2 * Copyright (C) 2008 Oracle. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
19 #include <linux/kernel.h>
20 #include <linux/bio.h>
21 #include <linux/buffer_head.h>
22 #include <linux/file.h>
24 #include <linux/pagemap.h>
25 #include <linux/highmem.h>
26 #include <linux/time.h>
27 #include <linux/init.h>
28 #include <linux/string.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mpage.h>
31 #include <linux/swap.h>
32 #include <linux/writeback.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/slab.h>
35 #include <linux/sched/mm.h>
36 #include <linux/log2.h>
39 #include "transaction.h"
40 #include "btrfs_inode.h"
42 #include "ordered-data.h"
43 #include "compression.h"
44 #include "extent_io.h"
45 #include "extent_map.h"
47 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
49 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
52 case BTRFS_COMPRESS_ZLIB:
53 case BTRFS_COMPRESS_LZO:
54 case BTRFS_COMPRESS_ZSTD:
55 case BTRFS_COMPRESS_NONE:
56 return btrfs_compress_types[type];
62 static int btrfs_decompress_bio(struct compressed_bio *cb);
64 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
65 unsigned long disk_size)
67 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
69 return sizeof(struct compressed_bio) +
70 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
73 static int check_compressed_csum(struct btrfs_inode *inode,
74 struct compressed_bio *cb,
82 u32 *cb_sum = &cb->sums;
84 if (inode->flags & BTRFS_INODE_NODATASUM)
87 for (i = 0; i < cb->nr_pages; i++) {
88 page = cb->compressed_pages[i];
91 kaddr = kmap_atomic(page);
92 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
93 btrfs_csum_final(csum, (u8 *)&csum);
96 if (csum != *cb_sum) {
97 btrfs_print_data_csum_error(inode, disk_start, csum,
98 *cb_sum, cb->mirror_num);
110 /* when we finish reading compressed pages from the disk, we
111 * decompress them and then run the bio end_io routines on the
112 * decompressed pages (in the inode address space).
114 * This allows the checksumming and other IO error handling routines
117 * The compressed pages are freed here, and it must be run
120 static void end_compressed_bio_read(struct bio *bio)
122 struct compressed_bio *cb = bio->bi_private;
126 unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
132 /* if there are more bios still pending for this compressed
135 if (!refcount_dec_and_test(&cb->pending_bios))
139 * Record the correct mirror_num in cb->orig_bio so that
140 * read-repair can work properly.
142 ASSERT(btrfs_io_bio(cb->orig_bio));
143 btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
144 cb->mirror_num = mirror;
147 * Some IO in this cb have failed, just skip checksum as there
148 * is no way it could be correct.
154 ret = check_compressed_csum(BTRFS_I(inode), cb,
155 (u64)bio->bi_iter.bi_sector << 9);
159 /* ok, we're the last bio for this extent, lets start
162 ret = btrfs_decompress_bio(cb);
168 /* release the compressed pages */
170 for (index = 0; index < cb->nr_pages; index++) {
171 page = cb->compressed_pages[index];
172 page->mapping = NULL;
176 /* do io completion on the original bio */
178 bio_io_error(cb->orig_bio);
181 struct bio_vec *bvec;
184 * we have verified the checksum already, set page
185 * checked so the end_io handlers know about it
187 ASSERT(!bio_flagged(bio, BIO_CLONED));
188 bio_for_each_segment_all(bvec, cb->orig_bio, i)
189 SetPageChecked(bvec->bv_page);
191 bio_endio(cb->orig_bio);
194 /* finally free the cb struct */
195 kfree(cb->compressed_pages);
202 * Clear the writeback bits on all of the file
203 * pages for a compressed write
205 static noinline void end_compressed_writeback(struct inode *inode,
206 const struct compressed_bio *cb)
208 unsigned long index = cb->start >> PAGE_SHIFT;
209 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
210 struct page *pages[16];
211 unsigned long nr_pages = end_index - index + 1;
216 mapping_set_error(inode->i_mapping, -EIO);
218 while (nr_pages > 0) {
219 ret = find_get_pages_contig(inode->i_mapping, index,
221 nr_pages, ARRAY_SIZE(pages)), pages);
227 for (i = 0; i < ret; i++) {
229 SetPageError(pages[i]);
230 end_page_writeback(pages[i]);
236 /* the inode may be gone now */
240 * do the cleanup once all the compressed pages hit the disk.
241 * This will clear writeback on the file pages and free the compressed
244 * This also calls the writeback end hooks for the file pages so that
245 * metadata and checksums can be updated in the file.
