1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Oracle. All rights reserved.
6 #include <linux/kernel.h>
8 #include <linux/file.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>
24 #include "transaction.h"
25 #include "btrfs_inode.h"
27 #include "ordered-data.h"
28 #include "compression.h"
29 #include "extent_io.h"
30 #include "extent_map.h"
32 int zlib_compress_pages(struct list_head *ws, struct address_space *mapping,
33 u64 start, struct page **pages, unsigned long *out_pages,
34 unsigned long *total_in, unsigned long *total_out);
35 int zlib_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
36 int zlib_decompress(struct list_head *ws, unsigned char *data_in,
37 struct page *dest_page, unsigned long start_byte, size_t srclen,
39 struct list_head *zlib_alloc_workspace(unsigned int level);
40 void zlib_free_workspace(struct list_head *ws);
41 struct list_head *zlib_get_workspace(unsigned int level);
43 int lzo_compress_pages(struct list_head *ws, struct address_space *mapping,
44 u64 start, struct page **pages, unsigned long *out_pages,
45 unsigned long *total_in, unsigned long *total_out);
46 int lzo_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
47 int lzo_decompress(struct list_head *ws, unsigned char *data_in,
48 struct page *dest_page, unsigned long start_byte, size_t srclen,
50 struct list_head *lzo_alloc_workspace(unsigned int level);
51 void lzo_free_workspace(struct list_head *ws);
53 int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
54 u64 start, struct page **pages, unsigned long *out_pages,
55 unsigned long *total_in, unsigned long *total_out);
56 int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
57 int zstd_decompress(struct list_head *ws, unsigned char *data_in,
58 struct page *dest_page, unsigned long start_byte, size_t srclen,
60 void zstd_init_workspace_manager(void);
61 void zstd_cleanup_workspace_manager(void);
62 struct list_head *zstd_alloc_workspace(unsigned int level);
63 void zstd_free_workspace(struct list_head *ws);
64 struct list_head *zstd_get_workspace(unsigned int level);
65 void zstd_put_workspace(struct list_head *ws);
67 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
69 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
72 case BTRFS_COMPRESS_ZLIB:
73 case BTRFS_COMPRESS_LZO:
74 case BTRFS_COMPRESS_ZSTD:
75 case BTRFS_COMPRESS_NONE:
76 return btrfs_compress_types[type];
84 bool btrfs_compress_is_valid_type(const char *str, size_t len)
88 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
89 size_t comp_len = strlen(btrfs_compress_types[i]);
94 if (!strncmp(btrfs_compress_types[i], str, comp_len))
100 static int compression_compress_pages(int type, struct list_head *ws,
101 struct address_space *mapping, u64 start, struct page **pages,
102 unsigned long *out_pages, unsigned long *total_in,
103 unsigned long *total_out)
106 case BTRFS_COMPRESS_ZLIB:
107 return zlib_compress_pages(ws, mapping, start, pages,
108 out_pages, total_in, total_out);
109 case BTRFS_COMPRESS_LZO:
110 return lzo_compress_pages(ws, mapping, start, pages,
111 out_pages, total_in, total_out);
112 case BTRFS_COMPRESS_ZSTD:
113 return zstd_compress_pages(ws, mapping, start, pages,
114 out_pages, total_in, total_out);
115 case BTRFS_COMPRESS_NONE:
118 * This can't happen, the type is validated several times
119 * before we get here. As a sane fallback, return what the
120 * callers will understand as 'no compression happened'.
126 static int compression_decompress_bio(int type, struct list_head *ws,
127 struct compressed_bio *cb)
130 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
131 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
132 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
133 case BTRFS_COMPRESS_NONE:
136 * This can't happen, the type is validated several times
137 * before we get here.
143 static int compression_decompress(int type, struct list_head *ws,
144 unsigned char *data_in, struct page *dest_page,
145 unsigned long start_byte, size_t srclen, size_t destlen)
148 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
149 start_byte, srclen, destlen);
150 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
151 start_byte, srclen, destlen);
152 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
153 start_byte, srclen, destlen);
154 case BTRFS_COMPRESS_NONE:
157 * This can't happen, the type is validated several times
158 * before we get here.
164 static int btrfs_decompress_bio(struct compressed_bio *cb);
166 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
167 unsigned long disk_size)
169 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
171 return sizeof(struct compressed_bio) +
172 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
175 static int check_compressed_csum(struct btrfs_inode *inode,
176 struct compressed_bio *cb,
179 struct btrfs_fs_info *fs_info = inode->root->fs_info;
180 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
181 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
186 u8 csum[BTRFS_CSUM_SIZE];
187 u8 *cb_sum = cb->sums;
189 if (inode->flags & BTRFS_INODE_NODATASUM)
192 shash->tfm = fs_info->csum_shash;
194 for (i = 0; i < cb->nr_pages; i++) {
195 page = cb->compressed_pages[i];
197 crypto_shash_init(shash);
198 kaddr = kmap_atomic(page);
199 crypto_shash_update(shash, kaddr, PAGE_SIZE);
200 kunmap_atomic(kaddr);
201 crypto_shash_final(shash, (u8 *)&csum);
203 if (memcmp(&csum, cb_sum, csum_size)) {
204 btrfs_print_data_csum_error(inode, disk_start,
205 csum, cb_sum, cb->mirror_num);
217 /* when we finish reading compressed pages from the disk, we
218 * decompress them and then run the bio end_io routines on the
219 * decompressed pages (in the inode address space).
