len, &prepared); 468 if (ret < 0) 469 goto out; 470 memcpy(prepared, name, name_len); 471 472 out: 473 return ret; 474 } 475 476 static int fs_path_add_path(struct fs_path *p, struct fs_path *p2) 477 { 478 int ret; 479 char *prepared; 480 481 ret = fs_path_prepare_for_add(p, p2->end - p2->start, &prepared); 482 if (ret < 0) 483 goto out; 484 memcpy(prepared, p2->start, p2->end - p2->start); 485 486 out: 487 return ret; 488 } 489 490 static int fs_path_add_from_extent_buffer(struct fs_path *p, 491 struct extent_buffer *eb, 492 unsigned long off, int len) 493 { 494 int ret; 495 char *prepared; 496 497 ret = fs_path_prepare_for_add(p, len, &prepared); 498 if (ret < 0) 499 goto out; 500 501 read_extent_buffer(eb, prepared, off, len); 502 503 out: 504 return ret; 505 } 506 507 static int fs_path_copy(struct fs_path *p, struct fs_path *from) 508 { 509 int ret; 510 511 p->reversed = from->reversed; 512 fs_path_reset(p); 513 514 ret = fs_path_add_path(p, from); 515 516 return ret; 517 } 518 519 520 static void fs_path_unreverse(struct fs_path *p) 521 { 522 char *tmp; 523 int len; 524 525 if (!p->reversed) 526 return; 527 528 tmp = p->start; 529 len = p->end - p->start; 530 p->start = p->buf; 531 p->end = p->start + len; 532 memmove(p->start, tmp, len + 1); 533 p->reversed = 0; 534 } 535 536 static struct btrfs_path *alloc_path_for_send(void) 537 { 538 struct btrfs_path *path; 539 540 path = btrfs_alloc_path(); 541 if (!path) 542 return NULL; 543 path->search_commit_root = 1; 544 path->skip_locking = 1; 545 path->need_commit_sem = 1; 546 return path; 547 } 548 549 static int write_buf(struct file *filp, const void *buf, u32 len, loff_t *off) 550 { 551 int ret; 552 u32 pos = 0; 553 554 while (pos < len) { 555 ret = kernel_write(filp, buf + pos, len - pos, off); 556 /* TODO handle that correctly */ 557 /*if (ret == -ERESTARTSYS) { 558 continue; 559 }*/ 560 if (ret < 0) 561 return ret; 562 if (ret == 0) { 563 return -EIO; 564 } 565 pos += ret; 566 } 567 568 return 0; 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 put_unaligned_le16(attr, &hdr->tlv_type); 582 put_unaligned_le16(len, &hdr->tlv_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, data, attrlen) \ 624 do { \ 625 ret = tlv_put(sctx, attrtype, data, attrlen); \ 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 put_unaligned_le16(cmd, &hdr->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 put_unaligned_le32(sctx->send_size - sizeof(*hdr), &hdr->len); 705 put_unaligned_le32(0, &hdr->crc); 706 707 crc = btrfs_crc32c(0, (unsigned char *)sctx->send_buf, sctx->send_size); 708 put_unaligned_le32(crc, &hdr->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[get_unaligned_le16(&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 iterate_dir_item_t iterate, void *ctx) 1006 { 1007 int ret = 0; 1008 struct extent_buffer *eb; 1009 struct btrfs_item *item; 1010 struct btrfs_dir_item *di; 1011 struct btrfs_key di_key; 1012 char *buf = NULL; 1013 int buf_len; 1014 u32 name_len; 1015 u32 data_len; 1016 u32 cur; 1017 u32 len; 1018 u32 total; 1019 int slot; 1020 int num; 1021 u8 type; 1022 1023 /* 1024 * Start with a small buffer (1 page). If later we end up needing more 1025 * space, which can happen for xattrs on a fs with a leaf size greater 1026 * then the page size, attempt to increase the buffer. Typically xattr 1027 * values are small. 1028 */ 1029 buf_len = PATH_MAX; 1030 buf = kmalloc(buf_len, GFP_KERNEL); 1031 if (!buf) { 1032 ret = -ENOMEM; 1033 goto out; 1034 } 1035 1036 eb = path->nodes[0]; 1037 slot = path->slots[0]; 1038 item = btrfs_item_nr(slot); 1039 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item); 1040 cur = 0; 1041 len = 0; 1042 total = btrfs_item_size(eb, item); 1043 1044 num = 0; 1045 while (cur < total) { 1046 name_len = btrfs_dir_name_len(eb, di); 1047 data_len = btrfs_dir_data_len(eb, di); 1048 type = btrfs_dir_type(eb, di); 1049 btrfs_dir_item_key_to_cpu(eb, di, &di_key); 1050 1051 if (type == BTRFS_FT_XATTR) { 1052 if (name_len > XATTR_NAME_MAX) { 1053 ret = -ENAMETOOLONG; 1054 goto out; 1055 } 1056 if (name_len + data_len > 1057 BTRFS_MAX_XATTR_SIZE(root->fs_info)) { 1058 ret = -E2BIG; 1059 goto out; 1060 } 1061 } else { 1062 /* 1063 * Path too long 1064 */ 1065 if (name_len + data_len > PATH_MAX) { 1066 ret = -ENAMETOOLONG; 1067 goto out; 1068 } 1069 } 1070 1071 if (name_len + data_len > buf_len) { 1072 buf_len = name_len + data_len; 1073 if (is_vmalloc_addr(buf)) { 1074 vfree(buf); 1075 buf = NULL; 1076 } else { 1077 char *tmp = krealloc(buf, buf_len, 1078 GFP_KERNEL | __GFP_NOWARN); 1079 1080 if (!tmp) 1081 kfree(buf); 1082 buf = tmp; 1083 } 1084 if (!buf) { 1085 buf = kvmalloc(buf_len, GFP_KERNEL); 1086 if (!buf) { 1087 ret = -ENOMEM; 1088 goto out; 1089 } 1090 } 1091 } 1092 1093 read_extent_buffer(eb, buf, (unsigned long)(di + 1), 1094 name_len + data_len); 1095 1096 len = sizeof(*di) + name_len + data_len; 1097 di = (struct btrfs_dir_item *)((char *)di + len); 1098 cur += len; 1099 1100 ret = iterate(num, &di_key, buf, name_len, buf + name_len, 1101 data_len, type, ctx); 1102 if (ret < 0) 1103 goto out; 1104 if (ret) { 1105 ret = 0; 1106 goto out; 1107 } 1108 1109 num++; 1110 } 1111 1112 out: 1113 kvfree(buf); 1114 return ret; 1115 } 1116 1117 static int __copy_first_ref(int num, u64 dir, int index, 1118 struct fs_path *p, void *ctx) 1119 { 1120 int ret; 1121 struct fs_path *pt = ctx; 1122 1123 ret = fs_path_copy(pt, p); 1124 if (ret < 0) 1125 return ret; 1126 1127 /* we want the first only */ 1128 return 1; 1129 } 1130 1131 /* 1132 * Retrieve the first path of an inode. If an inode has more then one 1133 * ref/hardlink, this is ignored. 1134 */ 1135 static int get_inode_path(struct btrfs_root *root, 1136 u64 ino, struct fs_path *path) 1137 { 1138 int ret; 1139 struct btrfs_key key, found_key; 1140 struct btrfs_path *p; 1141 1142 p = alloc_path_for_send(); 1143 if (!p) 1144 return -ENOMEM; 1145 1146 fs_path_reset(path); 1147 1148 key.objectid = ino; 1149 key.type = BTRFS_INODE_REF_KEY; 1150 key.offset = 0; 1151 1152 ret = btrfs_search_slot_for_read(root, &key, p, 1, 0); 1153 if (ret < 0) 1154 goto out; 1155 if (ret) { 1156 ret = 1; 1157 goto out; 1158 } 1159 btrfs_item_key_to_cpu(p->nodes[0], &found_key, p->slots[0]); 1160 if (found_key.objectid != ino || 1161 (found_key.type != BTRFS_INODE_REF_KEY && 1162 found_key.type != BTRFS_INODE_EXTREF_KEY)) { 1163 ret = -ENOENT; 1164 goto out; 1165 } 1166 1167 ret = iterate_inode_ref(root, p, &found_key, 1, 1168 __copy_first_ref, path); 1169 if (ret < 0) 1170 goto out; 1171 ret = 0; 1172 1173 out: 1174 btrfs_free_path(p); 1175 return ret; 1176 } 1177 1178 struct backref_ctx { 1179 struct send_ctx *sctx; 1180 1181 /* number of total found references */ 1182 u64 found; 1183 1184 /* 1185 * used for clones found in send_root. clones found behind cur_objectid 1186 * and cur_offset are not considered as allowed clones. 1187 */ 1188 u64 cur_objectid; 1189 u64 cur_offset; 1190 1191 /* may be truncated in case it's the last extent in a file */ 1192 u64 extent_len; 1193 1194 /* Just to check for bugs in backref resolving */ 1195 int found_itself; 1196 }; 1197 1198 static int __clone_root_cmp_bsearch(const void *key, const void *elt) 1199 { 1200 u64 root = (u64)(uintptr_t)key; 1201 struct clone_root *cr = (struct clone_root *)elt; 1202 1203 if (root < cr->root->root_key.objectid) 1204 return -1; 1205 if (root > cr->root->root_key.objectid) 1206 return 1; 1207 return 0; 1208 } 1209 1210 static int __clone_root_cmp_sort(const void *e1, const void *e2) 1211 { 1212 struct clone_root *cr1 = (struct clone_root *)e1; 1213 struct clone_root *cr2 = (struct clone_root *)e2; 1214 1215 if (cr1->root->root_key.objectid < cr2->root->root_key.objectid) 1216 return -1; 1217 if (cr1->root->root_key.objectid > cr2->root->root_key.objectid) 1218 return 1; 1219 return 0; 1220 } 1221 1222 /* 1223 * Called for every backref that is found for the current extent. 1224 * Results are collected in sctx->clone_roots->ino/offset/found_refs 1225 */ 1226 static int __iterate_backrefs(u64 ino, u64 offset, u64 root, void *ctx_) 1227 { 1228 struct backref_ctx *bctx = ctx_; 1229 struct clone_root *found; 1230 1231 /* First check if the root is in the list of accepted clone sources */ 1232 found = bsearch((void *)(uintptr_t)root, bctx->sctx->clone_roots, 1233 bctx->sctx->clone_roots_cnt, 1234 sizeof(struct clone_root), 1235 __clone_root_cmp_bsearch); 1236 if (!found) 1237 return 0; 1238 1239 if (found->root == bctx->sctx->send_root && 1240 ino == bctx->cur_objectid && 1241 offset == bctx->cur_offset) { 1242 bctx->found_itself = 1; 1243 } 1244 1245 /* 1246 * Make sure we don't consider clones from send_root that are 1247 * behind the current inode/offset. 1248 */ 1249 if (found->root == bctx->sctx->send_root) { 1250 /* 1251 * If the source inode was not yet processed we can't issue a 1252 * clone operation, as the source extent does not exist yet at 1253 * the destination of the stream. 1254 */ 1255 if (ino > bctx->cur_objectid) 1256 return 0; 1257 /* 1258 * We clone from the inode currently being sent as long as the 1259 * source extent is already processed, otherwise we could try 1260 * to clone from an extent that does not exist yet at the 1261 * destination of the stream. 1262 */ 1263 if (ino == bctx->cur_objectid && 1264 offset + bctx->extent_len > 1265 bctx->sctx->cur_inode_next_write_offset) 1266 return 0; 1267 } 1268 1269 bctx->found++; 1270 found->found_refs++; 1271 if (ino < found->ino) { 1272 found->ino = ino; 1273 found->offset = offset; 1274 } else if (found->ino == ino) { 1275 /* 1276 * same extent found more then once in the same file. 1277 */ 1278 if (found->offset > offset + bctx->extent_len) 1279 found->offset = offset; 1280 } 1281 1282 return 0; 1283 } 1284 1285 /* 1286 * Given an inode, offset and extent item, it finds a good clone for a clone 1287 * instruction. Returns -ENOENT when none could be found. The function makes 1288 * sure that the returned clone is usable at the point where sending is at the 1289 * moment. This means, that no clones are accepted which lie behind the current 1290 * inode+offset. 1291 * 1292 * path must point to the extent item when called. 1293 */ 1294 static int find_extent_clone(struct send_ctx *sctx, 1295 struct btrfs_path *path, 1296 u64 ino, u64 data_offset, 1297 u64 ino_size, 1298 struct clone_root **found) 1299 { 1300 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 1301 int ret; 1302 int extent_type; 1303 u64 logical; 1304 u64 disk_byte; 1305 u64 num_bytes; 1306 u64 extent_item_pos; 1307 u64 flags = 0; 1308 struct btrfs_file_extent_item *fi; 1309 struct extent_buffer *eb = path->nodes[0]; 1310 struct backref_ctx *backref_ctx = NULL; 1311 struct clone_root *cur_clone_root; 1312 struct btrfs_key found_key; 1313 struct btrfs_path *tmp_path; 1314 struct btrfs_extent_item *ei; 1315 int compressed; 1316 u32 i; 1317 1318 tmp_path = alloc_path_for_send(); 1319 if (!tmp_path) 1320 return -ENOMEM; 1321 1322 /* We only use this path under the commit sem */ 1323 tmp_path->need_commit_sem = 0; 1324 1325 backref_ctx = kmalloc(sizeof(*backref_ctx), GFP_KERNEL); 1326 if (!backref_ctx) { 1327 ret = -ENOMEM; 1328 goto out; 1329 } 1330 1331 if (data_offset >= ino_size) { 1332 /* 1333 * There may be extents that lie behind the file's size. 1334 * I at least had this in combination with snapshotting while 1335 * writing large files. 1336 */ 1337 ret = 0; 1338 goto out; 1339 } 1340 1341 fi = btrfs_item_ptr(eb, path->slots[0], 1342 struct btrfs_file_extent_item); 1343 extent_type = btrfs_file_extent_type(eb, fi); 1344 if (extent_type == BTRFS_FILE_EXTENT_INLINE) { 1345 ret = -ENOENT; 1346 goto out; 1347 } 1348 compressed = btrfs_file_extent_compression(eb, fi); 1349 1350 num_bytes = btrfs_file_extent_num_bytes(eb, fi); 1351 disk_byte = btrfs_file_extent_disk_bytenr(eb, fi); 1352 if (disk_byte == 0) { 1353 ret = -ENOENT; 1354 goto out; 1355 } 1356 logical = disk_byte + btrfs_file_extent_offset(eb, fi); 1357 1358 down_read(&fs_info->commit_root_sem); 1359 ret = extent_from_logical(fs_info, disk_byte, tmp_path, 1360 &found_key, &flags); 1361 up_read(&fs_info->commit_root_sem); 1362 1363 if (ret < 0) 1364 goto out; 1365 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 1366 ret = -EIO; 1367 goto out; 1368 } 1369 1370 ei = btrfs_item_ptr(tmp_path->nodes[0], tmp_path->slots[0], 1371 struct btrfs_extent_item); 1372 /* 1373 * Backreference walking (iterate_extent_inodes() below) is currently 1374 * too expensive when an extent has a large number of references, both 1375 * in time spent and used memory. So for now just fallback to write 1376 * operations instead of clone operations when an extent has more than 1377 * a certain amount of references. 1378 */ 1379 if (btrfs_extent_refs(tmp_path->nodes[0], ei) > SEND_MAX_EXTENT_REFS) { 1380 ret = -ENOENT; 1381 goto out; 1382 } 1383 btrfs_release_path(tmp_path); 1384 1385 /* 1386 * Setup the clone roots. 1387 */ 1388 for (i = 0; i < sctx->clone_roots_cnt; i++) { 1389 cur_clone_root = sctx->clone_roots + i; 1390 cur_clone_root->ino = (u64)-1; 1391 cur_clone_root->offset = 0; 1392 cur_clone_root->found_refs = 0; 1393 } 1394 1395 backref_ctx->sctx = sctx; 1396 backref_ctx->found = 0; 1397 backref_ctx->cur_objectid = ino; 1398 backref_ctx->cur_offset = data_offset; 1399 backref_ctx->found_itself = 0; 1400 backref_ctx->extent_len = num_bytes; 1401 1402 /* 1403 * The last extent of a file may be too large due to page alignment. 1404 * We need to adjust extent_len in this case so that the checks in 1405 * __iterate_backrefs work. 1406 */ 1407 if (data_offset + num_bytes >= ino_size) 1408 backref_ctx->extent_len = ino_size - data_offset; 1409 1410 /* 1411 * Now collect all backrefs. 1412 */ 1413 if (compressed == BTRFS_COMPRESS_NONE) 1414 extent_item_pos = logical - found_key.objectid; 1415 else 1416 extent_item_pos = 0; 1417 ret = iterate_extent_inodes(fs_info, found_key.objectid, 1418 extent_item_pos, 1, __iterate_backrefs, 1419 backref_ctx, false); 1420 1421 if (ret < 0) 1422 goto out; 1423 1424 if (!backref_ctx->found_itself) { 1425 /* found a bug in backref code? */ 1426 ret = -EIO; 1427 btrfs_err(fs_info, 1428 "did not find backref in send_root. inode=%llu, offset=%llu, disk_byte=%llu found extent=%llu", 1429 ino, data_offset, disk_byte, found_key.objectid); 1430 goto out; 1431 } 1432 1433 btrfs_debug(fs_info, 1434 "find_extent_clone: data_offset=%llu, ino=%llu, num_bytes=%llu, logical=%llu", 1435 data_offset, ino, num_bytes, logical); 1436 1437 if (!backref_ctx->found) 1438 btrfs_debug(fs_info, "no clones found"); 1439 1440 cur_clone_root = NULL; 1441 for (i = 0; i < sctx->clone_roots_cnt; i++) { 1442 if (sctx->clone_roots[i].found_refs) { 1443 if (!cur_clone_root) 1444 cur_clone_root = sctx->clone_roots + i; 1445 else if (sctx->clone_roots[i].root == sctx->send_root) 1446 /* prefer clones from send_root over others */ 1447 cur_clone_root = sctx->clone_roots + i; 1448 } 1449 1450 } 1451 1452 if (cur_clone_root) { 1453 *found = cur_clone_root; 1454 ret = 0; 1455 } else { 1456 ret = -ENOENT; 1457 } 1458 1459 out: 1460 btrfs_free_path(tmp_path); 1461 kfree(backref_ctx); 1462 return ret; 1463 } 1464 1465 static int read_symlink(struct btrfs_root *root, 1466 u64 ino, 1467 struct fs_path *dest) 1468 { 1469 int ret; 1470 struct btrfs_path *path; 1471 struct btrfs_key key; 1472 struct btrfs_file_extent_item *ei; 1473 u8 type; 1474 u8 compression; 1475 unsigned long off; 1476 int len; 1477 1478 path = alloc_path_for_send(); 1479 if (!path) 1480 return -ENOMEM; 1481 1482 key.objectid = ino; 1483 key.type = BTRFS_EXTENT_DATA_KEY; 1484 key.offset = 0; 1485 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 1486 if (ret < 0) 1487 goto out; 1488 if (ret) { 1489 /* 1490 * An empty symlink inode. Can happen in rare error paths when 1491 * creating a symlink (transaction committed before the inode 1492 * eviction handler removed the symlink inode items and a crash 1493 * happened in between or the subvol was snapshoted in between). 1494 * Print an informative message to dmesg/syslog so that the user 1495 * can delete the symlink. 1496 */ 1497 btrfs_err(root->fs_info, 1498 "Found empty symlink inode %llu at root %llu", 1499 ino, root->root_key.objectid); 1500 ret = -EIO; 1501 goto out; 1502 } 1503 1504 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 1505 struct btrfs_file_extent_item); 1506 type = btrfs_file_extent_type(path->nodes[0], ei); 1507 compression = btrfs_file_extent_compression(path->nodes[0], ei); 1508 BUG_ON(type != BTRFS_FILE_EXTENT_INLINE); 1509 BUG_ON(compression); 1510 1511 off = btrfs_file_extent_inline_start(ei); 1512 len = btrfs_file_extent_ram_bytes(path->nodes[0], ei); 1513 1514 ret = fs_path_add_from_extent_buffer(dest, path->nodes[0], off, len); 1515 1516 out: 1517 btrfs_free_path(path); 1518 return ret; 1519 } 1520 1521 /* 1522 * Helper function to generate a file name that is unique in the root of 1523 * send_root and parent_root. This is used to generate names for orphan inodes. 1524 */ 1525 static int gen_unique_name(struct send_ctx *sctx, 1526 u64 ino, u64 gen, 1527 struct fs_path *dest) 1528 { 1529 int ret = 0; 1530 struct btrfs_path *path; 1531 struct btrfs_dir_item *di; 1532 char tmp[64]; 1533 int len; 1534 u64 idx = 0; 1535 1536 path = alloc_path_for_send(); 1537 if (!path) 1538 return -ENOMEM; 1539 1540 while (1) { 1541 len = snprintf(tmp, sizeof(tmp), "o%llu-%llu-%llu", 1542 ino, gen, idx); 1543 ASSERT(len < sizeof(tmp)); 1544 1545 di = btrfs_lookup_dir_item(NULL, sctx->send_root, 1546 path, BTRFS_FIRST_FREE_OBJECTID, 1547 tmp, strlen(tmp), 0); 1548 btrfs_release_path(path); 1549 if (IS_ERR(di)) { 1550 ret = PTR_ERR(di); 1551 goto out; 1552 } 1553 if (di) { 1554 /* not unique, try again */ 1555 idx++; 1556 continue; 1557 } 1558 1559 if (!sctx->parent_root) { 1560 /* unique */ 1561 ret = 0; 1562 break; 1563 } 1564 1565 di = btrfs_lookup_dir_item(NULL, sctx->parent_root, 1566 path, BTRFS_FIRST_FREE_OBJECTID, 1567 tmp, strlen(tmp), 0); 1568 btrfs_release_path(path); 1569 if (IS_ERR(di)) { 1570 ret = PTR_ERR(di); 1571 goto out; 1572 } 1573 if (di) { 1574 /* not unique, try again */ 1575 idx++; 1576 continue; 1577 } 1578 /* unique */ 1579 break; 1580 } 1581 1582 ret = fs_path_add(dest, tmp, strlen(tmp)); 1583 1584 out: 1585 btrfs_free_path(path); 1586 return ret; 1587 } 1588 1589 enum inode_state { 1590 inode_state_no_change, 1591 inode_state_will_create, 1592 inode_state_did_create, 1593 inode_state_will_delete, 1594 inode_state_did_delete, 1595 }; 1596 1597 static int get_cur_inode_state(struct send_ctx *sctx, u64 ino, u64 gen) 1598 { 1599 int ret; 1600 int left_ret; 1601 int right_ret; 1602 u64 left_gen; 1603 u64 right_gen; 1604 1605 ret = get_inode_info(sctx->send_root, ino, NULL, &left_gen, NULL, NULL, 1606 NULL, NULL); 1607 if (ret < 0 && ret != -ENOENT) 1608 goto out; 1609 left_ret = ret; 1610 1611 if (!sctx->parent_root) { 1612 right_ret = -ENOENT; 1613 } else { 1614 ret = get_inode_info(sctx->parent_root, ino, NULL, &right_gen, 1615 NULL, NULL, NULL, NULL); 1616 if (ret < 0 && ret != -ENOENT) 1617 goto out; 1618 right_ret = ret; 1619 } 1620 1621 if (!left_ret && !right_ret) { 1622 if (left_gen == gen && right_gen == gen) { 1623 ret = inode_state_no_change; 1624 } else if (left_gen == gen) { 1625 if (ino < sctx->send_progress) 1626 ret = inode_state_did_create; 1627 else 1628 ret = inode_state_will_create; 1629 } else if (right_gen == gen) { 1630 if (ino < sctx->send_progress) 1631 ret = inode_state_did_delete; 1632 else 1633 ret = inode_state_will_delete; 1634 } else { 1635 ret = -ENOENT; 1636 } 1637 } else if (!left_ret) { 1638 if (left_gen == gen) { 1639 if (ino < sctx->send_progress) 1640 ret = inode_state_did_create; 1641 else 1642 ret = inode_state_will_create; 1643 } else { 1644 ret = -ENOENT; 1645 } 1646 } else if (!right_ret) { 1647 if (right_gen == gen) { 1648 if (ino < sctx->send_progress) 1649 ret = inode_state_did_delete; 1650 else 1651 ret = inode_state_will_delete; 1652 } else { 1653 ret = -ENOENT; 1654 } 1655 } else { 1656 ret = -ENOENT; 1657 } 1658 1659 out: 1660 return ret; 1661 } 1662 1663 static int is_inode_existent(struct send_ctx *sctx, u64 ino, u64 gen) 1664 { 1665 int ret; 1666 1667 if (ino == BTRFS_FIRST_FREE_OBJECTID) 1668 return 1; 1669 1670 ret = get_cur_inode_state(sctx, ino, gen); 1671 if (ret < 0) 1672 goto out; 1673 1674 if (ret == inode_state_no_change || 1675 ret == inode_state_did_create || 1676 ret == inode_state_will_delete) 1677 ret = 1; 1678 else 1679 ret = 0; 1680 1681 out: 1682 return ret; 1683 } 1684 1685 /* 1686 * Helper function to lookup a dir item in a dir. 1687 */ 1688 static int lookup_dir_item_inode(struct btrfs_root *root, 1689 u64 dir, const char *name, int name_len, 1690 u64 *found_inode, 1691 u8 *found_type) 1692 { 1693 int ret = 0; 1694 struct btrfs_dir_item *di; 1695 struct btrfs_key key; 1696 struct btrfs_path *path; 1697 1698 path = alloc_path_for_send(); 1699 if (!path) 1700 return -ENOMEM; 1701 1702 di = btrfs_lookup_dir_item(NULL, root, path, 1703 dir, name, name_len, 0); 1704 if (IS_ERR_OR_NULL(di)) { 1705 ret = di ? PTR_ERR(di) : -ENOENT; 1706 goto out; 1707 } 1708 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key); 1709 if (key.type == BTRFS_ROOT_ITEM_KEY) { 1710 ret = -ENOENT; 1711 goto out; 1712 } 1713 *found_inode = key.objectid; 1714 *found_type = btrfs_dir_type(path->nodes[0], di); 1715 1716 out: 1717 btrfs_free_path(path); 1718 return ret; 1719 } 1720 1721 /* 1722 * Looks up the first btrfs_inode_ref of a given ino. It returns the parent dir, 1723 * generation of the parent dir and the name of the dir entry. 