1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2008 Oracle. All rights reserved. 4 */ 5 6 #include <linux/kernel.h> 7 #include <linux/bio.h> 8 #include <linux/file.h> 9 #include <linux/fs.h> 10 #include <linux/pagemap.h> 11 #include <linux/highmem.h> 12 #include <linux/time.h> 13 #include <linux/init.h> 14 #include <linux/string.h> 15 #include <linux/backing-dev.h> 16 #include <linux/writeback.h> 17 #include <linux/slab.h> 18 #include <linux/sched/mm.h> 19 #include <linux/log2.h> 20 #include <crypto/hash.h> 21 #include "misc.h" 22 #include "ctree.h" 23 #include "disk-io.h" 24 #include "transaction.h" 25 #include "btrfs_inode.h" 26 #include "volumes.h" 27 #include "ordered-data.h" 28 #include "compression.h" 29 #include "extent_io.h" 30 #include "extent_map.h" 31 #include "subpage.h" 32 #include "zoned.h" 33 34 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; 35 36 const char* btrfs_compress_type2str(enum btrfs_compression_type type) 37 { 38 switch (type) { 39 case BTRFS_COMPRESS_ZLIB: 40 case BTRFS_COMPRESS_LZO: 41 case BTRFS_COMPRESS_ZSTD: 42 case BTRFS_COMPRESS_NONE: 43 return btrfs_compress_types[type]; 44 default: 45 break; 46 } 47 48 return NULL; 49 } 50 51 bool btrfs_compress_is_valid_type(const char *str, size_t len) 52 { 53 int i; 54 55 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { 56 size_t comp_len = strlen(btrfs_compress_types[i]); 57 58 if (len < comp_len) 59 continue; 60 61 if (!strncmp(btrfs_compress_types[i], str, comp_len)) 62 return true; 63 } 64 return false; 65 } 66 67 static int compression_compress_pages(int type, struct list_head *ws, 68 struct address_space *mapping, u64 start, struct page **pages, 69 unsigned long *out_pages, unsigned long *total_in, 70 unsigned long *total_out) 71 { 72 switch (type) { 73 case BTRFS_COMPRESS_ZLIB: 74 return zlib_compress_pages(ws, mapping, start, pages, 75 out_pages, total_in, total_out); 76 case BTRFS_COMPRESS_LZO: 77 return lzo_compress_pages(ws, mapping, start, pages, 78 out_pages, total_in, total_out); 79 case BTRFS_COMPRESS_ZSTD: 80 return zstd_compress_pages(ws, mapping, start, pages, 81 out_pages, total_in, total_out); 82 case BTRFS_COMPRESS_NONE: 83 default: 84 /* 85 * This can happen when compression races with remount setting 86 * it to 'no compress', while caller doesn't call 87 * inode_need_compress() to check if we really need to 88 * compress. 89 * 90 * Not a big deal, just need to inform caller that we 91 * haven't allocated any pages yet. 92 */ 93 *out_pages = 0; 94 return -E2BIG; 95 } 96 } 97 98 static int compression_decompress_bio(int type, struct list_head *ws, 99 struct compressed_bio *cb) 100 { 101 switch (type) { 102 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); 103 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); 104 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); 105 case BTRFS_COMPRESS_NONE: 106 default: 107 /* 108 * This can't happen, the type is validated several times 109 * before we get here. 110 */ 111 BUG(); 112 } 113 } 114 115 static int compression_decompress(int type, struct list_head *ws, 116 unsigned char *data_in, struct page *dest_page, 117 unsigned long start_byte, size_t srclen, size_t destlen) 118 { 119 switch (type) { 120 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, 121 start_byte, srclen, destlen); 122 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, 123 start_byte, srclen, destlen); 124 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, 125 start_byte, srclen, destlen); 126 case BTRFS_COMPRESS_NONE: 127 default: 128 /* 129 * This can't happen, the type is validated several times 130 * before we get here. 131 */ 132 BUG(); 133 } 134 } 135 136 static int btrfs_decompress_bio(struct compressed_bio *cb); 137 138 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, 139 unsigned long disk_size) 140 { 141 return sizeof(struct compressed_bio) + 142 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size; 143 } 144 145 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio, 146 u64 disk_start) 147 { 148 struct btrfs_fs_info *fs_info = inode->root->fs_info; 149 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash); 150 const u32 csum_size = fs_info->csum_size; 151 const u32 sectorsize = fs_info->sectorsize; 152 struct page *page; 153 unsigned int i; 154 char *kaddr; 155 u8 csum[BTRFS_CSUM_SIZE]; 156 struct compressed_bio *cb = bio->bi_private; 157 u8 *cb_sum = cb->sums; 158 159 if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM)) 160 return 0; 161 162 shash->tfm = fs_info->csum_shash; 163 164 for (i = 0; i < cb->nr_pages; i++) { 165 u32 pg_offset; 166 u32 bytes_left = PAGE_SIZE; 167 page = cb->compressed_pages[i]; 168 169 /* Determine the remaining bytes inside the page first */ 170 if (i == cb->nr_pages - 1) 171 bytes_left = cb->compressed_len - i * PAGE_SIZE; 172 173 /* Hash through the page sector by sector */ 174 for (pg_offset = 0; pg_offset < bytes_left; 175 pg_offset += sectorsize) { 176 kaddr = page_address(page); 177 crypto_shash_digest(shash, kaddr + pg_offset, 178 sectorsize, csum); 179 180 if (memcmp(&csum, cb_sum, csum_size) != 0) { 181 btrfs_print_data_csum_error(inode, disk_start, 182 csum, cb_sum, cb->mirror_num); 183 if (btrfs_bio(bio)->device) 184 btrfs_dev_stat_inc_and_print( 185 btrfs_bio(bio)->device, 186 BTRFS_DEV_STAT_CORRUPTION_ERRS); 187 return -EIO; 188 } 189 cb_sum += csum_size; 190 disk_start += sectorsize; 191 } 192 } 193 return 0; 194 } 195 196 /* when we finish reading compressed pages from the disk, we 197 * decompress them and then run the bio end_io routines on the 198 * decompressed pages (in the inode address space). 199 * 200 * This allows the checksumming and other IO error handling routines 201 * to work normally 202 * 203 * The compressed pages are freed here, and it must be run 204 * in process context 205 */ 206 static void end_compressed_bio_read(struct bio *bio) 207 { 208 struct compressed_bio *cb = bio->bi_private; 209 struct inode *inode; 210 struct page *page; 211 unsigned int index; 212 unsigned int mirror = btrfs_bio(bio)->mirror_num; 213 int ret = 0; 214 215 if (bio->bi_status) 216 cb->errors = 1; 217 218 /* if there are more bios still pending for this compressed 219 * extent, just exit 220 */ 221 if (!refcount_dec_and_test(&cb->pending_bios)) 222 goto out; 223 224 /* 225 * Record the correct mirror_num in cb->orig_bio so that 226 * read-repair can work properly. 