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 end_compressed_bio_read(struct btrfs_bio *bbio)
146 struct compressed_bio *cb = bbio->private;
150 if (bbio->bio.bi_status)
151 cb->status = bbio->bio.bi_status;
153 cb->status = errno_to_blk_status(btrfs_decompress_bio(cb));
155 /* Release the compressed pages */
156 for (index = 0; index < cb->nr_pages; index++) {
157 page = cb->compressed_pages[index];
158 page->mapping = NULL;
162 /* Do io completion on the original bio */
163 btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status);
165 /* Finally free the cb struct */
166 kfree(cb->compressed_pages);
172 * Clear the writeback bits on all of the file
173 * pages for a compressed write
175 static noinline void end_compressed_writeback(struct inode *inode,
176 const struct compressed_bio *cb)
178 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
179 unsigned long index = cb->start >> PAGE_SHIFT;
180 unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
181 struct folio_batch fbatch;
182 const int errno = blk_status_to_errno(cb->status);
187 mapping_set_error(inode->i_mapping, errno);
189 folio_batch_init(&fbatch);
190 while (index <= end_index) {
191 ret = filemap_get_folios(inode->i_mapping, &index, end_index,
197 for (i = 0; i < ret; i++) {
198 struct folio *folio = fbatch.folios[i];
201 folio_set_error(folio);
202 btrfs_page_clamp_clear_writeback(fs_info, &folio->page,
205 folio_batch_release(&fbatch);
207 /* the inode may be gone now */
210 static void finish_compressed_bio_write(struct compressed_bio *cb)
212 struct inode *inode = cb->inode;
216 * Ok, we're the last bio for this extent, step one is to call back
217 * into the FS and do all the end_io operations.
219 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
220 cb->start, cb->start + cb->len - 1,
221 cb->status == BLK_STS_OK);
224 end_compressed_writeback(inode, cb);
225 /* Note, our inode could be gone now */
228 * Release the compressed pages, these came from alloc_page and
229 * are not attached to the inode at all
231 for (index = 0; index < cb->nr_pages; index++) {
232 struct page *page = cb->compressed_pages[index];
234 page->mapping = NULL;
238 /* Finally free the cb struct */
239 kfree(cb->compressed_pages);
243 static void btrfs_finish_compressed_write_work(struct work_struct *work)
245 struct compressed_bio *cb =
246 container_of(work, struct compressed_bio, write_end_work);
248 finish_compressed_bio_write(cb);
252 * Do the cleanup once all the compressed pages hit the disk. This will clear
253 * writeback on the file pages and free the compressed pages.
255 * This also calls the writeback end hooks for the file pages so that metadata
256 * and checksums can be updated in the file.
258 static void end_compressed_bio_write(struct btrfs_bio *bbio)
260 struct compressed_bio *cb = bbio->private;
262 if (bbio->bio.bi_status)
263 cb->status = bbio->bio.bi_status;
265 if (refcount_dec_and_test(&cb->pending_ios)) {
266 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
268 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
274 * Allocate a compressed_bio, which will be used to read/write on-disk
275 * (aka, compressed) * data.
277 * @cb: The compressed_bio structure, which records all the needed
278 * information to bind the compressed data to the uncompressed
280 * @disk_byten: The logical bytenr where the compressed data will be read
281 * from or written to.
282 * @endio_func: The endio function to call after the IO for compressed data
285 static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr,
287 btrfs_bio_end_io_t endio_func)
291 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, BTRFS_I(cb->inode), endio_func,
293 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
295 if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
296 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
297 struct btrfs_device *device;
299 device = btrfs_zoned_get_device(fs_info, disk_bytenr,
300 fs_info->sectorsize);
301 if (IS_ERR(device)) {
303 return ERR_CAST(device);
306 bio_set_dev(bio, device->bdev);
308 refcount_inc(&cb->pending_ios);
313 * worker function to build and submit bios for previously compressed pages.
314 * The corresponding pages in the inode should be marked for writeback
315 * and the compressed pages should have a reference on them for dropping
316 * when the IO is complete.
