Loading...
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_bio(int type, struct page **pages_in,
85 u64 disk_start, struct bio *orig_bio,
86 size_t srclen);
87
88static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
89 unsigned long disk_size)
90{
91 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
92
93 return sizeof(struct compressed_bio) +
94 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
95}
96
97static struct bio *compressed_bio_alloc(struct block_device *bdev,
98 u64 first_byte, gfp_t gfp_flags)
99{
100 return btrfs_bio_alloc(bdev, first_byte >> 9, BIO_MAX_PAGES, gfp_flags);
101}
102
103static int check_compressed_csum(struct inode *inode,
104 struct compressed_bio *cb,
105 u64 disk_start)
106{
107 int ret;
108 struct page *page;
109 unsigned long i;
110 char *kaddr;
111 u32 csum;
112 u32 *cb_sum = &cb->sums;
113
114 if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
115 return 0;
116
117 for (i = 0; i < cb->nr_pages; i++) {
118 page = cb->compressed_pages[i];
119 csum = ~(u32)0;
120
121 kaddr = kmap_atomic(page);
122 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
123 btrfs_csum_final(csum, (u8 *)&csum);
124 kunmap_atomic(kaddr);
125
126 if (csum != *cb_sum) {
127 btrfs_info(BTRFS_I(inode)->root->fs_info,
128 "csum failed ino %llu extent %llu csum %u wanted %u mirror %d",
129 btrfs_ino(inode), disk_start, csum, *cb_sum,
130 cb->mirror_num);
131 ret = -EIO;
132 goto fail;
133 }
134 cb_sum++;
135
136 }
137 ret = 0;
138fail:
139 return ret;
140}
141
142/* when we finish reading compressed pages from the disk, we
143 * decompress them and then run the bio end_io routines on the
144 * decompressed pages (in the inode address space).
145 *
146 * This allows the checksumming and other IO error handling routines
147 * to work normally
148 *
149 * The compressed pages are freed here, and it must be run
150 * in process context
151 */
152static void end_compressed_bio_read(struct bio *bio)
153{
154 struct compressed_bio *cb = bio->bi_private;
155 struct inode *inode;
156 struct page *page;
157 unsigned long index;
158 int ret;
159
160 if (bio->bi_error)
161 cb->errors = 1;
162
163 /* if there are more bios still pending for this compressed
164 * extent, just exit
165 */
166 if (!atomic_dec_and_test(&cb->pending_bios))
167 goto out;
168
169 inode = cb->inode;
170 ret = check_compressed_csum(inode, cb,
171 (u64)bio->bi_iter.bi_sector << 9);
172 if (ret)
173 goto csum_failed;
174
175 /* ok, we're the last bio for this extent, lets start
176 * the decompression.
177 */
178 ret = btrfs_decompress_bio(cb->compress_type,
179 cb->compressed_pages,
180 cb->start,
181 cb->orig_bio,
182 cb->compressed_len);
183csum_failed:
184 if (ret)
185 cb->errors = 1;
186
187 /* release the compressed pages */
188 index = 0;
189 for (index = 0; index < cb->nr_pages; index++) {
190 page = cb->compressed_pages[index];
191 page->mapping = NULL;
192 put_page(page);
193 }
194
195 /* do io completion on the original bio */
196 if (cb->errors) {
197 bio_io_error(cb->orig_bio);
198 } else {
199 int i;
200 struct bio_vec *bvec;
201
202 /*
203 * we have verified the checksum already, set page
204 * checked so the end_io handlers know about it
205 */
206 bio_for_each_segment_all(bvec, cb->orig_bio, i)
207 SetPageChecked(bvec->bv_page);
208
209 bio_endio(cb->orig_bio);
210 }
211
212 /* finally free the cb struct */
213 kfree(cb->compressed_pages);
214 kfree(cb);
215out:
216 bio_put(bio);
217}
218
219/*
220 * Clear the writeback bits on all of the file
221 * pages for a compressed write
222 */
223static noinline void end_compressed_writeback(struct inode *inode,
224 const struct compressed_bio *cb)
225{
226 unsigned long index = cb->start >> PAGE_SHIFT;
227 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
228 struct page *pages[16];
229 unsigned long nr_pages = end_index - index + 1;
230 int i;
231 int ret;
232
233 if (cb->errors)
234 mapping_set_error(inode->i_mapping, -EIO);
235
236 while (nr_pages > 0) {
237 ret = find_get_pages_contig(inode->i_mapping, index,
238 min_t(unsigned long,
239 nr_pages, ARRAY_SIZE(pages)), pages);
240 if (ret == 0) {
241 nr_pages -= 1;
242 index += 1;
243 continue;
244 }
245 for (i = 0; i < ret; i++) {
246 if (cb->errors)
247 SetPageError(pages[i]);
248 end_page_writeback(pages[i]);
249 put_page(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)
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 (bio->bi_error)
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,
292 bio->bi_error ? 0 : 1);
293 cb->compressed_pages[0]->mapping = NULL;
294
295 end_compressed_writeback(inode, cb);
296 /* note, our inode could be gone now */
297
298 /*
299 * release the compressed pages, these came from alloc_page and
300 * are not attached to the inode at all
301 */
302 index = 0;
303 for (index = 0; index < cb->nr_pages; index++) {
304 page = cb->compressed_pages[index];
305 page->mapping = NULL;
306 put_page(page);
307 }
308
309 /* finally free the cb struct */
310 kfree(cb->compressed_pages);
311 kfree(cb);
312out:
313 bio_put(bio);
314}
315
316/*
317 * worker function to build and submit bios for previously compressed pages.
318 * The corresponding pages in the inode should be marked for writeback
319 * and the compressed pages should have a reference on them for dropping
320 * when the IO is complete.
321 *
322 * This also checksums the file bytes and gets things ready for
323 * the end io hooks.
324 */
325int btrfs_submit_compressed_write(struct inode *inode, u64 start,
326 unsigned long len, u64 disk_start,
327 unsigned long compressed_len,
328 struct page **compressed_pages,
329 unsigned long nr_pages)
330{
331 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
332 struct bio *bio = NULL;
333 struct compressed_bio *cb;
334 unsigned long bytes_left;
335 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
336 int pg_index = 0;
337 struct page *page;
338 u64 first_byte = disk_start;
339 struct block_device *bdev;
340 int ret;
341 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
342
343 WARN_ON(start & ((u64)PAGE_SIZE - 1));
344 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
345 if (!cb)
346 return -ENOMEM;
347 atomic_set(&cb->pending_bios, 0);
348 cb->errors = 0;
349 cb->inode = inode;
350 cb->start = start;
351 cb->len = len;
352 cb->mirror_num = 0;
353 cb->compressed_pages = compressed_pages;
354 cb->compressed_len = compressed_len;
355 cb->orig_bio = NULL;
356 cb->nr_pages = nr_pages;
357
358 bdev = fs_info->fs_devices->latest_bdev;
359
360 bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
361 if (!bio) {
362 kfree(cb);
363 return -ENOMEM;
364 }
365 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
366 bio->bi_private = cb;
367 bio->bi_end_io = end_compressed_bio_write;
368 atomic_inc(&cb->pending_bios);
369
370 /* create and submit bios for the compressed pages */
371 bytes_left = compressed_len;
372 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
373 page = compressed_pages[pg_index];
374 page->mapping = inode->i_mapping;
375 if (bio->bi_iter.bi_size)
376 ret = io_tree->ops->merge_bio_hook(page, 0,
377 PAGE_SIZE,
378 bio, 0);
379 else
380 ret = 0;
381
382 page->mapping = NULL;
383 if (ret || bio_add_page(bio, page, PAGE_SIZE, 0) <
384 PAGE_SIZE) {
385 bio_get(bio);
386
387 /*
388 * inc the count before we submit the bio so
389 * we know the end IO handler won't happen before
390 * we inc the count. Otherwise, the cb might get
391 * freed before we're done setting it up
392 */
393 atomic_inc(&cb->pending_bios);
394 ret = btrfs_bio_wq_end_io(fs_info, bio,
395 BTRFS_WQ_ENDIO_DATA);
396 BUG_ON(ret); /* -ENOMEM */
397
398 if (!skip_sum) {
399 ret = btrfs_csum_one_bio(inode, bio, start, 1);
400 BUG_ON(ret); /* -ENOMEM */
401 }
402
403 ret = btrfs_map_bio(fs_info, bio, 0, 1);
404 if (ret) {
405 bio->bi_error = ret;
406 bio_endio(bio);
407 }
408
409 bio_put(bio);
410
411 bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
412 BUG_ON(!bio);
413 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
414 bio->bi_private = cb;
415 bio->bi_end_io = end_compressed_bio_write;
416 bio_add_page(bio, page, PAGE_SIZE, 0);
417 }
418 if (bytes_left < PAGE_SIZE) {
419 btrfs_info(fs_info,
420 "bytes left %lu compress len %lu nr %lu",
421 bytes_left, cb->compressed_len, cb->nr_pages);
422 }
423 bytes_left -= PAGE_SIZE;
424 first_byte += PAGE_SIZE;
425 cond_resched();
426 }
427 bio_get(bio);
428
429 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
430 BUG_ON(ret); /* -ENOMEM */
431
432 if (!skip_sum) {
433 ret = btrfs_csum_one_bio(inode, bio, start, 1);
434 BUG_ON(ret); /* -ENOMEM */
435 }
436
437 ret = btrfs_map_bio(fs_info, bio, 0, 1);
438 if (ret) {
439 bio->bi_error = ret;
440 bio_endio(bio);
441 }
442
443 bio_put(bio);
444 return 0;
445}
446
447static u64 bio_end_offset(struct bio *bio)
448{
449 struct bio_vec *last = &bio->bi_io_vec[bio->bi_vcnt - 1];
450
451 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
452}
453
454static noinline int add_ra_bio_pages(struct inode *inode,
455 u64 compressed_end,
456 struct compressed_bio *cb)
457{
458 unsigned long end_index;
459 unsigned long pg_index;
460 u64 last_offset;
461 u64 isize = i_size_read(inode);
462 int ret;
463 struct page *page;
464 unsigned long nr_pages = 0;
465 struct extent_map *em;
466 struct address_space *mapping = inode->i_mapping;
467 struct extent_map_tree *em_tree;
468 struct extent_io_tree *tree;
469 u64 end;
470 int misses = 0;
471
472 last_offset = bio_end_offset(cb->orig_bio);
473 em_tree = &BTRFS_I(inode)->extent_tree;
474 tree = &BTRFS_I(inode)->io_tree;
475
476 if (isize == 0)
477 return 0;
478
479 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
480
481 while (last_offset < compressed_end) {
482 pg_index = last_offset >> PAGE_SHIFT;
483
484 if (pg_index > end_index)
485 break;
486
487 rcu_read_lock();
488 page = radix_tree_lookup(&mapping->page_tree, pg_index);
489 rcu_read_unlock();
490 if (page && !radix_tree_exceptional_entry(page)) {
491 misses++;
492 if (misses > 4)
493 break;
494 goto next;
495 }
496
497 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
498 ~__GFP_FS));
499 if (!page)
500 break;
501
502 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
503 put_page(page);
504 goto next;
505 }
506
507 end = last_offset + PAGE_SIZE - 1;
508 /*
509 * at this point, we have a locked page in the page cache
510 * for these bytes in the file. But, we have to make
511 * sure they map to this compressed extent on disk.
