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 happen when compression races with remount setting
84 * it to 'no compress', while caller doesn't call
85 * inode_need_compress() to check if we really need to
88 * Not a big deal, just need to inform caller that we
89 * haven't allocated any pages yet.
96 static int compression_decompress_bio(int type
, struct list_head
*ws
,
97 struct compressed_bio
*cb
)
100 case BTRFS_COMPRESS_ZLIB
: return zlib_decompress_bio(ws
, cb
);
101 case BTRFS_COMPRESS_LZO
: return lzo_decompress_bio(ws
, cb
);
102 case BTRFS_COMPRESS_ZSTD
: return zstd_decompress_bio(ws
, cb
);
103 case BTRFS_COMPRESS_NONE
:
106 * This can't happen, the type is validated several times
107 * before we get here.
113 static int compression_decompress(int type
, struct list_head
*ws
,
114 unsigned char *data_in
, struct page
*dest_page
,
115 unsigned long start_byte
, size_t srclen
, size_t destlen
)
118 case BTRFS_COMPRESS_ZLIB
: return zlib_decompress(ws
, data_in
, dest_page
,
119 start_byte
, srclen
, destlen
);
120 case BTRFS_COMPRESS_LZO
: return lzo_decompress(ws
, data_in
, dest_page
,
121 start_byte
, srclen
, destlen
);
122 case BTRFS_COMPRESS_ZSTD
: return zstd_decompress(ws
, data_in
, dest_page
,
123 start_byte
, srclen
, destlen
);
124 case BTRFS_COMPRESS_NONE
:
127 * This can't happen, the type is validated several times
128 * before we get here.
134 static int btrfs_decompress_bio(struct compressed_bio
*cb
);
136 static inline int compressed_bio_size(struct btrfs_fs_info
*fs_info
,
137 unsigned long disk_size
)
139 return sizeof(struct compressed_bio
) +
140 (DIV_ROUND_UP(disk_size
, fs_info
->sectorsize
)) * fs_info
->csum_size
;
143 static int check_compressed_csum(struct btrfs_inode
*inode
, struct bio
*bio
,
146 struct btrfs_fs_info
*fs_info
= inode
->root
->fs_info
;
147 SHASH_DESC_ON_STACK(shash
, fs_info
->csum_shash
);
148 const u32 csum_size
= fs_info
->csum_size
;
152 u8 csum
[BTRFS_CSUM_SIZE
];
153 struct compressed_bio
*cb
= bio
->bi_private
;
154 u8
*cb_sum
= cb
->sums
;
156 if (!fs_info
->csum_root
|| (inode
->flags
& BTRFS_INODE_NODATASUM
))
159 shash
->tfm
= fs_info
->csum_shash
;
161 for (i
= 0; i
< cb
->nr_pages
; i
++) {
162 page
= cb
->compressed_pages
[i
];
164 kaddr
= kmap_atomic(page
);
165 crypto_shash_digest(shash
, kaddr
, PAGE_SIZE
, csum
);
166 kunmap_atomic(kaddr
);
168 if (memcmp(&csum
, cb_sum
, csum_size
)) {
169 btrfs_print_data_csum_error(inode
, disk_start
,
170 csum
, cb_sum
, cb
->mirror_num
);
171 if (btrfs_io_bio(bio
)->device
)
172 btrfs_dev_stat_inc_and_print(
173 btrfs_io_bio(bio
)->device
,
174 BTRFS_DEV_STAT_CORRUPTION_ERRS
);
182 /* when we finish reading compressed pages from the disk, we
183 * decompress them and then run the bio end_io routines on the
184 * decompressed pages (in the inode address space).
186 * This allows the checksumming and other IO error handling routines
189 * The compressed pages are freed here, and it must be run
192 static void end_compressed_bio_read(struct bio
*bio
)
194 struct compressed_bio
*cb
= bio
->bi_private
;
198 unsigned int mirror
= btrfs_io_bio(bio
)->mirror_num
;
204 /* if there are more bios still pending for this compressed
207 if (!refcount_dec_and_test(&cb
->pending_bios
))
211 * Record the correct mirror_num in cb->orig_bio so that
212 * read-repair can work properly.
214 btrfs_io_bio(cb
->orig_bio
)->mirror_num
= mirror
;
215 cb
->mirror_num
= mirror
;
218 * Some IO in this cb have failed, just skip checksum as there
219 * is no way it could be correct.
225 ret
= check_compressed_csum(BTRFS_I(inode
), bio
,
226 bio
->bi_iter
.bi_sector
<< 9);
230 /* ok, we're the last bio for this extent, lets start
233 ret
= btrfs_decompress_bio(cb
);
239 /* release the compressed pages */
241 for (index
= 0; index
< cb
->nr_pages
; index
++) {
242 page
= cb
->compressed_pages
[index
];
243 page
->mapping
= NULL
;
247 /* do io completion on the original bio */
249 bio_io_error(cb
->orig_bio
);
251 struct bio_vec
*bvec
;
252 struct bvec_iter_all iter_all
;
255 * we have verified the checksum already, set page
256 * checked so the end_io handlers know about it
258 ASSERT(!bio_flagged(bio
, BIO_CLONED
));
259 bio_for_each_segment_all(bvec
, cb
->orig_bio
, iter_all
)
260 SetPageChecked(bvec
->bv_page
);
262 bio_endio(cb
->orig_bio
);
265 /* finally free the cb struct */
266 kfree(cb
->compressed_pages
);
273 * Clear the writeback bits on all of the file
274 * pages for a compressed write
276 static noinline
void end_compressed_writeback(struct inode
*inode
,
277 const struct compressed_bio
*cb
)
279 unsigned long index
= cb
->start
>> PAGE_SHIFT
;
280 unsigned long end_index
= (cb
->start
+ cb
->len
- 1) >> PAGE_SHIFT
;
281 struct page
*pages
[16];
282 unsigned long nr_pages
= end_index
- index
+ 1;
287 mapping_set_error(inode
->i_mapping
, -EIO
);
289 while (nr_pages
> 0) {
290 ret
= find_get_pages_contig(inode
->i_mapping
, index
,
292 nr_pages
, ARRAY_SIZE(pages
)), pages
);
298 for (i
= 0; i
< ret
; i
++) {
300 SetPageError(pages
[i
]);
301 end_page_writeback(pages
[i
]);
307 /* the inode may be gone now */
311 * do the cleanup once all the compressed pages hit the disk.
