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"
33 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
35 const char* btrfs_compress_type2str(enum btrfs_compression_type 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];
50 bool btrfs_compress_is_valid_type(const char *str, size_t len)
54 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
55 size_t comp_len = strlen(btrfs_compress_types[i]);
60 if (!strncmp(btrfs_compress_types[i], str, comp_len))
66 static 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)
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:
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
89 * Not a big deal, just need to inform caller that we
90 * haven't allocated any pages yet.
97 static int compression_decompress_bio(int type, struct list_head *ws,
98 struct compressed_bio *cb)
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:
107 * This can't happen, the type is validated several times
108 * before we get here.
114 static 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)
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:
128 * This can't happen, the type is validated several times
129 * before we get here.
135 static int btrfs_decompress_bio(struct compressed_bio *cb);
137 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
138 unsigned long disk_size)
140 return sizeof(struct compressed_bio) +
141 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size;
144 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
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;
154 u8 csum[BTRFS_CSUM_SIZE];
155 struct compressed_bio *cb = bio->bi_private;
156 u8 *cb_sum = cb->sums;
158 if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM))
161 shash->tfm = fs_info->csum_shash;
163 for (i = 0; i < cb->nr_pages; i++) {
165 u32 bytes_left = PAGE_SIZE;
166 page = cb->compressed_pages[i];
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;
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,
178 kunmap_atomic(kaddr);
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);
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_io_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_io_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 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
353 btrfs_record_physical_zoned(inode, cb->start, bio);
354 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
355 cb->start, cb->start + cb->len - 1,
357 cb->compressed_pages[0]->mapping = NULL;
359 end_compressed_writeback(inode, cb);
360 /* note, our inode could be gone now */
363 * release the compressed pages, these came from alloc_page and
364 * are not attached to the inode at all
367 for (index = 0; index < cb->nr_pages; index++) {
368 page = cb->compressed_pages[index];
369 page->mapping = NULL;
373 /* finally free the cb struct */
374 kfree(cb->compressed_pages);
381 * worker function to build and submit bios for previously compressed pages.
382 * The corresponding pages in the inode should be marked for writeback
383 * and the compressed pages should have a reference on them for dropping
384 * when the IO is complete.
386 * This also checksums the file bytes and gets things ready for
389 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
390 unsigned long len, u64 disk_start,
391 unsigned long compressed_len,
392 struct page **compressed_pages,
393 unsigned long nr_pages,
394 unsigned int write_flags,
395 struct cgroup_subsys_state *blkcg_css)
397 struct btrfs_fs_info *fs_info = inode->root->fs_info;
398 struct bio *bio = NULL;
399 struct compressed_bio *cb;
400 unsigned long bytes_left;
403 u64 first_byte = disk_start;
405 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
406 const bool use_append = btrfs_use_zone_append(inode, disk_start);
407 const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
409 WARN_ON(!PAGE_ALIGNED(start));
410 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
412 return BLK_STS_RESOURCE;
413 refcount_set(&cb->pending_bios, 0);
415 cb->inode = &inode->vfs_inode;
419 cb->compressed_pages = compressed_pages;
420 cb->compressed_len = compressed_len;
422 cb->nr_pages = nr_pages;
424 bio = btrfs_bio_alloc(first_byte);
425 bio->bi_opf = bio_op | write_flags;
426 bio->bi_private = cb;
427 bio->bi_end_io = end_compressed_bio_write;
430 struct extent_map *em;
431 struct map_lookup *map;
432 struct block_device *bdev;
