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/pagevec.h>
12 #include <linux/highmem.h>
13 #include <linux/kthread.h>
14 #include <linux/time.h>
15 #include <linux/init.h>
16 #include <linux/string.h>
17 #include <linux/backing-dev.h>
18 #include <linux/writeback.h>
19 #include <linux/psi.h>
20 #include <linux/slab.h>
21 #include <linux/sched/mm.h>
22 #include <linux/log2.h>
23 #include <linux/shrinker.h>
24 #include <crypto/hash.h>
29 #include "transaction.h"
30 #include "btrfs_inode.h"
32 #include "ordered-data.h"
33 #include "compression.h"
34 #include "extent_io.h"
35 #include "extent_map.h"
38 #include "file-item.h"
41 static struct bio_set btrfs_compressed_bioset;
43 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
45 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
48 case BTRFS_COMPRESS_ZLIB:
49 case BTRFS_COMPRESS_LZO:
50 case BTRFS_COMPRESS_ZSTD:
51 case BTRFS_COMPRESS_NONE:
52 return btrfs_compress_types[type];
60 static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)
62 return container_of(bbio, struct compressed_bio, bbio);
65 static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,
66 u64 start, blk_opf_t op,
67 btrfs_bio_end_io_t end_io)
69 struct btrfs_bio *bbio;
71 bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,
72 GFP_NOFS, &btrfs_compressed_bioset));
73 btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
75 bbio->file_offset = start;
76 return to_compressed_bio(bbio);
79 bool btrfs_compress_is_valid_type(const char *str, size_t len)
83 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
84 size_t comp_len = strlen(btrfs_compress_types[i]);
89 if (!strncmp(btrfs_compress_types[i], str, comp_len))
95 static int compression_compress_pages(int type, struct list_head *ws,
96 struct address_space *mapping, u64 start, struct page **pages,
97 unsigned long *out_pages, unsigned long *total_in,
98 unsigned long *total_out)
101 case BTRFS_COMPRESS_ZLIB:
102 return zlib_compress_pages(ws, mapping, start, pages,
103 out_pages, total_in, total_out);
104 case BTRFS_COMPRESS_LZO:
105 return lzo_compress_pages(ws, mapping, start, pages,
106 out_pages, total_in, total_out);
107 case BTRFS_COMPRESS_ZSTD:
108 return zstd_compress_pages(ws, mapping, start, pages,
109 out_pages, total_in, total_out);
110 case BTRFS_COMPRESS_NONE:
113 * This can happen when compression races with remount setting
114 * it to 'no compress', while caller doesn't call
115 * inode_need_compress() to check if we really need to
118 * Not a big deal, just need to inform caller that we
119 * haven't allocated any pages yet.
126 static int compression_decompress_bio(struct list_head *ws,
127 struct compressed_bio *cb)
129 switch (cb->compress_type) {
130 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
131 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
132 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
133 case BTRFS_COMPRESS_NONE:
136 * This can't happen, the type is validated several times
137 * before we get here.
143 static int compression_decompress(int type, struct list_head *ws,
144 const u8 *data_in, struct page *dest_page,
145 unsigned long start_byte, size_t srclen, size_t destlen)
148 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
149 start_byte, srclen, destlen);
150 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
151 start_byte, srclen, destlen);
152 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
153 start_byte, srclen, destlen);
154 case BTRFS_COMPRESS_NONE:
157 * This can't happen, the type is validated several times
158 * before we get here.
164 static void btrfs_free_compressed_pages(struct compressed_bio *cb)
166 for (unsigned int i = 0; i < cb->nr_pages; i++)
167 btrfs_free_compr_page(cb->compressed_pages[i]);
168 kfree(cb->compressed_pages);
171 static int btrfs_decompress_bio(struct compressed_bio *cb);
174 * Global cache of last unused pages for compression/decompression.
176 static struct btrfs_compr_pool {
177 struct shrinker *shrinker;
179 struct list_head list;
184 static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
189 * We must not read the values more than once if 'ret' gets expanded in
190 * the return statement so we don't accidentally return a negative
191 * number, even if the first condition finds it positive.
