1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Oracle. All rights reserved.
6 #include <linux/kernel.h>
8 #include <linux/file.h>
10 #include <linux/pagemap.h>
11 #include <linux/highmem.h>
12 #include <linux/time.h>
13 #include <linux/init.h>
14 #include <linux/string.h>
15 #include <linux/backing-dev.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sched/mm.h>
19 #include <linux/log2.h>
20 #include <crypto/hash.h>
24 #include "transaction.h"
25 #include "btrfs_inode.h"
27 #include "ordered-data.h"
28 #include "compression.h"
29 #include "extent_io.h"
30 #include "extent_map.h"
32 int zlib_compress_pages(struct list_head
*ws
, struct address_space
*mapping
,
33 u64 start
, struct page
**pages
, unsigned long *out_pages
,
34 unsigned long *total_in
, unsigned long *total_out
);
35 int zlib_decompress_bio(struct list_head
*ws
, struct compressed_bio
*cb
);
36 int zlib_decompress(struct list_head
*ws
, unsigned char *data_in
,
37 struct page
*dest_page
, unsigned long start_byte
, size_t srclen
,
39 struct list_head
*zlib_alloc_workspace(unsigned int level
);
40 void zlib_free_workspace(struct list_head
*ws
);
41 struct list_head
*zlib_get_workspace(unsigned int level
);
43 int lzo_compress_pages(struct list_head
*ws
, struct address_space
*mapping
,
44 u64 start
, struct page
**pages
, unsigned long *out_pages
,
45 unsigned long *total_in
, unsigned long *total_out
);
46 int lzo_decompress_bio(struct list_head
*ws
, struct compressed_bio
*cb
);
47 int lzo_decompress(struct list_head
*ws
, unsigned char *data_in
,
48 struct page
*dest_page
, unsigned long start_byte
, size_t srclen
,
50 struct list_head
*lzo_alloc_workspace(unsigned int level
);
51 void lzo_free_workspace(struct list_head
*ws
);
53 int zstd_compress_pages(struct list_head
*ws
, struct address_space
*mapping
,
54 u64 start
, struct page
**pages
, unsigned long *out_pages
,
55 unsigned long *total_in
, unsigned long *total_out
);
56 int zstd_decompress_bio(struct list_head
*ws
, struct compressed_bio
*cb
);
57 int zstd_decompress(struct list_head
*ws
, unsigned char *data_in
,
58 struct page
*dest_page
, unsigned long start_byte
, size_t srclen
,
60 void zstd_init_workspace_manager(void);
61 void zstd_cleanup_workspace_manager(void);
62 struct list_head
*zstd_alloc_workspace(unsigned int level
);
63 void zstd_free_workspace(struct list_head
*ws
);
64 struct list_head
*zstd_get_workspace(unsigned int level
);
65 void zstd_put_workspace(struct list_head
*ws
);
67 static const char* const btrfs_compress_types
[] = { "", "zlib", "lzo", "zstd" };
69 const char* btrfs_compress_type2str(enum btrfs_compression_type type
)
72 case BTRFS_COMPRESS_ZLIB
:
73 case BTRFS_COMPRESS_LZO
:
74 case BTRFS_COMPRESS_ZSTD
:
75 case BTRFS_COMPRESS_NONE
:
76 return btrfs_compress_types
[type
];
84 bool btrfs_compress_is_valid_type(const char *str
, size_t len
)
88 for (i
= 1; i
< ARRAY_SIZE(btrfs_compress_types
); i
++) {
89 size_t comp_len
= strlen(btrfs_compress_types
[i
]);
94 if (!strncmp(btrfs_compress_types
[i
], str
, comp_len
))
100 static int compression_compress_pages(int type
, struct list_head
*ws
,
101 struct address_space
*mapping
, u64 start
, struct page
**pages
,
102 unsigned long *out_pages
, unsigned long *total_in
,
103 unsigned long *total_out
)
106 case BTRFS_COMPRESS_ZLIB
:
107 return zlib_compress_pages(ws
, mapping
, start
, pages
,
108 out_pages
, total_in
, total_out
);
109 case BTRFS_COMPRESS_LZO
:
110 return lzo_compress_pages(ws
, mapping
, start
, pages
,
111 out_pages
, total_in
, total_out
);
112 case BTRFS_COMPRESS_ZSTD
:
113 return zstd_compress_pages(ws
, mapping
, start
, pages
,
114 out_pages
, total_in
, total_out
);
115 case BTRFS_COMPRESS_NONE
:
118 * This can't happen, the type is validated several times
119 * before we get here. As a sane fallback, return what the
120 * callers will understand as 'no compression happened'.
126 static int compression_decompress_bio(int type
, struct list_head
*ws
,
127 struct compressed_bio
*cb
)
130 case BTRFS_COMPRESS_ZLIB
: return zlib_decompress_bio(ws
, cb
);
131 case BTRFS_COMPRESS_LZO
: return lzo_decompress_bio(ws
, cb
);
132 case BTRFS_COMPRESS_ZSTD
: return zstd_decompress_bio(ws
, cb
);
133 case BTRFS_COMPRESS_NONE
:
136 * This can't happen, the type is validated several times
137 * before we get here.
143 static int compression_decompress(int type
, struct list_head
*ws
,
144 unsigned char *data_in
, struct page
*dest_page
,
145 unsigned long start_byte
, size_t srclen
, size_t destlen
)
148 case BTRFS_COMPRESS_ZLIB
: return zlib_decompress(ws
, data_in
, dest_page
,
149 start_byte
, srclen
, destlen
);
150 case BTRFS_COMPRESS_LZO
: return lzo_decompress(ws
, data_in
, dest_page
,
151 start_byte
, srclen
, destlen
);
152 case BTRFS_COMPRESS_ZSTD
: return zstd_decompress(ws
, data_in
, dest_page
,
153 start_byte
, srclen
, destlen
);
154 case BTRFS_COMPRESS_NONE
:
157 * This can't happen, the type is validated several times
158 * before we get here.