221 * This allows the checksumming and other IO error handling routines
224 * The compressed pages are freed here, and it must be run
227 static void end_compressed_bio_read(struct bio *bio)
229 struct compressed_bio *cb = bio->bi_private;
233 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
239 /* if there are more bios still pending for this compressed
242 if (!refcount_dec_and_test(&cb->pending_bios))
246 * Record the correct mirror_num in cb->orig_bio so that
247 * read-repair can work properly.
249 ASSERT(btrfs_io_bio(cb->orig_bio));
250 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
251 cb->mirror_num = mirror;
254 * Some IO in this cb have failed, just skip checksum as there
255 * is no way it could be correct.
261 ret = check_compressed_csum(BTRFS_I(inode), cb,
262 (u64)bio->bi_iter.bi_sector << 9);
266 /* ok, we're the last bio for this extent, lets start
269 ret = btrfs_decompress_bio(cb);
275 /* release the compressed pages */
277 for (index = 0; index < cb->nr_pages; index++) {
278 page = cb->compressed_pages[index];
279 page->mapping = NULL;
283 /* do io completion on the original bio */
285 bio_io_error(cb->orig_bio);
287 struct bio_vec *bvec;
288 struct bvec_iter_all iter_all;
291 * we have verified the checksum already, set page
292 * checked so the end_io handlers know about it
294 ASSERT(!bio_flagged(bio, BIO_CLONED));
295 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
296 SetPageChecked(bvec->bv_page);
298 bio_endio(cb->orig_bio);
301 /* finally free the cb struct */
302 kfree(cb->compressed_pages);
309 * Clear the writeback bits on all of the file
310 * pages for a compressed write
312 static noinline void end_compressed_writeback(struct inode *inode,
313 const struct compressed_bio *cb)
315 unsigned long index = cb->start >> PAGE_SHIFT;
316 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
317 struct page *pages[16];
318 unsigned long nr_pages = end_index - index + 1;
323 mapping_set_error(inode->i_mapping, -EIO);
325 while (nr_pages > 0) {
326 ret = find_get_pages_contig(inode->i_mapping, index,
328 nr_pages, ARRAY_SIZE(pages)), pages);
334 for (i = 0; i < ret; i++) {
336 SetPageError(pages[i]);
337 end_page_writeback(pages[i]);
343 /* the inode may be gone now */
347 * do the cleanup once all the compressed pages hit the disk.
348 * This will clear writeback on the file pages and free the compressed
351 * This also calls the writeback end hooks for the file pages so that
352 * metadata and checksums can be updated in the file.
354 static void end_compressed_bio_write(struct bio *bio)
356 struct compressed_bio *cb = bio->bi_private;
364 /* if there are more bios still pending for this compressed
367 if (!refcount_dec_and_test(&cb->pending_bios))
370 /* ok, we're the last bio for this extent, step one is to
371 * call back into the FS and do all the end_io operations
374 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
375 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
376 cb->start, cb->start + cb->len - 1,
377 bio->bi_status == BLK_STS_OK);
378 cb->compressed_pages[0]->mapping = NULL;
380 end_compressed_writeback(inode, cb);
381 /* note, our inode could be gone now */
384 * release the compressed pages, these came from alloc_page and
385 * are not attached to the inode at all
388 for (index = 0; index < cb->nr_pages; index++) {
389 page = cb->compressed_pages[index];
390 page->mapping = NULL;
394 /* finally free the cb struct */
395 kfree(cb->compressed_pages);
402 * worker function to build and submit bios for previously compressed pages.
403 * The corresponding pages in the inode should be marked for writeback
404 * and the compressed pages should have a reference on them for dropping
405 * when the IO is complete.
