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/slab.h>
20 #include <linux/sched/mm.h>
21 #include <linux/log2.h>
22 #include <crypto/hash.h>
26 #include "transaction.h"
27 #include "btrfs_inode.h"
29 #include "ordered-data.h"
30 #include "compression.h"
31 #include "extent_io.h"
32 #include "extent_map.h"
36 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
38 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
41 case BTRFS_COMPRESS_ZLIB:
42 case BTRFS_COMPRESS_LZO:
43 case BTRFS_COMPRESS_ZSTD:
44 case BTRFS_COMPRESS_NONE:
45 return btrfs_compress_types[type];
53 bool btrfs_compress_is_valid_type(const char *str, size_t len)
57 for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
58 size_t comp_len = strlen(btrfs_compress_types[i]);
63 if (!strncmp(btrfs_compress_types[i], str, comp_len))
69 static int compression_compress_pages(int type, struct list_head *ws,
70 struct address_space *mapping, u64 start, struct page **pages,
71 unsigned long *out_pages, unsigned long *total_in,
72 unsigned long *total_out)
75 case BTRFS_COMPRESS_ZLIB:
76 return zlib_compress_pages(ws, mapping, start, pages,
77 out_pages, total_in, total_out);
78 case BTRFS_COMPRESS_LZO:
79 return lzo_compress_pages(ws, mapping, start, pages,
80 out_pages, total_in, total_out);
81 case BTRFS_COMPRESS_ZSTD:
82 return zstd_compress_pages(ws, mapping, start, pages,
83 out_pages, total_in, total_out);
84 case BTRFS_COMPRESS_NONE:
87 * This can happen when compression races with remount setting
88 * it to 'no compress', while caller doesn't call
89 * inode_need_compress() to check if we really need to
92 * Not a big deal, just need to inform caller that we
93 * haven't allocated any pages yet.
100 static int compression_decompress_bio(struct list_head *ws,
101 struct compressed_bio *cb)
103 switch (cb->compress_type) {
104 case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
105 case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb);
106 case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
107 case BTRFS_COMPRESS_NONE:
110 * This can't happen, the type is validated several times
111 * before we get here.
117 static int compression_decompress(int type, struct list_head *ws,
118 unsigned char *data_in, struct page *dest_page,
119 unsigned long start_byte, size_t srclen, size_t destlen)
122 case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
123 start_byte, srclen, destlen);
124 case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page,
125 start_byte, srclen, destlen);
126 case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
127 start_byte, srclen, destlen);
128 case BTRFS_COMPRESS_NONE:
131 * This can't happen, the type is validated several times
132 * before we get here.
138 static int btrfs_decompress_bio(struct compressed_bio *cb);
140 static void finish_compressed_bio_read(struct compressed_bio *cb)
145 if (cb->status == BLK_STS_OK)
146 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
148 /* Release the compressed pages */
149 for (index = 0; index < cb->nr_pages; index++) {
150 page = cb->compressed_pages[index];
151 page->mapping = NULL;
155 /* Do io completion on the original bio */
156 if (cb->status != BLK_STS_OK)
157 cb->orig_bio->bi_status = cb->status;
158 bio_endio(cb->orig_bio);
160 /* Finally free the cb struct */
161 kfree(cb->compressed_pages);
166 * Verify the checksums and kick off repair if needed on the uncompressed data
167 * before decompressing it into the original bio and freeing the uncompressed
170 static void end_compressed_bio_read(struct bio *bio)
172 struct compressed_bio *cb = bio->bi_private;
173 struct inode *inode = cb->inode;
174 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
175 struct btrfs_inode *bi = BTRFS_I(inode);
176 bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) &&
177 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state);
178 blk_status_t status = bio->bi_status;
179 struct btrfs_bio *bbio = btrfs_bio(bio);
180 struct bvec_iter iter;
184 btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) {
185 u64 start = bbio->file_offset + offset;
188 (!csum || !btrfs_check_data_csum(inode, bbio, offset,
189 bv.bv_page, bv.bv_offset))) {
190 clean_io_failure(fs_info, &bi->io_failure_tree,
191 &bi->io_tree, start, bv.bv_page,
192 btrfs_ino(bi), bv.bv_offset);
196 refcount_inc(&cb->pending_ios);
197 ret = btrfs_repair_one_sector(inode, bbio, offset,
198 bv.bv_page, bv.bv_offset,
199 btrfs_submit_data_read_bio);
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 bio *bio)
305 struct compressed_bio *cb = bio->bi_private;
308 cb->status = 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, 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,
337 blk_opf_t opf, bio_end_io_t endio_func,
338 u64 *next_stripe_start)
340 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
341 struct btrfs_io_geometry geom;
342 struct extent_map *em;
346 bio = btrfs_bio_alloc(BIO_MAX_VECS);
348 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
350 bio->bi_private = cb;
351 bio->bi_end_io = endio_func;
353 em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize);
359 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
360 bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev);
362 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom);
368 *next_stripe_start = disk_bytenr + geom.len;
369 refcount_inc(&cb->pending_ios);
374 * worker function to build and submit bios for previously compressed pages.
