1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2008 Oracle. All rights reserved.
6 #include <linux/kernel.h>
8 #include <linux/file.h>
10 #include <linux/pagemap.h>
11 #include <linux/highmem.h>
12 #include <linux/kthread.h>
13 #include <linux/time.h>
14 #include <linux/init.h>
15 #include <linux/string.h>
16 #include <linux/backing-dev.h>
17 #include <linux/writeback.h>
18 #include <linux/slab.h>
19 #include <linux/sched/mm.h>
20 #include <linux/log2.h>
21 #include <crypto/hash.h>
25 #include "transaction.h"
26 #include "btrfs_inode.h"
28 #include "ordered-data.h"
29 #include "compression.h"
30 #include "extent_io.h"
31 #include "extent_map.h"
35 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
37 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
40 case BTRFS_COMPRESS_ZLIB:
41 case BTRFS_COMPRESS_LZO:
42 case BTRFS_COMPRESS_ZSTD:
43 case BTRFS_COMPRESS_NONE:
44 return btrfs_compress_types[type];
52 bool btrfs_compress_is_valid_type(const char *str, size_t len)
56 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
57 size_t comp_len = strlen(btrfs_compress_types[i]);
62 if (!strncmp(btrfs_compress_types[i], str, comp_len))
68 static int compression_compress_pages(int type, struct list_head *ws,
69 struct address_space *mapping, u64 start, struct page **pages,
70 unsigned long *out_pages, unsigned long *total_in,
71 unsigned long *total_out)
74 case BTRFS_COMPRESS_ZLIB:
75 return zlib_compress_pages(ws, mapping, start, pages,
76 out_pages, total_in, total_out);
77 case BTRFS_COMPRESS_LZO:
78 return lzo_compress_pages(ws, mapping, start, pages,
79 out_pages, total_in, total_out);
80 case BTRFS_COMPRESS_ZSTD:
81 return zstd_compress_pages(ws, mapping, start, pages,
82 out_pages, total_in, total_out);
83 case BTRFS_COMPRESS_NONE:
86 * This can happen when compression races with remount setting
87 * it to 'no compress', while caller doesn't call
88 * inode_need_compress() to check if we really need to
91 * Not a big deal, just need to inform caller that we
92 * haven't allocated any pages yet.
99 static int compression_decompress_bio(struct list_head *ws,
100 struct compressed_bio *cb)
102 switch (cb->compress_type) {
103 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
104 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
105 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
106 case BTRFS_COMPRESS_NONE:
109 * This can't happen, the type is validated several times
110 * before we get here.
116 static int compression_decompress(int type, struct list_head *ws,
117 unsigned char *data_in, struct page *dest_page,
118 unsigned long start_byte, size_t srclen, size_t destlen)
121 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
122 start_byte, srclen, destlen);
123 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
124 start_byte, srclen, destlen);
125 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
126 start_byte, srclen, destlen);
127 case BTRFS_COMPRESS_NONE:
130 * This can't happen, the type is validated several times
131 * before we get here.
137 static int btrfs_decompress_bio(struct compressed_bio *cb);
139 static void finish_compressed_bio_read(struct compressed_bio *cb)
144 if (cb->status == BLK_STS_OK)
145 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
147 /* Release the compressed pages */
148 for (index = 0; index < cb->nr_pages; index++) {
149 page = cb->compressed_pages[index];
150 page->mapping = NULL;
154 /* Do io completion on the original bio */
155 btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
157 /* Finally free the cb struct */
158 kfree(cb->compressed_pages);
163 * Verify the checksums and kick off repair if needed on the uncompressed data
164 * before decompressing it into the original bio and freeing the uncompressed
167 static void end_compressed_bio_read(struct btrfs_bio *bbio)
169 struct compressed_bio *cb = bbio->private;
170 struct inode *inode = cb->inode;
171 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
172 struct btrfs_inode *bi = BTRFS_I(inode);
173 bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
174 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
175 blk_status_t status = bbio->bio.bi_status;
176 struct bvec_iter iter;
180 btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
181 u64 start = bbio->file_offset + offset;
184 (!csum || !btrfs_check_data_csum(inode, bbio, offset,
185 bv.bv_page, bv.bv_offset))) {
186 btrfs_clean_io_failure(bi, start, bv.bv_page,
191 refcount_inc(&cb->pending_ios);
192 ret = btrfs_repair_one_sector(inode, bbio, offset,
193 bv.bv_page, bv.bv_offset,
194 btrfs_submit_data_read_bio);
196 refcount_dec(&cb->pending_ios);
197 status = errno_to_blk_status(ret);
205 if (refcount_dec_and_test(&cb->pending_ios))
206 finish_compressed_bio_read(cb);
207 btrfs_bio_free_csum(bbio);
212 * Clear the writeback bits on all of the file
213 * pages for a compressed write
215 static noinline void end_compressed_writeback(struct inode *inode,
216 const struct compressed_bio *cb)
218 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
219 unsigned long index = cb->start >> PAGE_SHIFT;
220 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
221 struct page *pages[16];
222 unsigned long nr_pages = end_index - index + 1;
223 const int errno = blk_status_to_errno(cb->status);
228 mapping_set_error(inode->i_mapping, errno);
230 while (nr_pages > 0) {
231 ret = find_get_pages_contig(inode->i_mapping, index,
233 nr_pages, ARRAY_SIZE(pages)), pages);
239 for (i = 0; i < ret; i++) {
241 SetPageError(pages[i]);
242 btrfs_page_clamp_clear_writeback(fs_info, pages[i],
249 /* the inode may be gone now */
252 static void finish_compressed_bio_write(struct compressed_bio *cb)
254 struct inode *inode = cb->inode;
258 * Ok, we're the last bio for this extent, step one is to call back
259 * into the FS and do all the end_io operations.
