1 // SPDX-License-Identifier: GPL-2.0 2 3 #include <linux/blkdev.h> 4 #include <linux/iversion.h> 5 #include "compression.h" 6 #include "ctree.h" 7 #include "delalloc-space.h" 8 #include "reflink.h" 9 #include "transaction.h" 10 #include "subpage.h" 11 12 #define BTRFS_MAX_DEDUPE_LEN SZ_16M 13 14 static int clone_finish_inode_update(struct btrfs_trans_handle *trans, 15 struct inode *inode, 16 u64 endoff, 17 const u64 destoff, 18 const u64 olen, 19 int no_time_update) 20 { 21 struct btrfs_root *root = BTRFS_I(inode)->root; 22 int ret; 23 24 inode_inc_iversion(inode); 25 if (!no_time_update) 26 inode->i_mtime = inode->i_ctime = current_time(inode); 27 /* 28 * We round up to the block size at eof when determining which 29 * extents to clone above, but shouldn't round up the file size. 30 */ 31 if (endoff > destoff + olen) 32 endoff = destoff + olen; 33 if (endoff > inode->i_size) { 34 i_size_write(inode, endoff); 35 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0); 36 } 37 38 ret = btrfs_update_inode(trans, root, BTRFS_I(inode)); 39 if (ret) { 40 btrfs_abort_transaction(trans, ret); 41 btrfs_end_transaction(trans); 42 goto out; 43 } 44 ret = btrfs_end_transaction(trans); 45 out: 46 return ret; 47 } 48 49 static int copy_inline_to_page(struct btrfs_inode *inode, 50 const u64 file_offset, 51 char *inline_data, 52 const u64 size, 53 const u64 datal, 54 const u8 comp_type) 55 { 56 struct btrfs_fs_info *fs_info = inode->root->fs_info; 57 const u32 block_size = fs_info->sectorsize; 58 const u64 range_end = file_offset + block_size - 1; 59 const size_t inline_size = size - btrfs_file_extent_calc_inline_size(0); 60 char *data_start = inline_data + btrfs_file_extent_calc_inline_size(0); 61 struct extent_changeset *data_reserved = NULL; 62 struct page *page = NULL; 63 struct address_space *mapping = inode->vfs_inode.i_mapping; 64 int ret; 65 66 ASSERT(IS_ALIGNED(file_offset, block_size)); 67 68 /* 69 * We have flushed and locked the ranges of the source and destination 70 * inodes, we also have locked the inodes, so we are safe to do a 71 * reservation here. Also we must not do the reservation while holding 72 * a transaction open, otherwise we would deadlock. 73 */ 74 ret = btrfs_delalloc_reserve_space(inode, &data_reserved, file_offset, 75 block_size); 76 if (ret) 77 goto out; 78 79 page = find_or_create_page(mapping, file_offset >> PAGE_SHIFT, 80 btrfs_alloc_write_mask(mapping)); 81 if (!page) { 82 ret = -ENOMEM; 83 goto out_unlock; 84 } 85 86 ret = set_page_extent_mapped(page); 87 if (ret < 0) 88 goto out_unlock; 89 90 clear_extent_bit(&inode->io_tree, file_offset, range_end, 91 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 92 0, 0, NULL); 93 ret = btrfs_set_extent_delalloc(inode, file_offset, range_end, 0, NULL); 94 if (ret) 95 goto out_unlock; 96 97 /* 98 * After dirtying the page our caller will need to start a transaction, 99 * and if we are low on metadata free space, that can cause flushing of 100 * delalloc for all inodes in order to get metadata space released. 101 * However we are holding the range locked for the whole duration of 102 * the clone/dedupe operation, so we may deadlock if that happens and no 103 * other task releases enough space. So mark this inode as not being 104 * possible to flush to avoid such deadlock. We will clear that flag 105 * when we finish cloning all extents, since a transaction is started 106 * after finding each extent to clone. 