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 btrfs_record_physical_zoned(inode, cb->start, bio);
353 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
354 cb->start, cb->start + cb->len - 1,
357 end_compressed_writeback(inode, cb);
358 /* note, our inode could be gone now */
361 * release the compressed pages, these came from alloc_page and
362 * are not attached to the inode at all
365 for (index = 0; index < cb->nr_pages; index++) {
366 page = cb->compressed_pages[index];
367 page->mapping = NULL;
371 /* finally free the cb struct */
372 kfree(cb->compressed_pages);
379 * worker function to build and submit bios for previously compressed pages.
380 * The corresponding pages in the inode should be marked for writeback
381 * and the compressed pages should have a reference on them for dropping
382 * when the IO is complete.
384 * This also checksums the file bytes and gets things ready for
387 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
388 unsigned int len, u64 disk_start,
389 unsigned int compressed_len,
390 struct page **compressed_pages,
391 unsigned int nr_pages,
392 unsigned int write_flags,
393 struct cgroup_subsys_state *blkcg_css)
395 struct btrfs_fs_info *fs_info = inode->root->fs_info;
396 struct bio *bio = NULL;
397 struct compressed_bio *cb;
398 unsigned long bytes_left;
401 u64 first_byte = disk_start;
403 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
404 const bool use_append = btrfs_use_zone_append(inode, disk_start);
405 const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
407 WARN_ON(!PAGE_ALIGNED(start));
408 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
410 return BLK_STS_RESOURCE;
411 refcount_set(&cb->pending_bios, 0);
413 cb->inode = &inode->vfs_inode;
417 cb->compressed_pages = compressed_pages;
418 cb->compressed_len = compressed_len;
420 cb->nr_pages = nr_pages;
422 bio = btrfs_bio_alloc(first_byte);
423 bio->bi_opf = bio_op | write_flags;
424 bio->bi_private = cb;
425 bio->bi_end_io = end_compressed_bio_write;
428 struct btrfs_device *device;
430 device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE);
431 if (IS_ERR(device)) {
434 return BLK_STS_NOTSUPP;
437 bio_set_dev(bio, device->bdev);
441 bio->bi_opf |= REQ_CGROUP_PUNT;
442 kthread_associate_blkcg(blkcg_css);
444 refcount_set(&cb->pending_bios, 1);
446 /* create and submit bios for the compressed pages */
447 bytes_left = compressed_len;
448 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
452 page = compressed_pages[pg_index];
453 page->mapping = inode->vfs_inode.i_mapping;
454 if (bio->bi_iter.bi_size)
455 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
459 * Page can only be added to bio if the current bio fits in
463 if (pg_index == 0 && use_append)
464 len = bio_add_zone_append_page(bio, page,
467 len = bio_add_page(bio, page, PAGE_SIZE, 0);
470 page->mapping = NULL;
471 if (submit || len < PAGE_SIZE) {
473 * inc the count before we submit the bio so
474 * we know the end IO handler won't happen before
475 * we inc the count. Otherwise, the cb might get
476 * freed before we're done setting it up
478 refcount_inc(&cb->pending_bios);
479 ret = btrfs_bio_wq_end_io(fs_info, bio,
480 BTRFS_WQ_ENDIO_DATA);
481 BUG_ON(ret); /* -ENOMEM */
484 ret = btrfs_csum_one_bio(inode, bio, start, 1);
485 BUG_ON(ret); /* -ENOMEM */
488 ret = btrfs_map_bio(fs_info, bio, 0);
490 bio->bi_status = ret;
494 bio = btrfs_bio_alloc(first_byte);
495 bio->bi_opf = bio_op | write_flags;
496 bio->bi_private = cb;
497 bio->bi_end_io = end_compressed_bio_write;
499 bio->bi_opf |= REQ_CGROUP_PUNT;
501 * Use bio_add_page() to ensure the bio has at least one
504 bio_add_page(bio, page, PAGE_SIZE, 0);
506 if (bytes_left < PAGE_SIZE) {
508 "bytes left %lu compress len %u nr %u",
509 bytes_left, cb->compressed_len, cb->nr_pages);
511 bytes_left -= PAGE_SIZE;
512 first_byte += PAGE_SIZE;
516 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
517 BUG_ON(ret); /* -ENOMEM */
520 ret = btrfs_csum_one_bio(inode, bio, start, 1);
521 BUG_ON(ret); /* -ENOMEM */
524 ret = btrfs_map_bio(fs_info, bio, 0);
526 bio->bi_status = ret;
531 kthread_associate_blkcg(NULL);
536 static u64 bio_end_offset(struct bio *bio)
538 struct bio_vec *last = bio_last_bvec_all(bio);
540 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
543 static noinline int add_ra_bio_pages(struct inode *inode,
545 struct compressed_bio *cb)
547 unsigned long end_index;
548 unsigned long pg_index;
550 u64 isize = i_size_read(inode);
553 unsigned long nr_pages = 0;
554 struct extent_map *em;
555 struct address_space *mapping = inode->i_mapping;
556 struct extent_map_tree *em_tree;
557 struct extent_io_tree *tree;
561 last_offset = bio_end_offset(cb->orig_bio);
562 em_tree = &BTRFS_I(inode)->extent_tree;
563 tree = &BTRFS_I(inode)->io_tree;
569 * For current subpage support, we only support 64K page size,
570 * which means maximum compressed extent size (128K) is just 2x page
572 * This makes readahead less effective, so here disable readahead for
573 * subpage for now, until full compressed write is supported.