312 * This will clear writeback on the file pages and free the compressed
315 * This also calls the writeback end hooks for the file pages so that
316 * metadata and checksums can be updated in the file.
318 static void end_compressed_bio_write(struct bio
*bio
)
320 struct compressed_bio
*cb
= bio
->bi_private
;
328 /* if there are more bios still pending for this compressed
331 if (!refcount_dec_and_test(&cb
->pending_bios
))
334 /* ok, we're the last bio for this extent, step one is to
335 * call back into the FS and do all the end_io operations
338 cb
->compressed_pages
[0]->mapping
= cb
->inode
->i_mapping
;
339 btrfs_writepage_endio_finish_ordered(cb
->compressed_pages
[0],
340 cb
->start
, cb
->start
+ cb
->len
- 1,
341 bio
->bi_status
== BLK_STS_OK
);
342 cb
->compressed_pages
[0]->mapping
= NULL
;
344 end_compressed_writeback(inode
, cb
);
345 /* note, our inode could be gone now */
348 * release the compressed pages, these came from alloc_page and
349 * are not attached to the inode at all
352 for (index
= 0; index
< cb
->nr_pages
; index
++) {
353 page
= cb
->compressed_pages
[index
];
354 page
->mapping
= NULL
;
358 /* finally free the cb struct */
359 kfree(cb
->compressed_pages
);
366 * worker function to build and submit bios for previously compressed pages.
367 * The corresponding pages in the inode should be marked for writeback
368 * and the compressed pages should have a reference on them for dropping
369 * when the IO is complete.
371 * This also checksums the file bytes and gets things ready for
374 blk_status_t
btrfs_submit_compressed_write(struct btrfs_inode
*inode
, u64 start
,
375 unsigned long len
, u64 disk_start
,
376 unsigned long compressed_len
,
377 struct page
**compressed_pages
,
378 unsigned long nr_pages
,
379 unsigned int write_flags
,
380 struct cgroup_subsys_state
*blkcg_css
)
382 struct btrfs_fs_info
*fs_info
= inode
->root
->fs_info
;
383 struct bio
*bio
= NULL
;
384 struct compressed_bio
*cb
;
385 unsigned long bytes_left
;
388 u64 first_byte
= disk_start
;
390 int skip_sum
= inode
->flags
& BTRFS_INODE_NODATASUM
;
392 WARN_ON(!PAGE_ALIGNED(start
));
393 cb
= kmalloc(compressed_bio_size(fs_info
, compressed_len
), GFP_NOFS
);
395 return BLK_STS_RESOURCE
;
396 refcount_set(&cb
->pending_bios
, 0);
398 cb
->inode
= &inode
->vfs_inode
;
402 cb
->compressed_pages
= compressed_pages
;
403 cb
->compressed_len
= compressed_len
;
405 cb
->nr_pages
= nr_pages
;
407 bio
= btrfs_bio_alloc(first_byte
);
408 bio
->bi_opf
= REQ_OP_WRITE
| write_flags
;
409 bio
->bi_private
= cb
;
410 bio
->bi_end_io
= end_compressed_bio_write
;
413 bio
->bi_opf
|= REQ_CGROUP_PUNT
;
414 kthread_associate_blkcg(blkcg_css
);
416 refcount_set(&cb
->pending_bios
, 1);
418 /* create and submit bios for the compressed pages */
419 bytes_left
= compressed_len
;
420 for (pg_index
= 0; pg_index
< cb
->nr_pages
; pg_index
++) {
423 page
= compressed_pages
[pg_index
];
424 page
->mapping
= inode
->vfs_inode
.i_mapping
;
425 if (bio
->bi_iter
.bi_size
)
426 submit
= btrfs_bio_fits_in_stripe(page
, PAGE_SIZE
, bio
,
429 page
->mapping
= NULL
;
430 if (submit
|| bio_add_page(bio
, page
, PAGE_SIZE
, 0) <
433 * inc the count before we submit the bio so
434 * we know the end IO handler won't happen before
435 * we inc the count. Otherwise, the cb might get
436 * freed before we're done setting it up
438 refcount_inc(&cb
->pending_bios
);
439 ret
= btrfs_bio_wq_end_io(fs_info
, bio
,
440 BTRFS_WQ_ENDIO_DATA
);
441 BUG_ON(ret
); /* -ENOMEM */
444 ret
= btrfs_csum_one_bio(inode
, bio
, start
, 1);
445 BUG_ON(ret
); /* -ENOMEM */
448 ret
= btrfs_map_bio(fs_info
, bio
, 0);
450 bio
->bi_status
= ret
;
454 bio
= btrfs_bio_alloc(first_byte
);
455 bio
->bi_opf
= REQ_OP_WRITE
| write_flags
;
456 bio
->bi_private
= cb
;
457 bio
->bi_end_io
= end_compressed_bio_write
;
459 bio
->bi_opf
|= REQ_CGROUP_PUNT
;
460 bio_add_page(bio
, page
, PAGE_SIZE
, 0);
462 if (bytes_left
< PAGE_SIZE
) {
464 "bytes left %lu compress len %lu nr %lu",
465 bytes_left
, cb
->compressed_len
, cb
->nr_pages
);
467 bytes_left
-= PAGE_SIZE
;
468 first_byte
+= PAGE_SIZE
;
472 ret
= btrfs_bio_wq_end_io(fs_info
, bio
, BTRFS_WQ_ENDIO_DATA
);
473 BUG_ON(ret
); /* -ENOMEM */
476 ret
= btrfs_csum_one_bio(inode
, bio
, start
, 1);
477 BUG_ON(ret
); /* -ENOMEM */
480 ret
= btrfs_map_bio(fs_info
, bio
, 0);
482 bio
->bi_status
= ret
;
487 kthread_associate_blkcg(NULL
);
492 static u64
bio_end_offset(struct bio
*bio
)
494 struct bio_vec
*last
= bio_last_bvec_all(bio
);
496 return page_offset(last
->bv_page
) + last
