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"
34 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
36 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
39 case BTRFS_COMPRESS_ZLIB:
40 case BTRFS_COMPRESS_LZO:
41 case BTRFS_COMPRESS_ZSTD:
42 case BTRFS_COMPRESS_NONE:
43 return btrfs_compress_types[type];
51 bool btrfs_compress_is_valid_type(const char *str, size_t len)
55 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
56 size_t comp_len = strlen(btrfs_compress_types[i]);
61 if (!strncmp(btrfs_compress_types[i], str, comp_len))
67 static int compression_compress_pages(int type, struct list_head *ws,
68 struct address_space *mapping, u64 start, struct page **pages,
69 unsigned long *out_pages, unsigned long *total_in,
70 unsigned long *total_out)
73 case BTRFS_COMPRESS_ZLIB:
74 return zlib_compress_pages(ws, mapping, start, pages,
75 out_pages, total_in, total_out);
76 case BTRFS_COMPRESS_LZO:
77 return lzo_compress_pages(ws, mapping, start, pages,
78 out_pages, total_in, total_out);
79 case BTRFS_COMPRESS_ZSTD:
80 return zstd_compress_pages(ws, mapping, start, pages,
81 out_pages, total_in, total_out);
82 case BTRFS_COMPRESS_NONE:
85 * This can happen when compression races with remount setting
86 * it to 'no compress', while caller doesn't call
87 * inode_need_compress() to check if we really need to
90 * Not a big deal, just need to inform caller that we
91 * haven't allocated any pages yet.
98 static int compression_decompress_bio(int type, struct list_head *ws,
99 struct compressed_bio *cb)
102 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
103 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
104 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
105 case BTRFS_COMPRESS_NONE:
108 * This can't happen, the type is validated several times
109 * before we get here.
115 static int compression_decompress(int type, struct list_head *ws,
116 unsigned char *data_in, struct page *dest_page,
117 unsigned long start_byte, size_t srclen, size_t destlen)
120 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
121 start_byte, srclen, destlen);
122 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
123 start_byte, srclen, destlen);
124 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
125 start_byte, srclen, destlen);
126 case BTRFS_COMPRESS_NONE:
129 * This can't happen, the type is validated several times
130 * before we get here.
136 static int btrfs_decompress_bio(struct compressed_bio *cb);
138 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
139 unsigned long disk_size)
141 return sizeof(struct compressed_bio) +
142 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size;
145 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
148 struct btrfs_fs_info *fs_info = inode->root->fs_info;
149 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
150 const u32 csum_size = fs_info->csum_size;
151 const u32 sectorsize = fs_info->sectorsize;
155 u8 csum[BTRFS_CSUM_SIZE];
156 struct compressed_bio *cb = bio->bi_private;
157 u8 *cb_sum = cb->sums;
159 if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM))
162 shash->tfm = fs_info->csum_shash;
164 for (i = 0; i < cb->nr_pages; i++) {
166 u32 bytes_left = PAGE_SIZE;
167 page = cb->compressed_pages[i];
169 /* Determine the remaining bytes inside the page first */
170 if (i == cb->nr_pages - 1)
171 bytes_left = cb->compressed_len - i * PAGE_SIZE;
173 /* Hash through the page sector by sector */
174 for (pg_offset = 0; pg_offset < bytes_left;
175 pg_offset += sectorsize) {
176 kaddr = page_address(page);
177 crypto_shash_digest(shash, kaddr + pg_offset,
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_bio(bio)->device)
184 btrfs_dev_stat_inc_and_print(
185 btrfs_bio(bio)->device,
186 BTRFS_DEV_STAT_CORRUPTION_ERRS);
190 disk_start += sectorsize;
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).
200 * This allows the checksumming and other IO error handling routines
203 * The compressed pages are freed here, and it must be run
206 static void end_compressed_bio_read(struct bio *bio)
208 struct compressed_bio *cb = bio->bi_private;
212 unsigned int mirror = btrfs_bio(bio)->mirror_num;
218 /* if there are more bios still pending for this compressed
221 if (!refcount_dec_and_test(&cb->pending_bios))
225 * Record the correct mirror_num in cb->orig_bio so that
226 * read-repair can work properly.
228 btrfs_bio(cb->orig_bio)->mirror_num = mirror;
229 cb->mirror_num = mirror;
232 * Some IO in this cb have failed, just skip checksum as there
233 * is no way it could be correct.
