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/pagevec.h> 12 #include <linux/highmem.h> 13 #include <linux/kthread.h> 14 #include <linux/time.h> 15 #include <linux/init.h> 16 #include <linux/string.h> 17 #include <linux/backing-dev.h> 18 #include <linux/writeback.h> 19 #include <linux/psi.h> 20 #include <linux/slab.h> 21 #include <linux/sched/mm.h> 22 #include <linux/log2.h> 23 #include <crypto/hash.h> 24 #include "misc.h" 25 #include "ctree.h" 26 #include "fs.h" 27 #include "disk-io.h" 28 #include "transaction.h" 29 #include "btrfs_inode.h" 30 #include "bio.h" 31 #include "ordered-data.h" 32 #include "compression.h" 33 #include "extent_io.h" 34 #include "extent_map.h" 35 #include "subpage.h" 36 #include "zoned.h" 37 #include "file-item.h" 38 #include "super.h" 39 40 struct bio_set btrfs_compressed_bioset; 41 42 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; 43 44 const char* btrfs_compress_type2str(enum btrfs_compression_type type) 45 { 46 switch (type) { 47 case BTRFS_COMPRESS_ZLIB: 48 case BTRFS_COMPRESS_LZO: 49 case BTRFS_COMPRESS_ZSTD: 50 case BTRFS_COMPRESS_NONE: 51 return btrfs_compress_types[type]; 52 default: 53 break; 54 } 55 56 return NULL; 57 } 58 59 static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio) 60 { 61 return container_of(bbio, struct compressed_bio, bbio); 62 } 63 64 static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode, 65 u64 start, blk_opf_t op, 66 btrfs_bio_end_io_t end_io) 67 { 68 struct btrfs_bio *bbio; 69 70 bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op, 71 GFP_NOFS, &btrfs_compressed_bioset)); 72 btrfs_bio_init(bbio, inode, end_io, NULL); 73 bbio->file_offset = start; 74 return to_compressed_bio(bbio); 75 } 76 77 bool btrfs_compress_is_valid_type(const char *str, size_t len) 78 { 79 int i; 80 81 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { 82 size_t comp_len = strlen(btrfs_compress_types[i]); 83 84 if (len < comp_len) 85 continue; 86 87 if (!strncmp(btrfs_compress_types[i], str, comp_len)) 88 return true; 89 } 90 return false; 91 } 92 93 static int compression_compress_pages(int type, struct list_head *ws, 94 struct address_space *mapping, u64 start, struct page **pages, 95 unsigned long *out_pages, unsigned long *total_in, 96 unsigned long *total_out) 97 { 98 switch (type) { 99 case BTRFS_COMPRESS_ZLIB: 100 return zlib_compress_pages(ws, mapping, start, pages, 101 out_pages, total_in, total_out); 102 case BTRFS_COMPRESS_LZO: 103 return lzo_compress_pages(ws, mapping, start, pages, 104 out_pages, total_in, total_out); 105 case BTRFS_COMPRESS_ZSTD: 106 return zstd_compress_pages(ws, mapping, start, pages, 107 out_pages, total_in, total_out); 108 case BTRFS_COMPRESS_NONE: 109 default: 110 /* 111 * This can happen when compression races with remount setting 112 * it to 'no compress', while caller doesn't call 113 * inode_need_compress() to check if we really need to 114 * compress. 115 * 116 * Not a big deal, just need to inform caller that we 117 * haven't allocated any pages yet. 118 */ 119 *out_pages = 0; 120 return -E2BIG; 121 } 122 } 123 124 static int compression_decompress_bio(struct list_head *ws, 125 struct compressed_bio *cb) 126 { 127 switch (cb->compress_type) { 128 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); 129 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); 130 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); 131 case BTRFS_COMPRESS_NONE: 132 default: 133 /* 134 * This can't happen, the type is validated several times 135 * before we get here. 136 */ 137 BUG(); 138 } 139 } 140 141 static int compression_decompress(int type, struct list_head *ws, 142 const u8 *data_in, struct page *dest_page, 143 unsigned long start_byte, size_t srclen, size_t destlen) 144 { 145 switch (type) { 146 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, 147 start_byte, srclen, destlen); 148 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, 149 start_byte, srclen, destlen); 150 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, 151 start_byte, srclen, destlen); 152 case BTRFS_COMPRESS_NONE: 153 default: 154 /* 155 * This can't happen, the type is validated several times 156 * before we get here. 157 */ 158 BUG(); 159 } 160 } 161 162 static void btrfs_free_compressed_pages(struct compressed_bio *cb) 163 { 164 for (unsigned int i = 0; i < cb->nr_pages; i++) 165 put_page(cb->compressed_pages[i]); 166 kfree(cb->compressed_pages); 167 } 168 169 static int btrfs_decompress_bio(struct compressed_bio *cb); 170 171 static void end_compressed_bio_read(struct btrfs_bio *bbio) 172 { 173 struct compressed_bio *cb = to_compressed_bio(bbio); 174 blk_status_t status = bbio->bio.bi_status; 175 176 if (!status) 177 status = errno_to_blk_status(btrfs_decompress_bio(cb)); 178 179 btrfs_free_compressed_pages(cb); 180 btrfs_bio_end_io(cb->orig_bbio, status); 181 bio_put(&bbio->bio); 182 } 183 184 /* 185 * Clear the writeback bits on all of the file 186 * pages for a compressed write 187 */ 188 static noinline void end_compressed_writeback(const struct compressed_bio *cb) 189 { 190 struct inode *inode = &cb->bbio.inode->vfs_inode; 191 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 192 unsigned long index = cb->start >> PAGE_SHIFT; 193 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; 194 struct folio_batch fbatch; 195 const int errno = blk_status_to_errno(cb->bbio.bio.bi_status); 196 int i; 197 int ret; 