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,
356 bio->bi_status == BLK_STS_OK);
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,
468 if (pg_index == 0 && use_append)
469 len = bio_add_zone_append_page(bio, page, PAGE_SIZE, 0);
471 len = bio_add_page(bio, page, PAGE_SIZE, 0);
473 page->mapping = NULL;
474 if (submit || len < PAGE_SIZE) {
476 * inc the count before we submit the bio so
477 * we know the end IO handler won't happen before
478 * we inc the count. Otherwise, the cb might get
479 * freed before we're done setting it up
481 refcount_inc(&cb->pending_bios);
482 ret = btrfs_bio_wq_end_io(fs_info, bio,
483 BTRFS_WQ_ENDIO_DATA);
484 BUG_ON(ret); /* -ENOMEM */
487 ret = btrfs_csum_one_bio(inode, bio, start, 1);
488 BUG_ON(ret); /* -ENOMEM */
491 ret = btrfs_map_bio(fs_info, bio, 0);
493 bio->bi_status = ret;
497 bio = btrfs_bio_alloc(first_byte);
498 bio->bi_opf = bio_op | write_flags;
499 bio->bi_private = cb;
500 bio->bi_end_io = end_compressed_bio_write;
502 bio->bi_opf |= REQ_CGROUP_PUNT;
504 * Use bio_add_page() to ensure the bio has at least one
507 bio_add_page(bio, page, PAGE_SIZE, 0);
509 if (bytes_left < PAGE_SIZE) {
511 "bytes left %lu compress len %lu nr %lu",
512 bytes_left, cb->compressed_len, cb->nr_pages);
514 bytes_left -= PAGE_SIZE;
515 first_byte += PAGE_SIZE;
519 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
520 BUG_ON(ret); /* -ENOMEM */
523 ret = btrfs_csum_one_bio(inode, bio, start, 1);
524 BUG_ON(ret); /* -ENOMEM */
527 ret = btrfs_map_bio(fs_info, bio, 0);
529 bio->bi_status = ret;
534 kthread_associate_blkcg(NULL);
539 static u64 bio_end_offset(struct bio *bio)
541 struct bio_vec *last = bio_last_bvec_all(bio);
543 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
546 static noinline int add_ra_bio_pages(struct inode *inode,
548 struct compressed_bio *cb)
550 unsigned long end_index;
551 unsigned long pg_index;
553 u64 isize = i_size_read(inode);
556 unsigned long nr_pages = 0;
557 struct extent_map *em;
558 struct address_space *mapping = inode->i_mapping;
559 struct extent_map_tree *em_tree;
560 struct extent_io_tree *tree;
564 last_offset = bio_end_offset(cb->orig_bio);
565 em_tree = &BTRFS_I(inode)->extent_tree;
566 tree = &BTRFS_I(inode)->io_tree;
571 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
573 while (last_offset < compressed_end) {
574 pg_index = last_offset >> PAGE_SHIFT;
576 if (pg_index > end_index)
579 page = xa_load(&mapping->i_pages, pg_index);
580 if (page && !xa_is_value(page)) {
587 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
592 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
598 * at this point, we have a locked page in the page cache
599 * for these bytes in the file. But, we have to make
600 * sure they map to this compressed extent on disk.
602 ret = set_page_extent_mapped(page);
609 end = last_offset + PAGE_SIZE - 1;
610 lock_extent(tree, last_offset, end);
611 read_lock(&em_tree->lock);
612 em = lookup_extent_mapping(em_tree, last_offset,
614 read_unlock(&em_tree->lock);
616 if (!em || last_offset < em->start ||
617 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
618 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
620 unlock_extent(tree, last_offset, end);
627 if (page->index == end_index) {
628 size_t zero_offset = offset_in_page(isize);
632 zeros = PAGE_SIZE - zero_offset;
633 memzero_page(page, zero_offset, zeros);
634 flush_dcache_page(page);
638 ret = bio_add_page(cb->orig_bio, page,
641 if (ret == PAGE_SIZE) {
645 unlock_extent(tree, last_offset, end);
651 last_offset += PAGE_SIZE;
657 * for a compressed read, the bio we get passed has all the inode pages
658 * in it. We don't actually do IO on those pages but allocate new ones
659 * to hold the compressed pages on disk.
