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 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
136 return sizeof(struct compressed_bio) +
137 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
140 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
143 struct btrfs_fs_info *fs_info = inode->root->fs_info;
144 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
145 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
149 u8 csum[BTRFS_CSUM_SIZE];
150 struct compressed_bio *cb = bio->bi_private;
151 u8 *cb_sum = cb->sums;
153 if (inode->flags & BTRFS_INODE_NODATASUM)
156 shash->tfm = fs_info->csum_shash;
158 for (i = 0; i < cb->nr_pages; i++) {
159 page = cb->compressed_pages[i];
161 kaddr = kmap_atomic(page);
162 crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum);
163 kunmap_atomic(kaddr);
165 if (memcmp(&csum, cb_sum, csum_size)) {
166 btrfs_print_data_csum_error(inode, disk_start,
167 csum, cb_sum, cb->mirror_num);
168 if (btrfs_io_bio(bio)->device)
169 btrfs_dev_stat_inc_and_print(
170 btrfs_io_bio(bio)->device,
171 BTRFS_DEV_STAT_CORRUPTION_ERRS);
179 /* when we finish reading compressed pages from the disk, we
180 * decompress them and then run the bio end_io routines on the
181 * decompressed pages (in the inode address space).
183 * This allows the checksumming and other IO error handling routines
186 * The compressed pages are freed here, and it must be run
189 static void end_compressed_bio_read(struct bio *bio)
191 struct compressed_bio *cb = bio->bi_private;
195 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
201 /* if there are more bios still pending for this compressed
204 if (!refcount_dec_and_test(&cb->pending_bios))
208 * Record the correct mirror_num in cb->orig_bio so that
209 * read-repair can work properly.
211 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
212 cb->mirror_num = mirror;
215 * Some IO in this cb have failed, just skip checksum as there
216 * is no way it could be correct.
222 ret = check_compressed_csum(BTRFS_I(inode), bio,
223 (u64)bio->bi_iter.bi_sector << 9);
227 /* ok, we're the last bio for this extent, lets start
230 ret = btrfs_decompress_bio(cb);
236 /* release the compressed pages */
238 for (index = 0; index < cb->nr_pages; index++) {
239 page = cb->compressed_pages[index];
240 page->mapping = NULL;
244 /* do io completion on the original bio */
246 bio_io_error(cb->orig_bio);
248 struct bio_vec *bvec;
249 struct bvec_iter_all iter_all;
252 * we have verified the checksum already, set page
253 * checked so the end_io handlers know about it
255 ASSERT(!bio_flagged(bio, BIO_CLONED));
256 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
257 SetPageChecked(bvec->bv_page);
259 bio_endio(cb->orig_bio);
262 /* finally free the cb struct */
263 kfree(cb->compressed_pages);
270 * Clear the writeback bits on all of the file
271 * pages for a compressed write
273 static noinline void end_compressed_writeback(struct inode *inode,
274 const struct compressed_bio *cb)
276 unsigned long index = cb->start >> PAGE_SHIFT;
277 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
278 struct page *pages[16];
279 unsigned long nr_pages = end_index - index + 1;
284 mapping_set_error(inode->i_mapping, -EIO);
286 while (nr_pages > 0) {
287 ret = find_get_pages_contig(inode->i_mapping, index,
289 nr_pages, ARRAY_SIZE(pages)), pages);
295 for (i = 0; i < ret; i++) {
297 SetPageError(pages[i]);
298 end_page_writeback(pages[i]);
304 /* the inode may be gone now */
308 * do the cleanup once all the compressed pages hit the disk.
309 * This will clear writeback on the file pages and free the compressed
312 * This also calls the writeback end hooks for the file pages so that
313 * metadata and checksums can be updated in the file.
315 static void end_compressed_bio_write(struct bio *bio)
317 struct compressed_bio *cb = bio->bi_private;
325 /* if there are more bios still pending for this compressed
328 if (!refcount_dec_and_test(&cb->pending_bios))
331 /* ok, we're the last bio for this extent, step one is to
332 * call back into the FS and do all the end_io operations
335 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
336 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
337 cb->start, cb->start + cb->len - 1,
338 bio->bi_status == BLK_STS_OK);
339 cb->compressed_pages[0]->mapping = NULL;
341 end_compressed_writeback(inode, cb);
342 /* note, our inode could be gone now */
345 * release the compressed pages, these came from alloc_page and
346 * are not attached to the inode at all
349 for (index = 0; index < cb->nr_pages; index++) {
350 page = cb->compressed_pages[index];
351 page->mapping = NULL;
355 /* finally free the cb struct */
356 kfree(cb->compressed_pages);
363 * worker function to build and submit bios for previously compressed pages.
