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>
22 #include "transaction.h"
23 #include "btrfs_inode.h"
25 #include "ordered-data.h"
26 #include "compression.h"
27 #include "extent_io.h"
28 #include "extent_map.h"
30 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
32 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
35 case BTRFS_COMPRESS_ZLIB:
36 case BTRFS_COMPRESS_LZO:
37 case BTRFS_COMPRESS_ZSTD:
38 case BTRFS_COMPRESS_NONE:
39 return btrfs_compress_types[type];
45 static int btrfs_decompress_bio(struct compressed_bio *cb);
47 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
48 unsigned long disk_size)
50 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
52 return sizeof(struct compressed_bio) +
53 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
56 static int check_compressed_csum(struct btrfs_inode *inode,
57 struct compressed_bio *cb,
60 struct btrfs_fs_info *fs_info = inode->root->fs_info;
61 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
67 u8 *cb_sum = cb->sums;
69 if (inode->flags & BTRFS_INODE_NODATASUM)
72 for (i = 0; i < cb->nr_pages; i++) {
73 page = cb->compressed_pages[i];
76 kaddr = kmap_atomic(page);
77 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
78 btrfs_csum_final(csum, (u8 *)&csum);
81 if (memcmp(&csum, cb_sum, csum_size)) {
82 btrfs_print_data_csum_error(inode, disk_start, csum,
83 *(u32 *)cb_sum, cb->mirror_num);
95 /* when we finish reading compressed pages from the disk, we
96 * decompress them and then run the bio end_io routines on the
97 * decompressed pages (in the inode address space).
99 * This allows the checksumming and other IO error handling routines
102 * The compressed pages are freed here, and it must be run
105 static void end_compressed_bio_read(struct bio *bio)
107 struct compressed_bio *cb = bio->bi_private;
111 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
117 /* if there are more bios still pending for this compressed
120 if (!refcount_dec_and_test(&cb->pending_bios))
124 * Record the correct mirror_num in cb->orig_bio so that
125 * read-repair can work properly.
127 ASSERT(btrfs_io_bio(cb->orig_bio));
128 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
129 cb->mirror_num = mirror;
132 * Some IO in this cb have failed, just skip checksum as there
133 * is no way it could be correct.
139 ret = check_compressed_csum(BTRFS_I(inode), cb,
140 (u64)bio->bi_iter.bi_sector << 9);
144 /* ok, we're the last bio for this extent, lets start
147 ret = btrfs_decompress_bio(cb);
153 /* release the compressed pages */
155 for (index = 0; index < cb->nr_pages; index++) {
156 page = cb->compressed_pages[index];
157 page->mapping = NULL;
161 /* do io completion on the original bio */
163 bio_io_error(cb->orig_bio);
165 struct bio_vec *bvec;
166 struct bvec_iter_all iter_all;
169 * we have verified the checksum already, set page
170 * checked so the end_io handlers know about it
172 ASSERT(!bio_flagged(bio, BIO_CLONED));
173 bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
174 SetPageChecked(bvec->bv_page);
176 bio_endio(cb->orig_bio);
179 /* finally free the cb struct */
180 kfree(cb->compressed_pages);
187 * Clear the writeback bits on all of the file
188 * pages for a compressed write
190 static noinline void end_compressed_writeback(struct inode *inode,
191 const struct compressed_bio *cb)
193 unsigned long index = cb->start >> PAGE_SHIFT;
194 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
195 struct page *pages[16];
196 unsigned long nr_pages = end_index - index + 1;
201 mapping_set_error(inode->i_mapping, -EIO);
203 while (nr_pages > 0) {
204 ret = find_get_pages_contig(inode->i_mapping, index,
206 nr_pages, ARRAY_SIZE(pages)), pages);
212 for (i = 0; i < ret; i++) {
214 SetPageError(pages[i]);
215 end_page_writeback(pages[i]);
221 /* the inode may be gone now */
225 * do the cleanup once all the compressed pages hit the disk.
226 * This will clear writeback on the file pages and free the compressed
229 * This also calls the writeback end hooks for the file pages so that
230 * metadata and checksums can be updated in the file.