1724 */ 1725 static int get_first_ref(struct btrfs_root *root, u64 ino, 1726 u64 *dir, u64 *dir_gen, struct fs_path *name) 1727 { 1728 int ret; 1729 struct btrfs_key key; 1730 struct btrfs_key found_key; 1731 struct btrfs_path *path; 1732 int len; 1733 u64 parent_dir; 1734 1735 path = alloc_path_for_send(); 1736 if (!path) 1737 return -ENOMEM; 1738 1739 key.objectid = ino; 1740 key.type = BTRFS_INODE_REF_KEY; 1741 key.offset = 0; 1742 1743 ret = btrfs_search_slot_for_read(root, &key, path, 1, 0); 1744 if (ret < 0) 1745 goto out; 1746 if (!ret) 1747 btrfs_item_key_to_cpu(path->nodes[0], &found_key, 1748 path->slots[0]); 1749 if (ret || found_key.objectid != ino || 1750 (found_key.type != BTRFS_INODE_REF_KEY && 1751 found_key.type != BTRFS_INODE_EXTREF_KEY)) { 1752 ret = -ENOENT; 1753 goto out; 1754 } 1755 1756 if (found_key.type == BTRFS_INODE_REF_KEY) { 1757 struct btrfs_inode_ref *iref; 1758 iref = btrfs_item_ptr(path->nodes[0], path->slots[0], 1759 struct btrfs_inode_ref); 1760 len = btrfs_inode_ref_name_len(path->nodes[0], iref); 1761 ret = fs_path_add_from_extent_buffer(name, path->nodes[0], 1762 (unsigned long)(iref + 1), 1763 len); 1764 parent_dir = found_key.offset; 1765 } else { 1766 struct btrfs_inode_extref *extref; 1767 extref = btrfs_item_ptr(path->nodes[0], path->slots[0], 1768 struct btrfs_inode_extref); 1769 len = btrfs_inode_extref_name_len(path->nodes[0], extref); 1770 ret = fs_path_add_from_extent_buffer(name, path->nodes[0], 1771 (unsigned long)&extref->name, len); 1772 parent_dir = btrfs_inode_extref_parent(path->nodes[0], extref); 1773 } 1774 if (ret < 0) 1775 goto out; 1776 btrfs_release_path(path); 1777 1778 if (dir_gen) { 1779 ret = get_inode_info(root, parent_dir, NULL, dir_gen, NULL, 1780 NULL, NULL, NULL); 1781 if (ret < 0) 1782 goto out; 1783 } 1784 1785 *dir = parent_dir; 1786 1787 out: 1788 btrfs_free_path(path); 1789 return ret; 1790 } 1791 1792 static int is_first_ref(struct btrfs_root *root, 1793 u64 ino, u64 dir, 1794 const char *name, int name_len) 1795 { 1796 int ret; 1797 struct fs_path *tmp_name; 1798 u64 tmp_dir; 1799 1800 tmp_name = fs_path_alloc(); 1801 if (!tmp_name) 1802 return -ENOMEM; 1803 1804 ret = get_first_ref(root, ino, &tmp_dir, NULL, tmp_name); 1805 if (ret < 0) 1806 goto out; 1807 1808 if (dir != tmp_dir || name_len != fs_path_len(tmp_name)) { 1809 ret = 0; 1810 goto out; 1811 } 1812 1813 ret = !memcmp(tmp_name->start, name, name_len); 1814 1815 out: 1816 fs_path_free(tmp_name); 1817 return ret; 1818 } 1819 1820 /* 1821 * Used by process_recorded_refs to determine if a new ref would overwrite an 1822 * already existing ref. In case it detects an overwrite, it returns the 1823 * inode/gen in who_ino/who_gen. 1824 * When an overwrite is detected, process_recorded_refs does proper orphanizing 1825 * to make sure later references to the overwritten inode are possible. 1826 * Orphanizing is however only required for the first ref of an inode. 1827 * process_recorded_refs does an additional is_first_ref check to see if 1828 * orphanizing is really required. 1829 */ 1830 static int will_overwrite_ref(struct send_ctx *sctx, u64 dir, u64 dir_gen, 1831 const char *name, int name_len, 1832 u64 *who_ino, u64 *who_gen, u64 *who_mode) 1833 { 1834 int ret = 0; 1835 u64 gen; 1836 u64 other_inode = 0; 1837 u8 other_type = 0; 1838 1839 if (!sctx->parent_root) 1840 goto out; 1841 1842 ret = is_inode_existent(sctx, dir, dir_gen); 1843 if (ret <= 0) 1844 goto out; 1845 1846 /* 1847 * If we have a parent root we need to verify that the parent dir was 1848 * not deleted and then re-created, if it was then we have no overwrite 1849 * and we can just unlink this entry. 1850 */ 1851 if (sctx->parent_root && dir != BTRFS_FIRST_FREE_OBJECTID) { 1852 ret = get_inode_info(sctx->parent_root, dir, NULL, &gen, NULL, 1853 NULL, NULL, NULL); 1854 if (ret < 0 && ret != -ENOENT) 1855 goto out; 1856 if (ret) { 1857 ret = 0; 1858 goto out; 1859 } 1860 if (gen != dir_gen) 1861 goto out; 1862 } 1863 1864 ret = lookup_dir_item_inode(sctx->parent_root, dir, name, name_len, 1865 &other_inode, &other_type); 1866 if (ret < 0 && ret != -ENOENT) 1867 goto out; 1868 if (ret) { 1869 ret = 0; 1870 goto out; 1871 } 1872 1873 /* 1874 * Check if the overwritten ref was already processed. If yes, the ref 1875 * was already unlinked/moved, so we can safely assume that we will not 1876 * overwrite anything at this point in time. 1877 */ 1878 if (other_inode > sctx->send_progress || 1879 is_waiting_for_move(sctx, other_inode)) { 1880 ret = get_inode_info(sctx->parent_root, other_inode, NULL, 1881 who_gen, who_mode, NULL, NULL, NULL); 1882 if (ret < 0) 1883 goto out; 1884 1885 ret = 1; 1886 *who_ino = other_inode; 1887 } else { 1888 ret = 0; 1889 } 1890 1891 out: 1892 return ret; 1893 } 1894 1895 /* 1896 * Checks if the ref was overwritten by an already processed inode. This is 1897 * used by __get_cur_name_and_parent to find out if the ref was orphanized and 1898 * thus the orphan name needs be used. 1899 * process_recorded_refs also uses it to avoid unlinking of refs that were 1900 * overwritten. 1901 */ 1902 static int did_overwrite_ref(struct send_ctx *sctx, 1903 u64 dir, u64 dir_gen, 1904 u64 ino, u64 ino_gen, 1905 const char *name, int name_len) 1906 { 1907 int ret = 0; 1908 u64 gen; 1909 u64 ow_inode; 1910 u8 other_type; 1911 1912 if (!sctx->parent_root) 1913 goto out; 1914 1915 ret = is_inode_existent(sctx, dir, dir_gen); 1916 if (ret <= 0) 1917 goto out; 1918 1919 if (dir != BTRFS_FIRST_FREE_OBJECTID) { 1920 ret = get_inode_info(sctx->send_root, dir, NULL, &gen, NULL, 1921 NULL, NULL, NULL); 1922 if (ret < 0 && ret != -ENOENT) 1923 goto out; 1924 if (ret) { 1925 ret = 0; 1926 goto out; 1927 } 1928 if (gen != dir_gen) 1929 goto out; 1930 } 1931 1932 /* check if the ref was overwritten by another ref */ 1933 ret = lookup_dir_item_inode(sctx->send_root, dir, name, name_len, 1934 &ow_inode, &other_type); 1935 if (ret < 0 && ret != -ENOENT) 1936 goto out; 1937 if (ret) { 1938 /* was never and will never be overwritten */ 1939 ret = 0; 1940 goto out; 1941 } 1942 1943 ret = get_inode_info(sctx->send_root, ow_inode, NULL, &gen, NULL, NULL, 1944 NULL, NULL); 1945 if (ret < 0) 1946 goto out; 1947 1948 if (ow_inode == ino && gen == ino_gen) { 1949 ret = 0; 1950 goto out; 1951 } 1952 1953 /* 1954 * We know that it is or will be overwritten. Check this now. 1955 * The current inode being processed might have been the one that caused 1956 * inode 'ino' to be orphanized, therefore check if ow_inode matches 1957 * the current inode being processed. 1958 */ 1959 if ((ow_inode < sctx->send_progress) || 1960 (ino != sctx->cur_ino && ow_inode == sctx->cur_ino && 1961 gen == sctx->cur_inode_gen)) 1962 ret = 1; 1963 else 1964 ret = 0; 1965 1966 out: 1967 return ret; 1968 } 1969 1970 /* 1971 * Same as did_overwrite_ref, but also checks if it is the first ref of an inode 1972 * that got overwritten. This is used by process_recorded_refs to determine 1973 * if it has to use the path as returned by get_cur_path or the orphan name. 1974 */ 1975 static int did_overwrite_first_ref(struct send_ctx *sctx, u64 ino, u64 gen) 1976 { 1977 int ret = 0; 1978 struct fs_path *name = NULL; 1979 u64 dir; 1980 u64 dir_gen; 1981 1982 if (!sctx->parent_root) 1983 goto out; 1984 1985 name = fs_path_alloc(); 1986 if (!name) 1987 return -ENOMEM; 1988 1989 ret = get_first_ref(sctx->parent_root, ino, &dir, &dir_gen, name); 1990 if (ret < 0) 1991 goto out; 1992 1993 ret = did_overwrite_ref(sctx, dir, dir_gen, ino, gen, 1994 name->start, fs_path_len(name)); 1995 1996 out: 1997 fs_path_free(name); 1998 return ret; 1999 } 2000 2001 /* 2002 * Insert a name cache entry. On 32bit kernels the radix tree index is 32bit, 2003 * so we need to do some special handling in case we have clashes. This function 2004 * takes care of this with the help of name_cache_entry::radix_list. 2005 * In case of error, nce is kfreed. 2006 */ 2007 static int name_cache_insert(struct send_ctx *sctx, 2008 struct name_cache_entry *nce) 2009 { 2010 int ret = 0; 2011 struct list_head *nce_head; 2012 2013 nce_head = radix_tree_lookup(&sctx->name_cache, 2014 (unsigned long)nce->ino); 2015 if (!nce_head) { 2016 nce_head = kmalloc(sizeof(*nce_head), GFP_KERNEL); 2017 if (!nce_head) { 2018 kfree(nce); 2019 return -ENOMEM; 2020 } 2021 INIT_LIST_HEAD(nce_head); 2022 2023 ret = radix_tree_insert(&sctx->name_cache, nce->ino, nce_head); 2024 if (ret < 0) { 2025 kfree(nce_head); 2026 kfree(nce); 2027 return ret; 2028 } 2029 } 2030 list_add_tail(&nce->radix_list, nce_head); 2031 list_add_tail(&nce->list, &sctx->name_cache_list); 2032 sctx->name_cache_size++; 2033 2034 return ret; 2035 } 2036 2037 static void name_cache_delete(struct send_ctx *sctx, 2038 struct name_cache_entry *nce) 2039 { 2040 struct list_head *nce_head; 2041 2042 nce_head = radix_tree_lookup(&sctx->name_cache, 2043 (unsigned long)nce->ino); 2044 if (!nce_head) { 2045 btrfs_err(sctx->send_root->fs_info, 2046 "name_cache_delete lookup failed ino %llu cache size %d, leaking memory", 2047 nce->ino, sctx->name_cache_size); 2048 } 2049 2050 list_del(&nce->radix_list); 2051 list_del(&nce->list); 2052 sctx->name_cache_size--; 2053 2054 /* 2055 * We may not get to the final release of nce_head if the lookup fails 2056 */ 2057 if (nce_head && list_empty(nce_head)) { 2058 radix_tree_delete(&sctx->name_cache, (unsigned long)nce->ino); 2059 kfree(nce_head); 2060 } 2061 } 2062 2063 static struct name_cache_entry *name_cache_search(struct send_ctx *sctx, 2064 u64 ino, u64 gen) 2065 { 2066 struct list_head *nce_head; 2067 struct name_cache_entry *cur; 2068 2069 nce_head = radix_tree_lookup(&sctx->name_cache, (unsigned long)ino); 2070 if (!nce_head) 2071 return NULL; 2072 2073 list_for_each_entry(cur, nce_head, radix_list) { 2074 if (cur->ino == ino && cur->gen == gen) 2075 return cur; 2076 } 2077 return NULL; 2078 } 2079 2080 /* 2081 * Removes the entry from the list and adds it back to the end. This marks the 2082 * entry as recently used so that name_cache_clean_unused does not remove it. 2083 */ 2084 static void name_cache_used(struct send_ctx *sctx, struct name_cache_entry *nce) 2085 { 2086 list_del(&nce->list); 2087 list_add_tail(&nce->list, &sctx->name_cache_list); 2088 } 2089 2090 /* 2091 * Remove some entries from the beginning of name_cache_list. 2092 */ 2093 static void name_cache_clean_unused(struct send_ctx *sctx) 2094 { 2095 struct name_cache_entry *nce; 2096 2097 if (sctx->name_cache_size < SEND_CTX_NAME_CACHE_CLEAN_SIZE) 2098 return; 2099 2100 while (sctx->name_cache_size > SEND_CTX_MAX_NAME_CACHE_SIZE) { 2101 nce = list_entry(sctx->name_cache_list.next, 2102 struct name_cache_entry, list); 2103 name_cache_delete(sctx, nce); 2104 kfree(nce); 2105 } 2106 } 2107 2108 static void name_cache_free(struct send_ctx *sctx) 2109 { 2110 struct name_cache_entry *nce; 2111 2112 while (!list_empty(&sctx->name_cache_list)) { 2113 nce = list_entry(sctx->name_cache_list.next, 2114 struct name_cache_entry, list); 2115 name_cache_delete(sctx, nce); 2116 kfree(nce); 2117 } 2118 } 2119 2120 /* 2121 * Used by get_cur_path for each ref up to the root. 2122 * Returns 0 if it succeeded. 2123 * Returns 1 if the inode is not existent or got overwritten. In that case, the 2124 * name is an orphan name. This instructs get_cur_path to stop iterating. If 1 2125 * is returned, parent_ino/parent_gen are not guaranteed to be valid. 2126 * Returns <0 in case of error. 2127 */ 2128 static int __get_cur_name_and_parent(struct send_ctx *sctx, 2129 u64 ino, u64 gen, 2130 u64 *parent_ino, 2131 u64 *parent_gen, 2132 struct fs_path *dest) 2133 { 2134 int ret; 2135 int nce_ret; 2136 struct name_cache_entry *nce = NULL; 2137 2138 /* 2139 * First check if we already did a call to this function with the same 2140 * ino/gen. If yes, check if the cache entry is still up-to-date. If yes 2141 * return the cached result. 2142 */ 2143 nce = name_cache_search(sctx, ino, gen); 2144 if (nce) { 2145 if (ino < sctx->send_progress && nce->need_later_update) { 2146 name_cache_delete(sctx, nce); 2147 kfree(nce); 2148 nce = NULL; 2149 } else { 2150 name_cache_used(sctx, nce); 2151 *parent_ino = nce->parent_ino; 2152 *parent_gen = nce->parent_gen; 2153 ret = fs_path_add(dest, nce->name, nce->name_len); 2154 if (ret < 0) 2155 goto out; 2156 ret = nce->ret; 2157 goto out; 2158 } 2159 } 2160 2161 /* 2162 * If the inode is not existent yet, add the orphan name and return 1. 2163 * This should only happen for the parent dir that we determine in 2164 * __record_new_ref 2165 */ 2166 ret = is_inode_existent(sctx, ino, gen); 2167 if (ret < 0) 2168 goto out; 2169 2170 if (!ret) { 2171 ret = gen_unique_name(sctx, ino, gen, dest); 2172 if (ret < 0) 2173 goto out; 2174 ret = 1; 2175 goto out_cache; 2176 } 2177 2178 /* 2179 * Depending on whether the inode was already processed or not, use 2180 * send_root or parent_root for ref lookup. 2181 */ 2182 if (ino < sctx->send_progress) 2183 ret = get_first_ref(sctx->send_root, ino, 2184 parent_ino, parent_gen, dest); 2185 else 2186 ret = get_first_ref(sctx->parent_root, ino, 2187 parent_ino, parent_gen, dest); 2188 if (ret < 0) 2189 goto out; 2190 2191 /* 2192 * Check if the ref was overwritten by an inode's ref that was processed 2193 * earlier. If yes, treat as orphan and return 1. 2194 */ 2195 ret = did_overwrite_ref(sctx, *parent_ino, *parent_gen, ino, gen, 2196 dest->start, dest->end - dest->start); 2197 if (ret < 0) 2198 goto out; 2199 if (ret) { 2200 fs_path_reset(dest); 2201 ret = gen_unique_name(sctx, ino, gen, dest); 2202 if (ret < 0) 2203 goto out; 2204 ret = 1; 2205 } 2206 2207 out_cache: 2208 /* 2209 * Store the result of the lookup in the name cache. 2210 */ 2211 nce = kmalloc(sizeof(*nce) + fs_path_len(dest) + 1, GFP_KERNEL); 2212 if (!nce) { 2213 ret = -ENOMEM; 2214 goto out; 2215 } 2216 2217 nce->ino = ino; 2218 nce->gen = gen; 2219 nce->parent_ino = *parent_ino; 2220 nce->parent_gen = *parent_gen; 2221 nce->name_len = fs_path_len(dest); 2222 nce->ret = ret; 2223 strcpy(nce->name, dest->start); 2224 2225 if (ino < sctx->send_progress) 2226 nce->need_later_update = 0; 2227 else 2228 nce->need_later_update = 1; 2229 2230 nce_ret = name_cache_insert(sctx, nce); 2231 if (nce_ret < 0) 2232 ret = nce_ret; 2233 name_cache_clean_unused(sctx); 2234 2235 out: 2236 return ret; 2237 } 2238 2239 /* 2240 * Magic happens here. This function returns the first ref to an inode as it 2241 * would look like while receiving the stream at this point in time. 2242 * We walk the path up to the root. For every inode in between, we check if it 2243 * was already processed/sent. If yes, we continue with the parent as found 2244 * in send_root. If not, we continue with the parent as found in parent_root. 2245 * If we encounter an inode that was deleted at this point in time, we use the 2246 * inodes "orphan" name instead of the real name and stop. Same with new inodes 2247 * that were not created yet and overwritten inodes/refs. 2248 * 2249 * When do we have orphan inodes: 2250 * 1. When an inode is freshly created and thus no valid refs are available yet 2251 * 2. When a directory lost all it's refs (deleted) but still has dir items 2252 * inside which were not processed yet (pending for move/delete). If anyone 2253 * tried to get the path to the dir items, it would get a path inside that 2254 * orphan directory. 2255 * 3. When an inode is moved around or gets new links, it may overwrite the ref 2256 * of an unprocessed inode. If in that case the first ref would be 2257 * overwritten, the overwritten inode gets "orphanized". Later when we 2258 * process this overwritten inode, it is restored at a new place by moving 2259 * the orphan inode. 2260 * 2261 * sctx->send_progress tells this function at which point in time receiving 2262 * would be. 2263 */ 2264 static int get_cur_path(struct send_ctx *sctx, u64 ino, u64 gen, 2265 struct fs_path *dest) 2266 { 2267 int ret = 0; 2268 struct fs_path *name = NULL; 2269 u64 parent_inode = 0; 2270 u64 parent_gen = 0; 2271 int stop = 0; 2272 2273 name = fs_path_alloc(); 2274 if (!name) { 2275 ret = -ENOMEM; 2276 goto out; 2277 } 2278 2279 dest->reversed = 1; 2280 fs_path_reset(dest); 2281 2282 while (!stop && ino != BTRFS_FIRST_FREE_OBJECTID) { 2283 struct waiting_dir_move *wdm; 2284 2285 fs_path_reset(name); 2286 2287 if (is_waiting_for_rm(sctx, ino, gen)) { 2288 ret = gen_unique_name(sctx, ino, gen, name); 2289 if (ret < 0) 2290 goto out; 2291 ret = fs_path_add_path(dest, name); 2292 break; 2293 } 2294 2295 wdm = get_waiting_dir_move(sctx, ino); 2296 if (wdm && wdm->orphanized) { 2297 ret = gen_unique_name(sctx, ino, gen, name); 2298 stop = 1; 2299 } else if (wdm) { 2300 ret = get_first_ref(sctx->parent_root, ino, 2301 &parent_inode, &parent_gen, name); 2302 } else { 2303 ret = __get_cur_name_and_parent(sctx, ino, gen, 2304 &parent_inode, 2305 &parent_gen, name); 2306 if (ret) 2307 stop = 1; 2308 } 2309 2310 if (ret < 0) 2311 goto out; 2312 2313 ret = fs_path_add_path(dest, name); 2314 if (ret < 0) 2315 goto out; 2316 2317 ino = parent_inode; 2318 gen = parent_gen; 2319 } 2320 2321 out: 2322 fs_path_free(name); 2323 if (!ret) 2324 fs_path_unreverse(dest); 2325 return ret; 2326 } 2327 2328 /* 2329 * Sends a BTRFS_SEND_C_SUBVOL command/item to userspace 2330 */ 2331 static int send_subvol_begin(struct send_ctx *sctx) 2332 { 2333 int ret; 2334 struct btrfs_root *send_root = sctx->send_root; 2335 struct btrfs_root *parent_root = sctx->parent_root; 2336 struct btrfs_path *path; 2337 struct btrfs_key key; 2338 struct btrfs_root_ref *ref; 2339 struct extent_buffer *leaf; 2340 char *name = NULL; 2341 int namelen; 2342 2343 path = btrfs_alloc_path(); 2344 if (!path) 2345 return -ENOMEM; 2346 2347 name = kmalloc(BTRFS_PATH_NAME_MAX, GFP_KERNEL); 2348 if (!name) { 2349 btrfs_free_path(path); 2350 return -ENOMEM; 2351 } 2352 2353 key.objectid = send_root->root_key.objectid; 2354 key.type = BTRFS_ROOT_BACKREF_KEY; 2355 key.offset = 0; 2356 2357 ret = btrfs_search_slot_for_read(send_root->fs_info->tree_root, 2358 &key, path, 1, 0); 2359 if (ret < 0) 2360 goto out; 2361 if (ret) { 2362 ret = -ENOENT; 2363 goto out; 2364 } 2365 2366 leaf = path->nodes[0]; 2367 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 2368 if (key.type != BTRFS_ROOT_BACKREF_KEY || 2369 key.objectid != send_root->root_key.objectid) { 2370 ret = -ENOENT; 2371 goto out; 2372 } 2373 ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref); 2374 namelen = btrfs_root_ref_name_len(leaf, ref); 2375 read_extent_buffer(leaf, name, (unsigned long)(ref + 1), namelen); 2376 btrfs_release_path(path); 2377 2378 if (parent_root) { 2379 ret = begin_cmd(sctx, BTRFS_SEND_C_SNAPSHOT); 2380 if (ret < 0) 2381 goto out; 2382 } else { 2383 ret = begin_cmd(sctx, BTRFS_SEND_C_SUBVOL); 2384 if (ret < 0) 2385 goto out; 2386 } 2387 2388 TLV_PUT_STRING(sctx, BTRFS_SEND_A_PATH, name, namelen); 2389 2390 if (!btrfs_is_empty_uuid(sctx->send_root->root_item.received_uuid)) 2391 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID, 2392 sctx->send_root->root_item.received_uuid); 2393 else 2394 TLV_PUT_UUID(sctx, BTRFS_SEND_A_UUID, 2395 sctx->send_root->root_item.uuid); 2396 2397 TLV_PUT_U64(sctx, BTRFS_SEND_A_CTRANSID, 2398 btrfs_root_ctransid(&sctx->send_root->root_item)); 2399 if (parent_root) { 2400 if (!btrfs_is_empty_uuid(parent_root->root_item.received_uuid)) 2401 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID, 2402 parent_root->root_item.received_uuid); 2403 else 2404 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID, 2405 parent_root->root_item.uuid); 2406 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID, 2407 btrfs_root_ctransid(&sctx->parent_root->root_item)); 2408 } 2409 2410 ret = send_cmd(sctx); 2411 2412 tlv_put_failure: 2413 out: 2414 btrfs_free_path(path); 2415 kfree(name); 2416 return ret; 2417 } 2418 2419 static int send_truncate(struct send_ctx *sctx, u64 ino, u64 gen, u64 size) 2420 { 2421 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 2422 int ret = 0; 2423 struct fs_path *p; 2424 2425 btrfs_debug(fs_info, "send_truncate %llu size=%llu", ino, size); 2426 2427 p = fs_path_alloc(); 2428 if (!p) 2429 return -ENOMEM; 2430 2431 ret = begin_cmd(sctx, BTRFS_SEND_C_TRUNCATE); 2432 if (ret < 0) 2433 goto out; 2434 2435 ret = get_cur_path(sctx, ino, gen, p); 2436 if (ret < 0) 2437 goto out; 2438 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 2439 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, size); 2440 2441 ret = send_cmd(sctx); 2442 2443 tlv_put_failure: 2444 out: 2445 fs_path_free(p); 2446 return ret; 2447 } 2448 2449 static int send_chmod(struct send_ctx *sctx, u64 ino, u64 gen, u64 mode) 2450 { 2451 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 2452 int ret = 0; 2453 struct fs_path *p; 2454 2455 btrfs_debug(fs_info, "send_chmod %llu mode=%llu", ino, mode); 2456 2457 p = fs_path_alloc(); 2458 if (!p) 2459 return -ENOMEM; 2460 2461 ret = begin_cmd(sctx, BTRFS_SEND_C_CHMOD); 2462 if (ret < 0) 2463 goto out; 2464 2465 ret = get_cur_path(sctx, ino, gen, p); 2466 if (ret < 0) 2467 goto out; 2468 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 2469 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode & 07777); 2470 2471 ret = send_cmd(sctx); 2472 2473 tlv_put_failure: 2474 out: 2475 fs_path_free(p); 2476 return ret; 2477 } 2478 2479 static int send_chown(struct send_ctx *sctx, u64 ino, u64 gen, u64 uid, u64 gid) 2480 { 2481 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 2482 int ret = 0; 2483 struct fs_path *p; 2484 2485 btrfs_debug(fs_info, "send_chown %llu uid=%llu, gid=%llu", 2486 ino, uid, gid); 2487 2488 p = fs_path_alloc(); 2489 if (!p) 2490 return -ENOMEM; 2491 2492 ret = begin_cmd(sctx, BTRFS_SEND_C_CHOWN); 2493 if (ret < 0) 2494 goto out; 2495 2496 ret = get_cur_path(sctx, ino, gen, p); 2497 if (ret < 0) 2498 goto out; 2499 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 2500 TLV_PUT_U64(sctx, BTRFS_SEND_A_UID, uid); 2501 TLV_PUT_U64(sctx, BTRFS_SEND_A_GID, gid); 2502 2503 ret = send_cmd(sctx); 2504 2505 tlv_put_failure: 2506 out: 2507 fs_path_free(p); 2508 return ret; 2509 } 2510 2511 static int send_utimes(struct send_ctx *sctx, u64 ino, u64 gen) 2512 { 2513 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 2514 int ret = 0; 2515 struct fs_path *p = NULL; 2516 struct btrfs_inode_item *ii; 2517 struct btrfs_path *path = NULL; 2518 struct extent_buffer *eb; 2519 struct btrfs_key key; 2520 int slot; 2521 2522 btrfs_debug(fs_info, "send_utimes %llu", ino); 2523 2524 p = fs_path_alloc(); 2525 if (!p) 2526 return -ENOMEM; 2527 2528 path = alloc_path_for_send(); 2529 if (!path) { 2530 ret = -ENOMEM; 2531 goto out; 2532 } 2533 2534 key.objectid = ino; 2535 key.type = BTRFS_INODE_ITEM_KEY; 2536 key.offset = 0; 2537 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0); 2538 if (ret > 0) 2539 ret = -ENOENT; 2540 if (ret < 0) 2541 goto out; 2542 2543 eb = path->nodes[0]; 2544 slot = path->slots[0]; 2545 ii = btrfs_item_ptr(eb, slot, struct btrfs_inode_item); 2546 2547 ret = begin_cmd(sctx, BTRFS_SEND_C_UTIMES); 2548 if (ret < 0) 2549 goto out; 2550 2551 ret = get_cur_path(sctx, ino, gen, p); 2552 if (ret < 0) 2553 goto out; 2554 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 2555 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_ATIME, eb, &ii->atime); 2556 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_MTIME, eb, &ii->mtime); 2557 TLV_PUT_BTRFS_TIMESPEC(sctx, BTRFS_SEND_A_CTIME, eb, &ii->ctime); 2558 /* TODO Add otime support when the otime patches get into upstream */ 2559 2560 ret = send_cmd(sctx); 2561 2562 tlv_put_failure: 2563 out: 2564 fs_path_free(p); 2565 btrfs_free_path(path); 2566 return ret; 2567 } 2568 2569 /* 2570 * Sends a BTRFS_SEND_C_MKXXX or SYMLINK command to user space. We don't have 2571 * a valid path yet because we did not process the refs yet. So, the inode 2572 * is created as orphan. 