239 ret = check_compressed_csum(BTRFS_I(inode), bio,
240 bio->bi_iter.bi_sector << 9);
244 /* ok, we're the last bio for this extent, lets start
247 ret = btrfs_decompress_bio(cb);
253 /* release the compressed pages */
255 for (index = 0; index < cb->nr_pages; index++) {
256 page = cb->compressed_pages[index];
257 page->mapping = NULL;
261 /* do io completion on the original bio */
263 bio_io_error(cb->orig_bio);
265 struct bio_vec *bvec;
266 struct bvec_iter_all iter_all;
269 * we have verified the checksum already, set page
270 * checked so the end_io handlers know about it
272 ASSERT(!bio_flagged(bio, BIO_CLONED));
273 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
274 SetPageChecked(bvec->bv_page);
276 bio_endio(cb->orig_bio);
279 /* finally free the cb struct */
280 kfree(cb->compressed_pages);
287 * Clear the writeback bits on all of the file
288 * pages for a compressed write
290 static noinline void end_compressed_writeback(struct inode *inode,
291 const struct compressed_bio *cb)
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;
301 mapping_set_error(inode->i_mapping, -EIO);
303 while (nr_pages > 0) {
304 ret = find_get_pages_contig(inode->i_mapping, index,
306 nr_pages, ARRAY_SIZE(pages)), pages);
312 for (i = 0; i < ret; i++) {
314 SetPageError(pages[i]);
315 end_page_writeback(pages[i]);
321 /* the inode may be gone now */
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
329 * This also calls the writeback end hooks for the file pages so that
330 * metadata and checksums can be updated in the file.
332 static void end_compressed_bio_write(struct bio *bio)
334 struct compressed_bio *cb = bio->bi_private;
342 /* if there are more bios still pending for this compressed
345 if (!refcount_dec_and_test(&cb->pending_bios))
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
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,
357 end_compressed_writeback(inode, cb);
358 /* note, our inode could be gone now */
361 * release the compressed pages, these came from alloc_page and
362 * are not attached to the inode at all
365 for (index = 0; index < cb->nr_pages; index++) {
366 page = cb->compressed_pages[index];
367 page->mapping = NULL;
371 /* finally free the cb struct */
372 kfree(cb->compressed_pages);
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.
384 * This also checksums the file bytes and gets things ready for
387 blk_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)
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;
401 u64 first_byte = disk_start;
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;
407 WARN_ON(!PAGE_ALIGNED(start));
408 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
410 return BLK_STS_RESOURCE;
411 refcount_set(&cb->pending_bios, 0);
413 cb->inode = &inode->vfs_inode;
417 cb->compressed_pages = compressed_pages;
418 cb->compressed_len = compressed_len;
420 cb->nr_pages = nr_pages;
422 bio = btrfs_bio_alloc(BIO_MAX_VECS);
423 bio->bi_iter.bi_sector = first_byte >> SECTOR_SHIFT;
424 bio->bi_opf = bio_op | write_flags;
425 bio->bi_private = cb;
426 bio->bi_end_io = end_compressed_bio_write;
429 struct btrfs_device *device;
431 device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE);
432 if (IS_ERR(device)) {
435 return BLK_STS_NOTSUPP;
438 bio_set_dev(bio, device->bdev);
442 bio->bi_opf |= REQ_CGROUP_PUNT;
443 kthread_associate_blkcg(blkcg_css);
445 refcount_set(&cb->pending_bios, 1);
447 /* create and submit bios for the compressed pages */
448 bytes_left = compressed_len;
449 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
453 page = compressed_pages[pg_index];
454 page->mapping = inode->vfs_inode.i_mapping;
455 if (bio->bi_iter.bi_size)
456 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
460 * Page can only be added to bio if the current bio fits in
464 if (pg_index == 0 && use_append)
465 len = bio_add_zone_append_page(bio, page,
468 len = bio_add_page(bio, page, PAGE_SIZE, 0);
471 page->mapping = NULL;
472 if (submit || len < PAGE_SIZE) {
474 * inc the count before we submit the bio so
475 * we know the end IO handler won't happen before
476 * we inc the count. Otherwise, the cb might get
477 * freed before we're done setting it up
479 refcount_inc(&cb->pending_bios);
480 ret = btrfs_bio_wq_end_io(fs_info, bio,
481 BTRFS_WQ_ENDIO_DATA);
482 BUG_ON(ret); /* -ENOMEM */
485 ret = btrfs_csum_one_bio(inode, bio, start, 1);
486 BUG_ON(ret); /* -ENOMEM */
489 ret = btrfs_map_bio(fs_info, bio, 0);
491 bio->bi_status = ret;
495 bio = btrfs_bio_alloc(BIO_MAX_VECS);
496 bio->bi_iter.bi_sector = first_byte >> SECTOR_SHIFT;
497 bio->bi_opf = bio_op | write_flags;
498 bio->bi_private = cb;
499 bio->bi_end_io = end_compressed_bio_write;
501 bio->bi_opf |= REQ_CGROUP_PUNT;
503 * Use bio_add_page() to ensure the bio has at least one
506 bio_add_page(bio, page, PAGE_SIZE, 0);
508 if (bytes_left < PAGE_SIZE) {
510 "bytes left %lu compress len %u nr %u",
511 bytes_left, cb->compressed_len, cb->nr_pages);
513 bytes_left -= PAGE_SIZE;
514 first_byte += PAGE_SIZE;
518 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
519 BUG_ON(ret); /* -ENOMEM */
522 ret = btrfs_csum_one_bio(inode, bio, start, 1);
523 BUG_ON(ret); /* -ENOMEM */
526 ret = btrfs_map_bio(fs_info, bio, 0);
528 bio->bi_status = ret;
533 kthread_associate_blkcg(NULL);
538 static u64 bio_end_offset(struct bio *bio)
540 struct bio_vec *last = bio_last_bvec_all(bio);
542 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
546 * Add extra pages in the same compressed file extent so that we don't need to
547 * re-read the same extent again and again.