193 ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
195 return ret > 0 ? ret : 0;
198 static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
200 struct list_head remove;
201 struct list_head *tmp, *next;
204 if (compr_pool.count == 0)
207 INIT_LIST_HEAD(&remove);
209 /* For now, just simply drain the whole list. */
210 spin_lock(&compr_pool.lock);
211 list_splice_init(&compr_pool.list, &remove);
212 freed = compr_pool.count;
213 compr_pool.count = 0;
214 spin_unlock(&compr_pool.lock);
216 list_for_each_safe(tmp, next, &remove) {
217 struct page *page = list_entry(tmp, struct page, lru);
219 ASSERT(page_ref_count(page) == 1);
227 * Common wrappers for page allocation from compression wrappers
229 struct page *btrfs_alloc_compr_page(void)
231 struct page *page = NULL;
233 spin_lock(&compr_pool.lock);
234 if (compr_pool.count > 0) {
235 page = list_first_entry(&compr_pool.list, struct page, lru);
236 list_del_init(&page->lru);
239 spin_unlock(&compr_pool.lock);
244 return alloc_page(GFP_NOFS);
247 void btrfs_free_compr_page(struct page *page)
249 bool do_free = false;
251 spin_lock(&compr_pool.lock);
252 if (compr_pool.count > compr_pool.thresh) {
255 list_add(&page->lru, &compr_pool.list);
258 spin_unlock(&compr_pool.lock);
263 ASSERT(page_ref_count(page) == 1);
267 static void end_bbio_comprssed_read(struct btrfs_bio *bbio)
269 struct compressed_bio *cb = to_compressed_bio(bbio);
270 blk_status_t status = bbio->bio.bi_status;
273 status = errno_to_blk_status(btrfs_decompress_bio(cb));
275 btrfs_free_compressed_pages(cb);
276 btrfs_bio_end_io(cb->orig_bbio, status);
281 * Clear the writeback bits on all of the file
282 * pages for a compressed write
284 static noinline void end_compressed_writeback(const struct compressed_bio *cb)
286 struct inode *inode = &cb->bbio.inode->vfs_inode;
287 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
288 unsigned long index = cb->start >> PAGE_SHIFT;
289 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
290 struct folio_batch fbatch;
291 const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
296 mapping_set_error(inode->i_mapping, error);
298 folio_batch_init(&fbatch);
299 while (index <= end_index) {
300 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
306 for (i = 0; i < ret; i++) {
307 struct folio *folio = fbatch.folios[i];
309 btrfs_folio_clamp_clear_writeback(fs_info, folio,
312 folio_batch_release(&fbatch);
314 /* the inode may be gone now */
317 static void btrfs_finish_compressed_write_work(struct work_struct *work)
319 struct compressed_bio *cb =
320 container_of(work, struct compressed_bio, write_end_work);
322 btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
323 cb->bbio.bio.bi_status == BLK_STS_OK);
326 end_compressed_writeback(cb);
327 /* Note, our inode could be gone now */
329 btrfs_free_compressed_pages(cb);
330 bio_put(&cb->bbio.bio);
334 * Do the cleanup once all the compressed pages hit the disk. This will clear
335 * writeback on the file pages and free the compressed pages.
337 * This also calls the writeback end hooks for the file pages so that metadata
338 * and checksums can be updated in the file.
340 static void end_bbio_comprssed_write(struct btrfs_bio *bbio)
342 struct compressed_bio *cb = to_compressed_bio(bbio);
343 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
345 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
348 static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb)
350 struct bio *bio = &cb->bbio.bio;
353 while (offset < cb->compressed_len) {
354 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
356 /* Maximum compressed extent is smaller than bio size limit. */
357 __bio_add_page(bio, cb->compressed_pages[offset >> PAGE_SHIFT],
364 * worker function to build and submit bios for previously compressed pages.
365 * The corresponding pages in the inode should be marked for writeback
366 * and the compressed pages should have a reference on them for dropping
367 * when the IO is complete.
369 * This also checksums the file bytes and gets things ready for
372 void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
373 struct page **compressed_pages,
374 unsigned int nr_pages,
375 blk_opf_t write_flags,
378 struct btrfs_inode *inode = BTRFS_I(ordered->inode);
379 struct btrfs_fs_info *fs_info = inode->root->fs_info;
380 struct compressed_bio *cb;
382 ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
383 ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
385 cb = alloc_compressed_bio(inode, ordered->file_offset,
386 REQ_OP_WRITE | write_flags,
387 end_bbio_comprssed_write);
388 cb->start = ordered->file_offset;
389 cb->len = ordered->num_bytes;
390 cb->compressed_pages = compressed_pages;
391 cb->compressed_len = ordered->disk_num_bytes;
392 cb->writeback = writeback;
393 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
394 cb->nr_pages = nr_pages;
395 cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
396 cb->bbio.ordered = ordered;
397 btrfs_add_compressed_bio_pages(cb);
399 btrfs_submit_bio(&cb->bbio, 0);
403 * Add extra pages in the same compressed file extent so that we don't need to
404 * re-read the same extent again and again.