407 * This also checksums the file bytes and gets things ready for
410 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
411 unsigned long len, u64 disk_start,
412 unsigned long compressed_len,
413 struct page **compressed_pages,
414 unsigned long nr_pages,
415 unsigned int write_flags,
416 struct cgroup_subsys_state *blkcg_css)
418 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
419 struct bio *bio = NULL;
420 struct compressed_bio *cb;
421 unsigned long bytes_left;
424 u64 first_byte = disk_start;
426 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
428 WARN_ON(!PAGE_ALIGNED(start));
429 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
431 return BLK_STS_RESOURCE;
432 refcount_set(&cb->pending_bios, 0);
438 cb->compressed_pages = compressed_pages;
439 cb->compressed_len = compressed_len;
441 cb->nr_pages = nr_pages;
443 bio = btrfs_bio_alloc(first_byte);
444 bio->bi_opf = REQ_OP_WRITE | write_flags;
445 bio->bi_private = cb;
446 bio->bi_end_io = end_compressed_bio_write;
449 bio->bi_opf |= REQ_CGROUP_PUNT;
450 bio_associate_blkg_from_css(bio, blkcg_css);
452 refcount_set(&cb->pending_bios, 1);
454 /* create and submit bios for the compressed pages */
455 bytes_left = compressed_len;
456 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
459 page = compressed_pages[pg_index];
460 page->mapping = inode->i_mapping;
461 if (bio->bi_iter.bi_size)
462 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
465 page->mapping = NULL;
466 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
469 * inc the count before we submit the bio so
470 * we know the end IO handler won't happen before
471 * we inc the count. Otherwise, the cb might get
472 * freed before we're done setting it up
474 refcount_inc(&cb->pending_bios);
475 ret = btrfs_bio_wq_end_io(fs_info, bio,
476 BTRFS_WQ_ENDIO_DATA);
477 BUG_ON(ret); /* -ENOMEM */
480 ret = btrfs_csum_one_bio(inode, bio, start, 1);
481 BUG_ON(ret); /* -ENOMEM */
484 ret = btrfs_map_bio(fs_info, bio, 0);
486 bio->bi_status = ret;
490 bio = btrfs_bio_alloc(first_byte);
491 bio->bi_opf = REQ_OP_WRITE | write_flags;
492 bio->bi_private = cb;
493 bio->bi_end_io = end_compressed_bio_write;
494 bio_add_page(bio, page, PAGE_SIZE, 0);
496 if (bytes_left < PAGE_SIZE) {
498 "bytes left %lu compress len %lu nr %lu",
499 bytes_left, cb->compressed_len, cb->nr_pages);
501 bytes_left -= PAGE_SIZE;
502 first_byte += PAGE_SIZE;
506 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
507 BUG_ON(ret); /* -ENOMEM */
510 ret = btrfs_csum_one_bio(inode, bio, start, 1);
511 BUG_ON(ret); /* -ENOMEM */
514 ret = btrfs_map_bio(fs_info, bio, 0);
516 bio->bi_status = ret;
523 static u64 bio_end_offset(struct bio *bio)
525 struct bio_vec *last = bio_last_bvec_all(bio);
527 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
530 static noinline int add_ra_bio_pages(struct inode *inode,
532 struct compressed_bio *cb)
534 unsigned long end_index;
535 unsigned long pg_index;
537 u64 isize = i_size_read(inode);
540 unsigned long nr_pages = 0;
541 struct extent_map *em;
542 struct address_space *mapping = inode->i_mapping;
543 struct extent_map_tree *em_tree;
544 struct extent_io_tree *tree;
548 last_offset = bio_end_offset(cb->orig_bio);
549 em_tree = &BTRFS_I(inode)->extent_tree;
550 tree = &BTRFS_I(inode)->io_tree;
555 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
557 while (last_offset < compressed_end) {
558 pg_index = last_offset >> PAGE_SHIFT;
560 if (pg_index > end_index)
563 page = xa_load(&mapping->i_pages, pg_index);
564 if (page && !xa_is_value(page)) {
571 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
576 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
581 end = last_offset + PAGE_SIZE - 1;
583 * at this point, we have a locked page in the page cache
584 * for these bytes in the file. But, we have to make
585 * sure they map to this compressed extent on disk.
587 set_page_extent_mapped(page);
588 lock_extent(tree, last_offset, end);
589 read_lock(&em_tree->lock);
590 em = lookup_extent_mapping(em_tree, last_offset,
592 read_unlock(&em_tree->lock);
594 if (!em || last_offset < em->start ||
595 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
596 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
598 unlock_extent(tree, last_offset, end);
605 if (page->index == end_index) {
607 size_t zero_offset = offset_in_page(isize);
611 zeros = PAGE_SIZE - zero_offset;
612 userpage = kmap_atomic(page);
613 memset(userpage + zero_offset, 0, zeros);
614 flush_dcache_page(page);
615 kunmap_atomic(userpage);
619 ret = bio_add_page(cb->orig_bio, page,
622 if (ret == PAGE_SIZE) {
626 unlock_extent(tree, last_offset, end);
632 last_offset += PAGE_SIZE;
638 * for a compressed read, the bio we get passed has all the inode pages
639 * in it. We don't actually do IO on those pages but allocate new ones
640 * to hold the compressed pages on disk.
