Loading...
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 <linux/shrinker.h>
24#include <crypto/hash.h>
25#include "misc.h"
26#include "ctree.h"
27#include "fs.h"
28#include "disk-io.h"
29#include "transaction.h"
30#include "btrfs_inode.h"
31#include "bio.h"
32#include "ordered-data.h"
33#include "compression.h"
34#include "extent_io.h"
35#include "extent_map.h"
36#include "subpage.h"
37#include "zoned.h"
38#include "file-item.h"
39#include "super.h"
40
41static struct bio_set btrfs_compressed_bioset;
42
43static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
44
45const char* btrfs_compress_type2str(enum btrfs_compression_type type)
46{
47 switch (type) {
48 case BTRFS_COMPRESS_ZLIB:
49 case BTRFS_COMPRESS_LZO:
50 case BTRFS_COMPRESS_ZSTD:
51 case BTRFS_COMPRESS_NONE:
52 return btrfs_compress_types[type];
53 default:
54 break;
55 }
56
57 return NULL;
58}
59
60static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)
61{
62 return container_of(bbio, struct compressed_bio, bbio);
63}
64
65static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,
66 u64 start, blk_opf_t op,
67 btrfs_bio_end_io_t end_io)
68{
69 struct btrfs_bio *bbio;
70
71 bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,
72 GFP_NOFS, &btrfs_compressed_bioset));
73 btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
74 bbio->inode = inode;
75 bbio->file_offset = start;
76 return to_compressed_bio(bbio);
77}
78
79bool btrfs_compress_is_valid_type(const char *str, size_t len)
80{
81 int i;
82
83 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
84 size_t comp_len = strlen(btrfs_compress_types[i]);
85
86 if (len < comp_len)
87 continue;
88
89 if (!strncmp(btrfs_compress_types[i], str, comp_len))
90 return true;
91 }
92 return false;
93}
94
95static int compression_compress_pages(int type, struct list_head *ws,
96 struct address_space *mapping, u64 start, struct page **pages,
97 unsigned long *out_pages, unsigned long *total_in,
98 unsigned long *total_out)
99{
100 switch (type) {
101 case BTRFS_COMPRESS_ZLIB:
102 return zlib_compress_pages(ws, mapping, start, pages,
103 out_pages, total_in, total_out);
104 case BTRFS_COMPRESS_LZO:
105 return lzo_compress_pages(ws, mapping, start, pages,
106 out_pages, total_in, total_out);
107 case BTRFS_COMPRESS_ZSTD:
108 return zstd_compress_pages(ws, mapping, start, pages,
109 out_pages, total_in, total_out);
110 case BTRFS_COMPRESS_NONE:
111 default:
112 /*
113 * This can happen when compression races with remount setting
114 * it to 'no compress', while caller doesn't call
115 * inode_need_compress() to check if we really need to
116 * compress.
117 *
118 * Not a big deal, just need to inform caller that we
119 * haven't allocated any pages yet.
120 */
121 *out_pages = 0;
122 return -E2BIG;
123 }
124}
125
126static int compression_decompress_bio(struct list_head *ws,
127 struct compressed_bio *cb)
128{
129 switch (cb->compress_type) {
130 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
131 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
132 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
133 case BTRFS_COMPRESS_NONE:
134 default:
135 /*
136 * This can't happen, the type is validated several times
137 * before we get here.
138 */
139 BUG();
140 }
141}
142
143static int compression_decompress(int type, struct list_head *ws,
144 const u8 *data_in, struct page *dest_page,
145 unsigned long dest_pgoff, size_t srclen, size_t destlen)
146{
147 switch (type) {
148 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
149 dest_pgoff, srclen, destlen);
150 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
151 dest_pgoff, srclen, destlen);
152 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
153 dest_pgoff, srclen, destlen);
154 case BTRFS_COMPRESS_NONE:
155 default:
156 /*
157 * This can't happen, the type is validated several times
158 * before we get here.
159 */
160 BUG();
161 }
162}
163
164static void btrfs_free_compressed_pages(struct compressed_bio *cb)
165{
166 for (unsigned int i = 0; i < cb->nr_pages; i++)
167 btrfs_free_compr_page(cb->compressed_pages[i]);
168 kfree(cb->compressed_pages);
169}
170
171static int btrfs_decompress_bio(struct compressed_bio *cb);
172
173/*
174 * Global cache of last unused pages for compression/decompression.
175 */
176static struct btrfs_compr_pool {
177 struct shrinker *shrinker;
178 spinlock_t lock;
179 struct list_head list;
180 int count;
181 int thresh;
182} compr_pool;
183
184static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
185{
186 int ret;
187
188 /*
189 * We must not read the values more than once if 'ret' gets expanded in
190 * the return statement so we don't accidentally return a negative
191 * number, even if the first condition finds it positive.
192 */
193 ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
194
195 return ret > 0 ? ret : 0;
196}
197
198static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
199{
200 struct list_head remove;
201 struct list_head *tmp, *next;
202 int freed;
203
204 if (compr_pool.count == 0)
205 return SHRINK_STOP;
206
207 INIT_LIST_HEAD(&remove);
208
209 /* For now, just simply drain the whole list. */
210 spin_lock(&compr_pool.lock);
211 list_splice_init(&compr_pool.list, &remove);
212 freed = compr_pool.count;
213 compr_pool.count = 0;
214 spin_unlock(&compr_pool.lock);
215
216 list_for_each_safe(tmp, next, &remove) {
217 struct page *page = list_entry(tmp, struct page, lru);
218
219 ASSERT(page_ref_count(page) == 1);
220 put_page(page);
221 }
222
223 return freed;
224}
225
226/*
227 * Common wrappers for page allocation from compression wrappers
228 */
229struct page *btrfs_alloc_compr_page(void)
230{
231 struct page *page = NULL;
232
233 spin_lock(&compr_pool.lock);
234 if (compr_pool.count > 0) {
235 page = list_first_entry(&compr_pool.list, struct page, lru);
236 list_del_init(&page->lru);
237 compr_pool.count--;
238 }
239 spin_unlock(&compr_pool.lock);
240
241 if (page)
242 return page;
243
244 return alloc_page(GFP_NOFS);
245}
246
247void btrfs_free_compr_page(struct page *page)
248{
249 bool do_free = false;
250
251 spin_lock(&compr_pool.lock);
252 if (compr_pool.count > compr_pool.thresh) {
253 do_free = true;
254 } else {
255 list_add(&page->lru, &compr_pool.list);
256 compr_pool.count++;
257 }
258 spin_unlock(&compr_pool.lock);
259
260 if (!do_free)
261 return;
262
263 ASSERT(page_ref_count(page) == 1);
264 put_page(page);
265}
266
267static void end_bbio_comprssed_read(struct btrfs_bio *bbio)
268{
269 struct compressed_bio *cb = to_compressed_bio(bbio);
270 blk_status_t status = bbio->bio.bi_status;
271
272 if (!status)
273 status = errno_to_blk_status(btrfs_decompress_bio(cb));
274
275 btrfs_free_compressed_pages(cb);
276 btrfs_bio_end_io(cb->orig_bbio, status);
277 bio_put(&bbio->bio);
278}
279
280/*
281 * Clear the writeback bits on all of the file
282 * pages for a compressed write
283 */
284static noinline void end_compressed_writeback(const struct compressed_bio *cb)
285{
286 struct inode *inode = &cb->bbio.inode->vfs_inode;
287 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
288 unsigned long index = cb->start >> PAGE_SHIFT;
289 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
290 struct folio_batch fbatch;
291 const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
292 int i;
293 int ret;
294
295 if (error)
296 mapping_set_error(inode->i_mapping, error);
297
298 folio_batch_init(&fbatch);
299 while (index <= end_index) {
300 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
301 &fbatch);
302
303 if (ret == 0)
304 return;
305
306 for (i = 0; i < ret; i++) {
307 struct folio *folio = fbatch.folios[i];
308
309 btrfs_folio_clamp_clear_writeback(fs_info, folio,
310 cb->start, cb->len);
311 }
312 folio_batch_release(&fbatch);
313 }
314 /* the inode may be gone now */
315}
316
317static void btrfs_finish_compressed_write_work(struct work_struct *work)
318{
319 struct compressed_bio *cb =
320 container_of(work, struct compressed_bio, write_end_work);
321
322 btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
323 cb->bbio.bio.bi_status == BLK_STS_OK);
324
325 if (cb->writeback)
326 end_compressed_writeback(cb);
327 /* Note, our inode could be gone now */
328
329 btrfs_free_compressed_pages(cb);
330 bio_put(&cb->bbio.bio);
331}
332
333/*
334 * Do the cleanup once all the compressed pages hit the disk. This will clear
335 * writeback on the file pages and free the compressed pages.
336 *
337 * This also calls the writeback end hooks for the file pages so that metadata
338 * and checksums can be updated in the file.
339 */
340static void end_bbio_comprssed_write(struct btrfs_bio *bbio)
341{
342 struct compressed_bio *cb = to_compressed_bio(bbio);
343 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
344
345 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
346}
347
348static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb)
349{
350 struct bio *bio = &cb->bbio.bio;
351 u32 offset = 0;
352
353 while (offset < cb->compressed_len) {
354 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
355
356 /* Maximum compressed extent is smaller than bio size limit. */
357 __bio_add_page(bio, cb->compressed_pages[offset >> PAGE_SHIFT],
358 len, 0);
359 offset += len;
360 }
361}
362
363/*
364 * worker function to build and submit bios for previously compressed pages.
365 * The corresponding pages in the inode should be marked for writeback
366 * and the compressed pages should have a reference on them for dropping
367 * when the IO is complete.
368 *
369 * This also checksums the file bytes and gets things ready for
370 * the end io hooks.
371 */
372void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
373 struct page **compressed_pages,
374 unsigned int nr_pages,
375 blk_opf_t write_flags,
376 bool writeback)
377{
378 struct btrfs_inode *inode = BTRFS_I(ordered->inode);
379 struct btrfs_fs_info *fs_info = inode->root->fs_info;
380 struct compressed_bio *cb;
381
382 ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
383 ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
384
385 cb = alloc_compressed_bio(inode, ordered->file_offset,
386 REQ_OP_WRITE | write_flags,
387 end_bbio_comprssed_write);
388 cb->start = ordered->file_offset;
389 cb->len = ordered->num_bytes;
390 cb->compressed_pages = compressed_pages;
391 cb->compressed_len = ordered->disk_num_bytes;
392 cb->writeback = writeback;
393 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
394 cb->nr_pages = nr_pages;
395 cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
396 cb->bbio.ordered = ordered;
397 btrfs_add_compressed_bio_pages(cb);
398
399 btrfs_submit_bio(&cb->bbio, 0);
400}
401
402/*
403 * Add extra pages in the same compressed file extent so that we don't need to
404 * re-read the same extent again and again.
405 *
406 * NOTE: this won't work well for subpage, as for subpage read, we lock the
407 * full page then submit bio for each compressed/regular extents.
408 *
409 * This means, if we have several sectors in the same page points to the same
410 * on-disk compressed data, we will re-read the same extent many times and
411 * this function can only help for the next page.
412 */
413static noinline int add_ra_bio_pages(struct inode *inode,
414 u64 compressed_end,
415 struct compressed_bio *cb,
416 int *memstall, unsigned long *pflags)
417{
418 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
419 unsigned long end_index;
420 struct bio *orig_bio = &cb->orig_bbio->bio;
421 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
422 u64 isize = i_size_read(inode);
423 int ret;
424 struct page *page;
425 struct extent_map *em;
426 struct address_space *mapping = inode->i_mapping;
427 struct extent_map_tree *em_tree;
428 struct extent_io_tree *tree;
429 int sectors_missed = 0;
430
431 em_tree = &BTRFS_I(inode)->extent_tree;
432 tree = &BTRFS_I(inode)->io_tree;
433
434 if (isize == 0)
435 return 0;
436
437 /*
438 * For current subpage support, we only support 64K page size,
439 * which means maximum compressed extent size (128K) is just 2x page
440 * size.
