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 compressed_bio *cb = bio->bi_private;
240 /* if there are more bios still pending for this compressed
243 if (!refcount_dec_and_test(&cb->pending_bios))
246 /* ok, we're the last bio for this extent, step one is to
247 * call back into the FS and do all the end_io operations
250 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
251 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
252 cb->start, cb->start + cb->len - 1,
253 bio->bi_status ? BLK_STS_OK : BLK_STS_NOTSUPP);
254 cb->compressed_pages[0]->mapping = NULL;
256 end_compressed_writeback(inode, cb);
257 /* note, our inode could be gone now */
260 * release the compressed pages, these came from alloc_page and
261 * are not attached to the inode at all
264 for (index = 0; index < cb->nr_pages; index++) {
265 page = cb->compressed_pages[index];
266 page->mapping = NULL;
270 /* finally free the cb struct */
271 kfree(cb->compressed_pages);
278 * worker function to build and submit bios for previously compressed pages.
279 * The corresponding pages in the inode should be marked for writeback
280 * and the compressed pages should have a reference on them for dropping
281 * when the IO is complete.
283 * This also checksums the file bytes and gets things ready for
286 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
287 unsigned long len, u64 disk_start,
288 unsigned long compressed_len,
289 struct page **compressed_pages,
290 unsigned long nr_pages,
291 unsigned int write_flags)
293 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
294 struct bio *bio = NULL;
295 struct compressed_bio *cb;
296 unsigned long bytes_left;
299 u64 first_byte = disk_start;
300 struct block_device *bdev;
302 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
304 WARN_ON(!PAGE_ALIGNED(start));
305 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
307 return BLK_STS_RESOURCE;
308 refcount_set(&cb->pending_bios, 0);
314 cb->compressed_pages = compressed_pages;
315 cb->compressed_len = compressed_len;
317 cb->nr_pages = nr_pages;
319 bdev = fs_info->fs_devices->latest_bdev;
321 bio = btrfs_bio_alloc(bdev, first_byte);
322 bio->bi_opf = REQ_OP_WRITE | write_flags;
323 bio->bi_private = cb;
324 bio->bi_end_io = end_compressed_bio_write;
325 refcount_set(&cb->pending_bios, 1);
327 /* create and submit bios for the compressed pages */
328 bytes_left = compressed_len;
329 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
332 page = compressed_pages[pg_index];
333 page->mapping = inode->i_mapping;
334 if (bio->bi_iter.bi_size)
335 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
338 page->mapping = NULL;
339 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
342 * inc the count before we submit the bio so
343 * we know the end IO handler won't happen before
344 * we inc the count. Otherwise, the cb might get
345 * freed before we're done setting it up
347 refcount_inc(&cb->pending_bios);
348 ret = btrfs_bio_wq_end_io(fs_info, bio,
349 BTRFS_WQ_ENDIO_DATA);
350 BUG_ON(ret); /* -ENOMEM */
353 ret = btrfs_csum_one_bio(inode, bio, start, 1);
354 BUG_ON(ret); /* -ENOMEM */
357 ret = btrfs_map_bio(fs_info, bio, 0, 1);
359 bio->bi_status = ret;
363 bio = btrfs_bio_alloc(bdev, first_byte);
364 bio->bi_opf = REQ_OP_WRITE | write_flags;
365 bio->bi_private = cb;
366 bio->bi_end_io = end_compressed_bio_write;
367 bio_add_page(bio, page, PAGE_SIZE, 0);
369 if (bytes_left < PAGE_SIZE) {
371 "bytes left %lu compress len %lu nr %lu",
372 bytes_left, cb->compressed_len, cb->nr_pages);
374 bytes_left -= PAGE_SIZE;
375 first_byte += PAGE_SIZE;
379 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
380 BUG_ON(ret); /* -ENOMEM */
383 ret = btrfs_csum_one_bio(inode, bio, start, 1);
384 BUG_ON(ret); /* -ENOMEM */
387 ret = btrfs_map_bio(fs_info, bio, 0, 1);
389 bio->bi_status = ret;
396 static u64 bio_end_offset(struct bio *bio)
398 struct bio_vec *last = bio_last_bvec_all(bio);
400 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
403 static noinline int add_ra_bio_pages(struct inode *inode,
405 struct compressed_bio *cb)
407 unsigned long end_index;
408 unsigned long pg_index;
410 u64 isize = i_size_read(inode);
413 unsigned long nr_pages = 0;
414 struct extent_map *em;
415 struct address_space *mapping = inode->i_mapping;
416 struct extent_map_tree *em_tree;
417 struct extent_io_tree *tree;
421 last_offset = bio_end_offset(cb->orig_bio);
422 em_tree = &BTRFS_I(inode)->extent_tree;
423 tree = &BTRFS_I(inode)->io_tree;
428 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
430 while (last_offset < compressed_end) {
431 pg_index = last_offset >> PAGE_SHIFT;
433 if (pg_index > end_index)
436 page = xa_load(&mapping->i_pages, pg_index);
437 if (page && !xa_is_value(page)) {
444 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
449 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
454 end = last_offset + PAGE_SIZE - 1;
456 * at this point, we have a locked page in the page cache
457 * for these bytes in the file. But, we have to make
458 * sure they map to this compressed extent on disk.