247 static void end_compressed_bio_write(struct bio *bio)
249 struct extent_io_tree *tree;
250 struct compressed_bio *cb = bio->bi_private;
258 /* if there are more bios still pending for this compressed
261 if (!refcount_dec_and_test(&cb->pending_bios))
264 /* ok, we're the last bio for this extent, step one is to
265 * call back into the FS and do all the end_io operations
268 tree = &BTRFS_I(inode)->io_tree;
269 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
270 tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
272 cb->start + cb->len - 1,
275 BLK_STS_OK : BLK_STS_NOTSUPP);
276 cb->compressed_pages[0]->mapping = NULL;
278 end_compressed_writeback(inode, cb);
279 /* note, our inode could be gone now */
282 * release the compressed pages, these came from alloc_page and
283 * are not attached to the inode at all
286 for (index = 0; index < cb->nr_pages; index++) {
287 page = cb->compressed_pages[index];
288 page->mapping = NULL;
292 /* finally free the cb struct */
293 kfree(cb->compressed_pages);
300 * worker function to build and submit bios for previously compressed pages.
301 * The corresponding pages in the inode should be marked for writeback
302 * and the compressed pages should have a reference on them for dropping
303 * when the IO is complete.
305 * This also checksums the file bytes and gets things ready for
308 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
309 unsigned long len, u64 disk_start,
310 unsigned long compressed_len,
311 struct page **compressed_pages,
312 unsigned long nr_pages,
313 unsigned int write_flags)
315 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
316 struct bio *bio = NULL;
317 struct compressed_bio *cb;
318 unsigned long bytes_left;
319 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
322 u64 first_byte = disk_start;
323 struct block_device *bdev;
325 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
327 WARN_ON(start & ((u64)PAGE_SIZE - 1));
328 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
330 return BLK_STS_RESOURCE;
331 refcount_set(&cb->pending_bios, 0);
337 cb->compressed_pages = compressed_pages;
338 cb->compressed_len = compressed_len;
340 cb->nr_pages = nr_pages;
342 bdev = fs_info->fs_devices->latest_bdev;
344 bio = btrfs_bio_alloc(bdev, first_byte);
345 bio->bi_opf = REQ_OP_WRITE | write_flags;
346 bio->bi_private = cb;
347 bio->bi_end_io = end_compressed_bio_write;
348 refcount_set(&cb->pending_bios, 1);
350 /* create and submit bios for the compressed pages */
351 bytes_left = compressed_len;
352 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
355 page = compressed_pages[pg_index];
356 page->mapping = inode->i_mapping;
357 if (bio->bi_iter.bi_size)
358 submit = io_tree->ops->merge_bio_hook(page, 0,
362 page->mapping = NULL;
363 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
368 * inc the count before we submit the bio so
369 * we know the end IO handler won't happen before
370 * we inc the count. Otherwise, the cb might get
371 * freed before we're done setting it up
373 refcount_inc(&cb->pending_bios);
374 ret = btrfs_bio_wq_end_io(fs_info, bio,
375 BTRFS_WQ_ENDIO_DATA);
376 BUG_ON(ret); /* -ENOMEM */
379 ret = btrfs_csum_one_bio(inode, bio, start, 1);
380 BUG_ON(ret); /* -ENOMEM */
383 ret = btrfs_map_bio(fs_info, bio, 0, 1);
385 bio->bi_status = ret;
391 bio = btrfs_bio_alloc(bdev, first_byte);
392 bio->bi_opf = REQ_OP_WRITE | write_flags;
393 bio->bi_private = cb;
394 bio->bi_end_io = end_compressed_bio_write;
395 bio_add_page(bio, page, PAGE_SIZE, 0);
397 if (bytes_left < PAGE_SIZE) {
399 "bytes left %lu compress len %lu nr %lu",
400 bytes_left, cb->compressed_len, cb->nr_pages);
402 bytes_left -= PAGE_SIZE;
403 first_byte += PAGE_SIZE;
408 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
409 BUG_ON(ret); /* -ENOMEM */
412 ret = btrfs_csum_one_bio(inode, bio, start, 1);
413 BUG_ON(ret); /* -ENOMEM */
416 ret = btrfs_map_bio(fs_info, bio, 0, 1);
418 bio->bi_status = ret;
426 static u64 bio_end_offset(struct bio *bio)
428 struct bio_vec *last = &bio->bi_io_vec[bio->bi_vcnt - 1];
430 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
433 static noinline int add_ra_bio_pages(struct inode *inode,
435 struct compressed_bio *cb)
437 unsigned long end_index;
438 unsigned long pg_index;
440 u64 isize = i_size_read(inode);
443 unsigned long nr_pages = 0;
444 struct extent_map *em;
445 struct address_space *mapping = inode->i_mapping;
446 struct extent_map_tree *em_tree;
447 struct extent_io_tree *tree;
451 last_offset = bio_end_offset(cb->orig_bio);
452 em_tree = &BTRFS_I(inode)->extent_tree;
453 tree = &BTRFS_I(inode)->io_tree;
458 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
460 while (last_offset < compressed_end) {
461 pg_index = last_offset >> PAGE_SHIFT;
463 if (pg_index > end_index)
467 page = radix_tree_lookup(&mapping->page_tree, pg_index);
469 if (page && !radix_tree_exceptional_entry(page)) {
476 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
481 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
486 end = last_offset + PAGE_SIZE - 1;
488 * at this point, we have a locked page in the page cache
489 * for these bytes in the file. But, we have to make
490 * sure they map to this compressed extent on disk.
