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,
65 u32 *cb_sum = &cb->sums;
67 if (inode->flags & BTRFS_INODE_NODATASUM)
70 for (i = 0; i < cb->nr_pages; i++) {
71 page = cb->compressed_pages[i];
74 kaddr = kmap_atomic(page);
75 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
76 btrfs_csum_final(csum, (u8 *)&csum);
79 if (csum != *cb_sum) {
80 btrfs_print_data_csum_error(inode, disk_start, csum,
81 *cb_sum, cb->mirror_num);
93 /* when we finish reading compressed pages from the disk, we
94 * decompress them and then run the bio end_io routines on the
95 * decompressed pages (in the inode address space).
97 * This allows the checksumming and other IO error handling routines
100 * The compressed pages are freed here, and it must be run
103 static void end_compressed_bio_read(struct bio *bio)
105 struct compressed_bio *cb = bio->bi_private;
109 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
115 /* if there are more bios still pending for this compressed
118 if (!refcount_dec_and_test(&cb->pending_bios))
122 * Record the correct mirror_num in cb->orig_bio so that
123 * read-repair can work properly.
125 ASSERT(btrfs_io_bio(cb->orig_bio));
126 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
127 cb->mirror_num = mirror;
130 * Some IO in this cb have failed, just skip checksum as there
131 * is no way it could be correct.
137 ret = check_compressed_csum(BTRFS_I(inode), cb,
138 (u64)bio->bi_iter.bi_sector << 9);
142 /* ok, we're the last bio for this extent, lets start
145 ret = btrfs_decompress_bio(cb);
151 /* release the compressed pages */
153 for (index = 0; index < cb->nr_pages; index++) {
154 page = cb->compressed_pages[index];
155 page->mapping = NULL;
159 /* do io completion on the original bio */
161 bio_io_error(cb->orig_bio);
164 struct bio_vec *bvec;
167 * we have verified the checksum already, set page
168 * checked so the end_io handlers know about it
170 ASSERT(!bio_flagged(bio, BIO_CLONED));
171 bio_for_each_segment_all(bvec, cb->orig_bio, i)
172 SetPageChecked(bvec->bv_page);
174 bio_endio(cb->orig_bio);
177 /* finally free the cb struct */
178 kfree(cb->compressed_pages);
185 * Clear the writeback bits on all of the file
186 * pages for a compressed write
188 static noinline void end_compressed_writeback(struct inode *inode,
189 const struct compressed_bio *cb)
191 unsigned long index = cb->start >> PAGE_SHIFT;
192 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
193 struct page *pages[16];
194 unsigned long nr_pages = end_index - index + 1;
199 mapping_set_error(inode->i_mapping, -EIO);
201 while (nr_pages > 0) {
202 ret = find_get_pages_contig(inode->i_mapping, index,
204 nr_pages, ARRAY_SIZE(pages)), pages);
210 for (i = 0; i < ret; i++) {
212 SetPageError(pages[i]);
213 end_page_writeback(pages[i]);
219 /* the inode may be gone now */
223 * do the cleanup once all the compressed pages hit the disk.
224 * This will clear writeback on the file pages and free the compressed
227 * This also calls the writeback end hooks for the file pages so that
228 * metadata and checksums can be updated in the file.
230 static void end_compressed_bio_write(struct bio *bio)
232 struct extent_io_tree *tree;
233 struct compressed_bio *cb = bio->bi_private;
241 /* if there are more bios still pending for this compressed
244 if (!refcount_dec_and_test(&cb->pending_bios))
247 /* ok, we're the last bio for this extent, step one is to
248 * call back into the FS and do all the end_io operations
251 tree = &BTRFS_I(inode)->io_tree;
252 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
253 tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
255 cb->start + cb->len - 1,
258 BLK_STS_OK : BLK_STS_NOTSUPP);
259 cb->compressed_pages[0]->mapping = NULL;
261 end_compressed_writeback(inode, cb);
262 /* note, our inode could be gone now */
265 * release the compressed pages, these came from alloc_page and
266 * are not attached to the inode at all
269 for (index = 0; index < cb->nr_pages; index++) {
270 page = cb->compressed_pages[index];
271 page->mapping = NULL;
275 /* finally free the cb struct */
276 kfree(cb->compressed_pages);
283 * worker function to build and submit bios for previously compressed pages.
284 * The corresponding pages in the inode should be marked for writeback
285 * and the compressed pages should have a reference on them for dropping
286 * when the IO is complete.
