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(btrfs_bio(cb->orig_bio), 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 finish_compressed_bio_write(struct compressed_bio *cb) 224 { 225 /* 226 * Ok, we're the last bio for this extent, step one is to call back 227 * into the FS and do all the end_io operations. 228 */ 229 btrfs_writepage_endio_finish_ordered(cb->bbio.inode, NULL, 230 cb->start, cb->start + cb->len - 1, 231 cb->bbio.bio.bi_status == BLK_STS_OK); 232 233 if (cb->writeback) 234 end_compressed_writeback(cb); 235 /* Note, our inode could be gone now */ 236 237 btrfs_free_compressed_pages(cb); 238 bio_put(&cb->bbio.bio); 239 } 240 241 static void btrfs_finish_compressed_write_work(struct work_struct *work) 242 { 243 struct compressed_bio *cb = 244 container_of(work, struct compressed_bio, write_end_work); 245 246 finish_compressed_bio_write(cb); 247 } 248 249 /* 250 * Do the cleanup once all the compressed pages hit the disk. This will clear 251 * writeback on the file pages and free the compressed pages. 252 * 253 * This also calls the writeback end hooks for the file pages so that metadata 254 * and checksums can be updated in the file. 255 */ 256 static void end_compressed_bio_write(struct btrfs_bio *bbio) 257 { 258 struct compressed_bio *cb = to_compressed_bio(bbio); 259 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info; 260 261 queue_work(fs_info->compressed_write_workers, &cb->write_end_work); 262 } 263 264 static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb, 265 u64 disk_bytenr) 266 { 267 struct btrfs_fs_info *fs_info = cb->bbio.inode->root->fs_info; 268 struct bio *bio = &cb->bbio.bio; 269 u64 cur_disk_byte = disk_bytenr; 270 271 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT; 272 while (cur_disk_byte < disk_bytenr + cb->compressed_len) { 273 u64 offset = cur_disk_byte - disk_bytenr; 274 unsigned int index = offset >> PAGE_SHIFT; 275 unsigned int real_size; 276 unsigned int added; 277 struct page *page = cb->compressed_pages[index]; 278 279 /* 280 * We have various limit on the real read size: 281 * - page boundary 282 * - compressed length boundary 283 */ 284 real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset)); 285 real_size = min_t(u64, real_size, cb->compressed_len - offset); 286 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); 287 288 added = bio_add_page(bio, page, real_size, offset_in_page(offset)); 289 /* 290 * Maximum compressed extent is smaller than bio size limit, 291 * thus bio_add_page() should always success. 292 */ 293 ASSERT(added == real_size); 294 cur_disk_byte += added; 295 } 296 297 ASSERT(bio->bi_iter.bi_size); 298 } 299 300 /* 301 * worker function to build and submit bios for previously compressed pages. 302 * The corresponding pages in the inode should be marked for writeback 303 * and the compressed pages should have a reference on them for dropping 304 * when the IO is complete. 305 * 306 * This also checksums the file bytes and gets things ready for 307 * the end io hooks. 308 */ 309 void btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, 310 unsigned int len, u64 disk_start, 311 unsigned int compressed_len, 312 struct page **compressed_pages, 313 unsigned int nr_pages, 314 blk_opf_t write_flags, 315 struct cgroup_subsys_state *blkcg_css, 316 bool writeback) 317 { 318 struct btrfs_fs_info *fs_info = inode->root->fs_info; 319 struct compressed_bio *cb; 320 321 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && 322 IS_ALIGNED(len, fs_info->sectorsize)); 323 324 if (blkcg_css) { 325 kthread_associate_blkcg(blkcg_css); 326 write_flags |= REQ_CGROUP_PUNT; 327 } 328 write_flags |= REQ_BTRFS_ONE_ORDERED; 329 330 cb = alloc_compressed_bio(inode, start, REQ_OP_WRITE | write_flags, 331 end_compressed_bio_write); 332 cb->start = start; 333 cb->len = len; 334 cb->compressed_pages = compressed_pages; 335 cb->compressed_len = compressed_len; 336 cb->writeback = writeback; 337 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); 338 cb->nr_pages = nr_pages; 339 340 btrfs_add_compressed_bio_pages(cb, disk_start); 341 btrfs_submit_bio(&cb->bbio.bio, 0); 342 343 if (blkcg_css) 344 kthread_associate_blkcg(NULL); 345 } 346 347 /* 348 * Add extra pages in the same compressed file extent so that we don't need to 349 * re-read the same extent again and again. 350 * 351 * NOTE: this won't work well for subpage, as for subpage read, we lock the 352 * full page then submit bio for each compressed/regular extents. 