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