512 */
513 set_page_extent_mapped(page);
514 lock_extent(tree, last_offset, end);
515 read_lock(&em_tree->lock);
516 em = lookup_extent_mapping(em_tree, last_offset,
517 PAGE_SIZE);
518 read_unlock(&em_tree->lock);
519
520 if (!em || last_offset < em->start ||
521 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
522 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
523 free_extent_map(em);
524 unlock_extent(tree, last_offset, end);
525 unlock_page(page);
526 put_page(page);
527 break;
528 }
529 free_extent_map(em);
530
531 if (page->index == end_index) {
532 char *userpage;
533 size_t zero_offset = isize & (PAGE_SIZE - 1);
534
535 if (zero_offset) {
536 int zeros;
537 zeros = PAGE_SIZE - zero_offset;
538 userpage = kmap_atomic(page);
539 memset(userpage + zero_offset, 0, zeros);
540 flush_dcache_page(page);
541 kunmap_atomic(userpage);
542 }
543 }
544
545 ret = bio_add_page(cb->orig_bio, page,
546 PAGE_SIZE, 0);
547
548 if (ret == PAGE_SIZE) {
549 nr_pages++;
550 put_page(page);
551 } else {
552 unlock_extent(tree, last_offset, end);
553 unlock_page(page);
554 put_page(page);
555 break;
556 }
557next:
558 last_offset += PAGE_SIZE;
559 }
560 return 0;
561}
562
563/*
564 * for a compressed read, the bio we get passed has all the inode pages
565 * in it. We don't actually do IO on those pages but allocate new ones
566 * to hold the compressed pages on disk.
567 *
568 * bio->bi_iter.bi_sector points to the compressed extent on disk
569 * bio->bi_io_vec points to all of the inode pages
570 *
571 * After the compressed pages are read, we copy the bytes into the
572 * bio we were passed and then call the bio end_io calls
573 */
574int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
575 int mirror_num, unsigned long bio_flags)
576{
577 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
578 struct extent_io_tree *tree;
579 struct extent_map_tree *em_tree;
580 struct compressed_bio *cb;
581 unsigned long compressed_len;
582 unsigned long nr_pages;
583 unsigned long pg_index;
584 struct page *page;
585 struct block_device *bdev;
586 struct bio *comp_bio;
587 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
588 u64 em_len;
589 u64 em_start;
590 struct extent_map *em;
591 int ret = -ENOMEM;
592 int faili = 0;
593 u32 *sums;
594
595 tree = &BTRFS_I(inode)->io_tree;
596 em_tree = &BTRFS_I(inode)->extent_tree;
597
598 /* we need the actual starting offset of this extent in the file */
599 read_lock(&em_tree->lock);
600 em = lookup_extent_mapping(em_tree,
601 page_offset(bio->bi_io_vec->bv_page),
602 PAGE_SIZE);
603 read_unlock(&em_tree->lock);
604 if (!em)
605 return -EIO;
606
607 compressed_len = em->block_len;
608 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
609 if (!cb)
610 goto out;
611
612 atomic_set(&cb->pending_bios, 0);
613 cb->errors = 0;
614 cb->inode = inode;
615 cb->mirror_num = mirror_num;
616 sums = &cb->sums;
617
618 cb->start = em->orig_start;
619 em_len = em->len;
620 em_start = em->start;
621
622 free_extent_map(em);
623 em = NULL;
624
625 cb->len = bio->bi_iter.bi_size;
626 cb->compressed_len = compressed_len;
627 cb->compress_type = extent_compress_type(bio_flags);
628 cb->orig_bio = bio;
629
630 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
631 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
632 GFP_NOFS);
633 if (!cb->compressed_pages)
634 goto fail1;
635
636 bdev = fs_info->fs_devices->latest_bdev;
637
638 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
639 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
640 __GFP_HIGHMEM);
641 if (!cb->compressed_pages[pg_index]) {
642 faili = pg_index - 1;
643 ret = -ENOMEM;
644 goto fail2;
645 }
646 }
647 faili = nr_pages - 1;
648 cb->nr_pages = nr_pages;
649
650 add_ra_bio_pages(inode, em_start + em_len, cb);
651
652 /* include any pages we added in add_ra-bio_pages */
653 cb->len = bio->bi_iter.bi_size;
654
655 comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
656 if (!comp_bio)
657 goto fail2;
658 bio_set_op_attrs (comp_bio, REQ_OP_READ, 0);
659 comp_bio->bi_private = cb;
660 comp_bio->bi_end_io = end_compressed_bio_read;
661 atomic_inc(&cb->pending_bios);
662
663 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
664 page = cb->compressed_pages[pg_index];
665 page->mapping = inode->i_mapping;
666 page->index = em_start >> PAGE_SHIFT;
667
668 if (comp_bio->bi_iter.bi_size)
669 ret = tree->ops->merge_bio_hook(page, 0,
670 PAGE_SIZE,
671 comp_bio, 0);
672 else
673 ret = 0;
674
675 page->mapping = NULL;
676 if (ret || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
677 PAGE_SIZE) {
678 bio_get(comp_bio);
679
680 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
681 BTRFS_WQ_ENDIO_DATA);
682 BUG_ON(ret); /* -ENOMEM */
683
684 /*
685 * inc the count before we submit the bio so
686 * we know the end IO handler won't happen before
687 * we inc the count. Otherwise, the cb might get
688 * freed before we're done setting it up
689 */
690 atomic_inc(&cb->pending_bios);
691
692 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
693 ret = btrfs_lookup_bio_sums(inode, comp_bio,
694 sums);
695 BUG_ON(ret); /* -ENOMEM */
696 }
697 sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
698 fs_info->sectorsize);
699
700 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
701 if (ret) {
702 comp_bio->bi_error = ret;
703 bio_endio(comp_bio);
704 }
705
706 bio_put(comp_bio);
707
708 comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
709 GFP_NOFS);
710 BUG_ON(!comp_bio);
711 bio_set_op_attrs(comp_bio, REQ_OP_READ, 0);
712 comp_bio->bi_private = cb;
713 comp_bio->bi_end_io = end_compressed_bio_read;
714
715 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
716 }
717 cur_disk_byte += PAGE_SIZE;
718 }
719 bio_get(comp_bio);
720
721 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
722 BUG_ON(ret); /* -ENOMEM */
723
724 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
725 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
726 BUG_ON(ret); /* -ENOMEM */
727 }
728
729 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
730 if (ret) {
731 comp_bio->bi_error = ret;
732 bio_endio(comp_bio);
733 }
734
735 bio_put(comp_bio);
736 return 0;
737
738fail2:
739 while (faili >= 0) {
740 __free_page(cb->compressed_pages[faili]);
741 faili--;
742 }
743
744 kfree(cb->compressed_pages);
745fail1:
746 kfree(cb);
747out:
748 free_extent_map(em);
749 return ret;
750}
751
752static struct {
753 struct list_head idle_ws;
754 spinlock_t ws_lock;
755 /* Number of free workspaces */
756 int free_ws;
757 /* Total number of allocated workspaces */
758 atomic_t total_ws;
759 /* Waiters for a free workspace */
760 wait_queue_head_t ws_wait;
761} btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
762
763static const struct btrfs_compress_op * const btrfs_compress_op[] = {
764 &btrfs_zlib_compress,
765 &btrfs_lzo_compress,
766};
767
768void __init btrfs_init_compress(void)
769{
770 int i;
771
772 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
773 struct list_head *workspace;
774
775 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
776 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
777 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
778 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
779
780 /*
781 * Preallocate one workspace for each compression type so
782 * we can guarantee forward progress in the worst case
783 */
784 workspace = btrfs_compress_op[i]->alloc_workspace();
785 if (IS_ERR(workspace)) {
786 pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n");
787 } else {
788 atomic_set(&btrfs_comp_ws[i].total_ws, 1);
789 btrfs_comp_ws[i].free_ws = 1;
790 list_add(workspace, &btrfs_comp_ws[i].idle_ws);
791 }
792 }
793}
794
795/*
796 * This finds an available workspace or allocates a new one.
797 * If it's not possible to allocate a new one, waits until there's one.
798 * Preallocation makes a forward progress guarantees and we do not return
799 * errors.