232 static void end_compressed_bio_write(struct bio *bio)
234 struct compressed_bio *cb = bio->bi_private;
242 /* if there are more bios still pending for this compressed
245 if (!refcount_dec_and_test(&cb->pending_bios))
248 /* ok, we're the last bio for this extent, step one is to
249 * call back into the FS and do all the end_io operations
252 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
253 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
254 cb->start, cb->start + cb->len - 1,
255 bio->bi_status == BLK_STS_OK);
256 cb->compressed_pages[0]->mapping = NULL;
258 end_compressed_writeback(inode, cb);
259 /* note, our inode could be gone now */
262 * release the compressed pages, these came from alloc_page and
263 * are not attached to the inode at all
266 for (index = 0; index < cb->nr_pages; index++) {
267 page = cb->compressed_pages[index];
268 page->mapping = NULL;
272 /* finally free the cb struct */
273 kfree(cb->compressed_pages);
280 * worker function to build and submit bios for previously compressed pages.
281 * The corresponding pages in the inode should be marked for writeback
282 * and the compressed pages should have a reference on them for dropping
283 * when the IO is complete.
285 * This also checksums the file bytes and gets things ready for
288 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
289 unsigned long len, u64 disk_start,
290 unsigned long compressed_len,
291 struct page **compressed_pages,
292 unsigned long nr_pages,
293 unsigned int write_flags)
295 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
296 struct bio *bio = NULL;
297 struct compressed_bio *cb;
298 unsigned long bytes_left;
301 u64 first_byte = disk_start;
302 struct block_device *bdev;
304 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
306 WARN_ON(!PAGE_ALIGNED(start));
307 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
309 return BLK_STS_RESOURCE;
310 refcount_set(&cb->pending_bios, 0);
316 cb->compressed_pages = compressed_pages;
317 cb->compressed_len = compressed_len;
319 cb->nr_pages = nr_pages;
321 bdev = fs_info->fs_devices->latest_bdev;
323 bio = btrfs_bio_alloc(bdev, first_byte);
324 bio->bi_opf = REQ_OP_WRITE | write_flags;
325 bio->bi_private = cb;
326 bio->bi_end_io = end_compressed_bio_write;
327 refcount_set(&cb->pending_bios, 1);
329 /* create and submit bios for the compressed pages */
330 bytes_left = compressed_len;
331 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
334 page = compressed_pages[pg_index];
335 page->mapping = inode->i_mapping;
336 if (bio->bi_iter.bi_size)
337 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
340 page->mapping = NULL;
341 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
344 * inc the count before we submit the bio so
345 * we know the end IO handler won't happen before
346 * we inc the count. Otherwise, the cb might get
347 * freed before we're done setting it up
349 refcount_inc(&cb->pending_bios);
350 ret = btrfs_bio_wq_end_io(fs_info, bio,
351 BTRFS_WQ_ENDIO_DATA);
352 BUG_ON(ret); /* -ENOMEM */
355 ret = btrfs_csum_one_bio(inode, bio, start, 1);
356 BUG_ON(ret); /* -ENOMEM */
359 ret = btrfs_map_bio(fs_info, bio, 0, 1);
361 bio->bi_status = ret;
365 bio = btrfs_bio_alloc(bdev, first_byte);
366 bio->bi_opf = REQ_OP_WRITE | write_flags;
367 bio->bi_private = cb;
368 bio->bi_end_io = end_compressed_bio_write;
369 bio_add_page(bio, page, PAGE_SIZE, 0);
371 if (bytes_left < PAGE_SIZE) {
373 "bytes left %lu compress len %lu nr %lu",
374 bytes_left, cb->compressed_len, cb->nr_pages);
376 bytes_left -= PAGE_SIZE;
377 first_byte += PAGE_SIZE;
381 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
382 BUG_ON(ret); /* -ENOMEM */
385 ret = btrfs_csum_one_bio(inode, bio, start, 1);
386 BUG_ON(ret); /* -ENOMEM */
389 ret = btrfs_map_bio(fs_info, bio, 0, 1);
391 bio->bi_status = ret;
398 static u64 bio_end_offset(struct bio *bio)
400 struct bio_vec *last = bio_last_bvec_all(bio);
402 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
405 static noinline int add_ra_bio_pages(struct inode *inode,
407 struct compressed_bio *cb)
409 unsigned long end_index;
410 unsigned long pg_index;
412 u64 isize = i_size_read(inode);
415 unsigned long nr_pages = 0;
416 struct extent_map *em;
417 struct address_space *mapping = inode->i_mapping;
418 struct extent_map_tree *em_tree;
419 struct extent_io_tree *tree;
423 last_offset = bio_end_offset(cb->orig_bio);
424 em_tree = &BTRFS_I(inode)->extent_tree;
425 tree = &BTRFS_I(inode)->io_tree;
430 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
432 while (last_offset < compressed_end) {
433 pg_index = last_offset >> PAGE_SHIFT;
435 if (pg_index > end_index)
438 page = xa_load(&mapping->i_pages, pg_index);
439 if (page && !xa_is_value(page)) {
446 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
451 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
456 end = last_offset + PAGE_SIZE - 1;
458 * at this point, we have a locked page in the page cache
459 * for these bytes in the file. But, we have to make
460 * sure they map to this compressed extent on disk.