661 * bio->bi_iter.bi_sector points to the compressed extent on disk
662 * bio->bi_io_vec points to all of the inode pages
664 * After the compressed pages are read, we copy the bytes into the
665 * bio we were passed and then call the bio end_io calls
667 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
668 int mirror_num, unsigned long bio_flags)
670 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
671 struct extent_map_tree *em_tree;
672 struct compressed_bio *cb;
673 unsigned long compressed_len;
674 unsigned long nr_pages;
675 unsigned long pg_index;
677 struct bio *comp_bio;
678 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
681 struct extent_map *em;
682 blk_status_t ret = BLK_STS_RESOURCE;
686 em_tree = &BTRFS_I(inode)->extent_tree;
688 /* we need the actual starting offset of this extent in the file */
689 read_lock(&em_tree->lock);
690 em = lookup_extent_mapping(em_tree,
691 page_offset(bio_first_page_all(bio)),
692 fs_info->sectorsize);
693 read_unlock(&em_tree->lock);
695 return BLK_STS_IOERR;
697 compressed_len = em->block_len;
698 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
702 refcount_set(&cb->pending_bios, 0);
705 cb->mirror_num = mirror_num;
708 cb->start = em->orig_start;
710 em_start = em->start;
715 cb->len = bio->bi_iter.bi_size;
716 cb->compressed_len = compressed_len;
717 cb->compress_type = extent_compress_type(bio_flags);
720 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
721 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
723 if (!cb->compressed_pages)
726 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
727 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
729 if (!cb->compressed_pages[pg_index]) {
730 faili = pg_index - 1;
731 ret = BLK_STS_RESOURCE;
735 faili = nr_pages - 1;
736 cb->nr_pages = nr_pages;
738 add_ra_bio_pages(inode, em_start + em_len, cb);
740 /* include any pages we added in add_ra-bio_pages */
741 cb->len = bio->bi_iter.bi_size;
743 comp_bio = btrfs_bio_alloc(cur_disk_byte);
744 comp_bio->bi_opf = REQ_OP_READ;
745 comp_bio->bi_private = cb;
746 comp_bio->bi_end_io = end_compressed_bio_read;
747 refcount_set(&cb->pending_bios, 1);
749 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
750 u32 pg_len = PAGE_SIZE;
754 * To handle subpage case, we need to make sure the bio only
755 * covers the range we need.
757 * If we're at the last page, truncate the length to only cover
758 * the remaining part.
760 if (pg_index == nr_pages - 1)
761 pg_len = min_t(u32, PAGE_SIZE,
762 compressed_len - pg_index * PAGE_SIZE);
764 page = cb->compressed_pages[pg_index];
765 page->mapping = inode->i_mapping;
766 page->index = em_start >> PAGE_SHIFT;
768 if (comp_bio->bi_iter.bi_size)
769 submit = btrfs_bio_fits_in_stripe(page, pg_len,
772 page->mapping = NULL;
773 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
774 unsigned int nr_sectors;
776 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
777 BTRFS_WQ_ENDIO_DATA);
778 BUG_ON(ret); /* -ENOMEM */
781 * inc the count before we submit the bio so
782 * we know the end IO handler won't happen before
783 * we inc the count. Otherwise, the cb might get
784 * freed before we're done setting it up
786 refcount_inc(&cb->pending_bios);
788 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
789 BUG_ON(ret); /* -ENOMEM */
791 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
792 fs_info->sectorsize);
793 sums += fs_info->csum_size * nr_sectors;
795 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
797 comp_bio->bi_status = ret;
801 comp_bio = btrfs_bio_alloc(cur_disk_byte);
802 comp_bio->bi_opf = REQ_OP_READ;
803 comp_bio->bi_private = cb;
804 comp_bio->bi_end_io = end_compressed_bio_read;
806 bio_add_page(comp_bio, page, pg_len, 0);
808 cur_disk_byte += pg_len;
811 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
812 BUG_ON(ret); /* -ENOMEM */
814 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
815 BUG_ON(ret); /* -ENOMEM */
817 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
819 comp_bio->bi_status = ret;
827 __free_page(cb->compressed_pages[faili]);
831 kfree(cb->compressed_pages);
840 * Heuristic uses systematic sampling to collect data from the input data
841 * range, the logic can be tuned by the following constants:
843 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
844 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
846 #define SAMPLING_READ_SIZE (16)
847 #define SAMPLING_INTERVAL (256)
850 * For statistical analysis of the input data we consider bytes that form a
851 * Galois Field of 256 objects. Each object has an attribute count, ie. how
852 * many times the object appeared in the sample.