227 */ 228 btrfs_bio(cb->orig_bio)->mirror_num = mirror; 229 cb->mirror_num = mirror; 230 231 /* 232 * Some IO in this cb have failed, just skip checksum as there 233 * is no way it could be correct. 234 */ 235 if (cb->errors == 1) 236 goto csum_failed; 237 238 inode = cb->inode; 239 ret = check_compressed_csum(BTRFS_I(inode), bio, 240 bio->bi_iter.bi_sector << 9); 241 if (ret) 242 goto csum_failed; 243 244 /* ok, we're the last bio for this extent, lets start 245 * the decompression. 246 */ 247 ret = btrfs_decompress_bio(cb); 248 249 csum_failed: 250 if (ret) 251 cb->errors = 1; 252 253 /* release the compressed pages */ 254 index = 0; 255 for (index = 0; index < cb->nr_pages; index++) { 256 page = cb->compressed_pages[index]; 257 page->mapping = NULL; 258 put_page(page); 259 } 260 261 /* do io completion on the original bio */ 262 if (cb->errors) { 263 bio_io_error(cb->orig_bio); 264 } else { 265 struct bio_vec *bvec; 266 struct bvec_iter_all iter_all; 267 268 /* 269 * we have verified the checksum already, set page 270 * checked so the end_io handlers know about it 271 */ 272 ASSERT(!bio_flagged(bio, BIO_CLONED)); 273 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) 274 SetPageChecked(bvec->bv_page); 275 276 bio_endio(cb->orig_bio); 277 } 278 279 /* finally free the cb struct */ 280 kfree(cb->compressed_pages); 281 kfree(cb); 282 out: 283 bio_put(bio); 284 } 285 286 /* 287 * Clear the writeback bits on all of the file 288 * pages for a compressed write 289 */ 290 static noinline void end_compressed_writeback(struct inode *inode, 291 const struct compressed_bio *cb) 292 { 293 unsigned long index = cb->start >> PAGE_SHIFT; 294 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; 295 struct page *pages[16]; 296 unsigned long nr_pages = end_index - index + 1; 297 int i; 298 int ret; 299 300 if (cb->errors) 301 mapping_set_error(inode->i_mapping, -EIO); 302 303 while (nr_pages > 0) { 304 ret = find_get_pages_contig(inode->i_mapping, index, 305 min_t(unsigned long, 306 nr_pages, ARRAY_SIZE(pages)), pages); 307 if (ret == 0) { 308 nr_pages -= 1; 309 index += 1; 310 continue; 311 } 312 for (i = 0; i < ret; i++) { 313 if (cb->errors) 314 SetPageError(pages[i]); 315 end_page_writeback(pages[i]); 316 put_page(pages[i]); 317 } 318 nr_pages -= ret; 319 index += ret; 320 } 321 /* the inode may be gone now */ 322 } 323 324 /* 325 * do the cleanup once all the compressed pages hit the disk. 326 * This will clear writeback on the file pages and free the compressed 327 * pages. 328 * 329 * This also calls the writeback end hooks for the file pages so that 330 * metadata and checksums can be updated in the file. 331 */ 332 static void end_compressed_bio_write(struct bio *bio) 333 { 334 struct compressed_bio *cb = bio->bi_private; 335 struct inode *inode; 336 struct page *page; 337 unsigned int index; 338 339 if (bio->bi_status) 340 cb->errors = 1; 341 342 /* if there are more bios still pending for this compressed 343 * extent, just exit 344 */ 345 if (!refcount_dec_and_test(&cb->pending_bios)) 346 goto out; 347 348 /* ok, we're the last bio for this extent, step one is to 349 * call back into the FS and do all the end_io operations 350 */ 351 inode = cb->inode; 352 btrfs_record_physical_zoned(inode, cb->start, bio); 353 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, 354 cb->start, cb->start + cb->len - 1, 355 !cb->errors); 356 357 end_compressed_writeback(inode, cb); 358 /* note, our inode could be gone now */ 359 360 /* 361 * release the compressed pages, these came from alloc_page and 362 * are not attached to the inode at all 363 */ 364 index = 0; 365 for (index = 0; index < cb->nr_pages; index++) { 366 page = cb->compressed_pages[index]; 367 page->mapping = NULL; 368 put_page(page); 369 } 370 371 /* finally free the cb struct */ 372 kfree(cb->compressed_pages); 373 kfree(cb); 374 out: 375 bio_put(bio); 376 } 377 378 /* 379 * worker function to build and submit bios for previously compressed pages. 380 * The corresponding pages in the inode should be marked for writeback 381 * and the compressed pages should have a reference on them for dropping 382 * when the IO is complete. 383 * 384 * This also checksums the file bytes and gets things ready for 385 * the end io hooks. 386 */ 387 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, 388 unsigned int len, u64 disk_start, 389 unsigned int compressed_len, 390 struct page **compressed_pages, 391 unsigned int nr_pages, 392 unsigned int write_flags, 393 struct cgroup_subsys_state *blkcg_css) 394 { 395 struct btrfs_fs_info *fs_info = inode->root->fs_info; 396 struct bio *bio = NULL; 397 struct compressed_bio *cb; 398 unsigned long bytes_left; 399 int pg_index = 0; 400 struct page *page; 401 u64 first_byte = disk_start; 402 blk_status_t ret; 403 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; 404 const bool use_append = btrfs_use_zone_append(inode, disk_start); 405 const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; 406 407 WARN_ON(!PAGE_ALIGNED(start)); 408 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); 409 if (!cb) 410 return BLK_STS_RESOURCE; 411 refcount_set(&cb->pending_bios, 0); 412 cb->errors = 0; 413 cb->inode = &inode->vfs_inode; 414 cb->start = start; 415 cb->len = len; 416 cb->mirror_num = 0; 417 cb->compressed_pages = compressed_pages; 418 cb->compressed_len = compressed_len; 419 cb->orig_bio = NULL; 420 cb->nr_pages = nr_pages; 421 422 bio = btrfs_bio_alloc(BIO_MAX_VECS); 423 bio->bi_iter.bi_sector = first_byte >> SECTOR_SHIFT; 424 bio->bi_opf = bio_op | write_flags; 425 bio->bi_private = cb; 426 bio->bi_end_io = end_compressed_bio_write; 427 428 if (use_append) { 429 struct btrfs_device *device; 430 431 device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE); 432 if (IS_ERR(device)) { 433 kfree(cb); 434 bio_put(bio); 435 return BLK_STS_NOTSUPP; 436 } 437 438 bio_set_dev(bio, device->bdev); 439 } 440 441 if (blkcg_css) { 442 bio->bi_opf |= REQ_CGROUP_PUNT; 443 kthread_associate_blkcg(blkcg_css); 444 } 445 refcount_set(&cb->pending_bios, 1); 446 447 /* create and submit bios for the compressed pages */ 448 bytes_left = compressed_len; 449 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { 450 int submit = 0; 451 int len = 0; 452 453 page = compressed_pages[pg_index]; 454 page->mapping = inode->vfs_inode.i_mapping; 455 if (bio->bi_iter.bi_size) 456 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, 457 0); 458 459 /* 460 * Page can only be added to bio if the current bio fits in 461 * stripe. 