318 * This also checksums the file bytes and gets things ready for
321 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
322 unsigned int len, u64 disk_start,
323 unsigned int compressed_len,
324 struct page **compressed_pages,
325 unsigned int nr_pages,
326 blk_opf_t write_flags,
327 struct cgroup_subsys_state *blkcg_css,
330 struct btrfs_fs_info *fs_info = inode->root->fs_info;
331 struct bio *bio = NULL;
332 struct compressed_bio *cb;
333 u64 cur_disk_bytenr = disk_start;
334 blk_status_t ret = BLK_STS_OK;
335 const bool use_append = btrfs_use_zone_append(inode, disk_start);
336 const enum req_op bio_op = REQ_BTRFS_ONE_ORDERED |
337 (use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE);
339 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
340 IS_ALIGNED(len, fs_info->sectorsize));
341 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
343 return BLK_STS_RESOURCE;
344 refcount_set(&cb->pending_ios, 1);
345 cb->status = BLK_STS_OK;
346 cb->inode = &inode->vfs_inode;
349 cb->compressed_pages = compressed_pages;
350 cb->compressed_len = compressed_len;
351 cb->writeback = writeback;
352 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
353 cb->nr_pages = nr_pages;
356 kthread_associate_blkcg(blkcg_css);
358 while (cur_disk_bytenr < disk_start + compressed_len) {
359 u64 offset = cur_disk_bytenr - disk_start;
360 unsigned int index = offset >> PAGE_SHIFT;
361 unsigned int real_size;
363 struct page *page = compressed_pages[index];
366 /* Allocate new bio if submitted or not yet allocated */
368 bio = alloc_compressed_bio(cb, cur_disk_bytenr,
369 bio_op | write_flags, end_compressed_bio_write);
371 ret = errno_to_blk_status(PTR_ERR(bio));
374 btrfs_bio(bio)->file_offset = start;
376 bio->bi_opf |= REQ_CGROUP_PUNT;
379 * We have various limits on the real read size:
381 * - compressed length boundary
383 real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset));
384 real_size = min_t(u64, real_size, compressed_len - offset);
385 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
388 added = bio_add_zone_append_page(bio, page, real_size,
389 offset_in_page(offset));
391 added = bio_add_page(bio, page, real_size,
392 offset_in_page(offset));
393 /* Reached zoned boundary */
397 cur_disk_bytenr += added;
399 /* Finished the range */
400 if (cur_disk_bytenr == disk_start + compressed_len)
404 ASSERT(bio->bi_iter.bi_size);
405 btrfs_submit_bio(bio, 0);
412 kthread_associate_blkcg(NULL);
414 if (refcount_dec_and_test(&cb->pending_ios))
415 finish_compressed_bio_write(cb);
419 static u64 bio_end_offset(struct bio *bio)
421 struct bio_vec *last = bio_last_bvec_all(bio);
423 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
427 * Add extra pages in the same compressed file extent so that we don't need to
428 * re-read the same extent again and again.
430 * NOTE: this won't work well for subpage, as for subpage read, we lock the
431 * full page then submit bio for each compressed/regular extents.
433 * This means, if we have several sectors in the same page points to the same
434 * on-disk compressed data, we will re-read the same extent many times and
435 * this function can only help for the next page.
437 static noinline int add_ra_bio_pages(struct inode *inode,
439 struct compressed_bio *cb,
440 int *memstall, unsigned long *pflags)
442 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
443 unsigned long end_index;
444 u64 cur = bio_end_offset(cb->orig_bio);
445 u64 isize = i_size_read(inode);
448 struct extent_map *em;
449 struct address_space *mapping = inode->i_mapping;
450 struct extent_map_tree *em_tree;
451 struct extent_io_tree *tree;
452 int sectors_missed = 0;
454 em_tree = &BTRFS_I(inode)->extent_tree;
455 tree = &BTRFS_I(inode)->io_tree;
461 * For current subpage support, we only support 64K page size,
462 * which means maximum compressed extent size (128K) is just 2x page
464 * This makes readahead less effective, so here disable readahead for
465 * subpage for now, until full compressed write is supported.