375 * The corresponding pages in the inode should be marked for writeback
376 * and the compressed pages should have a reference on them for dropping
377 * when the IO is complete.
379 * This also checksums the file bytes and gets things ready for
382 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
383 unsigned int len, u64 disk_start,
384 unsigned int compressed_len,
385 struct page **compressed_pages,
386 unsigned int nr_pages,
387 blk_opf_t write_flags,
388 struct cgroup_subsys_state *blkcg_css,
391 struct btrfs_fs_info *fs_info = inode->root->fs_info;
392 struct bio *bio = NULL;
393 struct compressed_bio *cb;
394 u64 cur_disk_bytenr = disk_start;
395 u64 next_stripe_start;
396 blk_status_t ret = BLK_STS_OK;
397 int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
398 const bool use_append = btrfs_use_zone_append(inode, disk_start);
399 const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
401 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
402 IS_ALIGNED(len, fs_info->sectorsize));
403 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
405 return BLK_STS_RESOURCE;
406 refcount_set(&cb->pending_ios, 1);
407 cb->status = BLK_STS_OK;
408 cb->inode = &inode->vfs_inode;
411 cb->compressed_pages = compressed_pages;
412 cb->compressed_len = compressed_len;
413 cb->writeback = writeback;
414 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
415 cb->nr_pages = nr_pages;
418 kthread_associate_blkcg(blkcg_css);
420 while (cur_disk_bytenr < disk_start + compressed_len) {
421 u64 offset = cur_disk_bytenr - disk_start;
422 unsigned int index = offset >> PAGE_SHIFT;
423 unsigned int real_size;
425 struct page *page = compressed_pages[index];
428 /* Allocate new bio if submitted or not yet allocated */
430 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
431 bio_op | write_flags, end_compressed_bio_write,
434 ret = errno_to_blk_status(PTR_ERR(bio));
438 bio->bi_opf |= REQ_CGROUP_PUNT;
441 * We should never reach next_stripe_start start as we will
442 * submit comp_bio when reach the boundary immediately.
444 ASSERT(cur_disk_bytenr != next_stripe_start);
447 * We have various limits on the real read size:
450 * - compressed length boundary
452 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr);
453 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
454 real_size = min_t(u64, real_size, compressed_len - offset);
455 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
458 added = bio_add_zone_append_page(bio, page, real_size,
459 offset_in_page(offset));
461 added = bio_add_page(bio, page, real_size,
462 offset_in_page(offset));
463 /* Reached zoned boundary */
467 cur_disk_bytenr += added;
468 /* Reached stripe boundary */
469 if (cur_disk_bytenr == next_stripe_start)
472 /* Finished the range */
473 if (cur_disk_bytenr == disk_start + compressed_len)
478 ret = btrfs_csum_one_bio(inode, bio, start, true);
480 bio->bi_status = ret;
486 ASSERT(bio->bi_iter.bi_size);
487 btrfs_submit_bio(fs_info, bio, 0);
494 kthread_associate_blkcg(NULL);
496 if (refcount_dec_and_test(&cb->pending_ios))
497 finish_compressed_bio_write(cb);
501 static u64 bio_end_offset(struct bio *bio)
503 struct bio_vec *last = bio_last_bvec_all(bio);
505 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
509 * Add extra pages in the same compressed file extent so that we don't need to
510 * re-read the same extent again and again.
512 * NOTE: this won't work well for subpage, as for subpage read, we lock the
513 * full page then submit bio for each compressed/regular extents.
515 * This means, if we have several sectors in the same page points to the same
516 * on-disk compressed data, we will re-read the same extent many times and
517 * this function can only help for the next page.
