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