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 "btrfs_inode.h"
29#include "bio.h"
30#include "ordered-data.h"
31#include "compression.h"
32#include "extent_io.h"
33#include "extent_map.h"
34#include "subpage.h"
35#include "messages.h"
36#include "super.h"
37
38static struct bio_set btrfs_compressed_bioset;
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
57static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)
58{
59 return container_of(bbio, struct compressed_bio, bbio);
60}
61
62static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,
63 u64 start, blk_opf_t op,
64 btrfs_bio_end_io_t end_io)
65{
66 struct btrfs_bio *bbio;
67
68 bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,
69 GFP_NOFS, &btrfs_compressed_bioset));
70 btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
71 bbio->inode = inode;
72 bbio->file_offset = start;
73 return to_compressed_bio(bbio);
74}
75
76bool btrfs_compress_is_valid_type(const char *str, size_t len)
77{
78 int i;
79
80 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
81 size_t comp_len = strlen(btrfs_compress_types[i]);
82
83 if (len < comp_len)
84 continue;
85
86 if (!strncmp(btrfs_compress_types[i], str, comp_len))
87 return true;
88 }
89 return false;
90}
91
92static int compression_compress_pages(int type, struct list_head *ws,
93 struct address_space *mapping, u64 start,
94 struct folio **folios, unsigned long *out_folios,
95 unsigned long *total_in, unsigned long *total_out)
96{
97 switch (type) {
98 case BTRFS_COMPRESS_ZLIB:
99 return zlib_compress_folios(ws, mapping, start, folios,
100 out_folios, total_in, total_out);
101 case BTRFS_COMPRESS_LZO:
102 return lzo_compress_folios(ws, mapping, start, folios,
103 out_folios, total_in, total_out);
104 case BTRFS_COMPRESS_ZSTD:
105 return zstd_compress_folios(ws, mapping, start, folios,
106 out_folios, total_in, total_out);
107 case BTRFS_COMPRESS_NONE:
108 default:
109 /*
110 * This can happen when compression races with remount setting
111 * it to 'no compress', while caller doesn't call
112 * inode_need_compress() to check if we really need to
113 * compress.
114 *
115 * Not a big deal, just need to inform caller that we
116 * haven't allocated any pages yet.
117 */
118 *out_folios = 0;
119 return -E2BIG;
120 }
121}
122
123static int compression_decompress_bio(struct list_head *ws,
124 struct compressed_bio *cb)
125{
126 switch (cb->compress_type) {
127 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
128 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
129 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
130 case BTRFS_COMPRESS_NONE:
131 default:
132 /*
133 * This can't happen, the type is validated several times
134 * before we get here.
135 */
136 BUG();
137 }
138}
139
140static int compression_decompress(int type, struct list_head *ws,
141 const u8 *data_in, struct folio *dest_folio,
142 unsigned long dest_pgoff, size_t srclen, size_t destlen)
143{
144 switch (type) {
145 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_folio,
146 dest_pgoff, srclen, destlen);
147 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_folio,
148 dest_pgoff, srclen, destlen);
149 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_folio,
150 dest_pgoff, srclen, destlen);
151 case BTRFS_COMPRESS_NONE:
152 default:
153 /*
154 * This can't happen, the type is validated several times
155 * before we get here.
156 */
157 BUG();
158 }
159}
160
161static void btrfs_free_compressed_folios(struct compressed_bio *cb)
162{
163 for (unsigned int i = 0; i < cb->nr_folios; i++)
164 btrfs_free_compr_folio(cb->compressed_folios[i]);
165 kfree(cb->compressed_folios);
166}
167
168static int btrfs_decompress_bio(struct compressed_bio *cb);
169
170/*
171 * Global cache of last unused pages for compression/decompression.
172 */
173static struct btrfs_compr_pool {
174 struct shrinker *shrinker;
175 spinlock_t lock;
176 struct list_head list;
177 int count;
178 int thresh;
179} compr_pool;
180
181static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
182{
183 int ret;
184
185 /*
186 * We must not read the values more than once if 'ret' gets expanded in
187 * the return statement so we don't accidentally return a negative
188 * number, even if the first condition finds it positive.
189 */
190 ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
191
192 return ret > 0 ? ret : 0;
193}
194
195static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
196{
197 struct list_head remove;
198 struct list_head *tmp, *next;
199 int freed;
200
201 if (compr_pool.count == 0)
202 return SHRINK_STOP;
203
204 INIT_LIST_HEAD(&remove);
205
206 /* For now, just simply drain the whole list. */
207 spin_lock(&compr_pool.lock);
208 list_splice_init(&compr_pool.list, &remove);
209 freed = compr_pool.count;
210 compr_pool.count = 0;
211 spin_unlock(&compr_pool.lock);
212
213 list_for_each_safe(tmp, next, &remove) {
214 struct page *page = list_entry(tmp, struct page, lru);
215
216 ASSERT(page_ref_count(page) == 1);
217 put_page(page);
218 }
219
220 return freed;
221}
222
223/*
224 * Common wrappers for page allocation from compression wrappers
225 */
226struct folio *btrfs_alloc_compr_folio(void)
227{
228 struct folio *folio = NULL;
229
230 spin_lock(&compr_pool.lock);
231 if (compr_pool.count > 0) {
232 folio = list_first_entry(&compr_pool.list, struct folio, lru);
233 list_del_init(&folio->lru);
234 compr_pool.count--;
235 }
236 spin_unlock(&compr_pool.lock);
237
238 if (folio)
239 return folio;
240
241 return folio_alloc(GFP_NOFS, 0);
242}
243
244void btrfs_free_compr_folio(struct folio *folio)
245{
246 bool do_free = false;
247
248 spin_lock(&compr_pool.lock);
249 if (compr_pool.count > compr_pool.thresh) {
250 do_free = true;
251 } else {
252 list_add(&folio->lru, &compr_pool.list);
253 compr_pool.count++;
254 }
255 spin_unlock(&compr_pool.lock);
256
257 if (!do_free)
258 return;
259
260 ASSERT(folio_ref_count(folio) == 1);
261 folio_put(folio);
262}
263
264static void end_bbio_compressed_read(struct btrfs_bio *bbio)
265{
266 struct compressed_bio *cb = to_compressed_bio(bbio);
267 blk_status_t status = bbio->bio.bi_status;
268
269 if (!status)
270 status = errno_to_blk_status(btrfs_decompress_bio(cb));
271
272 btrfs_free_compressed_folios(cb);
273 btrfs_bio_end_io(cb->orig_bbio, status);
274 bio_put(&bbio->bio);
275}
276
277/*
278 * Clear the writeback bits on all of the file
279 * pages for a compressed write
280 */
281static noinline void end_compressed_writeback(const struct compressed_bio *cb)
282{
283 struct inode *inode = &cb->bbio.inode->vfs_inode;
284 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
285 unsigned long index = cb->start >> PAGE_SHIFT;
286 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
287 struct folio_batch fbatch;
288 const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
289 int i;
290 int ret;
291
292 if (error)
293 mapping_set_error(inode->i_mapping, error);
294
295 folio_batch_init(&fbatch);
296 while (index <= end_index) {
297 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
298 &fbatch);
299
300 if (ret == 0)
301 return;
302
303 for (i = 0; i < ret; i++) {
304 struct folio *folio = fbatch.folios[i];
305
306 btrfs_folio_clamp_clear_writeback(fs_info, folio,
307 cb->start, cb->len);
308 }
309 folio_batch_release(&fbatch);
310 }
311 /* the inode may be gone now */
312}
313
314static void btrfs_finish_compressed_write_work(struct work_struct *work)
315{
316 struct compressed_bio *cb =
317 container_of(work, struct compressed_bio, write_end_work);
318
319 btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
320 cb->bbio.bio.bi_status == BLK_STS_OK);
321
322 if (cb->writeback)
323 end_compressed_writeback(cb);
324 /* Note, our inode could be gone now */
325
326 btrfs_free_compressed_folios(cb);
327 bio_put(&cb->bbio.bio);
328}
329
330/*
331 * Do the cleanup once all the compressed pages hit the disk. This will clear
332 * writeback on the file pages and free the compressed pages.
333 *
334 * This also calls the writeback end hooks for the file pages so that metadata
335 * and checksums can be updated in the file.
336 */
337static void end_bbio_compressed_write(struct btrfs_bio *bbio)
338{
339 struct compressed_bio *cb = to_compressed_bio(bbio);
340 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
341
342 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
343}
344
345static void btrfs_add_compressed_bio_folios(struct compressed_bio *cb)
346{
347 struct bio *bio = &cb->bbio.bio;
348 u32 offset = 0;
349
350 while (offset < cb->compressed_len) {
351 int ret;
352 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
353
354 /* Maximum compressed extent is smaller than bio size limit. */
355 ret = bio_add_folio(bio, cb->compressed_folios[offset >> PAGE_SHIFT],
356 len, 0);
357 ASSERT(ret);
358 offset += len;
359 }
360}
361
362/*
363 * worker function to build and submit bios for previously compressed pages.
364 * The corresponding pages in the inode should be marked for writeback
365 * and the compressed pages should have a reference on them for dropping
366 * when the IO is complete.
367 *
368 * This also checksums the file bytes and gets things ready for
369 * the end io hooks.