353 * 354 * This means, if we have several sectors in the same page points to the same 355 * on-disk compressed data, we will re-read the same extent many times and 356 * this function can only help for the next page. 357 */ 358 static noinline int add_ra_bio_pages(struct inode *inode, 359 u64 compressed_end, 360 struct compressed_bio *cb, 361 int *memstall, unsigned long *pflags) 362 { 363 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); 364 unsigned long end_index; 365 u64 cur = btrfs_bio(cb->orig_bio)->file_offset + cb->orig_bio->bi_iter.bi_size; 366 u64 isize = i_size_read(inode); 367 int ret; 368 struct page *page; 369 struct extent_map *em; 370 struct address_space *mapping = inode->i_mapping; 371 struct extent_map_tree *em_tree; 372 struct extent_io_tree *tree; 373 int sectors_missed = 0; 374 375 em_tree = &BTRFS_I(inode)->extent_tree; 376 tree = &BTRFS_I(inode)->io_tree; 377 378 if (isize == 0) 379 return 0; 380 381 /* 382 * For current subpage support, we only support 64K page size, 383 * which means maximum compressed extent size (128K) is just 2x page 384 * size. 385 * This makes readahead less effective, so here disable readahead for 386 * subpage for now, until full compressed write is supported. 387 */ 388 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) 389 return 0; 390 391 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; 392 393 while (cur < compressed_end) { 394 u64 page_end; 395 u64 pg_index = cur >> PAGE_SHIFT; 396 u32 add_size; 397 398 if (pg_index > end_index) 399 break; 400 401 page = xa_load(&mapping->i_pages, pg_index); 402 if (page && !xa_is_value(page)) { 403 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> 404 fs_info->sectorsize_bits; 405 406 /* Beyond threshold, no need to continue */ 407 if (sectors_missed > 4) 408 break; 409 410 /* 411 * Jump to next page start as we already have page for 412 * current offset. 413 */ 414 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 415 continue; 416 } 417 418 page = __page_cache_alloc(mapping_gfp_constraint(mapping, 419 ~__GFP_FS)); 420 if (!page) 421 break; 422 423 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { 424 put_page(page); 425 /* There is already a page, skip to page end */ 426 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; 427 continue; 428 } 429 430 if (!*memstall && PageWorkingset(page)) { 431 psi_memstall_enter(pflags); 432 *memstall = 1; 433 } 434 435 ret = set_page_extent_mapped(page); 436 if (ret < 0) { 437 unlock_page(page); 438 put_page(page); 439 break; 440 } 441 442 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; 443 lock_extent(tree, cur, page_end, NULL); 444 read_lock(&em_tree->lock); 445 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); 446 read_unlock(&em_tree->lock); 447 448 /* 449 * At this point, we have a locked page in the page cache for 450 * these bytes in the file. But, we have to make sure they map 451 * to this compressed extent on disk. 452 */ 453 if (!em || cur < em->start || 454 (cur + fs_info->sectorsize > extent_map_end(em)) || 455 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { 456 free_extent_map(em); 457 unlock_extent(tree, cur, page_end, NULL); 458 unlock_page(page); 459 put_page(page); 460 break; 461 } 462 free_extent_map(em); 463 464 if (page->index == end_index) { 465 size_t zero_offset = offset_in_page(isize); 466 467 if (zero_offset) { 468 int zeros; 469 zeros = PAGE_SIZE - zero_offset; 470 memzero_page(page, zero_offset, zeros); 471 } 472 } 473 474 add_size = min(em->start + em->len, page_end + 1) - cur; 475 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur)); 476 if (ret != add_size) { 477 unlock_extent(tree, cur, page_end, NULL); 478 unlock_page(page); 479 put_page(page); 480 break; 481 } 482 /* 483 * If it's subpage, we also need to increase its 484 * subpage::readers number, as at endio we will decrease 485 * subpage::readers and to unlock the page. 486 */ 487 if (fs_info->sectorsize < PAGE_SIZE) 488 btrfs_subpage_start_reader(fs_info, page, cur, add_size); 489 put_page(page); 490 cur += add_size; 491 } 492 return 0; 493 } 494 495 /* 496 * for a compressed read, the bio we get passed has all the inode pages 497 * in it. We don't actually do IO on those pages but allocate new ones 498 * to hold the compressed pages on disk. 499 * 500 * bio->bi_iter.bi_sector points to the compressed extent on disk 501 * bio->bi_io_vec points to all of the inode pages 502 * 503 * After the compressed pages are read, we copy the bytes into the 504 * bio we were passed and then call the bio end_io calls 505 */ 506 void btrfs_submit_compressed_read(struct bio *bio, int mirror_num) 507 { 508 struct btrfs_inode *inode = btrfs_bio(bio)->inode; 509 struct btrfs_fs_info *fs_info = inode->root->fs_info; 510 struct extent_map_tree *em_tree = &inode->extent_tree; 511 struct compressed_bio *cb; 512 unsigned int compressed_len; 513 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT; 514 u64 file_offset = btrfs_bio(bio)->file_offset; 515 u64 em_len; 516 u64 em_start; 517 struct extent_map *em; 518 unsigned long