575 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
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 int compressed_len;
681 unsigned int nr_pages;
682 unsigned int pg_index;
684 struct bio *comp_bio;
685 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
689 struct extent_map *em;
690 blk_status_t ret = BLK_STS_RESOURCE;
694 em_tree = &BTRFS_I(inode)->extent_tree;
696 file_offset = bio_first_bvec_all(bio)->bv_offset +
697 page_offset(bio_first_page_all(bio));
699 /* we need the actual starting offset of this extent in the file */
700 read_lock(&em_tree->lock);
701 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
702 read_unlock(&em_tree->lock);
704 return BLK_STS_IOERR;
706 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
707 compressed_len = em->block_len;
708 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
712 refcount_set(&cb->pending_bios, 0);
715 cb->mirror_num = mirror_num;
718 cb->start = em->orig_start;
720 em_start = em->start;
725 cb->len = bio->bi_iter.bi_size;
726 cb->compressed_len = compressed_len;
727 cb->compress_type = extent_compress_type(bio_flags);
730 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
731 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
733 if (!cb->compressed_pages)
736 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
737 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS);
738 if (!cb->compressed_pages[pg_index]) {
739 faili = pg_index - 1;
740 ret = BLK_STS_RESOURCE;
744 faili = nr_pages - 1;
745 cb->nr_pages = nr_pages;
747 add_ra_bio_pages(inode, em_start + em_len, cb);
749 /* include any pages we added in add_ra-bio_pages */
750 cb->len = bio->bi_iter.bi_size;
752 comp_bio = btrfs_bio_alloc(cur_disk_byte);
753 comp_bio->bi_opf = REQ_OP_READ;
754 comp_bio->bi_private = cb;
755 comp_bio->bi_end_io = end_compressed_bio_read;
756 refcount_set(&cb->pending_bios, 1);
758 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
759 u32 pg_len = PAGE_SIZE;
763 * To handle subpage case, we need to make sure the bio only
764 * covers the range we need.
766 * If we're at the last page, truncate the length to only cover
767 * the remaining part.
769 if (pg_index == nr_pages - 1)
770 pg_len = min_t(u32, PAGE_SIZE,
771 compressed_len - pg_index * PAGE_SIZE);
773 page = cb->compressed_pages[pg_index];
774 page->mapping = inode->i_mapping;
775 page->index = em_start >> PAGE_SHIFT;
777 if (comp_bio->bi_iter.bi_size)
778 submit = btrfs_bio_fits_in_stripe(page, pg_len,
781 page->mapping = NULL;
782 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
783 unsigned int nr_sectors;
785 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
786 BTRFS_WQ_ENDIO_DATA);
787 BUG_ON(ret); /* -ENOMEM */
790 * inc the count before we submit the bio so
791 * we know the end IO handler won't happen before
792 * we inc the count. Otherwise, the cb might get
793 * freed before we're done setting it up
795 refcount_inc(&cb->pending_bios);
797 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
798 BUG_ON(ret); /* -ENOMEM */
800 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
801 fs_info->sectorsize);
802 sums += fs_info->csum_size * nr_sectors;
804 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
806 comp_bio->bi_status = ret;
810 comp_bio = btrfs_bio_alloc(cur_disk_byte);
811 comp_bio->bi_opf = REQ_OP_READ;
812 comp_bio->bi_private = cb;
813 comp_bio->bi_end_io = end_compressed_bio_read;
815 bio_add_page(comp_bio, page, pg_len, 0);
817 cur_disk_byte += pg_len;
820 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
821 BUG_ON(ret); /* -ENOMEM */
823 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
824 BUG_ON(ret); /* -ENOMEM */
826 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
828 comp_bio->bi_status = ret;
836 __free_page(cb->compressed_pages[faili]);
840 kfree(cb->compressed_pages);
849 * Heuristic uses systematic sampling to collect data from the input data
850 * range, the logic can be tuned by the following constants:
852 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
853 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
855 #define SAMPLING_READ_SIZE (16)
856 #define SAMPLING_INTERVAL (256)
859 * For statistical analysis of the input data we consider bytes that form a
860 * Galois Field of 256 objects. Each object has an attribute count, ie. how
861 * many times the object appeared in the sample.