107 */ 108 set_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &inode->runtime_flags); 109 110 if (comp_type == BTRFS_COMPRESS_NONE) { 111 memcpy_to_page(page, offset_in_page(file_offset), data_start, 112 datal); 113 flush_dcache_page(page); 114 } else { 115 ret = btrfs_decompress(comp_type, data_start, page, 116 offset_in_page(file_offset), 117 inline_size, datal); 118 if (ret) 119 goto out_unlock; 120 flush_dcache_page(page); 121 } 122 123 /* 124 * If our inline data is smaller then the block/page size, then the 125 * remaining of the block/page is equivalent to zeroes. We had something 126 * like the following done: 127 * 128 * $ xfs_io -f -c "pwrite -S 0xab 0 500" file 129 * $ sync # (or fsync) 130 * $ xfs_io -c "falloc 0 4K" file 131 * $ xfs_io -c "pwrite -S 0xcd 4K 4K" 132 * 133 * So what's in the range [500, 4095] corresponds to zeroes. 134 */ 135 if (datal < block_size) { 136 memzero_page(page, datal, block_size - datal); 137 flush_dcache_page(page); 138 } 139 140 btrfs_page_set_uptodate(fs_info, page, file_offset, block_size); 141 btrfs_page_clear_checked(fs_info, page, file_offset, block_size); 142 btrfs_page_set_dirty(fs_info, page, file_offset, block_size); 143 out_unlock: 144 if (page) { 145 unlock_page(page); 146 put_page(page); 147 } 148 if (ret) 149 btrfs_delalloc_release_space(inode, data_reserved, file_offset, 150 block_size, true); 151 btrfs_delalloc_release_extents(inode, block_size); 152 out: 153 extent_changeset_free(data_reserved); 154 155 return ret; 156 } 157 158 /* 159 * Deal with cloning of inline extents. We try to copy the inline extent from 160 * the source inode to destination inode when possible. When not possible we 161 * copy the inline extent's data into the respective page of the inode. 162 */ 163 static int clone_copy_inline_extent(struct inode *dst, 164 struct btrfs_path *path, 165 struct btrfs_key *new_key, 166 const u64 drop_start, 167 const u64 datal, 168 const u64 size, 169 const u8 comp_type, 170 char *inline_data, 171 struct btrfs_trans_handle **trans_out) 172 { 173 struct btrfs_fs_info *fs_info = btrfs_sb(dst->i_sb); 174 struct btrfs_root *root = BTRFS_I(dst)->root; 175 const u64 aligned_end = ALIGN(new_key->offset + datal, 176 fs_info->sectorsize); 177 struct btrfs_trans_handle *trans = NULL; 178 struct btrfs_drop_extents_args drop_args = { 0 }; 179 int ret; 180 struct btrfs_key key; 181 182 if (new_key->offset > 0) { 183 ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset, 184 inline_data, size, datal, comp_type); 185 goto out; 186 } 187 188 key.objectid = btrfs_ino(BTRFS_I(dst)); 189 key.type = BTRFS_EXTENT_DATA_KEY; 190 key.offset = 0; 191 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 192 if (ret < 0) { 193 return ret; 194 } else if (ret > 0) { 195 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) { 196 ret = btrfs_next_leaf(root, path); 197 if (ret < 0) 198 return ret; 199 else if (ret > 0) 200 goto copy_inline_extent; 201 } 202 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 203 if (key.objectid == btrfs_ino(BTRFS_I(dst)) && 204 key.type == BTRFS_EXTENT_DATA_KEY) { 205 /* 206 * There's an implicit hole at file offset 0, copy the 207 * inline extent's data to the page. 208 */ 209 ASSERT(key.offset > 0); 210 goto copy_to_page; 211 } 212 } else if (i_size_read(dst) <= datal) { 213 struct btrfs_file_extent_item *ei; 214 215 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 216 struct btrfs_file_extent_item); 217 /* 218 * If it's an inline extent replace it with the source inline 219 * extent, otherwise copy the source inline extent data into 220 * the respective page at the destination inode. 221 */ 222 if (btrfs_file_extent_type(path->nodes[0], ei) == 223 BTRFS_FILE_EXTENT_INLINE) 224 goto copy_inline_extent; 225 226 goto copy_to_page; 227 } 228 229 copy_inline_extent: 230 /* 231 * We have no extent items, or we have an extent at offset 0 which may 232 * or may not be inlined. All these cases are dealt the same way. 233 */ 234 if (i_size_read(dst) > datal) { 235 /* 236 * At the destination offset 0 we have either a hole, a regular 237 * extent or an inline extent larger then the one we want to 238 * clone. Deal with all these cases by copying the inline extent 239 * data into the respective page at the destination inode. 