2573 */ 2574 static int send_create_inode(struct send_ctx *sctx, u64 ino) 2575 { 2576 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 2577 int ret = 0; 2578 struct fs_path *p; 2579 int cmd; 2580 u64 gen; 2581 u64 mode; 2582 u64 rdev; 2583 2584 btrfs_debug(fs_info, "send_create_inode %llu", ino); 2585 2586 p = fs_path_alloc(); 2587 if (!p) 2588 return -ENOMEM; 2589 2590 if (ino != sctx->cur_ino) { 2591 ret = get_inode_info(sctx->send_root, ino, NULL, &gen, &mode, 2592 NULL, NULL, &rdev); 2593 if (ret < 0) 2594 goto out; 2595 } else { 2596 gen = sctx->cur_inode_gen; 2597 mode = sctx->cur_inode_mode; 2598 rdev = sctx->cur_inode_rdev; 2599 } 2600 2601 if (S_ISREG(mode)) { 2602 cmd = BTRFS_SEND_C_MKFILE; 2603 } else if (S_ISDIR(mode)) { 2604 cmd = BTRFS_SEND_C_MKDIR; 2605 } else if (S_ISLNK(mode)) { 2606 cmd = BTRFS_SEND_C_SYMLINK; 2607 } else if (S_ISCHR(mode) || S_ISBLK(mode)) { 2608 cmd = BTRFS_SEND_C_MKNOD; 2609 } else if (S_ISFIFO(mode)) { 2610 cmd = BTRFS_SEND_C_MKFIFO; 2611 } else if (S_ISSOCK(mode)) { 2612 cmd = BTRFS_SEND_C_MKSOCK; 2613 } else { 2614 btrfs_warn(sctx->send_root->fs_info, "unexpected inode type %o", 2615 (int)(mode & S_IFMT)); 2616 ret = -EOPNOTSUPP; 2617 goto out; 2618 } 2619 2620 ret = begin_cmd(sctx, cmd); 2621 if (ret < 0) 2622 goto out; 2623 2624 ret = gen_unique_name(sctx, ino, gen, p); 2625 if (ret < 0) 2626 goto out; 2627 2628 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 2629 TLV_PUT_U64(sctx, BTRFS_SEND_A_INO, ino); 2630 2631 if (S_ISLNK(mode)) { 2632 fs_path_reset(p); 2633 ret = read_symlink(sctx->send_root, ino, p); 2634 if (ret < 0) 2635 goto out; 2636 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH_LINK, p); 2637 } else if (S_ISCHR(mode) || S_ISBLK(mode) || 2638 S_ISFIFO(mode) || S_ISSOCK(mode)) { 2639 TLV_PUT_U64(sctx, BTRFS_SEND_A_RDEV, new_encode_dev(rdev)); 2640 TLV_PUT_U64(sctx, BTRFS_SEND_A_MODE, mode); 2641 } 2642 2643 ret = send_cmd(sctx); 2644 if (ret < 0) 2645 goto out; 2646 2647 2648 tlv_put_failure: 2649 out: 2650 fs_path_free(p); 2651 return ret; 2652 } 2653 2654 /* 2655 * We need some special handling for inodes that get processed before the parent 2656 * directory got created. See process_recorded_refs for details. 2657 * This function does the check if we already created the dir out of order. 2658 */ 2659 static int did_create_dir(struct send_ctx *sctx, u64 dir) 2660 { 2661 int ret = 0; 2662 struct btrfs_path *path = NULL; 2663 struct btrfs_key key; 2664 struct btrfs_key found_key; 2665 struct btrfs_key di_key; 2666 struct extent_buffer *eb; 2667 struct btrfs_dir_item *di; 2668 int slot; 2669 2670 path = alloc_path_for_send(); 2671 if (!path) { 2672 ret = -ENOMEM; 2673 goto out; 2674 } 2675 2676 key.objectid = dir; 2677 key.type = BTRFS_DIR_INDEX_KEY; 2678 key.offset = 0; 2679 ret = btrfs_search_slot(NULL, sctx->send_root, &key, path, 0, 0); 2680 if (ret < 0) 2681 goto out; 2682 2683 while (1) { 2684 eb = path->nodes[0]; 2685 slot = path->slots[0]; 2686 if (slot >= btrfs_header_nritems(eb)) { 2687 ret = btrfs_next_leaf(sctx->send_root, path); 2688 if (ret < 0) { 2689 goto out; 2690 } else if (ret > 0) { 2691 ret = 0; 2692 break; 2693 } 2694 continue; 2695 } 2696 2697 btrfs_item_key_to_cpu(eb, &found_key, slot); 2698 if (found_key.objectid != key.objectid || 2699 found_key.type != key.type) { 2700 ret = 0; 2701 goto out; 2702 } 2703 2704 di = btrfs_item_ptr(eb, slot, struct btrfs_dir_item); 2705 btrfs_dir_item_key_to_cpu(eb, di, &di_key); 2706 2707 if (di_key.type != BTRFS_ROOT_ITEM_KEY && 2708 di_key.objectid < sctx->send_progress) { 2709 ret = 1; 2710 goto out; 2711 } 2712 2713 path->slots[0]++; 2714 } 2715 2716 out: 2717 btrfs_free_path(path); 2718 return ret; 2719 } 2720 2721 /* 2722 * Only creates the inode if it is: 2723 * 1. Not a directory 2724 * 2. Or a directory which was not created already due to out of order 2725 * directories. See did_create_dir and process_recorded_refs for details. 2726 */ 2727 static int send_create_inode_if_needed(struct send_ctx *sctx) 2728 { 2729 int ret; 2730 2731 if (S_ISDIR(sctx->cur_inode_mode)) { 2732 ret = did_create_dir(sctx, sctx->cur_ino); 2733 if (ret < 0) 2734 goto out; 2735 if (ret) { 2736 ret = 0; 2737 goto out; 2738 } 2739 } 2740 2741 ret = send_create_inode(sctx, sctx->cur_ino); 2742 if (ret < 0) 2743 goto out; 2744 2745 out: 2746 return ret; 2747 } 2748 2749 struct recorded_ref { 2750 struct list_head list; 2751 char *name; 2752 struct fs_path *full_path; 2753 u64 dir; 2754 u64 dir_gen; 2755 int name_len; 2756 }; 2757 2758 static void set_ref_path(struct recorded_ref *ref, struct fs_path *path) 2759 { 2760 ref->full_path = path; 2761 ref->name = (char *)kbasename(ref->full_path->start); 2762 ref->name_len = ref->full_path->end - ref->name; 2763 } 2764 2765 /* 2766 * We need to process new refs before deleted refs, but compare_tree gives us 2767 * everything mixed. So we first record all refs and later process them. 2768 * This function is a helper to record one ref. 2769 */ 2770 static int __record_ref(struct list_head *head, u64 dir, 2771 u64 dir_gen, struct fs_path *path) 2772 { 2773 struct recorded_ref *ref; 2774 2775 ref = kmalloc(sizeof(*ref), GFP_KERNEL); 2776 if (!ref) 2777 return -ENOMEM; 2778 2779 ref->dir = dir; 2780 ref->dir_gen = dir_gen; 2781 set_ref_path(ref, path); 2782 list_add_tail(&ref->list, head); 2783 return 0; 2784 } 2785 2786 static int dup_ref(struct recorded_ref *ref, struct list_head *list) 2787 { 2788 struct recorded_ref *new; 2789 2790 new = kmalloc(sizeof(*ref), GFP_KERNEL); 2791 if (!new) 2792 return -ENOMEM; 2793 2794 new->dir = ref->dir; 2795 new->dir_gen = ref->dir_gen; 2796 new->full_path = NULL; 2797 INIT_LIST_HEAD(&new->list); 2798 list_add_tail(&new->list, list); 2799 return 0; 2800 } 2801 2802 static void __free_recorded_refs(struct list_head *head) 2803 { 2804 struct recorded_ref *cur; 2805 2806 while (!list_empty(head)) { 2807 cur = list_entry(head->next, struct recorded_ref, list); 2808 fs_path_free(cur->full_path); 2809 list_del(&cur->list); 2810 kfree(cur); 2811 } 2812 } 2813 2814 static void free_recorded_refs(struct send_ctx *sctx) 2815 { 2816 __free_recorded_refs(&sctx->new_refs); 2817 __free_recorded_refs(&sctx->deleted_refs); 2818 } 2819 2820 /* 2821 * Renames/moves a file/dir to its orphan name. Used when the first 2822 * ref of an unprocessed inode gets overwritten and for all non empty 2823 * directories. 2824 */ 2825 static int orphanize_inode(struct send_ctx *sctx, u64 ino, u64 gen, 2826 struct fs_path *path) 2827 { 2828 int ret; 2829 struct fs_path *orphan; 2830 2831 orphan = fs_path_alloc(); 2832 if (!orphan) 2833 return -ENOMEM; 2834 2835 ret = gen_unique_name(sctx, ino, gen, orphan); 2836 if (ret < 0) 2837 goto out; 2838 2839 ret = send_rename(sctx, path, orphan); 2840 2841 out: 2842 fs_path_free(orphan); 2843 return ret; 2844 } 2845 2846 static struct orphan_dir_info *add_orphan_dir_info(struct send_ctx *sctx, 2847 u64 dir_ino, u64 dir_gen) 2848 { 2849 struct rb_node **p = &sctx->orphan_dirs.rb_node; 2850 struct rb_node *parent = NULL; 2851 struct orphan_dir_info *entry, *odi; 2852 2853 while (*p) { 2854 parent = *p; 2855 entry = rb_entry(parent, struct orphan_dir_info, node); 2856 if (dir_ino < entry->ino) 2857 p = &(*p)->rb_left; 2858 else if (dir_ino > entry->ino) 2859 p = &(*p)->rb_right; 2860 else if (dir_gen < entry->gen) 2861 p = &(*p)->rb_left; 2862 else if (dir_gen > entry->gen) 2863 p = &(*p)->rb_right; 2864 else 2865 return entry; 2866 } 2867 2868 odi = kmalloc(sizeof(*odi), GFP_KERNEL); 2869 if (!odi) 2870 return ERR_PTR(-ENOMEM); 2871 odi->ino = dir_ino; 2872 odi->gen = dir_gen; 2873 odi->last_dir_index_offset = 0; 2874 2875 rb_link_node(&odi->node, parent, p); 2876 rb_insert_color(&odi->node, &sctx->orphan_dirs); 2877 return odi; 2878 } 2879 2880 static struct orphan_dir_info *get_orphan_dir_info(struct send_ctx *sctx, 2881 u64 dir_ino, u64 gen) 2882 { 2883 struct rb_node *n = sctx->orphan_dirs.rb_node; 2884 struct orphan_dir_info *entry; 2885 2886 while (n) { 2887 entry = rb_entry(n, struct orphan_dir_info, node); 2888 if (dir_ino < entry->ino) 2889 n = n->rb_left; 2890 else if (dir_ino > entry->ino) 2891 n = n->rb_right; 2892 else if (gen < entry->gen) 2893 n = n->rb_left; 2894 else if (gen > entry->gen) 2895 n = n->rb_right; 2896 else 2897 return entry; 2898 } 2899 return NULL; 2900 } 2901 2902 static int is_waiting_for_rm(struct send_ctx *sctx, u64 dir_ino, u64 gen) 2903 { 2904 struct orphan_dir_info *odi = get_orphan_dir_info(sctx, dir_ino, gen); 2905 2906 return odi != NULL; 2907 } 2908 2909 static void free_orphan_dir_info(struct send_ctx *sctx, 2910 struct orphan_dir_info *odi) 2911 { 2912 if (!odi) 2913 return; 2914 rb_erase(&odi->node, &sctx->orphan_dirs); 2915 kfree(odi); 2916 } 2917 2918 /* 2919 * Returns 1 if a directory can be removed at this point in time. 2920 * We check this by iterating all dir items and checking if the inode behind 2921 * the dir item was already processed. 2922 */ 2923 static int can_rmdir(struct send_ctx *sctx, u64 dir, u64 dir_gen, 2924 u64 send_progress) 2925 { 2926 int ret = 0; 2927 struct btrfs_root *root = sctx->parent_root; 2928 struct btrfs_path *path; 2929 struct btrfs_key key; 2930 struct btrfs_key found_key; 2931 struct btrfs_key loc; 2932 struct btrfs_dir_item *di; 2933 struct orphan_dir_info *odi = NULL; 2934 2935 /* 2936 * Don't try to rmdir the top/root subvolume dir. 2937 */ 2938 if (dir == BTRFS_FIRST_FREE_OBJECTID) 2939 return 0; 2940 2941 path = alloc_path_for_send(); 2942 if (!path) 2943 return -ENOMEM; 2944 2945 key.objectid = dir; 2946 key.type = BTRFS_DIR_INDEX_KEY; 2947 key.offset = 0; 2948 2949 odi = get_orphan_dir_info(sctx, dir, dir_gen); 2950 if (odi) 2951 key.offset = odi->last_dir_index_offset; 2952 2953 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 2954 if (ret < 0) 2955 goto out; 2956 2957 while (1) { 2958 struct waiting_dir_move *dm; 2959 2960 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) { 2961 ret = btrfs_next_leaf(root, path); 2962 if (ret < 0) 2963 goto out; 2964 else if (ret > 0) 2965 break; 2966 continue; 2967 } 2968 btrfs_item_key_to_cpu(path->nodes[0], &found_key, 2969 path->slots[0]); 2970 if (found_key.objectid != key.objectid || 2971 found_key.type != key.type) 2972 break; 2973 2974 di = btrfs_item_ptr(path->nodes[0], path->slots[0], 2975 struct btrfs_dir_item); 2976 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &loc); 2977 2978 dm = get_waiting_dir_move(sctx, loc.objectid); 2979 if (dm) { 2980 odi = add_orphan_dir_info(sctx, dir, dir_gen); 2981 if (IS_ERR(odi)) { 2982 ret = PTR_ERR(odi); 2983 goto out; 2984 } 2985 odi->gen = dir_gen; 2986 odi->last_dir_index_offset = found_key.offset; 2987 dm->rmdir_ino = dir; 2988 dm->rmdir_gen = dir_gen; 2989 ret = 0; 2990 goto out; 2991 } 2992 2993 if (loc.objectid > send_progress) { 2994 odi = add_orphan_dir_info(sctx, dir, dir_gen); 2995 if (IS_ERR(odi)) { 2996 ret = PTR_ERR(odi); 2997 goto out; 2998 } 2999 odi->gen = dir_gen; 3000 odi->last_dir_index_offset = found_key.offset; 3001 ret = 0; 3002 goto out; 3003 } 3004 3005 path->slots[0]++; 3006 } 3007 free_orphan_dir_info(sctx, odi); 3008 3009 ret = 1; 3010 3011 out: 3012 btrfs_free_path(path); 3013 return ret; 3014 } 3015 3016 static int is_waiting_for_move(struct send_ctx *sctx, u64 ino) 3017 { 3018 struct waiting_dir_move *entry = get_waiting_dir_move(sctx, ino); 3019 3020 return entry != NULL; 3021 } 3022 3023 static int add_waiting_dir_move(struct send_ctx *sctx, u64 ino, bool orphanized) 3024 { 3025 struct rb_node **p = &sctx->waiting_dir_moves.rb_node; 3026 struct rb_node *parent = NULL; 3027 struct waiting_dir_move *entry, *dm; 3028 3029 dm = kmalloc(sizeof(*dm), GFP_KERNEL); 3030 if (!dm) 3031 return -ENOMEM; 3032 dm->ino = ino; 3033 dm->rmdir_ino = 0; 3034 dm->rmdir_gen = 0; 3035 dm->orphanized = orphanized; 3036 3037 while (*p) { 3038 parent = *p; 3039 entry = rb_entry(parent, struct waiting_dir_move, node); 3040 if (ino < entry->ino) { 3041 p = &(*p)->rb_left; 3042 } else if (ino > entry->ino) { 3043 p = &(*p)->rb_right; 3044 } else { 3045 kfree(dm); 3046 return -EEXIST; 3047 } 3048 } 3049 3050 rb_link_node(&dm->node, parent, p); 3051 rb_insert_color(&dm->node, &sctx->waiting_dir_moves); 3052 return 0; 3053 } 3054 3055 static struct waiting_dir_move * 3056 get_waiting_dir_move(struct send_ctx *sctx, u64 ino) 3057 { 3058 struct rb_node *n = sctx->waiting_dir_moves.rb_node; 3059 struct waiting_dir_move *entry; 3060 3061 while (n) { 3062 entry = rb_entry(n, struct waiting_dir_move, node); 3063 if (ino < entry->ino) 3064 n = n->rb_left; 3065 else if (ino > entry->ino) 3066 n = n->rb_right; 3067 else 3068 return entry; 3069 } 3070 return NULL; 3071 } 3072 3073 static void free_waiting_dir_move(struct send_ctx *sctx, 3074 struct waiting_dir_move *dm) 3075 { 3076 if (!dm) 3077 return; 3078 rb_erase(&dm->node, &sctx->waiting_dir_moves); 3079 kfree(dm); 3080 } 3081 3082 static int add_pending_dir_move(struct send_ctx *sctx, 3083 u64 ino, 3084 u64 ino_gen, 3085 u64 parent_ino, 3086 struct list_head *new_refs, 3087 struct list_head *deleted_refs, 3088 const bool is_orphan) 3089 { 3090 struct rb_node **p = &sctx->pending_dir_moves.rb_node; 3091 struct rb_node *parent = NULL; 3092 struct pending_dir_move *entry = NULL, *pm; 3093 struct recorded_ref *cur; 3094 int exists = 0; 3095 int ret; 3096 3097 pm = kmalloc(sizeof(*pm), GFP_KERNEL); 3098 if (!pm) 3099 return -ENOMEM; 3100 pm->parent_ino = parent_ino; 3101 pm->ino = ino; 3102 pm->gen = ino_gen; 3103 INIT_LIST_HEAD(&pm->list); 3104 INIT_LIST_HEAD(&pm->update_refs); 3105 RB_CLEAR_NODE(&pm->node); 3106 3107 while (*p) { 3108 parent = *p; 3109 entry = rb_entry(parent, struct pending_dir_move, node); 3110 if (parent_ino < entry->parent_ino) { 3111 p = &(*p)->rb_left; 3112 } else if (parent_ino > entry->parent_ino) { 3113 p = &(*p)->rb_right; 3114 } else { 3115 exists = 1; 3116 break; 3117 } 3118 } 3119 3120 list_for_each_entry(cur, deleted_refs, list) { 3121 ret = dup_ref(cur, &pm->update_refs); 3122 if (ret < 0) 3123 goto out; 3124 } 3125 list_for_each_entry(cur, new_refs, list) { 3126 ret = dup_ref(cur, &pm->update_refs); 3127 if (ret < 0) 3128 goto out; 3129 } 3130 3131 ret = add_waiting_dir_move(sctx, pm->ino, is_orphan); 3132 if (ret) 3133 goto out; 3134 3135 if (exists) { 3136 list_add_tail(&pm->list, &entry->list); 3137 } else { 3138 rb_link_node(&pm->node, parent, p); 3139 rb_insert_color(&pm->node, &sctx->pending_dir_moves); 3140 } 3141 ret = 0; 3142 out: 3143 if (ret) { 3144 __free_recorded_refs(&pm->update_refs); 3145 kfree(pm); 3146 } 3147 return ret; 3148 } 3149 3150 static struct pending_dir_move *get_pending_dir_moves(struct send_ctx *sctx, 3151 u64 parent_ino) 3152 { 3153 struct rb_node *n = sctx->pending_dir_moves.rb_node; 3154 struct pending_dir_move *entry; 3155 3156 while (n) { 3157 entry = rb_entry(n, struct pending_dir_move, node); 3158 if (parent_ino < entry->parent_ino) 3159 n = n->rb_left; 3160 else if (parent_ino > entry->parent_ino) 3161 n = n->rb_right; 3162 else 3163 return entry; 3164 } 3165 return NULL; 3166 } 3167 3168 static int path_loop(struct send_ctx *sctx, struct fs_path *name, 3169 u64 ino, u64 gen, u64 *ancestor_ino) 3170 { 3171 int ret = 0; 3172 u64 parent_inode = 0; 3173 u64 parent_gen = 0; 3174 u64 start_ino = ino; 3175 3176 *ancestor_ino = 0; 3177 while (ino != BTRFS_FIRST_FREE_OBJECTID) { 3178 fs_path_reset(name); 3179 3180 if (is_waiting_for_rm(sctx, ino, gen)) 3181 break; 3182 if (is_waiting_for_move(sctx, ino)) { 3183 if (*ancestor_ino == 0) 3184 *ancestor_ino = ino; 3185 ret = get_first_ref(sctx->parent_root, ino, 3186 &parent_inode, &parent_gen, name); 3187 } else { 3188 ret = __get_cur_name_and_parent(sctx, ino, gen, 3189 &parent_inode, 3190 &parent_gen, name); 3191 if (ret > 0) { 3192 ret = 0; 3193 break; 3194 } 3195 } 3196 if (ret < 0) 3197 break; 3198 if (parent_inode == start_ino) { 3199 ret = 1; 3200 if (*ancestor_ino == 0) 3201 *ancestor_ino = ino; 3202 break; 3203 } 3204 ino = parent_inode; 3205 gen = parent_gen; 3206 } 3207 return ret; 3208 } 3209 3210 static int apply_dir_move(struct send_ctx *sctx, struct pending_dir_move *pm) 3211 { 3212 struct fs_path *from_path = NULL; 3213 struct fs_path *to_path = NULL; 3214 struct fs_path *name = NULL; 3215 u64 orig_progress = sctx->send_progress; 3216 struct recorded_ref *cur; 3217 u64 parent_ino, parent_gen; 3218 struct waiting_dir_move *dm = NULL; 3219 u64 rmdir_ino = 0; 3220 u64 rmdir_gen; 3221 u64 ancestor; 3222 bool is_orphan; 3223 int ret; 3224 3225 name = fs_path_alloc(); 3226 from_path = fs_path_alloc(); 3227 if (!name || !from_path) { 3228 ret = -ENOMEM; 3229 goto out; 3230 } 3231 3232 dm = get_waiting_dir_move(sctx, pm->ino); 3233 ASSERT(dm); 3234 rmdir_ino = dm->rmdir_ino; 3235 rmdir_gen = dm->rmdir_gen; 3236 is_orphan = dm->orphanized; 3237 free_waiting_dir_move(sctx, dm); 3238 3239 if (is_orphan) { 3240 ret = gen_unique_name(sctx, pm->ino, 3241 pm->gen, from_path); 3242 } else { 3243 ret = get_first_ref(sctx->parent_root, pm->ino, 3244 &parent_ino, &parent_gen, name); 3245 if (ret < 0) 3246 goto out; 3247 ret = get_cur_path(sctx, parent_ino, parent_gen, 3248 from_path); 3249 if (ret < 0) 3250 goto out; 3251 ret = fs_path_add_path(from_path, name); 3252 } 3253 if (ret < 0) 3254 goto out; 3255 3256 sctx->send_progress = sctx->cur_ino + 1; 3257 ret = path_loop(sctx, name, pm->ino, pm->gen, &ancestor); 3258 if (ret < 0) 3259 goto out; 3260 if (ret) { 3261 LIST_HEAD(deleted_refs); 3262 ASSERT(ancestor > BTRFS_FIRST_FREE_OBJECTID); 3263 ret = add_pending_dir_move(sctx, pm->ino, pm->gen, ancestor, 3264 &pm->update_refs, &deleted_refs, 3265 is_orphan); 3266 if (ret < 0) 3267 goto out; 3268 if (rmdir_ino) { 3269 dm = get_waiting_dir_move(sctx, pm->ino); 3270 ASSERT(dm); 3271 dm->rmdir_ino = rmdir_ino; 3272 dm->rmdir_gen = rmdir_gen; 3273 } 3274 goto out; 3275 } 3276 fs_path_reset(name); 3277 to_path = name; 3278 name = NULL; 3279 ret = get_cur_path(sctx, pm->ino, pm->gen, to_path); 3280 if (ret < 0) 3281 goto out; 3282 3283 ret = send_rename(sctx, from_path, to_path); 3284 if (ret < 0) 3285 goto out; 3286 3287 if (rmdir_ino) { 3288 struct orphan_dir_info *odi; 3289 u64 gen; 3290 3291 odi = get_orphan_dir_info(sctx, rmdir_ino, rmdir_gen); 3292 if (!odi) { 3293 /* already deleted */ 3294 goto finish; 3295 } 3296 gen = odi->gen; 3297 3298 ret = can_rmdir(sctx, rmdir_ino, gen, sctx->cur_ino); 3299 if (ret < 0) 3300 goto out; 3301 if (!ret) 3302 goto finish; 3303 3304 name = fs_path_alloc(); 3305 if (!name) { 3306 ret = -ENOMEM; 3307 goto out; 3308 } 3309 ret = get_cur_path(sctx, rmdir_ino, gen, name); 3310 if (ret < 0) 3311 goto out; 3312 ret = send_rmdir(sctx, name); 3313 if (ret < 0) 3314 goto out; 3315 } 3316 3317 finish: 3318 ret = send_utimes(sctx, pm->ino, pm->gen); 3319 if (ret < 0) 3320 goto out; 3321 3322 /* 3323 * After rename/move, need to update the utimes of both new parent(s) 3324 * and old parent(s). 3325 */ 3326 list_for_each_entry(cur, &pm->update_refs, list) { 3327 /* 3328 * The parent inode might have been deleted in the send snapshot 3329 */ 3330 ret = get_inode_info(sctx->send_root, cur->dir, NULL, 3331 NULL, NULL, NULL, NULL, NULL); 3332 if (ret == -ENOENT) { 3333 ret = 0; 3334 continue; 3335 } 3336 if (ret < 0) 3337 goto out; 3338 3339 ret = send_utimes(sctx, cur->dir, cur->dir_gen); 3340 if (ret < 0) 3341 goto out; 3342 } 3343 3344 out: 3345 fs_path_free(name); 3346 fs_path_free(from_path); 3347 fs_path_free(to_path); 3348 sctx->send_progress = orig_progress; 3349 3350 return ret; 3351 } 3352 3353 static void free_pending_move(struct send_ctx *sctx, struct pending_dir_move *m) 3354 { 3355 if (!list_empty(&m->list)) 3356 list_del(&m->list); 3357 if (!RB_EMPTY_NODE(&m->node)) 3358 rb_erase(&m->node, &sctx->pending_dir_moves); 3359 __free_recorded_refs(&m->update_refs); 3360 kfree(m); 3361 } 3362 3363 static void tail_append_pending_moves(struct send_ctx *sctx, 3364 struct pending_dir_move *moves, 3365 struct list_head *stack) 3366 { 3367 if (list_empty(&moves->list)) { 3368 list_add_tail(&moves->list, stack); 3369 } else { 3370 LIST_HEAD(list); 3371 list_splice_init(&moves->list, &list); 3372 list_add_tail(&moves->list, stack); 3373 list_splice_tail(&list, stack); 3374 } 3375 if (!RB_EMPTY_NODE(&moves->node)) { 3376 rb_erase(&moves->node, &sctx->pending_dir_moves); 3377 RB_CLEAR_NODE(&moves->node); 3378 } 3379 } 3380 3381 static int apply_children_dir_moves(struct send_ctx *sctx) 3382 { 3383 struct pending_dir_move *pm; 3384 struct list_head stack; 3385 u64 parent_ino = sctx->cur_ino; 3386 int ret = 0; 3387 3388 pm = get_pending_dir_moves(sctx, parent_ino); 3389 if (!pm) 3390 return 0; 3391 3392 INIT_LIST_HEAD(&stack); 3393 tail_append_pending_moves(sctx, pm, &stack); 3394 3395 while (!list_empty(&stack)) { 3396 pm = list_first_entry(&stack, struct pending_dir_move, list); 3397 parent_ino = pm->ino; 3398 ret = apply_dir_move(sctx, pm); 3399 free_pending_move(sctx, pm); 3400 if (ret) 3401 goto out; 3402 pm = get_pending_dir_moves(sctx, parent_ino); 3403 if (pm) 3404 tail_append_pending_moves(sctx, pm, &stack); 3405 } 3406 return 0; 3407 3408 out: 3409 while (!list_empty(&stack)) { 3410 pm = list_first_entry(&stack, struct pending_dir_move, list); 3411 free_pending_move(sctx, pm); 3412 } 3413 return ret; 3414 } 3415 3416 /* 3417 * We might need to delay a directory rename even when no ancestor directory 3418 * (in the send root) with a higher inode number than ours (sctx->cur_ino) was 3419 * renamed. This happens when we rename a directory to the old name (the name 3420 * in the parent root) of some other unrelated directory that got its rename 3421 * delayed due to some ancestor with higher number that got renamed. 3422 * 3423 * Example: 3424 * 3425 * Parent snapshot: 3426 * . (ino 256) 3427 * |---- a/ (ino 257) 3428 * | |---- file (ino 260) 3429 * | 3430 * |---- b/ (ino 258) 3431 * |---- c/ (ino 259) 3432 * 3433 * Send snapshot: 3434 * . (ino 256) 3435 * |---- a/ (ino 258) 3436 * |---- x/ (ino 259) 3437 * |---- y/ (ino 257) 3438 * |----- file (ino 260) 3439 * 3440 * Here we can not rename 258 from 'b' to 'a' without the rename of inode 257 3441 * from 'a' to 'x/y' happening first, which in turn depends on the rename of 3442 * inode 259 from 'c' to 'x'. So the order of rename commands the send stream 3443 * must issue is: 3444 * 3445 * 1 - rename 259 from 'c' to 'x' 3446 * 2 - rename 257 from 'a' to 'x/y' 3447 * 3 - rename 258 from 'b' to 'a' 3448 * 3449 * Returns 1 if the rename of sctx->cur_ino needs to be delayed, 0 if it can 3450 * be done right away and < 0 on error. 