462 */ 463 if (!submit) { 464 if (pg_index == 0 && use_append) 465 len = bio_add_zone_append_page(bio, page, 466 PAGE_SIZE, 0); 467 else 468 len = bio_add_page(bio, page, PAGE_SIZE, 0); 469 } 470 471 page->mapping = NULL; 472 if (submit || len < PAGE_SIZE) { 473 /* 474 * inc the count before we submit the bio so 475 * we know the end IO handler won't happen before 476 * we inc the count. Otherwise, the cb might get 477 * freed before we're done setting it up 478 */ 479 refcount_inc(&cb->pending_bios); 480 ret = btrfs_bio_wq_end_io(fs_info, bio, 481 BTRFS_WQ_ENDIO_DATA); 482 BUG_ON(ret); /* -ENOMEM */ 483 484 if (!skip_sum) { 485 ret = btrfs_csum_one_bio(inode, bio, start, 1); 486 BUG_ON(ret); /* -ENOMEM */ 487 } 488 489 ret = btrfs_map_bio(fs_info, bio, 0); 490 if (ret) { 491 bio->bi_status = ret; 492 bio_endio(bio); 493 } 494 495 bio = btrfs_bio_alloc(BIO_MAX_VECS); 496 bio->bi_iter.bi_sector = first_byte >> SECTOR_SHIFT; 497 bio->bi_opf = bio_op | write_flags; 498 bio->bi_private = cb; 499 bio->bi_end_io = end_compressed_bio_write; 500 if (blkcg_css) 501 bio->bi_opf |= REQ_CGROUP_PUNT; 502 /* 503 * Use bio_add_page() to ensure the bio has at least one 504 * page. 505 */ 506 bio_add_page(bio, page, PAGE_SIZE, 0); 507 } 508 if (bytes_left < PAGE_SIZE) { 509 btrfs_info(fs_info, 510 "bytes left %lu compress len %u nr %u", 511 bytes_left, cb->compressed_len, cb->nr_pages); 512 } 513 bytes_left -= PAGE_SIZE; 514 first_byte += PAGE_SIZE; 515 cond_resched(); 516 } 517 518 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); 519 BUG_ON(ret); /* -ENOMEM */ 520 521 if (!skip_sum) { 522 ret = btrfs_csum_one_bio(inode, bio, start, 1); 523 BUG_ON(ret); /* -ENOMEM */ 524 } 525 526 ret = btrfs_map_bio(fs_info, bio, 0); 527 if (ret) { 528 bio->bi_status = ret; 529 bio_endio(bio); 530 } 531 532 if (blkcg_css) 533 kthread_associate_blkcg(NULL); 534 535 return 0; 536 } 537 538 static u64 bio_end_offset(struct bio *bio) 539 { 540 struct bio_vec *last = bio_last_bvec_all(bio); 541 542 return page_offset(last->bv_page) + last->bv_len + last->bv_offset; 543 } 544 545 /* 546 * Add extra pages in the same compressed file extent so that we don't need to 547 * re-read the same extent again and again. 548 * 549 * NOTE: this won't work well for subpage, as for subpage read, we lock the 550 * full page then submit bio for each compressed/regular extents. 551 * 552 * This means, if we have several sectors in the same page points to the same 553 * on-disk compressed data, we will re-read the same extent many times and 554 * this function can only help for the next page. 555 */ 556 static noinline int add_ra_bio_pages(struct inode *inode, 557 u64 compressed_end, 558 struct compressed_bio *cb) 559 { 560 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 561 unsigned long end_index; 562 u64 cur = bio_end_offset(cb->orig_bio); 563 u64 isize = i_size_read(inode); 564 int ret; 565 struct page *page; 566 struct extent_map *em; 567 struct address_space *mapping = inode->i_mapping; 568 struct extent_map_tree *em_tree; 569 struct extent_io_tree *tree; 570 int sectors_missed = 0; 571 572 em_tree = &BTRFS_I(inode)->extent_tree; 573 tree = &BTRFS_I(inode)->io_tree; 574 575 if (isize == 0) 576 return 0; 577 578 /* 579 * For current subpage support, we only support 64K page size, 580 * which means maximum compressed extent size (128K) is just 2x page 581 * size. 582 * This makes readahead less effective, so here disable readahead for 583 * subpage for now, until full compressed write is supported. 584 */ 585 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) 586 return 0; 587 588 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; 589 590 while (cur < compressed_end) { 591 u64 page_end; 592 u64 pg_index = cur >> PAGE_SHIFT; 593 u32 add_size; 594 595 if (pg_index > end_index) 596 break; 597 598 page = xa_load(&mapping->i_pages, pg_index); 599 if (page && !xa_is_value(page)) { 600 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> 601 fs_info->sectorsize_bits; 602 603 /* Beyond threshold, no need to continue */ 604 if (sectors_missed > 4) 605 break; 606 607 /* 608 * Jump to next page start as we already have page for 609 * current offset. 610 */ 611 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 612 continue; 613 } 614 615 page = __page_cache_alloc(mapping_gfp_constraint(mapping, 616 ~__GFP_FS)); 617 if (!page) 618 break; 619 620 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { 621 put_page(page); 622 /* There is already a page, skip to page end */ 623 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 624 continue; 625 } 626 627 ret = set_page_extent_mapped(page); 628 if (ret < 0) { 629 unlock_page(page); 630 put_page(page); 631 break; 632 } 633 634 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; 635 lock_extent(tree, cur, page_end); 636 read_lock(&em_tree->lock); 637 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); 638 read_unlock(&em_tree->lock); 639 640 /* 641 * At this point, we have a locked page in the page cache for 642 * these bytes in the file. But, we have to make sure they map 643 * to this compressed extent on disk. 644 */ 645 if (!em || cur < em->start || 646 (cur + fs_info->sectorsize > extent_map_end(em)) || 647 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { 648 free_extent_map(em); 649 unlock_extent(tree, cur, page_end); 650 unlock_page(page); 651 put_page(page); 652 break; 653 } 654 free_extent_map(em); 655 656 if (page->index == end_index) { 657 size_t zero_offset = offset_in_page(isize); 658 659 if (zero_offset) { 660 int zeros; 661 zeros = PAGE_SIZE - zero_offset; 662 memzero_page(page, zero_offset, zeros); 663 flush_dcache_page(page); 664 } 665 } 666 667 add_size = min(em->start + em->len, page_end + 1) - cur; 668 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur)); 669 if (ret != add_size) { 670 unlock_extent(tree, cur, page_end); 671 unlock_page(page); 672 put_page(page); 673 break; 674 } 675 /* 676 * If it's subpage, we also need to increase its 677 * subpage::readers number, as at endio we will decrease 678 * subpage::readers and to unlock the page. 679 */ 680 if (fs_info->sectorsize < PAGE_SIZE) 681 btrfs_subpage_start_reader(fs_info, page, cur, add_size); 682 put_page(page); 683 cur += add_size; 684 } 685 return 0; 686 } 687 688 /* 689 * for a compressed read, the bio we get passed has all the inode pages 690 * in it. We don't actually do IO on those pages but allocate new ones 691 * to hold the compressed pages on