549 * NOTE: this won't work well for subpage, as for subpage read, we lock the
550 * full page then submit bio for each compressed/regular extents.
552 * This means, if we have several sectors in the same page points to the same
553 * on-disk compressed data, we will re-read the same extent many times and
554 * this function can only help for the next page.
556 static noinline int add_ra_bio_pages(struct inode *inode,
558 struct compressed_bio *cb)
560 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
561 unsigned long end_index;
562 u64 cur = bio_end_offset(cb->orig_bio);
563 u64 isize = i_size_read(inode);
566 struct extent_map *em;
567 struct address_space *mapping = inode->i_mapping;
568 struct extent_map_tree *em_tree;
569 struct extent_io_tree *tree;
570 int sectors_missed = 0;
572 em_tree = &BTRFS_I(inode)->extent_tree;
573 tree = &BTRFS_I(inode)->io_tree;
579 * For current subpage support, we only support 64K page size,
580 * which means maximum compressed extent size (128K) is just 2x page
582 * This makes readahead less effective, so here disable readahead for
583 * subpage for now, until full compressed write is supported.
585 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
588 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
590 while (cur < compressed_end) {
592 u64 pg_index = cur >> PAGE_SHIFT;
595 if (pg_index > end_index)
598 page = xa_load(&mapping->i_pages, pg_index);
599 if (page && !xa_is_value(page)) {
600 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
601 fs_info->sectorsize_bits;
603 /* Beyond threshold, no need to continue */
604 if (sectors_missed > 4)
608 * Jump to next page start as we already have page for
611 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
615 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
620 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
622 /* There is already a page, skip to page end */
623 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
627 ret = set_page_extent_mapped(page);
634 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
635 lock_extent(tree, cur, page_end);
636 read_lock(&em_tree->lock);
637 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
638 read_unlock(&em_tree->lock);
641 * At this point, we have a locked page in the page cache for
642 * these bytes in the file. But, we have to make sure they map
643 * to this compressed extent on disk.
645 if (!em || cur < em->start ||
646 (cur + fs_info->sectorsize > extent_map_end(em)) ||
647 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
649 unlock_extent(tree, cur, page_end);
656 if (page->index == end_index) {
657 size_t zero_offset = offset_in_page(isize);
661 zeros = PAGE_SIZE - zero_offset;
662 memzero_page(page, zero_offset, zeros);
663 flush_dcache_page(page);
667 add_size = min(em->start + em->len, page_end + 1) - cur;
668 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
669 if (ret != add_size) {
670 unlock_extent(tree, cur, page_end);
676 * If it's subpage, we also need to increase its
677 * subpage::readers number, as at endio we will decrease
678 * subpage::readers and to unlock the page.
680 if (fs_info->sectorsize < PAGE_SIZE)
681 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
689 * for a compressed read, the bio we get passed has all the inode pages
690 * in it. We don't actually do IO on those pages but allocate new ones
691 * to hold the compressed pages on disk.