434 em = btrfs_get_chunk_map(fs_info, disk_start, PAGE_SIZE);
438 return BLK_STS_NOTSUPP;
441 map = em->map_lookup;
442 /* We only support single profile for now */
443 ASSERT(map->num_stripes == 1);
444 bdev = map->stripes[0].dev->bdev;
446 bio_set_dev(bio, bdev);
451 bio->bi_opf |= REQ_CGROUP_PUNT;
452 kthread_associate_blkcg(blkcg_css);
454 refcount_set(&cb->pending_bios, 1);
456 /* create and submit bios for the compressed pages */
457 bytes_left = compressed_len;
458 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
462 page = compressed_pages[pg_index];
463 page->mapping = inode->vfs_inode.i_mapping;
464 if (bio->bi_iter.bi_size)
465 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
469 * Page can only be added to bio if the current bio fits in
473 if (pg_index == 0 && use_append)
474 len = bio_add_zone_append_page(bio, page,
477 len = bio_add_page(bio, page, PAGE_SIZE, 0);
480 page->mapping = NULL;
481 if (submit || len < PAGE_SIZE) {
483 * inc the count before we submit the bio so
484 * we know the end IO handler won't happen before
485 * we inc the count. Otherwise, the cb might get
486 * freed before we're done setting it up
488 refcount_inc(&cb->pending_bios);
489 ret = btrfs_bio_wq_end_io(fs_info, bio,
490 BTRFS_WQ_ENDIO_DATA);
491 BUG_ON(ret); /* -ENOMEM */
494 ret = btrfs_csum_one_bio(inode, bio, start, 1);
495 BUG_ON(ret); /* -ENOMEM */
498 ret = btrfs_map_bio(fs_info, bio, 0);
500 bio->bi_status = ret;
504 bio = btrfs_bio_alloc(first_byte);
505 bio->bi_opf = bio_op | write_flags;
506 bio->bi_private = cb;
507 bio->bi_end_io = end_compressed_bio_write;
509 bio->bi_opf |= REQ_CGROUP_PUNT;
511 * Use bio_add_page() to ensure the bio has at least one
514 bio_add_page(bio, page, PAGE_SIZE, 0);
516 if (bytes_left < PAGE_SIZE) {
518 "bytes left %lu compress len %lu nr %lu",
519 bytes_left, cb->compressed_len, cb->nr_pages);
521 bytes_left -= PAGE_SIZE;
522 first_byte += PAGE_SIZE;
526 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
527 BUG_ON(ret); /* -ENOMEM */
530 ret = btrfs_csum_one_bio(inode, bio, start, 1);
531 BUG_ON(ret); /* -ENOMEM */
534 ret = btrfs_map_bio(fs_info, bio, 0);
536 bio->bi_status = ret;
541 kthread_associate_blkcg(NULL);
546 static u64 bio_end_offset(struct bio *bio)
548 struct bio_vec *last = bio_last_bvec_all(bio);
550 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
553 static noinline int add_ra_bio_pages(struct inode *inode,
555 struct compressed_bio *cb)
557 unsigned long end_index;
558 unsigned long pg_index;
560 u64 isize = i_size_read(inode);
563 unsigned long nr_pages = 0;
564 struct extent_map *em;
565 struct address_space *mapping = inode->i_mapping;
566 struct extent_map_tree *em_tree;
567 struct extent_io_tree *tree;
571 last_offset = bio_end_offset(cb->orig_bio);
572 em_tree = &BTRFS_I(inode)->extent_tree;
573 tree = &BTRFS_I(inode)->io_tree;
578 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
580 while (last_offset < compressed_end) {
581 pg_index = last_offset >> PAGE_SHIFT;
583 if (pg_index > end_index)
586 page = xa_load(&mapping->i_pages, pg_index);
587 if (page && !xa_is_value(page)) {
594 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
599 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
605 * at this point, we have a locked page in the page cache
606 * for these bytes in the file. But, we have to make
607 * sure they map to this compressed extent on disk.
609 ret = set_page_extent_mapped(page);
616 end = last_offset + PAGE_SIZE - 1;
617 lock_extent(tree, last_offset, end);
618 read_lock(&em_tree->lock);
619 em = lookup_extent_mapping(em_tree, last_offset,
621 read_unlock(&em_tree->lock);
623 if (!em || last_offset < em->start ||
624 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
625 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
627 unlock_extent(tree, last_offset, end);
634 if (page->index == end_index) {
635 size_t zero_offset = offset_in_page(isize);
639 zeros = PAGE_SIZE - zero_offset;
640 memzero_page(page, zero_offset, zeros);
641 flush_dcache_page(page);
645 ret = bio_add_page(cb->orig_bio, page,
648 if (ret == PAGE_SIZE) {
652 unlock_extent(tree, last_offset, end);
658 last_offset += PAGE_SIZE;
664 * for a compressed read, the bio we get passed has all the inode pages
665 * in it. We don't actually do IO on those pages but allocate new ones
666 * to hold the compressed pages on disk.