disk. 692 * 693 * bio->bi_iter.bi_sector points to the compressed extent on disk 694 * bio->bi_io_vec points to all of the inode pages 695 * 696 * After the compressed pages are read, we copy the bytes into the 697 * bio we were passed and then call the bio end_io calls 698 */ 699 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, 700 int mirror_num, unsigned long bio_flags) 701 { 702 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 703 struct extent_map_tree *em_tree; 704 struct compressed_bio *cb; 705 unsigned int compressed_len; 706 unsigned int nr_pages; 707 unsigned int pg_index; 708 struct page *page; 709 struct bio *comp_bio; 710 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9; 711 u64 file_offset; 712 u64 em_len; 713 u64 em_start; 714 struct extent_map *em; 715 blk_status_t ret = BLK_STS_RESOURCE; 716 int faili = 0; 717 u8 *sums; 718 719 em_tree = &BTRFS_I(inode)->extent_tree; 720 721 file_offset = bio_first_bvec_all(bio)->bv_offset + 722 page_offset(bio_first_page_all(bio)); 723 724 /* we need the actual starting offset of this extent in the file */ 725 read_lock(&em_tree->lock); 726 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); 727 read_unlock(&em_tree->lock); 728 if (!em) 729 return BLK_STS_IOERR; 730 731 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); 732 compressed_len = em->block_len; 733 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); 734 if (!cb) 735 goto out; 736 737 refcount_set(&cb->pending_bios, 0); 738 cb->errors = 0; 739 cb->inode = inode; 740 cb->mirror_num = mirror_num; 741 sums = cb->sums; 742 743 cb->start = em->orig_start; 744 em_len = em->len; 745 em_start = em->start; 746 747 free_extent_map(em); 748 em = NULL; 749 750 cb->len = bio->bi_iter.bi_size; 751 cb->compressed_len = compressed_len; 752 cb->compress_type = extent_compress_type(bio_flags); 753 cb->orig_bio = bio; 754 755 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); 756 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), 757 GFP_NOFS); 758 if (!cb->compressed_pages) 759 goto fail1; 760 761 for (pg_index = 0; pg_index < nr_pages; pg_index++) { 762 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS); 763 if (!cb->compressed_pages[pg_index]) { 764 faili = pg_index - 1; 765 ret = BLK_STS_RESOURCE; 766 goto fail2; 767 } 768 } 769 faili = nr_pages - 1; 770 cb->nr_pages = nr_pages; 771 772 add_ra_bio_pages(inode, em_start + em_len, cb); 773 774 /* include any pages we added in add_ra-bio_pages */ 775 cb->len = bio->bi_iter.bi_size; 776 777 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS); 778 comp_bio->bi_iter.bi_sector = cur_disk_byte >> SECTOR_SHIFT; 779 comp_bio->bi_opf = REQ_OP_READ; 780 comp_bio->bi_private = cb; 781 comp_bio->bi_end_io = end_compressed_bio_read; 782 refcount_set(&cb->pending_bios, 1); 783 784 for (pg_index = 0; pg_index < nr_pages; pg_index++) { 785 u32 pg_len = PAGE_SIZE; 786 int submit = 0; 787 788 /* 789 * To handle subpage case, we need to make sure the bio only 790 * covers the range we need. 791 * 792 * If we're at the last page, truncate the length to only cover 793 * the remaining part. 794 */ 795 if (pg_index == nr_pages - 1) 796 pg_len = min_t(u32, PAGE_SIZE, 797 compressed_len - pg_index * PAGE_SIZE); 798 799 page = cb->compressed_pages[pg_index]; 800 page->mapping = inode->i_mapping; 801 page->index = em_start >> PAGE_SHIFT; 802 803 if (comp_bio->bi_iter.bi_size) 804 submit = btrfs_bio_fits_in_stripe(page, pg_len, 805 comp_bio, 0); 806 807 page->mapping = NULL; 808 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) { 809 unsigned int nr_sectors; 810 811 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, 812 BTRFS_WQ_ENDIO_DATA); 813 BUG_ON(ret); /* -ENOMEM */ 814 815 /* 816 * inc the count before we submit the bio so 817 * we know the end IO handler won't happen before 818 * we inc the count. Otherwise, the cb might get 819 * freed before we're done setting it up 820 */ 821 refcount_inc(&cb->pending_bios); 822 823 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); 824 BUG_ON(ret); /* -ENOMEM */ 825 826 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, 827 fs_info->sectorsize); 828 sums += fs_info->csum_size * nr_sectors; 829 830 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); 831 if (ret) { 832 comp_bio->bi_status = ret; 833 bio_endio(comp_bio); 834 } 835 836 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS); 837 comp_bio->bi_iter.bi_sector = cur_disk_byte >> SECTOR_SHIFT; 838 comp_bio->bi_opf = REQ_OP_READ; 839 comp_bio->bi_private = cb; 840 comp_bio->bi_end_io = end_compressed_bio_read; 841 842 bio_add_page(comp_bio, page, pg_len, 0); 843 } 844 cur_disk_byte += pg_len; 845 } 846 847 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); 848 BUG_ON(ret); /* -ENOMEM */ 849 850 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); 851 BUG_ON(ret); /* -ENOMEM */ 852 853 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); 854 if (ret) { 855 comp_bio->bi_status = ret; 856 bio_endio(comp_bio); 857 } 858 859 return 0; 860 861 fail2: 862 while (faili >= 0) { 863 __free_page(cb->compressed_pages[faili]); 864 faili--; 865 } 866 867 kfree(cb->compressed_pages); 868 fail1: 869 kfree(cb); 870 out: 871 free_extent_map(em); 872 return ret; 873 } 874 875 /* 876 * Heuristic uses systematic sampling to collect data from the input data 877 * range, the logic can be tuned by the following constants: 878 * 879 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample 880 * @SAMPLING_INTERVAL - range from which the sampled data can be collected 881 */ 882 #define SAMPLING_READ_SIZE (16) 883 #define SAMPLING_INTERVAL (256) 884 885 /* 886 * For statistical analysis of the input data we consider bytes that form a 887 * Galois Field of 256 objects. Each object has an attribute count, ie. how 888 * many times the object appeared in the sample. 889 */ 890 #define BUCKET_SIZE (256) 891 892 /* 893 * The size of the sample is based on a statistical sampling rule of thumb. 894 * The common way is to perform sampling tests as long as the number of 895 * elements in each cell is at least 5. 896 * 897 * Instead of 5, we choose 32 to obtain more accurate results. 898 * If the data contain the maximum number of symbols, which is 256, we obtain a 899 * sample size bound by 8192. 900 * 901 * For a sample of at most 8KB of data per data range: 16 consecutive bytes 902 * from up to 512 locations. 