1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Oracle. All rights reserved.
6 #include <linux/kernel.h>
8 #include <linux/file.h>
10 #include <linux/pagemap.h>
11 #include <linux/highmem.h>
12 #include <linux/time.h>
13 #include <linux/init.h>
14 #include <linux/string.h>
15 #include <linux/backing-dev.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sched/mm.h>
19 #include <linux/log2.h>
20 #include <crypto/hash.h>
24 #include "transaction.h"
25 #include "btrfs_inode.h"
27 #include "ordered-data.h"
28 #include "compression.h"
29 #include "extent_io.h"
30 #include "extent_map.h"
32 int zlib_compress_pages(struct list_head *ws, struct address_space *mapping,
33 u64 start, struct page **pages, unsigned long *out_pages,
34 unsigned long *total_in, unsigned long *total_out);
35 int zlib_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
36 int zlib_decompress(struct list_head *ws, unsigned char *data_in,
37 struct page *dest_page, unsigned long start_byte, size_t srclen,
39 struct list_head *zlib_alloc_workspace(unsigned int level);
40 void zlib_free_workspace(struct list_head *ws);
41 struct list_head *zlib_get_workspace(unsigned int level);
43 int lzo_compress_pages(struct list_head *ws, struct address_space *mapping,
44 u64 start, struct page **pages, unsigned long *out_pages,
45 unsigned long *total_in, unsigned long *total_out);
46 int lzo_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
47 int lzo_decompress(struct list_head *ws, unsigned char *data_in,
48 struct page *dest_page, unsigned long start_byte, size_t srclen,
50 struct list_head *lzo_alloc_workspace(unsigned int level);
51 void lzo_free_workspace(struct list_head *ws);
53 int zstd_compress_pages(struct list_head *ws, struct address_space *mapping,
54 u64 start, struct page **pages, unsigned long *out_pages,
55 unsigned long *total_in, unsigned long *total_out);
56 int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb);
57 int zstd_decompress(struct list_head *ws, unsigned char *data_in,
58 struct page *dest_page, unsigned long start_byte, size_t srclen,
60 void zstd_init_workspace_manager(void);
61 void zstd_cleanup_workspace_manager(void);
62 struct list_head *zstd_alloc_workspace(unsigned int level);
63 void zstd_free_workspace(struct list_head *ws);
64 struct list_head *zstd_get_workspace(unsigned int level);
65 void zstd_put_workspace(struct list_head *ws);
67 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
69 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
72 case BTRFS_COMPRESS_ZLIB:
73 case BTRFS_COMPRESS_LZO:
74 case BTRFS_COMPRESS_ZSTD:
75 case BTRFS_COMPRESS_NONE:
76 return btrfs_compress_types[type];
84 bool btrfs_compress_is_valid_type(const char *str, size_t len)
88 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
89 size_t comp_len = strlen(btrfs_compress_types[i]);
94 if (!strncmp(btrfs_compress_types[i], str, comp_len))
100 static int compression_compress_pages(int type, struct list_head *ws,
101 struct address_space *mapping, u64 start, struct page **pages,
102 unsigned long *out_pages, unsigned long *total_in,
103 unsigned long *total_out)
106 case BTRFS_COMPRESS_ZLIB:
107 return zlib_compress_pages(ws, mapping, start, pages,
108 out_pages, total_in, total_out);
109 case BTRFS_COMPRESS_LZO:
110 return lzo_compress_pages(ws, mapping, start, pages,
111 out_pages, total_in, total_out);
112 case BTRFS_COMPRESS_ZSTD:
113 return zstd_compress_pages(ws, mapping, start, pages,
114 out_pages, total_in, total_out);
115 case BTRFS_COMPRESS_NONE:
118 * This can't happen, the type is validated several times
119 * before we get here. As a sane fallback, return what the
120 * callers will understand as 'no compression happened'.
126 static int compression_decompress_bio(int type, struct list_head *ws,
127 struct compressed_bio *cb)
130 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
131 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
132 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
133 case BTRFS_COMPRESS_NONE:
136 * This can't happen, the type is validated several times
137 * before we get here.
143 static int compression_decompress(int type, struct list_head *ws,
144 unsigned char *data_in, struct page *dest_page,
145 unsigned long start_byte, size_t srclen, size_t destlen)
148 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
149 start_byte, srclen, destlen);
150 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
151 start_byte, srclen, destlen);
152 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
153 start_byte, srclen, destlen);
154 case BTRFS_COMPRESS_NONE:
157 * This can't happen, the type is validated several times
158 * before we get here.