164 static int btrfs_decompress_bio(struct compressed_bio
*cb
);
166 static inline int compressed_bio_size(struct btrfs_fs_info
*fs_info
,
167 unsigned long disk_size
)
169 u16 csum_size
= btrfs_super_csum_size(fs_info
->super_copy
);
171 return sizeof(struct compressed_bio
) +
172 (DIV_ROUND_UP(disk_size
, fs_info
->sectorsize
)) * csum_size
;
175 static int check_compressed_csum(struct btrfs_inode
*inode
, struct bio
*bio
,
178 struct btrfs_fs_info
*fs_info
= inode
->root
->fs_info
;
179 SHASH_DESC_ON_STACK(shash
, fs_info
->csum_shash
);
180 const u16 csum_size
= btrfs_super_csum_size(fs_info
->super_copy
);
184 u8 csum
[BTRFS_CSUM_SIZE
];
185 struct compressed_bio
*cb
= bio
->bi_private
;
186 u8
*cb_sum
= cb
->sums
;
188 if (inode
->flags
& BTRFS_INODE_NODATASUM
)
191 shash
->tfm
= fs_info
->csum_shash
;
193 for (i
= 0; i
< cb
->nr_pages
; i
++) {
194 page
= cb
->compressed_pages
[i
];
196 kaddr
= kmap_atomic(page
);
197 crypto_shash_digest(shash
, kaddr
, PAGE_SIZE
, csum
);
198 kunmap_atomic(kaddr
);
200 if (memcmp(&csum
, cb_sum
, csum_size
)) {
201 btrfs_print_data_csum_error(inode
, disk_start
,
202 csum
, cb_sum
, cb
->mirror_num
);
203 if (btrfs_io_bio(bio
)->device
)
204 btrfs_dev_stat_inc_and_print(
205 btrfs_io_bio(bio
)->device
,
206 BTRFS_DEV_STAT_CORRUPTION_ERRS
);
214 /* when we finish reading compressed pages from the disk, we
215 * decompress them and then run the bio end_io routines on the
216 * decompressed pages (in the inode address space).
218 * This allows the checksumming and other IO error handling routines
221 * The compressed pages are freed here, and it must be run
224 static void end_compressed_bio_read(struct bio
*bio
)
226 struct compressed_bio
*cb
= bio
->bi_private
;
230 unsigned int mirror
= btrfs_io_bio(bio
)->mirror_num
;
236 /* if there are more bios still pending for this compressed
239 if (!refcount_dec_and_test(&cb
->pending_bios
))
243 * Record the correct mirror_num in cb->orig_bio so that
244 * read-repair can work properly.
246 btrfs_io_bio(cb
->orig_bio
)->mirror_num
= mirror
;
247 cb
->mirror_num
= mirror
;
250 * Some IO in this cb have failed, just skip checksum as there
251 * is no way it could be correct.
257 ret
= check_compressed_csum(BTRFS_I(inode
), bio
,
258 (u64
)bio
->bi_iter
.bi_sector
<< 9);
262 /* ok, we're the last bio for this extent, lets start
265 ret
= btrfs_decompress_bio(cb
);
271 /* release the compressed pages */
273 for (index
= 0; index
< cb
->nr_pages
; index
++) {
274 page
= cb
->compressed_pages
[index
];
275 page
->mapping
= NULL
;
279 /* do io completion on the original bio */
281 bio_io_error(cb
->orig_bio
);
283 struct bio_vec
*bvec
;
284 struct bvec_iter_all iter_all
;
287 * we have verified the checksum already, set page
288 * checked so the end_io handlers know about it
290 ASSERT(!bio_flagged(bio
, BIO_CLONED
));
291 bio_for_each_segment_all(bvec
, cb
->orig_bio
, iter_all
)
292 SetPageChecked(bvec
->bv_page
);
294 bio_endio(cb
->orig_bio
);
297 /* finally free the cb struct */
298 kfree(cb
->compressed_pages
);
305 * Clear the writeback bits on all of the file
306 * pages for a compressed write
308 static noinline
void end_compressed_writeback(struct inode
*inode
,
309 const struct compressed_bio
*cb
)
311 unsigned long index
= cb
->start
>> PAGE_SHIFT
;
312 unsigned long end_index
= (cb
->start
+ cb
->len
- 1) >> PAGE_SHIFT
;
313 struct page
*pages
[16];
314 unsigned long nr_pages
= end_index
- index
+ 1;
319 mapping_set_error(inode
->i_mapping
, -EIO
);
321 while (nr_pages
> 0) {
322 ret
= find_get_pages_contig(inode
->i_mapping
, index
,
324 nr_pages
, ARRAY_SIZE(pages
)), pages
);
330 for (i
= 0; i
< ret
; i
++) {
332 SetPageError(pages
[i
]);
333 end_page_writeback(pages
[i
]);
339 /* the inode may be gone now */
343 * do the cleanup once all the compressed pages hit the disk.
344 * This will clear writeback on the file pages and free the compressed
347 * This also calls the writeback end hooks for the file pages so that
348 * metadata and checksums can be updated in the file.
350 static void end_compressed_bio_write(struct bio
*bio
)
352 struct compressed_bio
*cb
= bio
->bi_private
;
360 /* if there are more bios still pending for this compressed
363 if (!refcount_dec_and_test(&cb
->pending_bios
))
366 /* ok, we're the last bio for this extent, step one is to
367 * call back into the FS and do all the end_io operations
370 cb
->compressed_pages
[0]->mapping
= cb
->inode
->i_mapping
;
371 btrfs_writepage_endio_finish_ordered(cb
->compressed_pages
[0],
372 cb
->start
, cb
->start
+ cb
->len
- 1,
373 bio
->bi_status
== BLK_STS_OK
);
374 cb
->compressed_pages
[0]->mapping
= NULL
;
376 end_compressed_writeback(inode
, cb
);
377 /* note, our inode could be gone now */
380 * release the compressed pages, these came from alloc_page and
381 * are not attached to the inode at all
384 for (index
= 0; index
< cb
->nr_pages
; index
++) {
385 page
= cb
->compressed_pages
[index
];
386 page
->mapping
= NULL
;
390 /* finally free the cb struct */
391 kfree(cb
->compressed_pages
);
398 * worker function to build and submit bios for previously compressed pages.
399 * The corresponding pages in the inode should be marked for writeback
400 * and the compressed pages should have a reference on them for dropping
401 * when the IO is complete.
