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 <crypto/hash.h>
28 #include "transaction.h"
29 #include "btrfs_inode.h"
31 #include "ordered-data.h"
32 #include "compression.h"
33 #include "extent_io.h"
34 #include "extent_map.h"
37 #include "file-item.h"
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 bool btrfs_compress_is_valid_type(const char *str, size_t len)
61 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
62 size_t comp_len = strlen(btrfs_compress_types[i]);
67 if (!strncmp(btrfs_compress_types[i], str, comp_len))
73 static int compression_compress_pages(int type, struct list_head *ws,
74 struct address_space *mapping, u64 start, struct page **pages,
75 unsigned long *out_pages, unsigned long *total_in,
76 unsigned long *total_out)
79 case BTRFS_COMPRESS_ZLIB:
80 return zlib_compress_pages(ws, mapping, start, pages,
81 out_pages, total_in, total_out);
82 case BTRFS_COMPRESS_LZO:
83 return lzo_compress_pages(ws, mapping, start, pages,
84 out_pages, total_in, total_out);
85 case BTRFS_COMPRESS_ZSTD:
86 return zstd_compress_pages(ws, mapping, start, pages,
87 out_pages, total_in, total_out);
88 case BTRFS_COMPRESS_NONE:
91 * This can happen when compression races with remount setting
92 * it to 'no compress', while caller doesn't call
93 * inode_need_compress() to check if we really need to
96 * Not a big deal, just need to inform caller that we
97 * haven't allocated any pages yet.
104 static int compression_decompress_bio(struct list_head *ws,
105 struct compressed_bio *cb)
107 switch (cb->compress_type) {
108 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
109 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
110 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
111 case BTRFS_COMPRESS_NONE:
114 * This can't happen, the type is validated several times
115 * before we get here.
121 static int compression_decompress(int type, struct list_head *ws,
122 const u8 *data_in, struct page *dest_page,
123 unsigned long start_byte, size_t srclen, size_t destlen)
126 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
127 start_byte, srclen, destlen);
128 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
129 start_byte, srclen, destlen);
130 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
131 start_byte, srclen, destlen);
132 case BTRFS_COMPRESS_NONE:
135 * This can't happen, the type is validated several times
136 * before we get here.
142 static int btrfs_decompress_bio(struct compressed_bio *cb);
144 static void finish_compressed_bio_read(struct compressed_bio *cb)
149 if (cb->status == BLK_STS_OK)
150 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
152 /* 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 */
160 btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
162 /* Finally free the cb struct */
163 kfree(cb->compressed_pages);
168 * Verify the checksums and kick off repair if needed on the uncompressed data
169 * before decompressing it into the original bio and freeing the uncompressed
172 static void end_compressed_bio_read(struct btrfs_bio *bbio)
174 struct compressed_bio *cb = bbio->private;
175 struct inode *inode = cb->inode;
176 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
177 struct btrfs_inode *bi = BTRFS_I(inode);
178 bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
179 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
180 blk_status_t status = bbio->bio.bi_status;
181 struct bvec_iter iter;
185 btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
186 u64 start = bbio->file_offset + offset;
189 (!csum || !btrfs_check_data_csum(bi, bbio, offset,
190 bv.bv_page, bv.bv_offset))) {
191 btrfs_clean_io_failure(bi, start, bv.bv_page,
196 refcount_inc(&cb->pending_ios);
197 ret = btrfs_repair_one_sector(BTRFS_I(inode), bbio, offset,
198 bv.bv_page, bv.bv_offset,
201 refcount_dec(&cb->pending_ios);
202 status = errno_to_blk_status(ret);
210 if (refcount_dec_and_test(&cb->pending_ios))
211 finish_compressed_bio_read(cb);
212 btrfs_bio_free_csum(bbio);
217 * Clear the writeback bits on all of the file
218 * pages for a compressed write
220 static noinline void end_compressed_writeback(struct inode *inode,
221 const struct compressed_bio *cb)
223 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
224 unsigned long index = cb->start >> PAGE_SHIFT;
225 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
226 struct folio_batch fbatch;
227 const int errno = blk_status_to_errno(cb->status);
232 mapping_set_error(inode->i_mapping, errno);
234 folio_batch_init(&fbatch);
235 while (index <= end_index) {
236 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
242 for (i = 0; i < ret; i++) {
243 struct folio *folio = fbatch.folios[i];
246 folio_set_error(folio);
247 btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
250 folio_batch_release(&fbatch);
252 /* the inode may be gone now */
255 static void finish_compressed_bio_write(struct compressed_bio *cb)
257 struct inode *inode = cb->inode;
261 * Ok, we're the last bio for this extent, step one is to call back
262 * into the FS and do all the end_io operations.
