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 11 #define BTRFS_MAX_DEDUPE_LEN SZ_16M 12 13 static int clone_finish_inode_update(struct btrfs_trans_handle *trans, 14 struct inode *inode, 15 u64 endoff, 16 const u64 destoff, 17 const u64 olen, 18 int no_time_update) 19 { 20 struct btrfs_root *root = BTRFS_I(inode)->root; 21 int ret; 22 23 inode_inc_iversion(inode); 24 if (!no_time_update) 25 inode->i_mtime = inode->i_ctime = current_time(inode); 26 /* 27 * We round up to the block size at eof when determining which 28 * extents to clone above, but shouldn't round up the file size. 29 */ 30 if (endoff > destoff + olen) 31 endoff = destoff + olen; 32 if (endoff > inode->i_size) { 33 i_size_write(inode, endoff); 34 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0); 35 } 36 37 ret = btrfs_update_inode(trans, root, BTRFS_I(inode)); 38 if (ret) { 39 btrfs_abort_transaction(trans, ret); 40 btrfs_end_transaction(trans); 41 goto out; 42 } 43 ret = btrfs_end_transaction(trans); 44 out: 45 return ret; 46 } 47 48 static int copy_inline_to_page(struct btrfs_inode *inode, 49 const u64 file_offset, 50 char *inline_data, 51 const u64 size, 52 const u64 datal, 53 const u8 comp_type) 54 { 55 const u64 block_size = btrfs_inode_sectorsize(inode); 56 const u64 range_end = file_offset + block_size - 1; 57 const size_t inline_size = size - btrfs_file_extent_calc_inline_size(0); 58 char *data_start = inline_data + btrfs_file_extent_calc_inline_size(0); 59 struct extent_changeset *data_reserved = NULL; 60 struct page *page = NULL; 61 struct address_space *mapping = inode->vfs_inode.i_mapping; 62 int ret; 63 64 ASSERT(IS_ALIGNED(file_offset, block_size)); 65 66 /* 67 * We have flushed and locked the ranges of the source and destination 68 * inodes, we also have locked the inodes, so we are safe to do a 69 * reservation here. Also we must not do the reservation while holding 70 * a transaction open, otherwise we would deadlock. 71 */ 72 ret = btrfs_delalloc_reserve_space(inode, &data_reserved, file_offset, 73 block_size); 74 if (ret) 75 goto out; 76 77 page = find_or_create_page(mapping, file_offset >> PAGE_SHIFT, 78 btrfs_alloc_write_mask(mapping)); 79 if (!page) { 80 ret = -ENOMEM; 81 goto out_unlock; 82 } 83 84 ret = set_page_extent_mapped(page); 85 if (ret < 0) 86 goto out_unlock; 87 88 clear_extent_bit(&inode->io_tree, file_offset, range_end, 89 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 90 0, 0, NULL); 91 ret = btrfs_set_extent_delalloc(inode, file_offset, range_end, 0, NULL); 92 if (ret) 93 goto out_unlock; 94 95 /* 96 * After dirtying the page our caller will need to start a transaction, 97 * and if we are low on metadata free space, that can cause flushing of 98 * delalloc for all inodes in order to get metadata space released. 99 * However we are holding the range locked for the whole duration of 100 * the clone/dedupe operation, so we may deadlock if that happens and no 101 * other task releases enough space. So mark this inode as not being 102 * possible to flush to avoid such deadlock. We will clear that flag 103 * when we finish cloning all extents, since a transaction is started 104 * after finding each extent to clone. 