403 * This also checksums the file bytes and gets things ready for
406 blk_status_t
btrfs_submit_compressed_write(struct btrfs_inode
*inode
, u64 start
,
407 unsigned long len
, u64 disk_start
,
408 unsigned long compressed_len
,
409 struct page
**compressed_pages
,
410 unsigned long nr_pages
,
411 unsigned int write_flags
,
412 struct cgroup_subsys_state
*blkcg_css
)
414 struct btrfs_fs_info
*fs_info
= inode
->root
->fs_info
;
415 struct bio
*bio
= NULL
;
416 struct compressed_bio
*cb
;
417 unsigned long bytes_left
;
420 u64 first_byte
= disk_start
;
422 int skip_sum
= inode
->flags
& BTRFS_INODE_NODATASUM
;
424 WARN_ON(!PAGE_ALIGNED(start
));
425 cb
= kmalloc(compressed_bio_size(fs_info
, compressed_len
), GFP_NOFS
);
427 return BLK_STS_RESOURCE
;
428 refcount_set(&cb
->pending_bios
, 0);
430 cb
->inode
= &inode
->vfs_inode
;
434 cb
->compressed_pages
= compressed_pages
;
435 cb
->compressed_len
= compressed_len
;
437 cb
->nr_pages
= nr_pages
;
439 bio
= btrfs_bio_alloc(first_byte
);
440 bio
->bi_opf
= REQ_OP_WRITE
| write_flags
;
441 bio
->bi_private
= cb
;
442 bio
->bi_end_io
= end_compressed_bio_write
;
445 bio
->bi_opf
|= REQ_CGROUP_PUNT
;
446 kthread_associate_blkcg(blkcg_css
);
448 refcount_set(&cb
->pending_bios
, 1);
450 /* create and submit bios for the compressed pages */
451 bytes_left
= compressed_len
;
452 for (pg_index
= 0; pg_index
< cb
->nr_pages
; pg_index
++) {
455 page
= compressed_pages
[pg_index
];
456 page
->mapping
= inode
->vfs_inode
.i_mapping
;
457 if (bio
->bi_iter
.bi_size
)
458 submit
= btrfs_bio_fits_in_stripe(page
, PAGE_SIZE
, bio
,
461 page
->mapping
= NULL
;
462 if (submit
|| bio_add_page(bio
, page
, PAGE_SIZE
, 0) <
465 * inc the count before we submit the bio so
466 * we know the end IO handler won't happen before
467 * we inc the count. Otherwise, the cb might get
468 * freed before we're done setting it up
470 refcount_inc(&cb
->pending_bios
);
471 ret
= btrfs_bio_wq_end_io(fs_info
, bio
,
472 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
;
486 bio
= btrfs_bio_alloc(first_byte
);
487 bio
->bi_opf
= REQ_OP_WRITE
| write_flags
;
488 bio
->bi_private
= cb
;
489 bio
->bi_end_io
= end_compressed_bio_write
;
491 bio
->bi_opf
|= REQ_CGROUP_PUNT
;
492 bio_add_page(bio
, page
, PAGE_SIZE
, 0);
494 if (bytes_left
< PAGE_SIZE
) {
496 "bytes left %lu compress len %lu nr %lu",
497 bytes_left
, cb
->compressed_len
, cb
->nr_pages
);
499 bytes_left
-= PAGE_SIZE
;
500 first_byte
+= PAGE_SIZE
;
504 ret
= btrfs_bio_wq_end_io(fs_info
, bio
, BTRFS_WQ_ENDIO_DATA
);
505 BUG_ON(ret
); /* -ENOMEM */
508 ret
= btrfs_csum_one_bio(inode
, bio
, start
, 1);
509 BUG_ON(ret
); /* -ENOMEM */
512 ret
= btrfs_map_bio(fs_info
, bio
, 0);
514 bio
->bi_status
= ret
;
519 kthread_associate_blkcg(NULL
);
524 static u64
bio_end_offset(struct bio
*bio
)
526 struct bio_vec
*last
= bio_last_bvec_all(bio
);
528 return page_offset(last
->bv_page
) + last
->bv_len
+ last
->bv_offset
;
531 static noinline
int add_ra_bio_pages(struct inode
*inode
,
533 struct compressed_bio
*cb
)
535 unsigned long end_index
;
536 unsigned long pg_index
;
538 u64 isize
= i_size_read(inode
);
541 unsigned long nr_pages
= 0;
542 struct extent_map
*em
;
543 struct address_space
*mapping
= inode
->i_mapping
;
544 struct extent_map_tree
*em_tree
;
545 struct extent_io_tree
*tree
;
549 last_offset
= bio_end_offset(cb
->orig_bio
);
550 em_tree
= &BTRFS_I(inode
)->extent_tree
;
551 tree
= &BTRFS_I(inode
)->io_tree
;
556 end_index
= (i_size_read(inode
) - 1) >> PAGE_SHIFT
;
558 while (last_offset
< compressed_end
) {
559 pg_index
= last_offset
>> PAGE_SHIFT
;
561 if (pg_index
> end_index
)
564 page
= xa_load(&mapping
->i_pages
, pg_index
);
565 if (page
&& !xa_is_value(page
)) {
572 page
= __page_cache_alloc(mapping_gfp_constraint(mapping
,
577 if (add_to_page_cache_lru(page
, mapping
, pg_index
, GFP_NOFS
)) {
582 end
= last_offset
+ PAGE_SIZE
- 1;
584 * at this point, we have a locked page in the page cache
585 * for these bytes in the file. But, we have to make
586 * sure they map to this compressed extent on disk.
588 set_page_extent_mapped(page
);
589 lock_extent(tree
, last_offset
, end
);
590 read_lock(&em_tree
->lock
);
591 em
= lookup_extent_mapping(em_tree
, last_offset
,
593 read_unlock(&em_tree
->lock
);
595 if (!em
|| last_offset
< em
->start
||
596 (last_offset
+ PAGE_SIZE
> extent_map_end(em
)) ||
597 (em
->block_start
>> 9) != cb
->orig_bio
->bi_iter
.bi_sector
) {
599 unlock_extent(tree
, last_offset
, end
);
606 if (page
->index
== end_index
) {
608 size_t zero_offset
= offset_in_page(isize
);
612 zeros
= PAGE_SIZE
- zero_offset
;
613 userpage
= kmap_atomic(page
);
614 memset(userpage
+ zero_offset
, 0, zeros
);
615 flush_dcache_page(page
);
616 kunmap_atomic(userpage
);
620 ret
= bio_add_page(cb
->orig_bio
, page
,
623 if (ret
== PAGE_SIZE
) {
627 unlock_extent(tree
, last_offset
, end
);
633 last_offset
+= PAGE_SIZE
;
639 * for a compressed read, the bio we get passed has all the inode pages
640 * in it. We don't actually do IO on those pages but allocate new ones
641 * to hold the compressed pages on disk.