364 * The corresponding pages in the inode should be marked for writeback
365 * and the compressed pages should have a reference on them for dropping
366 * when the IO is complete.
368 * This also checksums the file bytes and gets things ready for
371 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
372 unsigned long len, u64 disk_start,
373 unsigned long compressed_len,
374 struct page **compressed_pages,
375 unsigned long nr_pages,
376 unsigned int write_flags,
377 struct cgroup_subsys_state *blkcg_css)
379 struct btrfs_fs_info *fs_info = inode->root->fs_info;
380 struct bio *bio = NULL;
381 struct compressed_bio *cb;
382 unsigned long bytes_left;
385 u64 first_byte = disk_start;
387 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
389 WARN_ON(!PAGE_ALIGNED(start));
390 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
392 return BLK_STS_RESOURCE;
393 refcount_set(&cb->pending_bios, 0);
395 cb->inode = &inode->vfs_inode;
399 cb->compressed_pages = compressed_pages;
400 cb->compressed_len = compressed_len;
402 cb->nr_pages = nr_pages;
404 bio = btrfs_bio_alloc(first_byte);
405 bio->bi_opf = REQ_OP_WRITE | write_flags;
406 bio->bi_private = cb;
407 bio->bi_end_io = end_compressed_bio_write;
410 bio->bi_opf |= REQ_CGROUP_PUNT;
411 kthread_associate_blkcg(blkcg_css);
413 refcount_set(&cb->pending_bios, 1);
415 /* create and submit bios for the compressed pages */
416 bytes_left = compressed_len;
417 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
420 page = compressed_pages[pg_index];
421 page->mapping = inode->vfs_inode.i_mapping;
422 if (bio->bi_iter.bi_size)
423 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
426 page->mapping = NULL;
427 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
430 * inc the count before we submit the bio so
431 * we know the end IO handler won't happen before
432 * we inc the count. Otherwise, the cb might get
433 * freed before we're done setting it up
435 refcount_inc(&cb->pending_bios);
436 ret = btrfs_bio_wq_end_io(fs_info, bio,
437 BTRFS_WQ_ENDIO_DATA);
438 BUG_ON(ret); /* -ENOMEM */
441 ret = btrfs_csum_one_bio(inode, bio, start, 1);
442 BUG_ON(ret); /* -ENOMEM */
445 ret = btrfs_map_bio(fs_info, bio, 0);
447 bio->bi_status = ret;
451 bio = btrfs_bio_alloc(first_byte);
452 bio->bi_opf = REQ_OP_WRITE | write_flags;
453 bio->bi_private = cb;
454 bio->bi_end_io = end_compressed_bio_write;
456 bio->bi_opf |= REQ_CGROUP_PUNT;
457 bio_add_page(bio, page, PAGE_SIZE, 0);
459 if (bytes_left < PAGE_SIZE) {
461 "bytes left %lu compress len %lu nr %lu",
462 bytes_left, cb->compressed_len, cb->nr_pages);
464 bytes_left -= PAGE_SIZE;
465 first_byte += PAGE_SIZE;
469 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
470 BUG_ON(ret); /* -ENOMEM */
473 ret = btrfs_csum_one_bio(inode, bio, start, 1);
474 BUG_ON(ret); /* -ENOMEM */
477 ret = btrfs_map_bio(fs_info, bio, 0);
479 bio->bi_status = ret;
484 kthread_associate_blkcg(NULL);
489 static u64 bio_end_offset(struct bio *bio)
491 struct bio_vec *last = bio_last_bvec_all(bio);
493 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
496 static noinline int add_ra_bio_pages(struct inode *inode,
498 struct compressed_bio *cb)
500 unsigned long end_index;
501 unsigned long pg_index;
503 u64 isize = i_size_read(inode);
506 unsigned long nr_pages = 0;
507 struct extent_map *em;
508 struct address_space *mapping = inode->i_mapping;
509 struct extent_map_tree *em_tree;
510 struct extent_io_tree *tree;
514 last_offset = bio_end_offset(cb->orig_bio);
515 em_tree = &BTRFS_I(inode)->extent_tree;
516 tree = &BTRFS_I(inode)->io_tree;
521 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
523 while (last_offset < compressed_end) {
524 pg_index = last_offset >> PAGE_SHIFT;
526 if (pg_index > end_index)
529 page = xa_load(&mapping->i_pages, pg_index);
530 if (page && !xa_is_value(page)) {
537 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
542 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
547 end = last_offset + PAGE_SIZE - 1;
549 * at this point, we have a locked page in the page cache
550 * for these bytes in the file. But, we have to make
551 * sure they map to this compressed extent on disk.