642 * bio->bi_iter.bi_sector points to the compressed extent on disk
643 * bio->bi_io_vec points to all of the inode pages
645 * After the compressed pages are read, we copy the bytes into the
646 * bio we were passed and then call the bio end_io calls
648 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
649 int mirror_num, unsigned long bio_flags)
651 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
652 struct extent_map_tree *em_tree;
653 struct compressed_bio *cb;
654 unsigned long compressed_len;
655 unsigned long nr_pages;
656 unsigned long pg_index;
658 struct bio *comp_bio;
659 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
662 struct extent_map *em;
663 blk_status_t ret = BLK_STS_RESOURCE;
665 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
668 em_tree = &BTRFS_I(inode)->extent_tree;
670 /* we need the actual starting offset of this extent in the file */
671 read_lock(&em_tree->lock);
672 em = lookup_extent_mapping(em_tree,
673 page_offset(bio_first_page_all(bio)),
675 read_unlock(&em_tree->lock);
677 return BLK_STS_IOERR;
679 compressed_len = em->block_len;
680 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
684 refcount_set(&cb->pending_bios, 0);
687 cb->mirror_num = mirror_num;
690 cb->start = em->orig_start;
692 em_start = em->start;
697 cb->len = bio->bi_iter.bi_size;
698 cb->compressed_len = compressed_len;
699 cb->compress_type = extent_compress_type(bio_flags);
702 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
703 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
705 if (!cb->compressed_pages)
708 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
709 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
711 if (!cb->compressed_pages[pg_index]) {
712 faili = pg_index - 1;
713 ret = BLK_STS_RESOURCE;
717 faili = nr_pages - 1;
718 cb->nr_pages = nr_pages;
720 add_ra_bio_pages(inode, em_start + em_len, cb);
722 /* include any pages we added in add_ra-bio_pages */
723 cb->len = bio->bi_iter.bi_size;
725 comp_bio = btrfs_bio_alloc(cur_disk_byte);
726 comp_bio->bi_opf = REQ_OP_READ;
727 comp_bio->bi_private = cb;
728 comp_bio->bi_end_io = end_compressed_bio_read;
729 refcount_set(&cb->pending_bios, 1);
731 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
734 page = cb->compressed_pages[pg_index];
735 page->mapping = inode->i_mapping;
736 page->index = em_start >> PAGE_SHIFT;
738 if (comp_bio->bi_iter.bi_size)
739 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
742 page->mapping = NULL;
743 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
745 unsigned int nr_sectors;
747 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
748 BTRFS_WQ_ENDIO_DATA);
749 BUG_ON(ret); /* -ENOMEM */
752 * inc the count before we submit the bio so
753 * we know the end IO handler won't happen before
754 * we inc the count. Otherwise, the cb might get
755 * freed before we're done setting it up
757 refcount_inc(&cb->pending_bios);
759 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
760 ret = btrfs_lookup_bio_sums(inode, comp_bio,
762 BUG_ON(ret); /* -ENOMEM */
765 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
766 fs_info->sectorsize);
767 sums += csum_size * nr_sectors;
769 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
771 comp_bio->bi_status = ret;
775 comp_bio = btrfs_bio_alloc(cur_disk_byte);
776 comp_bio->bi_opf = REQ_OP_READ;
777 comp_bio->bi_private = cb;
778 comp_bio->bi_end_io = end_compressed_bio_read;
780 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
782 cur_disk_byte += PAGE_SIZE;
785 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
786 BUG_ON(ret); /* -ENOMEM */
788 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
789 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
790 BUG_ON(ret); /* -ENOMEM */
793 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
795 comp_bio->bi_status = ret;
803 __free_page(cb->compressed_pages[faili]);
807 kfree(cb->compressed_pages);
816 * Heuristic uses systematic sampling to collect data from the input data
817 * range, the logic can be tuned by the following constants:
819 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
820 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
822 #define SAMPLING_READ_SIZE (16)
823 #define SAMPLING_INTERVAL (256)
826 * For statistical analysis of the input data we consider bytes that form a
827 * Galois Field of 256 objects. Each object has an attribute count, ie. how
828 * many times the object appeared in the sample.
830 #define BUCKET_SIZE (256)
833 * The size of the sample is based on a statistical sampling rule of thumb.
834 * The common way is to perform sampling tests as long as the number of
835 * elements in each cell is at least 5.
837 * Instead of 5, we choose 32 to obtain more accurate results.
838 * If the data contain the maximum number of symbols, which is 256, we obtain a
839 * sample size bound by 8192.
841 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
842 * from up to 512 locations.
844 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
845 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
851 struct heuristic_ws {
852 /* Partial copy of input data */
855 /* Buckets store counters for each byte value */
856 struct bucket_item *bucket;
858 struct bucket_item *bucket_b;
859 struct list_head list;
862 static struct workspace_manager heuristic_wsm;
864 static void free_heuristic_ws(struct list_head *ws)
866 struct heuristic_ws *workspace;
868 workspace = list_entry(ws, struct heuristic_ws, list);
870 kvfree(workspace->sample);
871 kfree(workspace->bucket);
872 kfree(workspace->bucket_b);
876 static struct list_head *alloc_heuristic_ws(unsigned int level)
878 struct heuristic_ws *ws;
880 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
882 return ERR_PTR(-ENOMEM);
884 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
888 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
892 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
896 INIT_LIST_HEAD(&ws->list);
899 free_heuristic_ws(&ws->list);
900 return ERR_PTR(-ENOMEM);
903 const struct btrfs_compress_op btrfs_heuristic_compress = {
904 .workspace_manager = &heuristic_wsm,
907 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
908 /* The heuristic is represented as compression type 0 */
909 &btrfs_heuristic_compress,
910 &btrfs_zlib_compress,
912 &btrfs_zstd_compress,
915 static struct list_head *alloc_workspace(int type, unsigned int level)
918 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
919 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
920 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
921 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
924 * This can't happen, the type is validated several times
925 * before we get here.