643 * bio->bi_iter.bi_sector points to the compressed extent on disk
644 * bio->bi_io_vec points to all of the inode pages
646 * After the compressed pages are read, we copy the bytes into the
647 * bio we were passed and then call the bio end_io calls
649 blk_status_t
btrfs_submit_compressed_read(struct inode
*inode
, struct bio
*bio
,
650 int mirror_num
, unsigned long bio_flags
)
652 struct btrfs_fs_info
*fs_info
= btrfs_sb(inode
->i_sb
);
653 struct extent_map_tree
*em_tree
;
654 struct compressed_bio
*cb
;
655 unsigned long compressed_len
;
656 unsigned long nr_pages
;
657 unsigned long pg_index
;
659 struct bio
*comp_bio
;
660 u64 cur_disk_byte
= (u64
)bio
->bi_iter
.bi_sector
<< 9;
663 struct extent_map
*em
;
664 blk_status_t ret
= BLK_STS_RESOURCE
;
666 const u16 csum_size
= btrfs_super_csum_size(fs_info
->super_copy
);
669 em_tree
= &BTRFS_I(inode
)->extent_tree
;
671 /* we need the actual starting offset of this extent in the file */
672 read_lock(&em_tree
->lock
);
673 em
= lookup_extent_mapping(em_tree
,
674 page_offset(bio_first_page_all(bio
)),
676 read_unlock(&em_tree
->lock
);
678 return BLK_STS_IOERR
;
680 compressed_len
= em
->block_len
;
681 cb
= kmalloc(compressed_bio_size(fs_info
, compressed_len
), GFP_NOFS
);
685 refcount_set(&cb
->pending_bios
, 0);
688 cb
->mirror_num
= mirror_num
;
691 cb
->start
= em
->orig_start
;
693 em_start
= em
->start
;
698 cb
->len
= bio
->bi_iter
.bi_size
;
699 cb
->compressed_len
= compressed_len
;
700 cb
->compress_type
= extent_compress_type(bio_flags
);
703 nr_pages
= DIV_ROUND_UP(compressed_len
, PAGE_SIZE
);
704 cb
->compressed_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
706 if (!cb
->compressed_pages
)
709 for (pg_index
= 0; pg_index
< nr_pages
; pg_index
++) {
710 cb
->compressed_pages
[pg_index
] = alloc_page(GFP_NOFS
|
712 if (!cb
->compressed_pages
[pg_index
]) {
713 faili
= pg_index
- 1;
714 ret
= BLK_STS_RESOURCE
;
718 faili
= nr_pages
- 1;
719 cb
->nr_pages
= nr_pages
;
721 add_ra_bio_pages(inode
, em_start
+ em_len
, cb
);
723 /* include any pages we added in add_ra-bio_pages */
724 cb
->len
= bio
->bi_iter
.bi_size
;
726 comp_bio
= btrfs_bio_alloc(cur_disk_byte
);
727 comp_bio
->bi_opf
= REQ_OP_READ
;
728 comp_bio
->bi_private
= cb
;
729 comp_bio
->bi_end_io
= end_compressed_bio_read
;
730 refcount_set(&cb
->pending_bios
, 1);
732 for (pg_index
= 0; pg_index
< nr_pages
; pg_index
++) {
735 page
= cb
->compressed_pages
[pg_index
];
736 page
->mapping
= inode
->i_mapping
;
737 page
->index
= em_start
>> PAGE_SHIFT
;
739 if (comp_bio
->bi_iter
.bi_size
)
740 submit
= btrfs_bio_fits_in_stripe(page
, PAGE_SIZE
,
743 page
->mapping
= NULL
;
744 if (submit
|| bio_add_page(comp_bio
, page
, PAGE_SIZE
, 0) <
746 unsigned int nr_sectors
;
748 ret
= btrfs_bio_wq_end_io(fs_info
, comp_bio
,
749 BTRFS_WQ_ENDIO_DATA
);
750 BUG_ON(ret
); /* -ENOMEM */
753 * inc the count before we submit the bio so
754 * we know the end IO handler won't happen before
755 * we inc the count. Otherwise, the cb might get
756 * freed before we're done setting it up
758 refcount_inc(&cb
->pending_bios
);
760 if (!(BTRFS_I(inode
)->flags
& BTRFS_INODE_NODATASUM
)) {
761 ret
= btrfs_lookup_bio_sums(inode
, comp_bio
,
763 BUG_ON(ret
); /* -ENOMEM */
766 nr_sectors
= DIV_ROUND_UP(comp_bio
->bi_iter
.bi_size
,
767 fs_info
->sectorsize
);
768 sums
+= csum_size
* nr_sectors
;
770 ret
= btrfs_map_bio(fs_info
, comp_bio
, mirror_num
);
772 comp_bio
->bi_status
= ret
;
776 comp_bio
= btrfs_bio_alloc(cur_disk_byte
);
777 comp_bio
->bi_opf
= REQ_OP_READ
;
778 comp_bio
->bi_private
= cb
;
779 comp_bio
->bi_end_io
= end_compressed_bio_read
;
781 bio_add_page(comp_bio
, page
, PAGE_SIZE
, 0);
783 cur_disk_byte
+= PAGE_SIZE
;
786 ret
= btrfs_bio_wq_end_io(fs_info
, comp_bio
, BTRFS_WQ_ENDIO_DATA
);
787 BUG_ON(ret
); /* -ENOMEM */
789 if (!(BTRFS_I(inode
)->flags
& BTRFS_INODE_NODATASUM
)) {
790 ret
= btrfs_lookup_bio_sums(inode
, comp_bio
, (u64
)-1, sums
);
791 BUG_ON(ret
); /* -ENOMEM */
794 ret
= btrfs_map_bio(fs_info
, comp_bio
, mirror_num
);
796 comp_bio
->bi_status
= ret
;
804 __free_page(cb
->compressed_pages
[faili
]);
808 kfree(cb
->compressed_pages
);
817 * Heuristic uses systematic sampling to collect data from the input data
818 * range, the logic can be tuned by the following constants:
820 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
821 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
823 #define SAMPLING_READ_SIZE (16)
824 #define SAMPLING_INTERVAL (256)
827 * For statistical analysis of the input data we consider bytes that form a
828 * Galois Field of 256 objects. Each object has an attribute count, ie. how
829 * many times the object appeared in the sample.
831 #define BUCKET_SIZE (256)
834 * The size of the sample is based on a statistical sampling rule of thumb.
835 * The common way is to perform sampling tests as long as the number of
836 * elements in each cell is at least 5.
838 * Instead of 5, we choose 32 to obtain more accurate results.
839 * If the data contain the maximum number of symbols, which is 256, we obtain a
840 * sample size bound by 8192.
842 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
843 * from up to 512 locations.