800 */
801static struct list_head *find_workspace(int type)
802{
803 struct list_head *workspace;
804 int cpus = num_online_cpus();
805 int idx = type - 1;
806
807 struct list_head *idle_ws = &btrfs_comp_ws[idx].idle_ws;
808 spinlock_t *ws_lock = &btrfs_comp_ws[idx].ws_lock;
809 atomic_t *total_ws = &btrfs_comp_ws[idx].total_ws;
810 wait_queue_head_t *ws_wait = &btrfs_comp_ws[idx].ws_wait;
811 int *free_ws = &btrfs_comp_ws[idx].free_ws;
812again:
813 spin_lock(ws_lock);
814 if (!list_empty(idle_ws)) {
815 workspace = idle_ws->next;
816 list_del(workspace);
817 (*free_ws)--;
818 spin_unlock(ws_lock);
819 return workspace;
820
821 }
822 if (atomic_read(total_ws) > cpus) {
823 DEFINE_WAIT(wait);
824
825 spin_unlock(ws_lock);
826 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
827 if (atomic_read(total_ws) > cpus && !*free_ws)
828 schedule();
829 finish_wait(ws_wait, &wait);
830 goto again;
831 }
832 atomic_inc(total_ws);
833 spin_unlock(ws_lock);
834
835 workspace = btrfs_compress_op[idx]->alloc_workspace();
836 if (IS_ERR(workspace)) {
837 atomic_dec(total_ws);
838 wake_up(ws_wait);
839
840 /*
841 * Do not return the error but go back to waiting. There's a
842 * workspace preallocated for each type and the compression
843 * time is bounded so we get to a workspace eventually. This
844 * makes our caller's life easier.
845 *
846 * To prevent silent and low-probability deadlocks (when the
847 * initial preallocation fails), check if there are any
848 * workspaces at all.
849 */
850 if (atomic_read(total_ws) == 0) {
851 static DEFINE_RATELIMIT_STATE(_rs,
852 /* once per minute */ 60 * HZ,
853 /* no burst */ 1);
854
855 if (__ratelimit(&_rs)) {
856 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
857 }
858 }
859 goto again;
860 }
861 return workspace;
862}
863
864/*
865 * put a workspace struct back on the list or free it if we have enough
866 * idle ones sitting around
867 */
868static void free_workspace(int type, struct list_head *workspace)
869{
870 int idx = type - 1;
871 struct list_head *idle_ws = &btrfs_comp_ws[idx].idle_ws;
872 spinlock_t *ws_lock = &btrfs_comp_ws[idx].ws_lock;
873 atomic_t *total_ws = &btrfs_comp_ws[idx].total_ws;
874 wait_queue_head_t *ws_wait = &btrfs_comp_ws[idx].ws_wait;
875 int *free_ws = &btrfs_comp_ws[idx].free_ws;
876
877 spin_lock(ws_lock);
878 if (*free_ws < num_online_cpus()) {
879 list_add(workspace, idle_ws);
880 (*free_ws)++;
881 spin_unlock(ws_lock);
882 goto wake;
883 }
884 spin_unlock(ws_lock);
885
886 btrfs_compress_op[idx]->free_workspace(workspace);
887 atomic_dec(total_ws);
888wake:
889 /*
890 * Make sure counter is updated before we wake up waiters.
891 */
892 smp_mb();
893 if (waitqueue_active(ws_wait))
894 wake_up(ws_wait);
895}
896
897/*
898 * cleanup function for module exit
899 */
900static void free_workspaces(void)
901{
902 struct list_head *workspace;
903 int i;
904
905 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
906 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
907 workspace = btrfs_comp_ws[i].idle_ws.next;
908 list_del(workspace);
909 btrfs_compress_op[i]->free_workspace(workspace);
910 atomic_dec(&btrfs_comp_ws[i].total_ws);
911 }
912 }
913}
914
915/*
916 * given an address space and start/len, compress the bytes.
917 *
918 * pages are allocated to hold the compressed result and stored
919 * in 'pages'
920 *
921 * out_pages is used to return the number of pages allocated. There
922 * may be pages allocated even if we return an error
923 *
924 * total_in is used to return the number of bytes actually read. It
925 * may be smaller then len if we had to exit early because we
926 * ran out of room in the pages array or because we cross the
927 * max_out threshold.
928 *
929 * total_out is used to return the total number of compressed bytes
930 *
931 * max_out tells us the max number of bytes that we're allowed to
932 * stuff into pages
933 */
934int btrfs_compress_pages(int type, struct address_space *mapping,
935 u64 start, unsigned long len,
936 struct page **pages,
937 unsigned long nr_dest_pages,
938 unsigned long *out_pages,
939 unsigned long *total_in,
940 unsigned long *total_out,
941 unsigned long max_out)
942{
943 struct list_head *workspace;
944 int ret;
945
946 workspace = find_workspace(type);
947
948 ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
949 start, len, pages,
950 nr_dest_pages, out_pages,
951 total_in, total_out,
952 max_out);
953 free_workspace(type, workspace);
954 return ret;
955}
956
957/*
958 * pages_in is an array of pages with compressed data.
959 *
960 * disk_start is the starting logical offset of this array in the file
961 *
962 * orig_bio contains the pages from the file that we want to decompress into
963 *
964 * srclen is the number of bytes in pages_in
965 *
966 * The basic idea is that we have a bio that was created by readpages.
967 * The pages in the bio are for the uncompressed data, and they may not
968 * be contiguous. They all correspond to the range of bytes covered by
969 * the compressed extent.
970 */
971static int btrfs_decompress_bio(int type, struct page **pages_in,
972 u64 disk_start, struct bio *orig_bio,
973 size_t srclen)
974{
975 struct list_head *workspace;
976 int ret;
977
978 workspace = find_workspace(type);
979
980 ret = btrfs_compress_op[type-1]->decompress_bio(workspace, pages_in,
981 disk_start, orig_bio,
982 srclen);
983 free_workspace(type, workspace);
984 return ret;
985}
986
987/*
988 * a less complex decompression routine. Our compressed data fits in a
989 * single page, and we want to read a single page out of it.
990 * start_byte tells us the offset into the compressed data we're interested in
991 */
992int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
993 unsigned long start_byte, size_t srclen, size_t destlen)
994{
995 struct list_head *workspace;
996 int ret;
997
998 workspace = find_workspace(type);
999
1000 ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
1001 dest_page, start_byte,
1002 srclen, destlen);
1003
1004 free_workspace(type, workspace);
1005 return ret;
1006}
1007
1008void btrfs_exit_compress(void)
1009{
1010 free_workspaces();
1011}
1012
1013/*
1014 * Copy uncompressed data from working buffer to pages.
1015 *
1016 * buf_start is the byte offset we're of the start of our workspace buffer.
1017 *
1018 * total_out is the last byte of the buffer
1019 */
1020int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
1021 unsigned long total_out, u64 disk_start,
1022 struct bio *bio)
1023{
1024 unsigned long buf_offset;
1025 unsigned long current_buf_start;
1026 unsigned long start_byte;
1027 unsigned long prev_start_byte;
1028 unsigned long working_bytes = total_out - buf_start;
1029 unsigned long bytes;
1030 char *kaddr;
1031 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1032
1033 /*
1034 * start byte is the first byte of the page we're currently
1035 * copying into relative to the start of the compressed data.
1036 */
1037 start_byte = page_offset(bvec.bv_page) - disk_start;
1038
1039 /* we haven't yet hit data corresponding to this page */
1040 if (total_out <= start_byte)
1041 return 1;
1042
1043 /*
1044 * the start of the data we care about is offset into
1045 * the middle of our working buffer
1046 */
1047 if (total_out > start_byte && buf_start < start_byte) {
1048 buf_offset = start_byte - buf_start;
1049 working_bytes -= buf_offset;
1050 } else {
1051 buf_offset = 0;
1052 }
1053 current_buf_start = buf_start;
1054
1055 /* copy bytes from the working buffer into the pages */
1056 while (working_bytes > 0) {
1057 bytes = min_t(unsigned long, bvec.bv_len,
1058 PAGE_SIZE - buf_offset);
1059 bytes = min(bytes, working_bytes);
1060
1061 kaddr = kmap_atomic(bvec.bv_page);
1062 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1063 kunmap_atomic(kaddr);
1064 flush_dcache_page(bvec.bv_page);
1065
1066 buf_offset += bytes;
1067 working_bytes -= bytes;
1068 current_buf_start += bytes;
1069
1070 /* check if we need to pick another page */
1071 bio_advance(bio, bytes);
1072 if (!bio->bi_iter.bi_size)
1073 return 0;
1074 bvec = bio_iter_iovec(bio, bio->bi_iter);
1075 prev_start_byte = start_byte;
1076 start_byte = page_offset(bvec.bv_page) - disk_start;
1077
1078 /*
1079 * We need to make sure we're only adjusting
1080 * our offset into compression working buffer when
1081 * we're switching pages. Otherwise we can incorrectly
1082 * keep copying when we were actually done.
1083 */
1084 if (start_byte != prev_start_byte) {
1085 /*
1086 * make sure our new page is covered by this
1087 * working buffer
1088 */
1089 if (total_out <= start_byte)
1090 return 1;
1091
1092 /*
1093 * the next page in the biovec might not be adjacent
1094 * to the last page, but it might still be found
1095 * inside this working buffer. bump our offset pointer
1096 */
1097 if (total_out > start_byte &&
1098 current_buf_start < start_byte) {
1099 buf_offset = start_byte - buf_start;
1100 working_bytes = total_out - start_byte;
1101 current_buf_start = buf_start + buf_offset;
1102 }
1103 }
1104 }
1105
1106 return 1;
1107}
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/highmem.h>
12#include <linux/time.h>
13#include <linux/init.h>
14#include <linux/string.h>
15#include <linux/backing-dev.h>
16#include <linux/writeback.h>
17#include <linux/slab.h>
18#include <linux/sched/mm.h>
19#include <linux/log2.h>
20#include <crypto/hash.h>
21#include "misc.h"
22#include "ctree.h"
23#include "disk-io.h"
24#include "transaction.h"
25#include "btrfs_inode.h"
26#include "volumes.h"
27#include "ordered-data.h"
28#include "compression.h"
29#include "extent_io.h"
30#include "extent_map.h"
31#include "zoned.h"
32
33static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
34
35const char* btrfs_compress_type2str(enum btrfs_compression_type type)
36{
37 switch (type) {
38 case BTRFS_COMPRESS_ZLIB:
39 case BTRFS_COMPRESS_LZO:
40 case BTRFS_COMPRESS_ZSTD:
41 case BTRFS_COMPRESS_NONE:
42 return btrfs_compress_types[type];
43 default:
44 break;
45 }
46
47 return NULL;
48}
49
50bool btrfs_compress_is_valid_type(const char *str, size_t len)
51{
52 int i;
53
54 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
55 size_t comp_len = strlen(btrfs_compress_types[i]);
56
57 if (len < comp_len)
58 continue;
59
60 if (!strncmp(btrfs_compress_types[i], str, comp_len))
61 return true;
62 }
63 return false;
64}
65
66static int compression_compress_pages(int type, struct list_head *ws,
67 struct address_space *mapping, u64 start, struct page **pages,
68 unsigned long *out_pages, unsigned long *total_in,
69 unsigned long *total_out)
70{
71 switch (type) {
72 case BTRFS_COMPRESS_ZLIB:
73 return zlib_compress_pages(ws, mapping, start, pages,
74 out_pages, total_in, total_out);
75 case BTRFS_COMPRESS_LZO:
76 return lzo_compress_pages(ws, mapping, start, pages,
77 out_pages, total_in, total_out);
78 case BTRFS_COMPRESS_ZSTD:
79 return zstd_compress_pages(ws, mapping, start, pages,
80 out_pages, total_in, total_out);
81 case BTRFS_COMPRESS_NONE:
82 default:
83 /*
84 * This can happen when compression races with remount setting
85 * it to 'no compress', while caller doesn't call
86 * inode_need_compress() to check if we really need to
87 * compress.