264 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
265 cb->start, cb->start + cb->len - 1,
266 cb->status == BLK_STS_OK);
269 end_compressed_writeback(inode, cb);
270 /* Note, our inode could be gone now */
273 * Release the compressed pages, these came from alloc_page and
274 * are not attached to the inode at all
276 for (index = 0; index < cb->nr_pages; index++) {
277 struct page *page = cb->compressed_pages[index];
279 page->mapping = NULL;
283 /* Finally free the cb struct */
284 kfree(cb->compressed_pages);
288 static void btrfs_finish_compressed_write_work(struct work_struct *work)
290 struct compressed_bio *cb =
291 container_of(work, struct compressed_bio, write_end_work);
293 finish_compressed_bio_write(cb);
297 * Do the cleanup once all the compressed pages hit the disk. This will clear
298 * writeback on the file pages and free the compressed pages.
300 * This also calls the writeback end hooks for the file pages so that metadata
301 * and checksums can be updated in the file.
303 static void end_compressed_bio_write(struct btrfs_bio *bbio)
305 struct compressed_bio *cb = bbio->private;
307 if (bbio->bio.bi_status)
308 cb->status = bbio->bio.bi_status;
310 if (refcount_dec_and_test(&cb->pending_ios)) {
311 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
313 btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio);
314 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
320 * Allocate a compressed_bio, which will be used to read/write on-disk
321 * (aka, compressed) * data.
323 * @cb: The compressed_bio structure, which records all the needed
324 * information to bind the compressed data to the uncompressed
326 * @disk_byten: The logical bytenr where the compressed data will be read
327 * from or written to.
328 * @endio_func: The endio function to call after the IO for compressed data
330 * @next_stripe_start: Return value of logical bytenr of where next stripe starts.
331 * Let the caller know to only fill the bio up to the stripe
336 static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
338 btrfs_bio_end_io_t endio_func,
339 u64 *next_stripe_start)
341 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
342 struct btrfs_io_geometry geom;
343 struct extent_map *em;
347 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, BTRFS_I(cb->inode), endio_func,
349 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
351 em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
357 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
358 bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
360 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
366 *next_stripe_start = disk_bytenr + geom.len;
367 refcount_inc(&cb->pending_ios);
372 * worker function to build and submit bios for previously compressed pages.
373 * The corresponding pages in the inode should be marked for writeback
374 * and the compressed pages should have a reference on them for dropping
375 * when the IO is complete.
377 * This also checksums the file bytes and gets things ready for
380 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
381 unsigned int len, u64 disk_start,
382 unsigned int compressed_len,
383 struct page **compressed_pages,
384 unsigned int nr_pages,
385 blk_opf_t write_flags,
386 struct cgroup_subsys_state *blkcg_css,
389 struct btrfs_fs_info *fs_info = inode->root->fs_info;
390 struct bio *bio = NULL;
391 struct compressed_bio *cb;
392 u64 cur_disk_bytenr = disk_start;
393 u64 next_stripe_start;
394 blk_status_t ret = BLK_STS_OK;
395 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
396 const bool use_append = btrfs_use_zone_append(inode, disk_start);
397 const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
399 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
400 IS_ALIGNED(len, fs_info->sectorsize));
401 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
403 return BLK_STS_RESOURCE;
404 refcount_set(&cb->pending_ios, 1);
405 cb->status = BLK_STS_OK;
406 cb->inode = &inode->vfs_inode;
409 cb->compressed_pages = compressed_pages;
410 cb->compressed_len = compressed_len;
411 cb->writeback = writeback;
412 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
413 cb->nr_pages = nr_pages;
416 kthread_associate_blkcg(blkcg_css);
418 while (cur_disk_bytenr < disk_start + compressed_len) {
419 u64 offset = cur_disk_bytenr - disk_start;
420 unsigned int index = offset >> PAGE_SHIFT;
421 unsigned int real_size;
423 struct page *page = compressed_pages[index];
426 /* Allocate new bio if submitted or not yet allocated */
428 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
429 bio_op | write_flags, end_compressed_bio_write,
432 ret = errno_to_blk_status(PTR_ERR(bio));
436 bio->bi_opf |= REQ_CGROUP_PUNT;
439 * We should never reach next_stripe_start start as we will
440 * submit comp_bio when reach the boundary immediately.