854 #define BUCKET_SIZE (256)
857 * The size of the sample is based on a statistical sampling rule of thumb.
858 * The common way is to perform sampling tests as long as the number of
859 * elements in each cell is at least 5.
861 * Instead of 5, we choose 32 to obtain more accurate results.
862 * If the data contain the maximum number of symbols, which is 256, we obtain a
863 * sample size bound by 8192.
865 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
866 * from up to 512 locations.
868 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
869 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
875 struct heuristic_ws {
876 /* Partial copy of input data */
879 /* Buckets store counters for each byte value */
880 struct bucket_item *bucket;
882 struct bucket_item *bucket_b;
883 struct list_head list;
886 static struct workspace_manager heuristic_wsm;
888 static void free_heuristic_ws(struct list_head *ws)
890 struct heuristic_ws *workspace;
892 workspace = list_entry(ws, struct heuristic_ws, list);
894 kvfree(workspace->sample);
895 kfree(workspace->bucket);
896 kfree(workspace->bucket_b);
900 static struct list_head *alloc_heuristic_ws(unsigned int level)
902 struct heuristic_ws *ws;
904 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
906 return ERR_PTR(-ENOMEM);
908 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
912 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
916 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
920 INIT_LIST_HEAD(&ws->list);
923 free_heuristic_ws(&ws->list);
924 return ERR_PTR(-ENOMEM);
927 const struct btrfs_compress_op btrfs_heuristic_compress = {
928 .workspace_manager = &heuristic_wsm,
931 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
932 /* The heuristic is represented as compression type 0 */
933 &btrfs_heuristic_compress,
934 &btrfs_zlib_compress,
936 &btrfs_zstd_compress,
939 static struct list_head *alloc_workspace(int type, unsigned int level)
942 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
943 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
944 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
945 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
948 * This can't happen, the type is validated several times
949 * before we get here.
955 static void free_workspace(int type, struct list_head *ws)
958 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
959 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
960 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
961 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
964 * This can't happen, the type is validated several times
965 * before we get here.
971 static void btrfs_init_workspace_manager(int type)
973 struct workspace_manager *wsm;
974 struct list_head *workspace;
976 wsm = btrfs_compress_op[type]->workspace_manager;
977 INIT_LIST_HEAD(&wsm->idle_ws);
978 spin_lock_init(&wsm->ws_lock);
979 atomic_set(&wsm->total_ws, 0);
980 init_waitqueue_head(&wsm->ws_wait);
983 * Preallocate one workspace for each compression type so we can
984 * guarantee forward progress in the worst case
986 workspace = alloc_workspace(type, 0);
987 if (IS_ERR(workspace)) {
989 "BTRFS: cannot preallocate compression workspace, will try later\n");
991 atomic_set(&wsm->total_ws, 1);
993 list_add(workspace, &wsm->idle_ws);
997 static void btrfs_cleanup_workspace_manager(int type)
999 struct workspace_manager *wsman;
1000 struct list_head *ws;
1002 wsman = btrfs_compress_op[type]->workspace_manager;
1003 while (!list_empty(&wsman->idle_ws)) {
1004 ws = wsman->idle_ws.next;
1006 free_workspace(type, ws);
1007 atomic_dec(&wsman->total_ws);
1012 * This finds an available workspace or allocates a new one.
1013 * If it's not possible to allocate a new one, waits until there's one.