467 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
470 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
472 while (cur < compressed_end) {
474 u64 pg_index = cur >> PAGE_SHIFT;
477 if (pg_index > end_index)
480 page = xa_load(&mapping->i_pages, pg_index);
481 if (page && !xa_is_value(page)) {
482 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
483 fs_info->sectorsize_bits;
485 /* Beyond threshold, no need to continue */
486 if (sectors_missed > 4)
490 * Jump to next page start as we already have page for
493 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
497 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
502 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
504 /* There is already a page, skip to page end */
505 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
509 if (!*memstall && PageWorkingset(page)) {
510 psi_memstall_enter(pflags);
514 ret = set_page_extent_mapped(page);
521 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
522 lock_extent(tree, cur, page_end, NULL);
523 read_lock(&em_tree->lock);
524 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
525 read_unlock(&em_tree->lock);
528 * At this point, we have a locked page in the page cache for
529 * these bytes in the file. But, we have to make sure they map
530 * to this compressed extent on disk.
532 if (!em || cur < em->start ||
533 (cur + fs_info->sectorsize > extent_map_end(em)) ||
534 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
536 unlock_extent(tree, cur, page_end, NULL);
543 if (page->index == end_index) {
544 size_t zero_offset = offset_in_page(isize);
548 zeros = PAGE_SIZE - zero_offset;
549 memzero_page(page, zero_offset, zeros);
553 add_size = min(em->start + em->len, page_end + 1) - cur;
554 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
555 if (ret != add_size) {
556 unlock_extent(tree, cur, page_end, NULL);
562 * If it's subpage, we also need to increase its
563 * subpage::readers number, as at endio we will decrease
564 * subpage::readers and to unlock the page.
566 if (fs_info->sectorsize < PAGE_SIZE)
567 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
575 * for a compressed read, the bio we get passed has all the inode pages
576 * in it. We don't actually do IO on those pages but allocate new ones
577 * to hold the compressed pages on disk.
579 * bio->bi_iter.bi_sector points to the compressed extent on disk
580 * bio->bi_io_vec points to all of the inode pages
582 * After the compressed pages are read, we copy the bytes into the
583 * bio we were passed and then call the bio end_io calls
585 void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
588 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
589 struct extent_map_tree *em_tree;
590 struct compressed_bio *cb;
591 unsigned int compressed_len;
592 struct bio *comp_bio;
593 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
594 u64 cur_disk_byte = disk_bytenr;
598 struct extent_map *em;
599 unsigned long pflags;
605 em_tree = &BTRFS_I(inode)->extent_tree;
607 file_offset = bio_first_bvec_all(bio)->bv_offset +
608 page_offset(bio_first_page_all(bio));
610 /* we need the actual starting offset of this extent in the file */
611 read_lock(&em_tree->lock);
612 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
613 read_unlock(&em_tree->lock);
619 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
620 compressed_len = em->block_len;
621 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
623 ret = BLK_STS_RESOURCE;
627 refcount_set(&cb->pending_ios, 1);
628 cb->status = BLK_STS_OK;
631 cb->start = em->orig_start;
633 em_start = em->start;
635 cb->len = bio->bi_iter.bi_size;
636 cb->compressed_len = compressed_len;
637 cb->compress_type = em->compress_type;
643 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
644 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
645 if (!cb->compressed_pages) {
646 ret = BLK_STS_RESOURCE;
650 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
652 ret = BLK_STS_RESOURCE;
656 add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
658 /* include any pages we added in add_ra-bio_pages */
659 cb->len = bio->bi_iter.bi_size;
661 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS, REQ_OP_READ, BTRFS_I(cb->inode),
662 end_compressed_bio_read, cb);
663 comp_bio->bi_iter.bi_sector = (cur_disk_byte >> SECTOR_SHIFT);
665 while (cur_disk_byte < disk_bytenr + compressed_len) {
666 u64 offset = cur_disk_byte - disk_bytenr;
667 unsigned int index = offset >> PAGE_SHIFT;
668 unsigned int real_size;
670 struct page *page = cb->compressed_pages[index];
673 * We have various limit on the real read size:
675 * - compressed length boundary
677 real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset));
678 real_size = min_t(u64, real_size, compressed_len - offset);
679 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
681 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
683 * Maximum compressed extent is smaller than bio size limit,
684 * thus bio_add_page() should always success.
686 ASSERT(added == real_size);
687 cur_disk_byte += added;
691 psi_memstall_leave(&pflags);
694 * Stash the initial offset of this chunk, as there is no direct
695 * correlation between compressed pages and the original file offset.
696 * The field is only used for printing error messages anyway.