370 */
371void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
372 struct folio **compressed_folios,
373 unsigned int nr_folios,
374 blk_opf_t write_flags,
375 bool writeback)
376{
377 struct btrfs_inode *inode = ordered->inode;
378 struct btrfs_fs_info *fs_info = inode->root->fs_info;
379 struct compressed_bio *cb;
380
381 ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
382 ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
383
384 cb = alloc_compressed_bio(inode, ordered->file_offset,
385 REQ_OP_WRITE | write_flags,
386 end_bbio_compressed_write);
387 cb->start = ordered->file_offset;
388 cb->len = ordered->num_bytes;
389 cb->compressed_folios = compressed_folios;
390 cb->compressed_len = ordered->disk_num_bytes;
391 cb->writeback = writeback;
392 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
393 cb->nr_folios = nr_folios;
394 cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
395 cb->bbio.ordered = ordered;
396 btrfs_add_compressed_bio_folios(cb);
397
398 btrfs_submit_bbio(&cb->bbio, 0);
399}
400
401/*
402 * Add extra pages in the same compressed file extent so that we don't need to
403 * re-read the same extent again and again.
404 *
405 * NOTE: this won't work well for subpage, as for subpage read, we lock the
406 * full page then submit bio for each compressed/regular extents.
407 *
408 * This means, if we have several sectors in the same page points to the same
409 * on-disk compressed data, we will re-read the same extent many times and
410 * this function can only help for the next page.
411 */
412static noinline int add_ra_bio_pages(struct inode *inode,
413 u64 compressed_end,
414 struct compressed_bio *cb,
415 int *memstall, unsigned long *pflags)
416{
417 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
418 unsigned long end_index;
419 struct bio *orig_bio = &cb->orig_bbio->bio;
420 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
421 u64 isize = i_size_read(inode);
422 int ret;
423 struct folio *folio;
424 struct extent_map *em;
425 struct address_space *mapping = inode->i_mapping;
426 struct extent_map_tree *em_tree;
427 struct extent_io_tree *tree;
428 int sectors_missed = 0;
429
430 em_tree = &BTRFS_I(inode)->extent_tree;
431 tree = &BTRFS_I(inode)->io_tree;
432
433 if (isize == 0)
434 return 0;
435
436 /*
437 * For current subpage support, we only support 64K page size,
438 * which means maximum compressed extent size (128K) is just 2x page
439 * size.
440 * This makes readahead less effective, so here disable readahead for
441 * subpage for now, until full compressed write is supported.
442 */
443 if (fs_info->sectorsize < PAGE_SIZE)
444 return 0;
445
446 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
447
448 while (cur < compressed_end) {
449 u64 page_end;
450 u64 pg_index = cur >> PAGE_SHIFT;
451 u32 add_size;
452
453 if (pg_index > end_index)
454 break;
455
456 folio = filemap_get_folio(mapping, pg_index);
457 if (!IS_ERR(folio)) {
458 u64 folio_sz = folio_size(folio);
459 u64 offset = offset_in_folio(folio, cur);
460
461 folio_put(folio);
462 sectors_missed += (folio_sz - offset) >>
463 fs_info->sectorsize_bits;
464
465 /* Beyond threshold, no need to continue */
466 if (sectors_missed > 4)
467 break;
468
469 /*
470 * Jump to next page start as we already have page for
471 * current offset.
472 */
473 cur += (folio_sz - offset);
474 continue;
475 }
476
477 folio = filemap_alloc_folio(mapping_gfp_constraint(mapping,
478 ~__GFP_FS), 0);
479 if (!folio)
480 break;
481
482 if (filemap_add_folio(mapping, folio, pg_index, GFP_NOFS)) {
483 /* There is already a page, skip to page end */
484 cur += folio_size(folio);
485 folio_put(folio);
486 continue;
487 }
488
489 if (!*memstall && folio_test_workingset(folio)) {
490 psi_memstall_enter(pflags);
491 *memstall = 1;
492 }
493
494 ret = set_folio_extent_mapped(folio);
495 if (ret < 0) {
496 folio_unlock(folio);
497 folio_put(folio);
498 break;
499 }
500
501 page_end = (pg_index << PAGE_SHIFT) + folio_size(folio) - 1;
502 lock_extent(tree, cur, page_end, NULL);
503 read_lock(&em_tree->lock);
504 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
505 read_unlock(&em_tree->lock);
506
507 /*
508 * At this point, we have a locked page in the page cache for
509 * these bytes in the file. But, we have to make sure they map
510 * to this compressed extent on disk.
511 */
512 if (!em || cur < em->start ||
513 (cur + fs_info->sectorsize > extent_map_end(em)) ||
514 (extent_map_block_start(em) >> SECTOR_SHIFT) !=
515 orig_bio->bi_iter.bi_sector) {
516 free_extent_map(em);
517 unlock_extent(tree, cur, page_end, NULL);
518 folio_unlock(folio);
519 folio_put(folio);
520 break;
521 }
522 add_size = min(em->start + em->len, page_end + 1) - cur;
523 free_extent_map(em);
524 unlock_extent(tree, cur, page_end, NULL);
525
526 if (folio->index == end_index) {
527 size_t zero_offset = offset_in_folio(folio, isize);
528
529 if (zero_offset) {
530 int zeros;
531 zeros = folio_size(folio) - zero_offset;
532 folio_zero_range(folio, zero_offset, zeros);
533 }
534 }
535
536 if (!bio_add_folio(orig_bio, folio, add_size,
537 offset_in_folio(folio, cur))) {
538 folio_unlock(folio);
539 folio_put(folio);
540 break;
541 }
542 /*
543 * If it's subpage, we also need to increase its
544 * subpage::readers number, as at endio we will decrease
545 * subpage::readers and to unlock the page.
546 */
547 if (fs_info->sectorsize < PAGE_SIZE)
548 btrfs_folio_set_lock(fs_info, folio, cur, add_size);
549 folio_put(folio);
550 cur += add_size;
551 }
552 return 0;
553}
554
555/*
556 * for a compressed read, the bio we get passed has all the inode pages
557 * in it. We don't actually do IO on those pages but allocate new ones
558 * to hold the compressed pages on disk.
559 *
560 * bio->bi_iter.bi_sector points to the compressed extent on disk
561 * bio->bi_io_vec points to all of the inode pages
562 *
563 * After the compressed pages are read, we copy the bytes into the
564 * bio we were passed and then call the bio end_io calls
565 */
566void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
567{
568 struct btrfs_inode *inode = bbio->inode;
569 struct btrfs_fs_info *fs_info = inode->root->fs_info;
570 struct extent_map_tree *em_tree = &inode->extent_tree;
571 struct compressed_bio *cb;
572 unsigned int compressed_len;
573 u64 file_offset = bbio->file_offset;
574 u64 em_len;
575 u64 em_start;
576 struct extent_map *em;
577 unsigned long pflags;
578 int memstall = 0;
579 blk_status_t ret;
580 int ret2;
581
582 /* we need the actual starting offset of this extent in the file */
583 read_lock(&em_tree->lock);
584 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
585 read_unlock(&em_tree->lock);
586 if (!em) {
587 ret = BLK_STS_IOERR;
588 goto out;
589 }
590
591 ASSERT(extent_map_is_compressed(em));
592 compressed_len = em->disk_num_bytes;
593
594 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
595 end_bbio_compressed_read);
596
597 cb->start = em->start - em->offset;
598 em_len = em->len;
599 em_start = em->start;
600
601 cb->len = bbio->bio.bi_iter.bi_size;
602 cb->compressed_len = compressed_len;
603 cb->compress_type = extent_map_compression(em);
604 cb->orig_bbio = bbio;
605
606 free_extent_map(em);
607
608 cb->nr_folios = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
609 cb->compressed_folios = kcalloc(cb->nr_folios, sizeof(struct page *), GFP_NOFS);
610 if (!cb->compressed_folios) {
611 ret = BLK_STS_RESOURCE;
612 goto out_free_bio;
613 }
614
615 ret2 = btrfs_alloc_folio_array(cb->nr_folios, cb->compressed_folios);
616 if (ret2) {
617 ret = BLK_STS_RESOURCE;
618 goto out_free_compressed_pages;
619 }
620
621 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
622 &pflags);
623
624 /* include any pages we added in add_ra-bio_pages */
625 cb->len = bbio->bio.bi_iter.bi_size;
626 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
627 btrfs_add_compressed_bio_folios(cb);
628
629 if (memstall)
630 psi_memstall_leave(&pflags);
631
632 btrfs_submit_bbio(&cb->bbio, 0);
633 return;
634
635out_free_compressed_pages:
636 kfree(cb->compressed_folios);
637out_free_bio:
638 bio_put(&cb->bbio.bio);
639out:
640 btrfs_bio_end_io(bbio, ret);
641}
642
643/*
644 * Heuristic uses systematic sampling to collect data from the input data
645 * range, the logic can be tuned by the following constants:
646 *
647 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
648 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
649 */
650#define SAMPLING_READ_SIZE (16)
651#define SAMPLING_INTERVAL (256)
652
653/*
654 * For statistical analysis of the input data we consider bytes that form a
655 * Galois Field of 256 objects. Each object has an attribute count, ie. how
656 * many times the object appeared in the sample.
657 */
658#define BUCKET_SIZE (256)
659
660/*
661 * The size of the sample is based on a statistical sampling rule of thumb.
662 * The common way is to perform sampling tests as long as the number of
663 * elements in each cell is at least 5.
664 *
665 * Instead of 5, we choose 32 to obtain more accurate results.
666 * If the data contain the maximum number of symbols, which is 256, we obtain a
667 * sample size bound by 8192.
668 *
669 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
670 * from up to 512 locations.