462 set_page_extent_mapped(page);
463 lock_extent(tree, last_offset, end);
464 read_lock(&em_tree->lock);
465 em = lookup_extent_mapping(em_tree, last_offset,
467 read_unlock(&em_tree->lock);
469 if (!em || last_offset < em->start ||
470 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
471 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
473 unlock_extent(tree, last_offset, end);
480 if (page->index == end_index) {
482 size_t zero_offset = offset_in_page(isize);
486 zeros = PAGE_SIZE - zero_offset;
487 userpage = kmap_atomic(page);
488 memset(userpage + zero_offset, 0, zeros);
489 flush_dcache_page(page);
490 kunmap_atomic(userpage);
494 ret = bio_add_page(cb->orig_bio, page,
497 if (ret == PAGE_SIZE) {
501 unlock_extent(tree, last_offset, end);
507 last_offset += PAGE_SIZE;
513 * for a compressed read, the bio we get passed has all the inode pages
514 * in it. We don't actually do IO on those pages but allocate new ones
515 * to hold the compressed pages on disk.
517 * bio->bi_iter.bi_sector points to the compressed extent on disk
518 * bio->bi_io_vec points to all of the inode pages
520 * After the compressed pages are read, we copy the bytes into the
521 * bio we were passed and then call the bio end_io calls
523 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
524 int mirror_num, unsigned long bio_flags)
526 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
527 struct extent_map_tree *em_tree;
528 struct compressed_bio *cb;
529 unsigned long compressed_len;
530 unsigned long nr_pages;
531 unsigned long pg_index;
533 struct block_device *bdev;
534 struct bio *comp_bio;
535 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
538 struct extent_map *em;
539 blk_status_t ret = BLK_STS_RESOURCE;
541 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
544 em_tree = &BTRFS_I(inode)->extent_tree;
546 /* we need the actual starting offset of this extent in the file */
547 read_lock(&em_tree->lock);
548 em = lookup_extent_mapping(em_tree,
549 page_offset(bio_first_page_all(bio)),
551 read_unlock(&em_tree->lock);
553 return BLK_STS_IOERR;
555 compressed_len = em->block_len;
556 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
560 refcount_set(&cb->pending_bios, 0);
563 cb->mirror_num = mirror_num;
566 cb->start = em->orig_start;
568 em_start = em->start;
573 cb->len = bio->bi_iter.bi_size;
574 cb->compressed_len = compressed_len;
575 cb->compress_type = extent_compress_type(bio_flags);
578 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
579 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
581 if (!cb->compressed_pages)
584 bdev = fs_info->fs_devices->latest_bdev;
586 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
587 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
589 if (!cb->compressed_pages[pg_index]) {
590 faili = pg_index - 1;
591 ret = BLK_STS_RESOURCE;
595 faili = nr_pages - 1;
596 cb->nr_pages = nr_pages;
598 add_ra_bio_pages(inode, em_start + em_len, cb);
600 /* include any pages we added in add_ra-bio_pages */
601 cb->len = bio->bi_iter.bi_size;
603 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
604 comp_bio->bi_opf = REQ_OP_READ;
605 comp_bio->bi_private = cb;
606 comp_bio->bi_end_io = end_compressed_bio_read;
607 refcount_set(&cb->pending_bios, 1);
609 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
612 page = cb->compressed_pages[pg_index];
613 page->mapping = inode->i_mapping;
614 page->index = em_start >> PAGE_SHIFT;
616 if (comp_bio->bi_iter.bi_size)
617 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
620 page->mapping = NULL;
621 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
623 unsigned int nr_sectors;
625 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
626 BTRFS_WQ_ENDIO_DATA);
627 BUG_ON(ret); /* -ENOMEM */
630 * inc the count before we submit the bio so
631 * we know the end IO handler won't happen before
632 * we inc the count. Otherwise, the cb might get
633 * freed before we're done setting it up
635 refcount_inc(&cb->pending_bios);
637 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
638 ret = btrfs_lookup_bio_sums(inode, comp_bio,
640 BUG_ON(ret); /* -ENOMEM */
643 nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
644 fs_info->sectorsize);
645 sums += csum_size * nr_sectors;
647 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
649 comp_bio->bi_status = ret;
653 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
654 comp_bio->bi_opf = REQ_OP_READ;
655 comp_bio->bi_private = cb;
656 comp_bio->bi_end_io = end_compressed_bio_read;
658 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
660 cur_disk_byte += PAGE_SIZE;
663 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
664 BUG_ON(ret); /* -ENOMEM */
666 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
667 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
668 BUG_ON(ret); /* -ENOMEM */
671 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
673 comp_bio->bi_status = ret;
681 __free_page(cb->compressed_pages[faili]);
685 kfree(cb->compressed_pages);
694 * Heuristic uses systematic sampling to collect data from the input data
695 * range, the logic can be tuned by the following constants:
697 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
698 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
700 #define SAMPLING_READ_SIZE (16)
701 #define SAMPLING_INTERVAL (256)
704 * For statistical analysis of the input data we consider bytes that form a
705 * Galois Field of 256 objects. Each object has an attribute count, ie. how
706 * many times the object appeared in the sample.