105 */ 106 set_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &inode->runtime_flags); 107 108 if (comp_type == BTRFS_COMPRESS_NONE) { 109 memcpy_to_page(page, 0, data_start, datal); 110 flush_dcache_page(page); 111 } else { 112 ret = btrfs_decompress(comp_type, data_start, page, 0, 113 inline_size, datal); 114 if (ret) 115 goto out_unlock; 116 flush_dcache_page(page); 117 } 118 119 /* 120 * If our inline data is smaller then the block/page size, then the 121 * remaining of the block/page is equivalent to zeroes. We had something 122 * like the following done: 123 * 124 * $ xfs_io -f -c "pwrite -S 0xab 0 500" file 125 * $ sync # (or fsync) 126 * $ xfs_io -c "falloc 0 4K" file 127 * $ xfs_io -c "pwrite -S 0xcd 4K 4K" 128 * 129 * So what's in the range [500, 4095] corresponds to zeroes. 130 */ 131 if (datal < block_size) { 132 char *map; 133 134 map = kmap(page); 135 memset(map + datal, 0, block_size - datal); 136 flush_dcache_page(page); 137 kunmap(page); 138 } 139 140 SetPageUptodate(page); 141 ClearPageChecked(page); 142 set_page_dirty(page); 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 ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset, 211 inline_data, size, datal, 212 comp_type); 213 goto out; 214 } 215 } else if (i_size_read(dst) <= datal) { 216 struct btrfs_file_extent_item *ei; 217 218 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 219 struct btrfs_file_extent_item); 220 /* 221 * If it's an inline extent replace it with the source inline 222 * extent, otherwise copy the source inline extent data into 223 * the respective page at the destination inode. 224 */ 225 if (btrfs_file_extent_type(path->nodes[0], ei) == 226 BTRFS_FILE_EXTENT_INLINE) 227 goto copy_inline_extent; 228 229 ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset, 230 inline_data, size, datal, comp_type); 231 goto out; 232 } 233 234 copy_inline_extent: 235 ret = 0; 236 /* 237 * We have no extent items, or we have an extent at offset 0 which may 238 * or may not be inlined. All these cases are dealt the same way. 239 */ 240 if (i_size_read(dst) > datal) { 241 /* 242 * At the destination offset 0 we have either a hole, a regular 243 * extent or an inline extent larger then the one we want to 244 * clone. Deal with all these cases by copying the inline extent 245 * data into the respective page at the destination inode. 246 */ 247 ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset, 248 inline_data, size, datal, comp_type); 249 goto out; 250 } 251 252 btrfs_release_path(path); 253 /* 254 * If we end up here it means were copy the inline extent into a leaf 255 * of the destination inode. We know we will drop or adjust at most one 256 * extent item in the destination root. 257 * 258 * 1 unit - adjusting old extent (we may have to split it) 259 * 1 unit - add new extent 260 * 1 unit - inode update 261 */ 262 trans = btrfs_start_transaction(root, 3); 263 if (IS_ERR(trans)) { 264 ret = PTR_ERR(trans); 265 trans = NULL; 266 goto out; 267 } 268 drop_args.path = path; 269 drop_args.start = drop_start; 270 drop_args.end = aligned_end; 271 drop_args.drop_cache = true; 272 ret = btrfs_drop_extents(trans, root, BTRFS_I(dst), &drop_args); 273 if (ret) 274 goto out; 275 ret = btrfs_insert_empty_item(trans, root, path, new_key, size); 276 if (ret) 277 goto out; 278 279 write_extent_buffer(path->nodes[0], inline_data, 280 btrfs_item_ptr_offset(path->nodes[0], 281 path->slots[0]), 282 size); 283 btrfs_update_inode_bytes(BTRFS_I(dst), datal, drop_args.bytes_found); 284 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(dst)->runtime_flags); 285 ret = btrfs_inode_set_file_extent_range(BTRFS_I(dst), 0, aligned_end); 286 out: 287 if (!ret && !trans) { 288 /* 289 * Release path before starting a new transaction so we don't 290 * hold locks that would confuse lockdep. 