198 199 if (errno) 200 mapping_set_error(inode->i_mapping, errno); 201 202 folio_batch_init(&fbatch); 203 while (index <= end_index) { 204 ret = filemap_get_folios(inode->i_mapping, &index, end_index, 205 &fbatch); 206 207 if (ret == 0) 208 return; 209 210 for (i = 0; i < ret; i++) { 211 struct folio *folio = fbatch.folios[i]; 212 213 if (errno) 214 folio_set_error(folio); 215 btrfs_page_clamp_clear_writeback(fs_info, &folio->page, 216 cb->start, cb->len); 217 } 218 folio_batch_release(&fbatch); 219 } 220 /* the inode may be gone now */ 221 } 222 223 static void btrfs_finish_compressed_write_work(struct work_struct *work) 224 { 225 struct compressed_bio *cb = 226 container_of(work, struct compressed_bio, write_end_work); 227 228 /* 229 * Ok, we're the last bio for this extent, step one is to call back 230 * into the FS and do all the end_io operations. 231 */ 232 btrfs_writepage_endio_finish_ordered(cb->bbio.inode, NULL, 233 cb->start, cb->start + cb->len - 1, 234 cb->bbio.bio.bi_status == BLK_STS_OK); 235 236 if (cb->writeback) 237 end_compressed_writeback(cb); 238 /* Note, our inode could be gone now */ 239 240 btrfs_free_compressed_pages(cb); 241 bio_put(&cb->bbio.bio); 242 } 243 244 /* 245 * Do the cleanup once all the compressed pages hit the disk. This will clear 246 * writeback on the file pages and free the compressed pages. 247 * 248 * This also calls the writeback end hooks for the file pages so that metadata 249 * and checksums can be updated in the file. 250 */ 251 static void end_compressed_bio_write(struct btrfs_bio *bbio) 252 { 253 struct compressed_bio *cb = to_compressed_bio(bbio); 254 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info; 255 256 queue_work(fs_info->compressed_write_workers, &cb->write_end_work); 257 } 258 259 static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb) 260 { 261 struct bio *bio = &cb->bbio.bio; 262 u32 offset = 0; 263 264 while (offset < cb->compressed_len) { 265 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE); 266 267 /* Maximum compressed extent is smaller than bio size limit. */ 268 __bio_add_page(bio, cb->compressed_pages[offset >> PAGE_SHIFT], 269 len, 0); 270 offset += len; 271 } 272 } 273 274 /* 275 * worker function to build and submit bios for previously compressed pages. 276 * The corresponding pages in the inode should be marked for writeback 277 * and the compressed pages should have a reference on them for dropping 278 * when the IO is complete. 279 * 280 * This also checksums the file bytes and gets things ready for 281 * the end io hooks. 282 */ 283 void btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, 284 unsigned int len, u64 disk_start, 285 unsigned int compressed_len, 286 struct page **compressed_pages, 287 unsigned int nr_pages, 288 blk_opf_t write_flags, 289 struct cgroup_subsys_state *blkcg_css, 290 bool writeback) 291 { 292 struct btrfs_fs_info *fs_info = inode->root->fs_info; 293 struct compressed_bio *cb; 294 295 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && 296 IS_ALIGNED(len, fs_info->sectorsize)); 297 298 if (blkcg_css) { 299 kthread_associate_blkcg(blkcg_css); 300 write_flags |= REQ_CGROUP_PUNT; 301 } 302 write_flags |= REQ_BTRFS_ONE_ORDERED; 303 304 cb = alloc_compressed_bio(inode, start, REQ_OP_WRITE | write_flags, 305 end_compressed_bio_write); 306 cb->start = start; 307 cb->len = len; 308 cb->compressed_pages = compressed_pages; 309 cb->compressed_len = compressed_len; 310 cb->writeback = writeback; 311 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); 312 cb->nr_pages = nr_pages; 313 cb->bbio.bio.bi_iter.bi_sector = disk_start >> SECTOR_SHIFT; 314 btrfs_add_compressed_bio_pages(cb); 315 316 btrfs_submit_bio(&cb->bbio, 0); 317 318 if (blkcg_css) 319 kthread_associate_blkcg(NULL); 320 } 321 322 /* 323 * Add extra pages in the same compressed file extent so that we don't need to 324 * re-read the same extent again and again. 325 * 326 * NOTE: this won't work well for subpage, as for subpage read, we lock the 327 * full page then submit bio for each compressed/regular extents. 328 * 329 * This means, if we have several sectors in the same page points to the same 330 * on-disk compressed data, we will re-read the same extent many times and 331 * this function can only help for the next page. 332 */ 333 static noinline int add_ra_bio_pages(struct inode *inode, 334 u64 compressed_end, 335 struct compressed_bio *cb, 336 int *memstall, unsigned long *pflags) 337 { 338 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 339 unsigned long end_index; 340 struct bio *orig_bio = &cb->orig_bbio->bio; 341 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size; 342 u64 isize = i_size_read(inode); 343 int ret; 344 struct page *page; 345 struct extent_map *em; 346 struct address_space *mapping = inode->i_mapping; 347 struct extent_map_tree *em_tree; 348 struct extent_io_tree *tree; 349 int sectors_missed = 0; 350 351 em_tree = &BTRFS_I(inode)->extent_tree; 352 tree = &BTRFS_I(inode)->io_tree; 353 354 if (isize == 0) 355 return 0; 356 357 /* 358 * For current subpage support, we only support 64K page size, 359 * which means maximum compressed extent size (128K) is just 2x page 360 * size. 361 * This makes readahead less effective, so here disable readahead for 362 * subpage for now, until full compressed write is supported. 