708 #define BUCKET_SIZE (256)
711 * The size of the sample is based on a statistical sampling rule of thumb.
712 * The common way is to perform sampling tests as long as the number of
713 * elements in each cell is at least 5.
715 * Instead of 5, we choose 32 to obtain more accurate results.
716 * If the data contain the maximum number of symbols, which is 256, we obtain a
717 * sample size bound by 8192.
719 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
720 * from up to 512 locations.
722 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
723 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
729 struct heuristic_ws {
730 /* Partial copy of input data */
733 /* Buckets store counters for each byte value */
734 struct bucket_item *bucket;
736 struct bucket_item *bucket_b;
737 struct list_head list;
740 static struct workspace_manager heuristic_wsm;
742 static void heuristic_init_workspace_manager(void)
744 btrfs_init_workspace_manager(&heuristic_wsm, &btrfs_heuristic_compress);
747 static void heuristic_cleanup_workspace_manager(void)
749 btrfs_cleanup_workspace_manager(&heuristic_wsm);
752 static struct list_head *heuristic_get_workspace(unsigned int level)
754 return btrfs_get_workspace(&heuristic_wsm, level);
757 static void heuristic_put_workspace(struct list_head *ws)
759 btrfs_put_workspace(&heuristic_wsm, ws);
762 static void free_heuristic_ws(struct list_head *ws)
764 struct heuristic_ws *workspace;
766 workspace = list_entry(ws, struct heuristic_ws, list);
768 kvfree(workspace->sample);
769 kfree(workspace->bucket);
770 kfree(workspace->bucket_b);
774 static struct list_head *alloc_heuristic_ws(unsigned int level)
776 struct heuristic_ws *ws;
778 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
780 return ERR_PTR(-ENOMEM);
782 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
786 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
790 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
794 INIT_LIST_HEAD(&ws->list);
797 free_heuristic_ws(&ws->list);
798 return ERR_PTR(-ENOMEM);
801 const struct btrfs_compress_op btrfs_heuristic_compress = {
802 .init_workspace_manager = heuristic_init_workspace_manager,
803 .cleanup_workspace_manager = heuristic_cleanup_workspace_manager,
804 .get_workspace = heuristic_get_workspace,
805 .put_workspace = heuristic_put_workspace,
806 .alloc_workspace = alloc_heuristic_ws,
807 .free_workspace = free_heuristic_ws,
810 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
811 /* The heuristic is represented as compression type 0 */
812 &btrfs_heuristic_compress,
813 &btrfs_zlib_compress,
815 &btrfs_zstd_compress,
818 void btrfs_init_workspace_manager(struct workspace_manager *wsm,
819 const struct btrfs_compress_op *ops)
821 struct list_head *workspace;
825 INIT_LIST_HEAD(&wsm->idle_ws);
826 spin_lock_init(&wsm->ws_lock);
827 atomic_set(&wsm->total_ws, 0);
828 init_waitqueue_head(&wsm->ws_wait);
831 * Preallocate one workspace for each compression type so we can
832 * guarantee forward progress in the worst case
834 workspace = wsm->ops->alloc_workspace(0);
835 if (IS_ERR(workspace)) {
837 "BTRFS: cannot preallocate compression workspace, will try later\n");
839 atomic_set(&wsm->total_ws, 1);
841 list_add(workspace, &wsm->idle_ws);
845 void btrfs_cleanup_workspace_manager(struct workspace_manager *wsman)
847 struct list_head *ws;
849 while (!list_empty(&wsman->idle_ws)) {
850 ws = wsman->idle_ws.next;
852 wsman->ops->free_workspace(ws);
853 atomic_dec(&wsman->total_ws);
858 * This finds an available workspace or allocates a new one.