3451 */ 3452 static int wait_for_dest_dir_move(struct send_ctx *sctx, 3453 struct recorded_ref *parent_ref, 3454 const bool is_orphan) 3455 { 3456 struct btrfs_fs_info *fs_info = sctx->parent_root->fs_info; 3457 struct btrfs_path *path; 3458 struct btrfs_key key; 3459 struct btrfs_key di_key; 3460 struct btrfs_dir_item *di; 3461 u64 left_gen; 3462 u64 right_gen; 3463 int ret = 0; 3464 struct waiting_dir_move *wdm; 3465 3466 if (RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) 3467 return 0; 3468 3469 path = alloc_path_for_send(); 3470 if (!path) 3471 return -ENOMEM; 3472 3473 key.objectid = parent_ref->dir; 3474 key.type = BTRFS_DIR_ITEM_KEY; 3475 key.offset = btrfs_name_hash(parent_ref->name, parent_ref->name_len); 3476 3477 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0); 3478 if (ret < 0) { 3479 goto out; 3480 } else if (ret > 0) { 3481 ret = 0; 3482 goto out; 3483 } 3484 3485 di = btrfs_match_dir_item_name(fs_info, path, parent_ref->name, 3486 parent_ref->name_len); 3487 if (!di) { 3488 ret = 0; 3489 goto out; 3490 } 3491 /* 3492 * di_key.objectid has the number of the inode that has a dentry in the 3493 * parent directory with the same name that sctx->cur_ino is being 3494 * renamed to. We need to check if that inode is in the send root as 3495 * well and if it is currently marked as an inode with a pending rename, 3496 * if it is, we need to delay the rename of sctx->cur_ino as well, so 3497 * that it happens after that other inode is renamed. 3498 */ 3499 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &di_key); 3500 if (di_key.type != BTRFS_INODE_ITEM_KEY) { 3501 ret = 0; 3502 goto out; 3503 } 3504 3505 ret = get_inode_info(sctx->parent_root, di_key.objectid, NULL, 3506 &left_gen, NULL, NULL, NULL, NULL); 3507 if (ret < 0) 3508 goto out; 3509 ret = get_inode_info(sctx->send_root, di_key.objectid, NULL, 3510 &right_gen, NULL, NULL, NULL, NULL); 3511 if (ret < 0) { 3512 if (ret == -ENOENT) 3513 ret = 0; 3514 goto out; 3515 } 3516 3517 /* Different inode, no need to delay the rename of sctx->cur_ino */ 3518 if (right_gen != left_gen) { 3519 ret = 0; 3520 goto out; 3521 } 3522 3523 wdm = get_waiting_dir_move(sctx, di_key.objectid); 3524 if (wdm && !wdm->orphanized) { 3525 ret = add_pending_dir_move(sctx, 3526 sctx->cur_ino, 3527 sctx->cur_inode_gen, 3528 di_key.objectid, 3529 &sctx->new_refs, 3530 &sctx->deleted_refs, 3531 is_orphan); 3532 if (!ret) 3533 ret = 1; 3534 } 3535 out: 3536 btrfs_free_path(path); 3537 return ret; 3538 } 3539 3540 /* 3541 * Check if inode ino2, or any of its ancestors, is inode ino1. 3542 * Return 1 if true, 0 if false and < 0 on error. 3543 */ 3544 static int check_ino_in_path(struct btrfs_root *root, 3545 const u64 ino1, 3546 const u64 ino1_gen, 3547 const u64 ino2, 3548 const u64 ino2_gen, 3549 struct fs_path *fs_path) 3550 { 3551 u64 ino = ino2; 3552 3553 if (ino1 == ino2) 3554 return ino1_gen == ino2_gen; 3555 3556 while (ino > BTRFS_FIRST_FREE_OBJECTID) { 3557 u64 parent; 3558 u64 parent_gen; 3559 int ret; 3560 3561 fs_path_reset(fs_path); 3562 ret = get_first_ref(root, ino, &parent, &parent_gen, fs_path); 3563 if (ret < 0) 3564 return ret; 3565 if (parent == ino1) 3566 return parent_gen == ino1_gen; 3567 ino = parent; 3568 } 3569 return 0; 3570 } 3571 3572 /* 3573 * Check if ino ino1 is an ancestor of inode ino2 in the given root for any 3574 * possible path (in case ino2 is not a directory and has multiple hard links). 3575 * Return 1 if true, 0 if false and < 0 on error. 3576 */ 3577 static int is_ancestor(struct btrfs_root *root, 3578 const u64 ino1, 3579 const u64 ino1_gen, 3580 const u64 ino2, 3581 struct fs_path *fs_path) 3582 { 3583 bool free_fs_path = false; 3584 int ret = 0; 3585 struct btrfs_path *path = NULL; 3586 struct btrfs_key key; 3587 3588 if (!fs_path) { 3589 fs_path = fs_path_alloc(); 3590 if (!fs_path) 3591 return -ENOMEM; 3592 free_fs_path = true; 3593 } 3594 3595 path = alloc_path_for_send(); 3596 if (!path) { 3597 ret = -ENOMEM; 3598 goto out; 3599 } 3600 3601 key.objectid = ino2; 3602 key.type = BTRFS_INODE_REF_KEY; 3603 key.offset = 0; 3604 3605 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 3606 if (ret < 0) 3607 goto out; 3608 3609 while (true) { 3610 struct extent_buffer *leaf = path->nodes[0]; 3611 int slot = path->slots[0]; 3612 u32 cur_offset = 0; 3613 u32 item_size; 3614 3615 if (slot >= btrfs_header_nritems(leaf)) { 3616 ret = btrfs_next_leaf(root, path); 3617 if (ret < 0) 3618 goto out; 3619 if (ret > 0) 3620 break; 3621 continue; 3622 } 3623 3624 btrfs_item_key_to_cpu(leaf, &key, slot); 3625 if (key.objectid != ino2) 3626 break; 3627 if (key.type != BTRFS_INODE_REF_KEY && 3628 key.type != BTRFS_INODE_EXTREF_KEY) 3629 break; 3630 3631 item_size = btrfs_item_size_nr(leaf, slot); 3632 while (cur_offset < item_size) { 3633 u64 parent; 3634 u64 parent_gen; 3635 3636 if (key.type == BTRFS_INODE_EXTREF_KEY) { 3637 unsigned long ptr; 3638 struct btrfs_inode_extref *extref; 3639 3640 ptr = btrfs_item_ptr_offset(leaf, slot); 3641 extref = (struct btrfs_inode_extref *) 3642 (ptr + cur_offset); 3643 parent = btrfs_inode_extref_parent(leaf, 3644 extref); 3645 cur_offset += sizeof(*extref); 3646 cur_offset += btrfs_inode_extref_name_len(leaf, 3647 extref); 3648 } else { 3649 parent = key.offset; 3650 cur_offset = item_size; 3651 } 3652 3653 ret = get_inode_info(root, parent, NULL, &parent_gen, 3654 NULL, NULL, NULL, NULL); 3655 if (ret < 0) 3656 goto out; 3657 ret = check_ino_in_path(root, ino1, ino1_gen, 3658 parent, parent_gen, fs_path); 3659 if (ret) 3660 goto out; 3661 } 3662 path->slots[0]++; 3663 } 3664 ret = 0; 3665 out: 3666 btrfs_free_path(path); 3667 if (free_fs_path) 3668 fs_path_free(fs_path); 3669 return ret; 3670 } 3671 3672 static int wait_for_parent_move(struct send_ctx *sctx, 3673 struct recorded_ref *parent_ref, 3674 const bool is_orphan) 3675 { 3676 int ret = 0; 3677 u64 ino = parent_ref->dir; 3678 u64 ino_gen = parent_ref->dir_gen; 3679 u64 parent_ino_before, parent_ino_after; 3680 struct fs_path *path_before = NULL; 3681 struct fs_path *path_after = NULL; 3682 int len1, len2; 3683 3684 path_after = fs_path_alloc(); 3685 path_before = fs_path_alloc(); 3686 if (!path_after || !path_before) { 3687 ret = -ENOMEM; 3688 goto out; 3689 } 3690 3691 /* 3692 * Our current directory inode may not yet be renamed/moved because some 3693 * ancestor (immediate or not) has to be renamed/moved first. So find if 3694 * such ancestor exists and make sure our own rename/move happens after 3695 * that ancestor is processed to avoid path build infinite loops (done 3696 * at get_cur_path()). 3697 */ 3698 while (ino > BTRFS_FIRST_FREE_OBJECTID) { 3699 u64 parent_ino_after_gen; 3700 3701 if (is_waiting_for_move(sctx, ino)) { 3702 /* 3703 * If the current inode is an ancestor of ino in the 3704 * parent root, we need to delay the rename of the 3705 * current inode, otherwise don't delayed the rename 3706 * because we can end up with a circular dependency 3707 * of renames, resulting in some directories never 3708 * getting the respective rename operations issued in 3709 * the send stream or getting into infinite path build 3710 * loops. 3711 */ 3712 ret = is_ancestor(sctx->parent_root, 3713 sctx->cur_ino, sctx->cur_inode_gen, 3714 ino, path_before); 3715 if (ret) 3716 break; 3717 } 3718 3719 fs_path_reset(path_before); 3720 fs_path_reset(path_after); 3721 3722 ret = get_first_ref(sctx->send_root, ino, &parent_ino_after, 3723 &parent_ino_after_gen, path_after); 3724 if (ret < 0) 3725 goto out; 3726 ret = get_first_ref(sctx->parent_root, ino, &parent_ino_before, 3727 NULL, path_before); 3728 if (ret < 0 && ret != -ENOENT) { 3729 goto out; 3730 } else if (ret == -ENOENT) { 3731 ret = 0; 3732 break; 3733 } 3734 3735 len1 = fs_path_len(path_before); 3736 len2 = fs_path_len(path_after); 3737 if (ino > sctx->cur_ino && 3738 (parent_ino_before != parent_ino_after || len1 != len2 || 3739 memcmp(path_before->start, path_after->start, len1))) { 3740 u64 parent_ino_gen; 3741 3742 ret = get_inode_info(sctx->parent_root, ino, NULL, 3743 &parent_ino_gen, NULL, NULL, NULL, 3744 NULL); 3745 if (ret < 0) 3746 goto out; 3747 if (ino_gen == parent_ino_gen) { 3748 ret = 1; 3749 break; 3750 } 3751 } 3752 ino = parent_ino_after; 3753 ino_gen = parent_ino_after_gen; 3754 } 3755 3756 out: 3757 fs_path_free(path_before); 3758 fs_path_free(path_after); 3759 3760 if (ret == 1) { 3761 ret = add_pending_dir_move(sctx, 3762 sctx->cur_ino, 3763 sctx->cur_inode_gen, 3764 ino, 3765 &sctx->new_refs, 3766 &sctx->deleted_refs, 3767 is_orphan); 3768 if (!ret) 3769 ret = 1; 3770 } 3771 3772 return ret; 3773 } 3774 3775 static int update_ref_path(struct send_ctx *sctx, struct recorded_ref *ref) 3776 { 3777 int ret; 3778 struct fs_path *new_path; 3779 3780 /* 3781 * Our reference's name member points to its full_path member string, so 3782 * we use here a new path. 3783 */ 3784 new_path = fs_path_alloc(); 3785 if (!new_path) 3786 return -ENOMEM; 3787 3788 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, new_path); 3789 if (ret < 0) { 3790 fs_path_free(new_path); 3791 return ret; 3792 } 3793 ret = fs_path_add(new_path, ref->name, ref->name_len); 3794 if (ret < 0) { 3795 fs_path_free(new_path); 3796 return ret; 3797 } 3798 3799 fs_path_free(ref->full_path); 3800 set_ref_path(ref, new_path); 3801 3802 return 0; 3803 } 3804 3805 /* 3806 * When processing the new references for an inode we may orphanize an existing 3807 * directory inode because its old name conflicts with one of the new references 3808 * of the current inode. Later, when processing another new reference of our 3809 * inode, we might need to orphanize another inode, but the path we have in the 3810 * reference reflects the pre-orphanization name of the directory we previously 3811 * orphanized. For example: 3812 * 3813 * parent snapshot looks like: 3814 * 3815 * . (ino 256) 3816 * |----- f1 (ino 257) 3817 * |----- f2 (ino 258) 3818 * |----- d1/ (ino 259) 3819 * |----- d2/ (ino 260) 3820 * 3821 * send snapshot looks like: 3822 * 3823 * . (ino 256) 3824 * |----- d1 (ino 258) 3825 * |----- f2/ (ino 259) 3826 * |----- f2_link/ (ino 260) 3827 * | |----- f1 (ino 257) 3828 * | 3829 * |----- d2 (ino 258) 3830 * 3831 * When processing inode 257 we compute the name for inode 259 as "d1", and we 3832 * cache it in the name cache. Later when we start processing inode 258, when 3833 * collecting all its new references we set a full path of "d1/d2" for its new 3834 * reference with name "d2". When we start processing the new references we 3835 * start by processing the new reference with name "d1", and this results in 3836 * orphanizing inode 259, since its old reference causes a conflict. Then we 3837 * move on the next new reference, with name "d2", and we find out we must 3838 * orphanize inode 260, as its old reference conflicts with ours - but for the 3839 * orphanization we use a source path corresponding to the path we stored in the 3840 * new reference, which is "d1/d2" and not "o259-6-0/d2" - this makes the 3841 * receiver fail since the path component "d1/" no longer exists, it was renamed 3842 * to "o259-6-0/" when processing the previous new reference. So in this case we 3843 * must recompute the path in the new reference and use it for the new 3844 * orphanization operation. 3845 */ 3846 static int refresh_ref_path(struct send_ctx *sctx, struct recorded_ref *ref) 3847 { 3848 char *name; 3849 int ret; 3850 3851 name = kmemdup(ref->name, ref->name_len, GFP_KERNEL); 3852 if (!name) 3853 return -ENOMEM; 3854 3855 fs_path_reset(ref->full_path); 3856 ret = get_cur_path(sctx, ref->dir, ref->dir_gen, ref->full_path); 3857 if (ret < 0) 3858 goto out; 3859 3860 ret = fs_path_add(ref->full_path, name, ref->name_len); 3861 if (ret < 0) 3862 goto out; 3863 3864 /* Update the reference's base name pointer. */ 3865 set_ref_path(ref, ref->full_path); 3866 out: 3867 kfree(name); 3868 return ret; 3869 } 3870 3871 /* 3872 * This does all the move/link/unlink/rmdir magic. 3873 */ 3874 static int process_recorded_refs(struct send_ctx *sctx, int *pending_move) 3875 { 3876 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 3877 int ret = 0; 3878 struct recorded_ref *cur; 3879 struct recorded_ref *cur2; 3880 struct list_head check_dirs; 3881 struct fs_path *valid_path = NULL; 3882 u64 ow_inode = 0; 3883 u64 ow_gen; 3884 u64 ow_mode; 3885 int did_overwrite = 0; 3886 int is_orphan = 0; 3887 u64 last_dir_ino_rm = 0; 3888 bool can_rename = true; 3889 bool orphanized_dir = false; 3890 bool orphanized_ancestor = false; 3891 3892 btrfs_debug(fs_info, "process_recorded_refs %llu", sctx->cur_ino); 3893 3894 /* 3895 * This should never happen as the root dir always has the same ref 3896 * which is always '..' 3897 */ 3898 BUG_ON(sctx->cur_ino <= BTRFS_FIRST_FREE_OBJECTID); 3899 INIT_LIST_HEAD(&check_dirs); 3900 3901 valid_path = fs_path_alloc(); 3902 if (!valid_path) { 3903 ret = -ENOMEM; 3904 goto out; 3905 } 3906 3907 /* 3908 * First, check if the first ref of the current inode was overwritten 3909 * before. If yes, we know that the current inode was already orphanized 3910 * and thus use the orphan name. If not, we can use get_cur_path to 3911 * get the path of the first ref as it would like while receiving at 3912 * this point in time. 3913 * New inodes are always orphan at the beginning, so force to use the 3914 * orphan name in this case. 3915 * The first ref is stored in valid_path and will be updated if it 3916 * gets moved around. 3917 */ 3918 if (!sctx->cur_inode_new) { 3919 ret = did_overwrite_first_ref(sctx, sctx->cur_ino, 3920 sctx->cur_inode_gen); 3921 if (ret < 0) 3922 goto out; 3923 if (ret) 3924 did_overwrite = 1; 3925 } 3926 if (sctx->cur_inode_new || did_overwrite) { 3927 ret = gen_unique_name(sctx, sctx->cur_ino, 3928 sctx->cur_inode_gen, valid_path); 3929 if (ret < 0) 3930 goto out; 3931 is_orphan = 1; 3932 } else { 3933 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, 3934 valid_path); 3935 if (ret < 0) 3936 goto out; 3937 } 3938 3939 /* 3940 * Before doing any rename and link operations, do a first pass on the 3941 * new references to orphanize any unprocessed inodes that may have a 3942 * reference that conflicts with one of the new references of the current 3943 * inode. This needs to happen first because a new reference may conflict 3944 * with the old reference of a parent directory, so we must make sure 3945 * that the path used for link and rename commands don't use an 3946 * orphanized name when an ancestor was not yet orphanized. 3947 * 3948 * Example: 3949 * 3950 * Parent snapshot: 3951 * 3952 * . (ino 256) 3953 * |----- testdir/ (ino 259) 3954 * | |----- a (ino 257) 3955 * | 3956 * |----- b (ino 258) 3957 * 3958 * Send snapshot: 3959 * 3960 * . (ino 256) 3961 * |----- testdir_2/ (ino 259) 3962 * | |----- a (ino 260) 3963 * | 3964 * |----- testdir (ino 257) 3965 * |----- b (ino 257) 3966 * |----- b2 (ino 258) 3967 * 3968 * Processing the new reference for inode 257 with name "b" may happen 3969 * before processing the new reference with name "testdir". If so, we 3970 * must make sure that by the time we send a link command to create the 3971 * hard link "b", inode 259 was already orphanized, since the generated 3972 * path in "valid_path" already contains the orphanized name for 259. 3973 * We are processing inode 257, so only later when processing 259 we do 3974 * the rename operation to change its temporary (orphanized) name to 3975 * "testdir_2". 3976 */ 3977 list_for_each_entry(cur, &sctx->new_refs, list) { 3978 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen); 3979 if (ret < 0) 3980 goto out; 3981 if (ret == inode_state_will_create) 3982 continue; 3983 3984 /* 3985 * Check if this new ref would overwrite the first ref of another 3986 * unprocessed inode. If yes, orphanize the overwritten inode. 3987 * If we find an overwritten ref that is not the first ref, 3988 * simply unlink it. 3989 */ 3990 ret = will_overwrite_ref(sctx, cur->dir, cur->dir_gen, 3991 cur->name, cur->name_len, 3992 &ow_inode, &ow_gen, &ow_mode); 3993 if (ret < 0) 3994 goto out; 3995 if (ret) { 3996 ret = is_first_ref(sctx->parent_root, 3997 ow_inode, cur->dir, cur->name, 3998 cur->name_len); 3999 if (ret < 0) 4000 goto out; 4001 if (ret) { 4002 struct name_cache_entry *nce; 4003 struct waiting_dir_move *wdm; 4004 4005 if (orphanized_dir) { 4006 ret = refresh_ref_path(sctx, cur); 4007 if (ret < 0) 4008 goto out; 4009 } 4010 4011 ret = orphanize_inode(sctx, ow_inode, ow_gen, 4012 cur->full_path); 4013 if (ret < 0) 4014 goto out; 4015 if (S_ISDIR(ow_mode)) 4016 orphanized_dir = true; 4017 4018 /* 4019 * If ow_inode has its rename operation delayed 4020 * make sure that its orphanized name is used in 4021 * the source path when performing its rename 4022 * operation. 4023 */ 4024 if (is_waiting_for_move(sctx, ow_inode)) { 4025 wdm = get_waiting_dir_move(sctx, 4026 ow_inode); 4027 ASSERT(wdm); 4028 wdm->orphanized = true; 4029 } 4030 4031 /* 4032 * Make sure we clear our orphanized inode's 4033 * name from the name cache. This is because the 4034 * inode ow_inode might be an ancestor of some 4035 * other inode that will be orphanized as well 4036 * later and has an inode number greater than 4037 * sctx->send_progress. We need to prevent 4038 * future name lookups from using the old name 4039 * and get instead the orphan name. 4040 */ 4041 nce = name_cache_search(sctx, ow_inode, ow_gen); 4042 if (nce) { 4043 name_cache_delete(sctx, nce); 4044 kfree(nce); 4045 } 4046 4047 /* 4048 * ow_inode might currently be an ancestor of 4049 * cur_ino, therefore compute valid_path (the 4050 * current path of cur_ino) again because it 4051 * might contain the pre-orphanization name of 4052 * ow_inode, which is no longer valid. 4053 */ 4054 ret = is_ancestor(sctx->parent_root, 4055 ow_inode, ow_gen, 4056 sctx->cur_ino, NULL); 4057 if (ret > 0) { 4058 orphanized_ancestor = true; 4059 fs_path_reset(valid_path); 4060 ret = get_cur_path(sctx, sctx->cur_ino, 4061 sctx->cur_inode_gen, 4062 valid_path); 4063 } 4064 if (ret < 0) 4065 goto out; 4066 } else { 4067 ret = send_unlink(sctx, cur->full_path); 4068 if (ret < 0) 4069 goto out; 4070 } 4071 } 4072 4073 } 4074 4075 list_for_each_entry(cur, &sctx->new_refs, list) { 4076 /* 4077 * We may have refs where the parent directory does not exist 4078 * yet. This happens if the parent directories inum is higher 4079 * than the current inum. To handle this case, we create the 4080 * parent directory out of order. But we need to check if this 4081 * did already happen before due to other refs in the same dir. 4082 */ 4083 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen); 4084 if (ret < 0) 4085 goto out; 4086 if (ret == inode_state_will_create) { 4087 ret = 0; 4088 /* 4089 * First check if any of the current inodes refs did 4090 * already create the dir. 4091 */ 4092 list_for_each_entry(cur2, &sctx->new_refs, list) { 4093 if (cur == cur2) 4094 break; 4095 if (cur2->dir == cur->dir) { 4096 ret = 1; 4097 break; 4098 } 4099 } 4100 4101 /* 4102 * If that did not happen, check if a previous inode 4103 * did already create the dir. 4104 */ 4105 if (!ret) 4106 ret = did_create_dir(sctx, cur->dir); 4107 if (ret < 0) 4108 goto out; 4109 if (!ret) { 4110 ret = send_create_inode(sctx, cur->dir); 4111 if (ret < 0) 4112 goto out; 4113 } 4114 } 4115 4116 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root) { 4117 ret = wait_for_dest_dir_move(sctx, cur, is_orphan); 4118 if (ret < 0) 4119 goto out; 4120 if (ret == 1) { 4121 can_rename = false; 4122 *pending_move = 1; 4123 } 4124 } 4125 4126 if (S_ISDIR(sctx->cur_inode_mode) && sctx->parent_root && 4127 can_rename) { 4128 ret = wait_for_parent_move(sctx, cur, is_orphan); 4129 if (ret < 0) 4130 goto out; 4131 if (ret == 1) { 4132 can_rename = false; 4133 *pending_move = 1; 4134 } 4135 } 4136 4137 /* 4138 * link/move the ref to the new place. If we have an orphan 4139 * inode, move it and update valid_path. If not, link or move 4140 * it depending on the inode mode. 4141 */ 4142 if (is_orphan && can_rename) { 4143 ret = send_rename(sctx, valid_path, cur->full_path); 4144 if (ret < 0) 4145 goto out; 4146 is_orphan = 0; 4147 ret = fs_path_copy(valid_path, cur->full_path); 4148 if (ret < 0) 4149 goto out; 4150 } else if (can_rename) { 4151 if (S_ISDIR(sctx->cur_inode_mode)) { 4152 /* 4153 * Dirs can't be linked, so move it. For moved 4154 * dirs, we always have one new and one deleted 4155 * ref. The deleted ref is ignored later. 4156 */ 4157 ret = send_rename(sctx, valid_path, 4158 cur->full_path); 4159 if (!ret) 4160 ret = fs_path_copy(valid_path, 4161 cur->full_path); 4162 if (ret < 0) 4163 goto out; 4164 } else { 4165 /* 4166 * We might have previously orphanized an inode 4167 * which is an ancestor of our current inode, 4168 * so our reference's full path, which was 4169 * computed before any such orphanizations, must 4170 * be updated. 4171 */ 4172 if (orphanized_dir) { 4173 ret = update_ref_path(sctx, cur); 4174 if (ret < 0) 4175 goto out; 4176 } 4177 ret = send_link(sctx, cur->full_path, 4178 valid_path); 4179 if (ret < 0) 4180 goto out; 4181 } 4182 } 4183 ret = dup_ref(cur, &check_dirs); 4184 if (ret < 0) 4185 goto out; 4186 } 4187 4188 if (S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_deleted) { 4189 /* 4190 * Check if we can already rmdir the directory. If not, 4191 * orphanize it. For every dir item inside that gets deleted 4192 * later, we do this check again and rmdir it then if possible. 4193 * See the use of check_dirs for more details. 4194 */ 4195 ret = can_rmdir(sctx, sctx->cur_ino, sctx->cur_inode_gen, 4196 sctx->cur_ino); 4197 if (ret < 0) 4198 goto out; 4199 if (ret) { 4200 ret = send_rmdir(sctx, valid_path); 4201 if (ret < 0) 4202 goto out; 4203 } else if (!is_orphan) { 4204 ret = orphanize_inode(sctx, sctx->cur_ino, 4205 sctx->cur_inode_gen, valid_path); 4206 if (ret < 0) 4207 goto out; 4208 is_orphan = 1; 4209 } 4210 4211 list_for_each_entry(cur, &sctx->deleted_refs, list) { 4212 ret = dup_ref(cur, &check_dirs); 4213 if (ret < 0) 4214 goto out; 4215 } 4216 } else if (S_ISDIR(sctx->cur_inode_mode) && 4217 !list_empty(&sctx->deleted_refs)) { 4218 /* 4219 * We have a moved dir. Add the old parent to check_dirs 4220 */ 4221 cur = list_entry(sctx->deleted_refs.next, struct recorded_ref, 4222 list); 4223 ret = dup_ref(cur, &check_dirs); 4224 if (ret < 0) 4225 goto out; 4226 } else if (!S_ISDIR(sctx->cur_inode_mode)) { 4227 /* 4228 * We have a non dir inode. Go through all deleted refs and 4229 * unlink them if they were not already overwritten by other 4230 * inodes. 4231 */ 4232 list_for_each_entry(cur, &sctx->deleted_refs, list) { 4233 ret = did_overwrite_ref(sctx, cur->dir, cur->dir_gen, 4234 sctx->cur_ino, sctx->cur_inode_gen, 4235 cur->name, cur->name_len); 4236 if (ret < 0) 4237 goto out; 4238 if (!ret) { 4239 /* 4240 * If we orphanized any ancestor before, we need 4241 * to recompute the full path for deleted names, 4242 * since any such path was computed before we 4243 * processed any references and orphanized any 4244 * ancestor inode. 4245 */ 4246 if (orphanized_ancestor) { 4247 ret = update_ref_path(sctx, cur); 4248 if (ret < 0) 4249 goto out; 4250 } 4251 ret = send_unlink(sctx, cur->full_path); 4252 if (ret < 0) 4253 goto out; 4254 } 4255 ret = dup_ref(cur, &check_dirs); 4256 if (ret < 0) 4257 goto out; 4258 } 4259 /* 4260 * If the inode is still orphan, unlink the orphan. This may 4261 * happen when a previous inode did overwrite the first ref 4262 * of this inode and no new refs were added for the current 4263 * inode. Unlinking does not mean that the inode is deleted in 4264 * all cases. There may still be links to this inode in other 4265 * places. 