1014 * Preallocation makes a forward progress guarantees and we do not return
1017 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1019 struct workspace_manager *wsm;
1020 struct list_head *workspace;
1021 int cpus = num_online_cpus();
1023 struct list_head *idle_ws;
1024 spinlock_t *ws_lock;
1026 wait_queue_head_t *ws_wait;
1029 wsm = btrfs_compress_op[type]->workspace_manager;
1030 idle_ws = &wsm->idle_ws;
1031 ws_lock = &wsm->ws_lock;
1032 total_ws = &wsm->total_ws;
1033 ws_wait = &wsm->ws_wait;
1034 free_ws = &wsm->free_ws;
1038 if (!list_empty(idle_ws)) {
1039 workspace = idle_ws->next;
1040 list_del(workspace);
1042 spin_unlock(ws_lock);
1046 if (atomic_read(total_ws) > cpus) {
1049 spin_unlock(ws_lock);
1050 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1051 if (atomic_read(total_ws) > cpus && !*free_ws)
1053 finish_wait(ws_wait, &wait);
1056 atomic_inc(total_ws);
1057 spin_unlock(ws_lock);
1060 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1061 * to turn it off here because we might get called from the restricted
1062 * context of btrfs_compress_bio/btrfs_compress_pages
1064 nofs_flag = memalloc_nofs_save();
1065 workspace = alloc_workspace(type, level);
1066 memalloc_nofs_restore(nofs_flag);
1068 if (IS_ERR(workspace)) {
1069 atomic_dec(total_ws);
1073 * Do not return the error but go back to waiting. There's a
1074 * workspace preallocated for each type and the compression
1075 * time is bounded so we get to a workspace eventually. This
1076 * makes our caller's life easier.
1078 * To prevent silent and low-probability deadlocks (when the
1079 * initial preallocation fails), check if there are any
1080 * workspaces at all.
1082 if (atomic_read(total_ws) == 0) {
1083 static DEFINE_RATELIMIT_STATE(_rs,
1084 /* once per minute */ 60 * HZ,
1087 if (__ratelimit(&_rs)) {
1088 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1096 static struct list_head *get_workspace(int type, int level)
1099 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1100 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1101 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1102 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1105 * This can't happen, the type is validated several times
1106 * before we get here.
1113 * put a workspace struct back on the list or free it if we have enough
1114 * idle ones sitting around
1116 void btrfs_put_workspace(int type, struct list_head *ws)
1118 struct workspace_manager *wsm;
1119 struct list_head *idle_ws;
1120 spinlock_t *ws_lock;
1122 wait_queue_head_t *ws_wait;
1125 wsm = btrfs_compress_op[type]->workspace_manager;
1126 idle_ws = &wsm->idle_ws;
1127 ws_lock = &wsm->ws_lock;
1128 total_ws = &wsm->total_ws;
1129 ws_wait = &wsm->ws_wait;
1130 free_ws = &wsm->free_ws;
1133 if (*free_ws <= num_online_cpus()) {
1134 list_add(ws, idle_ws);
1136 spin_unlock(ws_lock);
1139 spin_unlock(ws_lock);
1141 free_workspace(type, ws);
1142 atomic_dec(total_ws);
1144 cond_wake_up(ws_wait);
1147 static void put_workspace(int type, struct list_head *ws)
1150 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1151 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1152 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1153 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1156 * This can't happen, the type is validated several times
1157 * before we get here.
1164 * Adjust @level according to the limits of the compression algorithm or
1165 * fallback to default
1167 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1169 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1172 level = ops->default_level;
1174 level = min(level, ops->max_level);
1180 * Given an address space and start and length, compress the bytes into @pages
1181 * that are allocated on demand.
1183 * @type_level is encoded algorithm and level, where level 0 means whatever
1184 * default the algorithm chooses and is opaque here;
1185 * - compression algo are 0-3
1186 * - the level are bits 4-7
1188 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1189 * and returns number of actually allocated pages
1191 * @total_in is used to return the number of bytes actually read. It
1192 * may be smaller than the input length if we had to exit early because we
1193 * ran out of room in the pages array or because we cross the
1194 * max_out threshold.
1196 * @total_out is an in/out parameter, must be set to the input length and will
1197 * be also used to return the total number of compressed bytes
1199 * @max_out tells us the max number of bytes that we're allowed to
1202 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1203 u64 start, struct page **pages,
1204 unsigned long *out_pages,
1205 unsigned long *total_in,
1206 unsigned long *total_out)
1208 int type = btrfs_compress_type(type_level);
1209 int level = btrfs_compress_level(type_level);
1210 struct list_head *workspace;
1213 level = btrfs_compress_set_level(type, level);
1214 workspace = get_workspace(type, level);
1215 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1216 out_pages, total_in, total_out);
1217 put_workspace(type, workspace);
1222 * pages_in is an array of pages with compressed data.
1224 * disk_start is the starting logical offset of this array in the file
1226 * orig_bio contains the pages from the file that we want to decompress into
1228 * srclen is the number of bytes in pages_in
1230 * The basic idea is that we have a bio that was created by readpages.