859 * If it's not possible to allocate a new one, waits until there's one.
860 * Preallocation makes a forward progress guarantees and we do not return
863 struct list_head *btrfs_get_workspace(struct workspace_manager *wsm,
866 struct list_head *workspace;
867 int cpus = num_online_cpus();
869 struct list_head *idle_ws;
872 wait_queue_head_t *ws_wait;
875 idle_ws = &wsm->idle_ws;
876 ws_lock = &wsm->ws_lock;
877 total_ws = &wsm->total_ws;
878 ws_wait = &wsm->ws_wait;
879 free_ws = &wsm->free_ws;
883 if (!list_empty(idle_ws)) {
884 workspace = idle_ws->next;
887 spin_unlock(ws_lock);
891 if (atomic_read(total_ws) > cpus) {
894 spin_unlock(ws_lock);
895 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
896 if (atomic_read(total_ws) > cpus && !*free_ws)
898 finish_wait(ws_wait, &wait);
901 atomic_inc(total_ws);
902 spin_unlock(ws_lock);
905 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
906 * to turn it off here because we might get called from the restricted
907 * context of btrfs_compress_bio/btrfs_compress_pages
909 nofs_flag = memalloc_nofs_save();
910 workspace = wsm->ops->alloc_workspace(level);
911 memalloc_nofs_restore(nofs_flag);
913 if (IS_ERR(workspace)) {
914 atomic_dec(total_ws);
918 * Do not return the error but go back to waiting. There's a
919 * workspace preallocated for each type and the compression
920 * time is bounded so we get to a workspace eventually. This
921 * makes our caller's life easier.
923 * To prevent silent and low-probability deadlocks (when the
924 * initial preallocation fails), check if there are any
927 if (atomic_read(total_ws) == 0) {
928 static DEFINE_RATELIMIT_STATE(_rs,
929 /* once per minute */ 60 * HZ,
932 if (__ratelimit(&_rs)) {
933 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
941 static struct list_head *get_workspace(int type, int level)
943 return btrfs_compress_op[type]->get_workspace(level);
947 * put a workspace struct back on the list or free it if we have enough
948 * idle ones sitting around
950 void btrfs_put_workspace(struct workspace_manager *wsm, struct list_head *ws)
952 struct list_head *idle_ws;
955 wait_queue_head_t *ws_wait;
958 idle_ws = &wsm->idle_ws;
959 ws_lock = &wsm->ws_lock;
960 total_ws = &wsm->total_ws;
961 ws_wait = &wsm->ws_wait;
962 free_ws = &wsm->free_ws;
965 if (*free_ws <= num_online_cpus()) {
966 list_add(ws, idle_ws);
968 spin_unlock(ws_lock);
971 spin_unlock(ws_lock);
973 wsm->ops->free_workspace(ws);
974 atomic_dec(total_ws);
976 cond_wake_up(ws_wait);
979 static void put_workspace(int type, struct list_head *ws)
981 return btrfs_compress_op[type]->put_workspace(ws);
985 * Given an address space and start and length, compress the bytes into @pages
986 * that are allocated on demand.
988 * @type_level is encoded algorithm and level, where level 0 means whatever
989 * default the algorithm chooses and is opaque here;
990 * - compression algo are 0-3
991 * - the level are bits 4-7
993 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
994 * and returns number of actually allocated pages
996 * @total_in is used to return the number of bytes actually read. It
997 * may be smaller than the input length if we had to exit early because we
998 * ran out of room in the pages array or because we cross the
1001 * @total_out is an in/out parameter, must be set to the input length and will
1002 * be also used to return the total number of compressed bytes
1004 * @max_out tells us the max number of bytes that we're allowed to
1007 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1008 u64 start, struct page **pages,
1009 unsigned long *out_pages,
1010 unsigned long *total_in,
1011 unsigned long *total_out)
1013 int type = btrfs_compress_type(type_level);
1014 int level = btrfs_compress_level(type_level);
1015 struct list_head *workspace;
1018 level = btrfs_compress_op[type]->set_level(level);
1019 workspace = get_workspace(type, level);
1020 ret = btrfs_compress_op[type]->compress_pages(workspace, mapping,
1023 total_in, total_out);
1024 put_workspace(type, workspace);
1029 * pages_in is an array of pages with compressed data.