pflags; 519 int memstall = 0; 520 blk_status_t ret; 521 int ret2; 522 523 /* we need the actual starting offset of this extent in the file */ 524 read_lock(&em_tree->lock); 525 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); 526 read_unlock(&em_tree->lock); 527 if (!em) { 528 ret = BLK_STS_IOERR; 529 goto out; 530 } 531 532 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); 533 compressed_len = em->block_len; 534 535 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ, 536 end_compressed_bio_read); 537 538 cb->start = em->orig_start; 539 em_len = em->len; 540 em_start = em->start; 541 542 cb->len = bio->bi_iter.bi_size; 543 cb->compressed_len = compressed_len; 544 cb->compress_type = em->compress_type; 545 cb->orig_bio = bio; 546 547 free_extent_map(em); 548 549 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); 550 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS); 551 if (!cb->compressed_pages) { 552 ret = BLK_STS_RESOURCE; 553 goto out_free_bio; 554 } 555 556 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages); 557 if (ret2) { 558 ret = BLK_STS_RESOURCE; 559 goto out_free_compressed_pages; 560 } 561 562 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall, 563 &pflags); 564 565 /* include any pages we added in add_ra-bio_pages */ 566 cb->len = bio->bi_iter.bi_size; 567 568 btrfs_add_compressed_bio_pages(cb, disk_bytenr); 569 570 if (memstall) 571 psi_memstall_leave(&pflags); 572 573 btrfs_submit_bio(&cb->bbio.bio, mirror_num); 574 return; 575 576 out_free_compressed_pages: 577 kfree(cb->compressed_pages); 578 out_free_bio: 579 bio_put(&cb->bbio.bio); 580 out: 581 btrfs_bio_end_io(btrfs_bio(bio), ret); 582 } 583 584 /* 585 * Heuristic uses systematic sampling to collect data from the input data 586 * range, the logic can be tuned by the following constants: 587 * 588 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample 589 * @SAMPLING_INTERVAL - range from which the sampled data can be collected 590 */ 591 #define SAMPLING_READ_SIZE (16) 592 #define SAMPLING_INTERVAL (256) 593 594 /* 595 * For statistical analysis of the input data we consider bytes that form a 596 * Galois Field of 256 objects. Each object has an attribute count, ie. how 597 * many times the object appeared in the sample. 598 */ 599 #define BUCKET_SIZE (256) 600 601 /* 602 * The size of the sample is based on a statistical sampling rule of thumb. 603 * The common way is to perform sampling tests as long as the number of 604 * elements in each cell is at least 5. 605 * 606 * Instead of 5, we choose 32 to obtain more accurate results. 607 * If the data contain the maximum number of symbols, which is 256, we obtain a 608 * sample size bound by 8192. 609 * 610 * For a sample of at most 8KB of data per data range: 16 consecutive bytes 611 * from up to 512 locations. 612 */ 613 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ 614 SAMPLING_READ_SIZE / SAMPLING_INTERVAL) 615 616 struct bucket_item { 617 u32 count; 618 }; 619 620 struct heuristic_ws { 621 /* Partial copy of input data */ 622 u8 *sample; 623 u32 sample_size; 624 /* Buckets store counters for each byte value */ 625 struct bucket_item *bucket; 626 /* Sorting buffer */ 627 struct bucket_item *bucket_b; 628 struct list_head list; 629 }; 630 631 static struct workspace_manager heuristic_wsm; 632 633 static void free_heuristic_ws(struct list_head *ws) 634 { 635 struct heuristic_ws *workspace; 636 637 workspace = list_entry(ws, struct heuristic_ws, list); 638 639 kvfree(workspace->sample); 640 kfree(workspace->bucket); 641 kfree(workspace->bucket_b); 642 kfree(workspace); 643 } 644 645 static struct list_head *alloc_heuristic_ws(unsigned int level) 646 { 647 struct heuristic_ws *ws; 648 649 ws = kzalloc(sizeof(*ws), GFP_KERNEL); 650 if (!ws) 651 return ERR_PTR(-ENOMEM); 652 653 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); 654 if (!ws->sample) 655 goto fail; 656 657 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); 658 if (!ws->bucket) 659 goto fail; 660 661 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); 662 if (!ws->bucket_b) 663 goto fail; 664 665 INIT_LIST_HEAD(&ws->list); 666 return &ws->list; 667 fail: 668 free_heuristic_ws(&ws->list); 669 return ERR_PTR(-ENOMEM); 670 } 671 672 const struct btrfs_compress_op btrfs_heuristic_compress = { 673 .workspace_manager = &heuristic_wsm, 674 }; 675 676 static const struct btrfs_compress_op * const btrfs_compress_op[] = { 677 /* The heuristic is represented as compression type 0 */ 678 &btrfs_heuristic_compress, 679 &btrfs_zlib_compress, 680 &btrfs_lzo_compress, 681 &btrfs_zstd_compress, 682 }; 683 684 static struct list_head *alloc_workspace(int type, unsigned int level) 685 { 686 switch (type) { 687 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); 