]> git.ipfire.org Git - thirdparty/linux.git/blame - fs/btrfs/compression.c
btrfs: use the flags of an extent map to identify the compression type
[thirdparty/linux.git] / fs / btrfs / compression.c
CommitLineData
c1d7c514 1// SPDX-License-Identifier: GPL-2.0
c8b97818
CM
2/*
3 * Copyright (C) 2008 Oracle. All rights reserved.
c8b97818
CM
4 */
5
6#include <linux/kernel.h>
7#include <linux/bio.h>
c8b97818
CM
8#include <linux/file.h>
9#include <linux/fs.h>
10#include <linux/pagemap.h>
a75b81c3 11#include <linux/pagevec.h>
c8b97818 12#include <linux/highmem.h>
e41d12f5 13#include <linux/kthread.h>
c8b97818
CM
14#include <linux/time.h>
15#include <linux/init.h>
16#include <linux/string.h>
c8b97818 17#include <linux/backing-dev.h>
c8b97818 18#include <linux/writeback.h>
4088a47e 19#include <linux/psi.h>
5a0e3ad6 20#include <linux/slab.h>
fe308533 21#include <linux/sched/mm.h>
19562430 22#include <linux/log2.h>
4cea422a 23#include <linux/shrinker.h>
d5178578 24#include <crypto/hash.h>
602cbe91 25#include "misc.h"
c8b97818 26#include "ctree.h"
ec8eb376 27#include "fs.h"
c8b97818
CM
28#include "disk-io.h"
29#include "transaction.h"
30#include "btrfs_inode.h"
103c1972 31#include "bio.h"
c8b97818 32#include "ordered-data.h"
c8b97818
CM
33#include "compression.h"
34#include "extent_io.h"
35#include "extent_map.h"
6a404910 36#include "subpage.h"
764c7c9a 37#include "zoned.h"
7c8ede16 38#include "file-item.h"
7f0add25 39#include "super.h"
c8b97818 40
e794203e 41static struct bio_set btrfs_compressed_bioset;
544fe4a9 42
e128f9c3
DS
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];
ce96b7ff
CX
53 default:
54 break;
e128f9c3
DS
55 }
56
57 return NULL;
58}
59
544fe4a9
CH
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));
4317ff00
QW
73 btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
74 bbio->inode = inode;
544fe4a9
CH
75 bbio->file_offset = start;
76 return to_compressed_bio(bbio);
77}
78
aa53e3bf
JT
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
1e4eb746
DS
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 /*
1d8ba9e7
QW
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.
1e4eb746 120 */
1d8ba9e7 121 *out_pages = 0;
1e4eb746
DS
122 return -E2BIG;
123 }
124}
125
4a9e803e
SY
126static int compression_decompress_bio(struct list_head *ws,
127 struct compressed_bio *cb)
1e4eb746 128{
4a9e803e 129 switch (cb->compress_type) {
1e4eb746
DS
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,
3e09b5b2 144 const u8 *data_in, struct page *dest_page,
1e4eb746
DS
145 unsigned long start_byte, size_t srclen, size_t destlen)
146{
147 switch (type) {
148 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
149 start_byte, srclen, destlen);
150 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
151 start_byte, srclen, destlen);
152 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
153 start_byte, 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
32586c5b
CH
164static void btrfs_free_compressed_pages(struct compressed_bio *cb)
165{
a959a174 166 for (unsigned int i = 0; i < cb->nr_pages; i++)
9ba965dc 167 btrfs_free_compr_page(cb->compressed_pages[i]);
32586c5b
CH
168 kfree(cb->compressed_pages);
169}
170
8140dc30 171static int btrfs_decompress_bio(struct compressed_bio *cb);
48a3b636 172
4cea422a
DS
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
9ba965dc
DS
226/*
227 * Common wrappers for page allocation from compression wrappers
228 */
229struct page *btrfs_alloc_compr_page(void)
230{
4cea422a
DS
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
9ba965dc
DS
244 return alloc_page(GFP_NOFS);
245}
246
247void btrfs_free_compr_page(struct page *page)
248{
4cea422a
DS
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
9ba965dc
DS
263 ASSERT(page_ref_count(page) == 1);
264 put_page(page);
265}
266
30493ff4 267static void end_compressed_bio_read(struct btrfs_bio *bbio)
86ccbb4d 268{
544fe4a9
CH
269 struct compressed_bio *cb = to_compressed_bio(bbio);
270 blk_status_t status = bbio->bio.bi_status;
86ccbb4d 271
544fe4a9
CH
272 if (!status)
273 status = errno_to_blk_status(btrfs_decompress_bio(cb));
81bd9328 274
32586c5b 275 btrfs_free_compressed_pages(cb);
b7d463a1 276 btrfs_bio_end_io(cb->orig_bbio, status);
917f32a2 277 bio_put(&bbio->bio);
c8b97818
CM
278}
279
280/*
281 * Clear the writeback bits on all of the file
282 * pages for a compressed write
283 */
544fe4a9 284static noinline void end_compressed_writeback(const struct compressed_bio *cb)
c8b97818 285{
544fe4a9 286 struct inode *inode = &cb->bbio.inode->vfs_inode;
741ec653 287 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
09cbfeaf
KS
288 unsigned long index = cb->start >> PAGE_SHIFT;
289 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
a75b81c3 290 struct folio_batch fbatch;
ed164802 291 const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
c8b97818
CM
292 int i;
293 int ret;
294
ed164802
DS
295 if (error)
296 mapping_set_error(inode->i_mapping, error);
7bdcefc1 297
a75b81c3
VMO
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
c8b97818 306 for (i = 0; i < ret; i++) {
a75b81c3
VMO
307 struct folio *folio = fbatch.folios[i];
308
a75b81c3 309 btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
741ec653 310 cb->start, cb->len);
c8b97818 311 }
a75b81c3 312 folio_batch_release(&fbatch);
c8b97818
CM
313 }
314 /* the inode may be gone now */
c8b97818
CM
315}
316
f9327a70 317static void btrfs_finish_compressed_write_work(struct work_struct *work)
c8b97818 318{
f9327a70
CH
319 struct compressed_bio *cb =
320 container_of(work, struct compressed_bio, write_end_work);
321
7dd43954
CH
322 btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
323 cb->bbio.bio.bi_status == BLK_STS_OK);
c8b97818 324
7c0c7269 325 if (cb->writeback)
544fe4a9 326 end_compressed_writeback(cb);
6853c64a 327 /* Note, our inode could be gone now */
c8b97818 328
32586c5b 329 btrfs_free_compressed_pages(cb);