553 set_page_extent_mapped(page);
554 lock_extent(tree, last_offset, end);
555 read_lock(&em_tree->lock);
556 em = lookup_extent_mapping(em_tree, last_offset,
558 read_unlock(&em_tree->lock);
560 if (!em || last_offset < em->start ||
561 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
562 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
564 unlock_extent(tree, last_offset, end);
571 if (page->index == end_index) {
573 size_t zero_offset = offset_in_page(isize);
577 zeros = PAGE_SIZE - zero_offset;
578 userpage = kmap_atomic(page);
579 memset(userpage + zero_offset, 0, zeros);
580 flush_dcache_page(page);
581 kunmap_atomic(userpage);
585 ret = bio_add_page(cb->orig_bio, page,
588 if (ret == PAGE_SIZE) {
592 unlock_extent(tree, last_offset, end);
598 last_offset += PAGE_SIZE;
604 * for a compressed read, the bio we get passed has all the inode pages
605 * in it. We don't actually do IO on those pages but allocate new ones
606 * to hold the compressed pages on disk.
608 * bio->bi_iter.bi_sector points to the compressed extent on disk
609 * bio->bi_io_vec points to all of the inode pages
611 * After the compressed pages are read, we copy the bytes into the
612 * bio we were passed and then call the bio end_io calls
614 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
615 int mirror_num, unsigned long bio_flags)
617 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
618 struct extent_map_tree *em_tree;
619 struct compressed_bio *cb;
620 unsigned long compressed_len;
621 unsigned long nr_pages;
622 unsigned long pg_index;
624 struct bio *comp_bio;
625 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
628 struct extent_map *em;
629 blk_status_t ret = BLK_STS_RESOURCE;
631 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
634 em_tree = &BTRFS_I(inode)->extent_tree;
636 /* we need the actual starting offset of this extent in the file */
637 read_lock(&em_tree->lock);
638 em = lookup_extent_mapping(em_tree,
639 page_offset(bio_first_page_all(bio)),
641 read_unlock(&em_tree->lock);
643 return BLK_STS_IOERR;
645 compressed_len = em->block_len;
646 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
650 refcount_set(&cb->pending_bios, 0);
653 cb->mirror_num = mirror_num;
656 cb->start = em->orig_start;
658 em_start = em->start;
663 cb->len = bio->bi_iter.bi_size;
664 cb->compressed_len = compressed_len;
665 cb->compress_type = extent_compress_type(bio_flags);
668 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
669 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
671 if (!cb->compressed_pages)
674 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
675 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
677 if (!cb->compressed_pages[pg_index]) {
678 faili = pg_index - 1;
679 ret = BLK_STS_RESOURCE;
683 faili = nr_pages - 1;
684 cb->nr_pages = nr_pages;
686 add_ra_bio_pages(inode, em_start + em_len, cb);
688 /* include any pages we added in add_ra-bio_pages */
689 cb->len = bio->bi_iter.bi_size;
691 comp_bio = btrfs_bio_alloc(cur_disk_byte);
692 comp_bio->bi_opf = REQ_OP_READ;
693 comp_bio->bi_private = cb;
694 comp_bio->bi_end_io = end_compressed_bio_read;
695 refcount_set(&cb->pending_bios, 1);
697 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
700 page = cb->compressed_pages[pg_index];
701 page->mapping = inode->i_mapping;
702 page->index = em_start >> PAGE_SHIFT;
704 if (comp_bio->bi_iter.bi_size)
705 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
708 page->mapping = NULL;
709 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
711 unsigned int nr_sectors;
713 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
714 BTRFS_WQ_ENDIO_DATA);
715 BUG_ON(ret); /* -ENOMEM */
718 * inc the count before we submit the bio so
719 * we know the end IO handler won't happen before
720 * we inc the count. Otherwise, the cb might get
721 * freed before we're done setting it up
723 refcount_inc(&cb->pending_bios);
725 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
726 ret = btrfs_lookup_bio_sums(inode, comp_bio,
728 BUG_ON(ret); /* -ENOMEM */
731 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
732 fs_info->sectorsize);
733 sums += csum_size * nr_sectors;
735 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
737 comp_bio->bi_status = ret;
741 comp_bio = btrfs_bio_alloc(cur_disk_byte);
742 comp_bio->bi_opf = REQ_OP_READ;
743 comp_bio->bi_private = cb;
744 comp_bio->bi_end_io = end_compressed_bio_read;
746 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
748 cur_disk_byte += PAGE_SIZE;
751 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
752 BUG_ON(ret); /* -ENOMEM */
754 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
755 ret = btrfs_lookup_bio_sums(inode, comp_bio, (u64)-1, sums);
756 BUG_ON(ret); /* -ENOMEM */
759 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
761 comp_bio->bi_status = ret;
769 __free_page(cb->compressed_pages[faili]);
773 kfree(cb->compressed_pages);
782 * Heuristic uses systematic sampling to collect data from the input data
783 * range, the logic can be tuned by the following constants:
785 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
786 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
788 #define SAMPLING_READ_SIZE (16)
789 #define SAMPLING_INTERVAL (256)
792 * For statistical analysis of the input data we consider bytes that form a
793 * Galois Field of 256 objects. Each object has an attribute count, ie. how
794 * many times the object appeared in the sample.