288 * This also checksums the file bytes and gets things ready for
291 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
292 unsigned long len, u64 disk_start,
293 unsigned long compressed_len,
294 struct page **compressed_pages,
295 unsigned long nr_pages,
296 unsigned int write_flags)
298 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
299 struct bio *bio = NULL;
300 struct compressed_bio *cb;
301 unsigned long bytes_left;
304 u64 first_byte = disk_start;
305 struct block_device *bdev;
307 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
309 WARN_ON(start & ((u64)PAGE_SIZE - 1));
310 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
312 return BLK_STS_RESOURCE;
313 refcount_set(&cb->pending_bios, 0);
319 cb->compressed_pages = compressed_pages;
320 cb->compressed_len = compressed_len;
322 cb->nr_pages = nr_pages;
324 bdev = fs_info->fs_devices->latest_bdev;
326 bio = btrfs_bio_alloc(bdev, first_byte);
327 bio->bi_opf = REQ_OP_WRITE | write_flags;
328 bio->bi_private = cb;
329 bio->bi_end_io = end_compressed_bio_write;
330 refcount_set(&cb->pending_bios, 1);
332 /* create and submit bios for the compressed pages */
333 bytes_left = compressed_len;
334 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
337 page = compressed_pages[pg_index];
338 page->mapping = inode->i_mapping;
339 if (bio->bi_iter.bi_size)
340 submit = btrfs_merge_bio_hook(page, 0, PAGE_SIZE, bio, 0);
342 page->mapping = NULL;
343 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
346 * inc the count before we submit the bio so
347 * we know the end IO handler won't happen before
348 * we inc the count. Otherwise, the cb might get
349 * freed before we're done setting it up
351 refcount_inc(&cb->pending_bios);
352 ret = btrfs_bio_wq_end_io(fs_info, bio,
353 BTRFS_WQ_ENDIO_DATA);
354 BUG_ON(ret); /* -ENOMEM */
357 ret = btrfs_csum_one_bio(inode, bio, start, 1);
358 BUG_ON(ret); /* -ENOMEM */
361 ret = btrfs_map_bio(fs_info, bio, 0, 1);
363 bio->bi_status = ret;
367 bio = btrfs_bio_alloc(bdev, first_byte);
368 bio->bi_opf = REQ_OP_WRITE | write_flags;
369 bio->bi_private = cb;
370 bio->bi_end_io = end_compressed_bio_write;
371 bio_add_page(bio, page, PAGE_SIZE, 0);
373 if (bytes_left < PAGE_SIZE) {
375 "bytes left %lu compress len %lu nr %lu",
376 bytes_left, cb->compressed_len, cb->nr_pages);
378 bytes_left -= PAGE_SIZE;
379 first_byte += PAGE_SIZE;
383 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
384 BUG_ON(ret); /* -ENOMEM */
387 ret = btrfs_csum_one_bio(inode, bio, start, 1);
388 BUG_ON(ret); /* -ENOMEM */
391 ret = btrfs_map_bio(fs_info, bio, 0, 1);
393 bio->bi_status = ret;
400 static u64 bio_end_offset(struct bio *bio)
402 struct bio_vec *last = bio_last_bvec_all(bio);
404 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
407 static noinline int add_ra_bio_pages(struct inode *inode,
409 struct compressed_bio *cb)
411 unsigned long end_index;
412 unsigned long pg_index;
414 u64 isize = i_size_read(inode);
417 unsigned long nr_pages = 0;
418 struct extent_map *em;
419 struct address_space *mapping = inode->i_mapping;
420 struct extent_map_tree *em_tree;
421 struct extent_io_tree *tree;
425 last_offset = bio_end_offset(cb->orig_bio);
426 em_tree = &BTRFS_I(inode)->extent_tree;
427 tree = &BTRFS_I(inode)->io_tree;
432 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
434 while (last_offset < compressed_end) {
435 pg_index = last_offset >> PAGE_SHIFT;
437 if (pg_index > end_index)
441 page = radix_tree_lookup(&mapping->i_pages, pg_index);
443 if (page && !radix_tree_exceptional_entry(page)) {
450 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
455 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
460 end = last_offset + PAGE_SIZE - 1;
462 * at this point, we have a locked page in the page cache
463 * for these bytes in the file. But, we have to make
464 * sure they map to this compressed extent on disk.