668 * bio->bi_iter.bi_sector points to the compressed extent on disk
669 * bio->bi_io_vec points to all of the inode pages
671 * After the compressed pages are read, we copy the bytes into the
672 * bio we were passed and then call the bio end_io calls
674 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
675 int mirror_num, unsigned long bio_flags)
677 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
678 struct extent_map_tree *em_tree;
679 struct compressed_bio *cb;
680 unsigned long compressed_len;
681 unsigned long nr_pages;
682 unsigned long pg_index;
684 struct bio *comp_bio;
685 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
688 struct extent_map *em;
689 blk_status_t ret = BLK_STS_RESOURCE;
693 em_tree = &BTRFS_I(inode)->extent_tree;
695 /* we need the actual starting offset of this extent in the file */
696 read_lock(&em_tree->lock);
697 em = lookup_extent_mapping(em_tree,
698 page_offset(bio_first_page_all(bio)),
699 fs_info->sectorsize);
700 read_unlock(&em_tree->lock);
702 return BLK_STS_IOERR;
704 compressed_len = em->block_len;
705 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
709 refcount_set(&cb->pending_bios, 0);
712 cb->mirror_num = mirror_num;
715 cb->start = em->orig_start;
717 em_start = em->start;
722 cb->len = bio->bi_iter.bi_size;
723 cb->compressed_len = compressed_len;
724 cb->compress_type = extent_compress_type(bio_flags);
727 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
728 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
730 if (!cb->compressed_pages)
733 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
734 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
736 if (!cb->compressed_pages[pg_index]) {
737 faili = pg_index - 1;
738 ret = BLK_STS_RESOURCE;
742 faili = nr_pages - 1;
743 cb->nr_pages = nr_pages;
745 add_ra_bio_pages(inode, em_start + em_len, cb);
747 /* include any pages we added in add_ra-bio_pages */
748 cb->len = bio->bi_iter.bi_size;
750 comp_bio = btrfs_bio_alloc(cur_disk_byte);
751 comp_bio->bi_opf = REQ_OP_READ;
752 comp_bio->bi_private = cb;
753 comp_bio->bi_end_io = end_compressed_bio_read;
754 refcount_set(&cb->pending_bios, 1);
756 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
757 u32 pg_len = PAGE_SIZE;
761 * To handle subpage case, we need to make sure the bio only
762 * covers the range we need.
764 * If we're at the last page, truncate the length to only cover
765 * the remaining part.
767 if (pg_index == nr_pages - 1)
768 pg_len = min_t(u32, PAGE_SIZE,
769 compressed_len - pg_index * PAGE_SIZE);
771 page = cb->compressed_pages[pg_index];
772 page->mapping = inode->i_mapping;
773 page->index = em_start >> PAGE_SHIFT;
775 if (comp_bio->bi_iter.bi_size)
776 submit = btrfs_bio_fits_in_stripe(page, pg_len,
779 page->mapping = NULL;
780 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
781 unsigned int nr_sectors;
783 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
784 BTRFS_WQ_ENDIO_DATA);
785 BUG_ON(ret); /* -ENOMEM */
788 * inc the count before we submit the bio so
789 * we know the end IO handler won't happen before
790 * we inc the count. Otherwise, the cb might get
791 * freed before we're done setting it up
793 refcount_inc(&cb->pending_bios);
795 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
796 BUG_ON(ret); /* -ENOMEM */
798 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
799 fs_info->sectorsize);
800 sums += fs_info->csum_size * nr_sectors;
802 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
804 comp_bio->bi_status = ret;
808 comp_bio = btrfs_bio_alloc(cur_disk_byte);
809 comp_bio->bi_opf = REQ_OP_READ;
810 comp_bio->bi_private = cb;
811 comp_bio->bi_end_io = end_compressed_bio_read;
813 bio_add_page(comp_bio, page, pg_len, 0);
815 cur_disk_byte += pg_len;
818 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
819 BUG_ON(ret); /* -ENOMEM */
821 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
822 BUG_ON(ret); /* -ENOMEM */
824 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
826 comp_bio->bi_status = ret;
834 __free_page(cb->compressed_pages[faili]);
838 kfree(cb->compressed_pages);
847 * Heuristic uses systematic sampling to collect data from the input data
848 * range, the logic can be tuned by the following constants:
850 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
851 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
853 #define SAMPLING_READ_SIZE (16)
854 #define SAMPLING_INTERVAL (256)
857 * For statistical analysis of the input data we consider bytes that form a
858 * Galois Field of 256 objects. Each object has an attribute count, ie. how
859 * many times the object appeared in the sample.