240 */ 241 goto copy_to_page; 242 } 243 244 /* 245 * Release path before starting a new transaction so we don't hold locks 246 * that would confuse lockdep. 247 */ 248 btrfs_release_path(path); 249 /* 250 * If we end up here it means were copy the inline extent into a leaf 251 * of the destination inode. We know we will drop or adjust at most one 252 * extent item in the destination root. 253 * 254 * 1 unit - adjusting old extent (we may have to split it) 255 * 1 unit - add new extent 256 * 1 unit - inode update 257 */ 258 trans = btrfs_start_transaction(root, 3); 259 if (IS_ERR(trans)) { 260 ret = PTR_ERR(trans); 261 trans = NULL; 262 goto out; 263 } 264 drop_args.path = path; 265 drop_args.start = drop_start; 266 drop_args.end = aligned_end; 267 drop_args.drop_cache = true; 268 ret = btrfs_drop_extents(trans, root, BTRFS_I(dst), &drop_args); 269 if (ret) 270 goto out; 271 ret = btrfs_insert_empty_item(trans, root, path, new_key, size); 272 if (ret) 273 goto out; 274 275 write_extent_buffer(path->nodes[0], inline_data, 276 btrfs_item_ptr_offset(path->nodes[0], 277 path->slots[0]), 278 size); 279 btrfs_update_inode_bytes(BTRFS_I(dst), datal, drop_args.bytes_found); 280 btrfs_set_inode_full_sync(BTRFS_I(dst)); 281 ret = btrfs_inode_set_file_extent_range(BTRFS_I(dst), 0, aligned_end); 282 out: 283 if (!ret && !trans) { 284 /* 285 * No transaction here means we copied the inline extent into a 286 * page of the destination inode. 287 * 288 * 1 unit to update inode item 289 */ 290 trans = btrfs_start_transaction(root, 1); 291 if (IS_ERR(trans)) { 292 ret = PTR_ERR(trans); 293 trans = NULL; 294 } 295 } 296 if (ret && trans) { 297 btrfs_abort_transaction(trans, ret); 298 btrfs_end_transaction(trans); 299 } 300 if (!ret) 301 *trans_out = trans; 302 303 return ret; 304 305 copy_to_page: 306 /* 307 * Release our path because we don't need it anymore and also because 308 * copy_inline_to_page() needs to reserve data and metadata, which may 309 * need to flush delalloc when we are low on available space and 310 * therefore cause a deadlock if writeback of an inline extent needs to 311 * write to the same leaf or an ordered extent completion needs to write 312 * to the same leaf. 313 */ 314 btrfs_release_path(path); 315 316 ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset, 317 inline_data, size, datal, comp_type); 318 goto out; 319 } 320 321 /** 322 * btrfs_clone() - clone a range from inode file to another 323 * 324 * @src: Inode to clone from 325 * @inode: Inode to clone to 326 * @off: Offset within source to start clone from 327 * @olen: Original length, passed by user, of range to clone 328 * @olen_aligned: Block-aligned value of olen 329 * @destoff: Offset within @inode to start clone 330 * @no_time_update: Whether to update mtime/ctime on the target inode 331 */ 332 static int btrfs_clone(struct inode *src, struct inode *inode, 333 const u64 off, const u64 olen, const u64 olen_aligned, 334 const u64 destoff, int no_time_update) 335 { 336 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 337 struct btrfs_path *path = NULL; 338 struct extent_buffer *leaf; 339 struct btrfs_trans_handle *trans; 340 char *buf = NULL; 341 struct btrfs_key key; 342 u32 nritems; 343 int slot; 344 int ret; 345 const u64 len = olen_aligned; 346 u64 last_dest_end = destoff; 347 348 ret = -ENOMEM; 349 buf = kvmalloc(fs_info->nodesize, GFP_KERNEL); 350 if (!buf) 351 return ret; 352 353 path = btrfs_alloc_path(); 354 if (!path) { 355 kvfree(buf); 356 return ret; 357 } 358 359 path->reada = READA_FORWARD; 360 /* Clone data */ 361 key.objectid = btrfs_ino(BTRFS_I(src)); 362 key.type = BTRFS_EXTENT_DATA_KEY; 363 key.offset = off; 364 365 while (1) { 366 u64 next_key_min_offset = key.offset + 1; 367 struct btrfs_file_extent_item *extent; 368 u64 extent_gen; 369 int type; 370 u32 size; 371 struct btrfs_key new_key; 372 u64 disko = 0, diskl = 0; 373 u64 datao = 0, datal = 0; 374 u8 comp; 375 u64 drop_start; 376 377 /* Note the key will change type as we walk through the tree */ 378 ret = btrfs_search_slot(NULL, BTRFS_I(src)->root, &key, path, 379 0, 0); 380 if (ret < 0) 381 goto out; 382 /* 383 * First search, if no extent item that starts at offset off was 384 * found but the previous item is an extent item, it's possible 385 * it might overlap our target range, therefore process it. 386 */ 387 if (key.offset == off && ret > 0 && path->slots[0] > 0) { 388 btrfs_item_key_to_cpu(path->nodes[0], &key, 389 path->slots[0] - 1); 390 if (key.type == BTRFS_EXTENT_DATA_KEY) 391 path->slots[0]--; 392 } 393 394 nritems = btrfs_header_nritems(path->nodes[0]); 395 process_slot: 396 if (path->slots[0] >= nritems) { 397 ret = btrfs_next_leaf(BTRFS_I(src)->root, path); 398 if (ret < 0) 399 goto out; 400 if (ret > 0) 401 break; 402 nritems = btrfs_header_nritems(path->nodes[0]); 403 } 404 leaf = path->nodes[0]; 405 slot = path->slots[0]; 406 407 btrfs_item_key_to_cpu(leaf, &key, slot); 408 if (key.type > BTRFS_EXTENT_DATA_KEY || 409 key.objectid != btrfs_ino(BTRFS_I(src))) 410 break; 411 412 ASSERT(key.type == BTRFS_EXTENT_DATA_KEY); 413 414 extent = btrfs_item_ptr(leaf, slot, 415 struct btrfs_file_extent_item); 416 extent_gen = btrfs_file_extent_generation(leaf, extent); 417 comp = btrfs_file_extent_compression(leaf, extent); 418 type = btrfs_file_extent_type(leaf, extent); 419 if (type == BTRFS_FILE_EXTENT_REG || 420 type == BTRFS_FILE_EXTENT_PREALLOC) { 421 disko = btrfs_file_extent_disk_bytenr(leaf, extent); 422 diskl = btrfs_file_extent_disk_num_bytes(leaf, extent); 423 datao = btrfs_file_extent_offset(leaf, extent); 424 datal = btrfs_file_extent_num_bytes(leaf, extent); 425 } else if (type == BTRFS_FILE_EXTENT_INLINE) { 426 /* Take upper bound, may be compressed */ 427 datal = btrfs_file_extent_ram_bytes(leaf, extent); 428 } 429 430 /* 431 * The first search might have left us at an extent item that 432 * ends before our target range's start, can happen if we have 433 * holes and NO_HOLES feature enabled. 434 */ 435 if (key.offset + datal <= off) { 436 path->slots[0]++; 437 goto process_slot; 438 } else if (key.offset >= off + len) { 439 break; 440 } 441 next_key_min_offset = key.offset + datal; 442 size = btrfs_item_size(leaf, slot); 443 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf, slot), 444 size); 445 446 btrfs_release_path(path); 447 448 memcpy(&new_key, &key, sizeof(new_key)); 449 new_key.objectid = btrfs_ino(BTRFS_I(inode)); 450 if (off <= key.offset) 451 new_key.offset = key.offset + destoff - off; 452 else 453 new_key.offset = destoff; 454 455 /* 456 * Deal with a hole that doesn't have an extent item that 457 * represents it (NO_HOLES feature enabled). 458 * This hole is either in the middle of the cloning range or at 459 * the beginning (fully overlaps it or partially overlaps it). 