466 set_page_extent_mapped(page);
467 lock_extent(tree, last_offset, end);
468 read_lock(&em_tree->lock);
469 em = lookup_extent_mapping(em_tree, last_offset,
471 read_unlock(&em_tree->lock);
473 if (!em || last_offset < em->start ||
474 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
475 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
477 unlock_extent(tree, last_offset, end);
484 if (page->index == end_index) {
486 size_t zero_offset = isize & (PAGE_SIZE - 1);
490 zeros = PAGE_SIZE - zero_offset;
491 userpage = kmap_atomic(page);
492 memset(userpage + zero_offset, 0, zeros);
493 flush_dcache_page(page);
494 kunmap_atomic(userpage);
498 ret = bio_add_page(cb->orig_bio, page,
501 if (ret == PAGE_SIZE) {
505 unlock_extent(tree, last_offset, end);
511 last_offset += PAGE_SIZE;
517 * for a compressed read, the bio we get passed has all the inode pages
518 * in it. We don't actually do IO on those pages but allocate new ones
519 * to hold the compressed pages on disk.
521 * bio->bi_iter.bi_sector points to the compressed extent on disk
522 * bio->bi_io_vec points to all of the inode pages
524 * After the compressed pages are read, we copy the bytes into the
525 * bio we were passed and then call the bio end_io calls
527 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
528 int mirror_num, unsigned long bio_flags)
530 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
531 struct extent_io_tree *tree;
532 struct extent_map_tree *em_tree;
533 struct compressed_bio *cb;
534 unsigned long compressed_len;
535 unsigned long nr_pages;
536 unsigned long pg_index;
538 struct block_device *bdev;
539 struct bio *comp_bio;
540 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
543 struct extent_map *em;
544 blk_status_t ret = BLK_STS_RESOURCE;
548 tree = &BTRFS_I(inode)->io_tree;
549 em_tree = &BTRFS_I(inode)->extent_tree;
551 /* we need the actual starting offset of this extent in the file */
552 read_lock(&em_tree->lock);
553 em = lookup_extent_mapping(em_tree,
554 page_offset(bio_first_page_all(bio)),
556 read_unlock(&em_tree->lock);
558 return BLK_STS_IOERR;
560 compressed_len = em->block_len;
561 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
565 refcount_set(&cb->pending_bios, 0);
568 cb->mirror_num = mirror_num;
571 cb->start = em->orig_start;
573 em_start = em->start;
578 cb->len = bio->bi_iter.bi_size;
579 cb->compressed_len = compressed_len;
580 cb->compress_type = extent_compress_type(bio_flags);
583 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
584 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
586 if (!cb->compressed_pages)
589 bdev = fs_info->fs_devices->latest_bdev;
591 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
592 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
594 if (!cb->compressed_pages[pg_index]) {
595 faili = pg_index - 1;
596 ret = BLK_STS_RESOURCE;
600 faili = nr_pages - 1;
601 cb->nr_pages = nr_pages;
603 add_ra_bio_pages(inode, em_start + em_len, cb);
605 /* include any pages we added in add_ra-bio_pages */
606 cb->len = bio->bi_iter.bi_size;
608 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
609 comp_bio->bi_opf = REQ_OP_READ;
610 comp_bio->bi_private = cb;
611 comp_bio->bi_end_io = end_compressed_bio_read;
612 refcount_set(&cb->pending_bios, 1);
614 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
617 page = cb->compressed_pages[pg_index];
618 page->mapping = inode->i_mapping;
619 page->index = em_start >> PAGE_SHIFT;
621 if (comp_bio->bi_iter.bi_size)
622 submit = btrfs_merge_bio_hook(page, 0, PAGE_SIZE,
625 page->mapping = NULL;
626 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
628 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
629 BTRFS_WQ_ENDIO_DATA);
630 BUG_ON(ret); /* -ENOMEM */
633 * inc the count before we submit the bio so
634 * we know the end IO handler won't happen before
635 * we inc the count. Otherwise, the cb might get
636 * freed before we're done setting it up
638 refcount_inc(&cb->pending_bios);
640 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
641 ret = btrfs_lookup_bio_sums(inode, comp_bio,
643 BUG_ON(ret); /* -ENOMEM */
645 sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
646 fs_info->sectorsize);
648 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
650 comp_bio->bi_status = ret;
654 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
655 comp_bio->bi_opf = REQ_OP_READ;
656 comp_bio->bi_private = cb;
657 comp_bio->bi_end_io = end_compressed_bio_read;
659 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
661 cur_disk_byte += PAGE_SIZE;
664 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
665 BUG_ON(ret); /* -ENOMEM */
667 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
668 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
669 BUG_ON(ret); /* -ENOMEM */
672 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
674 comp_bio->bi_status = ret;
682 __free_page(cb->compressed_pages[faili]);
686 kfree(cb->compressed_pages);
695 * Heuristic uses systematic sampling to collect data from the input data
696 * range, the logic can be tuned by the following constants:
698 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
699 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
701 #define SAMPLING_READ_SIZE (16)
702 #define SAMPLING_INTERVAL (256)
705 * For statistical analysis of the input data we consider bytes that form a
706 * Galois Field of 256 objects. Each object has an attribute count, ie. how
707 * many times the object appeared in the sample.