544fe4a9 330 bio_put(&cb->bbio.bio);
6853c64a
QW
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 */
917f32a2 340static void end_compressed_bio_write(struct btrfs_bio *bbio)
6853c64a 341{
544fe4a9
CH
342 struct compressed_bio *cb = to_compressed_bio(bbio);
343 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
6853c64a 344
d5e4377d 345 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
c8b97818
CM
346}
347
4513cb0c 348static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb)
10e924bc 349{
10e924bc 350 struct bio *bio = &cb->bbio.bio;
43fa4219 351 u32 offset = 0;
10e924bc 352
43fa4219
CH
353 while (offset < cb->compressed_len) {
354 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
10e924bc 355
43fa4219
CH
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;
10e924bc 360 }
10e924bc
CH
361}
362
c8b97818
CM
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 */
d611935b
CH
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)
c8b97818 377{
d611935b 378 struct btrfs_inode *inode = BTRFS_I(ordered->inode);
c7ee1819 379 struct btrfs_fs_info *fs_info = inode->root->fs_info;
c8b97818 380 struct compressed_bio *cb;
c8b97818 381
d611935b
CH
382 ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
383 ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
544fe4a9 384
d611935b
CH
385 cb = alloc_compressed_bio(inode, ordered->file_offset,
386 REQ_OP_WRITE | write_flags,
544fe4a9 387 end_compressed_bio_write);
d611935b
CH
388 cb->start = ordered->file_offset;
389 cb->len = ordered->num_bytes;
c8b97818 390 cb->compressed_pages = compressed_pages;
d611935b 391 cb->compressed_len = ordered->disk_num_bytes;
7c0c7269 392 cb->writeback = writeback;
fed8a72d 393 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
c8b97818 394 cb->nr_pages = nr_pages;
d611935b 395 cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
ec63b84d 396 cb->bbio.ordered = ordered;
4513cb0c 397 btrfs_add_compressed_bio_pages(cb);
c8b97818 398
ae42a154 399 btrfs_submit_bio(&cb->bbio, 0);
c8b97818
CM
400}
401
6a404910
QW
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 */
771ed689
CM
413static noinline int add_ra_bio_pages(struct inode *inode,
414 u64 compressed_end,
4088a47e 415 struct compressed_bio *cb,
82e60d00 416 int *memstall, unsigned long *pflags)
771ed689 417{
6a404910 418 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
771ed689 419 unsigned long end_index;
b7d463a1
CH
420 struct bio *orig_bio = &cb->orig_bbio->bio;
421 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
771ed689
CM
422 u64 isize = i_size_read(inode);
423 int ret;
424 struct page *page;
771ed689
CM
425 struct extent_map *em;
426 struct address_space *mapping = inode->i_mapping;
771ed689
CM
427 struct extent_map_tree *em_tree;
428 struct extent_io_tree *tree;
6a404910 429 int sectors_missed = 0;
771ed689 430
771ed689
CM
431 em_tree = &BTRFS_I(inode)->extent_tree;
432 tree = &BTRFS_I(inode)->io_tree;
433
434 if (isize == 0)
435 return 0;
436
ca62e85d
QW
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
09cbfeaf 447 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
771ed689 448
6a404910
QW
449 while (cur < compressed_end) {
450 u64 page_end;
451 u64 pg_index = cur >> PAGE_SHIFT;
452 u32 add_size;
771ed689 453
306e16ce 454 if (pg_index > end_index)
771ed689
CM
455 break;
456
0a943c65 457 page = xa_load(&mapping->i_pages, pg_index);
3159f943 458 if (page && !xa_is_value(page)) {
6a404910
QW
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)
771ed689 464 break;
6a404910
QW
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;
771ed689
CM
472 }
473
c62d2555
MH
474 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
475 ~__GFP_FS));
771ed689
CM
476 if (!page)
477 break;
478
c62d2555 479 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
09cbfeaf 480 put_page(page);
6a404910
QW
481 /* There is already a page, skip to page end */
482 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
483 continue;
771ed689
CM
484 }
485
82e60d00 486 if (!*memstall && PageWorkingset(page)) {
4088a47e 487 psi_memstall_enter(pflags);
82e60d00
JW
488 *memstall = 1;
489 }
4088a47e 490
32443de3
QW
491 ret = set_page_extent_mapped(page);
492 if (ret < 0) {
493 unlock_page(page);
494 put_page(page);
495 break;
496 }
497
6a404910 498 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
570eb97b 499 lock_extent(tree, cur, page_end, NULL);
890871be 500 read_lock(&em_tree->lock);
6a404910 501 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
890871be 502 read_unlock(&em_tree->lock);
771ed689 503
6a404910
QW
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)) ||
29e70be2 511 (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) {
771ed689 512 free_extent_map(em);
570eb97b 513 unlock_extent(tree, cur, page_end, NULL);
771ed689 514 unlock_page(page);
09cbfeaf 515 put_page(page);
771ed689
CM
516 break;
517 }
518 free_extent_map(em);
519
520 if (page->index == end_index) {
7073017a 521 size_t zero_offset = offset_in_page(isize);
771ed689
CM
522
523 if (zero_offset) {
524 int zeros;
09cbfeaf 525 zeros = PAGE_SIZE - zero_offset;
d048b9c2 526 memzero_page(page, zero_offset, zeros);
771ed689
CM
527 }
528 }
529
6a404910 530 add_size = min(em->start + em->len, page_end + 1) - cur;
b7d463a1 531 ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
6a404910 532 if (ret != add_size) {
570eb97b 533 unlock_extent(tree, cur, page_end, NULL);
771ed689 534 unlock_page(page);
09cbfeaf 535 put_page(page);
771ed689
CM
536 break;
537 }
6a404910
QW
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, cur, add_size);
545 put_page(page);
546 cur += add_size;
771ed689 547 }
771ed689
CM
548 return 0;
549}
550
c8b97818
CM
551/*
552 * for a compressed read, the bio we get passed has all the inode pages
553 * in it. We don't actually do IO on those pages but allocate new ones
554 * to hold the compressed pages on disk.