863 #define BUCKET_SIZE (256)
866 * The size of the sample is based on a statistical sampling rule of thumb.
867 * The common way is to perform sampling tests as long as the number of
868 * elements in each cell is at least 5.
870 * Instead of 5, we choose 32 to obtain more accurate results.
871 * If the data contain the maximum number of symbols, which is 256, we obtain a
872 * sample size bound by 8192.
874 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
875 * from up to 512 locations.
877 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
878 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
884 struct heuristic_ws {
885 /* Partial copy of input data */
888 /* Buckets store counters for each byte value */
889 struct bucket_item *bucket;
891 struct bucket_item *bucket_b;
892 struct list_head list;
895 static struct workspace_manager heuristic_wsm;
897 static void free_heuristic_ws(struct list_head *ws)
899 struct heuristic_ws *workspace;
901 workspace = list_entry(ws, struct heuristic_ws, list);
903 kvfree(workspace->sample);
904 kfree(workspace->bucket);
905 kfree(workspace->bucket_b);
909 static struct list_head *alloc_heuristic_ws(unsigned int level)
911 struct heuristic_ws *ws;
913 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
915 return ERR_PTR(-ENOMEM);
917 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
921 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
925 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
929 INIT_LIST_HEAD(&ws->list);
932 free_heuristic_ws(&ws->list);
933 return ERR_PTR(-ENOMEM);
936 const struct btrfs_compress_op btrfs_heuristic_compress = {
937 .workspace_manager = &heuristic_wsm,
940 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
941 /* The heuristic is represented as compression type 0 */
942 &btrfs_heuristic_compress,
943 &btrfs_zlib_compress,
945 &btrfs_zstd_compress,
948 static struct list_head *alloc_workspace(int type, unsigned int level)
951 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
952 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
953 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
954 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
957 * This can't happen, the type is validated several times
958 * before we get here.
964 static void free_workspace(int type, struct list_head *ws)
967 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
968 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
969 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
970 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
973 * This can't happen, the type is validated several times
974 * before we get here.
980 static void btrfs_init_workspace_manager(int type)
982 struct workspace_manager *wsm;
983 struct list_head *workspace;
985 wsm = btrfs_compress_op[type]->workspace_manager;
986 INIT_LIST_HEAD(&wsm->idle_ws);
987 spin_lock_init(&wsm->ws_lock);
988 atomic_set(&wsm->total_ws, 0);
989 init_waitqueue_head(&wsm->ws_wait);
992 * Preallocate one workspace for each compression type so we can
993 * guarantee forward progress in the worst case
995 workspace = alloc_workspace(type, 0);
996 if (IS_ERR(workspace)) {
998 "BTRFS: cannot preallocate compression workspace, will try later\n");
1000 atomic_set(&wsm->total_ws, 1);
1002 list_add(workspace, &wsm->idle_ws);
1006 static void btrfs_cleanup_workspace_manager(int type)
1008 struct workspace_manager *wsman;
1009 struct list_head *ws;
1011 wsman = btrfs_compress_op[type]->workspace_manager;
1012 while (!list_empty(&wsman->idle_ws)) {
1013 ws = wsman->idle_ws.next;
1015 free_workspace(type, ws);
1016 atomic_dec(&wsman->total_ws);
1021 * This finds an available workspace or allocates a new one.