460 set_page_extent_mapped(page);
461 lock_extent(tree, last_offset, end);
462 read_lock(&em_tree->lock);
463 em = lookup_extent_mapping(em_tree, last_offset,
465 read_unlock(&em_tree->lock);
467 if (!em || last_offset < em->start ||
468 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
469 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
471 unlock_extent(tree, last_offset, end);
478 if (page->index == end_index) {
480 size_t zero_offset = offset_in_page(isize);
484 zeros = PAGE_SIZE - zero_offset;
485 userpage = kmap_atomic(page);
486 memset(userpage + zero_offset, 0, zeros);
487 flush_dcache_page(page);
488 kunmap_atomic(userpage);
492 ret = bio_add_page(cb->orig_bio, page,
495 if (ret == PAGE_SIZE) {
499 unlock_extent(tree, last_offset, end);
505 last_offset += PAGE_SIZE;
511 * for a compressed read, the bio we get passed has all the inode pages
512 * in it. We don't actually do IO on those pages but allocate new ones
513 * to hold the compressed pages on disk.
515 * bio->bi_iter.bi_sector points to the compressed extent on disk
516 * bio->bi_io_vec points to all of the inode pages
518 * After the compressed pages are read, we copy the bytes into the
519 * bio we were passed and then call the bio end_io calls
521 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
522 int mirror_num, unsigned long bio_flags)
524 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
525 struct extent_map_tree *em_tree;
526 struct compressed_bio *cb;
527 unsigned long compressed_len;
528 unsigned long nr_pages;
529 unsigned long pg_index;
531 struct block_device *bdev;
532 struct bio *comp_bio;
533 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
536 struct extent_map *em;
537 blk_status_t ret = BLK_STS_RESOURCE;
541 em_tree = &BTRFS_I(inode)->extent_tree;
543 /* we need the actual starting offset of this extent in the file */
544 read_lock(&em_tree->lock);
545 em = lookup_extent_mapping(em_tree,
546 page_offset(bio_first_page_all(bio)),
548 read_unlock(&em_tree->lock);
550 return BLK_STS_IOERR;
552 compressed_len = em->block_len;
553 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
557 refcount_set(&cb->pending_bios, 0);
560 cb->mirror_num = mirror_num;
563 cb->start = em->orig_start;
565 em_start = em->start;
570 cb->len = bio->bi_iter.bi_size;
571 cb->compressed_len = compressed_len;
572 cb->compress_type = extent_compress_type(bio_flags);
575 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
576 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
578 if (!cb->compressed_pages)
581 bdev = fs_info->fs_devices->latest_bdev;
583 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
584 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
586 if (!cb->compressed_pages[pg_index]) {
587 faili = pg_index - 1;
588 ret = BLK_STS_RESOURCE;
592 faili = nr_pages - 1;
593 cb->nr_pages = nr_pages;
595 add_ra_bio_pages(inode, em_start + em_len, cb);
597 /* include any pages we added in add_ra-bio_pages */
598 cb->len = bio->bi_iter.bi_size;
600 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
601 comp_bio->bi_opf = REQ_OP_READ;
602 comp_bio->bi_private = cb;
603 comp_bio->bi_end_io = end_compressed_bio_read;
604 refcount_set(&cb->pending_bios, 1);
606 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
609 page = cb->compressed_pages[pg_index];
610 page->mapping = inode->i_mapping;
611 page->index = em_start >> PAGE_SHIFT;
613 if (comp_bio->bi_iter.bi_size)
614 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
617 page->mapping = NULL;
618 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
620 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
621 BTRFS_WQ_ENDIO_DATA);
622 BUG_ON(ret); /* -ENOMEM */
625 * inc the count before we submit the bio so
626 * we know the end IO handler won't happen before
627 * we inc the count. Otherwise, the cb might get
628 * freed before we're done setting it up
630 refcount_inc(&cb->pending_bios);
632 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
633 ret = btrfs_lookup_bio_sums(inode, comp_bio,
635 BUG_ON(ret); /* -ENOMEM */
637 sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
638 fs_info->sectorsize);
640 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
642 comp_bio->bi_status = ret;
646 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
647 comp_bio->bi_opf = REQ_OP_READ;