671 */
672#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
673 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
674
675struct bucket_item {
676 u32 count;
677};
678
679struct heuristic_ws {
680 /* Partial copy of input data */
681 u8 *sample;
682 u32 sample_size;
683 /* Buckets store counters for each byte value */
684 struct bucket_item *bucket;
685 /* Sorting buffer */
686 struct bucket_item *bucket_b;
687 struct list_head list;
688};
689
690static struct workspace_manager heuristic_wsm;
691
692static void free_heuristic_ws(struct list_head *ws)
693{
694 struct heuristic_ws *workspace;
695
696 workspace = list_entry(ws, struct heuristic_ws, list);
697
698 kvfree(workspace->sample);
699 kfree(workspace->bucket);
700 kfree(workspace->bucket_b);
701 kfree(workspace);
702}
703
704static struct list_head *alloc_heuristic_ws(void)
705{
706 struct heuristic_ws *ws;
707
708 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
709 if (!ws)
710 return ERR_PTR(-ENOMEM);
711
712 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
713 if (!ws->sample)
714 goto fail;
715
716 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
717 if (!ws->bucket)
718 goto fail;
719
720 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
721 if (!ws->bucket_b)
722 goto fail;
723
724 INIT_LIST_HEAD(&ws->list);
725 return &ws->list;
726fail:
727 free_heuristic_ws(&ws->list);
728 return ERR_PTR(-ENOMEM);
729}
730
731const struct btrfs_compress_op btrfs_heuristic_compress = {
732 .workspace_manager = &heuristic_wsm,
733};
734
735static const struct btrfs_compress_op * const btrfs_compress_op[] = {
736 /* The heuristic is represented as compression type 0 */
737 &btrfs_heuristic_compress,
738 &btrfs_zlib_compress,
739 &btrfs_lzo_compress,
740 &btrfs_zstd_compress,
741};
742
743static struct list_head *alloc_workspace(int type, unsigned int level)
744{
745 switch (type) {
746 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws();
747 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
748 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace();
749 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
750 default:
751 /*
752 * This can't happen, the type is validated several times
753 * before we get here.
754 */
755 BUG();
756 }
757}
758
759static void free_workspace(int type, struct list_head *ws)
760{
761 switch (type) {
762 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
763 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
764 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
765 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
766 default:
767 /*
768 * This can't happen, the type is validated several times
769 * before we get here.
770 */
771 BUG();
772 }
773}
774
775static void btrfs_init_workspace_manager(int type)
776{
777 struct workspace_manager *wsm;
778 struct list_head *workspace;
779
780 wsm = btrfs_compress_op[type]->workspace_manager;
781 INIT_LIST_HEAD(&wsm->idle_ws);
782 spin_lock_init(&wsm->ws_lock);
783 atomic_set(&wsm->total_ws, 0);
784 init_waitqueue_head(&wsm->ws_wait);
785
786 /*
787 * Preallocate one workspace for each compression type so we can
788 * guarantee forward progress in the worst case
789 */
790 workspace = alloc_workspace(type, 0);
791 if (IS_ERR(workspace)) {
792 pr_warn(
793 "BTRFS: cannot preallocate compression workspace, will try later\n");
794 } else {
795 atomic_set(&wsm->total_ws, 1);
796 wsm->free_ws = 1;
797 list_add(workspace, &wsm->idle_ws);
798 }
799}
800
801static void btrfs_cleanup_workspace_manager(int type)
802{
803 struct workspace_manager *wsman;
804 struct list_head *ws;
805
806 wsman = btrfs_compress_op[type]->workspace_manager;
807 while (!list_empty(&wsman->idle_ws)) {
808 ws = wsman->idle_ws.next;
809 list_del(ws);
810 free_workspace(type, ws);
811 atomic_dec(&wsman->total_ws);
812 }
813}
814
815/*
816 * This finds an available workspace or allocates a new one.
817 * If it's not possible to allocate a new one, waits until there's one.
818 * Preallocation makes a forward progress guarantees and we do not return
819 * errors.
820 */
821struct list_head *btrfs_get_workspace(int type, unsigned int level)
822{
823 struct workspace_manager *wsm;
824 struct list_head *workspace;
825 int cpus = num_online_cpus();
826 unsigned nofs_flag;
827 struct list_head *idle_ws;
828 spinlock_t *ws_lock;
829 atomic_t *total_ws;
830 wait_queue_head_t *ws_wait;
831 int *free_ws;
832
833 wsm = btrfs_compress_op[type]->workspace_manager;
834 idle_ws = &wsm->idle_ws;
835 ws_lock = &wsm->ws_lock;
836 total_ws = &wsm->total_ws;
837 ws_wait = &wsm->ws_wait;
838 free_ws = &wsm->free_ws;
839
840again:
841 spin_lock(ws_lock);
842 if (!list_empty(idle_ws)) {
843 workspace = idle_ws->next;
844 list_del(workspace);
845 (*free_ws)--;
846 spin_unlock(ws_lock);
847 return workspace;
848
849 }
850 if (atomic_read(total_ws) > cpus) {
851 DEFINE_WAIT(wait);
852
853 spin_unlock(ws_lock);
854 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
855 if (atomic_read(total_ws) > cpus && !*free_ws)
856 schedule();
857 finish_wait(ws_wait, &wait);
858 goto again;
859 }
860 atomic_inc(total_ws);
861 spin_unlock(ws_lock);
862
863 /*
864 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
865 * to turn it off here because we might get called from the restricted
866 * context of btrfs_compress_bio/btrfs_compress_pages
867 */
868 nofs_flag = memalloc_nofs_save();
869 workspace = alloc_workspace(type, level);
870 memalloc_nofs_restore(nofs_flag);
871
872 if (IS_ERR(workspace)) {
873 atomic_dec(total_ws);
874 wake_up(ws_wait);
875
876 /*
877 * Do not return the error but go back to waiting. There's a
878 * workspace preallocated for each type and the compression
879 * time is bounded so we get to a workspace eventually. This
880 * makes our caller's life easier.
881 *
882 * To prevent silent and low-probability deadlocks (when the
883 * initial preallocation fails), check if there are any
884 * workspaces at all.
885 */
886 if (atomic_read(total_ws) == 0) {
887 static DEFINE_RATELIMIT_STATE(_rs,
888 /* once per minute */ 60 * HZ,
889 /* no burst */ 1);
890
891 if (__ratelimit(&_rs)) {
892 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
893 }
894 }
895 goto again;
896 }
897 return workspace;
898}
899
900static struct list_head *get_workspace(int type, int level)
901{
902 switch (type) {
903 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
904 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
905 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
906 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
907 default:
908 /*
909 * This can't happen, the type is validated several times
910 * before we get here.
911 */
912 BUG();
913 }
914}
915
916/*
917 * put a workspace struct back on the list or free it if we have enough
918 * idle ones sitting around
919 */
920void btrfs_put_workspace(int type, struct list_head *ws)
921{
922 struct workspace_manager *wsm;
923 struct list_head *idle_ws;
924 spinlock_t *ws_lock;
925 atomic_t *total_ws;
926 wait_queue_head_t *ws_wait;
927 int *free_ws;
928
929 wsm = btrfs_compress_op[type]->workspace_manager;
930 idle_ws = &wsm->idle_ws;
931 ws_lock = &wsm->ws_lock;
932 total_ws = &wsm->total_ws;
933 ws_wait = &wsm->ws_wait;
934 free_ws = &wsm->free_ws;
935
936 spin_lock(ws_lock);
937 if (*free_ws <= num_online_cpus()) {
938 list_add(ws, idle_ws);
939 (*free_ws)++;
940 spin_unlock(ws_lock);
941 goto wake;
942 }
943 spin_unlock(ws_lock);
944
945 free_workspace(type, ws);
946 atomic_dec(total_ws);
947wake:
948 cond_wake_up(ws_wait);
949}
950
951static void put_workspace(int type, struct list_head *ws)
952{
953 switch (type) {
954 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
955 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
956 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
957 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
958 default:
959 /*
960 * This can't happen, the type is validated several times
961 * before we get here.
962 */
963 BUG();
964 }
965}
966
967/*
968 * Adjust @level according to the limits of the compression algorithm or
969 * fallback to default
970 */
971static unsigned int btrfs_compress_set_level(int type, unsigned level)
972{
973 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
974
975 if (level == 0)
976 level = ops->default_level;
977 else
978 level = min(level, ops->max_level);
979
980 return level;
981}
982
983/* Wrapper around find_get_page(), with extra error message. */
984int btrfs_compress_filemap_get_folio(struct address_space *mapping, u64 start,
985 struct folio **in_folio_ret)
986{
987 struct folio *in_folio;
988
989 /*
990 * The compressed write path should have the folio locked already, thus
991 * we only need to grab one reference.
992 */
993 in_folio = filemap_get_folio(mapping, start >> PAGE_SHIFT);
994 if (IS_ERR(in_folio)) {
995 struct btrfs_inode *inode = BTRFS_I(mapping->host);
996
997 btrfs_crit(inode->root->fs_info,
998 "failed to get page cache, root %lld ino %llu file offset %llu",
999 btrfs_root_id(inode->root), btrfs_ino(inode), start);
1000 return -ENOENT;
1001 }
1002 *in_folio_ret = in_folio;
1003 return 0;
1004}
1005
1006/*
1007 * Given an address space and start and length, compress the bytes into @pages
1008 * that are allocated on demand.
1009 *
1010 * @type_level is encoded algorithm and level, where level 0 means whatever
1011 * default the algorithm chooses and is opaque here;
1012 * - compression algo are 0-3
1013 * - the level are bits 4-7
1014 *
1015 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1016 * and returns number of actually allocated pages
1017 *
1018 * @total_in is used to return the number of bytes actually read. It
1019 * may be smaller than the input length if we had to exit early because we
1020 * ran out of room in the pages array or because we cross the
1021 * max_out threshold.