845 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
846 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
852 struct heuristic_ws
{
853 /* Partial copy of input data */
856 /* Buckets store counters for each byte value */
857 struct bucket_item
*bucket
;
859 struct bucket_item
*bucket_b
;
860 struct list_head list
;
863 static struct workspace_manager heuristic_wsm
;
865 static void free_heuristic_ws(struct list_head
*ws
)
867 struct heuristic_ws
*workspace
;
869 workspace
= list_entry(ws
, struct heuristic_ws
, list
);
871 kvfree(workspace
->sample
);
872 kfree(workspace
->bucket
);
873 kfree(workspace
->bucket_b
);
877 static struct list_head
*alloc_heuristic_ws(unsigned int level
)
879 struct heuristic_ws
*ws
;
881 ws
= kzalloc(sizeof(*ws
), GFP_KERNEL
);
883 return ERR_PTR(-ENOMEM
);
885 ws
->sample
= kvmalloc(MAX_SAMPLE_SIZE
, GFP_KERNEL
);
889 ws
->bucket
= kcalloc(BUCKET_SIZE
, sizeof(*ws
->bucket
), GFP_KERNEL
);
893 ws
->bucket_b
= kcalloc(BUCKET_SIZE
, sizeof(*ws
->bucket_b
), GFP_KERNEL
);
897 INIT_LIST_HEAD(&ws
->list
);
900 free_heuristic_ws(&ws
->list
);
901 return ERR_PTR(-ENOMEM
);
904 const struct btrfs_compress_op btrfs_heuristic_compress
= {
905 .workspace_manager
= &heuristic_wsm
,
908 static const struct btrfs_compress_op
* const btrfs_compress_op
[] = {
909 /* The heuristic is represented as compression type 0 */
910 &btrfs_heuristic_compress
,
911 &btrfs_zlib_compress
,
913 &btrfs_zstd_compress
,
916 static struct list_head
*alloc_workspace(int type
, unsigned int level
)
919 case BTRFS_COMPRESS_NONE
: return alloc_heuristic_ws(level
);
920 case BTRFS_COMPRESS_ZLIB
: return zlib_alloc_workspace(level
);
921 case BTRFS_COMPRESS_LZO
: return lzo_alloc_workspace(level
);
922 case BTRFS_COMPRESS_ZSTD
: return zstd_alloc_workspace(level
);
925 * This can't happen, the type is validated several times
926 * before we get here.
932 static void free_workspace(int type
, struct list_head
*ws
)
935 case BTRFS_COMPRESS_NONE
: return free_heuristic_ws(ws
);
936 case BTRFS_COMPRESS_ZLIB
: return zlib_free_workspace(ws
);
937 case BTRFS_COMPRESS_LZO
: return lzo_free_workspace(ws
);
938 case BTRFS_COMPRESS_ZSTD
: return zstd_free_workspace(ws
);
941 * This can't happen, the type is validated several times
942 * before we get here.
948 static void btrfs_init_workspace_manager(int type
)
950 struct workspace_manager
*wsm
;
951 struct list_head
*workspace
;
953 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
954 INIT_LIST_HEAD(&wsm
->idle_ws
);
955 spin_lock_init(&wsm
->ws_lock
);
956 atomic_set(&wsm
->total_ws
, 0);
957 init_waitqueue_head(&wsm
->ws_wait
);
960 * Preallocate one workspace for each compression type so we can
961 * guarantee forward progress in the worst case
963 workspace
= alloc_workspace(type
, 0);
964 if (IS_ERR(workspace
)) {
966 "BTRFS: cannot preallocate compression workspace, will try later\n");
968 atomic_set(&wsm
->total_ws
, 1);
970 list_add(workspace
, &wsm
->idle_ws
);
974 static void btrfs_cleanup_workspace_manager(int type
)
976 struct workspace_manager
*wsman
;
977 struct list_head
*ws
;
979 wsman
= btrfs_compress_op
[type
]->workspace_manager
;
980 while (!list_empty(&wsman
->idle_ws
)) {
981 ws
= wsman
->idle_ws
.next
;
983 free_workspace(type
, ws
);
984 atomic_dec(&wsman
->total_ws
);
989 * This finds an available workspace or allocates a new one.
990 * If it's not possible to allocate a new one, waits until there's one.
991 * Preallocation makes a forward progress guarantees and we do not return
994 struct list_head
*btrfs_get_workspace(int type
, unsigned int level
)
996 struct workspace_manager
*wsm
;
997 struct list_head
*workspace
;
998 int cpus
= num_online_cpus();
1000 struct list_head
*idle_ws
;
1001 spinlock_t
*ws_lock
;
1003 wait_queue_head_t
*ws_wait
;
1006 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
1007 idle_ws
= &wsm
->idle_ws
;
1008 ws_lock
= &wsm
->ws_lock
;
1009 total_ws
= &wsm
->total_ws
;
1010 ws_wait
= &wsm
->ws_wait
;
1011 free_ws
= &wsm
->free_ws
;
1015 if (!list_empty(idle_ws
)) {
1016 workspace
= idle_ws
->next
;
1017 list_del(workspace
);
1019 spin_unlock(ws_lock
);
1023 if (atomic_read(total_ws
) > cpus
) {
1026 spin_unlock(ws_lock
);
1027 prepare_to_wait(ws_wait
, &wait
, TASK_UNINTERRUPTIBLE
);
1028 if (atomic_read(total_ws
) > cpus
&& !*free_ws
)
1030 finish_wait(ws_wait
, &wait
);
1033 atomic_inc(total_ws
);
1034 spin_unlock(ws_lock
);
1037 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1038 * to turn it off here because we might get called from the restricted
1039 * context of btrfs_compress_bio/btrfs_compress_pages
1041 nofs_flag
= memalloc_nofs_save();
1042 workspace
= alloc_workspace(type
, level
);
1043 memalloc_nofs_restore(nofs_flag
);
1045 if (IS_ERR(workspace
)) {
1046 atomic_dec(total_ws
);
1050 * Do not return the error but go back to waiting. There's a
1051 * workspace preallocated for each type and the compression
1052 * time is bounded so we get to a workspace eventually. This
1053 * makes our caller's life easier.
1055 * To prevent silent and low-probability deadlocks (when the
1056 * initial preallocation fails), check if there are any
1057 * workspaces at all.
1059 if (atomic_read(total_ws
) == 0) {
1060 static DEFINE_RATELIMIT_STATE(_rs
,
1061 /* once per minute */ 60 * HZ
,
1064 if (__ratelimit(&_rs
)) {
1065 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1073 static struct list_head
*get_workspace(int type
, int level
)
1076 case BTRFS_COMPRESS_NONE
: return btrfs_get_workspace(type
, level
);
1077 case BTRFS_COMPRESS_ZLIB
: return zlib_get_workspace(level
);
1078 case BTRFS_COMPRESS_LZO
: return btrfs_get_workspace(type
, level
);
1079 case BTRFS_COMPRESS_ZSTD
: return zstd_get_workspace(level
);
1082 * This can't happen, the type is validated several times
1083 * before we get here.