291 */ 292 btrfs_release_path(path); 293 /* 294 * No transaction here means we copied the inline extent into a 295 * page of the destination inode. 296 * 297 * 1 unit to update inode item 298 */ 299 trans = btrfs_start_transaction(root, 1); 300 if (IS_ERR(trans)) { 301 ret = PTR_ERR(trans); 302 trans = NULL; 303 } 304 } 305 if (ret && trans) { 306 btrfs_abort_transaction(trans, ret); 307 btrfs_end_transaction(trans); 308 } 309 if (!ret) 310 *trans_out = trans; 311 312 return ret; 313 } 314 315 /** 316 * btrfs_clone() - clone a range from inode file to another 317 * 318 * @src: Inode to clone from 319 * @inode: Inode to clone to 320 * @off: Offset within source to start clone from 321 * @olen: Original length, passed by user, of range to clone 322 * @olen_aligned: Block-aligned value of olen 323 * @destoff: Offset within @inode to start clone 324 * @no_time_update: Whether to update mtime/ctime on the target inode 325 */ 326 static int btrfs_clone(struct inode *src, struct inode *inode, 327 const u64 off, const u64 olen, const u64 olen_aligned, 328 const u64 destoff, int no_time_update) 329 { 330 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 331 struct btrfs_path *path = NULL; 332 struct extent_buffer *leaf; 333 struct btrfs_trans_handle *trans; 334 char *buf = NULL; 335 struct btrfs_key key; 336 u32 nritems; 337 int slot; 338 int ret; 339 const u64 len = olen_aligned; 340 u64 last_dest_end = destoff; 341 342 ret = -ENOMEM; 343 buf = kvmalloc(fs_info->nodesize, GFP_KERNEL); 344 if (!buf) 345 return ret; 346 347 path = btrfs_alloc_path(); 348 if (!path) { 349 kvfree(buf); 350 return ret; 351 } 352 353 path->reada = READA_FORWARD; 354 /* Clone data */ 355 key.objectid = btrfs_ino(BTRFS_I(src)); 356 key.type = BTRFS_EXTENT_DATA_KEY; 357 key.offset = off; 358 359 while (1) { 360 u64 next_key_min_offset = key.offset + 1; 361 struct btrfs_file_extent_item *extent; 362 u64 extent_gen; 363 int type; 364 u32 size; 365 struct btrfs_key new_key; 366 u64 disko = 0, diskl = 0; 367 u64 datao = 0, datal = 0; 368 u8 comp; 369 u64 drop_start; 370 371 /* Note the key will change type as we walk through the tree */ 372 ret = btrfs_search_slot(NULL, BTRFS_I(src)->root, &key, path, 373 0, 0); 374 if (ret < 0) 375 goto out; 376 /* 377 * First search, if no extent item that starts at offset off was 378 * found but the previous item is an extent item, it's possible 379 * it might overlap our target range, therefore process it. 380 */ 381 if (key.offset == off && ret > 0 && path->slots[0] > 0) { 382 btrfs_item_key_to_cpu(path->nodes[0], &key, 383 path->slots[0] - 1); 384 if (key.type == BTRFS_EXTENT_DATA_KEY) 385 path->slots[0]--; 386 } 387 388 nritems = btrfs_header_nritems(path->nodes[0]); 389 process_slot: 390 if (path->slots[0] >= nritems) { 391 ret = btrfs_next_leaf(BTRFS_I(src)->root, path); 392 if (ret < 0) 393 goto out; 394 if (ret > 0) 395 break; 396 nritems = btrfs_header_nritems(path->nodes[0]); 397 } 398 leaf = path->nodes[0]; 399 slot = path->slots[0]; 400 401 btrfs_item_key_to_cpu(leaf, &key, slot); 402 if (key.type > BTRFS_EXTENT_DATA_KEY || 403 key.objectid != btrfs_ino(BTRFS_I(src))) 