903 */ 904 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ 905 SAMPLING_READ_SIZE / SAMPLING_INTERVAL) 906 907 struct bucket_item { 908 u32 count; 909 }; 910 911 struct heuristic_ws { 912 /* Partial copy of input data */ 913 u8 *sample; 914 u32 sample_size; 915 /* Buckets store counters for each byte value */ 916 struct bucket_item *bucket; 917 /* Sorting buffer */ 918 struct bucket_item *bucket_b; 919 struct list_head list; 920 }; 921 922 static struct workspace_manager heuristic_wsm; 923 924 static void free_heuristic_ws(struct list_head *ws) 925 { 926 struct heuristic_ws *workspace; 927 928 workspace = list_entry(ws, struct heuristic_ws, list); 929 930 kvfree(workspace->sample); 931 kfree(workspace->bucket); 932 kfree(workspace->bucket_b); 933 kfree(workspace); 934 } 935 936 static struct list_head *alloc_heuristic_ws(unsigned int level) 937 { 938 struct heuristic_ws *ws; 939 940 ws = kzalloc(sizeof(*ws), GFP_KERNEL); 941 if (!ws) 942 return ERR_PTR(-ENOMEM); 943 944 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); 945 if (!ws->sample) 946 goto fail; 947 948 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); 949 if (!ws->bucket) 950 goto fail; 951 952 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); 953 if (!ws->bucket_b) 954 goto fail; 955 956 INIT_LIST_HEAD(&ws->list); 957 return &ws->list; 958 fail: 959 free_heuristic_ws(&ws->list); 960 return ERR_PTR(-ENOMEM); 961 } 962 963 const struct btrfs_compress_op btrfs_heuristic_compress = { 964 .workspace_manager = &heuristic_wsm, 965 }; 966 967 static const struct btrfs_compress_op * const btrfs_compress_op[] = { 968 /* The heuristic is represented as compression type 0 */ 969 &btrfs_heuristic_compress, 970 &btrfs_zlib_compress, 971 &btrfs_lzo_compress, 972 &btrfs_zstd_compress, 973 }; 974 975 static struct list_head *alloc_workspace(int type, unsigned int level) 976 { 977 switch (type) { 978 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); 979 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); 980 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); 981 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); 982 default: 983 /* 984 * This can't happen, the type is validated several times 985 * before we get here. 986 */ 987 BUG(); 988 } 989 } 990 991 static void free_workspace(int type, struct list_head *ws) 992 { 993 switch (type) { 994 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); 995 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); 996 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); 997 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); 998 default: 999 /* 1000 * This can't happen, the type is validated several times 1001 * before we get here. 1002 */ 1003 BUG(); 1004 } 1005 } 1006 1007 static void btrfs_init_workspace_manager(int type) 1008 { 1009 struct workspace_manager *wsm; 1010 struct list_head *workspace; 1011 1012 wsm = btrfs_compress_op[type]->workspace_manager; 1013 INIT_LIST_HEAD(&wsm->idle_ws); 1014 spin_lock_init(&wsm->ws_lock); 1015 atomic_set(&wsm->total_ws, 0); 1016 init_waitqueue_head(&wsm->ws_wait); 1017 1018 /* 1019 * Preallocate one workspace for each compression type so we can 1020 * guarantee forward progress in the worst case 1021 */ 1022 workspace = alloc_workspace(type, 0); 1023 if (IS_ERR(workspace)) { 1024 pr_warn( 1025 "BTRFS: cannot preallocate compression workspace, will try later\n"); 1026 } else { 1027 atomic_set(&wsm->total_ws, 1); 1028 wsm->free_ws = 1; 1029 list_add(workspace, &wsm->idle_ws); 1030 } 1031 } 1032 1033 static void btrfs_cleanup_workspace_manager(int type) 1034 { 1035 struct workspace_manager *wsman; 1036 struct list_head *ws; 1037 1038 wsman = btrfs_compress_op[type]->workspace_manager; 1039 while (!list_empty(&wsman->idle_ws)) { 1040 ws = wsman->idle_ws.next; 1041 list_del(ws); 1042 free_workspace(type, ws); 1043 atomic_dec(&wsman->total_ws); 1044 } 1045 } 1046 1047 /* 1048 * This finds an available workspace or allocates a new one. 1049 * If it's not possible to allocate a new one, waits until there's one. 1050 * Preallocation makes a forward progress guarantees and we do not return 1051 * errors. 1052 */ 1053 struct list_head *btrfs_get_workspace(int type, unsigned int level) 1054 { 1055 struct workspace_manager *wsm; 1056 struct list_head *workspace; 1057 int cpus = num_online_cpus(); 1058 unsigned nofs_flag; 1059 struct list_head *idle_ws; 1060 spinlock_t *ws_lock; 1061 atomic_t *total_ws; 1062 wait_queue_head_t *ws_wait; 1063 int *free_ws; 1064 1065 wsm = btrfs_compress_op[type]->workspace_manager; 1066 idle_ws = &wsm->idle_ws; 1067 ws_lock = &wsm->ws_lock; 1068 total_ws = &wsm->total_ws; 1069 ws_wait = &wsm->ws_wait; 1070 free_ws = &wsm->free_ws; 1071 1072 again: 1073 spin_lock(ws_lock); 1074 if (!list_empty(idle_ws)) { 1075 workspace = idle_ws->next; 1076 list_del(workspace); 1077 (*free_ws)--; 1078 spin_unlock(ws_lock); 1079 return workspace; 1080 1081 } 1082 if (atomic_read(total_ws) > cpus) { 1083 DEFINE_WAIT(wait); 1084 1085 spin_unlock(ws_lock); 1086 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); 1087 if (atomic_read(total_ws) > cpus && !*free_ws) 1088 schedule(); 1089 finish_wait(ws_wait, &wait); 1090 goto again; 1091 } 1092 atomic_inc(total_ws); 1093 spin_unlock(ws_lock); 1094 1095 /* 1096 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have 1097 * to turn it off here because we might get called from the restricted 1098 * context of btrfs_compress_bio/btrfs_compress_pages 1099 */ 1100 nofs_flag = memalloc_nofs_save(); 1101 workspace = alloc_workspace(type, level); 1102 memalloc_nofs_restore(nofs_flag); 1103 1104 if (IS_ERR(workspace)) { 1105 atomic_dec(total_ws); 1106 wake_up(ws_wait); 1107 1108 /* 1109 * Do not return the error but go back to waiting. There's a 1110 * workspace preallocated for each type and the compression 1111 * time is bounded so we get to a workspace eventually. This 1112 * makes our caller's life easier. 