796 #define BUCKET_SIZE (256)
799 * The size of the sample is based on a statistical sampling rule of thumb.
800 * The common way is to perform sampling tests as long as the number of
801 * elements in each cell is at least 5.
803 * Instead of 5, we choose 32 to obtain more accurate results.
804 * If the data contain the maximum number of symbols, which is 256, we obtain a
805 * sample size bound by 8192.
807 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
808 * from up to 512 locations.
810 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
811 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
817 struct heuristic_ws {
818 /* Partial copy of input data */
821 /* Buckets store counters for each byte value */
822 struct bucket_item *bucket;
824 struct bucket_item *bucket_b;
825 struct list_head list;
828 static struct workspace_manager heuristic_wsm;
830 static void free_heuristic_ws(struct list_head *ws)
832 struct heuristic_ws *workspace;
834 workspace = list_entry(ws, struct heuristic_ws, list);
836 kvfree(workspace->sample);
837 kfree(workspace->bucket);
838 kfree(workspace->bucket_b);
842 static struct list_head *alloc_heuristic_ws(unsigned int level)
844 struct heuristic_ws *ws;
846 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
848 return ERR_PTR(-ENOMEM);
850 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
854 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
858 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
862 INIT_LIST_HEAD(&ws->list);
865 free_heuristic_ws(&ws->list);
866 return ERR_PTR(-ENOMEM);
869 const struct btrfs_compress_op btrfs_heuristic_compress = {
870 .workspace_manager = &heuristic_wsm,
873 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
874 /* The heuristic is represented as compression type 0 */
875 &btrfs_heuristic_compress,
876 &btrfs_zlib_compress,
878 &btrfs_zstd_compress,
881 static struct list_head *alloc_workspace(int type, unsigned int level)
884 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
885 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
886 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
887 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
890 * This can't happen, the type is validated several times
891 * before we get here.
897 static void free_workspace(int type, struct list_head *ws)
900 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
901 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
902 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
903 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
906 * This can't happen, the type is validated several times
907 * before we get here.
913 static void btrfs_init_workspace_manager(int type)
915 struct workspace_manager *wsm;
916 struct list_head *workspace;
918 wsm = btrfs_compress_op[type]->workspace_manager;
919 INIT_LIST_HEAD(&wsm->idle_ws);
920 spin_lock_init(&wsm->ws_lock);
921 atomic_set(&wsm->total_ws, 0);
922 init_waitqueue_head(&wsm->ws_wait);
925 * Preallocate one workspace for each compression type so we can
926 * guarantee forward progress in the worst case
928 workspace = alloc_workspace(type, 0);
929 if (IS_ERR(workspace)) {
931 "BTRFS: cannot preallocate compression workspace, will try later\n");
933 atomic_set(&wsm->total_ws, 1);
935 list_add(workspace, &wsm->idle_ws);
939 static void btrfs_cleanup_workspace_manager(int type)
941 struct workspace_manager *wsman;
942 struct list_head *ws;
944 wsman = btrfs_compress_op[type]->workspace_manager;
945 while (!list_empty(&wsman->idle_ws)) {
946 ws = wsman->idle_ws.next;
948 free_workspace(type, ws);
949 atomic_dec(&wsman->total_ws);
954 * This finds an available workspace or allocates a new one.
955 * If it's not possible to allocate a new one, waits until there's one.