492 set_page_extent_mapped(page);
493 lock_extent(tree, last_offset, end);
494 read_lock(&em_tree->lock);
495 em = lookup_extent_mapping(em_tree, last_offset,
497 read_unlock(&em_tree->lock);
499 if (!em || last_offset < em->start ||
500 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
501 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
503 unlock_extent(tree, last_offset, end);
510 if (page->index == end_index) {
512 size_t zero_offset = isize & (PAGE_SIZE - 1);
516 zeros = PAGE_SIZE - zero_offset;
517 userpage = kmap_atomic(page);
518 memset(userpage + zero_offset, 0, zeros);
519 flush_dcache_page(page);
520 kunmap_atomic(userpage);
524 ret = bio_add_page(cb->orig_bio, page,
527 if (ret == PAGE_SIZE) {
531 unlock_extent(tree, last_offset, end);
537 last_offset += PAGE_SIZE;
543 * for a compressed read, the bio we get passed has all the inode pages
544 * in it. We don't actually do IO on those pages but allocate new ones
545 * to hold the compressed pages on disk.
547 * bio->bi_iter.bi_sector points to the compressed extent on disk
548 * bio->bi_io_vec points to all of the inode pages
550 * After the compressed pages are read, we copy the bytes into the
551 * bio we were passed and then call the bio end_io calls
553 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
554 int mirror_num, unsigned long bio_flags)
556 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
557 struct extent_io_tree *tree;
558 struct extent_map_tree *em_tree;
559 struct compressed_bio *cb;
560 unsigned long compressed_len;
561 unsigned long nr_pages;
562 unsigned long pg_index;
564 struct block_device *bdev;
565 struct bio *comp_bio;
566 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
569 struct extent_map *em;
570 blk_status_t ret = BLK_STS_RESOURCE;
574 tree = &BTRFS_I(inode)->io_tree;
575 em_tree = &BTRFS_I(inode)->extent_tree;
577 /* we need the actual starting offset of this extent in the file */
578 read_lock(&em_tree->lock);
579 em = lookup_extent_mapping(em_tree,
580 page_offset(bio->bi_io_vec->bv_page),
582 read_unlock(&em_tree->lock);
584 return BLK_STS_IOERR;
586 compressed_len = em->block_len;
587 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
591 refcount_set(&cb->pending_bios, 0);
594 cb->mirror_num = mirror_num;
597 cb->start = em->orig_start;
599 em_start = em->start;
604 cb->len = bio->bi_iter.bi_size;
605 cb->compressed_len = compressed_len;
606 cb->compress_type = extent_compress_type(bio_flags);
609 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
610 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
612 if (!cb->compressed_pages)
615 bdev = fs_info->fs_devices->latest_bdev;
617 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
618 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
620 if (!cb->compressed_pages[pg_index]) {
621 faili = pg_index - 1;
622 ret = BLK_STS_RESOURCE;
626 faili = nr_pages - 1;
627 cb->nr_pages = nr_pages;
629 add_ra_bio_pages(inode, em_start + em_len, cb);
631 /* include any pages we added in add_ra-bio_pages */
632 cb->len = bio->bi_iter.bi_size;
634 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
635 bio_set_op_attrs (comp_bio, REQ_OP_READ, 0);
636 comp_bio->bi_private = cb;
637 comp_bio->bi_end_io = end_compressed_bio_read;
638 refcount_set(&cb->pending_bios, 1);
640 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
643 page = cb->compressed_pages[pg_index];
644 page->mapping = inode->i_mapping;
645 page->index = em_start >> PAGE_SHIFT;
647 if (comp_bio->bi_iter.bi_size)
648 submit = tree->ops->merge_bio_hook(page, 0,
652 page->mapping = NULL;
653 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
657 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
658 BTRFS_WQ_ENDIO_DATA);
659 BUG_ON(ret); /* -ENOMEM */
662 * inc the count before we submit the bio so
663 * we know the end IO handler won't happen before
664 * we inc the count. Otherwise, the cb might get
665 * freed before we're done setting it up
667 refcount_inc(&cb->pending_bios);
669 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
670 ret = btrfs_lookup_bio_sums(inode, comp_bio,
672 BUG_ON(ret); /* -ENOMEM */
674 sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
675 fs_info->sectorsize);
677 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
679 comp_bio->bi_status = ret;
685 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
686 bio_set_op_attrs(comp_bio, REQ_OP_READ, 0);
687 comp_bio->bi_private = cb;
688 comp_bio->bi_end_io = end_compressed_bio_read;
690 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
692 cur_disk_byte += PAGE_SIZE;
696 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
697 BUG_ON(ret); /* -ENOMEM */
699 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
700 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
701 BUG_ON(ret); /* -ENOMEM */
704 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
706 comp_bio->bi_status = ret;
715 __free_page(cb->compressed_pages[faili]);
719 kfree(cb->compressed_pages);
728 * Heuristic uses systematic sampling to collect data from the input data
729 * range, the logic can be tuned by the following constants:
731 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
732 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
734 #define SAMPLING_READ_SIZE (16)
735 #define SAMPLING_INTERVAL (256)
738 * For statistical analysis of the input data we consider bytes that form a
739 * Galois Field of 256 objects. Each object has an attribute count, ie. how
740 * many times the object appeared in the sample.