698 btrfs_bio(comp_bio)->file_offset = file_offset;
700 ASSERT(comp_bio->bi_iter.bi_size);
701 btrfs_submit_bio(comp_bio, mirror_num);
705 if (cb->compressed_pages) {
706 for (i = 0; i < cb->nr_pages; i++) {
707 if (cb->compressed_pages[i])
708 __free_page(cb->compressed_pages[i]);
712 kfree(cb->compressed_pages);
716 btrfs_bio_end_io(btrfs_bio(bio), ret);
721 * Heuristic uses systematic sampling to collect data from the input data
722 * range, the logic can be tuned by the following constants:
724 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
725 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
727 #define SAMPLING_READ_SIZE (16)
728 #define SAMPLING_INTERVAL (256)
731 * For statistical analysis of the input data we consider bytes that form a
732 * Galois Field of 256 objects. Each object has an attribute count, ie. how
733 * many times the object appeared in the sample.
735 #define BUCKET_SIZE (256)
738 * The size of the sample is based on a statistical sampling rule of thumb.
739 * The common way is to perform sampling tests as long as the number of
740 * elements in each cell is at least 5.
742 * Instead of 5, we choose 32 to obtain more accurate results.
743 * If the data contain the maximum number of symbols, which is 256, we obtain a
744 * sample size bound by 8192.
746 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
747 * from up to 512 locations.
749 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
750 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
756 struct heuristic_ws {
757 /* Partial copy of input data */
760 /* Buckets store counters for each byte value */
761 struct bucket_item *bucket;
763 struct bucket_item *bucket_b;
764 struct list_head list;
767 static struct workspace_manager heuristic_wsm;
769 static void free_heuristic_ws(struct list_head *ws)
771 struct heuristic_ws *workspace;
773 workspace = list_entry(ws, struct heuristic_ws, list);
775 kvfree(workspace->sample);
776 kfree(workspace->bucket);
777 kfree(workspace->bucket_b);
781 static struct list_head *alloc_heuristic_ws(unsigned int level)
783 struct heuristic_ws *ws;
785 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
787 return ERR_PTR(-ENOMEM);
789 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
793 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
797 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
801 INIT_LIST_HEAD(&ws->list);
804 free_heuristic_ws(&ws->list);
805 return ERR_PTR(-ENOMEM);
808 const struct btrfs_compress_op btrfs_heuristic_compress = {
809 .workspace_manager = &heuristic_wsm,
812 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
813 /* The heuristic is represented as compression type 0 */
814 &btrfs_heuristic_compress,
815 &btrfs_zlib_compress,
817 &btrfs_zstd_compress,
820 static struct list_head *alloc_workspace(int type, unsigned int level)
823 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
824 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
825 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
826 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
829 * This can't happen, the type is validated several times
830 * before we get here.
836 static void free_workspace(int type, struct list_head *ws)
839 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
840 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
841 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
842 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
845 * This can't happen, the type is validated several times
846 * before we get here.
852 static void btrfs_init_workspace_manager(int type)
854 struct workspace_manager *wsm;
855 struct list_head *workspace;
857 wsm = btrfs_compress_op[type]->workspace_manager;
858 INIT_LIST_HEAD(&wsm->idle_ws);
859 spin_lock_init(&wsm->ws_lock);
860 atomic_set(&wsm->total_ws, 0);
861 init_waitqueue_head(&wsm->ws_wait);
864 * Preallocate one workspace for each compression type so we can
865 * guarantee forward progress in the worst case
867 workspace = alloc_workspace(type, 0);
868 if (IS_ERR(workspace)) {
870 "BTRFS: cannot preallocate compression workspace, will try later\n");
872 atomic_set(&wsm->total_ws, 1);
874 list_add(workspace, &wsm->idle_ws);
878 static void btrfs_cleanup_workspace_manager(int type)
880 struct workspace_manager *wsman;
881 struct list_head *ws;
883 wsman = btrfs_compress_op[type]->workspace_manager;
884 while (!list_empty(&wsman->idle_ws)) {
885 ws = wsman->idle_ws.next;
887 free_workspace(type, ws);
888 atomic_dec(&wsman->total_ws);
893 * This finds an available workspace or allocates a new one.
894 * If it's not possible to allocate a new one, waits until there's one.
