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>
28 #include "btrfs_inode.h"
30 #include "ordered-data.h"
31 #include "compression.h"
32 #include "extent_io.h"
33 #include "extent_map.h"
38 static struct bio_set btrfs_compressed_bioset;
40 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
42 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
45 case BTRFS_COMPRESS_ZLIB:
46 case BTRFS_COMPRESS_LZO:
47 case BTRFS_COMPRESS_ZSTD:
48 case BTRFS_COMPRESS_NONE:
49 return btrfs_compress_types[type];
57 static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)
59 return container_of(bbio, struct compressed_bio, bbio);
62 static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,
63 u64 start, blk_opf_t op,
64 btrfs_bio_end_io_t end_io)
66 struct btrfs_bio *bbio;
68 bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,
69 GFP_NOFS, &btrfs_compressed_bioset));
70 btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
72 bbio->file_offset = start;
73 return to_compressed_bio(bbio);
76 bool btrfs_compress_is_valid_type(const char *str, size_t len)
80 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
81 size_t comp_len = strlen(btrfs_compress_types[i]);
86 if (!strncmp(btrfs_compress_types[i], str, comp_len))
92 static int compression_compress_pages(int type, struct list_head *ws,
93 struct address_space *mapping, u64 start,
94 struct folio **folios, unsigned long *out_folios,
95 unsigned long *total_in, unsigned long *total_out)
98 case BTRFS_COMPRESS_ZLIB:
99 return zlib_compress_folios(ws, mapping, start, folios,
100 out_folios, total_in, total_out);
101 case BTRFS_COMPRESS_LZO:
102 return lzo_compress_folios(ws, mapping, start, folios,
103 out_folios, total_in, total_out);
104 case BTRFS_COMPRESS_ZSTD:
105 return zstd_compress_folios(ws, mapping, start, folios,
106 out_folios, total_in, total_out);
107 case BTRFS_COMPRESS_NONE:
110 * This can happen when compression races with remount setting
111 * it to 'no compress', while caller doesn't call
112 * inode_need_compress() to check if we really need to
115 * Not a big deal, just need to inform caller that we
116 * haven't allocated any pages yet.
123 static int compression_decompress_bio(struct list_head *ws,
124 struct compressed_bio *cb)
126 switch (cb->compress_type) {
127 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
128 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
129 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
130 case BTRFS_COMPRESS_NONE:
133 * This can't happen, the type is validated several times
134 * before we get here.
140 static int compression_decompress(int type, struct list_head *ws,
141 const u8 *data_in, struct page *dest_page,
142 unsigned long dest_pgoff, size_t srclen, size_t destlen)
145 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
146 dest_pgoff, srclen, destlen);
147 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
148 dest_pgoff, srclen, destlen);
149 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
150 dest_pgoff, srclen, destlen);
151 case BTRFS_COMPRESS_NONE:
154 * This can't happen, the type is validated several times
155 * before we get here.
161 static void btrfs_free_compressed_folios(struct compressed_bio *cb)
163 for (unsigned int i = 0; i < cb->nr_folios; i++)
164 btrfs_free_compr_folio(cb->compressed_folios[i]);
165 kfree(cb->compressed_folios);
168 static int btrfs_decompress_bio(struct compressed_bio *cb);
171 * Global cache of last unused pages for compression/decompression.
173 static struct btrfs_compr_pool {
174 struct shrinker *shrinker;
176 struct list_head list;
181 static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
186 * We must not read the values more than once if 'ret' gets expanded in
187 * the return statement so we don't accidentally return a negative
188 * number, even if the first condition finds it positive.