363 */ 364 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) 365 return 0; 366 367 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; 368 369 while (cur < compressed_end) { 370 u64 page_end; 371 u64 pg_index = cur >> PAGE_SHIFT; 372 u32 add_size; 373 374 if (pg_index > end_index) 375 break; 376 377 page = xa_load(&mapping->i_pages, pg_index); 378 if (page && !xa_is_value(page)) { 379 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> 380 fs_info->sectorsize_bits; 381 382 /* Beyond threshold, no need to continue */ 383 if (sectors_missed > 4) 384 break; 385 386 /* 387 * Jump to next page start as we already have page for 388 * current offset. 389 */ 390 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 391 continue; 392 } 393 394 page = __page_cache_alloc(mapping_gfp_constraint(mapping, 395 ~__GFP_FS)); 396 if (!page) 397 break; 398 399 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { 400 put_page(page); 401 /* There is already a page, skip to page end */ 402 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 403 continue; 404 } 405 406 if (!*memstall && PageWorkingset(page)) { 407 psi_memstall_enter(pflags); 408 *memstall = 1; 409 } 410 411 ret = set_page_extent_mapped(page); 412 if (ret < 0) { 413 unlock_page(page); 414 put_page(page); 415 break; 416 } 417 418 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; 419 lock_extent(tree, cur, page_end, NULL); 420 read_lock(&em_tree->lock); 421 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); 422 read_unlock(&em_tree->lock); 423 424 /* 425 * At this point, we have a locked page in the page cache for 426 * these bytes in the file. But, we have to make sure they map 427 * to this compressed extent on disk. 428 */ 429 if (!em || cur < em->start || 430 (cur + fs_info->sectorsize > extent_map_end(em)) || 431 (em->block_start >> 9) != orig_bio->bi_iter.bi_sector) { 432 free_extent_map(em); 433 unlock_extent(tree, cur, page_end, NULL); 434 unlock_page(page); 435 put_page(page); 436 break; 437 } 438 free_extent_map(em); 439 440 if (page->index == end_index) { 441 size_t zero_offset = offset_in_page(isize); 442 443 if (zero_offset) { 444 int zeros; 445 zeros = PAGE_SIZE - zero_offset; 446 memzero_page(page, zero_offset, zeros); 447 } 448 } 449 450 add_size = min(em->start + em->len, page_end + 1) - cur; 451 ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur)); 452 if (ret != add_size) { 453 unlock_extent(tree, cur, page_end, NULL); 454 unlock_page(page); 455 put_page(page); 456 break; 457 } 458 /* 459 * If it's subpage, we also need to increase its 460 * subpage::readers number, as at endio we will decrease 461 * subpage::readers and to unlock the page. 462 */ 463 if (fs_info->sectorsize < PAGE_SIZE) 464 btrfs_subpage_start_reader(fs_info, page, cur, add_size); 465 put_page(page); 466 cur += add_size; 467 } 468 return 0; 469 } 470 471 /* 472 * for a compressed read, the bio we get passed has all the inode pages 473 * in it. We don't actually do IO on those pages but allocate new ones 474 * to hold the compressed pages on disk. 475 * 476 * bio->bi_iter.bi_sector points to the compressed extent on disk 477 * bio->bi_io_vec points to all of the inode pages 478 * 479 * After the compressed pages are read, we copy the bytes into the 480 * bio we were passed and then call the bio end_io calls 481 */ 482 void btrfs_submit_compressed_read(struct btrfs_bio *bbio, int mirror_num) 483 { 484 struct btrfs_inode *inode = bbio->inode; 485 struct btrfs_fs_info *fs_info = inode->root->fs_info; 486 struct extent_map_tree *em_tree = &inode->extent_tree; 487 struct compressed_bio *cb; 488 unsigned int compressed_len; 489 u64 file_offset = bbio->file_offset; 490 u64 em_len; 491 u64 em_start; 492 struct extent_map *em; 493 unsigned long pflags; 494 int memstall = 0; 495 blk_status_t ret; 496 int ret2; 497 498 /* we need the actual starting offset of this extent in the file */ 499 read_lock(&em_tree->lock); 500 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); 501 read_unlock(&em_tree->lock); 502 if (!em) { 503 ret = BLK_STS_IOERR; 504 goto out; 505 } 506 507 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); 508 compressed_len = em->block_len; 509 510 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ, 511 end_compressed_bio_read); 512 513 cb->start = em->orig_start; 514 em_len = em->len; 515 em_start = em->start; 516 517 cb->len = bbio->bio.bi_iter.bi_size; 518 cb->compressed_len = compressed_len; 519 cb->compress_type = em->compress_type; 520 cb->orig_bbio = bbio; 521 522 free_extent_map(em); 523 524 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); 525 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS); 526 if (!cb->compressed_pages) { 527 ret = BLK_STS_RESOURCE; 528 goto out_free_bio; 529 } 530 531 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages); 532 if (ret2) { 533 ret = BLK_STS_RESOURCE; 534 goto out_free_compressed_pages; 535 } 536 537 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall, 538 &pflags); 539 540 /* include any pages we added in add_ra-bio_pages */ 541 cb->len = bbio->bio.bi_iter.bi_size; 542 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector; 543 btrfs_add_compressed_bio_pages(cb); 544 545 if (memstall) 546 psi_memstall_leave(&pflags); 547 548 btrfs_submit_bio(&cb->bbio, mirror_num); 549 return; 550 551 out_free_compressed_pages: 552 kfree(cb->compressed_pages); 553 out_free_bio: 554 bio_put(&cb->bbio.bio); 555 out: 556 btrfs_bio_end_io(bbio, ret); 557 } 558 559 /* 560 * Heuristic uses systematic sampling to collect data from the input data 561 * range, the logic can be tuned by the following constants: 562 * 563 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample 564 * @SAMPLING_INTERVAL - range from which the sampled data can be collected 565 */ 566 #define SAMPLING_READ_SIZE (16) 567 #define SAMPLING_INTERVAL (256) 568 569 /* 570 * For statistical analysis of the input data we consider bytes that form a 571 * Galois Field of 256 objects. Each object has an attribute count, ie. how 572 * many times the object appeared in the sample. 