261 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
262 cb->start, cb->start + cb->len - 1,
263 cb->status == BLK_STS_OK);
266 end_compressed_writeback(inode, cb);
267 /* Note, our inode could be gone now */
270 * Release the compressed pages, these came from alloc_page and
271 * are not attached to the inode at all
273 for (index = 0; index < cb->nr_pages; index++) {
274 struct page *page = cb->compressed_pages[index];
276 page->mapping = NULL;
280 /* Finally free the cb struct */
281 kfree(cb->compressed_pages);
285 static void btrfs_finish_compressed_write_work(struct work_struct *work)
287 struct compressed_bio *cb =
288 container_of(work, struct compressed_bio, write_end_work);
290 finish_compressed_bio_write(cb);
294 * Do the cleanup once all the compressed pages hit the disk. This will clear
295 * writeback on the file pages and free the compressed pages.
297 * This also calls the writeback end hooks for the file pages so that metadata
298 * and checksums can be updated in the file.
300 static void end_compressed_bio_write(struct btrfs_bio *bbio)
302 struct compressed_bio *cb = bbio->private;
304 if (bbio->bio.bi_status)
305 cb->status = bbio->bio.bi_status;
307 if (refcount_dec_and_test(&cb->pending_ios)) {
308 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
310 btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio);
311 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
317 * Allocate a compressed_bio, which will be used to read/write on-disk
318 * (aka, compressed) * data.
320 * @cb: The compressed_bio structure, which records all the needed
321 * information to bind the compressed data to the uncompressed
323 * @disk_byten: The logical bytenr where the compressed data will be read
324 * from or written to.
325 * @endio_func: The endio function to call after the IO for compressed data
327 * @next_stripe_start: Return value of logical bytenr of where next stripe starts.
328 * Let the caller know to only fill the bio up to the stripe
333 static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
335 btrfs_bio_end_io_t endio_func,
336 u64 *next_stripe_start)
338 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
339 struct btrfs_io_geometry geom;
340 struct extent_map *em;
344 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb);
345 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
347 em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
353 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
354 bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
356 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
362 *next_stripe_start = disk_bytenr + geom.len;
363 refcount_inc(&cb->pending_ios);
368 * worker function to build and submit bios for previously compressed pages.
369 * The corresponding pages in the inode should be marked for writeback
370 * and the compressed pages should have a reference on them for dropping
371 * when the IO is complete.
373 * This also checksums the file bytes and gets things ready for
376 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
377 unsigned int len, u64 disk_start,
378 unsigned int compressed_len,
379 struct page **compressed_pages,
380 unsigned int nr_pages,
381 blk_opf_t write_flags,
382 struct cgroup_subsys_state *blkcg_css,
385 struct btrfs_fs_info *fs_info = inode->root->fs_info;
386 struct bio *bio = NULL;
387 struct compressed_bio *cb;
388 u64 cur_disk_bytenr = disk_start;
389 u64 next_stripe_start;
390 blk_status_t ret = BLK_STS_OK;
391 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
392 const bool use_append = btrfs_use_zone_append(inode, disk_start);
393 const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
395 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
396 IS_ALIGNED(len, fs_info->sectorsize));
397 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
399 return BLK_STS_RESOURCE;
400 refcount_set(&cb->pending_ios, 1);
401 cb->status = BLK_STS_OK;
402 cb->inode = &inode->vfs_inode;
405 cb->compressed_pages = compressed_pages;
406 cb->compressed_len = compressed_len;
407 cb->writeback = writeback;
408 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
409 cb->nr_pages = nr_pages;
412 kthread_associate_blkcg(blkcg_css);
414 while (cur_disk_bytenr < disk_start + compressed_len) {
415 u64 offset = cur_disk_bytenr - disk_start;
416 unsigned int index = offset >> PAGE_SHIFT;
417 unsigned int real_size;
419 struct page *page = compressed_pages[index];
422 /* Allocate new bio if submitted or not yet allocated */
424 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
425 bio_op | write_flags, end_compressed_bio_write,
428 ret = errno_to_blk_status(PTR_ERR(bio));
432 bio->bi_opf |= REQ_CGROUP_PUNT;
435 * We should never reach next_stripe_start start as we will
436 * submit comp_bio when reach the boundary immediately.