1113 * 1114 * To prevent silent and low-probability deadlocks (when the 1115 * initial preallocation fails), check if there are any 1116 * workspaces at all. 1117 */ 1118 if (atomic_read(total_ws) == 0) { 1119 static DEFINE_RATELIMIT_STATE(_rs, 1120 /* once per minute */ 60 * HZ, 1121 /* no burst */ 1); 1122 1123 if (__ratelimit(&_rs)) { 1124 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); 1125 } 1126 } 1127 goto again; 1128 } 1129 return workspace; 1130 } 1131 1132 static struct list_head *get_workspace(int type, int level) 1133 { 1134 switch (type) { 1135 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); 1136 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); 1137 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); 1138 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); 1139 default: 1140 /* 1141 * This can't happen, the type is validated several times 1142 * before we get here. 1143 */ 1144 BUG(); 1145 } 1146 } 1147 1148 /* 1149 * put a workspace struct back on the list or free it if we have enough 1150 * idle ones sitting around 1151 */ 1152 void btrfs_put_workspace(int type, struct list_head *ws) 1153 { 1154 struct workspace_manager *wsm; 1155 struct list_head *idle_ws; 1156 spinlock_t *ws_lock; 1157 atomic_t *total_ws; 1158 wait_queue_head_t *ws_wait; 1159 int *free_ws; 1160 1161 wsm = btrfs_compress_op[type]->workspace_manager; 1162 idle_ws = &wsm->idle_ws; 1163 ws_lock = &wsm->ws_lock; 1164 total_ws = &wsm->total_ws; 1165 ws_wait = &wsm->ws_wait; 1166 free_ws = &wsm->free_ws; 1167 1168 spin_lock(ws_lock); 1169 if (*free_ws <= num_online_cpus()) { 1170 list_add(ws, idle_ws); 1171 (*free_ws)++; 1172 spin_unlock(ws_lock); 1173 goto wake; 1174 } 1175 spin_unlock(ws_lock); 1176 1177 free_workspace(type, ws); 1178 atomic_dec(total_ws); 1179 wake: 1180 cond_wake_up(ws_wait); 1181 } 1182 1183 static void put_workspace(int type, struct list_head *ws) 1184 { 1185 switch (type) { 1186 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); 1187 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); 1188 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); 1189 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); 1190 default: 1191 /* 1192 * This can't happen, the type is validated several times 1193 * before we get here. 1194 */ 1195 BUG(); 1196 } 1197 } 1198 1199 /* 1200 * Adjust @level according to the limits of the compression algorithm or 1201 * fallback to default 1202 */ 1203 static unsigned int btrfs_compress_set_level(int type, unsigned level) 1204 { 1205 const struct btrfs_compress_op *ops = btrfs_compress_op[type]; 1206 1207 if (level == 0) 1208 level = ops->default_level; 1209 else 1210 level = min(level, ops->max_level); 1211 1212 return level; 1213 } 1214 1215 /* 1216 * Given an address space and start and length, compress the bytes into @pages 1217 * that are allocated on demand. 1218 * 1219 * @type_level is encoded algorithm and level, where level 0 means whatever 1220 * default the algorithm chooses and is opaque here; 1221 * - compression algo are 0-3 1222 * - the level are bits 4-7 1223 * 1224 * @out_pages is an in/out parameter, holds maximum number of pages to allocate 1225 * and returns number of actually allocated pages 1226 * 1227 * @total_in is used to return the number of bytes actually read. It 1228 * may be smaller than the input length if we had to exit early because we 1229 * ran out of room in the pages array or because we cross the 1230 * max_out threshold. 1231 * 1232 * @total_out is an in/out parameter, must be set to the input length and will 1233 * be also used to return the total number of compressed bytes 1234 */ 1235 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, 1236 u64 start, struct page **pages, 1237 unsigned long *out_pages, 1238 unsigned long *total_in, 1239 unsigned long *total_out) 1240 { 1241 int type = btrfs_compress_type(type_level); 1242 int level = btrfs_compress_level(type_level); 1243 struct list_head *workspace; 1244 int ret; 1245 1246 level = btrfs_compress_set_level(type, level); 1247 workspace = get_workspace(type, level); 1248 ret = compression_compress_pages(type, workspace, mapping, start, pages, 1249 out_pages, total_in, total_out); 1250 put_workspace(type, workspace); 1251 return ret; 1252 } 1253 1254 static int btrfs_decompress_bio(struct compressed_bio *cb) 1255 { 1256 struct list_head *workspace; 1257 int ret; 1258 int type = cb->compress_type; 1259 1260 workspace = get_workspace(type, 0); 1261 ret = compression_decompress_bio(type, workspace, cb); 1262 put_workspace(type, workspace); 1263 1264 return ret; 1265 } 1266 1267 /* 1268 * a less complex decompression routine. Our compressed data fits in a 1269 * single page, and we want to read a single page out of it. 1270 * start_byte tells us the offset into the compressed data we're interested in 1271 */ 1272 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, 1273 unsigned long start_byte, size_t srclen, size_t destlen) 1274 { 1275 struct list_head *workspace; 1276 int ret; 1277 1278 workspace = get_workspace(type, 0); 1279 ret = compression_decompress(type, workspace, data_in, dest_page, 1280 start_byte, srclen, destlen); 1281 put_workspace(type, workspace); 1282 1283 return ret; 1284 } 1285 1286 void __init btrfs_init_compress(void) 1287 { 1288 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); 1289 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); 1290 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); 1291 zstd_init_workspace_manager(); 1292 } 1293 1294 void __cold btrfs_exit_compress(void) 1295 { 1296 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); 1297 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); 1298 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); 1299 zstd_cleanup_workspace_manager(); 1300 } 1301 1302 /* 1303 * Copy decompressed data from working buffer to pages. 1304 * 1305 * @buf: The decompressed data buffer 1306 * @buf_len: The decompressed data length 1307 * @decompressed: Number of bytes that are already decompressed inside the 1308 * compressed extent 1309 * @cb: The compressed extent descriptor 1310 * @orig_bio: The original bio that the caller wants to read for 1311 * 1312 * An easier to understand graph is like below: 1313 * 1314 * |<- orig_bio ->| |<- orig_bio->| 1315 * |<------- full decompressed extent ----->| 1316 * |<----------- @cb range ---->| 1317 * | |<-- @buf_len -->| 1318 * |<--- @decompressed --->| 1319 * 1320 * Note that, @cb can be a subpage of the full decompressed extent, but 1321 * @cb->start always has the same as the orig_file_offset value of the full 1322 * decompressed extent. 1323 * 1324 * When reading compressed extent, we have to read the full compressed extent, 1325 * while @orig_bio may only want part of the range. 1326 * Thus this function will ensure only data covered by @orig_bio will be copied 1327 * to. 1328 * 1329 * Return 0 if we have copied all needed contents for @orig_bio. 1330 * Return >0 if we need continue decompress. 