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;
165 struct bvec_iter_all iter_all;
168 * we have verified the checksum already, set page
169 * checked so the end_io handlers know about it
171 ASSERT(!bio_flagged(bio, BIO_CLONED));
172 bio_for_each_segment_all(bvec, cb->orig_bio, i, iter_all)
173 SetPageChecked(bvec->bv_page);
175 bio_endio(cb->orig_bio);
178 /* finally free the cb struct */
179 kfree(cb->compressed_pages);
186 * Clear the writeback bits on all of the file
187 * pages for a compressed write
189 static noinline void end_compressed_writeback(struct inode *inode,
190 const struct compressed_bio *cb)
192 unsigned long index = cb->start >> PAGE_SHIFT;
193 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
194 struct page *pages[16];
195 unsigned long nr_pages = end_index - index + 1;
200 mapping_set_error(inode->i_mapping, -EIO);
202 while (nr_pages > 0) {
203 ret = find_get_pages_contig(inode->i_mapping, index,
205 nr_pages, ARRAY_SIZE(pages)), pages);
211 for (i = 0; i < ret; i++) {
213 SetPageError(pages[i]);
214 end_page_writeback(pages[i]);
220 /* the inode may be gone now */
224 * do the cleanup once all the compressed pages hit the disk.
225 * This will clear writeback on the file pages and free the compressed
228 * This also calls the writeback end hooks for the file pages so that
229 * metadata and checksums can be updated in the file.
231 static void end_compressed_bio_write(struct bio *bio)
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 cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
252 btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0],
253 cb->start, cb->start + cb->len - 1,
254 bio->bi_status ? BLK_STS_OK : BLK_STS_NOTSUPP);
255 cb->compressed_pages[0]->mapping = NULL;
257 end_compressed_writeback(inode, cb);
258 /* note, our inode could be gone now */
261 * release the compressed pages, these came from alloc_page and
262 * are not attached to the inode at all
265 for (index = 0; index < cb->nr_pages; index++) {
266 page = cb->compressed_pages[index];
267 page->mapping = NULL;
271 /* finally free the cb struct */
272 kfree(cb->compressed_pages);
279 * worker function to build and submit bios for previously compressed pages.
280 * The corresponding pages in the inode should be marked for writeback
281 * and the compressed pages should have a reference on them for dropping
282 * when the IO is complete.
284 * This also checksums the file bytes and gets things ready for
287 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
288 unsigned long len, u64 disk_start,
289 unsigned long compressed_len,
290 struct page **compressed_pages,
291 unsigned long nr_pages,
292 unsigned int write_flags)
294 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
295 struct bio *bio = NULL;
296 struct compressed_bio *cb;
297 unsigned long bytes_left;
300 u64 first_byte = disk_start;
301 struct block_device *bdev;
303 int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
305 WARN_ON(!PAGE_ALIGNED(start));
306 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
308 return BLK_STS_RESOURCE;
309 refcount_set(&cb->pending_bios, 0);
315 cb->compressed_pages = compressed_pages;
316 cb->compressed_len = compressed_len;
318 cb->nr_pages = nr_pages;
320 bdev = fs_info->fs_devices->latest_bdev;
322 bio = btrfs_bio_alloc(bdev, first_byte);
323 bio->bi_opf = REQ_OP_WRITE | write_flags;
324 bio->bi_private = cb;
325 bio->bi_end_io = end_compressed_bio_write;
326 refcount_set(&cb->pending_bios, 1);
328 /* create and submit bios for the compressed pages */
329 bytes_left = compressed_len;
330 for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