1022 * If it's not possible to allocate a new one, waits until there's one.
1023 * Preallocation makes a forward progress guarantees and we do not return
1026 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1028 struct workspace_manager *wsm;
1029 struct list_head *workspace;
1030 int cpus = num_online_cpus();
1032 struct list_head *idle_ws;
1033 spinlock_t *ws_lock;
1035 wait_queue_head_t *ws_wait;
1038 wsm = btrfs_compress_op[type]->workspace_manager;
1039 idle_ws = &wsm->idle_ws;
1040 ws_lock = &wsm->ws_lock;
1041 total_ws = &wsm->total_ws;
1042 ws_wait = &wsm->ws_wait;
1043 free_ws = &wsm->free_ws;
1047 if (!list_empty(idle_ws)) {
1048 workspace = idle_ws->next;
1049 list_del(workspace);
1051 spin_unlock(ws_lock);
1055 if (atomic_read(total_ws) > cpus) {
1058 spin_unlock(ws_lock);
1059 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1060 if (atomic_read(total_ws) > cpus && !*free_ws)
1062 finish_wait(ws_wait, &wait);
1065 atomic_inc(total_ws);
1066 spin_unlock(ws_lock);
1069 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1070 * to turn it off here because we might get called from the restricted
1071 * context of btrfs_compress_bio/btrfs_compress_pages
1073 nofs_flag = memalloc_nofs_save();
1074 workspace = alloc_workspace(type, level);
1075 memalloc_nofs_restore(nofs_flag);
1077 if (IS_ERR(workspace)) {
1078 atomic_dec(total_ws);
1082 * Do not return the error but go back to waiting. There's a
1083 * workspace preallocated for each type and the compression
1084 * time is bounded so we get to a workspace eventually. This
1085 * makes our caller's life easier.
1087 * To prevent silent and low-probability deadlocks (when the
1088 * initial preallocation fails), check if there are any
1089 * workspaces at all.
1091 if (atomic_read(total_ws) == 0) {
1092 static DEFINE_RATELIMIT_STATE(_rs,
1093 /* once per minute */ 60 * HZ,
1096 if (__ratelimit(&_rs)) {
1097 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1105 static struct list_head *get_workspace(int type, int level)
1108 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1109 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1110 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1111 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1114 * This can't happen, the type is validated several times
1115 * before we get here.
1122 * put a workspace struct back on the list or free it if we have enough
1123 * idle ones sitting around
1125 void btrfs_put_workspace(int type, struct list_head *ws)
1127 struct workspace_manager *wsm;
1128 struct list_head *idle_ws;
1129 spinlock_t *ws_lock;
1131 wait_queue_head_t *ws_wait;
1134 wsm = btrfs_compress_op[type]->workspace_manager;
1135 idle_ws = &wsm->idle_ws;
1136 ws_lock = &wsm->ws_lock;
1137 total_ws = &wsm->total_ws;
1138 ws_wait = &wsm->ws_wait;
1139 free_ws = &wsm->free_ws;
1142 if (*free_ws <= num_online_cpus()) {
1143 list_add(ws, idle_ws);
1145 spin_unlock(ws_lock);
1148 spin_unlock(ws_lock);
1150 free_workspace(type, ws);
1151 atomic_dec(total_ws);
1153 cond_wake_up(ws_wait);
1156 static void put_workspace(int type, struct list_head *ws)
1159 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1160 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1161 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1162 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1165 * This can't happen, the type is validated several times
1166 * before we get here.
1173 * Adjust @level according to the limits of the compression algorithm or
1174 * fallback to default
1176 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1178 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1181 level = ops->default_level;
1183 level = min(level, ops->max_level);
1189 * Given an address space and start and length, compress the bytes into @pages
1190 * that are allocated on demand.
1192 * @type_level is encoded algorithm and level, where level 0 means whatever
1193 * default the algorithm chooses and is opaque here;
1194 * - compression algo are 0-3
1195 * - the level are bits 4-7
1197 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1198 * and returns number of actually allocated pages
1200 * @total_in is used to return the number of bytes actually read. It
1201 * may be smaller than the input length if we had to exit early because we
1202 * ran out of room in the pages array or because we cross the
1203 * max_out threshold.