555 *
4f024f37 556 * bio->bi_iter.bi_sector points to the compressed extent on disk
c8b97818 557 * bio->bi_io_vec points to all of the inode pages
c8b97818
CM
558 *
559 * After the compressed pages are read, we copy the bytes into the
560 * bio we were passed and then call the bio end_io calls
561 */
e1949310 562void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
c8b97818 563{
690834e4 564 struct btrfs_inode *inode = bbio->inode;
544fe4a9
CH
565 struct btrfs_fs_info *fs_info = inode->root->fs_info;
566 struct extent_map_tree *em_tree = &inode->extent_tree;
c8b97818 567 struct compressed_bio *cb;
356b4a2d 568 unsigned int compressed_len;
690834e4 569 u64 file_offset = bbio->file_offset;
e04ca626
CM
570 u64 em_len;
571 u64 em_start;
c8b97818 572 struct extent_map *em;
82e60d00
JW
573 unsigned long pflags;
574 int memstall = 0;
f9f15de8 575 blk_status_t ret;
dd137dd1 576 int ret2;
c8b97818
CM
577
578 /* we need the actual starting offset of this extent in the file */
890871be 579 read_lock(&em_tree->lock);
557023ea 580 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
890871be 581 read_unlock(&em_tree->lock);
f9f15de8
JB
582 if (!em) {
583 ret = BLK_STS_IOERR;
584 goto out;
585 }
c8b97818 586
f86f7a75 587 ASSERT(extent_map_is_compressed(em));
d20f7043 588 compressed_len = em->block_len;
6b82ce8d 589
544fe4a9
CH
590 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
591 end_compressed_bio_read);
c8b97818 592
ff5b7ee3 593 cb->start = em->orig_start;
e04ca626
CM
594 em_len = em->len;
595 em_start = em->start;
d20f7043 596
690834e4 597 cb->len = bbio->bio.bi_iter.bi_size;
c8b97818 598 cb->compressed_len = compressed_len;
f86f7a75 599 cb->compress_type = extent_map_compression(em);
b7d463a1 600 cb->orig_bbio = bbio;
c8b97818 601
1d8fa2e2 602 free_extent_map(em);
1d8fa2e2 603
dd137dd1
STD
604 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
605 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
f9f15de8
JB
606 if (!cb->compressed_pages) {
607 ret = BLK_STS_RESOURCE;
544fe4a9 608 goto out_free_bio;
f9f15de8 609 }
6b82ce8d 610
dd137dd1
STD
611 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
612 if (ret2) {
613 ret = BLK_STS_RESOURCE;
544fe4a9 614 goto out_free_compressed_pages;
c8b97818 615 }
c8b97818 616
544fe4a9
CH
617 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
618 &pflags);
771ed689 619
771ed689 620 /* include any pages we added in add_ra-bio_pages */
690834e4 621 cb->len = bbio->bio.bi_iter.bi_size;
4513cb0c
CH
622 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
623 btrfs_add_compressed_bio_pages(cb);
524bcd1e 624
82e60d00 625 if (memstall)
4088a47e
CH
626 psi_memstall_leave(&pflags);
627
e1949310 628 btrfs_submit_bio(&cb->bbio, 0);
cb4411dd 629 return;
6b82ce8d 630
544fe4a9 631out_free_compressed_pages:
6b82ce8d 632 kfree(cb->compressed_pages);
544fe4a9 633out_free_bio:
10e924bc 634 bio_put(&cb->bbio.bio);
544fe4a9 635out:
690834e4 636 btrfs_bio_end_io(bbio, ret);
c8b97818 637}
261507a0 638
17b5a6c1
TT
639/*
640 * Heuristic uses systematic sampling to collect data from the input data
641 * range, the logic can be tuned by the following constants:
642 *
643 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
644 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
645 */
646#define SAMPLING_READ_SIZE (16)
647#define SAMPLING_INTERVAL (256)
648
649/*
650 * For statistical analysis of the input data we consider bytes that form a
651 * Galois Field of 256 objects. Each object has an attribute count, ie. how
652 * many times the object appeared in the sample.
653 */
654#define BUCKET_SIZE (256)
655
656/*
657 * The size of the sample is based on a statistical sampling rule of thumb.
658 * The common way is to perform sampling tests as long as the number of
659 * elements in each cell is at least 5.
660 *
661 * Instead of 5, we choose 32 to obtain more accurate results.
662 * If the data contain the maximum number of symbols, which is 256, we obtain a
663 * sample size bound by 8192.
664 *
665 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
666 * from up to 512 locations.
667 */
668#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
669 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
670
671struct bucket_item {
672 u32 count;
673};
4e439a0b
TT
674
675struct heuristic_ws {
17b5a6c1
TT
676 /* Partial copy of input data */
677 u8 *sample;
a440d48c 678 u32 sample_size;
17b5a6c1
TT
679 /* Buckets store counters for each byte value */
680 struct bucket_item *bucket;
440c840c
TT
681 /* Sorting buffer */
682 struct bucket_item *bucket_b;
4e439a0b
TT
683 struct list_head list;
684};
685
92ee5530
DZ
686static struct workspace_manager heuristic_wsm;
687
4e439a0b
TT
688static void free_heuristic_ws(struct list_head *ws)
689{
690 struct heuristic_ws *workspace;
691
692 workspace = list_entry(ws, struct heuristic_ws, list);
693
17b5a6c1
TT
694 kvfree(workspace->sample);
695 kfree(workspace->bucket);
440c840c 696 kfree(workspace->bucket_b);
4e439a0b
TT
697 kfree(workspace);
698}
699
7bf49943 700static struct list_head *alloc_heuristic_ws(unsigned int level)
4e439a0b
TT
701{
702 struct heuristic_ws *ws;
703
704 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
705 if (!ws)
706 return ERR_PTR(-ENOMEM);
707
17b5a6c1
TT
708 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
709 if (!ws->sample)
710 goto fail;
711
712 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
713 if (!ws->bucket)
714 goto fail;
4e439a0b 715
440c840c
TT
716 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
717 if (!ws->bucket_b)
718 goto fail;
719
17b5a6c1 720 INIT_LIST_HEAD(&ws->list);
4e439a0b 721 return &ws->list;
17b5a6c1
TT
722fail:
723 free_heuristic_ws(&ws->list);
724 return ERR_PTR(-ENOMEM);
4e439a0b
TT
725}
726
ca4ac360 727const struct btrfs_compress_op btrfs_heuristic_compress = {
be951045 728 .workspace_manager = &heuristic_wsm,
ca4ac360
DZ
729};
730
e8c9f186 731static const struct btrfs_compress_op * const btrfs_compress_op[] = {
ca4ac360
DZ
732 /* The heuristic is represented as compression type 0 */
733 &btrfs_heuristic_compress,
261507a0 734 &btrfs_zlib_compress,
a6fa6fae 735 &btrfs_lzo_compress,
5c1aab1d 736 &btrfs_zstd_compress,
261507a0
LZ
737};
738
c778df14
DS
739static struct list_head *alloc_workspace(int type, unsigned int level)
740{
741 switch (type) {
742 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
743 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
744 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
745 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
746 default:
747 /*
748 * This can't happen, the type is validated several times
749 * before we get here.
