1 // SPDX-License-Identifier: GPL-2.0 2 3 #include <linux/blkdev.h> 4 #include <linux/iversion.h> 5 #include "ctree.h" 6 #include "fs.h" 7 #include "messages.h" 8 #include "compression.h" 9 #include "delalloc-space.h" 10 #include "disk-io.h" 11 #include "reflink.h" 12 #include "transaction.h" 13 #include "subpage.h" 14 #include "accessors.h" 15 #include "file-item.h" 16 #include "file.h" 17 #include "super.h" 18 19 #define BTRFS_MAX_DEDUPE_LEN SZ_16M 20 21 static int clone_finish_inode_update(struct btrfs_trans_handle *trans, 22 struct inode *inode, 23 u64 endoff, 24 const u64 destoff, 25 const u64 olen, 26 int no_time_update) 27 { 28 int ret; 29 30 inode_inc_iversion(inode); 31 if (!no_time_update) { 32 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode)); 33 } 34 /* 35 * We round up to the block size at eof when determining which 36 * extents to clone above, but shouldn't round up the file size. 37 */ 38 if (endoff > destoff + olen) 39 endoff = destoff + olen; 40 if (endoff > inode->i_size) { 41 i_size_write(inode, endoff); 42 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0); 43 } 44 45 ret = btrfs_update_inode(trans, BTRFS_I(inode)); 46 if (ret) { 47 btrfs_abort_transaction(trans, ret); 48 btrfs_end_transaction(trans); 49 return ret; 50 } 51 return btrfs_end_transaction(trans); 52 } 53 54 static int copy_inline_to_page(struct btrfs_inode *inode, 55 const u64 file_offset, 56 char *inline_data, 57 const u64 size, 58 const u64 datal, 59 const u8 comp_type) 60 { 61 struct btrfs_fs_info *fs_info = inode->root->fs_info; 62 const u32 block_size = fs_info->sectorsize; 63 const u64 range_end = file_offset + block_size - 1; 64 const size_t inline_size = size - btrfs_file_extent_calc_inline_size(0); 65 char *data_start = inline_data + btrfs_file_extent_calc_inline_size(0); 66 struct extent_changeset *data_reserved = NULL; 67 struct folio *folio = NULL; 68 struct address_space *mapping = inode->vfs_inode.i_mapping; 69 int ret; 70 71 ASSERT(IS_ALIGNED(file_offset, block_size)); 72 73 /* 74 * We have flushed and locked the ranges of the source and destination 75 * inodes, we also have locked the inodes, so we are safe to do a 76 * reservation here. Also we must not do the reservation while holding 77 * a transaction open, otherwise we would deadlock. 78 */ 79 ret = btrfs_delalloc_reserve_space(inode, &data_reserved, file_offset, 80 block_size); 81 if (ret) 82 goto out; 83 84 folio = __filemap_get_folio(mapping, file_offset >> PAGE_SHIFT, 85 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, 86 btrfs_alloc_write_mask(mapping)); 87 if (IS_ERR(folio)) { 88 ret = PTR_ERR(folio); 89 goto out_unlock; 90 } 91 92 ret = set_folio_extent_mapped(folio); 93 if (ret < 0) 94 goto out_unlock; 95 96 btrfs_clear_extent_bit(&inode->io_tree, file_offset, range_end, 97 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, NULL); 98 ret = btrfs_set_extent_delalloc(inode, file_offset, range_end, 0, NULL); 99 if (ret) 100 goto out_unlock; 101 102 /* 103 * After dirtying the page our caller will need to start a transaction, 104 * and if we are low on metadata free space, that can cause flushing of 105 * delalloc for all inodes in order to get metadata space released. 106 * However we are holding the range locked for the whole duration of 107 * the clone/dedupe operation, so we may deadlock if that happens and no 108 * other task releases enough space. So mark this inode as not being 109 * possible to flush to avoid such deadlock. We will clear that flag 110 * when we finish cloning all extents, since a transaction is started 111 * after finding each extent to clone. 