441 * This makes readahead less effective, so here disable readahead for
442 * subpage for now, until full compressed write is supported.
443 */
444 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
445 return 0;
446
447 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
448
449 while (cur < compressed_end) {
450 u64 page_end;
451 u64 pg_index = cur >> PAGE_SHIFT;
452 u32 add_size;
453
454 if (pg_index > end_index)
455 break;
456
457 page = xa_load(&mapping->i_pages, pg_index);
458 if (page && !xa_is_value(page)) {
459 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
460 fs_info->sectorsize_bits;
461
462 /* Beyond threshold, no need to continue */
463 if (sectors_missed > 4)
464 break;
465
466 /*
467 * Jump to next page start as we already have page for
468 * current offset.
469 */
470 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
471 continue;
472 }
473
474 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
475 ~__GFP_FS));
476 if (!page)
477 break;
478
479 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
480 put_page(page);
481 /* There is already a page, skip to page end */
482 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
483 continue;
484 }
485
486 if (!*memstall && PageWorkingset(page)) {
487 psi_memstall_enter(pflags);
488 *memstall = 1;
489 }
490
491 ret = set_page_extent_mapped(page);
492 if (ret < 0) {
493 unlock_page(page);
494 put_page(page);
495 break;
496 }
497
498 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
499 lock_extent(tree, cur, page_end, NULL);
500 read_lock(&em_tree->lock);
501 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
502 read_unlock(&em_tree->lock);
503
504 /*
505 * At this point, we have a locked page in the page cache for
506 * these bytes in the file. But, we have to make sure they map
507 * to this compressed extent on disk.
508 */
509 if (!em || cur < em->start ||
510 (cur + fs_info->sectorsize > extent_map_end(em)) ||
511 (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) {
512 free_extent_map(em);
513 unlock_extent(tree, cur, page_end, NULL);
514 unlock_page(page);
515 put_page(page);
516 break;
517 }
518 free_extent_map(em);
519
520 if (page->index == end_index) {
521 size_t zero_offset = offset_in_page(isize);
522
523 if (zero_offset) {
524 int zeros;
525 zeros = PAGE_SIZE - zero_offset;
526 memzero_page(page, zero_offset, zeros);
527 }
528 }
529
530 add_size = min(em->start + em->len, page_end + 1) - cur;
531 ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
532 if (ret != add_size) {
533 unlock_extent(tree, cur, page_end, NULL);
534 unlock_page(page);
535 put_page(page);
536 break;
537 }
538 /*
539 * If it's subpage, we also need to increase its
540 * subpage::readers number, as at endio we will decrease
541 * subpage::readers and to unlock the page.
542 */
543 if (fs_info->sectorsize < PAGE_SIZE)
544 btrfs_subpage_start_reader(fs_info, page_folio(page),
545 cur, add_size);
546 put_page(page);
547 cur += add_size;
548 }
549 return 0;
550}
551
552/*
553 * for a compressed read, the bio we get passed has all the inode pages
554 * in it. We don't actually do IO on those pages but allocate new ones
555 * to hold the compressed pages on disk.
556 *
557 * bio->bi_iter.bi_sector points to the compressed extent on disk
558 * bio->bi_io_vec points to all of the inode pages
559 *
560 * After the compressed pages are read, we copy the bytes into the
561 * bio we were passed and then call the bio end_io calls
562 */
563void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
564{
565 struct btrfs_inode *inode = bbio->inode;
566 struct btrfs_fs_info *fs_info = inode->root->fs_info;
567 struct extent_map_tree *em_tree = &inode->extent_tree;
568 struct compressed_bio *cb;
569 unsigned int compressed_len;
570 u64 file_offset = bbio->file_offset;
571 u64 em_len;
572 u64 em_start;
573 struct extent_map *em;
574 unsigned long pflags;
575 int memstall = 0;
576 blk_status_t ret;
577 int ret2;
578
579 /* we need the actual starting offset of this extent in the file */
580 read_lock(&em_tree->lock);
581 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
582 read_unlock(&em_tree->lock);
583 if (!em) {
584 ret = BLK_STS_IOERR;
585 goto out;
586 }
587
588 ASSERT(extent_map_is_compressed(em));
589 compressed_len = em->block_len;
590
591 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
592 end_bbio_comprssed_read);
593
594 cb->start = em->orig_start;
595 em_len = em->len;
596 em_start = em->start;
597
598 cb->len = bbio->bio.bi_iter.bi_size;
599 cb->compressed_len = compressed_len;
600 cb->compress_type = extent_map_compression(em);
601 cb->orig_bbio = bbio;
602
603 free_extent_map(em);
604
605 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
606 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
607 if (!cb->compressed_pages) {
608 ret = BLK_STS_RESOURCE;
609 goto out_free_bio;
610 }
611
612 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages, 0);
613 if (ret2) {
614 ret = BLK_STS_RESOURCE;
615 goto out_free_compressed_pages;
616 }
617
618 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
619 &pflags);
620
621 /* include any pages we added in add_ra-bio_pages */
622 cb->len = bbio->bio.bi_iter.bi_size;
623 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
624 btrfs_add_compressed_bio_pages(cb);
625
626 if (memstall)
627 psi_memstall_leave(&pflags);
628
629 btrfs_submit_bio(&cb->bbio, 0);
630 return;
631
632out_free_compressed_pages:
633 kfree(cb->compressed_pages);
634out_free_bio:
635 bio_put(&cb->bbio.bio);
636out:
637 btrfs_bio_end_io(bbio, ret);
638}
639
640/*
641 * Heuristic uses systematic sampling to collect data from the input data
642 * range, the logic can be tuned by the following constants:
643 *
644 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
645 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
646 */
647#define SAMPLING_READ_SIZE (16)
648#define SAMPLING_INTERVAL (256)
649
650/*
651 * For statistical analysis of the input data we consider bytes that form a
652 * Galois Field of 256 objects. Each object has an attribute count, ie. how
653 * many times the object appeared in the sample.
654 */
655#define BUCKET_SIZE (256)
656
657/*
658 * The size of the sample is based on a statistical sampling rule of thumb.
659 * The common way is to perform sampling tests as long as the number of
660 * elements in each cell is at least 5.
661 *
662 * Instead of 5, we choose 32 to obtain more accurate results.
663 * If the data contain the maximum number of symbols, which is 256, we obtain a
664 * sample size bound by 8192.
665 *
666 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
667 * from up to 512 locations.
668 */
669#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
670 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
671
672struct bucket_item {
673 u32 count;
674};
675
676struct heuristic_ws {
677 /* Partial copy of input data */
678 u8 *sample;
679 u32 sample_size;
680 /* Buckets store counters for each byte value */
681 struct bucket_item *bucket;
682 /* Sorting buffer */
683 struct bucket_item *bucket_b;
684 struct list_head list;
685};
686
687static struct workspace_manager heuristic_wsm;
688
689static void free_heuristic_ws(struct list_head *ws)
690{
691 struct heuristic_ws *workspace;
692
693 workspace = list_entry(ws, struct heuristic_ws, list);
694
695 kvfree(workspace->sample);
696 kfree(workspace->bucket);
697 kfree(workspace->bucket_b);
698 kfree(workspace);
699}
700
701static struct list_head *alloc_heuristic_ws(unsigned int level)
702{
703 struct heuristic_ws *ws;
704
705 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
706 if (!ws)
707 return ERR_PTR(-ENOMEM);
708
709 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
710 if (!ws->sample)
711 goto fail;
712
713 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
714 if (!ws->bucket)
715 goto fail;
716
717 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
718 if (!ws->bucket_b)
719 goto fail;
720
721 INIT_LIST_HEAD(&ws->list);
722 return &ws->list;
723fail:
724 free_heuristic_ws(&ws->list);
725 return ERR_PTR(-ENOMEM);
726}
727
728const struct btrfs_compress_op btrfs_heuristic_compress = {
729 .workspace_manager = &heuristic_wsm,
730};
731
732static const struct btrfs_compress_op * const btrfs_compress_op[] = {
733 /* The heuristic is represented as compression type 0 */
734 &btrfs_heuristic_compress,
735 &btrfs_zlib_compress,
736 &btrfs_lzo_compress,
737 &btrfs_zstd_compress,
738};
739
740static struct list_head *alloc_workspace(int type, unsigned int level)
741{
742 switch (type) {
743 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
744 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
745 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
746 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
747 default:
748 /*
749 * This can't happen, the type is validated several times
750 * before we get here.
751 */
752 BUG();
753 }
754}
755
756static void free_workspace(int type, struct list_head *ws)
757{
758 switch (type) {
759 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
760 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
761 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
762 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
763 default:
764 /*
765 * This can't happen, the type is validated several times
766 * before we get here.
767 */
768 BUG();
769 }
770}
771
772static void btrfs_init_workspace_manager(int type)
773{
774 struct workspace_manager *wsm;
775 struct list_head *workspace;
776
777 wsm = btrfs_compress_op[type]->workspace_manager;
778 INIT_LIST_HEAD(&wsm->idle_ws);
779 spin_lock_init(&wsm->ws_lock);
780 atomic_set(&wsm->total_ws, 0);
781 init_waitqueue_head(&wsm->ws_wait);
782
783 /*
784 * Preallocate one workspace for each compression type so we can
785 * guarantee forward progress in the worst case
786 */
787 workspace = alloc_workspace(type, 0);
788 if (IS_ERR(workspace)) {
789 pr_warn(
790 "BTRFS: cannot preallocate compression workspace, will try later\n");
791 } else {
792 atomic_set(&wsm->total_ws, 1);
793 wsm->free_ws = 1;
794 list_add(workspace, &wsm->idle_ws);
795 }
796}
797
798static void btrfs_cleanup_workspace_manager(int type)
799{
800 struct workspace_manager *wsman;
801 struct list_head *ws;
802
803 wsman = btrfs_compress_op[type]->workspace_manager;
804 while (!list_empty(&wsman->idle_ws)) {
805 ws = wsman->idle_ws.next;
806 list_del(ws);
807 free_workspace(type, ws);
808 atomic_dec(&wsman->total_ws);
809 }
810}
811
812/*
813 * This finds an available workspace or allocates a new one.
814 * If it's not possible to allocate a new one, waits until there's one.
815 * Preallocation makes a forward progress guarantees and we do not return
816 * errors.
817 */
818struct list_head *btrfs_get_workspace(int type, unsigned int level)
819{
820 struct workspace_manager *wsm;
821 struct list_head *workspace;
822 int cpus = num_online_cpus();
823 unsigned nofs_flag;
824 struct list_head *idle_ws;
825 spinlock_t *ws_lock;
826 atomic_t *total_ws;
827 wait_queue_head_t *ws_wait;
828 int *free_ws;
829
830 wsm = btrfs_compress_op[type]->workspace_manager;
831 idle_ws = &wsm->idle_ws;
832 ws_lock = &wsm->ws_lock;
833 total_ws = &wsm->total_ws;
834 ws_wait = &wsm->ws_wait;
835 free_ws = &wsm->free_ws;
836
837again:
838 spin_lock(ws_lock);
839 if (!list_empty(idle_ws)) {
840 workspace = idle_ws->next;
841 list_del(workspace);
842 (*free_ws)--;
843 spin_unlock(ws_lock);
844 return workspace;
845
846 }
847 if (atomic_read(total_ws) > cpus) {
848 DEFINE_WAIT(wait);
849
850 spin_unlock(ws_lock);
851 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
852 if (atomic_read(total_ws) > cpus && !*free_ws)
853 schedule();
854 finish_wait(ws_wait, &wait);
855 goto again;
856 }
857 atomic_inc(total_ws);
858 spin_unlock(ws_lock);
859
860 /*
861 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
862 * to turn it off here because we might get called from the restricted
863 * context of btrfs_compress_bio/btrfs_compress_pages
864 */
865 nofs_flag = memalloc_nofs_save();
866 workspace = alloc_workspace(type, level);
867 memalloc_nofs_restore(nofs_flag);
868
869 if (IS_ERR(workspace)) {
870 atomic_dec(total_ws);
871 wake_up(ws_wait);
872
873 /*
874 * Do not return the error but go back to waiting. There's a
875 * workspace preallocated for each type and the compression
876 * time is bounded so we get to a workspace eventually. This
877 * makes our caller's life easier.