438 ASSERT(cur_disk_bytenr != next_stripe_start);
441 * We have various limits on the real read size:
444 * - compressed length boundary
446 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
447 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
448 real_size = min_t(u64, real_size, compressed_len - offset);
449 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
452 added = bio_add_zone_append_page(bio, page, real_size,
453 offset_in_page(offset));
455 added = bio_add_page(bio, page, real_size,
456 offset_in_page(offset));
457 /* Reached zoned boundary */
461 cur_disk_bytenr += added;
462 /* Reached stripe boundary */
463 if (cur_disk_bytenr == next_stripe_start)
466 /* Finished the range */
467 if (cur_disk_bytenr == disk_start + compressed_len)
472 ret = btrfs_csum_one_bio(inode, bio, start, true);
474 btrfs_bio_end_io(btrfs_bio(bio), ret);
479 ASSERT(bio->bi_iter.bi_size);
480 btrfs_submit_bio(fs_info, bio, 0);
487 kthread_associate_blkcg(NULL);
489 if (refcount_dec_and_test(&cb->pending_ios))
490 finish_compressed_bio_write(cb);
494 static u64 bio_end_offset(struct bio *bio)
496 struct bio_vec *last = bio_last_bvec_all(bio);
498 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
502 * Add extra pages in the same compressed file extent so that we don't need to
503 * re-read the same extent again and again.
505 * NOTE: this won't work well for subpage, as for subpage read, we lock the
506 * full page then submit bio for each compressed/regular extents.
508 * This means, if we have several sectors in the same page points to the same
509 * on-disk compressed data, we will re-read the same extent many times and
510 * this function can only help for the next page.
512 static noinline int add_ra_bio_pages(struct inode *inode,
514 struct compressed_bio *cb)
516 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
517 unsigned long end_index;
518 u64 cur = bio_end_offset(cb->orig_bio);
519 u64 isize = i_size_read(inode);
522 struct extent_map *em;
523 struct address_space *mapping = inode->i_mapping;
524 struct extent_map_tree *em_tree;
525 struct extent_io_tree *tree;
526 int sectors_missed = 0;
528 em_tree = &BTRFS_I(inode)->extent_tree;
529 tree = &BTRFS_I(inode)->io_tree;
535 * For current subpage support, we only support 64K page size,
536 * which means maximum compressed extent size (128K) is just 2x page
538 * This makes readahead less effective, so here disable readahead for
539 * subpage for now, until full compressed write is supported.
541 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
544 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
546 while (cur < compressed_end) {
548 u64 pg_index = cur >> PAGE_SHIFT;
551 if (pg_index > end_index)
554 page = xa_load(&mapping->i_pages, pg_index);
555 if (page && !xa_is_value(page)) {
556 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
557 fs_info->sectorsize_bits;
559 /* Beyond threshold, no need to continue */
560 if (sectors_missed > 4)
564 * Jump to next page start as we already have page for
567 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
571 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
576 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
578 /* There is already a page, skip to page end */
579 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
583 ret = set_page_extent_mapped(page);
590 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
591 lock_extent(tree, cur, page_end, NULL);
592 read_lock(&em_tree->lock);
593 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
594 read_unlock(&em_tree->lock);
597 * At this point, we have a locked page in the page cache for
598 * these bytes in the file. But, we have to make sure they map
599 * to this compressed extent on disk.
601 if (!em || cur < em->start ||
602 (cur + fs_info->sectorsize > extent_map_end(em)) ||
603 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
605 unlock_extent(tree, cur, page_end, NULL);
612 if (page->index == end_index) {
613 size_t zero_offset = offset_in_page(isize);
617 zeros = PAGE_SIZE - zero_offset;
618 memzero_page(page, zero_offset, zeros);
622 add_size = min(em->start + em->len, page_end + 1) - cur;
623 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
624 if (ret != add_size) {
625 unlock_extent(tree, cur, page_end, NULL);
631 * If it's subpage, we also need to increase its
632 * subpage::readers number, as at endio we will decrease
633 * subpage::readers and to unlock the page.
635 if (fs_info->sectorsize < PAGE_SIZE)
636 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
644 * for a compressed read, the bio we get passed has all the inode pages
645 * in it. We don't actually do IO on those pages but allocate new ones
646 * to hold the compressed pages on disk.