1205 * @total_out is an in/out parameter, must be set to the input length and will
1206 * be also used to return the total number of compressed bytes
1208 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1209 u64 start, struct page **pages,
1210 unsigned long *out_pages,
1211 unsigned long *total_in,
1212 unsigned long *total_out)
1214 int type = btrfs_compress_type(type_level);
1215 int level = btrfs_compress_level(type_level);
1216 struct list_head *workspace;
1219 level = btrfs_compress_set_level(type, level);
1220 workspace = get_workspace(type, level);
1221 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1222 out_pages, total_in, total_out);
1223 put_workspace(type, workspace);
1227 static int btrfs_decompress_bio(struct compressed_bio *cb)
1229 struct list_head *workspace;
1231 int type = cb->compress_type;
1233 workspace = get_workspace(type, 0);
1234 ret = compression_decompress_bio(type, workspace, cb);
1235 put_workspace(type, workspace);
1241 * a less complex decompression routine. Our compressed data fits in a
1242 * single page, and we want to read a single page out of it.
1243 * start_byte tells us the offset into the compressed data we're interested in
1245 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1246 unsigned long start_byte, size_t srclen, size_t destlen)
1248 struct list_head *workspace;
1251 workspace = get_workspace(type, 0);
1252 ret = compression_decompress(type, workspace, data_in, dest_page,
1253 start_byte, srclen, destlen);
1254 put_workspace(type, workspace);
1259 void __init btrfs_init_compress(void)
1261 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1262 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1263 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1264 zstd_init_workspace_manager();
1267 void __cold btrfs_exit_compress(void)
1269 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1270 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1271 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1272 zstd_cleanup_workspace_manager();
1276 * Copy decompressed data from working buffer to pages.
1278 * @buf: The decompressed data buffer
1279 * @buf_len: The decompressed data length
1280 * @decompressed: Number of bytes that are already decompressed inside the
1282 * @cb: The compressed extent descriptor
1283 * @orig_bio: The original bio that the caller wants to read for
1285 * An easier to understand graph is like below:
1287 * |<- orig_bio ->| |<- orig_bio->|
1288 * |<------- full decompressed extent ----->|
1289 * |<----------- @cb range ---->|
1290 * | |<-- @buf_len -->|
1291 * |<--- @decompressed --->|
1293 * Note that, @cb can be a subpage of the full decompressed extent, but
1294 * @cb->start always has the same as the orig_file_offset value of the full
1295 * decompressed extent.
1297 * When reading compressed extent, we have to read the full compressed extent,
1298 * while @orig_bio may only want part of the range.
1299 * Thus this function will ensure only data covered by @orig_bio will be copied
1302 * Return 0 if we have copied all needed contents for @orig_bio.
1303 * Return >0 if we need continue decompress.
1305 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1306 struct compressed_bio *cb, u32 decompressed)
1308 struct bio *orig_bio = cb->orig_bio;
1309 /* Offset inside the full decompressed extent */
1312 cur_offset = decompressed;
1313 /* The main loop to do the copy */
1314 while (cur_offset < decompressed + buf_len) {
1315 struct bio_vec bvec;
1318 /* Offset inside the full decompressed extent */
1321 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1323 * cb->start may underflow, but subtracting that value can still
1324 * give us correct offset inside the full decompressed extent.
1326 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1328 /* Haven't reached the bvec range, exit */
1329 if (decompressed + buf_len <= bvec_offset)
1332 copy_start = max(cur_offset, bvec_offset);
1333 copy_len = min(bvec_offset + bvec.bv_len,
1334 decompressed + buf_len) - copy_start;
1338 * Extra range check to ensure we didn't go beyond
1341 ASSERT(copy_start - decompressed < buf_len);
1342 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1343 buf + copy_start - decompressed, copy_len);
1344 flush_dcache_page(bvec.bv_page);
1345 cur_offset += copy_len;
1347 bio_advance(orig_bio, copy_len);
1348 /* Finished the bio */
1349 if (!orig_bio->bi_iter.bi_size)
1356 * Shannon Entropy calculation
1358 * Pure byte distribution analysis fails to determine compressibility of data.
1359 * Try calculating entropy to estimate the average minimum number of bits
1360 * needed to encode the sampled data.