709 #define BUCKET_SIZE (256)
712 * The size of the sample is based on a statistical sampling rule of thumb.
713 * The common way is to perform sampling tests as long as the number of
714 * elements in each cell is at least 5.
716 * Instead of 5, we choose 32 to obtain more accurate results.
717 * If the data contain the maximum number of symbols, which is 256, we obtain a
718 * sample size bound by 8192.
720 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
721 * from up to 512 locations.
723 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
724 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
730 struct heuristic_ws {
731 /* Partial copy of input data */
734 /* Buckets store counters for each byte value */
735 struct bucket_item *bucket;
737 struct bucket_item *bucket_b;
738 struct list_head list;
741 static void free_heuristic_ws(struct list_head *ws)
743 struct heuristic_ws *workspace;
745 workspace = list_entry(ws, struct heuristic_ws, list);
747 kvfree(workspace->sample);
748 kfree(workspace->bucket);
749 kfree(workspace->bucket_b);
753 static struct list_head *alloc_heuristic_ws(void)
755 struct heuristic_ws *ws;
757 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
759 return ERR_PTR(-ENOMEM);
761 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
765 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
769 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
773 INIT_LIST_HEAD(&ws->list);
776 free_heuristic_ws(&ws->list);
777 return ERR_PTR(-ENOMEM);
780 struct workspaces_list {
781 struct list_head idle_ws;
783 /* Number of free workspaces */
785 /* Total number of allocated workspaces */
787 /* Waiters for a free workspace */
788 wait_queue_head_t ws_wait;
791 static struct workspaces_list btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
793 static struct workspaces_list btrfs_heuristic_ws;
795 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
796 &btrfs_zlib_compress,
798 &btrfs_zstd_compress,
801 void __init btrfs_init_compress(void)
803 struct list_head *workspace;
806 INIT_LIST_HEAD(&btrfs_heuristic_ws.idle_ws);
807 spin_lock_init(&btrfs_heuristic_ws.ws_lock);
808 atomic_set(&btrfs_heuristic_ws.total_ws, 0);
809 init_waitqueue_head(&btrfs_heuristic_ws.ws_wait);
811 workspace = alloc_heuristic_ws();
812 if (IS_ERR(workspace)) {
814 "BTRFS: cannot preallocate heuristic workspace, will try later\n");
816 atomic_set(&btrfs_heuristic_ws.total_ws, 1);
817 btrfs_heuristic_ws.free_ws = 1;
818 list_add(workspace, &btrfs_heuristic_ws.idle_ws);
821 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
822 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
823 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
824 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
825 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
828 * Preallocate one workspace for each compression type so
829 * we can guarantee forward progress in the worst case
831 workspace = btrfs_compress_op[i]->alloc_workspace();
832 if (IS_ERR(workspace)) {
833 pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n");
835 atomic_set(&btrfs_comp_ws[i].total_ws, 1);
836 btrfs_comp_ws[i].free_ws = 1;
837 list_add(workspace, &btrfs_comp_ws[i].idle_ws);
843 * This finds an available workspace or allocates a new one.
844 * If it's not possible to allocate a new one, waits until there's one.