404 break; 405 406 ASSERT(key.type == BTRFS_EXTENT_DATA_KEY); 407 408 extent = btrfs_item_ptr(leaf, slot, 409 struct btrfs_file_extent_item); 410 extent_gen = btrfs_file_extent_generation(leaf, extent); 411 comp = btrfs_file_extent_compression(leaf, extent); 412 type = btrfs_file_extent_type(leaf, extent); 413 if (type == BTRFS_FILE_EXTENT_REG || 414 type == BTRFS_FILE_EXTENT_PREALLOC) { 415 disko = btrfs_file_extent_disk_bytenr(leaf, extent); 416 diskl = btrfs_file_extent_disk_num_bytes(leaf, extent); 417 datao = btrfs_file_extent_offset(leaf, extent); 418 datal = btrfs_file_extent_num_bytes(leaf, extent); 419 } else if (type == BTRFS_FILE_EXTENT_INLINE) { 420 /* Take upper bound, may be compressed */ 421 datal = btrfs_file_extent_ram_bytes(leaf, extent); 422 } 423 424 /* 425 * The first search might have left us at an extent item that 426 * ends before our target range's start, can happen if we have 427 * holes and NO_HOLES feature enabled. 428 */ 429 if (key.offset + datal <= off) { 430 path->slots[0]++; 431 goto process_slot; 432 } else if (key.offset >= off + len) { 433 break; 434 } 435 next_key_min_offset = key.offset + datal; 436 size = btrfs_item_size_nr(leaf, slot); 437 read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf, slot), 438 size); 439 440 btrfs_release_path(path); 441 442 memcpy(&new_key, &key, sizeof(new_key)); 443 new_key.objectid = btrfs_ino(BTRFS_I(inode)); 444 if (off <= key.offset) 445 new_key.offset = key.offset + destoff - off; 446 else 447 new_key.offset = destoff; 448 449 /* 450 * Deal with a hole that doesn't have an extent item that 451 * represents it (NO_HOLES feature enabled). 452 * This hole is either in the middle of the cloning range or at 453 * the beginning (fully overlaps it or partially overlaps it). 454 */ 455 if (new_key.offset != last_dest_end) 456 drop_start = last_dest_end; 457 else 458 drop_start = new_key.offset; 459 460 if (type == BTRFS_FILE_EXTENT_REG || 461 type == BTRFS_FILE_EXTENT_PREALLOC) { 462 struct btrfs_replace_extent_info clone_info; 463 464 /* 465 * a | --- range to clone ---| b 466 * | ------------- extent ------------- | 467 */ 468 469 /* Subtract range b */ 470 if (key.offset + datal > off + len) 471 datal = off + len - key.offset; 472 473 /* Subtract range a */ 474 if (off > key.offset) { 475 datao += off - key.offset; 476 datal -= off - key.offset; 477 } 478 479 clone_info.disk_offset = disko; 480 clone_info.disk_len = diskl; 481 clone_info.data_offset = datao; 482 clone_info.data_len = datal; 483 clone_info.file_offset = new_key.offset; 484 clone_info.extent_buf = buf; 485 clone_info.is_new_extent = false; 486 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 487 drop_start, new_key.offset + datal - 1, 488 &clone_info, &trans); 489 if (ret) 490 goto out; 491 } else if (type == BTRFS_FILE_EXTENT_INLINE) { 492 /* 493 * Inline extents always have to start at file offset 0 494 * and can never be bigger then the sector size. We can 495 * never clone only parts of an inline extent, since all 496 * reflink operations must start at a sector size aligned 497 * offset, and the length must be aligned too or end at 498 * the i_size (which implies the whole inlined data). 