573 */ 574 #define BUCKET_SIZE (256) 575 576 /* 577 * The size of the sample is based on a statistical sampling rule of thumb. 578 * The common way is to perform sampling tests as long as the number of 579 * elements in each cell is at least 5. 580 * 581 * Instead of 5, we choose 32 to obtain more accurate results. 582 * If the data contain the maximum number of symbols, which is 256, we obtain a 583 * sample size bound by 8192. 584 * 585 * For a sample of at most 8KB of data per data range: 16 consecutive bytes 586 * from up to 512 locations. 587 */ 588 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ 589 SAMPLING_READ_SIZE / SAMPLING_INTERVAL) 590 591 struct bucket_item { 592 u32 count; 593 }; 594 595 struct heuristic_ws { 596 /* Partial copy of input data */ 597 u8 *sample; 598 u32 sample_size; 599 /* Buckets store counters for each byte value */ 600 struct bucket_item *bucket; 601 /* Sorting buffer */ 602 struct bucket_item *bucket_b; 603 struct list_head list; 604 }; 605 606 static struct workspace_manager heuristic_wsm; 607 608 static void free_heuristic_ws(struct list_head *ws) 609 { 610 struct heuristic_ws *workspace; 611 612 workspace = list_entry(ws, struct heuristic_ws, list); 613 614 kvfree(workspace->sample); 615 kfree(workspace->bucket); 616 kfree(workspace->bucket_b); 617 kfree(workspace); 618 } 619 620 static struct list_head *alloc_heuristic_ws(unsigned int level) 621 { 622 struct heuristic_ws *ws; 623 624 ws = kzalloc(sizeof(*ws), GFP_KERNEL); 625 if (!ws) 626 return ERR_PTR(-ENOMEM); 627 628 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); 629 if (!ws->sample) 630 goto fail; 631 632 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); 633 if (!ws->bucket) 634 goto fail; 635 636 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); 637 if (!ws->bucket_b) 638 goto fail; 639 640 INIT_LIST_HEAD(&ws->list); 641 return &ws->list; 642 fail: 643 free_heuristic_ws(&ws->list); 644 return ERR_PTR(-ENOMEM); 645 } 646 647 const struct btrfs_compress_op btrfs_heuristic_compress = { 648 .workspace_manager = &heuristic_wsm, 649 }; 650 651 static const struct btrfs_compress_op * const btrfs_compress_op[] = { 652 /* The heuristic is represented as compression type 0 */ 653 &btrfs_heuristic_compress, 654 &btrfs_zlib_compress, 655 &btrfs_lzo_compress, 656 &btrfs_zstd_compress, 657 }; 658 659 static struct list_head *alloc_workspace(int type, unsigned int level) 660 { 661 switch (type) { 662 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); 663 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); 664 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); 665 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); 666 default: 667 /* 668 * This can't happen, the type is validated several times 669 * before we get here. 670 */ 671 BUG(); 672 } 673 } 674 675 static void free_workspace(int type, struct list_head *ws) 676 { 677 switch (type) { 678 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); 679 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); 680 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); 681 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); 682 default: 683 /* 684 * This can't happen, the type is validated several times 685 * before we get here. 686 */ 687 BUG(); 688 } 689 } 690 691 static void btrfs_init_workspace_manager(int type) 692 { 693 struct workspace_manager *wsm; 694 struct list_head *workspace; 695 696 wsm = btrfs_compress_op[type]->workspace_manager; 697 INIT_LIST_HEAD(&wsm->idle_ws); 698 spin_lock_init(&wsm->ws_lock); 699 atomic_set(&wsm->total_ws, 0); 700 init_waitqueue_head(&wsm->ws_wait); 701 702 /* 703 * Preallocate one workspace for each compression type so we can 704 * guarantee forward progress in the worst case 705 */ 706 workspace = alloc_workspace(type, 0); 707 if (IS_ERR(workspace)) { 708 pr_warn( 709 "BTRFS: cannot preallocate compression workspace, will try later\n"); 710 } else { 711 atomic_set(&wsm->total_ws, 1); 712 wsm->free_ws = 1; 713 list_add(workspace, &wsm->idle_ws); 714 } 715 } 716 717 static void btrfs_cleanup_workspace_manager(int type) 718 { 719 struct workspace_manager *wsman; 720 struct list_head *ws; 721 722 wsman = btrfs_compress_op[type]->workspace_manager; 723 while (!list_empty(&wsman->idle_ws)) { 724 ws = wsman->idle_ws.next; 725 list_del(ws); 726 free_workspace(type, ws); 727 atomic_dec(&wsman->total_ws); 728 } 729 } 730 731 /* 732 * This finds an available workspace or allocates a new one. 733 * If it's not possible to allocate a new one, waits until there's one. 734 * Preallocation makes a forward progress guarantees and we do not return 735 * errors. 