931 static void free_workspace(int type, struct list_head *ws)
934 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
935 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
936 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
937 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
940 * This can't happen, the type is validated several times
941 * before we get here.
947 static void btrfs_init_workspace_manager(int type)
949 struct workspace_manager *wsm;
950 struct list_head *workspace;
952 wsm = btrfs_compress_op[type]->workspace_manager;
953 INIT_LIST_HEAD(&wsm->idle_ws);
954 spin_lock_init(&wsm->ws_lock);
955 atomic_set(&wsm->total_ws, 0);
956 init_waitqueue_head(&wsm->ws_wait);
959 * Preallocate one workspace for each compression type so we can
960 * guarantee forward progress in the worst case
962 workspace = alloc_workspace(type, 0);
963 if (IS_ERR(workspace)) {
965 "BTRFS: cannot preallocate compression workspace, will try later\n");
967 atomic_set(&wsm->total_ws, 1);
969 list_add(workspace, &wsm->idle_ws);
973 static void btrfs_cleanup_workspace_manager(int type)
975 struct workspace_manager *wsman;
976 struct list_head *ws;
978 wsman = btrfs_compress_op[type]->workspace_manager;
979 while (!list_empty(&wsman->idle_ws)) {
980 ws = wsman->idle_ws.next;
982 free_workspace(type, ws);
983 atomic_dec(&wsman->total_ws);
988 * This finds an available workspace or allocates a new one.
989 * If it's not possible to allocate a new one, waits until there's one.
990 * Preallocation makes a forward progress guarantees and we do not return
993 struct list_head *btrfs_get_workspace(int type, unsigned int level)
995 struct workspace_manager *wsm;
996 struct list_head *workspace;
997 int cpus = num_online_cpus();
999 struct list_head *idle_ws;
1000 spinlock_t *ws_lock;
1002 wait_queue_head_t *ws_wait;
1005 wsm = btrfs_compress_op[type]->workspace_manager;
1006 idle_ws = &wsm->idle_ws;
1007 ws_lock = &wsm->ws_lock;
1008 total_ws = &wsm->total_ws;
1009 ws_wait = &wsm->ws_wait;
1010 free_ws = &wsm->free_ws;
1014 if (!list_empty(idle_ws)) {
1015 workspace = idle_ws->next;
1016 list_del(workspace);
1018 spin_unlock(ws_lock);
1022 if (atomic_read(total_ws) > cpus) {
1025 spin_unlock(ws_lock);
1026 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1027 if (atomic_read(total_ws) > cpus && !*free_ws)
1029 finish_wait(ws_wait, &wait);
1032 atomic_inc(total_ws);
1033 spin_unlock(ws_lock);
1036 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1037 * to turn it off here because we might get called from the restricted
1038 * context of btrfs_compress_bio/btrfs_compress_pages
1040 nofs_flag = memalloc_nofs_save();
1041 workspace = alloc_workspace(type, level);
1042 memalloc_nofs_restore(nofs_flag);
1044 if (IS_ERR(workspace)) {
1045 atomic_dec(total_ws);
1049 * Do not return the error but go back to waiting. There's a
1050 * workspace preallocated for each type and the compression
1051 * time is bounded so we get to a workspace eventually. This
1052 * makes our caller's life easier.
1054 * To prevent silent and low-probability deadlocks (when the
1055 * initial preallocation fails), check if there are any
1056 * workspaces at all.
1058 if (atomic_read(total_ws) == 0) {
1059 static DEFINE_RATELIMIT_STATE(_rs,
1060 /* once per minute */ 60 * HZ,
1063 if (__ratelimit(&_rs)) {
1064 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1072 static struct list_head *get_workspace(int type, int level)
1075 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1076 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1077 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1078 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1081 * This can't happen, the type is validated several times
1082 * before we get here.
1089 * put a workspace struct back on the list or free it if we have enough
1090 * idle ones sitting around
1092 void btrfs_put_workspace(int type, struct list_head *ws)
1094 struct workspace_manager *wsm;
1095 struct list_head *idle_ws;
1096 spinlock_t *ws_lock;
1098 wait_queue_head_t *ws_wait;
1101 wsm = btrfs_compress_op[type]->workspace_manager;
1102 idle_ws = &wsm->idle_ws;
1103 ws_lock = &wsm->ws_lock;
1104 total_ws = &wsm->total_ws;
1105 ws_wait = &wsm->ws_wait;
1106 free_ws = &wsm->free_ws;
1109 if (*free_ws <= num_online_cpus()) {
1110 list_add(ws, idle_ws);
1112 spin_unlock(ws_lock);
1115 spin_unlock(ws_lock);
1117 free_workspace(type, ws);
1118 atomic_dec(total_ws);
1120 cond_wake_up(ws_wait);
1123 static void put_workspace(int type, struct list_head *ws)
1126 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1127 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1128 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1129 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1132 * This can't happen, the type is validated several times
1133 * before we get here.