112 */ 113 set_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &inode->runtime_flags); 114 115 if (comp_type == BTRFS_COMPRESS_NONE) { 116 memcpy_to_folio(folio, offset_in_folio(folio, file_offset), data_start, 117 datal); 118 } else { 119 ret = btrfs_decompress(comp_type, data_start, folio, 120 offset_in_folio(folio, file_offset), 121 inline_size, datal); 122 if (ret) 123 goto out_unlock; 124 flush_dcache_folio(folio); 125 } 126 127 /* 128 * If our inline data is smaller then the block/page size, then the 129 * remaining of the block/page is equivalent to zeroes. We had something 130 * like the following done: 131 * 132 * $ xfs_io -f -c "pwrite -S 0xab 0 500" file 133 * $ sync # (or fsync) 134 * $ xfs_io -c "falloc 0 4K" file 135 * $ xfs_io -c "pwrite -S 0xcd 4K 4K" 136 * 137 * So what's in the range [500, 4095] corresponds to zeroes. 138 */ 139 if (datal < block_size) 140 folio_zero_range(folio, datal, block_size - datal); 141 142 btrfs_folio_set_uptodate(fs_info, folio, file_offset, block_size); 143 btrfs_folio_clear_checked(fs_info, folio, file_offset, block_size); 144 btrfs_folio_set_dirty(fs_info, folio, file_offset, block_size); 145 out_unlock: 146 if (!IS_ERR(folio)) { 147 folio_unlock(folio); 148 folio_put(folio); 149 } 150 if (ret) 151 btrfs_delalloc_release_space(inode, data_reserved, file_offset, 152 block_size, true); 153 btrfs_delalloc_release_extents(inode, block_size); 154 out: 155 extent_changeset_free(data_reserved); 156 157 return ret; 158 } 159 160 /* 161 * Deal with cloning of inline extents. We try to copy the inline extent from 162 * the source inode to destination inode when possible. When not possible we 163 * copy the inline extent's data into the respective page of the inode. 164 */ 165 static int clone_copy_inline_extent(struct btrfs_inode *inode, 166 struct btrfs_path *path, 167 struct btrfs_key *new_key, 168 const u64 drop_start, 169 const u64 datal, 170 const u64 size, 171 const u8 comp_type, 172 char *inline_data, 173 struct btrfs_trans_handle **trans_out) 174 { 175 struct btrfs_root *root = inode->root; 176 struct btrfs_fs_info *fs_info = root->fs_info; 177 const u64 aligned_end = ALIGN(new_key->offset + datal, 178 fs_info->sectorsize); 179 struct btrfs_trans_handle *trans = NULL; 180 struct btrfs_drop_extents_args drop_args = { 0 }; 181 int ret; 182 struct btrfs_key key; 183 184 if (new_key->offset > 0) { 185 ret = copy_inline_to_page(inode, new_key->offset, 186 inline_data, size, datal, comp_type); 187 goto out; 188 } 189 190 key.objectid = btrfs_ino(inode); 191 key.type = BTRFS_EXTENT_DATA_KEY; 192 key.offset = 0; 193 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 194 if (ret < 0) { 195 return ret; 196 } else if (ret > 0) { 197 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) { 198 ret = btrfs_next_leaf(root, path); 199 if (ret < 0) 200 return ret; 201 else if (ret > 0) 202 goto copy_inline_extent; 203 } 204 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 205 if (key.objectid == btrfs_ino(inode) && 206 key.type == BTRFS_EXTENT_DATA_KEY) { 207 /* 208 * There's an implicit hole at file offset 0, copy the 209 * inline extent's data to the page. 210 */ 211 ASSERT(key.offset > 0); 212 goto copy_to_page; 213 } 214 } else if (i_size_read(&inode->vfs_inode) <= datal) { 215 struct btrfs_file_extent_item *ei; 216 217 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 218 struct btrfs_file_extent_item); 219 /* 220 * If it's an inline extent replace it with the source inline 221 * extent, otherwise copy the source inline extent data into 222 * the respective page at the destination inode. 223 */ 224 if (btrfs_file_extent_type(path->nodes[0], ei) == 225 BTRFS_FILE_EXTENT_INLINE) 226 goto copy_inline_extent; 227 228 goto copy_to_page; 229 } 230 231 copy_inline_extent: 232 /* 233 * We have no extent items, or we have an extent at offset 0 which may 234 * or may not be inlined. All these cases are dealt the same way. 235 */ 236 if (i_size_read(&inode->vfs_inode) > datal) { 237 /* 238 * At the destination offset 0 we have either a hole, a regular 239 * extent or an inline extent larger then the one we want to 240 * clone. Deal with all these cases by copying the inline extent 241 * data into the respective page at the destination inode. 