895 * Preallocation makes a forward progress guarantees and we do not return
898 struct list_head *btrfs_get_workspace(int type, unsigned int level)
900 struct workspace_manager *wsm;
901 struct list_head *workspace;
902 int cpus = num_online_cpus();
904 struct list_head *idle_ws;
907 wait_queue_head_t *ws_wait;
910 wsm = btrfs_compress_op[type]->workspace_manager;
911 idle_ws = &wsm->idle_ws;
912 ws_lock = &wsm->ws_lock;
913 total_ws = &wsm->total_ws;
914 ws_wait = &wsm->ws_wait;
915 free_ws = &wsm->free_ws;
919 if (!list_empty(idle_ws)) {
920 workspace = idle_ws->next;
923 spin_unlock(ws_lock);
927 if (atomic_read(total_ws) > cpus) {
930 spin_unlock(ws_lock);
931 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
932 if (atomic_read(total_ws) > cpus && !*free_ws)
934 finish_wait(ws_wait, &wait);
937 atomic_inc(total_ws);
938 spin_unlock(ws_lock);
941 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
942 * to turn it off here because we might get called from the restricted
943 * context of btrfs_compress_bio/btrfs_compress_pages
945 nofs_flag = memalloc_nofs_save();
946 workspace = alloc_workspace(type, level);
947 memalloc_nofs_restore(nofs_flag);
949 if (IS_ERR(workspace)) {
950 atomic_dec(total_ws);
954 * Do not return the error but go back to waiting. There's a
955 * workspace preallocated for each type and the compression
956 * time is bounded so we get to a workspace eventually. This
957 * makes our caller's life easier.
959 * To prevent silent and low-probability deadlocks (when the
960 * initial preallocation fails), check if there are any
963 if (atomic_read(total_ws) == 0) {
964 static DEFINE_RATELIMIT_STATE(_rs,
965 /* once per minute */ 60 * HZ,
968 if (__ratelimit(&_rs)) {
969 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
977 static struct list_head *get_workspace(int type, int level)
980 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
981 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
982 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
983 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
986 * This can't happen, the type is validated several times
987 * before we get here.
994 * put a workspace struct back on the list or free it if we have enough
995 * idle ones sitting around
997 void btrfs_put_workspace(int type, struct list_head *ws)
999 struct workspace_manager *wsm;
1000 struct list_head *idle_ws;
1001 spinlock_t *ws_lock;
1003 wait_queue_head_t *ws_wait;
1006 wsm = btrfs_compress_op[type]->workspace_manager;
1007 idle_ws = &wsm->idle_ws;
1008 ws_lock = &wsm->ws_lock;
1009 total_ws = &wsm->total_ws;
1010 ws_wait = &wsm->ws_wait;
1011 free_ws = &wsm->free_ws;
1014 if (*free_ws <= num_online_cpus()) {
1015 list_add(ws, idle_ws);
1017 spin_unlock(ws_lock);
1020 spin_unlock(ws_lock);
1022 free_workspace(type, ws);
1023 atomic_dec(total_ws);
1025 cond_wake_up(ws_wait);
1028 static void put_workspace(int type, struct list_head *ws)
1031 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1032 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1033 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
1034 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1037 * This can't happen, the type is validated several times
1038 * before we get here.
1045 * Adjust @level according to the limits of the compression algorithm or
1046 * fallback to default
1048 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1050 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1053 level = ops->default_level;
1055 level = min(level, ops->max_level);
1061 * Given an address space and start and length, compress the bytes into @pages
1062 * that are allocated on demand.
1064 * @type_level is encoded algorithm and level, where level 0 means whatever
1065 * default the algorithm chooses and is opaque here;
1066 * - compression algo are 0-3
1067 * - the level are bits 4-7
1069 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1070 * and returns number of actually allocated pages
1072 * @total_in is used to return the number of bytes actually read. It
1073 * may be smaller than the input length if we had to exit early because we
1074 * ran out of room in the pages array or because we cross the
1075 * max_out threshold.