648 * bio->bi_iter.bi_sector points to the compressed extent on disk
649 * bio->bi_io_vec points to all of the inode pages
651 * After the compressed pages are read, we copy the bytes into the
652 * bio we were passed and then call the bio end_io calls
654 void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
657 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
658 struct extent_map_tree *em_tree;
659 struct compressed_bio *cb;
660 unsigned int compressed_len;
661 struct bio *comp_bio = NULL;
662 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
663 u64 cur_disk_byte = disk_bytenr;
664 u64 next_stripe_start;
668 struct extent_map *em;
673 em_tree = &BTRFS_I(inode)->extent_tree;
675 file_offset = bio_first_bvec_all(bio)->bv_offset +
676 page_offset(bio_first_page_all(bio));
678 /* we need the actual starting offset of this extent in the file */
679 read_lock(&em_tree->lock);
680 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
681 read_unlock(&em_tree->lock);
687 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
688 compressed_len = em->block_len;
689 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
691 ret = BLK_STS_RESOURCE;
695 refcount_set(&cb->pending_ios, 1);
696 cb->status = BLK_STS_OK;
699 cb->start = em->orig_start;
701 em_start = em->start;
703 cb->len = bio->bi_iter.bi_size;
704 cb->compressed_len = compressed_len;
705 cb->compress_type = em->compress_type;
711 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
712 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
713 if (!cb->compressed_pages) {
714 ret = BLK_STS_RESOURCE;
718 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
720 ret = BLK_STS_RESOURCE;
724 add_ra_bio_pages(inode, em_start + em_len, cb);
726 /* include any pages we added in add_ra-bio_pages */
727 cb->len = bio->bi_iter.bi_size;
729 while (cur_disk_byte < disk_bytenr + compressed_len) {
730 u64 offset = cur_disk_byte - disk_bytenr;
731 unsigned int index = offset >> PAGE_SHIFT;
732 unsigned int real_size;
734 struct page *page = cb->compressed_pages[index];
737 /* Allocate new bio if submitted or not yet allocated */
739 comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
740 REQ_OP_READ, end_compressed_bio_read,
742 if (IS_ERR(comp_bio)) {
743 cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
748 * We should never reach next_stripe_start start as we will
749 * submit comp_bio when reach the boundary immediately.
751 ASSERT(cur_disk_byte != next_stripe_start);
753 * We have various limit on the real read size:
756 * - compressed length boundary
758 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
759 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
760 real_size = min_t(u64, real_size, compressed_len - offset);
761 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
763 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
765 * Maximum compressed extent is smaller than bio size limit,
766 * thus bio_add_page() should always success.
768 ASSERT(added == real_size);
769 cur_disk_byte += added;
771 /* Reached stripe boundary, need to submit */
772 if (cur_disk_byte == next_stripe_start)
775 /* Has finished the range, need to submit */
776 if (cur_disk_byte == disk_bytenr + compressed_len)
780 /* Save the original iter for read repair */
781 if (bio_op(comp_bio) == REQ_OP_READ)
782 btrfs_bio(comp_bio)->iter = comp_bio->bi_iter;
785 * Save the initial offset of this chunk, as there
786 * is no direct correlation between compressed pages and
787 * the original file offset. The field is only used for
788 * priting error messages.
790 btrfs_bio(comp_bio)->file_offset = file_offset;
792 ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL);
794 btrfs_bio_end_io(btrfs_bio(comp_bio), ret);
798 ASSERT(comp_bio->bi_iter.bi_size);
799 btrfs_submit_bio(fs_info, comp_bio, mirror_num);
804 if (refcount_dec_and_test(&cb->pending_ios))
805 finish_compressed_bio_read(cb);
809 if (cb->compressed_pages) {
810 for (i = 0; i < cb->nr_pages; i++) {
811 if (cb->compressed_pages[i])
812 __free_page(cb->compressed_pages[i]);
816 kfree(cb->compressed_pages);
820 btrfs_bio_end_io(btrfs_bio(bio), ret);
825 * Heuristic uses systematic sampling to collect data from the input data
826 * range, the logic can be tuned by the following constants:
828 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
829 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
831 #define SAMPLING_READ_SIZE (16)
832 #define SAMPLING_INTERVAL (256)
835 * For statistical analysis of the input data we consider bytes that form a
836 * Galois Field of 256 objects. Each object has an attribute count, ie. how
837 * many times the object appeared in the sample.
839 #define BUCKET_SIZE (256)
842 * The size of the sample is based on a statistical sampling rule of thumb.
843 * The common way is to perform sampling tests as long as the number of
844 * elements in each cell is at least 5.
846 * Instead of 5, we choose 32 to obtain more accurate results.
847 * If the data contain the maximum number of symbols, which is 256, we obtain a
848 * sample size bound by 8192.
850 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
851 * from up to 512 locations.