442 ASSERT(cur_disk_bytenr != next_stripe_start);
445 * We have various limits on the real read size:
448 * - compressed length boundary
450 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
451 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
452 real_size = min_t(u64, real_size, compressed_len - offset);
453 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
456 added = bio_add_zone_append_page(bio, page, real_size,
457 offset_in_page(offset));
459 added = bio_add_page(bio, page, real_size,
460 offset_in_page(offset));
461 /* Reached zoned boundary */
465 cur_disk_bytenr += added;
466 /* Reached stripe boundary */
467 if (cur_disk_bytenr == next_stripe_start)
470 /* Finished the range */
471 if (cur_disk_bytenr == disk_start + compressed_len)
476 ret = btrfs_csum_one_bio(inode, bio, start, true);
478 btrfs_bio_end_io(btrfs_bio(bio), ret);
483 ASSERT(bio->bi_iter.bi_size);
484 btrfs_submit_bio(fs_info, bio, 0);
491 kthread_associate_blkcg(NULL);
493 if (refcount_dec_and_test(&cb->pending_ios))
494 finish_compressed_bio_write(cb);
498 static u64 bio_end_offset(struct bio *bio)
500 struct bio_vec *last = bio_last_bvec_all(bio);
502 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
506 * Add extra pages in the same compressed file extent so that we don't need to
507 * re-read the same extent again and again.
509 * NOTE: this won't work well for subpage, as for subpage read, we lock the
510 * full page then submit bio for each compressed/regular extents.
512 * This means, if we have several sectors in the same page points to the same
513 * on-disk compressed data, we will re-read the same extent many times and
514 * this function can only help for the next page.
516 static noinline int add_ra_bio_pages(struct inode *inode,
518 struct compressed_bio *cb,
519 int *memstall, unsigned long *pflags)
521 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
522 unsigned long end_index;
523 u64 cur = bio_end_offset(cb->orig_bio);
524 u64 isize = i_size_read(inode);
527 struct extent_map *em;
528 struct address_space *mapping = inode->i_mapping;
529 struct extent_map_tree *em_tree;
530 struct extent_io_tree *tree;
531 int sectors_missed = 0;
533 em_tree = &BTRFS_I(inode)->extent_tree;
534 tree = &BTRFS_I(inode)->io_tree;
540 * For current subpage support, we only support 64K page size,
541 * which means maximum compressed extent size (128K) is just 2x page
543 * This makes readahead less effective, so here disable readahead for
544 * subpage for now, until full compressed write is supported.
546 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
549 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
551 while (cur < compressed_end) {
553 u64 pg_index = cur >> PAGE_SHIFT;
556 if (pg_index > end_index)
559 page = xa_load(&mapping->i_pages, pg_index);
560 if (page && !xa_is_value(page)) {
561 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
562 fs_info->sectorsize_bits;
564 /* Beyond threshold, no need to continue */
565 if (sectors_missed > 4)
569 * Jump to next page start as we already have page for
572 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
576 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
581 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
583 /* There is already a page, skip to page end */
584 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
588 if (!*memstall && PageWorkingset(page)) {
589 psi_memstall_enter(pflags);
593 ret = set_page_extent_mapped(page);
600 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
601 lock_extent(tree, cur, page_end, NULL);
602 read_lock(&em_tree->lock);
603 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
604 read_unlock(&em_tree->lock);
607 * At this point, we have a locked page in the page cache for
608 * these bytes in the file. But, we have to make sure they map
609 * to this compressed extent on disk.
611 if (!em || cur < em->start ||
612 (cur + fs_info->sectorsize > extent_map_end(em)) ||
613 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
615 unlock_extent(tree, cur, page_end, NULL);
622 if (page->index == end_index) {
623 size_t zero_offset = offset_in_page(isize);
627 zeros = PAGE_SIZE - zero_offset;
628 memzero_page(page, zero_offset, zeros);
632 add_size = min(em->start + em->len, page_end + 1) - cur;
633 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
634 if (ret != add_size) {
635 unlock_extent(tree, cur, page_end, NULL);
641 * If it's subpage, we also need to increase its
642 * subpage::readers number, as at endio we will decrease
643 * subpage::readers and to unlock the page.
645 if (fs_info->sectorsize < PAGE_SIZE)
646 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
654 * for a compressed read, the bio we get passed has all the inode pages
655 * in it. We don't actually do IO on those pages but allocate new ones
656 * to hold the compressed pages on disk.