460 */ 461 if (new_key.offset != last_dest_end) 462 drop_start = last_dest_end; 463 else 464 drop_start = new_key.offset; 465 466 if (type == BTRFS_FILE_EXTENT_REG || 467 type == BTRFS_FILE_EXTENT_PREALLOC) { 468 struct btrfs_replace_extent_info clone_info; 469 470 /* 471 * a | --- range to clone ---| b 472 * | ------------- extent ------------- | 473 */ 474 475 /* Subtract range b */ 476 if (key.offset + datal > off + len) 477 datal = off + len - key.offset; 478 479 /* Subtract range a */ 480 if (off > key.offset) { 481 datao += off - key.offset; 482 datal -= off - key.offset; 483 } 484 485 clone_info.disk_offset = disko; 486 clone_info.disk_len = diskl; 487 clone_info.data_offset = datao; 488 clone_info.data_len = datal; 489 clone_info.file_offset = new_key.offset; 490 clone_info.extent_buf = buf; 491 clone_info.is_new_extent = false; 492 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 493 drop_start, new_key.offset + datal - 1, 494 &clone_info, &trans); 495 if (ret) 496 goto out; 497 } else { 498 ASSERT(type == BTRFS_FILE_EXTENT_INLINE); 499 /* 500 * Inline extents always have to start at file offset 0 501 * and can never be bigger then the sector size. We can 502 * never clone only parts of an inline extent, since all 503 * reflink operations must start at a sector size aligned 504 * offset, and the length must be aligned too or end at 505 * the i_size (which implies the whole inlined data). 506 */ 507 ASSERT(key.offset == 0); 508 ASSERT(datal <= fs_info->sectorsize); 509 if (WARN_ON(type != BTRFS_FILE_EXTENT_INLINE) || 510 WARN_ON(key.offset != 0) || 511 WARN_ON(datal > fs_info->sectorsize)) { 512 ret = -EUCLEAN; 513 goto out; 514 } 515 516 ret = clone_copy_inline_extent(inode, path, &new_key, 517 drop_start, datal, size, 518 comp, buf, &trans); 519 if (ret) 520 goto out; 521 } 522 523 btrfs_release_path(path); 524 525 /* 526 * Whenever we share an extent we update the last_reflink_trans 527 * of each inode to the current transaction. This is needed to 528 * make sure fsync does not log multiple checksum items with 529 * overlapping ranges (because some extent items might refer 530 * only to sections of the original extent). For the destination 531 * inode we do this regardless of the generation of the extents 532 * or even if they are inline extents or explicit holes, to make 533 * sure a full fsync does not skip them. For the source inode, 534 * we only need to update last_reflink_trans in case it's a new 535 * extent that is not a hole or an inline extent, to deal with 536 * the checksums problem on fsync. 537 */ 538 if (extent_gen == trans->transid && disko > 0) 539 BTRFS_I(src)->last_reflink_trans = trans->transid; 540 541 BTRFS_I(inode)->last_reflink_trans = trans->transid; 542 543 last_dest_end = ALIGN(new_key.offset + datal, 544 fs_info->sectorsize); 545 ret = clone_finish_inode_update(trans, inode, last_dest_end, 546 destoff, olen, no_time_update); 547 if (ret) 548 goto out; 549 if (new_key.offset + datal >= destoff + len) 550 break; 551 552 btrfs_release_path(path); 553 key.offset = next_key_min_offset; 554 555 if (fatal_signal_pending(current)) { 556 ret = -EINTR; 557 goto out; 558 } 559 560 cond_resched(); 561 } 562 ret = 0; 563 564 if (last_dest_end < destoff + len) { 565 /* 566 * We have an implicit hole that fully or partially overlaps our 567 * cloning range at its end. This means that we either have the 568 * NO_HOLES feature enabled or the implicit hole happened due to 569 * mixing buffered and direct IO writes against this file. 570 */ 571 btrfs_release_path(path); 572 573 /* 574 * When using NO_HOLES and we are cloning a range that covers 575 * only a hole (no extents) into a range beyond the current 576 * i_size, punching a hole in the target range will not create 577 * an extent map defining a hole, because the range starts at or 578 * beyond current i_size. If the file previously had an i_size 579 * greater than the new i_size set by this clone operation, we 580 * need to make sure the next fsync is a full fsync, so that it 581 * detects and logs a hole covering a range from the current 582 * i_size to the new i_size. If the clone range covers extents, 583 * besides a hole, then we know the full sync flag was already 584 * set by previous calls to btrfs_replace_file_extents() that 585 * replaced file extent items. 