688 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); 689 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); 690 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); 691 default: 692 /* 693 * This can't happen, the type is validated several times 694 * before we get here. 695 */ 696 BUG(); 697 } 698 } 699 700 static void free_workspace(int type, struct list_head *ws) 701 { 702 switch (type) { 703 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); 704 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); 705 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); 706 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); 707 default: 708 /* 709 * This can't happen, the type is validated several times 710 * before we get here. 711 */ 712 BUG(); 713 } 714 } 715 716 static void btrfs_init_workspace_manager(int type) 717 { 718 struct workspace_manager *wsm; 719 struct list_head *workspace; 720 721 wsm = btrfs_compress_op[type]->workspace_manager; 722 INIT_LIST_HEAD(&wsm->idle_ws); 723 spin_lock_init(&wsm->ws_lock); 724 atomic_set(&wsm->total_ws, 0); 725 init_waitqueue_head(&wsm->ws_wait); 726 727 /* 728 * Preallocate one workspace for each compression type so we can 729 * guarantee forward progress in the worst case 730 */ 731 workspace = alloc_workspace(type, 0); 732 if (IS_ERR(workspace)) { 733 pr_warn( 734 "BTRFS: cannot preallocate compression workspace, will try later\n"); 735 } else { 736 atomic_set(&wsm->total_ws, 1); 737 wsm->free_ws = 1; 738 list_add(workspace, &wsm->idle_ws); 739 } 740 } 741 742 static void btrfs_cleanup_workspace_manager(int type) 743 { 744 struct workspace_manager *wsman; 745 struct list_head *ws; 746 747 wsman = btrfs_compress_op[type]->workspace_manager; 748 while (!list_empty(&wsman->idle_ws)) { 749 ws = wsman->idle_ws.next; 750 list_del(ws); 751 free_workspace(type, ws); 752 atomic_dec(&wsman->total_ws); 753 } 754 } 755 756 /* 757 * This finds an available workspace or allocates a new one. 758 * If it's not possible to allocate a new one, waits until there's one. 759 * Preallocation makes a forward progress guarantees and we do not return 760 * errors. 761 */ 762 struct list_head *btrfs_get_workspace(int type, unsigned int level) 763 { 764 struct workspace_manager *wsm; 765 struct list_head *workspace; 766 int cpus = num_online_cpus(); 767 unsigned nofs_flag; 768 struct list_head *idle_ws; 769 spinlock_t *ws_lock; 770 atomic_t *total_ws; 771 wait_queue_head_t *ws_wait; 772 int *free_ws; 773 774 wsm = btrfs_compress_op[type]->workspace_manager; 775 idle_ws = &wsm->idle_ws; 776 ws_lock = &wsm->ws_lock; 777 total_ws = &wsm->total_ws; 778 ws_wait = &wsm->ws_wait; 779 free_ws = &wsm->free_ws; 780 781 again: 782 spin_lock(ws_lock); 783 if (!list_empty(idle_ws)) { 784 workspace = idle_ws->next; 785 list_del(workspace); 786 (*free_ws)--; 787 spin_unlock(ws_lock); 788 return workspace; 789 790 } 791 if (atomic_read(total_ws) > cpus) { 792 DEFINE_WAIT(wait); 793 794 spin_unlock(ws_lock); 795 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); 796 if (atomic_read(total_ws) > cpus && !*free_ws) 797 schedule(); 798 finish_wait(ws_wait, &wait); 799 goto again; 800 } 801 atomic_inc(total_ws); 802 spin_unlock(ws_lock); 803 804 /* 805 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have 806 * to turn it off here because we might get called from the restricted 807 * context of btrfs_compress_bio/btrfs_compress_pages 808 */ 809 nofs_flag = memalloc_nofs_save(); 810 workspace = alloc_workspace(type, level); 811 memalloc_nofs_restore(nofs_flag); 812 813 if (IS_ERR(workspace)) { 814 atomic_dec(total_ws); 815 wake_up(ws_wait); 816 817 /* 818 * Do not return the error but go back to waiting. There's a 819 * workspace preallocated for each type and the compression 820 * time is bounded so we get to a workspace eventually. This 821 * makes our caller's life easier. 822 * 823 * To prevent silent and low-probability deadlocks (when the 824 * initial preallocation fails), check if there are any 825 * workspaces at all. 826 */ 827 if (atomic_read(total_ws) == 0) { 828 static DEFINE_RATELIMIT_STATE(_rs, 829 /* once per minute */ 60 * HZ, 830 /* no burst */ 1); 831 832 if (__ratelimit(&_rs)) { 833 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); 834 } 835 } 836 goto again; 837 } 838 return workspace; 839 } 840 841 static struct list_head *get_workspace(int type, int level) 842 { 843 switch (type) { 844 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); 845 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); 846 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); 847 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); 848 default: 849 /* 850 * This can't happen, the type is validated several times 851 * before we get here. 852 */ 