658 * bio->bi_iter.bi_sector points to the compressed extent on disk
659 * bio->bi_io_vec points to all of the inode pages
661 * After the compressed pages are read, we copy the bytes into the
662 * bio we were passed and then call the bio end_io calls
664 void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
667 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
668 struct extent_map_tree *em_tree;
669 struct compressed_bio *cb;
670 unsigned int compressed_len;
671 struct bio *comp_bio = NULL;
672 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
673 u64 cur_disk_byte = disk_bytenr;
674 u64 next_stripe_start;
678 struct extent_map *em;
679 unsigned long pflags;
685 em_tree = &BTRFS_I(inode)->extent_tree;
687 file_offset = bio_first_bvec_all(bio)->bv_offset +
688 page_offset(bio_first_page_all(bio));
690 /* we need the actual starting offset of this extent in the file */
691 read_lock(&em_tree->lock);
692 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
693 read_unlock(&em_tree->lock);
699 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
700 compressed_len = em->block_len;
701 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
703 ret = BLK_STS_RESOURCE;
707 refcount_set(&cb->pending_ios, 1);
708 cb->status = BLK_STS_OK;
711 cb->start = em->orig_start;
713 em_start = em->start;
715 cb->len = bio->bi_iter.bi_size;
716 cb->compressed_len = compressed_len;
717 cb->compress_type = em->compress_type;
723 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
724 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
725 if (!cb->compressed_pages) {
726 ret = BLK_STS_RESOURCE;
730 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
732 ret = BLK_STS_RESOURCE;
736 add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
738 /* include any pages we added in add_ra-bio_pages */
739 cb->len = bio->bi_iter.bi_size;
741 while (cur_disk_byte < disk_bytenr + compressed_len) {
742 u64 offset = cur_disk_byte - disk_bytenr;
743 unsigned int index = offset >> PAGE_SHIFT;
744 unsigned int real_size;
746 struct page *page = cb->compressed_pages[index];
749 /* Allocate new bio if submitted or not yet allocated */
751 comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
752 REQ_OP_READ, end_compressed_bio_read,
754 if (IS_ERR(comp_bio)) {
755 cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
760 * We should never reach next_stripe_start start as we will
761 * submit comp_bio when reach the boundary immediately.
763 ASSERT(cur_disk_byte != next_stripe_start);
765 * We have various limit on the real read size:
768 * - compressed length boundary
770 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
771 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
772 real_size = min_t(u64, real_size, compressed_len - offset);
773 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
775 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
777 * Maximum compressed extent is smaller than bio size limit,
778 * thus bio_add_page() should always success.
780 ASSERT(added == real_size);
781 cur_disk_byte += added;
783 /* Reached stripe boundary, need to submit */
784 if (cur_disk_byte == next_stripe_start)
787 /* Has finished the range, need to submit */
788 if (cur_disk_byte == disk_bytenr + compressed_len)
793 * Save the initial offset of this chunk, as there
794 * is no direct correlation between compressed pages and
795 * the original file offset. The field is only used for
796 * printing error messages.
798 btrfs_bio(comp_bio)->file_offset = file_offset;
800 ASSERT(comp_bio->bi_iter.bi_size);
801 btrfs_submit_bio(fs_info, comp_bio, mirror_num);
807 psi_memstall_leave(&pflags);
809 if (refcount_dec_and_test(&cb->pending_ios))
810 finish_compressed_bio_read(cb);
814 if (cb->compressed_pages) {
815 for (i = 0; i < cb->nr_pages; i++) {
816 if (cb->compressed_pages[i])
817 __free_page(cb->compressed_pages[i]);
821 kfree(cb->compressed_pages);
825 btrfs_bio_end_io(btrfs_bio(bio), ret);
830 * Heuristic uses systematic sampling to collect data from the input data
831 * range, the logic can be tuned by the following constants:
833 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
834 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
836 #define SAMPLING_READ_SIZE (16)
837 #define SAMPLING_INTERVAL (256)
840 * For statistical analysis of the input data we consider bytes that form a
841 * Galois Field of 256 objects. Each object has an attribute count, ie. how
842 * many times the object appeared in the sample.
844 #define BUCKET_SIZE (256)
847 * The size of the sample is based on a statistical sampling rule of thumb.
848 * The common way is to perform sampling tests as long as the number of
849 * elements in each cell is at least 5.
851 * Instead of 5, we choose 32 to obtain more accurate results.
852 * If the data contain the maximum number of symbols, which is 256, we obtain a
853 * sample size bound by 8192.
855 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
856 * from up to 512 locations.
858 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
859 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
865 struct heuristic_ws {
866 /* Partial copy of input data */
869 /* Buckets store counters for each byte value */
870 struct bucket_item *bucket;
872 struct bucket_item *bucket_b;
873 struct list_head list;
876 static struct workspace_manager heuristic_wsm;
878 static void free_heuristic_ws(struct list_head *ws)
880 struct heuristic_ws *workspace;
882 workspace = list_entry(ws, struct heuristic_ws, list);
884 kvfree(workspace->sample);
885 kfree(workspace->bucket);
886 kfree(workspace->bucket_b);
890 static struct list_head *alloc_heuristic_ws(unsigned int level)
892 struct heuristic_ws *ws;
894 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
896 return ERR_PTR(-ENOMEM);
898 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
902 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
906 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
910 INIT_LIST_HEAD(&ws->list);
913 free_heuristic_ws(&ws->list);
914 return ERR_PTR(-ENOMEM);
917 const struct btrfs_compress_op btrfs_heuristic_compress = {
918 .workspace_manager = &heuristic_wsm,
921 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
922 /* The heuristic is represented as compression type 0 */
923 &btrfs_heuristic_compress,
924 &btrfs_zlib_compress,
926 &btrfs_zstd_compress,
929 static struct list_head *alloc_workspace(int type, unsigned int level)
932 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
933 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
934 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
935 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
938 * This can't happen, the type is validated several times
939 * before we get here.