190 ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
192 return ret > 0 ? ret : 0;
195 static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
197 struct list_head remove;
198 struct list_head *tmp, *next;
201 if (compr_pool.count == 0)
204 INIT_LIST_HEAD(&remove);
206 /* For now, just simply drain the whole list. */
207 spin_lock(&compr_pool.lock);
208 list_splice_init(&compr_pool.list, &remove);
209 freed = compr_pool.count;
210 compr_pool.count = 0;
211 spin_unlock(&compr_pool.lock);
213 list_for_each_safe(tmp, next, &remove) {
214 struct page *page = list_entry(tmp, struct page, lru);
216 ASSERT(page_ref_count(page) == 1);
224 * Common wrappers for page allocation from compression wrappers
226 struct folio *btrfs_alloc_compr_folio(void)
228 struct folio *folio = NULL;
230 spin_lock(&compr_pool.lock);
231 if (compr_pool.count > 0) {
232 folio = list_first_entry(&compr_pool.list, struct folio, lru);
233 list_del_init(&folio->lru);
236 spin_unlock(&compr_pool.lock);
241 return folio_alloc(GFP_NOFS, 0);
244 void btrfs_free_compr_folio(struct folio *folio)
246 bool do_free = false;
248 spin_lock(&compr_pool.lock);
249 if (compr_pool.count > compr_pool.thresh) {
252 list_add(&folio->lru, &compr_pool.list);
255 spin_unlock(&compr_pool.lock);
260 ASSERT(folio_ref_count(folio) == 1);
264 static void end_bbio_comprssed_read(struct btrfs_bio *bbio)
266 struct compressed_bio *cb = to_compressed_bio(bbio);
267 blk_status_t status = bbio->bio.bi_status;
270 status = errno_to_blk_status(btrfs_decompress_bio(cb));
272 btrfs_free_compressed_folios(cb);
273 btrfs_bio_end_io(cb->orig_bbio, status);
278 * Clear the writeback bits on all of the file
279 * pages for a compressed write
281 static noinline void end_compressed_writeback(const struct compressed_bio *cb)
283 struct inode *inode = &cb->bbio.inode->vfs_inode;
284 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
285 unsigned long index = cb->start >> PAGE_SHIFT;
286 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
287 struct folio_batch fbatch;
288 const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
293 mapping_set_error(inode->i_mapping, error);
295 folio_batch_init(&fbatch);
296 while (index <= end_index) {
297 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
303 for (i = 0; i < ret; i++) {
304 struct folio *folio = fbatch.folios[i];
306 btrfs_folio_clamp_clear_writeback(fs_info, folio,
309 folio_batch_release(&fbatch);
311 /* the inode may be gone now */
314 static void btrfs_finish_compressed_write_work(struct work_struct *work)
316 struct compressed_bio *cb =
317 container_of(work, struct compressed_bio, write_end_work);
319 btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
320 cb->bbio.bio.bi_status == BLK_STS_OK);
323 end_compressed_writeback(cb);
324 /* Note, our inode could be gone now */
326 btrfs_free_compressed_folios(cb);
327 bio_put(&cb->bbio.bio);
331 * Do the cleanup once all the compressed pages hit the disk. This will clear
332 * writeback on the file pages and free the compressed pages.
334 * This also calls the writeback end hooks for the file pages so that metadata
335 * and checksums can be updated in the file.
337 static void end_bbio_comprssed_write(struct btrfs_bio *bbio)
339 struct compressed_bio *cb = to_compressed_bio(bbio);
340 struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
342 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
345 static void btrfs_add_compressed_bio_folios(struct compressed_bio *cb)
347 struct bio *bio = &cb->bbio.bio;
350 while (offset < cb->compressed_len) {
352 u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
354 /* Maximum compressed extent is smaller than bio size limit. */
355 ret = bio_add_folio(bio, cb->compressed_folios[offset >> PAGE_SHIFT],
363 * worker function to build and submit bios for previously compressed pages.
364 * The corresponding pages in the inode should be marked for writeback
365 * and the compressed pages should have a reference on them for dropping
366 * when the IO is complete.
368 * This also checksums the file bytes and gets things ready for
371 void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
372 struct folio **compressed_folios,
373 unsigned int nr_folios,
374 blk_opf_t write_flags,
377 struct btrfs_inode *inode = BTRFS_I(ordered->inode);
378 struct btrfs_fs_info *fs_info = inode->root->fs_info;
379 struct compressed_bio *cb;
381 ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
382 ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
384 cb = alloc_compressed_bio(inode, ordered->file_offset,
385 REQ_OP_WRITE | write_flags,
386 end_bbio_comprssed_write);
387 cb->start = ordered->file_offset;
388 cb->len = ordered->num_bytes;
389 cb->compressed_folios = compressed_folios;
390 cb->compressed_len = ordered->disk_num_bytes;
391 cb->writeback = writeback;
392 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
393 cb->nr_folios = nr_folios;
394 cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
395 cb->bbio.ordered = ordered;
396 btrfs_add_compressed_bio_folios(cb);
398 btrfs_submit_bio(&cb->bbio, 0);
402 * Add extra pages in the same compressed file extent so that we don't need to
403 * re-read the same extent again and again.