750 */
751 BUG();
752 }
753}
754
1e002351
DS
755static void free_workspace(int type, struct list_head *ws)
756{
757 switch (type) {
758 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
759 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
760 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
761 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
762 default:
763 /*
764 * This can't happen, the type is validated several times
765 * before we get here.
766 */
767 BUG();
768 }
769}
770
d5517033 771static void btrfs_init_workspace_manager(int type)
261507a0 772{
0cf25213 773 struct workspace_manager *wsm;
4e439a0b 774 struct list_head *workspace;
261507a0 775
0cf25213 776 wsm = btrfs_compress_op[type]->workspace_manager;
92ee5530
DZ
777 INIT_LIST_HEAD(&wsm->idle_ws);
778 spin_lock_init(&wsm->ws_lock);
779 atomic_set(&wsm->total_ws, 0);
780 init_waitqueue_head(&wsm->ws_wait);
f77dd0d6 781
1666edab
DZ
782 /*
783 * Preallocate one workspace for each compression type so we can
784 * guarantee forward progress in the worst case
785 */
c778df14 786 workspace = alloc_workspace(type, 0);
1666edab
DZ
787 if (IS_ERR(workspace)) {
788 pr_warn(
789 "BTRFS: cannot preallocate compression workspace, will try later\n");
790 } else {
92ee5530
DZ
791 atomic_set(&wsm->total_ws, 1);
792 wsm->free_ws = 1;
793 list_add(workspace, &wsm->idle_ws);
1666edab
DZ
794 }
795}
796
2510307e 797static void btrfs_cleanup_workspace_manager(int type)
1666edab 798{
2dba7143 799 struct workspace_manager *wsman;
1666edab
DZ
800 struct list_head *ws;
801
2dba7143 802 wsman = btrfs_compress_op[type]->workspace_manager;
1666edab
DZ
803 while (!list_empty(&wsman->idle_ws)) {
804 ws = wsman->idle_ws.next;
805 list_del(ws);
1e002351 806 free_workspace(type, ws);
1666edab 807 atomic_dec(&wsman->total_ws);
261507a0 808 }
261507a0
LZ
809}
810
811/*
e721e49d
DS
812 * This finds an available workspace or allocates a new one.
813 * If it's not possible to allocate a new one, waits until there's one.
814 * Preallocation makes a forward progress guarantees and we do not return
815 * errors.
261507a0 816 */
5907a9bb 817struct list_head *btrfs_get_workspace(int type, unsigned int level)
261507a0 818{
5907a9bb 819 struct workspace_manager *wsm;
261507a0
LZ
820 struct list_head *workspace;
821 int cpus = num_online_cpus();
fe308533 822 unsigned nofs_flag;
4e439a0b
TT
823 struct list_head *idle_ws;
824 spinlock_t *ws_lock;
825 atomic_t *total_ws;
826 wait_queue_head_t *ws_wait;
827 int *free_ws;
828
5907a9bb 829 wsm = btrfs_compress_op[type]->workspace_manager;
92ee5530
DZ
830 idle_ws = &wsm->idle_ws;
831 ws_lock = &wsm->ws_lock;
832 total_ws = &wsm->total_ws;
833 ws_wait = &wsm->ws_wait;
834 free_ws = &wsm->free_ws;
261507a0 835
261507a0 836again:
d9187649
BL
837 spin_lock(ws_lock);
838 if (!list_empty(idle_ws)) {
839 workspace = idle_ws->next;
261507a0 840 list_del(workspace);
6ac10a6a 841 (*free_ws)--;
d9187649 842 spin_unlock(ws_lock);
261507a0
LZ
843 return workspace;
844
845 }
6ac10a6a 846 if (atomic_read(total_ws) > cpus) {
261507a0
LZ
847 DEFINE_WAIT(wait);
848
d9187649
BL
849 spin_unlock(ws_lock);
850 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
6ac10a6a 851 if (atomic_read(total_ws) > cpus && !*free_ws)
261507a0 852 schedule();
d9187649 853 finish_wait(ws_wait, &wait);
261507a0
LZ
854 goto again;
855 }
6ac10a6a 856 atomic_inc(total_ws);
d9187649 857 spin_unlock(ws_lock);
261507a0 858
fe308533
DS
859 /*
860 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
861 * to turn it off here because we might get called from the restricted
862 * context of btrfs_compress_bio/btrfs_compress_pages
863 */
864 nofs_flag = memalloc_nofs_save();
c778df14 865 workspace = alloc_workspace(type, level);
fe308533
DS
866 memalloc_nofs_restore(nofs_flag);
867
261507a0 868 if (IS_ERR(workspace)) {
6ac10a6a 869 atomic_dec(total_ws);
d9187649 870 wake_up(ws_wait);
e721e49d
DS
871
872 /*
873 * Do not return the error but go back to waiting. There's a
874 * workspace preallocated for each type and the compression
875 * time is bounded so we get to a workspace eventually. This
876 * makes our caller's life easier.