333 page = compressed_pages[pg_index];
334 page->mapping = inode->i_mapping;
335 if (bio->bi_iter.bi_size)
336 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
339 page->mapping = NULL;
340 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
343 * inc the count before we submit the bio so
344 * we know the end IO handler won't happen before
345 * we inc the count. Otherwise, the cb might get
346 * freed before we're done setting it up
348 refcount_inc(&cb->pending_bios);
349 ret = btrfs_bio_wq_end_io(fs_info, bio,
350 BTRFS_WQ_ENDIO_DATA);
351 BUG_ON(ret); /* -ENOMEM */
354 ret = btrfs_csum_one_bio(inode, bio, start, 1);
355 BUG_ON(ret); /* -ENOMEM */
358 ret = btrfs_map_bio(fs_info, bio, 0, 1);
360 bio->bi_status = ret;
364 bio = btrfs_bio_alloc(bdev, first_byte);
365 bio->bi_opf = REQ_OP_WRITE | write_flags;
366 bio->bi_private = cb;
367 bio->bi_end_io = end_compressed_bio_write;
368 bio_add_page(bio, page, PAGE_SIZE, 0);
370 if (bytes_left < PAGE_SIZE) {
372 "bytes left %lu compress len %lu nr %lu",
373 bytes_left, cb->compressed_len, cb->nr_pages);
375 bytes_left -= PAGE_SIZE;
376 first_byte += PAGE_SIZE;
380 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
381 BUG_ON(ret); /* -ENOMEM */
384 ret = btrfs_csum_one_bio(inode, bio, start, 1);
385 BUG_ON(ret); /* -ENOMEM */
388 ret = btrfs_map_bio(fs_info, bio, 0, 1);
390 bio->bi_status = ret;
397 static u64 bio_end_offset(struct bio *bio)
399 struct bio_vec *last = bio_last_bvec_all(bio);
401 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
404 static noinline int add_ra_bio_pages(struct inode *inode,
406 struct compressed_bio *cb)
408 unsigned long end_index;
409 unsigned long pg_index;
411 u64 isize = i_size_read(inode);
414 unsigned long nr_pages = 0;
415 struct extent_map *em;
416 struct address_space *mapping = inode->i_mapping;
417 struct extent_map_tree *em_tree;
418 struct extent_io_tree *tree;
422 last_offset = bio_end_offset(cb->orig_bio);
423 em_tree = &BTRFS_I(inode)->extent_tree;
424 tree = &BTRFS_I(inode)->io_tree;
429 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
431 while (last_offset < compressed_end) {
432 pg_index = last_offset >> PAGE_SHIFT;
434 if (pg_index > end_index)
437 page = xa_load(&mapping->i_pages, pg_index);
438 if (page && !xa_is_value(page)) {
445 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
450 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
455 end = last_offset + PAGE_SIZE - 1;
457 * at this point, we have a locked page in the page cache
458 * for these bytes in the file. But, we have to make
459 * sure they map to this compressed extent on disk.
461 set_page_extent_mapped(page);
462 lock_extent(tree, last_offset, end);
463 read_lock(&em_tree->lock);
464 em = lookup_extent_mapping(em_tree, last_offset,
466 read_unlock(&em_tree->lock);
468 if (!em || last_offset < em->start ||
469 (last_offset + PAGE_SIZE > extent_map_end(em)) ||
470 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
472 unlock_extent(tree, last_offset, end);
479 if (page->index == end_index) {
481 size_t zero_offset = offset_in_page(isize);
485 zeros = PAGE_SIZE - zero_offset;
486 userpage = kmap_atomic(page);
487 memset(userpage + zero_offset, 0, zeros);
488 flush_dcache_page(page);
489 kunmap_atomic(userpage);
493 ret = bio_add_page(cb->orig_bio, page,
496 if (ret == PAGE_SIZE) {
500 unlock_extent(tree, last_offset, end);
506 last_offset += PAGE_SIZE;
512 * for a compressed read, the bio we get passed has all the inode pages
513 * in it. We don't actually do IO on those pages but allocate new ones
514 * to hold the compressed pages on disk.