1022 *
1023 * @total_out is an in/out parameter, must be set to the input length and will
1024 * be also used to return the total number of compressed bytes
1025 */
1026int btrfs_compress_folios(unsigned int type_level, struct address_space *mapping,
1027 u64 start, struct folio **folios, unsigned long *out_folios,
1028 unsigned long *total_in, unsigned long *total_out)
1029{
1030 int type = btrfs_compress_type(type_level);
1031 int level = btrfs_compress_level(type_level);
1032 const unsigned long orig_len = *total_out;
1033 struct list_head *workspace;
1034 int ret;
1035
1036 level = btrfs_compress_set_level(type, level);
1037 workspace = get_workspace(type, level);
1038 ret = compression_compress_pages(type, workspace, mapping, start, folios,
1039 out_folios, total_in, total_out);
1040 /* The total read-in bytes should be no larger than the input. */
1041 ASSERT(*total_in <= orig_len);
1042 put_workspace(type, workspace);
1043 return ret;
1044}
1045
1046static int btrfs_decompress_bio(struct compressed_bio *cb)
1047{
1048 struct list_head *workspace;
1049 int ret;
1050 int type = cb->compress_type;
1051
1052 workspace = get_workspace(type, 0);
1053 ret = compression_decompress_bio(workspace, cb);
1054 put_workspace(type, workspace);
1055
1056 if (!ret)
1057 zero_fill_bio(&cb->orig_bbio->bio);
1058 return ret;
1059}
1060
1061/*
1062 * a less complex decompression routine. Our compressed data fits in a
1063 * single page, and we want to read a single page out of it.
1064 * start_byte tells us the offset into the compressed data we're interested in
1065 */
1066int btrfs_decompress(int type, const u8 *data_in, struct folio *dest_folio,
1067 unsigned long dest_pgoff, size_t srclen, size_t destlen)
1068{
1069 struct btrfs_fs_info *fs_info = folio_to_fs_info(dest_folio);
1070 struct list_head *workspace;
1071 const u32 sectorsize = fs_info->sectorsize;
1072 int ret;
1073
1074 /*
1075 * The full destination page range should not exceed the page size.
1076 * And the @destlen should not exceed sectorsize, as this is only called for
1077 * inline file extents, which should not exceed sectorsize.
1078 */
1079 ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);
1080
1081 workspace = get_workspace(type, 0);
1082 ret = compression_decompress(type, workspace, data_in, dest_folio,
1083 dest_pgoff, srclen, destlen);
1084 put_workspace(type, workspace);
1085
1086 return ret;
1087}
1088
1089int __init btrfs_init_compress(void)
1090{
1091 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1092 offsetof(struct compressed_bio, bbio.bio),
1093 BIOSET_NEED_BVECS))
1094 return -ENOMEM;
1095
1096 compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1097 if (!compr_pool.shrinker)
1098 return -ENOMEM;
1099
1100 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1101 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1102 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1103 zstd_init_workspace_manager();
1104
1105 spin_lock_init(&compr_pool.lock);
1106 INIT_LIST_HEAD(&compr_pool.list);
1107 compr_pool.count = 0;
1108 /* 128K / 4K = 32, for 8 threads is 256 pages. */
1109 compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1110 compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1111 compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1112 compr_pool.shrinker->batch = 32;
1113 compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1114 shrinker_register(compr_pool.shrinker);
1115
1116 return 0;
1117}
1118
1119void __cold btrfs_exit_compress(void)
1120{
1121 /* For now scan drains all pages and does not touch the parameters. */
1122 btrfs_compr_pool_scan(NULL, NULL);
1123 shrinker_free(compr_pool.shrinker);
1124
1125 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1126 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1127 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1128 zstd_cleanup_workspace_manager();
1129 bioset_exit(&btrfs_compressed_bioset);
1130}
1131
1132/*
1133 * Copy decompressed data from working buffer to pages.
1134 *
1135 * @buf: The decompressed data buffer
1136 * @buf_len: The decompressed data length
1137 * @decompressed: Number of bytes that are already decompressed inside the
1138 * compressed extent
1139 * @cb: The compressed extent descriptor
1140 * @orig_bio: The original bio that the caller wants to read for
1141 *
1142 * An easier to understand graph is like below:
1143 *
1144 * |<- orig_bio ->| |<- orig_bio->|
1145 * |<------- full decompressed extent ----->|
1146 * |<----------- @cb range ---->|
1147 * | |<-- @buf_len -->|
1148 * |<--- @decompressed --->|
1149 *
1150 * Note that, @cb can be a subpage of the full decompressed extent, but
1151 * @cb->start always has the same as the orig_file_offset value of the full
1152 * decompressed extent.
1153 *
1154 * When reading compressed extent, we have to read the full compressed extent,
1155 * while @orig_bio may only want part of the range.
1156 * Thus this function will ensure only data covered by @orig_bio will be copied
1157 * to.
1158 *
1159 * Return 0 if we have copied all needed contents for @orig_bio.
1160 * Return >0 if we need continue decompress.
1161 */
1162int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1163 struct compressed_bio *cb, u32 decompressed)
1164{
1165 struct bio *orig_bio = &cb->orig_bbio->bio;
1166 /* Offset inside the full decompressed extent */
1167 u32 cur_offset;
1168
1169 cur_offset = decompressed;
1170 /* The main loop to do the copy */
1171 while (cur_offset < decompressed + buf_len) {
1172 struct bio_vec bvec;
1173 size_t copy_len;
1174 u32 copy_start;
1175 /* Offset inside the full decompressed extent */
1176 u32 bvec_offset;
1177
1178 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1179 /*
1180 * cb->start may underflow, but subtracting that value can still
1181 * give us correct offset inside the full decompressed extent.
1182 */
1183 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1184
1185 /* Haven't reached the bvec range, exit */
1186 if (decompressed + buf_len <= bvec_offset)
1187 return 1;
1188
1189 copy_start = max(cur_offset, bvec_offset);
1190 copy_len = min(bvec_offset + bvec.bv_len,
1191 decompressed + buf_len) - copy_start;
1192 ASSERT(copy_len);
1193
1194 /*
1195 * Extra range check to ensure we didn't go beyond
1196 * @buf + @buf_len.
1197 */
1198 ASSERT(copy_start - decompressed < buf_len);
1199 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1200 buf + copy_start - decompressed, copy_len);
1201 cur_offset += copy_len;
1202
1203 bio_advance(orig_bio, copy_len);
1204 /* Finished the bio */
1205 if (!orig_bio->bi_iter.bi_size)
1206 return 0;
1207 }
1208 return 1;
1209}
1210
1211/*
1212 * Shannon Entropy calculation
1213 *
1214 * Pure byte distribution analysis fails to determine compressibility of data.
1215 * Try calculating entropy to estimate the average minimum number of bits
1216 * needed to encode the sampled data.
1217 *
1218 * For convenience, return the percentage of needed bits, instead of amount of
1219 * bits directly.
1220 *
1221 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1222 * and can be compressible with high probability
1223 *
1224 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1225 *
1226 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1227 */
1228#define ENTROPY_LVL_ACEPTABLE (65)
1229#define ENTROPY_LVL_HIGH (80)
1230
1231/*
1232 * For increasead precision in shannon_entropy calculation,
1233 * let's do pow(n, M) to save more digits after comma:
1234 *
1235 * - maximum int bit length is 64
1236 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1237 * - 13 * 4 = 52 < 64 -> M = 4
1238 *
1239 * So use pow(n, 4).
1240 */
1241static inline u32 ilog2_w(u64 n)
1242{
1243 return ilog2(n * n * n * n);
1244}
1245
1246static u32 shannon_entropy(struct heuristic_ws *ws)
1247{
1248 const u32 entropy_max = 8 * ilog2_w(2);
1249 u32 entropy_sum = 0;
1250 u32 p, p_base, sz_base;
1251 u32 i;
1252
1253 sz_base = ilog2_w(ws->sample_size);
1254 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1255 p = ws->bucket[i].count;
1256 p_base = ilog2_w(p);
1257 entropy_sum += p * (sz_base - p_base);
1258 }
1259
1260 entropy_sum /= ws->sample_size;
1261 return entropy_sum * 100 / entropy_max;
1262}
1263
1264#define RADIX_BASE 4U
1265#define COUNTERS_SIZE (1U << RADIX_BASE)
1266
1267static u8 get4bits(u64 num, int shift) {
1268 u8 low4bits;
1269
1270 num >>= shift;
1271 /* Reverse order */
1272 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1273 return low4bits;
1274}
1275
1276/*
1277 * Use 4 bits as radix base
1278 * Use 16 u32 counters for calculating new position in buf array
1279 *
1280 * @array - array that will be sorted
1281 * @array_buf - buffer array to store sorting results
1282 * must be equal in size to @array
1283 * @num - array size
1284 */
1285static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1286 int num)
1287{
1288 u64 max_num;
1289 u64 buf_num;
1290 u32 counters[COUNTERS_SIZE];
1291 u32 new_addr;
1292 u32 addr;
1293 int bitlen;
1294 int shift;
1295 int i;
1296
1297 /*
1298 * Try avoid useless loop iterations for small numbers stored in big
1299 * counters. Example: 48 33 4 ... in 64bit array
1300 */
1301 max_num = array[0].count;
1302 for (i = 1; i < num; i++) {
1303 buf_num = array[i].count;
1304 if (buf_num > max_num)
1305 max_num = buf_num;
1306 }
1307
1308 buf_num = ilog2(max_num);
1309 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1310
1311 shift = 0;
1312 while (shift < bitlen) {
1313 memset(counters, 0, sizeof(counters));
1314
1315 for (i = 0; i < num; i++) {
1316 buf_num = array[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[i].count;
1326 addr = get4bits(buf_num, shift);
1327 counters[addr]--;
1328 new_addr = counters[addr];
1329 array_buf[new_addr] = array[i];
1330 }
1331
1332 shift += RADIX_BASE;
1333
1334 /*
1335 * Normal radix expects to move data from a temporary array, to
1336 * the main one. But that requires some CPU time. Avoid that
1337 * by doing another sort iteration to original array instead of
1338 * memcpy()
1339 */
1340 memset(counters, 0, sizeof(counters));
1341
1342 for (i = 0; i < num; i ++) {
1343 buf_num = array_buf[i].count;
1344 addr = get4bits(buf_num, shift);
1345 counters[addr]++;
1346 }
1347
1348 for (i = 1; i < COUNTERS_SIZE; i++)
1349 counters[i] += counters[i - 1];
1350
1351 for (i = num - 1; i >= 0; i--) {
1352 buf_num = array_buf[i].count;
1353 addr = get4bits(buf_num, shift);
1354 counters[addr]--;
1355 new_addr = counters[addr];
1356 array[new_addr] = array_buf[i];
1357 }
1358
1359 shift += RADIX_BASE;
1360 }
1361}
1362
1363/*
1364 * Size of the core byte set - how many bytes cover 90% of the sample
1365 *
1366 * There are several types of structured binary data that use nearly all byte
1367 * values. The distribution can be uniform and counts in all buckets will be
1368 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1369 *
1370 * Other possibility is normal (Gaussian) distribution, where the data could
1371 * be potentially compressible, but we have to take a few more steps to decide
1372 * how much.