878 *
879 * To prevent silent and low-probability deadlocks (when the
880 * initial preallocation fails), check if there are any
881 * workspaces at all.
882 */
883 if (atomic_read(total_ws) == 0) {
884 static DEFINE_RATELIMIT_STATE(_rs,
885 /* once per minute */ 60 * HZ,
886 /* no burst */ 1);
887
888 if (__ratelimit(&_rs)) {
889 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
890 }
891 }
892 goto again;
893 }
894 return workspace;
895}
896
897static struct list_head *get_workspace(int type, int level)
898{
899 switch (type) {
900 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
901 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
902 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
903 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
904 default:
905 /*
906 * This can't happen, the type is validated several times
907 * before we get here.
908 */
909 BUG();
910 }
911}
912
913/*
914 * put a workspace struct back on the list or free it if we have enough
915 * idle ones sitting around
916 */
917void btrfs_put_workspace(int type, struct list_head *ws)
918{
919 struct workspace_manager *wsm;
920 struct list_head *idle_ws;
921 spinlock_t *ws_lock;
922 atomic_t *total_ws;
923 wait_queue_head_t *ws_wait;
924 int *free_ws;
925
926 wsm = btrfs_compress_op[type]->workspace_manager;
927 idle_ws = &wsm->idle_ws;
928 ws_lock = &wsm->ws_lock;
929 total_ws = &wsm->total_ws;
930 ws_wait = &wsm->ws_wait;
931 free_ws = &wsm->free_ws;
932
933 spin_lock(ws_lock);
934 if (*free_ws <= num_online_cpus()) {
935 list_add(ws, idle_ws);
936 (*free_ws)++;
937 spin_unlock(ws_lock);
938 goto wake;
939 }
940 spin_unlock(ws_lock);
941
942 free_workspace(type, ws);
943 atomic_dec(total_ws);
944wake:
945 cond_wake_up(ws_wait);
946}
947
948static void put_workspace(int type, struct list_head *ws)
949{
950 switch (type) {
951 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
952 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
953 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
954 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
955 default:
956 /*
957 * This can't happen, the type is validated several times
958 * before we get here.
959 */
960 BUG();
961 }
962}
963
964/*
965 * Adjust @level according to the limits of the compression algorithm or
966 * fallback to default
967 */
968static unsigned int btrfs_compress_set_level(int type, unsigned level)
969{
970 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
971
972 if (level == 0)
973 level = ops->default_level;
974 else
975 level = min(level, ops->max_level);
976
977 return level;
978}
979
980/*
981 * Given an address space and start and length, compress the bytes into @pages
982 * that are allocated on demand.
983 *
984 * @type_level is encoded algorithm and level, where level 0 means whatever
985 * default the algorithm chooses and is opaque here;
986 * - compression algo are 0-3
987 * - the level are bits 4-7
988 *
989 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
990 * and returns number of actually allocated pages
991 *
992 * @total_in is used to return the number of bytes actually read. It
993 * may be smaller than the input length if we had to exit early because we
994 * ran out of room in the pages array or because we cross the
995 * max_out threshold.
996 *
997 * @total_out is an in/out parameter, must be set to the input length and will
998 * be also used to return the total number of compressed bytes
999 */
1000int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1001 u64 start, struct page **pages,
1002 unsigned long *out_pages,
1003 unsigned long *total_in,
1004 unsigned long *total_out)
1005{
1006 int type = btrfs_compress_type(type_level);
1007 int level = btrfs_compress_level(type_level);
1008 struct list_head *workspace;
1009 int ret;
1010
1011 level = btrfs_compress_set_level(type, level);
1012 workspace = get_workspace(type, level);
1013 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1014 out_pages, total_in, total_out);
1015 put_workspace(type, workspace);
1016 return ret;
1017}
1018
1019static int btrfs_decompress_bio(struct compressed_bio *cb)
1020{
1021 struct list_head *workspace;
1022 int ret;
1023 int type = cb->compress_type;
1024
1025 workspace = get_workspace(type, 0);
1026 ret = compression_decompress_bio(workspace, cb);
1027 put_workspace(type, workspace);
1028
1029 if (!ret)
1030 zero_fill_bio(&cb->orig_bbio->bio);
1031 return ret;
1032}
1033
1034/*
1035 * a less complex decompression routine. Our compressed data fits in a
1036 * single page, and we want to read a single page out of it.
1037 * start_byte tells us the offset into the compressed data we're interested in
1038 */
1039int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1040 unsigned long dest_pgoff, size_t srclen, size_t destlen)
1041{
1042 struct btrfs_fs_info *fs_info = btrfs_sb(dest_page->mapping->host->i_sb);
1043 struct list_head *workspace;
1044 const u32 sectorsize = fs_info->sectorsize;
1045 int ret;
1046
1047 /*
1048 * The full destination page range should not exceed the page size.
1049 * And the @destlen should not exceed sectorsize, as this is only called for
1050 * inline file extents, which should not exceed sectorsize.
1051 */
1052 ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);
1053
1054 workspace = get_workspace(type, 0);
1055 ret = compression_decompress(type, workspace, data_in, dest_page,
1056 dest_pgoff, srclen, destlen);
1057 put_workspace(type, workspace);
1058
1059 return ret;
1060}
1061
1062int __init btrfs_init_compress(void)
1063{
1064 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1065 offsetof(struct compressed_bio, bbio.bio),
1066 BIOSET_NEED_BVECS))
1067 return -ENOMEM;
1068
1069 compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1070 if (!compr_pool.shrinker)
1071 return -ENOMEM;
1072
1073 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1074 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1075 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1076 zstd_init_workspace_manager();
1077
1078 spin_lock_init(&compr_pool.lock);
1079 INIT_LIST_HEAD(&compr_pool.list);
1080 compr_pool.count = 0;
1081 /* 128K / 4K = 32, for 8 threads is 256 pages. */
1082 compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1083 compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1084 compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1085 compr_pool.shrinker->batch = 32;
1086 compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1087 shrinker_register(compr_pool.shrinker);
1088
1089 return 0;
1090}
1091
1092void __cold btrfs_exit_compress(void)
1093{
1094 /* For now scan drains all pages and does not touch the parameters. */
1095 btrfs_compr_pool_scan(NULL, NULL);
1096 shrinker_free(compr_pool.shrinker);
1097
1098 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1099 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1100 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1101 zstd_cleanup_workspace_manager();
1102 bioset_exit(&btrfs_compressed_bioset);
1103}
1104
1105/*
1106 * Copy decompressed data from working buffer to pages.
1107 *
1108 * @buf: The decompressed data buffer
1109 * @buf_len: The decompressed data length
1110 * @decompressed: Number of bytes that are already decompressed inside the
1111 * compressed extent
1112 * @cb: The compressed extent descriptor
1113 * @orig_bio: The original bio that the caller wants to read for
1114 *
1115 * An easier to understand graph is like below:
1116 *
1117 * |<- orig_bio ->| |<- orig_bio->|
1118 * |<------- full decompressed extent ----->|
1119 * |<----------- @cb range ---->|
1120 * | |<-- @buf_len -->|
1121 * |<--- @decompressed --->|
1122 *
1123 * Note that, @cb can be a subpage of the full decompressed extent, but
1124 * @cb->start always has the same as the orig_file_offset value of the full
1125 * decompressed extent.
1126 *
1127 * When reading compressed extent, we have to read the full compressed extent,
1128 * while @orig_bio may only want part of the range.
1129 * Thus this function will ensure only data covered by @orig_bio will be copied
1130 * to.
1131 *
1132 * Return 0 if we have copied all needed contents for @orig_bio.
1133 * Return >0 if we need continue decompress.
1134 */
1135int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1136 struct compressed_bio *cb, u32 decompressed)
1137{
1138 struct bio *orig_bio = &cb->orig_bbio->bio;
1139 /* Offset inside the full decompressed extent */
1140 u32 cur_offset;
1141
1142 cur_offset = decompressed;
1143 /* The main loop to do the copy */
1144 while (cur_offset < decompressed + buf_len) {
1145 struct bio_vec bvec;
1146 size_t copy_len;
1147 u32 copy_start;
1148 /* Offset inside the full decompressed extent */
1149 u32 bvec_offset;
1150
1151 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1152 /*
1153 * cb->start may underflow, but subtracting that value can still
1154 * give us correct offset inside the full decompressed extent.
1155 */
1156 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1157
1158 /* Haven't reached the bvec range, exit */
1159 if (decompressed + buf_len <= bvec_offset)
1160 return 1;
1161
1162 copy_start = max(cur_offset, bvec_offset);
1163 copy_len = min(bvec_offset + bvec.bv_len,
1164 decompressed + buf_len) - copy_start;
1165 ASSERT(copy_len);
1166
1167 /*
1168 * Extra range check to ensure we didn't go beyond
1169 * @buf + @buf_len.
1170 */
1171 ASSERT(copy_start - decompressed < buf_len);
1172 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1173 buf + copy_start - decompressed, copy_len);
1174 cur_offset += copy_len;
1175
1176 bio_advance(orig_bio, copy_len);
1177 /* Finished the bio */
1178 if (!orig_bio->bi_iter.bi_size)
1179 return 0;
1180 }
1181 return 1;
1182}
1183
1184/*
1185 * Shannon Entropy calculation
1186 *
1187 * Pure byte distribution analysis fails to determine compressibility of data.
1188 * Try calculating entropy to estimate the average minimum number of bits
1189 * needed to encode the sampled data.
1190 *
1191 * For convenience, return the percentage of needed bits, instead of amount of
1192 * bits directly.
1193 *
1194 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1195 * and can be compressible with high probability
1196 *
1197 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1198 *
1199 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1200 */
1201#define ENTROPY_LVL_ACEPTABLE (65)
1202#define ENTROPY_LVL_HIGH (80)
1203
1204/*
1205 * For increasead precision in shannon_entropy calculation,
1206 * let's do pow(n, M) to save more digits after comma:
1207 *
1208 * - maximum int bit length is 64
1209 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1210 * - 13 * 4 = 52 < 64 -> M = 4
1211 *
1212 * So use pow(n, 4).