516 * bio->bi_iter.bi_sector points to the compressed extent on disk
517 * bio->bi_io_vec points to all of the inode pages
519 * After the compressed pages are read, we copy the bytes into the
520 * bio we were passed and then call the bio end_io calls
522 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
523 int mirror_num, unsigned long bio_flags)
525 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
526 struct extent_map_tree *em_tree;
527 struct compressed_bio *cb;
528 unsigned long compressed_len;
529 unsigned long nr_pages;
530 unsigned long pg_index;
532 struct block_device *bdev;
533 struct bio *comp_bio;
534 u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
537 struct extent_map *em;
538 blk_status_t ret = BLK_STS_RESOURCE;
542 em_tree = &BTRFS_I(inode)->extent_tree;
544 /* we need the actual starting offset of this extent in the file */
545 read_lock(&em_tree->lock);
546 em = lookup_extent_mapping(em_tree,
547 page_offset(bio_first_page_all(bio)),
549 read_unlock(&em_tree->lock);
551 return BLK_STS_IOERR;
553 compressed_len = em->block_len;
554 cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
558 refcount_set(&cb->pending_bios, 0);
561 cb->mirror_num = mirror_num;
564 cb->start = em->orig_start;
566 em_start = em->start;
571 cb->len = bio->bi_iter.bi_size;
572 cb->compressed_len = compressed_len;
573 cb->compress_type = extent_compress_type(bio_flags);
576 nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
577 cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
579 if (!cb->compressed_pages)
582 bdev = fs_info->fs_devices->latest_bdev;
584 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
585 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
587 if (!cb->compressed_pages[pg_index]) {
588 faili = pg_index - 1;
589 ret = BLK_STS_RESOURCE;
593 faili = nr_pages - 1;
594 cb->nr_pages = nr_pages;
596 add_ra_bio_pages(inode, em_start + em_len, cb);
598 /* include any pages we added in add_ra-bio_pages */
599 cb->len = bio->bi_iter.bi_size;
601 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
602 comp_bio->bi_opf = REQ_OP_READ;
603 comp_bio->bi_private = cb;
604 comp_bio->bi_end_io = end_compressed_bio_read;
605 refcount_set(&cb->pending_bios, 1);
607 for (pg_index = 0; pg_index < nr_pages; pg_index++) {
610 page = cb->compressed_pages[pg_index];
611 page->mapping = inode->i_mapping;
612 page->index = em_start >> PAGE_SHIFT;
614 if (comp_bio->bi_iter.bi_size)
615 submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE,
618 page->mapping = NULL;
619 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
621 ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
622 BTRFS_WQ_ENDIO_DATA);
623 BUG_ON(ret); /* -ENOMEM */
626 * inc the count before we submit the bio so
627 * we know the end IO handler won't happen before
628 * we inc the count. Otherwise, the cb might get
629 * freed before we're done setting it up
631 refcount_inc(&cb->pending_bios);
633 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
634 ret = btrfs_lookup_bio_sums(inode, comp_bio,
636 BUG_ON(ret); /* -ENOMEM */
638 sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
639 fs_info->sectorsize);
641 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
643 comp_bio->bi_status = ret;
647 comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
648 comp_bio->bi_opf = REQ_OP_READ;
649 comp_bio->bi_private = cb;
650 comp_bio->bi_end_io = end_compressed_bio_read;
652 bio_add_page(comp_bio, page, PAGE_SIZE, 0);
654 cur_disk_byte += PAGE_SIZE;
657 ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
658 BUG_ON(ret); /* -ENOMEM */
660 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
661 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
662 BUG_ON(ret); /* -ENOMEM */
665 ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
667 comp_bio->bi_status = ret;
675 __free_page(cb->compressed_pages[faili]);
679 kfree(cb->compressed_pages);
688 * Heuristic uses systematic sampling to collect data from the input data
689 * range, the logic can be tuned by the following constants:
691 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
692 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
694 #define SAMPLING_READ_SIZE (16)
695 #define SAMPLING_INTERVAL (256)
698 * For statistical analysis of the input data we consider bytes that form a
699 * Galois Field of 256 objects. Each object has an attribute count, ie. how
700 * many times the object appeared in the sample.
702 #define BUCKET_SIZE (256)
705 * The size of the sample is based on a statistical sampling rule of thumb.
706 * The common way is to perform sampling tests as long as the number of
707 * elements in each cell is at least 5.
709 * Instead of 5, we choose 32 to obtain more accurate results.
710 * If the data contain the maximum number of symbols, which is 256, we obtain a
711 * sample size bound by 8192.
713 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
714 * from up to 512 locations.