499 */ 500 ASSERT(key.offset == 0); 501 ASSERT(datal <= fs_info->sectorsize); 502 if (key.offset != 0 || datal > fs_info->sectorsize) 503 return -EUCLEAN; 504 505 ret = clone_copy_inline_extent(inode, path, &new_key, 506 drop_start, datal, size, 507 comp, buf, &trans); 508 if (ret) 509 goto out; 510 } 511 512 btrfs_release_path(path); 513 514 /* 515 * If this is a new extent update the last_reflink_trans of both 516 * inodes. This is used by fsync to make sure it does not log 517 * multiple checksum items with overlapping ranges. For older 518 * extents we don't need to do it since inode logging skips the 519 * checksums for older extents. Also ignore holes and inline 520 * extents because they don't have checksums in the csum tree. 521 */ 522 if (extent_gen == trans->transid && disko > 0) { 523 BTRFS_I(src)->last_reflink_trans = trans->transid; 524 BTRFS_I(inode)->last_reflink_trans = trans->transid; 525 } 526 527 last_dest_end = ALIGN(new_key.offset + datal, 528 fs_info->sectorsize); 529 ret = clone_finish_inode_update(trans, inode, last_dest_end, 530 destoff, olen, no_time_update); 531 if (ret) 532 goto out; 533 if (new_key.offset + datal >= destoff + len) 534 break; 535 536 btrfs_release_path(path); 537 key.offset = next_key_min_offset; 538 539 if (fatal_signal_pending(current)) { 540 ret = -EINTR; 541 goto out; 542 } 543 544 cond_resched(); 545 } 546 ret = 0; 547 548 if (last_dest_end < destoff + len) { 549 /* 550 * We have an implicit hole that fully or partially overlaps our 551 * cloning range at its end. This means that we either have the 552 * NO_HOLES feature enabled or the implicit hole happened due to 553 * mixing buffered and direct IO writes against this file. 554 */ 555 btrfs_release_path(path); 556 557 /* 558 * When using NO_HOLES and we are cloning a range that covers 559 * only a hole (no extents) into a range beyond the current 560 * i_size, punching a hole in the target range will not create 561 * an extent map defining a hole, because the range starts at or 562 * beyond current i_size. If the file previously had an i_size 563 * greater than the new i_size set by this clone operation, we 564 * need to make sure the next fsync is a full fsync, so that it 565 * detects and logs a hole covering a range from the current 566 * i_size to the new i_size. If the clone range covers extents, 567 * besides a hole, then we know the full sync flag was already 568 * set by previous calls to btrfs_replace_file_extents() that 569 * replaced file extent items. 570 */ 571 if (last_dest_end >= i_size_read(inode)) 572 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, 573 &BTRFS_I(inode)->runtime_flags); 574 575 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, 576 last_dest_end, destoff + len - 1, NULL, &trans); 577 if (ret) 578 goto out; 579 580 ret = clone_finish_inode_update(trans, inode, destoff + len, 581 destoff, olen, no_time_update); 582 } 583 584 out: 585 btrfs_free_path(path); 586 kvfree(buf); 587 clear_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &BTRFS_I(inode)->runtime_flags); 588 589 return ret; 590 } 591 592 static void btrfs_double_extent_unlock(struct inode *inode1, u64 loff1, 593 struct inode *inode2, u64 loff2, u64 len) 594 { 595 unlock_extent(&BTRFS_I(inode1)->io_tree, loff1, loff1 + len - 1); 596 unlock_extent(&BTRFS_I(inode2)->io_tree, loff2, loff2 + len - 1); 597 } 598 599 static void btrfs_double_extent_lock(struct inode *inode1, u64 loff1, 600 struct inode *inode2, u64 loff2, u64 len) 601 { 602 if (inode1 < inode2) { 603 swap(inode1, inode2); 