845 * Preallocation makes a forward progress guarantees and we do not return
848 static struct list_head *__find_workspace(int type, bool heuristic)
850 struct list_head *workspace;
851 int cpus = num_online_cpus();
854 struct list_head *idle_ws;
857 wait_queue_head_t *ws_wait;
861 idle_ws = &btrfs_heuristic_ws.idle_ws;
862 ws_lock = &btrfs_heuristic_ws.ws_lock;
863 total_ws = &btrfs_heuristic_ws.total_ws;
864 ws_wait = &btrfs_heuristic_ws.ws_wait;
865 free_ws = &btrfs_heuristic_ws.free_ws;
867 idle_ws = &btrfs_comp_ws[idx].idle_ws;
868 ws_lock = &btrfs_comp_ws[idx].ws_lock;
869 total_ws = &btrfs_comp_ws[idx].total_ws;
870 ws_wait = &btrfs_comp_ws[idx].ws_wait;
871 free_ws = &btrfs_comp_ws[idx].free_ws;
876 if (!list_empty(idle_ws)) {
877 workspace = idle_ws->next;
880 spin_unlock(ws_lock);
884 if (atomic_read(total_ws) > cpus) {
887 spin_unlock(ws_lock);
888 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
889 if (atomic_read(total_ws) > cpus && !*free_ws)
891 finish_wait(ws_wait, &wait);
894 atomic_inc(total_ws);
895 spin_unlock(ws_lock);
898 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
899 * to turn it off here because we might get called from the restricted
900 * context of btrfs_compress_bio/btrfs_compress_pages
902 nofs_flag = memalloc_nofs_save();
904 workspace = alloc_heuristic_ws();
906 workspace = btrfs_compress_op[idx]->alloc_workspace();
907 memalloc_nofs_restore(nofs_flag);
909 if (IS_ERR(workspace)) {
910 atomic_dec(total_ws);
914 * Do not return the error but go back to waiting. There's a
915 * workspace preallocated for each type and the compression
916 * time is bounded so we get to a workspace eventually. This
917 * makes our caller's life easier.
919 * To prevent silent and low-probability deadlocks (when the
920 * initial preallocation fails), check if there are any
923 if (atomic_read(total_ws) == 0) {
924 static DEFINE_RATELIMIT_STATE(_rs,
925 /* once per minute */ 60 * HZ,
928 if (__ratelimit(&_rs)) {
929 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
937 static struct list_head *find_workspace(int type)
939 return __find_workspace(type, false);
943 * put a workspace struct back on the list or free it if we have enough
944 * idle ones sitting around
946 static void __free_workspace(int type, struct list_head *workspace,
950 struct list_head *idle_ws;
953 wait_queue_head_t *ws_wait;
957 idle_ws = &btrfs_heuristic_ws.idle_ws;
958 ws_lock = &btrfs_heuristic_ws.ws_lock;
959 total_ws = &btrfs_heuristic_ws.total_ws;
960 ws_wait = &btrfs_heuristic_ws.ws_wait;
961 free_ws = &btrfs_heuristic_ws.free_ws;
963 idle_ws = &btrfs_comp_ws[idx].idle_ws;
964 ws_lock = &btrfs_comp_ws[idx].ws_lock;
965 total_ws = &btrfs_comp_ws[idx].total_ws;
966 ws_wait = &btrfs_comp_ws[idx].ws_wait;
967 free_ws = &btrfs_comp_ws[idx].free_ws;
971 if (*free_ws <= num_online_cpus()) {
972 list_add(workspace, idle_ws);
974 spin_unlock(ws_lock);
977 spin_unlock(ws_lock);
980 free_heuristic_ws(workspace);
982 btrfs_compress_op[idx]->free_workspace(workspace);
983 atomic_dec(total_ws);
985 cond_wake_up(ws_wait);
988 static void free_workspace(int type, struct list_head *ws)
990 return __free_workspace(type, ws, false);
994 * cleanup function for module exit
996 static void free_workspaces(void)
998 struct list_head *workspace;
1001 while (!list_empty(&btrfs_heuristic_ws.idle_ws)) {
1002 workspace = btrfs_heuristic_ws.idle_ws.next;
1003 list_del(workspace);
1004 free_heuristic_ws(workspace);
1005 atomic_dec(&btrfs_heuristic_ws.total_ws);
1008 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
1009 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
1010 workspace = btrfs_comp_ws[i].idle_ws.next;
1011 list_del(workspace);
1012 btrfs_compress_op[i]->free_workspace(workspace);
1013 atomic_dec(&btrfs_comp_ws[i].total_ws);
1019 * Given an address space and start and length, compress the bytes into @pages
1020 * that are allocated on demand.
1022 * @type_level is encoded algorithm and level, where level 0 means whatever
1023 * default the algorithm chooses and is opaque here;
1024 * - compression algo are 0-3
1025 * - the level are bits 4-7
1027 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1028 * and returns number of actually allocated pages
1030 * @total_in is used to return the number of bytes actually read. It
1031 * may be smaller than the input length if we had to exit early because we
1032 * ran out of room in the pages array or because we cross the
1033 * max_out threshold.