405 * NOTE: this won't work well for subpage, as for subpage read, we lock the
406 * full page then submit bio for each compressed/regular extents.
408 * This means, if we have several sectors in the same page points to the same
409 * on-disk compressed data, we will re-read the same extent many times and
410 * this function can only help for the next page.
412 static noinline int add_ra_bio_pages(struct inode *inode,
414 struct compressed_bio *cb,
415 int *memstall, unsigned long *pflags)
417 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
418 unsigned long end_index;
419 struct bio *orig_bio = &cb->orig_bbio->bio;
420 u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
421 u64 isize = i_size_read(inode);
424 struct extent_map *em;
425 struct address_space *mapping = inode->i_mapping;
426 struct extent_map_tree *em_tree;
427 struct extent_io_tree *tree;
428 int sectors_missed = 0;
430 em_tree = &BTRFS_I(inode)->extent_tree;
431 tree = &BTRFS_I(inode)->io_tree;
437 * For current subpage support, we only support 64K page size,
438 * which means maximum compressed extent size (128K) is just 2x page
440 * This makes readahead less effective, so here disable readahead for
441 * subpage for now, until full compressed write is supported.
443 if (fs_info->sectorsize < PAGE_SIZE)
446 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
448 while (cur < compressed_end) {
450 u64 pg_index = cur >> PAGE_SHIFT;
453 if (pg_index > end_index)
456 page = xa_load(&mapping->i_pages, pg_index);
457 if (page && !xa_is_value(page)) {
458 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
459 fs_info->sectorsize_bits;
461 /* Beyond threshold, no need to continue */
462 if (sectors_missed > 4)
466 * Jump to next page start as we already have page for
469 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
473 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
478 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
480 /* There is already a page, skip to page end */
481 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
485 if (!*memstall && PageWorkingset(page)) {
486 psi_memstall_enter(pflags);
490 ret = set_page_extent_mapped(page);
497 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
498 lock_extent(tree, cur, page_end, NULL);
499 read_lock(&em_tree->lock);
500 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
501 read_unlock(&em_tree->lock);
504 * At this point, we have a locked page in the page cache for
505 * these bytes in the file. But, we have to make sure they map
506 * to this compressed extent on disk.
508 if (!em || cur < em->start ||
509 (cur + fs_info->sectorsize > extent_map_end(em)) ||
510 (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) {
512 unlock_extent(tree, cur, page_end, NULL);
519 if (page->index == end_index) {
520 size_t zero_offset = offset_in_page(isize);
524 zeros = PAGE_SIZE - zero_offset;
525 memzero_page(page, zero_offset, zeros);
529 add_size = min(em->start + em->len, page_end + 1) - cur;
530 ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
531 if (ret != add_size) {
532 unlock_extent(tree, cur, page_end, NULL);
538 * If it's subpage, we also need to increase its
539 * subpage::readers number, as at endio we will decrease
540 * subpage::readers and to unlock the page.
542 if (fs_info->sectorsize < PAGE_SIZE)
543 btrfs_subpage_start_reader(fs_info, page_folio(page),
552 * for a compressed read, the bio we get passed has all the inode pages
553 * in it. We don't actually do IO on those pages but allocate new ones
554 * to hold the compressed pages on disk.