1331 */ 1332 int btrfs_decompress_buf2page(const char *buf, u32 buf_len, 1333 struct compressed_bio *cb, u32 decompressed) 1334 { 1335 struct bio *orig_bio = cb->orig_bio; 1336 /* Offset inside the full decompressed extent */ 1337 u32 cur_offset; 1338 1339 cur_offset = decompressed; 1340 /* The main loop to do the copy */ 1341 while (cur_offset < decompressed + buf_len) { 1342 struct bio_vec bvec; 1343 size_t copy_len; 1344 u32 copy_start; 1345 /* Offset inside the full decompressed extent */ 1346 u32 bvec_offset; 1347 1348 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); 1349 /* 1350 * cb->start may underflow, but subtracting that value can still 1351 * give us correct offset inside the full decompressed extent. 1352 */ 1353 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; 1354 1355 /* Haven't reached the bvec range, exit */ 1356 if (decompressed + buf_len <= bvec_offset) 1357 return 1; 1358 1359 copy_start = max(cur_offset, bvec_offset); 1360 copy_len = min(bvec_offset + bvec.bv_len, 1361 decompressed + buf_len) - copy_start; 1362 ASSERT(copy_len); 1363 1364 /* 1365 * Extra range check to ensure we didn't go beyond 1366 * @buf + @buf_len. 1367 */ 1368 ASSERT(copy_start - decompressed < buf_len); 1369 memcpy_to_page(bvec.bv_page, bvec.bv_offset, 1370 buf + copy_start - decompressed, copy_len); 1371 flush_dcache_page(bvec.bv_page); 1372 cur_offset += copy_len; 1373 1374 bio_advance(orig_bio, copy_len); 1375 /* Finished the bio */ 1376 if (!orig_bio->bi_iter.bi_size) 1377 return 0; 1378 } 1379 return 1; 1380 } 1381 1382 /* 1383 * Shannon Entropy calculation 1384 * 1385 * Pure byte distribution analysis fails to determine compressibility of data. 1386 * Try calculating entropy to estimate the average minimum number of bits 1387 * needed to encode the sampled data. 1388 * 1389 * For convenience, return the percentage of needed bits, instead of amount of 1390 * bits directly. 1391 * 1392 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy 1393 * and can be compressible with high probability 1394 * 1395 * @ENTROPY_LVL_HIGH - data are not compressible with high probability 1396 * 1397 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. 1398 */ 1399 #define ENTROPY_LVL_ACEPTABLE (65) 1400 #define ENTROPY_LVL_HIGH (80) 1401 1402 /* 1403 * For increasead precision in shannon_entropy calculation, 1404 * let's do pow(n, M) to save more digits after comma: 1405 * 1406 * - maximum int bit length is 64 1407 * - ilog2(MAX_SAMPLE_SIZE) -> 13 1408 * - 13 * 4 = 52 < 64 -> M = 4 1409 * 1410 * So use pow(n, 4). 1411 */ 1412 static inline u32 ilog2_w(u64 n) 1413 { 1414 return ilog2(n * n * n * n); 1415 } 1416 1417 static u32 shannon_entropy(struct heuristic_ws *ws) 1418 { 1419 const u32 entropy_max = 8 * ilog2_w(2); 1420 u32 entropy_sum = 0; 1421 u32 p, p_base, sz_base; 1422 u32 i; 1423 1424 sz_base = ilog2_w(ws->sample_size); 1425 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { 1426 p = ws->bucket[i].count; 1427 p_base = ilog2_w(p); 1428 entropy_sum += p * (sz_base - p_base); 1429 } 1430 1431 entropy_sum /= ws->sample_size; 1432 return entropy_sum * 100 / entropy_max; 1433 } 1434 1435 #define RADIX_BASE 4U 1436 #define COUNTERS_SIZE (1U << RADIX_BASE) 1437 1438 static u8 get4bits(u64 num, int shift) { 1439 u8 low4bits; 1440 1441 num >>= shift; 1442 /* Reverse order */ 1443 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); 1444 return low4bits; 1445 } 1446 1447 /* 1448 * Use 4 bits as radix base 1449 * Use 16 u32 counters for calculating new position in buf array 1450 * 1451 * @array - array that will be sorted 1452 * @array_buf - buffer array to store sorting results 1453 * must be equal in size to @array 1454 * @num - array size 1455 */ 1456 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, 1457 int num) 1458 { 1459 u64 max_num; 1460 u64 buf_num; 1461 u32 counters[COUNTERS_SIZE]; 1462 u32 new_addr; 1463 u32 addr; 1464 int bitlen; 1465 int shift; 1466 int i; 1467 1468 /* 1469 * Try avoid useless loop iterations for small numbers stored in big 1470 * counters. Example: 48 33 4 ... in 64bit array 1471 */ 1472 max_num = array[0].count; 1473 for (i = 1; i < num; i++) { 1474 buf_num = array[i].count; 1475 if (buf_num > max_num) 1476 max_num = buf_num; 1477 } 1478 1479 buf_num = ilog2(max_num); 1480 bitlen = ALIGN(buf_num, RADIX_BASE * 2); 1481 1482 shift = 0; 1483 while (shift < bitlen) { 1484 memset(counters, 0, sizeof(counters)); 1485 1486 for (i = 0; i < num; i++) { 1487 buf_num = array[i].count; 1488 addr = get4bits(buf_num, shift); 1489 counters[addr]++; 1490 } 1491 1492 for (i = 1; i < COUNTERS_SIZE; i++) 1493 counters[i] += counters[i - 1]; 1494 1495 for (i = num - 1; i >= 0; i--) { 1496 buf_num = array[i].count; 1497 addr = get4bits(buf_num, shift); 1498 counters[addr]--; 1499 new_addr = counters[addr]; 1500 array_buf[new_addr] = array[i]; 1501 } 1502 1503 shift += RADIX_BASE; 1504 1505 /* 1506 * Normal radix expects to move data from a temporary array, to 1507 * the main one. But that requires some CPU time. Avoid that 1508 * by doing another sort iteration to original array instead of 1509 * memcpy() 1510 */ 1511 memset(counters, 0, sizeof(counters)); 1512 1513 for (i = 0; i < num; i ++) { 1514 buf_num = array_buf[i].count; 1515 addr = get4bits(buf_num, shift); 1516 counters[addr]++; 1517 } 1518 1519 for (i = 1; i < COUNTERS_SIZE; i++) 1520 counters[i] += counters[i - 1]; 1521 1522 for (i = num - 1; i >= 0; i--) { 1523 buf_num = array_buf[i].count; 1524 addr = get4bits(buf_num, shift); 1525 counters[addr]--; 1526 new_addr = counters[addr]; 1527 array[new_addr] = array_buf[i]; 1528 } 1529 1530 shift += RADIX_BASE; 1531 } 1532 } 1533 1534 /* 1535 * Size of the core byte set - how many bytes cover 90% of the sample 1536 * 1537 * There are several types of structured binary data that use nearly all byte 1538 * values. The distribution can be uniform and counts in all buckets will be 1539 * nearly the same (eg. encrypted data). Unlikely to be compressible. 