1362 * For convenience, return the percentage of needed bits, instead of amount of
1365 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1366 * and can be compressible with high probability
1368 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1370 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1372 #define ENTROPY_LVL_ACEPTABLE (65)
1373 #define ENTROPY_LVL_HIGH (80)
1376 * For increasead precision in shannon_entropy calculation,
1377 * let's do pow(n, M) to save more digits after comma:
1379 * - maximum int bit length is 64
1380 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1381 * - 13 * 4 = 52 < 64 -> M = 4
1385 static inline u32 ilog2_w(u64 n)
1387 return ilog2(n * n * n * n);
1390 static u32 shannon_entropy(struct heuristic_ws *ws)
1392 const u32 entropy_max = 8 * ilog2_w(2);
1393 u32 entropy_sum = 0;
1394 u32 p, p_base, sz_base;
1397 sz_base = ilog2_w(ws->sample_size);
1398 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1399 p = ws->bucket[i].count;
1400 p_base = ilog2_w(p);
1401 entropy_sum += p * (sz_base - p_base);
1404 entropy_sum /= ws->sample_size;
1405 return entropy_sum * 100 / entropy_max;
1408 #define RADIX_BASE 4U
1409 #define COUNTERS_SIZE (1U << RADIX_BASE)
1411 static u8 get4bits(u64 num, int shift) {
1416 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1421 * Use 4 bits as radix base
1422 * Use 16 u32 counters for calculating new position in buf array
1424 * @array - array that will be sorted
1425 * @array_buf - buffer array to store sorting results
1426 * must be equal in size to @array
1429 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1434 u32 counters[COUNTERS_SIZE];
1442 * Try avoid useless loop iterations for small numbers stored in big
1443 * counters. Example: 48 33 4 ... in 64bit array
1445 max_num = array[0].count;
1446 for (i = 1; i < num; i++) {
1447 buf_num = array[i].count;
1448 if (buf_num > max_num)
1452 buf_num = ilog2(max_num);
1453 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1456 while (shift < bitlen) {
1457 memset(counters, 0, sizeof(counters));
1459 for (i = 0; i < num; i++) {
1460 buf_num = array[i].count;
1461 addr = get4bits(buf_num, shift);
1465 for (i = 1; i < COUNTERS_SIZE; i++)
1466 counters[i] += counters[i - 1];
1468 for (i = num - 1; i >= 0; i--) {
1469 buf_num = array[i].count;
1470 addr = get4bits(buf_num, shift);
1472 new_addr = counters[addr];
1473 array_buf[new_addr] = array[i];
1476 shift += RADIX_BASE;
1479 * Normal radix expects to move data from a temporary array, to
1480 * the main one. But that requires some CPU time. Avoid that
1481 * by doing another sort iteration to original array instead of
1484 memset(counters, 0, sizeof(counters));
1486 for (i = 0; i < num; i ++) {
1487 buf_num = array_buf[i].count;
1488 addr = get4bits(buf_num, shift);
1492 for (i = 1; i < COUNTERS_SIZE; i++)
1493 counters[i] += counters[i - 1];
1495 for (i = num - 1; i >= 0; i--) {
1496 buf_num = array_buf[i].count;
1497 addr = get4bits(buf_num, shift);
1499 new_addr = counters[addr];
1500 array[new_addr] = array_buf[i];
1503 shift += RADIX_BASE;
1508 * Size of the core byte set - how many bytes cover 90% of the sample
1510 * There are several types of structured binary data that use nearly all byte
1511 * values. The distribution can be uniform and counts in all buckets will be
1512 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1514 * Other possibility is normal (Gaussian) distribution, where the data could
1515 * be potentially compressible, but we have to take a few more steps to decide
1518 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1519 * compression algo can easy fix that
1520 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1521 * probability is not compressible
1523 #define BYTE_CORE_SET_LOW (64)
1524 #define BYTE_CORE_SET_HIGH (200)
1526 static int byte_core_set_size(struct heuristic_ws *ws)
1529 u32 coreset_sum = 0;
1530 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1531 struct bucket_item *bucket = ws->bucket;
1533 /* Sort in reverse order */
1534 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1536 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1537 coreset_sum += bucket[i].count;
1539 if (coreset_sum > core_set_threshold)
1542 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1543 coreset_sum += bucket[i].count;
1544 if (coreset_sum > core_set_threshold)
1552 * Count byte values in buckets.
1553 * This heuristic can detect textual data (configs, xml, json, html, etc).