693 * bio->bi_iter.bi_sector points to the compressed extent on disk
694 * bio->bi_io_vec points to all of the inode pages
696 * After the compressed pages are read, we copy the bytes into the
697 * bio we were passed and then call the bio end_io calls
699 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
700 int mirror_num, unsigned long bio_flags)
702 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
703 struct extent_map_tree *em_tree;
704 struct compressed_bio *cb;
705 unsigned int compressed_len;
706 unsigned int nr_pages;
707 unsigned int pg_index;
709 struct bio *comp_bio;
710 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
714 struct extent_map *em;
715 blk_status_t ret = BLK_STS_RESOURCE;
719 em_tree = &BTRFS_I(inode)->extent_tree;
721 file_offset = bio_first_bvec_all(bio)->bv_offset +
722 page_offset(bio_first_page_all(bio));
724 /* we need the actual starting offset of this extent in the file */
725 read_lock(&em_tree->lock);
726 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
727 read_unlock(&em_tree->lock);
729 return BLK_STS_IOERR;
731 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
732 compressed_len = em->block_len;
733 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
737 refcount_set(&cb->pending_bios, 0);
740 cb->mirror_num = mirror_num;
743 cb->start = em->orig_start;
745 em_start = em->start;
750 cb->len = bio->bi_iter.bi_size;
751 cb->compressed_len = compressed_len;
752 cb->compress_type = extent_compress_type(bio_flags);
755 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
756 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
758 if (!cb->compressed_pages)
761 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
762 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS);
763 if (!cb->compressed_pages[pg_index]) {
764 faili = pg_index - 1;
765 ret = BLK_STS_RESOURCE;
769 faili = nr_pages - 1;
770 cb->nr_pages = nr_pages;
772 add_ra_bio_pages(inode, em_start + em_len, cb);
774 /* include any pages we added in add_ra-bio_pages */
775 cb->len = bio->bi_iter.bi_size;
777 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS);
778 comp_bio->bi_iter.bi_sector = cur_disk_byte >> SECTOR_SHIFT;
779 comp_bio->bi_opf = REQ_OP_READ;
780 comp_bio->bi_private = cb;
781 comp_bio->bi_end_io = end_compressed_bio_read;
782 refcount_set(&cb->pending_bios, 1);
784 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
785 u32 pg_len = PAGE_SIZE;
789 * To handle subpage case, we need to make sure the bio only
790 * covers the range we need.
792 * If we're at the last page, truncate the length to only cover
793 * the remaining part.
795 if (pg_index == nr_pages - 1)
796 pg_len = min_t(u32, PAGE_SIZE,
797 compressed_len - pg_index * PAGE_SIZE);
799 page = cb->compressed_pages[pg_index];
800 page->mapping = inode->i_mapping;
801 page->index = em_start >> PAGE_SHIFT;
803 if (comp_bio->bi_iter.bi_size)
804 submit = btrfs_bio_fits_in_stripe(page, pg_len,
807 page->mapping = NULL;
808 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
809 unsigned int nr_sectors;
811 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
812 BTRFS_WQ_ENDIO_DATA);
813 BUG_ON(ret); /* -ENOMEM */
816 * inc the count before we submit the bio so
817 * we know the end IO handler won't happen before
818 * we inc the count. Otherwise, the cb might get
819 * freed before we're done setting it up
821 refcount_inc(&cb->pending_bios);
823 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
824 BUG_ON(ret); /* -ENOMEM */
826 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
827 fs_info->sectorsize);
828 sums += fs_info->csum_size * nr_sectors;
830 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
832 comp_bio->bi_status = ret;
836 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS);
837 comp_bio->bi_iter.bi_sector = cur_disk_byte >> SECTOR_SHIFT;
838 comp_bio->bi_opf = REQ_OP_READ;
839 comp_bio->bi_private = cb;
840 comp_bio->bi_end_io = end_compressed_bio_read;
842 bio_add_page(comp_bio, page, pg_len, 0);
844 cur_disk_byte += pg_len;
847 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
848 BUG_ON(ret); /* -ENOMEM */
850 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
851 BUG_ON(ret); /* -ENOMEM */
853 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
855 comp_bio->bi_status = ret;
863 __free_page(cb->compressed_pages[faili]);
867 kfree(cb->compressed_pages);
876 * Heuristic uses systematic sampling to collect data from the input data
877 * range, the logic can be tuned by the following constants:
879 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
880 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
882 #define SAMPLING_READ_SIZE (16)
883 #define SAMPLING_INTERVAL (256)
886 * For statistical analysis of the input data we consider bytes that form a
887 * Galois Field of 256 objects. Each object has an attribute count, ie. how
888 * many times the object appeared in the sample.
890 #define BUCKET_SIZE (256)
893 * The size of the sample is based on a statistical sampling rule of thumb.
894 * The common way is to perform sampling tests as long as the number of
895 * elements in each cell is at least 5.
897 * Instead of 5, we choose 32 to obtain more accurate results.
898 * If the data contain the maximum number of symbols, which is 256, we obtain a
899 * sample size bound by 8192.
901 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
902 * from up to 512 locations.
904 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
905 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
911 struct heuristic_ws {
912 /* Partial copy of input data */
915 /* Buckets store counters for each byte value */
916 struct bucket_item *bucket;
918 struct bucket_item *bucket_b;
919 struct list_head list;
922 static struct workspace_manager heuristic_wsm;
924 static void free_heuristic_ws(struct list_head *ws)
926 struct heuristic_ws *workspace;
928 workspace = list_entry(ws, struct heuristic_ws, list);
930 kvfree(workspace->sample);
931 kfree(workspace->bucket);
932 kfree(workspace->bucket_b);
936 static struct list_head *alloc_heuristic_ws(unsigned int level)
938 struct heuristic_ws *ws;
940 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
942 return ERR_PTR(-ENOMEM);
944 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
948 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
952 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
956 INIT_LIST_HEAD(&ws->list);
959 free_heuristic_ws(&ws->list);
960 return ERR_PTR(-ENOMEM);
963 const struct btrfs_compress_op btrfs_heuristic_compress = {
964 .workspace_manager = &heuristic_wsm,
967 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
968 /* The heuristic is represented as compression type 0 */
969 &btrfs_heuristic_compress,
970 &btrfs_zlib_compress,
972 &btrfs_zstd_compress,
975 static struct list_head *alloc_workspace(int type, unsigned int level)
978 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
979 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
980 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
981 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
984 * This can't happen, the type is validated several times
985 * before we get here.
