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 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
34 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
37 case BTRFS_COMPRESS_ZLIB:
38 case BTRFS_COMPRESS_LZO:
39 case BTRFS_COMPRESS_ZSTD:
40 case BTRFS_COMPRESS_NONE:
41 return btrfs_compress_types[type];
49 bool btrfs_compress_is_valid_type(const char *str, size_t len)
53 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
54 size_t comp_len = strlen(btrfs_compress_types[i]);
59 if (!strncmp(btrfs_compress_types[i], str, comp_len))
65 static int compression_compress_pages(int type, struct list_head *ws,
66 struct address_space *mapping, u64 start, struct page **pages,
67 unsigned long *out_pages, unsigned long *total_in,
68 unsigned long *total_out)
71 case BTRFS_COMPRESS_ZLIB:
72 return zlib_compress_pages(ws, mapping, start, pages,
73 out_pages, total_in, total_out);
74 case BTRFS_COMPRESS_LZO:
75 return lzo_compress_pages(ws, mapping, start, pages,
76 out_pages, total_in, total_out);
77 case BTRFS_COMPRESS_ZSTD:
78 return zstd_compress_pages(ws, mapping, start, pages,
79 out_pages, total_in, total_out);
80 case BTRFS_COMPRESS_NONE:
83 * This can't happen, the type is validated several times
84 * before we get here. As a sane fallback, return what the
85 * callers will understand as 'no compression happened'.
91 static int compression_decompress_bio(int type, struct list_head *ws,
92 struct compressed_bio *cb)
95 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
96 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
97 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
98 case BTRFS_COMPRESS_NONE:
101 * This can't happen, the type is validated several times
102 * before we get here.
108 static int compression_decompress(int type, struct list_head *ws,
109 unsigned char *data_in, struct page *dest_page,
110 unsigned long start_byte, size_t srclen, size_t destlen)
113 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
114 start_byte, srclen, destlen);
115 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
116 start_byte, srclen, destlen);
117 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
118 start_byte, srclen, destlen);
119 case BTRFS_COMPRESS_NONE:
122 * This can't happen, the type is validated several times
123 * before we get here.
129 static int btrfs_decompress_bio(struct compressed_bio *cb);
131 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
132 unsigned long disk_size)
134 return sizeof(struct compressed_bio) +
135 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size;
138 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
141 struct btrfs_fs_info *fs_info = inode->root->fs_info;
142 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
143 const u32 csum_size = fs_info->csum_size;
147 u8 csum[BTRFS_CSUM_SIZE];
148 struct compressed_bio *cb = bio->bi_private;
149 u8 *cb_sum = cb->sums;
151 if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM))
154 shash->tfm = fs_info->csum_shash;
156 for (i = 0; i < cb->nr_pages; i++) {
157 page = cb->compressed_pages[i];
159 kaddr = kmap_atomic(page);
160 crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum);
161 kunmap_atomic(kaddr);
163 if (memcmp(&csum, cb_sum, csum_size)) {
164 btrfs_print_data_csum_error(inode, disk_start,
165 csum, cb_sum, cb->mirror_num);
166 if (btrfs_io_bio(bio)->device)
167 btrfs_dev_stat_inc_and_print(
168 btrfs_io_bio(bio)->device,
169 BTRFS_DEV_STAT_CORRUPTION_ERRS);
177 /* when we finish reading compressed pages from the disk, we
178 * decompress them and then run the bio end_io routines on the
179 * decompressed pages (in the inode address space).
181 * This allows the checksumming and other IO error handling routines
184 * The compressed pages are freed here, and it must be run
187 static void end_compressed_bio_read(struct bio *bio)
189 struct compressed_bio *cb = bio->bi_private;
193 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
199 /* if there are more bios still pending for this compressed
202 if (!refcount_dec_and_test(&cb->pending_bios))
206 * Record the correct mirror_num in cb->orig_bio so that
207 * read-repair can work properly.
209 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
210 cb->mirror_num = mirror;
213 * Some IO in this cb have failed, just skip checksum as there
214 * is no way it could be correct.