861 #define BUCKET_SIZE (256)
864 * The size of the sample is based on a statistical sampling rule of thumb.
865 * The common way is to perform sampling tests as long as the number of
866 * elements in each cell is at least 5.
868 * Instead of 5, we choose 32 to obtain more accurate results.
869 * If the data contain the maximum number of symbols, which is 256, we obtain a
870 * sample size bound by 8192.
872 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
873 * from up to 512 locations.
875 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
876 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
882 struct heuristic_ws {
883 /* Partial copy of input data */
886 /* Buckets store counters for each byte value */
887 struct bucket_item *bucket;
889 struct bucket_item *bucket_b;
890 struct list_head list;
893 static struct workspace_manager heuristic_wsm;
895 static void free_heuristic_ws(struct list_head *ws)
897 struct heuristic_ws *workspace;
899 workspace = list_entry(ws, struct heuristic_ws, list);
901 kvfree(workspace->sample);
902 kfree(workspace->bucket);
903 kfree(workspace->bucket_b);
907 static struct list_head *alloc_heuristic_ws(unsigned int level)
909 struct heuristic_ws *ws;
911 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
913 return ERR_PTR(-ENOMEM);
915 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
919 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
923 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
927 INIT_LIST_HEAD(&ws->list);
930 free_heuristic_ws(&ws->list);
931 return ERR_PTR(-ENOMEM);
934 const struct btrfs_compress_op btrfs_heuristic_compress = {
935 .workspace_manager = &heuristic_wsm,
938 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
939 /* The heuristic is represented as compression type 0 */
940 &btrfs_heuristic_compress,
941 &btrfs_zlib_compress,
943 &btrfs_zstd_compress,
946 static struct list_head *alloc_workspace(int type, unsigned int level)
949 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
950 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
951 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
952 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
955 * This can't happen, the type is validated several times
956 * before we get here.
962 static void free_workspace(int type, struct list_head *ws)
965 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
966 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
967 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
968 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
971 * This can't happen, the type is validated several times
972 * before we get here.
978 static void btrfs_init_workspace_manager(int type)
980 struct workspace_manager *wsm;
981 struct list_head *workspace;
983 wsm = btrfs_compress_op[type]->workspace_manager;
984 INIT_LIST_HEAD(&wsm->idle_ws);
985 spin_lock_init(&wsm->ws_lock);
986 atomic_set(&wsm->total_ws, 0);
987 init_waitqueue_head(&wsm->ws_wait);
990 * Preallocate one workspace for each compression type so we can
991 * guarantee forward progress in the worst case
993 workspace = alloc_workspace(type, 0);
994 if (IS_ERR(workspace)) {
996 "BTRFS: cannot preallocate compression workspace, will try later\n");
998 atomic_set(&wsm->total_ws, 1);
1000 list_add(workspace, &wsm->idle_ws);
1004 static void btrfs_cleanup_workspace_manager(int type)
1006 struct workspace_manager *wsman;
1007 struct list_head *ws;
1009 wsman = btrfs_compress_op[type]->workspace_manager;
1010 while (!list_empty(&wsman->idle_ws)) {
1011 ws = wsman->idle_ws.next;
1013 free_workspace(type, ws);
1014 atomic_dec(&wsman->total_ws);
1019 * This finds an available workspace or allocates a new one.
1020 * If it's not possible to allocate a new one, waits until there's one.