164 static int btrfs_decompress_bio(struct compressed_bio *cb);
166 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
167 unsigned long disk_size)
169 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
171 return sizeof(struct compressed_bio) +
172 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
175 static int check_compressed_csum(struct btrfs_inode *inode,
176 struct compressed_bio *cb,
179 struct btrfs_fs_info *fs_info = inode->root->fs_info;
180 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
181 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
186 u8 csum[BTRFS_CSUM_SIZE];
187 u8 *cb_sum = cb->sums;
189 if (inode->flags & BTRFS_INODE_NODATASUM)
192 shash->tfm = fs_info->csum_shash;
194 for (i = 0; i < cb->nr_pages; i++) {
195 page = cb->compressed_pages[i];
197 kaddr = kmap_atomic(page);
198 crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum);
199 kunmap_atomic(kaddr);
201 if (memcmp(&csum, cb_sum, csum_size)) {
202 btrfs_print_data_csum_error(inode, disk_start,
203 csum, cb_sum, cb->mirror_num);
215 /* when we finish reading compressed pages from the disk, we
216 * decompress them and then run the bio end_io routines on the
217 * decompressed pages (in the inode address space).
219 * This allows the checksumming and other IO error handling routines
222 * The compressed pages are freed here, and it must be run
225 static void end_compressed_bio_read(struct bio *bio)
227 struct compressed_bio *cb = bio->bi_private;
231 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
237 /* if there are more bios still pending for this compressed
240 if (!refcount_dec_and_test(&cb->pending_bios))
244 * Record the correct mirror_num in cb->orig_bio so that
245 * read-repair can work properly.
247 ASSERT(btrfs_io_bio(cb->orig_bio));
248 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
249 cb->mirror_num = mirror;
252 * Some IO in this cb have failed, just skip checksum as there
253 * is no way it could be correct.
259 ret = check_compressed_csum(BTRFS_I(inode), cb,
260 (u64)bio->bi_iter.bi_sector << 9);
264 /* ok, we're the last bio for this extent, lets start
267 ret = btrfs_decompress_bio(cb);
273 /* release the compressed pages */
275 for (index = 0; index < cb->nr_pages; index++) {
276 page = cb->compressed_pages[index];
277 page->mapping = NULL;
281 /* do io completion on the original bio */
283 bio_io_error(cb->orig_bio);
285 struct bio_vec *bvec;
286 struct bvec_iter_all iter_all;
289 * we have verified the checksum already, set page
290 * checked so the end_io handlers know about it
292 ASSERT(!bio_flagged(bio, BIO_CLONED));
293 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
294 SetPageChecked(bvec->bv_page);
296 bio_endio(cb->orig_bio);
299 /* finally free the cb struct */
300 kfree(cb->compressed_pages);
307 * Clear the writeback bits on all of the file
308 * pages for a compressed write
310 static noinline void end_compressed_writeback(struct inode *inode,
311 const struct compressed_bio *cb)
313 unsigned long index = cb->start >> PAGE_SHIFT;
314 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
315 struct page *pages[16];
316 unsigned long nr_pages = end_index - index + 1;
321 mapping_set_error(inode->i_mapping, -EIO);
323 while (nr_pages > 0) {
324 ret = find_get_pages_contig(inode->i_mapping, index,
326 nr_pages, ARRAY_SIZE(pages)), pages);
332 for (i = 0; i < ret; i++) {
334 SetPageError(pages[i]);
335 end_page_writeback(pages[i]);
341 /* the inode may be gone now */
345 * do the cleanup once all the compressed pages hit the disk.
346 * This will clear writeback on the file pages and free the compressed
349 * This also calls the writeback end hooks for the file pages so that
350 * metadata and checksums can be updated in the file.
352 static void end_compressed_bio_write(struct bio *bio)
354 struct compressed_bio *cb = bio->bi_private;
362 /* if there are more bios still pending for this compressed
365 if (!refcount_dec_and_test(&cb->pending_bios))
368 /* ok, we're the last bio for this extent, step one is to
369 * call back into the FS and do all the end_io operations
372 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
373 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
374 cb->start, cb->start + cb->len - 1,
375 bio->bi_status == BLK_STS_OK);
376 cb->compressed_pages[0]->mapping = NULL;
378 end_compressed_writeback(inode, cb);
379 /* note, our inode could be gone now */
382 * release the compressed pages, these came from alloc_page and
383 * are not attached to the inode at all
386 for (index = 0; index < cb->nr_pages; index++) {
387 page = cb->compressed_pages[index];
388 page->mapping = NULL;
392 /* finally free the cb struct */
393 kfree(cb->compressed_pages);
400 * worker function to build and submit bios for previously compressed pages.
401 * The corresponding pages in the inode should be marked for writeback
402 * and the compressed pages should have a reference on them for dropping
403 * when the IO is complete.