1140 * Given an address space and start and length, compress the bytes into @pages
1141 * that are allocated on demand.
1143 * @type_level is encoded algorithm and level, where level 0 means whatever
1144 * default the algorithm chooses and is opaque here;
1145 * - compression algo are 0-3
1146 * - the level are bits 4-7
1148 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1149 * and returns number of actually allocated pages
1151 * @total_in is used to return the number of bytes actually read. It
1152 * may be smaller than the input length if we had to exit early because we
1153 * ran out of room in the pages array or because we cross the
1154 * max_out threshold.
1156 * @total_out is an in/out parameter, must be set to the input length and will
1157 * be also used to return the total number of compressed bytes
1159 * @max_out tells us the max number of bytes that we're allowed to
1162 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1163 u64 start, struct page **pages,
1164 unsigned long *out_pages,
1165 unsigned long *total_in,
1166 unsigned long *total_out)
1168 int type = btrfs_compress_type(type_level);
1169 int level = btrfs_compress_level(type_level);
1170 struct list_head *workspace;
1173 level = btrfs_compress_set_level(type, level);
1174 workspace = get_workspace(type, level);
1175 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1176 out_pages, total_in, total_out);
1177 put_workspace(type, workspace);
1182 * pages_in is an array of pages with compressed data.
1184 * disk_start is the starting logical offset of this array in the file
1186 * orig_bio contains the pages from the file that we want to decompress into
1188 * srclen is the number of bytes in pages_in
1190 * The basic idea is that we have a bio that was created by readpages.
1191 * The pages in the bio are for the uncompressed data, and they may not
1192 * be contiguous. They all correspond to the range of bytes covered by
1193 * the compressed extent.
1195 static int btrfs_decompress_bio(struct compressed_bio *cb)
1197 struct list_head *workspace;
1199 int type = cb->compress_type;
1201 workspace = get_workspace(type, 0);
1202 ret = compression_decompress_bio(type, workspace, cb);
1203 put_workspace(type, workspace);
1209 * a less complex decompression routine. Our compressed data fits in a
1210 * single page, and we want to read a single page out of it.
1211 * start_byte tells us the offset into the compressed data we're interested in
1213 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1214 unsigned long start_byte, size_t srclen, size_t destlen)
1216 struct list_head *workspace;
1219 workspace = get_workspace(type, 0);
1220 ret = compression_decompress(type, workspace, data_in, dest_page,
1221 start_byte, srclen, destlen);
1222 put_workspace(type, workspace);
1227 void __init btrfs_init_compress(void)
1229 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1230 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1231 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1232 zstd_init_workspace_manager();
1235 void __cold btrfs_exit_compress(void)
1237 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1238 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1239 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1240 zstd_cleanup_workspace_manager();
1244 * Copy uncompressed data from working buffer to pages.
1246 * buf_start is the byte offset we're of the start of our workspace buffer.
1248 * total_out is the last byte of the buffer
1250 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1251 unsigned long total_out, u64 disk_start,
1254 unsigned long buf_offset;
1255 unsigned long current_buf_start;
1256 unsigned long start_byte;
1257 unsigned long prev_start_byte;
1258 unsigned long working_bytes = total_out - buf_start;
1259 unsigned long bytes;
1261 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1264 * start byte is the first byte of the page we're currently
1265 * copying into relative to the start of the compressed data.
1267 start_byte = page_offset(bvec.bv_page) - disk_start;
1269 /* we haven't yet hit data corresponding to this page */
1270 if (total_out <= start_byte)
1274 * the start of the data we care about is offset into
1275 * the middle of our working buffer
1277 if (total_out > start_byte && buf_start < start_byte) {
1278 buf_offset = start_byte - buf_start;
1279 working_bytes -= buf_offset;
1283 current_buf_start = buf_start;
1285 /* copy bytes from the working buffer into the pages */
1286 while (working_bytes > 0) {
1287 bytes = min_t(unsigned long, bvec.bv_len,
1288 PAGE_SIZE - buf_offset);
1289 bytes = min(bytes, working_bytes);
1291 kaddr = kmap_atomic(bvec.bv_page);
1292 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1293 kunmap_atomic(kaddr);
1294 flush_dcache_page(bvec.bv_page);
1296 buf_offset += bytes;
1297 working_bytes -= bytes;
1298 current_buf_start += bytes;
1300 /* check if we need to pick another page */
1301 bio_advance(bio, bytes);
1302 if (!bio->bi_iter.bi_size)
1304 bvec = bio_iter_iovec(bio, bio->bi_iter);
1305 prev_start_byte = start_byte;
1306 start_byte = page_offset(bvec.bv_page) - disk_start;
1309 * We need to make sure we're only adjusting
1310 * our offset into compression working buffer when
1311 * we're switching pages. Otherwise we can incorrectly
1312 * keep copying when we were actually done.