1373 *
1374 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1375 * compression algo can easy fix that
1376 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1377 * probability is not compressible
1378 */
1379#define BYTE_CORE_SET_LOW (64)
1380#define BYTE_CORE_SET_HIGH (200)
1381
1382static int byte_core_set_size(struct heuristic_ws *ws)
1383{
1384 u32 i;
1385 u32 coreset_sum = 0;
1386 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1387 struct bucket_item *bucket = ws->bucket;
1388
1389 /* Sort in reverse order */
1390 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1391
1392 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1393 coreset_sum += bucket[i].count;
1394
1395 if (coreset_sum > core_set_threshold)
1396 return i;
1397
1398 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1399 coreset_sum += bucket[i].count;
1400 if (coreset_sum > core_set_threshold)
1401 break;
1402 }
1403
1404 return i;
1405}
1406
1407/*
1408 * Count byte values in buckets.
1409 * This heuristic can detect textual data (configs, xml, json, html, etc).
1410 * Because in most text-like data byte set is restricted to limited number of
1411 * possible characters, and that restriction in most cases makes data easy to
1412 * compress.
1413 *
1414 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1415 * less - compressible
1416 * more - need additional analysis
1417 */
1418#define BYTE_SET_THRESHOLD (64)
1419
1420static u32 byte_set_size(const struct heuristic_ws *ws)
1421{
1422 u32 i;
1423 u32 byte_set_size = 0;
1424
1425 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1426 if (ws->bucket[i].count > 0)
1427 byte_set_size++;
1428 }
1429
1430 /*
1431 * Continue collecting count of byte values in buckets. If the byte
1432 * set size is bigger then the threshold, it's pointless to continue,
1433 * the detection technique would fail for this type of data.
1434 */
1435 for (; i < BUCKET_SIZE; i++) {
1436 if (ws->bucket[i].count > 0) {
1437 byte_set_size++;
1438 if (byte_set_size > BYTE_SET_THRESHOLD)
1439 return byte_set_size;
1440 }
1441 }
1442
1443 return byte_set_size;
1444}
1445
1446static bool sample_repeated_patterns(struct heuristic_ws *ws)
1447{
1448 const u32 half_of_sample = ws->sample_size / 2;
1449 const u8 *data = ws->sample;
1450
1451 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1452}
1453
1454static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1455 struct heuristic_ws *ws)
1456{
1457 struct page *page;
1458 u64 index, index_end;
1459 u32 i, curr_sample_pos;
1460 u8 *in_data;
1461
1462 /*
1463 * Compression handles the input data by chunks of 128KiB
1464 * (defined by BTRFS_MAX_UNCOMPRESSED)
1465 *
1466 * We do the same for the heuristic and loop over the whole range.
1467 *
1468 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1469 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1470 */
1471 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1472 end = start + BTRFS_MAX_UNCOMPRESSED;
1473
1474 index = start >> PAGE_SHIFT;
1475 index_end = end >> PAGE_SHIFT;
1476
1477 /* Don't miss unaligned end */
1478 if (!PAGE_ALIGNED(end))
1479 index_end++;
1480
1481 curr_sample_pos = 0;
1482 while (index < index_end) {
1483 page = find_get_page(inode->i_mapping, index);
1484 in_data = kmap_local_page(page);
1485 /* Handle case where the start is not aligned to PAGE_SIZE */
1486 i = start % PAGE_SIZE;
1487 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1488 /* Don't sample any garbage from the last page */
1489 if (start > end - SAMPLING_READ_SIZE)
1490 break;
1491 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1492 SAMPLING_READ_SIZE);
1493 i += SAMPLING_INTERVAL;
1494 start += SAMPLING_INTERVAL;
1495 curr_sample_pos += SAMPLING_READ_SIZE;
1496 }
1497 kunmap_local(in_data);
1498 put_page(page);
1499
1500 index++;
1501 }
1502
1503 ws->sample_size = curr_sample_pos;
1504}
1505
1506/*
1507 * Compression heuristic.
1508 *
1509 * The following types of analysis can be performed:
1510 * - detect mostly zero data
1511 * - detect data with low "byte set" size (text, etc)
1512 * - detect data with low/high "core byte" set
1513 *
1514 * Return non-zero if the compression should be done, 0 otherwise.
1515 */
1516int btrfs_compress_heuristic(struct btrfs_inode *inode, u64 start, u64 end)
1517{
1518 struct list_head *ws_list = get_workspace(0, 0);
1519 struct heuristic_ws *ws;
1520 u32 i;
1521 u8 byte;
1522 int ret = 0;
1523
1524 ws = list_entry(ws_list, struct heuristic_ws, list);
1525
1526 heuristic_collect_sample(&inode->vfs_inode, start, end, ws);
1527
1528 if (sample_repeated_patterns(ws)) {
1529 ret = 1;
1530 goto out;
1531 }
1532
1533 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1534
1535 for (i = 0; i < ws->sample_size; i++) {
1536 byte = ws->sample[i];
1537 ws->bucket[byte].count++;
1538 }
1539
1540 i = byte_set_size(ws);
1541 if (i < BYTE_SET_THRESHOLD) {
1542 ret = 2;
1543 goto out;
1544 }
1545
1546 i = byte_core_set_size(ws);
1547 if (i <= BYTE_CORE_SET_LOW) {
1548 ret = 3;
1549 goto out;
1550 }
1551
1552 if (i >= BYTE_CORE_SET_HIGH) {
1553 ret = 0;
1554 goto out;
1555 }
1556
1557 i = shannon_entropy(ws);
1558 if (i <= ENTROPY_LVL_ACEPTABLE) {
1559 ret = 4;
1560 goto out;
1561 }
1562
1563 /*
1564 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1565 * needed to give green light to compression.
1566 *
1567 * For now just assume that compression at that level is not worth the
1568 * resources because:
1569 *
1570 * 1. it is possible to defrag the data later
1571 *
1572 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1573 * values, every bucket has counter at level ~54. The heuristic would
1574 * be confused. This can happen when data have some internal repeated
1575 * patterns like "abbacbbc...". This can be detected by analyzing
1576 * pairs of bytes, which is too costly.
1577 */
1578 if (i < ENTROPY_LVL_HIGH) {
1579 ret = 5;
1580 goto out;
1581 } else {
1582 ret = 0;
1583 goto out;
1584 }
1585
1586out:
1587 put_workspace(0, ws_list);
1588 return ret;
1589}
1590
1591/*
1592 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1593 * level, unrecognized string will set the default level
1594 */
1595unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1596{
1597 unsigned int level = 0;
1598 int ret;
1599
1600 if (!type)
1601 return 0;
1602
1603 if (str[0] == ':') {
1604 ret = kstrtouint(str + 1, 10, &level);
1605 if (ret)
1606 level = 0;
1607 }
1608
1609 level = btrfs_compress_set_level(type, level);
1610
1611 return level;
1612}
1/*
2 * Copyright (C) 2008 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/kernel.h>
20#include <linux/bio.h>
21#include <linux/buffer_head.h>
22#include <linux/file.h>
23#include <linux/fs.h>
24#include <linux/pagemap.h>
25#include <linux/highmem.h>
26#include <linux/time.h>
27#include <linux/init.h>
28#include <linux/string.h>
29#include <linux/backing-dev.h>
30#include <linux/mpage.h>
31#include <linux/swap.h>
32#include <linux/writeback.h>
33#include <linux/bit_spinlock.h>
34#include <linux/slab.h>
35#include "ctree.h"
36#include "disk-io.h"
37#include "transaction.h"
38#include "btrfs_inode.h"
39#include "volumes.h"
40#include "ordered-data.h"
41#include "compression.h"
42#include "extent_io.h"
43#include "extent_map.h"
44
45struct compressed_bio {
46 /* number of bios pending for this compressed extent */
47 atomic_t pending_bios;
48
49 /* the pages with the compressed data on them */
50 struct page **compressed_pages;
51
52 /* inode that owns this data */
53 struct inode *inode;
54
55 /* starting offset in the inode for our pages */
56 u64 start;
57
58 /* number of bytes in the inode we're working on */
59 unsigned long len;
60
61 /* number of bytes on disk */
62 unsigned long compressed_len;
63
64 /* the compression algorithm for this bio */
65 int compress_type;
66
67 /* number of compressed pages in the array */
68 unsigned long nr_pages;
69
70 /* IO errors */
71 int errors;
72 int mirror_num;
73
74 /* for reads, this is the bio we are copying the data into */
75 struct bio *orig_bio;
76
77 /*
78 * the start of a variable length array of checksums only
79 * used by reads
80 */
81 u32 sums;
82};
83
84static int btrfs_decompress_biovec(int type, struct page **pages_in,
85 u64 disk_start, struct bio_vec *bvec,
86 int vcnt, size_t srclen);
87
88static inline int compressed_bio_size(struct btrfs_root *root,
89 unsigned long disk_size)
90{
91 u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
92
93 return sizeof(struct compressed_bio) +
94 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
95 csum_size;
96}
97
98static struct bio *compressed_bio_alloc(struct block_device *bdev,
99 u64 first_byte, gfp_t gfp_flags)
100{
101 int nr_vecs;
102
103 nr_vecs = bio_get_nr_vecs(bdev);
104 return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
105}
106
107static int check_compressed_csum(struct inode *inode,
108 struct compressed_bio *cb,
109 u64 disk_start)
110{
111 int ret;
112 struct page *page;
113 unsigned long i;
114 char *kaddr;
115 u32 csum;
116 u32 *cb_sum = &cb->sums;
117
118 if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
119 return 0;
120
121 for (i = 0; i < cb->nr_pages; i++) {
122 page = cb->compressed_pages[i];
123 csum = ~(u32)0;
124
125 kaddr = kmap_atomic(page);
126 csum = btrfs_csum_data(kaddr, csum, PAGE_CACHE_SIZE);
127 btrfs_csum_final(csum, (char *)&csum);
128 kunmap_atomic(kaddr);
129
130 if (csum != *cb_sum) {
131 btrfs_info(BTRFS_I(inode)->root->fs_info,
132 "csum failed ino %llu extent %llu csum %u wanted %u mirror %d",
133 btrfs_ino(inode), disk_start, csum, *cb_sum,
134 cb->mirror_num);
135 ret = -EIO;
136 goto fail;
137 }
138 cb_sum++;
139
140 }
141 ret = 0;
142fail:
143 return ret;
144}
145
146/* when we finish reading compressed pages from the disk, we
147 * decompress them and then run the bio end_io routines on the
148 * decompressed pages (in the inode address space).