242 */ 243 goto copy_to_page; 244 } 245 246 /* 247 * Release path before starting a new transaction so we don't hold locks 248 * that would confuse lockdep. 249 */ 250 btrfs_release_path(path); 251 /* 252 * If we end up here it means were copy the inline extent into a leaf 253 * of the destination inode. We know we will drop or adjust at most one 254 * extent item in the destination root. 255 * 256 * 1 unit - adjusting old extent (we may have to split it) 257 * 1 unit - add new extent 258 * 1 unit - inode update 259 */ 260 trans = btrfs_start_transaction(root, 3); 261 if (IS_ERR(trans)) { 262 ret = PTR_ERR(trans); 263 trans = NULL; 264 goto out; 265 } 266 drop_args.path = path; 267 drop_args.start = drop_start; 268 drop_args.end = aligned_end; 269 drop_args.drop_cache = true; 270 ret = btrfs_drop_extents(trans, root, inode, &drop_args); 271 if (ret) { 272 btrfs_abort_transaction(trans, ret); 273 goto out; 274 } 275 ret = btrfs_insert_empty_item(trans, root, path, new_key, size); 276 if (ret) { 277 btrfs_abort_transaction(trans, ret); 278 goto out; 279 } 280 281 write_extent_buffer(path->nodes[0], inline_data, 282 btrfs_item_ptr_offset(path->nodes[0], 283 path->slots[0]), 284 size); 285 btrfs_update_inode_bytes(inode, datal, drop_args.bytes_found); 286 btrfs_set_inode_full_sync(inode); 287 ret = btrfs_inode_set_file_extent_range(inode, 0, aligned_end); 288 if (ret) 289 btrfs_abort_transaction(trans, ret); 290 out: 291 if (!ret && !trans) { 292 /* 293 * No transaction here means we copied the inline extent into a 294 * page of the destination inode. 295 * 296 * 1 unit to update inode item 297 */ 298 trans = btrfs_start_transaction(root, 1); 299 if (IS_ERR(trans)) { 300 ret = PTR_ERR(trans); 301 trans = NULL; 302 } 303 } 304 if (ret && trans) 305 btrfs_end_transaction(trans); 306 if (!ret) 307 *trans_out = trans; 308 309 return ret; 310 311 copy_to_page: 312 /* 313 * Release our path because we don't need it anymore and also because 314 * copy_inline_to_page() needs to reserve data and metadata, which may 315 * need to flush delalloc when we are low on available space and 316 * therefore cause a deadlock if writeback of an inline extent needs to 317 * write to the same leaf or an ordered extent completion needs to write 318 * to the same leaf. 319 */ 320 btrfs_release_path(path); 321 322 ret = copy_inline_to_page(inode, new_key->offset, 323 inline_data, size, datal, comp_type); 324 goto out; 325 } 326 327 /* 328 * Clone a range from inode file to another. 329 * 330 * @src: Inode to clone from 331 * @inode: Inode to clone to 332 * @off: Offset within source to start clone from 333 * @olen: Original length, passed by user, of range to clone 334 * @olen_aligned: Block-aligned value of olen 335 * @destoff: Offset within @inode to start clone 336 * @no_time_update: Whether to update mtime/ctime on the target inode 337 */ 338 static int btrfs_clone(struct inode *src, struct inode *inode, 339 const u64 off, const u64 olen, const u64 olen_aligned, 340 const u64 destoff, int no_time_update) 341 { 342 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode); 343 struct btrfs_path *path = NULL; 344 struct extent_buffer *leaf; 345 struct btrfs_trans_handle *trans; 346 char *buf = NULL; 347 struct btrfs_key key; 348 u32 nritems; 349 int slot; 350 int ret; 351 const u64 len = olen_aligned; 352 u64 last_dest_end = destoff; 353 u64 prev_extent_end = off; 354 355 ret = -ENOMEM; 356 buf = kvmalloc(fs_info->nodesize, GFP_KERNEL); 