853 BUG(); 854 } 855 } 856 857 /* 858 * put a workspace struct back on the list or free it if we have enough 859 * idle ones sitting around 860 */ 861 void btrfs_put_workspace(int type, struct list_head *ws) 862 { 863 struct workspace_manager *wsm; 864 struct list_head *idle_ws; 865 spinlock_t *ws_lock; 866 atomic_t *total_ws; 867 wait_queue_head_t *ws_wait; 868 int *free_ws; 869 870 wsm = btrfs_compress_op[type]->workspace_manager; 871 idle_ws = &wsm->idle_ws; 872 ws_lock = &wsm->ws_lock; 873 total_ws = &wsm->total_ws; 874 ws_wait = &wsm->ws_wait; 875 free_ws = &wsm->free_ws; 876 877 spin_lock(ws_lock); 878 if (*free_ws <= num_online_cpus()) { 879 list_add(ws, idle_ws); 880 (*free_ws)++; 881 spin_unlock(ws_lock); 882 goto wake; 883 } 884 spin_unlock(ws_lock); 885 886 free_workspace(type, ws); 887 atomic_dec(total_ws); 888 wake: 889 cond_wake_up(ws_wait); 890 } 891 892 static void put_workspace(int type, struct list_head *ws) 893 { 894 switch (type) { 895 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); 896 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); 897 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); 898 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); 899 default: 900 /* 901 * This can't happen, the type is validated several times 902 * before we get here. 903 */ 904 BUG(); 905 } 906 } 907 908 /* 909 * Adjust @level according to the limits of the compression algorithm or 910 * fallback to default 911 */ 912 static unsigned int btrfs_compress_set_level(int type, unsigned level) 913 { 914 const struct btrfs_compress_op *ops = btrfs_compress_op[type]; 915 916 if (level == 0) 917 level = ops->default_level; 918 else 919 level = min(level, ops->max_level); 920 921 return level; 922 } 923 924 /* 925 * Given an address space and start and length, compress the bytes into @pages 926 * that are allocated on demand. 927 * 928 * @type_level is encoded algorithm and level, where level 0 means whatever 929 * default the algorithm chooses and is opaque here; 930 * - compression algo are 0-3 931 * - the level are bits 4-7 932 * 933 * @out_pages is an in/out parameter, holds maximum number of pages to allocate 934 * and returns number of actually allocated pages 935 * 936 * @total_in is used to return the number of bytes actually read. It 937 * may be smaller than the input length if we had to exit early because we 938 * ran out of room in the pages array or because we cross the 939 * max_out threshold. 940 * 941 * @total_out is an in/out parameter, must be set to the input length and will 942 * be also used to return the total number of compressed bytes 943 */ 944 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, 945 u64 start, struct page **pages, 946 unsigned long *out_pages, 947 unsigned long *total_in, 948 unsigned long *total_out) 949 { 950 int type = btrfs_compress_type(type_level); 951 int level = btrfs_compress_level(type_level); 952 struct list_head *workspace; 953 int ret; 954 955 level = btrfs_compress_set_level(type, level); 956 workspace = get_workspace(type, level); 957 ret = compression_compress_pages(type, workspace, mapping, start, pages, 958 out_pages, total_in, total_out); 959 put_workspace(type, workspace); 960 return ret; 961 } 962 963 static int btrfs_decompress_bio(struct compressed_bio *cb) 964 { 965 struct list_head *workspace; 966 int ret; 967 int type = cb->compress_type; 968 969 workspace = get_workspace(type, 0); 970 ret = compression_decompress_bio(workspace, cb); 971 put_workspace(type, workspace); 972 973 return ret; 974 } 975 976 /* 977 * a less complex decompression routine. Our compressed data fits in a 978 * single page, and we want to read a single page out of it. 979 * start_byte tells us the offset into the compressed data we're interested in 980 */ 981 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page, 982 unsigned long start_byte, size_t srclen, size_t destlen) 983 { 984 struct list_head *workspace; 985 int ret; 986 987 workspace = get_workspace(type, 0); 988 ret = compression_decompress(type, workspace, data_in, dest_page, 989 start_byte, srclen, destlen); 990 put_workspace(type, workspace); 991 992 return ret; 993 } 994 995 int __init btrfs_init_compress(void) 996 { 997 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE, 998 offsetof(struct compressed_bio, bbio.bio), 999 BIOSET_NEED_BVECS)) 1000 return -ENOMEM; 1001 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); 1002 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); 1003 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); 1004 zstd_init_workspace_manager(); 1005 return 0; 1006 } 1007 1008 void __cold btrfs_exit_compress(void) 1009 { 1010 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); 1011 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); 1012 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); 1013 zstd_cleanup_workspace_manager(); 1014 bioset_exit(&btrfs_compressed_bioset); 1015 } 1016 1017 /* 1018 * Copy decompressed data from working buffer to pages. 