406 * NOTE: this won't work well for subpage, as for subpage read, we lock the
407 * full page then submit bio for each compressed/regular extents.
409 * This means, if we have several sectors in the same page points to the same
410 * on-disk compressed data, we will re-read the same extent many times and
411 * this function can only help for the next page.
413 static noinline int add_ra_bio_pages(struct inode *inode,
415 struct compressed_bio *cb,
416 int *memstall, unsigned long *pflags)
418 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
419 unsigned long end_index;
420 struct bio *orig_bio = &cb->orig_bbio->bio;
421 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
422 u64 isize = i_size_read(inode);
425 struct extent_map *em;
426 struct address_space *mapping = inode->i_mapping;
427 struct extent_map_tree *em_tree;
428 struct extent_io_tree *tree;
429 int sectors_missed = 0;
431 em_tree = &BTRFS_I(inode)->extent_tree;
432 tree = &BTRFS_I(inode)->io_tree;
438 * For current subpage support, we only support 64K page size,
439 * which means maximum compressed extent size (128K) is just 2x page
441 * This makes readahead less effective, so here disable readahead for
442 * subpage for now, until full compressed write is supported.
444 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
447 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
449 while (cur < compressed_end) {
451 u64 pg_index = cur >> PAGE_SHIFT;
454 if (pg_index > end_index)
457 page = xa_load(&mapping->i_pages, pg_index);
458 if (page && !xa_is_value(page)) {
459 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
460 fs_info->sectorsize_bits;
462 /* Beyond threshold, no need to continue */
463 if (sectors_missed > 4)
467 * Jump to next page start as we already have page for
470 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
474 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
479 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
481 /* There is already a page, skip to page end */
482 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
486 if (!*memstall && PageWorkingset(page)) {
487 psi_memstall_enter(pflags);
491 ret = set_page_extent_mapped(page);
498 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
499 lock_extent(tree, cur, page_end, NULL);
500 read_lock(&em_tree->lock);
501 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
502 read_unlock(&em_tree->lock);
505 * At this point, we have a locked page in the page cache for
506 * these bytes in the file. But, we have to make sure they map
507 * to this compressed extent on disk.
509 if (!em || cur < em->start ||
510 (cur + fs_info->sectorsize > extent_map_end(em)) ||
511 (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) {
513 unlock_extent(tree, cur, page_end, NULL);
520 if (page->index == end_index) {
521 size_t zero_offset = offset_in_page(isize);
525 zeros = PAGE_SIZE - zero_offset;
526 memzero_page(page, zero_offset, zeros);
530 add_size = min(em->start + em->len, page_end + 1) - cur;
531 ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
532 if (ret != add_size) {
533 unlock_extent(tree, cur, page_end, NULL);
539 * If it's subpage, we also need to increase its
540 * subpage::readers number, as at endio we will decrease
541 * subpage::readers and to unlock the page.
543 if (fs_info->sectorsize < PAGE_SIZE)
544 btrfs_subpage_start_reader(fs_info, page_folio(page),
553 * for a compressed read, the bio we get passed has all the inode pages
554 * in it. We don't actually do IO on those pages but allocate new ones
555 * to hold the compressed pages on disk.