556 * bio->bi_iter.bi_sector points to the compressed extent on disk
557 * bio->bi_io_vec points to all of the inode pages
559 * After the compressed pages are read, we copy the bytes into the
560 * bio we were passed and then call the bio end_io calls
562 void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
564 struct btrfs_inode *inode = bbio->inode;
565 struct btrfs_fs_info *fs_info = inode->root->fs_info;
566 struct extent_map_tree *em_tree = &inode->extent_tree;
567 struct compressed_bio *cb;
568 unsigned int compressed_len;
569 u64 file_offset = bbio->file_offset;
572 struct extent_map *em;
573 unsigned long pflags;
578 /* we need the actual starting offset of this extent in the file */
579 read_lock(&em_tree->lock);
580 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
581 read_unlock(&em_tree->lock);
587 ASSERT(extent_map_is_compressed(em));
588 compressed_len = em->block_len;
590 cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
591 end_bbio_comprssed_read);
593 cb->start = em->orig_start;
595 em_start = em->start;
597 cb->len = bbio->bio.bi_iter.bi_size;
598 cb->compressed_len = compressed_len;
599 cb->compress_type = extent_map_compression(em);
600 cb->orig_bbio = bbio;
604 cb->nr_folios = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
605 cb->compressed_folios = kcalloc(cb->nr_folios, sizeof(struct page *), GFP_NOFS);
606 if (!cb->compressed_folios) {
607 ret = BLK_STS_RESOURCE;
611 ret2 = btrfs_alloc_folio_array(cb->nr_folios, cb->compressed_folios, 0);
613 ret = BLK_STS_RESOURCE;
614 goto out_free_compressed_pages;
617 add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
620 /* include any pages we added in add_ra-bio_pages */
621 cb->len = bbio->bio.bi_iter.bi_size;
622 cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
623 btrfs_add_compressed_bio_folios(cb);
626 psi_memstall_leave(&pflags);
628 btrfs_submit_bio(&cb->bbio, 0);
631 out_free_compressed_pages:
632 kfree(cb->compressed_folios);
634 bio_put(&cb->bbio.bio);
636 btrfs_bio_end_io(bbio, ret);
640 * Heuristic uses systematic sampling to collect data from the input data
641 * range, the logic can be tuned by the following constants:
643 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
644 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
646 #define SAMPLING_READ_SIZE (16)
647 #define SAMPLING_INTERVAL (256)
650 * For statistical analysis of the input data we consider bytes that form a
651 * Galois Field of 256 objects. Each object has an attribute count, ie. how
652 * many times the object appeared in the sample.
654 #define BUCKET_SIZE (256)
657 * The size of the sample is based on a statistical sampling rule of thumb.
658 * The common way is to perform sampling tests as long as the number of
659 * elements in each cell is at least 5.
661 * Instead of 5, we choose 32 to obtain more accurate results.
662 * If the data contain the maximum number of symbols, which is 256, we obtain a
663 * sample size bound by 8192.
665 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
666 * from up to 512 locations.
668 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
669 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
675 struct heuristic_ws {
676 /* Partial copy of input data */
679 /* Buckets store counters for each byte value */
680 struct bucket_item *bucket;
682 struct bucket_item *bucket_b;
683 struct list_head list;
686 static struct workspace_manager heuristic_wsm;
688 static void free_heuristic_ws(struct list_head *ws)
690 struct heuristic_ws *workspace;
692 workspace = list_entry(ws, struct heuristic_ws, list);
694 kvfree(workspace->sample);
695 kfree(workspace->bucket);
696 kfree(workspace->bucket_b);
700 static struct list_head *alloc_heuristic_ws(unsigned int level)
702 struct heuristic_ws *ws;
704 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
706 return ERR_PTR(-ENOMEM);
708 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
712 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
716 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
720 INIT_LIST_HEAD(&ws->list);
723 free_heuristic_ws(&ws->list);
724 return ERR_PTR(-ENOMEM);
727 const struct btrfs_compress_op btrfs_heuristic_compress = {
728 .workspace_manager = &heuristic_wsm,
731 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
732 /* The heuristic is represented as compression type 0 */
733 &btrfs_heuristic_compress,
734 &btrfs_zlib_compress,
736 &btrfs_zstd_compress,
739 static struct list_head *alloc_workspace(int type, unsigned int level)
742 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
743 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
744 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
745 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
748 * This can't happen, the type is validated several times
749 * before we get here.
755 static void free_workspace(int type, struct list_head *ws)
758 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
759 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
760 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
761 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
764 * This can't happen, the type is validated several times
765 * before we get here.
771 static void btrfs_init_workspace_manager(int type)
773 struct workspace_manager *wsm;
774 struct list_head *workspace;
776 wsm = btrfs_compress_op[type]->workspace_manager;
777 INIT_LIST_HEAD(&wsm->idle_ws);
778 spin_lock_init(&wsm->ws_lock);
779 atomic_set(&wsm->total_ws, 0);
780 init_waitqueue_head(&wsm->ws_wait);
783 * Preallocate one workspace for each compression type so we can
784 * guarantee forward progress in the worst case
786 workspace = alloc_workspace(type, 0);
787 if (IS_ERR(workspace)) {
789 "BTRFS: cannot preallocate compression workspace, will try later\n");
791 atomic_set(&wsm->total_ws, 1);
793 list_add(workspace, &wsm->idle_ws);
797 static void btrfs_cleanup_workspace_manager(int type)
799 struct workspace_manager *wsman;
800 struct list_head *ws;
802 wsman = btrfs_compress_op[type]->workspace_manager;
803 while (!list_empty(&wsman->idle_ws)) {
804 ws = wsman->idle_ws.next;
806 free_workspace(type, ws);
807 atomic_dec(&wsman->total_ws);
812 * This finds an available workspace or allocates a new one.