149 *
150 * This allows the checksumming and other IO error handling routines
151 * to work normally
152 *
153 * The compressed pages are freed here, and it must be run
154 * in process context
155 */
156static void end_compressed_bio_read(struct bio *bio, int err)
157{
158 struct compressed_bio *cb = bio->bi_private;
159 struct inode *inode;
160 struct page *page;
161 unsigned long index;
162 int ret;
163
164 if (err)
165 cb->errors = 1;
166
167 /* if there are more bios still pending for this compressed
168 * extent, just exit
169 */
170 if (!atomic_dec_and_test(&cb->pending_bios))
171 goto out;
172
173 inode = cb->inode;
174 ret = check_compressed_csum(inode, cb,
175 (u64)bio->bi_iter.bi_sector << 9);
176 if (ret)
177 goto csum_failed;
178
179 /* ok, we're the last bio for this extent, lets start
180 * the decompression.
181 */
182 ret = btrfs_decompress_biovec(cb->compress_type,
183 cb->compressed_pages,
184 cb->start,
185 cb->orig_bio->bi_io_vec,
186 cb->orig_bio->bi_vcnt,
187 cb->compressed_len);
188csum_failed:
189 if (ret)
190 cb->errors = 1;
191
192 /* release the compressed pages */
193 index = 0;
194 for (index = 0; index < cb->nr_pages; index++) {
195 page = cb->compressed_pages[index];
196 page->mapping = NULL;
197 page_cache_release(page);
198 }
199
200 /* do io completion on the original bio */
201 if (cb->errors) {
202 bio_io_error(cb->orig_bio);
203 } else {
204 int i;
205 struct bio_vec *bvec;
206
207 /*
208 * we have verified the checksum already, set page
209 * checked so the end_io handlers know about it
210 */
211 bio_for_each_segment_all(bvec, cb->orig_bio, i)
212 SetPageChecked(bvec->bv_page);
213
214 bio_endio(cb->orig_bio, 0);
215 }
216
217 /* finally free the cb struct */
218 kfree(cb->compressed_pages);
219 kfree(cb);
220out:
221 bio_put(bio);
222}
223
224/*
225 * Clear the writeback bits on all of the file
226 * pages for a compressed write
227 */
228static noinline void end_compressed_writeback(struct inode *inode, u64 start,
229 unsigned long ram_size)
230{
231 unsigned long index = start >> PAGE_CACHE_SHIFT;
232 unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
233 struct page *pages[16];
234 unsigned long nr_pages = end_index - index + 1;
235 int i;
236 int ret;
237
238 while (nr_pages > 0) {
239 ret = find_get_pages_contig(inode->i_mapping, index,
240 min_t(unsigned long,
241 nr_pages, ARRAY_SIZE(pages)), pages);
242 if (ret == 0) {
243 nr_pages -= 1;
244 index += 1;
245 continue;
246 }
247 for (i = 0; i < ret; i++) {
248 end_page_writeback(pages[i]);
249 page_cache_release(pages[i]);
250 }
251 nr_pages -= ret;
252 index += ret;
253 }
254 /* the inode may be gone now */
255}
256
257/*
258 * do the cleanup once all the compressed pages hit the disk.
259 * This will clear writeback on the file pages and free the compressed
260 * pages.
261 *
262 * This also calls the writeback end hooks for the file pages so that
263 * metadata and checksums can be updated in the file.
264 */
265static void end_compressed_bio_write(struct bio *bio, int err)
266{
267 struct extent_io_tree *tree;
268 struct compressed_bio *cb = bio->bi_private;
269 struct inode *inode;
270 struct page *page;
271 unsigned long index;
272
273 if (err)
274 cb->errors = 1;
275
276 /* if there are more bios still pending for this compressed
277 * extent, just exit
278 */
279 if (!atomic_dec_and_test(&cb->pending_bios))
280 goto out;
281
282 /* ok, we're the last bio for this extent, step one is to
283 * call back into the FS and do all the end_io operations
284 */
285 inode = cb->inode;
286 tree = &BTRFS_I(inode)->io_tree;
287 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
288 tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
289 cb->start,
290 cb->start + cb->len - 1,
291 NULL, 1);
292 cb->compressed_pages[0]->mapping = NULL;
293
294 end_compressed_writeback(inode, cb->start, cb->len);
295 /* note, our inode could be gone now */
296
297 /*
298 * release the compressed pages, these came from alloc_page and
299 * are not attached to the inode at all
300 */
301 index = 0;
302 for (index = 0; index < cb->nr_pages; index++) {
303 page = cb->compressed_pages[index];
304 page->mapping = NULL;
305 page_cache_release(page);
306 }
307
308 /* finally free the cb struct */
309 kfree(cb->compressed_pages);
310 kfree(cb);
311out:
312 bio_put(bio);
313}
314
315/*
316 * worker function to build and submit bios for previously compressed pages.
317 * The corresponding pages in the inode should be marked for writeback
318 * and the compressed pages should have a reference on them for dropping
319 * when the IO is complete.
320 *
321 * This also checksums the file bytes and gets things ready for
322 * the end io hooks.