519 static noinline int add_ra_bio_pages(struct inode *inode,
521 struct compressed_bio *cb)
523 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
524 unsigned long end_index;
525 u64 cur = bio_end_offset(cb->orig_bio);
526 u64 isize = i_size_read(inode);
529 struct extent_map *em;
530 struct address_space *mapping = inode->i_mapping;
531 struct extent_map_tree *em_tree;
532 struct extent_io_tree *tree;
533 int sectors_missed = 0;
535 em_tree = &BTRFS_I(inode)->extent_tree;
536 tree = &BTRFS_I(inode)->io_tree;
542 * For current subpage support, we only support 64K page size,
543 * which means maximum compressed extent size (128K) is just 2x page
545 * This makes readahead less effective, so here disable readahead for
546 * subpage for now, until full compressed write is supported.
548 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
551 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
553 while (cur < compressed_end) {
555 u64 pg_index = cur >> PAGE_SHIFT;
558 if (pg_index > end_index)
561 page = xa_load(&mapping->i_pages, pg_index);
562 if (page && !xa_is_value(page)) {
563 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
564 fs_info->sectorsize_bits;
566 /* Beyond threshold, no need to continue */
567 if (sectors_missed > 4)
571 * Jump to next page start as we already have page for
574 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
578 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
583 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
585 /* There is already a page, skip to page end */
586 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
590 ret = set_page_extent_mapped(page);
597 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
598 lock_extent(tree, cur, page_end);
599 read_lock(&em_tree->lock);
600 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
601 read_unlock(&em_tree->lock);
604 * At this point, we have a locked page in the page cache for
605 * these bytes in the file. But, we have to make sure they map
606 * to this compressed extent on disk.
608 if (!em || cur < em->start ||
609 (cur + fs_info->sectorsize > extent_map_end(em)) ||
610 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
612 unlock_extent(tree, cur, page_end);
619 if (page->index == end_index) {
620 size_t zero_offset = offset_in_page(isize);
624 zeros = PAGE_SIZE - zero_offset;
625 memzero_page(page, zero_offset, zeros);
629 add_size = min(em->start + em->len, page_end + 1) - cur;
630 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
631 if (ret != add_size) {
632 unlock_extent(tree, cur, page_end);
638 * If it's subpage, we also need to increase its
639 * subpage::readers number, as at endio we will decrease
640 * subpage::readers and to unlock the page.
642 if (fs_info->sectorsize < PAGE_SIZE)
643 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
651 * for a compressed read, the bio we get passed has all the inode pages
652 * in it. We don't actually do IO on those pages but allocate new ones
653 * to hold the compressed pages on disk.
655 * bio->bi_iter.bi_sector points to the compressed extent on disk
656 * bio->bi_io_vec points to all of the inode pages
658 * After the compressed pages are read, we copy the bytes into the
659 * bio we were passed and then call the bio end_io calls
661 void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
664 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
665 struct extent_map_tree *em_tree;
666 struct compressed_bio *cb;
667 unsigned int compressed_len;
668 struct bio *comp_bio = NULL;
669 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
670 u64 cur_disk_byte = disk_bytenr;
671 u64 next_stripe_start;
675 struct extent_map *em;
680 em_tree = &BTRFS_I(inode)->extent_tree;
682 file_offset = bio_first_bvec_all(bio)->bv_offset +
683 page_offset(bio_first_page_all(bio));
685 /* we need the actual starting offset of this extent in the file */
686 read_lock(&em_tree->lock);
687 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
688 read_unlock(&em_tree->lock);
694 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
695 compressed_len = em->block_len;
696 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
698 ret = BLK_STS_RESOURCE;
702 refcount_set(&cb->pending_ios, 1);
703 cb->status = BLK_STS_OK;
706 cb->start = em->orig_start;
708 em_start = em->start;
710 cb->len = bio->bi_iter.bi_size;
711 cb->compressed_len = compressed_len;
712 cb->compress_type = em->compress_type;
718 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
719 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
720 if (!cb->compressed_pages) {
721 ret = BLK_STS_RESOURCE;
725 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
727 ret = BLK_STS_RESOURCE;
731 add_ra_bio_pages(inode, em_start + em_len, cb);
733 /* include any pages we added in add_ra-bio_pages */
734 cb->len = bio->bi_iter.bi_size;
736 while (cur_disk_byte < disk_bytenr + compressed_len) {
737 u64 offset = cur_disk_byte - disk_bytenr;
738 unsigned int index = offset >> PAGE_SHIFT;
739 unsigned int real_size;
741 struct page *page = cb->compressed_pages[index];
744 /* Allocate new bio if submitted or not yet allocated */
746 comp_bio = alloc_compressed_bio(cb, cur_disk_byte,
747 REQ_OP_READ, end_compressed_bio_read,
749 if (IS_ERR(comp_bio)) {
750 cb->status = errno_to_blk_status(PTR_ERR(comp_bio));
755 * We should never reach next_stripe_start start as we will
756 * submit comp_bio when reach the boundary immediately.