604 swap(loff1, loff2); 605 } else if (inode1 == inode2 && loff2 < loff1) { 606 swap(loff1, loff2); 607 } 608 lock_extent(&BTRFS_I(inode1)->io_tree, loff1, loff1 + len - 1); 609 lock_extent(&BTRFS_I(inode2)->io_tree, loff2, loff2 + len - 1); 610 } 611 612 static void btrfs_double_mmap_lock(struct inode *inode1, struct inode *inode2) 613 { 614 if (inode1 < inode2) 615 swap(inode1, inode2); 616 down_write(&BTRFS_I(inode1)->i_mmap_lock); 617 down_write_nested(&BTRFS_I(inode2)->i_mmap_lock, SINGLE_DEPTH_NESTING); 618 } 619 620 static void btrfs_double_mmap_unlock(struct inode *inode1, struct inode *inode2) 621 { 622 up_write(&BTRFS_I(inode1)->i_mmap_lock); 623 up_write(&BTRFS_I(inode2)->i_mmap_lock); 624 } 625 626 static int btrfs_extent_same_range(struct inode *src, u64 loff, u64 len, 627 struct inode *dst, u64 dst_loff) 628 { 629 const u64 bs = BTRFS_I(src)->root->fs_info->sb->s_blocksize; 630 int ret; 631 632 /* 633 * Lock destination range to serialize with concurrent readpages() and 634 * source range to serialize with relocation. 635 */ 636 btrfs_double_extent_lock(src, loff, dst, dst_loff, len); 637 ret = btrfs_clone(src, dst, loff, len, ALIGN(len, bs), dst_loff, 1); 638 btrfs_double_extent_unlock(src, loff, dst, dst_loff, len); 639 640 return ret; 641 } 642 643 static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen, 644 struct inode *dst, u64 dst_loff) 645 { 646 int ret; 647 u64 i, tail_len, chunk_count; 648 struct btrfs_root *root_dst = BTRFS_I(dst)->root; 649 650 spin_lock(&root_dst->root_item_lock); 651 if (root_dst->send_in_progress) { 652 btrfs_warn_rl(root_dst->fs_info, 653 "cannot deduplicate to root %llu while send operations are using it (%d in progress)", 654 root_dst->root_key.objectid, 655 root_dst->send_in_progress); 656 spin_unlock(&root_dst->root_item_lock); 657 return -EAGAIN; 658 } 659 root_dst->dedupe_in_progress++; 660 spin_unlock(&root_dst->root_item_lock); 661 662 tail_len = olen % BTRFS_MAX_DEDUPE_LEN; 663 chunk_count = div_u64(olen, BTRFS_MAX_DEDUPE_LEN); 664 665 for (i = 0; i < chunk_count; i++) { 666 ret = btrfs_extent_same_range(src, loff, BTRFS_MAX_DEDUPE_LEN, 667 dst, dst_loff); 668 if (ret) 669 goto out; 670 671 loff += BTRFS_MAX_DEDUPE_LEN; 672 dst_loff += BTRFS_MAX_DEDUPE_LEN; 673 } 674 675 if (tail_len > 0) 676 ret = btrfs_extent_same_range(src, loff, tail_len, dst, dst_loff); 677 out: 678 spin_lock(&root_dst->root_item_lock); 679 root_dst->dedupe_in_progress--; 680 spin_unlock(&root_dst->root_item_lock); 681 682 return ret; 683 } 684 685 static noinline int btrfs_clone_files(struct file *file, struct file *file_src, 686 u64 off, u64 olen, u64 destoff) 687 { 688 struct inode *inode = file_inode(file); 689 struct inode *src = file_inode(file_src); 690 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 691 int ret; 692 int wb_ret; 693 u64 len = olen; 694 u64 bs = fs_info->sb->s_blocksize; 695 696 /* 697 * VFS's generic_remap_file_range_prep() protects us from cloning the 698 * eof block into the middle of a file, which would result in corruption 699 * if the file size is not blocksize aligned. So we don't need to check 700 * for that case here. 701 */ 702 if (off + len == src->i_size) 703 len = ALIGN(src->i_size, bs) - off; 704 705 if (destoff > inode->i_size) { 706 const u64 wb_start = ALIGN_DOWN(inode->i_size, bs); 707 708 ret = btrfs_cont_expand(BTRFS_I(inode), inode->i_size, destoff); 709 if (ret) 710 return ret; 711 /* 712 * We may have truncated the last block if the inode's size is 713 * not sector size aligned, so we need to wait for writeback to 714 * complete before proceeding further, otherwise we can race 715 * with cloning and attempt to increment a reference to an 716 * extent that no longer exists (writeback completed right after 717 * we found the previous extent covering eof and before we 718 * attempted to increment its reference count). 