->bv_len
+ last
->bv_offset
;
499 static noinline
int add_ra_bio_pages(struct inode
*inode
,
501 struct compressed_bio
*cb
)
503 unsigned long end_index
;
504 unsigned long pg_index
;
506 u64 isize
= i_size_read(inode
);
509 unsigned long nr_pages
= 0;
510 struct extent_map
*em
;
511 struct address_space
*mapping
= inode
->i_mapping
;
512 struct extent_map_tree
*em_tree
;
513 struct extent_io_tree
*tree
;
517 last_offset
= bio_end_offset(cb
->orig_bio
);
518 em_tree
= &BTRFS_I(inode
)->extent_tree
;
519 tree
= &BTRFS_I(inode
)->io_tree
;
524 end_index
= (i_size_read(inode
) - 1) >> PAGE_SHIFT
;
526 while (last_offset
< compressed_end
) {
527 pg_index
= last_offset
>> PAGE_SHIFT
;
529 if (pg_index
> end_index
)
532 page
= xa_load(&mapping
->i_pages
, pg_index
);
533 if (page
&& !xa_is_value(page
)) {
540 page
= __page_cache_alloc(mapping_gfp_constraint(mapping
,
545 if (add_to_page_cache_lru(page
, mapping
, pg_index
, GFP_NOFS
)) {
550 end
= last_offset
+ PAGE_SIZE
- 1;
552 * at this point, we have a locked page in the page cache
553 * for these bytes in the file. But, we have to make
554 * sure they map to this compressed extent on disk.
556 set_page_extent_mapped(page
);
557 lock_extent(tree
, last_offset
, end
);
558 read_lock(&em_tree
->lock
);
559 em
= lookup_extent_mapping(em_tree
, last_offset
,
561 read_unlock(&em_tree
->lock
);
563 if (!em
|| last_offset
< em
->start
||
564 (last_offset
+ PAGE_SIZE
> extent_map_end(em
)) ||
565 (em
->block_start
>> 9) != cb
->orig_bio
->bi_iter
.bi_sector
) {
567 unlock_extent(tree
, last_offset
, end
);
574 if (page
->index
== end_index
) {
576 size_t zero_offset
= offset_in_page(isize
);
580 zeros
= PAGE_SIZE
- zero_offset
;
581 userpage
= kmap_atomic(page
);
582 memset(userpage
+ zero_offset
, 0, zeros
);
583 flush_dcache_page(page
);
584 kunmap_atomic(userpage
);
588 ret
= bio_add_page(cb
->orig_bio
, page
,
591 if (ret
== PAGE_SIZE
) {
595 unlock_extent(tree
, last_offset
, end
);
601 last_offset
+= PAGE_SIZE
;
607 * for a compressed read, the bio we get passed has all the inode pages
608 * in it. We don't actually do IO on those pages but allocate new ones
609 * to hold the compressed pages on disk.
611 * bio->bi_iter.bi_sector points to the compressed extent on disk
612 * bio->bi_io_vec points to all of the inode pages
614 * After the compressed pages are read, we copy the bytes into the
615 * bio we were passed and then call the bio end_io calls
617 blk_status_t
btrfs_submit_compressed_read(struct inode
*inode
, struct bio
*bio
,
618 int mirror_num
, unsigned long bio_flags
)
620 struct btrfs_fs_info
*fs_info
= btrfs_sb(inode
->i_sb
);
621 struct extent_map_tree
*em_tree
;
622 struct compressed_bio
*cb
;
623 unsigned long compressed_len
;
624 unsigned long nr_pages
;
625 unsigned long pg_index
;
627 struct bio
*comp_bio
;
628 u64 cur_disk_byte
= bio
->bi_iter
.bi_sector
<< 9;
631 struct extent_map
*em
;
632 blk_status_t ret
= BLK_STS_RESOURCE
;
636 em_tree
= &BTRFS_I(inode
)->extent_tree
;
638 /* we need the actual starting offset of this extent in the file */
639 read_lock(&em_tree
->lock
);
640 em
= lookup_extent_mapping(em_tree
,
641 page_offset(bio_first_page_all(bio
)),
643 read_unlock(&em_tree
->lock
);
645 return BLK_STS_IOERR
;
647 compressed_len
= em
->block_len
;
648 cb
= kmalloc(compressed_bio_size(fs_info
, compressed_len
), GFP_NOFS
);
652 refcount_set(&cb
->pending_bios
, 0);
655 cb
->mirror_num
= mirror_num
;
658 cb
->start
= em
->orig_start
;
660 em_start
= em
->start
;
665 cb
->len
= bio
->bi_iter
.bi_size
;
666 cb
->compressed_len
= compressed_len
;
667 cb
->compress_type
= extent_compress_type(bio_flags
);
670 nr_pages
= DIV_ROUND_UP(compressed_len
, PAGE_SIZE
);
671 cb
->compressed_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
673 if (!cb
->compressed_pages
)
676 for (pg_index
= 0; pg_index
< nr_pages
; pg_index
++) {
677 cb
->compressed_pages
[pg_index
] = alloc_page(GFP_NOFS
|
679 if (!cb
->compressed_pages
[pg_index
]) {
680 faili
= pg_index
- 1;
681 ret
= BLK_STS_RESOURCE
;
685 faili
= nr_pages
- 1;
686 cb
->nr_pages
= nr_pages
;
688 add_ra_bio_pages(inode
, em_start
+ em_len
, cb
);
690 /* include any pages we added in add_ra-bio_pages */
691 cb
->len
= bio
->bi_iter
.bi_size
;
693 comp_bio
= btrfs_bio_alloc(cur_disk_byte
);
694 comp_bio
->bi_opf
= REQ_OP_READ
;
695 comp_bio
->bi_private
= cb
;
696 comp_bio
->bi_end_io
= end_compressed_bio_read
;
697 refcount_set(&cb
->pending_bios
, 1);
699 for (pg_index
= 0; pg_index
< nr_pages
; pg_index
++) {
702 page
= cb
->compressed_pages