1035 * @total_out is an in/out parameter, must be set to the input length and will
1036 * be also used to return the total number of compressed bytes
1038 * @max_out tells us the max number of bytes that we're allowed to
1041 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1042 u64 start, struct page **pages,
1043 unsigned long *out_pages,
1044 unsigned long *total_in,
1045 unsigned long *total_out)
1047 struct list_head *workspace;
1049 int type = type_level & 0xF;
1051 workspace = find_workspace(type);
1053 btrfs_compress_op[type - 1]->set_level(workspace, type_level);
1054 ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
1057 total_in, total_out);
1058 free_workspace(type, workspace);
1063 * pages_in is an array of pages with compressed data.
1065 * disk_start is the starting logical offset of this array in the file
1067 * orig_bio contains the pages from the file that we want to decompress into
1069 * srclen is the number of bytes in pages_in
1071 * The basic idea is that we have a bio that was created by readpages.
1072 * The pages in the bio are for the uncompressed data, and they may not
1073 * be contiguous. They all correspond to the range of bytes covered by
1074 * the compressed extent.
1076 static int btrfs_decompress_bio(struct compressed_bio *cb)
1078 struct list_head *workspace;
1080 int type = cb->compress_type;
1082 workspace = find_workspace(type);
1083 ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb);
1084 free_workspace(type, workspace);
1090 * a less complex decompression routine. Our compressed data fits in a
1091 * single page, and we want to read a single page out of it.
1092 * start_byte tells us the offset into the compressed data we're interested in
1094 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1095 unsigned long start_byte, size_t srclen, size_t destlen)
1097 struct list_head *workspace;
1100 workspace = find_workspace(type);
1102 ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
1103 dest_page, start_byte,
1106 free_workspace(type, workspace);
1110 void __cold btrfs_exit_compress(void)
1116 * Copy uncompressed data from working buffer to pages.
1118 * buf_start is the byte offset we're of the start of our workspace buffer.
1120 * total_out is the last byte of the buffer
1122 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1123 unsigned long total_out, u64 disk_start,
1126 unsigned long buf_offset;
1127 unsigned long current_buf_start;
1128 unsigned long start_byte;
1129 unsigned long prev_start_byte;
1130 unsigned long working_bytes = total_out - buf_start;
1131 unsigned long bytes;
1133 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1136 * start byte is the first byte of the page we're currently
1137 * copying into relative to the start of the compressed data.
1139 start_byte = page_offset(bvec.bv_page) - disk_start;
1141 /* we haven't yet hit data corresponding to this page */
1142 if (total_out <= start_byte)
1146 * the start of the data we care about is offset into
1147 * the middle of our working buffer
1149 if (total_out > start_byte && buf_start < start_byte) {
1150 buf_offset = start_byte - buf_start;
1151 working_bytes -= buf_offset;
1155 current_buf_start = buf_start;
1157 /* copy bytes from the working buffer into the pages */
1158 while (working_bytes > 0) {
1159 bytes = min_t(unsigned long, bvec.bv_len,
1160 PAGE_SIZE - buf_offset);
1161 bytes = min(bytes, working_bytes);
1163 kaddr = kmap_atomic(bvec.bv_page);
1164 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1165 kunmap_atomic(kaddr);
1166 flush_dcache_page(bvec.bv_page);
1168 buf_offset += bytes;
1169 working_bytes -= bytes;
1170 current_buf_start += bytes;
1172 /* check if we need to pick another page */
1173 bio_advance(bio, bytes);
1174 if (!bio->bi_iter.bi_size)
1176 bvec = bio_iter_iovec(bio, bio->bi_iter);
1177 prev_start_byte = start_byte;
1178 start_byte = page_offset(bvec.bv_page) - disk_start;
1181 * We need to make sure we're only adjusting
1182 * our offset into compression working buffer when
1183 * we're switching pages. Otherwise we can incorrectly
1184 * keep copying when we were actually done.
1186 if (start_byte != prev_start_byte) {
1188 * make sure our new page is covered by this
1191 if (total_out <= start_byte)
1195 * the next page in the biovec might not be adjacent
1196 * to the last page, but it might still be found
1197 * inside this working buffer. bump our offset pointer
1199 if (total_out > start_byte &&
1200 current_buf_start < start_byte) {
1201 buf_offset = start_byte - buf_start;
1202 working_bytes = total_out - start_byte;
1203 current_buf_start = buf_start + buf_offset;
1212 * Shannon Entropy calculation
1214 * Pure byte distribution analysis fails to determine compressiability of data.