220 ret = check_compressed_csum(BTRFS_I(inode), bio,
221 bio->bi_iter.bi_sector << 9);
225 /* ok, we're the last bio for this extent, lets start
228 ret = btrfs_decompress_bio(cb);
234 /* release the compressed pages */
236 for (index = 0; index < cb->nr_pages; index++) {
237 page = cb->compressed_pages[index];
238 page->mapping = NULL;
242 /* do io completion on the original bio */
244 bio_io_error(cb->orig_bio);
246 struct bio_vec *bvec;
247 struct bvec_iter_all iter_all;
250 * we have verified the checksum already, set page
251 * checked so the end_io handlers know about it
253 ASSERT(!bio_flagged(bio, BIO_CLONED));
254 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
255 SetPageChecked(bvec->bv_page);
257 bio_endio(cb->orig_bio);
260 /* finally free the cb struct */
261 kfree(cb->compressed_pages);
268 * Clear the writeback bits on all of the file
269 * pages for a compressed write
271 static noinline void end_compressed_writeback(struct inode *inode,
272 const struct compressed_bio *cb)
274 unsigned long index = cb->start >> PAGE_SHIFT;
275 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
276 struct page *pages[16];
277 unsigned long nr_pages = end_index - index + 1;
282 mapping_set_error(inode->i_mapping, -EIO);
284 while (nr_pages > 0) {
285 ret = find_get_pages_contig(inode->i_mapping, index,
287 nr_pages, ARRAY_SIZE(pages)), pages);
293 for (i = 0; i < ret; i++) {
295 SetPageError(pages[i]);
296 end_page_writeback(pages[i]);
302 /* the inode may be gone now */
306 * do the cleanup once all the compressed pages hit the disk.
307 * This will clear writeback on the file pages and free the compressed
310 * This also calls the writeback end hooks for the file pages so that
311 * metadata and checksums can be updated in the file.
313 static void end_compressed_bio_write(struct bio *bio)
315 struct compressed_bio *cb = bio->bi_private;
323 /* if there are more bios still pending for this compressed
326 if (!refcount_dec_and_test(&cb->pending_bios))
329 /* ok, we're the last bio for this extent, step one is to
330 * call back into the FS and do all the end_io operations
333 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
334 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
335 cb->start, cb->start + cb->len - 1,
336 bio->bi_status == BLK_STS_OK);
337 cb->compressed_pages[0]->mapping = NULL;
339 end_compressed_writeback(inode, cb);
340 /* note, our inode could be gone now */
343 * release the compressed pages, these came from alloc_page and
344 * are not attached to the inode at all
347 for (index = 0; index < cb->nr_pages; index++) {
348 page = cb->compressed_pages[index];
349 page->mapping = NULL;
353 /* finally free the cb struct */
354 kfree(cb->compressed_pages);
361 * worker function to build and submit bios for previously compressed pages.
362 * The corresponding pages in the inode should be marked for writeback
363 * and the compressed pages should have a reference on them for dropping
364 * when the IO is complete.
366 * This also checksums the file bytes and gets things ready for
369 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
370 unsigned long len, u64 disk_start,
371 unsigned long compressed_len,
372 struct page **compressed_pages,
373 unsigned long nr_pages,
374 unsigned int write_flags,
375 struct cgroup_subsys_state *blkcg_css)
377 struct btrfs_fs_info *fs_info = inode->root->fs_info;
378 struct bio *bio = NULL;
379 struct compressed_bio *cb;
380 unsigned long bytes_left;
383 u64 first_byte = disk_start;
385 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
387 WARN_ON(!PAGE_ALIGNED(start));
388 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
390 return BLK_STS_RESOURCE;
391 refcount_set(&cb->pending_bios, 0);
393 cb->inode = &inode->vfs_inode;
397 cb->compressed_pages = compressed_pages;
398 cb->compressed_len = compressed_len;
400 cb->nr_pages = nr_pages;
402 bio = btrfs_bio_alloc(first_byte);
403 bio->bi_opf = REQ_OP_WRITE | write_flags;
404 bio->bi_private = cb;
405 bio->bi_end_io = end_compressed_bio_write;
408 bio->bi_opf |= REQ_CGROUP_PUNT;
409 kthread_associate_blkcg(blkcg_css);
411 refcount_set(&cb->pending_bios, 1);
413 /* create and submit bios for the compressed pages */
414 bytes_left = compressed_len;
415 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
418 page = compressed_pages[pg_index];
419 page->mapping = inode->vfs_inode.i_mapping;
420 if (bio->bi_iter.bi_size)
421 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
424 page->mapping = NULL;
425 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
428 * inc the count before we submit the bio so
429 * we know the end IO handler won't happen before
430 * we inc the count. Otherwise, the cb might get
431 * freed before we're done setting it up
433 refcount_inc(&cb->pending_bios);
434 ret = btrfs_bio_wq_end_io(fs_info, bio,
435 BTRFS_WQ_ENDIO_DATA);
436 BUG_ON(ret); /* -ENOMEM */
439 ret = btrfs_csum_one_bio(inode, bio, start, 1);
440 BUG_ON(ret); /* -ENOMEM */
443 ret = btrfs_map_bio(fs_info, bio, 0);
445 bio->bi_status = ret;
449 bio = btrfs_bio_alloc(first_byte);
450 bio->bi_opf = REQ_OP_WRITE | write_flags;
451 bio->bi_private = cb;