52356716
DS
877 *
878 * To prevent silent and low-probability deadlocks (when the
879 * initial preallocation fails), check if there are any
880 * workspaces at all.
e721e49d 881 */
52356716
DS
882 if (atomic_read(total_ws) == 0) {
883 static DEFINE_RATELIMIT_STATE(_rs,
884 /* once per minute */ 60 * HZ,
885 /* no burst */ 1);
886
887 if (__ratelimit(&_rs)) {
ab8d0fc4 888 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
52356716
DS
889 }
890 }
e721e49d 891 goto again;
261507a0
LZ
892 }
893 return workspace;
894}
895
7bf49943 896static struct list_head *get_workspace(int type, int level)
929f4baf 897{
6a0d1272 898 switch (type) {
5907a9bb 899 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
6a0d1272 900 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
5907a9bb 901 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
6a0d1272
DS
902 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
903 default:
904 /*
905 * This can't happen, the type is validated several times
906 * before we get here.
907 */
908 BUG();
909 }
929f4baf
DZ
910}
911
261507a0
LZ
912/*
913 * put a workspace struct back on the list or free it if we have enough
914 * idle ones sitting around
915 */
a3bbd2a9 916void btrfs_put_workspace(int type, struct list_head *ws)
261507a0 917{
a3bbd2a9 918 struct workspace_manager *wsm;
4e439a0b
TT
919 struct list_head *idle_ws;
920 spinlock_t *ws_lock;
921 atomic_t *total_ws;
922 wait_queue_head_t *ws_wait;
923 int *free_ws;
924
a3bbd2a9 925 wsm = btrfs_compress_op[type]->workspace_manager;
92ee5530
DZ
926 idle_ws = &wsm->idle_ws;
927 ws_lock = &wsm->ws_lock;
928 total_ws = &wsm->total_ws;
929 ws_wait = &wsm->ws_wait;
930 free_ws = &wsm->free_ws;
d9187649
BL
931
932 spin_lock(ws_lock);
26b28dce 933 if (*free_ws <= num_online_cpus()) {
929f4baf 934 list_add(ws, idle_ws);
6ac10a6a 935 (*free_ws)++;
d9187649 936 spin_unlock(ws_lock);
261507a0
LZ
937 goto wake;
938 }
d9187649 939 spin_unlock(ws_lock);
261507a0 940
1e002351 941 free_workspace(type, ws);
6ac10a6a 942 atomic_dec(total_ws);
261507a0 943wake:
093258e6 944 cond_wake_up(ws_wait);
261507a0
LZ
945}
946
929f4baf
DZ
947static void put_workspace(int type, struct list_head *ws)
948{
bd3a5287 949 switch (type) {
a3bbd2a9
DS
950 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
951 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
952 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
bd3a5287
DS
953 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
954 default:
955 /*
956 * This can't happen, the type is validated several times
957 * before we get here.
958 */
959 BUG();
960 }
929f4baf
DZ
961}
962
adbab642
AJ
963/*
964 * Adjust @level according to the limits of the compression algorithm or
965 * fallback to default
966 */
967static unsigned int btrfs_compress_set_level(int type, unsigned level)
968{
969 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
970
971 if (level == 0)
972 level = ops->default_level;
973 else
974 level = min(level, ops->max_level);
975
976 return level;
977}
978
261507a0 979/*
38c31464
DS
980 * Given an address space and start and length, compress the bytes into @pages
981 * that are allocated on demand.
261507a0 982 *
f51d2b59
DS
983 * @type_level is encoded algorithm and level, where level 0 means whatever
984 * default the algorithm chooses and is opaque here;
985 * - compression algo are 0-3
986 * - the level are bits 4-7
987 *
4d3a800e
DS
988 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
989 * and returns number of actually allocated pages
261507a0 990 *
38c31464
DS
991 * @total_in is used to return the number of bytes actually read. It
992 * may be smaller than the input length if we had to exit early because we
261507a0
LZ
993 * ran out of room in the pages array or because we cross the
994 * max_out threshold.
995 *
38c31464
DS
996 * @total_out is an in/out parameter, must be set to the input length and will
997 * be also used to return the total number of compressed bytes
261507a0 998 */
f51d2b59 999int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
38c31464 1000 u64 start, struct page **pages,
261507a0
LZ
1001 unsigned long *out_pages,
1002 unsigned long *total_in,
e5d74902 1003 unsigned long *total_out)
261507a0 1004{
1972708a 1005 int type = btrfs_compress_type(type_level);
7bf49943 1006 int level = btrfs_compress_level(type_level);
261507a0
LZ
1007 struct list_head *workspace;
1008 int ret;
1009
b0c1fe1e 1010 level = btrfs_compress_set_level(type, level);
7bf49943 1011 workspace = get_workspace(type, level);
1e4eb746
DS
1012 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1013 out_pages, total_in, total_out);
929f4baf 1014 put_workspace(type, workspace);
261507a0
LZ
1015 return ret;
1016}
1017
8140dc30 1018static int btrfs_decompress_bio(struct compressed_bio *cb)
261507a0
LZ
1019{
1020 struct list_head *workspace;
1021 int ret;
8140dc30 1022 int type = cb->compress_type;
261507a0 1023
7bf49943 1024 workspace = get_workspace(type, 0);
4a9e803e 1025 ret = compression_decompress_bio(workspace, cb);
929f4baf 1026 put_workspace(type, workspace);
e1ddce71 1027
7edb9a3e 1028 if (!ret)
b7d463a1 1029 zero_fill_bio(&cb->orig_bbio->bio);
261507a0
LZ
1030 return ret;
1031}
1032
1033/*
1034 * a less complex decompression routine. Our compressed data fits in a
1035 * single page, and we want to read a single page out of it.