88 *
89 * Not a big deal, just need to inform caller that we
90 * haven't allocated any pages yet.
91 */
92 *out_pages = 0;
93 return -E2BIG;
94 }
95}
96
97static int compression_decompress_bio(int type, struct list_head *ws,
98 struct compressed_bio *cb)
99{
100 switch (type) {
101 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
102 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
103 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
104 case BTRFS_COMPRESS_NONE:
105 default:
106 /*
107 * This can't happen, the type is validated several times
108 * before we get here.
109 */
110 BUG();
111 }
112}
113
114static int compression_decompress(int type, struct list_head *ws,
115 unsigned char *data_in, struct page *dest_page,
116 unsigned long start_byte, size_t srclen, size_t destlen)
117{
118 switch (type) {
119 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
120 start_byte, srclen, destlen);
121 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
122 start_byte, srclen, destlen);
123 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
124 start_byte, srclen, destlen);
125 case BTRFS_COMPRESS_NONE:
126 default:
127 /*
128 * This can't happen, the type is validated several times
129 * before we get here.
130 */
131 BUG();
132 }
133}
134
135static int btrfs_decompress_bio(struct compressed_bio *cb);
136
137static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
138 unsigned long disk_size)
139{
140 return sizeof(struct compressed_bio) +
141 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size;
142}
143
144static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
145 u64 disk_start)
146{
147 struct btrfs_fs_info *fs_info = inode->root->fs_info;
148 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
149 const u32 csum_size = fs_info->csum_size;
150 const u32 sectorsize = fs_info->sectorsize;
151 struct page *page;
152 unsigned int i;
153 char *kaddr;
154 u8 csum[BTRFS_CSUM_SIZE];
155 struct compressed_bio *cb = bio->bi_private;
156 u8 *cb_sum = cb->sums;
157
158 if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM))
159 return 0;
160
161 shash->tfm = fs_info->csum_shash;
162
163 for (i = 0; i < cb->nr_pages; i++) {
164 u32 pg_offset;
165 u32 bytes_left = PAGE_SIZE;
166 page = cb->compressed_pages[i];
167
168 /* Determine the remaining bytes inside the page first */
169 if (i == cb->nr_pages - 1)
170 bytes_left = cb->compressed_len - i * PAGE_SIZE;
171
172 /* Hash through the page sector by sector */
173 for (pg_offset = 0; pg_offset < bytes_left;
174 pg_offset += sectorsize) {
175 kaddr = kmap_atomic(page);
176 crypto_shash_digest(shash, kaddr + pg_offset,
177 sectorsize, csum);
178 kunmap_atomic(kaddr);
179
180 if (memcmp(&csum, cb_sum, csum_size) != 0) {
181 btrfs_print_data_csum_error(inode, disk_start,
182 csum, cb_sum, cb->mirror_num);
183 if (btrfs_io_bio(bio)->device)
184 btrfs_dev_stat_inc_and_print(
185 btrfs_io_bio(bio)->device,
186 BTRFS_DEV_STAT_CORRUPTION_ERRS);
187 return -EIO;
188 }
189 cb_sum += csum_size;
190 disk_start += sectorsize;
191 }
192 }
193 return 0;
194}
195
196/* when we finish reading compressed pages from the disk, we
197 * decompress them and then run the bio end_io routines on the
198 * decompressed pages (in the inode address space).
199 *
200 * This allows the checksumming and other IO error handling routines
201 * to work normally
202 *
203 * The compressed pages are freed here, and it must be run
204 * in process context
205 */
206static void end_compressed_bio_read(struct bio *bio)
207{
208 struct compressed_bio *cb = bio->bi_private;
209 struct inode *inode;
210 struct page *page;
211 unsigned int index;
212 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
213 int ret = 0;
214
215 if (bio->bi_status)
216 cb->errors = 1;
217
218 /* if there are more bios still pending for this compressed
219 * extent, just exit
220 */
221 if (!refcount_dec_and_test(&cb->pending_bios))
222 goto out;
223
224 /*
225 * Record the correct mirror_num in cb->orig_bio so that
226 * read-repair can work properly.
227 */
228 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
229 cb->mirror_num = mirror;
230
231 /*
232 * Some IO in this cb have failed, just skip checksum as there
233 * is no way it could be correct.
234 */
235 if (cb->errors == 1)
236 goto csum_failed;
237
238 inode = cb->inode;
239 ret = check_compressed_csum(BTRFS_I(inode), bio,
240 bio->bi_iter.bi_sector << 9);
241 if (ret)
242 goto csum_failed;
243
244 /* ok, we're the last bio for this extent, lets start
245 * the decompression.
246 */
247 ret = btrfs_decompress_bio(cb);
248
249csum_failed:
250 if (ret)
251 cb->errors = 1;
252
253 /* release the compressed pages */
254 index = 0;
255 for (index = 0; index < cb->nr_pages; index++) {
256 page = cb->compressed_pages[index];
257 page->mapping = NULL;
258 put_page(page);
259 }
260
261 /* do io completion on the original bio */
262 if (cb->errors) {
263 bio_io_error(cb->orig_bio);
264 } else {
265 struct bio_vec *bvec;
266 struct bvec_iter_all iter_all;
267
268 /*
269 * we have verified the checksum already, set page
270 * checked so the end_io handlers know about it
271 */
272 ASSERT(!bio_flagged(bio, BIO_CLONED));
273 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
274 SetPageChecked(bvec->bv_page);
275
276 bio_endio(cb->orig_bio);
277 }
278
279 /* finally free the cb struct */
280 kfree(cb->compressed_pages);
281 kfree(cb);
282out:
283 bio_put(bio);
284}
285
286/*
287 * Clear the writeback bits on all of the file
288 * pages for a compressed write
289 */
290static noinline void end_compressed_writeback(struct inode *inode,
291 const struct compressed_bio *cb)
292{
293 unsigned long index = cb->start >> PAGE_SHIFT;
294 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
295 struct page *pages[16];
296 unsigned long nr_pages = end_index - index + 1;
297 int i;
298 int ret;
299
300 if (cb->errors)
301 mapping_set_error(inode->i_mapping, -EIO);
302
303 while (nr_pages > 0) {
304 ret = find_get_pages_contig(inode->i_mapping, index,
305 min_t(unsigned long,
306 nr_pages, ARRAY_SIZE(pages)), pages);
307 if (ret == 0) {
308 nr_pages -= 1;
309 index += 1;
310 continue;
311 }
312 for (i = 0; i < ret; i++) {
313 if (cb->errors)
314 SetPageError(pages[i]);
315 end_page_writeback(pages[i]);
316 put_page(pages[i]);
317 }
318 nr_pages -= ret;
319 index += ret;
320 }
321 /* the inode may be gone now */
322}
323
324/*
325 * do the cleanup once all the compressed pages hit the disk.
326 * This will clear writeback on the file pages and free the compressed
327 * pages.
328 *
329 * This also calls the writeback end hooks for the file pages so that
330 * metadata and checksums can be updated in the file.
331 */
332static void end_compressed_bio_write(struct bio *bio)
333{
334 struct compressed_bio *cb = bio->bi_private;
335 struct inode *inode;
336 struct page *page;
337 unsigned int index;
338
339 if (bio->bi_status)
340 cb->errors = 1;
341
342 /* if there are more bios still pending for this compressed
343 * extent, just exit
344 */
345 if (!refcount_dec_and_test(&cb->pending_bios))
346 goto out;
347
348 /* ok, we're the last bio for this extent, step one is to
349 * call back into the FS and do all the end_io operations
350 */
351 inode = cb->inode;
352 btrfs_record_physical_zoned(inode, cb->start, bio);
353 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
354 cb->start, cb->start + cb->len - 1,
355 !cb->errors);
356
357 end_compressed_writeback(inode, cb);
358 /* note, our inode could be gone now */
359
360 /*
361 * release the compressed pages, these came from alloc_page and
362 * are not attached to the inode at all
363 */
364 index = 0;
365 for (index = 0; index < cb->nr_pages; index++) {
366 page = cb->compressed_pages[index];
367 page->mapping = NULL;
368 put_page(page);
369 }
370
371 /* finally free the cb struct */
372 kfree(cb->compressed_pages);
373 kfree(cb);
374out:
375 bio_put(bio);
376}
377
378/*
379 * worker function to build and submit bios for previously compressed pages.
380 * The corresponding pages in the inode should be marked for writeback
381 * and the compressed pages should have a reference on them for dropping
382 * when the IO is complete.
383 *
384 * This also checksums the file bytes and gets things ready for
385 * the end io hooks.