1314 if (start_byte != prev_start_byte) {
1316 * make sure our new page is covered by this
1319 if (total_out <= start_byte)
1323 * the next page in the biovec might not be adjacent
1324 * to the last page, but it might still be found
1325 * inside this working buffer. bump our offset pointer
1327 if (total_out > start_byte &&
1328 current_buf_start < start_byte) {
1329 buf_offset = start_byte - buf_start;
1330 working_bytes = total_out - start_byte;
1331 current_buf_start = buf_start + buf_offset;
1340 * Shannon Entropy calculation
1342 * Pure byte distribution analysis fails to determine compressibility of data.
1343 * Try calculating entropy to estimate the average minimum number of bits
1344 * needed to encode the sampled data.
1346 * For convenience, return the percentage of needed bits, instead of amount of
1349 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1350 * and can be compressible with high probability
1352 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1354 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1356 #define ENTROPY_LVL_ACEPTABLE (65)
1357 #define ENTROPY_LVL_HIGH (80)
1360 * For increasead precision in shannon_entropy calculation,
1361 * let's do pow(n, M) to save more digits after comma:
1363 * - maximum int bit length is 64
1364 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1365 * - 13 * 4 = 52 < 64 -> M = 4
1369 static inline u32 ilog2_w(u64 n)
1371 return ilog2(n * n * n * n);
1374 static u32 shannon_entropy(struct heuristic_ws *ws)
1376 const u32 entropy_max = 8 * ilog2_w(2);
1377 u32 entropy_sum = 0;
1378 u32 p, p_base, sz_base;
1381 sz_base = ilog2_w(ws->sample_size);
1382 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1383 p = ws->bucket[i].count;
1384 p_base = ilog2_w(p);
1385 entropy_sum += p * (sz_base - p_base);
1388 entropy_sum /= ws->sample_size;
1389 return entropy_sum * 100 / entropy_max;
1392 #define RADIX_BASE 4U
1393 #define COUNTERS_SIZE (1U << RADIX_BASE)
1395 static u8 get4bits(u64 num, int shift) {
1400 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1405 * Use 4 bits as radix base
1406 * Use 16 u32 counters for calculating new position in buf array
1408 * @array - array that will be sorted
1409 * @array_buf - buffer array to store sorting results
1410 * must be equal in size to @array
1413 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1418 u32 counters[COUNTERS_SIZE];
1426 * Try avoid useless loop iterations for small numbers stored in big
1427 * counters. Example: 48 33 4 ... in 64bit array
1429 max_num = array[0].count;
1430 for (i = 1; i < num; i++) {
1431 buf_num = array[i].count;
1432 if (buf_num > max_num)
1436 buf_num = ilog2(max_num);
1437 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1440 while (shift < bitlen) {
1441 memset(counters, 0, sizeof(counters));
1443 for (i = 0; i < num; i++) {
1444 buf_num = array[i].count;
1445 addr = get4bits(buf_num, shift);
1449 for (i = 1; i < COUNTERS_SIZE; i++)
1450 counters[i] += counters[i - 1];
1452 for (i = num - 1; i >= 0; i--) {
1453 buf_num = array[i].count;
1454 addr = get4bits(buf_num, shift);
1456 new_addr = counters[addr];
1457 array_buf[new_addr] = array[i];
1460 shift += RADIX_BASE;
1463 * Normal radix expects to move data from a temporary array, to
1464 * the main one. But that requires some CPU time. Avoid that
1465 * by doing another sort iteration to original array instead of
1468 memset(counters, 0, sizeof(counters));
1470 for (i = 0; i < num; i ++) {
1471 buf_num = array_buf[i].count;
1472 addr = get4bits(buf_num, shift);
1476 for (i = 1; i < COUNTERS_SIZE; i++)
1477 counters[i] += counters[i - 1];
1479 for (i = num - 1; i >= 0; i--) {
1480 buf_num = array_buf[i].count;
1481 addr = get4bits(buf_num, shift);
1483 new_addr = counters[addr];
1484 array[new_addr] = array_buf[i];
1487 shift += RADIX_BASE;
1492 * Size of the core byte set - how many bytes cover 90% of the sample
1494 * There are several types of structured binary data that use nearly all byte
1495 * values. The distribution can be uniform and counts in all buckets will be
1496 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1498 * Other possibility is normal (Gaussian) distribution, where the data could
1499 * be potentially compressible, but we have to take a few more steps to decide
1502 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1503 * compression algo can easy fix that
1504 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1505 * probability is not compressible
1507 #define BYTE_CORE_SET_LOW (64)
1508 #define BYTE_CORE_SET_HIGH (200)
1510 static int byte_core_set_size(struct heuristic_ws *ws)
1513 u32 coreset_sum = 0;
1514 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1515 struct bucket_item *bucket = ws->bucket;
1517 /* Sort in reverse order */
1518 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1520 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1521 coreset_sum += bucket[i].count;
1523 if (coreset_sum > core_set_threshold)
1526 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1527 coreset_sum += bucket[i].count;
1528 if (coreset_sum > core_set_threshold)
1536 * Count byte values in buckets.