1036 * start_byte tells us the offset into the compressed data we're interested in
1037 */
3e09b5b2 1038int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
261507a0
LZ
1039 unsigned long start_byte, size_t srclen, size_t destlen)
1040{
1041 struct list_head *workspace;
1042 int ret;
1043
7bf49943 1044 workspace = get_workspace(type, 0);
1e4eb746
DS
1045 ret = compression_decompress(type, workspace, data_in, dest_page,
1046 start_byte, srclen, destlen);
929f4baf 1047 put_workspace(type, workspace);
7bf49943 1048
261507a0
LZ
1049 return ret;
1050}
1051
5565b8e0 1052int __init btrfs_init_compress(void)
1666edab 1053{
544fe4a9
CH
1054 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1055 offsetof(struct compressed_bio, bbio.bio),
1056 BIOSET_NEED_BVECS))
1057 return -ENOMEM;
4cea422a
DS
1058
1059 compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1060 if (!compr_pool.shrinker)
1061 return -ENOMEM;
1062
d5517033
DS
1063 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1064 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1065 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1066 zstd_init_workspace_manager();
4cea422a
DS
1067
1068 spin_lock_init(&compr_pool.lock);
1069 INIT_LIST_HEAD(&compr_pool.list);
1070 compr_pool.count = 0;
1071 /* 128K / 4K = 32, for 8 threads is 256 pages. */
1072 compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1073 compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1074 compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1075 compr_pool.shrinker->batch = 32;
1076 compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1077 shrinker_register(compr_pool.shrinker);
1078
5565b8e0 1079 return 0;
1666edab
DZ
1080}
1081
e67c718b 1082void __cold btrfs_exit_compress(void)
261507a0 1083{
4cea422a
DS
1084 /* For now scan drains all pages and does not touch the parameters. */
1085 btrfs_compr_pool_scan(NULL, NULL);
1086 shrinker_free(compr_pool.shrinker);
1087
2510307e
DS
1088 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1089 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1090 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1091 zstd_cleanup_workspace_manager();
544fe4a9 1092 bioset_exit(&btrfs_compressed_bioset);
261507a0 1093}
3a39c18d
LZ
1094
1095/*
1c3dc173 1096 * Copy decompressed data from working buffer to pages.
3a39c18d 1097 *
1c3dc173
QW
1098 * @buf: The decompressed data buffer
1099 * @buf_len: The decompressed data length
1100 * @decompressed: Number of bytes that are already decompressed inside the
1101 * compressed extent
1102 * @cb: The compressed extent descriptor
1103 * @orig_bio: The original bio that the caller wants to read for
3a39c18d 1104 *
1c3dc173
QW
1105 * An easier to understand graph is like below:
1106 *
1107 * |<- orig_bio ->| |<- orig_bio->|
1108 * |<------- full decompressed extent ----->|
1109 * |<----------- @cb range ---->|
1110 * | |<-- @buf_len -->|
1111 * |<--- @decompressed --->|
1112 *
1113 * Note that, @cb can be a subpage of the full decompressed extent, but
1114 * @cb->start always has the same as the orig_file_offset value of the full
1115 * decompressed extent.
1116 *
1117 * When reading compressed extent, we have to read the full compressed extent,
1118 * while @orig_bio may only want part of the range.
1119 * Thus this function will ensure only data covered by @orig_bio will be copied
1120 * to.
1121 *
1122 * Return 0 if we have copied all needed contents for @orig_bio.
1123 * Return >0 if we need continue decompress.
3a39c18d 1124 */
1c3dc173
QW
1125int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1126 struct compressed_bio *cb, u32 decompressed)
3a39c18d 1127{
b7d463a1 1128 struct bio *orig_bio = &cb->orig_bbio->bio;
1c3dc173
QW
1129 /* Offset inside the full decompressed extent */
1130 u32 cur_offset;
1131
1132 cur_offset = decompressed;
1133 /* The main loop to do the copy */
1134 while (cur_offset < decompressed + buf_len) {
1135 struct bio_vec bvec;
1136 size_t copy_len;
1137 u32 copy_start;
1138 /* Offset inside the full decompressed extent */
1139 u32 bvec_offset;
1140
1141 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1142 /*
1143 * cb->start may underflow, but subtracting that value can still
1144 * give us correct offset inside the full decompressed extent.
1145 */
1146 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
974b1adc 1147
1c3dc173
QW
1148 /* Haven't reached the bvec range, exit */
1149 if (decompressed + buf_len <= bvec_offset)
1150 return 1;
3a39c18d 1151
1c3dc173
QW
1152 copy_start = max(cur_offset, bvec_offset);
1153 copy_len = min(bvec_offset + bvec.bv_len,
1154 decompressed + buf_len) - copy_start;
1155 ASSERT(copy_len);
3a39c18d 1156
974b1adc 1157 /*
1c3dc173
QW
1158 * Extra range check to ensure we didn't go beyond
1159 * @buf + @buf_len.
974b1adc 1160 */
1c3dc173
QW
1161 ASSERT(copy_start - decompressed < buf_len);
1162 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1163 buf + copy_start - decompressed, copy_len);
1c3dc173 1164 cur_offset += copy_len;
3a39c18d 1165
1c3dc173
QW
1166 bio_advance(orig_bio, copy_len);
1167 /* Finished the bio */
1168 if (!orig_bio->bi_iter.bi_size)
1169 return 0;
3a39c18d 1170 }
3a39c18d
LZ
1171 return 1;
1172}
c2fcdcdf 1173
19562430
TT
1174/*
1175 * Shannon Entropy calculation
1176 *
52042d8e 1177 * Pure byte distribution analysis fails to determine compressibility of data.
19562430
TT
1178 * Try calculating entropy to estimate the average minimum number of bits
1179 * needed to encode the sampled data.
1180 *
1181 * For convenience, return the percentage of needed bits, instead of amount of
1182 * bits directly.
1183 *
1184 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1185 * and can be compressible with high probability
1186 *
1187 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1188 *
1189 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1190 */
1191#define ENTROPY_LVL_ACEPTABLE (65)
1192#define ENTROPY_LVL_HIGH (80)
1193
1194/*
1195 * For increasead precision in shannon_entropy calculation,
1196 * let's do pow(n, M) to save more digits after comma:
1197 *
1198 * - maximum int bit length is 64
1199 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1200 * - 13 * 4 = 52 < 64 -> M = 4
1201 *
1202 * So use pow(n, 4).