1019 * 1020 * @buf: The decompressed data buffer 1021 * @buf_len: The decompressed data length 1022 * @decompressed: Number of bytes that are already decompressed inside the 1023 * compressed extent 1024 * @cb: The compressed extent descriptor 1025 * @orig_bio: The original bio that the caller wants to read for 1026 * 1027 * An easier to understand graph is like below: 1028 * 1029 * |<- orig_bio ->| |<- orig_bio->| 1030 * |<------- full decompressed extent ----->| 1031 * |<----------- @cb range ---->| 1032 * | |<-- @buf_len -->| 1033 * |<--- @decompressed --->| 1034 * 1035 * Note that, @cb can be a subpage of the full decompressed extent, but 1036 * @cb->start always has the same as the orig_file_offset value of the full 1037 * decompressed extent. 1038 * 1039 * When reading compressed extent, we have to read the full compressed extent, 1040 * while @orig_bio may only want part of the range. 1041 * Thus this function will ensure only data covered by @orig_bio will be copied 1042 * to. 1043 * 1044 * Return 0 if we have copied all needed contents for @orig_bio. 1045 * Return >0 if we need continue decompress. 1046 */ 1047 int btrfs_decompress_buf2page(const char *buf, u32 buf_len, 1048 struct compressed_bio *cb, u32 decompressed) 1049 { 1050 struct bio *orig_bio = cb->orig_bio; 1051 /* Offset inside the full decompressed extent */ 1052 u32 cur_offset; 1053 1054 cur_offset = decompressed; 1055 /* The main loop to do the copy */ 1056 while (cur_offset < decompressed + buf_len) { 1057 struct bio_vec bvec; 1058 size_t copy_len; 1059 u32 copy_start; 1060 /* Offset inside the full decompressed extent */ 1061 u32 bvec_offset; 1062 1063 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); 1064 /* 1065 * cb->start may underflow, but subtracting that value can still 1066 * give us correct offset inside the full decompressed extent. 1067 */ 1068 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; 1069 1070 /* Haven't reached the bvec range, exit */ 1071 if (decompressed + buf_len <= bvec_offset) 1072 return 1; 1073 1074 copy_start = max(cur_offset, bvec_offset); 1075 copy_len = min(bvec_offset + bvec.bv_len, 1076 decompressed + buf_len) - copy_start; 1077 ASSERT(copy_len); 1078 1079 /* 1080 * Extra range check to ensure we didn't go beyond 1081 * @buf + @buf_len. 1082 */ 1083 ASSERT(copy_start - decompressed < buf_len); 1084 memcpy_to_page(bvec.bv_page, bvec.bv_offset, 1085 buf + copy_start - decompressed, copy_len); 1086 cur_offset += copy_len; 1087 1088 bio_advance(orig_bio, copy_len); 1089 /* Finished the bio */ 1090 if (!orig_bio->bi_iter.bi_size) 1091 return 0; 1092 } 1093 return 1; 1094 } 1095 1096 /* 1097 * Shannon Entropy calculation 1098 * 1099 * Pure byte distribution analysis fails to determine compressibility of data. 1100 * Try calculating entropy to estimate the average minimum number of bits 1101 * needed to encode the sampled data. 1102 * 1103 * For convenience, return the percentage of needed bits, instead of amount of 1104 * bits directly. 1105 * 1106 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy 1107 * and can be compressible with high probability 1108 * 1109 * @ENTROPY_LVL_HIGH - data are not compressible with high probability 1110 * 1111 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. 1112 */ 1113 #define ENTROPY_LVL_ACEPTABLE (65) 1114 #define ENTROPY_LVL_HIGH (80) 1115 1116 /* 1117 * For increasead precision in shannon_entropy calculation, 1118 * let's do pow(n, M) to save more digits after comma: 1119 * 1120 * - maximum int bit length is 64 1121 * - ilog2(MAX_SAMPLE_SIZE) -> 13 1122 * - 13 * 4 = 52 < 64 -> M = 4 1123 * 1124 * So use pow(n, 4). 1125 */ 1126 static inline u32 ilog2_w(u64 n) 1127 { 1128 return ilog2(n * n * n * n); 1129 } 1130 1131 static u32 shannon_entropy(struct heuristic_ws *ws) 1132 { 1133 const u32 entropy_max = 8 * ilog2_w(2); 1134 u32 entropy_sum = 0; 1135 u32 p, p_base, sz_base; 1136 u32 i; 1137 1138 sz_base = ilog2_w(ws->sample_size); 1139 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { 1140 p = ws->bucket[i].count; 1141 p_base = ilog2_w(p); 1142 entropy_sum += p * (sz_base - p_base); 1143 } 1144 1145 entropy_sum /= ws->sample_size; 1146 return entropy_sum * 100 / entropy_max; 1147 } 1148 1149 #define RADIX_BASE 4U 1150 #define COUNTERS_SIZE (1U << RADIX_BASE) 1151 1152 static u8 get4bits(u64 num, int shift) { 1153 u8 low4bits; 1154 1155 num >>= shift; 1156 /* Reverse order */ 1157 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); 1158 return low4bits; 1159 } 1160 1161 /* 1162 * Use 4 