452 bio->bi_end_io = end_compressed_bio_write;
454 bio->bi_opf |= REQ_CGROUP_PUNT;
455 bio_add_page(bio, page, PAGE_SIZE, 0);
457 if (bytes_left < PAGE_SIZE) {
459 "bytes left %lu compress len %lu nr %lu",
460 bytes_left, cb->compressed_len, cb->nr_pages);
462 bytes_left -= PAGE_SIZE;
463 first_byte += PAGE_SIZE;
467 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
468 BUG_ON(ret); /* -ENOMEM */
471 ret = btrfs_csum_one_bio(inode, bio, start, 1);
472 BUG_ON(ret); /* -ENOMEM */
475 ret = btrfs_map_bio(fs_info, bio, 0);
477 bio->bi_status = ret;
482 kthread_associate_blkcg(NULL);
487 static u64 bio_end_offset(struct bio *bio)
489 struct bio_vec *last = bio_last_bvec_all(bio);
491 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
494 static noinline int add_ra_bio_pages(struct inode *inode,
496 struct compressed_bio *cb)
498 unsigned long end_index;
499 unsigned long pg_index;
501 u64 isize = i_size_read(inode);
504 unsigned long nr_pages = 0;
505 struct extent_map *em;
506 struct address_space *mapping = inode->i_mapping;
507 struct extent_map_tree *em_tree;
508 struct extent_io_tree *tree;
512 last_offset = bio_end_offset(cb->orig_bio);
513 em_tree = &BTRFS_I(inode)->extent_tree;
514 tree = &BTRFS_I(inode)->io_tree;
519 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
521 while (last_offset < compressed_end) {
522 pg_index = last_offset >> PAGE_SHIFT;
524 if (pg_index > end_index)
527 page = xa_load(&mapping->i_pages, pg_index);
528 if (page && !xa_is_value(page)) {
535 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
540 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
546 * at this point, we have a locked page in the page cache
547 * for these bytes in the file. But, we have to make
548 * sure they map to this compressed extent on disk.
550 ret = set_page_extent_mapped(page);
557 end = last_offset + PAGE_SIZE - 1;
558 lock_extent(tree, last_offset, end);
559 read_lock(&em_tree->lock);
560 em = lookup_extent_mapping(em_tree, last_offset,
562 read_unlock(&em_tree->lock);
564 if (!em || last_offset < em->start ||
565 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
566 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
568 unlock_extent(tree, last_offset, end);
575 if (page->index == end_index) {
577 size_t zero_offset = offset_in_page(isize);
581 zeros = PAGE_SIZE - zero_offset;
582 userpage = kmap_atomic(page);
583 memset(userpage + zero_offset, 0, zeros);
584 flush_dcache_page(page);
585 kunmap_atomic(userpage);
589 ret = bio_add_page(cb->orig_bio, page,
592 if (ret == PAGE_SIZE) {
596 unlock_extent(tree, last_offset, end);
602 last_offset += PAGE_SIZE;
608 * for a compressed read, the bio we get passed has all the inode pages
609 * in it. We don't actually do IO on those pages but allocate new ones
610 * to hold the compressed pages on disk.
612 * bio->bi_iter.bi_sector points to the compressed extent on disk
613 * bio->bi_io_vec points to all of the inode pages
615 * After the compressed pages are read, we copy the bytes into the
616 * bio we were passed and then call the bio end_io calls
618 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
619 int mirror_num, unsigned long bio_flags)
621 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
622 struct extent_map_tree *em_tree;
623 struct compressed_bio *cb;
624 unsigned long compressed_len;
625 unsigned long nr_pages;
626 unsigned long pg_index;
628 struct bio *comp_bio;
629 u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
632 struct extent_map *em;
633 blk_status_t ret = BLK_STS_RESOURCE;
637 em_tree = &BTRFS_I(inode)->extent_tree;
639 /* we need the actual starting offset of this extent in the file */
640 read_lock(&em_tree->lock);
641 em = lookup_extent_mapping(em_tree,
642 page_offset(bio_first_page_all(bio)),
643 fs_info->sectorsize);
644 read_unlock(&em_tree->lock);
646 return BLK_STS_IOERR;
648 compressed_len = em->block_len;
649 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
653 refcount_set(&cb->pending_bios, 0);
656 cb->mirror_num = mirror_num;
659 cb->start = em->orig_start;
661 em_start = em->start;
666 cb->len = bio->bi_iter.bi_size;
667 cb->compressed_len = compressed_len;
668 cb->compress_type = extent_compress_type(bio_flags);
671 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
672 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
674 if (!cb->compressed_pages)
677 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
678 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
680 if (!cb->compressed_pages[pg_index]) {
681 faili = pg_index - 1;
682 ret = BLK_STS_RESOURCE;
686 faili = nr_pages - 1;
687 cb->nr_pages = nr_pages;
689 add_ra_bio_pages(inode, em_start + em_len, cb);
691 /* include any pages we added in add_ra-bio_pages */
692 cb->len = bio->bi_iter.bi_size;
694 comp_bio = btrfs_bio_alloc(cur_disk_byte);
695 comp_bio->bi_opf = REQ_OP_READ;
696 comp_bio->bi_private = cb;
697 comp_bio->bi_end_io = end_compressed_bio_read;
698 refcount_set(&cb->pending_bios, 1);
700 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
701 u32 pg_len = PAGE_SIZE;
705 * To handle subpage case, we need to make sure the bio only
706 * covers the range we need.