742 #define BUCKET_SIZE (256)
745 * The size of the sample is based on a statistical sampling rule of thumb.
746 * The common way is to perform sampling tests as long as the number of
747 * elements in each cell is at least 5.
749 * Instead of 5, we choose 32 to obtain more accurate results.
750 * If the data contain the maximum number of symbols, which is 256, we obtain a
751 * sample size bound by 8192.
753 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
754 * from up to 512 locations.
756 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
757 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
763 struct heuristic_ws {
764 /* Partial copy of input data */
767 /* Buckets store counters for each byte value */
768 struct bucket_item *bucket;
770 struct bucket_item *bucket_b;
771 struct list_head list;
774 static void free_heuristic_ws(struct list_head *ws)
776 struct heuristic_ws *workspace;
778 workspace = list_entry(ws, struct heuristic_ws, list);
780 kvfree(workspace->sample);
781 kfree(workspace->bucket);
782 kfree(workspace->bucket_b);
786 static struct list_head *alloc_heuristic_ws(void)
788 struct heuristic_ws *ws;
790 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
792 return ERR_PTR(-ENOMEM);
794 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
798 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
802 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
806 INIT_LIST_HEAD(&ws->list);
809 free_heuristic_ws(&ws->list);
810 return ERR_PTR(-ENOMEM);
813 struct workspaces_list {
814 struct list_head idle_ws;
816 /* Number of free workspaces */
818 /* Total number of allocated workspaces */
820 /* Waiters for a free workspace */
821 wait_queue_head_t ws_wait;
824 static struct workspaces_list btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
826 static struct workspaces_list btrfs_heuristic_ws;
828 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
829 &btrfs_zlib_compress,
831 &btrfs_zstd_compress,
834 void __init btrfs_init_compress(void)
836 struct list_head *workspace;
839 INIT_LIST_HEAD(&btrfs_heuristic_ws.idle_ws);
840 spin_lock_init(&btrfs_heuristic_ws.ws_lock);
841 atomic_set(&btrfs_heuristic_ws.total_ws, 0);
842 init_waitqueue_head(&btrfs_heuristic_ws.ws_wait);
844 workspace = alloc_heuristic_ws();
845 if (IS_ERR(workspace)) {
847 "BTRFS: cannot preallocate heuristic workspace, will try later\n");
849 atomic_set(&btrfs_heuristic_ws.total_ws, 1);
850 btrfs_heuristic_ws.free_ws = 1;
851 list_add(workspace, &btrfs_heuristic_ws.idle_ws);
854 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
855 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
856 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
857 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
858 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
861 * Preallocate one workspace for each compression type so
862 * we can guarantee forward progress in the worst case
864 workspace = btrfs_compress_op[i]->alloc_workspace();
865 if (IS_ERR(workspace)) {
866 pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n");
868 atomic_set(&btrfs_comp_ws[i].total_ws, 1);
869 btrfs_comp_ws[i].free_ws = 1;
870 list_add(workspace, &btrfs_comp_ws[i].idle_ws);
876 * This finds an available workspace or allocates a new one.
877 * If it's not possible to allocate a new one, waits until there's one.