1031 * disk_start is the starting logical offset of this array in the file
1033 * orig_bio contains the pages from the file that we want to decompress into
1035 * srclen is the number of bytes in pages_in
1037 * The basic idea is that we have a bio that was created by readpages.
1038 * The pages in the bio are for the uncompressed data, and they may not
1039 * be contiguous. They all correspond to the range of bytes covered by
1040 * the compressed extent.
1042 static int btrfs_decompress_bio(struct compressed_bio *cb)
1044 struct list_head *workspace;
1046 int type = cb->compress_type;
1048 workspace = get_workspace(type, 0);
1049 ret = btrfs_compress_op[type]->decompress_bio(workspace, cb);
1050 put_workspace(type, workspace);
1056 * a less complex decompression routine. Our compressed data fits in a
1057 * single page, and we want to read a single page out of it.
1058 * start_byte tells us the offset into the compressed data we're interested in
1060 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1061 unsigned long start_byte, size_t srclen, size_t destlen)
1063 struct list_head *workspace;
1066 workspace = get_workspace(type, 0);
1067 ret = btrfs_compress_op[type]->decompress(workspace, data_in,
1068 dest_page, start_byte,
1070 put_workspace(type, workspace);
1075 void __init btrfs_init_compress(void)
1079 for (i = 0; i < BTRFS_NR_WORKSPACE_MANAGERS; i++)
1080 btrfs_compress_op[i]->init_workspace_manager();
1083 void __cold btrfs_exit_compress(void)
1087 for (i = 0; i < BTRFS_NR_WORKSPACE_MANAGERS; i++)
1088 btrfs_compress_op[i]->cleanup_workspace_manager();
1092 * Copy uncompressed data from working buffer to pages.
1094 * buf_start is the byte offset we're of the start of our workspace buffer.
1096 * total_out is the last byte of the buffer
1098 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1099 unsigned long total_out, u64 disk_start,
1102 unsigned long buf_offset;
1103 unsigned long current_buf_start;
1104 unsigned long start_byte;
1105 unsigned long prev_start_byte;
1106 unsigned long working_bytes = total_out - buf_start;
1107 unsigned long bytes;
1109 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1112 * start byte is the first byte of the page we're currently
1113 * copying into relative to the start of the compressed data.
1115 start_byte = page_offset(bvec.bv_page) - disk_start;
1117 /* we haven't yet hit data corresponding to this page */
1118 if (total_out <= start_byte)
1122 * the start of the data we care about is offset into
1123 * the middle of our working buffer
1125 if (total_out > start_byte && buf_start < start_byte) {
1126 buf_offset = start_byte - buf_start;
1127 working_bytes -= buf_offset;
1131 current_buf_start = buf_start;
1133 /* copy bytes from the working buffer into the pages */
1134 while (working_bytes > 0) {
1135 bytes = min_t(unsigned long, bvec.bv_len,
1136 PAGE_SIZE - buf_offset);
1137 bytes = min(bytes, working_bytes);
1139 kaddr = kmap_atomic(bvec.bv_page);
1140 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1141 kunmap_atomic(kaddr);
1142 flush_dcache_page(bvec.bv_page);
1144 buf_offset += bytes;
1145 working_bytes -= bytes;
1146 current_buf_start += bytes;
1148 /* check if we need to pick another page */
1149 bio_advance(bio, bytes);
1150 if (!bio->bi_iter.bi_size)
1152 bvec = bio_iter_iovec(bio, bio->bi_iter);
1153 prev_start_byte = start_byte;
1154 start_byte = page_offset(bvec.bv_page) - disk_start;
1157 * We need to make sure we're only adjusting
1158 * our offset into compression working buffer when
1159 * we're switching pages. Otherwise we can incorrectly
1160 * keep copying when we were actually done.
1162 if (start_byte != prev_start_byte) {
1164 * make sure our new page is covered by this
1167 if (total_out <= start_byte)
1171 * the next page in the biovec might not be adjacent
1172 * to the last page, but it might still be found
1173 * inside this working buffer. bump our offset pointer
1175 if (total_out > start_byte &&
1176 current_buf_start < start_byte) {
1177 buf_offset = start_byte - buf_start;
1178 working_bytes = total_out - start_byte;
1179 current_buf_start = buf_start + buf_offset;
1188 * Shannon Entropy calculation
1190 * Pure byte distribution analysis fails to determine compressibility of data.