1213 */
1214static inline u32 ilog2_w(u64 n)
1215{
1216 return ilog2(n * n * n * n);
1217}
1218
1219static u32 shannon_entropy(struct heuristic_ws *ws)
1220{
1221 const u32 entropy_max = 8 * ilog2_w(2);
1222 u32 entropy_sum = 0;
1223 u32 p, p_base, sz_base;
1224 u32 i;
1225
1226 sz_base = ilog2_w(ws->sample_size);
1227 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1228 p = ws->bucket[i].count;
1229 p_base = ilog2_w(p);
1230 entropy_sum += p * (sz_base - p_base);
1231 }
1232
1233 entropy_sum /= ws->sample_size;
1234 return entropy_sum * 100 / entropy_max;
1235}
1236
1237#define RADIX_BASE 4U
1238#define COUNTERS_SIZE (1U << RADIX_BASE)
1239
1240static u8 get4bits(u64 num, int shift) {
1241 u8 low4bits;
1242
1243 num >>= shift;
1244 /* Reverse order */
1245 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1246 return low4bits;
1247}
1248
1249/*
1250 * Use 4 bits as radix base
1251 * Use 16 u32 counters for calculating new position in buf array
1252 *
1253 * @array - array that will be sorted
1254 * @array_buf - buffer array to store sorting results
1255 * must be equal in size to @array
1256 * @num - array size
1257 */
1258static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1259 int num)
1260{
1261 u64 max_num;
1262 u64 buf_num;
1263 u32 counters[COUNTERS_SIZE];
1264 u32 new_addr;
1265 u32 addr;
1266 int bitlen;
1267 int shift;
1268 int i;
1269
1270 /*
1271 * Try avoid useless loop iterations for small numbers stored in big
1272 * counters. Example: 48 33 4 ... in 64bit array
1273 */
1274 max_num = array[0].count;
1275 for (i = 1; i < num; i++) {
1276 buf_num = array[i].count;
1277 if (buf_num > max_num)
1278 max_num = buf_num;
1279 }
1280
1281 buf_num = ilog2(max_num);
1282 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1283
1284 shift = 0;
1285 while (shift < bitlen) {
1286 memset(counters, 0, sizeof(counters));
1287
1288 for (i = 0; i < num; i++) {
1289 buf_num = array[i].count;
1290 addr = get4bits(buf_num, shift);
1291 counters[addr]++;
1292 }
1293
1294 for (i = 1; i < COUNTERS_SIZE; i++)
1295 counters[i] += counters[i - 1];
1296
1297 for (i = num - 1; i >= 0; i--) {
1298 buf_num = array[i].count;
1299 addr = get4bits(buf_num, shift);
1300 counters[addr]--;
1301 new_addr = counters[addr];
1302 array_buf[new_addr] = array[i];
1303 }
1304
1305 shift += RADIX_BASE;
1306
1307 /*
1308 * Normal radix expects to move data from a temporary array, to
1309 * the main one. But that requires some CPU time. Avoid that
1310 * by doing another sort iteration to original array instead of
1311 * memcpy()
1312 */
1313 memset(counters, 0, sizeof(counters));
1314
1315 for (i = 0; i < num; i ++) {
1316 buf_num = array_buf[i].count;
1317 addr = get4bits(buf_num, shift);
1318 counters[addr]++;
1319 }
1320
1321 for (i = 1; i < COUNTERS_SIZE; i++)
1322 counters[i] += counters[i - 1];
1323
1324 for (i = num - 1; i >= 0; i--) {
1325 buf_num = array_buf[i].count;
1326 addr = get4bits(buf_num, shift);
1327 counters[addr]--;
1328 new_addr = counters[addr];
1329 array[new_addr] = array_buf[i];
1330 }
1331
1332 shift += RADIX_BASE;
1333 }
1334}
1335
1336/*
1337 * Size of the core byte set - how many bytes cover 90% of the sample
1338 *
1339 * There are several types of structured binary data that use nearly all byte
1340 * values. The distribution can be uniform and counts in all buckets will be
1341 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1342 *
1343 * Other possibility is normal (Gaussian) distribution, where the data could
1344 * be potentially compressible, but we have to take a few more steps to decide
1345 * how much.
1346 *
1347 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1348 * compression algo can easy fix that
1349 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1350 * probability is not compressible
1351 */
1352#define BYTE_CORE_SET_LOW (64)
1353#define BYTE_CORE_SET_HIGH (200)
1354
1355static int byte_core_set_size(struct heuristic_ws *ws)
1356{
1357 u32 i;
1358 u32 coreset_sum = 0;
1359 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1360 struct bucket_item *bucket = ws->bucket;
1361
1362 /* Sort in reverse order */
1363 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1364
1365 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1366 coreset_sum += bucket[i].count;
1367
1368 if (coreset_sum > core_set_threshold)
1369 return i;
1370
1371 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1372 coreset_sum += bucket[i].count;
1373 if (coreset_sum > core_set_threshold)
1374 break;
1375 }
1376
1377 return i;
1378}
1379
1380/*
1381 * Count byte values in buckets.
1382 * This heuristic can detect textual data (configs, xml, json, html, etc).
1383 * Because in most text-like data byte set is restricted to limited number of
1384 * possible characters, and that restriction in most cases makes data easy to
1385 * compress.
1386 *
1387 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1388 * less - compressible
1389 * more - need additional analysis
1390 */
1391#define BYTE_SET_THRESHOLD (64)
1392
1393static u32 byte_set_size(const struct heuristic_ws *ws)
1394{
1395 u32 i;
1396 u32 byte_set_size = 0;
1397
1398 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1399 if (ws->bucket[i].count > 0)
1400 byte_set_size++;
1401 }
1402
1403 /*
1404 * Continue collecting count of byte values in buckets. If the byte
1405 * set size is bigger then the threshold, it's pointless to continue,
1406 * the detection technique would fail for this type of data.
1407 */
1408 for (; i < BUCKET_SIZE; i++) {
1409 if (ws->bucket[i].count > 0) {
1410 byte_set_size++;
1411 if (byte_set_size > BYTE_SET_THRESHOLD)
1412 return byte_set_size;
1413 }
1414 }
1415
1416 return byte_set_size;
1417}
1418
1419static bool sample_repeated_patterns(struct heuristic_ws *ws)
1420{
1421 const u32 half_of_sample = ws->sample_size / 2;
1422 const u8 *data = ws->sample;
1423
1424 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1425}
1426
1427static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1428 struct heuristic_ws *ws)
1429{
1430 struct page *page;
1431 u64 index, index_end;
1432 u32 i, curr_sample_pos;
1433 u8 *in_data;
1434
1435 /*
1436 * Compression handles the input data by chunks of 128KiB
1437 * (defined by BTRFS_MAX_UNCOMPRESSED)
1438 *
1439 * We do the same for the heuristic and loop over the whole range.
1440 *
1441 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1442 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1443 */
1444 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1445 end = start + BTRFS_MAX_UNCOMPRESSED;
1446
1447 index = start >> PAGE_SHIFT;
1448 index_end = end >> PAGE_SHIFT;
1449
1450 /* Don't miss unaligned end */
1451 if (!PAGE_ALIGNED(end))
1452 index_end++;
1453
1454 curr_sample_pos = 0;
1455 while (index < index_end) {
1456 page = find_get_page(inode->i_mapping, index);
1457 in_data = kmap_local_page(page);
1458 /* Handle case where the start is not aligned to PAGE_SIZE */
1459 i = start % PAGE_SIZE;
1460 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1461 /* Don't sample any garbage from the last page */
1462 if (start > end - SAMPLING_READ_SIZE)
1463 break;
1464 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1465 SAMPLING_READ_SIZE);
1466 i += SAMPLING_INTERVAL;
1467 start += SAMPLING_INTERVAL;
1468 curr_sample_pos += SAMPLING_READ_SIZE;
1469 }
1470 kunmap_local(in_data);
1471 put_page(page);
1472
1473 index++;
1474 }
1475
1476 ws->sample_size = curr_sample_pos;
1477}
1478
1479/*
1480 * Compression heuristic.
1481 *
1482 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1483 * quickly (compared to direct compression) detect data characteristics
1484 * (compressible/incompressible) to avoid wasting CPU time on incompressible
1485 * data.
1486 *
1487 * The following types of analysis can be performed:
1488 * - detect mostly zero data
1489 * - detect data with low "byte set" size (text, etc)
1490 * - detect data with low/high "core byte" set
1491 *
1492 * Return non-zero if the compression should be done, 0 otherwise.
1493 */
1494int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1495{
1496 struct list_head *ws_list = get_workspace(0, 0);
1497 struct heuristic_ws *ws;
1498 u32 i;
1499 u8 byte;
1500 int ret = 0;
1501
1502 ws = list_entry(ws_list, struct heuristic_ws, list);
1503
1504 heuristic_collect_sample(inode, start, end, ws);
1505
1506 if (sample_repeated_patterns(ws)) {
1507 ret = 1;
1508 goto out;
1509 }
1510
1511 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1512
1513 for (i = 0; i < ws->sample_size; i++) {
1514 byte = ws->sample[i];
1515 ws->bucket[byte].count++;
1516 }
1517
1518 i = byte_set_size(ws);
1519 if (i < BYTE_SET_THRESHOLD) {
1520 ret = 2;
1521 goto out;
1522 }
1523
1524 i = byte_core_set_size(ws);
1525 if (i <= BYTE_CORE_SET_LOW) {
1526 ret = 3;
1527 goto out;
1528 }
1529
1530 if (i >= BYTE_CORE_SET_HIGH) {
1531 ret = 0;
1532 goto out;
1533 }
1534
1535 i = shannon_entropy(ws);
1536 if (i <= ENTROPY_LVL_ACEPTABLE) {
1537 ret = 4;
1538 goto out;
1539 }
1540
1541 /*
1542 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1543 * needed to give green light to compression.
1544 *
1545 * For now just assume that compression at that level is not worth the
1546 * resources because:
1547 *
1548 * 1. it is possible to defrag the data later
1549 *
1550 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1551 * values, every bucket has counter at level ~54. The heuristic would
1552 * be confused. This can happen when data have some internal repeated
1553 * patterns like "abbacbbc...". This can be detected by analyzing
1554 * pairs of bytes, which is too costly.
1555 */
1556 if (i < ENTROPY_LVL_HIGH) {
1557 ret = 5;
1558 goto out;
1559 } else {
1560 ret = 0;
1561 goto out;
1562 }
1563
1564out:
1565 put_workspace(0, ws_list);
1566 return ret;
1567}
1568
1569/*
1570 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1571 * level, unrecognized string will set the default level
1572 */
1573unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1574{
1575 unsigned int level = 0;
1576 int ret;
1577
1578 if (!type)
1579 return 0;
1580
1581 if (str[0] == ':') {
1582 ret = kstrtouint(str + 1, 10, &level);
1583 if (ret)
1584 level = 0;
1585 }
1586
1587 level = btrfs_compress_set_level(type, level);
1588
1589 return level;
1590}
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
40static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
41
42const char* btrfs_compress_type2str(enum btrfs_compression_type type)
43{
44 switch (type) {
45 case BTRFS_COMPRESS_ZLIB:
46 case BTRFS_COMPRESS_LZO:
47 case BTRFS_COMPRESS_ZSTD:
48 case BTRFS_COMPRESS_NONE:
49 return btrfs_compress_types[type];
50 default:
51 break;
52 }
53
54 return NULL;
55}
56
57bool btrfs_compress_is_valid_type(const char *str, size_t len)
58{
59 int i;
60
61 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
62 size_t comp_len = strlen(btrfs_compress_types[i]);
63
64 if (len < comp_len)
65 continue;
66
67 if (!strncmp(btrfs_compress_types[i], str, comp_len))
68 return true;
69 }
70 return false;
71}
72
73static int compression_compress_pages(int type, struct list_head *ws,
74 struct address_space *mapping, u64 start, struct page **pages,
75 unsigned long *out_pages, unsigned long *total_in,
76 unsigned long *total_out)
77{
78 switch (type) {
79 case BTRFS_COMPRESS_ZLIB:
80 return zlib_compress_pages(ws, mapping, start, pages,
81 out_pages, total_in, total_out);
82 case BTRFS_COMPRESS_LZO:
83 return lzo_compress_pages(ws, mapping, start, pages,
84 out_pages, total_in, total_out);
85 case BTRFS_COMPRESS_ZSTD:
86 return zstd_compress_pages(ws, mapping, start, pages,
87 out_pages, total_in, total_out);
88 case BTRFS_COMPRESS_NONE:
89 default:
90 /*
91 * This can happen when compression races with remount setting
92 * it to 'no compress', while caller doesn't call
93 * inode_need_compress() to check if we really need to
94 * compress.
95 *
96 * Not a big deal, just need to inform caller that we
97 * haven't allocated any pages yet.
98 */
99 *out_pages = 0;
100 return -E2BIG;
101 }
102}
103
104static int compression_decompress_bio(struct list_head *ws,
105 struct compressed_bio *cb)
106{
107 switch (cb->compress_type) {
108 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
109 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
110 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
111 case BTRFS_COMPRESS_NONE:
112 default:
113 /*
114 * This can't happen, the type is validated several times
115 * before we get here.