736 */ 737 struct list_head *btrfs_get_workspace(int type, unsigned int level) 738 { 739 struct workspace_manager *wsm; 740 struct list_head *workspace; 741 int cpus = num_online_cpus(); 742 unsigned nofs_flag; 743 struct list_head *idle_ws; 744 spinlock_t *ws_lock; 745 atomic_t *total_ws; 746 wait_queue_head_t *ws_wait; 747 int *free_ws; 748 749 wsm = btrfs_compress_op[type]->workspace_manager; 750 idle_ws = &wsm->idle_ws; 751 ws_lock = &wsm->ws_lock; 752 total_ws = &wsm->total_ws; 753 ws_wait = &wsm->ws_wait; 754 free_ws = &wsm->free_ws; 755 756 again: 757 spin_lock(ws_lock); 758 if (!list_empty(idle_ws)) { 759 workspace = idle_ws->next; 760 list_del(workspace); 761 (*free_ws)--; 762 spin_unlock(ws_lock); 763 return workspace; 764 765 } 766 if (atomic_read(total_ws) > cpus) { 767 DEFINE_WAIT(wait); 768 769 spin_unlock(ws_lock); 770 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); 771 if (atomic_read(total_ws) > cpus && !*free_ws) 772 schedule(); 773 finish_wait(ws_wait, &wait); 774 goto again; 775 } 776 atomic_inc(total_ws); 777 spin_unlock(ws_lock); 778 779 /* 780 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have 781 * to turn it off here because we might get called from the restricted 782 * context of btrfs_compress_bio/btrfs_compress_pages 783 */ 784 nofs_flag = memalloc_nofs_save(); 785 workspace = alloc_workspace(type, level); 786 memalloc_nofs_restore(nofs_flag); 787 788 if (IS_ERR(workspace)) { 789 atomic_dec(total_ws); 790 wake_up(ws_wait); 791 792 /* 793 * Do not return the error but go back to waiting. There's a 794 * workspace preallocated for each type and the compression 795 * time is bounded so we get to a workspace eventually. This 796 * makes our caller's life easier. 797 * 798 * To prevent silent and low-probability deadlocks (when the 799 * initial preallocation fails), check if there are any 800 * workspaces at all. 801 */ 802 if (atomic_read(total_ws) == 0) { 803 static DEFINE_RATELIMIT_STATE(_rs, 804 /* once per minute */ 60 * HZ, 805 /* no burst */ 1); 806 807 if (__ratelimit(&_rs)) { 808 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); 809 } 810 } 811 goto again; 812 } 813 return workspace; 814 } 815 816 static struct list_head *get_workspace(int type, int level) 817 { 818 switch (type) { 819 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); 820 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); 821 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); 822 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); 823 default: 824 /* 825 * This can't happen, the type is validated several times 826 * before we get here. 827 */ 828 BUG(); 829 } 830 } 831 832 /* 833 * put a workspace struct back on the list or free it if we have enough 834 * idle ones sitting around 835 */ 836 void btrfs_put_workspace(int type, struct list_head *ws) 837 { 838 struct workspace_manager *wsm; 839 struct list_head *idle_ws; 840 spinlock_t *ws_lock; 841 atomic_t *total_ws; 842 wait_queue_head_t *ws_wait; 843 int *free_ws; 844 845 wsm = btrfs_compress_op[type]->workspace_manager; 846 idle_ws = &wsm->idle_ws; 847 ws_lock = &wsm->ws_lock; 848 total_ws = &wsm->total_ws; 849 ws_wait = &wsm->ws_wait; 850 free_ws = &wsm->free_ws; 851 852 spin_lock(ws_lock); 853 if (*free_ws <= num_online_cpus()) { 854 list_add(ws, idle_ws); 855 (*free_ws)++; 856 spin_unlock(ws_lock); 857 goto wake; 858 } 859 spin_unlock(ws_lock); 860 861 free_workspace(type, ws); 862 atomic_dec(total_ws); 863 wake: 864 cond_wake_up(ws_wait); 865 } 866 867 static void put_workspace(int type, struct list_head *ws) 868 { 869 switch (type) { 870 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); 871 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); 872 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); 873 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); 874 default: 875 /* 876 * This can't happen, the type is validated several times 877 * before we get here. 878 */ 879 BUG(); 880 } 881 } 882 883 /* 884 * Adjust @level according to the limits of the compression algorithm or 885 * fallback to default 886 */ 887 static unsigned int btrfs_compress_set_level(int type, unsigned level) 888 { 889 const struct btrfs_compress_op *ops = btrfs_compress_op[type]; 890 891 if (level == 0) 892 level = ops->default_level; 893 else 894 level = min(level, ops->max_level); 895 896 return level; 897 } 898 899 /* 900 * Given an address space and start and length, compress the bytes into @pages 901 * that are allocated on demand. 902 * 903 * @type_level is encoded algorithm and level, where level 0 means whatever 904 * default the algorithm chooses and is opaque here; 905 * - compression algo are 0-3 906 * - the level are bits 4-7 907 * 908 * @out_pages is an in/out parameter, holds maximum number of pages to allocate 909 * and returns number of actually allocated pages 910 * 911 * @total_in is used to return the number of bytes actually read. It 912 * may be smaller than the input length if we had to exit early because we 913 * ran out of room in the pages array or because we cross the 914 * max_out threshold. 