878 * Preallocation makes a forward progress guarantees and we do not return
881 static struct list_head *__find_workspace(int type, bool heuristic)
883 struct list_head *workspace;
884 int cpus = num_online_cpus();
887 struct list_head *idle_ws;
890 wait_queue_head_t *ws_wait;
894 idle_ws = &btrfs_heuristic_ws.idle_ws;
895 ws_lock = &btrfs_heuristic_ws.ws_lock;
896 total_ws = &btrfs_heuristic_ws.total_ws;
897 ws_wait = &btrfs_heuristic_ws.ws_wait;
898 free_ws = &btrfs_heuristic_ws.free_ws;
900 idle_ws = &btrfs_comp_ws[idx].idle_ws;
901 ws_lock = &btrfs_comp_ws[idx].ws_lock;
902 total_ws = &btrfs_comp_ws[idx].total_ws;
903 ws_wait = &btrfs_comp_ws[idx].ws_wait;
904 free_ws = &btrfs_comp_ws[idx].free_ws;
909 if (!list_empty(idle_ws)) {
910 workspace = idle_ws->next;
913 spin_unlock(ws_lock);
917 if (atomic_read(total_ws) > cpus) {
920 spin_unlock(ws_lock);
921 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
922 if (atomic_read(total_ws) > cpus && !*free_ws)
924 finish_wait(ws_wait, &wait);
927 atomic_inc(total_ws);
928 spin_unlock(ws_lock);
931 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
932 * to turn it off here because we might get called from the restricted
933 * context of btrfs_compress_bio/btrfs_compress_pages
935 nofs_flag = memalloc_nofs_save();
937 workspace = alloc_heuristic_ws();
939 workspace = btrfs_compress_op[idx]->alloc_workspace();
940 memalloc_nofs_restore(nofs_flag);
942 if (IS_ERR(workspace)) {
943 atomic_dec(total_ws);
947 * Do not return the error but go back to waiting. There's a
948 * workspace preallocated for each type and the compression
949 * time is bounded so we get to a workspace eventually. This
950 * makes our caller's life easier.
952 * To prevent silent and low-probability deadlocks (when the
953 * initial preallocation fails), check if there are any
956 if (atomic_read(total_ws) == 0) {
957 static DEFINE_RATELIMIT_STATE(_rs,
958 /* once per minute */ 60 * HZ,
961 if (__ratelimit(&_rs)) {
962 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
970 static struct list_head *find_workspace(int type)
972 return __find_workspace(type, false);
976 * put a workspace struct back on the list or free it if we have enough
977 * idle ones sitting around
979 static void __free_workspace(int type, struct list_head *workspace,
983 struct list_head *idle_ws;
986 wait_queue_head_t *ws_wait;
990 idle_ws = &btrfs_heuristic_ws.idle_ws;
991 ws_lock = &btrfs_heuristic_ws.ws_lock;
992 total_ws = &btrfs_heuristic_ws.total_ws;
993 ws_wait = &btrfs_heuristic_ws.ws_wait;
994 free_ws = &btrfs_heuristic_ws.free_ws;
996 idle_ws = &btrfs_comp_ws[idx].idle_ws;
997 ws_lock = &btrfs_comp_ws[idx].ws_lock;
998 total_ws = &btrfs_comp_ws[idx].total_ws;
999 ws_wait = &btrfs_comp_ws[idx].ws_wait;
1000 free_ws = &btrfs_comp_ws[idx].free_ws;
1004 if (*free_ws <= num_online_cpus()) {
1005 list_add(workspace, idle_ws);
1007 spin_unlock(ws_lock);
1010 spin_unlock(ws_lock);
1013 free_heuristic_ws(workspace);
1015 btrfs_compress_op[idx]->free_workspace(workspace);
1016 atomic_dec(total_ws);
1019 * Make sure counter is updated before we wake up waiters.
1022 if (waitqueue_active(ws_wait))
1026 static void free_workspace(int type, struct list_head *ws)
1028 return __free_workspace(type, ws, false);
1032 * cleanup function for module exit
1034 static void free_workspaces(void)
1036 struct list_head *workspace;
1039 while (!list_empty(&btrfs_heuristic_ws.idle_ws)) {
1040 workspace = btrfs_heuristic_ws.idle_ws.next;
1041 list_del(workspace);
1042 free_heuristic_ws(workspace);
1043 atomic_dec(&btrfs_heuristic_ws.total_ws);
1046 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
1047 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
1048 workspace = btrfs_comp_ws[i].idle_ws.next;
1049 list_del(workspace);
1050 btrfs_compress_op[i]->free_workspace(workspace);
1051 atomic_dec(&btrfs_comp_ws[i].total_ws);
1057 * Given an address space and start and length, compress the bytes into @pages
1058 * that are allocated on demand.
1060 * @type_level is encoded algorithm and level, where level 0 means whatever
1061 * default the algorithm chooses and is opaque here;
1062 * - compression algo are 0-3
1063 * - the level are bits 4-7
1065 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1066 * and returns number of actually allocated pages
1068 * @total_in is used to return the number of bytes actually read. It
1069 * may be smaller than the input length if we had to exit early because we
1070 * ran out of room in the pages array or because we cross the
1071 * max_out threshold.