1540 * 1541 * Other possibility is normal (Gaussian) distribution, where the data could 1542 * be potentially compressible, but we have to take a few more steps to decide 1543 * how much. 1544 * 1545 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, 1546 * compression algo can easy fix that 1547 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high 1548 * probability is not compressible 1549 */ 1550 #define BYTE_CORE_SET_LOW (64) 1551 #define BYTE_CORE_SET_HIGH (200) 1552 1553 static int byte_core_set_size(struct heuristic_ws *ws) 1554 { 1555 u32 i; 1556 u32 coreset_sum = 0; 1557 const u32 core_set_threshold = ws->sample_size * 90 / 100; 1558 struct bucket_item *bucket = ws->bucket; 1559 1560 /* Sort in reverse order */ 1561 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); 1562 1563 for (i = 0; i < BYTE_CORE_SET_LOW; i++) 1564 coreset_sum += bucket[i].count; 1565 1566 if (coreset_sum > core_set_threshold) 1567 return i; 1568 1569 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { 1570 coreset_sum += bucket[i].count; 1571 if (coreset_sum > core_set_threshold) 1572 break; 1573 } 1574 1575 return i; 1576 } 1577 1578 /* 1579 * Count byte values in buckets. 1580 * This heuristic can detect textual data (configs, xml, json, html, etc). 1581 * Because in most text-like data byte set is restricted to limited number of 1582 * possible characters, and that restriction in most cases makes data easy to 1583 * compress. 1584 * 1585 * @BYTE_SET_THRESHOLD - consider all data within this byte set size: 1586 * less - compressible 1587 * more - need additional analysis 1588 */ 1589 #define BYTE_SET_THRESHOLD (64) 1590 1591 static u32 byte_set_size(const struct heuristic_ws *ws) 1592 { 1593 u32 i; 1594 u32 byte_set_size = 0; 1595 1596 for (i = 0; i < BYTE_SET_THRESHOLD; i++) { 1597 if (ws->bucket[i].count > 0) 1598 byte_set_size++; 1599 } 1600 1601 /* 1602 * Continue collecting count of byte values in buckets. If the byte 1603 * set size is bigger then the threshold, it's pointless to continue, 1604 * the detection technique would fail for this type of data. 1605 */ 1606 for (; i < BUCKET_SIZE; i++) { 1607 if (ws->bucket[i].count > 0) { 1608 byte_set_size++; 1609 if (byte_set_size > BYTE_SET_THRESHOLD) 1610 return byte_set_size; 1611 } 1612 } 1613 1614 return byte_set_size; 1615 } 1616 1617 static bool sample_repeated_patterns(struct heuristic_ws *ws) 1618 { 1619 const u32 half_of_sample = ws->sample_size / 2; 1620 const u8 *data = ws->sample; 1621 1622 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; 1623 } 1624 1625 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, 1626 struct heuristic_ws *ws) 1627 { 1628 struct page *page; 1629 u64 index, index_end; 1630 u32 i, curr_sample_pos; 1631 u8 *in_data; 1632 1633 /* 1634 * Compression handles the input data by chunks of 128KiB 1635 * (defined by BTRFS_MAX_UNCOMPRESSED) 1636 * 1637 * We do the same for the heuristic and loop over the whole range. 1638 * 1639 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will 1640 * process no more than BTRFS_MAX_UNCOMPRESSED at a time. 1641 */ 1642 if (end - start > BTRFS_MAX_UNCOMPRESSED) 1643 end = start + BTRFS_MAX_UNCOMPRESSED; 1644 1645 index = start >> PAGE_SHIFT; 1646 index_end = end >> PAGE_SHIFT; 1647 1648 /* Don't miss unaligned end */ 1649 if (!IS_ALIGNED(end, PAGE_SIZE)) 1650 index_end++; 1651 1652 curr_sample_pos = 0; 1653 while (index < index_end) { 1654 page = find_get_page(inode->i_mapping, index); 1655 in_data = kmap_local_page(page); 1656 /* Handle case where the start is not aligned to PAGE_SIZE */ 1657 i = start % PAGE_SIZE; 1658 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { 1659 /* Don't sample any garbage from the last page */ 1660 if (start > end - SAMPLING_READ_SIZE) 1661 break; 1662 memcpy(&ws->sample[curr_sample_pos], &in_data[i], 1663 SAMPLING_READ_SIZE); 1664 i += SAMPLING_INTERVAL; 1665 start += SAMPLING_INTERVAL; 1666 curr_sample_pos += SAMPLING_READ_SIZE; 1667 } 1668 kunmap_local(in_data); 1669 put_page(page); 1670 1671 index++; 1672 } 1673 1674 ws->sample_size = curr_sample_pos; 1675 } 1676 1677 /* 1678 * Compression heuristic. 1679 * 1680 * For now is's a naive and optimistic 'return true', we'll extend the logic to 1681 * quickly (compared to direct compression) detect data characteristics 1682 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible 1683 * data. 1684 * 1685 * The following types of analysis can be performed: 1686 * - detect mostly zero data 1687 * - detect data with low "byte set" size (text, etc) 1688 * - detect data with low/high "core byte" set 1689 * 1690 * Return non-zero if the compression should be done, 0 otherwise. 1691 */ 1692 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) 1693 { 1694 struct list_head *ws_list = get_workspace(0, 0); 1695 struct heuristic_ws *ws; 1696 u32 i; 1697 u8 byte; 1698 int ret = 0; 1699 1700 ws = list_entry(ws_list, struct heuristic_ws, list); 1701 1702 heuristic_collect_sample(inode, start, end, ws); 1703 1704 if (sample_repeated_patterns(ws)) { 1705 ret = 1; 1706 goto out; 1707 } 1708 1709 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); 1710 1711 for (i = 0; i < ws->sample_size; i++) { 1712 byte = ws->sample[i]; 1713 ws->bucket[byte].count++; 1714 } 1715 1716 i = byte_set_size(ws); 1717 if (i < BYTE_SET_THRESHOLD) { 1718 ret = 2; 1719 goto out; 1720 } 1721 1722 i = byte_core_set_size(ws); 1723 if (i <= BYTE_CORE_SET_LOW) { 1724 ret = 3; 1725 goto out; 1726 } 1727 1728 if (i >= BYTE_CORE_SET_HIGH) { 1729 ret = 0; 1730 goto out; 1731 } 1732 1733 i = shannon_entropy(ws); 1734 if (i <= ENTROPY_LVL_ACEPTABLE) { 1735 ret = 4; 1736 goto out; 1737 } 1738 1739 /* 1740 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be 1741 * needed to give green light to compression. 1742 * 1743 * For now just assume that compression at that level is not worth the 1744 * resources because: 1745 * 1746 * 1. it is possible to defrag the data later 1747 * 1748 * 2. the data would turn out to be hardly compressible, eg. 150 byte 1749 * values, every bucket has counter at level ~54. The heuristic would 1750 * be confused. This can happen when data have some internal repeated 1751 * patterns like "abbacbbc...". This can be detected by analyzing 1752 * pairs of bytes, which is too costly. 1753 */ 1754 if (i < ENTROPY_LVL_HIGH) { 1755 ret = 5; 1756 goto out; 1757 } else { 1758 ret = 0; 1759 goto out; 1760 } 1761 1762 out: 1763 put_workspace(0, ws_list); 1764 return ret; 1765 } 1766 1767 /* 1768 * Convert the compression suffix (eg. after "zlib" starting with ":") to 1769 * level, unrecognized string will set the default level 1770 */ 1771 unsigned int btrfs_compress_str2level(unsigned int type, const char *str) 1772 { 1773 unsigned int level = 0; 1774 int ret; 1775 1776 if (!type) 1777 return 0; 1778 1779 if (str[0] == ':') { 1780 ret = kstrtouint(str + 1, 10, &level); 1781 if (ret) 1782 level = 0; 1783 } 1784 1785 level = btrfs_compress_set_level(type, level); 1786 1787 return level; 1788 } 1789