708 * If we're at the last page, truncate the length to only cover
709 * the remaining part.
711 if (pg_index == nr_pages - 1)
712 pg_len = min_t(u32, PAGE_SIZE,
713 compressed_len - pg_index * PAGE_SIZE);
715 page = cb->compressed_pages[pg_index];
716 page->mapping = inode->i_mapping;
717 page->index = em_start >> PAGE_SHIFT;
719 if (comp_bio->bi_iter.bi_size)
720 submit = btrfs_bio_fits_in_stripe(page, pg_len,
723 page->mapping = NULL;
724 if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
725 unsigned int nr_sectors;
727 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
728 BTRFS_WQ_ENDIO_DATA);
729 BUG_ON(ret); /* -ENOMEM */
732 * inc the count before we submit the bio so
733 * we know the end IO handler won't happen before
734 * we inc the count. Otherwise, the cb might get
735 * freed before we're done setting it up
737 refcount_inc(&cb->pending_bios);
739 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
740 BUG_ON(ret); /* -ENOMEM */
742 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
743 fs_info->sectorsize);
744 sums += fs_info->csum_size * nr_sectors;
746 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
748 comp_bio->bi_status = ret;
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;
757 bio_add_page(comp_bio, page, pg_len, 0);
759 cur_disk_byte += pg_len;
762 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
763 BUG_ON(ret); /* -ENOMEM */
765 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
766 BUG_ON(ret); /* -ENOMEM */
768 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
770 comp_bio->bi_status = ret;
778 __free_page(cb->compressed_pages[faili]);
782 kfree(cb->compressed_pages);
791 * Heuristic uses systematic sampling to collect data from the input data
792 * range, the logic can be tuned by the following constants:
794 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
795 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
797 #define SAMPLING_READ_SIZE (16)
798 #define SAMPLING_INTERVAL (256)
801 * For statistical analysis of the input data we consider bytes that form a
802 * Galois Field of 256 objects. Each object has an attribute count, ie. how
803 * many times the object appeared in the sample.
805 #define BUCKET_SIZE (256)
808 * The size of the sample is based on a statistical sampling rule of thumb.
809 * The common way is to perform sampling tests as long as the number of
810 * elements in each cell is at least 5.
812 * Instead of 5, we choose 32 to obtain more accurate results.
813 * If the data contain the maximum number of symbols, which is 256, we obtain a
814 * sample size bound by 8192.
816 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
817 * from up to 512 locations.
819 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
820 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
826 struct heuristic_ws {
827 /* Partial copy of input data */
830 /* Buckets store counters for each byte value */
831 struct bucket_item *bucket;
833 struct bucket_item *bucket_b;
834 struct list_head list;
837 static struct workspace_manager heuristic_wsm;
839 static void free_heuristic_ws(struct list_head *ws)
841 struct heuristic_ws *workspace;
843 workspace = list_entry(ws, struct heuristic_ws, list);
845 kvfree(workspace->sample);
846 kfree(workspace->bucket);
847 kfree(workspace->bucket_b);
851 static struct list_head *alloc_heuristic_ws(unsigned int level)
853 struct heuristic_ws *ws;
855 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
857 return ERR_PTR(-ENOMEM);
859 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
863 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
867 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
871 INIT_LIST_HEAD(&ws->list);
874 free_heuristic_ws(&ws->list);
875 return ERR_PTR(-ENOMEM);
878 const struct btrfs_compress_op btrfs_heuristic_compress = {
879 .workspace_manager = &heuristic_wsm,
882 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
883 /* The heuristic is represented as compression type 0 */
884 &btrfs_heuristic_compress,
885 &btrfs_zlib_compress,
887 &btrfs_zstd_compress,
890 static struct list_head *alloc_workspace(int type, unsigned int level)
893 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
894 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
895 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
896 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
899 * This can't happen, the type is validated several times
900 * before we get here.
906 static void free_workspace(int type, struct list_head *ws)
909 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
910 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
911 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
912 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
915 * This can't happen, the type is validated several times
916 * before we get here.