1215 * Try calculating entropy to estimate the average minimum number of bits
1216 * needed to encode the sampled data.
1218 * For convenience, return the percentage of needed bits, instead of amount of
1221 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1222 * and can be compressible with high probability
1224 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1226 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1228 #define ENTROPY_LVL_ACEPTABLE (65)
1229 #define ENTROPY_LVL_HIGH (80)
1232 * For increasead precision in shannon_entropy calculation,
1233 * let's do pow(n, M) to save more digits after comma:
1235 * - maximum int bit length is 64
1236 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1237 * - 13 * 4 = 52 < 64 -> M = 4
1241 static inline u32 ilog2_w(u64 n)
1243 return ilog2(n * n * n * n);
1246 static u32 shannon_entropy(struct heuristic_ws *ws)
1248 const u32 entropy_max = 8 * ilog2_w(2);
1249 u32 entropy_sum = 0;
1250 u32 p, p_base, sz_base;
1253 sz_base = ilog2_w(ws->sample_size);
1254 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1255 p = ws->bucket[i].count;
1256 p_base = ilog2_w(p);
1257 entropy_sum += p * (sz_base - p_base);
1260 entropy_sum /= ws->sample_size;
1261 return entropy_sum * 100 / entropy_max;
1264 #define RADIX_BASE 4U
1265 #define COUNTERS_SIZE (1U << RADIX_BASE)
1267 static u8 get4bits(u64 num, int shift) {
1272 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1277 * Use 4 bits as radix base
1278 * Use 16 u32 counters for calculating new possition in buf array
1280 * @array - array that will be sorted
1281 * @array_buf - buffer array to store sorting results
1282 * must be equal in size to @array
1285 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1290 u32 counters[COUNTERS_SIZE];
1298 * Try avoid useless loop iterations for small numbers stored in big
1299 * counters. Example: 48 33 4 ... in 64bit array
1301 max_num = array[0].count;
1302 for (i = 1; i < num; i++) {
1303 buf_num = array[i].count;
1304 if (buf_num > max_num)
1308 buf_num = ilog2(max_num);
1309 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1312 while (shift < bitlen) {
1313 memset(counters, 0, sizeof(counters));
1315 for (i = 0; i < num; i++) {
1316 buf_num = array[i].count;
1317 addr = get4bits(buf_num, shift);
1321 for (i = 1; i < COUNTERS_SIZE; i++)
1322 counters[i] += counters[i - 1];
1324 for (i = num - 1; i >= 0; i--) {
1325 buf_num = array[i].count;
1326 addr = get4bits(buf_num, shift);
1328 new_addr = counters[addr];
1329 array_buf[new_addr] = array[i];
1332 shift += RADIX_BASE;
1335 * Normal radix expects to move data from a temporary array, to
1336 * the main one. But that requires some CPU time. Avoid that
1337 * by doing another sort iteration to original array instead of
1340 memset(counters, 0, sizeof(counters));
1342 for (i = 0; i < num; i ++) {
1343 buf_num = array_buf[i].count;
1344 addr = get4bits(buf_num, shift);
1348 for (i = 1; i < COUNTERS_SIZE; i++)
1349 counters[i] += counters[i - 1];
1351 for (i = num - 1; i >= 0; i--) {
1352 buf_num = array_buf[i].count;
1353 addr = get4bits(buf_num, shift);
1355 new_addr = counters[addr];
1356 array[new_addr] = array_buf[i];
1359 shift += RADIX_BASE;
1364 * Size of the core byte set - how many bytes cover 90% of the sample
1366 * There are several types of structured binary data that use nearly all byte
1367 * values. The distribution can be uniform and counts in all buckets will be
1368 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1370 * Other possibility is normal (Gaussian) distribution, where the data could
1371 * be potentially compressible, but we have to take a few more steps to decide
1374 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1375 * compression algo can easy fix that
1376 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1377 * probability is not compressible
1379 #define BYTE_CORE_SET_LOW (64)
1380 #define BYTE_CORE_SET_HIGH (200)
1382 static int byte_core_set_size(struct heuristic_ws *ws)
1385 u32 coreset_sum = 0;
1386 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1387 struct bucket_item *bucket = ws->bucket;
1389 /* Sort in reverse order */
1390 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1392 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1393 coreset_sum += bucket[i].count;
1395 if (coreset_sum > core_set_threshold)
1398 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1399 coreset_sum += bucket[i].count;
1400 if (coreset_sum > core_set_threshold)
1408 * Count byte values in buckets.