648 comp_bio->bi_private = cb;
649 comp_bio->bi_end_io = end_compressed_bio_read;
651 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
653 cur_disk_byte += PAGE_SIZE;
656 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
657 BUG_ON(ret); /* -ENOMEM */
659 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
660 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
661 BUG_ON(ret); /* -ENOMEM */
664 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
666 comp_bio->bi_status = ret;
674 __free_page(cb->compressed_pages[faili]);
678 kfree(cb->compressed_pages);
687 * Heuristic uses systematic sampling to collect data from the input data
688 * range, the logic can be tuned by the following constants:
690 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
691 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
693 #define SAMPLING_READ_SIZE (16)
694 #define SAMPLING_INTERVAL (256)
697 * For statistical analysis of the input data we consider bytes that form a
698 * Galois Field of 256 objects. Each object has an attribute count, ie. how
699 * many times the object appeared in the sample.
701 #define BUCKET_SIZE (256)
704 * The size of the sample is based on a statistical sampling rule of thumb.
705 * The common way is to perform sampling tests as long as the number of
706 * elements in each cell is at least 5.
708 * Instead of 5, we choose 32 to obtain more accurate results.
709 * If the data contain the maximum number of symbols, which is 256, we obtain a
710 * sample size bound by 8192.
712 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
713 * from up to 512 locations.
715 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
716 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
722 struct heuristic_ws {
723 /* Partial copy of input data */
726 /* Buckets store counters for each byte value */
727 struct bucket_item *bucket;
729 struct bucket_item *bucket_b;
730 struct list_head list;
733 static void free_heuristic_ws(struct list_head *ws)
735 struct heuristic_ws *workspace;
737 workspace = list_entry(ws, struct heuristic_ws, list);
739 kvfree(workspace->sample);
740 kfree(workspace->bucket);
741 kfree(workspace->bucket_b);
745 static struct list_head *alloc_heuristic_ws(void)
747 struct heuristic_ws *ws;
749 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
751 return ERR_PTR(-ENOMEM);
753 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
757 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
761 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
765 INIT_LIST_HEAD(&ws->list);
768 free_heuristic_ws(&ws->list);
769 return ERR_PTR(-ENOMEM);
772 const struct btrfs_compress_op btrfs_heuristic_compress = {
773 .alloc_workspace = alloc_heuristic_ws,
774 .free_workspace = free_heuristic_ws,
777 struct workspace_manager {
778 const struct btrfs_compress_op *ops;
779 struct list_head idle_ws;
781 /* Number of free workspaces */
783 /* Total number of allocated workspaces */
785 /* Waiters for a free workspace */
786 wait_queue_head_t ws_wait;
789 static struct workspace_manager wsm[BTRFS_NR_WORKSPACE_MANAGERS];
791 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
792 /* The heuristic is represented as compression type 0 */
793 &btrfs_heuristic_compress,
794 &btrfs_zlib_compress,
796 &btrfs_zstd_compress,
799 static void btrfs_init_workspace_manager(int type)
801 struct workspace_manager *wsman = &wsm[type];
802 struct list_head *workspace;
804 wsman->ops = btrfs_compress_op[type];
806 INIT_LIST_HEAD(&wsman->idle_ws);
807 spin_lock_init(&wsman->ws_lock);
808 atomic_set(&wsman->total_ws, 0);
809 init_waitqueue_head(&wsman->ws_wait);
812 * Preallocate one workspace for each compression type so we can
813 * guarantee forward progress in the worst case
815 workspace = wsman->ops->alloc_workspace();
816 if (IS_ERR(workspace)) {
818 "BTRFS: cannot preallocate compression workspace, will try later\n");
820 atomic_set(&wsman->total_ws, 1);
822 list_add(workspace, &wsman->idle_ws);
826 static void btrfs_cleanup_workspace_manager(struct workspace_manager *wsman)
828 struct list_head *ws;
830 while (!list_empty(&wsman->idle_ws)) {
831 ws = wsman->idle_ws.next;
833 wsman->ops->free_workspace(ws);
834 atomic_dec(&wsman->total_ws);
839 * This finds an available workspace or allocates a new one.