357 if (!buf) 358 return ret; 359 360 path = btrfs_alloc_path(); 361 if (!path) { 362 kvfree(buf); 363 return ret; 364 } 365 366 path->reada = READA_FORWARD; 367 /* Clone data */ 368 key.objectid = btrfs_ino(BTRFS_I(src)); 369 key.type = BTRFS_EXTENT_DATA_KEY; 370 key.offset = off; 371 372 while (1) { 373 struct btrfs_file_extent_item *extent; 374 u64 extent_gen; 375 int type; 376 u32 size; 377 struct btrfs_key new_key; 378 u64 disko = 0, diskl = 0; 379 u64 datao = 0, datal = 0; 380 u8 comp; 381 u64 drop_start; 382 383 /* Note the key will change type as we walk through the tree */ 384 ret = btrfs_search_slot(NULL, BTRFS_I(src)->root, &key, path, 385 0, 0); 386 if (ret < 0) 387 goto out; 388 /* 389 * First search, if no extent item that starts at offset off was 390 * found but the previous item is an extent item, it's possible 391 * it might overlap our target range, therefore process it. 392 */ 393 if (key.offset == off && ret > 0 && path->slots[0] > 0) { 394 btrfs_item_key_to_cpu(path->nodes[0], &key, 395 path->slots[0] - 1); 396 if (key.type == BTRFS_EXTENT_DATA_KEY) 397 path->slots[0]--; 398 } 399 400 nritems = btrfs_header_nritems(path->nodes[0]); 401 process_slot: 402 if (path->slots[0] >= nritems) { 403 ret = btrfs_next_leaf(BTRFS_I(src)->root, path); 404 if (ret < 0) 405 goto out; 406 if (ret > 0) 407 break; 408 nritems = btrfs_header_nritems(path->nodes[0]); 409 } 410 leaf = path->nodes[0]; 411 slot = path->slots[0]; 412 413 btrfs_item_key_to_cpu(leaf, &key, slot); 414 if (key.type > BTRFS_EXTENT_DATA_KEY || 415 key.objectid != btrfs_ino(BTRFS_I(src))) 416 break; 417 418 ASSERT(key.type == BTRFS_EXTENT_DATA_KEY); 419 420 extent = btrfs_item_ptr(leaf, slot, 421 struct btrfs_file_extent_item); 422 extent_gen = btrfs_file_extent_generation(leaf, extent); 423 comp = btrfs_file_extent_compression(leaf, extent); 424 type = btrfs_file_extent_type(leaf, extent); 425 if (type == BTRFS_FILE_EXTENT_REG || 426 type == BTRFS_FILE_EXTENT_PREALLOC) { 427 disko = btrfs_file_extent_disk_bytenr(leaf, extent); 428 diskl = btrfs_file_extent_disk_num_bytes(leaf, extent); 429 datao = btrfs_file_extent_offset(leaf, extent); 430 datal = btrfs_file_extent_num_bytes(leaf, extent); 431 } else if (type == BTRFS_FILE_EXTENT_INLINE) { 432 /* Take upper bound, may be compressed */ 433 datal = btrfs_file_extent_ram_bytes(leaf, extent); 434 } 435 436 /* 437 * The first search might have left us at an extent item that 438 * ends before our target range's start, can happen if we have 439 * holes and NO_HOLES feature enabled. 440 * 441 * Subsequent searches may leave us on a file range we have 442 * processed before - this happens due to a race with ordered 443 * extent completion for a file range that is outside our source 444 * range, but that range was part of a file extent item that 445 * also covered a leading part of our source range. 446 */ 447 if (key.offset + datal <= prev_extent_end) { 448 path->slots[0]++; 449 goto process_slot; 450 } else if (key.offset >= off + len) { 451 break; 452 } 453 454 prev_extent_end = key.offset + datal; 455 size = btrfs_item_size(leaf, slot); 456 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf, slot), 457 size); 458 459 btrfs_release_path(path); 460 461 memcpy(&new_key, &key, sizeof(new_key)); 462 new_key.objectid = btrfs_ino(BTRFS_I(inode)); 463 if (off <= key.offset) 464 new_key.offset = key.offset + destoff - off; 465 else 466 new_key.offset = destoff; 467 468 /* 469 * Deal with a hole that doesn't have an extent item that 470 * represents it (NO_HOLES feature enabled). 471 * This hole is either in the middle of the cloning range or at 472 * the beginning (fully overlaps it or partially overlaps it). 