945 static void free_workspace(int type, struct list_head *ws)
948 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
949 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
950 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
951 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
954 * This can't happen, the type is validated several times
955 * before we get here.
961 static void btrfs_init_workspace_manager(int type)
963 struct workspace_manager *wsm;
964 struct list_head *workspace;
966 wsm = btrfs_compress_op[type]->workspace_manager;
967 INIT_LIST_HEAD(&wsm->idle_ws);
968 spin_lock_init(&wsm->ws_lock);
969 atomic_set(&wsm->total_ws, 0);
970 init_waitqueue_head(&wsm->ws_wait);
973 * Preallocate one workspace for each compression type so we can
974 * guarantee forward progress in the worst case
976 workspace = alloc_workspace(type, 0);
977 if (IS_ERR(workspace)) {
979 "BTRFS: cannot preallocate compression workspace, will try later\n");
981 atomic_set(&wsm->total_ws, 1);
983 list_add(workspace, &wsm->idle_ws);
987 static void btrfs_cleanup_workspace_manager(int type)
989 struct workspace_manager *wsman;
990 struct list_head *ws;
992 wsman = btrfs_compress_op[type]->workspace_manager;
993 while (!list_empty(&wsman->idle_ws)) {
994 ws = wsman->idle_ws.next;
996 free_workspace(type, ws);
997 atomic_dec(&wsman->total_ws);
1002 * This finds an available workspace or allocates a new one.
1003 * If it's not possible to allocate a new one, waits until there's one.
1004 * Preallocation makes a forward progress guarantees and we do not return
1007 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1009 struct workspace_manager *wsm;
1010 struct list_head *workspace;
1011 int cpus = num_online_cpus();
1013 struct list_head *idle_ws;
1014 spinlock_t *ws_lock;
1016 wait_queue_head_t *ws_wait;
1019 wsm = btrfs_compress_op[type]->workspace_manager;
1020 idle_ws = &wsm->idle_ws;
1021 ws_lock = &wsm->ws_lock;
1022 total_ws = &wsm->total_ws;
1023 ws_wait = &wsm->ws_wait;
1024 free_ws = &wsm->free_ws;
1028 if (!list_empty(idle_ws)) {
1029 workspace = idle_ws->next;
1030 list_del(workspace);
1032 spin_unlock(ws_lock);
1036 if (atomic_read(total_ws) > cpus) {
1039 spin_unlock(ws_lock);
1040 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1041 if (atomic_read(total_ws) > cpus && !*free_ws)
1043 finish_wait(ws_wait, &wait);
1046 atomic_inc(total_ws);
1047 spin_unlock(ws_lock);
1050 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1051 * to turn it off here because we might get called from the restricted
1052 * context of btrfs_compress_bio/btrfs_compress_pages
1054 nofs_flag = memalloc_nofs_save();
1055 workspace = alloc_workspace(type, level);
1056 memalloc_nofs_restore(nofs_flag);
1058 if (IS_ERR(workspace)) {
1059 atomic_dec(total_ws);
1063 * Do not return the error but go back to waiting. There's a
1064 * workspace preallocated for each type and the compression
1065 * time is bounded so we get to a workspace eventually. This
1066 * makes our caller's life easier.
1068 * To prevent silent and low-probability deadlocks (when the
1069 * initial preallocation fails), check if there are any
1070 * workspaces at all.
1072 if (atomic_read(total_ws) == 0) {
1073 static DEFINE_RATELIMIT_STATE(_rs,
1074 /* once per minute */ 60 * HZ,
1077 if (__ratelimit(&_rs)) {
1078 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1086 static struct list_head *get_workspace(int type, int level)
1089 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1090 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1091 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1092 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1095 * This can't happen, the type is validated several times
1096 * before we get here.
1103 * put a workspace struct back on the list or free it if we have enough
1104 * idle ones sitting around
1106 void btrfs_put_workspace(int type, struct list_head *ws)
1108 struct workspace_manager *wsm;
1109 struct list_head *idle_ws;
1110 spinlock_t *ws_lock;
1112 wait_queue_head_t *ws_wait;
1115 wsm = btrfs_compress_op[type]->workspace_manager;
1116 idle_ws = &wsm->idle_ws;
1117 ws_lock = &wsm->ws_lock;
1118 total_ws = &wsm->total_ws;
1119 ws_wait = &wsm->ws_wait;
1120 free_ws = &wsm->free_ws;
1123 if (*free_ws <= num_online_cpus()) {
1124 list_add(ws, idle_ws);
1126 spin_unlock(ws_lock);
1129 spin_unlock(ws_lock);
1131 free_workspace(type, ws);
1132 atomic_dec(total_ws);
1134 cond_wake_up(ws_wait);
1137 static void put_workspace(int type, struct list_head *ws)
1140 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1141 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1142 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1143 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1146 * This can't happen, the type is validated several times
1147 * before we get here.