557 * bio->bi_iter.bi_sector points to the compressed extent on disk
558 * bio->bi_io_vec points to all of the inode pages
560 * After the compressed pages are read, we copy the bytes into the
561 * bio we were passed and then call the bio end_io calls
563 void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
565 struct btrfs_inode *inode = bbio->inode;
566 struct btrfs_fs_info *fs_info = inode->root->fs_info;
567 struct extent_map_tree *em_tree = &inode->extent_tree;
568 struct compressed_bio *cb;
569 unsigned int compressed_len;
570 u64 file_offset = bbio->file_offset;
573 struct extent_map *em;
574 unsigned long pflags;
579 /* we need the actual starting offset of this extent in the file */
580 read_lock(&em_tree->lock);
581 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
582 read_unlock(&em_tree->lock);
588 ASSERT(extent_map_is_compressed(em));
589 compressed_len = em->block_len;
591 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
592 end_bbio_comprssed_read);
594 cb->start = em->orig_start;
596 em_start = em->start;
598 cb->len = bbio->bio.bi_iter.bi_size;
599 cb->compressed_len = compressed_len;
600 cb->compress_type = extent_map_compression(em);
601 cb->orig_bbio = bbio;
605 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
606 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
607 if (!cb->compressed_pages) {
608 ret = BLK_STS_RESOURCE;
612 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages, 0);
614 ret = BLK_STS_RESOURCE;
615 goto out_free_compressed_pages;
618 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
621 /* include any pages we added in add_ra-bio_pages */
622 cb->len = bbio->bio.bi_iter.bi_size;
623 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
624 btrfs_add_compressed_bio_pages(cb);
627 psi_memstall_leave(&pflags);
629 btrfs_submit_bio(&cb->bbio, 0);
632 out_free_compressed_pages:
633 kfree(cb->compressed_pages);
635 bio_put(&cb->bbio.bio);
637 btrfs_bio_end_io(bbio, ret);
641 * Heuristic uses systematic sampling to collect data from the input data
642 * range, the logic can be tuned by the following constants:
644 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
645 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
647 #define SAMPLING_READ_SIZE (16)
648 #define SAMPLING_INTERVAL (256)
651 * For statistical analysis of the input data we consider bytes that form a
652 * Galois Field of 256 objects. Each object has an attribute count, ie. how
653 * many times the object appeared in the sample.
655 #define BUCKET_SIZE (256)
658 * The size of the sample is based on a statistical sampling rule of thumb.
659 * The common way is to perform sampling tests as long as the number of
660 * elements in each cell is at least 5.
662 * Instead of 5, we choose 32 to obtain more accurate results.
663 * If the data contain the maximum number of symbols, which is 256, we obtain a
664 * sample size bound by 8192.
666 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
667 * from up to 512 locations.
669 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
670 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
676 struct heuristic_ws {
677 /* Partial copy of input data */
680 /* Buckets store counters for each byte value */
681 struct bucket_item *bucket;
683 struct bucket_item *bucket_b;
684 struct list_head list;
687 static struct workspace_manager heuristic_wsm;
689 static void free_heuristic_ws(struct list_head *ws)
691 struct heuristic_ws *workspace;
693 workspace = list_entry(ws, struct heuristic_ws, list);
695 kvfree(workspace->sample);
696 kfree(workspace->bucket);
697 kfree(workspace->bucket_b);
701 static struct list_head *alloc_heuristic_ws(unsigned int level)
703 struct heuristic_ws *ws;
705 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
707 return ERR_PTR(-ENOMEM);
709 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
713 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
717 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
721 INIT_LIST_HEAD(&ws->list);
724 free_heuristic_ws(&ws->list);
725 return ERR_PTR(-ENOMEM);
728 const struct btrfs_compress_op btrfs_heuristic_compress = {
729 .workspace_manager = &heuristic_wsm,
732 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
733 /* The heuristic is represented as compression type 0 */
734 &btrfs_heuristic_compress,
735 &btrfs_zlib_compress,
737 &btrfs_zstd_compress,
740 static struct list_head *alloc_workspace(int type, unsigned int level)
743 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
744 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
745 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
746 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
749 * This can't happen, the type is validated several times
750 * before we get here.
756 static void free_workspace(int type, struct list_head *ws)
759 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
760 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
761 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
762 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
765 * This can't happen, the type is validated several times
766 * before we get here.
772 static void btrfs_init_workspace_manager(int type)
774 struct workspace_manager *wsm;
775 struct list_head *workspace;
777 wsm = btrfs_compress_op[type]->workspace_manager;
778 INIT_LIST_HEAD(&wsm->idle_ws);
779 spin_lock_init(&wsm->ws_lock);
780 atomic_set(&wsm->total_ws, 0);
781 init_waitqueue_head(&wsm->ws_wait);
784 * Preallocate one workspace for each compression type so we can
785 * guarantee forward progress in the worst case
787 workspace = alloc_workspace(type, 0);
788 if (IS_ERR(workspace)) {
790 "BTRFS: cannot preallocate compression workspace, will try later\n");
792 atomic_set(&wsm->total_ws, 1);
794 list_add(workspace, &wsm->idle_ws);
798 static void btrfs_cleanup_workspace_manager(int type)
800 struct workspace_manager *wsman;
801 struct list_head *ws;
803 wsman = btrfs_compress_op[type]->workspace_manager;
804 while (!list_empty(&wsman->idle_ws)) {
805 ws = wsman->idle_ws.next;
807 free_workspace(type, ws);
808 atomic_dec(&wsman->total_ws);
813 * This finds an available workspace or allocates a new one.