840 * If it's not possible to allocate a new one, waits until there's one.
841 * Preallocation makes a forward progress guarantees and we do not return
844 static struct list_head *find_workspace(int type)
846 struct list_head *workspace;
847 int cpus = num_online_cpus();
849 struct list_head *idle_ws;
852 wait_queue_head_t *ws_wait;
855 idle_ws = &wsm[type].idle_ws;
856 ws_lock = &wsm[type].ws_lock;
857 total_ws = &wsm[type].total_ws;
858 ws_wait = &wsm[type].ws_wait;
859 free_ws = &wsm[type].free_ws;
863 if (!list_empty(idle_ws)) {
864 workspace = idle_ws->next;
867 spin_unlock(ws_lock);
871 if (atomic_read(total_ws) > cpus) {
874 spin_unlock(ws_lock);
875 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
876 if (atomic_read(total_ws) > cpus && !*free_ws)
878 finish_wait(ws_wait, &wait);
881 atomic_inc(total_ws);
882 spin_unlock(ws_lock);
885 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
886 * to turn it off here because we might get called from the restricted
887 * context of btrfs_compress_bio/btrfs_compress_pages
889 nofs_flag = memalloc_nofs_save();
890 workspace = wsm[type].ops->alloc_workspace();
891 memalloc_nofs_restore(nofs_flag);
893 if (IS_ERR(workspace)) {
894 atomic_dec(total_ws);
898 * Do not return the error but go back to waiting. There's a
899 * workspace preallocated for each type and the compression
900 * time is bounded so we get to a workspace eventually. This
901 * makes our caller's life easier.
903 * To prevent silent and low-probability deadlocks (when the
904 * initial preallocation fails), check if there are any
907 if (atomic_read(total_ws) == 0) {
908 static DEFINE_RATELIMIT_STATE(_rs,
909 /* once per minute */ 60 * HZ,
912 if (__ratelimit(&_rs)) {
913 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
922 * put a workspace struct back on the list or free it if we have enough
923 * idle ones sitting around
925 static void free_workspace(int type, struct list_head *workspace)
927 struct list_head *idle_ws;
930 wait_queue_head_t *ws_wait;
933 idle_ws = &wsm[type].idle_ws;
934 ws_lock = &wsm[type].ws_lock;
935 total_ws = &wsm[type].total_ws;
936 ws_wait = &wsm[type].ws_wait;
937 free_ws = &wsm[type].free_ws;
940 if (*free_ws <= num_online_cpus()) {
941 list_add(workspace, idle_ws);
943 spin_unlock(ws_lock);
946 spin_unlock(ws_lock);
948 wsm[type].ops->free_workspace(workspace);
949 atomic_dec(total_ws);
951 cond_wake_up(ws_wait);
955 * Given an address space and start and length, compress the bytes into @pages
956 * that are allocated on demand.
958 * @type_level is encoded algorithm and level, where level 0 means whatever
959 * default the algorithm chooses and is opaque here;
960 * - compression algo are 0-3
961 * - the level are bits 4-7
963 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
964 * and returns number of actually allocated pages
966 * @total_in is used to return the number of bytes actually read. It
967 * may be smaller than the input length if we had to exit early because we
968 * ran out of room in the pages array or because we cross the
971 * @total_out is an in/out parameter, must be set to the input length and will
972 * be also used to return the total number of compressed bytes
974 * @max_out tells us the max number of bytes that we're allowed to
977 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
978 u64 start, struct page **pages,
979 unsigned long *out_pages,
980 unsigned long *total_in,
981 unsigned long *total_out)
983 int type = btrfs_compress_type(type_level);
984 struct list_head *workspace;
987 workspace = find_workspace(type);
989 btrfs_compress_op[type]->set_level(workspace, type_level);
990 ret = btrfs_compress_op[type]->compress_pages(workspace, mapping,
993 total_in, total_out);
994 free_workspace(type, workspace);
999 * pages_in is an array of pages with compressed data.