[pg_index
];
703 page
->mapping
= inode
->i_mapping
;
704 page
->index
= em_start
>> PAGE_SHIFT
;
706 if (comp_bio
->bi_iter
.bi_size
)
707 submit
= btrfs_bio_fits_in_stripe(page
, PAGE_SIZE
,
710 page
->mapping
= NULL
;
711 if (submit
|| bio_add_page(comp_bio
, page
, PAGE_SIZE
, 0) <
713 unsigned int nr_sectors
;
715 ret
= btrfs_bio_wq_end_io(fs_info
, comp_bio
,
716 BTRFS_WQ_ENDIO_DATA
);
717 BUG_ON(ret
); /* -ENOMEM */
720 * inc the count before we submit the bio so
721 * we know the end IO handler won't happen before
722 * we inc the count. Otherwise, the cb might get
723 * freed before we're done setting it up
725 refcount_inc(&cb
->pending_bios
);
727 ret
= btrfs_lookup_bio_sums(inode
, comp_bio
, sums
);
728 BUG_ON(ret
); /* -ENOMEM */
730 nr_sectors
= DIV_ROUND_UP(comp_bio
->bi_iter
.bi_size
,
731 fs_info
->sectorsize
);
732 sums
+= fs_info
->csum_size
* nr_sectors
;
734 ret
= btrfs_map_bio(fs_info
, comp_bio
, mirror_num
);
736 comp_bio
->bi_status
= ret
;
740 comp_bio
= btrfs_bio_alloc(cur_disk_byte
);
741 comp_bio
->bi_opf
= REQ_OP_READ
;
742 comp_bio
->bi_private
= cb
;
743 comp_bio
->bi_end_io
= end_compressed_bio_read
;
745 bio_add_page(comp_bio
, page
, PAGE_SIZE
, 0);
747 cur_disk_byte
+= PAGE_SIZE
;
750 ret
= btrfs_bio_wq_end_io(fs_info
, comp_bio
, BTRFS_WQ_ENDIO_DATA
);
751 BUG_ON(ret
); /* -ENOMEM */
753 ret
= btrfs_lookup_bio_sums(inode
, comp_bio
, sums
);
754 BUG_ON(ret
); /* -ENOMEM */
756 ret
= btrfs_map_bio(fs_info
, comp_bio
, mirror_num
);
758 comp_bio
->bi_status
= ret
;
766 __free_page(cb
->compressed_pages
[faili
]);
770 kfree(cb
->compressed_pages
);
779 * Heuristic uses systematic sampling to collect data from the input data
780 * range, the logic can be tuned by the following constants:
782 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
783 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
785 #define SAMPLING_READ_SIZE (16)
786 #define SAMPLING_INTERVAL (256)
789 * For statistical analysis of the input data we consider bytes that form a
790 * Galois Field of 256 objects. Each object has an attribute count, ie. how
791 * many times the object appeared in the sample.
793 #define BUCKET_SIZE (256)
796 * The size of the sample is based on a statistical sampling rule of thumb.
797 * The common way is to perform sampling tests as long as the number of
798 * elements in each cell is at least 5.
800 * Instead of 5, we choose 32 to obtain more accurate results.
801 * If the data contain the maximum number of symbols, which is 256, we obtain a
802 * sample size bound by 8192.
804 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
805 * from up to 512 locations.
807 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
808 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
814 struct heuristic_ws
{
815 /* Partial copy of input data */
818 /* Buckets store counters for each byte value */
819 struct bucket_item
*bucket
;
821 struct bucket_item
*bucket_b
;
822 struct list_head list
;
825 static struct workspace_manager heuristic_wsm
;
827 static void free_heuristic_ws(struct list_head
*ws
)
829 struct heuristic_ws
*workspace
;
831 workspace
= list_entry(ws
, struct heuristic_ws
, list
);
833 kvfree(workspace
->sample
);
834 kfree(workspace
->bucket
);
835 kfree(workspace
->bucket_b
);
839 static struct list_head
*alloc_heuristic_ws(unsigned int level
)
841 struct heuristic_ws
*ws
;
843 ws
= kzalloc(sizeof(*ws
), GFP_KERNEL
);
845 return ERR_PTR(-ENOMEM
);
847 ws
->sample
= kvmalloc(MAX_SAMPLE_SIZE
, GFP_KERNEL
);
851 ws
->bucket
= kcalloc(BUCKET_SIZE
, sizeof(*ws
->bucket
), GFP_KERNEL
);
855 ws
->bucket_b
= kcalloc(BUCKET_SIZE
, sizeof(*ws
->bucket_b
), GFP_KERNEL
);
859 INIT_LIST_HEAD(&ws
->list
);
862 free_heuristic_ws(&ws
->list
);
863 return ERR_PTR(-ENOMEM
);
866 const struct btrfs_compress_op btrfs_heuristic_compress
= {
867 .workspace_manager
= &heuristic_wsm
,
870 static const struct btrfs_compress_op
* const btrfs_compress_op
[] = {
871 /* The heuristic is represented as compression type 0 */
872 &btrfs_heuristic_compress
,
873 &btrfs_zlib_compress
,
875 &btrfs_zstd_compress
,
878 static struct list_head
*alloc_workspace(int type
, unsigned int level
)
881 case BTRFS_COMPRESS_NONE
: return alloc_heuristic_ws(level
);
882 case BTRFS_COMPRESS_ZLIB
: return zlib_alloc_workspace(level
);
883 case BTRFS_COMPRESS_LZO
: return lzo_alloc_workspace(level
);
884 case BTRFS_COMPRESS_ZSTD
: return zstd_alloc_workspace(level
);
887 * This can't happen, the type is validated several times
888 * before we get here.