1409 * This heuristic can detect textual data (configs, xml, json, html, etc).
1410 * Because in most text-like data byte set is restricted to limited number of
1411 * possible characters, and that restriction in most cases makes data easy to
1414 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1415 * less - compressible
1416 * more - need additional analysis
1418 #define BYTE_SET_THRESHOLD (64)
1420 static u32 byte_set_size(const struct heuristic_ws *ws)
1423 u32 byte_set_size = 0;
1425 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1426 if (ws->bucket[i].count > 0)
1431 * Continue collecting count of byte values in buckets. If the byte
1432 * set size is bigger then the threshold, it's pointless to continue,
1433 * the detection technique would fail for this type of data.
1435 for (; i < BUCKET_SIZE; i++) {
1436 if (ws->bucket[i].count > 0) {
1438 if (byte_set_size > BYTE_SET_THRESHOLD)
1439 return byte_set_size;
1443 return byte_set_size;
1446 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1448 const u32 half_of_sample = ws->sample_size / 2;
1449 const u8 *data = ws->sample;
1451 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1454 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1455 struct heuristic_ws *ws)
1458 u64 index, index_end;
1459 u32 i, curr_sample_pos;
1463 * Compression handles the input data by chunks of 128KiB
1464 * (defined by BTRFS_MAX_UNCOMPRESSED)
1466 * We do the same for the heuristic and loop over the whole range.
1468 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1469 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1471 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1472 end = start + BTRFS_MAX_UNCOMPRESSED;
1474 index = start >> PAGE_SHIFT;
1475 index_end = end >> PAGE_SHIFT;
1477 /* Don't miss unaligned end */
1478 if (!IS_ALIGNED(end, PAGE_SIZE))
1481 curr_sample_pos = 0;
1482 while (index < index_end) {
1483 page = find_get_page(inode->i_mapping, index);
1484 in_data = kmap(page);
1485 /* Handle case where the start is not aligned to PAGE_SIZE */
1486 i = start % PAGE_SIZE;
1487 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1488 /* Don't sample any garbage from the last page */
1489 if (start > end - SAMPLING_READ_SIZE)
1491 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1492 SAMPLING_READ_SIZE);
1493 i += SAMPLING_INTERVAL;
1494 start += SAMPLING_INTERVAL;
1495 curr_sample_pos += SAMPLING_READ_SIZE;
1503 ws->sample_size = curr_sample_pos;
1507 * Compression heuristic.
1509 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1510 * quickly (compared to direct compression) detect data characteristics
1511 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1514 * The following types of analysis can be performed:
1515 * - detect mostly zero data
1516 * - detect data with low "byte set" size (text, etc)
1517 * - detect data with low/high "core byte" set
1519 * Return non-zero if the compression should be done, 0 otherwise.
1521 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1523 struct list_head *ws_list = __find_workspace(0, true);
1524 struct heuristic_ws *ws;
1529 ws = list_entry(ws_list, struct heuristic_ws, list);
1531 heuristic_collect_sample(inode, start, end, ws);
1533 if (sample_repeated_patterns(ws)) {
1538 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1540 for (i = 0; i < ws->sample_size; i++) {
1541 byte = ws->sample[i];
1542 ws->bucket[byte].count++;
1545 i = byte_set_size(ws);
1546 if (i < BYTE_SET_THRESHOLD) {
1551 i = byte_core_set_size(ws);
1552 if (i <= BYTE_CORE_SET_LOW) {
1557 if (i >= BYTE_CORE_SET_HIGH) {
1562 i = shannon_entropy(ws);
1563 if (i <= ENTROPY_LVL_ACEPTABLE) {
1569 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1570 * needed to give green light to compression.
1572 * For now just assume that compression at that level is not worth the
1573 * resources because:
1575 * 1. it is possible to defrag the data later
1577 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1578 * values, every bucket has counter at level ~54. The heuristic would
1579 * be confused. This can happen when data have some internal repeated
1580 * patterns like "abbacbbc...". This can be detected by analyzing
1581 * pairs of bytes, which is too costly.
1583 if (i < ENTROPY_LVL_HIGH) {
1592 __free_workspace(0, ws_list, true);
1596 unsigned int btrfs_compress_str2level(const char *str)
1598 if (strncmp(str, "zlib", 4) != 0)
1601 /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */
1602 if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0)
1603 return str[5] - '0';
1605 return BTRFS_ZLIB_DEFAULT_LEVEL;