386 */
387blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
388 unsigned int len, u64 disk_start,
389 unsigned int compressed_len,
390 struct page **compressed_pages,
391 unsigned int nr_pages,
392 unsigned int write_flags,
393 struct cgroup_subsys_state *blkcg_css)
394{
395 struct btrfs_fs_info *fs_info = inode->root->fs_info;
396 struct bio *bio = NULL;
397 struct compressed_bio *cb;
398 unsigned long bytes_left;
399 int pg_index = 0;
400 struct page *page;
401 u64 first_byte = disk_start;
402 blk_status_t ret;
403 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
404 const bool use_append = btrfs_use_zone_append(inode, disk_start);
405 const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
406
407 WARN_ON(!PAGE_ALIGNED(start));
408 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
409 if (!cb)
410 return BLK_STS_RESOURCE;
411 refcount_set(&cb->pending_bios, 0);
412 cb->errors = 0;
413 cb->inode = &inode->vfs_inode;
414 cb->start = start;
415 cb->len = len;
416 cb->mirror_num = 0;
417 cb->compressed_pages = compressed_pages;
418 cb->compressed_len = compressed_len;
419 cb->orig_bio = NULL;
420 cb->nr_pages = nr_pages;
421
422 bio = btrfs_bio_alloc(first_byte);
423 bio->bi_opf = bio_op | write_flags;
424 bio->bi_private = cb;
425 bio->bi_end_io = end_compressed_bio_write;
426
427 if (use_append) {
428 struct btrfs_device *device;
429
430 device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE);
431 if (IS_ERR(device)) {
432 kfree(cb);
433 bio_put(bio);
434 return BLK_STS_NOTSUPP;
435 }
436
437 bio_set_dev(bio, device->bdev);
438 }
439
440 if (blkcg_css) {
441 bio->bi_opf |= REQ_CGROUP_PUNT;
442 kthread_associate_blkcg(blkcg_css);
443 }
444 refcount_set(&cb->pending_bios, 1);
445
446 /* create and submit bios for the compressed pages */
447 bytes_left = compressed_len;
448 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
449 int submit = 0;
450 int len = 0;
451
452 page = compressed_pages[pg_index];
453 page->mapping = inode->vfs_inode.i_mapping;
454 if (bio->bi_iter.bi_size)
455 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
456 0);
457
458 /*
459 * Page can only be added to bio if the current bio fits in
460 * stripe.
461 */
462 if (!submit) {
463 if (pg_index == 0 && use_append)
464 len = bio_add_zone_append_page(bio, page,
465 PAGE_SIZE, 0);
466 else
467 len = bio_add_page(bio, page, PAGE_SIZE, 0);
468 }
469
470 page->mapping = NULL;
471 if (submit || len < PAGE_SIZE) {
472 /*
473 * inc the count before we submit the bio so
474 * we know the end IO handler won't happen before
475 * we inc the count. Otherwise, the cb might get
476 * freed before we're done setting it up
477 */
478 refcount_inc(&cb->pending_bios);
479 ret = btrfs_bio_wq_end_io(fs_info, bio,
480 BTRFS_WQ_ENDIO_DATA);
481 BUG_ON(ret); /* -ENOMEM */
482
483 if (!skip_sum) {
484 ret = btrfs_csum_one_bio(inode, bio, start, 1);
485 BUG_ON(ret); /* -ENOMEM */
486 }
487
488 ret = btrfs_map_bio(fs_info, bio, 0);
489 if (ret) {
490 bio->bi_status = ret;
491 bio_endio(bio);
492 }
493
494 bio = btrfs_bio_alloc(first_byte);
495 bio->bi_opf = bio_op | write_flags;
496 bio->bi_private = cb;
497 bio->bi_end_io = end_compressed_bio_write;
498 if (blkcg_css)
499 bio->bi_opf |= REQ_CGROUP_PUNT;
500 /*
501 * Use bio_add_page() to ensure the bio has at least one
502 * page.
503 */
504 bio_add_page(bio, page, PAGE_SIZE, 0);
505 }
506 if (bytes_left < PAGE_SIZE) {
507 btrfs_info(fs_info,
508 "bytes left %lu compress len %u nr %u",
509 bytes_left, cb->compressed_len, cb->nr_pages);
510 }
511 bytes_left -= PAGE_SIZE;
512 first_byte += PAGE_SIZE;
513 cond_resched();
514 }
515
516 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
517 BUG_ON(ret); /* -ENOMEM */
518
519 if (!skip_sum) {
520 ret = btrfs_csum_one_bio(inode, bio, start, 1);
521 BUG_ON(ret); /* -ENOMEM */
522 }
523
524 ret = btrfs_map_bio(fs_info, bio, 0);
525 if (ret) {
526 bio->bi_status = ret;
527 bio_endio(bio);
528 }
529
530 if (blkcg_css)
531 kthread_associate_blkcg(NULL);
532
533 return 0;
534}
535
536static u64 bio_end_offset(struct bio *bio)
537{
538 struct bio_vec *last = bio_last_bvec_all(bio);
539
540 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
541}
542
543static noinline int add_ra_bio_pages(struct inode *inode,
544 u64 compressed_end,
545 struct compressed_bio *cb)
546{
547 unsigned long end_index;
548 unsigned long pg_index;
549 u64 last_offset;
550 u64 isize = i_size_read(inode);
551 int ret;
552 struct page *page;
553 unsigned long nr_pages = 0;
554 struct extent_map *em;
555 struct address_space *mapping = inode->i_mapping;
556 struct extent_map_tree *em_tree;
557 struct extent_io_tree *tree;
558 u64 end;
559 int misses = 0;
560
561 last_offset = bio_end_offset(cb->orig_bio);
562 em_tree = &BTRFS_I(inode)->extent_tree;
563 tree = &BTRFS_I(inode)->io_tree;
564
565 if (isize == 0)
566 return 0;
567
568 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
569
570 while (last_offset < compressed_end) {
571 pg_index = last_offset >> PAGE_SHIFT;
572
573 if (pg_index > end_index)
574 break;
575
576 page = xa_load(&mapping->i_pages, pg_index);
577 if (page && !xa_is_value(page)) {
578 misses++;
579 if (misses > 4)
580 break;
581 goto next;
582 }
583
584 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
585 ~__GFP_FS));
586 if (!page)
587 break;
588
589 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
590 put_page(page);
591 goto next;
592 }
593
594 /*
595 * at this point, we have a locked page in the page cache
596 * for these bytes in the file. But, we have to make
597 * sure they map to this compressed extent on disk.
598 */
599 ret = set_page_extent_mapped(page);
600 if (ret < 0) {
601 unlock_page(page);
602 put_page(page);
603 break;
604 }
605
606 end = last_offset + PAGE_SIZE - 1;
607 lock_extent(tree, last_offset, end);
608 read_lock(&em_tree->lock);
609 em = lookup_extent_mapping(em_tree, last_offset,
610 PAGE_SIZE);
611 read_unlock(&em_tree->lock);
612
613 if (!em || last_offset < em->start ||
614 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
615 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
616 free_extent_map(em);
617 unlock_extent(tree, last_offset, end);
618 unlock_page(page);
619 put_page(page);
620 break;
621 }
622 free_extent_map(em);
623
624 if (page->index == end_index) {
625 size_t zero_offset = offset_in_page(isize);
626
627 if (zero_offset) {
628 int zeros;
629 zeros = PAGE_SIZE - zero_offset;
630 memzero_page(page, zero_offset, zeros);
631 flush_dcache_page(page);
632 }
633 }
634
635 ret = bio_add_page(cb->orig_bio, page,
636 PAGE_SIZE, 0);
637
638 if (ret == PAGE_SIZE) {
639 nr_pages++;
640 put_page(page);
641 } else {
642 unlock_extent(tree, last_offset, end);
643 unlock_page(page);
644 put_page(page);
645 break;
646 }
647next:
648 last_offset += PAGE_SIZE;
649 }
650 return 0;
651}
652
653/*
654 * for a compressed read, the bio we get passed has all the inode pages
655 * in it. We don't actually do IO on those pages but allocate new ones
656 * to hold the compressed pages on disk.
657 *
658 * bio->bi_iter.bi_sector points to the compressed extent on disk
659 * bio->bi_io_vec points to all of the inode pages
660 *
661 * After the compressed pages are read, we copy the bytes into the
662 * bio we were passed and then call the bio end_io calls
663 */
664blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
665 int mirror_num, unsigned long bio_flags)
666{
667 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
668 struct extent_map_tree *em_tree;
669 struct compressed_bio *cb;
670 unsigned int compressed_len;
671 unsigned int nr_pages;
672 unsigned int pg_index;
673 struct page *page;
674 struct bio *comp_bio;
675 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
676 u64 em_len;
677 u64 em_start;
678 struct extent_map *em;
679 blk_status_t ret = BLK_STS_RESOURCE;
680 int faili = 0;
681 u8 *sums;
682
683 em_tree = &BTRFS_I(inode)->extent_tree;
684
685 /* we need the actual starting offset of this extent in the file */
686 read_lock(&em_tree->lock);
687 em = lookup_extent_mapping(em_tree,
688 page_offset(bio_first_page_all(bio)),
689 fs_info->sectorsize);
690 read_unlock(&em_tree->lock);
691 if (!em)
692 return BLK_STS_IOERR;
693
694 compressed_len = em->block_len;
695 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
696 if (!cb)
697 goto out;
698
699 refcount_set(&cb->pending_bios, 0);
700 cb->errors = 0;
701 cb->inode = inode;
702 cb->mirror_num = mirror_num;
703 sums = cb->sums;
704
705 cb->start = em->orig_start;
706 em_len = em->len;
707 em_start = em->start;
708
709 free_extent_map(em);
710 em = NULL;
711
712 cb->len = bio->bi_iter.bi_size;
713 cb->compressed_len = compressed_len;
714 cb->compress_type = extent_compress_type(bio_flags);
715 cb->orig_bio = bio;
716
717 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
718 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
719 GFP_NOFS);
720 if (!cb->compressed_pages)
721 goto fail1;
722
723 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
724 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
725 __GFP_HIGHMEM);
726 if (!cb->compressed_pages[pg_index]) {
727 faili = pg_index - 1;
728 ret = BLK_STS_RESOURCE;
729 goto fail2;
730 }
731 }
732 faili = nr_pages - 1;
733 cb->nr_pages = nr_pages;
734
735 add_ra_bio_pages(inode, em_start + em_len, cb);
736
737 /* include any pages we added in add_ra-bio_pages */
738 cb->len = bio->bi_iter.bi_size;
739
740 comp_bio = btrfs_bio_alloc(cur_disk_byte);
741 comp_bio->bi_opf = REQ_OP_READ;
742 comp_bio->bi_private = cb;
743 comp_bio->bi_end_io = end_compressed_bio_read;
744 refcount_set(&cb->pending_bios, 1);
745
746 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
747 u32 pg_len = PAGE_SIZE;
748 int submit = 0;
749
750 /*
751 * To handle subpage case, we need to make sure the bio only
752 * covers the range we need.