956 * Preallocation makes a forward progress guarantees and we do not return
959 struct list_head *btrfs_get_workspace(int type, unsigned int level)
961 struct workspace_manager *wsm;
962 struct list_head *workspace;
963 int cpus = num_online_cpus();
965 struct list_head *idle_ws;
968 wait_queue_head_t *ws_wait;
971 wsm = btrfs_compress_op[type]->workspace_manager;
972 idle_ws = &wsm->idle_ws;
973 ws_lock = &wsm->ws_lock;
974 total_ws = &wsm->total_ws;
975 ws_wait = &wsm->ws_wait;
976 free_ws = &wsm->free_ws;
980 if (!list_empty(idle_ws)) {
981 workspace = idle_ws->next;
984 spin_unlock(ws_lock);
988 if (atomic_read(total_ws) > cpus) {
991 spin_unlock(ws_lock);
992 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
993 if (atomic_read(total_ws) > cpus && !*free_ws)
995 finish_wait(ws_wait, &wait);
998 atomic_inc(total_ws);
999 spin_unlock(ws_lock);
1002 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1003 * to turn it off here because we might get called from the restricted
1004 * context of btrfs_compress_bio/btrfs_compress_pages
1006 nofs_flag = memalloc_nofs_save();
1007 workspace = alloc_workspace(type, level);
1008 memalloc_nofs_restore(nofs_flag);
1010 if (IS_ERR(workspace)) {
1011 atomic_dec(total_ws);
1015 * Do not return the error but go back to waiting. There's a
1016 * workspace preallocated for each type and the compression
1017 * time is bounded so we get to a workspace eventually. This
1018 * makes our caller's life easier.
1020 * To prevent silent and low-probability deadlocks (when the
1021 * initial preallocation fails), check if there are any
1022 * workspaces at all.
1024 if (atomic_read(total_ws) == 0) {
1025 static DEFINE_RATELIMIT_STATE(_rs,
1026 /* once per minute */ 60 * HZ,
1029 if (__ratelimit(&_rs)) {
1030 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1038 static struct list_head *get_workspace(int type, int level)
1041 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1042 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1043 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1044 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1047 * This can't happen, the type is validated several times
1048 * before we get here.
1055 * put a workspace struct back on the list or free it if we have enough
1056 * idle ones sitting around
1058 void btrfs_put_workspace(int type, struct list_head *ws)
1060 struct workspace_manager *wsm;
1061 struct list_head *idle_ws;
1062 spinlock_t *ws_lock;
1064 wait_queue_head_t *ws_wait;
1067 wsm = btrfs_compress_op[type]->workspace_manager;
1068 idle_ws = &wsm->idle_ws;
1069 ws_lock = &wsm->ws_lock;
1070 total_ws = &wsm->total_ws;
1071 ws_wait = &wsm->ws_wait;
1072 free_ws = &wsm->free_ws;
1075 if (*free_ws <= num_online_cpus()) {
1076 list_add(ws, idle_ws);
1078 spin_unlock(ws_lock);
1081 spin_unlock(ws_lock);
1083 free_workspace(type, ws);
1084 atomic_dec(total_ws);
1086 cond_wake_up(ws_wait);
1089 static void put_workspace(int type, struct list_head *ws)
1092 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1093 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1094 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1095 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1098 * This can't happen, the type is validated several times
1099 * before we get here.
1106 * Adjust @level according to the limits of the compression algorithm or
1107 * fallback to default
1109 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1111 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1114 level = ops->default_level;
1116 level = min(level, ops->max_level);
1122 * Given an address space and start and length, compress the bytes into @pages
1123 * that are allocated on demand.
1125 * @type_level is encoded algorithm and level, where level 0 means whatever
1126 * default the algorithm chooses and is opaque here;
1127 * - compression algo are 0-3
1128 * - the level are bits 4-7
1130 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1131 * and returns number of actually allocated pages
1133 * @total_in is used to return the number of bytes actually read. It
1134 * may be smaller than the input length if we had to exit early because we
1135 * ran out of room in the pages array or because we cross the
1136 * max_out threshold.
1138 * @total_out is an in/out parameter, must be set to the input length and will
1139 * be also used to return the total number of compressed bytes
1141 * @max_out tells us the max number of bytes that we're allowed to
1144 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1145 u64 start, struct page **pages,
1146 unsigned long *out_pages,
1147 unsigned long *total_in,
1148 unsigned long *total_out)
1150 int type = btrfs_compress_type(type_level);
1151 int level = btrfs_compress_level(type_level);
1152 struct list_head *workspace;
1155 level = btrfs_compress_set_level(type, level);
1156 workspace = get_workspace(type, level);
1157 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1158 out_pages, total_in, total_out);
1159 put_workspace(type, workspace);
1164 * pages_in is an array of pages with compressed data.