915 * 916 * @total_out is an in/out parameter, must be set to the input length and will 917 * be also used to return the total number of compressed bytes 918 */ 919 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, 920 u64 start, struct page **pages, 921 unsigned long *out_pages, 922 unsigned long *total_in, 923 unsigned long *total_out) 924 { 925 int type = btrfs_compress_type(type_level); 926 int level = btrfs_compress_level(type_level); 927 struct list_head *workspace; 928 int ret; 929 930 level = btrfs_compress_set_level(type, level); 931 workspace = get_workspace(type, level); 932 ret = compression_compress_pages(type, workspace, mapping, start, pages, 933 out_pages, total_in, total_out); 934 put_workspace(type, workspace); 935 return ret; 936 } 937 938 static int btrfs_decompress_bio(struct compressed_bio *cb) 939 { 940 struct list_head *workspace; 941 int ret; 942 int type = cb->compress_type; 943 944 workspace = get_workspace(type, 0); 945 ret = compression_decompress_bio(workspace, cb); 946 put_workspace(type, workspace); 947 948 if (!ret) 949 zero_fill_bio(&cb->orig_bbio->bio); 950 return ret; 951 } 952 953 /* 954 * a less complex decompression routine. Our compressed data fits in a 955 * single page, and we want to read a single page out of it. 956 * start_byte tells us the offset into the compressed data we're interested in 957 */ 958 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page, 959 unsigned long start_byte, size_t srclen, size_t destlen) 960 { 961 struct list_head *workspace; 962 int ret; 963 964 workspace = get_workspace(type, 0); 965 ret = compression_decompress(type, workspace, data_in, dest_page, 966 start_byte, srclen, destlen); 967 put_workspace(type, workspace); 968 969 return ret; 970 } 971 972 int __init btrfs_init_compress(void) 973 { 974 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE, 975 offsetof(struct compressed_bio, bbio.bio), 976 BIOSET_NEED_BVECS)) 977 return -ENOMEM; 978 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); 979 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); 980 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); 981 zstd_init_workspace_manager(); 982 return 0; 983 } 984 985 void __cold btrfs_exit_compress(void) 986 { 987 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); 988 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); 989 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); 990 zstd_cleanup_workspace_manager(); 991 bioset_exit(&btrfs_compressed_bioset); 992 } 993 994 /* 995 * Copy decompressed data from working buffer to pages. 996 * 997 * @buf: The decompressed data buffer 998 * @buf_len: The decompressed data length 999 * @decompressed: Number of bytes that are already decompressed inside the 1000 * compressed extent 1001 * @cb: The compressed extent descriptor 1002 * @orig_bio: The original bio that the caller wants to read for 1003 * 1004 * An easier to understand graph is like below: 1005 * 1006 * |<- orig_bio ->| |<- orig_bio->| 1007 * |<------- full decompressed extent ----->| 1008 * |<----------- @cb range ---->| 1009 * | |<-- @buf_len -->| 1010 * |<--- @decompressed --->| 1011 * 1012 * Note that, @cb can be a subpage of the full decompressed extent, but 1013 * @cb->start always has the same as the orig_file_offset value of the full 1014 * decompressed extent. 1015 * 1016 * When reading compressed extent, we have to read the full compressed extent, 1017 * while @orig_bio may only want part of the range. 1018 * Thus this function will ensure only data covered by @orig_bio will be copied 1019 * to. 1020 * 1021 * Return 0 if we have copied all needed contents for @orig_bio. 1022 * Return >0 if we need continue decompress. 1023 */ 1024 int btrfs_decompress_buf2page(const char *buf, u32 buf_len, 1025 struct compressed_bio *cb, u32 decompressed) 1026 { 1027 struct bio *orig_bio = &cb->orig_bbio->bio; 1028 /* Offset inside the full decompressed extent */ 1029 u32 cur_offset; 1030 1031 cur_offset = decompressed; 1032 /* The main loop to do the copy */ 1033 while (cur_offset < decompressed + buf_len) { 1034 struct bio_vec bvec; 1035 size_t copy_len; 1036 u32 copy_start; 1037 /* Offset inside the full decompressed extent */ 1038 u32 bvec_offset; 1039 1040 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); 1041 /* 1042 * cb->start may underflow, but subtracting that value can still 1043 * give us correct offset inside the full decompressed extent. 1044 */ 1045 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; 1046 1047 /* Haven't reached the bvec range, exit */ 1048 if (decompressed + buf_len <= bvec_offset) 1049 return 1; 1050 1051 copy_start = max(cur_offset, bvec_offset); 1052 copy_len = min(bvec_offset + bvec.bv_len, 1053 decompressed + buf_len) - copy_start; 1054 ASSERT(copy_len); 1055 1056 /* 1057 * Extra range check to ensure we didn't go beyond 1058 * @buf + @buf_len. 1059 */ 1060 ASSERT(copy_start - decompressed < buf_len); 1061 memcpy_to_page(bvec.bv_page, bvec.bv_offset, 1062 buf + copy_start - decompressed, copy_len); 1063 cur_offset += copy_len; 1064 1065 bio_advance(orig_bio, copy_len); 1066 /* Finished the bio */ 1067 if (!orig_bio->bi_iter.bi_size) 1068 return 0; 1069 } 1070 return 1; 1071 } 1072 1073 /* 1074 * Shannon Entropy calculation 1075 * 1076 * Pure byte distribution analysis fails to determine compressibility of data. 1077 * Try calculating entropy to estimate the average minimum number of bits 1078 * needed to encode the sampled data. 1079 * 1080 * For convenience, return the percentage of needed bits, instead of amount of 1081 * bits directly. 1082 * 1083 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy 1084 * and can be compressible with high probability 1085 * 1086 * @ENTROPY_LVL_HIGH - data are not compressible with high probability 1087 * 1088 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. 