1001 * disk_start is the starting logical offset of this array in the file
1003 * orig_bio contains the pages from the file that we want to decompress into
1005 * srclen is the number of bytes in pages_in
1007 * The basic idea is that we have a bio that was created by readpages.
1008 * The pages in the bio are for the uncompressed data, and they may not
1009 * be contiguous. They all correspond to the range of bytes covered by
1010 * the compressed extent.
1012 static int btrfs_decompress_bio(struct compressed_bio *cb)
1014 struct list_head *workspace;
1016 int type = cb->compress_type;
1018 workspace = find_workspace(type);
1019 ret = btrfs_compress_op[type]->decompress_bio(workspace, cb);
1020 free_workspace(type, workspace);
1026 * a less complex decompression routine. Our compressed data fits in a
1027 * single page, and we want to read a single page out of it.
1028 * start_byte tells us the offset into the compressed data we're interested in
1030 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1031 unsigned long start_byte, size_t srclen, size_t destlen)
1033 struct list_head *workspace;
1036 workspace = find_workspace(type);
1038 ret = btrfs_compress_op[type]->decompress(workspace, data_in,
1039 dest_page, start_byte,
1042 free_workspace(type, workspace);
1046 void __init btrfs_init_compress(void)
1050 for (i = 0; i < BTRFS_NR_WORKSPACE_MANAGERS; i++)
1051 btrfs_init_workspace_manager(i);
1054 void __cold btrfs_exit_compress(void)
1058 for (i = 0; i < BTRFS_NR_WORKSPACE_MANAGERS; i++)
1059 btrfs_cleanup_workspace_manager(&wsm[i]);
1063 * Copy uncompressed data from working buffer to pages.
1065 * buf_start is the byte offset we're of the start of our workspace buffer.
1067 * total_out is the last byte of the buffer
1069 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1070 unsigned long total_out, u64 disk_start,
1073 unsigned long buf_offset;
1074 unsigned long current_buf_start;
1075 unsigned long start_byte;
1076 unsigned long prev_start_byte;
1077 unsigned long working_bytes = total_out - buf_start;
1078 unsigned long bytes;
1080 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1083 * start byte is the first byte of the page we're currently
1084 * copying into relative to the start of the compressed data.
1086 start_byte = page_offset(bvec.bv_page) - disk_start;
1088 /* we haven't yet hit data corresponding to this page */
1089 if (total_out <= start_byte)
1093 * the start of the data we care about is offset into
1094 * the middle of our working buffer
1096 if (total_out > start_byte && buf_start < start_byte) {
1097 buf_offset = start_byte - buf_start;
1098 working_bytes -= buf_offset;
1102 current_buf_start = buf_start;
1104 /* copy bytes from the working buffer into the pages */
1105 while (working_bytes > 0) {
1106 bytes = min_t(unsigned long, bvec.bv_len,
1107 PAGE_SIZE - buf_offset);
1108 bytes = min(bytes, working_bytes);
1110 kaddr = kmap_atomic(bvec.bv_page);
1111 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1112 kunmap_atomic(kaddr);
1113 flush_dcache_page(bvec.bv_page);
1115 buf_offset += bytes;
1116 working_bytes -= bytes;
1117 current_buf_start += bytes;
1119 /* check if we need to pick another page */
1120 bio_advance(bio, bytes);
1121 if (!bio->bi_iter.bi_size)
1123 bvec = bio_iter_iovec(bio, bio->bi_iter);
1124 prev_start_byte = start_byte;
1125 start_byte = page_offset(bvec.bv_page) - disk_start;
1128 * We need to make sure we're only adjusting
1129 * our offset into compression working buffer when
1130 * we're switching pages. Otherwise we can incorrectly
1131 * keep copying when we were actually done.
1133 if (start_byte != prev_start_byte) {
1135 * make sure our new page is covered by this
1138 if (total_out <= start_byte)
1142 * the next page in the biovec might not be adjacent
1143 * to the last page, but it might still be found
1144 * inside this working buffer. bump our offset pointer
1146 if (total_out > start_byte &&
1147 current_buf_start < start_byte) {
1148 buf_offset = start_byte - buf_start;
1149 working_bytes = total_out - start_byte;
1150 current_buf_start = buf_start + buf_offset;
1159 * Shannon Entropy calculation
1161 * Pure byte distribution analysis fails to determine compressibility of data.
1162 * Try calculating entropy to estimate the average minimum number of bits
1163 * needed to encode the sampled data.