1090 * put a workspace struct back on the list or free it if we have enough
1091 * idle ones sitting around
1093 void btrfs_put_workspace(int type
, struct list_head
*ws
)
1095 struct workspace_manager
*wsm
;
1096 struct list_head
*idle_ws
;
1097 spinlock_t
*ws_lock
;
1099 wait_queue_head_t
*ws_wait
;
1102 wsm
= btrfs_compress_op
[type
]->workspace_manager
;
1103 idle_ws
= &wsm
->idle_ws
;
1104 ws_lock
= &wsm
->ws_lock
;
1105 total_ws
= &wsm
->total_ws
;
1106 ws_wait
= &wsm
->ws_wait
;
1107 free_ws
= &wsm
->free_ws
;
1110 if (*free_ws
<= num_online_cpus()) {
1111 list_add(ws
, idle_ws
);
1113 spin_unlock(ws_lock
);
1116 spin_unlock(ws_lock
);
1118 free_workspace(type
, ws
);
1119 atomic_dec(total_ws
);
1121 cond_wake_up(ws_wait
);
1124 static void put_workspace(int type
, struct list_head
*ws
)
1127 case BTRFS_COMPRESS_NONE
: return btrfs_put_workspace(type
, ws
);
1128 case BTRFS_COMPRESS_ZLIB
: return btrfs_put_workspace(type
, ws
);
1129 case BTRFS_COMPRESS_LZO
: return btrfs_put_workspace(type
, ws
);
1130 case BTRFS_COMPRESS_ZSTD
: return zstd_put_workspace(ws
);
1133 * This can't happen, the type is validated several times
1134 * before we get here.
1141 * Adjust @level according to the limits of the compression algorithm or
1142 * fallback to default
1144 static unsigned int btrfs_compress_set_level(int type
, unsigned level
)
1146 const struct btrfs_compress_op
*ops
= btrfs_compress_op
[type
];
1149 level
= ops
->default_level
;
1151 level
= min(level
, ops
->max_level
);
1157 * Given an address space and start and length, compress the bytes into @pages
1158 * that are allocated on demand.
1160 * @type_level is encoded algorithm and level, where level 0 means whatever
1161 * default the algorithm chooses and is opaque here;
1162 * - compression algo are 0-3
1163 * - the level are bits 4-7
1165 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1166 * and returns number of actually allocated pages
1168 * @total_in is used to return the number of bytes actually read. It
1169 * may be smaller than the input length if we had to exit early because we
1170 * ran out of room in the pages array or because we cross the
1171 * max_out threshold.
1173 * @total_out is an in/out parameter, must be set to the input length and will
1174 * be also used to return the total number of compressed bytes
1176 * @max_out tells us the max number of bytes that we're allowed to
1179 int btrfs_compress_pages(unsigned int type_level
, struct address_space
*mapping
,
1180 u64 start
, struct page
**pages
,
1181 unsigned long *out_pages
,
1182 unsigned long *total_in
,
1183 unsigned long *total_out
)
1185 int type
= btrfs_compress_type(type_level
);
1186 int level
= btrfs_compress_level(type_level
);
1187 struct list_head
*workspace
;
1190 level
= btrfs_compress_set_level(type
, level
);
1191 workspace
= get_workspace(type
, level
);
1192 ret
= compression_compress_pages(type
, workspace
, mapping
, start
, pages
,
1193 out_pages
, total_in
, total_out
);
1194 put_workspace(type
, workspace
);
1199 * pages_in is an array of pages with compressed data.
1201 * disk_start is the starting logical offset of this array in the file
1203 * orig_bio contains the pages from the file that we want to decompress into
1205 * srclen is the number of bytes in pages_in
1207 * The basic idea is that we have a bio that was created by readpages.
1208 * The pages in the bio are for the uncompressed data, and they may not
1209 * be contiguous. They all correspond to the range of bytes covered by
1210 * the compressed extent.
1212 static int btrfs_decompress_bio(struct compressed_bio
*cb
)
1214 struct list_head
*workspace
;
1216 int type
= cb
->compress_type
;
1218 workspace
= get_workspace(type
, 0);
1219 ret
= compression_decompress_bio(type
, workspace
, cb
);
1220 put_workspace(type
, workspace
);
1226 * a less complex decompression routine. Our compressed data fits in a
1227 * single page, and we want to read a single page out of it.
1228 * start_byte tells us the offset into the compressed data we're interested in
1230 int btrfs_decompress(int type
, unsigned char *data_in
, struct page
*dest_page
,
1231 unsigned long start_byte
, size_t srclen
, size_t destlen
)
1233 struct list_head
*workspace
;
1236 workspace
= get_workspace(type
, 0);
1237 ret
= compression_decompress(type
, workspace
, data_in
, dest_page
,
1238 start_byte
, srclen
, destlen
);
1239 put_workspace(type
, workspace
);
1244 void __init
btrfs_init_compress(void)
1246 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE
);
1247 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB
);
1248 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO
);
1249 zstd_init_workspace_manager();
1252 void __cold
btrfs_exit_compress(void)
1254 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE
);
1255 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB
);
1256 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO
);
1257 zstd_cleanup_workspace_manager();
1261 * Copy uncompressed data from working buffer to pages.
1263 * buf_start is the byte offset we're of the start of our workspace buffer.
1265 * total_out is the last byte of the buffer
1267 int btrfs_decompress_buf2page(const char *buf
, unsigned long buf_start
,
1268 unsigned long total_out
, u64 disk_start
,
1271 unsigned long buf_offset
;
1272 unsigned long current_buf_start
;
1273 unsigned long start_byte
;
1274 unsigned long prev_start_byte
;
1275 unsigned long working_bytes
= total_out
- buf_start
;
1276 unsigned long bytes
;
1278 struct bio_vec bvec
= bio_iter_iovec(bio
, bio
->bi_iter
);
1281 * start byte is the first byte of the page we're currently
1282 * copying into relative to the start of the compressed data.