323 */
324int btrfs_submit_compressed_write(struct inode *inode, u64 start,
325 unsigned long len, u64 disk_start,
326 unsigned long compressed_len,
327 struct page **compressed_pages,
328 unsigned long nr_pages)
329{
330 struct bio *bio = NULL;
331 struct btrfs_root *root = BTRFS_I(inode)->root;
332 struct compressed_bio *cb;
333 unsigned long bytes_left;
334 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
335 int pg_index = 0;
336 struct page *page;
337 u64 first_byte = disk_start;
338 struct block_device *bdev;
339 int ret;
340 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
341
342 WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
343 cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
344 if (!cb)
345 return -ENOMEM;
346 atomic_set(&cb->pending_bios, 0);
347 cb->errors = 0;
348 cb->inode = inode;
349 cb->start = start;
350 cb->len = len;
351 cb->mirror_num = 0;
352 cb->compressed_pages = compressed_pages;
353 cb->compressed_len = compressed_len;
354 cb->orig_bio = NULL;
355 cb->nr_pages = nr_pages;
356
357 bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
358
359 bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
360 if (!bio) {
361 kfree(cb);
362 return -ENOMEM;
363 }
364 bio->bi_private = cb;
365 bio->bi_end_io = end_compressed_bio_write;
366 atomic_inc(&cb->pending_bios);
367
368 /* create and submit bios for the compressed pages */
369 bytes_left = compressed_len;
370 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
371 page = compressed_pages[pg_index];
372 page->mapping = inode->i_mapping;
373 if (bio->bi_iter.bi_size)
374 ret = io_tree->ops->merge_bio_hook(WRITE, page, 0,
375 PAGE_CACHE_SIZE,
376 bio, 0);
377 else
378 ret = 0;
379
380 page->mapping = NULL;
381 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
382 PAGE_CACHE_SIZE) {
383 bio_get(bio);
384
385 /*
386 * inc the count before we submit the bio so
387 * we know the end IO handler won't happen before
388 * we inc the count. Otherwise, the cb might get
389 * freed before we're done setting it up
390 */
391 atomic_inc(&cb->pending_bios);
392 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
393 BUG_ON(ret); /* -ENOMEM */
394
395 if (!skip_sum) {
396 ret = btrfs_csum_one_bio(root, inode, bio,
397 start, 1);
398 BUG_ON(ret); /* -ENOMEM */
399 }
400
401 ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
402 BUG_ON(ret); /* -ENOMEM */
403
404 bio_put(bio);
405
406 bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
407 BUG_ON(!bio);
408 bio->bi_private = cb;
409 bio->bi_end_io = end_compressed_bio_write;
410 bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
411 }
412 if (bytes_left < PAGE_CACHE_SIZE) {
413 btrfs_info(BTRFS_I(inode)->root->fs_info,
414 "bytes left %lu compress len %lu nr %lu",
415 bytes_left, cb->compressed_len, cb->nr_pages);
416 }
417 bytes_left -= PAGE_CACHE_SIZE;
418 first_byte += PAGE_CACHE_SIZE;
419 cond_resched();
420 }
421 bio_get(bio);
422
423 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
424 BUG_ON(ret); /* -ENOMEM */
425
426 if (!skip_sum) {
427 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
428 BUG_ON(ret); /* -ENOMEM */
429 }
430
431 ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
432 BUG_ON(ret); /* -ENOMEM */
433
434 bio_put(bio);
435 return 0;
436}
437
438static noinline int add_ra_bio_pages(struct inode *inode,
439 u64 compressed_end,
440 struct compressed_bio *cb)
441{
442 unsigned long end_index;
443 unsigned long pg_index;
444 u64 last_offset;
445 u64 isize = i_size_read(inode);
446 int ret;
447 struct page *page;
448 unsigned long nr_pages = 0;
449 struct extent_map *em;
450 struct address_space *mapping = inode->i_mapping;
451 struct extent_map_tree *em_tree;
452 struct extent_io_tree *tree;
453 u64 end;
454 int misses = 0;
455
456 page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
457 last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
458 em_tree = &BTRFS_I(inode)->extent_tree;
459 tree = &BTRFS_I(inode)->io_tree;
460
461 if (isize == 0)
462 return 0;
463
464 end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
465
466 while (last_offset < compressed_end) {
467 pg_index = last_offset >> PAGE_CACHE_SHIFT;
468
469 if (pg_index > end_index)
470 break;
471
472 rcu_read_lock();
473 page = radix_tree_lookup(&mapping->page_tree, pg_index);
474 rcu_read_unlock();
475 if (page && !radix_tree_exceptional_entry(page)) {
476 misses++;
477 if (misses > 4)
478 break;
479 goto next;
480 }
481
482 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
483 ~__GFP_FS);
484 if (!page)
485 break;
486
487 if (add_to_page_cache_lru(page, mapping, pg_index,
488 GFP_NOFS)) {
489 page_cache_release(page);
490 goto next;
491 }
492
493 end = last_offset + PAGE_CACHE_SIZE - 1;
494 /*
495 * at this point, we have a locked page in the page cache
496 * for these bytes in the file. But, we have to make
497 * sure they map to this compressed extent on disk.
498 */
499 set_page_extent_mapped(page);
500 lock_extent(tree, last_offset, end);
501 read_lock(&em_tree->lock);
502 em = lookup_extent_mapping(em_tree, last_offset,
503 PAGE_CACHE_SIZE);
504 read_unlock(&em_tree->lock);
505
506 if (!em || last_offset < em->start ||
507 (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
508 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
509 free_extent_map(em);
510 unlock_extent(tree, last_offset, end);
511 unlock_page(page);
512 page_cache_release(page);
513 break;
514 }
515 free_extent_map(em);
516
517 if (page->index == end_index) {
518 char *userpage;
519 size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
520
521 if (zero_offset) {
522 int zeros;
523 zeros = PAGE_CACHE_SIZE - zero_offset;
524 userpage = kmap_atomic(page);
525 memset(userpage + zero_offset, 0, zeros);
526 flush_dcache_page(page);
527 kunmap_atomic(userpage);
528 }
529 }
530
531 ret = bio_add_page(cb->orig_bio, page,
532 PAGE_CACHE_SIZE, 0);
533
534 if (ret == PAGE_CACHE_SIZE) {
535 nr_pages++;
536 page_cache_release(page);
537 } else {
538 unlock_extent(tree, last_offset, end);
539 unlock_page(page);
540 page_cache_release(page);
541 break;
542 }
543next:
544 last_offset += PAGE_CACHE_SIZE;
545 }
546 return 0;
547}
548
549/*
550 * for a compressed read, the bio we get passed has all the inode pages
551 * in it. We don't actually do IO on those pages but allocate new ones
552 * to hold the compressed pages on disk.
553 *
554 * bio->bi_iter.bi_sector points to the compressed extent on disk
555 * bio->bi_io_vec points to all of the inode pages
556 * bio->bi_vcnt is a count of pages
557 *
558 * After the compressed pages are read, we copy the bytes into the
559 * bio we were passed and then call the bio end_io calls
560 */
561int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
562 int mirror_num, unsigned long bio_flags)
563{
564 struct extent_io_tree *tree;
565 struct extent_map_tree *em_tree;
566 struct compressed_bio *cb;
567 struct btrfs_root *root = BTRFS_I(inode)->root;
568 unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
569 unsigned long compressed_len;
570 unsigned long nr_pages;
571 unsigned long pg_index;
572 struct page *page;
573 struct block_device *bdev;
574 struct bio *comp_bio;
575 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
576 u64 em_len;
577 u64 em_start;
578 struct extent_map *em;
579 int ret = -ENOMEM;
580 int faili = 0;
581 u32 *sums;
582
583 tree = &BTRFS_I(inode)->io_tree;
584 em_tree = &BTRFS_I(inode)->extent_tree;
585
586 /* we need the actual starting offset of this extent in the file */
587 read_lock(&em_tree->lock);
588 em = lookup_extent_mapping(em_tree,
589 page_offset(bio->bi_io_vec->bv_page),
590 PAGE_CACHE_SIZE);
591 read_unlock(&em_tree->lock);
592 if (!em)
593 return -EIO;
594
595 compressed_len = em->block_len;
596 cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
597 if (!cb)
598 goto out;
599
600 atomic_set(&cb->pending_bios, 0);
601 cb->errors = 0;
602 cb->inode = inode;
603 cb->mirror_num = mirror_num;
604 sums = &cb->sums;
605
606 cb->start = em->orig_start;
607 em_len = em->len;
608 em_start = em->start;
609
610 free_extent_map(em);
611 em = NULL;
612
613 cb->len = uncompressed_len;
614 cb->compressed_len = compressed_len;
615 cb->compress_type = extent_compress_type(bio_flags);
616 cb->orig_bio = bio;
617
618 nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
619 PAGE_CACHE_SIZE;
620 cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
621 GFP_NOFS);
622 if (!cb->compressed_pages)
623 goto fail1;
624
625 bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
626
627 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
628 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
629 __GFP_HIGHMEM);
630 if (!cb->compressed_pages[pg_index]) {
631 faili = pg_index - 1;
632 ret = -ENOMEM;
633 goto fail2;
634 }
635 }
636 faili = nr_pages - 1;
637 cb->nr_pages = nr_pages;
638
639 /* In the parent-locked case, we only locked the range we are
640 * interested in. In all other cases, we can opportunistically
641 * cache decompressed data that goes beyond the requested range. */
642 if (!(bio_flags & EXTENT_BIO_PARENT_LOCKED))
643 add_ra_bio_pages(inode, em_start + em_len, cb);
644
645 /* include any pages we added in add_ra-bio_pages */
646 uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
647 cb->len = uncompressed_len;
648
649 comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
650 if (!comp_bio)
651 goto fail2;
652 comp_bio->bi_private = cb;
653 comp_bio->bi_end_io = end_compressed_bio_read;
654 atomic_inc(&cb->pending_bios);
655
656 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
657 page = cb->compressed_pages[pg_index];
658 page->mapping = inode->i_mapping;
659 page->index = em_start >> PAGE_CACHE_SHIFT;
660
661 if (comp_bio->bi_iter.bi_size)
662 ret = tree->ops->merge_bio_hook(READ, page, 0,
663 PAGE_CACHE_SIZE,
664 comp_bio, 0);
665 else
666 ret = 0;
667
668 page->mapping = NULL;
669 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
670 PAGE_CACHE_SIZE) {
671 bio_get(comp_bio);
672
673 ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
674 BUG_ON(ret); /* -ENOMEM */
675
676 /*
677 * inc the count before we submit the bio so
678 * we know the end IO handler won't happen before
679 * we inc the count. Otherwise, the cb might get
680 * freed before we're done setting it up
681 */
682 atomic_inc(&cb->pending_bios);
683
684 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
685 ret = btrfs_lookup_bio_sums(root, inode,
686 comp_bio, sums);
687 BUG_ON(ret); /* -ENOMEM */
688 }
689 sums += (comp_bio->bi_iter.bi_size +
690 root->sectorsize - 1) / root->sectorsize;
691
692 ret = btrfs_map_bio(root, READ, comp_bio,
693 mirror_num, 0);
694 if (ret)
695 bio_endio(comp_bio, ret);
696
697 bio_put(comp_bio);
698
699 comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
700 GFP_NOFS);
701 BUG_ON(!comp_bio);
702 comp_bio->bi_private = cb;
703 comp_bio->bi_end_io = end_compressed_bio_read;
704
705 bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
706 }
707 cur_disk_byte += PAGE_CACHE_SIZE;
708 }
709 bio_get(comp_bio);
710
711 ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
712 BUG_ON(ret); /* -ENOMEM */
713
714 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
715 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
716 BUG_ON(ret); /* -ENOMEM */
717 }
718
719 ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
720 if (ret)
721 bio_endio(comp_bio, ret);
722
723 bio_put(comp_bio);
724 return 0;
725
726fail2:
727 while (faili >= 0) {
728 __free_page(cb->compressed_pages[faili]);
729 faili--;
730 }
731
732 kfree(cb->compressed_pages);
733fail1:
734 kfree(cb);
735out:
736 free_extent_map(em);
737 return ret;
738}
739
740static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
741static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
742static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
743static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
744static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
745
746static struct btrfs_compress_op *btrfs_compress_op[] = {
747 &btrfs_zlib_compress,
748 &btrfs_lzo_compress,
749};
750
751void __init btrfs_init_compress(void)
752{
753 int i;
754
755 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
756 INIT_LIST_HEAD(&comp_idle_workspace[i]);
757 spin_lock_init(&comp_workspace_lock[i]);
758 atomic_set(&comp_alloc_workspace[i], 0);
759 init_waitqueue_head(&comp_workspace_wait[i]);
760 }
761}
762
763/*
764 * this finds an available workspace or allocates a new one
765 * ERR_PTR is returned if things go bad.