813 * If it's not possible to allocate a new one, waits until there's one.
814 * Preallocation makes a forward progress guarantees and we do not return
817 struct list_head *btrfs_get_workspace(int type, unsigned int level)
819 struct workspace_manager *wsm;
820 struct list_head *workspace;
821 int cpus = num_online_cpus();
823 struct list_head *idle_ws;
826 wait_queue_head_t *ws_wait;
829 wsm = btrfs_compress_op[type]->workspace_manager;
830 idle_ws = &wsm->idle_ws;
831 ws_lock = &wsm->ws_lock;
832 total_ws = &wsm->total_ws;
833 ws_wait = &wsm->ws_wait;
834 free_ws = &wsm->free_ws;
838 if (!list_empty(idle_ws)) {
839 workspace = idle_ws->next;
842 spin_unlock(ws_lock);
846 if (atomic_read(total_ws) > cpus) {
849 spin_unlock(ws_lock);
850 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
851 if (atomic_read(total_ws) > cpus && !*free_ws)
853 finish_wait(ws_wait, &wait);
856 atomic_inc(total_ws);
857 spin_unlock(ws_lock);
860 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
861 * to turn it off here because we might get called from the restricted
862 * context of btrfs_compress_bio/btrfs_compress_pages
864 nofs_flag = memalloc_nofs_save();
865 workspace = alloc_workspace(type, level);
866 memalloc_nofs_restore(nofs_flag);
868 if (IS_ERR(workspace)) {
869 atomic_dec(total_ws);
873 * Do not return the error but go back to waiting. There's a
874 * workspace preallocated for each type and the compression
875 * time is bounded so we get to a workspace eventually. This
876 * makes our caller's life easier.
878 * To prevent silent and low-probability deadlocks (when the
879 * initial preallocation fails), check if there are any
882 if (atomic_read(total_ws) == 0) {
883 static DEFINE_RATELIMIT_STATE(_rs,
884 /* once per minute */ 60 * HZ,
887 if (__ratelimit(&_rs)) {
888 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
896 static struct list_head *get_workspace(int type, int level)
899 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
900 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
901 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
902 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
905 * This can't happen, the type is validated several times
906 * before we get here.
913 * put a workspace struct back on the list or free it if we have enough
914 * idle ones sitting around
916 void btrfs_put_workspace(int type, struct list_head *ws)
918 struct workspace_manager *wsm;
919 struct list_head *idle_ws;
922 wait_queue_head_t *ws_wait;
925 wsm = btrfs_compress_op[type]->workspace_manager;
926 idle_ws = &wsm->idle_ws;
927 ws_lock = &wsm->ws_lock;
928 total_ws = &wsm->total_ws;
929 ws_wait = &wsm->ws_wait;
930 free_ws = &wsm->free_ws;
933 if (*free_ws <= num_online_cpus()) {
934 list_add(ws, idle_ws);
936 spin_unlock(ws_lock);
939 spin_unlock(ws_lock);
941 free_workspace(type, ws);
942 atomic_dec(total_ws);
944 cond_wake_up(ws_wait);
947 static void put_workspace(int type, struct list_head *ws)
950 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
951 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
952 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
953 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
956 * This can't happen, the type is validated several times
957 * before we get here.
964 * Adjust @level according to the limits of the compression algorithm or
965 * fallback to default
967 static unsigned int btrfs_compress_set_level(int type, unsigned level)
969 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
972 level = ops->default_level;
974 level = min(level, ops->max_level);
979 /* Wrapper around find_get_page(), with extra error message. */
980 int btrfs_compress_filemap_get_folio(struct address_space *mapping, u64 start,
981 struct folio **in_folio_ret)
983 struct folio *in_folio;
986 * The compressed write path should have the folio locked already, thus
987 * we only need to grab one reference.