1231 * The pages in the bio are for the uncompressed data, and they may not
1232 * be contiguous. They all correspond to the range of bytes covered by
1233 * the compressed extent.
1235 static int btrfs_decompress_bio(struct compressed_bio *cb)
1237 struct list_head *workspace;
1239 int type = cb->compress_type;
1241 workspace = get_workspace(type, 0);
1242 ret = compression_decompress_bio(type, workspace, cb);
1243 put_workspace(type, workspace);
1249 * a less complex decompression routine. Our compressed data fits in a
1250 * single page, and we want to read a single page out of it.
1251 * start_byte tells us the offset into the compressed data we're interested in
1253 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1254 unsigned long start_byte, size_t srclen, size_t destlen)
1256 struct list_head *workspace;
1259 workspace = get_workspace(type, 0);
1260 ret = compression_decompress(type, workspace, data_in, dest_page,
1261 start_byte, srclen, destlen);
1262 put_workspace(type, workspace);
1267 void __init btrfs_init_compress(void)
1269 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1270 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1271 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1272 zstd_init_workspace_manager();
1275 void __cold btrfs_exit_compress(void)
1277 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1278 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1279 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1280 zstd_cleanup_workspace_manager();
1284 * Copy uncompressed data from working buffer to pages.
1286 * buf_start is the byte offset we're of the start of our workspace buffer.
1288 * total_out is the last byte of the buffer
1290 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1291 unsigned long total_out, u64 disk_start,
1294 unsigned long buf_offset;
1295 unsigned long current_buf_start;
1296 unsigned long start_byte;
1297 unsigned long prev_start_byte;
1298 unsigned long working_bytes = total_out - buf_start;
1299 unsigned long bytes;
1300 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1303 * start byte is the first byte of the page we're currently
1304 * copying into relative to the start of the compressed data.
1306 start_byte = page_offset(bvec.bv_page) - disk_start;
1308 /* we haven't yet hit data corresponding to this page */
1309 if (total_out <= start_byte)
1313 * the start of the data we care about is offset into
1314 * the middle of our working buffer
1316 if (total_out > start_byte && buf_start < start_byte) {
1317 buf_offset = start_byte - buf_start;
1318 working_bytes -= buf_offset;
1322 current_buf_start = buf_start;
1324 /* copy bytes from the working buffer into the pages */
1325 while (working_bytes > 0) {
1326 bytes = min_t(unsigned long, bvec.bv_len,
1327 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1328 bytes = min(bytes, working_bytes);
1330 memcpy_to_page(bvec.bv_page, bvec.bv_offset, buf + buf_offset,
1332 flush_dcache_page(bvec.bv_page);
1334 buf_offset += bytes;
1335 working_bytes -= bytes;
1336 current_buf_start += bytes;
1338 /* check if we need to pick another page */
1339 bio_advance(bio, bytes);
1340 if (!bio->bi_iter.bi_size)
1342 bvec = bio_iter_iovec(bio, bio->bi_iter);
1343 prev_start_byte = start_byte;
1344 start_byte = page_offset(bvec.bv_page) - disk_start;
1347 * We need to make sure we're only adjusting
1348 * our offset into compression working buffer when
1349 * we're switching pages. Otherwise we can incorrectly
1350 * keep copying when we were actually done.
1352 if (start_byte != prev_start_byte) {
1354 * make sure our new page is covered by this
1357 if (total_out <= start_byte)
1361 * the next page in the biovec might not be adjacent
1362 * to the last page, but it might still be found
1363 * inside this working buffer. bump our offset pointer
1365 if (total_out > start_byte &&
1366 current_buf_start < start_byte) {
1367 buf_offset = start_byte - buf_start;
1368 working_bytes = total_out - start_byte;
1369 current_buf_start = buf_start + buf_offset;
1378 * Shannon Entropy calculation
1380 * Pure byte distribution analysis fails to determine compressibility of data.
1381 * Try calculating entropy to estimate the average minimum number of bits
1382 * needed to encode the sampled data.