1166 * disk_start is the starting logical offset of this array in the file
1168 * orig_bio contains the pages from the file that we want to decompress into
1170 * srclen is the number of bytes in pages_in
1172 * The basic idea is that we have a bio that was created by readpages.
1173 * The pages in the bio are for the uncompressed data, and they may not
1174 * be contiguous. They all correspond to the range of bytes covered by
1175 * the compressed extent.
1177 static int btrfs_decompress_bio(struct compressed_bio *cb)
1179 struct list_head *workspace;
1181 int type = cb->compress_type;
1183 workspace = get_workspace(type, 0);
1184 ret = compression_decompress_bio(type, workspace, cb);
1185 put_workspace(type, workspace);
1191 * a less complex decompression routine. Our compressed data fits in a
1192 * single page, and we want to read a single page out of it.
1193 * start_byte tells us the offset into the compressed data we're interested in
1195 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1196 unsigned long start_byte, size_t srclen, size_t destlen)
1198 struct list_head *workspace;
1201 workspace = get_workspace(type, 0);
1202 ret = compression_decompress(type, workspace, data_in, dest_page,
1203 start_byte, srclen, destlen);
1204 put_workspace(type, workspace);
1209 void __init btrfs_init_compress(void)
1211 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1212 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1213 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1214 zstd_init_workspace_manager();
1217 void __cold btrfs_exit_compress(void)
1219 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1220 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1221 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1222 zstd_cleanup_workspace_manager();
1226 * Copy uncompressed data from working buffer to pages.
1228 * buf_start is the byte offset we're of the start of our workspace buffer.
1230 * total_out is the last byte of the buffer
1232 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1233 unsigned long total_out, u64 disk_start,
1236 unsigned long buf_offset;
1237 unsigned long current_buf_start;
1238 unsigned long start_byte;
1239 unsigned long prev_start_byte;
1240 unsigned long working_bytes = total_out - buf_start;
1241 unsigned long bytes;
1243 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1246 * start byte is the first byte of the page we're currently
1247 * copying into relative to the start of the compressed data.
1249 start_byte = page_offset(bvec.bv_page) - disk_start;
1251 /* we haven't yet hit data corresponding to this page */
1252 if (total_out <= start_byte)
1256 * the start of the data we care about is offset into
1257 * the middle of our working buffer
1259 if (total_out > start_byte && buf_start < start_byte) {
1260 buf_offset = start_byte - buf_start;
1261 working_bytes -= buf_offset;
1265 current_buf_start = buf_start;
1267 /* copy bytes from the working buffer into the pages */
1268 while (working_bytes > 0) {
1269 bytes = min_t(unsigned long, bvec.bv_len,
1270 PAGE_SIZE - (buf_offset % PAGE_SIZE));
1271 bytes = min(bytes, working_bytes);
1273 kaddr = kmap_atomic(bvec.bv_page);
1274 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1275 kunmap_atomic(kaddr);
1276 flush_dcache_page(bvec.bv_page);
1278 buf_offset += bytes;
1279 working_bytes -= bytes;
1280 current_buf_start += bytes;
1282 /* check if we need to pick another page */
1283 bio_advance(bio, bytes);
1284 if (!bio->bi_iter.bi_size)
1286 bvec = bio_iter_iovec(bio, bio->bi_iter);
1287 prev_start_byte = start_byte;
1288 start_byte = page_offset(bvec.bv_page) - disk_start;
1291 * We need to make sure we're only adjusting
1292 * our offset into compression working buffer when
1293 * we're switching pages. Otherwise we can incorrectly
1294 * keep copying when we were actually done.
1296 if (start_byte != prev_start_byte) {
1298 * make sure our new page is covered by this
1301 if (total_out <= start_byte)
1305 * the next page in the biovec might not be adjacent
1306 * to the last page, but it might still be found
1307 * inside this working buffer. bump our offset pointer
1309 if (total_out > start_byte &&
1310 current_buf_start < start_byte) {
1311 buf_offset = start_byte - buf_start;
1312 working_bytes = total_out - start_byte;
1313 current_buf_start = buf_start + buf_offset;
1322 * Shannon Entropy calculation
1324 * Pure byte distribution analysis fails to determine compressibility of data.