4266 */ 4267 if (is_orphan) { 4268 ret = send_unlink(sctx, valid_path); 4269 if (ret < 0) 4270 goto out; 4271 } 4272 } 4273 4274 /* 4275 * We did collect all parent dirs where cur_inode was once located. We 4276 * now go through all these dirs and check if they are pending for 4277 * deletion and if it's finally possible to perform the rmdir now. 4278 * We also update the inode stats of the parent dirs here. 4279 */ 4280 list_for_each_entry(cur, &check_dirs, list) { 4281 /* 4282 * In case we had refs into dirs that were not processed yet, 4283 * we don't need to do the utime and rmdir logic for these dirs. 4284 * The dir will be processed later. 4285 */ 4286 if (cur->dir > sctx->cur_ino) 4287 continue; 4288 4289 ret = get_cur_inode_state(sctx, cur->dir, cur->dir_gen); 4290 if (ret < 0) 4291 goto out; 4292 4293 if (ret == inode_state_did_create || 4294 ret == inode_state_no_change) { 4295 /* TODO delayed utimes */ 4296 ret = send_utimes(sctx, cur->dir, cur->dir_gen); 4297 if (ret < 0) 4298 goto out; 4299 } else if (ret == inode_state_did_delete && 4300 cur->dir != last_dir_ino_rm) { 4301 ret = can_rmdir(sctx, cur->dir, cur->dir_gen, 4302 sctx->cur_ino); 4303 if (ret < 0) 4304 goto out; 4305 if (ret) { 4306 ret = get_cur_path(sctx, cur->dir, 4307 cur->dir_gen, valid_path); 4308 if (ret < 0) 4309 goto out; 4310 ret = send_rmdir(sctx, valid_path); 4311 if (ret < 0) 4312 goto out; 4313 last_dir_ino_rm = cur->dir; 4314 } 4315 } 4316 } 4317 4318 ret = 0; 4319 4320 out: 4321 __free_recorded_refs(&check_dirs); 4322 free_recorded_refs(sctx); 4323 fs_path_free(valid_path); 4324 return ret; 4325 } 4326 4327 static int record_ref(struct btrfs_root *root, u64 dir, struct fs_path *name, 4328 void *ctx, struct list_head *refs) 4329 { 4330 int ret = 0; 4331 struct send_ctx *sctx = ctx; 4332 struct fs_path *p; 4333 u64 gen; 4334 4335 p = fs_path_alloc(); 4336 if (!p) 4337 return -ENOMEM; 4338 4339 ret = get_inode_info(root, dir, NULL, &gen, NULL, NULL, 4340 NULL, NULL); 4341 if (ret < 0) 4342 goto out; 4343 4344 ret = get_cur_path(sctx, dir, gen, p); 4345 if (ret < 0) 4346 goto out; 4347 ret = fs_path_add_path(p, name); 4348 if (ret < 0) 4349 goto out; 4350 4351 ret = __record_ref(refs, dir, gen, p); 4352 4353 out: 4354 if (ret) 4355 fs_path_free(p); 4356 return ret; 4357 } 4358 4359 static int __record_new_ref(int num, u64 dir, int index, 4360 struct fs_path *name, 4361 void *ctx) 4362 { 4363 struct send_ctx *sctx = ctx; 4364 return record_ref(sctx->send_root, dir, name, ctx, &sctx->new_refs); 4365 } 4366 4367 4368 static int __record_deleted_ref(int num, u64 dir, int index, 4369 struct fs_path *name, 4370 void *ctx) 4371 { 4372 struct send_ctx *sctx = ctx; 4373 return record_ref(sctx->parent_root, dir, name, ctx, 4374 &sctx->deleted_refs); 4375 } 4376 4377 static int record_new_ref(struct send_ctx *sctx) 4378 { 4379 int ret; 4380 4381 ret = iterate_inode_ref(sctx->send_root, sctx->left_path, 4382 sctx->cmp_key, 0, __record_new_ref, sctx); 4383 if (ret < 0) 4384 goto out; 4385 ret = 0; 4386 4387 out: 4388 return ret; 4389 } 4390 4391 static int record_deleted_ref(struct send_ctx *sctx) 4392 { 4393 int ret; 4394 4395 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path, 4396 sctx->cmp_key, 0, __record_deleted_ref, sctx); 4397 if (ret < 0) 4398 goto out; 4399 ret = 0; 4400 4401 out: 4402 return ret; 4403 } 4404 4405 struct find_ref_ctx { 4406 u64 dir; 4407 u64 dir_gen; 4408 struct btrfs_root *root; 4409 struct fs_path *name; 4410 int found_idx; 4411 }; 4412 4413 static int __find_iref(int num, u64 dir, int index, 4414 struct fs_path *name, 4415 void *ctx_) 4416 { 4417 struct find_ref_ctx *ctx = ctx_; 4418 u64 dir_gen; 4419 int ret; 4420 4421 if (dir == ctx->dir && fs_path_len(name) == fs_path_len(ctx->name) && 4422 strncmp(name->start, ctx->name->start, fs_path_len(name)) == 0) { 4423 /* 4424 * To avoid doing extra lookups we'll only do this if everything 4425 * else matches. 4426 */ 4427 ret = get_inode_info(ctx->root, dir, NULL, &dir_gen, NULL, 4428 NULL, NULL, NULL); 4429 if (ret) 4430 return ret; 4431 if (dir_gen != ctx->dir_gen) 4432 return 0; 4433 ctx->found_idx = num; 4434 return 1; 4435 } 4436 return 0; 4437 } 4438 4439 static int find_iref(struct btrfs_root *root, 4440 struct btrfs_path *path, 4441 struct btrfs_key *key, 4442 u64 dir, u64 dir_gen, struct fs_path *name) 4443 { 4444 int ret; 4445 struct find_ref_ctx ctx; 4446 4447 ctx.dir = dir; 4448 ctx.name = name; 4449 ctx.dir_gen = dir_gen; 4450 ctx.found_idx = -1; 4451 ctx.root = root; 4452 4453 ret = iterate_inode_ref(root, path, key, 0, __find_iref, &ctx); 4454 if (ret < 0) 4455 return ret; 4456 4457 if (ctx.found_idx == -1) 4458 return -ENOENT; 4459 4460 return ctx.found_idx; 4461 } 4462 4463 static int __record_changed_new_ref(int num, u64 dir, int index, 4464 struct fs_path *name, 4465 void *ctx) 4466 { 4467 u64 dir_gen; 4468 int ret; 4469 struct send_ctx *sctx = ctx; 4470 4471 ret = get_inode_info(sctx->send_root, dir, NULL, &dir_gen, NULL, 4472 NULL, NULL, NULL); 4473 if (ret) 4474 return ret; 4475 4476 ret = find_iref(sctx->parent_root, sctx->right_path, 4477 sctx->cmp_key, dir, dir_gen, name); 4478 if (ret == -ENOENT) 4479 ret = __record_new_ref(num, dir, index, name, sctx); 4480 else if (ret > 0) 4481 ret = 0; 4482 4483 return ret; 4484 } 4485 4486 static int __record_changed_deleted_ref(int num, u64 dir, int index, 4487 struct fs_path *name, 4488 void *ctx) 4489 { 4490 u64 dir_gen; 4491 int ret; 4492 struct send_ctx *sctx = ctx; 4493 4494 ret = get_inode_info(sctx->parent_root, dir, NULL, &dir_gen, NULL, 4495 NULL, NULL, NULL); 4496 if (ret) 4497 return ret; 4498 4499 ret = find_iref(sctx->send_root, sctx->left_path, sctx->cmp_key, 4500 dir, dir_gen, name); 4501 if (ret == -ENOENT) 4502 ret = __record_deleted_ref(num, dir, index, name, sctx); 4503 else if (ret > 0) 4504 ret = 0; 4505 4506 return ret; 4507 } 4508 4509 static int record_changed_ref(struct send_ctx *sctx) 4510 { 4511 int ret = 0; 4512 4513 ret = iterate_inode_ref(sctx->send_root, sctx->left_path, 4514 sctx->cmp_key, 0, __record_changed_new_ref, sctx); 4515 if (ret < 0) 4516 goto out; 4517 ret = iterate_inode_ref(sctx->parent_root, sctx->right_path, 4518 sctx->cmp_key, 0, __record_changed_deleted_ref, sctx); 4519 if (ret < 0) 4520 goto out; 4521 ret = 0; 4522 4523 out: 4524 return ret; 4525 } 4526 4527 /* 4528 * Record and process all refs at once. Needed when an inode changes the 4529 * generation number, which means that it was deleted and recreated. 4530 */ 4531 static int process_all_refs(struct send_ctx *sctx, 4532 enum btrfs_compare_tree_result cmd) 4533 { 4534 int ret; 4535 struct btrfs_root *root; 4536 struct btrfs_path *path; 4537 struct btrfs_key key; 4538 struct btrfs_key found_key; 4539 struct extent_buffer *eb; 4540 int slot; 4541 iterate_inode_ref_t cb; 4542 int pending_move = 0; 4543 4544 path = alloc_path_for_send(); 4545 if (!path) 4546 return -ENOMEM; 4547 4548 if (cmd == BTRFS_COMPARE_TREE_NEW) { 4549 root = sctx->send_root; 4550 cb = __record_new_ref; 4551 } else if (cmd == BTRFS_COMPARE_TREE_DELETED) { 4552 root = sctx->parent_root; 4553 cb = __record_deleted_ref; 4554 } else { 4555 btrfs_err(sctx->send_root->fs_info, 4556 "Wrong command %d in process_all_refs", cmd); 4557 ret = -EINVAL; 4558 goto out; 4559 } 4560 4561 key.objectid = sctx->cmp_key->objectid; 4562 key.type = BTRFS_INODE_REF_KEY; 4563 key.offset = 0; 4564 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 4565 if (ret < 0) 4566 goto out; 4567 4568 while (1) { 4569 eb = path->nodes[0]; 4570 slot = path->slots[0]; 4571 if (slot >= btrfs_header_nritems(eb)) { 4572 ret = btrfs_next_leaf(root, path); 4573 if (ret < 0) 4574 goto out; 4575 else if (ret > 0) 4576 break; 4577 continue; 4578 } 4579 4580 btrfs_item_key_to_cpu(eb, &found_key, slot); 4581 4582 if (found_key.objectid != key.objectid || 4583 (found_key.type != BTRFS_INODE_REF_KEY && 4584 found_key.type != BTRFS_INODE_EXTREF_KEY)) 4585 break; 4586 4587 ret = iterate_inode_ref(root, path, &found_key, 0, cb, sctx); 4588 if (ret < 0) 4589 goto out; 4590 4591 path->slots[0]++; 4592 } 4593 btrfs_release_path(path); 4594 4595 /* 4596 * We don't actually care about pending_move as we are simply 4597 * re-creating this inode and will be rename'ing it into place once we 4598 * rename the parent directory. 4599 */ 4600 ret = process_recorded_refs(sctx, &pending_move); 4601 out: 4602 btrfs_free_path(path); 4603 return ret; 4604 } 4605 4606 static int send_set_xattr(struct send_ctx *sctx, 4607 struct fs_path *path, 4608 const char *name, int name_len, 4609 const char *data, int data_len) 4610 { 4611 int ret = 0; 4612 4613 ret = begin_cmd(sctx, BTRFS_SEND_C_SET_XATTR); 4614 if (ret < 0) 4615 goto out; 4616 4617 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path); 4618 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len); 4619 TLV_PUT(sctx, BTRFS_SEND_A_XATTR_DATA, data, data_len); 4620 4621 ret = send_cmd(sctx); 4622 4623 tlv_put_failure: 4624 out: 4625 return ret; 4626 } 4627 4628 static int send_remove_xattr(struct send_ctx *sctx, 4629 struct fs_path *path, 4630 const char *name, int name_len) 4631 { 4632 int ret = 0; 4633 4634 ret = begin_cmd(sctx, BTRFS_SEND_C_REMOVE_XATTR); 4635 if (ret < 0) 4636 goto out; 4637 4638 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, path); 4639 TLV_PUT_STRING(sctx, BTRFS_SEND_A_XATTR_NAME, name, name_len); 4640 4641 ret = send_cmd(sctx); 4642 4643 tlv_put_failure: 4644 out: 4645 return ret; 4646 } 4647 4648 static int __process_new_xattr(int num, struct btrfs_key *di_key, 4649 const char *name, int name_len, 4650 const char *data, int data_len, 4651 u8 type, void *ctx) 4652 { 4653 int ret; 4654 struct send_ctx *sctx = ctx; 4655 struct fs_path *p; 4656 struct posix_acl_xattr_header dummy_acl; 4657 4658 /* Capabilities are emitted by finish_inode_if_needed */ 4659 if (!strncmp(name, XATTR_NAME_CAPS, name_len)) 4660 return 0; 4661 4662 p = fs_path_alloc(); 4663 if (!p) 4664 return -ENOMEM; 4665 4666 /* 4667 * This hack is needed because empty acls are stored as zero byte 4668 * data in xattrs. Problem with that is, that receiving these zero byte 4669 * acls will fail later. To fix this, we send a dummy acl list that 4670 * only contains the version number and no entries. 4671 */ 4672 if (!strncmp(name, XATTR_NAME_POSIX_ACL_ACCESS, name_len) || 4673 !strncmp(name, XATTR_NAME_POSIX_ACL_DEFAULT, name_len)) { 4674 if (data_len == 0) { 4675 dummy_acl.a_version = 4676 cpu_to_le32(POSIX_ACL_XATTR_VERSION); 4677 data = (char *)&dummy_acl; 4678 data_len = sizeof(dummy_acl); 4679 } 4680 } 4681 4682 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 4683 if (ret < 0) 4684 goto out; 4685 4686 ret = send_set_xattr(sctx, p, name, name_len, data, data_len); 4687 4688 out: 4689 fs_path_free(p); 4690 return ret; 4691 } 4692 4693 static int __process_deleted_xattr(int num, struct btrfs_key *di_key, 4694 const char *name, int name_len, 4695 const char *data, int data_len, 4696 u8 type, void *ctx) 4697 { 4698 int ret; 4699 struct send_ctx *sctx = ctx; 4700 struct fs_path *p; 4701 4702 p = fs_path_alloc(); 4703 if (!p) 4704 return -ENOMEM; 4705 4706 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 4707 if (ret < 0) 4708 goto out; 4709 4710 ret = send_remove_xattr(sctx, p, name, name_len); 4711 4712 out: 4713 fs_path_free(p); 4714 return ret; 4715 } 4716 4717 static int process_new_xattr(struct send_ctx *sctx) 4718 { 4719 int ret = 0; 4720 4721 ret = iterate_dir_item(sctx->send_root, sctx->left_path, 4722 __process_new_xattr, sctx); 4723 4724 return ret; 4725 } 4726 4727 static int process_deleted_xattr(struct send_ctx *sctx) 4728 { 4729 return iterate_dir_item(sctx->parent_root, sctx->right_path, 4730 __process_deleted_xattr, sctx); 4731 } 4732 4733 struct find_xattr_ctx { 4734 const char *name; 4735 int name_len; 4736 int found_idx; 4737 char *found_data; 4738 int found_data_len; 4739 }; 4740 4741 static int __find_xattr(int num, struct btrfs_key *di_key, 4742 const char *name, int name_len, 4743 const char *data, int data_len, 4744 u8 type, void *vctx) 4745 { 4746 struct find_xattr_ctx *ctx = vctx; 4747 4748 if (name_len == ctx->name_len && 4749 strncmp(name, ctx->name, name_len) == 0) { 4750 ctx->found_idx = num; 4751 ctx->found_data_len = data_len; 4752 ctx->found_data = kmemdup(data, data_len, GFP_KERNEL); 4753 if (!ctx->found_data) 4754 return -ENOMEM; 4755 return 1; 4756 } 4757 return 0; 4758 } 4759 4760 static int find_xattr(struct btrfs_root *root, 4761 struct btrfs_path *path, 4762 struct btrfs_key *key, 4763 const char *name, int name_len, 4764 char **data, int *data_len) 4765 { 4766 int ret; 4767 struct find_xattr_ctx ctx; 4768 4769 ctx.name = name; 4770 ctx.name_len = name_len; 4771 ctx.found_idx = -1; 4772 ctx.found_data = NULL; 4773 ctx.found_data_len = 0; 4774 4775 ret = iterate_dir_item(root, path, __find_xattr, &ctx); 4776 if (ret < 0) 4777 return ret; 4778 4779 if (ctx.found_idx == -1) 4780 return -ENOENT; 4781 if (data) { 4782 *data = ctx.found_data; 4783 *data_len = ctx.found_data_len; 4784 } else { 4785 kfree(ctx.found_data); 4786 } 4787 return ctx.found_idx; 4788 } 4789 4790 4791 static int __process_changed_new_xattr(int num, struct btrfs_key *di_key, 4792 const char *name, int name_len, 4793 const char *data, int data_len, 4794 u8 type, void *ctx) 4795 { 4796 int ret; 4797 struct send_ctx *sctx = ctx; 4798 char *found_data = NULL; 4799 int found_data_len = 0; 4800 4801 ret = find_xattr(sctx->parent_root, sctx->right_path, 4802 sctx->cmp_key, name, name_len, &found_data, 4803 &found_data_len); 4804 if (ret == -ENOENT) { 4805 ret = __process_new_xattr(num, di_key, name, name_len, data, 4806 data_len, type, ctx); 4807 } else if (ret >= 0) { 4808 if (data_len != found_data_len || 4809 memcmp(data, found_data, data_len)) { 4810 ret = __process_new_xattr(num, di_key, name, name_len, 4811 data, data_len, type, ctx); 4812 } else { 4813 ret = 0; 4814 } 4815 } 4816 4817 kfree(found_data); 4818 return ret; 4819 } 4820 4821 static int __process_changed_deleted_xattr(int num, struct btrfs_key *di_key, 4822 const char *name, int name_len, 4823 const char *data, int data_len, 4824 u8 type, void *ctx) 4825 { 4826 int ret; 4827 struct send_ctx *sctx = ctx; 4828 4829 ret = find_xattr(sctx->send_root, sctx->left_path, sctx->cmp_key, 4830 name, name_len, NULL, NULL); 4831 if (ret == -ENOENT) 4832 ret = __process_deleted_xattr(num, di_key, name, name_len, data, 4833 data_len, type, ctx); 4834 else if (ret >= 0) 4835 ret = 0; 4836 4837 return ret; 4838 } 4839 4840 static int process_changed_xattr(struct send_ctx *sctx) 4841 { 4842 int ret = 0; 4843 4844 ret = iterate_dir_item(sctx->send_root, sctx->left_path, 4845 __process_changed_new_xattr, sctx); 4846 if (ret < 0) 4847 goto out; 4848 ret = iterate_dir_item(sctx->parent_root, sctx->right_path, 4849 __process_changed_deleted_xattr, sctx); 4850 4851 out: 4852 return ret; 4853 } 4854 4855 static int process_all_new_xattrs(struct send_ctx *sctx) 4856 { 4857 int ret; 4858 struct btrfs_root *root; 4859 struct btrfs_path *path; 4860 struct btrfs_key key; 4861 struct btrfs_key found_key; 4862 struct extent_buffer *eb; 4863 int slot; 4864 4865 path = alloc_path_for_send(); 4866 if (!path) 4867 return -ENOMEM; 4868 4869 root = sctx->send_root; 4870 4871 key.objectid = sctx->cmp_key->objectid; 4872 key.type = BTRFS_XATTR_ITEM_KEY; 4873 key.offset = 0; 4874 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 4875 if (ret < 0) 4876 goto out; 4877 4878 while (1) { 4879 eb = path->nodes[0]; 4880 slot = path->slots[0]; 4881 if (slot >= btrfs_header_nritems(eb)) { 4882 ret = btrfs_next_leaf(root, path); 4883 if (ret < 0) { 4884 goto out; 4885 } else if (ret > 0) { 4886 ret = 0; 4887 break; 4888 } 4889 continue; 4890 } 4891 4892 btrfs_item_key_to_cpu(eb, &found_key, slot); 4893 if (found_key.objectid != key.objectid || 4894 found_key.type != key.type) { 4895 ret = 0; 4896 goto out; 4897 } 4898 4899 ret = iterate_dir_item(root, path, __process_new_xattr, sctx); 4900 if (ret < 0) 4901 goto out; 4902 4903 path->slots[0]++; 4904 } 4905 4906 out: 4907 btrfs_free_path(path); 4908 return ret; 4909 } 4910 4911 static inline u64 max_send_read_size(const struct send_ctx *sctx) 4912 { 4913 return sctx->send_max_size - SZ_16K; 4914 } 4915 4916 static int put_data_header(struct send_ctx *sctx, u32 len) 4917 { 4918 struct btrfs_tlv_header *hdr; 4919 4920 if (sctx->send_max_size - sctx->send_size < sizeof(*hdr) + len) 4921 return -EOVERFLOW; 4922 hdr = (struct btrfs_tlv_header *)(sctx->send_buf + sctx->send_size); 4923 put_unaligned_le16(BTRFS_SEND_A_DATA, &hdr->tlv_type); 4924 put_unaligned_le16(len, &hdr->tlv_len); 4925 sctx->send_size += sizeof(*hdr); 4926 return 0; 4927 } 4928 4929 static int put_file_data(struct send_ctx *sctx, u64 offset, u32 len) 4930 { 4931 struct btrfs_root *root = sctx->send_root; 4932 struct btrfs_fs_info *fs_info = root->fs_info; 4933 struct inode *inode; 4934 struct page *page; 4935 char *addr; 4936 pgoff_t index = offset >> PAGE_SHIFT; 4937 pgoff_t last_index; 4938 unsigned pg_offset = offset_in_page(offset); 4939 int ret; 4940 4941 ret = put_data_header(sctx, len); 4942 if (ret) 4943 return ret; 4944 4945 inode = btrfs_iget(fs_info->sb, sctx->cur_ino, root); 4946 if (IS_ERR(inode)) 4947 return PTR_ERR(inode); 4948 4949 last_index = (offset + len - 1) >> PAGE_SHIFT; 4950 4951 /* initial readahead */ 4952 memset(&sctx->ra, 0, sizeof(struct file_ra_state)); 4953 file_ra_state_init(&sctx->ra, inode->i_mapping); 4954 4955 while (index <= last_index) { 4956 unsigned cur_len = min_t(unsigned, len, 4957 PAGE_SIZE - pg_offset); 4958 4959 page = find_lock_page(inode->i_mapping, index); 4960 if (!page) { 4961 page_cache_sync_readahead(inode->i_mapping, &sctx->ra, 4962 NULL, index, last_index + 1 - index); 4963 4964 page = find_or_create_page(inode->i_mapping, index, 4965 GFP_KERNEL); 4966 if (!page) { 4967 ret = -ENOMEM; 4968 break; 4969 } 4970 } 4971 4972 if (PageReadahead(page)) { 4973 page_cache_async_readahead(inode->i_mapping, &sctx->ra, 4974 NULL, page, index, last_index + 1 - index); 4975 } 4976 4977 if (!PageUptodate(page)) { 4978 btrfs_readpage(NULL, page); 4979 lock_page(page); 4980 if (!PageUptodate(page)) { 4981 unlock_page(page); 4982 put_page(page); 4983 ret = -EIO; 4984 break; 4985 } 4986 } 4987 4988 addr = kmap(page); 4989 memcpy(sctx->send_buf + sctx->send_size, addr + pg_offset, 4990 cur_len); 4991 kunmap(page); 4992 unlock_page(page); 4993 put_page(page); 4994 index++; 4995 pg_offset = 0; 4996 len -= cur_len; 4997 sctx->send_size += cur_len; 4998 } 4999 iput(inode); 5000 return ret; 5001 } 5002 5003 /* 5004 * Read some bytes from the current inode/file and send a write command to 5005 * user space. 5006 */ 5007 static int send_write(struct send_ctx *sctx, u64 offset, u32 len) 5008 { 5009 struct btrfs_fs_info *fs_info = sctx->send_root->fs_info; 5010 int ret = 0; 5011 struct fs_path *p; 5012 5013 p = fs_path_alloc(); 5014 if (!p) 5015 return -ENOMEM; 5016 5017 btrfs_debug(fs_info, "send_write offset=%llu, len=%d", offset, len); 5018 5019 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE); 5020 if (ret < 0) 5021 goto out; 5022 5023 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 5024 if (ret < 0) 5025 goto out; 5026 5027 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 5028 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset); 5029 ret = put_file_data(sctx, offset, len); 5030 if (ret < 0) 5031 goto out; 5032 5033 ret = send_cmd(sctx); 5034 5035 tlv_put_failure: 5036 out: 5037 fs_path_free(p); 5038 return ret; 5039 } 5040 5041 /* 5042 * Send a clone command to user space. 5043 */ 5044 static int send_clone(struct send_ctx *sctx, 5045 u64 offset, u32 len, 5046 struct clone_root *clone_root) 5047 { 5048 int ret = 0; 5049 struct fs_path *p; 5050 u64 gen; 5051 5052 btrfs_debug(sctx->send_root->fs_info, 5053 "send_clone offset=%llu, len=%d, clone_root=%llu, clone_inode=%llu, clone_offset=%llu", 5054 offset, len, clone_root->root->root_key.objectid, 5055 clone_root->ino, clone_root->offset); 5056 5057 p = fs_path_alloc(); 5058 if (!p) 5059 return -ENOMEM; 5060 5061 ret = begin_cmd(sctx, BTRFS_SEND_C_CLONE); 5062 if (ret < 0) 5063 goto out; 5064 5065 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 5066 if (ret < 0) 5067 goto out; 5068 5069 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset); 5070 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_LEN, len); 5071 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 5072 5073 if (clone_root->root == sctx->send_root) { 5074 ret = get_inode_info(sctx->send_root, clone_root->ino, NULL, 5075 &gen, NULL, NULL, NULL, NULL); 5076 if (ret < 0) 5077 goto out; 5078 ret = get_cur_path(sctx, clone_root->ino, gen, p); 5079 } else { 5080 ret = get_inode_path(clone_root->root, clone_root->ino, p); 5081 } 5082 if (ret < 0) 5083 goto out; 5084 5085 /* 5086 * If the parent we're using has a received_uuid set then use that as 5087 * our clone source as that is what we will look for when doing a 5088 * receive. 5089 * 5090 * This covers the case that we create a snapshot off of a received 5091 * subvolume and then use that as the parent and try to receive on a 5092 * different host. 5093 */ 5094 if (!btrfs_is_empty_uuid(clone_root->root->root_item.received_uuid)) 5095 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID, 5096 clone_root->root->root_item.received_uuid); 5097 else 5098 TLV_PUT_UUID(sctx, BTRFS_SEND_A_CLONE_UUID, 5099 clone_root->root->root_item.uuid); 5100 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_CTRANSID, 5101 btrfs_root_ctransid(&clone_root->root->root_item)); 5102 TLV_PUT_PATH(sctx, BTRFS_SEND_A_CLONE_PATH, p); 5103 TLV_PUT_U64(sctx, BTRFS_SEND_A_CLONE_OFFSET, 5104 clone_root->offset); 5105 5106 ret = send_cmd(sctx); 5107 5108 tlv_put_failure: 5109 out: 5110 fs_path_free(p); 5111 return ret; 5112 } 5113 5114 /* 5115 * Send an update extent command to user space. 5116 */ 5117 static int send_update_extent(struct send_ctx *sctx, 5118 u64 offset, u32 len) 5119 { 5120 int ret = 0; 5121 struct fs_path *p; 5122 5123 p = fs_path_alloc(); 5124 if (!p) 5125 return -ENOMEM; 5126 5127 ret = begin_cmd(sctx, BTRFS_SEND_C_UPDATE_EXTENT); 5128 if (ret < 0) 5129 goto out; 5130 5131 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 5132 if (ret < 0) 5133 goto out; 5134 5135 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 5136 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset); 5137 TLV_PUT_U64(sctx, BTRFS_SEND_A_SIZE, len); 5138 5139 ret = send_cmd(sctx); 5140 5141 tlv_put_failure: 5142 out: 5143 fs_path_free(p); 5144 return ret; 5145 } 5146 5147 static int send_hole(struct send_ctx *sctx, u64 end) 5148 { 5149 struct fs_path *p = NULL; 5150 u64 read_size = max_send_read_size(sctx); 5151 u64 offset = sctx->cur_inode_last_extent; 5152 int ret = 0; 5153 5154 /* 5155 * A hole that starts at EOF or beyond it. Since we do not yet support 5156 * fallocate (for extent preallocation and hole punching), sending a 5157 * write of zeroes starting at EOF or beyond would later require issuing 5158 * a truncate operation which would undo the write and achieve nothing. 