716 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
717 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
723 struct heuristic_ws {
724 /* Partial copy of input data */
727 /* Buckets store counters for each byte value */
728 struct bucket_item *bucket;
730 struct bucket_item *bucket_b;
731 struct list_head list;
734 static void free_heuristic_ws(struct list_head *ws)
736 struct heuristic_ws *workspace;
738 workspace = list_entry(ws, struct heuristic_ws, list);
740 kvfree(workspace->sample);
741 kfree(workspace->bucket);
742 kfree(workspace->bucket_b);
746 static struct list_head *alloc_heuristic_ws(void)
748 struct heuristic_ws *ws;
750 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
752 return ERR_PTR(-ENOMEM);
754 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
758 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
762 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
766 INIT_LIST_HEAD(&ws->list);
769 free_heuristic_ws(&ws->list);
770 return ERR_PTR(-ENOMEM);
773 struct workspaces_list {
774 struct list_head idle_ws;
776 /* Number of free workspaces */
778 /* Total number of allocated workspaces */
780 /* Waiters for a free workspace */
781 wait_queue_head_t ws_wait;
784 static struct workspaces_list btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
786 static struct workspaces_list btrfs_heuristic_ws;
788 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
789 &btrfs_zlib_compress,
791 &btrfs_zstd_compress,
794 void __init btrfs_init_compress(void)
796 struct list_head *workspace;
799 INIT_LIST_HEAD(&btrfs_heuristic_ws.idle_ws);
800 spin_lock_init(&btrfs_heuristic_ws.ws_lock);
801 atomic_set(&btrfs_heuristic_ws.total_ws, 0);
802 init_waitqueue_head(&btrfs_heuristic_ws.ws_wait);
804 workspace = alloc_heuristic_ws();
805 if (IS_ERR(workspace)) {
807 "BTRFS: cannot preallocate heuristic workspace, will try later\n");
809 atomic_set(&btrfs_heuristic_ws.total_ws, 1);
810 btrfs_heuristic_ws.free_ws = 1;
811 list_add(workspace, &btrfs_heuristic_ws.idle_ws);
814 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
815 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
816 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
817 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
818 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
821 * Preallocate one workspace for each compression type so
822 * we can guarantee forward progress in the worst case
824 workspace = btrfs_compress_op[i]->alloc_workspace();
825 if (IS_ERR(workspace)) {
826 pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n");
828 atomic_set(&btrfs_comp_ws[i].total_ws, 1);
829 btrfs_comp_ws[i].free_ws = 1;
830 list_add(workspace, &btrfs_comp_ws[i].idle_ws);
836 * This finds an available workspace or allocates a new one.
837 * If it's not possible to allocate a new one, waits until there's one.
838 * Preallocation makes a forward progress guarantees and we do not return
841 static struct list_head *__find_workspace(int type, bool heuristic)
843 struct list_head *workspace;
844 int cpus = num_online_cpus();
847 struct list_head *idle_ws;
850 wait_queue_head_t *ws_wait;
854 idle_ws = &btrfs_heuristic_ws.idle_ws;
855 ws_lock = &btrfs_heuristic_ws.ws_lock;
856 total_ws = &btrfs_heuristic_ws.total_ws;
857 ws_wait = &btrfs_heuristic_ws.ws_wait;
858 free_ws = &btrfs_heuristic_ws.free_ws;
860 idle_ws = &btrfs_comp_ws[idx].idle_ws;
861 ws_lock = &btrfs_comp_ws[idx].ws_lock;
862 total_ws = &btrfs_comp_ws[idx].total_ws;
863 ws_wait = &btrfs_comp_ws[idx].ws_wait;
864 free_ws = &btrfs_comp_ws[idx].free_ws;
869 if (!list_empty(idle_ws)) {
870 workspace = idle_ws->next;
873 spin_unlock(ws_lock);
877 if (atomic_read(total_ws) > cpus) {
880 spin_unlock(ws_lock);
881 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
882 if (atomic_read(total_ws) > cpus && !*free_ws)
884 finish_wait(ws_wait, &wait);
887 atomic_inc(total_ws);
888 spin_unlock(ws_lock);
891 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
892 * to turn it off here because we might get called from the restricted
893 * context of btrfs_compress_bio/btrfs_compress_pages
895 nofs_flag = memalloc_nofs_save();
897 workspace = alloc_heuristic_ws();
899 workspace = btrfs_compress_op[idx]->alloc_workspace();
900 memalloc_nofs_restore(nofs_flag);
902 if (IS_ERR(workspace)) {
903 atomic_dec(total_ws);
907 * Do not return the error but go back to waiting. There's a
908 * workspace preallocated for each type and the compression
909 * time is bounded so we get to a workspace eventually. This
910 * makes our caller's life easier.