1165 * For convenience, return the percentage of needed bits, instead of amount of
1168 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1169 * and can be compressible with high probability
1171 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1173 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1175 #define ENTROPY_LVL_ACEPTABLE (65)
1176 #define ENTROPY_LVL_HIGH (80)
1179 * For increasead precision in shannon_entropy calculation,
1180 * let's do pow(n, M) to save more digits after comma:
1182 * - maximum int bit length is 64
1183 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1184 * - 13 * 4 = 52 < 64 -> M = 4
1188 static inline u32 ilog2_w(u64 n)
1190 return ilog2(n * n * n * n);
1193 static u32 shannon_entropy(struct heuristic_ws *ws)
1195 const u32 entropy_max = 8 * ilog2_w(2);
1196 u32 entropy_sum = 0;
1197 u32 p, p_base, sz_base;
1200 sz_base = ilog2_w(ws->sample_size);
1201 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1202 p = ws->bucket[i].count;
1203 p_base = ilog2_w(p);
1204 entropy_sum += p * (sz_base - p_base);
1207 entropy_sum /= ws->sample_size;
1208 return entropy_sum * 100 / entropy_max;
1211 #define RADIX_BASE 4U
1212 #define COUNTERS_SIZE (1U << RADIX_BASE)
1214 static u8 get4bits(u64 num, int shift) {
1219 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1224 * Use 4 bits as radix base
1225 * Use 16 u32 counters for calculating new position in buf array
1227 * @array - array that will be sorted
1228 * @array_buf - buffer array to store sorting results
1229 * must be equal in size to @array
1232 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1237 u32 counters[COUNTERS_SIZE];
1245 * Try avoid useless loop iterations for small numbers stored in big
1246 * counters. Example: 48 33 4 ... in 64bit array
1248 max_num = array[0].count;
1249 for (i = 1; i < num; i++) {
1250 buf_num = array[i].count;
1251 if (buf_num > max_num)
1255 buf_num = ilog2(max_num);
1256 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1259 while (shift < bitlen) {
1260 memset(counters, 0, sizeof(counters));
1262 for (i = 0; i < num; i++) {
1263 buf_num = array[i].count;
1264 addr = get4bits(buf_num, shift);
1268 for (i = 1; i < COUNTERS_SIZE; i++)
1269 counters[i] += counters[i - 1];
1271 for (i = num - 1; i >= 0; i--) {
1272 buf_num = array[i].count;
1273 addr = get4bits(buf_num, shift);
1275 new_addr = counters[addr];
1276 array_buf[new_addr] = array[i];
1279 shift += RADIX_BASE;
1282 * Normal radix expects to move data from a temporary array, to
1283 * the main one. But that requires some CPU time. Avoid that
1284 * by doing another sort iteration to original array instead of
1287 memset(counters, 0, sizeof(counters));
1289 for (i = 0; i < num; i ++) {
1290 buf_num = array_buf[i].count;
1291 addr = get4bits(buf_num, shift);
1295 for (i = 1; i < COUNTERS_SIZE; i++)
1296 counters[i] += counters[i - 1];
1298 for (i = num - 1; i >= 0; i--) {
1299 buf_num = array_buf[i].count;
1300 addr = get4bits(buf_num, shift);
1302 new_addr = counters[addr];
1303 array[new_addr] = array_buf[i];
1306 shift += RADIX_BASE;
1311 * Size of the core byte set - how many bytes cover 90% of the sample
1313 * There are several types of structured binary data that use nearly all byte
1314 * values. The distribution can be uniform and counts in all buckets will be
1315 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1317 * Other possibility is normal (Gaussian) distribution, where the data could
1318 * be potentially compressible, but we have to take a few more steps to decide
1321 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1322 * compression algo can easy fix that
1323 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1324 * probability is not compressible
1326 #define BYTE_CORE_SET_LOW (64)
1327 #define BYTE_CORE_SET_HIGH (200)
1329 static int byte_core_set_size(struct heuristic_ws *ws)
1332 u32 coreset_sum = 0;
1333 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1334 struct bucket_item *bucket = ws->bucket;
1336 /* Sort in reverse order */
1337 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1339 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1340 coreset_sum += bucket[i].count;
1342 if (coreset_sum > core_set_threshold)
1345 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1346 coreset_sum += bucket[i].count;
1347 if (coreset_sum > core_set_threshold)
1355 * Count byte values in buckets.