1554 * Because in most text-like data byte set is restricted to limited number of
1555 * possible characters, and that restriction in most cases makes data easy to
1558 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1559 * less - compressible
1560 * more - need additional analysis
1562 #define BYTE_SET_THRESHOLD (64)
1564 static u32 byte_set_size(const struct heuristic_ws *ws)
1567 u32 byte_set_size = 0;
1569 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1570 if (ws->bucket[i].count > 0)
1575 * Continue collecting count of byte values in buckets. If the byte
1576 * set size is bigger then the threshold, it's pointless to continue,
1577 * the detection technique would fail for this type of data.
1579 for (; i < BUCKET_SIZE; i++) {
1580 if (ws->bucket[i].count > 0) {
1582 if (byte_set_size > BYTE_SET_THRESHOLD)
1583 return byte_set_size;
1587 return byte_set_size;
1590 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1592 const u32 half_of_sample = ws->sample_size / 2;
1593 const u8 *data = ws->sample;
1595 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1598 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1599 struct heuristic_ws *ws)
1602 u64 index, index_end;
1603 u32 i, curr_sample_pos;
1607 * Compression handles the input data by chunks of 128KiB
1608 * (defined by BTRFS_MAX_UNCOMPRESSED)
1610 * We do the same for the heuristic and loop over the whole range.
1612 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1613 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1615 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1616 end = start + BTRFS_MAX_UNCOMPRESSED;
1618 index = start >> PAGE_SHIFT;
1619 index_end = end >> PAGE_SHIFT;
1621 /* Don't miss unaligned end */
1622 if (!IS_ALIGNED(end, PAGE_SIZE))
1625 curr_sample_pos = 0;
1626 while (index < index_end) {
1627 page = find_get_page(inode->i_mapping, index);
1628 in_data = kmap_local_page(page);
1629 /* Handle case where the start is not aligned to PAGE_SIZE */
1630 i = start % PAGE_SIZE;
1631 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1632 /* Don't sample any garbage from the last page */
1633 if (start > end - SAMPLING_READ_SIZE)
1635 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1636 SAMPLING_READ_SIZE);
1637 i += SAMPLING_INTERVAL;
1638 start += SAMPLING_INTERVAL;
1639 curr_sample_pos += SAMPLING_READ_SIZE;
1641 kunmap_local(in_data);
1647 ws->sample_size = curr_sample_pos;
1651 * Compression heuristic.
1653 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1654 * quickly (compared to direct compression) detect data characteristics
1655 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1658 * The following types of analysis can be performed:
1659 * - detect mostly zero data
1660 * - detect data with low "byte set" size (text, etc)
1661 * - detect data with low/high "core byte" set
1663 * Return non-zero if the compression should be done, 0 otherwise.
1665 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1667 struct list_head *ws_list = get_workspace(0, 0);
1668 struct heuristic_ws *ws;
1673 ws = list_entry(ws_list, struct heuristic_ws, list);
1675 heuristic_collect_sample(inode, start, end, ws);
1677 if (sample_repeated_patterns(ws)) {
1682 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1684 for (i = 0; i < ws->sample_size; i++) {
1685 byte = ws->sample[i];
1686 ws->bucket[byte].count++;
1689 i = byte_set_size(ws);
1690 if (i < BYTE_SET_THRESHOLD) {
1695 i = byte_core_set_size(ws);
1696 if (i <= BYTE_CORE_SET_LOW) {
1701 if (i >= BYTE_CORE_SET_HIGH) {
1706 i = shannon_entropy(ws);
1707 if (i <= ENTROPY_LVL_ACEPTABLE) {
1713 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1714 * needed to give green light to compression.
1716 * For now just assume that compression at that level is not worth the
1717 * resources because:
1719 * 1. it is possible to defrag the data later
1721 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1722 * values, every bucket has counter at level ~54. The heuristic would
1723 * be confused. This can happen when data have some internal repeated
1724 * patterns like "abbacbbc...". This can be detected by analyzing
1725 * pairs of bytes, which is too costly.
1727 if (i < ENTROPY_LVL_HIGH) {
1736 put_workspace(0, ws_list);
1741 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1742 * level, unrecognized string will set the default level
1744 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1746 unsigned int level = 0;
1752 if (str[0] == ':') {
1753 ret = kstrtouint(str + 1, 10, &level);
1758 level = btrfs_compress_set_level(type, level);