853 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
854 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
860 struct heuristic_ws {
861 /* Partial copy of input data */
864 /* Buckets store counters for each byte value */
865 struct bucket_item *bucket;
867 struct bucket_item *bucket_b;
868 struct list_head list;
871 static struct workspace_manager heuristic_wsm;
873 static void free_heuristic_ws(struct list_head *ws)
875 struct heuristic_ws *workspace;
877 workspace = list_entry(ws, struct heuristic_ws, list);
879 kvfree(workspace->sample);
880 kfree(workspace->bucket);
881 kfree(workspace->bucket_b);
885 static struct list_head *alloc_heuristic_ws(unsigned int level)
887 struct heuristic_ws *ws;
889 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
891 return ERR_PTR(-ENOMEM);
893 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
897 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
901 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
905 INIT_LIST_HEAD(&ws->list);
908 free_heuristic_ws(&ws->list);
909 return ERR_PTR(-ENOMEM);
912 const struct btrfs_compress_op btrfs_heuristic_compress = {
913 .workspace_manager = &heuristic_wsm,
916 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
917 /* The heuristic is represented as compression type 0 */
918 &btrfs_heuristic_compress,
919 &btrfs_zlib_compress,
921 &btrfs_zstd_compress,
924 static struct list_head *alloc_workspace(int type, unsigned int level)
927 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
928 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
929 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
930 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
933 * This can't happen, the type is validated several times
934 * before we get here.
940 static void free_workspace(int type, struct list_head *ws)
943 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
944 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
945 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
946 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
949 * This can't happen, the type is validated several times
950 * before we get here.
956 static void btrfs_init_workspace_manager(int type)
958 struct workspace_manager *wsm;
959 struct list_head *workspace;
961 wsm = btrfs_compress_op[type]->workspace_manager;
962 INIT_LIST_HEAD(&wsm->idle_ws);
963 spin_lock_init(&wsm->ws_lock);
964 atomic_set(&wsm->total_ws, 0);
965 init_waitqueue_head(&wsm->ws_wait);
968 * Preallocate one workspace for each compression type so we can
969 * guarantee forward progress in the worst case
971 workspace = alloc_workspace(type, 0);
972 if (IS_ERR(workspace)) {
974 "BTRFS: cannot preallocate compression workspace, will try later\n");
976 atomic_set(&wsm->total_ws, 1);
978 list_add(workspace, &wsm->idle_ws);
982 static void btrfs_cleanup_workspace_manager(int type)
984 struct workspace_manager *wsman;
985 struct list_head *ws;
987 wsman = btrfs_compress_op[type]->workspace_manager;
988 while (!list_empty(&wsman->idle_ws)) {
989 ws = wsman->idle_ws.next;
991 free_workspace(type, ws);
992 atomic_dec(&wsman->total_ws);
997 * This finds an available workspace or allocates a new one.
998 * If it's not possible to allocate a new one, waits until there's one.
999 * Preallocation makes a forward progress guarantees and we do not return
1002 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1004 struct workspace_manager *wsm;
1005 struct list_head *workspace;
1006 int cpus = num_online_cpus();
1008 struct list_head *idle_ws;
1009 spinlock_t *ws_lock;
1011 wait_queue_head_t *ws_wait;
1014 wsm = btrfs_compress_op[type]->workspace_manager;
1015 idle_ws = &wsm->idle_ws;
1016 ws_lock = &wsm->ws_lock;
1017 total_ws = &wsm->total_ws;
1018 ws_wait = &wsm->ws_wait;
1019 free_ws = &wsm->free_ws;
1023 if (!list_empty(idle_ws)) {
1024 workspace = idle_ws->next;
1025 list_del(workspace);
1027 spin_unlock(ws_lock);
1031 if (atomic_read(total_ws) > cpus) {
1034 spin_unlock(ws_lock);
1035 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1036 if (atomic_read(total_ws) > cpus && !*free_ws)
1038 finish_wait(ws_wait, &wait);
1041 atomic_inc(total_ws);
1042 spin_unlock(ws_lock);
1045 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1046 * to turn it off here because we might get called from the restricted
1047 * context of btrfs_compress_bio/btrfs_compress_pages
1049 nofs_flag = memalloc_nofs_save();
1050 workspace = alloc_workspace(type, level);
1051 memalloc_nofs_restore(nofs_flag);
1053 if (IS_ERR(workspace)) {
1054 atomic_dec(total_ws);
1058 * Do not return the error but go back to waiting. There's a
1059 * workspace preallocated for each type and the compression
1060 * time is bounded so we get to a workspace eventually. This
1061 * makes our caller's life easier.
1063 * To prevent silent and low-probability deadlocks (when the
1064 * initial preallocation fails), check if there are any
1065 * workspaces at all.
1067 if (atomic_read(total_ws) == 0) {
1068 static DEFINE_RATELIMIT_STATE(_rs,
1069 /* once per minute */ 60 * HZ,
1072 if (__ratelimit(&_rs)) {
1073 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1081 static struct list_head *get_workspace(int type, int level)
1084 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1085 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1086 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1087 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1090 * This can't happen, the type is validated several times
1091 * before we get here.