1154 * Adjust @level according to the limits of the compression algorithm or
1155 * fallback to default
1157 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1159 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1162 level = ops->default_level;
1164 level = min(level, ops->max_level);
1170 * Given an address space and start and length, compress the bytes into @pages
1171 * that are allocated on demand.
1173 * @type_level is encoded algorithm and level, where level 0 means whatever
1174 * default the algorithm chooses and is opaque here;
1175 * - compression algo are 0-3
1176 * - the level are bits 4-7
1178 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1179 * and returns number of actually allocated pages
1181 * @total_in is used to return the number of bytes actually read. It
1182 * may be smaller than the input length if we had to exit early because we
1183 * ran out of room in the pages array or because we cross the
1184 * max_out threshold.
1186 * @total_out is an in/out parameter, must be set to the input length and will
1187 * be also used to return the total number of compressed bytes
1189 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1190 u64 start, struct page **pages,
1191 unsigned long *out_pages,
1192 unsigned long *total_in,
1193 unsigned long *total_out)
1195 int type = btrfs_compress_type(type_level);
1196 int level = btrfs_compress_level(type_level);
1197 struct list_head *workspace;
1200 level = btrfs_compress_set_level(type, level);
1201 workspace = get_workspace(type, level);
1202 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1203 out_pages, total_in, total_out);
1204 put_workspace(type, workspace);
1208 static int btrfs_decompress_bio(struct compressed_bio *cb)
1210 struct list_head *workspace;
1212 int type = cb->compress_type;
1214 workspace = get_workspace(type, 0);
1215 ret = compression_decompress_bio(workspace, cb);
1216 put_workspace(type, workspace);
1222 * a less complex decompression routine. Our compressed data fits in a
1223 * single page, and we want to read a single page out of it.
1224 * start_byte tells us the offset into the compressed data we're interested in
1226 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1227 unsigned long start_byte, size_t srclen, size_t destlen)
1229 struct list_head *workspace;
1232 workspace = get_workspace(type, 0);
1233 ret = compression_decompress(type, workspace, data_in, dest_page,
1234 start_byte, srclen, destlen);
1235 put_workspace(type, workspace);
1240 int __init btrfs_init_compress(void)
1242 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1243 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1244 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1245 zstd_init_workspace_manager();
1249 void __cold btrfs_exit_compress(void)
1251 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1252 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1253 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1254 zstd_cleanup_workspace_manager();
1258 * Copy decompressed data from working buffer to pages.
1260 * @buf: The decompressed data buffer
1261 * @buf_len: The decompressed data length
1262 * @decompressed: Number of bytes that are already decompressed inside the
1264 * @cb: The compressed extent descriptor
1265 * @orig_bio: The original bio that the caller wants to read for
1267 * An easier to understand graph is like below:
1269 * |<- orig_bio ->| |<- orig_bio->|
1270 * |<------- full decompressed extent ----->|
1271 * |<----------- @cb range ---->|
1272 * | |<-- @buf_len -->|
1273 * |<--- @decompressed --->|
1275 * Note that, @cb can be a subpage of the full decompressed extent, but
1276 * @cb->start always has the same as the orig_file_offset value of the full
1277 * decompressed extent.
1279 * When reading compressed extent, we have to read the full compressed extent,
1280 * while @orig_bio may only want part of the range.
1281 * Thus this function will ensure only data covered by @orig_bio will be copied
1284 * Return 0 if we have copied all needed contents for @orig_bio.
1285 * Return >0 if we need continue decompress.
1287 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1288 struct compressed_bio *cb, u32 decompressed)
1290 struct bio *orig_bio = cb->orig_bio;
1291 /* Offset inside the full decompressed extent */
1294 cur_offset = decompressed;
1295 /* The main loop to do the copy */
1296 while (cur_offset < decompressed + buf_len) {
1297 struct bio_vec bvec;
1300 /* Offset inside the full decompressed extent */
1303 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1305 * cb->start may underflow, but subtracting that value can still
1306 * give us correct offset inside the full decompressed extent.