753 *
754 * If we're at the last page, truncate the length to only cover
755 * the remaining part.
756 */
757 if (pg_index == nr_pages - 1)
758 pg_len = min_t(u32, PAGE_SIZE,
759 compressed_len - pg_index * PAGE_SIZE);
760
761 page = cb->compressed_pages[pg_index];
762 page->mapping = inode->i_mapping;
763 page->index = em_start >> PAGE_SHIFT;
764
765 if (comp_bio->bi_iter.bi_size)
766 submit = btrfs_bio_fits_in_stripe(page, pg_len,
767 comp_bio, 0);
768
769 page->mapping = NULL;
770 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
771 unsigned int nr_sectors;
772
773 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
774 BTRFS_WQ_ENDIO_DATA);
775 BUG_ON(ret); /* -ENOMEM */
776
777 /*
778 * inc the count before we submit the bio so
779 * we know the end IO handler won't happen before
780 * we inc the count. Otherwise, the cb might get
781 * freed before we're done setting it up
782 */
783 refcount_inc(&cb->pending_bios);
784
785 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
786 BUG_ON(ret); /* -ENOMEM */
787
788 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
789 fs_info->sectorsize);
790 sums += fs_info->csum_size * nr_sectors;
791
792 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
793 if (ret) {
794 comp_bio->bi_status = ret;
795 bio_endio(comp_bio);
796 }
797
798 comp_bio = btrfs_bio_alloc(cur_disk_byte);
799 comp_bio->bi_opf = REQ_OP_READ;
800 comp_bio->bi_private = cb;
801 comp_bio->bi_end_io = end_compressed_bio_read;
802
803 bio_add_page(comp_bio, page, pg_len, 0);
804 }
805 cur_disk_byte += pg_len;
806 }
807
808 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
809 BUG_ON(ret); /* -ENOMEM */
810
811 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
812 BUG_ON(ret); /* -ENOMEM */
813
814 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
815 if (ret) {
816 comp_bio->bi_status = ret;
817 bio_endio(comp_bio);
818 }
819
820 return 0;
821
822fail2:
823 while (faili >= 0) {
824 __free_page(cb->compressed_pages[faili]);
825 faili--;
826 }
827
828 kfree(cb->compressed_pages);
829fail1:
830 kfree(cb);
831out:
832 free_extent_map(em);
833 return ret;
834}
835
836/*
837 * Heuristic uses systematic sampling to collect data from the input data
838 * range, the logic can be tuned by the following constants:
839 *
840 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
841 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
842 */
843#define SAMPLING_READ_SIZE (16)
844#define SAMPLING_INTERVAL (256)
845
846/*
847 * For statistical analysis of the input data we consider bytes that form a
848 * Galois Field of 256 objects. Each object has an attribute count, ie. how
849 * many times the object appeared in the sample.
850 */
851#define BUCKET_SIZE (256)
852
853/*
854 * The size of the sample is based on a statistical sampling rule of thumb.
855 * The common way is to perform sampling tests as long as the number of
856 * elements in each cell is at least 5.
857 *
858 * Instead of 5, we choose 32 to obtain more accurate results.
859 * If the data contain the maximum number of symbols, which is 256, we obtain a
860 * sample size bound by 8192.
861 *
862 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
863 * from up to 512 locations.
864 */
865#define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
866 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
867
868struct bucket_item {
869 u32 count;
870};
871
872struct heuristic_ws {
873 /* Partial copy of input data */
874 u8 *sample;
875 u32 sample_size;
876 /* Buckets store counters for each byte value */
877 struct bucket_item *bucket;
878 /* Sorting buffer */
879 struct bucket_item *bucket_b;
880 struct list_head list;
881};
882
883static struct workspace_manager heuristic_wsm;
884
885static void free_heuristic_ws(struct list_head *ws)
886{
887 struct heuristic_ws *workspace;
888
889 workspace = list_entry(ws, struct heuristic_ws, list);
890
891 kvfree(workspace->sample);
892 kfree(workspace->bucket);
893 kfree(workspace->bucket_b);
894 kfree(workspace);
895}
896
897static struct list_head *alloc_heuristic_ws(unsigned int level)
898{
899 struct heuristic_ws *ws;
900
901 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
902 if (!ws)
903 return ERR_PTR(-ENOMEM);
904
905 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
906 if (!ws->sample)
907 goto fail;
908
909 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
910 if (!ws->bucket)
911 goto fail;
912
913 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
914 if (!ws->bucket_b)
915 goto fail;
916
917 INIT_LIST_HEAD(&ws->list);
918 return &ws->list;
919fail:
920 free_heuristic_ws(&ws->list);
921 return ERR_PTR(-ENOMEM);
922}
923
924const struct btrfs_compress_op btrfs_heuristic_compress = {
925 .workspace_manager = &heuristic_wsm,
926};
927
928static const struct btrfs_compress_op * const btrfs_compress_op[] = {
929 /* The heuristic is represented as compression type 0 */
930 &btrfs_heuristic_compress,
931 &btrfs_zlib_compress,
932 &btrfs_lzo_compress,
933 &btrfs_zstd_compress,
934};
935
936static struct list_head *alloc_workspace(int type, unsigned int level)
937{
938 switch (type) {
939 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
940 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
941 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
942 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
943 default:
944 /*
945 * This can't happen, the type is validated several times
946 * before we get here.
947 */
948 BUG();
949 }
950}
951
952static void free_workspace(int type, struct list_head *ws)
953{
954 switch (type) {
955 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
956 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
957 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
958 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
959 default:
960 /*
961 * This can't happen, the type is validated several times
962 * before we get here.
963 */
964 BUG();
965 }
966}
967
968static void btrfs_init_workspace_manager(int type)
969{
970 struct workspace_manager *wsm;
971 struct list_head *workspace;
972
973 wsm = btrfs_compress_op[type]->workspace_manager;
974 INIT_LIST_HEAD(&wsm->idle_ws);
975 spin_lock_init(&wsm->ws_lock);
976 atomic_set(&wsm->total_ws, 0);
977 init_waitqueue_head(&wsm->ws_wait);
978
979 /*
980 * Preallocate one workspace for each compression type so we can
981 * guarantee forward progress in the worst case
982 */
983 workspace = alloc_workspace(type, 0);
984 if (IS_ERR(workspace)) {
985 pr_warn(
986 "BTRFS: cannot preallocate compression workspace, will try later\n");
987 } else {
988 atomic_set(&wsm->total_ws, 1);
989 wsm->free_ws = 1;
990 list_add(workspace, &wsm->idle_ws);
991 }
992}
993
994static void btrfs_cleanup_workspace_manager(int type)
995{
996 struct workspace_manager *wsman;
997 struct list_head *ws;
998
999 wsman = btrfs_compress_op[type]->workspace_manager;
1000 while (!list_empty(&wsman->idle_ws)) {
1001 ws = wsman->idle_ws.next;
1002 list_del(ws);
1003 free_workspace(type, ws);
1004 atomic_dec(&wsman->total_ws);
1005 }
1006}
1007
1008/*
1009 * This finds an available workspace or allocates a new one.
1010 * If it's not possible to allocate a new one, waits until there's one.
1011 * Preallocation makes a forward progress guarantees and we do not return
1012 * errors.
1013 */
1014struct list_head *btrfs_get_workspace(int type, unsigned int level)
1015{
1016 struct workspace_manager *wsm;
1017 struct list_head *workspace;
1018 int cpus = num_online_cpus();
1019 unsigned nofs_flag;
1020 struct list_head *idle_ws;
1021 spinlock_t *ws_lock;
1022 atomic_t *total_ws;
1023 wait_queue_head_t *ws_wait;
1024 int *free_ws;
1025
1026 wsm = btrfs_compress_op[type]->workspace_manager;
1027 idle_ws = &wsm->idle_ws;
1028 ws_lock = &wsm->ws_lock;
1029 total_ws = &wsm->total_ws;
1030 ws_wait = &wsm->ws_wait;
1031 free_ws = &wsm->free_ws;
1032
1033again:
1034 spin_lock(ws_lock);
1035 if (!list_empty(idle_ws)) {
1036 workspace = idle_ws->next;
1037 list_del(workspace);
1038 (*free_ws)--;
1039 spin_unlock(ws_lock);
1040 return workspace;
1041
1042 }
1043 if (atomic_read(total_ws) > cpus) {
1044 DEFINE_WAIT(wait);
1045
1046 spin_unlock(ws_lock);
1047 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1048 if (atomic_read(total_ws) > cpus && !*free_ws)
1049 schedule();
1050 finish_wait(ws_wait, &wait);
1051 goto again;
1052 }
1053 atomic_inc(total_ws);
1054 spin_unlock(ws_lock);
1055
1056 /*
1057 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1058 * to turn it off here because we might get called from the restricted
1059 * context of btrfs_compress_bio/btrfs_compress_pages
1060 */
1061 nofs_flag = memalloc_nofs_save();
1062 workspace = alloc_workspace(type, level);
1063 memalloc_nofs_restore(nofs_flag);
1064
1065 if (IS_ERR(workspace)) {
1066 atomic_dec(total_ws);
1067 wake_up(ws_wait);
1068
1069 /*
1070 * Do not return the error but go back to waiting. There's a
1071 * workspace preallocated for each type and the compression
1072 * time is bounded so we get to a workspace eventually. This
1073 * makes our caller's life easier.