1325 * Try calculating entropy to estimate the average minimum number of bits
1326 * needed to encode the sampled data.
1328 * For convenience, return the percentage of needed bits, instead of amount of
1331 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1332 * and can be compressible with high probability
1334 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1336 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1338 #define ENTROPY_LVL_ACEPTABLE (65)
1339 #define ENTROPY_LVL_HIGH (80)
1342 * For increasead precision in shannon_entropy calculation,
1343 * let's do pow(n, M) to save more digits after comma:
1345 * - maximum int bit length is 64
1346 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1347 * - 13 * 4 = 52 < 64 -> M = 4
1351 static inline u32 ilog2_w(u64 n)
1353 return ilog2(n * n * n * n);
1356 static u32 shannon_entropy(struct heuristic_ws *ws)
1358 const u32 entropy_max = 8 * ilog2_w(2);
1359 u32 entropy_sum = 0;
1360 u32 p, p_base, sz_base;
1363 sz_base = ilog2_w(ws->sample_size);
1364 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1365 p = ws->bucket[i].count;
1366 p_base = ilog2_w(p);
1367 entropy_sum += p * (sz_base - p_base);
1370 entropy_sum /= ws->sample_size;
1371 return entropy_sum * 100 / entropy_max;
1374 #define RADIX_BASE 4U
1375 #define COUNTERS_SIZE (1U << RADIX_BASE)
1377 static u8 get4bits(u64 num, int shift) {
1382 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1387 * Use 4 bits as radix base
1388 * Use 16 u32 counters for calculating new position in buf array
1390 * @array - array that will be sorted
1391 * @array_buf - buffer array to store sorting results
1392 * must be equal in size to @array
1395 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1400 u32 counters[COUNTERS_SIZE];
1408 * Try avoid useless loop iterations for small numbers stored in big
1409 * counters. Example: 48 33 4 ... in 64bit array
1411 max_num = array[0].count;
1412 for (i = 1; i < num; i++) {
1413 buf_num = array[i].count;
1414 if (buf_num > max_num)
1418 buf_num = ilog2(max_num);
1419 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1422 while (shift < bitlen) {
1423 memset(counters, 0, sizeof(counters));
1425 for (i = 0; i < num; i++) {
1426 buf_num = array[i].count;
1427 addr = get4bits(buf_num, shift);
1431 for (i = 1; i < COUNTERS_SIZE; i++)
1432 counters[i] += counters[i - 1];
1434 for (i = num - 1; i >= 0; i--) {
1435 buf_num = array[i].count;
1436 addr = get4bits(buf_num, shift);
1438 new_addr = counters[addr];
1439 array_buf[new_addr] = array[i];
1442 shift += RADIX_BASE;
1445 * Normal radix expects to move data from a temporary array, to
1446 * the main one. But that requires some CPU time. Avoid that
1447 * by doing another sort iteration to original array instead of
1450 memset(counters, 0, sizeof(counters));
1452 for (i = 0; i < num; i ++) {
1453 buf_num = array_buf[i].count;
1454 addr = get4bits(buf_num, shift);
1458 for (i = 1; i < COUNTERS_SIZE; i++)
1459 counters[i] += counters[i - 1];
1461 for (i = num - 1; i >= 0; i--) {
1462 buf_num = array_buf[i].count;
1463 addr = get4bits(buf_num, shift);
1465 new_addr = counters[addr];
1466 array[new_addr] = array_buf[i];
1469 shift += RADIX_BASE;
1474 * Size of the core byte set - how many bytes cover 90% of the sample
1476 * There are several types of structured binary data that use nearly all byte
1477 * values. The distribution can be uniform and counts in all buckets will be
1478 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1480 * Other possibility is normal (Gaussian) distribution, where the data could
1481 * be potentially compressible, but we have to take a few more steps to decide
1484 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1485 * compression algo can easy fix that
1486 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1487 * probability is not compressible
1489 #define BYTE_CORE_SET_LOW (64)
1490 #define BYTE_CORE_SET_HIGH (200)
1492 static int byte_core_set_size(struct heuristic_ws *ws)
1495 u32 coreset_sum = 0;
1496 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1497 struct bucket_item *bucket = ws->bucket;
1499 /* Sort in reverse order */
1500 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1502 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1503 coreset_sum += bucket[i].count;
1505 if (coreset_sum > core_set_threshold)
1508 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1509 coreset_sum += bucket[i].count;
1510 if (coreset_sum > core_set_threshold)
1518 * Count byte values in buckets.