1308 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1310 /* Haven't reached the bvec range, exit */
1311 if (decompressed + buf_len <= bvec_offset)
1314 copy_start = max(cur_offset, bvec_offset);
1315 copy_len = min(bvec_offset + bvec.bv_len,
1316 decompressed + buf_len) - copy_start;
1320 * Extra range check to ensure we didn't go beyond
1323 ASSERT(copy_start - decompressed < buf_len);
1324 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1325 buf + copy_start - decompressed, copy_len);
1326 cur_offset += copy_len;
1328 bio_advance(orig_bio, copy_len);
1329 /* Finished the bio */
1330 if (!orig_bio->bi_iter.bi_size)
1337 * Shannon Entropy calculation
1339 * Pure byte distribution analysis fails to determine compressibility of data.
1340 * Try calculating entropy to estimate the average minimum number of bits
1341 * needed to encode the sampled data.
1343 * For convenience, return the percentage of needed bits, instead of amount of
1346 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1347 * and can be compressible with high probability
1349 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1351 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1353 #define ENTROPY_LVL_ACEPTABLE (65)
1354 #define ENTROPY_LVL_HIGH (80)
1357 * For increasead precision in shannon_entropy calculation,
1358 * let's do pow(n, M) to save more digits after comma:
1360 * - maximum int bit length is 64
1361 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1362 * - 13 * 4 = 52 < 64 -> M = 4
1366 static inline u32 ilog2_w(u64 n)
1368 return ilog2(n * n * n * n);
1371 static u32 shannon_entropy(struct heuristic_ws *ws)
1373 const u32 entropy_max = 8 * ilog2_w(2);
1374 u32 entropy_sum = 0;
1375 u32 p, p_base, sz_base;
1378 sz_base = ilog2_w(ws->sample_size);
1379 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1380 p = ws->bucket[i].count;
1381 p_base = ilog2_w(p);
1382 entropy_sum += p * (sz_base - p_base);
1385 entropy_sum /= ws->sample_size;
1386 return entropy_sum * 100 / entropy_max;
1389 #define RADIX_BASE 4U
1390 #define COUNTERS_SIZE (1U << RADIX_BASE)
1392 static u8 get4bits(u64 num, int shift) {
1397 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1402 * Use 4 bits as radix base
1403 * Use 16 u32 counters for calculating new position in buf array
1405 * @array - array that will be sorted
1406 * @array_buf - buffer array to store sorting results
1407 * must be equal in size to @array
1410 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1415 u32 counters[COUNTERS_SIZE];
1423 * Try avoid useless loop iterations for small numbers stored in big
1424 * counters. Example: 48 33 4 ... in 64bit array
1426 max_num = array[0].count;
1427 for (i = 1; i < num; i++) {
1428 buf_num = array[i].count;
1429 if (buf_num > max_num)
1433 buf_num = ilog2(max_num);
1434 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1437 while (shift < bitlen) {
1438 memset(counters, 0, sizeof(counters));
1440 for (i = 0; i < num; i++) {
1441 buf_num = array[i].count;
1442 addr = get4bits(buf_num, shift);
1446 for (i = 1; i < COUNTERS_SIZE; i++)
1447 counters[i] += counters[i - 1];
1449 for (i = num - 1; i >= 0; i--) {
1450 buf_num = array[i].count;
1451 addr = get4bits(buf_num, shift);
1453 new_addr = counters[addr];
1454 array_buf[new_addr] = array[i];
1457 shift += RADIX_BASE;
1460 * Normal radix expects to move data from a temporary array, to
1461 * the main one. But that requires some CPU time. Avoid that
1462 * by doing another sort iteration to original array instead of
1465 memset(counters, 0, sizeof(counters));
1467 for (i = 0; i < num; i ++) {
1468 buf_num = array_buf[i].count;
1469 addr = get4bits(buf_num, shift);
1473 for (i = 1; i < COUNTERS_SIZE; i++)
1474 counters[i] += counters[i - 1];
1476 for (i = num - 1; i >= 0; i--) {
1477 buf_num = array_buf[i].count;
1478 addr = get4bits(buf_num, shift);
1480 new_addr = counters[addr];
1481 array[new_addr] = array_buf[i];
1484 shift += RADIX_BASE;
1489 * Size of the core byte set - how many bytes cover 90% of the sample
1491 * There are several types of structured binary data that use nearly all byte
1492 * values. The distribution can be uniform and counts in all buckets will be
1493 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1495 * Other possibility is normal (Gaussian) distribution, where the data could
1496 * be potentially compressible, but we have to take a few more steps to decide
1499 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1500 * compression algo can easy fix that
1501 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1502 * probability is not compressible
1504 #define BYTE_CORE_SET_LOW (64)
1505 #define BYTE_CORE_SET_HIGH (200)
1507 static int byte_core_set_size(struct heuristic_ws *ws)
1510 u32 coreset_sum = 0;
1511 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1512 struct bucket_item *bucket = ws->bucket;
1514 /* Sort in reverse order */
1515 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1517 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1518 coreset_sum += bucket[i].count;
1520 if (coreset_sum > core_set_threshold)
1523 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1524 coreset_sum += bucket[i].count;
1525 if (coreset_sum > core_set_threshold)
1533 * Count byte values in buckets.