912 * To prevent silent and low-probability deadlocks (when the
913 * initial preallocation fails), check if there are any
916 if (atomic_read(total_ws) == 0) {
917 static DEFINE_RATELIMIT_STATE(_rs,
918 /* once per minute */ 60 * HZ,
921 if (__ratelimit(&_rs)) {
922 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
930 static struct list_head *find_workspace(int type)
932 return __find_workspace(type, false);
936 * put a workspace struct back on the list or free it if we have enough
937 * idle ones sitting around
939 static void __free_workspace(int type, struct list_head *workspace,
943 struct list_head *idle_ws;
946 wait_queue_head_t *ws_wait;
950 idle_ws = &btrfs_heuristic_ws.idle_ws;
951 ws_lock = &btrfs_heuristic_ws.ws_lock;
952 total_ws = &btrfs_heuristic_ws.total_ws;
953 ws_wait = &btrfs_heuristic_ws.ws_wait;
954 free_ws = &btrfs_heuristic_ws.free_ws;
956 idle_ws = &btrfs_comp_ws[idx].idle_ws;
957 ws_lock = &btrfs_comp_ws[idx].ws_lock;
958 total_ws = &btrfs_comp_ws[idx].total_ws;
959 ws_wait = &btrfs_comp_ws[idx].ws_wait;
960 free_ws = &btrfs_comp_ws[idx].free_ws;
964 if (*free_ws <= num_online_cpus()) {
965 list_add(workspace, idle_ws);
967 spin_unlock(ws_lock);
970 spin_unlock(ws_lock);
973 free_heuristic_ws(workspace);
975 btrfs_compress_op[idx]->free_workspace(workspace);
976 atomic_dec(total_ws);
978 cond_wake_up(ws_wait);
981 static void free_workspace(int type, struct list_head *ws)
983 return __free_workspace(type, ws, false);
987 * cleanup function for module exit
989 static void free_workspaces(void)
991 struct list_head *workspace;
994 while (!list_empty(&btrfs_heuristic_ws.idle_ws)) {
995 workspace = btrfs_heuristic_ws.idle_ws.next;
997 free_heuristic_ws(workspace);
998 atomic_dec(&btrfs_heuristic_ws.total_ws);
1001 for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
1002 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
1003 workspace = btrfs_comp_ws[i].idle_ws.next;
1004 list_del(workspace);
1005 btrfs_compress_op[i]->free_workspace(workspace);
1006 atomic_dec(&btrfs_comp_ws[i].total_ws);
1012 * Given an address space and start and length, compress the bytes into @pages
1013 * that are allocated on demand.
1015 * @type_level is encoded algorithm and level, where level 0 means whatever
1016 * default the algorithm chooses and is opaque here;
1017 * - compression algo are 0-3
1018 * - the level are bits 4-7
1020 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1021 * and returns number of actually allocated pages
1023 * @total_in is used to return the number of bytes actually read. It
1024 * may be smaller than the input length if we had to exit early because we
1025 * ran out of room in the pages array or because we cross the
1026 * max_out threshold.
1028 * @total_out is an in/out parameter, must be set to the input length and will
1029 * be also used to return the total number of compressed bytes
1031 * @max_out tells us the max number of bytes that we're allowed to
1034 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1035 u64 start, struct page **pages,
1036 unsigned long *out_pages,
1037 unsigned long *total_in,
1038 unsigned long *total_out)
1040 struct list_head *workspace;
1042 int type = type_level & 0xF;
1044 workspace = find_workspace(type);
1046 btrfs_compress_op[type - 1]->set_level(workspace, type_level);
1047 ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
1050 total_in, total_out);
1051 free_workspace(type, workspace);
1056 * pages_in is an array of pages with compressed data.
1058 * disk_start is the starting logical offset of this array in the file
1060 * orig_bio contains the pages from the file that we want to decompress into
1062 * srclen is the number of bytes in pages_in
1064 * The basic idea is that we have a bio that was created by readpages.