1021 * Preallocation makes a forward progress guarantees and we do not return
1024 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1026 struct workspace_manager *wsm;
1027 struct list_head *workspace;
1028 int cpus = num_online_cpus();
1030 struct list_head *idle_ws;
1031 spinlock_t *ws_lock;
1033 wait_queue_head_t *ws_wait;
1036 wsm = btrfs_compress_op[type]->workspace_manager;
1037 idle_ws = &wsm->idle_ws;
1038 ws_lock = &wsm->ws_lock;
1039 total_ws = &wsm->total_ws;
1040 ws_wait = &wsm->ws_wait;
1041 free_ws = &wsm->free_ws;
1045 if (!list_empty(idle_ws)) {
1046 workspace = idle_ws->next;
1047 list_del(workspace);
1049 spin_unlock(ws_lock);
1053 if (atomic_read(total_ws) > cpus) {
1056 spin_unlock(ws_lock);
1057 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1058 if (atomic_read(total_ws) > cpus && !*free_ws)
1060 finish_wait(ws_wait, &wait);
1063 atomic_inc(total_ws);
1064 spin_unlock(ws_lock);
1067 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1068 * to turn it off here because we might get called from the restricted
1069 * context of btrfs_compress_bio/btrfs_compress_pages
1071 nofs_flag = memalloc_nofs_save();
1072 workspace = alloc_workspace(type, level);
1073 memalloc_nofs_restore(nofs_flag);
1075 if (IS_ERR(workspace)) {
1076 atomic_dec(total_ws);
1080 * Do not return the error but go back to waiting. There's a
1081 * workspace preallocated for each type and the compression
1082 * time is bounded so we get to a workspace eventually. This
1083 * makes our caller's life easier.
1085 * To prevent silent and low-probability deadlocks (when the
1086 * initial preallocation fails), check if there are any
1087 * workspaces at all.
1089 if (atomic_read(total_ws) == 0) {
1090 static DEFINE_RATELIMIT_STATE(_rs,
1091 /* once per minute */ 60 * HZ,
1094 if (__ratelimit(&_rs)) {
1095 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1103 static struct list_head *get_workspace(int type, int level)
1106 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1107 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1108 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1109 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1112 * This can't happen, the type is validated several times
1113 * before we get here.
1120 * put a workspace struct back on the list or free it if we have enough
1121 * idle ones sitting around
1123 void btrfs_put_workspace(int type, struct list_head *ws)
1125 struct workspace_manager *wsm;
1126 struct list_head *idle_ws;
1127 spinlock_t *ws_lock;
1129 wait_queue_head_t *ws_wait;
1132 wsm = btrfs_compress_op[type]->workspace_manager;
1133 idle_ws = &wsm->idle_ws;
1134 ws_lock = &wsm->ws_lock;
1135 total_ws = &wsm->total_ws;
1136 ws_wait = &wsm->ws_wait;
1137 free_ws = &wsm->free_ws;
1140 if (*free_ws <= num_online_cpus()) {
1141 list_add(ws, idle_ws);
1143 spin_unlock(ws_lock);
1146 spin_unlock(ws_lock);
1148 free_workspace(type, ws);
1149 atomic_dec(total_ws);
1151 cond_wake_up(ws_wait);
1154 static void put_workspace(int type, struct list_head *ws)
1157 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1158 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1159 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1160 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1163 * This can't happen, the type is validated several times
1164 * before we get here.
1171 * Adjust @level according to the limits of the compression algorithm or
1172 * fallback to default
1174 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1176 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1179 level = ops->default_level;
1181 level = min(level, ops->max_level);
1187 * Given an address space and start and length, compress the bytes into @pages
1188 * that are allocated on demand.
1190 * @type_level is encoded algorithm and level, where level 0 means whatever
1191 * default the algorithm chooses and is opaque here;
1192 * - compression algo are 0-3
1193 * - the level are bits 4-7
1195 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1196 * and returns number of actually allocated pages
1198 * @total_in is used to return the number of bytes actually read. It
1199 * may be smaller than the input length if we had to exit early because we
1200 * ran out of room in the pages array or because we cross the
1201 * max_out threshold.
1203 * @total_out is an in/out parameter, must be set to the input length and will
1204 * be also used to return the total number of compressed bytes
1206 * @max_out tells us the max number of bytes that we're allowed to
1209 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1210 u64 start, struct page **pages,
1211 unsigned long *out_pages,
1212 unsigned long *total_in,
1213 unsigned long *total_out)
1215 int type = btrfs_compress_type(type_level);
1216 int level = btrfs_compress_level(type_level);
1217 struct list_head *workspace;
1220 level = btrfs_compress_set_level(type, level);
1221 workspace = get_workspace(type, level);
1222 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1223 out_pages, total_in, total_out);
1224 put_workspace(type, workspace);
1229 * pages_in is an array of pages with compressed data.
1231 * disk_start is the starting logical offset of this array in the file
1233 * orig_bio contains the pages from the file that we want to decompress into
1235 * srclen is the number of bytes in pages_in
1237 * The basic idea is that we have a bio that was created by readpages.