1534 * This heuristic can detect textual data (configs, xml, json, html, etc).
1535 * Because in most text-like data byte set is restricted to limited number of
1536 * possible characters, and that restriction in most cases makes data easy to
1539 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1540 * less - compressible
1541 * more - need additional analysis
1543 #define BYTE_SET_THRESHOLD (64)
1545 static u32 byte_set_size(const struct heuristic_ws *ws)
1548 u32 byte_set_size = 0;
1550 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1551 if (ws->bucket[i].count > 0)
1556 * Continue collecting count of byte values in buckets. If the byte
1557 * set size is bigger then the threshold, it's pointless to continue,
1558 * the detection technique would fail for this type of data.
1560 for (; i < BUCKET_SIZE; i++) {
1561 if (ws->bucket[i].count > 0) {
1563 if (byte_set_size > BYTE_SET_THRESHOLD)
1564 return byte_set_size;
1568 return byte_set_size;
1571 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1573 const u32 half_of_sample = ws->sample_size / 2;
1574 const u8 *data = ws->sample;
1576 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1579 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1580 struct heuristic_ws *ws)
1583 u64 index, index_end;
1584 u32 i, curr_sample_pos;
1588 * Compression handles the input data by chunks of 128KiB
1589 * (defined by BTRFS_MAX_UNCOMPRESSED)
1591 * We do the same for the heuristic and loop over the whole range.
1593 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1594 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1596 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1597 end = start + BTRFS_MAX_UNCOMPRESSED;
1599 index = start >> PAGE_SHIFT;
1600 index_end = end >> PAGE_SHIFT;
1602 /* Don't miss unaligned end */
1603 if (!PAGE_ALIGNED(end))
1606 curr_sample_pos = 0;
1607 while (index < index_end) {
1608 page = find_get_page(inode->i_mapping, index);
1609 in_data = kmap_local_page(page);
1610 /* Handle case where the start is not aligned to PAGE_SIZE */
1611 i = start % PAGE_SIZE;
1612 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1613 /* Don't sample any garbage from the last page */
1614 if (start > end - SAMPLING_READ_SIZE)
1616 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1617 SAMPLING_READ_SIZE);
1618 i += SAMPLING_INTERVAL;
1619 start += SAMPLING_INTERVAL;
1620 curr_sample_pos += SAMPLING_READ_SIZE;
1622 kunmap_local(in_data);
1628 ws->sample_size = curr_sample_pos;
1632 * Compression heuristic.
1634 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1635 * quickly (compared to direct compression) detect data characteristics
1636 * (compressible/incompressible) to avoid wasting CPU time on incompressible
1639 * The following types of analysis can be performed:
1640 * - detect mostly zero data
1641 * - detect data with low "byte set" size (text, etc)
1642 * - detect data with low/high "core byte" set
1644 * Return non-zero if the compression should be done, 0 otherwise.
1646 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1648 struct list_head *ws_list = get_workspace(0, 0);
1649 struct heuristic_ws *ws;
1654 ws = list_entry(ws_list, struct heuristic_ws, list);
1656 heuristic_collect_sample(inode, start, end, ws);
1658 if (sample_repeated_patterns(ws)) {
1663 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1665 for (i = 0; i < ws->sample_size; i++) {
1666 byte = ws->sample[i];
1667 ws->bucket[byte].count++;
1670 i = byte_set_size(ws);
1671 if (i < BYTE_SET_THRESHOLD) {
1676 i = byte_core_set_size(ws);
1677 if (i <= BYTE_CORE_SET_LOW) {
1682 if (i >= BYTE_CORE_SET_HIGH) {
1687 i = shannon_entropy(ws);
1688 if (i <= ENTROPY_LVL_ACEPTABLE) {
1694 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1695 * needed to give green light to compression.
1697 * For now just assume that compression at that level is not worth the
1698 * resources because:
1700 * 1. it is possible to defrag the data later
1702 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1703 * values, every bucket has counter at level ~54. The heuristic would
1704 * be confused. This can happen when data have some internal repeated
1705 * patterns like "abbacbbc...". This can be detected by analyzing
1706 * pairs of bytes, which is too costly.
1708 if (i < ENTROPY_LVL_HIGH) {
1717 put_workspace(0, ws_list);
1722 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1723 * level, unrecognized string will set the default level
1725 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1727 unsigned int level = 0;
1733 if (str[0] == ':') {
1734 ret = kstrtouint(str + 1, 10, &level);
1739 level = btrfs_compress_set_level(type, level);