814 * If it's not possible to allocate a new one, waits until there's one.
815 * Preallocation makes a forward progress guarantees and we do not return
818 struct list_head *btrfs_get_workspace(int type, unsigned int level)
820 struct workspace_manager *wsm;
821 struct list_head *workspace;
822 int cpus = num_online_cpus();
824 struct list_head *idle_ws;
827 wait_queue_head_t *ws_wait;
830 wsm = btrfs_compress_op[type]->workspace_manager;
831 idle_ws = &wsm->idle_ws;
832 ws_lock = &wsm->ws_lock;
833 total_ws = &wsm->total_ws;
834 ws_wait = &wsm->ws_wait;
835 free_ws = &wsm->free_ws;
839 if (!list_empty(idle_ws)) {
840 workspace = idle_ws->next;
843 spin_unlock(ws_lock);
847 if (atomic_read(total_ws) > cpus) {
850 spin_unlock(ws_lock);
851 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
852 if (atomic_read(total_ws) > cpus && !*free_ws)
854 finish_wait(ws_wait, &wait);
857 atomic_inc(total_ws);
858 spin_unlock(ws_lock);
861 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
862 * to turn it off here because we might get called from the restricted
863 * context of btrfs_compress_bio/btrfs_compress_pages
865 nofs_flag = memalloc_nofs_save();
866 workspace = alloc_workspace(type, level);
867 memalloc_nofs_restore(nofs_flag);
869 if (IS_ERR(workspace)) {
870 atomic_dec(total_ws);
874 * Do not return the error but go back to waiting. There's a
875 * workspace preallocated for each type and the compression
876 * time is bounded so we get to a workspace eventually. This
877 * makes our caller's life easier.
879 * To prevent silent and low-probability deadlocks (when the
880 * initial preallocation fails), check if there are any
883 if (atomic_read(total_ws) == 0) {
884 static DEFINE_RATELIMIT_STATE(_rs,
885 /* once per minute */ 60 * HZ,
888 if (__ratelimit(&_rs)) {
889 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
897 static struct list_head *get_workspace(int type, int level)
900 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
901 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
902 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
903 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
906 * This can't happen, the type is validated several times
907 * before we get here.
914 * put a workspace struct back on the list or free it if we have enough
915 * idle ones sitting around
917 void btrfs_put_workspace(int type, struct list_head *ws)
919 struct workspace_manager *wsm;
920 struct list_head *idle_ws;
923 wait_queue_head_t *ws_wait;
926 wsm = btrfs_compress_op[type]->workspace_manager;
927 idle_ws = &wsm->idle_ws;
928 ws_lock = &wsm->ws_lock;
929 total_ws = &wsm->total_ws;
930 ws_wait = &wsm->ws_wait;
931 free_ws = &wsm->free_ws;
934 if (*free_ws <= num_online_cpus()) {
935 list_add(ws, idle_ws);
937 spin_unlock(ws_lock);
940 spin_unlock(ws_lock);
942 free_workspace(type, ws);
943 atomic_dec(total_ws);
945 cond_wake_up(ws_wait);
948 static void put_workspace(int type, struct list_head *ws)
951 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
952 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
953 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
954 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
957 * This can't happen, the type is validated several times
958 * before we get here.
965 * Adjust @level according to the limits of the compression algorithm or
966 * fallback to default
968 static unsigned int btrfs_compress_set_level(int type, unsigned level)
970 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
973 level = ops->default_level;
975 level = min(level, ops->max_level);
981 * Given an address space and start and length, compress the bytes into @pages
982 * that are allocated on demand.
984 * @type_level is encoded algorithm and level, where level 0 means whatever
985 * default the algorithm chooses and is opaque here;
986 * - compression algo are 0-3
987 * - the level are bits 4-7
989 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
990 * and returns number of actually allocated pages
992 * @total_in is used to return the number of bytes actually read. It
993 * may be smaller than the input length if we had to exit early because we
994 * ran out of room in the pages array or because we cross the
997 * @total_out is an in/out parameter, must be set to the input length and will
998 * be also used to return the total number of compressed bytes
1000 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1001 u64 start, struct page **pages,
1002 unsigned long *out_pages,
1003 unsigned long *total_in,
1004 unsigned long *total_out)
1006 int type = btrfs_compress_type(type_level);
1007 int level = btrfs_compress_level(type_level);
1008 struct list_head *workspace;
1011 level = btrfs_compress_set_level(type, level);
1012 workspace = get_workspace(type, level);
1013 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1014 out_pages, total_in, total_out);
1015 put_workspace(type, workspace);
1019 static int btrfs_decompress_bio(struct compressed_bio *cb)
1021 struct list_head *workspace;
1023 int type = cb->compress_type;
1025 workspace = get_workspace(type, 0);
1026 ret = compression_decompress_bio(workspace, cb);
1027 put_workspace(type, workspace);
1030 zero_fill_bio(&cb->orig_bbio->bio);
1035 * a less complex decompression routine. Our compressed data fits in a
1036 * single page, and we want to read a single page out of it.