1356 * This heuristic can detect textual data (configs, xml, json, html, etc).
1357 * Because in most text-like data byte set is restricted to limited number of
1358 * possible characters, and that restriction in most cases makes data easy to
1361 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1362 * less - compressible
1363 * more - need additional analysis
1365 #define BYTE_SET_THRESHOLD (64)
1367 static u32 byte_set_size(const struct heuristic_ws *ws)
1370 u32 byte_set_size = 0;
1372 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1373 if (ws->bucket[i].count > 0)
1378 * Continue collecting count of byte values in buckets. If the byte
1379 * set size is bigger then the threshold, it's pointless to continue,
1380 * the detection technique would fail for this type of data.
1382 for (; i < BUCKET_SIZE; i++) {
1383 if (ws->bucket[i].count > 0) {
1385 if (byte_set_size > BYTE_SET_THRESHOLD)
1386 return byte_set_size;
1390 return byte_set_size;
1393 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1395 const u32 half_of_sample = ws->sample_size / 2;
1396 const u8 *data = ws->sample;
1398 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1401 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1402 struct heuristic_ws *ws)
1405 u64 index, index_end;
1406 u32 i, curr_sample_pos;
1410 * Compression handles the input data by chunks of 128KiB
1411 * (defined by BTRFS_MAX_UNCOMPRESSED)
1413 * We do the same for the heuristic and loop over the whole range.
1415 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1416 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1418 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1419 end = start + BTRFS_MAX_UNCOMPRESSED;
1421 index = start >> PAGE_SHIFT;
1422 index_end = end >> PAGE_SHIFT;
1424 /* Don't miss unaligned end */
1425 if (!IS_ALIGNED(end, PAGE_SIZE))
1428 curr_sample_pos = 0;
1429 while (index < index_end) {
1430 page = find_get_page(inode->i_mapping, index);
1431 in_data = kmap(page);
1432 /* Handle case where the start is not aligned to PAGE_SIZE */
1433 i = start % PAGE_SIZE;
1434 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1435 /* Don't sample any garbage from the last page */
1436 if (start > end - SAMPLING_READ_SIZE)
1438 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1439 SAMPLING_READ_SIZE);
1440 i += SAMPLING_INTERVAL;
1441 start += SAMPLING_INTERVAL;
1442 curr_sample_pos += SAMPLING_READ_SIZE;
1450 ws->sample_size = curr_sample_pos;
1454 * Compression heuristic.
1456 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1457 * quickly (compared to direct compression) detect data characteristics
1458 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1461 * The following types of analysis can be performed:
1462 * - detect mostly zero data
1463 * - detect data with low "byte set" size (text, etc)
1464 * - detect data with low/high "core byte" set
1466 * Return non-zero if the compression should be done, 0 otherwise.
1468 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1470 struct list_head *ws_list = find_workspace(0);
1471 struct heuristic_ws *ws;
1476 ws = list_entry(ws_list, struct heuristic_ws, list);
1478 heuristic_collect_sample(inode, start, end, ws);
1480 if (sample_repeated_patterns(ws)) {
1485 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1487 for (i = 0; i < ws->sample_size; i++) {
1488 byte = ws->sample[i];
1489 ws->bucket[byte].count++;
1492 i = byte_set_size(ws);
1493 if (i < BYTE_SET_THRESHOLD) {
1498 i = byte_core_set_size(ws);
1499 if (i <= BYTE_CORE_SET_LOW) {
1504 if (i >= BYTE_CORE_SET_HIGH) {
1509 i = shannon_entropy(ws);
1510 if (i <= ENTROPY_LVL_ACEPTABLE) {
1516 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1517 * needed to give green light to compression.
1519 * For now just assume that compression at that level is not worth the
1520 * resources because:
1522 * 1. it is possible to defrag the data later
1524 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1525 * values, every bucket has counter at level ~54. The heuristic would
1526 * be confused. This can happen when data have some internal repeated
1527 * patterns like "abbacbbc...". This can be detected by analyzing
1528 * pairs of bytes, which is too costly.
1530 if (i < ENTROPY_LVL_HIGH) {
1539 free_workspace(0, ws_list);
1543 unsigned int btrfs_compress_str2level(const char *str)
1545 if (strncmp(str, "zlib", 4) != 0)
1548 /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */
1549 if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0)
1550 return str[5] - '0';
1552 return BTRFS_ZLIB_DEFAULT_LEVEL;