473 */ 474 if (new_key.offset != last_dest_end) 475 drop_start = last_dest_end; 476 else 477 drop_start = new_key.offset; 478 479 if (type == BTRFS_FILE_EXTENT_REG || 480 type == BTRFS_FILE_EXTENT_PREALLOC) { 481 struct btrfs_replace_extent_info clone_info; 482 483 /* 484 * a | --- range to clone ---| b 485 * | ------------- extent ------------- | 486 */ 487 488 /* Subtract range b */ 489 if (key.offset + datal > off + len) 490 datal = off + len - key.offset; 491 492 /* Subtract range a */ 493 if (off > key.offset) { 494 datao += off - key.offset; 495 datal -= off - key.offset; 496 } 497 498 clone_info.disk_offset = disko; 499 clone_info.disk_len = diskl; 500 clone_info.data_offset = datao; 501 clone_info.data_len = datal; 502 clone_info.file_offset = new_key.offset; 503 clone_info.extent_buf = buf; 504 clone_info.is_new_extent = false; 505 clone_info.update_times = !no_time_update; 506 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 507 drop_start, new_key.offset + datal - 1, 508 &clone_info, &trans); 509 if (ret) 510 goto out; 511 } else { 512 ASSERT(type == BTRFS_FILE_EXTENT_INLINE); 513 /* 514 * Inline extents always have to start at file offset 0 515 * and can never be bigger then the sector size. We can 516 * never clone only parts of an inline extent, since all 517 * reflink operations must start at a sector size aligned 518 * offset, and the length must be aligned too or end at 519 * the i_size (which implies the whole inlined data). 520 */ 521 ASSERT(key.offset == 0); 522 ASSERT(datal <= fs_info->sectorsize); 523 if (WARN_ON(type != BTRFS_FILE_EXTENT_INLINE) || 524 WARN_ON(key.offset != 0) || 525 WARN_ON(datal > fs_info->sectorsize)) { 526 ret = -EUCLEAN; 527 goto out; 528 } 529 530 ret = clone_copy_inline_extent(BTRFS_I(inode), path, &new_key, 531 drop_start, datal, size, 532 comp, buf, &trans); 533 if (ret) 534 goto out; 535 } 536 537 btrfs_release_path(path); 538 539 /* 540 * Whenever we share an extent we update the last_reflink_trans 541 * of each inode to the current transaction. This is needed to 542 * make sure fsync does not log multiple checksum items with 543 * overlapping ranges (because some extent items might refer 544 * only to sections of the original extent). For the destination 545 * inode we do this regardless of the generation of the extents 546 * or even if they are inline extents or explicit holes, to make 547 * sure a full fsync does not skip them. For the source inode, 548 * we only need to update last_reflink_trans in case it's a new 549 * extent that is not a hole or an inline extent, to deal with 550 * the checksums problem on fsync. 551 */ 552 if (extent_gen == trans->transid && disko > 0) 553 BTRFS_I(src)->last_reflink_trans = trans->transid; 554 555 BTRFS_I(inode)->last_reflink_trans = trans->transid; 556 557 last_dest_end = ALIGN(new_key.offset + datal, 558 fs_info->sectorsize); 559 ret = clone_finish_inode_update(trans, inode, last_dest_end, 560 destoff, olen, no_time_update); 561 if (ret) 562 goto out; 563 if (new_key.offset + datal >= destoff + len) 564 break; 565 566 btrfs_release_path(path); 567 key.offset = prev_extent_end; 568 569 if (fatal_signal_pending(current)) { 570 ret = -EINTR; 571 goto out; 572 } 573 574 cond_resched(); 575 } 576 ret = 0; 577 578 if (last_dest_end < destoff + len) { 579 /* 580 * We have an implicit hole that fully or partially overlaps our 581 * cloning range at its end. This means that we either have the 582 * NO_HOLES feature enabled or the implicit hole happened due to 583 * mixing buffered and direct IO writes against this file. 584 */ 585 btrfs_release_path(path); 586 587 /* 588 * When using NO_HOLES and we are cloning a range that covers 589 * only a hole (no extents) into a range beyond the current 590 * i_size, punching a hole in the target range will not create 591 * an extent map defining a hole, because the range starts at or 592 * beyond current i_size. If the file previously had an i_size 593 * greater than the new i_size set by this clone operation, we 594 * need to make sure the next fsync is a full fsync, so that it 595 * detects and logs a hole covering a range from the current 596 * i_size to the new i_size. If the clone range covers extents, 597 * besides a hole, then we know the full sync flag was already 598 * set by previous calls to btrfs_replace_file_extents() that 599 * replaced file extent items. 