1238 * The pages in the bio are for the uncompressed data, and they may not
1239 * be contiguous. They all correspond to the range of bytes covered by
1240 * the compressed extent.
1242 static int btrfs_decompress_bio(struct compressed_bio *cb)
1244 struct list_head *workspace;
1246 int type = cb->compress_type;
1248 workspace = get_workspace(type, 0);
1249 ret = compression_decompress_bio(type, workspace, cb);
1250 put_workspace(type, workspace);
1256 * a less complex decompression routine. Our compressed data fits in a
1257 * single page, and we want to read a single page out of it.
1258 * start_byte tells us the offset into the compressed data we're interested in
1260 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1261 unsigned long start_byte, size_t srclen, size_t destlen)
1263 struct list_head *workspace;
1266 workspace = get_workspace(type, 0);
1267 ret = compression_decompress(type, workspace, data_in, dest_page,
1268 start_byte, srclen, destlen);
1269 put_workspace(type, workspace);
1274 void __init btrfs_init_compress(void)
1276 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1277 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1278 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1279 zstd_init_workspace_manager();
1282 void __cold btrfs_exit_compress(void)
1284 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1285 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1286 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1287 zstd_cleanup_workspace_manager();
1291 * Copy uncompressed data from working buffer to pages.
1293 * buf_start is the byte offset we're of the start of our workspace buffer.
1295 * total_out is the last byte of the buffer
1297 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1298 unsigned long total_out, u64 disk_start,
1301 unsigned long buf_offset;
1302 unsigned long current_buf_start;
1303 unsigned long start_byte;
1304 unsigned long prev_start_byte;
1305 unsigned long working_bytes = total_out - buf_start;
1306 unsigned long bytes;
1307 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1310 * start byte is the first byte of the page we're currently
1311 * copying into relative to the start of the compressed data.
1313 start_byte = page_offset(bvec.bv_page) - disk_start;
1315 /* we haven't yet hit data corresponding to this page */
1316 if (total_out <= start_byte)
1320 * the start of the data we care about is offset into
1321 * the middle of our working buffer
1323 if (total_out > start_byte && buf_start < start_byte) {
1324 buf_offset = start_byte - buf_start;
1325 working_bytes -= buf_offset;
1329 current_buf_start = buf_start;
1331 /* copy bytes from the working buffer into the pages */
1332 while (working_bytes > 0) {
1333 bytes = min_t(unsigned long, bvec.bv_len,
1334 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1335 bytes = min(bytes, working_bytes);
1337 memcpy_to_page(bvec.bv_page, bvec.bv_offset, buf + buf_offset,
1339 flush_dcache_page(bvec.bv_page);
1341 buf_offset += bytes;
1342 working_bytes -= bytes;
1343 current_buf_start += bytes;
1345 /* check if we need to pick another page */
1346 bio_advance(bio, bytes);
1347 if (!bio->bi_iter.bi_size)
1349 bvec = bio_iter_iovec(bio, bio->bi_iter);
1350 prev_start_byte = start_byte;
1351 start_byte = page_offset(bvec.bv_page) - disk_start;
1354 * We need to make sure we're only adjusting
1355 * our offset into compression working buffer when
1356 * we're switching pages. Otherwise we can incorrectly
1357 * keep copying when we were actually done.
1359 if (start_byte != prev_start_byte) {
1361 * make sure our new page is covered by this
1364 if (total_out <= start_byte)
1368 * the next page in the biovec might not be adjacent
1369 * to the last page, but it might still be found
1370 * inside this working buffer. bump our offset pointer
1372 if (total_out > start_byte &&
1373 current_buf_start < start_byte) {
1374 buf_offset = start_byte - buf_start;
1375 working_bytes = total_out - start_byte;
1376 current_buf_start = buf_start + buf_offset;
1385 * Shannon Entropy calculation
1387 * Pure byte distribution analysis fails to determine compressibility of data.
1388 * Try calculating entropy to estimate the average minimum number of bits
1389 * needed to encode the sampled data.