991 static void free_workspace(int type, struct list_head *ws)
994 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
995 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
996 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
997 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
1000 * This can't happen, the type is validated several times
1001 * before we get here.
1007 static void btrfs_init_workspace_manager(int type)
1009 struct workspace_manager *wsm;
1010 struct list_head *workspace;
1012 wsm = btrfs_compress_op[type]->workspace_manager;
1013 INIT_LIST_HEAD(&wsm->idle_ws);
1014 spin_lock_init(&wsm->ws_lock);
1015 atomic_set(&wsm->total_ws, 0);
1016 init_waitqueue_head(&wsm->ws_wait);
1019 * Preallocate one workspace for each compression type so we can
1020 * guarantee forward progress in the worst case
1022 workspace = alloc_workspace(type, 0);
1023 if (IS_ERR(workspace)) {
1025 "BTRFS: cannot preallocate compression workspace, will try later\n");
1027 atomic_set(&wsm->total_ws, 1);
1029 list_add(workspace, &wsm->idle_ws);
1033 static void btrfs_cleanup_workspace_manager(int type)
1035 struct workspace_manager *wsman;
1036 struct list_head *ws;
1038 wsman = btrfs_compress_op[type]->workspace_manager;
1039 while (!list_empty(&wsman->idle_ws)) {
1040 ws = wsman->idle_ws.next;
1042 free_workspace(type, ws);
1043 atomic_dec(&wsman->total_ws);
1048 * This finds an available workspace or allocates a new one.
1049 * If it's not possible to allocate a new one, waits until there's one.
1050 * Preallocation makes a forward progress guarantees and we do not return
1053 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1055 struct workspace_manager *wsm;
1056 struct list_head *workspace;
1057 int cpus = num_online_cpus();
1059 struct list_head *idle_ws;
1060 spinlock_t *ws_lock;
1062 wait_queue_head_t *ws_wait;
1065 wsm = btrfs_compress_op[type]->workspace_manager;
1066 idle_ws = &wsm->idle_ws;
1067 ws_lock = &wsm->ws_lock;
1068 total_ws = &wsm->total_ws;
1069 ws_wait = &wsm->ws_wait;
1070 free_ws = &wsm->free_ws;
1074 if (!list_empty(idle_ws)) {
1075 workspace = idle_ws->next;
1076 list_del(workspace);
1078 spin_unlock(ws_lock);
1082 if (atomic_read(total_ws) > cpus) {
1085 spin_unlock(ws_lock);
1086 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1087 if (atomic_read(total_ws) > cpus && !*free_ws)
1089 finish_wait(ws_wait, &wait);
1092 atomic_inc(total_ws);
1093 spin_unlock(ws_lock);
1096 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1097 * to turn it off here because we might get called from the restricted
1098 * context of btrfs_compress_bio/btrfs_compress_pages
1100 nofs_flag = memalloc_nofs_save();
1101 workspace = alloc_workspace(type, level);
1102 memalloc_nofs_restore(nofs_flag);
1104 if (IS_ERR(workspace)) {
1105 atomic_dec(total_ws);
1109 * Do not return the error but go back to waiting. There's a
1110 * workspace preallocated for each type and the compression
1111 * time is bounded so we get to a workspace eventually. This
1112 * makes our caller's life easier.
1114 * To prevent silent and low-probability deadlocks (when the
1115 * initial preallocation fails), check if there are any
1116 * workspaces at all.
1118 if (atomic_read(total_ws) == 0) {
1119 static DEFINE_RATELIMIT_STATE(_rs,
1120 /* once per minute */ 60 * HZ,
1123 if (__ratelimit(&_rs)) {
1124 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1132 static struct list_head *get_workspace(int type, int level)
1135 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1136 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1137 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1138 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1141 * This can't happen, the type is validated several times
1142 * before we get here.