1384 * For convenience, return the percentage of needed bits, instead of amount of
1387 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1388 * and can be compressible with high probability
1390 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1392 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1394 #define ENTROPY_LVL_ACEPTABLE (65)
1395 #define ENTROPY_LVL_HIGH (80)
1398 * For increasead precision in shannon_entropy calculation,
1399 * let's do pow(n, M) to save more digits after comma:
1401 * - maximum int bit length is 64
1402 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1403 * - 13 * 4 = 52 < 64 -> M = 4
1407 static inline u32 ilog2_w(u64 n)
1409 return ilog2(n * n * n * n);
1412 static u32 shannon_entropy(struct heuristic_ws *ws)
1414 const u32 entropy_max = 8 * ilog2_w(2);
1415 u32 entropy_sum = 0;
1416 u32 p, p_base, sz_base;
1419 sz_base = ilog2_w(ws->sample_size);
1420 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1421 p = ws->bucket[i].count;
1422 p_base = ilog2_w(p);
1423 entropy_sum += p * (sz_base - p_base);
1426 entropy_sum /= ws->sample_size;
1427 return entropy_sum * 100 / entropy_max;
1430 #define RADIX_BASE 4U
1431 #define COUNTERS_SIZE (1U << RADIX_BASE)
1433 static u8 get4bits(u64 num, int shift) {
1438 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1443 * Use 4 bits as radix base
1444 * Use 16 u32 counters for calculating new position in buf array
1446 * @array - array that will be sorted
1447 * @array_buf - buffer array to store sorting results
1448 * must be equal in size to @array
1451 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1456 u32 counters[COUNTERS_SIZE];
1464 * Try avoid useless loop iterations for small numbers stored in big
1465 * counters. Example: 48 33 4 ... in 64bit array
1467 max_num = array[0].count;
1468 for (i = 1; i < num; i++) {
1469 buf_num = array[i].count;
1470 if (buf_num > max_num)
1474 buf_num = ilog2(max_num);
1475 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1478 while (shift < bitlen) {
1479 memset(counters, 0, sizeof(counters));
1481 for (i = 0; i < num; i++) {
1482 buf_num = array[i].count;
1483 addr = get4bits(buf_num, shift);
1487 for (i = 1; i < COUNTERS_SIZE; i++)
1488 counters[i] += counters[i - 1];
1490 for (i = num - 1; i >= 0; i--) {
1491 buf_num = array[i].count;
1492 addr = get4bits(buf_num, shift);
1494 new_addr = counters[addr];
1495 array_buf[new_addr] = array[i];
1498 shift += RADIX_BASE;
1501 * Normal radix expects to move data from a temporary array, to
1502 * the main one. But that requires some CPU time. Avoid that
1503 * by doing another sort iteration to original array instead of
1506 memset(counters, 0, sizeof(counters));
1508 for (i = 0; i < num; i ++) {
1509 buf_num = array_buf[i].count;
1510 addr = get4bits(buf_num, shift);
1514 for (i = 1; i < COUNTERS_SIZE; i++)
1515 counters[i] += counters[i - 1];
1517 for (i = num - 1; i >= 0; i--) {
1518 buf_num = array_buf[i].count;
1519 addr = get4bits(buf_num, shift);
1521 new_addr = counters[addr];
1522 array[new_addr] = array_buf[i];
1525 shift += RADIX_BASE;
1530 * Size of the core byte set - how many bytes cover 90% of the sample
1532 * There are several types of structured binary data that use nearly all byte
1533 * values. The distribution can be uniform and counts in all buckets will be
1534 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1536 * Other possibility is normal (Gaussian) distribution, where the data could
1537 * be potentially compressible, but we have to take a few more steps to decide
1540 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1541 * compression algo can easy fix that
1542 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1543 * probability is not compressible
1545 #define BYTE_CORE_SET_LOW (64)
1546 #define BYTE_CORE_SET_HIGH (200)
1548 static int byte_core_set_size(struct heuristic_ws *ws)
1551 u32 coreset_sum = 0;
1552 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1553 struct bucket_item *bucket = ws->bucket;
1555 /* Sort in reverse order */
1556 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1558 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1559 coreset_sum += bucket[i].count;
1561 if (coreset_sum > core_set_threshold)
1564 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1565 coreset_sum += bucket[i].count;
1566 if (coreset_sum > core_set_threshold)
1574 * Count byte values in buckets.
1575 * This heuristic can detect textual data (configs, xml, json, html, etc).