719 */ 720 ret = btrfs_wait_ordered_range(inode, wb_start, 721 destoff - wb_start); 722 if (ret) 723 return ret; 724 } 725 726 /* 727 * Lock destination range to serialize with concurrent readpages() and 728 * source range to serialize with relocation. 729 */ 730 btrfs_double_extent_lock(src, off, inode, destoff, len); 731 ret = btrfs_clone(src, inode, off, olen, len, destoff, 0); 732 btrfs_double_extent_unlock(src, off, inode, destoff, len); 733 734 /* 735 * We may have copied an inline extent into a page of the destination 736 * range, so wait for writeback to complete before truncating pages 737 * from the page cache. This is a rare case. 738 */ 739 wb_ret = btrfs_wait_ordered_range(inode, destoff, len); 740 ret = ret ? ret : wb_ret; 741 /* 742 * Truncate page cache pages so that future reads will see the cloned 743 * data immediately and not the previous data. 744 */ 745 truncate_inode_pages_range(&inode->i_data, 746 round_down(destoff, PAGE_SIZE), 747 round_up(destoff + len, PAGE_SIZE) - 1); 748 749 return ret; 750 } 751 752 static int btrfs_remap_file_range_prep(struct file *file_in, loff_t pos_in, 753 struct file *file_out, loff_t pos_out, 754 loff_t *len, unsigned int remap_flags) 755 { 756 struct inode *inode_in = file_inode(file_in); 757 struct inode *inode_out = file_inode(file_out); 758 u64 bs = BTRFS_I(inode_out)->root->fs_info->sb->s_blocksize; 759 bool same_inode = inode_out == inode_in; 760 u64 wb_len; 761 int ret; 762 763 if (!(remap_flags & REMAP_FILE_DEDUP)) { 764 struct btrfs_root *root_out = BTRFS_I(inode_out)->root; 765 766 if (btrfs_root_readonly(root_out)) 767 return -EROFS; 768 769 if (file_in->f_path.mnt != file_out->f_path.mnt || 770 inode_in->i_sb != inode_out->i_sb) 771 return -EXDEV; 772 } 773 774 /* Don't make the dst file partly checksummed */ 775 if ((BTRFS_I(inode_in)->flags & BTRFS_INODE_NODATASUM) != 776 (BTRFS_I(inode_out)->flags & BTRFS_INODE_NODATASUM)) { 777 return -EINVAL; 778 } 779 780 /* 781 * Now that the inodes are locked, we need to start writeback ourselves 782 * and can not rely on the writeback from the VFS's generic helper 783 * generic_remap_file_range_prep() because: 784 * 785 * 1) For compression we must call filemap_fdatawrite_range() range 786 * twice (btrfs_fdatawrite_range() does it for us), and the generic 787 * helper only calls it once; 788 * 789 * 2) filemap_fdatawrite_range(), called by the generic helper only 790 * waits for the writeback to complete, i.e. for IO to be done, and 791 * not for the ordered extents to complete. We need to wait for them 792 * to complete so that new file extent items are in the fs tree. 793 */ 794 if (*len == 0 && !(remap_flags & REMAP_FILE_DEDUP)) 795 wb_len = ALIGN(inode_in->i_size, bs) - ALIGN_DOWN(pos_in, bs); 796 else 797 wb_len = ALIGN(*len, bs); 798 799 /* 800 * Since we don't lock ranges, wait for ongoing lockless dio writes (as 801 * any in progress could create its ordered extents after we wait for 802 * existing ordered extents below). 