bits as radix base 1163 * Use 16 u32 counters for calculating new position in buf array 1164 * 1165 * @array - array that will be sorted 1166 * @array_buf - buffer array to store sorting results 1167 * must be equal in size to @array 1168 * @num - array size 1169 */ 1170 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, 1171 int num) 1172 { 1173 u64 max_num; 1174 u64 buf_num; 1175 u32 counters[COUNTERS_SIZE]; 1176 u32 new_addr; 1177 u32 addr; 1178 int bitlen; 1179 int shift; 1180 int i; 1181 1182 /* 1183 * Try avoid useless loop iterations for small numbers stored in big 1184 * counters. Example: 48 33 4 ... in 64bit array 1185 */ 1186 max_num = array[0].count; 1187 for (i = 1; i < num; i++) { 1188 buf_num = array[i].count; 1189 if (buf_num > max_num) 1190 max_num = buf_num; 1191 } 1192 1193 buf_num = ilog2(max_num); 1194 bitlen = ALIGN(buf_num, RADIX_BASE * 2); 1195 1196 shift = 0; 1197 while (shift < bitlen) { 1198 memset(counters, 0, sizeof(counters)); 1199 1200 for (i = 0; i < num; i++) { 1201 buf_num = array[i].count; 1202 addr = get4bits(buf_num, shift); 1203 counters[addr]++; 1204 } 1205 1206 for (i = 1; i < COUNTERS_SIZE; i++) 1207 counters[i] += counters[i - 1]; 1208 1209 for (i = num - 1; i >= 0; i--) { 1210 buf_num = array[i].count; 1211 addr = get4bits(buf_num, shift); 1212 counters[addr]--; 1213 new_addr = counters[addr]; 1214 array_buf[new_addr] = array[i]; 1215 } 1216 1217 shift += RADIX_BASE; 1218 1219 /* 1220 * Normal radix expects to move data from a temporary array, to 1221 * the main one. But that requires some CPU time. Avoid that 1222 * by doing another sort iteration to original array instead of 1223 * memcpy() 1224 */ 1225 memset(counters, 0, sizeof(counters)); 1226 1227 for (i = 0; i < num; i ++) { 1228 buf_num = array_buf[i].count; 1229 addr = get4bits(buf_num, shift); 1230 counters[addr]++; 1231 } 1232 1233 for (i = 1; i < COUNTERS_SIZE; i++) 1234 counters[i] += counters[i - 1]; 1235 1236 for (i = num - 1; i >= 0; i--) { 1237 buf_num = array_buf[i].count; 1238 addr = get4bits(buf_num, shift); 1239 counters[addr]--; 1240 new_addr = counters[addr]; 1241 array[new_addr] = array_buf[i]; 1242 } 1243 1244 shift += RADIX_BASE; 1245 } 1246 } 1247 1248 /* 1249 * Size of the core byte set - how many bytes cover 90% of the sample 1250 * 1251 * There are several types of structured binary data that use nearly all byte 1252 * values. The distribution can be uniform and counts in all buckets will be 1253 * nearly the same (eg. encrypted data). Unlikely to be compressible. 1254 * 1255 * Other possibility is normal (Gaussian) distribution, where the data could 1256 * be potentially compressible, but we have to take a few more steps to decide 1257 * how much. 1258 * 1259 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, 1260 * compression algo can easy fix that 1261 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high 1262 * probability is not compressible 1263 */ 1264 #define BYTE_CORE_SET_LOW (64) 1265 #define BYTE_CORE_SET_HIGH (200) 1266 1267 static int byte_core_set_size(struct heuristic_ws *ws) 1268 { 1269 u32 i; 1270 u32 coreset_sum = 0; 1271 const u32 core_set_threshold = ws->sample_size * 90 / 100; 1272 struct bucket_item *bucket = ws->bucket; 1273 1274 /* Sort in reverse order */ 1275 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); 1276 1277 for (i = 0; i < BYTE_CORE_SET_LOW; i++) 1278 coreset_sum += bucket[i].count; 1279 1280 if (coreset_sum > core_set_threshold) 1281 return i; 1282 1283 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { 1284 coreset_sum += bucket[i].count; 1285 if (coreset_sum > core_set_threshold) 1286 break; 1287 } 1288 1289 return i; 1290 } 1291 1292 /* 1293 * Count byte values in buckets. 1294 * This heuristic can detect textual data (configs, xml, json, html, etc). 1295 * Because in most text-like data byte set is restricted to limited number of 1296 * possible characters, and that restriction in most cases makes data easy to 1297 * compress. 1298 * 1299 * @BYTE_SET_THRESHOLD - consider all data within this byte set size: 1300 * less - compressible 1301 * more - need additional analysis 1302 */ 1303 #define BYTE_SET_THRESHOLD (64) 1304 1305 static u32 byte_set_size(const struct heuristic_ws *ws) 1306 { 1307 u32 i; 1308 u32 byte_set_size = 0; 1309 1310 for (i = 0; i < BYTE_SET_THRESHOLD; i++) { 1311 if (ws->bucket[i].count > 0) 1312 byte_set_size++; 1313 } 1314 1315 /* 1316 * Continue collecting count of byte values in buckets. If the byte 1317 * set size is bigger then the threshold, it's pointless to continue, 1318 * the detection technique would fail for this type of data. 1319 */ 1320 for (; i < BUCKET_SIZE; i++) { 1321 if (ws->bucket[i].count > 0) { 1322 byte_set_size++; 1323 if (byte_set_size > BYTE_SET_THRESHOLD) 1324 return byte_set_size; 1325 } 1326 } 1327 1328 return byte_set_size; 1329 } 1330 1331 static bool sample_repeated_patterns(struct heuristic_ws *ws) 1332 { 1333 const u32 half_of_sample = ws->sample_size / 2; 1334 const u8 *data = ws->sample; 1335 1336 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; 1337 } 1338 1339 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, 1340 struct heuristic_ws *ws) 1341 { 1342 struct page *page; 1343 u64 index, index_end; 1344 u32 i, curr_sample_pos; 1345 u8 *in_data; 1346 1347 /* 1348 * Compression handles the input data by chunks of 128KiB 1349 * (defined by BTRFS_MAX_UNCOMPRESSED) 1350 * 1351 * We do the same for the heuristic and loop over the whole range. 