989 in_folio = filemap_get_folio(mapping, start >> PAGE_SHIFT);
990 if (IS_ERR(in_folio)) {
991 struct btrfs_inode *inode = BTRFS_I(mapping->host);
993 btrfs_crit(inode->root->fs_info,
994 "failed to get page cache, root %lld ino %llu file offset %llu",
995 btrfs_root_id(inode->root), btrfs_ino(inode), start);
998 *in_folio_ret = in_folio;
1003 * Given an address space and start and length, compress the bytes into @pages
1004 * that are allocated on demand.
1006 * @type_level is encoded algorithm and level, where level 0 means whatever
1007 * default the algorithm chooses and is opaque here;
1008 * - compression algo are 0-3
1009 * - the level are bits 4-7
1011 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1012 * and returns number of actually allocated pages
1014 * @total_in is used to return the number of bytes actually read. It
1015 * may be smaller than the input length if we had to exit early because we
1016 * ran out of room in the pages array or because we cross the
1017 * max_out threshold.
1019 * @total_out is an in/out parameter, must be set to the input length and will
1020 * be also used to return the total number of compressed bytes
1022 int btrfs_compress_folios(unsigned int type_level, struct address_space *mapping,
1023 u64 start, struct folio **folios, unsigned long *out_folios,
1024 unsigned long *total_in, unsigned long *total_out)
1026 int type = btrfs_compress_type(type_level);
1027 int level = btrfs_compress_level(type_level);
1028 struct list_head *workspace;
1031 level = btrfs_compress_set_level(type, level);
1032 workspace = get_workspace(type, level);
1033 ret = compression_compress_pages(type, workspace, mapping, start, folios,
1034 out_folios, total_in, total_out);
1035 put_workspace(type, workspace);
1039 static int btrfs_decompress_bio(struct compressed_bio *cb)
1041 struct list_head *workspace;
1043 int type = cb->compress_type;
1045 workspace = get_workspace(type, 0);
1046 ret = compression_decompress_bio(workspace, cb);
1047 put_workspace(type, workspace);
1050 zero_fill_bio(&cb->orig_bbio->bio);
1055 * a less complex decompression routine. Our compressed data fits in a
1056 * single page, and we want to read a single page out of it.
1057 * start_byte tells us the offset into the compressed data we're interested in
1059 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1060 unsigned long dest_pgoff, size_t srclen, size_t destlen)
1062 struct btrfs_fs_info *fs_info = page_to_fs_info(dest_page);
1063 struct list_head *workspace;
1064 const u32 sectorsize = fs_info->sectorsize;
1068 * The full destination page range should not exceed the page size.
1069 * And the @destlen should not exceed sectorsize, as this is only called for
1070 * inline file extents, which should not exceed sectorsize.
1072 ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);
1074 workspace = get_workspace(type, 0);
1075 ret = compression_decompress(type, workspace, data_in, dest_page,
1076 dest_pgoff, srclen, destlen);
1077 put_workspace(type, workspace);
1082 int __init btrfs_init_compress(void)
1084 if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1085 offsetof(struct compressed_bio, bbio.bio),
1089 compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1090 if (!compr_pool.shrinker)
1093 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1094 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1095 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1096 zstd_init_workspace_manager();
1098 spin_lock_init(&compr_pool.lock);
1099 INIT_LIST_HEAD(&compr_pool.list);
1100 compr_pool.count = 0;
1101 /* 128K / 4K = 32, for 8 threads is 256 pages. */
1102 compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1103 compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1104 compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1105 compr_pool.shrinker->batch = 32;
1106 compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1107 shrinker_register(compr_pool.shrinker);
1112 void __cold btrfs_exit_compress(void)
1114 /* For now scan drains all pages and does not touch the parameters. */
1115 btrfs_compr_pool_scan(NULL, NULL);
1116 shrinker_free(compr_pool.shrinker);
1118 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1119 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1120 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1121 zstd_cleanup_workspace_manager();
1122 bioset_exit(&btrfs_compressed_bioset);
1126 * Copy decompressed data from working buffer to pages.