1203 */
1204static inline u32 ilog2_w(u64 n)
1205{
1206 return ilog2(n * n * n * n);
1207}
1208
1209static u32 shannon_entropy(struct heuristic_ws *ws)
1210{
1211 const u32 entropy_max = 8 * ilog2_w(2);
1212 u32 entropy_sum = 0;
1213 u32 p, p_base, sz_base;
1214 u32 i;
1215
1216 sz_base = ilog2_w(ws->sample_size);
1217 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1218 p = ws->bucket[i].count;
1219 p_base = ilog2_w(p);
1220 entropy_sum += p * (sz_base - p_base);
1221 }
1222
1223 entropy_sum /= ws->sample_size;
1224 return entropy_sum * 100 / entropy_max;
1225}
1226
440c840c
TT
1227#define RADIX_BASE 4U
1228#define COUNTERS_SIZE (1U << RADIX_BASE)
1229
1230static u8 get4bits(u64 num, int shift) {
1231 u8 low4bits;
1232
1233 num >>= shift;
1234 /* Reverse order */
1235 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1236 return low4bits;
1237}
1238
440c840c
TT
1239/*
1240 * Use 4 bits as radix base
52042d8e 1241 * Use 16 u32 counters for calculating new position in buf array
440c840c
TT
1242 *
1243 * @array - array that will be sorted
1244 * @array_buf - buffer array to store sorting results
1245 * must be equal in size to @array
1246 * @num - array size
440c840c 1247 */
23ae8c63 1248static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
36243c91 1249 int num)
858177d3 1250{
440c840c
TT
1251 u64 max_num;
1252 u64 buf_num;
1253 u32 counters[COUNTERS_SIZE];
1254 u32 new_addr;
1255 u32 addr;
1256 int bitlen;
1257 int shift;
1258 int i;
858177d3 1259
440c840c
TT
1260 /*
1261 * Try avoid useless loop iterations for small numbers stored in big
1262 * counters. Example: 48 33 4 ... in 64bit array
1263 */
23ae8c63 1264 max_num = array[0].count;
440c840c 1265 for (i = 1; i < num; i++) {
23ae8c63 1266 buf_num = array[i].count;
440c840c
TT
1267 if (buf_num > max_num)
1268 max_num = buf_num;
1269 }
1270
1271 buf_num = ilog2(max_num);
1272 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1273
1274 shift = 0;
1275 while (shift < bitlen) {
1276 memset(counters, 0, sizeof(counters));
1277
1278 for (i = 0; i < num; i++) {
23ae8c63 1279 buf_num = array[i].count;
440c840c
TT
1280 addr = get4bits(buf_num, shift);
1281 counters[addr]++;
1282 }
1283
1284 for (i = 1; i < COUNTERS_SIZE; i++)
1285 counters[i] += counters[i - 1];
1286
1287 for (i = num - 1; i >= 0; i--) {
23ae8c63 1288 buf_num = array[i].count;
440c840c
TT
1289 addr = get4bits(buf_num, shift);
1290 counters[addr]--;
1291 new_addr = counters[addr];
7add17be 1292 array_buf[new_addr] = array[i];
440c840c
TT
1293 }
1294
1295 shift += RADIX_BASE;
1296
1297 /*
1298 * Normal radix expects to move data from a temporary array, to
1299 * the main one. But that requires some CPU time. Avoid that
1300 * by doing another sort iteration to original array instead of
1301 * memcpy()
1302 */
1303 memset(counters, 0, sizeof(counters));
1304
1305 for (i = 0; i < num; i ++) {
23ae8c63 1306 buf_num = array_buf[i].count;
440c840c
TT
1307 addr = get4bits(buf_num, shift);
1308 counters[addr]++;
1309 }
1310
1311 for (i = 1; i < COUNTERS_SIZE; i++)
1312 counters[i] += counters[i - 1];
1313
1314 for (i = num - 1; i >= 0; i--) {
23ae8c63 1315 buf_num = array_buf[i].count;
440c840c
TT
1316 addr = get4bits(buf_num, shift);
1317 counters[addr]--;
1318 new_addr = counters[addr];
7add17be 1319 array[new_addr] = array_buf[i];
440c840c
TT
1320 }
1321
1322 shift += RADIX_BASE;
1323 }
858177d3
TT
1324}
1325
1326/*
1327 * Size of the core byte set - how many bytes cover 90% of the sample
1328 *
1329 * There are several types of structured binary data that use nearly all byte
1330 * values. The distribution can be uniform and counts in all buckets will be
1331 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1332 *
1333 * Other possibility is normal (Gaussian) distribution, where the data could
1334 * be potentially compressible, but we have to take a few more steps to decide
1335 * how much.
1336 *
1337 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1338 * compression algo can easy fix that
1339 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1340 * probability is not compressible
1341 */
1342#define BYTE_CORE_SET_LOW (64)
1343#define BYTE_CORE_SET_HIGH (200)
1344
1345static int byte_core_set_size(struct heuristic_ws *ws)
1346{
1347 u32 i;
1348 u32 coreset_sum = 0;
1349 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1350 struct bucket_item *bucket = ws->bucket;
1351
1352 /* Sort in reverse order */
36243c91 1353 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
858177d3
TT
1354
1355 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1356 coreset_sum += bucket[i].count;
1357
1358 if (coreset_sum > core_set_threshold)
1359 return i;
1360
1361 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1362 coreset_sum += bucket[i].count;
1363 if (coreset_sum > core_set_threshold)
1364 break;
1365 }
1366
1367 return i;
1368}
1369
a288e92c
TT
1370/*
1371 * Count byte values in buckets.
1372 * This heuristic can detect textual data (configs, xml, json, html, etc).
1373 * Because in most text-like data byte set is restricted to limited number of
1374 * possible characters, and that restriction in most cases makes data easy to
1375 * compress.
1376 *
1377 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1378 * less - compressible
1379 * more - need additional analysis
1380 */
1381#define BYTE_SET_THRESHOLD (64)
1382
1383static u32 byte_set_size(const struct heuristic_ws *ws)
1384{
1385 u32 i;
1386 u32 byte_set_size = 0;
1387
1388 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1389 if (ws->bucket[i].count > 0)
1390 byte_set_size++;
1391 }
1392
1393 /*
1394 * Continue collecting count of byte values in buckets. If the byte
1395 * set size is bigger then the threshold, it's pointless to continue,
1396 * the detection technique would fail for this type of data.