1391 * For convenience, return the percentage of needed bits, instead of amount of
1394 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1395 * and can be compressible with high probability
1397 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1399 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1401 #define ENTROPY_LVL_ACEPTABLE (65)
1402 #define ENTROPY_LVL_HIGH (80)
1405 * For increasead precision in shannon_entropy calculation,
1406 * let's do pow(n, M) to save more digits after comma:
1408 * - maximum int bit length is 64
1409 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1410 * - 13 * 4 = 52 < 64 -> M = 4
1414 static inline u32 ilog2_w(u64 n)
1416 return ilog2(n * n * n * n);
1419 static u32 shannon_entropy(struct heuristic_ws *ws)
1421 const u32 entropy_max = 8 * ilog2_w(2);
1422 u32 entropy_sum = 0;
1423 u32 p, p_base, sz_base;
1426 sz_base = ilog2_w(ws->sample_size);
1427 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1428 p = ws->bucket[i].count;
1429 p_base = ilog2_w(p);
1430 entropy_sum += p * (sz_base - p_base);
1433 entropy_sum /= ws->sample_size;
1434 return entropy_sum * 100 / entropy_max;
1437 #define RADIX_BASE 4U
1438 #define COUNTERS_SIZE (1U << RADIX_BASE)
1440 static u8 get4bits(u64 num, int shift) {
1445 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1450 * Use 4 bits as radix base
1451 * Use 16 u32 counters for calculating new position in buf array
1453 * @array - array that will be sorted
1454 * @array_buf - buffer array to store sorting results
1455 * must be equal in size to @array
1458 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1463 u32 counters[COUNTERS_SIZE];
1471 * Try avoid useless loop iterations for small numbers stored in big
1472 * counters. Example: 48 33 4 ... in 64bit array
1474 max_num = array[0].count;
1475 for (i = 1; i < num; i++) {
1476 buf_num = array[i].count;
1477 if (buf_num > max_num)
1481 buf_num = ilog2(max_num);
1482 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1485 while (shift < bitlen) {
1486 memset(counters, 0, sizeof(counters));
1488 for (i = 0; i < num; i++) {
1489 buf_num = array[i].count;
1490 addr = get4bits(buf_num, shift);
1494 for (i = 1; i < COUNTERS_SIZE; i++)
1495 counters[i] += counters[i - 1];
1497 for (i = num - 1; i >= 0; i--) {
1498 buf_num = array[i].count;
1499 addr = get4bits(buf_num, shift);
1501 new_addr = counters[addr];
1502 array_buf[new_addr] = array[i];
1505 shift += RADIX_BASE;
1508 * Normal radix expects to move data from a temporary array, to
1509 * the main one. But that requires some CPU time. Avoid that
1510 * by doing another sort iteration to original array instead of
1513 memset(counters, 0, sizeof(counters));
1515 for (i = 0; i < num; i ++) {
1516 buf_num = array_buf[i].count;
1517 addr = get4bits(buf_num, shift);
1521 for (i = 1; i < COUNTERS_SIZE; i++)
1522 counters[i] += counters[i - 1];
1524 for (i = num - 1; i >= 0; i--) {
1525 buf_num = array_buf[i].count;
1526 addr = get4bits(buf_num, shift);
1528 new_addr = counters[addr];
1529 array[new_addr] = array_buf[i];
1532 shift += RADIX_BASE;
1537 * Size of the core byte set - how many bytes cover 90% of the sample
1539 * There are several types of structured binary data that use nearly all byte
1540 * values. The distribution can be uniform and counts in all buckets will be
1541 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1543 * Other possibility is normal (Gaussian) distribution, where the data could
1544 * be potentially compressible, but we have to take a few more steps to decide
1547 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1548 * compression algo can easy fix that
1549 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1550 * probability is not compressible
1552 #define BYTE_CORE_SET_LOW (64)
1553 #define BYTE_CORE_SET_HIGH (200)
1555 static int byte_core_set_size(struct heuristic_ws *ws)
1558 u32 coreset_sum = 0;
1559 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1560 struct bucket_item *bucket = ws->bucket;
1562 /* Sort in reverse order */
1563 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1565 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1566 coreset_sum += bucket[i].count;
1568 if (coreset_sum > core_set_threshold)
1571 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1572 coreset_sum += bucket[i].count;
1573 if (coreset_sum > core_set_threshold)
1581 * Count byte values in buckets.
1582 * This heuristic can detect textual data (configs, xml, json, html, etc).