1284 start_byte
= page_offset(bvec
.bv_page
) - disk_start
;
1286 /* we haven't yet hit data corresponding to this page */
1287 if (total_out
<= start_byte
)
1291 * the start of the data we care about is offset into
1292 * the middle of our working buffer
1294 if (total_out
> start_byte
&& buf_start
< start_byte
) {
1295 buf_offset
= start_byte
- buf_start
;
1296 working_bytes
-= buf_offset
;
1300 current_buf_start
= buf_start
;
1302 /* copy bytes from the working buffer into the pages */
1303 while (working_bytes
> 0) {
1304 bytes
= min_t(unsigned long, bvec
.bv_len
,
1305 PAGE_SIZE
- (buf_offset
% PAGE_SIZE
));
1306 bytes
= min(bytes
, working_bytes
);
1308 kaddr
= kmap_atomic(bvec
.bv_page
);
1309 memcpy(kaddr
+ bvec
.bv_offset
, buf
+ buf_offset
, bytes
);
1310 kunmap_atomic(kaddr
);
1311 flush_dcache_page(bvec
.bv_page
);
1313 buf_offset
+= bytes
;
1314 working_bytes
-= bytes
;
1315 current_buf_start
+= bytes
;
1317 /* check if we need to pick another page */
1318 bio_advance(bio
, bytes
);
1319 if (!bio
->bi_iter
.bi_size
)
1321 bvec
= bio_iter_iovec(bio
, bio
->bi_iter
);
1322 prev_start_byte
= start_byte
;
1323 start_byte
= page_offset(bvec
.bv_page
) - disk_start
;
1326 * We need to make sure we're only adjusting
1327 * our offset into compression working buffer when
1328 * we're switching pages. Otherwise we can incorrectly
1329 * keep copying when we were actually done.
1331 if (start_byte
!= prev_start_byte
) {
1333 * make sure our new page is covered by this
1336 if (total_out
<= start_byte
)
1340 * the next page in the biovec might not be adjacent
1341 * to the last page, but it might still be found
1342 * inside this working buffer. bump our offset pointer
1344 if (total_out
> start_byte
&&
1345 current_buf_start
< start_byte
) {
1346 buf_offset
= start_byte
- buf_start
;
1347 working_bytes
= total_out
- start_byte
;
1348 current_buf_start
= buf_start
+ buf_offset
;
1357 * Shannon Entropy calculation
1359 * Pure byte distribution analysis fails to determine compressibility of data.
1360 * Try calculating entropy to estimate the average minimum number of bits
1361 * needed to encode the sampled data.
1363 * For convenience, return the percentage of needed bits, instead of amount of
1366 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1367 * and can be compressible with high probability
1369 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1371 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1373 #define ENTROPY_LVL_ACEPTABLE (65)
1374 #define ENTROPY_LVL_HIGH (80)
1377 * For increasead precision in shannon_entropy calculation,
1378 * let's do pow(n, M) to save more digits after comma:
1380 * - maximum int bit length is 64
1381 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1382 * - 13 * 4 = 52 < 64 -> M = 4
1386 static inline u32
ilog2_w(u64 n
)
1388 return ilog2(n
* n
* n
* n
);
1391 static u32
shannon_entropy(struct heuristic_ws
*ws
)
1393 const u32 entropy_max
= 8 * ilog2_w(2);
1394 u32 entropy_sum
= 0;
1395 u32 p
, p_base
, sz_base
;
1398 sz_base
= ilog2_w(ws
->sample_size
);
1399 for (i
= 0; i
< BUCKET_SIZE
&& ws
->bucket
[i
].count
> 0; i
++) {
1400 p
= ws
->bucket
[i
].count
;
1401 p_base
= ilog2_w(p
);
1402 entropy_sum
+= p
* (sz_base
- p_base
);
1405 entropy_sum
/= ws
->sample_size
;
1406 return entropy_sum
* 100 / entropy_max
;
1409 #define RADIX_BASE 4U
1410 #define COUNTERS_SIZE (1U << RADIX_BASE)
1412 static u8
get4bits(u64 num
, int shift
) {
1417 low4bits
= (COUNTERS_SIZE
- 1) - (num
% COUNTERS_SIZE
);
1422 * Use 4 bits as radix base
1423 * Use 16 u32 counters for calculating new position in buf array
1425 * @array - array that will be sorted
1426 * @array_buf - buffer array to store sorting results
1427 * must be equal in size to @array
1430 static void radix_sort(struct bucket_item
*array
, struct bucket_item
*array_buf
,
1435 u32 counters
[COUNTERS_SIZE
];
1443 * Try avoid useless loop iterations for small numbers stored in big
1444 * counters. Example: 48 33 4 ... in 64bit array
1446 max_num
= array
[0].count
;
1447 for (i
= 1; i
< num
; i
++) {
1448 buf_num
= array
[i
].count
;
1449 if (buf_num
> max_num
)
1453 buf_num
= ilog2(max_num
);
1454 bitlen
= ALIGN(buf_num
, RADIX_BASE
* 2);
1457 while (shift
< bitlen
) {
1458 memset(counters
, 0, sizeof(counters
));
1460 for (i
= 0; i
< num
; i
++) {
1461 buf_num
= array
[i
].count
;
1462 addr
= get4bits(buf_num
, shift
);
1466 for (i
= 1; i
< COUNTERS_SIZE
; i
++)
1467 counters
[i
] += counters
[i
- 1];
1469 for (i
= num
- 1; i
>= 0; i
--) {
1470 buf_num
= array
[i
].count
;
1471 addr
= get4bits(buf_num
, shift
);
1473 new_addr
= counters
[addr
];
1474 array_buf
[new_addr
] = array
[i
];
1477 shift
+= RADIX_BASE
;
1480 * Normal radix expects to move data from a temporary array, to
1481 * the main one. But that requires some CPU time. Avoid that
1482 * by doing another sort iteration to original array instead of
1485 memset(counters
, 0, sizeof(counters
));
1487 for (i
= 0; i
< num
; i
++) {
1488 buf_num
= array_buf
[i
].count
;
1489 addr
= get4bits(buf_num
, shift
);
1493 for (i
= 1; i
< COUNTERS_SIZE
; i
++)
1494 counters
[i
] += counters
[i
- 1];
1496 for (i
= num
- 1; i
>= 0; i
--) {
1497 buf_num
= array_buf
[i
].count
;
1498 addr
= get4bits(buf_num
, shift
);
1500 new_addr
= counters
[addr
];
1501 array
[new_addr
] = array_buf
[i
];
1504 shift
+= RADIX_BASE
;
1509 * Size of the core byte set - how many bytes cover 90% of the sample
1511 * There are several types of structured binary data that use nearly all byte
1512 * values. The distribution can be uniform and counts in all buckets will be
1513 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1515 * Other possibility is normal (Gaussian) distribution, where the data could
1516 * be potentially compressible, but we have to take a few more steps to decide
1519 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1520 * compression algo can easy fix that
1521 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1522 * probability is not compressible
1524 #define BYTE_CORE_SET_LOW (64)
1525 #define BYTE_CORE_SET_HIGH (200)
1527 static int byte_core_set_size(struct heuristic_ws
*ws
)
1530 u32 coreset_sum
= 0;
1531 const u32 core_set_threshold
= ws
->sample_size
* 90 / 100;
1532 struct bucket_item
*bucket
= ws
->bucket
;
1534 /* Sort in reverse order */
1535 radix_sort(ws
->bucket
, ws
->bucket_b
, BUCKET_SIZE
);
1537 for (i
= 0; i
< BYTE_CORE_SET_LOW
; i
++)
1538 coreset_sum
+= bucket
[i
].count
;
1540 if (coreset_sum
> core_set_threshold
)
1543 for (; i
< BYTE_CORE_SET_HIGH
&& bucket
[i
].count
> 0; i
++) {
1544 coreset_sum
+= bucket
[i
].count
;
1545 if (coreset_sum
> core_set_threshold
)
1553 * Count byte values in buckets.