922 static void btrfs_init_workspace_manager(int type)
924 struct workspace_manager *wsm;
925 struct list_head *workspace;
927 wsm = btrfs_compress_op[type]->workspace_manager;
928 INIT_LIST_HEAD(&wsm->idle_ws);
929 spin_lock_init(&wsm->ws_lock);
930 atomic_set(&wsm->total_ws, 0);
931 init_waitqueue_head(&wsm->ws_wait);
934 * Preallocate one workspace for each compression type so we can
935 * guarantee forward progress in the worst case
937 workspace = alloc_workspace(type, 0);
938 if (IS_ERR(workspace)) {
940 "BTRFS: cannot preallocate compression workspace, will try later\n");
942 atomic_set(&wsm->total_ws, 1);
944 list_add(workspace, &wsm->idle_ws);
948 static void btrfs_cleanup_workspace_manager(int type)
950 struct workspace_manager *wsman;
951 struct list_head *ws;
953 wsman = btrfs_compress_op[type]->workspace_manager;
954 while (!list_empty(&wsman->idle_ws)) {
955 ws = wsman->idle_ws.next;
957 free_workspace(type, ws);
958 atomic_dec(&wsman->total_ws);
963 * This finds an available workspace or allocates a new one.
964 * If it's not possible to allocate a new one, waits until there's one.
965 * Preallocation makes a forward progress guarantees and we do not return
968 struct list_head *btrfs_get_workspace(int type, unsigned int level)
970 struct workspace_manager *wsm;
971 struct list_head *workspace;
972 int cpus = num_online_cpus();
974 struct list_head *idle_ws;
977 wait_queue_head_t *ws_wait;
980 wsm = btrfs_compress_op[type]->workspace_manager;
981 idle_ws = &wsm->idle_ws;
982 ws_lock = &wsm->ws_lock;
983 total_ws = &wsm->total_ws;
984 ws_wait = &wsm->ws_wait;
985 free_ws = &wsm->free_ws;
989 if (!list_empty(idle_ws)) {
990 workspace = idle_ws->next;
993 spin_unlock(ws_lock);
997 if (atomic_read(total_ws) > cpus) {
1000 spin_unlock(ws_lock);
1001 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1002 if (atomic_read(total_ws) > cpus && !*free_ws)
1004 finish_wait(ws_wait, &wait);
1007 atomic_inc(total_ws);
1008 spin_unlock(ws_lock);
1011 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1012 * to turn it off here because we might get called from the restricted
1013 * context of btrfs_compress_bio/btrfs_compress_pages
1015 nofs_flag = memalloc_nofs_save();
1016 workspace = alloc_workspace(type, level);
1017 memalloc_nofs_restore(nofs_flag);
1019 if (IS_ERR(workspace)) {
1020 atomic_dec(total_ws);
1024 * Do not return the error but go back to waiting. There's a
1025 * workspace preallocated for each type and the compression
1026 * time is bounded so we get to a workspace eventually. This
1027 * makes our caller's life easier.
1029 * To prevent silent and low-probability deadlocks (when the
1030 * initial preallocation fails), check if there are any
1031 * workspaces at all.
1033 if (atomic_read(total_ws) == 0) {
1034 static DEFINE_RATELIMIT_STATE(_rs,
1035 /* once per minute */ 60 * HZ,
1038 if (__ratelimit(&_rs)) {
1039 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1047 static struct list_head *get_workspace(int type, int level)
1050 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1051 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1052 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1053 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1056 * This can't happen, the type is validated several times
1057 * before we get here.
1064 * put a workspace struct back on the list or free it if we have enough
1065 * idle ones sitting around
1067 void btrfs_put_workspace(int type, struct list_head *ws)
1069 struct workspace_manager *wsm;
1070 struct list_head *idle_ws;
1071 spinlock_t *ws_lock;
1073 wait_queue_head_t *ws_wait;
1076 wsm = btrfs_compress_op[type]->workspace_manager;
1077 idle_ws = &wsm->idle_ws;
1078 ws_lock = &wsm->ws_lock;
1079 total_ws = &wsm->total_ws;
1080 ws_wait = &wsm->ws_wait;
1081 free_ws = &wsm->free_ws;
1084 if (*free_ws <= num_online_cpus()) {
1085 list_add(ws, idle_ws);
1087 spin_unlock(ws_lock);
1090 spin_unlock(ws_lock);
1092 free_workspace(type, ws);
1093 atomic_dec(total_ws);
1095 cond_wake_up(ws_wait);
1098 static void put_workspace(int type, struct list_head *ws)
1101 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1102 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1103 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1104 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1107 * This can't happen, the type is validated several times
1108 * before we get here.
1115 * Adjust @level according to the limits of the compression algorithm or
1116 * fallback to default
1118 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1120 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1123 level = ops->default_level;
1125 level = min(level, ops->max_level);
1131 * Given an address space and start and length, compress the bytes into @pages
1132 * that are allocated on demand.