1065 * The pages in the bio are for the uncompressed data, and they may not
1066 * be contiguous. They all correspond to the range of bytes covered by
1067 * the compressed extent.
1069 static int btrfs_decompress_bio(struct compressed_bio *cb)
1071 struct list_head *workspace;
1073 int type = cb->compress_type;
1075 workspace = find_workspace(type);
1076 ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb);
1077 free_workspace(type, workspace);
1083 * a less complex decompression routine. Our compressed data fits in a
1084 * single page, and we want to read a single page out of it.
1085 * start_byte tells us the offset into the compressed data we're interested in
1087 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1088 unsigned long start_byte, size_t srclen, size_t destlen)
1090 struct list_head *workspace;
1093 workspace = find_workspace(type);
1095 ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
1096 dest_page, start_byte,
1099 free_workspace(type, workspace);
1103 void __cold btrfs_exit_compress(void)
1109 * Copy uncompressed data from working buffer to pages.
1111 * buf_start is the byte offset we're of the start of our workspace buffer.
1113 * total_out is the last byte of the buffer
1115 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
1116 unsigned long total_out, u64 disk_start,
1119 unsigned long buf_offset;
1120 unsigned long current_buf_start;
1121 unsigned long start_byte;
1122 unsigned long prev_start_byte;
1123 unsigned long working_bytes = total_out - buf_start;
1124 unsigned long bytes;
1126 struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
1129 * start byte is the first byte of the page we're currently
1130 * copying into relative to the start of the compressed data.
1132 start_byte = page_offset(bvec.bv_page) - disk_start;
1134 /* we haven't yet hit data corresponding to this page */
1135 if (total_out <= start_byte)
1139 * the start of the data we care about is offset into
1140 * the middle of our working buffer
1142 if (total_out > start_byte && buf_start < start_byte) {
1143 buf_offset = start_byte - buf_start;
1144 working_bytes -= buf_offset;
1148 current_buf_start = buf_start;
1150 /* copy bytes from the working buffer into the pages */
1151 while (working_bytes > 0) {
1152 bytes = min_t(unsigned long, bvec.bv_len,
1153 PAGE_SIZE - buf_offset);
1154 bytes = min(bytes, working_bytes);
1156 kaddr = kmap_atomic(bvec.bv_page);
1157 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1158 kunmap_atomic(kaddr);
1159 flush_dcache_page(bvec.bv_page);
1161 buf_offset += bytes;
1162 working_bytes -= bytes;
1163 current_buf_start += bytes;
1165 /* check if we need to pick another page */
1166 bio_advance(bio, bytes);
1167 if (!bio->bi_iter.bi_size)
1169 bvec = bio_iter_iovec(bio, bio->bi_iter);
1170 prev_start_byte = start_byte;
1171 start_byte = page_offset(bvec.bv_page) - disk_start;
1174 * We need to make sure we're only adjusting
1175 * our offset into compression working buffer when
1176 * we're switching pages. Otherwise we can incorrectly
1177 * keep copying when we were actually done.
1179 if (start_byte != prev_start_byte) {
1181 * make sure our new page is covered by this
1184 if (total_out <= start_byte)
1188 * the next page in the biovec might not be adjacent
1189 * to the last page, but it might still be found
1190 * inside this working buffer. bump our offset pointer
1192 if (total_out > start_byte &&
1193 current_buf_start < start_byte) {
1194 buf_offset = start_byte - buf_start;
1195 working_bytes = total_out - start_byte;
1196 current_buf_start = buf_start + buf_offset;
1205 * Shannon Entropy calculation
1207 * Pure byte distribution analysis fails to determine compressibility of data.
1208 * Try calculating entropy to estimate the average minimum number of bits
1209 * needed to encode the sampled data.