586 */ 587 if (last_dest_end >= i_size_read(inode)) 588 btrfs_set_inode_full_sync(BTRFS_I(inode)); 589 590 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 591 last_dest_end, destoff + len - 1, NULL, &trans); 592 if (ret) 593 goto out; 594 595 ret = clone_finish_inode_update(trans, inode, destoff + len, 596 destoff, olen, no_time_update); 597 } 598 599 out: 600 btrfs_free_path(path); 601 kvfree(buf); 602 clear_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &BTRFS_I(inode)->runtime_flags); 603 604 return ret; 605 } 606 607 static void btrfs_double_extent_unlock(struct inode *inode1, u64 loff1, 608 struct inode *inode2, u64 loff2, u64 len) 609 { 610 unlock_extent(&BTRFS_I(inode1)->io_tree, loff1, loff1 + len - 1); 611 unlock_extent(&BTRFS_I(inode2)->io_tree, loff2, loff2 + len - 1); 612 } 613 614 static void btrfs_double_extent_lock(struct inode *inode1, u64 loff1, 615 struct inode *inode2, u64 loff2, u64 len) 616 { 617 u64 range1_end = loff1 + len - 1; 618 u64 range2_end = loff2 + len - 1; 619 620 if (inode1 < inode2) { 621 swap(inode1, inode2); 622 swap(loff1, loff2); 623 swap(range1_end, range2_end); 624 } else if (inode1 == inode2 && loff2 < loff1) { 625 swap(loff1, loff2); 626 swap(range1_end, range2_end); 627 } 628 629 lock_extent(&BTRFS_I(inode1)->io_tree, loff1, range1_end); 630 lock_extent(&BTRFS_I(inode2)->io_tree, loff2, range2_end); 631 632 btrfs_assert_inode_range_clean(BTRFS_I(inode1), loff1, range1_end); 633 btrfs_assert_inode_range_clean(BTRFS_I(inode2), loff2, range2_end); 634 } 635 636 static void btrfs_double_mmap_lock(struct inode *inode1, struct inode *inode2) 637 { 638 if (inode1 < inode2) 639 swap(inode1, inode2); 640 down_write(&BTRFS_I(inode1)->i_mmap_lock); 641 down_write_nested(&BTRFS_I(inode2)->i_mmap_lock, SINGLE_DEPTH_NESTING); 642 } 643 644 static void btrfs_double_mmap_unlock(struct inode *inode1, struct inode *inode2) 645 { 646 up_write(&BTRFS_I(inode1)->i_mmap_lock); 647 up_write(&BTRFS_I(inode2)->i_mmap_lock); 648 } 649 650 static int btrfs_extent_same_range(struct inode *src, u64 loff, u64 len, 651 struct inode *dst, u64 dst_loff) 652 { 653 const u64 bs = BTRFS_I(src)->root->fs_info->sb->s_blocksize; 654 int ret; 655 656 /* 657 * Lock destination range to serialize with concurrent readahead() and 658 * source range to serialize with relocation. 659 */ 660 btrfs_double_extent_lock(src, loff, dst, dst_loff, len); 661 ret = btrfs_clone(src, dst, loff, len, ALIGN(len, bs), dst_loff, 1); 662 btrfs_double_extent_unlock(src, loff, dst, dst_loff, len); 663 664 return ret; 665 } 666 667 static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen, 668 struct inode *dst, u64 dst_loff) 669 { 670 int ret = 0; 671 u64 i, tail_len, chunk_count; 672 struct btrfs_root *root_dst = BTRFS_I(dst)->root; 673 674 spin_lock(&root_dst->root_item_lock); 675 if (root_dst->send_in_progress) { 676 btrfs_warn_rl(root_dst->fs_info, 677 "cannot deduplicate to root %llu while send operations are using it (%d in progress)", 678 root_dst->root_key.objectid, 679 root_dst->send_in_progress); 680 spin_unlock(&root_dst->root_item_lock); 681 return -EAGAIN; 682 } 683 root_dst->dedupe_in_progress++; 684 spin_unlock(&root_dst->root_item_lock); 685 686 tail_len = olen % BTRFS_MAX_DEDUPE_LEN; 687 chunk_count = div_u64(olen, BTRFS_MAX_DEDUPE_LEN); 688 689 for (i = 0; i < chunk_count; i++) { 690 ret = btrfs_extent_same_range(src, loff, BTRFS_MAX_DEDUPE_LEN, 691 dst, dst_loff); 692 if (ret) 693 goto out; 694 695 loff += BTRFS_MAX_DEDUPE_LEN; 696 dst_loff += BTRFS_MAX_DEDUPE_LEN; 697 } 698 699 if (tail_len > 0) 700 ret = btrfs_extent_same_range(src, loff, tail_len, dst, dst_loff); 701 out: 702 spin_lock(&root_dst->root_item_lock); 703 root_dst->dedupe_in_progress--; 704 spin_unlock(&root_dst->root_item_lock); 705 706 return ret; 707 } 708 709 static noinline int btrfs_clone_files(struct file *file, struct file *file_src, 710 u64 off, u64 olen, u64 destoff) 711 { 712 struct inode *inode = file_inode(file); 713 struct inode *src = file_inode(file_src); 714 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 715 int ret; 716 int wb_ret; 717 u64 len = olen; 718 u64 bs = fs_info->sb->s_blocksize; 719 720 /* 721 * VFS's generic_remap_file_range_prep() protects us from cloning the 722 * eof block into the middle of a file, which would result in corruption 723 * if the file size is not blocksize aligned. So we don't need to check 724 * for that case here. 725 */ 726 if (off + len == src->i_size) 727 len = ALIGN(src->i_size, bs) - off; 728 729 if (destoff > inode->i_size) { 730 const u64 wb_start = ALIGN_DOWN(inode->i_size, bs); 731 732 ret = btrfs_cont_expand(BTRFS_I(inode), inode->i_size, destoff); 733 if (ret) 734 return ret; 735 /* 736 * We may have truncated the last block if the inode's size is 737 * not sector size aligned, so we need to wait for writeback to 738 * complete before proceeding further, otherwise we can race 739 * with cloning and attempt to increment a reference to an 740 * extent that no longer exists (writeback completed right after 741 * we found the previous extent covering eof and before we 742 * attempted to increment its reference count). 743 */ 744 ret = btrfs_wait_ordered_range(inode, wb_start, 745 destoff - wb_start); 746 if (ret) 747 return ret; 748 } 749 750 /* 751 * Lock destination range to serialize with concurrent readahead() and 752 * source range to serialize with relocation. 753 */ 754 btrfs_double_extent_lock(src, off, inode, destoff, len); 755 ret = btrfs_clone(src, inode, off, olen, len, destoff, 0); 756 btrfs_double_extent_unlock(src, off, inode, destoff, len); 757 758 /* 759 * We may have copied an inline extent into a page of the destination 760 * range, so wait for writeback to complete before truncating pages 761 * from the page cache. This is a rare case. 762 */ 763 wb_ret = btrfs_wait_ordered_range(inode, destoff, len); 764 ret = ret ? ret : wb_ret; 765 /* 766 * Truncate page cache pages so that future reads will see the cloned 767 * data immediately and not the previous data. 768 */ 769 truncate_inode_pages_range(&inode->i_data, 770 round_down(destoff, PAGE_SIZE), 771 round_up(destoff + len, PAGE_SIZE) - 1); 772 773 return ret; 774 } 775 776 static int btrfs_remap_file_range_prep(struct file *file_in, loff_t pos_in, 777 struct file *file_out, loff_t pos_out, 778 loff_t *len, unsigned int remap_flags) 779 { 780 struct inode *inode_in = file_inode(file_in); 781 struct inode *inode_out = file_inode(file_out); 782 u64 bs = BTRFS_I(inode_out)->root->fs_info->sb->s_blocksize; 783 u64 wb_len; 784 int ret; 785 786 if (!