1134 * @type_level is encoded algorithm and level, where level 0 means whatever
1135 * default the algorithm chooses and is opaque here;
1136 * - compression algo are 0-3
1137 * - the level are bits 4-7
1139 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1140 * and returns number of actually allocated pages
1142 * @total_in is used to return the number of bytes actually read. It
1143 * may be smaller than the input length if we had to exit early because we
1144 * ran out of room in the pages array or because we cross the
1145 * max_out threshold.
1147 * @total_out is an in/out parameter, must be set to the input length and will
1148 * be also used to return the total number of compressed bytes
1150 * @max_out tells us the max number of bytes that we're allowed to
1153 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1154 u64 start, struct page **pages,
1155 unsigned long *out_pages,
1156 unsigned long *total_in,
1157 unsigned long *total_out)
1159 int type = btrfs_compress_type(type_level);
1160 int level = btrfs_compress_level(type_level);
1161 struct list_head *workspace;
1164 level = btrfs_compress_set_level(type, level);
1165 workspace = get_workspace(type, level);
1166 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1167 out_pages, total_in, total_out);
1168 put_workspace(type, workspace);
1173 * pages_in is an array of pages with compressed data.
1175 * disk_start is the starting logical offset of this array in the file
1177 * orig_bio contains the pages from the file that we want to decompress into
1179 * srclen is the number of bytes in pages_in
1181 * The basic idea is that we have a bio that was created by readpages.
1182 * The pages in the bio are for the uncompressed data, and they may not
1183 * be contiguous. They all correspond to the range of bytes covered by
1184 * the compressed extent.
1186 static int btrfs_decompress_bio(struct compressed_bio *cb)
1188 struct list_head *workspace;
1190 int type = cb->compress_type;
1192 workspace = get_workspace(type, 0);
1193 ret = compression_decompress_bio(type, workspace, cb);
1194 put_workspace(type, workspace);
1200 * a less complex decompression routine. Our compressed data fits in a
1201 * single page, and we want to read a single page out of it.
1202 * start_byte tells us the offset into the compressed data we're interested in
1204 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1205 unsigned long start_byte, size_t srclen, size_t destlen)
1207 struct list_head *workspace;
1210 workspace = get_workspace(type, 0);
1211 ret = compression_decompress(type, workspace, data_in, dest_page,
1212 start_byte, srclen, destlen);
1213 put_workspace(type, workspace);
1218 void __init btrfs_init_compress(void)
1220 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1221 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1222 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1223 zstd_init_workspace_manager();
1226 void __cold btrfs_exit_compress(void)
1228 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1229 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1230 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1231 zstd_cleanup_workspace_manager();
1235 * Copy uncompressed data from working buffer to pages.
1237 * buf_start is the byte offset we're of the start of our workspace buffer.
1239 * total_out is the last byte of the buffer
1241 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1242 unsigned long total_out, u64 disk_start,
1245 unsigned long buf_offset;
1246 unsigned long current_buf_start;
1247 unsigned long start_byte;
1248 unsigned long prev_start_byte;
1249 unsigned long working_bytes = total_out - buf_start;
1250 unsigned long bytes;
1252 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1255 * start byte is the first byte of the page we're currently
1256 * copying into relative to the start of the compressed data.
1258 start_byte = page_offset(bvec.bv_page) - disk_start;
1260 /* we haven't yet hit data corresponding to this page */
1261 if (total_out <= start_byte)
1265 * the start of the data we care about is offset into
1266 * the middle of our working buffer
1268 if (total_out > start_byte && buf_start < start_byte) {
1269 buf_offset = start_byte - buf_start;
1270 working_bytes -= buf_offset;
1274 current_buf_start = buf_start;
1276 /* copy bytes from the working buffer into the pages */
1277 while (working_bytes > 0) {
1278 bytes = min_t(unsigned long, bvec.bv_len,
1279 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1280 bytes = min(bytes, working_bytes);
1282 kaddr = kmap_atomic(bvec.bv_page);
1283 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1284 kunmap_atomic(kaddr);
1285 flush_dcache_page(bvec.bv_page);
1287 buf_offset += bytes;
1288 working_bytes -= bytes;
1289 current_buf_start += bytes;
1291 /* check if we need to pick another page */
1292 bio_advance(bio, bytes);
1293 if (!bio->bi_iter.bi_size)
1295 bvec = bio_iter_iovec(bio, bio->bi_iter);
1296 prev_start_byte = start_byte;
1297 start_byte = page_offset(bvec.bv_page) - disk_start;
1300 * We need to make sure we're only adjusting
1301 * our offset into compression working buffer when
1302 * we're switching pages. Otherwise we can incorrectly
1303 * keep copying when we were actually done.