1519 * This heuristic can detect textual data (configs, xml, json, html, etc).
1520 * Because in most text-like data byte set is restricted to limited number of
1521 * possible characters, and that restriction in most cases makes data easy to
1524 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1525 * less - compressible
1526 * more - need additional analysis
1528 #define BYTE_SET_THRESHOLD (64)
1530 static u32 byte_set_size(const struct heuristic_ws *ws)
1533 u32 byte_set_size = 0;
1535 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1536 if (ws->bucket[i].count > 0)
1541 * Continue collecting count of byte values in buckets. If the byte
1542 * set size is bigger then the threshold, it's pointless to continue,
1543 * the detection technique would fail for this type of data.
1545 for (; i < BUCKET_SIZE; i++) {
1546 if (ws->bucket[i].count > 0) {
1548 if (byte_set_size > BYTE_SET_THRESHOLD)
1549 return byte_set_size;
1553 return byte_set_size;
1556 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1558 const u32 half_of_sample = ws->sample_size / 2;
1559 const u8 *data = ws->sample;
1561 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1564 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1565 struct heuristic_ws *ws)
1568 u64 index, index_end;
1569 u32 i, curr_sample_pos;
1573 * Compression handles the input data by chunks of 128KiB
1574 * (defined by BTRFS_MAX_UNCOMPRESSED)
1576 * We do the same for the heuristic and loop over the whole range.
1578 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1579 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1581 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1582 end = start + BTRFS_MAX_UNCOMPRESSED;
1584 index = start >> PAGE_SHIFT;
1585 index_end = end >> PAGE_SHIFT;
1587 /* Don't miss unaligned end */
1588 if (!IS_ALIGNED(end, PAGE_SIZE))
1591 curr_sample_pos = 0;
1592 while (index < index_end) {
1593 page = find_get_page(inode->i_mapping, index);
1594 in_data = kmap(page);
1595 /* Handle case where the start is not aligned to PAGE_SIZE */
1596 i = start % PAGE_SIZE;
1597 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1598 /* Don't sample any garbage from the last page */
1599 if (start > end - SAMPLING_READ_SIZE)
1601 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1602 SAMPLING_READ_SIZE);
1603 i += SAMPLING_INTERVAL;
1604 start += SAMPLING_INTERVAL;
1605 curr_sample_pos += SAMPLING_READ_SIZE;
1613 ws->sample_size = curr_sample_pos;
1617 * Compression heuristic.
1619 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1620 * quickly (compared to direct compression) detect data characteristics
1621 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1624 * The following types of analysis can be performed:
1625 * - detect mostly zero data
1626 * - detect data with low "byte set" size (text, etc)
1627 * - detect data with low/high "core byte" set
1629 * Return non-zero if the compression should be done, 0 otherwise.
1631 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1633 struct list_head *ws_list = get_workspace(0, 0);
1634 struct heuristic_ws *ws;
1639 ws = list_entry(ws_list, struct heuristic_ws, list);
1641 heuristic_collect_sample(inode, start, end, ws);
1643 if (sample_repeated_patterns(ws)) {
1648 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1650 for (i = 0; i < ws->sample_size; i++) {
1651 byte = ws->sample[i];
1652 ws->bucket[byte].count++;
1655 i = byte_set_size(ws);
1656 if (i < BYTE_SET_THRESHOLD) {
1661 i = byte_core_set_size(ws);
1662 if (i <= BYTE_CORE_SET_LOW) {
1667 if (i >= BYTE_CORE_SET_HIGH) {
1672 i = shannon_entropy(ws);
1673 if (i <= ENTROPY_LVL_ACEPTABLE) {
1679 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1680 * needed to give green light to compression.
1682 * For now just assume that compression at that level is not worth the
1683 * resources because:
1685 * 1. it is possible to defrag the data later
1687 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1688 * values, every bucket has counter at level ~54. The heuristic would
1689 * be confused. This can happen when data have some internal repeated
1690 * patterns like "abbacbbc...". This can be detected by analyzing
1691 * pairs of bytes, which is too costly.
1693 if (i < ENTROPY_LVL_HIGH) {
1702 put_workspace(0, ws_list);
1707 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1708 * level, unrecognized string will set the default level
1710 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1712 unsigned int level = 0;
1718 if (str[0] == ':') {
1719 ret = kstrtouint(str + 1, 10, &level);
1724 level = btrfs_compress_set_level(type, level);