1073 * @total_out is an in/out parameter, must be set to the input length and will
1074 * be also used to return the total number of compressed bytes
1076 * @max_out tells us the max number of bytes that we're allowed to
1079 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1080 u64 start, struct page **pages,
1081 unsigned long *out_pages,
1082 unsigned long *total_in,
1083 unsigned long *total_out)
1085 struct list_head *workspace;
1087 int type = type_level & 0xF;
1089 workspace = find_workspace(type);
1091 btrfs_compress_op[type - 1]->set_level(workspace, type_level);
1092 ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
1095 total_in, total_out);
1096 free_workspace(type, workspace);
1101 * pages_in is an array of pages with compressed data.
1103 * disk_start is the starting logical offset of this array in the file
1105 * orig_bio contains the pages from the file that we want to decompress into
1107 * srclen is the number of bytes in pages_in
1109 * The basic idea is that we have a bio that was created by readpages.
1110 * The pages in the bio are for the uncompressed data, and they may not
1111 * be contiguous. They all correspond to the range of bytes covered by
1112 * the compressed extent.
1114 static int btrfs_decompress_bio(struct compressed_bio *cb)
1116 struct list_head *workspace;
1118 int type = cb->compress_type;
1120 workspace = find_workspace(type);
1121 ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb);
1122 free_workspace(type, workspace);
1128 * a less complex decompression routine. Our compressed data fits in a
1129 * single page, and we want to read a single page out of it.
1130 * start_byte tells us the offset into the compressed data we're interested in
1132 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1133 unsigned long start_byte, size_t srclen, size_t destlen)
1135 struct list_head *workspace;
1138 workspace = find_workspace(type);
1140 ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
1141 dest_page, start_byte,
1144 free_workspace(type, workspace);
1148 void btrfs_exit_compress(void)
1154 * Copy uncompressed data from working buffer to pages.
1156 * buf_start is the byte offset we're of the start of our workspace buffer.
1158 * total_out is the last byte of the buffer
1160 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1161 unsigned long total_out, u64 disk_start,
1164 unsigned long buf_offset;
1165 unsigned long current_buf_start;
1166 unsigned long start_byte;
1167 unsigned long prev_start_byte;
1168 unsigned long working_bytes = total_out - buf_start;
1169 unsigned long bytes;
1171 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1174 * start byte is the first byte of the page we're currently
1175 * copying into relative to the start of the compressed data.
1177 start_byte = page_offset(bvec.bv_page) - disk_start;
1179 /* we haven't yet hit data corresponding to this page */
1180 if (total_out <= start_byte)
1184 * the start of the data we care about is offset into
1185 * the middle of our working buffer
1187 if (total_out > start_byte && buf_start < start_byte) {
1188 buf_offset = start_byte - buf_start;
1189 working_bytes -= buf_offset;
1193 current_buf_start = buf_start;
1195 /* copy bytes from the working buffer into the pages */
1196 while (working_bytes > 0) {
1197 bytes = min_t(unsigned long, bvec.bv_len,
1198 PAGE_SIZE - buf_offset);
1199 bytes = min(bytes, working_bytes);
1201 kaddr = kmap_atomic(bvec.bv_page);
1202 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1203 kunmap_atomic(kaddr);
1204 flush_dcache_page(bvec.bv_page);
1206 buf_offset += bytes;
1207 working_bytes -= bytes;
1208 current_buf_start += bytes;
1210 /* check if we need to pick another page */
1211 bio_advance(bio, bytes);
1212 if (!bio->bi_iter.bi_size)
1214 bvec = bio_iter_iovec(bio, bio->bi_iter);
1215 prev_start_byte = start_byte;
1216 start_byte = page_offset(bvec.bv_page) - disk_start;
1219 * We need to make sure we're only adjusting
1220 * our offset into compression working buffer when
1221 * we're switching pages. Otherwise we can incorrectly
1222 * keep copying when we were actually done.
1224 if (start_byte != prev_start_byte) {
1226 * make sure our new page is covered by this
1229 if (total_out <= start_byte)
1233 * the next page in the biovec might not be adjacent
1234 * to the last page, but it might still be found
1235 * inside this working buffer. bump our offset pointer
1237 if (total_out > start_byte &&
1238 current_buf_start < start_byte) {
1239 buf_offset = start_byte - buf_start;
1240 working_bytes = total_out - start_byte;
1241 current_buf_start = buf_start + buf_offset;
1250 * Shannon Entropy calculation
1252 * Pure byte distribution analysis fails to determine compressiability of data.
1253 * Try calculating entropy to estimate the average minimum number of bits
1254 * needed to encode the sampled data.