1211 * For convenience, return the percentage of needed bits, instead of amount of
1214 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1215 * and can be compressible with high probability
1217 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1219 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1221 #define ENTROPY_LVL_ACEPTABLE (65)
1222 #define ENTROPY_LVL_HIGH (80)
1225 * For increasead precision in shannon_entropy calculation,
1226 * let's do pow(n, M) to save more digits after comma:
1228 * - maximum int bit length is 64
1229 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1230 * - 13 * 4 = 52 < 64 -> M = 4
1234 static inline u32 ilog2_w(u64 n)
1236 return ilog2(n * n * n * n);
1239 static u32 shannon_entropy(struct heuristic_ws *ws)
1241 const u32 entropy_max = 8 * ilog2_w(2);
1242 u32 entropy_sum = 0;
1243 u32 p, p_base, sz_base;
1246 sz_base = ilog2_w(ws->sample_size);
1247 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1248 p = ws->bucket[i].count;
1249 p_base = ilog2_w(p);
1250 entropy_sum += p * (sz_base - p_base);
1253 entropy_sum /= ws->sample_size;
1254 return entropy_sum * 100 / entropy_max;
1257 #define RADIX_BASE 4U
1258 #define COUNTERS_SIZE (1U << RADIX_BASE)
1260 static u8 get4bits(u64 num, int shift) {
1265 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1270 * Use 4 bits as radix base
1271 * Use 16 u32 counters for calculating new position in buf array
1273 * @array - array that will be sorted
1274 * @array_buf - buffer array to store sorting results
1275 * must be equal in size to @array
1278 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1283 u32 counters[COUNTERS_SIZE];
1291 * Try avoid useless loop iterations for small numbers stored in big
1292 * counters. Example: 48 33 4 ... in 64bit array
1294 max_num = array[0].count;
1295 for (i = 1; i < num; i++) {
1296 buf_num = array[i].count;
1297 if (buf_num > max_num)
1301 buf_num = ilog2(max_num);
1302 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1305 while (shift < bitlen) {
1306 memset(counters, 0, sizeof(counters));
1308 for (i = 0; i < num; i++) {
1309 buf_num = array[i].count;
1310 addr = get4bits(buf_num, shift);
1314 for (i = 1; i < COUNTERS_SIZE; i++)
1315 counters[i] += counters[i - 1];
1317 for (i = num - 1; i >= 0; i--) {
1318 buf_num = array[i].count;
1319 addr = get4bits(buf_num, shift);
1321 new_addr = counters[addr];
1322 array_buf[new_addr] = array[i];
1325 shift += RADIX_BASE;
1328 * Normal radix expects to move data from a temporary array, to
1329 * the main one. But that requires some CPU time. Avoid that
1330 * by doing another sort iteration to original array instead of
1333 memset(counters, 0, sizeof(counters));
1335 for (i = 0; i < num; i ++) {
1336 buf_num = array_buf[i].count;
1337 addr = get4bits(buf_num, shift);
1341 for (i = 1; i < COUNTERS_SIZE; i++)
1342 counters[i] += counters[i - 1];
1344 for (i = num - 1; i >= 0; i--) {
1345 buf_num = array_buf[i].count;
1346 addr = get4bits(buf_num, shift);
1348 new_addr = counters[addr];
1349 array[new_addr] = array_buf[i];
1352 shift += RADIX_BASE;
1357 * Size of the core byte set - how many bytes cover 90% of the sample
1359 * There are several types of structured binary data that use nearly all byte
1360 * values. The distribution can be uniform and counts in all buckets will be
1361 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1363 * Other possibility is normal (Gaussian) distribution, where the data could
1364 * be potentially compressible, but we have to take a few more steps to decide
1367 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1368 * compression algo can easy fix that
1369 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1370 * probability is not compressible
1372 #define BYTE_CORE_SET_LOW (64)
1373 #define BYTE_CORE_SET_HIGH (200)
1375 static int byte_core_set_size(struct heuristic_ws *ws)
1378 u32 coreset_sum = 0;
1379 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1380 struct bucket_item *bucket = ws->bucket;
1382 /* Sort in reverse order */
1383 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1385 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1386 coreset_sum += bucket[i].count;
1388 if (coreset_sum > core_set_threshold)
1391 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1392 coreset_sum += bucket[i].count;
1393 if (coreset_sum > core_set_threshold)
1401 * Count byte values in buckets.
1402 * This heuristic can detect textual data (configs, xml, json, html, etc).