758 ASSERT(cur_disk_byte != next_stripe_start);
760 * We have various limit on the real read size:
763 * - compressed length boundary
765 real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte);
766 real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset));
767 real_size = min_t(u64, real_size, compressed_len - offset);
768 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
770 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
772 * Maximum compressed extent is smaller than bio size limit,
773 * thus bio_add_page() should always success.
775 ASSERT(added == real_size);
776 cur_disk_byte += added;
778 /* Reached stripe boundary, need to submit */
779 if (cur_disk_byte == next_stripe_start)
782 /* Has finished the range, need to submit */
783 if (cur_disk_byte == disk_bytenr + compressed_len)
787 /* Save the original iter for read repair */
788 if (bio_op(comp_bio) == REQ_OP_READ)
789 btrfs_bio(comp_bio)->iter = comp_bio->bi_iter;
792 * Save the initial offset of this chunk, as there
793 * is no direct correlation between compressed pages and
794 * the original file offset. The field is only used for
795 * priting error messages.
797 btrfs_bio(comp_bio)->file_offset = file_offset;
799 ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL);
801 comp_bio->bi_status = ret;
806 ASSERT(comp_bio->bi_iter.bi_size);
807 btrfs_submit_bio(fs_info, comp_bio, mirror_num);
812 if (refcount_dec_and_test(&cb->pending_ios))
813 finish_compressed_bio_read(cb);
817 if (cb->compressed_pages) {
818 for (i = 0; i < cb->nr_pages; i++) {
819 if (cb->compressed_pages[i])
820 __free_page(cb->compressed_pages[i]);
824 kfree(cb->compressed_pages);
828 bio->bi_status = ret;
834 * Heuristic uses systematic sampling to collect data from the input data
835 * range, the logic can be tuned by the following constants:
837 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
838 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
840 #define SAMPLING_READ_SIZE (16)
841 #define SAMPLING_INTERVAL (256)
844 * For statistical analysis of the input data we consider bytes that form a
845 * Galois Field of 256 objects. Each object has an attribute count, ie. how
846 * many times the object appeared in the sample.
848 #define BUCKET_SIZE (256)
851 * The size of the sample is based on a statistical sampling rule of thumb.
852 * The common way is to perform sampling tests as long as the number of
853 * elements in each cell is at least 5.
855 * Instead of 5, we choose 32 to obtain more accurate results.
856 * If the data contain the maximum number of symbols, which is 256, we obtain a
857 * sample size bound by 8192.
859 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
860 * from up to 512 locations.
862 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
863 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
869 struct heuristic_ws {
870 /* Partial copy of input data */
873 /* Buckets store counters for each byte value */
874 struct bucket_item *bucket;
876 struct bucket_item *bucket_b;
877 struct list_head list;
880 static struct workspace_manager heuristic_wsm;
882 static void free_heuristic_ws(struct list_head *ws)
884 struct heuristic_ws *workspace;
886 workspace = list_entry(ws, struct heuristic_ws, list);
888 kvfree(workspace->sample);
889 kfree(workspace->bucket);
890 kfree(workspace->bucket_b);
894 static struct list_head *alloc_heuristic_ws(unsigned int level)
896 struct heuristic_ws *ws;
898 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
900 return ERR_PTR(-ENOMEM);
902 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
906 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
910 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
914 INIT_LIST_HEAD(&ws->list);
917 free_heuristic_ws(&ws->list);
918 return ERR_PTR(-ENOMEM);
921 const struct btrfs_compress_op btrfs_heuristic_compress = {
922 .workspace_manager = &heuristic_wsm,
925 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
926 /* The heuristic is represented as compression type 0 */
927 &btrfs_heuristic_compress,
928 &btrfs_zlib_compress,
930 &btrfs_zstd_compress,
933 static struct list_head *alloc_workspace(int type, unsigned int level)
936 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
937 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
938 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
939 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
942 * This can't happen, the type is validated several times
943 * before we get here.
949 static void free_workspace(int type, struct list_head *ws)
952 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
953 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
954 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
955 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
958 * This can't happen, the type is validated several times
959 * before we get here.