405 * This also checksums the file bytes and gets things ready for
408 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
409 unsigned long len, u64 disk_start,
410 unsigned long compressed_len,
411 struct page **compressed_pages,
412 unsigned long nr_pages,
413 unsigned int write_flags,
414 struct cgroup_subsys_state *blkcg_css)
416 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
417 struct bio *bio = NULL;
418 struct compressed_bio *cb;
419 unsigned long bytes_left;
422 u64 first_byte = disk_start;
424 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
426 WARN_ON(!PAGE_ALIGNED(start));
427 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
429 return BLK_STS_RESOURCE;
430 refcount_set(&cb->pending_bios, 0);
436 cb->compressed_pages = compressed_pages;
437 cb->compressed_len = compressed_len;
439 cb->nr_pages = nr_pages;
441 bio = btrfs_bio_alloc(first_byte);
442 bio->bi_opf = REQ_OP_WRITE | write_flags;
443 bio->bi_private = cb;
444 bio->bi_end_io = end_compressed_bio_write;
447 bio->bi_opf |= REQ_CGROUP_PUNT;
448 kthread_associate_blkcg(blkcg_css);
450 refcount_set(&cb->pending_bios, 1);
452 /* create and submit bios for the compressed pages */
453 bytes_left = compressed_len;
454 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
457 page = compressed_pages[pg_index];
458 page->mapping = inode->i_mapping;
459 if (bio->bi_iter.bi_size)
460 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
463 page->mapping = NULL;
464 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
467 * inc the count before we submit the bio so
468 * we know the end IO handler won't happen before
469 * we inc the count. Otherwise, the cb might get
470 * freed before we're done setting it up
472 refcount_inc(&cb->pending_bios);
473 ret = btrfs_bio_wq_end_io(fs_info, bio,
474 BTRFS_WQ_ENDIO_DATA);
475 BUG_ON(ret); /* -ENOMEM */
478 ret = btrfs_csum_one_bio(inode, bio, start, 1);
479 BUG_ON(ret); /* -ENOMEM */
482 ret = btrfs_map_bio(fs_info, bio, 0);
484 bio->bi_status = ret;
488 bio = btrfs_bio_alloc(first_byte);
489 bio->bi_opf = REQ_OP_WRITE | write_flags;
490 bio->bi_private = cb;
491 bio->bi_end_io = end_compressed_bio_write;
493 bio->bi_opf |= REQ_CGROUP_PUNT;
494 bio_add_page(bio, page, PAGE_SIZE, 0);
496 if (bytes_left < PAGE_SIZE) {
498 "bytes left %lu compress len %lu nr %lu",
499 bytes_left, cb->compressed_len, cb->nr_pages);
501 bytes_left -= PAGE_SIZE;
502 first_byte += PAGE_SIZE;
506 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
507 BUG_ON(ret); /* -ENOMEM */
510 ret = btrfs_csum_one_bio(inode, bio, start, 1);
511 BUG_ON(ret); /* -ENOMEM */
514 ret = btrfs_map_bio(fs_info, bio, 0);
516 bio->bi_status = ret;
521 kthread_associate_blkcg(NULL);
526 static u64 bio_end_offset(struct bio *bio)
528 struct bio_vec *last = bio_last_bvec_all(bio);
530 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
533 static noinline int add_ra_bio_pages(struct inode *inode,
535 struct compressed_bio *cb)
537 unsigned long end_index;
538 unsigned long pg_index;
540 u64 isize = i_size_read(inode);
543 unsigned long nr_pages = 0;
544 struct extent_map *em;
545 struct address_space *mapping = inode->i_mapping;
546 struct extent_map_tree *em_tree;
547 struct extent_io_tree *tree;
551 last_offset = bio_end_offset(cb->orig_bio);
552 em_tree = &BTRFS_I(inode)->extent_tree;
553 tree = &BTRFS_I(inode)->io_tree;
558 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
560 while (last_offset < compressed_end) {
561 pg_index = last_offset >> PAGE_SHIFT;
563 if (pg_index > end_index)
566 page = xa_load(&mapping->i_pages, pg_index);
567 if (page && !xa_is_value(page)) {
574 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
579 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
584 end = last_offset + PAGE_SIZE - 1;
586 * at this point, we have a locked page in the page cache
587 * for these bytes in the file. But, we have to make
588 * sure they map to this compressed extent on disk.
590 set_page_extent_mapped(page);
591 lock_extent(tree, last_offset, end);
592 read_lock(&em_tree->lock);
593 em = lookup_extent_mapping(em_tree, last_offset,
595 read_unlock(&em_tree->lock);
597 if (!em || last_offset < em->start ||
598 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
599 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
601 unlock_extent(tree, last_offset, end);
608 if (page->index == end_index) {
610 size_t zero_offset = offset_in_page(isize);
614 zeros = PAGE_SIZE - zero_offset;
615 userpage = kmap_atomic(page);
616 memset(userpage + zero_offset, 0, zeros);
617 flush_dcache_page(page);
618 kunmap_atomic(userpage);
622 ret = bio_add_page(cb->orig_bio, page,
625 if (ret == PAGE_SIZE) {
629 unlock_extent(tree, last_offset, end);
635 last_offset += PAGE_SIZE;
641 * for a compressed read, the bio we get passed has all the inode pages
642 * in it. We don't actually do IO on those pages but allocate new ones
643 * to hold the compressed pages on disk.