1128 * @buf: The decompressed data buffer
1129 * @buf_len: The decompressed data length
1130 * @decompressed: Number of bytes that are already decompressed inside the
1132 * @cb: The compressed extent descriptor
1133 * @orig_bio: The original bio that the caller wants to read for
1135 * An easier to understand graph is like below:
1137 * |<- orig_bio ->| |<- orig_bio->|
1138 * |<------- full decompressed extent ----->|
1139 * |<----------- @cb range ---->|
1140 * | |<-- @buf_len -->|
1141 * |<--- @decompressed --->|
1143 * Note that, @cb can be a subpage of the full decompressed extent, but
1144 * @cb->start always has the same as the orig_file_offset value of the full
1145 * decompressed extent.
1147 * When reading compressed extent, we have to read the full compressed extent,
1148 * while @orig_bio may only want part of the range.
1149 * Thus this function will ensure only data covered by @orig_bio will be copied
1152 * Return 0 if we have copied all needed contents for @orig_bio.
1153 * Return >0 if we need continue decompress.
1155 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1156 struct compressed_bio *cb, u32 decompressed)
1158 struct bio *orig_bio = &cb->orig_bbio->bio;
1159 /* Offset inside the full decompressed extent */
1162 cur_offset = decompressed;
1163 /* The main loop to do the copy */
1164 while (cur_offset < decompressed + buf_len) {
1165 struct bio_vec bvec;
1168 /* Offset inside the full decompressed extent */
1171 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1173 * cb->start may underflow, but subtracting that value can still
1174 * give us correct offset inside the full decompressed extent.
1176 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1178 /* Haven't reached the bvec range, exit */
1179 if (decompressed + buf_len <= bvec_offset)
1182 copy_start = max(cur_offset, bvec_offset);
1183 copy_len = min(bvec_offset + bvec.bv_len,
1184 decompressed + buf_len) - copy_start;
1188 * Extra range check to ensure we didn't go beyond
1191 ASSERT(copy_start - decompressed < buf_len);
1192 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1193 buf + copy_start - decompressed, copy_len);
1194 cur_offset += copy_len;
1196 bio_advance(orig_bio, copy_len);
1197 /* Finished the bio */
1198 if (!orig_bio->bi_iter.bi_size)
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 (!PAGE_ALIGNED(end))
1474 curr_sample_pos = 0;
1475 while (index < index_end) {
1476 page = find_get_page(inode->i_mapping, index);
1477 in_data = kmap_local_page(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;
1490 kunmap_local(in_data);
1496 ws->sample_size = curr_sample_pos;
1500 * Compression heuristic.
1502 * The following types of analysis can be performed:
1503 * - detect mostly zero data
1504 * - detect data with low "byte set" size (text, etc)
1505 * - detect data with low/high "core byte" set
1507 * Return non-zero if the compression should be done, 0 otherwise.
1509 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1511 struct list_head *ws_list = get_workspace(0, 0);
1512 struct heuristic_ws *ws;
1517 ws = list_entry(ws_list, struct heuristic_ws, list);
1519 heuristic_collect_sample(inode, start, end, ws);
1521 if (sample_repeated_patterns(ws)) {
1526 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1528 for (i = 0; i < ws->sample_size; i++) {
1529 byte = ws->sample[i];
1530 ws->bucket[byte].count++;
1533 i = byte_set_size(ws);
1534 if (i < BYTE_SET_THRESHOLD) {
1539 i = byte_core_set_size(ws);
1540 if (i <= BYTE_CORE_SET_LOW) {
1545 if (i >= BYTE_CORE_SET_HIGH) {
1550 i = shannon_entropy(ws);
1551 if (i <= ENTROPY_LVL_ACEPTABLE) {
1557 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1558 * needed to give green light to compression.
1560 * For now just assume that compression at that level is not worth the
1561 * resources because:
1563 * 1. it is possible to defrag the data later
1565 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1566 * values, every bucket has counter at level ~54. The heuristic would
1567 * be confused. This can happen when data have some internal repeated
1568 * patterns like "abbacbbc...". This can be detected by analyzing
1569 * pairs of bytes, which is too costly.
1571 if (i < ENTROPY_LVL_HIGH) {
1580 put_workspace(0, ws_list);
1585 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1586 * level, unrecognized string will set the default level
1588 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1590 unsigned int level = 0;
1596 if (str[0] == ':') {
1597 ret = kstrtouint(str + 1, 10, &level);
1602 level = btrfs_compress_set_level(type, level);