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