1256 * For convenience, return the percentage of needed bits, instead of amount of
1259 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1260 * and can be compressible with high probability
1262 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1264 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1266 #define ENTROPY_LVL_ACEPTABLE (65)
1267 #define ENTROPY_LVL_HIGH (80)
1270 * For increasead precision in shannon_entropy calculation,
1271 * let's do pow(n, M) to save more digits after comma:
1273 * - maximum int bit length is 64
1274 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1275 * - 13 * 4 = 52 < 64 -> M = 4
1279 static inline u32 ilog2_w(u64 n)
1281 return ilog2(n * n * n * n);
1284 static u32 shannon_entropy(struct heuristic_ws *ws)
1286 const u32 entropy_max = 8 * ilog2_w(2);
1287 u32 entropy_sum = 0;
1288 u32 p, p_base, sz_base;
1291 sz_base = ilog2_w(ws->sample_size);
1292 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1293 p = ws->bucket[i].count;
1294 p_base = ilog2_w(p);
1295 entropy_sum += p * (sz_base - p_base);
1298 entropy_sum /= ws->sample_size;
1299 return entropy_sum * 100 / entropy_max;
1302 #define RADIX_BASE 4U
1303 #define COUNTERS_SIZE (1U << RADIX_BASE)
1305 static u8 get4bits(u64 num, int shift) {
1310 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1314 static void copy_cell(void *dst, int dest_i, void *src, int src_i)
1316 struct bucket_item *dstv = (struct bucket_item *)dst;
1317 struct bucket_item *srcv = (struct bucket_item *)src;
1318 dstv[dest_i] = srcv[src_i];
1322 * Use 4 bits as radix base
1323 * Use 16 u32 counters for calculating new possition in buf array
1325 * @array - array that will be sorted
1326 * @array_buf - buffer array to store sorting results
1327 * must be equal in size to @array
1329 * @copy_cell - function to copy data from array to array_buf and vice versa
1330 * @get4bits - function to get 4 bits from number at specified offset
1332 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1334 void (*copy_cell)(void *dest, int dest_i,
1335 void* src, int src_i),
1336 u8 (*get4bits)(u64 num, int shift))
1340 u32 counters[COUNTERS_SIZE];
1348 * Try avoid useless loop iterations for small numbers stored in big
1349 * counters. Example: 48 33 4 ... in 64bit array
1351 max_num = array[0].count;
1352 for (i = 1; i < num; i++) {
1353 buf_num = array[i].count;
1354 if (buf_num > max_num)
1358 buf_num = ilog2(max_num);
1359 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1362 while (shift < bitlen) {
1363 memset(counters, 0, sizeof(counters));
1365 for (i = 0; i < num; i++) {
1366 buf_num = array[i].count;
1367 addr = get4bits(buf_num, shift);
1371 for (i = 1; i < COUNTERS_SIZE; i++)
1372 counters[i] += counters[i - 1];
1374 for (i = num - 1; i >= 0; i--) {
1375 buf_num = array[i].count;
1376 addr = get4bits(buf_num, shift);
1378 new_addr = counters[addr];
1379 copy_cell(array_buf, new_addr, array, i);
1382 shift += RADIX_BASE;
1385 * Normal radix expects to move data from a temporary array, to
1386 * the main one. But that requires some CPU time. Avoid that
1387 * by doing another sort iteration to original array instead of
1390 memset(counters, 0, sizeof(counters));
1392 for (i = 0; i < num; i ++) {
1393 buf_num = array_buf[i].count;
1394 addr = get4bits(buf_num, shift);
1398 for (i = 1; i < COUNTERS_SIZE; i++)
1399 counters[i] += counters[i - 1];
1401 for (i = num - 1; i >= 0; i--) {
1402 buf_num = array_buf[i].count;
1403 addr = get4bits(buf_num, shift);
1405 new_addr = counters[addr];
1406 copy_cell(array, new_addr, array_buf, i);
1409 shift += RADIX_BASE;
1414 * Size of the core byte set - how many bytes cover 90% of the sample
1416 * There are several types of structured binary data that use nearly all byte
1417 * values. The distribution can be uniform and counts in all buckets will be
1418 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1420 * Other possibility is normal (Gaussian) distribution, where the data could
1421 * be potentially compressible, but we have to take a few more steps to decide
1424 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1425 * compression algo can easy fix that
1426 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1427 * probability is not compressible
1429 #define BYTE_CORE_SET_LOW (64)
1430 #define BYTE_CORE_SET_HIGH (200)
1432 static int byte_core_set_size(struct heuristic_ws *ws)
1435 u32 coreset_sum = 0;
1436 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1437 struct bucket_item *bucket = ws->bucket;
1439 /* Sort in reverse order */
1440 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE, copy_cell,
1443 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1444 coreset_sum += bucket[i].count;
1446 if (coreset_sum > core_set_threshold)
1449 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1450 coreset_sum += bucket[i].count;
1451 if (coreset_sum > core_set_threshold)
1459 * Count byte values in buckets.