766 */
767static struct list_head *find_workspace(int type)
768{
769 struct list_head *workspace;
770 int cpus = num_online_cpus();
771 int idx = type - 1;
772
773 struct list_head *idle_workspace = &comp_idle_workspace[idx];
774 spinlock_t *workspace_lock = &comp_workspace_lock[idx];
775 atomic_t *alloc_workspace = &comp_alloc_workspace[idx];
776 wait_queue_head_t *workspace_wait = &comp_workspace_wait[idx];
777 int *num_workspace = &comp_num_workspace[idx];
778again:
779 spin_lock(workspace_lock);
780 if (!list_empty(idle_workspace)) {
781 workspace = idle_workspace->next;
782 list_del(workspace);
783 (*num_workspace)--;
784 spin_unlock(workspace_lock);
785 return workspace;
786
787 }
788 if (atomic_read(alloc_workspace) > cpus) {
789 DEFINE_WAIT(wait);
790
791 spin_unlock(workspace_lock);
792 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
793 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
794 schedule();
795 finish_wait(workspace_wait, &wait);
796 goto again;
797 }
798 atomic_inc(alloc_workspace);
799 spin_unlock(workspace_lock);
800
801 workspace = btrfs_compress_op[idx]->alloc_workspace();
802 if (IS_ERR(workspace)) {
803 atomic_dec(alloc_workspace);
804 wake_up(workspace_wait);
805 }
806 return workspace;
807}
808
809/*
810 * put a workspace struct back on the list or free it if we have enough
811 * idle ones sitting around
812 */
813static void free_workspace(int type, struct list_head *workspace)
814{
815 int idx = type - 1;
816 struct list_head *idle_workspace = &comp_idle_workspace[idx];
817 spinlock_t *workspace_lock = &comp_workspace_lock[idx];
818 atomic_t *alloc_workspace = &comp_alloc_workspace[idx];
819 wait_queue_head_t *workspace_wait = &comp_workspace_wait[idx];
820 int *num_workspace = &comp_num_workspace[idx];
821
822 spin_lock(workspace_lock);
823 if (*num_workspace < num_online_cpus()) {
824 list_add_tail(workspace, idle_workspace);
825 (*num_workspace)++;
826 spin_unlock(workspace_lock);
827 goto wake;
828 }
829 spin_unlock(workspace_lock);
830
831 btrfs_compress_op[idx]->free_workspace(workspace);
832 atomic_dec(alloc_workspace);
833wake:
834 smp_mb();
835 if (waitqueue_active(workspace_wait))
836 wake_up(workspace_wait);
837}
838
839/*
840 * cleanup function for module exit
841 */
842static void free_workspaces(void)
843{
844 struct list_head *workspace;
845 int i;
846
847 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
848 while (!list_empty(&comp_idle_workspace[i])) {
849 workspace = comp_idle_workspace[i].next;
850 list_del(workspace);
851 btrfs_compress_op[i]->free_workspace(workspace);
852 atomic_dec(&comp_alloc_workspace[i]);
853 }
854 }
855}
856
857/*
858 * given an address space and start/len, compress the bytes.
859 *
860 * pages are allocated to hold the compressed result and stored
861 * in 'pages'
862 *
863 * out_pages is used to return the number of pages allocated. There
864 * may be pages allocated even if we return an error
865 *
866 * total_in is used to return the number of bytes actually read. It
867 * may be smaller then len if we had to exit early because we
868 * ran out of room in the pages array or because we cross the
869 * max_out threshold.
870 *
871 * total_out is used to return the total number of compressed bytes
872 *
873 * max_out tells us the max number of bytes that we're allowed to
874 * stuff into pages
875 */
876int btrfs_compress_pages(int type, struct address_space *mapping,
877 u64 start, unsigned long len,
878 struct page **pages,
879 unsigned long nr_dest_pages,
880 unsigned long *out_pages,
881 unsigned long *total_in,
882 unsigned long *total_out,
883 unsigned long max_out)
884{
885 struct list_head *workspace;
886 int ret;
887
888 workspace = find_workspace(type);
889 if (IS_ERR(workspace))
890 return -1;
891
892 ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
893 start, len, pages,
894 nr_dest_pages, out_pages,
895 total_in, total_out,
896 max_out);
897 free_workspace(type, workspace);
898 return ret;
899}
900
901/*
902 * pages_in is an array of pages with compressed data.
903 *
904 * disk_start is the starting logical offset of this array in the file
905 *
906 * bvec is a bio_vec of pages from the file that we want to decompress into
907 *
908 * vcnt is the count of pages in the biovec
909 *
910 * srclen is the number of bytes in pages_in
911 *
912 * The basic idea is that we have a bio that was created by readpages.
913 * The pages in the bio are for the uncompressed data, and they may not
914 * be contiguous. They all correspond to the range of bytes covered by
915 * the compressed extent.
916 */
917static int btrfs_decompress_biovec(int type, struct page **pages_in,
918 u64 disk_start, struct bio_vec *bvec,
919 int vcnt, size_t srclen)
920{
921 struct list_head *workspace;
922 int ret;
923
924 workspace = find_workspace(type);
925 if (IS_ERR(workspace))
926 return -ENOMEM;
927
928 ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
929 disk_start,
930 bvec, vcnt, srclen);
931 free_workspace(type, workspace);
932 return ret;
933}
934
935/*
936 * a less complex decompression routine. Our compressed data fits in a
937 * single page, and we want to read a single page out of it.
938 * start_byte tells us the offset into the compressed data we're interested in
939 */
940int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
941 unsigned long start_byte, size_t srclen, size_t destlen)
942{
943 struct list_head *workspace;
944 int ret;
945
946 workspace = find_workspace(type);
947 if (IS_ERR(workspace))
948 return -ENOMEM;
949
950 ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
951 dest_page, start_byte,
952 srclen, destlen);
953
954 free_workspace(type, workspace);
955 return ret;
956}
957
958void btrfs_exit_compress(void)
959{
960 free_workspaces();
961}
962
963/*
964 * Copy uncompressed data from working buffer to pages.
965 *
966 * buf_start is the byte offset we're of the start of our workspace buffer.
967 *
968 * total_out is the last byte of the buffer
969 */
970int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
971 unsigned long total_out, u64 disk_start,
972 struct bio_vec *bvec, int vcnt,
973 unsigned long *pg_index,
974 unsigned long *pg_offset)
975{
976 unsigned long buf_offset;
977 unsigned long current_buf_start;
978 unsigned long start_byte;
979 unsigned long working_bytes = total_out - buf_start;
980 unsigned long bytes;
981 char *kaddr;
982 struct page *page_out = bvec[*pg_index].bv_page;
983
984 /*
985 * start byte is the first byte of the page we're currently
986 * copying into relative to the start of the compressed data.
987 */
988 start_byte = page_offset(page_out) - disk_start;
989
990 /* we haven't yet hit data corresponding to this page */
991 if (total_out <= start_byte)
992 return 1;
993
994 /*
995 * the start of the data we care about is offset into
996 * the middle of our working buffer
997 */
998 if (total_out > start_byte && buf_start < start_byte) {
999 buf_offset = start_byte - buf_start;
1000 working_bytes -= buf_offset;
1001 } else {
1002 buf_offset = 0;
1003 }
1004 current_buf_start = buf_start;
1005
1006 /* copy bytes from the working buffer into the pages */
1007 while (working_bytes > 0) {
1008 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
1009 PAGE_CACHE_SIZE - buf_offset);
1010 bytes = min(bytes, working_bytes);
1011 kaddr = kmap_atomic(page_out);
1012 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1013 if (*pg_index == (vcnt - 1) && *pg_offset == 0)
1014 memset(kaddr + bytes, 0, PAGE_CACHE_SIZE - bytes);
1015 kunmap_atomic(kaddr);
1016 flush_dcache_page(page_out);
1017
1018 *pg_offset += bytes;
1019 buf_offset += bytes;
1020 working_bytes -= bytes;
1021 current_buf_start += bytes;
1022
1023 /* check if we need to pick another page */
1024 if (*pg_offset == PAGE_CACHE_SIZE) {
1025 (*pg_index)++;
1026 if (*pg_index >= vcnt)
1027 return 0;
1028
1029 page_out = bvec[*pg_index].bv_page;
1030 *pg_offset = 0;
1031 start_byte = page_offset(page_out) - disk_start;
1032
1033 /*
1034 * make sure our new page is covered by this
1035 * working buffer
1036 */
1037 if (total_out <= start_byte)
1038 return 1;
1039
1040 /*
1041 * the next page in the biovec might not be adjacent
1042 * to the last page, but it might still be found
1043 * inside this working buffer. bump our offset pointer
1044 */
1045 if (total_out > start_byte &&
1046 current_buf_start < start_byte) {
1047 buf_offset = start_byte - buf_start;
1048 working_bytes = total_out - start_byte;
1049 current_buf_start = buf_start + buf_offset;
1050 }
1051 }
1052 }
1053
1054 return 1;
1055}