1074 *
1075 * To prevent silent and low-probability deadlocks (when the
1076 * initial preallocation fails), check if there are any
1077 * workspaces at all.
1078 */
1079 if (atomic_read(total_ws) == 0) {
1080 static DEFINE_RATELIMIT_STATE(_rs,
1081 /* once per minute */ 60 * HZ,
1082 /* no burst */ 1);
1083
1084 if (__ratelimit(&_rs)) {
1085 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1086 }
1087 }
1088 goto again;
1089 }
1090 return workspace;
1091}
1092
1093static struct list_head *get_workspace(int type, int level)
1094{
1095 switch (type) {
1096 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1097 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1098 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1099 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1100 default:
1101 /*
1102 * This can't happen, the type is validated several times
1103 * before we get here.
1104 */
1105 BUG();
1106 }
1107}
1108
1109/*
1110 * put a workspace struct back on the list or free it if we have enough
1111 * idle ones sitting around
1112 */
1113void btrfs_put_workspace(int type, struct list_head *ws)
1114{
1115 struct workspace_manager *wsm;
1116 struct list_head *idle_ws;
1117 spinlock_t *ws_lock;
1118 atomic_t *total_ws;
1119 wait_queue_head_t *ws_wait;
1120 int *free_ws;
1121
1122 wsm = btrfs_compress_op[type]->workspace_manager;
1123 idle_ws = &wsm->idle_ws;
1124 ws_lock = &wsm->ws_lock;
1125 total_ws = &wsm->total_ws;
1126 ws_wait = &wsm->ws_wait;
1127 free_ws = &wsm->free_ws;
1128
1129 spin_lock(ws_lock);
1130 if (*free_ws <= num_online_cpus()) {
1131 list_add(ws, idle_ws);
1132 (*free_ws)++;
1133 spin_unlock(ws_lock);
1134 goto wake;
1135 }
1136 spin_unlock(ws_lock);
1137
1138 free_workspace(type, ws);
1139 atomic_dec(total_ws);
1140wake:
1141 cond_wake_up(ws_wait);
1142}
1143
1144static void put_workspace(int type, struct list_head *ws)
1145{
1146 switch (type) {
1147 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1148 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1149 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1150 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1151 default:
1152 /*
1153 * This can't happen, the type is validated several times
1154 * before we get here.
1155 */
1156 BUG();
1157 }
1158}
1159
1160/*
1161 * Adjust @level according to the limits of the compression algorithm or
1162 * fallback to default
1163 */
1164static unsigned int btrfs_compress_set_level(int type, unsigned level)
1165{
1166 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1167
1168 if (level == 0)
1169 level = ops->default_level;
1170 else
1171 level = min(level, ops->max_level);
1172
1173 return level;
1174}
1175
1176/*
1177 * Given an address space and start and length, compress the bytes into @pages
1178 * that are allocated on demand.
1179 *
1180 * @type_level is encoded algorithm and level, where level 0 means whatever
1181 * default the algorithm chooses and is opaque here;
1182 * - compression algo are 0-3
1183 * - the level are bits 4-7
1184 *
1185 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1186 * and returns number of actually allocated pages
1187 *
1188 * @total_in is used to return the number of bytes actually read. It
1189 * may be smaller than the input length if we had to exit early because we
1190 * ran out of room in the pages array or because we cross the
1191 * max_out threshold.
1192 *
1193 * @total_out is an in/out parameter, must be set to the input length and will
1194 * be also used to return the total number of compressed bytes
1195 */
1196int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1197 u64 start, struct page **pages,
1198 unsigned long *out_pages,
1199 unsigned long *total_in,
1200 unsigned long *total_out)
1201{
1202 int type = btrfs_compress_type(type_level);
1203 int level = btrfs_compress_level(type_level);
1204 struct list_head *workspace;
1205 int ret;
1206
1207 level = btrfs_compress_set_level(type, level);
1208 workspace = get_workspace(type, level);
1209 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1210 out_pages, total_in, total_out);
1211 put_workspace(type, workspace);
1212 return ret;
1213}
1214
1215static int btrfs_decompress_bio(struct compressed_bio *cb)
1216{
1217 struct list_head *workspace;
1218 int ret;
1219 int type = cb->compress_type;
1220
1221 workspace = get_workspace(type, 0);
1222 ret = compression_decompress_bio(type, workspace, cb);
1223 put_workspace(type, workspace);
1224
1225 return ret;
1226}
1227
1228/*
1229 * a less complex decompression routine. Our compressed data fits in a
1230 * single page, and we want to read a single page out of it.
1231 * start_byte tells us the offset into the compressed data we're interested in
1232 */
1233int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1234 unsigned long start_byte, size_t srclen, size_t destlen)
1235{
1236 struct list_head *workspace;
1237 int ret;
1238
1239 workspace = get_workspace(type, 0);
1240 ret = compression_decompress(type, workspace, data_in, dest_page,
1241 start_byte, srclen, destlen);
1242 put_workspace(type, workspace);
1243
1244 return ret;
1245}
1246
1247void __init btrfs_init_compress(void)
1248{
1249 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1250 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1251 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1252 zstd_init_workspace_manager();
1253}
1254
1255void __cold btrfs_exit_compress(void)
1256{
1257 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1258 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1259 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1260 zstd_cleanup_workspace_manager();
1261}
1262
1263/*
1264 * Copy uncompressed data from working buffer to pages.
1265 *
1266 * buf_start is the byte offset we're of the start of our workspace buffer.
1267 *
1268 * total_out is the last byte of the buffer
1269 */
1270int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1271 unsigned long total_out, u64 disk_start,
1272 struct bio *bio)
1273{
1274 unsigned long buf_offset;
1275 unsigned long current_buf_start;
1276 unsigned long start_byte;
1277 unsigned long prev_start_byte;
1278 unsigned long working_bytes = total_out - buf_start;
1279 unsigned long bytes;
1280 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1281
1282 /*
1283 * start byte is the first byte of the page we're currently
1284 * copying into relative to the start of the compressed data.
1285 */
1286 start_byte = page_offset(bvec.bv_page) - disk_start;
1287
1288 /* we haven't yet hit data corresponding to this page */
1289 if (total_out <= start_byte)
1290 return 1;
1291
1292 /*
1293 * the start of the data we care about is offset into
1294 * the middle of our working buffer
1295 */
1296 if (total_out > start_byte && buf_start < start_byte) {
1297 buf_offset = start_byte - buf_start;
1298 working_bytes -= buf_offset;
1299 } else {
1300 buf_offset = 0;
1301 }
1302 current_buf_start = buf_start;
1303
1304 /* copy bytes from the working buffer into the pages */
1305 while (working_bytes > 0) {
1306 bytes = min_t(unsigned long, bvec.bv_len,
1307 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1308 bytes = min(bytes, working_bytes);
1309
1310 memcpy_to_page(bvec.bv_page, bvec.bv_offset, buf + buf_offset,
1311 bytes);
1312 flush_dcache_page(bvec.bv_page);
1313
1314 buf_offset += bytes;
1315 working_bytes -= bytes;
1316 current_buf_start += bytes;
1317
1318 /* check if we need to pick another page */
1319 bio_advance(bio, bytes);
1320 if (!bio->bi_iter.bi_size)
1321 return 0;
1322 bvec = bio_iter_iovec(bio, bio->bi_iter);
1323 prev_start_byte = start_byte;
1324 start_byte = page_offset(bvec.bv_page) - disk_start;
1325
1326 /*
1327 * We need to make sure we're only adjusting
1328 * our offset into compression working buffer when
1329 * we're switching pages. Otherwise we can incorrectly
1330 * keep copying when we were actually done.
1331 */
1332 if (start_byte != prev_start_byte) {
1333 /*
1334 * make sure our new page is covered by this
1335 * working buffer
1336 */
1337 if (total_out <= start_byte)
1338 return 1;
1339
1340 /*
1341 * the next page in the biovec might not be adjacent
1342 * to the last page, but it might still be found
1343 * inside this working buffer. bump our offset pointer
1344 */
1345 if (total_out > start_byte &&
1346 current_buf_start < start_byte) {
1347 buf_offset = start_byte - buf_start;
1348 working_bytes = total_out - start_byte;
1349 current_buf_start = buf_start + buf_offset;
1350 }
1351 }
1352 }
1353
1354 return 1;
1355}
1356
1357/*
1358 * Shannon Entropy calculation
1359 *
1360 * Pure byte distribution analysis fails to determine compressibility of data.
1361 * Try calculating entropy to estimate the average minimum number of bits
1362 * needed to encode the sampled data.
1363 *
1364 * For convenience, return the percentage of needed bits, instead of amount of
1365 * bits directly.
1366 *
1367 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1368 * and can be compressible with high probability
1369 *
1370 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1371 *
1372 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1373 */
1374#define ENTROPY_LVL_ACEPTABLE (65)
1375#define ENTROPY_LVL_HIGH (80)
1376
1377/*
1378 * For increasead precision in shannon_entropy calculation,
1379 * let's do pow(n, M) to save more digits after comma:
1380 *
1381 * - maximum int bit length is 64
1382 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1383 * - 13 * 4 = 52 < 64 -> M = 4
1384 *
1385 * So use pow(n, 4).