1460 * This heuristic can detect textual data (configs, xml, json, html, etc).
1461 * Because in most text-like data byte set is restricted to limited number of
1462 * possible characters, and that restriction in most cases makes data easy to
1465 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1466 * less - compressible
1467 * more - need additional analysis
1469 #define BYTE_SET_THRESHOLD (64)
1471 static u32 byte_set_size(const struct heuristic_ws *ws)
1474 u32 byte_set_size = 0;
1476 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1477 if (ws->bucket[i].count > 0)
1482 * Continue collecting count of byte values in buckets. If the byte
1483 * set size is bigger then the threshold, it's pointless to continue,
1484 * the detection technique would fail for this type of data.
1486 for (; i < BUCKET_SIZE; i++) {
1487 if (ws->bucket[i].count > 0) {
1489 if (byte_set_size > BYTE_SET_THRESHOLD)
1490 return byte_set_size;
1494 return byte_set_size;
1497 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1499 const u32 half_of_sample = ws->sample_size / 2;
1500 const u8 *data = ws->sample;
1502 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1505 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1506 struct heuristic_ws *ws)
1509 u64 index, index_end;
1510 u32 i, curr_sample_pos;
1514 * Compression handles the input data by chunks of 128KiB
1515 * (defined by BTRFS_MAX_UNCOMPRESSED)
1517 * We do the same for the heuristic and loop over the whole range.
1519 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1520 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1522 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1523 end = start + BTRFS_MAX_UNCOMPRESSED;
1525 index = start >> PAGE_SHIFT;
1526 index_end = end >> PAGE_SHIFT;
1528 /* Don't miss unaligned end */
1529 if (!IS_ALIGNED(end, PAGE_SIZE))
1532 curr_sample_pos = 0;
1533 while (index < index_end) {
1534 page = find_get_page(inode->i_mapping, index);
1535 in_data = kmap(page);
1536 /* Handle case where the start is not aligned to PAGE_SIZE */
1537 i = start % PAGE_SIZE;
1538 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1539 /* Don't sample any garbage from the last page */
1540 if (start > end - SAMPLING_READ_SIZE)
1542 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1543 SAMPLING_READ_SIZE);
1544 i += SAMPLING_INTERVAL;
1545 start += SAMPLING_INTERVAL;
1546 curr_sample_pos += SAMPLING_READ_SIZE;
1554 ws->sample_size = curr_sample_pos;
1558 * Compression heuristic.
1560 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1561 * quickly (compared to direct compression) detect data characteristics
1562 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1565 * The following types of analysis can be performed:
1566 * - detect mostly zero data
1567 * - detect data with low "byte set" size (text, etc)
1568 * - detect data with low/high "core byte" set
1570 * Return non-zero if the compression should be done, 0 otherwise.
1572 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1574 struct list_head *ws_list = __find_workspace(0, true);
1575 struct heuristic_ws *ws;
1580 ws = list_entry(ws_list, struct heuristic_ws, list);
1582 heuristic_collect_sample(inode, start, end, ws);
1584 if (sample_repeated_patterns(ws)) {
1589 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1591 for (i = 0; i < ws->sample_size; i++) {
1592 byte = ws->sample[i];
1593 ws->bucket[byte].count++;
1596 i = byte_set_size(ws);
1597 if (i < BYTE_SET_THRESHOLD) {
1602 i = byte_core_set_size(ws);
1603 if (i <= BYTE_CORE_SET_LOW) {
1608 if (i >= BYTE_CORE_SET_HIGH) {
1613 i = shannon_entropy(ws);
1614 if (i <= ENTROPY_LVL_ACEPTABLE) {
1620 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1621 * needed to give green light to compression.
1623 * For now just assume that compression at that level is not worth the
1624 * resources because:
1626 * 1. it is possible to defrag the data later
1628 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1629 * values, every bucket has counter at level ~54. The heuristic would
1630 * be confused. This can happen when data have some internal repeated
1631 * patterns like "abbacbbc...". This can be detected by analyzing
1632 * pairs of bytes, which is too costly.
1634 if (i < ENTROPY_LVL_HIGH) {
1643 __free_workspace(0, ws_list, true);
1647 unsigned int btrfs_compress_str2level(const char *str)
1649 if (strncmp(str, "zlib", 4) != 0)
1652 /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */
1653 if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0)
1654 return str[5] - '0';
1656 return BTRFS_ZLIB_DEFAULT_LEVEL;