1191 * Try calculating entropy to estimate the average minimum number of bits
1192 * needed to encode the sampled data.
1194 * For convenience, return the percentage of needed bits, instead of amount of
1197 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1198 * and can be compressible with high probability
1200 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1202 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1204 #define ENTROPY_LVL_ACEPTABLE (65)
1205 #define ENTROPY_LVL_HIGH (80)
1208 * For increasead precision in shannon_entropy calculation,
1209 * let's do pow(n, M) to save more digits after comma:
1211 * - maximum int bit length is 64
1212 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1213 * - 13 * 4 = 52 < 64 -> M = 4
1217 static inline u32 ilog2_w(u64 n)
1219 return ilog2(n * n * n * n);
1222 static u32 shannon_entropy(struct heuristic_ws *ws)
1224 const u32 entropy_max = 8 * ilog2_w(2);
1225 u32 entropy_sum = 0;
1226 u32 p, p_base, sz_base;
1229 sz_base = ilog2_w(ws->sample_size);
1230 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1231 p = ws->bucket[i].count;
1232 p_base = ilog2_w(p);
1233 entropy_sum += p * (sz_base - p_base);
1236 entropy_sum /= ws->sample_size;
1237 return entropy_sum * 100 / entropy_max;
1240 #define RADIX_BASE 4U
1241 #define COUNTERS_SIZE (1U << RADIX_BASE)
1243 static u8 get4bits(u64 num, int shift) {
1248 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1253 * Use 4 bits as radix base
1254 * Use 16 u32 counters for calculating new position in buf array
1256 * @array - array that will be sorted
1257 * @array_buf - buffer array to store sorting results
1258 * must be equal in size to @array
1261 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1266 u32 counters[COUNTERS_SIZE];
1274 * Try avoid useless loop iterations for small numbers stored in big
1275 * counters. Example: 48 33 4 ... in 64bit array
1277 max_num = array[0].count;
1278 for (i = 1; i < num; i++) {
1279 buf_num = array[i].count;
1280 if (buf_num > max_num)
1284 buf_num = ilog2(max_num);
1285 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1288 while (shift < bitlen) {
1289 memset(counters, 0, sizeof(counters));
1291 for (i = 0; i < num; i++) {
1292 buf_num = array[i].count;
1293 addr = get4bits(buf_num, shift);
1297 for (i = 1; i < COUNTERS_SIZE; i++)
1298 counters[i] += counters[i - 1];
1300 for (i = num - 1; i >= 0; i--) {
1301 buf_num = array[i].count;
1302 addr = get4bits(buf_num, shift);
1304 new_addr = counters[addr];
1305 array_buf[new_addr] = array[i];
1308 shift += RADIX_BASE;
1311 * Normal radix expects to move data from a temporary array, to
1312 * the main one. But that requires some CPU time. Avoid that
1313 * by doing another sort iteration to original array instead of
1316 memset(counters, 0, sizeof(counters));
1318 for (i = 0; i < num; i ++) {
1319 buf_num = array_buf[i].count;
1320 addr = get4bits(buf_num, shift);
1324 for (i = 1; i < COUNTERS_SIZE; i++)
1325 counters[i] += counters[i - 1];
1327 for (i = num - 1; i >= 0; i--) {
1328 buf_num = array_buf[i].count;
1329 addr = get4bits(buf_num, shift);
1331 new_addr = counters[addr];
1332 array[new_addr] = array_buf[i];
1335 shift += RADIX_BASE;
1340 * Size of the core byte set - how many bytes cover 90% of the sample
1342 * There are several types of structured binary data that use nearly all byte
1343 * values. The distribution can be uniform and counts in all buckets will be
1344 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1346 * Other possibility is normal (Gaussian) distribution, where the data could
1347 * be potentially compressible, but we have to take a few more steps to decide
1350 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1351 * compression algo can easy fix that
1352 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1353 * probability is not compressible
1355 #define BYTE_CORE_SET_LOW (64)
1356 #define BYTE_CORE_SET_HIGH (200)
1358 static int byte_core_set_size(struct heuristic_ws *ws)
1361 u32 coreset_sum = 0;
1362 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1363 struct bucket_item *bucket = ws->bucket;
1365 /* Sort in reverse order */
1366 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1368 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1369 coreset_sum += bucket[i].count;
1371 if (coreset_sum > core_set_threshold)
1374 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1375 coreset_sum += bucket[i].count;
1376 if (coreset_sum > core_set_threshold)
1384 * Count byte values in buckets.