1149 * put a workspace struct back on the list or free it if we have enough
1150 * idle ones sitting around
1152 void btrfs_put_workspace(int type, struct list_head *ws)
1154 struct workspace_manager *wsm;
1155 struct list_head *idle_ws;
1156 spinlock_t *ws_lock;
1158 wait_queue_head_t *ws_wait;
1161 wsm = btrfs_compress_op[type]->workspace_manager;
1162 idle_ws = &wsm->idle_ws;
1163 ws_lock = &wsm->ws_lock;
1164 total_ws = &wsm->total_ws;
1165 ws_wait = &wsm->ws_wait;
1166 free_ws = &wsm->free_ws;
1169 if (*free_ws <= num_online_cpus()) {
1170 list_add(ws, idle_ws);
1172 spin_unlock(ws_lock);
1175 spin_unlock(ws_lock);
1177 free_workspace(type, ws);
1178 atomic_dec(total_ws);
1180 cond_wake_up(ws_wait);
1183 static void put_workspace(int type, struct list_head *ws)
1186 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1187 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1188 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1189 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1192 * This can't happen, the type is validated several times
1193 * before we get here.
1200 * Adjust @level according to the limits of the compression algorithm or
1201 * fallback to default
1203 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1205 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1208 level = ops->default_level;
1210 level = min(level, ops->max_level);
1216 * Given an address space and start and length, compress the bytes into @pages
1217 * that are allocated on demand.
1219 * @type_level is encoded algorithm and level, where level 0 means whatever
1220 * default the algorithm chooses and is opaque here;
1221 * - compression algo are 0-3
1222 * - the level are bits 4-7
1224 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1225 * and returns number of actually allocated pages
1227 * @total_in is used to return the number of bytes actually read. It
1228 * may be smaller than the input length if we had to exit early because we
1229 * ran out of room in the pages array or because we cross the
1230 * max_out threshold.
1232 * @total_out is an in/out parameter, must be set to the input length and will
1233 * be also used to return the total number of compressed bytes
1235 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1236 u64 start, struct page **pages,
1237 unsigned long *out_pages,
1238 unsigned long *total_in,
1239 unsigned long *total_out)
1241 int type = btrfs_compress_type(type_level);
1242 int level = btrfs_compress_level(type_level);
1243 struct list_head *workspace;
1246 level = btrfs_compress_set_level(type, level);
1247 workspace = get_workspace(type, level);
1248 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1249 out_pages, total_in, total_out);
1250 put_workspace(type, workspace);
1254 static int btrfs_decompress_bio(struct compressed_bio *cb)
1256 struct list_head *workspace;
1258 int type = cb->compress_type;
1260 workspace = get_workspace(type, 0);
1261 ret = compression_decompress_bio(type, workspace, cb);
1262 put_workspace(type, workspace);
1268 * a less complex decompression routine. Our compressed data fits in a
1269 * single page, and we want to read a single page out of it.
1270 * start_byte tells us the offset into the compressed data we're interested in
1272 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1273 unsigned long start_byte, size_t srclen, size_t destlen)
1275 struct list_head *workspace;
1278 workspace = get_workspace(type, 0);
1279 ret = compression_decompress(type, workspace, data_in, dest_page,
1280 start_byte, srclen, destlen);
1281 put_workspace(type, workspace);
1286 void __init btrfs_init_compress(void)
1288 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1289 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1290 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1291 zstd_init_workspace_manager();
1294 void __cold btrfs_exit_compress(void)
1296 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1297 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1298 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1299 zstd_cleanup_workspace_manager();
1303 * Copy decompressed data from working buffer to pages.
1305 * @buf: The decompressed data buffer
1306 * @buf_len: The decompressed data length
1307 * @decompressed: Number of bytes that are already decompressed inside the
1309 * @cb: The compressed extent descriptor
1310 * @orig_bio: The original bio that the caller wants to read for
1312 * An easier to understand graph is like below:
1314 * |<- orig_bio ->| |<- orig_bio->|
1315 * |<------- full decompressed extent ----->|
1316 * |<----------- @cb range ---->|
1317 * | |<-- @buf_len -->|
1318 * |<--- @decompressed --->|
1320 * Note that, @cb can be a subpage of the full decompressed extent, but
1321 * @cb->start always has the same as the orig_file_offset value of the full
1322 * decompressed extent.
1324 * When reading compressed extent, we have to read the full compressed extent,
1325 * while @orig_bio may only want part of the range.
1326 * Thus this function will ensure only data covered by @orig_bio will be copied
1329 * Return 0 if we have copied all needed contents for @orig_bio.
1330 * Return >0 if we need continue decompress.
1332 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1333 struct compressed_bio *cb, u32 decompressed)
1335 struct bio *orig_bio = cb->orig_bio;
1336 /* Offset inside the full decompressed extent */
1339 cur_offset = decompressed;
1340 /* The main loop to do the copy */
1341 while (cur_offset < decompressed + buf_len) {
1342 struct bio_vec bvec;
1345 /* Offset inside the full decompressed extent */
1348 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1350 * cb->start may underflow, but subtracting that value can still
1351 * give us correct offset inside the full decompressed extent.