600 */ 601 if (last_dest_end >= i_size_read(inode)) 602 btrfs_set_inode_full_sync(BTRFS_I(inode)); 603 604 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 605 last_dest_end, destoff + len - 1, NULL, &trans); 606 if (ret) 607 goto out; 608 609 ret = clone_finish_inode_update(trans, inode, destoff + len, 610 destoff, olen, no_time_update); 611 } 612 613 out: 614 btrfs_free_path(path); 615 kvfree(buf); 616 clear_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &BTRFS_I(inode)->runtime_flags); 617 618 return ret; 619 } 620 621 static void btrfs_double_mmap_lock(struct btrfs_inode *inode1, struct btrfs_inode *inode2) 622 { 623 if (inode1 < inode2) 624 swap(inode1, inode2); 625 down_write(&inode1->i_mmap_lock); 626 down_write_nested(&inode2->i_mmap_lock, SINGLE_DEPTH_NESTING); 627 } 628 629 static void btrfs_double_mmap_unlock(struct btrfs_inode *inode1, struct btrfs_inode *inode2) 630 { 631 up_write(&inode1->i_mmap_lock); 632 up_write(&inode2->i_mmap_lock); 633 } 634 635 static int btrfs_extent_same_range(struct btrfs_inode *src, u64 loff, u64 len, 636 struct btrfs_inode *dst, u64 dst_loff) 637 { 638 const u64 end = dst_loff + len - 1; 639 struct extent_state *cached_state = NULL; 640 struct btrfs_fs_info *fs_info = src->root->fs_info; 641 const u64 bs = fs_info->sectorsize; 642 int ret; 643 644 /* 645 * Lock destination range to serialize with concurrent readahead(), and 646 * we are safe from concurrency with relocation of source extents 647 * because we have already locked the inode's i_mmap_lock in exclusive 648 * mode. 649 */ 650 btrfs_lock_extent(&dst->io_tree, dst_loff, end, &cached_state); 651 ret = btrfs_clone(&src->vfs_inode, &dst->vfs_inode, loff, len, 652 ALIGN(len, bs), dst_loff, 1); 653 btrfs_unlock_extent(&dst->io_tree, dst_loff, end, &cached_state); 654 655 btrfs_btree_balance_dirty(fs_info); 656 657 return ret; 658 } 659 660 static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen, 661 struct inode *dst, u64 dst_loff) 662 { 663 int ret = 0; 664 u64 i, tail_len, chunk_count; 665 struct btrfs_root *root_dst = BTRFS_I(dst)->root; 666 667 spin_lock(&root_dst->root_item_lock); 668 if (root_dst->send_in_progress) { 669 btrfs_warn_rl(root_dst->fs_info, 670 "cannot deduplicate to root %llu while send operations are using it (%d in progress)", 671 btrfs_root_id(root_dst), 672 root_dst->send_in_progress); 673 spin_unlock(&root_dst->root_item_lock); 674 return -EAGAIN; 675 } 676 root_dst->dedupe_in_progress++; 677 spin_unlock(&root_dst->root_item_lock); 678 679 tail_len = olen % BTRFS_MAX_DEDUPE_LEN; 680 chunk_count = div_u64(olen, BTRFS_MAX_DEDUPE_LEN); 681 682 for (i = 0; i < chunk_count; i++) { 683 ret = btrfs_extent_same_range(BTRFS_I(src), loff, BTRFS_MAX_DEDUPE_LEN, 684 BTRFS_I(dst), dst_loff); 685 if (ret) 686 goto out; 687 688 loff += BTRFS_MAX_DEDUPE_LEN; 689 dst_loff += BTRFS_MAX_DEDUPE_LEN; 690 } 691 692 if (tail_len > 0) 693 ret = btrfs_extent_same_range(BTRFS_I(src), loff, tail_len, 694 BTRFS_I(dst), dst_loff); 695 out: 696 spin_lock(&root_dst->root_item_lock); 697 root_dst->dedupe_in_progress--; 698 spin_unlock(&root_dst->root_item_lock); 699 700 return ret; 701 } 702 703 static noinline int btrfs_clone_files(struct file *file, struct file *file_src, 704 