1386 */
1387static inline u32 ilog2_w(u64 n)
1388{
1389 return ilog2(n * n * n * n);
1390}
1391
1392static u32 shannon_entropy(struct heuristic_ws *ws)
1393{
1394 const u32 entropy_max = 8 * ilog2_w(2);
1395 u32 entropy_sum = 0;
1396 u32 p, p_base, sz_base;
1397 u32 i;
1398
1399 sz_base = ilog2_w(ws->sample_size);
1400 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1401 p = ws->bucket[i].count;
1402 p_base = ilog2_w(p);
1403 entropy_sum += p * (sz_base - p_base);
1404 }
1405
1406 entropy_sum /= ws->sample_size;
1407 return entropy_sum * 100 / entropy_max;
1408}
1409
1410#define RADIX_BASE 4U
1411#define COUNTERS_SIZE (1U << RADIX_BASE)
1412
1413static u8 get4bits(u64 num, int shift) {
1414 u8 low4bits;
1415
1416 num >>= shift;
1417 /* Reverse order */
1418 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1419 return low4bits;
1420}
1421
1422/*
1423 * Use 4 bits as radix base
1424 * Use 16 u32 counters for calculating new position in buf array
1425 *
1426 * @array - array that will be sorted
1427 * @array_buf - buffer array to store sorting results
1428 * must be equal in size to @array
1429 * @num - array size
1430 */
1431static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1432 int num)
1433{
1434 u64 max_num;
1435 u64 buf_num;
1436 u32 counters[COUNTERS_SIZE];
1437 u32 new_addr;
1438 u32 addr;
1439 int bitlen;
1440 int shift;
1441 int i;
1442
1443 /*
1444 * Try avoid useless loop iterations for small numbers stored in big
1445 * counters. Example: 48 33 4 ... in 64bit array
1446 */
1447 max_num = array[0].count;
1448 for (i = 1; i < num; i++) {
1449 buf_num = array[i].count;
1450 if (buf_num > max_num)
1451 max_num = buf_num;
1452 }
1453
1454 buf_num = ilog2(max_num);
1455 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1456
1457 shift = 0;
1458 while (shift < bitlen) {
1459 memset(counters, 0, sizeof(counters));
1460
1461 for (i = 0; i < num; i++) {
1462 buf_num = array[i].count;
1463 addr = get4bits(buf_num, shift);
1464 counters[addr]++;
1465 }
1466
1467 for (i = 1; i < COUNTERS_SIZE; i++)
1468 counters[i] += counters[i - 1];
1469
1470 for (i = num - 1; i >= 0; i--) {
1471 buf_num = array[i].count;
1472 addr = get4bits(buf_num, shift);
1473 counters[addr]--;
1474 new_addr = counters[addr];
1475 array_buf[new_addr] = array[i];
1476 }
1477
1478 shift += RADIX_BASE;
1479
1480 /*
1481 * Normal radix expects to move data from a temporary array, to
1482 * the main one. But that requires some CPU time. Avoid that
1483 * by doing another sort iteration to original array instead of
1484 * memcpy()
1485 */
1486 memset(counters, 0, sizeof(counters));
1487
1488 for (i = 0; i < num; i ++) {
1489 buf_num = array_buf[i].count;
1490 addr = get4bits(buf_num, shift);
1491 counters[addr]++;
1492 }
1493
1494 for (i = 1; i < COUNTERS_SIZE; i++)
1495 counters[i] += counters[i - 1];
1496
1497 for (i = num - 1; i >= 0; i--) {
1498 buf_num = array_buf[i].count;
1499 addr = get4bits(buf_num, shift);
1500 counters[addr]--;
1501 new_addr = counters[addr];
1502 array[new_addr] = array_buf[i];
1503 }
1504
1505 shift += RADIX_BASE;
1506 }
1507}
1508
1509/*
1510 * Size of the core byte set - how many bytes cover 90% of the sample
1511 *
1512 * There are several types of structured binary data that use nearly all byte
1513 * values. The distribution can be uniform and counts in all buckets will be
1514 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1515 *
1516 * Other possibility is normal (Gaussian) distribution, where the data could
1517 * be potentially compressible, but we have to take a few more steps to decide
1518 * how much.
1519 *
1520 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1521 * compression algo can easy fix that
1522 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1523 * probability is not compressible
1524 */
1525#define BYTE_CORE_SET_LOW (64)
1526#define BYTE_CORE_SET_HIGH (200)
1527
1528static int byte_core_set_size(struct heuristic_ws *ws)
1529{
1530 u32 i;
1531 u32 coreset_sum = 0;
1532 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1533 struct bucket_item *bucket = ws->bucket;
1534
1535 /* Sort in reverse order */
1536 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1537
1538 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1539 coreset_sum += bucket[i].count;
1540
1541 if (coreset_sum > core_set_threshold)
1542 return i;
1543
1544 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1545 coreset_sum += bucket[i].count;
1546 if (coreset_sum > core_set_threshold)
1547 break;
1548 }
1549
1550 return i;
1551}
1552
1553/*
1554 * Count byte values in buckets.
1555 * This heuristic can detect textual data (configs, xml, json, html, etc).
1556 * Because in most text-like data byte set is restricted to limited number of
1557 * possible characters, and that restriction in most cases makes data easy to
1558 * compress.
1559 *
1560 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1561 * less - compressible
1562 * more - need additional analysis
1563 */
1564#define BYTE_SET_THRESHOLD (64)
1565
1566static u32 byte_set_size(const struct heuristic_ws *ws)
1567{
1568 u32 i;
1569 u32 byte_set_size = 0;
1570
1571 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1572 if (ws->bucket[i].count > 0)
1573 byte_set_size++;
1574 }
1575
1576 /*
1577 * Continue collecting count of byte values in buckets. If the byte
1578 * set size is bigger then the threshold, it's pointless to continue,
1579 * the detection technique would fail for this type of data.
1580 */
1581 for (; i < BUCKET_SIZE; i++) {
1582 if (ws->bucket[i].count > 0) {
1583 byte_set_size++;
1584 if (byte_set_size > BYTE_SET_THRESHOLD)
1585 return byte_set_size;
1586 }
1587 }
1588
1589 return byte_set_size;
1590}
1591
1592static bool sample_repeated_patterns(struct heuristic_ws *ws)
1593{
1594 const u32 half_of_sample = ws->sample_size / 2;
1595 const u8 *data = ws->sample;
1596
1597 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1598}
1599
1600static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1601 struct heuristic_ws *ws)
1602{
1603 struct page *page;
1604 u64 index, index_end;
1605 u32 i, curr_sample_pos;
1606 u8 *in_data;
1607
1608 /*
1609 * Compression handles the input data by chunks of 128KiB
1610 * (defined by BTRFS_MAX_UNCOMPRESSED)
1611 *
1612 * We do the same for the heuristic and loop over the whole range.
1613 *
1614 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1615 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1616 */
1617 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1618 end = start + BTRFS_MAX_UNCOMPRESSED;
1619
1620 index = start >> PAGE_SHIFT;
1621 index_end = end >> PAGE_SHIFT;
1622
1623 /* Don't miss unaligned end */
1624 if (!IS_ALIGNED(end, PAGE_SIZE))
1625 index_end++;
1626
1627 curr_sample_pos = 0;
1628 while (index < index_end) {
1629 page = find_get_page(inode->i_mapping, index);
1630 in_data = kmap_local_page(page);
1631 /* Handle case where the start is not aligned to PAGE_SIZE */
1632 i = start % PAGE_SIZE;
1633 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1634 /* Don't sample any garbage from the last page */
1635 if (start > end - SAMPLING_READ_SIZE)
1636 break;
1637 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1638 SAMPLING_READ_SIZE);
1639 i += SAMPLING_INTERVAL;
1640 start += SAMPLING_INTERVAL;
1641 curr_sample_pos += SAMPLING_READ_SIZE;
1642 }
1643 kunmap_local(in_data);
1644 put_page(page);
1645
1646 index++;
1647 }
1648
1649 ws->sample_size = curr_sample_pos;
1650}
1651
1652/*
1653 * Compression heuristic.
1654 *
1655 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1656 * quickly (compared to direct compression) detect data characteristics
1657 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1658 * data.
1659 *
1660 * The following types of analysis can be performed:
1661 * - detect mostly zero data
1662 * - detect data with low "byte set" size (text, etc)
1663 * - detect data with low/high "core byte" set
1664 *
1665 * Return non-zero if the compression should be done, 0 otherwise.
1666 */
1667int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1668{
1669 struct list_head *ws_list = get_workspace(0, 0);
1670 struct heuristic_ws *ws;
1671 u32 i;
1672 u8 byte;
1673 int ret = 0;
1674
1675 ws = list_entry(ws_list, struct heuristic_ws, list);
1676
1677 heuristic_collect_sample(inode, start, end, ws);
1678
1679 if (sample_repeated_patterns(ws)) {
1680 ret = 1;
1681 goto out;
1682 }
1683
1684 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1685
1686 for (i = 0; i < ws->sample_size; i++) {
1687 byte = ws->sample[i];
1688 ws->bucket[byte].count++;
1689 }
1690
1691 i = byte_set_size(ws);
1692 if (i < BYTE_SET_THRESHOLD) {
1693 ret = 2;
1694 goto out;
1695 }
1696
1697 i = byte_core_set_size(ws);
1698 if (i <= BYTE_CORE_SET_LOW) {
1699 ret = 3;
1700 goto out;
1701 }
1702
1703 if (i >= BYTE_CORE_SET_HIGH) {
1704 ret = 0;
1705 goto out;
1706 }
1707
1708 i = shannon_entropy(ws);
1709 if (i <= ENTROPY_LVL_ACEPTABLE) {
1710 ret = 4;
1711 goto out;
1712 }
1713
1714 /*
1715 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1716 * needed to give green light to compression.
1717 *
1718 * For now just assume that compression at that level is not worth the
1719 * resources because:
1720 *
1721 * 1. it is possible to defrag the data later
1722 *
1723 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1724 * values, every bucket has counter at level ~54. The heuristic would
1725 * be confused. This can happen when data have some internal repeated
1726 * patterns like "abbacbbc...". This can be detected by analyzing
1727 * pairs of bytes, which is too costly.
1728 */
1729 if (i < ENTROPY_LVL_HIGH) {
1730 ret = 5;
1731 goto out;
1732 } else {
1733 ret = 0;
1734 goto out;
1735 }
1736
1737out:
1738 put_workspace(0, ws_list);
1739 return ret;
1740}
1741
1742/*
1743 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1744 * level, unrecognized string will set the default level
1745 */
1746unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1747{
1748 unsigned int level = 0;
1749 int ret;
1750
1751 if (!type)
1752 return 0;
1753
1754 if (str[0] == ':') {
1755 ret = kstrtouint(str + 1, 10, &level);
1756 if (ret)
1757 level = 0;
1758 }
1759
1760 level = btrfs_compress_set_level(type, level);
1761
1762 return level;
1763}