1403 * Because in most text-like data byte set is restricted to limited number of
1404 * possible characters, and that restriction in most cases makes data easy to
1407 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1408 * less - compressible
1409 * more - need additional analysis
1411 #define BYTE_SET_THRESHOLD (64)
1413 static u32 byte_set_size(const struct heuristic_ws *ws)
1416 u32 byte_set_size = 0;
1418 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1419 if (ws->bucket[i].count > 0)
1424 * Continue collecting count of byte values in buckets. If the byte
1425 * set size is bigger then the threshold, it's pointless to continue,
1426 * the detection technique would fail for this type of data.
1428 for (; i < BUCKET_SIZE; i++) {
1429 if (ws->bucket[i].count > 0) {
1431 if (byte_set_size > BYTE_SET_THRESHOLD)
1432 return byte_set_size;
1436 return byte_set_size;
1439 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1441 const u32 half_of_sample = ws->sample_size / 2;
1442 const u8 *data = ws->sample;
1444 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1447 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1448 struct heuristic_ws *ws)
1451 u64 index, index_end;
1452 u32 i, curr_sample_pos;
1456 * Compression handles the input data by chunks of 128KiB
1457 * (defined by BTRFS_MAX_UNCOMPRESSED)
1459 * We do the same for the heuristic and loop over the whole range.
1461 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1462 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1464 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1465 end = start + BTRFS_MAX_UNCOMPRESSED;
1467 index = start >> PAGE_SHIFT;
1468 index_end = end >> PAGE_SHIFT;
1470 /* Don't miss unaligned end */
1471 if (!IS_ALIGNED(end, PAGE_SIZE))
1474 curr_sample_pos = 0;
1475 while (index < index_end) {
1476 page = find_get_page(inode->i_mapping, index);
1477 in_data = kmap(page);
1478 /* Handle case where the start is not aligned to PAGE_SIZE */
1479 i = start % PAGE_SIZE;
1480 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1481 /* Don't sample any garbage from the last page */
1482 if (start > end - SAMPLING_READ_SIZE)
1484 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1485 SAMPLING_READ_SIZE);
1486 i += SAMPLING_INTERVAL;
1487 start += SAMPLING_INTERVAL;
1488 curr_sample_pos += SAMPLING_READ_SIZE;
1496 ws->sample_size = curr_sample_pos;
1500 * Compression heuristic.
1502 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1503 * quickly (compared to direct compression) detect data characteristics
1504 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1507 * The following types of analysis can be performed:
1508 * - detect mostly zero data
1509 * - detect data with low "byte set" size (text, etc)
1510 * - detect data with low/high "core byte" set
1512 * Return non-zero if the compression should be done, 0 otherwise.
1514 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1516 struct list_head *ws_list = __find_workspace(0, true);
1517 struct heuristic_ws *ws;
1522 ws = list_entry(ws_list, struct heuristic_ws, list);
1524 heuristic_collect_sample(inode, start, end, ws);
1526 if (sample_repeated_patterns(ws)) {
1531 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1533 for (i = 0; i < ws->sample_size; i++) {
1534 byte = ws->sample[i];
1535 ws->bucket[byte].count++;
1538 i = byte_set_size(ws);
1539 if (i < BYTE_SET_THRESHOLD) {
1544 i = byte_core_set_size(ws);
1545 if (i <= BYTE_CORE_SET_LOW) {
1550 if (i >= BYTE_CORE_SET_HIGH) {
1555 i = shannon_entropy(ws);
1556 if (i <= ENTROPY_LVL_ACEPTABLE) {
1562 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1563 * needed to give green light to compression.
1565 * For now just assume that compression at that level is not worth the
1566 * resources because:
1568 * 1. it is possible to defrag the data later
1570 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1571 * values, every bucket has counter at level ~54. The heuristic would
1572 * be confused. This can happen when data have some internal repeated
1573 * patterns like "abbacbbc...". This can be detected by analyzing
1574 * pairs of bytes, which is too costly.
1576 if (i < ENTROPY_LVL_HIGH) {
1585 __free_workspace(0, ws_list, true);
1589 unsigned int btrfs_compress_str2level(const char *str)
1591 if (strncmp(str, "zlib", 4) != 0)
1594 /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */
1595 if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0)
1596 return str[5] - '0';
1598 return BTRFS_ZLIB_DEFAULT_LEVEL;