1305 if (start_byte != prev_start_byte) {
1307 * make sure our new page is covered by this
1310 if (total_out <= start_byte)
1314 * the next page in the biovec might not be adjacent
1315 * to the last page, but it might still be found
1316 * inside this working buffer. bump our offset pointer
1318 if (total_out > start_byte &&
1319 current_buf_start < start_byte) {
1320 buf_offset = start_byte - buf_start;
1321 working_bytes = total_out - start_byte;
1322 current_buf_start = buf_start + buf_offset;
1331 * Shannon Entropy calculation
1333 * Pure byte distribution analysis fails to determine compressibility of data.
1334 * Try calculating entropy to estimate the average minimum number of bits
1335 * needed to encode the sampled data.
1337 * For convenience, return the percentage of needed bits, instead of amount of
1340 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1341 * and can be compressible with high probability
1343 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1345 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1347 #define ENTROPY_LVL_ACEPTABLE (65)
1348 #define ENTROPY_LVL_HIGH (80)
1351 * For increasead precision in shannon_entropy calculation,
1352 * let's do pow(n, M) to save more digits after comma:
1354 * - maximum int bit length is 64
1355 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1356 * - 13 * 4 = 52 < 64 -> M = 4
1360 static inline u32 ilog2_w(u64 n)
1362 return ilog2(n * n * n * n);
1365 static u32 shannon_entropy(struct heuristic_ws *ws)
1367 const u32 entropy_max = 8 * ilog2_w(2);
1368 u32 entropy_sum = 0;
1369 u32 p, p_base, sz_base;
1372 sz_base = ilog2_w(ws->sample_size);
1373 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1374 p = ws->bucket[i].count;
1375 p_base = ilog2_w(p);
1376 entropy_sum += p * (sz_base - p_base);
1379 entropy_sum /= ws->sample_size;
1380 return entropy_sum * 100 / entropy_max;
1383 #define RADIX_BASE 4U
1384 #define COUNTERS_SIZE (1U << RADIX_BASE)
1386 static u8 get4bits(u64 num, int shift) {
1391 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1396 * Use 4 bits as radix base
1397 * Use 16 u32 counters for calculating new position in buf array
1399 * @array - array that will be sorted
1400 * @array_buf - buffer array to store sorting results
1401 * must be equal in size to @array
1404 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1409 u32 counters[COUNTERS_SIZE];
1417 * Try avoid useless loop iterations for small numbers stored in big
1418 * counters. Example: 48 33 4 ... in 64bit array
1420 max_num = array[0].count;
1421 for (i = 1; i < num; i++) {
1422 buf_num = array[i].count;
1423 if (buf_num > max_num)
1427 buf_num = ilog2(max_num);
1428 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1431 while (shift < bitlen) {
1432 memset(counters, 0, sizeof(counters));
1434 for (i = 0; i < num; i++) {
1435 buf_num = array[i].count;
1436 addr = get4bits(buf_num, shift);
1440 for (i = 1; i < COUNTERS_SIZE; i++)
1441 counters[i] += counters[i - 1];
1443 for (i = num - 1; i >= 0; i--) {
1444 buf_num = array[i].count;
1445 addr = get4bits(buf_num, shift);
1447 new_addr = counters[addr];
1448 array_buf[new_addr] = array[i];
1451 shift += RADIX_BASE;
1454 * Normal radix expects to move data from a temporary array, to
1455 * the main one. But that requires some CPU time. Avoid that
1456 * by doing another sort iteration to original array instead of
1459 memset(counters, 0, sizeof(counters));
1461 for (i = 0; i < num; i ++) {
1462 buf_num = array_buf[i].count;
1463 addr = get4bits(buf_num, shift);
1467 for (i = 1; i < COUNTERS_SIZE; i++)
1468 counters[i] += counters[i - 1];
1470 for (i = num - 1; i >= 0; i--) {
1471 buf_num = array_buf[i].count;
1472 addr = get4bits(buf_num, shift);
1474 new_addr = counters[addr];
1475 array[new_addr] = array_buf[i];
1478 shift += RADIX_BASE;
1483 * Size of the core byte set - how many bytes cover 90% of the sample
1485 * There are several types of structured binary data that use nearly all byte
1486 * values. The distribution can be uniform and counts in all buckets will be
1487 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1489 * Other possibility is normal (Gaussian) distribution, where the data could
1490 * be potentially compressible, but we have to take a few more steps to decide
1493 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1494 * compression algo can easy fix that
1495 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1496 * probability is not compressible
1498 #define BYTE_CORE_SET_LOW (64)
1499 #define BYTE_CORE_SET_HIGH (200)
1501 static int byte_core_set_size(struct heuristic_ws *ws)
1504 u32 coreset_sum = 0;
1505 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1506 struct bucket_item *bucket = ws->bucket;
1508 /* Sort in reverse order */
1509 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1511 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1512 coreset_sum += bucket[i].count;
1514 if (coreset_sum > core_set_threshold)
1517 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1518 coreset_sum += bucket[i].count;
1519 if (coreset_sum > core_set_threshold)
1527 * Count byte values in buckets.