1037 * start_byte tells us the offset into the compressed data we're interested in
1039 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1040 unsigned long start_byte, size_t srclen, size_t destlen)
1042 struct list_head *workspace;
1045 workspace = get_workspace(type, 0);
1046 ret = compression_decompress(type, workspace, data_in, dest_page,
1047 start_byte, srclen, destlen);
1048 put_workspace(type, workspace);
1053 int __init btrfs_init_compress(void)
1055 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1056 offsetof(struct compressed_bio, bbio.bio),
1060 compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1061 if (!compr_pool.shrinker)
1064 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1065 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1066 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1067 zstd_init_workspace_manager();
1069 spin_lock_init(&compr_pool.lock);
1070 INIT_LIST_HEAD(&compr_pool.list);
1071 compr_pool.count = 0;
1072 /* 128K / 4K = 32, for 8 threads is 256 pages. */
1073 compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1074 compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1075 compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1076 compr_pool.shrinker->batch = 32;
1077 compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1078 shrinker_register(compr_pool.shrinker);
1083 void __cold btrfs_exit_compress(void)
1085 /* For now scan drains all pages and does not touch the parameters. */
1086 btrfs_compr_pool_scan(NULL, NULL);
1087 shrinker_free(compr_pool.shrinker);
1089 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1090 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1091 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1092 zstd_cleanup_workspace_manager();
1093 bioset_exit(&btrfs_compressed_bioset);
1097 * Copy decompressed data from working buffer to pages.
1099 * @buf: The decompressed data buffer
1100 * @buf_len: The decompressed data length
1101 * @decompressed: Number of bytes that are already decompressed inside the
1103 * @cb: The compressed extent descriptor
1104 * @orig_bio: The original bio that the caller wants to read for
1106 * An easier to understand graph is like below:
1108 * |<- orig_bio ->| |<- orig_bio->|
1109 * |<------- full decompressed extent ----->|
1110 * |<----------- @cb range ---->|
1111 * | |<-- @buf_len -->|
1112 * |<--- @decompressed --->|
1114 * Note that, @cb can be a subpage of the full decompressed extent, but
1115 * @cb->start always has the same as the orig_file_offset value of the full
1116 * decompressed extent.
1118 * When reading compressed extent, we have to read the full compressed extent,
1119 * while @orig_bio may only want part of the range.
1120 * Thus this function will ensure only data covered by @orig_bio will be copied
1123 * Return 0 if we have copied all needed contents for @orig_bio.
1124 * Return >0 if we need continue decompress.
1126 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1127 struct compressed_bio *cb, u32 decompressed)
1129 struct bio *orig_bio = &cb->orig_bbio->bio;
1130 /* Offset inside the full decompressed extent */
1133 cur_offset = decompressed;
1134 /* The main loop to do the copy */
1135 while (cur_offset < decompressed + buf_len) {
1136 struct bio_vec bvec;
1139 /* Offset inside the full decompressed extent */
1142 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1144 * cb->start may underflow, but subtracting that value can still
1145 * give us correct offset inside the full decompressed extent.
1147 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1149 /* Haven't reached the bvec range, exit */
1150 if (decompressed + buf_len <= bvec_offset)
1153 copy_start = max(cur_offset, bvec_offset);
1154 copy_len = min(bvec_offset + bvec.bv_len,
1155 decompressed + buf_len) - copy_start;
1159 * Extra range check to ensure we didn't go beyond
1162 ASSERT(copy_start - decompressed < buf_len);
1163 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1164 buf + copy_start - decompressed, copy_len);
1165 cur_offset += copy_len;
1167 bio_advance(orig_bio, copy_len);
1168 /* Finished the bio */
1169 if (!orig_bio->bi_iter.bi_size)
1176 * Shannon Entropy calculation
1178 * Pure byte distribution analysis fails to determine compressibility of data.