965 static void btrfs_init_workspace_manager(int type)
967 struct workspace_manager *wsm;
968 struct list_head *workspace;
970 wsm = btrfs_compress_op[type]->workspace_manager;
971 INIT_LIST_HEAD(&wsm->idle_ws);
972 spin_lock_init(&wsm->ws_lock);
973 atomic_set(&wsm->total_ws, 0);
974 init_waitqueue_head(&wsm->ws_wait);
977 * Preallocate one workspace for each compression type so we can
978 * guarantee forward progress in the worst case
980 workspace = alloc_workspace(type, 0);
981 if (IS_ERR(workspace)) {
983 "BTRFS: cannot preallocate compression workspace, will try later\n");
985 atomic_set(&wsm->total_ws, 1);
987 list_add(workspace, &wsm->idle_ws);
991 static void btrfs_cleanup_workspace_manager(int type)
993 struct workspace_manager *wsman;
994 struct list_head *ws;
996 wsman = btrfs_compress_op[type]->workspace_manager;
997 while (!list_empty(&wsman->idle_ws)) {
998 ws = wsman->idle_ws.next;
1000 free_workspace(type, ws);
1001 atomic_dec(&wsman->total_ws);
1006 * This finds an available workspace or allocates a new one.
1007 * If it's not possible to allocate a new one, waits until there's one.
1008 * Preallocation makes a forward progress guarantees and we do not return
1011 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1013 struct workspace_manager *wsm;
1014 struct list_head *workspace;
1015 int cpus = num_online_cpus();
1017 struct list_head *idle_ws;
1018 spinlock_t *ws_lock;
1020 wait_queue_head_t *ws_wait;
1023 wsm = btrfs_compress_op[type]->workspace_manager;
1024 idle_ws = &wsm->idle_ws;
1025 ws_lock = &wsm->ws_lock;
1026 total_ws = &wsm->total_ws;
1027 ws_wait = &wsm->ws_wait;
1028 free_ws = &wsm->free_ws;
1032 if (!list_empty(idle_ws)) {
1033 workspace = idle_ws->next;
1034 list_del(workspace);
1036 spin_unlock(ws_lock);
1040 if (atomic_read(total_ws) > cpus) {
1043 spin_unlock(ws_lock);
1044 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1045 if (atomic_read(total_ws) > cpus && !*free_ws)
1047 finish_wait(ws_wait, &wait);
1050 atomic_inc(total_ws);
1051 spin_unlock(ws_lock);
1054 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1055 * to turn it off here because we might get called from the restricted
1056 * context of btrfs_compress_bio/btrfs_compress_pages
1058 nofs_flag = memalloc_nofs_save();
1059 workspace = alloc_workspace(type, level);
1060 memalloc_nofs_restore(nofs_flag);
1062 if (IS_ERR(workspace)) {
1063 atomic_dec(total_ws);
1067 * Do not return the error but go back to waiting. There's a
1068 * workspace preallocated for each type and the compression
1069 * time is bounded so we get to a workspace eventually. This
1070 * makes our caller's life easier.
1072 * To prevent silent and low-probability deadlocks (when the
1073 * initial preallocation fails), check if there are any
1074 * workspaces at all.
1076 if (atomic_read(total_ws) == 0) {
1077 static DEFINE_RATELIMIT_STATE(_rs,
1078 /* once per minute */ 60 * HZ,
1081 if (__ratelimit(&_rs)) {
1082 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1090 static struct list_head *get_workspace(int type, int level)
1093 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1094 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1095 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
1096 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1099 * This can't happen, the type is validated several times
1100 * before we get here.
1107 * put a workspace struct back on the list or free it if we have enough
1108 * idle ones sitting around
1110 void btrfs_put_workspace(int type, struct list_head *ws)
1112 struct workspace_manager *wsm;
1113 struct list_head *idle_ws;
1114 spinlock_t *ws_lock;
1116 wait_queue_head_t *ws_wait;
1119 wsm = btrfs_compress_op[type]->workspace_manager;
1120 idle_ws = &wsm->idle_ws;
1121 ws_lock = &wsm->ws_lock;
1122 total_ws = &wsm->total_ws;
1123 ws_wait = &wsm->ws_wait;
1124 free_ws = &wsm->free_ws;
1127 if (*free_ws <= num_online_cpus()) {
1128 list_add(ws, idle_ws);
1130 spin_unlock(ws_lock);
1133 spin_unlock(ws_lock);
1135 free_workspace(type, ws);
1136 atomic_dec(total_ws);
1138 cond_wake_up(ws_wait);
1141 static void put_workspace(int type, struct list_head *ws)
1144 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1145 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1146 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1147 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1150 * This can't happen, the type is validated several times
1151 * before we get here.