1583 * Because in most text-like data byte set is restricted to limited number of
1584 * possible characters, and that restriction in most cases makes data easy to
1587 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1588 * less - compressible
1589 * more - need additional analysis
1591 #define BYTE_SET_THRESHOLD (64)
1593 static u32 byte_set_size(const struct heuristic_ws *ws)
1596 u32 byte_set_size = 0;
1598 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1599 if (ws->bucket[i].count > 0)
1604 * Continue collecting count of byte values in buckets. If the byte
1605 * set size is bigger then the threshold, it's pointless to continue,
1606 * the detection technique would fail for this type of data.
1608 for (; i < BUCKET_SIZE; i++) {
1609 if (ws->bucket[i].count > 0) {
1611 if (byte_set_size > BYTE_SET_THRESHOLD)
1612 return byte_set_size;
1616 return byte_set_size;
1619 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1621 const u32 half_of_sample = ws->sample_size / 2;
1622 const u8 *data = ws->sample;
1624 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1627 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1628 struct heuristic_ws *ws)
1631 u64 index, index_end;
1632 u32 i, curr_sample_pos;
1636 * Compression handles the input data by chunks of 128KiB
1637 * (defined by BTRFS_MAX_UNCOMPRESSED)
1639 * We do the same for the heuristic and loop over the whole range.
1641 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1642 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1644 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1645 end = start + BTRFS_MAX_UNCOMPRESSED;
1647 index = start >> PAGE_SHIFT;
1648 index_end = end >> PAGE_SHIFT;
1650 /* Don't miss unaligned end */
1651 if (!IS_ALIGNED(end, PAGE_SIZE))
1654 curr_sample_pos = 0;
1655 while (index < index_end) {
1656 page = find_get_page(inode->i_mapping, index);
1657 in_data = kmap_local_page(page);
1658 /* Handle case where the start is not aligned to PAGE_SIZE */
1659 i = start % PAGE_SIZE;
1660 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1661 /* Don't sample any garbage from the last page */
1662 if (start > end - SAMPLING_READ_SIZE)
1664 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1665 SAMPLING_READ_SIZE);
1666 i += SAMPLING_INTERVAL;
1667 start += SAMPLING_INTERVAL;
1668 curr_sample_pos += SAMPLING_READ_SIZE;
1670 kunmap_local(in_data);
1676 ws->sample_size = curr_sample_pos;
1680 * Compression heuristic.
1682 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1683 * quickly (compared to direct compression) detect data characteristics
1684 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1687 * The following types of analysis can be performed:
1688 * - detect mostly zero data
1689 * - detect data with low "byte set" size (text, etc)
1690 * - detect data with low/high "core byte" set
1692 * Return non-zero if the compression should be done, 0 otherwise.
1694 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1696 struct list_head *ws_list = get_workspace(0, 0);
1697 struct heuristic_ws *ws;
1702 ws = list_entry(ws_list, struct heuristic_ws, list);
1704 heuristic_collect_sample(inode, start, end, ws);
1706 if (sample_repeated_patterns(ws)) {
1711 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1713 for (i = 0; i < ws->sample_size; i++) {
1714 byte = ws->sample[i];
1715 ws->bucket[byte].count++;
1718 i = byte_set_size(ws);
1719 if (i < BYTE_SET_THRESHOLD) {
1724 i = byte_core_set_size(ws);
1725 if (i <= BYTE_CORE_SET_LOW) {
1730 if (i >= BYTE_CORE_SET_HIGH) {
1735 i = shannon_entropy(ws);
1736 if (i <= ENTROPY_LVL_ACEPTABLE) {
1742 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1743 * needed to give green light to compression.
1745 * For now just assume that compression at that level is not worth the
1746 * resources because:
1748 * 1. it is possible to defrag the data later
1750 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1751 * values, every bucket has counter at level ~54. The heuristic would
1752 * be confused. This can happen when data have some internal repeated
1753 * patterns like "abbacbbc...". This can be detected by analyzing
1754 * pairs of bytes, which is too costly.
1756 if (i < ENTROPY_LVL_HIGH) {
1765 put_workspace(0, ws_list);
1770 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1771 * level, unrecognized string will set the default level
1773 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1775 unsigned int level = 0;
1781 if (str[0] == ':') {
1782 ret = kstrtouint(str + 1, 10, &level);
1787 level = btrfs_compress_set_level(type, level);