1385 * This heuristic can detect textual data (configs, xml, json, html, etc).
1386 * Because in most text-like data byte set is restricted to limited number of
1387 * possible characters, and that restriction in most cases makes data easy to
1390 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1391 * less - compressible
1392 * more - need additional analysis
1394 #define BYTE_SET_THRESHOLD (64)
1396 static u32 byte_set_size(const struct heuristic_ws *ws)
1399 u32 byte_set_size = 0;
1401 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1402 if (ws->bucket[i].count > 0)
1407 * Continue collecting count of byte values in buckets. If the byte
1408 * set size is bigger then the threshold, it's pointless to continue,
1409 * the detection technique would fail for this type of data.
1411 for (; i < BUCKET_SIZE; i++) {
1412 if (ws->bucket[i].count > 0) {
1414 if (byte_set_size > BYTE_SET_THRESHOLD)
1415 return byte_set_size;
1419 return byte_set_size;
1422 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1424 const u32 half_of_sample = ws->sample_size / 2;
1425 const u8 *data = ws->sample;
1427 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1430 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1431 struct heuristic_ws *ws)
1434 u64 index, index_end;
1435 u32 i, curr_sample_pos;
1439 * Compression handles the input data by chunks of 128KiB
1440 * (defined by BTRFS_MAX_UNCOMPRESSED)
1442 * We do the same for the heuristic and loop over the whole range.
1444 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1445 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1447 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1448 end = start + BTRFS_MAX_UNCOMPRESSED;
1450 index = start >> PAGE_SHIFT;
1451 index_end = end >> PAGE_SHIFT;
1453 /* Don't miss unaligned end */
1454 if (!IS_ALIGNED(end, PAGE_SIZE))
1457 curr_sample_pos = 0;
1458 while (index < index_end) {
1459 page = find_get_page(inode->i_mapping, index);
1460 in_data = kmap(page);
1461 /* Handle case where the start is not aligned to PAGE_SIZE */
1462 i = start % PAGE_SIZE;
1463 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1464 /* Don't sample any garbage from the last page */
1465 if (start > end - SAMPLING_READ_SIZE)
1467 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1468 SAMPLING_READ_SIZE);
1469 i += SAMPLING_INTERVAL;
1470 start += SAMPLING_INTERVAL;
1471 curr_sample_pos += SAMPLING_READ_SIZE;
1479 ws->sample_size = curr_sample_pos;
1483 * Compression heuristic.
1485 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1486 * quickly (compared to direct compression) detect data characteristics
1487 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1490 * The following types of analysis can be performed:
1491 * - detect mostly zero data
1492 * - detect data with low "byte set" size (text, etc)
1493 * - detect data with low/high "core byte" set
1495 * Return non-zero if the compression should be done, 0 otherwise.
1497 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1499 struct list_head *ws_list = get_workspace(0, 0);
1500 struct heuristic_ws *ws;
1505 ws = list_entry(ws_list, struct heuristic_ws, list);
1507 heuristic_collect_sample(inode, start, end, ws);
1509 if (sample_repeated_patterns(ws)) {
1514 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1516 for (i = 0; i < ws->sample_size; i++) {
1517 byte = ws->sample[i];
1518 ws->bucket[byte].count++;
1521 i = byte_set_size(ws);
1522 if (i < BYTE_SET_THRESHOLD) {
1527 i = byte_core_set_size(ws);
1528 if (i <= BYTE_CORE_SET_LOW) {
1533 if (i >= BYTE_CORE_SET_HIGH) {
1538 i = shannon_entropy(ws);
1539 if (i <= ENTROPY_LVL_ACEPTABLE) {
1545 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1546 * needed to give green light to compression.
1548 * For now just assume that compression at that level is not worth the
1549 * resources because:
1551 * 1. it is possible to defrag the data later
1553 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1554 * values, every bucket has counter at level ~54. The heuristic would
1555 * be confused. This can happen when data have some internal repeated
1556 * patterns like "abbacbbc...". This can be detected by analyzing
1557 * pairs of bytes, which is too costly.
1559 if (i < ENTROPY_LVL_HIGH) {
1568 put_workspace(0, ws_list);
1573 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1574 * level, unrecognized string will set the default level
1576 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1578 unsigned int level = 0;
1584 if (str[0] == ':') {
1585 ret = kstrtouint(str + 1, 10, &level);
1590 level = btrfs_compress_op[type]->set_level(level);