803 */ 804 inode_dio_wait(inode_in); 805 if (!same_inode) 806 inode_dio_wait(inode_out); 807 808 /* 809 * Workaround to make sure NOCOW buffered write reach disk as NOCOW. 810 * 811 * Btrfs' back references do not have a block level granularity, they 812 * work at the whole extent level. 813 * NOCOW buffered write without data space reserved may not be able 814 * to fall back to CoW due to lack of data space, thus could cause 815 * data loss. 816 * 817 * Here we take a shortcut by flushing the whole inode, so that all 818 * nocow write should reach disk as nocow before we increase the 819 * reference of the extent. We could do better by only flushing NOCOW 820 * data, but that needs extra accounting. 821 * 822 * Also we don't need to check ASYNC_EXTENT, as async extent will be 823 * CoWed anyway, not affecting nocow part. 824 */ 825 ret = filemap_flush(inode_in->i_mapping); 826 if (ret < 0) 827 return ret; 828 829 ret = btrfs_wait_ordered_range(inode_in, ALIGN_DOWN(pos_in, bs), 830 wb_len); 831 if (ret < 0) 832 return ret; 833 ret = btrfs_wait_ordered_range(inode_out, ALIGN_DOWN(pos_out, bs), 834 wb_len); 835 if (ret < 0) 836 return ret; 837 838 return generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out, 839 len, remap_flags); 840 } 841 842 static bool file_sync_write(const struct file *file) 843 { 844 if (file->f_flags & (__O_SYNC | O_DSYNC)) 845 return true; 846 if (IS_SYNC(file_inode(file))) 847 return true; 848 849 return false; 850 } 851 852 loff_t btrfs_remap_file_range(struct file *src_file, loff_t off, 853 struct file *dst_file, loff_t destoff, loff_t len, 854 unsigned int remap_flags) 855 { 856 struct inode *src_inode = file_inode(src_file); 857 struct inode *dst_inode = file_inode(dst_file); 858 bool same_inode = dst_inode == src_inode; 859 int ret; 860 861 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY)) 862 return -EINVAL; 863 864 if (same_inode) { 865 btrfs_inode_lock(src_inode, BTRFS_ILOCK_MMAP); 866 } else { 867 lock_two_nondirectories(src_inode, dst_inode); 868 btrfs_double_mmap_lock(src_inode, dst_inode); 869 } 870 871 ret = btrfs_remap_file_range_prep(src_file, off, dst_file, destoff, 872 &len, remap_flags); 873 if (ret < 0 || len == 0) 874 goto out_unlock; 875 876 if (remap_flags & REMAP_FILE_DEDUP) 877 ret = btrfs_extent_same(src_inode, off, len, dst_inode, destoff); 878 else 879 ret = btrfs_clone_files(dst_file, src_file, off, len, destoff); 880 881 out_unlock: 882 if (same_inode) { 883 btrfs_inode_unlock(src_inode, BTRFS_ILOCK_MMAP); 884 } else { 885 btrfs_double_mmap_unlock(src_inode, dst_inode); 886 unlock_two_nondirectories(src_inode, dst_inode); 887 } 888 889 /* 890 * If either the source or the destination file was opened with O_SYNC, 891 * O_DSYNC or has the S_SYNC attribute, fsync both the destination and 892 * source files/ranges, so that after a successful return (0) followed 893 * by a power failure results in the reflinked data to be readable from 894 * both files/ranges. 895 */ 896 if (ret == 0 && len > 0 && 897 (file_sync_write(src_file) || file_sync_write(dst_file))) { 898 ret = btrfs_sync_file(src_file, off, off + len - 1, 0); 899 if (ret == 0) 900 ret = btrfs_sync_file(dst_file, destoff, 901 destoff + len - 1, 0); 902 } 903 904 return ret < 0 ? ret : len; 905 } 906