645 * bio->bi_iter.bi_sector points to the compressed extent on disk
646 * bio->bi_io_vec points to all of the inode pages
648 * After the compressed pages are read, we copy the bytes into the
649 * bio we were passed and then call the bio end_io calls
651 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
652 int mirror_num, unsigned long bio_flags)
654 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
655 struct extent_map_tree *em_tree;
656 struct compressed_bio *cb;
657 unsigned long compressed_len;
658 unsigned long nr_pages;
659 unsigned long pg_index;
661 struct bio *comp_bio;
662 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
665 struct extent_map *em;
666 blk_status_t ret = BLK_STS_RESOURCE;
668 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
671 em_tree = &BTRFS_I(inode)->extent_tree;
673 /* we need the actual starting offset of this extent in the file */
674 read_lock(&em_tree->lock);
675 em = lookup_extent_mapping(em_tree,
676 page_offset(bio_first_page_all(bio)),
678 read_unlock(&em_tree->lock);
680 return BLK_STS_IOERR;
682 compressed_len = em->block_len;
683 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
687 refcount_set(&cb->pending_bios, 0);
690 cb->mirror_num = mirror_num;
693 cb->start = em->orig_start;
695 em_start = em->start;
700 cb->len = bio->bi_iter.bi_size;
701 cb->compressed_len = compressed_len;
702 cb->compress_type = extent_compress_type(bio_flags);
705 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
706 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
708 if (!cb->compressed_pages)
711 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
712 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
714 if (!cb->compressed_pages[pg_index]) {
715 faili = pg_index - 1;
716 ret = BLK_STS_RESOURCE;
720 faili = nr_pages - 1;
721 cb->nr_pages = nr_pages;
723 add_ra_bio_pages(inode, em_start + em_len, cb);
725 /* include any pages we added in add_ra-bio_pages */
726 cb->len = bio->bi_iter.bi_size;
728 comp_bio = btrfs_bio_alloc(cur_disk_byte);
729 comp_bio->bi_opf = REQ_OP_READ;
730 comp_bio->bi_private = cb;
731 comp_bio->bi_end_io = end_compressed_bio_read;
732 refcount_set(&cb->pending_bios, 1);
734 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
737 page = cb->compressed_pages[pg_index];
738 page->mapping = inode->i_mapping;
739 page->index = em_start >> PAGE_SHIFT;
741 if (comp_bio->bi_iter.bi_size)
742 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
745 page->mapping = NULL;
746 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
748 unsigned int nr_sectors;
750 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
751 BTRFS_WQ_ENDIO_DATA);
752 BUG_ON(ret); /* -ENOMEM */
755 * inc the count before we submit the bio so
756 * we know the end IO handler won't happen before
757 * we inc the count. Otherwise, the cb might get
758 * freed before we're done setting it up
760 refcount_inc(&cb->pending_bios);
762 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
763 ret = btrfs_lookup_bio_sums(inode, comp_bio,
765 BUG_ON(ret); /* -ENOMEM */
768 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
769 fs_info->sectorsize);
770 sums += csum_size * nr_sectors;
772 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
774 comp_bio->bi_status = ret;
778 comp_bio = btrfs_bio_alloc(cur_disk_byte);
779 comp_bio->bi_opf = REQ_OP_READ;
780 comp_bio->bi_private = cb;
781 comp_bio->bi_end_io = end_compressed_bio_read;
783 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
785 cur_disk_byte += PAGE_SIZE;
788 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
789 BUG_ON(ret); /* -ENOMEM */
791 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
792 ret = btrfs_lookup_bio_sums(inode, comp_bio, (u64)-1, sums);
793 BUG_ON(ret); /* -ENOMEM */
796 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
798 comp_bio->bi_status = ret;
806 __free_page(cb->compressed_pages[faili]);
810 kfree(cb->compressed_pages);
819 * Heuristic uses systematic sampling to collect data from the input data
820 * range, the logic can be tuned by the following constants:
822 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
823 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
825 #define SAMPLING_READ_SIZE (16)
826 #define SAMPLING_INTERVAL (256)
829 * For statistical analysis of the input data we consider bytes that form a
830 * Galois Field of 256 objects. Each object has an attribute count, ie. how
831 * many times the object appeared in the sample.
833 #define BUCKET_SIZE (256)
836 * The size of the sample is based on a statistical sampling rule of thumb.
837 * The common way is to perform sampling tests as long as the number of
838 * elements in each cell is at least 5.
840 * Instead of 5, we choose 32 to obtain more accurate results.
841 * If the data contain the maximum number of symbols, which is 256, we obtain a
842 * sample size bound by 8192.
844 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
845 * from up to 512 locations.