1554 * This heuristic can detect textual data (configs, xml, json, html, etc).
1555 * Because in most text-like data byte set is restricted to limited number of
1556 * possible characters, and that restriction in most cases makes data easy to
1559 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1560 * less - compressible
1561 * more - need additional analysis
1563 #define BYTE_SET_THRESHOLD (64)
1565 static u32
byte_set_size(const struct heuristic_ws
*ws
)
1568 u32 byte_set_size
= 0;
1570 for (i
= 0; i
< BYTE_SET_THRESHOLD
; i
++) {
1571 if (ws
->bucket
[i
].count
> 0)
1576 * Continue collecting count of byte values in buckets. If the byte
1577 * set size is bigger then the threshold, it's pointless to continue,
1578 * the detection technique would fail for this type of data.
1580 for (; i
< BUCKET_SIZE
; i
++) {
1581 if (ws
->bucket
[i
].count
> 0) {
1583 if (byte_set_size
> BYTE_SET_THRESHOLD
)
1584 return byte_set_size
;
1588 return byte_set_size
;
1591 static bool sample_repeated_patterns(struct heuristic_ws
*ws
)
1593 const u32 half_of_sample
= ws
->sample_size
/ 2;
1594 const u8
*data
= ws
->sample
;
1596 return memcmp(&data
[0], &data
[half_of_sample
], half_of_sample
) == 0;
1599 static void heuristic_collect_sample(struct inode
*inode
, u64 start
, u64 end
,
1600 struct heuristic_ws
*ws
)
1603 u64 index
, index_end
;
1604 u32 i
, curr_sample_pos
;
1608 * Compression handles the input data by chunks of 128KiB
1609 * (defined by BTRFS_MAX_UNCOMPRESSED)
1611 * We do the same for the heuristic and loop over the whole range.
1613 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1614 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1616 if (end
- start
> BTRFS_MAX_UNCOMPRESSED
)
1617 end
= start
+ BTRFS_MAX_UNCOMPRESSED
;
1619 index
= start
>> PAGE_SHIFT
;
1620 index_end
= end
>> PAGE_SHIFT
;
1622 /* Don't miss unaligned end */
1623 if (!IS_ALIGNED(end
, PAGE_SIZE
))
1626 curr_sample_pos
= 0;
1627 while (index
< index_end
) {
1628 page
= find_get_page(inode
->i_mapping
, index
);
1629 in_data
= kmap(page
);
1630 /* Handle case where the start is not aligned to PAGE_SIZE */
1631 i
= start
% PAGE_SIZE
;
1632 while (i
< PAGE_SIZE
- SAMPLING_READ_SIZE
) {
1633 /* Don't sample any garbage from the last page */
1634 if (start
> end
- SAMPLING_READ_SIZE
)
1636 memcpy(&ws
->sample
[curr_sample_pos
], &in_data
[i
],
1637 SAMPLING_READ_SIZE
);
1638 i
+= SAMPLING_INTERVAL
;
1639 start
+= SAMPLING_INTERVAL
;
1640 curr_sample_pos
+= SAMPLING_READ_SIZE
;
1648 ws
->sample_size
= curr_sample_pos
;
1652 * Compression heuristic.
1654 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1655 * quickly (compared to direct compression) detect data characteristics
1656 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1659 * The following types of analysis can be performed:
1660 * - detect mostly zero data
1661 * - detect data with low "byte set" size (text, etc)
1662 * - detect data with low/high "core byte" set
1664 * Return non-zero if the compression should be done, 0 otherwise.
1666 int btrfs_compress_heuristic(struct inode
*inode
, u64 start
, u64 end
)
1668 struct list_head
*ws_list
= get_workspace(0, 0);
1669 struct heuristic_ws
*ws
;
1674 ws
= list_entry(ws_list
, struct heuristic_ws
, list
);
1676 heuristic_collect_sample(inode
, start
, end
, ws
);
1678 if (sample_repeated_patterns(ws
)) {
1683 memset(ws
->bucket
, 0, sizeof(*ws
->bucket
)*BUCKET_SIZE
);
1685 for (i
= 0; i
< ws
->sample_size
; i
++) {
1686 byte
= ws
->sample
[i
];
1687 ws
->bucket
[byte
].count
++;
1690 i
= byte_set_size(ws
);
1691 if (i
< BYTE_SET_THRESHOLD
) {
1696 i
= byte_core_set_size(ws
);
1697 if (i
<= BYTE_CORE_SET_LOW
) {
1702 if (i
>= BYTE_CORE_SET_HIGH
) {
1707 i
= shannon_entropy(ws
);
1708 if (i
<= ENTROPY_LVL_ACEPTABLE
) {
1714 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1715 * needed to give green light to compression.
1717 * For now just assume that compression at that level is not worth the
1718 * resources because:
1720 * 1. it is possible to defrag the data later
1722 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1723 * values, every bucket has counter at level ~54. The heuristic would
1724 * be confused. This can happen when data have some internal repeated
1725 * patterns like "abbacbbc...". This can be detected by analyzing
1726 * pairs of bytes, which is too costly.
1728 if (i
< ENTROPY_LVL_HIGH
) {
1737 put_workspace(0, ws_list
);
1742 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1743 * level, unrecognized string will set the default level
1745 unsigned int btrfs_compress_str2level(unsigned int type
, const char *str
)
1747 unsigned int level
= 0;
1753 if (str
[0] == ':') {
1754 ret
= kstrtouint(str
+ 1, 10, &level
);
1759 level
= btrfs_compress_set_level(type
, level
);