1353 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1355 /* Haven't reached the bvec range, exit */
1356 if (decompressed + buf_len <= bvec_offset)
1359 copy_start = max(cur_offset, bvec_offset);
1360 copy_len = min(bvec_offset + bvec.bv_len,
1361 decompressed + buf_len) - copy_start;
1365 * Extra range check to ensure we didn't go beyond
1368 ASSERT(copy_start - decompressed < buf_len);
1369 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1370 buf + copy_start - decompressed, copy_len);
1371 flush_dcache_page(bvec.bv_page);
1372 cur_offset += copy_len;
1374 bio_advance(orig_bio, copy_len);
1375 /* Finished the bio */
1376 if (!orig_bio->bi_iter.bi_size)
1383 * Shannon Entropy calculation
1385 * Pure byte distribution analysis fails to determine compressibility of data.
1386 * Try calculating entropy to estimate the average minimum number of bits
1387 * needed to encode the sampled data.
1389 * For convenience, return the percentage of needed bits, instead of amount of
1392 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1393 * and can be compressible with high probability
1395 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1397 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1399 #define ENTROPY_LVL_ACEPTABLE (65)
1400 #define ENTROPY_LVL_HIGH (80)
1403 * For increasead precision in shannon_entropy calculation,
1404 * let's do pow(n, M) to save more digits after comma:
1406 * - maximum int bit length is 64
1407 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1408 * - 13 * 4 = 52 < 64 -> M = 4
1412 static inline u32 ilog2_w(u64 n)
1414 return ilog2(n * n * n * n);
1417 static u32 shannon_entropy(struct heuristic_ws *ws)
1419 const u32 entropy_max = 8 * ilog2_w(2);
1420 u32 entropy_sum = 0;
1421 u32 p, p_base, sz_base;
1424 sz_base = ilog2_w(ws->sample_size);
1425 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1426 p = ws->bucket[i].count;
1427 p_base = ilog2_w(p);
1428 entropy_sum += p * (sz_base - p_base);
1431 entropy_sum /= ws->sample_size;
1432 return entropy_sum * 100 / entropy_max;
1435 #define RADIX_BASE 4U
1436 #define COUNTERS_SIZE (1U << RADIX_BASE)
1438 static u8 get4bits(u64 num, int shift) {
1443 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1448 * Use 4 bits as radix base
1449 * Use 16 u32 counters for calculating new position in buf array
1451 * @array - array that will be sorted
1452 * @array_buf - buffer array to store sorting results
1453 * must be equal in size to @array
1456 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1461 u32 counters[COUNTERS_SIZE];
1469 * Try avoid useless loop iterations for small numbers stored in big
1470 * counters. Example: 48 33 4 ... in 64bit array
1472 max_num = array[0].count;
1473 for (i = 1; i < num; i++) {
1474 buf_num = array[i].count;
1475 if (buf_num > max_num)
1479 buf_num = ilog2(max_num);
1480 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1483 while (shift < bitlen) {
1484 memset(counters, 0, sizeof(counters));
1486 for (i = 0; i < num; i++) {
1487 buf_num = array[i].count;
1488 addr = get4bits(buf_num, shift);
1492 for (i = 1; i < COUNTERS_SIZE; i++)
1493 counters[i] += counters[i - 1];
1495 for (i = num - 1; i >= 0; i--) {
1496 buf_num = array[i].count;
1497 addr = get4bits(buf_num, shift);
1499 new_addr = counters[addr];
1500 array_buf[new_addr] = array[i];
1503 shift += RADIX_BASE;
1506 * Normal radix expects to move data from a temporary array, to
1507 * the main one. But that requires some CPU time. Avoid that
1508 * by doing another sort iteration to original array instead of
1511 memset(counters, 0, sizeof(counters));
1513 for (i = 0; i < num; i ++) {
1514 buf_num = array_buf[i].count;
1515 addr = get4bits(buf_num, shift);
1519 for (i = 1; i < COUNTERS_SIZE; i++)
1520 counters[i] += counters[i - 1];
1522 for (i = num - 1; i >= 0; i--) {
1523 buf_num = array_buf[i].count;
1524 addr = get4bits(buf_num, shift);
1526 new_addr = counters[addr];
1527 array[new_addr] = array_buf[i];
1530 shift += RADIX_BASE;
1535 * Size of the core byte set - how many bytes cover 90% of the sample
1537 * There are several types of structured binary data that use nearly all byte
1538 * values. The distribution can be uniform and counts in all buckets will be
1539 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1541 * Other possibility is normal (Gaussian) distribution, where the data could
1542 * be potentially compressible, but we have to take a few more steps to decide
1545 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1546 * compression algo can easy fix that
1547 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1548 * probability is not compressible
1550 #define BYTE_CORE_SET_LOW (64)
1551 #define BYTE_CORE_SET_HIGH (200)
1553 static int byte_core_set_size(struct heuristic_ws *ws)
1556 u32 coreset_sum = 0;
1557 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1558 struct bucket_item *bucket = ws->bucket;
1560 /* Sort in reverse order */
1561 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1563 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1564 coreset_sum += bucket[i].count;
1566 if (coreset_sum > core_set_threshold)
1569 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1570 coreset_sum += bucket[i].count;
1571 if (coreset_sum > core_set_threshold)
1579 * Count byte values in buckets.