1179 * Try calculating entropy to estimate the average minimum number of bits
1180 * needed to encode the sampled data.
1182 * For convenience, return the percentage of needed bits, instead of amount of
1185 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1186 * and can be compressible with high probability
1188 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1190 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1192 #define ENTROPY_LVL_ACEPTABLE (65)
1193 #define ENTROPY_LVL_HIGH (80)
1196 * For increasead precision in shannon_entropy calculation,
1197 * let's do pow(n, M) to save more digits after comma:
1199 * - maximum int bit length is 64
1200 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1201 * - 13 * 4 = 52 < 64 -> M = 4
1205 static inline u32 ilog2_w(u64 n)
1207 return ilog2(n * n * n * n);
1210 static u32 shannon_entropy(struct heuristic_ws *ws)
1212 const u32 entropy_max = 8 * ilog2_w(2);
1213 u32 entropy_sum = 0;
1214 u32 p, p_base, sz_base;
1217 sz_base = ilog2_w(ws->sample_size);
1218 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1219 p = ws->bucket[i].count;
1220 p_base = ilog2_w(p);
1221 entropy_sum += p * (sz_base - p_base);
1224 entropy_sum /= ws->sample_size;
1225 return entropy_sum * 100 / entropy_max;
1228 #define RADIX_BASE 4U
1229 #define COUNTERS_SIZE (1U << RADIX_BASE)
1231 static u8 get4bits(u64 num, int shift) {
1236 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1241 * Use 4 bits as radix base
1242 * Use 16 u32 counters for calculating new position in buf array
1244 * @array - array that will be sorted
1245 * @array_buf - buffer array to store sorting results
1246 * must be equal in size to @array
1249 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1254 u32 counters[COUNTERS_SIZE];
1262 * Try avoid useless loop iterations for small numbers stored in big
1263 * counters. Example: 48 33 4 ... in 64bit array
1265 max_num = array[0].count;
1266 for (i = 1; i < num; i++) {
1267 buf_num = array[i].count;
1268 if (buf_num > max_num)
1272 buf_num = ilog2(max_num);
1273 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1276 while (shift < bitlen) {
1277 memset(counters, 0, sizeof(counters));
1279 for (i = 0; i < num; i++) {
1280 buf_num = array[i].count;
1281 addr = get4bits(buf_num, shift);
1285 for (i = 1; i < COUNTERS_SIZE; i++)
1286 counters[i] += counters[i - 1];
1288 for (i = num - 1; i >= 0; i--) {
1289 buf_num = array[i].count;
1290 addr = get4bits(buf_num, shift);
1292 new_addr = counters[addr];
1293 array_buf[new_addr] = array[i];
1296 shift += RADIX_BASE;
1299 * Normal radix expects to move data from a temporary array, to
1300 * the main one. But that requires some CPU time. Avoid that
1301 * by doing another sort iteration to original array instead of
1304 memset(counters, 0, sizeof(counters));
1306 for (i = 0; i < num; i ++) {
1307 buf_num = array_buf[i].count;
1308 addr = get4bits(buf_num, shift);
1312 for (i = 1; i < COUNTERS_SIZE; i++)
1313 counters[i] += counters[i - 1];
1315 for (i = num - 1; i >= 0; i--) {
1316 buf_num = array_buf[i].count;
1317 addr = get4bits(buf_num, shift);
1319 new_addr = counters[addr];
1320 array[new_addr] = array_buf[i];
1323 shift += RADIX_BASE;
1328 * Size of the core byte set - how many bytes cover 90% of the sample
1330 * There are several types of structured binary data that use nearly all byte
1331 * values. The distribution can be uniform and counts in all buckets will be
1332 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1334 * Other possibility is normal (Gaussian) distribution, where the data could
1335 * be potentially compressible, but we have to take a few more steps to decide
1338 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1339 * compression algo can easy fix that
1340 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1341 * probability is not compressible
1343 #define BYTE_CORE_SET_LOW (64)
1344 #define BYTE_CORE_SET_HIGH (200)
1346 static int byte_core_set_size(struct heuristic_ws *ws)
1349 u32 coreset_sum = 0;
1350 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1351 struct bucket_item *bucket = ws->bucket;
1353 /* Sort in reverse order */
1354 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1356 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1357 coreset_sum += bucket[i].count;
1359 if (coreset_sum > core_set_threshold)
1362 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1363 coreset_sum += bucket[i].count;
1364 if (coreset_sum > core_set_threshold)
1372 * Count byte values in buckets.