1158 * Adjust @level according to the limits of the compression algorithm or
1159 * fallback to default
1161 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1163 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1166 level = ops->default_level;
1168 level = min(level, ops->max_level);
1174 * Given an address space and start and length, compress the bytes into @pages
1175 * that are allocated on demand.
1177 * @type_level is encoded algorithm and level, where level 0 means whatever
1178 * default the algorithm chooses and is opaque here;
1179 * - compression algo are 0-3
1180 * - the level are bits 4-7
1182 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1183 * and returns number of actually allocated pages
1185 * @total_in is used to return the number of bytes actually read. It
1186 * may be smaller than the input length if we had to exit early because we
1187 * ran out of room in the pages array or because we cross the
1188 * max_out threshold.
1190 * @total_out is an in/out parameter, must be set to the input length and will
1191 * be also used to return the total number of compressed bytes
1193 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1194 u64 start, struct page **pages,
1195 unsigned long *out_pages,
1196 unsigned long *total_in,
1197 unsigned long *total_out)
1199 int type = btrfs_compress_type(type_level);
1200 int level = btrfs_compress_level(type_level);
1201 struct list_head *workspace;
1204 level = btrfs_compress_set_level(type, level);
1205 workspace = get_workspace(type, level);
1206 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1207 out_pages, total_in, total_out);
1208 put_workspace(type, workspace);
1212 static int btrfs_decompress_bio(struct compressed_bio *cb)
1214 struct list_head *workspace;
1216 int type = cb->compress_type;
1218 workspace = get_workspace(type, 0);
1219 ret = compression_decompress_bio(workspace, cb);
1220 put_workspace(type, workspace);
1226 * a less complex decompression routine. Our compressed data fits in a
1227 * single page, and we want to read a single page out of it.
1228 * start_byte tells us the offset into the compressed data we're interested in
1230 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1231 unsigned long start_byte, size_t srclen, size_t destlen)
1233 struct list_head *workspace;
1236 workspace = get_workspace(type, 0);
1237 ret = compression_decompress(type, workspace, data_in, dest_page,
1238 start_byte, srclen, destlen);
1239 put_workspace(type, workspace);
1244 void __init btrfs_init_compress(void)
1246 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1247 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1248 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1249 zstd_init_workspace_manager();
1252 void __cold btrfs_exit_compress(void)
1254 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1255 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1256 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1257 zstd_cleanup_workspace_manager();
1261 * Copy decompressed data from working buffer to pages.
1263 * @buf: The decompressed data buffer
1264 * @buf_len: The decompressed data length
1265 * @decompressed: Number of bytes that are already decompressed inside the
1267 * @cb: The compressed extent descriptor
1268 * @orig_bio: The original bio that the caller wants to read for
1270 * An easier to understand graph is like below:
1272 * |<- orig_bio ->| |<- orig_bio->|
1273 * |<------- full decompressed extent ----->|
1274 * |<----------- @cb range ---->|
1275 * | |<-- @buf_len -->|
1276 * |<--- @decompressed --->|
1278 * Note that, @cb can be a subpage of the full decompressed extent, but
1279 * @cb->start always has the same as the orig_file_offset value of the full
1280 * decompressed extent.
1282 * When reading compressed extent, we have to read the full compressed extent,
1283 * while @orig_bio may only want part of the range.
1284 * Thus this function will ensure only data covered by @orig_bio will be copied
1287 * Return 0 if we have copied all needed contents for @orig_bio.
1288 * Return >0 if we need continue decompress.
1290 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1291 struct compressed_bio *cb, u32 decompressed)
1293 struct bio *orig_bio = cb->orig_bio;
1294 /* Offset inside the full decompressed extent */
1297 cur_offset = decompressed;
1298 /* The main loop to do the copy */
1299 while (cur_offset < decompressed + buf_len) {
1300 struct bio_vec bvec;
1303 /* Offset inside the full decompressed extent */
1306 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1308 * cb->start may underflow, but subtracting that value can still
1309 * give us correct offset inside the full decompressed extent.
1311 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1313 /* Haven't reached the bvec range, exit */
1314 if (decompressed + buf_len <= bvec_offset)
1317 copy_start = max(cur_offset, bvec_offset);
1318 copy_len = min(bvec_offset + bvec.bv_len,
1319 decompressed + buf_len) - copy_start;
1323 * Extra range check to ensure we didn't go beyond
1326 ASSERT(copy_start - decompressed < buf_len);
1327 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1328 buf + copy_start - decompressed, copy_len);
1329 cur_offset += copy_len;
1331 bio_advance(orig_bio, copy_len);
1332 /* Finished the bio */
1333 if (!orig_bio->bi_iter.bi_size)
1340 * Shannon Entropy calculation
1342 * Pure byte distribution analysis fails to determine compressibility of data.