u64 off, u64 olen, u64 destoff) 705 { 706 struct extent_state *cached_state = NULL; 707 struct inode *inode = file_inode(file); 708 struct inode *src = file_inode(file_src); 709 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode); 710 int ret; 711 int wb_ret; 712 u64 len = olen; 713 u64 bs = fs_info->sectorsize; 714 u64 end; 715 716 /* 717 * VFS's generic_remap_file_range_prep() protects us from cloning the 718 * eof block into the middle of a file, which would result in corruption 719 * if the file size is not blocksize aligned. So we don't need to check 720 * for that case here. 721 */ 722 if (off + len == src->i_size) 723 len = ALIGN(src->i_size, bs) - off; 724 725 if (destoff > inode->i_size) { 726 const u64 wb_start = ALIGN_DOWN(inode->i_size, bs); 727 728 ret = btrfs_cont_expand(BTRFS_I(inode), inode->i_size, destoff); 729 if (ret) 730 return ret; 731 /* 732 * We may have truncated the last block if the inode's size is 733 * not sector size aligned, so we need to wait for writeback to 734 * complete before proceeding further, otherwise we can race 735 * with cloning and attempt to increment a reference to an 736 * extent that no longer exists (writeback completed right after 737 * we found the previous extent covering eof and before we 738 * attempted to increment its reference count). 739 */ 740 ret = btrfs_wait_ordered_range(BTRFS_I(inode), wb_start, 741 destoff - wb_start); 742 if (ret) 743 return ret; 744 } 745 746 /* 747 * Lock destination range to serialize with concurrent readahead(), and 748 * we are safe from concurrency with relocation of source extents 749 * because we have already locked the inode's i_mmap_lock in exclusive 750 * mode. 751 */ 752 end = destoff + len - 1; 753 btrfs_lock_extent(&BTRFS_I(inode)->io_tree, destoff, end, &cached_state); 754 ret = btrfs_clone(src, inode, off, olen, len, destoff, 0); 755 btrfs_unlock_extent(&BTRFS_I(inode)->io_tree, destoff, end, &cached_state); 756 757 /* 758 * We may have copied an inline extent into a page of the destination 759 * range, so wait for writeback to complete before truncating pages 760 * from the page cache. This is a rare case. 761 */ 762 wb_ret = btrfs_wait_ordered_range(BTRFS_I(inode), destoff, len); 763 ret = ret ? ret : wb_ret; 764 /* 765 * Truncate page cache pages so that future reads will see the cloned 766 * data immediately and not the previous data. 767 */ 768 truncate_inode_pages_range(&inode->i_data, 769 round_down(destoff, PAGE_SIZE), 770 round_up(destoff + len, PAGE_SIZE) - 1); 771 772 btrfs_btree_balance_dirty(fs_info); 773 774 return ret; 775 } 776 777 static int btrfs_remap_file_range_prep(struct file *file_in, loff_t pos_in, 778 struct file *file_out, loff_t pos_out, 779 loff_t *len, unsigned int remap_flags) 780 { 781 struct btrfs_inode *inode_in = BTRFS_I(file_inode(file_in)); 782 struct btrfs_inode *inode_out = BTRFS_I(file_inode(file_out)); 783 u64 bs = inode_out->root->fs_info->sectorsize; 784 u64 wb_len; 785 int ret; 786 787 if (!(remap_flags & REMAP_FILE_DEDUP)) { 788 struct btrfs_root *root_out = inode_out->root; 789 790 if (btrfs_root_readonly(root_out)) 791 return -EROFS; 792 793 ASSERT(inode_in->vfs_inode.i_sb == inode_out->vfs_inode.i_sb); 794 } 795 796 /* Don't make the dst file partly checksummed */ 797 if ((inode_in->flags & BTRFS_INODE_NODATASUM) != 798 (inode_out->flags & BTRFS_INODE_NODATASUM)) { 799 return -EINVAL; 800 } 801 802 /* 803 * Now that the inodes are locked, we need to start writeback ourselves 804 * and can not rely on the writeback from the VFS's generic helper 805 * generic_remap_file_range_prep() because: 806 * 807 * 1) For compression we must call filemap_fdatawrite_range() range 808 * twice (btrfs_fdatawrite_range() does it for us), and the generic 809 * helper only calls it once; 810 * 811 * 2) filemap_fdatawrite_range(), called by the generic helper only 812 * waits for the writeback to complete, i.e. for IO to be done, and 813 * not for the ordered extents to complete. We need to wait for them 814 * to complete so that new file extent items are in the fs tree. 815 */ 816 if (*len == 0 && !