894 static void free_workspace(int type
, struct list_head
*ws
)
897 case BTRFS_COMPRESS_NONE
: return free_heuristic_ws(ws
);
898 case BTRFS_COMPRESS_ZLIB
: return zlib_free_workspace(ws
);
899 case BTRFS_COMPRESS_LZO
: return lzo_free_workspace(ws
);
900 case BTRFS_COMPRESS_ZSTD
: return zstd_free_workspace(ws
);
903 * This can't happen, the type is validated several times
904 * before we get here.
910 static void btrfs_init_workspace_manager(int type
)
912 struct workspace_manager
*wsm
;
913 struct list_head
*workspace
;
915 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
916 INIT_LIST_HEAD(&wsm
->idle_ws
);
917 spin_lock_init(&wsm
->ws_lock
);
918 atomic_set(&wsm
->total_ws
, 0);
919 init_waitqueue_head(&wsm
->ws_wait
);
922 * Preallocate one workspace for each compression type so we can
923 * guarantee forward progress in the worst case
925 workspace
= alloc_workspace(type
, 0);
926 if (IS_ERR(workspace
)) {
928 "BTRFS: cannot preallocate compression workspace, will try later\n");
930 atomic_set(&wsm
->total_ws
, 1);
932 list_add(workspace
, &wsm
->idle_ws
);
936 static void btrfs_cleanup_workspace_manager(int type
)
938 struct workspace_manager
*wsman
;
939 struct list_head
*ws
;
941 wsman
= btrfs_compress_op
[type
]->workspace_manager
;
942 while (!list_empty(&wsman
->idle_ws
)) {
943 ws
= wsman
->idle_ws
.next
;
945 free_workspace(type
, ws
);
946 atomic_dec(&wsman
->total_ws
);
951 * This finds an available workspace or allocates a new one.
952 * If it's not possible to allocate a new one, waits until there's one.
953 * Preallocation makes a forward progress guarantees and we do not return
956 struct list_head
*btrfs_get_workspace(int type
, unsigned int level
)
958 struct workspace_manager
*wsm
;
959 struct list_head
*workspace
;
960 int cpus
= num_online_cpus();
962 struct list_head
*idle_ws
;
965 wait_queue_head_t
*ws_wait
;
968 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
969 idle_ws
= &wsm
->idle_ws
;
970 ws_lock
= &wsm
->ws_lock
;
971 total_ws
= &wsm
->total_ws
;
972 ws_wait
= &wsm
->ws_wait
;
973 free_ws
= &wsm
->free_ws
;
977 if (!list_empty(idle_ws
)) {
978 workspace
= idle_ws
->next
;
981 spin_unlock(ws_lock
);
985 if (atomic_read(total_ws
) > cpus
) {
988 spin_unlock(ws_lock
);
989 prepare_to_wait(ws_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
990 if (atomic_read(total_ws
) > cpus
&& !*free_ws
)
992 finish_wait(ws_wait
, &wait
);
995 atomic_inc(total_ws
);
996 spin_unlock(ws_lock
);
999 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1000 * to turn it off here because we might get called from the restricted
1001 * context of btrfs_compress_bio/btrfs_compress_pages
1003 nofs_flag
= memalloc_nofs_save();
1004 workspace
= alloc_workspace(type
, level
);
1005 memalloc_nofs_restore(nofs_flag
);
1007 if (IS_ERR(workspace
)) {
1008 atomic_dec(total_ws
);
1012 * Do not return the error but go back to waiting. There's a
1013 * workspace preallocated for each type and the compression
1014 * time is bounded so we get to a workspace eventually. This
1015 * makes our caller's life easier.
1017 * To prevent silent and low-probability deadlocks (when the
1018 * initial preallocation fails), check if there are any
1019 * workspaces at all.
1021 if (atomic_read(total_ws
) == 0) {
1022 static DEFINE_RATELIMIT_STATE(_rs
,
1023 /* once per minute */ 60 * HZ
,
1026 if (__ratelimit(&_rs
)) {
1027 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1035 static struct list_head
*get_workspace(int type
, int level
)
1038 case BTRFS_COMPRESS_NONE
: return btrfs_get_workspace(type
, level
);
1039 case BTRFS_COMPRESS_ZLIB
: return zlib_get_workspace(level
);
1040 case BTRFS_COMPRESS_LZO
: return btrfs_get_workspace(type
, level
);
1041 case BTRFS_COMPRESS_ZSTD
: return zstd_get_workspace(level
);
1044 * This can't happen, the type is validated several times
1045 * before we get here.
1052 * put a workspace struct back on the list or free it if we have enough
1053 * idle ones sitting around
1055 void btrfs_put_workspace(int type
, struct list_head
*ws
)
1057 struct workspace_manager
*wsm
;
1058 struct list_head
*idle_ws
;
1059 spinlock_t
*ws_lock
;
1061 wait_queue_head_t
*ws_wait
;
1064 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
1065 idle_ws
= &wsm
->idle_ws
;
1066 ws_lock
= &wsm
->ws_lock
;
1067 total_ws
= &wsm
->total_ws
;
1068 ws_wait
= &wsm
->ws_wait
;
1069 free_ws
= &wsm
->free_ws
;
1072 if (*free_ws
<= num_online_cpus()) {
1073 list_add(ws
, idle_ws
);
1075 spin_unlock(ws_lock
);
1078 spin_unlock(ws_lock
);
1080 free_workspace(type
, ws
);
1081 atomic_dec(total_ws
);
1083 cond_wake_up(ws_wait
);
1086 static void put_workspace(int type
, struct list_head
*ws
)
1089 case BTRFS_COMPRESS_NONE
: return btrfs_put_workspace(type
, ws
);
1090 case BTRFS_COMPRESS_ZLIB
: return btrfs_put_workspace(type
, ws
);
1091 case BTRFS_COMPRESS_LZO
: return btrfs_put_workspace(type
, ws
);
1092 case BTRFS_COMPRESS_ZSTD
: return zstd_put_workspace(ws
);
1095 * This can't happen, the type is validated several times
1096 * before we get here.