1089 */ 1090 #define ENTROPY_LVL_ACEPTABLE (65) 1091 #define ENTROPY_LVL_HIGH (80) 1092 1093 /* 1094 * For increasead precision in shannon_entropy calculation, 1095 * let's do pow(n, M) to save more digits after comma: 1096 * 1097 * - maximum int bit length is 64 1098 * - ilog2(MAX_SAMPLE_SIZE) -> 13 1099 * - 13 * 4 = 52 < 64 -> M = 4 1100 * 1101 * So use pow(n, 4). 1102 */ 1103 static inline u32 ilog2_w(u64 n) 1104 { 1105 return ilog2(n * n * n * n); 1106 } 1107 1108 static u32 shannon_entropy(struct heuristic_ws *ws) 1109 { 1110 const u32 entropy_max = 8 * ilog2_w(2); 1111 u32 entropy_sum = 0; 1112 u32 p, p_base, sz_base; 1113 u32 i; 1114 1115 sz_base = ilog2_w(ws->sample_size); 1116 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { 1117 p = ws->bucket[i].count; 1118 p_base = ilog2_w(p); 1119 entropy_sum += p * (sz_base - p_base); 1120 } 1121 1122 entropy_sum /= ws->sample_size; 1123 return entropy_sum * 100 / entropy_max; 1124 } 1125 1126 #define RADIX_BASE 4U 1127 #define COUNTERS_SIZE (1U << RADIX_BASE) 1128 1129 static u8 get4bits(u64 num, int shift) { 1130 u8 low4bits; 1131 1132 num >>= shift; 1133 /* Reverse order */ 1134 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); 1135 return low4bits; 1136 } 1137 1138 /* 1139 * Use 4 bits as radix base 1140 * Use 16 u32 counters for calculating new position in buf array 1141 * 1142 * @array - array that will be sorted 1143 * @array_buf - buffer array to store sorting results 1144 * must be equal in size to @array 1145 * @num - array size 1146 */ 1147 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, 1148 int num) 1149 { 1150 u64 max_num; 1151 u64 buf_num; 1152 u32 counters[COUNTERS_SIZE]; 1153 u32 new_addr; 1154 u32 addr; 1155 int bitlen; 1156 int shift; 1157 int i; 1158 1159 /* 1160 * Try avoid useless loop iterations for small numbers stored in big 1161 * counters. Example: 48 33 4 ... in 64bit array 1162 */ 1163 max_num = array[0].count; 1164 for (i = 1; i < num; i++) { 1165 buf_num = array[i].count; 1166 if (buf_num > max_num) 1167 max_num = buf_num; 1168 } 1169 1170 buf_num = ilog2(max_num); 1171 bitlen = ALIGN(buf_num, RADIX_BASE * 2); 1172 1173 shift = 0; 1174 while (shift < bitlen) { 1175 memset(counters, 0, sizeof(counters)); 1176 1177 for (i = 0; i < num; i++) { 1178 buf_num = array[i].count; 1179 addr = get4bits(buf_num, shift); 1180 counters[addr]++; 1181 } 1182 1183 for (i = 1; i < COUNTERS_SIZE; i++) 1184 counters[i] += counters[i - 1]; 1185 1186 for (i = num - 1; i >= 0; i--) { 1187 buf_num = array[i].count; 1188 addr = get4bits(buf_num, shift); 1189 counters[addr]--; 1190 new_addr = counters[addr]; 1191 array_buf[new_addr] = array[i]; 1192 } 1193 1194 shift += RADIX_BASE; 1195 1196 /* 1197 * Normal radix expects to move data from a temporary array, to 1198 * the main one. But that requires some CPU time. Avoid that 1199 * by doing another sort iteration to original array instead of 1200 * memcpy() 1201 */ 1202 memset(counters, 0, sizeof(counters)); 1203 1204 for (i = 0; i < num; i ++) { 1205 buf_num = array_buf[i].count; 1206 addr = get4bits(buf_num, shift); 1207 counters[addr]++; 1208 } 1209 1210 for (i = 1; i < COUNTERS_SIZE; i++) 1211 counters[i] += counters[i - 1]; 1212 1213 for (i = num - 1; i >= 0; i--) { 1214 buf_num = array_buf[i].count; 1215 addr = get4bits(buf_num, shift); 1216 counters[addr]--; 1217 new_addr = counters[addr]; 1218 array[new_addr] = array_buf[i]; 1219 } 1220 1221 shift += RADIX_BASE; 1222 } 1223 } 1224 1225 /* 1226 * Size of the core byte set - how many bytes cover 90% of the sample 1227 * 1228 * There are several types of structured binary data that use nearly all byte 1229 * values. The distribution can be uniform and counts in all buckets will be 1230 * nearly the same (eg. encrypted data). Unlikely to be compressible. 1231 * 1232 * Other possibility is normal (Gaussian) distribution, where the data could 1233 * be potentially compressible, but we have to take a few more steps to decide 1234 * how much. 1235 * 1236 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, 1237 * compression algo can easy fix that 1238 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high 1239 * probability is not compressible 1240 */ 1241 #define BYTE_CORE_SET_LOW (64) 1242 #define BYTE_CORE_SET_HIGH (200) 1243 1244 static int byte_core_set_size(struct heuristic_ws *ws) 1245 { 1246 u32 i; 1247 u32 coreset_sum = 0; 1248 const u32 core_set_threshold = ws->sample_size * 90 / 100; 1249 struct bucket_item *bucket = ws->bucket; 1250 1251 /* Sort in reverse order */ 1252 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); 1253 1254 for (i = 0; i < BYTE_CORE_SET_LOW; i++) 1255 coreset_sum += bucket[i].count; 1256 1257 if (coreset_sum > core_set_threshold) 1258 return i; 1259 1260 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { 1261 coreset_sum += bucket[i].count; 1262 if (coreset_sum > core_set_threshold) 1263 break; 1264 } 1265 1266 return i; 1267 } 1268 1269 /* 1270 * Count byte values in buckets. 1271 * This heuristic can detect textual data (configs, xml, json, html, etc). 1272 * Because in most text-like data byte set is restricted to limited number of 1273 * possible characters, and that restriction in most cases makes data easy to 1274 * compress. 