1098 * put a workspace struct back on the list or free it if we have enough
1099 * idle ones sitting around
1101 void btrfs_put_workspace(int type, struct list_head *ws)
1103 struct workspace_manager *wsm;
1104 struct list_head *idle_ws;
1105 spinlock_t *ws_lock;
1107 wait_queue_head_t *ws_wait;
1110 wsm = btrfs_compress_op[type]->workspace_manager;
1111 idle_ws = &wsm->idle_ws;
1112 ws_lock = &wsm->ws_lock;
1113 total_ws = &wsm->total_ws;
1114 ws_wait = &wsm->ws_wait;
1115 free_ws = &wsm->free_ws;
1118 if (*free_ws <= num_online_cpus()) {
1119 list_add(ws, idle_ws);
1121 spin_unlock(ws_lock);
1124 spin_unlock(ws_lock);
1126 free_workspace(type, ws);
1127 atomic_dec(total_ws);
1129 cond_wake_up(ws_wait);
1132 static void put_workspace(int type, struct list_head *ws)
1135 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1136 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1137 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1138 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1141 * This can't happen, the type is validated several times
1142 * before we get here.
1149 * Adjust @level according to the limits of the compression algorithm or
1150 * fallback to default
1152 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1154 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1157 level = ops->default_level;
1159 level = min(level, ops->max_level);
1165 * Given an address space and start and length, compress the bytes into @pages
1166 * that are allocated on demand.
1168 * @type_level is encoded algorithm and level, where level 0 means whatever
1169 * default the algorithm chooses and is opaque here;
1170 * - compression algo are 0-3
1171 * - the level are bits 4-7
1173 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1174 * and returns number of actually allocated pages
1176 * @total_in is used to return the number of bytes actually read. It
1177 * may be smaller than the input length if we had to exit early because we
1178 * ran out of room in the pages array or because we cross the
1179 * max_out threshold.
1181 * @total_out is an in/out parameter, must be set to the input length and will
1182 * be also used to return the total number of compressed bytes
1184 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1185 u64 start, struct page **pages,
1186 unsigned long *out_pages,
1187 unsigned long *total_in,
1188 unsigned long *total_out)
1190 int type = btrfs_compress_type(type_level);
1191 int level = btrfs_compress_level(type_level);
1192 struct list_head *workspace;
1195 level = btrfs_compress_set_level(type, level);
1196 workspace = get_workspace(type, level);
1197 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1198 out_pages, total_in, total_out);
1199 put_workspace(type, workspace);
1203 static int btrfs_decompress_bio(struct compressed_bio *cb)
1205 struct list_head *workspace;
1207 int type = cb->compress_type;
1209 workspace = get_workspace(type, 0);
1210 ret = compression_decompress_bio(workspace, cb);
1211 put_workspace(type, workspace);
1217 * a less complex decompression routine. Our compressed data fits in a
1218 * single page, and we want to read a single page out of it.
1219 * start_byte tells us the offset into the compressed data we're interested in
1221 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1222 unsigned long start_byte, size_t srclen, size_t destlen)
1224 struct list_head *workspace;
1227 workspace = get_workspace(type, 0);
1228 ret = compression_decompress(type, workspace, data_in, dest_page,
1229 start_byte, srclen, destlen);
1230 put_workspace(type, workspace);
1235 void __init btrfs_init_compress(void)
1237 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1238 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1239 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1240 zstd_init_workspace_manager();
1243 void __cold btrfs_exit_compress(void)
1245 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1246 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1247 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1248 zstd_cleanup_workspace_manager();
1252 * Copy decompressed data from working buffer to pages.
1254 * @buf: The decompressed data buffer
1255 * @buf_len: The decompressed data length
1256 * @decompressed: Number of bytes that are already decompressed inside the
1258 * @cb: The compressed extent descriptor
1259 * @orig_bio: The original bio that the caller wants to read for
1261 * An easier to understand graph is like below:
1263 * |<- orig_bio ->| |<- orig_bio->|
1264 * |<------- full decompressed extent ----->|
1265 * |<----------- @cb range ---->|
1266 * | |<-- @buf_len -->|
1267 * |<--- @decompressed --->|
1269 * Note that, @cb can be a subpage of the full decompressed extent, but
1270 * @cb->start always has the same as the orig_file_offset value of the full
1271 * decompressed extent.
1273 * When reading compressed extent, we have to read the full compressed extent,
1274 * while @orig_bio may only want part of the range.
1275 * Thus this function will ensure only data covered by @orig_bio will be copied
1278 * Return 0 if we have copied all needed contents for @orig_bio.
1279 * Return >0 if we need continue decompress.
1281 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1282 struct compressed_bio *cb, u32 decompressed)
1284 struct bio *orig_bio = cb->orig_bio;
1285 /* Offset inside the full decompressed extent */
1288 cur_offset = decompressed;
1289 /* The main loop to do the copy */
1290 while (cur_offset < decompressed + buf_len) {
1291 struct bio_vec bvec;
1294 /* Offset inside the full decompressed extent */
1297 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1299 * cb->start may underflow, but subtracting that value can still
1300 * give us correct offset inside the full decompressed extent.