1103 * Adjust @level according to the limits of the compression algorithm or
1104 * fallback to default
1106 static unsigned int btrfs_compress_set_level(int type
, unsigned level
)
1108 const struct btrfs_compress_op
*ops
= btrfs_compress_op
[type
];
1111 level
= ops
->default_level
;
1113 level
= min(level
, ops
->max_level
);
1119 * Given an address space and start and length, compress the bytes into @pages
1120 * that are allocated on demand.
1122 * @type_level is encoded algorithm and level, where level 0 means whatever
1123 * default the algorithm chooses and is opaque here;
1124 * - compression algo are 0-3
1125 * - the level are bits 4-7
1127 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1128 * and returns number of actually allocated pages
1130 * @total_in is used to return the number of bytes actually read. It
1131 * may be smaller than the input length if we had to exit early because we
1132 * ran out of room in the pages array or because we cross the
1133 * max_out threshold.
1135 * @total_out is an in/out parameter, must be set to the input length and will
1136 * be also used to return the total number of compressed bytes
1138 * @max_out tells us the max number of bytes that we're allowed to
1141 int btrfs_compress_pages(unsigned int type_level
, struct address_space
*mapping
,
1142 u64 start
, struct page
**pages
,
1143 unsigned long *out_pages
,
1144 unsigned long *total_in
,
1145 unsigned long *total_out
)
1147 int type
= btrfs_compress_type(type_level
);
1148 int level
= btrfs_compress_level(type_level
);
1149 struct list_head
*workspace
;
1152 level
= btrfs_compress_set_level(type
, level
);
1153 workspace
= get_workspace(type
, level
);
1154 ret
= compression_compress_pages(type
, workspace
, mapping
, start
, pages
,
1155 out_pages
, total_in
, total_out
);
1156 put_workspace(type
, workspace
);
1161 * pages_in is an array of pages with compressed data.
1163 * disk_start is the starting logical offset of this array in the file
1165 * orig_bio contains the pages from the file that we want to decompress into
1167 * srclen is the number of bytes in pages_in
1169 * The basic idea is that we have a bio that was created by readpages.
1170 * The pages in the bio are for the uncompressed data, and they may not
1171 * be contiguous. They all correspond to the range of bytes covered by
1172 * the compressed extent.
1174 static int btrfs_decompress_bio(struct compressed_bio
*cb
)
1176 struct list_head
*workspace
;
1178 int type
= cb
->compress_type
;
1180 workspace
= get_workspace(type
, 0);
1181 ret
= compression_decompress_bio(type
, workspace
, cb
);
1182 put_workspace(type
, workspace
);
1188 * a less complex decompression routine. Our compressed data fits in a
1189 * single page, and we want to read a single page out of it.
1190 * start_byte tells us the offset into the compressed data we're interested in
1192 int btrfs_decompress(int type
, unsigned char *data_in
, struct page
*dest_page
,
1193 unsigned long start_byte
, size_t srclen
, size_t destlen
)
1195 struct list_head
*workspace
;
1198 workspace
= get_workspace(type
, 0);
1199 ret
= compression_decompress(type
, workspace
, data_in
, dest_page
,
1200 start_byte
, srclen
, destlen
);
1201 put_workspace(type
, workspace
);
1206 void __init
btrfs_init_compress(void)
1208 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE
);
1209 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB
);
1210 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO
);
1211 zstd_init_workspace_manager();
1214 void __cold
btrfs_exit_compress(void)
1216 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE
);
1217 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB
);
1218 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO
);
1219 zstd_cleanup_workspace_manager();
1223 * Copy uncompressed data from working buffer to pages.
1225 * buf_start is the byte offset we're of the start of our workspace buffer.
1227 * total_out is the last byte of the buffer
1229 int btrfs_decompress_buf2page(const char *buf
, unsigned long buf_start
,
1230 unsigned long total_out
, u64 disk_start
,
1233 unsigned long buf_offset
;
1234 unsigned long current_buf_start
;
1235 unsigned long start_byte
;
1236 unsigned long prev_start_byte
;
1237 unsigned long working_bytes
= total_out
- buf_start
;
1238 unsigned long bytes
;
1240 struct bio_vec bvec
= bio_iter_iovec(bio
, bio
->bi_iter
);
1243 * start byte is the first byte of the page we're currently
1244 * copying into relative to the start of the compressed data.
1246 start_byte
= page_offset(bvec
.bv_page
) - disk_start
;
1248 /* we haven't yet hit data corresponding to this page */
1249 if (total_out
<= start_byte
)
1253 * the start of the data we care about is offset into
1254 * the middle of our working buffer
1256 if (total_out
> start_byte
&& buf_start
< start_byte
) {
1257 buf_offset
= start_byte
- buf_start
;
1258 working_bytes
-= buf_offset
;
1262 current_buf_start
= buf_start
;
1264 /* copy bytes from the working buffer into the pages */
1265 while (working_bytes
> 0) {
1266 bytes
= min_t(unsigned long, bvec
.bv_len
,
1267 PAGE_SIZE
- (buf_offset
% PAGE_SIZE
));
1268 bytes
= min(bytes
, working_bytes
);
1270 kaddr
= kmap_atomic(bvec
.bv_page
);
1271 memcpy(kaddr
+ bvec
.bv_offset
, buf
+ buf_offset
, bytes
);
1272 kunmap_atomic(kaddr
);
1273 flush_dcache_page(bvec
.bv_page
);
1275 buf_offset
+= bytes
;
1276 working_bytes
-= bytes
;
1277 current_buf_start
+= bytes
;
1279 /* check if we need to pick another page */
1280 bio_advance(bio
, bytes
);
1281 if (!bio
->bi_iter
.bi_size
)
1283 bvec
= bio_iter_iovec(bio
, bio
->bi_iter
);
1284 prev_start_byte
= start_byte
;
1285 start_byte
= page_offset(bvec
.bv_page
) - disk_start
;
1288 * We need to make sure we're only adjusting
1289 * our offset into compression working buffer when
1290 * we're switching pages. Otherwise we can incorrectly
1291 * keep copying when we were actually done.