1077 * @total_out is an in/out parameter, must be set to the input length and will
1078 * be also used to return the total number of compressed bytes
1080 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1081 u64 start, struct page **pages,
1082 unsigned long *out_pages,
1083 unsigned long *total_in,
1084 unsigned long *total_out)
1086 int type = btrfs_compress_type(type_level);
1087 int level = btrfs_compress_level(type_level);
1088 struct list_head *workspace;
1091 level = btrfs_compress_set_level(type, level);
1092 workspace = get_workspace(type, level);
1093 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1094 out_pages, total_in, total_out);
1095 put_workspace(type, workspace);
1099 static int btrfs_decompress_bio(struct compressed_bio *cb)
1101 struct list_head *workspace;
1103 int type = cb->compress_type;
1105 workspace = get_workspace(type, 0);
1106 ret = compression_decompress_bio(workspace, cb);
1107 put_workspace(type, workspace);
1113 * a less complex decompression routine. Our compressed data fits in a
1114 * single page, and we want to read a single page out of it.
1115 * start_byte tells us the offset into the compressed data we're interested in
1117 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1118 unsigned long start_byte, size_t srclen, size_t destlen)
1120 struct list_head *workspace;
1123 workspace = get_workspace(type, 0);
1124 ret = compression_decompress(type, workspace, data_in, dest_page,
1125 start_byte, srclen, destlen);
1126 put_workspace(type, workspace);
1131 int __init btrfs_init_compress(void)
1133 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1134 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1135 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1136 zstd_init_workspace_manager();
1140 void __cold btrfs_exit_compress(void)
1142 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1143 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1144 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1145 zstd_cleanup_workspace_manager();
1149 * Copy decompressed data from working buffer to pages.
1151 * @buf: The decompressed data buffer
1152 * @buf_len: The decompressed data length
1153 * @decompressed: Number of bytes that are already decompressed inside the
1155 * @cb: The compressed extent descriptor
1156 * @orig_bio: The original bio that the caller wants to read for
1158 * An easier to understand graph is like below:
1160 * |<- orig_bio ->| |<- orig_bio->|
1161 * |<------- full decompressed extent ----->|
1162 * |<----------- @cb range ---->|
1163 * | |<-- @buf_len -->|
1164 * |<--- @decompressed --->|
1166 * Note that, @cb can be a subpage of the full decompressed extent, but
1167 * @cb->start always has the same as the orig_file_offset value of the full
1168 * decompressed extent.
1170 * When reading compressed extent, we have to read the full compressed extent,
1171 * while @orig_bio may only want part of the range.
1172 * Thus this function will ensure only data covered by @orig_bio will be copied
1175 * Return 0 if we have copied all needed contents for @orig_bio.
1176 * Return >0 if we need continue decompress.
1178 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1179 struct compressed_bio *cb, u32 decompressed)
1181 struct bio *orig_bio = cb->orig_bio;
1182 /* Offset inside the full decompressed extent */
1185 cur_offset = decompressed;
1186 /* The main loop to do the copy */
1187 while (cur_offset < decompressed + buf_len) {
1188 struct bio_vec bvec;
1191 /* Offset inside the full decompressed extent */
1194 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1196 * cb->start may underflow, but subtracting that value can still
1197 * give us correct offset inside the full decompressed extent.
1199 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1201 /* Haven't reached the bvec range, exit */
1202 if (decompressed + buf_len <= bvec_offset)
1205 copy_start = max(cur_offset, bvec_offset);
1206 copy_len = min(bvec_offset + bvec.bv_len,
1207 decompressed + buf_len) - copy_start;
1211 * Extra range check to ensure we didn't go beyond
1214 ASSERT(copy_start - decompressed < buf_len);
1215 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1216 buf + copy_start - decompressed, copy_len);
1217 cur_offset += copy_len;
1219 bio_advance(orig_bio, copy_len);
1220 /* Finished the bio */
1221 if (!orig_bio->bi_iter.bi_size)
1228 * Shannon Entropy calculation
1230 * Pure byte distribution analysis fails to determine compressibility of data.
1231 * Try calculating entropy to estimate the average minimum number of bits
1232 * needed to encode the sampled data.