847 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
848 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
854 struct heuristic_ws {
855 /* Partial copy of input data */
858 /* Buckets store counters for each byte value */
859 struct bucket_item *bucket;
861 struct bucket_item *bucket_b;
862 struct list_head list;
865 static struct workspace_manager heuristic_wsm;
867 static void free_heuristic_ws(struct list_head *ws)
869 struct heuristic_ws *workspace;
871 workspace = list_entry(ws, struct heuristic_ws, list);
873 kvfree(workspace->sample);
874 kfree(workspace->bucket);
875 kfree(workspace->bucket_b);
879 static struct list_head *alloc_heuristic_ws(unsigned int level)
881 struct heuristic_ws *ws;
883 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
885 return ERR_PTR(-ENOMEM);
887 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
891 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
895 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
899 INIT_LIST_HEAD(&ws->list);
902 free_heuristic_ws(&ws->list);
903 return ERR_PTR(-ENOMEM);
906 const struct btrfs_compress_op btrfs_heuristic_compress = {
907 .workspace_manager = &heuristic_wsm,
910 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
911 /* The heuristic is represented as compression type 0 */
912 &btrfs_heuristic_compress,
913 &btrfs_zlib_compress,
915 &btrfs_zstd_compress,
918 static struct list_head *alloc_workspace(int type, unsigned int level)
921 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
922 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
923 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
924 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
927 * This can't happen, the type is validated several times
928 * before we get here.
934 static void free_workspace(int type, struct list_head *ws)
937 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
938 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
939 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
940 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
943 * This can't happen, the type is validated several times
944 * before we get here.
950 static void btrfs_init_workspace_manager(int type)
952 struct workspace_manager *wsm;
953 struct list_head *workspace;
955 wsm = btrfs_compress_op[type]->workspace_manager;
956 INIT_LIST_HEAD(&wsm->idle_ws);
957 spin_lock_init(&wsm->ws_lock);
958 atomic_set(&wsm->total_ws, 0);
959 init_waitqueue_head(&wsm->ws_wait);
962 * Preallocate one workspace for each compression type so we can
963 * guarantee forward progress in the worst case
965 workspace = alloc_workspace(type, 0);
966 if (IS_ERR(workspace)) {
968 "BTRFS: cannot preallocate compression workspace, will try later\n");
970 atomic_set(&wsm->total_ws, 1);
972 list_add(workspace, &wsm->idle_ws);
976 static void btrfs_cleanup_workspace_manager(int type)
978 struct workspace_manager *wsman;
979 struct list_head *ws;
981 wsman = btrfs_compress_op[type]->workspace_manager;
982 while (!list_empty(&wsman->idle_ws)) {
983 ws = wsman->idle_ws.next;
985 free_workspace(type, ws);
986 atomic_dec(&wsman->total_ws);
991 * This finds an available workspace or allocates a new one.
992 * If it's not possible to allocate a new one, waits until there's one.
993 * Preallocation makes a forward progress guarantees and we do not return
996 struct list_head *btrfs_get_workspace(int type, unsigned int level)
998 struct workspace_manager *wsm;
999 struct list_head *workspace;
1000 int cpus = num_online_cpus();
1002 struct list_head *idle_ws;
1003 spinlock_t *ws_lock;
1005 wait_queue_head_t *ws_wait;
1008 wsm = btrfs_compress_op[type]->workspace_manager;
1009 idle_ws = &wsm->idle_ws;
1010 ws_lock = &wsm->ws_lock;
1011 total_ws = &wsm->total_ws;
1012 ws_wait = &wsm->ws_wait;
1013 free_ws = &wsm->free_ws;
1017 if (!list_empty(idle_ws)) {
1018 workspace = idle_ws->next;
1019 list_del(workspace);
1021 spin_unlock(ws_lock);
1025 if (atomic_read(total_ws) > cpus) {
1028 spin_unlock(ws_lock);
1029 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1030 if (atomic_read(total_ws) > cpus && !*free_ws)
1032 finish_wait(ws_wait, &wait);
1035 atomic_inc(total_ws);
1036 spin_unlock(ws_lock);
1039 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1040 * to turn it off here because we might get called from the restricted
1041 * context of btrfs_compress_bio/btrfs_compress_pages
1043 nofs_flag = memalloc_nofs_save();
1044 workspace = alloc_workspace(type, level);
1045 memalloc_nofs_restore(nofs_flag);
1047 if (IS_ERR(workspace)) {
1048 atomic_dec(total_ws);
1052 * Do not return the error but go back to waiting. There's a
1053 * workspace preallocated for each type and the compression
1054 * time is bounded so we get to a workspace eventually. This
1055 * makes our caller's life easier.
1057 * To prevent silent and low-probability deadlocks (when the
1058 * initial preallocation fails), check if there are any
1059 * workspaces at all.
1061 if (atomic_read(total_ws) == 0) {
1062 static DEFINE_RATELIMIT_STATE(_rs,
1063 /* once per minute */ 60 * HZ,
1066 if (__ratelimit(&_rs)) {
1067 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1075 static struct list_head *get_workspace(int type, int level)
1078 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1079 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1080 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1081 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1084 * This can't happen, the type is validated several times
1085 * before we get here.