1275 * 1276 * @BYTE_SET_THRESHOLD - consider all data within this byte set size: 1277 * less - compressible 1278 * more - need additional analysis 1279 */ 1280 #define BYTE_SET_THRESHOLD (64) 1281 1282 static u32 byte_set_size(const struct heuristic_ws *ws) 1283 { 1284 u32 i; 1285 u32 byte_set_size = 0; 1286 1287 for (i = 0; i < BYTE_SET_THRESHOLD; i++) { 1288 if (ws->bucket[i].count > 0) 1289 byte_set_size++; 1290 } 1291 1292 /* 1293 * Continue collecting count of byte values in buckets. If the byte 1294 * set size is bigger then the threshold, it's pointless to continue, 1295 * the detection technique would fail for this type of data. 1296 */ 1297 for (; i < BUCKET_SIZE; i++) { 1298 if (ws->bucket[i].count > 0) { 1299 byte_set_size++; 1300 if (byte_set_size > BYTE_SET_THRESHOLD) 1301 return byte_set_size; 1302 } 1303 } 1304 1305 return byte_set_size; 1306 } 1307 1308 static bool sample_repeated_patterns(struct heuristic_ws *ws) 1309 { 1310 const u32 half_of_sample = ws->sample_size / 2; 1311 const u8 *data = ws->sample; 1312 1313 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; 1314 } 1315 1316 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, 1317 struct heuristic_ws *ws) 1318 { 1319 struct page *page; 1320 u64 index, index_end; 1321 u32 i, curr_sample_pos; 1322 u8 *in_data; 1323 1324 /* 1325 * Compression handles the input data by chunks of 128KiB 1326 * (defined by BTRFS_MAX_UNCOMPRESSED) 1327 * 1328 * We do the same for the heuristic and loop over the whole range. 1329 * 1330 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will 1331 * process no more than BTRFS_MAX_UNCOMPRESSED at a time. 1332 */ 1333 if (end - start > BTRFS_MAX_UNCOMPRESSED) 1334 end = start + BTRFS_MAX_UNCOMPRESSED; 1335 1336 index = start >> PAGE_SHIFT; 1337 index_end = end >> PAGE_SHIFT; 1338 1339 /* Don't miss unaligned end */ 1340 if (!PAGE_ALIGNED(end)) 1341 index_end++; 1342 1343 curr_sample_pos = 0; 1344 while (index < index_end) { 1345 page = find_get_page(inode->i_mapping, index); 1346 in_data = kmap_local_page(page); 1347 /* Handle case where the start is not aligned to PAGE_SIZE */ 1348 i = start % PAGE_SIZE; 1349 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { 1350 /* Don't sample any garbage from the last page */ 1351 if (start > end - SAMPLING_READ_SIZE) 1352 break; 1353 memcpy(&ws->sample[curr_sample_pos], &in_data[i], 1354 SAMPLING_READ_SIZE); 1355 i += SAMPLING_INTERVAL; 1356 start += SAMPLING_INTERVAL; 1357 curr_sample_pos += SAMPLING_READ_SIZE; 1358 } 1359 kunmap_local(in_data); 1360 put_page(page); 1361 1362 index++; 1363 } 1364 1365 ws->sample_size = curr_sample_pos; 1366 } 1367 1368 /* 1369 * Compression heuristic. 1370 * 1371 * For now is's a naive and optimistic 'return true', we'll extend the logic to 1372 * quickly (compared to direct compression) detect data characteristics 1373 * (compressible/incompressible) to avoid wasting CPU time on incompressible 1374 * data. 1375 * 1376 * The following types of analysis can be performed: 1377 * - detect mostly zero data 1378 * - detect data with low "byte set" size (text, etc) 1379 * - detect data with low/high "core byte" set 1380 * 1381 * Return non-zero if the compression should be done, 0 otherwise. 1382 */ 1383 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) 1384 { 1385 struct list_head *ws_list = get_workspace(0, 0); 1386 struct heuristic_ws *ws; 1387 u32 i; 1388 u8 byte; 1389 int ret = 0; 1390 1391 ws = list_entry(ws_list, struct heuristic_ws, list); 1392 1393 heuristic_collect_sample(inode, start, end, ws); 1394 1395 if (sample_repeated_patterns(ws)) { 1396 ret = 1; 1397 goto out; 1398 } 1399 1400 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); 1401 1402 for (i = 0; i < ws->sample_size; i++) { 1403 byte = ws->sample[i]; 1404 ws->bucket[byte].count++; 1405 } 1406 1407 i = byte_set_size(ws); 1408 if (i < BYTE_SET_THRESHOLD) { 1409 ret = 2; 1410 goto out; 1411 } 1412 1413 i = byte_core_set_size(ws); 1414 if (i <= BYTE_CORE_SET_LOW) { 1415 ret = 3; 1416 goto out; 1417 } 1418 1419 if (i >= BYTE_CORE_SET_HIGH) { 1420 ret = 0; 1421 goto out; 1422 } 1423 1424 i = shannon_entropy(ws); 1425 if (i <= ENTROPY_LVL_ACEPTABLE) { 1426 ret = 4; 1427 goto out; 1428 } 1429 1430 /* 1431 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be 1432 * needed to give green light to compression. 1433 * 1434 * For now just assume that compression at that level is not worth the 1435 * resources because: 1436 * 1437 * 1. it is possible to defrag the data later 1438 * 1439 * 2. the data would turn out to be hardly compressible, eg. 150 byte 1440 * values, every bucket has counter at level ~54. The heuristic would 1441 * be confused. This can happen when data have some internal repeated 1442 * patterns like "abbacbbc...". This can be detected by analyzing 1443 * pairs of bytes, which is too costly. 1444 */ 1445 if (i < ENTROPY_LVL_HIGH) { 1446 ret = 5; 1447 goto out; 1448 } else { 1449 ret = 0; 1450 goto out; 1451 } 1452 1453 out: 1454 put_workspace(0, ws_list); 1455 return ret; 1456 } 1457 1458 /* 1459 * Convert the compression suffix (eg. after "zlib" starting with ":") to 1460 * level, unrecognized string will set the default level 1461 */ 1462 unsigned int btrfs_compress_str2level(unsigned int type, const char *str) 1463 { 1464 unsigned int level = 0; 1465 int ret; 1466 1467 if (!type) 1468 return 0; 1469 1470 if (str[0] == ':') { 1471 ret = kstrtouint(str + 1, 10, &level); 1472 if (ret) 1473 level = 0; 1474 } 1475 1476 level = btrfs_compress_set_level(type, level); 1477 1478 return level; 1479 } 1480