1234 * For convenience, return the percentage of needed bits, instead of amount of
1237 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1238 * and can be compressible with high probability
1240 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1242 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1244 #define ENTROPY_LVL_ACEPTABLE (65)
1245 #define ENTROPY_LVL_HIGH (80)
1248 * For increasead precision in shannon_entropy calculation,
1249 * let's do pow(n, M) to save more digits after comma:
1251 * - maximum int bit length is 64
1252 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1253 * - 13 * 4 = 52 < 64 -> M = 4
1257 static inline u32 ilog2_w(u64 n)
1259 return ilog2(n * n * n * n);
1262 static u32 shannon_entropy(struct heuristic_ws *ws)
1264 const u32 entropy_max = 8 * ilog2_w(2);
1265 u32 entropy_sum = 0;
1266 u32 p, p_base, sz_base;
1269 sz_base = ilog2_w(ws->sample_size);
1270 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1271 p = ws->bucket[i].count;
1272 p_base = ilog2_w(p);
1273 entropy_sum += p * (sz_base - p_base);
1276 entropy_sum /= ws->sample_size;
1277 return entropy_sum * 100 / entropy_max;
1280 #define RADIX_BASE 4U
1281 #define COUNTERS_SIZE (1U << RADIX_BASE)
1283 static u8 get4bits(u64 num, int shift) {
1288 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1293 * Use 4 bits as radix base
1294 * Use 16 u32 counters for calculating new position in buf array
1296 * @array - array that will be sorted
1297 * @array_buf - buffer array to store sorting results
1298 * must be equal in size to @array
1301 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1306 u32 counters[COUNTERS_SIZE];
1314 * Try avoid useless loop iterations for small numbers stored in big
1315 * counters. Example: 48 33 4 ... in 64bit array
1317 max_num = array[0].count;
1318 for (i = 1; i < num; i++) {
1319 buf_num = array[i].count;
1320 if (buf_num > max_num)
1324 buf_num = ilog2(max_num);
1325 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1328 while (shift < bitlen) {
1329 memset(counters, 0, sizeof(counters));
1331 for (i = 0; i < num; i++) {
1332 buf_num = array[i].count;
1333 addr = get4bits(buf_num, shift);
1337 for (i = 1; i < COUNTERS_SIZE; i++)
1338 counters[i] += counters[i - 1];
1340 for (i = num - 1; i >= 0; i--) {
1341 buf_num = array[i].count;
1342 addr = get4bits(buf_num, shift);
1344 new_addr = counters[addr];
1345 array_buf[new_addr] = array[i];
1348 shift += RADIX_BASE;
1351 * Normal radix expects to move data from a temporary array, to
1352 * the main one. But that requires some CPU time. Avoid that
1353 * by doing another sort iteration to original array instead of
1356 memset(counters, 0, sizeof(counters));
1358 for (i = 0; i < num; i ++) {
1359 buf_num = array_buf[i].count;
1360 addr = get4bits(buf_num, shift);
1364 for (i = 1; i < COUNTERS_SIZE; i++)
1365 counters[i] += counters[i - 1];
1367 for (i = num - 1; i >= 0; i--) {
1368 buf_num = array_buf[i].count;
1369 addr = get4bits(buf_num, shift);
1371 new_addr = counters[addr];
1372 array[new_addr] = array_buf[i];
1375 shift += RADIX_BASE;
1380 * Size of the core byte set - how many bytes cover 90% of the sample
1382 * There are several types of structured binary data that use nearly all byte
1383 * values. The distribution can be uniform and counts in all buckets will be
1384 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1386 * Other possibility is normal (Gaussian) distribution, where the data could
1387 * be potentially compressible, but we have to take a few more steps to decide
1390 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1391 * compression algo can easy fix that
1392 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1393 * probability is not compressible
1395 #define BYTE_CORE_SET_LOW (64)
1396 #define BYTE_CORE_SET_HIGH (200)
1398 static int byte_core_set_size(struct heuristic_ws *ws)
1401 u32 coreset_sum = 0;
1402 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1403 struct bucket_item *bucket = ws->bucket;
1405 /* Sort in reverse order */
1406 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1408 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1409 coreset_sum += bucket[i].count;
1411 if (coreset_sum > core_set_threshold)
1414 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1415 coreset_sum += bucket[i].count;
1416 if (coreset_sum > core_set_threshold)
1424 * Count byte values in buckets.
1425 * This heuristic can detect textual data (configs, xml, json, html, etc).