(remap_flags & REMAP_FILE_DEDUP)) { 787 struct btrfs_root *root_out = BTRFS_I(inode_out)->root; 788 789 if (btrfs_root_readonly(root_out)) 790 return -EROFS; 791 792 ASSERT(inode_in->i_sb == inode_out->i_sb); 793 } 794 795 /* Don't make the dst file partly checksummed */ 796 if ((BTRFS_I(inode_in)->flags & BTRFS_INODE_NODATASUM) != 797 (BTRFS_I(inode_out)->flags & BTRFS_INODE_NODATASUM)) { 798 return -EINVAL; 799 } 800 801 /* 802 * Now that the inodes are locked, we need to start writeback ourselves 803 * and can not rely on the writeback from the VFS's generic helper 804 * generic_remap_file_range_prep() because: 805 * 806 * 1) For compression we must call filemap_fdatawrite_range() range 807 * twice (btrfs_fdatawrite_range() does it for us), and the generic 808 * helper only calls it once; 809 * 810 * 2) filemap_fdatawrite_range(), called by the generic helper only 811 * waits for the writeback to complete, i.e. for IO to be done, and 812 * not for the ordered extents to complete. We need to wait for them 813 * to complete so that new file extent items are in the fs tree. 814 */ 815 if (*len == 0 && !(remap_flags & REMAP_FILE_DEDUP)) 816 wb_len = ALIGN(inode_in->i_size, bs) - ALIGN_DOWN(pos_in, bs); 817 else 818 wb_len = ALIGN(*len, bs); 819 820 /* 821 * Workaround to make sure NOCOW buffered write reach disk as NOCOW. 822 * 823 * Btrfs' back references do not have a block level granularity, they 824 * work at the whole extent level. 825 * NOCOW buffered write without data space reserved may not be able 826 * to fall back to CoW due to lack of data space, thus could cause 827 * data loss. 828 * 829 * Here we take a shortcut by flushing the whole inode, so that all 830 * nocow write should reach disk as nocow before we increase the 831 * reference of the extent. We could do better by only flushing NOCOW 832 * data, but that needs extra accounting. 833 * 834 * Also we don't need to check ASYNC_EXTENT, as async extent will be 835 * CoWed anyway, not affecting nocow part. 836 */ 837 ret = filemap_flush(inode_in->i_mapping); 838 if (ret < 0) 839 return ret; 840 841 ret = btrfs_wait_ordered_range(inode_in, ALIGN_DOWN(pos_in, bs), 842 wb_len); 843 if (ret < 0) 844 return ret; 845 ret = btrfs_wait_ordered_range(inode_out, ALIGN_DOWN(pos_out, bs), 846 wb_len); 847 if (ret < 0) 848 return ret; 849 850 return generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out, 851 len, remap_flags); 852 } 853 854 static bool file_sync_write(const struct file *file) 855 { 856 if (file->f_flags & (__O_SYNC | O_DSYNC)) 857 return true; 858 if (IS_SYNC(file_inode(file))) 859 return true; 860 861 return false; 862 } 863 864 loff_t btrfs_remap_file_range(struct file *src_file, loff_t off, 865 struct file *dst_file, loff_t destoff, loff_t len, 866 unsigned int remap_flags) 867 { 868 struct inode *src_inode = file_inode(src_file); 869 struct inode *dst_inode = file_inode(dst_file); 870 bool same_inode = dst_inode == src_inode; 871 int ret; 872 873 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY)) 874 return -EINVAL; 875 876 if (same_inode) { 877 btrfs_inode_lock(src_inode, BTRFS_ILOCK_MMAP); 878 } else { 879 lock_two_nondirectories(src_inode, dst_inode); 880 btrfs_double_mmap_lock(src_inode, dst_inode); 881 } 882 883 ret = btrfs_remap_file_range_prep(src_file, off, dst_file, destoff, 884 &len, remap_flags); 885 if (ret < 0 || len == 0) 886 goto out_unlock; 887 888 if (remap_flags & REMAP_FILE_DEDUP) 889 ret = btrfs_extent_same(src_inode, off, len, dst_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, dst_inode); 899 } 900 901 /* 902 * If either the source or the destination file was opened with O_SYNC, 903 * O_DSYNC or has the S_SYNC attribute, fsync both the destination and 904 * source files/ranges, so that after a successful return (0) followed 905 * by a power failure results in the reflinked data to be readable from 906 * both files/ranges. 907 */ 908 if (ret == 0 && len > 0 && 909 (file_sync_write(src_file) || file_sync_write(dst_file))) { 910 ret = btrfs_sync_file(src_file, off, off + len - 1, 0); 911 if (ret == 0) 912 ret = btrfs_sync_file(dst_file, destoff, 913 destoff + len - 1, 0); 914 } 915 916 return ret < 0 ? ret : len; 917 } 918