1537 * This heuristic can detect textual data (configs, xml, json, html, etc).
1538 * Because in most text-like data byte set is restricted to limited number of
1539 * possible characters, and that restriction in most cases makes data easy to
1542 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1543 * less - compressible
1544 * more - need additional analysis
1546 #define BYTE_SET_THRESHOLD (64)
1548 static u32 byte_set_size(const struct heuristic_ws *ws)
1551 u32 byte_set_size = 0;
1553 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1554 if (ws->bucket[i].count > 0)
1559 * Continue collecting count of byte values in buckets. If the byte
1560 * set size is bigger then the threshold, it's pointless to continue,
1561 * the detection technique would fail for this type of data.
1563 for (; i < BUCKET_SIZE; i++) {
1564 if (ws->bucket[i].count > 0) {
1566 if (byte_set_size > BYTE_SET_THRESHOLD)
1567 return byte_set_size;
1571 return byte_set_size;
1574 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1576 const u32 half_of_sample = ws->sample_size / 2;
1577 const u8 *data = ws->sample;
1579 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1582 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1583 struct heuristic_ws *ws)
1586 u64 index, index_end;
1587 u32 i, curr_sample_pos;
1591 * Compression handles the input data by chunks of 128KiB
1592 * (defined by BTRFS_MAX_UNCOMPRESSED)
1594 * We do the same for the heuristic and loop over the whole range.
1596 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1597 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1599 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1600 end = start + BTRFS_MAX_UNCOMPRESSED;
1602 index = start >> PAGE_SHIFT;
1603 index_end = end >> PAGE_SHIFT;
1605 /* Don't miss unaligned end */
1606 if (!IS_ALIGNED(end, PAGE_SIZE))
1609 curr_sample_pos = 0;
1610 while (index < index_end) {
1611 page = find_get_page(inode->i_mapping, index);
1612 in_data = kmap(page);
1613 /* Handle case where the start is not aligned to PAGE_SIZE */
1614 i = start % PAGE_SIZE;
1615 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1616 /* Don't sample any garbage from the last page */
1617 if (start > end - SAMPLING_READ_SIZE)
1619 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1620 SAMPLING_READ_SIZE);
1621 i += SAMPLING_INTERVAL;
1622 start += SAMPLING_INTERVAL;
1623 curr_sample_pos += SAMPLING_READ_SIZE;
1631 ws->sample_size = curr_sample_pos;
1635 * Compression heuristic.
1637 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1638 * quickly (compared to direct compression) detect data characteristics
1639 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1642 * The following types of analysis can be performed:
1643 * - detect mostly zero data
1644 * - detect data with low "byte set" size (text, etc)
1645 * - detect data with low/high "core byte" set
1647 * Return non-zero if the compression should be done, 0 otherwise.
1649 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1651 struct list_head *ws_list = get_workspace(0, 0);
1652 struct heuristic_ws *ws;
1657 ws = list_entry(ws_list, struct heuristic_ws, list);
1659 heuristic_collect_sample(inode, start, end, ws);
1661 if (sample_repeated_patterns(ws)) {
1666 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1668 for (i = 0; i < ws->sample_size; i++) {
1669 byte = ws->sample[i];
1670 ws->bucket[byte].count++;
1673 i = byte_set_size(ws);
1674 if (i < BYTE_SET_THRESHOLD) {
1679 i = byte_core_set_size(ws);
1680 if (i <= BYTE_CORE_SET_LOW) {
1685 if (i >= BYTE_CORE_SET_HIGH) {
1690 i = shannon_entropy(ws);
1691 if (i <= ENTROPY_LVL_ACEPTABLE) {
1697 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1698 * needed to give green light to compression.
1700 * For now just assume that compression at that level is not worth the
1701 * resources because:
1703 * 1. it is possible to defrag the data later
1705 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1706 * values, every bucket has counter at level ~54. The heuristic would
1707 * be confused. This can happen when data have some internal repeated
1708 * patterns like "abbacbbc...". This can be detected by analyzing
1709 * pairs of bytes, which is too costly.
1711 if (i < ENTROPY_LVL_HIGH) {
1720 put_workspace(0, ws_list);
1725 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1726 * level, unrecognized string will set the default level
1728 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1730 unsigned int level = 0;
1736 if (str[0] == ':') {
1737 ret = kstrtouint(str + 1, 10, &level);
1742 level = btrfs_compress_set_level(type, level);
1748 * Adjust @level according to the limits of the compression algorithm or
1749 * fallback to default
1751 unsigned int btrfs_compress_set_level(int type, unsigned level)
1753 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1756 level = ops->default_level;
1758 level = min(level, ops->max_level);