116 */
117 BUG();
118 }
119}
120
121static int compression_decompress(int type, struct list_head *ws,
122 const u8 *data_in, struct page *dest_page,
123 unsigned long start_byte, size_t srclen, size_t destlen)
124{
125 switch (type) {
126 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
127 start_byte, srclen, destlen);
128 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
129 start_byte, srclen, destlen);
130 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
131 start_byte, srclen, destlen);
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
142static int btrfs_decompress_bio(struct compressed_bio *cb);
143
144static void finish_compressed_bio_read(struct compressed_bio *cb)
145{
146 unsigned int index;
147 struct page *page;
148
149 if (cb->status == BLK_STS_OK)
150 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
151
152 /* Release the compressed pages */
153 for (index = 0; index < cb->nr_pages; index++) {
154 page = cb->compressed_pages[index];
155 page->mapping = NULL;
156 put_page(page);
157 }
158
159 /* Do io completion on the original bio */
160 btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
161
162 /* Finally free the cb struct */
163 kfree(cb->compressed_pages);
164 kfree(cb);
165}
166
167/*
168 * Verify the checksums and kick off repair if needed on the uncompressed data
169 * before decompressing it into the original bio and freeing the uncompressed
170 * pages.
171 */
172static void end_compressed_bio_read(struct btrfs_bio *bbio)
173{
174 struct compressed_bio *cb = bbio->private;
175 struct inode *inode = cb->inode;
176 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
177 struct btrfs_inode *bi = BTRFS_I(inode);
178 bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
179 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
180 blk_status_t status = bbio->bio.bi_status;
181 struct bvec_iter iter;
182 struct bio_vec bv;
183 u32 offset;
184
185 btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
186 u64 start = bbio->file_offset + offset;
187
188 if (!status &&
189 (!csum || !btrfs_check_data_csum(bi, bbio, offset,
190 bv.bv_page, bv.bv_offset))) {
191 btrfs_clean_io_failure(bi, start, bv.bv_page,
192 bv.bv_offset);
193 } else {
194 int ret;
195
196 refcount_inc(&cb->pending_ios);
197 ret = btrfs_repair_one_sector(BTRFS_I(inode), bbio, offset,
198 bv.bv_page, bv.bv_offset,
199 true);
200 if (ret) {
201 refcount_dec(&cb->pending_ios);
202 status = errno_to_blk_status(ret);
203 }
204 }
205 }
206
207 if (status)
208 cb->status = status;
209
210 if (refcount_dec_and_test(&cb->pending_ios))
211 finish_compressed_bio_read(cb);
212 btrfs_bio_free_csum(bbio);
213 bio_put(&bbio->bio);
214}
215
216/*
217 * Clear the writeback bits on all of the file
218 * pages for a compressed write
219 */
220static noinline void end_compressed_writeback(struct inode *inode,
221 const struct compressed_bio *cb)
222{
223 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
224 unsigned long index = cb->start >> PAGE_SHIFT;
225 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
226 struct folio_batch fbatch;
227 const int errno = blk_status_to_errno(cb->status);
228 int i;
229 int ret;
230
231 if (errno)
232 mapping_set_error(inode->i_mapping, errno);
233
234 folio_batch_init(&fbatch);
235 while (index <= end_index) {
236 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
237 &fbatch);
238
239 if (ret == 0)
240 return;
241
242 for (i = 0; i < ret; i++) {
243 struct folio *folio = fbatch.folios[i];
244
245 if (errno)
246 folio_set_error(folio);
247 btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
248 cb->start, cb->len);
249 }
250 folio_batch_release(&fbatch);
251 }
252 /* the inode may be gone now */
253}
254
255static void finish_compressed_bio_write(struct compressed_bio *cb)
256{
257 struct inode *inode = cb->inode;
258 unsigned int index;
259
260 /*
261 * Ok, we're the last bio for this extent, step one is to call back
262 * into the FS and do all the end_io operations.
263 */
264 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
265 cb->start, cb->start + cb->len - 1,
266 cb->status == BLK_STS_OK);
267
268 if (cb->writeback)
269 end_compressed_writeback(inode, cb);
270 /* Note, our inode could be gone now */
271
272 /*
273 * Release the compressed pages, these came from alloc_page and
274 * are not attached to the inode at all
275 */
276 for (index = 0; index < cb->nr_pages; index++) {
277 struct page *page = cb->compressed_pages[index];
278
279 page->mapping = NULL;
280 put_page(page);
281 }
282
283 /* Finally free the cb struct */
284 kfree(cb->compressed_pages);
285 kfree(cb);
286}
287
288static void btrfs_finish_compressed_write_work(struct work_struct *work)
289{
290 struct compressed_bio *cb =
291 container_of(work, struct compressed_bio, write_end_work);
292
293 finish_compressed_bio_write(cb);
294}
295
296/*
297 * Do the cleanup once all the compressed pages hit the disk. This will clear
298 * writeback on the file pages and free the compressed pages.
299 *
300 * This also calls the writeback end hooks for the file pages so that metadata
301 * and checksums can be updated in the file.
302 */
303static void end_compressed_bio_write(struct btrfs_bio *bbio)
304{
305 struct compressed_bio *cb = bbio->private;
306
307 if (bbio->bio.bi_status)
308 cb->status = bbio->bio.bi_status;
309
310 if (refcount_dec_and_test(&cb->pending_ios)) {
311 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
312
313 btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio);
314 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
315 }
316 bio_put(&bbio->bio);
317}
318
319/*
320 * Allocate a compressed_bio, which will be used to read/write on-disk
321 * (aka, compressed) * data.
322 *
323 * @cb: The compressed_bio structure, which records all the needed
324 * information to bind the compressed data to the uncompressed
325 * page cache.
326 * @disk_byten: The logical bytenr where the compressed data will be read
327 * from or written to.
328 * @endio_func: The endio function to call after the IO for compressed data
329 * is finished.
330 * @next_stripe_start: Return value of logical bytenr of where next stripe starts.
331 * Let the caller know to only fill the bio up to the stripe
332 * boundary.
333 */
334
335
336static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
337 blk_opf_t opf,
338 btrfs_bio_end_io_t endio_func,
339 u64 *next_stripe_start)
340{
341 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
342 struct btrfs_io_geometry geom;
343 struct extent_map *em;
344 struct bio *bio;
345 int ret;
346
347 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb);
348 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
349
350 em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
351 if (IS_ERR(em)) {
352 bio_put(bio);
353 return ERR_CAST(em);
354 }
355
356 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
357 bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
358
359 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
360 free_extent_map(em);
361 if (ret < 0) {
362 bio_put(bio);
363 return ERR_PTR(ret);
364 }
365 *next_stripe_start = disk_bytenr + geom.len;
366 refcount_inc(&cb->pending_ios);
367 return bio;
368}
369
370/*
371 * worker function to build and submit bios for previously compressed pages.
372 * The corresponding pages in the inode should be marked for writeback
373 * and the compressed pages should have a reference on them for dropping
374 * when the IO is complete.
375 *
376 * This also checksums the file bytes and gets things ready for
377 * the end io hooks.
378 */
379blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
380 unsigned int len, u64 disk_start,
381 unsigned int compressed_len,
382 struct page **compressed_pages,
383 unsigned int nr_pages,
384 blk_opf_t write_flags,
385 struct cgroup_subsys_state *blkcg_css,
386 bool writeback)
387{
388 struct btrfs_fs_info *fs_info = inode->root->fs_info;
389 struct bio *bio = NULL;
390 struct compressed_bio *cb;
391 u64 cur_disk_bytenr = disk_start;
392 u64 next_stripe_start;
393 blk_status_t ret = BLK_STS_OK;
394 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
395 const bool use_append = btrfs_use_zone_append(inode, disk_start);
396 const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
397
398 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
399 IS_ALIGNED(len, fs_info->sectorsize));
400 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
401 if (!cb)
402 return BLK_STS_RESOURCE;
403 refcount_set(&cb->pending_ios, 1);
404 cb->status = BLK_STS_OK;
405 cb->inode = &inode->vfs_inode;
406 cb->start = start;
407 cb->len = len;
408 cb->compressed_pages = compressed_pages;
409 cb->compressed_len = compressed_len;
410 cb->writeback = writeback;
411 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
412 cb->nr_pages = nr_pages;
413
414 if (blkcg_css)
415 kthread_associate_blkcg(blkcg_css);
416
417 while (cur_disk_bytenr < disk_start + compressed_len) {
418 u64 offset = cur_disk_bytenr - disk_start;
419 unsigned int index = offset >> PAGE_SHIFT;
420 unsigned int real_size;
421 unsigned int added;
422 struct page *page = compressed_pages[index];
423 bool submit = false;
424
425 /* Allocate new bio if submitted or not yet allocated */
426 if (!bio) {
427 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
428 bio_op | write_flags, end_compressed_bio_write,
429 &next_stripe_start);
430 if (IS_ERR(bio)) {
431 ret = errno_to_blk_status(PTR_ERR(bio));
432 break;
433 }
434 if (blkcg_css)
435 bio->bi_opf |= REQ_CGROUP_PUNT;
436 }
437 /*
438 * We should never reach next_stripe_start start as we will
439 * submit comp_bio when reach the boundary immediately.
440 */
441 ASSERT(cur_disk_bytenr != next_stripe_start);
442
443 /*
444 * We have various limits on the real read size:
445 * - stripe boundary
446 * - page boundary
447 * - compressed length boundary
448 */
449 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
450 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
451 real_size = min_t(u64, real_size, compressed_len - offset);
452 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
453
454 if (use_append)
455 added = bio_add_zone_append_page(bio, page, real_size,
456 offset_in_page(offset));
457 else
458 added = bio_add_page(bio, page, real_size,
459 offset_in_page(offset));
460 /* Reached zoned boundary */
461 if (added == 0)
462 submit = true;
463
464 cur_disk_bytenr += added;
465 /* Reached stripe boundary */
466 if (cur_disk_bytenr == next_stripe_start)
467 submit = true;
468
469 /* Finished the range */
470 if (cur_disk_bytenr == disk_start + compressed_len)
471 submit = true;
472
473 if (submit) {
474 if (!skip_sum) {
475 ret = btrfs_csum_one_bio(inode, bio, start, true);
476 if (ret) {
477 btrfs_bio_end_io(btrfs_bio(bio), ret);
478 break;
479 }
480 }
481
482 ASSERT(bio->bi_iter.bi_size);
483 btrfs_submit_bio(fs_info, bio, 0);
484 bio = NULL;
485 }
486 cond_resched();
487 }
488
489 if (blkcg_css)
490 kthread_associate_blkcg(NULL);
491
492 if (refcount_dec_and_test(&cb->pending_ios))
493 finish_compressed_bio_write(cb);
494 return ret;
495}
496
497static u64 bio_end_offset(struct bio *bio)
498{
499 struct bio_vec *last = bio_last_bvec_all(bio);
500
501 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
502}
503
504/*
505 * Add extra pages in the same compressed file extent so that we don't need to
506 * re-read the same extent again and again.
507 *
508 * NOTE: this won't work well for subpage, as for subpage read, we lock the
509 * full page then submit bio for each compressed/regular extents.
510 *
511 * This means, if we have several sectors in the same page points to the same
512 * on-disk compressed data, we will re-read the same extent many times and
513 * this function can only help for the next page.