1373 * This heuristic can detect textual data (configs, xml, json, html, etc).
1374 * Because in most text-like data byte set is restricted to limited number of
1375 * possible characters, and that restriction in most cases makes data easy to
1378 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1379 * less - compressible
1380 * more - need additional analysis
1382 #define BYTE_SET_THRESHOLD (64)
1384 static u32 byte_set_size(const struct heuristic_ws *ws)
1387 u32 byte_set_size = 0;
1389 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1390 if (ws->bucket[i].count > 0)
1395 * Continue collecting count of byte values in buckets. If the byte
1396 * set size is bigger then the threshold, it's pointless to continue,
1397 * the detection technique would fail for this type of data.
1399 for (; i < BUCKET_SIZE; i++) {
1400 if (ws->bucket[i].count > 0) {
1402 if (byte_set_size > BYTE_SET_THRESHOLD)
1403 return byte_set_size;
1407 return byte_set_size;
1410 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1412 const u32 half_of_sample = ws->sample_size / 2;
1413 const u8 *data = ws->sample;
1415 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1418 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1419 struct heuristic_ws *ws)
1422 u64 index, index_end;
1423 u32 i, curr_sample_pos;
1427 * Compression handles the input data by chunks of 128KiB
1428 * (defined by BTRFS_MAX_UNCOMPRESSED)
1430 * We do the same for the heuristic and loop over the whole range.
1432 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1433 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1435 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1436 end = start + BTRFS_MAX_UNCOMPRESSED;
1438 index = start >> PAGE_SHIFT;
1439 index_end = end >> PAGE_SHIFT;
1441 /* Don't miss unaligned end */
1442 if (!PAGE_ALIGNED(end))
1445 curr_sample_pos = 0;
1446 while (index < index_end) {
1447 page = find_get_page(inode->i_mapping, index);
1448 in_data = kmap_local_page(page);
1449 /* Handle case where the start is not aligned to PAGE_SIZE */
1450 i = start % PAGE_SIZE;
1451 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1452 /* Don't sample any garbage from the last page */
1453 if (start > end - SAMPLING_READ_SIZE)
1455 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1456 SAMPLING_READ_SIZE);
1457 i += SAMPLING_INTERVAL;
1458 start += SAMPLING_INTERVAL;
1459 curr_sample_pos += SAMPLING_READ_SIZE;
1461 kunmap_local(in_data);
1467 ws->sample_size = curr_sample_pos;
1471 * Compression heuristic.
1473 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1474 * quickly (compared to direct compression) detect data characteristics
1475 * (compressible/incompressible) to avoid wasting CPU time on incompressible
1478 * The following types of analysis can be performed:
1479 * - detect mostly zero data
1480 * - detect data with low "byte set" size (text, etc)
1481 * - detect data with low/high "core byte" set
1483 * Return non-zero if the compression should be done, 0 otherwise.
1485 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1487 struct list_head *ws_list = get_workspace(0, 0);
1488 struct heuristic_ws *ws;
1493 ws = list_entry(ws_list, struct heuristic_ws, list);
1495 heuristic_collect_sample(inode, start, end, ws);
1497 if (sample_repeated_patterns(ws)) {
1502 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1504 for (i = 0; i < ws->sample_size; i++) {
1505 byte = ws->sample[i];
1506 ws->bucket[byte].count++;
1509 i = byte_set_size(ws);
1510 if (i < BYTE_SET_THRESHOLD) {
1515 i = byte_core_set_size(ws);
1516 if (i <= BYTE_CORE_SET_LOW) {
1521 if (i >= BYTE_CORE_SET_HIGH) {
1526 i = shannon_entropy(ws);
1527 if (i <= ENTROPY_LVL_ACEPTABLE) {
1533 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1534 * needed to give green light to compression.
1536 * For now just assume that compression at that level is not worth the
1537 * resources because:
1539 * 1. it is possible to defrag the data later
1541 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1542 * values, every bucket has counter at level ~54. The heuristic would
1543 * be confused. This can happen when data have some internal repeated
1544 * patterns like "abbacbbc...". This can be detected by analyzing
1545 * pairs of bytes, which is too costly.
1547 if (i < ENTROPY_LVL_HIGH) {
1556 put_workspace(0, ws_list);
1561 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1562 * level, unrecognized string will set the default level
1564 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1566 unsigned int level = 0;
1572 if (str[0] == ':') {
1573 ret = kstrtouint(str + 1, 10, &level);
1578 level = btrfs_compress_set_level(type, level);