5159 */ 5160 if (offset >= sctx->cur_inode_size) 5161 return 0; 5162 5163 /* 5164 * Don't go beyond the inode's i_size due to prealloc extents that start 5165 * after the i_size. 5166 */ 5167 end = min_t(u64, end, sctx->cur_inode_size); 5168 5169 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA) 5170 return send_update_extent(sctx, offset, end - offset); 5171 5172 p = fs_path_alloc(); 5173 if (!p) 5174 return -ENOMEM; 5175 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, p); 5176 if (ret < 0) 5177 goto tlv_put_failure; 5178 while (offset < end) { 5179 u64 len = min(end - offset, read_size); 5180 5181 ret = begin_cmd(sctx, BTRFS_SEND_C_WRITE); 5182 if (ret < 0) 5183 break; 5184 TLV_PUT_PATH(sctx, BTRFS_SEND_A_PATH, p); 5185 TLV_PUT_U64(sctx, BTRFS_SEND_A_FILE_OFFSET, offset); 5186 ret = put_data_header(sctx, len); 5187 if (ret < 0) 5188 break; 5189 memset(sctx->send_buf + sctx->send_size, 0, len); 5190 sctx->send_size += len; 5191 ret = send_cmd(sctx); 5192 if (ret < 0) 5193 break; 5194 offset += len; 5195 } 5196 sctx->cur_inode_next_write_offset = offset; 5197 tlv_put_failure: 5198 fs_path_free(p); 5199 return ret; 5200 } 5201 5202 static int send_extent_data(struct send_ctx *sctx, 5203 const u64 offset, 5204 const u64 len) 5205 { 5206 u64 read_size = max_send_read_size(sctx); 5207 u64 sent = 0; 5208 5209 if (sctx->flags & BTRFS_SEND_FLAG_NO_FILE_DATA) 5210 return send_update_extent(sctx, offset, len); 5211 5212 while (sent < len) { 5213 u64 size = min(len - sent, read_size); 5214 int ret; 5215 5216 ret = send_write(sctx, offset + sent, size); 5217 if (ret < 0) 5218 return ret; 5219 sent += size; 5220 } 5221 return 0; 5222 } 5223 5224 /* 5225 * Search for a capability xattr related to sctx->cur_ino. If the capability is 5226 * found, call send_set_xattr function to emit it. 5227 * 5228 * Return 0 if there isn't a capability, or when the capability was emitted 5229 * successfully, or < 0 if an error occurred. 5230 */ 5231 static int send_capabilities(struct send_ctx *sctx) 5232 { 5233 struct fs_path *fspath = NULL; 5234 struct btrfs_path *path; 5235 struct btrfs_dir_item *di; 5236 struct extent_buffer *leaf; 5237 unsigned long data_ptr; 5238 char *buf = NULL; 5239 int buf_len; 5240 int ret = 0; 5241 5242 path = alloc_path_for_send(); 5243 if (!path) 5244 return -ENOMEM; 5245 5246 di = btrfs_lookup_xattr(NULL, sctx->send_root, path, sctx->cur_ino, 5247 XATTR_NAME_CAPS, strlen(XATTR_NAME_CAPS), 0); 5248 if (!di) { 5249 /* There is no xattr for this inode */ 5250 goto out; 5251 } else if (IS_ERR(di)) { 5252 ret = PTR_ERR(di); 5253 goto out; 5254 } 5255 5256 leaf = path->nodes[0]; 5257 buf_len = btrfs_dir_data_len(leaf, di); 5258 5259 fspath = fs_path_alloc(); 5260 buf = kmalloc(buf_len, GFP_KERNEL); 5261 if (!fspath || !buf) { 5262 ret = -ENOMEM; 5263 goto out; 5264 } 5265 5266 ret = get_cur_path(sctx, sctx->cur_ino, sctx->cur_inode_gen, fspath); 5267 if (ret < 0) 5268 goto out; 5269 5270 data_ptr = (unsigned long)(di + 1) + btrfs_dir_name_len(leaf, di); 5271 read_extent_buffer(leaf, buf, data_ptr, buf_len); 5272 5273 ret = send_set_xattr(sctx, fspath, XATTR_NAME_CAPS, 5274 strlen(XATTR_NAME_CAPS), buf, buf_len); 5275 out: 5276 kfree(buf); 5277 fs_path_free(fspath); 5278 btrfs_free_path(path); 5279 return ret; 5280 } 5281 5282 static int clone_range(struct send_ctx *sctx, 5283 struct clone_root *clone_root, 5284 const u64 disk_byte, 5285 u64 data_offset, 5286 u64 offset, 5287 u64 len) 5288 { 5289 struct btrfs_path *path; 5290 struct btrfs_key key; 5291 int ret; 5292 u64 clone_src_i_size = 0; 5293 5294 /* 5295 * Prevent cloning from a zero offset with a length matching the sector 5296 * size because in some scenarios this will make the receiver fail. 5297 * 5298 * For example, if in the source filesystem the extent at offset 0 5299 * has a length of sectorsize and it was written using direct IO, then 5300 * it can never be an inline extent (even if compression is enabled). 5301 * Then this extent can be cloned in the original filesystem to a non 5302 * zero file offset, but it may not be possible to clone in the 5303 * destination filesystem because it can be inlined due to compression 5304 * on the destination filesystem (as the receiver's write operations are 5305 * always done using buffered IO). The same happens when the original 5306 * filesystem does not have compression enabled but the destination 5307 * filesystem has. 5308 */ 5309 if (clone_root->offset == 0 && 5310 len == sctx->send_root->fs_info->sectorsize) 5311 return send_extent_data(sctx, offset, len); 5312 5313 path = alloc_path_for_send(); 5314 if (!path) 5315 return -ENOMEM; 5316 5317 /* 5318 * There are inodes that have extents that lie behind its i_size. Don't 5319 * accept clones from these extents. 5320 */ 5321 ret = __get_inode_info(clone_root->root, path, clone_root->ino, 5322 &clone_src_i_size, NULL, NULL, NULL, NULL, NULL); 5323 btrfs_release_path(path); 5324 if (ret < 0) 5325 goto out; 5326 5327 /* 5328 * We can't send a clone operation for the entire range if we find 5329 * extent items in the respective range in the source file that 5330 * refer to different extents or if we find holes. 5331 * So check for that and do a mix of clone and regular write/copy 5332 * operations if needed. 5333 * 5334 * Example: 5335 * 5336 * mkfs.btrfs -f /dev/sda 5337 * mount /dev/sda /mnt 5338 * xfs_io -f -c "pwrite -S 0xaa 0K 100K" /mnt/foo 5339 * cp --reflink=always /mnt/foo /mnt/bar 5340 * xfs_io -c "pwrite -S 0xbb 50K 50K" /mnt/foo 5341 * btrfs subvolume snapshot -r /mnt /mnt/snap 5342 * 5343 * If when we send the snapshot and we are processing file bar (which 5344 * has a higher inode number than foo) we blindly send a clone operation 5345 * for the [0, 100K[ range from foo to bar, the receiver ends up getting 5346 * a file bar that matches the content of file foo - iow, doesn't match 5347 * the content from bar in the original filesystem. 5348 */ 5349 key.objectid = clone_root->ino; 5350 key.type = BTRFS_EXTENT_DATA_KEY; 5351 key.offset = clone_root->offset; 5352 ret = btrfs_search_slot(NULL, clone_root->root, &key, path, 0, 0); 5353 if (ret < 0) 5354 goto out; 5355 if (ret > 0 && path->slots[0] > 0) { 5356 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1); 5357 if (key.objectid == clone_root->ino && 5358 key.type == BTRFS_EXTENT_DATA_KEY) 5359 path->slots[0]--; 5360 } 5361 5362 while (true) { 5363 struct extent_buffer *leaf = path->nodes[0]; 5364 int slot = path->slots[0]; 5365 struct btrfs_file_extent_item *ei; 5366 u8 type; 5367 u64 ext_len; 5368 u64 clone_len; 5369 u64 clone_data_offset; 5370 5371 if (slot >= btrfs_header_nritems(leaf)) { 5372 ret = btrfs_next_leaf(clone_root->root, path); 5373 if (ret < 0) 5374 goto out; 5375 else if (ret > 0) 5376 break; 5377 continue; 5378 } 5379 5380 btrfs_item_key_to_cpu(leaf, &key, slot); 5381 5382 /* 5383 * We might have an implicit trailing hole (NO_HOLES feature 5384 * enabled). We deal with it after leaving this loop. 5385 */ 5386 if (key.objectid != clone_root->ino || 5387 key.type != BTRFS_EXTENT_DATA_KEY) 5388 break; 5389 5390 ei = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item); 5391 type = btrfs_file_extent_type(leaf, ei); 5392 if (type == BTRFS_FILE_EXTENT_INLINE) { 5393 ext_len = btrfs_file_extent_ram_bytes(leaf, ei); 5394 ext_len = PAGE_ALIGN(ext_len); 5395 } else { 5396 ext_len = btrfs_file_extent_num_bytes(leaf, ei); 5397 } 5398 5399 if (key.offset + ext_len <= clone_root->offset) 5400 goto next; 5401 5402 if (key.offset > clone_root->offset) { 5403 /* Implicit hole, NO_HOLES feature enabled. */ 5404 u64 hole_len = key.offset - clone_root->offset; 5405 5406 if (hole_len > len) 5407 hole_len = len; 5408 ret = send_extent_data(sctx, offset, hole_len); 5409 if (ret < 0) 5410 goto out; 5411 5412 len -= hole_len; 5413 if (len == 0) 5414 break; 5415 offset += hole_len; 5416 clone_root->offset += hole_len; 5417 data_offset += hole_len; 5418 } 5419 5420 if (key.offset >= clone_root->offset + len) 5421 break; 5422 5423 if (key.offset >= clone_src_i_size) 5424 break; 5425 5426 if (key.offset + ext_len > clone_src_i_size) 5427 ext_len = clone_src_i_size - key.offset; 5428 5429 clone_data_offset = btrfs_file_extent_offset(leaf, ei); 5430 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte) { 5431 clone_root->offset = key.offset; 5432 if (clone_data_offset < data_offset && 5433 clone_data_offset + ext_len > data_offset) { 5434 u64 extent_offset; 5435 5436 extent_offset = data_offset - clone_data_offset; 5437 ext_len -= extent_offset; 5438 clone_data_offset += extent_offset; 5439 clone_root->offset += extent_offset; 5440 } 5441 } 5442 5443 clone_len = min_t(u64, ext_len, len); 5444 5445 if (btrfs_file_extent_disk_bytenr(leaf, ei) == disk_byte && 5446 clone_data_offset == data_offset) { 5447 const u64 src_end = clone_root->offset + clone_len; 5448 const u64 sectorsize = SZ_64K; 5449 5450 /* 5451 * We can't clone the last block, when its size is not 5452 * sector size aligned, into the middle of a file. If we 5453 * do so, the receiver will get a failure (-EINVAL) when 5454 * trying to clone or will silently corrupt the data in 5455 * the destination file if it's on a kernel without the 5456 * fix introduced by commit ac765f83f1397646 5457 * ("Btrfs: fix data corruption due to cloning of eof 5458 * block). 5459 * 5460 * So issue a clone of the aligned down range plus a 5461 * regular write for the eof block, if we hit that case. 5462 * 5463 * Also, we use the maximum possible sector size, 64K, 5464 * because we don't know what's the sector size of the 5465 * filesystem that receives the stream, so we have to 5466 * assume the largest possible sector size. 5467 */ 5468 if (src_end == clone_src_i_size && 5469 !IS_ALIGNED(src_end, sectorsize) && 5470 offset + clone_len < sctx->cur_inode_size) { 5471 u64 slen; 5472 5473 slen = ALIGN_DOWN(src_end - clone_root->offset, 5474 sectorsize); 5475 if (slen > 0) { 5476 ret = send_clone(sctx, offset, slen, 5477 clone_root); 5478 if (ret < 0) 5479 goto out; 5480 } 5481 ret = send_extent_data(sctx, offset + slen, 5482 clone_len - slen); 5483 } else { 5484 ret = send_clone(sctx, offset, clone_len, 5485 clone_root); 5486 } 5487 } else { 5488 ret = send_extent_data(sctx, offset, clone_len); 5489 } 5490 5491 if (ret < 0) 5492 goto out; 5493 5494 len -= clone_len; 5495 if (len == 0) 5496 break; 5497 offset += clone_len; 5498 clone_root->offset += clone_len; 5499 5500 /* 5501 * If we are cloning from the file we are currently processing, 5502 * and using the send root as the clone root, we must stop once 5503 * the current clone offset reaches the current eof of the file 5504 * at the receiver, otherwise we would issue an invalid clone 5505 * operation (source range going beyond eof) and cause the 5506 * receiver to fail. So if we reach the current eof, bail out 5507 * and fallback to a regular write. 5508 */ 5509 if (clone_root->root == sctx->send_root && 5510 clone_root->ino == sctx->cur_ino && 5511 clone_root->offset >= sctx->cur_inode_next_write_offset) 5512 break; 5513 5514 data_offset += clone_len; 5515 next: 5516 path->slots[0]++; 5517 } 5518 5519 if (len > 0) 5520 ret = send_extent_data(sctx, offset, len); 5521 else 5522 ret = 0; 5523 out: 5524 btrfs_free_path(path); 5525 return ret; 5526 } 5527 5528 static int send_write_or_clone(struct send_ctx *sctx, 5529 struct btrfs_path *path, 5530 struct btrfs_key *key, 5531 struct clone_root *clone_root) 5532 { 5533 int ret = 0; 5534 u64 offset = key->offset; 5535 u64 end; 5536 u64 bs = sctx->send_root->fs_info->sb->s_blocksize; 5537 5538 end = min_t(u64, btrfs_file_extent_end(path), sctx->cur_inode_size); 5539 if (offset >= end) 5540 return 0; 5541 5542 if (clone_root && IS_ALIGNED(end, bs)) { 5543 struct btrfs_file_extent_item *ei; 5544 u64 disk_byte; 5545 u64 data_offset; 5546 5547 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 5548 struct btrfs_file_extent_item); 5549 disk_byte = btrfs_file_extent_disk_bytenr(path->nodes[0], ei); 5550 data_offset = btrfs_file_extent_offset(path->nodes[0], ei); 5551 ret = clone_range(sctx, clone_root, disk_byte, data_offset, 5552 offset, end - offset); 5553 } else { 5554 ret = send_extent_data(sctx, offset, end - offset); 5555 } 5556 sctx->cur_inode_next_write_offset = end; 5557 return ret; 5558 } 5559 5560 static int is_extent_unchanged(struct send_ctx *sctx, 5561 struct btrfs_path *left_path, 5562 struct btrfs_key *ekey) 5563 { 5564 int ret = 0; 5565 struct btrfs_key key; 5566 struct btrfs_path *path = NULL; 5567 struct extent_buffer *eb; 5568 int slot; 5569 struct btrfs_key found_key; 5570 struct btrfs_file_extent_item *ei; 5571 u64 left_disknr; 5572 u64 right_disknr; 5573 u64 left_offset; 5574 u64 right_offset; 5575 u64 left_offset_fixed; 5576 u64 left_len; 5577 u64 right_len; 5578 u64 left_gen; 5579 u64 right_gen; 5580 u8 left_type; 5581 u8 right_type; 5582 5583 path = alloc_path_for_send(); 5584 if (!path) 5585 return -ENOMEM; 5586 5587 eb = left_path->nodes[0]; 5588 slot = left_path->slots[0]; 5589 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item); 5590 left_type = btrfs_file_extent_type(eb, ei); 5591 5592 if (left_type != BTRFS_FILE_EXTENT_REG) { 5593 ret = 0; 5594 goto out; 5595 } 5596 left_disknr = btrfs_file_extent_disk_bytenr(eb, ei); 5597 left_len = btrfs_file_extent_num_bytes(eb, ei); 5598 left_offset = btrfs_file_extent_offset(eb, ei); 5599 left_gen = btrfs_file_extent_generation(eb, ei); 5600 5601 /* 5602 * Following comments will refer to these graphics. L is the left 5603 * extents which we are checking at the moment. 1-8 are the right 5604 * extents that we iterate. 5605 * 5606 * |-----L-----| 5607 * |-1-|-2a-|-3-|-4-|-5-|-6-| 5608 * 5609 * |-----L-----| 5610 * |--1--|-2b-|...(same as above) 5611 * 5612 * Alternative situation. Happens on files where extents got split. 5613 * |-----L-----| 5614 * |-----------7-----------|-6-| 5615 * 5616 * Alternative situation. Happens on files which got larger. 5617 * |-----L-----| 5618 * |-8-| 5619 * Nothing follows after 8. 5620 */ 5621 5622 key.objectid = ekey->objectid; 5623 key.type = BTRFS_EXTENT_DATA_KEY; 5624 key.offset = ekey->offset; 5625 ret = btrfs_search_slot_for_read(sctx->parent_root, &key, path, 0, 0); 5626 if (ret < 0) 5627 goto out; 5628 if (ret) { 5629 ret = 0; 5630 goto out; 5631 } 5632 5633 /* 5634 * Handle special case where the right side has no extents at all. 5635 */ 5636 eb = path->nodes[0]; 5637 slot = path->slots[0]; 5638 btrfs_item_key_to_cpu(eb, &found_key, slot); 5639 if (found_key.objectid != key.objectid || 5640 found_key.type != key.type) { 5641 /* If we're a hole then just pretend nothing changed */ 5642 ret = (left_disknr) ? 0 : 1; 5643 goto out; 5644 } 5645 5646 /* 5647 * We're now on 2a, 2b or 7. 5648 */ 5649 key = found_key; 5650 while (key.offset < ekey->offset + left_len) { 5651 ei = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item); 5652 right_type = btrfs_file_extent_type(eb, ei); 5653 if (right_type != BTRFS_FILE_EXTENT_REG && 5654 right_type != BTRFS_FILE_EXTENT_INLINE) { 5655 ret = 0; 5656 goto out; 5657 } 5658 5659 if (right_type == BTRFS_FILE_EXTENT_INLINE) { 5660 right_len = btrfs_file_extent_ram_bytes(eb, ei); 5661 right_len = PAGE_ALIGN(right_len); 5662 } else { 5663 right_len = btrfs_file_extent_num_bytes(eb, ei); 5664 } 5665 5666 /* 5667 * Are we at extent 8? If yes, we know the extent is changed. 5668 * This may only happen on the first iteration. 5669 */ 5670 if (found_key.offset + right_len <= ekey->offset) { 5671 /* If we're a hole just pretend nothing changed */ 5672 ret = (left_disknr) ? 0 : 1; 5673 goto out; 5674 } 5675 5676 /* 5677 * We just wanted to see if when we have an inline extent, what 5678 * follows it is a regular extent (wanted to check the above 5679 * condition for inline extents too). This should normally not 5680 * happen but it's possible for example when we have an inline 5681 * compressed extent representing data with a size matching 5682 * the page size (currently the same as sector size). 5683 */ 5684 if (right_type == BTRFS_FILE_EXTENT_INLINE) { 5685 ret = 0; 5686 goto out; 5687 } 5688 5689 right_disknr = btrfs_file_extent_disk_bytenr(eb, ei); 5690 right_offset = btrfs_file_extent_offset(eb, ei); 5691 right_gen = btrfs_file_extent_generation(eb, ei); 5692 5693 left_offset_fixed = left_offset; 5694 if (key.offset < ekey->offset) { 5695 /* Fix the right offset for 2a and 7. */ 5696 right_offset += ekey->offset - key.offset; 5697 } else { 5698 /* Fix the left offset for all behind 2a and 2b */ 5699 left_offset_fixed += key.offset - ekey->offset; 5700 } 5701 5702 /* 5703 * Check if we have the same extent. 5704 */ 5705 if (left_disknr != right_disknr || 5706 left_offset_fixed != right_offset || 5707 left_gen != right_gen) { 5708 ret = 0; 5709 goto out; 5710 } 5711 5712 /* 5713 * Go to the next extent. 5714 */ 5715 ret = btrfs_next_item(sctx->parent_root, path); 5716 if (ret < 0) 5717 goto out; 5718 if (!ret) { 5719 eb = path->nodes[0]; 5720 slot = path->slots[0]; 5721 btrfs_item_key_to_cpu(eb, &found_key, slot); 5722 } 5723 if (ret || found_key.objectid != key.objectid || 5724 found_key.type != key.type) { 5725 key.offset += right_len; 5726 break; 5727 } 5728 if (found_key.offset != key.offset + right_len) { 5729 ret = 0; 5730 goto out; 5731 } 5732 key = found_key; 5733 } 5734 5735 /* 5736 * We're now behind the left extent (treat as unchanged) or at the end 5737 * of the right side (treat as changed). 5738 */ 5739 if (key.offset >= ekey->offset + left_len) 5740 ret = 1; 5741 else 5742 ret = 0; 5743 5744 5745 out: 5746 btrfs_free_path(path); 5747 return ret; 5748 } 5749 5750 static int get_last_extent(struct send_ctx *sctx, u64 offset) 5751 { 5752 struct btrfs_path *path; 5753 struct btrfs_root *root = sctx->send_root; 5754 struct btrfs_key key; 5755 int ret; 5756 5757 path = alloc_path_for_send(); 5758 if (!path) 5759 return -ENOMEM; 5760 5761 sctx->cur_inode_last_extent = 0; 5762 5763 key.objectid = sctx->cur_ino; 5764 key.type = BTRFS_EXTENT_DATA_KEY; 5765 key.offset = offset; 5766 ret = btrfs_search_slot_for_read(root, &key, path, 0, 1); 5767 if (ret < 0) 5768 goto out; 5769 ret = 0; 5770 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 5771 if (key.objectid != sctx->cur_ino || key.type != BTRFS_EXTENT_DATA_KEY) 5772 goto out; 5773 5774 sctx->cur_inode_last_extent = btrfs_file_extent_end(path); 5775 out: 5776 btrfs_free_path(path); 5777 return ret; 5778 } 5779 5780 static int range_is_hole_in_parent(struct send_ctx *sctx, 5781 const u64 start, 5782 const u64 end) 5783 { 5784 struct btrfs_path *path; 5785 struct btrfs_key key; 5786 struct btrfs_root *root = sctx->parent_root; 5787 u64 search_start = start; 5788 int ret; 5789 5790 path = alloc_path_for_send(); 5791 if (!path) 5792 return -ENOMEM; 5793 5794 key.objectid = sctx->cur_ino; 5795 key.type = BTRFS_EXTENT_DATA_KEY; 5796 key.offset = search_start; 5797 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 5798 if (ret < 0) 5799 goto out; 5800 if (ret > 0 && path->slots[0] > 0) 5801 path->slots[0]--; 5802 5803 while (search_start < end) { 5804 struct extent_buffer *leaf = path->nodes[0]; 5805 int slot = path->slots[0]; 5806 struct btrfs_file_extent_item *fi; 5807 u64 extent_end; 5808 5809 if (slot >= btrfs_header_nritems(leaf)) { 5810 ret = btrfs_next_leaf(root, path); 5811 if (ret < 0) 5812 goto out; 5813 else if (ret > 0) 5814 break; 5815 continue; 5816 } 5817 5818 btrfs_item_key_to_cpu(leaf, &key, slot); 5819 if (key.objectid < sctx->cur_ino || 5820 key.type < BTRFS_EXTENT_DATA_KEY) 5821 goto next; 5822 if (key.objectid > sctx->cur_ino || 5823 key.type > BTRFS_EXTENT_DATA_KEY || 5824 key.offset >= end) 5825 break; 5826 5827 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item); 5828 extent_end = btrfs_file_extent_end(path); 5829 if (extent_end <= start) 5830 goto next; 5831 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0) { 5832 search_start = extent_end; 5833 goto next; 5834 } 5835 ret = 0; 5836 goto out; 5837 next: 5838 path->slots[0]++; 5839 } 5840 ret = 1; 5841 out: 5842 btrfs_free_path(path); 5843 return ret; 5844 } 5845 5846 static int maybe_send_hole(struct send_ctx *sctx, struct btrfs_path *path, 5847 struct btrfs_key *key) 5848 { 5849 int ret = 0; 5850 5851 if (sctx->cur_ino != key->objectid || !need_send_hole(sctx)) 5852 return 0; 5853 5854 if (sctx->cur_inode_last_extent == (u64)-1) { 5855 ret = get_last_extent(sctx, key->offset - 1); 5856 if (ret) 5857 return ret; 5858 } 5859 5860 if (path->slots[0] == 0 && 5861 sctx->cur_inode_last_extent < key->offset) { 5862 /* 5863 * We might have skipped entire leafs that contained only 5864 * file extent items for our current inode. These leafs have 5865 * a generation number smaller (older) than the one in the 5866 * current leaf and the leaf our last extent came from, and 5867 * are located between these 2 leafs. 