514 */
515static noinline int add_ra_bio_pages(struct inode *inode,
516 u64 compressed_end,
517 struct compressed_bio *cb,
518 int *memstall, unsigned long *pflags)
519{
520 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
521 unsigned long end_index;
522 u64 cur = bio_end_offset(cb->orig_bio);
523 u64 isize = i_size_read(inode);
524 int ret;
525 struct page *page;
526 struct extent_map *em;
527 struct address_space *mapping = inode->i_mapping;
528 struct extent_map_tree *em_tree;
529 struct extent_io_tree *tree;
530 int sectors_missed = 0;
531
532 em_tree = &BTRFS_I(inode)->extent_tree;
533 tree = &BTRFS_I(inode)->io_tree;
534
535 if (isize == 0)
536 return 0;
537
538 /*
539 * For current subpage support, we only support 64K page size,
540 * which means maximum compressed extent size (128K) is just 2x page
541 * size.
542 * This makes readahead less effective, so here disable readahead for
543 * subpage for now, until full compressed write is supported.
544 */
545 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
546 return 0;
547
548 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
549
550 while (cur < compressed_end) {
551 u64 page_end;
552 u64 pg_index = cur >> PAGE_SHIFT;
553 u32 add_size;
554
555 if (pg_index > end_index)
556 break;
557
558 page = xa_load(&mapping->i_pages, pg_index);
559 if (page && !xa_is_value(page)) {
560 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
561 fs_info->sectorsize_bits;
562
563 /* Beyond threshold, no need to continue */
564 if (sectors_missed > 4)
565 break;
566
567 /*
568 * Jump to next page start as we already have page for
569 * current offset.
570 */
571 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
572 continue;
573 }
574
575 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
576 ~__GFP_FS));
577 if (!page)
578 break;
579
580 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
581 put_page(page);
582 /* There is already a page, skip to page end */
583 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
584 continue;
585 }
586
587 if (!*memstall && PageWorkingset(page)) {
588 psi_memstall_enter(pflags);
589 *memstall = 1;
590 }
591
592 ret = set_page_extent_mapped(page);
593 if (ret < 0) {
594 unlock_page(page);
595 put_page(page);
596 break;
597 }
598
599 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
600 lock_extent(tree, cur, page_end, NULL);
601 read_lock(&em_tree->lock);
602 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
603 read_unlock(&em_tree->lock);
604
605 /*
606 * At this point, we have a locked page in the page cache for
607 * these bytes in the file. But, we have to make sure they map
608 * to this compressed extent on disk.
609 */
610 if (!em || cur < em->start ||
611 (cur + fs_info->sectorsize > extent_map_end(em)) ||
612 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
613 free_extent_map(em);
614 unlock_extent(tree, cur, page_end, NULL);
615 unlock_page(page);
616 put_page(page);
617 break;
618 }
619 free_extent_map(em);
620
621 if (page->index == end_index) {
622 size_t zero_offset = offset_in_page(isize);
623
624 if (zero_offset) {
625 int zeros;
626 zeros = PAGE_SIZE - zero_offset;
627 memzero_page(page, zero_offset, zeros);
628 }
629 }
630
631 add_size = min(em->start + em->len, page_end + 1) - cur;
632 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
633 if (ret != add_size) {
634 unlock_extent(tree, cur, page_end, NULL);
635 unlock_page(page);
636 put_page(page);
637 break;
638 }
639 /*
640 * If it's subpage, we also need to increase its
641 * subpage::readers number, as at endio we will decrease
642 * subpage::readers and to unlock the page.
643 */
644 if (fs_info->sectorsize < PAGE_SIZE)
645 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
646 put_page(page);
647 cur += add_size;
648 }
649 return 0;
650}
651
652/*
653 * for a compressed read, the bio we get passed has all the inode pages
654 * in it. We don't actually do IO on those pages but allocate new ones
655 * to hold the compressed pages on disk.
656 *
657 * bio->bi_iter.bi_sector points to the compressed extent on disk
658 * bio->bi_io_vec points to all of the inode pages
659 *
660 * After the compressed pages are read, we copy the bytes into the
661 * bio we were passed and then call the bio end_io calls
662 */
663void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
664 int mirror_num)
665{
666 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
667 struct extent_map_tree *em_tree;
668 struct compressed_bio *cb;
669 unsigned int compressed_len;
670 struct bio *comp_bio = NULL;
671 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
672 u64 cur_disk_byte = disk_bytenr;
673 u64 next_stripe_start;
674 u64 file_offset;
675 u64 em_len;
676 u64 em_start;
677 struct extent_map *em;
678 unsigned long pflags;
679 int memstall = 0;
680 blk_status_t ret;
681 int ret2;
682 int i;
683
684 em_tree = &BTRFS_I(inode)->extent_tree;
685
686 file_offset = bio_first_bvec_all(bio)->bv_offset +
687 page_offset(bio_first_page_all(bio));
688
689 /* we need the actual starting offset of this extent in the file */
690 read_lock(&em_tree->lock);
691 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
692 read_unlock(&em_tree->lock);
693 if (!em) {
694 ret = BLK_STS_IOERR;
695 goto out;
696 }
697
698 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
699 compressed_len = em->block_len;
700 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
701 if (!cb) {
702 ret = BLK_STS_RESOURCE;
703 goto out;
704 }
705
706 refcount_set(&cb->pending_ios, 1);
707 cb->status = BLK_STS_OK;
708 cb->inode = inode;
709
710 cb->start = em->orig_start;
711 em_len = em->len;
712 em_start = em->start;
713
714 cb->len = bio->bi_iter.bi_size;
715 cb->compressed_len = compressed_len;
716 cb->compress_type = em->compress_type;
717 cb->orig_bio = bio;
718
719 free_extent_map(em);
720 em = NULL;
721
722 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
723 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
724 if (!cb->compressed_pages) {
725 ret = BLK_STS_RESOURCE;
726 goto fail;
727 }
728
729 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
730 if (ret2) {
731 ret = BLK_STS_RESOURCE;
732 goto fail;
733 }
734
735 add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
736
737 /* include any pages we added in add_ra-bio_pages */
738 cb->len = bio->bi_iter.bi_size;
739
740 while (cur_disk_byte < disk_bytenr + compressed_len) {
741 u64 offset = cur_disk_byte - disk_bytenr;
742 unsigned int index = offset >> PAGE_SHIFT;
743 unsigned int real_size;
744 unsigned int added;
745 struct page *page = cb->compressed_pages[index];
746 bool submit = false;
747
748 /* Allocate new bio if submitted or not yet allocated */
749 if (!comp_bio) {
750 comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
751 REQ_OP_READ, end_compressed_bio_read,
752 &next_stripe_start);
753 if (IS_ERR(comp_bio)) {
754 cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
755 break;
756 }
757 }
758 /*
759 * We should never reach next_stripe_start start as we will
760 * submit comp_bio when reach the boundary immediately.
761 */
762 ASSERT(cur_disk_byte != next_stripe_start);
763 /*
764 * We have various limit on the real read size:
765 * - stripe boundary
766 * - page boundary
767 * - compressed length boundary
768 */
769 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
770 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
771 real_size = min_t(u64, real_size, compressed_len - offset);
772 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
773
774 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
775 /*
776 * Maximum compressed extent is smaller than bio size limit,
777 * thus bio_add_page() should always success.
778 */
779 ASSERT(added == real_size);
780 cur_disk_byte += added;
781
782 /* Reached stripe boundary, need to submit */
783 if (cur_disk_byte == next_stripe_start)
784 submit = true;
785
786 /* Has finished the range, need to submit */
787 if (cur_disk_byte == disk_bytenr + compressed_len)
788 submit = true;
789
790 if (submit) {
791 /* Save the original iter for read repair */
792 if (bio_op(comp_bio) == REQ_OP_READ)
793 btrfs_bio(comp_bio)->iter = comp_bio->bi_iter;
794
795 /*
796 * Save the initial offset of this chunk, as there
797 * is no direct correlation between compressed pages and
798 * the original file offset. The field is only used for
799 * priting error messages.
800 */
801 btrfs_bio(comp_bio)->file_offset = file_offset;
802
803 ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL);
804 if (ret) {
805 btrfs_bio_end_io(btrfs_bio(comp_bio), ret);
806 break;
807 }
808
809 ASSERT(comp_bio->bi_iter.bi_size);
810 btrfs_submit_bio(fs_info, comp_bio, mirror_num);
811 comp_bio = NULL;
812 }
813 }
814
815 if (memstall)
816 psi_memstall_leave(&pflags);
817
818 if (refcount_dec_and_test(&cb->pending_ios))
819 finish_compressed_bio_read(cb);
820 return;
821
822fail:
823 if (cb->compressed_pages) {
824 for (i = 0; i < cb->nr_pages; i++) {
825 if (cb->compressed_pages[i])
826 __free_page(cb->compressed_pages[i]);
827 }
828 }
829
830 kfree(cb->compressed_pages);
831 kfree(cb);
832out:
833 free_extent_map(em);
834 btrfs_bio_end_io(btrfs_bio(bio), ret);
835 return;
836}
837
838/*
839 * Heuristic uses systematic sampling to collect data from the input data
840 * range, the logic can be tuned by the following constants:
841 *
842 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
843 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
844 */
845#define SAMPLING_READ_SIZE (16)
846#define SAMPLING_INTERVAL (256)
847
848/*
849 * For statistical analysis of the input data we consider bytes that form a
850 * Galois Field of 256 objects. Each object has an attribute count, ie. how
851 * many times the object appeared in the sample.
852 */
853#define BUCKET_SIZE (256)
854
855/*
856 * The size of the sample is based on a statistical sampling rule of thumb.
857 * The common way is to perform sampling tests as long as the number of
858 * elements in each cell is at least 5.
859 *
860 * Instead of 5, we choose 32 to obtain more accurate results.
861 * If the data contain the maximum number of symbols, which is 256, we obtain a
862 * sample size bound by 8192.
863 *
864 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
865 * from up to 512 locations.
866 */
867#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
868 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
869
870struct bucket_item {
871 u32 count;
872};
873
874struct heuristic_ws {
875 /* Partial copy of input data */
876 u8 *sample;
877 u32 sample_size;
878 /* Buckets store counters for each byte value */
879 struct bucket_item *bucket;
880 /* Sorting buffer */
881 struct bucket_item *bucket_b;
882 struct list_head list;
883};
884
885static struct workspace_manager heuristic_wsm;
886
887static void free_heuristic_ws(struct list_head *ws)
888{
889 struct heuristic_ws *workspace;
890
891 workspace = list_entry(ws, struct heuristic_ws, list);
892
893 kvfree(workspace->sample);
894 kfree(workspace->bucket);
895 kfree(workspace->bucket_b);
896 kfree(workspace);
897}
898
899static struct list_head *alloc_heuristic_ws(unsigned int level)
900{
901 struct heuristic_ws *ws;
902
903 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
904 if (!ws)
905 return ERR_PTR(-ENOMEM);
906
907 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
908 if (!ws->sample)
909 goto fail;
910
911 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
912 if (!ws->bucket)
913 goto fail;
914
915 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
916 if (!ws->bucket_b)
917 goto fail;
918
919 INIT_LIST_HEAD(&ws->list);
920 return &ws->list;
921fail:
922 free_heuristic_ws(&ws->list);
923 return ERR_PTR(-ENOMEM);
924}
925
926const struct btrfs_compress_op btrfs_heuristic_compress = {
927 .workspace_manager = &heuristic_wsm,
928};
929
930static const struct btrfs_compress_op * const btrfs_compress_op[] = {
931 /* The heuristic is represented as compression type 0 */
932 &btrfs_heuristic_compress,
933 &btrfs_zlib_compress,
934 &btrfs_lzo_compress,
935 &btrfs_zstd_compress,
936};
937
938static struct list_head *alloc_workspace(int type, unsigned int level)
939{
940 switch (type) {
941 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
942 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
943 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
944 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
945 default:
946 /*
947 * This can't happen, the type is validated several times
948 * before we get here.
949 */
950 BUG();
951 }
952}
953
954static void free_workspace(int type, struct list_head *ws)
955{
956 switch (type) {
957 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
958 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
959 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
960 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
961 default:
962 /*
963 * This can't happen, the type is validated several times
964 * before we get here.