1293 if (start_byte
!= prev_start_byte
) {
1295 * make sure our new page is covered by this
1298 if (total_out
<= start_byte
)
1302 * the next page in the biovec might not be adjacent
1303 * to the last page, but it might still be found
1304 * inside this working buffer. bump our offset pointer
1306 if (total_out
> start_byte
&&
1307 current_buf_start
< start_byte
) {
1308 buf_offset
= start_byte
- buf_start
;
1309 working_bytes
= total_out
- start_byte
;
1310 current_buf_start
= buf_start
+ buf_offset
;
1319 * Shannon Entropy calculation
1321 * Pure byte distribution analysis fails to determine compressibility of data.
1322 * Try calculating entropy to estimate the average minimum number of bits
1323 * needed to encode the sampled data.
1325 * For convenience, return the percentage of needed bits, instead of amount of
1328 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1329 * and can be compressible with high probability
1331 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1333 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1335 #define ENTROPY_LVL_ACEPTABLE (65)
1336 #define ENTROPY_LVL_HIGH (80)
1339 * For increasead precision in shannon_entropy calculation,
1340 * let's do pow(n, M) to save more digits after comma:
1342 * - maximum int bit length is 64
1343 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1344 * - 13 * 4 = 52 < 64 -> M = 4
1348 static inline u32
ilog2_w(u64 n
)
1350 return ilog2(n
* n
* n
* n
);
1353 static u32
shannon_entropy(struct heuristic_ws
*ws
)
1355 const u32 entropy_max
= 8 * ilog2_w(2);
1356 u32 entropy_sum
= 0;
1357 u32 p
, p_base
, sz_base
;
1360 sz_base
= ilog2_w(ws
->sample_size
);
1361 for (i
= 0; i
< BUCKET_SIZE
&& ws
->bucket
[i
].count
> 0; i
++) {
1362 p
= ws
->bucket
[i
].count
;
1363 p_base
= ilog2_w(p
);
1364 entropy_sum
+= p
* (sz_base
- p_base
);
1367 entropy_sum
/= ws
->sample_size
;
1368 return entropy_sum
* 100 / entropy_max
;
1371 #define RADIX_BASE 4U
1372 #define COUNTERS_SIZE (1U << RADIX_BASE)
1374 static u8
get4bits(u64 num
, int shift
) {
1379 low4bits
= (COUNTERS_SIZE
- 1) - (num
% COUNTERS_SIZE
);
1384 * Use 4 bits as radix base
1385 * Use 16 u32 counters for calculating new position in buf array
1387 * @array - array that will be sorted
1388 * @array_buf - buffer array to store sorting results
1389 * must be equal in size to @array
1392 static void radix_sort(struct bucket_item
*array
, struct bucket_item
*array_buf
,
1397 u32 counters
[COUNTERS_SIZE
];
1405 * Try avoid useless loop iterations for small numbers stored in big
1406 * counters. Example: 48 33 4 ... in 64bit array
1408 max_num
= array
[0].count
;
1409 for (i
= 1; i
< num
; i
++) {
1410 buf_num
= array
[i
].count
;
1411 if (buf_num
> max_num
)
1415 buf_num
= ilog2(max_num
);
1416 bitlen
= ALIGN(buf_num
, RADIX_BASE
* 2);
1419 while (shift
< bitlen
) {
1420 memset(counters
, 0, sizeof(counters
));
1422 for (i
= 0; i
< num
; i
++) {
1423 buf_num
= array
[i
].count
;
1424 addr
= get4bits(buf_num
, shift
);
1428 for (i
= 1; i
< COUNTERS_SIZE
; i
++)
1429 counters
[i
] += counters
[i
- 1];
1431 for (i
= num
- 1; i
>= 0; i
--) {
1432 buf_num
= array
[i
].count
;
1433 addr
= get4bits(buf_num
, shift
);
1435 new_addr
= counters
[addr
];
1436 array_buf
[new_addr
] = array
[i
];
1439 shift
+= RADIX_BASE
;
1442 * Normal radix expects to move data from a temporary array, to
1443 * the main one. But that requires some CPU time. Avoid that
1444 * by doing another sort iteration to original array instead of
1447 memset(counters
, 0, sizeof(counters
));
1449 for (i
= 0; i
< num
; i
++) {
1450 buf_num
= array_buf
[i
].count
;
1451 addr
= get4bits(buf_num
, shift
);
1455 for (i
= 1; i
< COUNTERS_SIZE
; i
++)
1456 counters
[i
] += counters
[i
- 1];
1458 for (i
= num
- 1; i
>= 0; i
--) {
1459 buf_num
= array_buf
[i
].count
;
1460 addr
= get4bits(buf_num
, shift
);
1462 new_addr
= counters
[addr
];
1463 array
[new_addr
] = array_buf
[i
];
1466 shift
+= RADIX_BASE
;
1471 * Size of the core byte set - how many bytes cover 90% of the sample
1473 * There are several types of structured binary data that use nearly all byte
1474 * values. The distribution can be uniform and counts in all buckets will be
1475 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1477 * Other possibility is normal (Gaussian) distribution, where the data could
1478 * be potentially compressible, but we have to take a few more steps to decide
1481 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1482 * compression algo can easy fix that
1483 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1484 * probability is not compressible
1486 #define BYTE_CORE_SET_LOW (64)
1487 #define BYTE_CORE_SET_HIGH (200)
1489 static int byte_core_set_size(struct heuristic_ws
*ws
)
1492 u32 coreset_sum
= 0;
1493 const u32 core_set_threshold
= ws
->sample_size
* 90 / 100;
1494 struct bucket_item
*bucket
= ws
->bucket
;
1496 /* Sort in reverse order */
1497 radix_sort(ws
->bucket
, ws
->bucket_b
, BUCKET_SIZE
);
1499 for (i
= 0; i
< BYTE_CORE_SET_LOW
; i
++)
1500 coreset_sum
+= bucket
[i
].count
;
1502 if (coreset_sum
> core_set_threshold
)
1505 for (; i
< BYTE_CORE_SET_HIGH
&& bucket
[i
].count
> 0; i
++) {
1506 coreset_sum
+= bucket
[i
].count
;
1507 if (coreset_sum
> core_set_threshold
)
1515 * Count byte values in buckets.