1343 * Try calculating entropy to estimate the average minimum number of bits
1344 * needed to encode the sampled data.
1346 * For convenience, return the percentage of needed bits, instead of amount of
1349 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1350 * and can be compressible with high probability
1352 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1354 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1356 #define ENTROPY_LVL_ACEPTABLE (65)
1357 #define ENTROPY_LVL_HIGH (80)
1360 * For increasead precision in shannon_entropy calculation,
1361 * let's do pow(n, M) to save more digits after comma:
1363 * - maximum int bit length is 64
1364 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1365 * - 13 * 4 = 52 < 64 -> M = 4
1369 static inline u32 ilog2_w(u64 n)
1371 return ilog2(n * n * n * n);
1374 static u32 shannon_entropy(struct heuristic_ws *ws)
1376 const u32 entropy_max = 8 * ilog2_w(2);
1377 u32 entropy_sum = 0;
1378 u32 p, p_base, sz_base;
1381 sz_base = ilog2_w(ws->sample_size);
1382 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1383 p = ws->bucket[i].count;
1384 p_base = ilog2_w(p);
1385 entropy_sum += p * (sz_base - p_base);
1388 entropy_sum /= ws->sample_size;
1389 return entropy_sum * 100 / entropy_max;
1392 #define RADIX_BASE 4U
1393 #define COUNTERS_SIZE (1U << RADIX_BASE)
1395 static u8 get4bits(u64 num, int shift) {
1400 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1405 * Use 4 bits as radix base
1406 * Use 16 u32 counters for calculating new position in buf array
1408 * @array - array that will be sorted
1409 * @array_buf - buffer array to store sorting results
1410 * must be equal in size to @array
1413 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1418 u32 counters[COUNTERS_SIZE];
1426 * Try avoid useless loop iterations for small numbers stored in big
1427 * counters. Example: 48 33 4 ... in 64bit array
1429 max_num = array[0].count;
1430 for (i = 1; i < num; i++) {
1431 buf_num = array[i].count;
1432 if (buf_num > max_num)
1436 buf_num = ilog2(max_num);
1437 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1440 while (shift < bitlen) {
1441 memset(counters, 0, sizeof(counters));
1443 for (i = 0; i < num; i++) {
1444 buf_num = array[i].count;
1445 addr = get4bits(buf_num, shift);
1449 for (i = 1; i < COUNTERS_SIZE; i++)
1450 counters[i] += counters[i - 1];
1452 for (i = num - 1; i >= 0; i--) {
1453 buf_num = array[i].count;
1454 addr = get4bits(buf_num, shift);
1456 new_addr = counters[addr];
1457 array_buf[new_addr] = array[i];
1460 shift += RADIX_BASE;
1463 * Normal radix expects to move data from a temporary array, to
1464 * the main one. But that requires some CPU time. Avoid that
1465 * by doing another sort iteration to original array instead of
1468 memset(counters, 0, sizeof(counters));
1470 for (i = 0; i < num; i ++) {
1471 buf_num = array_buf[i].count;
1472 addr = get4bits(buf_num, shift);
1476 for (i = 1; i < COUNTERS_SIZE; i++)
1477 counters[i] += counters[i - 1];
1479 for (i = num - 1; i >= 0; i--) {
1480 buf_num = array_buf[i].count;
1481 addr = get4bits(buf_num, shift);
1483 new_addr = counters[addr];
1484 array[new_addr] = array_buf[i];
1487 shift += RADIX_BASE;
1492 * Size of the core byte set - how many bytes cover 90% of the sample
1494 * There are several types of structured binary data that use nearly all byte
1495 * values. The distribution can be uniform and counts in all buckets will be
1496 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1498 * Other possibility is normal (Gaussian) distribution, where the data could
1499 * be potentially compressible, but we have to take a few more steps to decide
1502 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1503 * compression algo can easy fix that
1504 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1505 * probability is not compressible
1507 #define BYTE_CORE_SET_LOW (64)
1508 #define BYTE_CORE_SET_HIGH (200)
1510 static int byte_core_set_size(struct heuristic_ws *ws)
1513 u32 coreset_sum = 0;
1514 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1515 struct bucket_item *bucket = ws->bucket;
1517 /* Sort in reverse order */
1518 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1520 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1521 coreset_sum += bucket[i].count;
1523 if (coreset_sum > core_set_threshold)
1526 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1527 coreset_sum += bucket[i].count;
1528 if (coreset_sum > core_set_threshold)
1536 * Count byte values in buckets.