965 */
966 BUG();
967 }
968}
969
970static void btrfs_init_workspace_manager(int type)
971{
972 struct workspace_manager *wsm;
973 struct list_head *workspace;
974
975 wsm = btrfs_compress_op[type]->workspace_manager;
976 INIT_LIST_HEAD(&wsm->idle_ws);
977 spin_lock_init(&wsm->ws_lock);
978 atomic_set(&wsm->total_ws, 0);
979 init_waitqueue_head(&wsm->ws_wait);
980
981 /*
982 * Preallocate one workspace for each compression type so we can
983 * guarantee forward progress in the worst case
984 */
985 workspace = alloc_workspace(type, 0);
986 if (IS_ERR(workspace)) {
987 pr_warn(
988 "BTRFS: cannot preallocate compression workspace, will try later\n");
989 } else {
990 atomic_set(&wsm->total_ws, 1);
991 wsm->free_ws = 1;
992 list_add(workspace, &wsm->idle_ws);
993 }
994}
995
996static void btrfs_cleanup_workspace_manager(int type)
997{
998 struct workspace_manager *wsman;
999 struct list_head *ws;
1000
1001 wsman = btrfs_compress_op[type]->workspace_manager;
1002 while (!list_empty(&wsman->idle_ws)) {
1003 ws = wsman->idle_ws.next;
1004 list_del(ws);
1005 free_workspace(type, ws);
1006 atomic_dec(&wsman->total_ws);
1007 }
1008}
1009
1010/*
1011 * This finds an available workspace or allocates a new one.
1012 * If it's not possible to allocate a new one, waits until there's one.
1013 * Preallocation makes a forward progress guarantees and we do not return
1014 * errors.
1015 */
1016struct list_head *btrfs_get_workspace(int type, unsigned int level)
1017{
1018 struct workspace_manager *wsm;
1019 struct list_head *workspace;
1020 int cpus = num_online_cpus();
1021 unsigned nofs_flag;
1022 struct list_head *idle_ws;
1023 spinlock_t *ws_lock;
1024 atomic_t *total_ws;
1025 wait_queue_head_t *ws_wait;
1026 int *free_ws;
1027
1028 wsm = btrfs_compress_op[type]->workspace_manager;
1029 idle_ws = &wsm->idle_ws;
1030 ws_lock = &wsm->ws_lock;
1031 total_ws = &wsm->total_ws;
1032 ws_wait = &wsm->ws_wait;
1033 free_ws = &wsm->free_ws;
1034
1035again:
1036 spin_lock(ws_lock);
1037 if (!list_empty(idle_ws)) {
1038 workspace = idle_ws->next;
1039 list_del(workspace);
1040 (*free_ws)--;
1041 spin_unlock(ws_lock);
1042 return workspace;
1043
1044 }
1045 if (atomic_read(total_ws) > cpus) {
1046 DEFINE_WAIT(wait);
1047
1048 spin_unlock(ws_lock);
1049 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1050 if (atomic_read(total_ws) > cpus && !*free_ws)
1051 schedule();
1052 finish_wait(ws_wait, &wait);
1053 goto again;
1054 }
1055 atomic_inc(total_ws);
1056 spin_unlock(ws_lock);
1057
1058 /*
1059 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1060 * to turn it off here because we might get called from the restricted
1061 * context of btrfs_compress_bio/btrfs_compress_pages
1062 */
1063 nofs_flag = memalloc_nofs_save();
1064 workspace = alloc_workspace(type, level);
1065 memalloc_nofs_restore(nofs_flag);
1066
1067 if (IS_ERR(workspace)) {
1068 atomic_dec(total_ws);
1069 wake_up(ws_wait);
1070
1071 /*
1072 * Do not return the error but go back to waiting. There's a
1073 * workspace preallocated for each type and the compression
1074 * time is bounded so we get to a workspace eventually. This
1075 * makes our caller's life easier.
1076 *
1077 * To prevent silent and low-probability deadlocks (when the
1078 * initial preallocation fails), check if there are any
1079 * workspaces at all.
1080 */
1081 if (atomic_read(total_ws) == 0) {
1082 static DEFINE_RATELIMIT_STATE(_rs,
1083 /* once per minute */ 60 * HZ,
1084 /* no burst */ 1);
1085
1086 if (__ratelimit(&_rs)) {
1087 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1088 }
1089 }
1090 goto again;
1091 }
1092 return workspace;
1093}
1094
1095static struct list_head *get_workspace(int type, int level)
1096{
1097 switch (type) {
1098 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1099 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1100 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1101 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1102 default:
1103 /*
1104 * This can't happen, the type is validated several times
1105 * before we get here.
1106 */
1107 BUG();
1108 }
1109}
1110
1111/*
1112 * put a workspace struct back on the list or free it if we have enough
1113 * idle ones sitting around
1114 */
1115void btrfs_put_workspace(int type, struct list_head *ws)
1116{
1117 struct workspace_manager *wsm;
1118 struct list_head *idle_ws;
1119 spinlock_t *ws_lock;
1120 atomic_t *total_ws;
1121 wait_queue_head_t *ws_wait;
1122 int *free_ws;
1123
1124 wsm = btrfs_compress_op[type]->workspace_manager;
1125 idle_ws = &wsm->idle_ws;
1126 ws_lock = &wsm->ws_lock;
1127 total_ws = &wsm->total_ws;
1128 ws_wait = &wsm->ws_wait;
1129 free_ws = &wsm->free_ws;
1130
1131 spin_lock(ws_lock);
1132 if (*free_ws <= num_online_cpus()) {
1133 list_add(ws, idle_ws);
1134 (*free_ws)++;
1135 spin_unlock(ws_lock);
1136 goto wake;
1137 }
1138 spin_unlock(ws_lock);
1139
1140 free_workspace(type, ws);
1141 atomic_dec(total_ws);
1142wake:
1143 cond_wake_up(ws_wait);
1144}
1145
1146static void put_workspace(int type, struct list_head *ws)
1147{
1148 switch (type) {
1149 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1150 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1151 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1152 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1153 default:
1154 /*
1155 * This can't happen, the type is validated several times
1156 * before we get here.
1157 */
1158 BUG();
1159 }
1160}
1161
1162/*
1163 * Adjust @level according to the limits of the compression algorithm or
1164 * fallback to default
1165 */
1166static unsigned int btrfs_compress_set_level(int type, unsigned level)
1167{
1168 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1169
1170 if (level == 0)
1171 level = ops->default_level;
1172 else
1173 level = min(level, ops->max_level);
1174
1175 return level;
1176}
1177
1178/*
1179 * Given an address space and start and length, compress the bytes into @pages
1180 * that are allocated on demand.
1181 *
1182 * @type_level is encoded algorithm and level, where level 0 means whatever
1183 * default the algorithm chooses and is opaque here;
1184 * - compression algo are 0-3
1185 * - the level are bits 4-7
1186 *
1187 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1188 * and returns number of actually allocated pages
1189 *
1190 * @total_in is used to return the number of bytes actually read. It
1191 * may be smaller than the input length if we had to exit early because we
1192 * ran out of room in the pages array or because we cross the
1193 * max_out threshold.
1194 *
1195 * @total_out is an in/out parameter, must be set to the input length and will
1196 * be also used to return the total number of compressed bytes
1197 */
1198int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1199 u64 start, struct page **pages,
1200 unsigned long *out_pages,
1201 unsigned long *total_in,
1202 unsigned long *total_out)
1203{
1204 int type = btrfs_compress_type(type_level);
1205 int level = btrfs_compress_level(type_level);
1206 struct list_head *workspace;
1207 int ret;
1208
1209 level = btrfs_compress_set_level(type, level);
1210 workspace = get_workspace(type, level);
1211 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1212 out_pages, total_in, total_out);
1213 put_workspace(type, workspace);
1214 return ret;
1215}
1216
1217static int btrfs_decompress_bio(struct compressed_bio *cb)
1218{
1219 struct list_head *workspace;
1220 int ret;
1221 int type = cb->compress_type;
1222
1223 workspace = get_workspace(type, 0);
1224 ret = compression_decompress_bio(workspace, cb);
1225 put_workspace(type, workspace);
1226
1227 return ret;
1228}
1229
1230/*
1231 * a less complex decompression routine. Our compressed data fits in a
1232 * single page, and we want to read a single page out of it.
1233 * start_byte tells us the offset into the compressed data we're interested in
1234 */
1235int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1236 unsigned long start_byte, size_t srclen, size_t destlen)
1237{
1238 struct list_head *workspace;
1239 int ret;
1240
1241 workspace = get_workspace(type, 0);
1242 ret = compression_decompress(type, workspace, data_in, dest_page,
1243 start_byte, srclen, destlen);
1244 put_workspace(type, workspace);
1245
1246 return ret;
1247}
1248
1249int __init btrfs_init_compress(void)
1250{
1251 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1252 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1253 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1254 zstd_init_workspace_manager();
1255 return 0;
1256}
1257
1258void __cold btrfs_exit_compress(void)
1259{
1260 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1261 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1262 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1263 zstd_cleanup_workspace_manager();
1264}
1265
1266/*
1267 * Copy decompressed data from working buffer to pages.
1268 *
1269 * @buf: The decompressed data buffer
1270 * @buf_len: The decompressed data length
1271 * @decompressed: Number of bytes that are already decompressed inside the
1272 * compressed extent
1273 * @cb: The compressed extent descriptor
1274 * @orig_bio: The original bio that the caller wants to read for
1275 *
1276 * An easier to understand graph is like below:
1277 *
1278 * |<- orig_bio ->| |<- orig_bio->|
1279 * |<------- full decompressed extent ----->|
1280 * |<----------- @cb range ---->|
1281 * | |<-- @buf_len -->|
1282 * |<--- @decompressed --->|
1283 *
1284 * Note that, @cb can be a subpage of the full decompressed extent, but
1285 * @cb->start always has the same as the orig_file_offset value of the full
1286 * decompressed extent.
1287 *
1288 * When reading compressed extent, we have to read the full compressed extent,
1289 * while @orig_bio may only want part of the range.
1290 * Thus this function will ensure only data covered by @orig_bio will be copied
1291 * to.
1292 *
1293 * Return 0 if we have copied all needed contents for @orig_bio.
1294 * Return >0 if we need continue decompress.
1295 */
1296int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1297 struct compressed_bio *cb, u32 decompressed)
1298{
1299 struct bio *orig_bio = cb->orig_bio;
1300 /* Offset inside the full decompressed extent */
1301 u32 cur_offset;
1302
1303 cur_offset = decompressed;
1304 /* The main loop to do the copy */
1305 while (cur_offset < decompressed + buf_len) {
1306 struct bio_vec bvec;
1307 size_t copy_len;
1308 u32 copy_start;
1309 /* Offset inside the full decompressed extent */
1310 u32 bvec_offset;
1311
1312 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1313 /*
1314 * cb->start may underflow, but subtracting that value can still
1315 * give us correct offset inside the full decompressed extent.
1316 */
1317 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1318
1319 /* Haven't reached the bvec range, exit */
1320 if (decompressed + buf_len <= bvec_offset)
1321 return 1;
1322
1323 copy_start = max(cur_offset, bvec_offset);
1324 copy_len = min(bvec_offset + bvec.bv_len,
1325 decompressed + buf_len) - copy_start;
1326 ASSERT(copy_len);
1327
1328 /*
1329 * Extra range check to ensure we didn't go beyond
1330 * @buf + @buf_len.
1331 */
1332 ASSERT(copy_start - decompressed < buf_len);
1333 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1334 buf + copy_start - decompressed, copy_len);
1335 cur_offset += copy_len;
1336
1337 bio_advance(orig_bio, copy_len);
1338 /* Finished the bio */
1339 if (!orig_bio->bi_iter.bi_size)
1340 return 0;
1341 }
1342 return 1;
1343}
1344
1345/*
1346 * Shannon Entropy calculation
1347 *
1348 * Pure byte distribution analysis fails to determine compressibility of data.