1092 * put a workspace struct back on the list or free it if we have enough
1093 * idle ones sitting around
1095 void btrfs_put_workspace(int type, struct list_head *ws)
1097 struct workspace_manager *wsm;
1098 struct list_head *idle_ws;
1099 spinlock_t *ws_lock;
1101 wait_queue_head_t *ws_wait;
1104 wsm = btrfs_compress_op[type]->workspace_manager;
1105 idle_ws = &wsm->idle_ws;
1106 ws_lock = &wsm->ws_lock;
1107 total_ws = &wsm->total_ws;
1108 ws_wait = &wsm->ws_wait;
1109 free_ws = &wsm->free_ws;
1112 if (*free_ws <= num_online_cpus()) {
1113 list_add(ws, idle_ws);
1115 spin_unlock(ws_lock);
1118 spin_unlock(ws_lock);
1120 free_workspace(type, ws);
1121 atomic_dec(total_ws);
1123 cond_wake_up(ws_wait);
1126 static void put_workspace(int type, struct list_head *ws)
1129 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1130 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1131 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1132 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1135 * This can't happen, the type is validated several times
1136 * before we get here.
1143 * Adjust @level according to the limits of the compression algorithm or
1144 * fallback to default
1146 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1148 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1151 level = ops->default_level;
1153 level = min(level, ops->max_level);
1159 * Given an address space and start and length, compress the bytes into @pages
1160 * that are allocated on demand.
1162 * @type_level is encoded algorithm and level, where level 0 means whatever
1163 * default the algorithm chooses and is opaque here;
1164 * - compression algo are 0-3
1165 * - the level are bits 4-7
1167 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1168 * and returns number of actually allocated pages
1170 * @total_in is used to return the number of bytes actually read. It
1171 * may be smaller than the input length if we had to exit early because we
1172 * ran out of room in the pages array or because we cross the
1173 * max_out threshold.
1175 * @total_out is an in/out parameter, must be set to the input length and will
1176 * be also used to return the total number of compressed bytes
1178 * @max_out tells us the max number of bytes that we're allowed to
1181 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1182 u64 start, struct page **pages,
1183 unsigned long *out_pages,
1184 unsigned long *total_in,
1185 unsigned long *total_out)
1187 int type = btrfs_compress_type(type_level);
1188 int level = btrfs_compress_level(type_level);
1189 struct list_head *workspace;
1192 level = btrfs_compress_set_level(type, level);
1193 workspace = get_workspace(type, level);
1194 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1195 out_pages, total_in, total_out);
1196 put_workspace(type, workspace);
1201 * pages_in is an array of pages with compressed data.
1203 * disk_start is the starting logical offset of this array in the file
1205 * orig_bio contains the pages from the file that we want to decompress into
1207 * srclen is the number of bytes in pages_in
1209 * The basic idea is that we have a bio that was created by readpages.
1210 * The pages in the bio are for the uncompressed data, and they may not
1211 * be contiguous. They all correspond to the range of bytes covered by
1212 * the compressed extent.
1214 static int btrfs_decompress_bio(struct compressed_bio *cb)
1216 struct list_head *workspace;
1218 int type = cb->compress_type;
1220 workspace = get_workspace(type, 0);
1221 ret = compression_decompress_bio(type, workspace, cb);
1222 put_workspace(type, workspace);
1228 * a less complex decompression routine. Our compressed data fits in a
1229 * single page, and we want to read a single page out of it.
1230 * start_byte tells us the offset into the compressed data we're interested in
1232 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1233 unsigned long start_byte, size_t srclen, size_t destlen)
1235 struct list_head *workspace;
1238 workspace = get_workspace(type, 0);
1239 ret = compression_decompress(type, workspace, data_in, dest_page,
1240 start_byte, srclen, destlen);
1241 put_workspace(type, workspace);
1246 void __init btrfs_init_compress(void)
1248 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1249 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1250 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1251 zstd_init_workspace_manager();
1254 void __cold btrfs_exit_compress(void)
1256 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1257 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1258 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1259 zstd_cleanup_workspace_manager();
1263 * Copy uncompressed data from working buffer to pages.
1265 * buf_start is the byte offset we're of the start of our workspace buffer.
1267 * total_out is the last byte of the buffer
1269 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1270 unsigned long total_out, u64 disk_start,
1273 unsigned long buf_offset;
1274 unsigned long current_buf_start;
1275 unsigned long start_byte;
1276 unsigned long prev_start_byte;
1277 unsigned long working_bytes = total_out - buf_start;
1278 unsigned long bytes;
1280 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1283 * start byte is the first byte of the page we're currently
1284 * copying into relative to the start of the compressed data.
1286 start_byte = page_offset(bvec.bv_page) - disk_start;
1288 /* we haven't yet hit data corresponding to this page */
1289 if (total_out <= start_byte)
1293 * the start of the data we care about is offset into
1294 * the middle of our working buffer
1296 if (total_out > start_byte && buf_start < start_byte) {
1297 buf_offset = start_byte - buf_start;
1298 working_bytes -= buf_offset;
1302 current_buf_start = buf_start;
1304 /* copy bytes from the working buffer into the pages */
1305 while (working_bytes > 0) {
1306 bytes = min_t(unsigned long, bvec.bv_len,
1307 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1308 bytes = min(bytes, working_bytes);
1310 kaddr = kmap_atomic(bvec.bv_page);
1311 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1312 kunmap_atomic(kaddr);
1313 flush_dcache_page(bvec.bv_page);
1315 buf_offset += bytes;
1316 working_bytes -= bytes;
1317 current_buf_start += bytes;
1319 /* check if we need to pick another page */
1320 bio_advance(bio, bytes);
1321 if (!bio->bi_iter.bi_size)
1323 bvec = bio_iter_iovec(bio, bio->bi_iter);
1324 prev_start_byte = start_byte;
1325 start_byte = page_offset(bvec.bv_page) - disk_start;
1328 * We need to make sure we're only adjusting
1329 * our offset into compression working buffer when
1330 * we're switching pages. Otherwise we can incorrectly
1331 * keep copying when we were actually done.