1302 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1304 /* Haven't reached the bvec range, exit */
1305 if (decompressed + buf_len <= bvec_offset)
1308 copy_start = max(cur_offset, bvec_offset);
1309 copy_len = min(bvec_offset + bvec.bv_len,
1310 decompressed + buf_len) - copy_start;
1314 * Extra range check to ensure we didn't go beyond
1317 ASSERT(copy_start - decompressed < buf_len);
1318 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1319 buf + copy_start - decompressed, copy_len);
1320 cur_offset += copy_len;
1322 bio_advance(orig_bio, copy_len);
1323 /* Finished the bio */
1324 if (!orig_bio->bi_iter.bi_size)
1331 * Shannon Entropy calculation
1333 * Pure byte distribution analysis fails to determine compressibility of data.
1334 * Try calculating entropy to estimate the average minimum number of bits
1335 * needed to encode the sampled data.
1337 * For convenience, return the percentage of needed bits, instead of amount of
1340 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1341 * and can be compressible with high probability
1343 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1345 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1347 #define ENTROPY_LVL_ACEPTABLE (65)
1348 #define ENTROPY_LVL_HIGH (80)
1351 * For increasead precision in shannon_entropy calculation,
1352 * let's do pow(n, M) to save more digits after comma:
1354 * - maximum int bit length is 64
1355 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1356 * - 13 * 4 = 52 < 64 -> M = 4
1360 static inline u32 ilog2_w(u64 n)
1362 return ilog2(n * n * n * n);
1365 static u32 shannon_entropy(struct heuristic_ws *ws)
1367 const u32 entropy_max = 8 * ilog2_w(2);
1368 u32 entropy_sum = 0;
1369 u32 p, p_base, sz_base;
1372 sz_base = ilog2_w(ws->sample_size);
1373 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1374 p = ws->bucket[i].count;
1375 p_base = ilog2_w(p);
1376 entropy_sum += p * (sz_base - p_base);
1379 entropy_sum /= ws->sample_size;
1380 return entropy_sum * 100 / entropy_max;
1383 #define RADIX_BASE 4U
1384 #define COUNTERS_SIZE (1U << RADIX_BASE)
1386 static u8 get4bits(u64 num, int shift) {
1391 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1396 * Use 4 bits as radix base
1397 * Use 16 u32 counters for calculating new position in buf array
1399 * @array - array that will be sorted
1400 * @array_buf - buffer array to store sorting results
1401 * must be equal in size to @array
1404 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1409 u32 counters[COUNTERS_SIZE];
1417 * Try avoid useless loop iterations for small numbers stored in big
1418 * counters. Example: 48 33 4 ... in 64bit array
1420 max_num = array[0].count;
1421 for (i = 1; i < num; i++) {
1422 buf_num = array[i].count;
1423 if (buf_num > max_num)
1427 buf_num = ilog2(max_num);
1428 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1431 while (shift < bitlen) {
1432 memset(counters, 0, sizeof(counters));
1434 for (i = 0; i < num; i++) {
1435 buf_num = array[i].count;
1436 addr = get4bits(buf_num, shift);
1440 for (i = 1; i < COUNTERS_SIZE; i++)
1441 counters[i] += counters[i - 1];
1443 for (i = num - 1; i >= 0; i--) {
1444 buf_num = array[i].count;
1445 addr = get4bits(buf_num, shift);
1447 new_addr = counters[addr];
1448 array_buf[new_addr] = array[i];
1451 shift += RADIX_BASE;
1454 * Normal radix expects to move data from a temporary array, to
1455 * the main one. But that requires some CPU time. Avoid that
1456 * by doing another sort iteration to original array instead of
1459 memset(counters, 0, sizeof(counters));
1461 for (i = 0; i < num; i ++) {
1462 buf_num = array_buf[i].count;
1463 addr = get4bits(buf_num, shift);
1467 for (i = 1; i < COUNTERS_SIZE; i++)
1468 counters[i] += counters[i - 1];
1470 for (i = num - 1; i >= 0; i--) {
1471 buf_num = array_buf[i].count;
1472 addr = get4bits(buf_num, shift);
1474 new_addr = counters[addr];
1475 array[new_addr] = array_buf[i];
1478 shift += RADIX_BASE;
1483 * Size of the core byte set - how many bytes cover 90% of the sample
1485 * There are several types of structured binary data that use nearly all byte
1486 * values. The distribution can be uniform and counts in all buckets will be
1487 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1489 * Other possibility is normal (Gaussian) distribution, where the data could
1490 * be potentially compressible, but we have to take a few more steps to decide
1493 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1494 * compression algo can easy fix that
1495 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1496 * probability is not compressible
1498 #define BYTE_CORE_SET_LOW (64)
1499 #define BYTE_CORE_SET_HIGH (200)
1501 static int byte_core_set_size(struct heuristic_ws *ws)
1504 u32 coreset_sum = 0;
1505 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1506 struct bucket_item *bucket = ws->bucket;
1508 /* Sort in reverse order */
1509 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1511 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1512 coreset_sum += bucket[i].count;
1514 if (coreset_sum > core_set_threshold)
1517 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1518 coreset_sum += bucket[i].count;
1519 if (coreset_sum > core_set_threshold)
1527 * Count byte values in buckets.