1349 * Try calculating entropy to estimate the average minimum number of bits
1350 * needed to encode the sampled data.
1351 *
1352 * For convenience, return the percentage of needed bits, instead of amount of
1353 * bits directly.
1354 *
1355 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1356 * and can be compressible with high probability
1357 *
1358 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1359 *
1360 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1361 */
1362#define ENTROPY_LVL_ACEPTABLE (65)
1363#define ENTROPY_LVL_HIGH (80)
1364
1365/*
1366 * For increasead precision in shannon_entropy calculation,
1367 * let's do pow(n, M) to save more digits after comma:
1368 *
1369 * - maximum int bit length is 64
1370 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1371 * - 13 * 4 = 52 < 64 -> M = 4
1372 *
1373 * So use pow(n, 4).
1374 */
1375static inline u32 ilog2_w(u64 n)
1376{
1377 return ilog2(n * n * n * n);
1378}
1379
1380static u32 shannon_entropy(struct heuristic_ws *ws)
1381{
1382 const u32 entropy_max = 8 * ilog2_w(2);
1383 u32 entropy_sum = 0;
1384 u32 p, p_base, sz_base;
1385 u32 i;
1386
1387 sz_base = ilog2_w(ws->sample_size);
1388 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1389 p = ws->bucket[i].count;
1390 p_base = ilog2_w(p);
1391 entropy_sum += p * (sz_base - p_base);
1392 }
1393
1394 entropy_sum /= ws->sample_size;
1395 return entropy_sum * 100 / entropy_max;
1396}
1397
1398#define RADIX_BASE 4U
1399#define COUNTERS_SIZE (1U << RADIX_BASE)
1400
1401static u8 get4bits(u64 num, int shift) {
1402 u8 low4bits;
1403
1404 num >>= shift;
1405 /* Reverse order */
1406 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1407 return low4bits;
1408}
1409
1410/*
1411 * Use 4 bits as radix base
1412 * Use 16 u32 counters for calculating new position in buf array
1413 *
1414 * @array - array that will be sorted
1415 * @array_buf - buffer array to store sorting results
1416 * must be equal in size to @array
1417 * @num - array size
1418 */
1419static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1420 int num)
1421{
1422 u64 max_num;
1423 u64 buf_num;
1424 u32 counters[COUNTERS_SIZE];
1425 u32 new_addr;
1426 u32 addr;
1427 int bitlen;
1428 int shift;
1429 int i;
1430
1431 /*
1432 * Try avoid useless loop iterations for small numbers stored in big
1433 * counters. Example: 48 33 4 ... in 64bit array
1434 */
1435 max_num = array[0].count;
1436 for (i = 1; i < num; i++) {
1437 buf_num = array[i].count;
1438 if (buf_num > max_num)
1439 max_num = buf_num;
1440 }
1441
1442 buf_num = ilog2(max_num);
1443 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1444
1445 shift = 0;
1446 while (shift < bitlen) {
1447 memset(counters, 0, sizeof(counters));
1448
1449 for (i = 0; i < num; i++) {
1450 buf_num = array[i].count;
1451 addr = get4bits(buf_num, shift);
1452 counters[addr]++;
1453 }
1454
1455 for (i = 1; i < COUNTERS_SIZE; i++)
1456 counters[i] += counters[i - 1];
1457
1458 for (i = num - 1; i >= 0; i--) {
1459 buf_num = array[i].count;
1460 addr = get4bits(buf_num, shift);
1461 counters[addr]--;
1462 new_addr = counters[addr];
1463 array_buf[new_addr] = array[i];
1464 }
1465
1466 shift += RADIX_BASE;
1467
1468 /*
1469 * Normal radix expects to move data from a temporary array, to
1470 * the main one. But that requires some CPU time. Avoid that
1471 * by doing another sort iteration to original array instead of
1472 * memcpy()
1473 */
1474 memset(counters, 0, sizeof(counters));
1475
1476 for (i = 0; i < num; i ++) {
1477 buf_num = array_buf[i].count;
1478 addr = get4bits(buf_num, shift);
1479 counters[addr]++;
1480 }
1481
1482 for (i = 1; i < COUNTERS_SIZE; i++)
1483 counters[i] += counters[i - 1];
1484
1485 for (i = num - 1; i >= 0; i--) {
1486 buf_num = array_buf[i].count;
1487 addr = get4bits(buf_num, shift);
1488 counters[addr]--;
1489 new_addr = counters[addr];
1490 array[new_addr] = array_buf[i];
1491 }
1492
1493 shift += RADIX_BASE;
1494 }
1495}
1496
1497/*
1498 * Size of the core byte set - how many bytes cover 90% of the sample
1499 *
1500 * There are several types of structured binary data that use nearly all byte
1501 * values. The distribution can be uniform and counts in all buckets will be
1502 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1503 *
1504 * Other possibility is normal (Gaussian) distribution, where the data could
1505 * be potentially compressible, but we have to take a few more steps to decide
1506 * how much.
1507 *
1508 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1509 * compression algo can easy fix that
1510 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1511 * probability is not compressible
1512 */
1513#define BYTE_CORE_SET_LOW (64)
1514#define BYTE_CORE_SET_HIGH (200)
1515
1516static int byte_core_set_size(struct heuristic_ws *ws)
1517{
1518 u32 i;
1519 u32 coreset_sum = 0;
1520 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1521 struct bucket_item *bucket = ws->bucket;
1522
1523 /* Sort in reverse order */
1524 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1525
1526 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1527 coreset_sum += bucket[i].count;
1528
1529 if (coreset_sum > core_set_threshold)
1530 return i;
1531
1532 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1533 coreset_sum += bucket[i].count;
1534 if (coreset_sum > core_set_threshold)
1535 break;
1536 }
1537
1538 return i;
1539}
1540
1541/*
1542 * Count byte values in buckets.
1543 * This heuristic can detect textual data (configs, xml, json, html, etc).
1544 * Because in most text-like data byte set is restricted to limited number of
1545 * possible characters, and that restriction in most cases makes data easy to
1546 * compress.
1547 *
1548 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1549 * less - compressible
1550 * more - need additional analysis
1551 */
1552#define BYTE_SET_THRESHOLD (64)
1553
1554static u32 byte_set_size(const struct heuristic_ws *ws)
1555{
1556 u32 i;
1557 u32 byte_set_size = 0;
1558
1559 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1560 if (ws->bucket[i].count > 0)
1561 byte_set_size++;
1562 }
1563
1564 /*
1565 * Continue collecting count of byte values in buckets. If the byte
1566 * set size is bigger then the threshold, it's pointless to continue,
1567 * the detection technique would fail for this type of data.
1568 */
1569 for (; i < BUCKET_SIZE; i++) {
1570 if (ws->bucket[i].count > 0) {
1571 byte_set_size++;
1572 if (byte_set_size > BYTE_SET_THRESHOLD)
1573 return byte_set_size;
1574 }
1575 }
1576
1577 return byte_set_size;
1578}
1579
1580static bool sample_repeated_patterns(struct heuristic_ws *ws)
1581{
1582 const u32 half_of_sample = ws->sample_size / 2;
1583 const u8 *data = ws->sample;
1584
1585 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1586}
1587
1588static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1589 struct heuristic_ws *ws)
1590{
1591 struct page *page;
1592 u64 index, index_end;
1593 u32 i, curr_sample_pos;
1594 u8 *in_data;
1595
1596 /*
1597 * Compression handles the input data by chunks of 128KiB
1598 * (defined by BTRFS_MAX_UNCOMPRESSED)
1599 *
1600 * We do the same for the heuristic and loop over the whole range.
1601 *
1602 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1603 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1604 */
1605 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1606 end = start + BTRFS_MAX_UNCOMPRESSED;
1607
1608 index = start >> PAGE_SHIFT;
1609 index_end = end >> PAGE_SHIFT;
1610
1611 /* Don't miss unaligned end */
1612 if (!IS_ALIGNED(end, PAGE_SIZE))
1613 index_end++;
1614
1615 curr_sample_pos = 0;
1616 while (index < index_end) {
1617 page = find_get_page(inode->i_mapping, index);
1618 in_data = kmap_local_page(page);
1619 /* Handle case where the start is not aligned to PAGE_SIZE */
1620 i = start % PAGE_SIZE;
1621 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1622 /* Don't sample any garbage from the last page */
1623 if (start > end - SAMPLING_READ_SIZE)
1624 break;
1625 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1626 SAMPLING_READ_SIZE);
1627 i += SAMPLING_INTERVAL;
1628 start += SAMPLING_INTERVAL;
1629 curr_sample_pos += SAMPLING_READ_SIZE;
1630 }
1631 kunmap_local(in_data);
1632 put_page(page);
1633
1634 index++;
1635 }
1636
1637 ws->sample_size = curr_sample_pos;
1638}
1639
1640/*
1641 * Compression heuristic.
1642 *
1643 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1644 * quickly (compared to direct compression) detect data characteristics
1645 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1646 * data.
1647 *
1648 * The following types of analysis can be performed:
1649 * - detect mostly zero data
1650 * - detect data with low "byte set" size (text, etc)
1651 * - detect data with low/high "core byte" set
1652 *
1653 * Return non-zero if the compression should be done, 0 otherwise.
1654 */
1655int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1656{
1657 struct list_head *ws_list = get_workspace(0, 0);
1658 struct heuristic_ws *ws;
1659 u32 i;
1660 u8 byte;
1661 int ret = 0;
1662
1663 ws = list_entry(ws_list, struct heuristic_ws, list);
1664
1665 heuristic_collect_sample(inode, start, end, ws);
1666
1667 if (sample_repeated_patterns(ws)) {
1668 ret = 1;
1669 goto out;
1670 }
1671
1672 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1673
1674 for (i = 0; i < ws->sample_size; i++) {
1675 byte = ws->sample[i];
1676 ws->bucket[byte].count++;
1677 }
1678
1679 i = byte_set_size(ws);
1680 if (i < BYTE_SET_THRESHOLD) {
1681 ret = 2;
1682 goto out;
1683 }
1684
1685 i = byte_core_set_size(ws);
1686 if (i <= BYTE_CORE_SET_LOW) {
1687 ret = 3;
1688 goto out;
1689 }
1690
1691 if (i >= BYTE_CORE_SET_HIGH) {
1692 ret = 0;
1693 goto out;
1694 }
1695
1696 i = shannon_entropy(ws);
1697 if (i <= ENTROPY_LVL_ACEPTABLE) {
1698 ret = 4;
1699 goto out;
1700 }
1701
1702 /*
1703 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1704 * needed to give green light to compression.
1705 *
1706 * For now just assume that compression at that level is not worth the
1707 * resources because:
1708 *
1709 * 1. it is possible to defrag the data later
1710 *
1711 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1712 * values, every bucket has counter at level ~54. The heuristic would
1713 * be confused. This can happen when data have some internal repeated
1714 * patterns like "abbacbbc...". This can be detected by analyzing
1715 * pairs of bytes, which is too costly.
1716 */
1717 if (i < ENTROPY_LVL_HIGH) {
1718 ret = 5;
1719 goto out;
1720 } else {
1721 ret = 0;
1722 goto out;
1723 }
1724
1725out:
1726 put_workspace(0, ws_list);
1727 return ret;
1728}
1729
1730/*
1731 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1732 * level, unrecognized string will set the default level
1733 */
1734unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1735{
1736 unsigned int level = 0;
1737 int ret;
1738
1739 if (!type)
1740 return 0;
1741
1742 if (str[0] == ':') {
1743 ret = kstrtouint(str + 1, 10, &level);
1744 if (ret)
1745 level = 0;
1746 }
1747
1748 level = btrfs_compress_set_level(type, level);
1749
1750 return level;
1751}