5868 */ 5869 ret = get_last_extent(sctx, key->offset - 1); 5870 if (ret) 5871 return ret; 5872 } 5873 5874 if (sctx->cur_inode_last_extent < key->offset) { 5875 ret = range_is_hole_in_parent(sctx, 5876 sctx->cur_inode_last_extent, 5877 key->offset); 5878 if (ret < 0) 5879 return ret; 5880 else if (ret == 0) 5881 ret = send_hole(sctx, key->offset); 5882 else 5883 ret = 0; 5884 } 5885 sctx->cur_inode_last_extent = btrfs_file_extent_end(path); 5886 return ret; 5887 } 5888 5889 static int process_extent(struct send_ctx *sctx, 5890 struct btrfs_path *path, 5891 struct btrfs_key *key) 5892 { 5893 struct clone_root *found_clone = NULL; 5894 int ret = 0; 5895 5896 if (S_ISLNK(sctx->cur_inode_mode)) 5897 return 0; 5898 5899 if (sctx->parent_root && !sctx->cur_inode_new) { 5900 ret = is_extent_unchanged(sctx, path, key); 5901 if (ret < 0) 5902 goto out; 5903 if (ret) { 5904 ret = 0; 5905 goto out_hole; 5906 } 5907 } else { 5908 struct btrfs_file_extent_item *ei; 5909 u8 type; 5910 5911 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 5912 struct btrfs_file_extent_item); 5913 type = btrfs_file_extent_type(path->nodes[0], ei); 5914 if (type == BTRFS_FILE_EXTENT_PREALLOC || 5915 type == BTRFS_FILE_EXTENT_REG) { 5916 /* 5917 * The send spec does not have a prealloc command yet, 5918 * so just leave a hole for prealloc'ed extents until 5919 * we have enough commands queued up to justify rev'ing 5920 * the send spec. 5921 */ 5922 if (type == BTRFS_FILE_EXTENT_PREALLOC) { 5923 ret = 0; 5924 goto out; 5925 } 5926 5927 /* Have a hole, just skip it. */ 5928 if (btrfs_file_extent_disk_bytenr(path->nodes[0], ei) == 0) { 5929 ret = 0; 5930 goto out; 5931 } 5932 } 5933 } 5934 5935 ret = find_extent_clone(sctx, path, key->objectid, key->offset, 5936 sctx->cur_inode_size, &found_clone); 5937 if (ret != -ENOENT && ret < 0) 5938 goto out; 5939 5940 ret = send_write_or_clone(sctx, path, key, found_clone); 5941 if (ret) 5942 goto out; 5943 out_hole: 5944 ret = maybe_send_hole(sctx, path, key); 5945 out: 5946 return ret; 5947 } 5948 5949 static int process_all_extents(struct send_ctx *sctx) 5950 { 5951 int ret; 5952 struct btrfs_root *root; 5953 struct btrfs_path *path; 5954 struct btrfs_key key; 5955 struct btrfs_key found_key; 5956 struct extent_buffer *eb; 5957 int slot; 5958 5959 root = sctx->send_root; 5960 path = alloc_path_for_send(); 5961 if (!path) 5962 return -ENOMEM; 5963 5964 key.objectid = sctx->cmp_key->objectid; 5965 key.type = BTRFS_EXTENT_DATA_KEY; 5966 key.offset = 0; 5967 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 5968 if (ret < 0) 5969 goto out; 5970 5971 while (1) { 5972 eb = path->nodes[0]; 5973 slot = path->slots[0]; 5974 5975 if (slot >= btrfs_header_nritems(eb)) { 5976 ret = btrfs_next_leaf(root, path); 5977 if (ret < 0) { 5978 goto out; 5979 } else if (ret > 0) { 5980 ret = 0; 5981 break; 5982 } 5983 continue; 5984 } 5985 5986 btrfs_item_key_to_cpu(eb, &found_key, slot); 5987 5988 if (found_key.objectid != key.objectid || 5989 found_key.type != key.type) { 5990 ret = 0; 5991 goto out; 5992 } 5993 5994 ret = process_extent(sctx, path, &found_key); 5995 if (ret < 0) 5996 goto out; 5997 5998 path->slots[0]++; 5999 } 6000 6001 out: 6002 btrfs_free_path(path); 6003 return ret; 6004 } 6005 6006 static int process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end, 6007 int *pending_move, 6008 int *refs_processed) 6009 { 6010 int ret = 0; 6011 6012 if (sctx->cur_ino == 0) 6013 goto out; 6014 if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid && 6015 sctx->cmp_key->type <= BTRFS_INODE_EXTREF_KEY) 6016 goto out; 6017 if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs)) 6018 goto out; 6019 6020 ret = process_recorded_refs(sctx, pending_move); 6021 if (ret < 0) 6022 goto out; 6023 6024 *refs_processed = 1; 6025 out: 6026 return ret; 6027 } 6028 6029 static int finish_inode_if_needed(struct send_ctx *sctx, int at_end) 6030 { 6031 int ret = 0; 6032 u64 left_mode; 6033 u64 left_uid; 6034 u64 left_gid; 6035 u64 right_mode; 6036 u64 right_uid; 6037 u64 right_gid; 6038 int need_chmod = 0; 6039 int need_chown = 0; 6040 int need_truncate = 1; 6041 int pending_move = 0; 6042 int refs_processed = 0; 6043 6044 if (sctx->ignore_cur_inode) 6045 return 0; 6046 6047 ret = process_recorded_refs_if_needed(sctx, at_end, &pending_move, 6048 &refs_processed); 6049 if (ret < 0) 6050 goto out; 6051 6052 /* 6053 * We have processed the refs and thus need to advance send_progress. 6054 * Now, calls to get_cur_xxx will take the updated refs of the current 6055 * inode into account. 6056 * 6057 * On the other hand, if our current inode is a directory and couldn't 6058 * be moved/renamed because its parent was renamed/moved too and it has 6059 * a higher inode number, we can only move/rename our current inode 6060 * after we moved/renamed its parent. Therefore in this case operate on 6061 * the old path (pre move/rename) of our current inode, and the 6062 * move/rename will be performed later. 6063 */ 6064 if (refs_processed && !pending_move) 6065 sctx->send_progress = sctx->cur_ino + 1; 6066 6067 if (sctx->cur_ino == 0 || sctx->cur_inode_deleted) 6068 goto out; 6069 if (!at_end && sctx->cmp_key->objectid == sctx->cur_ino) 6070 goto out; 6071 6072 ret = get_inode_info(sctx->send_root, sctx->cur_ino, NULL, NULL, 6073 &left_mode, &left_uid, &left_gid, NULL); 6074 if (ret < 0) 6075 goto out; 6076 6077 if (!sctx->parent_root || sctx->cur_inode_new) { 6078 need_chown = 1; 6079 if (!S_ISLNK(sctx->cur_inode_mode)) 6080 need_chmod = 1; 6081 if (sctx->cur_inode_next_write_offset == sctx->cur_inode_size) 6082 need_truncate = 0; 6083 } else { 6084 u64 old_size; 6085 6086 ret = get_inode_info(sctx->parent_root, sctx->cur_ino, 6087 &old_size, NULL, &right_mode, &right_uid, 6088 &right_gid, NULL); 6089 if (ret < 0) 6090 goto out; 6091 6092 if (left_uid != right_uid || left_gid != right_gid) 6093 need_chown = 1; 6094 if (!S_ISLNK(sctx->cur_inode_mode) && left_mode != right_mode) 6095 need_chmod = 1; 6096 if ((old_size == sctx->cur_inode_size) || 6097 (sctx->cur_inode_size > old_size && 6098 sctx->cur_inode_next_write_offset == sctx->cur_inode_size)) 6099 need_truncate = 0; 6100 } 6101 6102 if (S_ISREG(sctx->cur_inode_mode)) { 6103 if (need_send_hole(sctx)) { 6104 if (sctx->cur_inode_last_extent == (u64)-1 || 6105 sctx->cur_inode_last_extent < 6106 sctx->cur_inode_size) { 6107 ret = get_last_extent(sctx, (u64)-1); 6108 if (ret) 6109 goto out; 6110 } 6111 if (sctx->cur_inode_last_extent < 6112 sctx->cur_inode_size) { 6113 ret = send_hole(sctx, sctx->cur_inode_size); 6114 if (ret) 6115 goto out; 6116 } 6117 } 6118 if (need_truncate) { 6119 ret = send_truncate(sctx, sctx->cur_ino, 6120 sctx->cur_inode_gen, 6121 sctx->cur_inode_size); 6122 if (ret < 0) 6123 goto out; 6124 } 6125 } 6126 6127 if (need_chown) { 6128 ret = send_chown(sctx, sctx->cur_ino, sctx->cur_inode_gen, 6129 left_uid, left_gid); 6130 if (ret < 0) 6131 goto out; 6132 } 6133 if (need_chmod) { 6134 ret = send_chmod(sctx, sctx->cur_ino, sctx->cur_inode_gen, 6135 left_mode); 6136 if (ret < 0) 6137 goto out; 6138 } 6139 6140 ret = send_capabilities(sctx); 6141 if (ret < 0) 6142 goto out; 6143 6144 /* 6145 * If other directory inodes depended on our current directory 6146 * inode's move/rename, now do their move/rename operations. 6147 */ 6148 if (!is_waiting_for_move(sctx, sctx->cur_ino)) { 6149 ret = apply_children_dir_moves(sctx); 6150 if (ret) 6151 goto out; 6152 /* 6153 * Need to send that every time, no matter if it actually 6154 * changed between the two trees as we have done changes to 6155 * the inode before. If our inode is a directory and it's 6156 * waiting to be moved/renamed, we will send its utimes when 6157 * it's moved/renamed, therefore we don't need to do it here. 6158 */ 6159 sctx->send_progress = sctx->cur_ino + 1; 6160 ret = send_utimes(sctx, sctx->cur_ino, sctx->cur_inode_gen); 6161 if (ret < 0) 6162 goto out; 6163 } 6164 6165 out: 6166 return ret; 6167 } 6168 6169 struct parent_paths_ctx { 6170 struct list_head *refs; 6171 struct send_ctx *sctx; 6172 }; 6173 6174 static int record_parent_ref(int num, u64 dir, int index, struct fs_path *name, 6175 void *ctx) 6176 { 6177 struct parent_paths_ctx *ppctx = ctx; 6178 6179 return record_ref(ppctx->sctx->parent_root, dir, name, ppctx->sctx, 6180 ppctx->refs); 6181 } 6182 6183 /* 6184 * Issue unlink operations for all paths of the current inode found in the 6185 * parent snapshot. 6186 */ 6187 static int btrfs_unlink_all_paths(struct send_ctx *sctx) 6188 { 6189 LIST_HEAD(deleted_refs); 6190 struct btrfs_path *path; 6191 struct btrfs_key key; 6192 struct parent_paths_ctx ctx; 6193 int ret; 6194 6195 path = alloc_path_for_send(); 6196 if (!path) 6197 return -ENOMEM; 6198 6199 key.objectid = sctx->cur_ino; 6200 key.type = BTRFS_INODE_REF_KEY; 6201 key.offset = 0; 6202 ret = btrfs_search_slot(NULL, sctx->parent_root, &key, path, 0, 0); 6203 if (ret < 0) 6204 goto out; 6205 6206 ctx.refs = &deleted_refs; 6207 ctx.sctx = sctx; 6208 6209 while (true) { 6210 struct extent_buffer *eb = path->nodes[0]; 6211 int slot = path->slots[0]; 6212 6213 if (slot >= btrfs_header_nritems(eb)) { 6214 ret = btrfs_next_leaf(sctx->parent_root, path); 6215 if (ret < 0) 6216 goto out; 6217 else if (ret > 0) 6218 break; 6219 continue; 6220 } 6221 6222 btrfs_item_key_to_cpu(eb, &key, slot); 6223 if (key.objectid != sctx->cur_ino) 6224 break; 6225 if (key.type != BTRFS_INODE_REF_KEY && 6226 key.type != BTRFS_INODE_EXTREF_KEY) 6227 break; 6228 6229 ret = iterate_inode_ref(sctx->parent_root, path, &key, 1, 6230 record_parent_ref, &ctx); 6231 if (ret < 0) 6232 goto out; 6233 6234 path->slots[0]++; 6235 } 6236 6237 while (!list_empty(&deleted_refs)) { 6238 struct recorded_ref *ref; 6239 6240 ref = list_first_entry(&deleted_refs, struct recorded_ref, list); 6241 ret = send_unlink(sctx, ref->full_path); 6242 if (ret < 0) 6243 goto out; 6244 fs_path_free(ref->full_path); 6245 list_del(&ref->list); 6246 kfree(ref); 6247 } 6248 ret = 0; 6249 out: 6250 btrfs_free_path(path); 6251 if (ret) 6252 __free_recorded_refs(&deleted_refs); 6253 return ret; 6254 } 6255 6256 static int changed_inode(struct send_ctx *sctx, 6257 enum btrfs_compare_tree_result result) 6258 { 6259 int ret = 0; 6260 struct btrfs_key *key = sctx->cmp_key; 6261 struct btrfs_inode_item *left_ii = NULL; 6262 struct btrfs_inode_item *right_ii = NULL; 6263 u64 left_gen = 0; 6264 u64 right_gen = 0; 6265 6266 sctx->cur_ino = key->objectid; 6267 sctx->cur_inode_new_gen = 0; 6268 sctx->cur_inode_last_extent = (u64)-1; 6269 sctx->cur_inode_next_write_offset = 0; 6270 sctx->ignore_cur_inode = false; 6271 6272 /* 6273 * Set send_progress to current inode. This will tell all get_cur_xxx 6274 * functions that the current inode's refs are not updated yet. Later, 6275 * when process_recorded_refs is finished, it is set to cur_ino + 1. 6276 */ 6277 sctx->send_progress = sctx->cur_ino; 6278 6279 if (result == BTRFS_COMPARE_TREE_NEW || 6280 result == BTRFS_COMPARE_TREE_CHANGED) { 6281 left_ii = btrfs_item_ptr(sctx->left_path->nodes[0], 6282 sctx->left_path->slots[0], 6283 struct btrfs_inode_item); 6284 left_gen = btrfs_inode_generation(sctx->left_path->nodes[0], 6285 left_ii); 6286 } else { 6287 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0], 6288 sctx->right_path->slots[0], 6289 struct btrfs_inode_item); 6290 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0], 6291 right_ii); 6292 } 6293 if (result == BTRFS_COMPARE_TREE_CHANGED) { 6294 right_ii = btrfs_item_ptr(sctx->right_path->nodes[0], 6295 sctx->right_path->slots[0], 6296 struct btrfs_inode_item); 6297 6298 right_gen = btrfs_inode_generation(sctx->right_path->nodes[0], 6299 right_ii); 6300 6301 /* 6302 * The cur_ino = root dir case is special here. We can't treat 6303 * the inode as deleted+reused because it would generate a 6304 * stream that tries to delete/mkdir the root dir. 6305 */ 6306 if (left_gen != right_gen && 6307 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) 6308 sctx->cur_inode_new_gen = 1; 6309 } 6310 6311 /* 6312 * Normally we do not find inodes with a link count of zero (orphans) 6313 * because the most common case is to create a snapshot and use it 6314 * for a send operation. However other less common use cases involve 6315 * using a subvolume and send it after turning it to RO mode just 6316 * after deleting all hard links of a file while holding an open 6317 * file descriptor against it or turning a RO snapshot into RW mode, 6318 * keep an open file descriptor against a file, delete it and then 6319 * turn the snapshot back to RO mode before using it for a send 6320 * operation. So if we find such cases, ignore the inode and all its 6321 * items completely if it's a new inode, or if it's a changed inode 6322 * make sure all its previous paths (from the parent snapshot) are all 6323 * unlinked and all other the inode items are ignored. 6324 */ 6325 if (result == BTRFS_COMPARE_TREE_NEW || 6326 result == BTRFS_COMPARE_TREE_CHANGED) { 6327 u32 nlinks; 6328 6329 nlinks = btrfs_inode_nlink(sctx->left_path->nodes[0], left_ii); 6330 if (nlinks == 0) { 6331 sctx->ignore_cur_inode = true; 6332 if (result == BTRFS_COMPARE_TREE_CHANGED) 6333 ret = btrfs_unlink_all_paths(sctx); 6334 goto out; 6335 } 6336 } 6337 6338 if (result == BTRFS_COMPARE_TREE_NEW) { 6339 sctx->cur_inode_gen = left_gen; 6340 sctx->cur_inode_new = 1; 6341 sctx->cur_inode_deleted = 0; 6342 sctx->cur_inode_size = btrfs_inode_size( 6343 sctx->left_path->nodes[0], left_ii); 6344 sctx->cur_inode_mode = btrfs_inode_mode( 6345 sctx->left_path->nodes[0], left_ii); 6346 sctx->cur_inode_rdev = btrfs_inode_rdev( 6347 sctx->left_path->nodes[0], left_ii); 6348 if (sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) 6349 ret = send_create_inode_if_needed(sctx); 6350 } else if (result == BTRFS_COMPARE_TREE_DELETED) { 6351 sctx->cur_inode_gen = right_gen; 6352 sctx->cur_inode_new = 0; 6353 sctx->cur_inode_deleted = 1; 6354 sctx->cur_inode_size = btrfs_inode_size( 6355 sctx->right_path->nodes[0], right_ii); 6356 sctx->cur_inode_mode = btrfs_inode_mode( 6357 sctx->right_path->nodes[0], right_ii); 6358 } else if (result == BTRFS_COMPARE_TREE_CHANGED) { 6359 /* 6360 * We need to do some special handling in case the inode was 6361 * reported as changed with a changed generation number. This 6362 * means that the original inode was deleted and new inode 6363 * reused the same inum. So we have to treat the old inode as 6364 * deleted and the new one as new. 6365 */ 6366 if (sctx->cur_inode_new_gen) { 6367 /* 6368 * First, process the inode as if it was deleted. 6369 */ 6370 sctx->cur_inode_gen = right_gen; 6371 sctx->cur_inode_new = 0; 6372 sctx->cur_inode_deleted = 1; 6373 sctx->cur_inode_size = btrfs_inode_size( 6374 sctx->right_path->nodes[0], right_ii); 6375 sctx->cur_inode_mode = btrfs_inode_mode( 6376 sctx->right_path->nodes[0], right_ii); 6377 ret = process_all_refs(sctx, 6378 BTRFS_COMPARE_TREE_DELETED); 6379 if (ret < 0) 6380 goto out; 6381 6382 /* 6383 * Now process the inode as if it was new. 6384 */ 6385 sctx->cur_inode_gen = left_gen; 6386 sctx->cur_inode_new = 1; 6387 sctx->cur_inode_deleted = 0; 6388 sctx->cur_inode_size = btrfs_inode_size( 6389 sctx->left_path->nodes[0], left_ii); 6390 sctx->cur_inode_mode = btrfs_inode_mode( 6391 sctx->left_path->nodes[0], left_ii); 6392 sctx->cur_inode_rdev = btrfs_inode_rdev( 6393 sctx->left_path->nodes[0], left_ii); 6394 ret = send_create_inode_if_needed(sctx); 6395 if (ret < 0) 6396 goto out; 6397 6398 ret = process_all_refs(sctx, BTRFS_COMPARE_TREE_NEW); 6399 if (ret < 0) 6400 goto out; 6401 /* 6402 * Advance send_progress now as we did not get into 6403 * process_recorded_refs_if_needed in the new_gen case. 6404 */ 6405 sctx->send_progress = sctx->cur_ino + 1; 6406 6407 /* 6408 * Now process all extents and xattrs of the inode as if 6409 * they were all new. 6410 */ 6411 ret = process_all_extents(sctx); 6412 if (ret < 0) 6413 goto out; 6414 ret = process_all_new_xattrs(sctx); 6415 if (ret < 0) 6416 goto out; 6417 } else { 6418 sctx->cur_inode_gen = left_gen; 6419 sctx->cur_inode_new = 0; 6420 sctx->cur_inode_new_gen = 0; 6421 sctx->cur_inode_deleted = 0; 6422 sctx->cur_inode_size = btrfs_inode_size( 6423 sctx->left_path->nodes[0], left_ii); 6424 sctx->cur_inode_mode = btrfs_inode_mode( 6425 sctx->left_path->nodes[0], left_ii); 6426 } 6427 } 6428 6429 out: 6430 return ret; 6431 } 6432 6433 /* 6434 * We have to process new refs before deleted refs, but compare_trees gives us 6435 * the new and deleted refs mixed. To fix this, we record the new/deleted refs 6436 * first and later process them in process_recorded_refs. 6437 * For the cur_inode_new_gen case, we skip recording completely because 6438 * changed_inode did already initiate processing of refs. The reason for this is 6439 * that in this case, compare_tree actually compares the refs of 2 different 6440 * inodes. To fix this, process_all_refs is used in changed_inode to handle all 6441 * refs of the right tree as deleted and all refs of the left tree as new. 6442 */ 6443 static int changed_ref(struct send_ctx *sctx, 6444 enum btrfs_compare_tree_result result) 6445 { 6446 int ret = 0; 6447 6448 if (sctx->cur_ino != sctx->cmp_key->objectid) { 6449 inconsistent_snapshot_error(sctx, result, "reference"); 6450 return -EIO; 6451 } 6452 6453 if (!sctx->cur_inode_new_gen && 6454 sctx->cur_ino != BTRFS_FIRST_FREE_OBJECTID) { 6455 if (result == BTRFS_COMPARE_TREE_NEW) 6456 ret = record_new_ref(sctx); 6457 else if (result == BTRFS_COMPARE_TREE_DELETED) 6458 ret = record_deleted_ref(sctx); 6459 else if (result == BTRFS_COMPARE_TREE_CHANGED) 6460 ret = record_changed_ref(sctx); 6461 } 6462 6463 return ret; 6464 } 6465 6466 /* 6467 * Process new/deleted/changed xattrs. We skip processing in the 6468 * cur_inode_new_gen case because changed_inode did already initiate processing 6469 * of xattrs. The reason is the same as in changed_ref 6470 */ 6471 static int changed_xattr(struct send_ctx *sctx, 6472 enum btrfs_compare_tree_result result) 6473 { 6474 int ret = 0; 6475 6476 if (sctx->cur_ino != sctx->cmp_key->objectid) { 6477 inconsistent_snapshot_error(sctx, result, "xattr"); 6478 return -EIO; 6479 } 6480 6481 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) { 6482 if (result == BTRFS_COMPARE_TREE_NEW) 6483 ret = process_new_xattr(sctx); 6484 else if (result == BTRFS_COMPARE_TREE_DELETED) 6485 ret = process_deleted_xattr(sctx); 6486 else if (result == BTRFS_COMPARE_TREE_CHANGED) 6487 ret = process_changed_xattr(sctx); 6488 } 6489 6490 return ret; 6491 } 6492 6493 /* 6494 * Process new/deleted/changed extents. We skip processing in the 6495 * cur_inode_new_gen case because changed_inode did already initiate processing 6496 * of extents. The reason is the same as in changed_ref 6497 */ 6498 static int changed_extent(struct send_ctx *sctx, 6499 enum btrfs_compare_tree_result result) 6500 { 6501 int ret = 0; 6502 6503 /* 6504 * We have found an extent item that changed without the inode item 6505 * having changed. This can happen either after relocation (where the 6506 * disk_bytenr of an extent item is replaced at 6507 * relocation.c:replace_file_extents()) or after deduplication into a 6508 * file in both the parent and send snapshots (where an extent item can 6509 * get modified or replaced with a new one). Note that deduplication 6510 * updates the inode item, but it only changes the iversion (sequence 6511 * field in the inode item) of the inode, so if a file is deduplicated 6512 * the same amount of times in both the parent and send snapshots, its 6513 * iversion becames the same in both snapshots, whence the inode item is 6514 * the same on both snapshots. 6515 */ 6516 if (sctx->cur_ino != sctx->cmp_key->objectid) 6517 return 0; 6518 6519 if (!sctx->cur_inode_new_gen && !sctx->cur_inode_deleted) { 6520 if (result != BTRFS_COMPARE_TREE_DELETED) 6521 ret = process_extent(sctx, sctx->left_path, 6522 sctx->cmp_key); 6523 } 6524 6525 return ret; 6526 } 6527 6528 static int dir_changed(struct send_ctx *sctx, u64 dir) 6529 { 6530 u64 orig_gen, new_gen; 6531 int ret; 6532 6533 ret = get_inode_info(sctx->send_root, dir, NULL, &new_gen, NULL, NULL, 6534 NULL, NULL); 6535 if (ret) 6536 return ret; 6537 6538 ret = get_inode_info(sctx->parent_root, dir, NULL, &orig_gen, NULL, 6539 NULL, NULL, NULL); 6540 if (ret) 6541 return ret; 6542 6543 return (orig_gen != new_gen) ? 1 : 0; 6544 } 6545 6546 static int compare_refs(struct send_ctx *sctx, struct btrfs_path *path, 6547 struct btrfs_key *key) 6548 { 6549 struct btrfs_inode_extref *extref; 6550 struct extent_buffer *leaf; 6551 u64 dirid = 0, last_dirid = 0; 6552 unsigned long ptr; 6553 u32 item_size; 6554 u32 cur_offset = 0; 6555 int ref_name_len; 6556 int ret = 0; 6557 6558 /* Easy case, just check this one dirid */ 6559 if (key->type == BTRFS_INODE_REF_KEY) { 6560 dirid = key->offset; 6561 6562 ret = dir_changed(sctx, dirid); 6563 goto out; 6564 } 6565 6566 leaf = path->nodes[0]; 6567 item_size = btrfs_item_size_nr(leaf, path->slots[0]); 6568 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]); 6569 while (cur_offset < item_size) { 6570 extref = (struct btrfs_inode_extref *)(ptr + 6571 cur_offset); 6572 dirid = btrfs_inode_extref_parent(leaf, extref); 6573 ref_name_len = btrfs_inode_extref_name_len(leaf, extref); 6574 cur_offset += ref_name_len + sizeof(*extref); 6575 if (dirid == last_dirid) 6576 continue; 6577 ret = dir_changed(sctx, dirid); 6578 if (ret) 6579 break; 6580 last_dirid = dirid; 6581 } 6582 out: 6583 return ret; 6584 } 6585 6586 /* 6587 * Updates compare related fields in sctx and simply forwards to the actual 6588 * changed_xxx functions. 6589 */ 6590 static int changed_cb(struct btrfs_path *left_path, 6591 struct btrfs_path *right_path, 6592 struct btrfs_key *key, 6593 enum btrfs_compare_tree_result result, 6594 struct send_ctx *sctx) 6595 { 6596 int ret = 0; 6597 6598 if (result == BTRFS_COMPARE_TREE_SAME) { 6599 if (key->type == BTRFS_INODE_REF_KEY || 6600 key->type == BTRFS_INODE_EXTREF_KEY) { 6601 ret = compare_refs(sctx, left_path, key); 6602 if (!ret) 6603 return 0; 6604 if (ret < 0) 6605 return ret; 6606 } else if (key->type == BTRFS_EXTENT_DATA_KEY) { 6607 return maybe_send_hole(sctx, left_path, key); 6608 } else { 6609 return 0; 6610 } 6611 result = BTRFS_COMPARE_TREE_CHANGED; 6612 ret = 0; 6613 } 6614 6615 sctx->left_path = left_path; 6616 sctx->right_path = right_path; 6617 sctx->cmp_key = key; 6618 6619 ret = finish_inode_if_needed(sctx, 0); 6620 if (ret < 0) 6621 goto out; 6622 6623 /* Ignore non-FS objects */ 6624 if (key->objectid == BTRFS_FREE_INO_OBJECTID || 6625 key->objectid == BTRFS_FREE_SPACE_OBJECTID) 6626 goto out; 6627 6628 if (key->type == BTRFS_INODE_ITEM_KEY) { 6629 ret = changed_inode(sctx, result); 6630 } else if (!sctx->ignore_cur_inode) { 6631 if (key->type == BTRFS_INODE_REF_KEY || 6632 key->type == BTRFS_INODE_EXTREF_KEY) 6633 ret = changed_ref(sctx, result); 6634 else if (key->type == BTRFS_XATTR_ITEM_KEY) 6635 ret = changed_xattr(sctx, result); 6636 else if (key->type == BTRFS_EXTENT_DATA_KEY) 6637 ret = changed_extent(sctx, result); 6638 } 6639 6640 out: 6641 return ret; 6642 } 6643 6644 static int full_send_tree(struct send_ctx *sctx) 6645 { 6646 int ret; 6647 struct btrfs_root *send_root = sctx->send_root; 6648 struct btrfs_key key; 6649 struct btrfs_path *path; 6650 struct extent_buffer *eb; 6651 int slot; 6652 6653 path = alloc_path_for_send(); 6654 if (!path) 6655 return -ENOMEM; 6656 6657 key.objectid = BTRFS_FIRST_FREE_OBJECTID; 6658 key.type = BTRFS_INODE_ITEM_KEY; 6659 key.offset = 0; 6660 6661 ret = btrfs_search_slot_for_read(send_root, &key, path, 1, 0); 6662 if (ret < 0) 6663 goto out; 6664 if (ret) 6665 goto out_finish; 6666 6667 while (1) { 6668 eb = path->nodes[0]; 6669 slot = path->slots[0]; 6670 btrfs_item_key_to_cpu(eb, &key, slot); 6671 6672 ret = changed_cb(path, NULL, &key, 6673 BTRFS_COMPARE_TREE_NEW, sctx); 6674 if (ret < 0) 6675 goto out; 6676 6677 ret = btrfs_next_item(send_root, path); 6678 if (ret < 0) 6679 goto out; 6680 if (ret) { 6681 ret = 0; 6682 break; 6683 } 6684 } 6685 6686 out_finish: 6687 ret = finish_inode_if_needed(sctx, 1); 6688 6689 out: 6690 btrfs_free_path(path); 6691 return ret; 6692 } 6693 6694 static int tree_move_down(struct btrfs_path *path, int *level) 6695 { 6696 struct extent_buffer *eb; 6697 6698 BUG_ON(*level == 0); 6699 eb = btrfs_read_node_slot(path->nodes[*level], path->slots[*level]); 6700 if (IS_ERR(eb)) 6701 return PTR_ERR(eb); 6702 6703 path->nodes[*level - 1] = eb; 6704 path->slots[*level - 1] = 0; 6705 (*level)--; 6706 return 0; 6707 } 6708 6709 static int tree_move_next_or_upnext(struct btrfs_path *path, 6710 int *level, int root_level) 6711 { 6712 int ret = 0; 6713 int nritems; 6714 nritems = btrfs_header_nritems(path->nodes[*level]); 6715 6716 path->slots[*level]++; 6717 6718 while (path->slots[*level] >= nritems) { 6719 if (*level == root_level) 6720 return -1; 6721 6722 /* move upnext */ 6723 path->slots[*level] = 0; 6724 free_extent_buffer(path->nodes[*level]); 6725 path->nodes[*level] = NULL; 6726 (*level)++; 6727 path->slots[*level]++; 6728 6729 nritems = btrfs_header_nritems(path->nodes[*level]); 6730 ret = 1; 6731 } 6732 return ret; 6733 } 6734 6735 /* 6736 * Returns 1 if it had to move up and next. 0 is returned if it moved only next 6737 * or down. 6738 */ 6739 static int tree_advance(struct btrfs_path *path, 6740 int *level, int root_level, 6741 int allow_down, 6742 struct btrfs_key *key) 6743 { 6744 int ret; 6745 6746 if (*level == 0 || !allow_down) { 6747 ret = tree_move_next_or_upnext(path, level, root_level); 6748 } else { 6749 ret = tree_move_down(path, level); 6750 } 6751 if (ret >= 0) { 6752 if (*level == 0) 6753 btrfs_item_key_to_cpu(path->nodes[*level], key, 6754 path->slots[*level]); 6755 else 6756 btrfs_node_key_to_cpu(path->nodes[*level], key, 6757 path->slots[*level]); 6758 } 6759 return ret; 6760 } 6761 6762 static int tree_compare_item(struct btrfs_path *left_path, 6763 struct btrfs_path *right_path, 6764 char *tmp_buf) 6765 { 6766 int cmp; 6767 int len1, len2; 6768 unsigned long off1, off2; 6769 6770 len1 = btrfs_item_size_nr(left_path->nodes[0], left_path->slots[0]); 6771 len2 = btrfs_item_size_nr(right_path->nodes[0], right_path->slots[0]); 6772 if (len1 != len2) 6773 return 1; 6774 6775 off1 = btrfs_item_ptr_offset(left_path->nodes[0], left_path->slots[0]); 6776 off2 = btrfs_item_ptr_offset(right_path->nodes[0], 6777 right_path->slots[0]); 6778 6779 read_extent_buffer(left_path->nodes[0], tmp_buf, off1, len1); 6780 6781 cmp = memcmp_extent_buffer(right_path->nodes[0], tmp_buf, off2, len1); 6782 if (cmp) 6783 return 1; 6784 return 0; 6785 } 6786 6787 /* 6788 * This function compares two trees and calls the provided callback for 6789 * every changed/new/deleted item it finds. 