1516 * This heuristic can detect textual data (configs, xml, json, html, etc).
1517 * Because in most text-like data byte set is restricted to limited number of
1518 * possible characters, and that restriction in most cases makes data easy to
1521 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1522 * less - compressible
1523 * more - need additional analysis
1525 #define BYTE_SET_THRESHOLD (64)
1527 static u32
byte_set_size(const struct heuristic_ws
*ws
)
1530 u32 byte_set_size
= 0;
1532 for (i
= 0; i
< BYTE_SET_THRESHOLD
; i
++) {
1533 if (ws
->bucket
[i
].count
> 0)
1538 * Continue collecting count of byte values in buckets. If the byte
1539 * set size is bigger then the threshold, it's pointless to continue,
1540 * the detection technique would fail for this type of data.
1542 for (; i
< BUCKET_SIZE
; i
++) {
1543 if (ws
->bucket
[i
].count
> 0) {
1545 if (byte_set_size
> BYTE_SET_THRESHOLD
)
1546 return byte_set_size
;
1550 return byte_set_size
;
1553 static bool sample_repeated_patterns(struct heuristic_ws
*ws
)
1555 const u32 half_of_sample
= ws
->sample_size
/ 2;
1556 const u8
*data
= ws
->sample
;
1558 return memcmp(&data
[0], &data
[half_of_sample
], half_of_sample
) == 0;
1561 static void heuristic_collect_sample(struct inode
*inode
, u64 start
, u64 end
,
1562 struct heuristic_ws
*ws
)
1565 u64 index
, index_end
;
1566 u32 i
, curr_sample_pos
;
1570 * Compression handles the input data by chunks of 128KiB
1571 * (defined by BTRFS_MAX_UNCOMPRESSED)
1573 * We do the same for the heuristic and loop over the whole range.
1575 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1576 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1578 if (end
- start
> BTRFS_MAX_UNCOMPRESSED
)
1579 end
= start
+ BTRFS_MAX_UNCOMPRESSED
;
1581 index
= start
>> PAGE_SHIFT
;
1582 index_end
= end
>> PAGE_SHIFT
;
1584 /* Don't miss unaligned end */
1585 if (!IS_ALIGNED(end
, PAGE_SIZE
))
1588 curr_sample_pos
= 0;
1589 while (index
< index_end
) {
1590 page
= find_get_page(inode
->i_mapping
, index
);
1591 in_data
= kmap(page
);
1592 /* Handle case where the start is not aligned to PAGE_SIZE */
1593 i
= start
% PAGE_SIZE
;
1594 while (i
< PAGE_SIZE
- SAMPLING_READ_SIZE
) {
1595 /* Don't sample any garbage from the last page */
1596 if (start
> end
- SAMPLING_READ_SIZE
)
1598 memcpy(&ws
->sample
[curr_sample_pos
], &in_data
[i
],
1599 SAMPLING_READ_SIZE
);
1600 i
+= SAMPLING_INTERVAL
;
1601 start
+= SAMPLING_INTERVAL
;
1602 curr_sample_pos
+= SAMPLING_READ_SIZE
;
1610 ws
->sample_size
= curr_sample_pos
;
1614 * Compression heuristic.
1616 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1617 * quickly (compared to direct compression) detect data characteristics
1618 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1621 * The following types of analysis can be performed:
1622 * - detect mostly zero data
1623 * - detect data with low "byte set" size (text, etc)
1624 * - detect data with low/high "core byte" set
1626 * Return non-zero if the compression should be done, 0 otherwise.
1628 int btrfs_compress_heuristic(struct inode
*inode
, u64 start
, u64 end
)
1630 struct list_head
*ws_list
= get_workspace(0, 0);
1631 struct heuristic_ws
*ws
;
1636 ws
= list_entry(ws_list
, struct heuristic_ws
, list
);
1638 heuristic_collect_sample(inode
, start
, end
, ws
);
1640 if (sample_repeated_patterns(ws
)) {
1645 memset(ws
->bucket
, 0, sizeof(*ws
->bucket
)*BUCKET_SIZE
);
1647 for (i
= 0; i
< ws
->sample_size
; i
++) {
1648 byte
= ws
->sample
[i
];
1649 ws
->bucket
[byte
].count
++;
1652 i
= byte_set_size(ws
);
1653 if (i
< BYTE_SET_THRESHOLD
) {
1658 i
= byte_core_set_size(ws
);
1659 if (i
<= BYTE_CORE_SET_LOW
) {
1664 if (i
>= BYTE_CORE_SET_HIGH
) {
1669 i
= shannon_entropy(ws
);
1670 if (i
<= ENTROPY_LVL_ACEPTABLE
) {
1676 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1677 * needed to give green light to compression.
1679 * For now just assume that compression at that level is not worth the
1680 * resources because:
1682 * 1. it is possible to defrag the data later
1684 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1685 * values, every bucket has counter at level ~54. The heuristic would
1686 * be confused. This can happen when data have some internal repeated
1687 * patterns like "abbacbbc...". This can be detected by analyzing
1688 * pairs of bytes, which is too costly.
1690 if (i
< ENTROPY_LVL_HIGH
) {
1699 put_workspace(0, ws_list
);
1704 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1705 * level, unrecognized string will set the default level
1707 unsigned int btrfs_compress_str2level(unsigned int type
, const char *str
)
1709 unsigned int level
= 0;
1715 if (str
[0] == ':') {
1716 ret
= kstrtouint(str
+ 1, 10, &level
);
1721 level
= btrfs_compress_set_level(type
, level
);