1352 * 1353 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will 1354 * process no more than BTRFS_MAX_UNCOMPRESSED at a time. 1355 */ 1356 if (end - start > BTRFS_MAX_UNCOMPRESSED) 1357 end = start + BTRFS_MAX_UNCOMPRESSED; 1358 1359 index = start >> PAGE_SHIFT; 1360 index_end = end >> PAGE_SHIFT; 1361 1362 /* Don't miss unaligned end */ 1363 if (!PAGE_ALIGNED(end)) 1364 index_end++; 1365 1366 curr_sample_pos = 0; 1367 while (index < index_end) { 1368 page = find_get_page(inode->i_mapping, index); 1369 in_data = kmap_local_page(page); 1370 /* Handle case where the start is not aligned to PAGE_SIZE */ 1371 i = start % PAGE_SIZE; 1372 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { 1373 /* Don't sample any garbage from the last page */ 1374 if (start > end - SAMPLING_READ_SIZE) 1375 break; 1376 memcpy(&ws->sample[curr_sample_pos], &in_data[i], 1377 SAMPLING_READ_SIZE); 1378 i += SAMPLING_INTERVAL; 1379 start += SAMPLING_INTERVAL; 1380 curr_sample_pos += SAMPLING_READ_SIZE; 1381 } 1382 kunmap_local(in_data); 1383 put_page(page); 1384 1385 index++; 1386 } 1387 1388 ws->sample_size = curr_sample_pos; 1389 } 1390 1391 /* 1392 * Compression heuristic. 1393 * 1394 * For now is's a naive and optimistic 'return true', we'll extend the logic to 1395 * quickly (compared to direct compression) detect data characteristics 1396 * (compressible/incompressible) to avoid wasting CPU time on incompressible 1397 * data. 1398 * 1399 * The following types of analysis can be performed: 1400 * - detect mostly zero data 1401 * - detect data with low "byte set" size (text, etc) 1402 * - detect data with low/high "core byte" set 1403 * 1404 * Return non-zero if the compression should be done, 0 otherwise. 1405 */ 1406 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) 1407 { 1408 struct list_head *ws_list = get_workspace(0, 0); 1409 struct heuristic_ws *ws; 1410 u32 i; 1411 u8 byte; 1412 int ret = 0; 1413 1414 ws = list_entry(ws_list, struct heuristic_ws, list); 1415 1416 heuristic_collect_sample(inode, start, end, ws); 1417 1418 if (sample_repeated_patterns(ws)) { 1419 ret = 1; 1420 goto out; 1421 } 1422 1423 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); 1424 1425 for (i = 0; i < ws->sample_size; i++) { 1426 byte = ws->sample[i]; 1427 ws->bucket[byte].count++; 1428 } 1429 1430 i = byte_set_size(ws); 1431 if (i < BYTE_SET_THRESHOLD) { 1432 ret = 2; 1433 goto out; 1434 } 1435 1436 i = byte_core_set_size(ws); 1437 if (i <= BYTE_CORE_SET_LOW) { 1438 ret = 3; 1439 goto out; 1440 } 1441 1442 if (i >= BYTE_CORE_SET_HIGH) { 1443 ret = 0; 1444 goto out; 1445 } 1446 1447 i = shannon_entropy(ws); 1448 if (i <= ENTROPY_LVL_ACEPTABLE) { 1449 ret = 4; 1450 goto out; 1451 } 1452 1453 /* 1454 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be 1455 * needed to give green light to compression. 1456 * 1457 * For now just assume that compression at that level is not worth the 1458 * resources because: 1459 * 1460 * 1. it is possible to defrag the data later 1461 * 1462 * 2. the data would turn out to be hardly compressible, eg. 150 byte 1463 * values, every bucket has counter at level ~54. The heuristic would 1464 * be confused. This can happen when data have some internal repeated 1465 * patterns like "abbacbbc...". This can be detected by analyzing 1466 * pairs of bytes, which is too costly. 1467 */ 1468 if (i < ENTROPY_LVL_HIGH) { 1469 ret = 5; 1470 goto out; 1471 } else { 1472 ret = 0; 1473 goto out; 1474 } 1475 1476 out: 1477 put_workspace(0, ws_list); 1478 return ret; 1479 } 1480 1481 /* 1482 * Convert the compression suffix (eg. after "zlib" starting with ":") to 1483 * level, unrecognized string will set the default level 1484 */ 1485 unsigned int btrfs_compress_str2level(unsigned int type, const char *str) 1486 { 1487 unsigned int level = 0; 1488 int ret; 1489 1490 if (!type) 1491 return 0; 1492 1493 if (str[0] == ':') { 1494 ret = kstrtouint(str + 1, 10, &level); 1495 if (ret) 1496 level = 0; 1497 } 1498 1499 level = btrfs_compress_set_level(type, level); 1500 1501 return level; 1502 } 1503