1576 * Because in most text-like data byte set is restricted to limited number of
1577 * possible characters, and that restriction in most cases makes data easy to
1580 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1581 * less - compressible
1582 * more - need additional analysis
1584 #define BYTE_SET_THRESHOLD (64)
1586 static u32 byte_set_size(const struct heuristic_ws *ws)
1589 u32 byte_set_size = 0;
1591 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1592 if (ws->bucket[i].count > 0)
1597 * Continue collecting count of byte values in buckets. If the byte
1598 * set size is bigger then the threshold, it's pointless to continue,
1599 * the detection technique would fail for this type of data.
1601 for (; i < BUCKET_SIZE; i++) {
1602 if (ws->bucket[i].count > 0) {
1604 if (byte_set_size > BYTE_SET_THRESHOLD)
1605 return byte_set_size;
1609 return byte_set_size;
1612 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1614 const u32 half_of_sample = ws->sample_size / 2;
1615 const u8 *data = ws->sample;
1617 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1620 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1621 struct heuristic_ws *ws)
1624 u64 index, index_end;
1625 u32 i, curr_sample_pos;
1629 * Compression handles the input data by chunks of 128KiB
1630 * (defined by BTRFS_MAX_UNCOMPRESSED)
1632 * We do the same for the heuristic and loop over the whole range.
1634 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1635 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1637 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1638 end = start + BTRFS_MAX_UNCOMPRESSED;
1640 index = start >> PAGE_SHIFT;
1641 index_end = end >> PAGE_SHIFT;
1643 /* Don't miss unaligned end */
1644 if (!IS_ALIGNED(end, PAGE_SIZE))
1647 curr_sample_pos = 0;
1648 while (index < index_end) {
1649 page = find_get_page(inode->i_mapping, index);
1650 in_data = kmap_local_page(page);
1651 /* Handle case where the start is not aligned to PAGE_SIZE */
1652 i = start % PAGE_SIZE;
1653 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1654 /* Don't sample any garbage from the last page */
1655 if (start > end - SAMPLING_READ_SIZE)
1657 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1658 SAMPLING_READ_SIZE);
1659 i += SAMPLING_INTERVAL;
1660 start += SAMPLING_INTERVAL;
1661 curr_sample_pos += SAMPLING_READ_SIZE;
1663 kunmap_local(in_data);
1669 ws->sample_size = curr_sample_pos;
1673 * Compression heuristic.
1675 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1676 * quickly (compared to direct compression) detect data characteristics
1677 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1680 * The following types of analysis can be performed:
1681 * - detect mostly zero data
1682 * - detect data with low "byte set" size (text, etc)
1683 * - detect data with low/high "core byte" set
1685 * Return non-zero if the compression should be done, 0 otherwise.
1687 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1689 struct list_head *ws_list = get_workspace(0, 0);
1690 struct heuristic_ws *ws;
1695 ws = list_entry(ws_list, struct heuristic_ws, list);
1697 heuristic_collect_sample(inode, start, end, ws);
1699 if (sample_repeated_patterns(ws)) {
1704 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1706 for (i = 0; i < ws->sample_size; i++) {
1707 byte = ws->sample[i];
1708 ws->bucket[byte].count++;
1711 i = byte_set_size(ws);
1712 if (i < BYTE_SET_THRESHOLD) {
1717 i = byte_core_set_size(ws);
1718 if (i <= BYTE_CORE_SET_LOW) {
1723 if (i >= BYTE_CORE_SET_HIGH) {
1728 i = shannon_entropy(ws);
1729 if (i <= ENTROPY_LVL_ACEPTABLE) {
1735 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1736 * needed to give green light to compression.
1738 * For now just assume that compression at that level is not worth the
1739 * resources because:
1741 * 1. it is possible to defrag the data later
1743 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1744 * values, every bucket has counter at level ~54. The heuristic would
1745 * be confused. This can happen when data have some internal repeated
1746 * patterns like "abbacbbc...". This can be detected by analyzing
1747 * pairs of bytes, which is too costly.
1749 if (i < ENTROPY_LVL_HIGH) {
1758 put_workspace(0, ws_list);
1763 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1764 * level, unrecognized string will set the default level
1766 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1768 unsigned int level = 0;
1774 if (str[0] == ':') {
1775 ret = kstrtouint(str + 1, 10, &level);
1780 level = btrfs_compress_set_level(type, level);