1537 * This heuristic can detect textual data (configs, xml, json, html, etc).
1538 * Because in most text-like data byte set is restricted to limited number of
1539 * possible characters, and that restriction in most cases makes data easy to
1542 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1543 * less - compressible
1544 * more - need additional analysis
1546 #define BYTE_SET_THRESHOLD (64)
1548 static u32 byte_set_size(const struct heuristic_ws *ws)
1551 u32 byte_set_size = 0;
1553 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1554 if (ws->bucket[i].count > 0)
1559 * Continue collecting count of byte values in buckets. If the byte
1560 * set size is bigger then the threshold, it's pointless to continue,
1561 * the detection technique would fail for this type of data.
1563 for (; i < BUCKET_SIZE; i++) {
1564 if (ws->bucket[i].count > 0) {
1566 if (byte_set_size > BYTE_SET_THRESHOLD)
1567 return byte_set_size;
1571 return byte_set_size;
1574 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1576 const u32 half_of_sample = ws->sample_size / 2;
1577 const u8 *data = ws->sample;
1579 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1582 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1583 struct heuristic_ws *ws)
1586 u64 index, index_end;
1587 u32 i, curr_sample_pos;
1591 * Compression handles the input data by chunks of 128KiB
1592 * (defined by BTRFS_MAX_UNCOMPRESSED)
1594 * We do the same for the heuristic and loop over the whole range.
1596 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1597 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1599 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1600 end = start + BTRFS_MAX_UNCOMPRESSED;
1602 index = start >> PAGE_SHIFT;
1603 index_end = end >> PAGE_SHIFT;
1605 /* Don't miss unaligned end */
1606 if (!IS_ALIGNED(end, PAGE_SIZE))
1609 curr_sample_pos = 0;
1610 while (index < index_end) {
1611 page = find_get_page(inode->i_mapping, index);
1612 in_data = kmap_local_page(page);
1613 /* Handle case where the start is not aligned to PAGE_SIZE */
1614 i = start % PAGE_SIZE;
1615 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1616 /* Don't sample any garbage from the last page */
1617 if (start > end - SAMPLING_READ_SIZE)
1619 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1620 SAMPLING_READ_SIZE);
1621 i += SAMPLING_INTERVAL;
1622 start += SAMPLING_INTERVAL;
1623 curr_sample_pos += SAMPLING_READ_SIZE;
1625 kunmap_local(in_data);
1631 ws->sample_size = curr_sample_pos;
1635 * Compression heuristic.
1637 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1638 * quickly (compared to direct compression) detect data characteristics
1639 * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1642 * The following types of analysis can be performed:
1643 * - detect mostly zero data
1644 * - detect data with low "byte set" size (text, etc)
1645 * - detect data with low/high "core byte" set
1647 * Return non-zero if the compression should be done, 0 otherwise.
1649 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1651 struct list_head *ws_list = get_workspace(0, 0);
1652 struct heuristic_ws *ws;
1657 ws = list_entry(ws_list, struct heuristic_ws, list);
1659 heuristic_collect_sample(inode, start, end, ws);
1661 if (sample_repeated_patterns(ws)) {
1666 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1668 for (i = 0; i < ws->sample_size; i++) {
1669 byte = ws->sample[i];
1670 ws->bucket[byte].count++;
1673 i = byte_set_size(ws);
1674 if (i < BYTE_SET_THRESHOLD) {
1679 i = byte_core_set_size(ws);
1680 if (i <= BYTE_CORE_SET_LOW) {
1685 if (i >= BYTE_CORE_SET_HIGH) {
1690 i = shannon_entropy(ws);
1691 if (i <= ENTROPY_LVL_ACEPTABLE) {
1697 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1698 * needed to give green light to compression.
1700 * For now just assume that compression at that level is not worth the
1701 * resources because:
1703 * 1. it is possible to defrag the data later
1705 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1706 * values, every bucket has counter at level ~54. The heuristic would
1707 * be confused. This can happen when data have some internal repeated
1708 * patterns like "abbacbbc...". This can be detected by analyzing
1709 * pairs of bytes, which is too costly.
1711 if (i < ENTROPY_LVL_HIGH) {
1720 put_workspace(0, ws_list);
1725 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1726 * level, unrecognized string will set the default level
1728 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1730 unsigned int level = 0;
1736 if (str[0] == ':') {
1737 ret = kstrtouint(str + 1, 10, &level);
1742 level = btrfs_compress_set_level(type, level);