1333 if (start_byte != prev_start_byte) {
1335 * make sure our new page is covered by this
1338 if (total_out <= start_byte)
1342 * the next page in the biovec might not be adjacent
1343 * to the last page, but it might still be found
1344 * inside this working buffer. bump our offset pointer
1346 if (total_out > start_byte &&
1347 current_buf_start < start_byte) {
1348 buf_offset = start_byte - buf_start;
1349 working_bytes = total_out - start_byte;
1350 current_buf_start = buf_start + buf_offset;
1359 * Shannon Entropy calculation
1361 * Pure byte distribution analysis fails to determine compressibility of data.
1362 * Try calculating entropy to estimate the average minimum number of bits
1363 * needed to encode the sampled data.
1365 * For convenience, return the percentage of needed bits, instead of amount of
1368 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1369 * and can be compressible with high probability
1371 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1373 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1375 #define ENTROPY_LVL_ACEPTABLE (65)
1376 #define ENTROPY_LVL_HIGH (80)
1379 * For increasead precision in shannon_entropy calculation,
1380 * let's do pow(n, M) to save more digits after comma:
1382 * - maximum int bit length is 64
1383 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1384 * - 13 * 4 = 52 < 64 -> M = 4
1388 static inline u32 ilog2_w(u64 n)
1390 return ilog2(n * n * n * n);
1393 static u32 shannon_entropy(struct heuristic_ws *ws)
1395 const u32 entropy_max = 8 * ilog2_w(2);
1396 u32 entropy_sum = 0;
1397 u32 p, p_base, sz_base;
1400 sz_base = ilog2_w(ws->sample_size);
1401 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1402 p = ws->bucket[i].count;
1403 p_base = ilog2_w(p);
1404 entropy_sum += p * (sz_base - p_base);
1407 entropy_sum /= ws->sample_size;
1408 return entropy_sum * 100 / entropy_max;
1411 #define RADIX_BASE 4U
1412 #define COUNTERS_SIZE (1U << RADIX_BASE)
1414 static u8 get4bits(u64 num, int shift) {
1419 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1424 * Use 4 bits as radix base
1425 * Use 16 u32 counters for calculating new position in buf array
1427 * @array - array that will be sorted
1428 * @array_buf - buffer array to store sorting results
1429 * must be equal in size to @array
1432 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1437 u32 counters[COUNTERS_SIZE];
1445 * Try avoid useless loop iterations for small numbers stored in big
1446 * counters. Example: 48 33 4 ... in 64bit array
1448 max_num = array[0].count;
1449 for (i = 1; i < num; i++) {
1450 buf_num = array[i].count;
1451 if (buf_num > max_num)
1455 buf_num = ilog2(max_num);
1456 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1459 while (shift < bitlen) {
1460 memset(counters, 0, sizeof(counters));
1462 for (i = 0; i < num; i++) {
1463 buf_num = array[i].count;
1464 addr = get4bits(buf_num, shift);
1468 for (i = 1; i < COUNTERS_SIZE; i++)
1469 counters[i] += counters[i - 1];
1471 for (i = num - 1; i >= 0; i--) {
1472 buf_num = array[i].count;
1473 addr = get4bits(buf_num, shift);
1475 new_addr = counters[addr];
1476 array_buf[new_addr] = array[i];
1479 shift += RADIX_BASE;
1482 * Normal radix expects to move data from a temporary array, to
1483 * the main one. But that requires some CPU time. Avoid that
1484 * by doing another sort iteration to original array instead of
1487 memset(counters, 0, sizeof(counters));
1489 for (i = 0; i < num; i ++) {
1490 buf_num = array_buf[i].count;
1491 addr = get4bits(buf_num, shift);
1495 for (i = 1; i < COUNTERS_SIZE; i++)
1496 counters[i] += counters[i - 1];
1498 for (i = num - 1; i >= 0; i--) {
1499 buf_num = array_buf[i].count;
1500 addr = get4bits(buf_num, shift);
1502 new_addr = counters[addr];
1503 array[new_addr] = array_buf[i];
1506 shift += RADIX_BASE;
1511 * Size of the core byte set - how many bytes cover 90% of the sample
1513 * There are several types of structured binary data that use nearly all byte
1514 * values. The distribution can be uniform and counts in all buckets will be
1515 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1517 * Other possibility is normal (Gaussian) distribution, where the data could
1518 * be potentially compressible, but we have to take a few more steps to decide
1521 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1522 * compression algo can easy fix that
1523 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1524 * probability is not compressible
1526 #define BYTE_CORE_SET_LOW (64)
1527 #define BYTE_CORE_SET_HIGH (200)
1529 static int byte_core_set_size(struct heuristic_ws *ws)
1532 u32 coreset_sum = 0;
1533 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1534 struct bucket_item *bucket = ws->bucket;
1536 /* Sort in reverse order */
1537 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1539 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1540 coreset_sum += bucket[i].count;
1542 if (coreset_sum > core_set_threshold)
1545 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1546 coreset_sum += bucket[i].count;
1547 if (coreset_sum > core_set_threshold)
1555 * Count byte values in buckets.