(remap_flags & REMAP_FILE_DEDUP)) 817 wb_len = ALIGN(inode_in->vfs_inode.i_size, bs) - ALIGN_DOWN(pos_in, bs); 818 else 819 wb_len = ALIGN(*len, bs); 820 821 /* 822 * Workaround to make sure NOCOW buffered write reach disk as NOCOW. 823 * 824 * Btrfs' back references do not have a block level granularity, they 825 * work at the whole extent level. 826 * NOCOW buffered write without data space reserved may not be able 827 * to fall back to CoW due to lack of data space, thus could cause 828 * data loss. 829 * 830 * Here we take a shortcut by flushing the whole inode, so that all 831 * nocow write should reach disk as nocow before we increase the 832 * reference of the extent. We could do better by only flushing NOCOW 833 * data, but that needs extra accounting. 834 * 835 * Also we don't need to check ASYNC_EXTENT, as async extent will be 836 * CoWed anyway, not affecting nocow part. 837 */ 838 ret = filemap_flush(inode_in->vfs_inode.i_mapping); 839 if (ret < 0) 840 return ret; 841 842 ret = btrfs_wait_ordered_range(inode_in, ALIGN_DOWN(pos_in, bs), wb_len); 843 if (ret < 0) 844 return ret; 845 ret = btrfs_wait_ordered_range(inode_out, ALIGN_DOWN(pos_out, bs), wb_len); 846 if (ret < 0) 847 return ret; 848 849 return generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out, 850 len, remap_flags); 851 } 852 853 static bool file_sync_write(const struct file *file) 854 { 855 if (file->f_flags & (__O_SYNC | O_DSYNC)) 856 return true; 857 if (IS_SYNC(file_inode(file))) 858 return true; 859 860 return false; 861 } 862 863 loff_t btrfs_remap_file_range(struct file *src_file, loff_t off, 864 struct file *dst_file, loff_t destoff, loff_t len, 865 unsigned int remap_flags) 866 { 867 struct btrfs_inode *src_inode = BTRFS_I(file_inode(src_file)); 868 struct btrfs_inode *dst_inode = BTRFS_I(file_inode(dst_file)); 869 bool same_inode = dst_inode == src_inode; 870 int ret; 871 872 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY)) 873 return -EINVAL; 874 875 if (same_inode) { 876 btrfs_inode_lock(src_inode, BTRFS_ILOCK_MMAP); 877 } else { 878 lock_two_nondirectories(&src_inode->vfs_inode, &dst_inode->vfs_inode); 879 btrfs_double_mmap_lock(src_inode, dst_inode); 880 } 881 882 ret = btrfs_remap_file_range_prep(src_file, off, dst_file, destoff, 883 &len, remap_flags); 884 if (ret < 0 || len == 0) 885 goto out_unlock; 886 887 if (remap_flags & REMAP_FILE_DEDUP) 888 ret = btrfs_extent_same(&src_inode->vfs_inode, off, len, 889 &dst_inode->vfs_inode, destoff); 890 else 891 ret = btrfs_clone_files(dst_file, src_file, off, len, destoff); 892 893 out_unlock: 894 if (same_inode) { 895 btrfs_inode_unlock(src_inode, BTRFS_ILOCK_MMAP); 896 } else { 897 btrfs_double_mmap_unlock(src_inode, dst_inode); 898 unlock_two_nondirectories(&src_inode->vfs_inode, 899 &dst_inode->vfs_inode); 900 } 901 902 /* 903 * If either the source or the destination file was opened with O_SYNC, 904 * O_DSYNC or has the S_SYNC attribute, fsync both the destination and 905 * source files/ranges, so that after a successful return (0) followed 906 * by a power failure results in the reflinked data to be readable from 907 * both files/ranges. 908 */ 909 if (ret == 0 && len > 0 && 910 (file_sync_write(src_file) || file_sync_write(dst_file))) { 911 ret = btrfs_sync_file(src_file, off, off + len - 1, 0); 912 if (ret == 0) 913 ret = btrfs_sync_file(dst_file, destoff, 914 destoff + len - 1, 0); 915 } 916 917 return ret < 0 ? ret : len; 918 } 919