1397 */
1398 for (; i < BUCKET_SIZE; i++) {
1399 if (ws->bucket[i].count > 0) {
1400 byte_set_size++;
1401 if (byte_set_size > BYTE_SET_THRESHOLD)
1402 return byte_set_size;
1403 }
1404 }
1405
1406 return byte_set_size;
1407}
1408
1fe4f6fa
TT
1409static bool sample_repeated_patterns(struct heuristic_ws *ws)
1410{
1411 const u32 half_of_sample = ws->sample_size / 2;
1412 const u8 *data = ws->sample;
1413
1414 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1415}
1416
a440d48c
TT
1417static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1418 struct heuristic_ws *ws)
1419{
1420 struct page *page;
1421 u64 index, index_end;
1422 u32 i, curr_sample_pos;
1423 u8 *in_data;
1424
1425 /*
1426 * Compression handles the input data by chunks of 128KiB
1427 * (defined by BTRFS_MAX_UNCOMPRESSED)
1428 *
1429 * We do the same for the heuristic and loop over the whole range.
1430 *
1431 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1432 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1433 */
1434 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1435 end = start + BTRFS_MAX_UNCOMPRESSED;
1436
1437 index = start >> PAGE_SHIFT;
1438 index_end = end >> PAGE_SHIFT;
1439
1440 /* Don't miss unaligned end */
ce394a7f 1441 if (!PAGE_ALIGNED(end))
a440d48c
TT
1442 index_end++;
1443
1444 curr_sample_pos = 0;
1445 while (index < index_end) {
1446 page = find_get_page(inode->i_mapping, index);
58c1a35c 1447 in_data = kmap_local_page(page);
a440d48c
TT
1448 /* Handle case where the start is not aligned to PAGE_SIZE */
1449 i = start % PAGE_SIZE;
1450 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1451 /* Don't sample any garbage from the last page */
1452 if (start > end - SAMPLING_READ_SIZE)
1453 break;
1454 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1455 SAMPLING_READ_SIZE);
1456 i += SAMPLING_INTERVAL;
1457 start += SAMPLING_INTERVAL;
1458 curr_sample_pos += SAMPLING_READ_SIZE;
1459 }
58c1a35c 1460 kunmap_local(in_data);
a440d48c
TT
1461 put_page(page);
1462
1463 index++;
1464 }
1465
1466 ws->sample_size = curr_sample_pos;
1467}
1468
c2fcdcdf
TT
1469/*
1470 * Compression heuristic.
1471 *
1472 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1473 * quickly (compared to direct compression) detect data characteristics
67da05b3 1474 * (compressible/incompressible) to avoid wasting CPU time on incompressible
c2fcdcdf
TT
1475 * data.
1476 *
1477 * The following types of analysis can be performed:
1478 * - detect mostly zero data
1479 * - detect data with low "byte set" size (text, etc)
1480 * - detect data with low/high "core byte" set
1481 *
1482 * Return non-zero if the compression should be done, 0 otherwise.
1483 */
1484int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1485{
7bf49943 1486 struct list_head *ws_list = get_workspace(0, 0);
4e439a0b 1487 struct heuristic_ws *ws;
a440d48c
TT
1488 u32 i;
1489 u8 byte;
19562430 1490 int ret = 0;
c2fcdcdf 1491
4e439a0b
TT
1492 ws = list_entry(ws_list, struct heuristic_ws, list);
1493
a440d48c
TT
1494 heuristic_collect_sample(inode, start, end, ws);
1495
1fe4f6fa
TT
1496 if (sample_repeated_patterns(ws)) {
1497 ret = 1;
1498 goto out;
1499 }
1500
a440d48c
TT
1501 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1502
1503 for (i = 0; i < ws->sample_size; i++) {
1504 byte = ws->sample[i];
1505 ws->bucket[byte].count++;
c2fcdcdf
TT
1506 }
1507
a288e92c
TT
1508 i = byte_set_size(ws);
1509 if (i < BYTE_SET_THRESHOLD) {
1510 ret = 2;
1511 goto out;
1512 }
1513
858177d3
TT
1514 i = byte_core_set_size(ws);
1515 if (i <= BYTE_CORE_SET_LOW) {
1516 ret = 3;
1517 goto out;
1518 }
1519
1520 if (i >= BYTE_CORE_SET_HIGH) {
1521 ret = 0;
1522 goto out;
1523 }
1524
19562430
TT
1525 i = shannon_entropy(ws);
1526 if (i <= ENTROPY_LVL_ACEPTABLE) {
1527 ret = 4;
1528 goto out;
1529 }
1530
1531 /*
1532 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1533 * needed to give green light to compression.
1534 *
1535 * For now just assume that compression at that level is not worth the
1536 * resources because:
1537 *
1538 * 1. it is possible to defrag the data later
1539 *
1540 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1541 * values, every bucket has counter at level ~54. The heuristic would
1542 * be confused. This can happen when data have some internal repeated
1543 * patterns like "abbacbbc...". This can be detected by analyzing
1544 * pairs of bytes, which is too costly.
1545 */
1546 if (i < ENTROPY_LVL_HIGH) {
1547 ret = 5;
1548 goto out;
1549 } else {
1550 ret = 0;
1551 goto out;
1552 }
1553
1fe4f6fa 1554out:
929f4baf 1555 put_workspace(0, ws_list);
c2fcdcdf
TT
1556 return ret;
1557}
f51d2b59 1558
d0ab62ce
DZ
1559/*
1560 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1561 * level, unrecognized string will set the default level
1562 */
1563unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
f51d2b59 1564{
d0ab62ce
DZ
1565 unsigned int level = 0;
1566 int ret;
1567
1568 if (!type)
f51d2b59
DS
1569 return 0;
1570
d0ab62ce
DZ
1571 if (str[0] == ':') {
1572 ret = kstrtouint(str + 1, 10, &level);
1573 if (ret)
1574 level = 0;
1575 }
1576
b0c1fe1e
DS
1577 level = btrfs_compress_set_level(type, level);
1578
1579 return level;
1580}