1556 * This heuristic can detect textual data (configs, xml, json, html, etc).
1557 * Because in most text-like data byte set is restricted to limited number of
1558 * possible characters, and that restriction in most cases makes data easy to
1561 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1562 * less - compressible
1563 * more - need additional analysis
1565 #define BYTE_SET_THRESHOLD (64)
1567 static u32 byte_set_size(const struct heuristic_ws *ws)
1570 u32 byte_set_size = 0;
1572 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1573 if (ws->bucket[i].count > 0)
1578 * Continue collecting count of byte values in buckets. If the byte
1579 * set size is bigger then the threshold, it's pointless to continue,
1580 * the detection technique would fail for this type of data.
1582 for (; i < BUCKET_SIZE; i++) {
1583 if (ws->bucket[i].count > 0) {
1585 if (byte_set_size > BYTE_SET_THRESHOLD)
1586 return byte_set_size;
1590 return byte_set_size;
1593 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1595 const u32 half_of_sample = ws->sample_size / 2;
1596 const u8 *data = ws->sample;
1598 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1601 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1602 struct heuristic_ws *ws)
1605 u64 index, index_end;
1606 u32 i, curr_sample_pos;
1610 * Compression handles the input data by chunks of 128KiB
1611 * (defined by BTRFS_MAX_UNCOMPRESSED)
1613 * We do the same for the heuristic and loop over the whole range.
1615 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1616 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1618 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1619 end = start + BTRFS_MAX_UNCOMPRESSED;
1621 index = start >> PAGE_SHIFT;
1622 index_end = end >> PAGE_SHIFT;
1624 /* Don't miss unaligned end */
1625 if (!IS_ALIGNED(end, PAGE_SIZE))
1628 curr_sample_pos = 0;
1629 while (index < index_end) {
1630 page = find_get_page(inode->i_mapping, index);
1631 in_data = kmap(page);
1632 /* Handle case where the start is not aligned to PAGE_SIZE */
1633 i = start % PAGE_SIZE;
1634 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1635 /* Don't sample any garbage from the last page */
1636 if (start > end - SAMPLING_READ_SIZE)
1638 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1639 SAMPLING_READ_SIZE);
1640 i += SAMPLING_INTERVAL;
1641 start += SAMPLING_INTERVAL;
1642 curr_sample_pos += SAMPLING_READ_SIZE;
1650 ws->sample_size = curr_sample_pos;
1654 * Compression heuristic.
1656 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1657 * quickly (compared to direct compression) detect data characteristics
1658 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1661 * The following types of analysis can be performed:
1662 * - detect mostly zero data
1663 * - detect data with low "byte set" size (text, etc)
1664 * - detect data with low/high "core byte" set
1666 * Return non-zero if the compression should be done, 0 otherwise.
1668 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1670 struct list_head *ws_list = get_workspace(0, 0);
1671 struct heuristic_ws *ws;
1676 ws = list_entry(ws_list, struct heuristic_ws, list);
1678 heuristic_collect_sample(inode, start, end, ws);
1680 if (sample_repeated_patterns(ws)) {
1685 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1687 for (i = 0; i < ws->sample_size; i++) {
1688 byte = ws->sample[i];
1689 ws->bucket[byte].count++;
1692 i = byte_set_size(ws);
1693 if (i < BYTE_SET_THRESHOLD) {
1698 i = byte_core_set_size(ws);
1699 if (i <= BYTE_CORE_SET_LOW) {
1704 if (i >= BYTE_CORE_SET_HIGH) {
1709 i = shannon_entropy(ws);
1710 if (i <= ENTROPY_LVL_ACEPTABLE) {
1716 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1717 * needed to give green light to compression.
1719 * For now just assume that compression at that level is not worth the
1720 * resources because:
1722 * 1. it is possible to defrag the data later
1724 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1725 * values, every bucket has counter at level ~54. The heuristic would
1726 * be confused. This can happen when data have some internal repeated
1727 * patterns like "abbacbbc...". This can be detected by analyzing
1728 * pairs of bytes, which is too costly.
1730 if (i < ENTROPY_LVL_HIGH) {
1739 put_workspace(0, ws_list);
1744 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1745 * level, unrecognized string will set the default level
1747 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1749 unsigned int level = 0;
1755 if (str[0] == ':') {
1756 ret = kstrtouint(str + 1, 10, &level);
1761 level = btrfs_compress_set_level(type, level);