6790 * If shared tree blocks are encountered, whole subtrees are skipped, making 6791 * the compare pretty fast on snapshotted subvolumes. 6792 * 6793 * This currently works on commit roots only. As commit roots are read only, 6794 * we don't do any locking. The commit roots are protected with transactions. 6795 * Transactions are ended and rejoined when a commit is tried in between. 6796 * 6797 * This function checks for modifications done to the trees while comparing. 6798 * If it detects a change, it aborts immediately. 6799 */ 6800 static int btrfs_compare_trees(struct btrfs_root *left_root, 6801 struct btrfs_root *right_root, struct send_ctx *sctx) 6802 { 6803 struct btrfs_fs_info *fs_info = left_root->fs_info; 6804 int ret; 6805 int cmp; 6806 struct btrfs_path *left_path = NULL; 6807 struct btrfs_path *right_path = NULL; 6808 struct btrfs_key left_key; 6809 struct btrfs_key right_key; 6810 char *tmp_buf = NULL; 6811 int left_root_level; 6812 int right_root_level; 6813 int left_level; 6814 int right_level; 6815 int left_end_reached; 6816 int right_end_reached; 6817 int advance_left; 6818 int advance_right; 6819 u64 left_blockptr; 6820 u64 right_blockptr; 6821 u64 left_gen; 6822 u64 right_gen; 6823 6824 left_path = btrfs_alloc_path(); 6825 if (!left_path) { 6826 ret = -ENOMEM; 6827 goto out; 6828 } 6829 right_path = btrfs_alloc_path(); 6830 if (!right_path) { 6831 ret = -ENOMEM; 6832 goto out; 6833 } 6834 6835 tmp_buf = kvmalloc(fs_info->nodesize, GFP_KERNEL); 6836 if (!tmp_buf) { 6837 ret = -ENOMEM; 6838 goto out; 6839 } 6840 6841 left_path->search_commit_root = 1; 6842 left_path->skip_locking = 1; 6843 right_path->search_commit_root = 1; 6844 right_path->skip_locking = 1; 6845 6846 /* 6847 * Strategy: Go to the first items of both trees. Then do 6848 * 6849 * If both trees are at level 0 6850 * Compare keys of current items 6851 * If left < right treat left item as new, advance left tree 6852 * and repeat 6853 * If left > right treat right item as deleted, advance right tree 6854 * and repeat 6855 * If left == right do deep compare of items, treat as changed if 6856 * needed, advance both trees and repeat 6857 * If both trees are at the same level but not at level 0 6858 * Compare keys of current nodes/leafs 6859 * If left < right advance left tree and repeat 6860 * If left > right advance right tree and repeat 6861 * If left == right compare blockptrs of the next nodes/leafs 6862 * If they match advance both trees but stay at the same level 6863 * and repeat 6864 * If they don't match advance both trees while allowing to go 6865 * deeper and repeat 6866 * If tree levels are different 6867 * Advance the tree that needs it and repeat 6868 * 6869 * Advancing a tree means: 6870 * If we are at level 0, try to go to the next slot. If that's not 6871 * possible, go one level up and repeat. Stop when we found a level 6872 * where we could go to the next slot. We may at this point be on a 6873 * node or a leaf. 6874 * 6875 * If we are not at level 0 and not on shared tree blocks, go one 6876 * level deeper. 6877 * 6878 * If we are not at level 0 and on shared tree blocks, go one slot to 6879 * the right if possible or go up and right. 6880 */ 6881 6882 down_read(&fs_info->commit_root_sem); 6883 left_level = btrfs_header_level(left_root->commit_root); 6884 left_root_level = left_level; 6885 left_path->nodes[left_level] = 6886 btrfs_clone_extent_buffer(left_root->commit_root); 6887 if (!left_path->nodes[left_level]) { 6888 up_read(&fs_info->commit_root_sem); 6889 ret = -ENOMEM; 6890 goto out; 6891 } 6892 6893 right_level = btrfs_header_level(right_root->commit_root); 6894 right_root_level = right_level; 6895 right_path->nodes[right_level] = 6896 btrfs_clone_extent_buffer(right_root->commit_root); 6897 if (!right_path->nodes[right_level]) { 6898 up_read(&fs_info->commit_root_sem); 6899 ret = -ENOMEM; 6900 goto out; 6901 } 6902 up_read(&fs_info->commit_root_sem); 6903 6904 if (left_level == 0) 6905 btrfs_item_key_to_cpu(left_path->nodes[left_level], 6906 &left_key, left_path->slots[left_level]); 6907 else 6908 btrfs_node_key_to_cpu(left_path->nodes[left_level], 6909 &left_key, left_path->slots[left_level]); 6910 if (right_level == 0) 6911 btrfs_item_key_to_cpu(right_path->nodes[right_level], 6912 &right_key, right_path->slots[right_level]); 6913 else 6914 btrfs_node_key_to_cpu(right_path->nodes[right_level], 6915 &right_key, right_path->slots[right_level]); 6916 6917 left_end_reached = right_end_reached = 0; 6918 advance_left = advance_right = 0; 6919 6920 while (1) { 6921 cond_resched(); 6922 if (advance_left && !left_end_reached) { 6923 ret = tree_advance(left_path, &left_level, 6924 left_root_level, 6925 advance_left != ADVANCE_ONLY_NEXT, 6926 &left_key); 6927 if (ret == -1) 6928 left_end_reached = ADVANCE; 6929 else if (ret < 0) 6930 goto out; 6931 advance_left = 0; 6932 } 6933 if (advance_right && !right_end_reached) { 6934 ret = tree_advance(right_path, &right_level, 6935 right_root_level, 6936 advance_right != ADVANCE_ONLY_NEXT, 6937 &right_key); 6938 if (ret == -1) 6939 right_end_reached = ADVANCE; 6940 else if (ret < 0) 6941 goto out; 6942 advance_right = 0; 6943 } 6944 6945 if (left_end_reached && right_end_reached) { 6946 ret = 0; 6947 goto out; 6948 } else if (left_end_reached) { 6949 if (right_level == 0) { 6950 ret = changed_cb(left_path, right_path, 6951 &right_key, 6952 BTRFS_COMPARE_TREE_DELETED, 6953 sctx); 6954 if (ret < 0) 6955 goto out; 6956 } 6957 advance_right = ADVANCE; 6958 continue; 6959 } else if (right_end_reached) { 6960 if (left_level == 0) { 6961 ret = changed_cb(left_path, right_path, 6962 &left_key, 6963 BTRFS_COMPARE_TREE_NEW, 6964 sctx); 6965 if (ret < 0) 6966 goto out; 6967 } 6968 advance_left = ADVANCE; 6969 continue; 6970 } 6971 6972 if (left_level == 0 && right_level == 0) { 6973 cmp = btrfs_comp_cpu_keys(&left_key, &right_key); 6974 if (cmp < 0) { 6975 ret = changed_cb(left_path, right_path, 6976 &left_key, 6977 BTRFS_COMPARE_TREE_NEW, 6978 sctx); 6979 if (ret < 0) 6980 goto out; 6981 advance_left = ADVANCE; 6982 } else if (cmp > 0) { 6983 ret = changed_cb(left_path, right_path, 6984 &right_key, 6985 BTRFS_COMPARE_TREE_DELETED, 6986 sctx); 6987 if (ret < 0) 6988 goto out; 6989 advance_right = ADVANCE; 6990 } else { 6991 enum btrfs_compare_tree_result result; 6992 6993 WARN_ON(!extent_buffer_uptodate(left_path->nodes[0])); 6994 ret = tree_compare_item(left_path, right_path, 6995 tmp_buf); 6996 if (ret) 6997 result = BTRFS_COMPARE_TREE_CHANGED; 6998 else 6999 result = BTRFS_COMPARE_TREE_SAME; 7000 ret = changed_cb(left_path, right_path, 7001 &left_key, result, sctx); 7002 if (ret < 0) 7003 goto out; 7004 advance_left = ADVANCE; 7005 advance_right = ADVANCE; 7006 } 7007 } else if (left_level == right_level) { 7008 cmp = btrfs_comp_cpu_keys(&left_key, &right_key); 7009 if (cmp < 0) { 7010 advance_left = ADVANCE; 7011 } else if (cmp > 0) { 7012 advance_right = ADVANCE; 7013 } else { 7014 left_blockptr = btrfs_node_blockptr( 7015 left_path->nodes[left_level], 7016 left_path->slots[left_level]); 7017 right_blockptr = btrfs_node_blockptr( 7018 right_path->nodes[right_level], 7019 right_path->slots[right_level]); 7020 left_gen = btrfs_node_ptr_generation( 7021 left_path->nodes[left_level], 7022 left_path->slots[left_level]); 7023 right_gen = btrfs_node_ptr_generation( 7024 right_path->nodes[right_level], 7025 right_path->slots[right_level]); 7026 if (left_blockptr == right_blockptr && 7027 left_gen == right_gen) { 7028 /* 7029 * As we're on a shared block, don't 7030 * allow to go deeper. 7031 */ 7032 advance_left = ADVANCE_ONLY_NEXT; 7033 advance_right = ADVANCE_ONLY_NEXT; 7034 } else { 7035 advance_left = ADVANCE; 7036 advance_right = ADVANCE; 7037 } 7038 } 7039 } else if (left_level < right_level) { 7040 advance_right = ADVANCE; 7041 } else { 7042 advance_left = ADVANCE; 7043 } 7044 } 7045 7046 out: 7047 btrfs_free_path(left_path); 7048 btrfs_free_path(right_path); 7049 kvfree(tmp_buf); 7050 return ret; 7051 } 7052 7053 static int send_subvol(struct send_ctx *sctx) 7054 { 7055 int ret; 7056 7057 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_STREAM_HEADER)) { 7058 ret = send_header(sctx); 7059 if (ret < 0) 7060 goto out; 7061 } 7062 7063 ret = send_subvol_begin(sctx); 7064 if (ret < 0) 7065 goto out; 7066 7067 if (sctx->parent_root) { 7068 ret = btrfs_compare_trees(sctx->send_root, sctx->parent_root, sctx); 7069 if (ret < 0) 7070 goto out; 7071 ret = finish_inode_if_needed(sctx, 1); 7072 if (ret < 0) 7073 goto out; 7074 } else { 7075 ret = full_send_tree(sctx); 7076 if (ret < 0) 7077 goto out; 7078 } 7079 7080 out: 7081 free_recorded_refs(sctx); 7082 return ret; 7083 } 7084 7085 /* 7086 * If orphan cleanup did remove any orphans from a root, it means the tree 7087 * was modified and therefore the commit root is not the same as the current 7088 * root anymore. This is a problem, because send uses the commit root and 7089 * therefore can see inode items that don't exist in the current root anymore, 7090 * and for example make calls to btrfs_iget, which will do tree lookups based 7091 * on the current root and not on the commit root. Those lookups will fail, 7092 * returning a -ESTALE error, and making send fail with that error. So make 7093 * sure a send does not see any orphans we have just removed, and that it will 7094 * see the same inodes regardless of whether a transaction commit happened 7095 * before it started (meaning that the commit root will be the same as the 7096 * current root) or not. 7097 */ 7098 static int ensure_commit_roots_uptodate(struct send_ctx *sctx) 7099 { 7100 int i; 7101 struct btrfs_trans_handle *trans = NULL; 7102 7103 again: 7104 if (sctx->parent_root && 7105 sctx->parent_root->node != sctx->parent_root->commit_root) 7106 goto commit_trans; 7107 7108 for (i = 0; i < sctx->clone_roots_cnt; i++) 7109 if (sctx->clone_roots[i].root->node != 7110 sctx->clone_roots[i].root->commit_root) 7111 goto commit_trans; 7112 7113 if (trans) 7114 return btrfs_end_transaction(trans); 7115 7116 return 0; 7117 7118 commit_trans: 7119 /* Use any root, all fs roots will get their commit roots updated. */ 7120 if (!trans) { 7121 trans = btrfs_join_transaction(sctx->send_root); 7122 if (IS_ERR(trans)) 7123 return PTR_ERR(trans); 7124 goto again; 7125 } 7126 7127 return btrfs_commit_transaction(trans); 7128 } 7129 7130 /* 7131 * Make sure any existing dellaloc is flushed for any root used by a send 7132 * operation so that we do not miss any data and we do not race with writeback 7133 * finishing and changing a tree while send is using the tree. This could 7134 * happen if a subvolume is in RW mode, has delalloc, is turned to RO mode and 7135 * a send operation then uses the subvolume. 7136 * After flushing delalloc ensure_commit_roots_uptodate() must be called. 7137 */ 7138 static int flush_delalloc_roots(struct send_ctx *sctx) 7139 { 7140 struct btrfs_root *root = sctx->parent_root; 7141 int ret; 7142 int i; 7143 7144 if (root) { 7145 ret = btrfs_start_delalloc_snapshot(root); 7146 if (ret) 7147 return ret; 7148 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX); 7149 } 7150 7151 for (i = 0; i < sctx->clone_roots_cnt; i++) { 7152 root = sctx->clone_roots[i].root; 7153 ret = btrfs_start_delalloc_snapshot(root); 7154 if (ret) 7155 return ret; 7156 btrfs_wait_ordered_extents(root, U64_MAX, 0, U64_MAX); 7157 } 7158 7159 return 0; 7160 } 7161 7162 static void btrfs_root_dec_send_in_progress(struct btrfs_root* root) 7163 { 7164 spin_lock(&root->root_item_lock); 7165 root->send_in_progress--; 7166 /* 7167 * Not much left to do, we don't know why it's unbalanced and 7168 * can't blindly reset it to 0. 7169 */ 7170 if (root->send_in_progress < 0) 7171 btrfs_err(root->fs_info, 7172 "send_in_progress unbalanced %d root %llu", 7173 root->send_in_progress, root->root_key.objectid); 7174 spin_unlock(&root->root_item_lock); 7175 } 7176 7177 static void dedupe_in_progress_warn(const struct btrfs_root *root) 7178 { 7179 btrfs_warn_rl(root->fs_info, 7180 "cannot use root %llu for send while deduplications on it are in progress (%d in progress)", 7181 root->root_key.objectid, root->dedupe_in_progress); 7182 } 7183 7184 long btrfs_ioctl_send(struct file *mnt_file, struct btrfs_ioctl_send_args *arg) 7185 { 7186 int ret = 0; 7187 struct btrfs_root *send_root = BTRFS_I(file_inode(mnt_file))->root; 7188 struct btrfs_fs_info *fs_info = send_root->fs_info; 7189 struct btrfs_root *clone_root; 7190 struct send_ctx *sctx = NULL; 7191 u32 i; 7192 u64 *clone_sources_tmp = NULL; 7193 int clone_sources_to_rollback = 0; 7194 size_t alloc_size; 7195 int sort_clone_roots = 0; 7196 7197 if (!capable(CAP_SYS_ADMIN)) 7198 return -EPERM; 7199 7200 /* 7201 * The subvolume must remain read-only during send, protect against 7202 * making it RW. This also protects against deletion. 7203 */ 7204 spin_lock(&send_root->root_item_lock); 7205 if (btrfs_root_readonly(send_root) && send_root->dedupe_in_progress) { 7206 dedupe_in_progress_warn(send_root); 7207 spin_unlock(&send_root->root_item_lock); 7208 return -EAGAIN; 7209 } 7210 send_root->send_in_progress++; 7211 spin_unlock(&send_root->root_item_lock); 7212 7213 /* 7214 * Userspace tools do the checks and warn the user if it's 7215 * not RO. 7216 */ 7217 if (!btrfs_root_readonly(send_root)) { 7218 ret = -EPERM; 7219 goto out; 7220 } 7221 7222 /* 7223 * Check that we don't overflow at later allocations, we request 7224 * clone_sources_count + 1 items, and compare to unsigned long inside 7225 * access_ok. 7226 */ 7227 if (arg->clone_sources_count > 7228 ULONG_MAX / sizeof(struct clone_root) - 1) { 7229 ret = -EINVAL; 7230 goto out; 7231 } 7232 7233 if (arg->flags & ~BTRFS_SEND_FLAG_MASK) { 7234 ret = -EINVAL; 7235 goto out; 7236 } 7237 7238 sctx = kzalloc(sizeof(struct send_ctx), GFP_KERNEL); 7239 if (!sctx) { 7240 ret = -ENOMEM; 7241 goto out; 7242 } 7243 7244 INIT_LIST_HEAD(&sctx->new_refs); 7245 INIT_LIST_HEAD(&sctx->deleted_refs); 7246 INIT_RADIX_TREE(&sctx->name_cache, GFP_KERNEL); 7247 INIT_LIST_HEAD(&sctx->name_cache_list); 7248 7249 sctx->flags = arg->flags; 7250 7251 sctx->send_filp = fget(arg->send_fd); 7252 if (!sctx->send_filp) { 7253 ret = -EBADF; 7254 goto out; 7255 } 7256 7257 sctx->send_root = send_root; 7258 /* 7259 * Unlikely but possible, if the subvolume is marked for deletion but 7260 * is slow to remove the directory entry, send can still be started 7261 */ 7262 if (btrfs_root_dead(sctx->send_root)) { 7263 ret = -EPERM; 7264 goto out; 7265 } 7266 7267 sctx->clone_roots_cnt = arg->clone_sources_count; 7268 7269 sctx->send_max_size = BTRFS_SEND_BUF_SIZE; 7270 sctx->send_buf = kvmalloc(sctx->send_max_size, GFP_KERNEL); 7271 if (!sctx->send_buf) { 7272 ret = -ENOMEM; 7273 goto out; 7274 } 7275 7276 sctx->pending_dir_moves = RB_ROOT; 7277 sctx->waiting_dir_moves = RB_ROOT; 7278 sctx->orphan_dirs = RB_ROOT; 7279 7280 sctx->clone_roots = kvcalloc(sizeof(*sctx->clone_roots), 7281 arg->clone_sources_count + 1, 7282 GFP_KERNEL); 7283 if (!sctx->clone_roots) { 7284 ret = -ENOMEM; 7285 goto out; 7286 } 7287 7288 alloc_size = array_size(sizeof(*arg->clone_sources), 7289 arg->clone_sources_count); 7290 7291 if (arg->clone_sources_count) { 7292 clone_sources_tmp = kvmalloc(alloc_size, GFP_KERNEL); 7293 if (!clone_sources_tmp) { 7294 ret = -ENOMEM; 7295 goto out; 7296 } 7297 7298 ret = copy_from_user(clone_sources_tmp, arg->clone_sources, 7299 alloc_size); 7300 if (ret) { 7301 ret = -EFAULT; 7302 goto out; 7303 } 7304 7305 for (i = 0; i < arg->clone_sources_count; i++) { 7306 clone_root = btrfs_get_fs_root(fs_info, 7307 clone_sources_tmp[i], true); 7308 if (IS_ERR(clone_root)) { 7309 ret = PTR_ERR(clone_root); 7310 goto out; 7311 } 7312 spin_lock(&clone_root->root_item_lock); 7313 if (!btrfs_root_readonly(clone_root) || 7314 btrfs_root_dead(clone_root)) { 7315 spin_unlock(&clone_root->root_item_lock); 7316 btrfs_put_root(clone_root); 7317 ret = -EPERM; 7318 goto out; 7319 } 7320 if (clone_root->dedupe_in_progress) { 7321 dedupe_in_progress_warn(clone_root); 7322 spin_unlock(&clone_root->root_item_lock); 7323 btrfs_put_root(clone_root); 7324 ret = -EAGAIN; 7325 goto out; 7326 } 7327 clone_root->send_in_progress++; 7328 spin_unlock(&clone_root->root_item_lock); 7329 7330 sctx->clone_roots[i].root = clone_root; 7331 clone_sources_to_rollback = i + 1; 7332 } 7333 kvfree(clone_sources_tmp); 7334 clone_sources_tmp = NULL; 7335 } 7336 7337 if (arg->parent_root) { 7338 sctx->parent_root = btrfs_get_fs_root(fs_info, arg->parent_root, 7339 true); 7340 if (IS_ERR(sctx->parent_root)) { 7341 ret = PTR_ERR(sctx->parent_root); 7342 goto out; 7343 } 7344 7345 spin_lock(&sctx->parent_root->root_item_lock); 7346 sctx->parent_root->send_in_progress++; 7347 if (!btrfs_root_readonly(sctx->parent_root) || 7348 btrfs_root_dead(sctx->parent_root)) { 7349 spin_unlock(&sctx->parent_root->root_item_lock); 7350 ret = -EPERM; 7351 goto out; 7352 } 7353 if (sctx->parent_root->dedupe_in_progress) { 7354 dedupe_in_progress_warn(sctx->parent_root); 7355 spin_unlock(&sctx->parent_root->root_item_lock); 7356 ret = -EAGAIN; 7357 goto out; 7358 } 7359 spin_unlock(&sctx->parent_root->root_item_lock); 7360 } 7361 7362 /* 7363 * Clones from send_root are allowed, but only if the clone source 7364 * is behind the current send position. This is checked while searching 7365 * for possible clone sources. 7366 */ 7367 sctx->clone_roots[sctx->clone_roots_cnt++].root = 7368 btrfs_grab_root(sctx->send_root); 7369 7370 /* We do a bsearch later */ 7371 sort(sctx->clone_roots, sctx->clone_roots_cnt, 7372 sizeof(*sctx->clone_roots), __clone_root_cmp_sort, 7373 NULL); 7374 sort_clone_roots = 1; 7375 7376 ret = flush_delalloc_roots(sctx); 7377 if (ret) 7378 goto out; 7379 7380 ret = ensure_commit_roots_uptodate(sctx); 7381 if (ret) 7382 goto out; 7383 7384 mutex_lock(&fs_info->balance_mutex); 7385 if (test_bit(BTRFS_FS_BALANCE_RUNNING, &fs_info->flags)) { 7386 mutex_unlock(&fs_info->balance_mutex); 7387 btrfs_warn_rl(fs_info, 7388 "cannot run send because a balance operation is in progress"); 7389 ret = -EAGAIN; 7390 goto out; 7391 } 7392 fs_info->send_in_progress++; 7393 mutex_unlock(&fs_info->balance_mutex); 7394 7395 current->journal_info = BTRFS_SEND_TRANS_STUB; 7396 ret = send_subvol(sctx); 7397 current->journal_info = NULL; 7398 mutex_lock(&fs_info->balance_mutex); 7399 fs_info->send_in_progress--; 7400 mutex_unlock(&fs_info->balance_mutex); 7401 if (ret < 0) 7402 goto out; 7403 7404 if (!(sctx->flags & BTRFS_SEND_FLAG_OMIT_END_CMD)) { 7405 ret = begin_cmd(sctx, BTRFS_SEND_C_END); 7406 if (ret < 0) 7407 goto out; 7408 ret = send_cmd(sctx); 7409 if (ret < 0) 7410 goto out; 7411 } 7412 7413 out: 7414 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)); 7415 while (sctx && !RB_EMPTY_ROOT(&sctx->pending_dir_moves)) { 7416 struct rb_node *n; 7417 struct pending_dir_move *pm; 7418 7419 n = rb_first(&sctx->pending_dir_moves); 7420 pm = rb_entry(n, struct pending_dir_move, node); 7421 while (!list_empty(&pm->list)) { 7422 struct pending_dir_move *pm2; 7423 7424 pm2 = list_first_entry(&pm->list, 7425 struct pending_dir_move, list); 7426 free_pending_move(sctx, pm2); 7427 } 7428 free_pending_move(sctx, pm); 7429 } 7430 7431 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)); 7432 while (sctx && !RB_EMPTY_ROOT(&sctx->waiting_dir_moves)) { 7433 struct rb_node *n; 7434 struct waiting_dir_move *dm; 7435 7436 n = rb_first(&sctx->waiting_dir_moves); 7437 dm = rb_entry(n, struct waiting_dir_move, node); 7438 rb_erase(&dm->node, &sctx->waiting_dir_moves); 7439 kfree(dm); 7440 } 7441 7442 WARN_ON(sctx && !ret && !RB_EMPTY_ROOT(&sctx->orphan_dirs)); 7443 while (sctx && !RB_EMPTY_ROOT(&sctx->orphan_dirs)) { 7444 struct rb_node *n; 7445 struct orphan_dir_info *odi; 7446 7447 n = rb_first(&sctx->orphan_dirs); 7448 odi = rb_entry(n, struct orphan_dir_info, node); 7449 free_orphan_dir_info(sctx, odi); 7450 } 7451 7452 if (sort_clone_roots) { 7453 for (i = 0; i < sctx->clone_roots_cnt; i++) { 7454 btrfs_root_dec_send_in_progress( 7455 sctx->clone_roots[i].root); 7456 btrfs_put_root(sctx->clone_roots[i].root); 7457 } 7458 } else { 7459 for (i = 0; sctx && i < clone_sources_to_rollback; i++) { 7460 btrfs_root_dec_send_in_progress( 7461 sctx->clone_roots[i].root); 7462 btrfs_put_root(sctx->clone_roots[i].root); 7463 } 7464 7465 btrfs_root_dec_send_in_progress(send_root); 7466 } 7467 if (sctx && !IS_ERR_OR_NULL(sctx->parent_root)) { 7468 btrfs_root_dec_send_in_progress(sctx->parent_root); 7469 btrfs_put_root(sctx->parent_root); 7470 } 7471 7472 kvfree(clone_sources_tmp); 7473 7474 if (sctx) { 7475 if (sctx->send_filp) 7476 fput(sctx->send_filp); 7477 7478 kvfree(sctx->clone_roots); 7479 kvfree(sctx->send_buf); 7480 7481 name_cache_free(sctx); 7482 7483 kfree(sctx); 7484 } 7485 7486 return ret; 7487 } 7488