1580 * This heuristic can detect textual data (configs, xml, json, html, etc).
1581 * Because in most text-like data byte set is restricted to limited number of
1582 * possible characters, and that restriction in most cases makes data easy to
1585 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1586 * less - compressible
1587 * more - need additional analysis
1589 #define BYTE_SET_THRESHOLD (64)
1591 static u32 byte_set_size(const struct heuristic_ws *ws)
1594 u32 byte_set_size = 0;
1596 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1597 if (ws->bucket[i].count > 0)
1602 * Continue collecting count of byte values in buckets. If the byte
1603 * set size is bigger then the threshold, it's pointless to continue,
1604 * the detection technique would fail for this type of data.
1606 for (; i < BUCKET_SIZE; i++) {
1607 if (ws->bucket[i].count > 0) {
1609 if (byte_set_size > BYTE_SET_THRESHOLD)
1610 return byte_set_size;
1614 return byte_set_size;
1617 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1619 const u32 half_of_sample = ws->sample_size / 2;
1620 const u8 *data = ws->sample;
1622 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1625 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1626 struct heuristic_ws *ws)
1629 u64 index, index_end;
1630 u32 i, curr_sample_pos;
1634 * Compression handles the input data by chunks of 128KiB
1635 * (defined by BTRFS_MAX_UNCOMPRESSED)
1637 * We do the same for the heuristic and loop over the whole range.
1639 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1640 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1642 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1643 end = start + BTRFS_MAX_UNCOMPRESSED;
1645 index = start >> PAGE_SHIFT;
1646 index_end = end >> PAGE_SHIFT;
1648 /* Don't miss unaligned end */
1649 if (!IS_ALIGNED(end, PAGE_SIZE))
1652 curr_sample_pos = 0;
1653 while (index < index_end) {
1654 page = find_get_page(inode->i_mapping, index);
1655 in_data = kmap_local_page(page);
1656 /* Handle case where the start is not aligned to PAGE_SIZE */
1657 i = start % PAGE_SIZE;
1658 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1659 /* Don't sample any garbage from the last page */
1660 if (start > end - SAMPLING_READ_SIZE)
1662 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1663 SAMPLING_READ_SIZE);
1664 i += SAMPLING_INTERVAL;
1665 start += SAMPLING_INTERVAL;
1666 curr_sample_pos += SAMPLING_READ_SIZE;
1668 kunmap_local(in_data);
1674 ws->sample_size = curr_sample_pos;
1678 * Compression heuristic.
1680 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1681 * quickly (compared to direct compression) detect data characteristics
1682 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1685 * The following types of analysis can be performed:
1686 * - detect mostly zero data
1687 * - detect data with low "byte set" size (text, etc)
1688 * - detect data with low/high "core byte" set
1690 * Return non-zero if the compression should be done, 0 otherwise.
1692 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1694 struct list_head *ws_list = get_workspace(0, 0);
1695 struct heuristic_ws *ws;
1700 ws = list_entry(ws_list, struct heuristic_ws, list);
1702 heuristic_collect_sample(inode, start, end, ws);
1704 if (sample_repeated_patterns(ws)) {
1709 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1711 for (i = 0; i < ws->sample_size; i++) {
1712 byte = ws->sample[i];
1713 ws->bucket[byte].count++;
1716 i = byte_set_size(ws);
1717 if (i < BYTE_SET_THRESHOLD) {
1722 i = byte_core_set_size(ws);
1723 if (i <= BYTE_CORE_SET_LOW) {
1728 if (i >= BYTE_CORE_SET_HIGH) {
1733 i = shannon_entropy(ws);
1734 if (i <= ENTROPY_LVL_ACEPTABLE) {
1740 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1741 * needed to give green light to compression.
1743 * For now just assume that compression at that level is not worth the
1744 * resources because:
1746 * 1. it is possible to defrag the data later
1748 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1749 * values, every bucket has counter at level ~54. The heuristic would
1750 * be confused. This can happen when data have some internal repeated
1751 * patterns like "abbacbbc...". This can be detected by analyzing
1752 * pairs of bytes, which is too costly.
1754 if (i < ENTROPY_LVL_HIGH) {
1763 put_workspace(0, ws_list);
1768 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1769 * level, unrecognized string will set the default level
1771 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1773 unsigned int level = 0;
1779 if (str[0] == ':') {
1780 ret = kstrtouint(str + 1, 10, &level);
1785 level = btrfs_compress_set_level(type, level);