1528 * This heuristic can detect textual data (configs, xml, json, html, etc).
1529 * Because in most text-like data byte set is restricted to limited number of
1530 * possible characters, and that restriction in most cases makes data easy to
1533 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1534 * less - compressible
1535 * more - need additional analysis
1537 #define BYTE_SET_THRESHOLD (64)
1539 static u32 byte_set_size(const struct heuristic_ws *ws)
1542 u32 byte_set_size = 0;
1544 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1545 if (ws->bucket[i].count > 0)
1550 * Continue collecting count of byte values in buckets. If the byte
1551 * set size is bigger then the threshold, it's pointless to continue,
1552 * the detection technique would fail for this type of data.
1554 for (; i < BUCKET_SIZE; i++) {
1555 if (ws->bucket[i].count > 0) {
1557 if (byte_set_size > BYTE_SET_THRESHOLD)
1558 return byte_set_size;
1562 return byte_set_size;
1565 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1567 const u32 half_of_sample = ws->sample_size / 2;
1568 const u8 *data = ws->sample;
1570 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1573 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1574 struct heuristic_ws *ws)
1577 u64 index, index_end;
1578 u32 i, curr_sample_pos;
1582 * Compression handles the input data by chunks of 128KiB
1583 * (defined by BTRFS_MAX_UNCOMPRESSED)
1585 * We do the same for the heuristic and loop over the whole range.
1587 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1588 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1590 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1591 end = start + BTRFS_MAX_UNCOMPRESSED;
1593 index = start >> PAGE_SHIFT;
1594 index_end = end >> PAGE_SHIFT;
1596 /* Don't miss unaligned end */
1597 if (!IS_ALIGNED(end, PAGE_SIZE))
1600 curr_sample_pos = 0;
1601 while (index < index_end) {
1602 page = find_get_page(inode->i_mapping, index);
1603 in_data = kmap_local_page(page);
1604 /* Handle case where the start is not aligned to PAGE_SIZE */
1605 i = start % PAGE_SIZE;
1606 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1607 /* Don't sample any garbage from the last page */
1608 if (start > end - SAMPLING_READ_SIZE)
1610 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1611 SAMPLING_READ_SIZE);
1612 i += SAMPLING_INTERVAL;
1613 start += SAMPLING_INTERVAL;
1614 curr_sample_pos += SAMPLING_READ_SIZE;
1616 kunmap_local(in_data);
1622 ws->sample_size = curr_sample_pos;
1626 * Compression heuristic.
1628 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1629 * quickly (compared to direct compression) detect data characteristics
1630 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1633 * The following types of analysis can be performed:
1634 * - detect mostly zero data
1635 * - detect data with low "byte set" size (text, etc)
1636 * - detect data with low/high "core byte" set
1638 * Return non-zero if the compression should be done, 0 otherwise.
1640 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1642 struct list_head *ws_list = get_workspace(0, 0);
1643 struct heuristic_ws *ws;
1648 ws = list_entry(ws_list, struct heuristic_ws, list);
1650 heuristic_collect_sample(inode, start, end, ws);
1652 if (sample_repeated_patterns(ws)) {
1657 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1659 for (i = 0; i < ws->sample_size; i++) {
1660 byte = ws->sample[i];
1661 ws->bucket[byte].count++;
1664 i = byte_set_size(ws);
1665 if (i < BYTE_SET_THRESHOLD) {
1670 i = byte_core_set_size(ws);
1671 if (i <= BYTE_CORE_SET_LOW) {
1676 if (i >= BYTE_CORE_SET_HIGH) {
1681 i = shannon_entropy(ws);
1682 if (i <= ENTROPY_LVL_ACEPTABLE) {
1688 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1689 * needed to give green light to compression.
1691 * For now just assume that compression at that level is not worth the
1692 * resources because:
1694 * 1. it is possible to defrag the data later
1696 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1697 * values, every bucket has counter at level ~54. The heuristic would
1698 * be confused. This can happen when data have some internal repeated
1699 * patterns like "abbacbbc...". This can be detected by analyzing
1700 * pairs of bytes, which is too costly.
1702 if (i < ENTROPY_LVL_HIGH) {
1711 put_workspace(0, ws_list);
1716 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1717 * level, unrecognized string will set the default level
1719 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1721 unsigned int level = 0;
1727 if (str[0] == ':') {
1728 ret = kstrtouint(str + 1, 10, &level);
1733 level = btrfs_compress_set_level(type, level);