1528 * This heuristic can detect textual data (configs, xml, json, html, etc).
1529 * Because in most text-like data byte set is restricted to limited number of
1530 * possible characters, and that restriction in most cases makes data easy to
1533 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1534 * less - compressible
1535 * more - need additional analysis
1537 #define BYTE_SET_THRESHOLD (64)
1539 static u32 byte_set_size(const struct heuristic_ws *ws)
1542 u32 byte_set_size = 0;
1544 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1545 if (ws->bucket[i].count > 0)
1550 * Continue collecting count of byte values in buckets. If the byte
1551 * set size is bigger then the threshold, it's pointless to continue,
1552 * the detection technique would fail for this type of data.
1554 for (; i < BUCKET_SIZE; i++) {
1555 if (ws->bucket[i].count > 0) {
1557 if (byte_set_size > BYTE_SET_THRESHOLD)
1558 return byte_set_size;
1562 return byte_set_size;
1565 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1567 const u32 half_of_sample = ws->sample_size / 2;
1568 const u8 *data = ws->sample;
1570 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1573 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1574 struct heuristic_ws *ws)
1577 u64 index, index_end;
1578 u32 i, curr_sample_pos;
1582 * Compression handles the input data by chunks of 128KiB
1583 * (defined by BTRFS_MAX_UNCOMPRESSED)
1585 * We do the same for the heuristic and loop over the whole range.
1587 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1588 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1590 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1591 end = start + BTRFS_MAX_UNCOMPRESSED;
1593 index = start >> PAGE_SHIFT;
1594 index_end = end >> PAGE_SHIFT;
1596 /* Don't miss unaligned end */
1597 if (!IS_ALIGNED(end, PAGE_SIZE))
1600 curr_sample_pos = 0;
1601 while (index < index_end) {
1602 page = find_get_page(inode->i_mapping, index);
1603 in_data = kmap(page);
1604 /* Handle case where the start is not aligned to PAGE_SIZE */
1605 i = start % PAGE_SIZE;
1606 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1607 /* Don't sample any garbage from the last page */
1608 if (start > end - SAMPLING_READ_SIZE)
1610 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1611 SAMPLING_READ_SIZE);
1612 i += SAMPLING_INTERVAL;
1613 start += SAMPLING_INTERVAL;
1614 curr_sample_pos += SAMPLING_READ_SIZE;
1622 ws->sample_size = curr_sample_pos;
1626 * Compression heuristic.
1628 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1629 * quickly (compared to direct compression) detect data characteristics
1630 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1633 * The following types of analysis can be performed:
1634 * - detect mostly zero data
1635 * - detect data with low "byte set" size (text, etc)
1636 * - detect data with low/high "core byte" set
1638 * Return non-zero if the compression should be done, 0 otherwise.
1640 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1642 struct list_head *ws_list = get_workspace(0, 0);
1643 struct heuristic_ws *ws;
1648 ws = list_entry(ws_list, struct heuristic_ws, list);
1650 heuristic_collect_sample(inode, start, end, ws);
1652 if (sample_repeated_patterns(ws)) {
1657 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1659 for (i = 0; i < ws->sample_size; i++) {
1660 byte = ws->sample[i];
1661 ws->bucket[byte].count++;
1664 i = byte_set_size(ws);
1665 if (i < BYTE_SET_THRESHOLD) {
1670 i = byte_core_set_size(ws);
1671 if (i <= BYTE_CORE_SET_LOW) {
1676 if (i >= BYTE_CORE_SET_HIGH) {
1681 i = shannon_entropy(ws);
1682 if (i <= ENTROPY_LVL_ACEPTABLE) {
1688 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1689 * needed to give green light to compression.
1691 * For now just assume that compression at that level is not worth the
1692 * resources because:
1694 * 1. it is possible to defrag the data later
1696 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1697 * values, every bucket has counter at level ~54. The heuristic would
1698 * be confused. This can happen when data have some internal repeated
1699 * patterns like "abbacbbc...". This can be detected by analyzing
1700 * pairs of bytes, which is too costly.
1702 if (i < ENTROPY_LVL_HIGH) {
1711 put_workspace(0, ws_list);
1716 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1717 * level, unrecognized string will set the default level
1719 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1721 unsigned int level = 0;
1727 if (str[0] == ':') {
1728 ret = kstrtouint(str + 1, 10, &level);
1733 level = btrfs_compress_set_level(type, level);