1426 * Because in most text-like data byte set is restricted to limited number of
1427 * possible characters, and that restriction in most cases makes data easy to
1430 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1431 * less - compressible
1432 * more - need additional analysis
1434 #define BYTE_SET_THRESHOLD (64)
1436 static u32 byte_set_size(const struct heuristic_ws *ws)
1439 u32 byte_set_size = 0;
1441 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1442 if (ws->bucket[i].count > 0)
1447 * Continue collecting count of byte values in buckets. If the byte
1448 * set size is bigger then the threshold, it's pointless to continue,
1449 * the detection technique would fail for this type of data.
1451 for (; i < BUCKET_SIZE; i++) {
1452 if (ws->bucket[i].count > 0) {
1454 if (byte_set_size > BYTE_SET_THRESHOLD)
1455 return byte_set_size;
1459 return byte_set_size;
1462 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1464 const u32 half_of_sample = ws->sample_size / 2;
1465 const u8 *data = ws->sample;
1467 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1470 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1471 struct heuristic_ws *ws)
1474 u64 index, index_end;
1475 u32 i, curr_sample_pos;
1479 * Compression handles the input data by chunks of 128KiB
1480 * (defined by BTRFS_MAX_UNCOMPRESSED)
1482 * We do the same for the heuristic and loop over the whole range.
1484 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1485 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1487 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1488 end = start + BTRFS_MAX_UNCOMPRESSED;
1490 index = start >> PAGE_SHIFT;
1491 index_end = end >> PAGE_SHIFT;
1493 /* Don't miss unaligned end */
1494 if (!PAGE_ALIGNED(end))
1497 curr_sample_pos = 0;
1498 while (index < index_end) {
1499 page = find_get_page(inode->i_mapping, index);
1500 in_data = kmap_local_page(page);
1501 /* Handle case where the start is not aligned to PAGE_SIZE */
1502 i = start % PAGE_SIZE;
1503 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1504 /* Don't sample any garbage from the last page */
1505 if (start > end - SAMPLING_READ_SIZE)
1507 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1508 SAMPLING_READ_SIZE);
1509 i += SAMPLING_INTERVAL;
1510 start += SAMPLING_INTERVAL;
1511 curr_sample_pos += SAMPLING_READ_SIZE;
1513 kunmap_local(in_data);
1519 ws->sample_size = curr_sample_pos;
1523 * Compression heuristic.
1525 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1526 * quickly (compared to direct compression) detect data characteristics
1527 * (compressible/incompressible) to avoid wasting CPU time on incompressible
1530 * The following types of analysis can be performed:
1531 * - detect mostly zero data
1532 * - detect data with low "byte set" size (text, etc)
1533 * - detect data with low/high "core byte" set
1535 * Return non-zero if the compression should be done, 0 otherwise.
1537 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1539 struct list_head *ws_list = get_workspace(0, 0);
1540 struct heuristic_ws *ws;
1545 ws = list_entry(ws_list, struct heuristic_ws, list);
1547 heuristic_collect_sample(inode, start, end, ws);
1549 if (sample_repeated_patterns(ws)) {
1554 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1556 for (i = 0; i < ws->sample_size; i++) {
1557 byte = ws->sample[i];
1558 ws->bucket[byte].count++;
1561 i = byte_set_size(ws);
1562 if (i < BYTE_SET_THRESHOLD) {
1567 i = byte_core_set_size(ws);
1568 if (i <= BYTE_CORE_SET_LOW) {
1573 if (i >= BYTE_CORE_SET_HIGH) {
1578 i = shannon_entropy(ws);
1579 if (i <= ENTROPY_LVL_ACEPTABLE) {
1585 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1586 * needed to give green light to compression.
1588 * For now just assume that compression at that level is not worth the
1589 * resources because:
1591 * 1. it is possible to defrag the data later
1593 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1594 * values, every bucket has counter at level ~54. The heuristic would
1595 * be confused. This can happen when data have some internal repeated
1596 * patterns like "abbacbbc...". This can be detected by analyzing
1597 * pairs of bytes, which is too costly.
1599 if (i < ENTROPY_LVL_HIGH) {
1608 put_workspace(0, ws_list);
1613 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1614 * level, unrecognized string will set the default level
1616 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1618 unsigned int level = 0;
1624 if (str[0] == ':') {
1625 ret = kstrtouint(str + 1, 10, &level);
1630 level = btrfs_compress_set_level(type, level);