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;
261 struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb);
263 cb->status = bbio->bio.bi_status;
264 queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
270 * worker function to build and submit bios for previously compressed pages.
271 * The corresponding pages in the inode should be marked for writeback
272 * and the compressed pages should have a reference on them for dropping
273 * when the IO is complete.
275 * This also checksums the file bytes and gets things ready for
278 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
279 unsigned int len, u64 disk_start,
280 unsigned int compressed_len,
281 struct page **compressed_pages,
282 unsigned int nr_pages,
283 blk_opf_t write_flags,
284 struct cgroup_subsys_state *blkcg_css,
287 struct btrfs_fs_info *fs_info = inode->root->fs_info;
288 struct bio *bio = NULL;
289 struct compressed_bio *cb;
290 u64 cur_disk_bytenr = disk_start;
291 blk_status_t ret = BLK_STS_OK;
293 ASSERT(IS_ALIGNED(start, fs_info->sectorsize) &&
294 IS_ALIGNED(len, fs_info->sectorsize));
295 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
297 return BLK_STS_RESOURCE;
298 cb->status = BLK_STS_OK;
299 cb->inode = &inode->vfs_inode;
302 cb->compressed_pages = compressed_pages;
303 cb->compressed_len = compressed_len;
304 cb->writeback = writeback;
305 INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
306 cb->nr_pages = nr_pages;
309 kthread_associate_blkcg(blkcg_css);
310 write_flags |= REQ_CGROUP_PUNT;
313 write_flags |= REQ_BTRFS_ONE_ORDERED;
314 bio = btrfs_bio_alloc(BIO_MAX_VECS, REQ_OP_WRITE | write_flags,
315 BTRFS_I(cb->inode), end_compressed_bio_write, cb);
316 bio->bi_iter.bi_sector = cur_disk_bytenr >> SECTOR_SHIFT;
317 btrfs_bio(bio)->file_offset = start;
319 while (cur_disk_bytenr < disk_start + compressed_len) {
320 u64 offset = cur_disk_bytenr - disk_start;
321 unsigned int index = offset >> PAGE_SHIFT;
322 unsigned int real_size;
324 struct page *page = compressed_pages[index];
327 * We have various limits on the real read size:
329 * - compressed length boundary
331 real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset));
332 real_size = min_t(u64, real_size, compressed_len - offset);
333 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
335 added = bio_add_page(bio, page, real_size, offset_in_page(offset));
337 * Maximum compressed extent is smaller than bio size limit,
338 * thus bio_add_page() should always success.
340 ASSERT(added == real_size);
341 cur_disk_bytenr += added;
344 /* Finished the range. */
345 ASSERT(bio->bi_iter.bi_size);
346 btrfs_submit_bio(bio, 0);
348 kthread_associate_blkcg(NULL);
352 static u64 bio_end_offset(struct bio *bio)
354 struct bio_vec *last = bio_last_bvec_all(bio);
356 return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
360 * Add extra pages in the same compressed file extent so that we don't need to
361 * re-read the same extent again and again.
363 * NOTE: this won't work well for subpage, as for subpage read, we lock the
364 * full page then submit bio for each compressed/regular extents.
366 * This means, if we have several sectors in the same page points to the same
367 * on-disk compressed data, we will re-read the same extent many times and
368 * this function can only help for the next page.
370 static noinline int add_ra_bio_pages(struct inode *inode,
372 struct compressed_bio *cb,
373 int *memstall, unsigned long *pflags)
375 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
376 unsigned long end_index;
377 u64 cur = bio_end_offset(cb->orig_bio);
378 u64 isize = i_size_read(inode);
381 struct extent_map *em;
382 struct address_space *mapping = inode->i_mapping;
383 struct extent_map_tree *em_tree;
384 struct extent_io_tree *tree;
385 int sectors_missed = 0;
387 em_tree = &BTRFS_I(inode)->extent_tree;
388 tree = &BTRFS_I(inode)->io_tree;
394 * For current subpage support, we only support 64K page size,
395 * which means maximum compressed extent size (128K) is just 2x page
397 * This makes readahead less effective, so here disable readahead for
398 * subpage for now, until full compressed write is supported.
400 if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
403 end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
405 while (cur < compressed_end) {
407 u64 pg_index = cur >> PAGE_SHIFT;
410 if (pg_index > end_index)
413 page = xa_load(&mapping->i_pages, pg_index);
414 if (page && !xa_is_value(page)) {
415 sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
416 fs_info->sectorsize_bits;
418 /* Beyond threshold, no need to continue */
419 if (sectors_missed > 4)
423 * Jump to next page start as we already have page for
426 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
430 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
435 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
437 /* There is already a page, skip to page end */
438 cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
442 if (!*memstall && PageWorkingset(page)) {
443 psi_memstall_enter(pflags);
447 ret = set_page_extent_mapped(page);
454 page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
455 lock_extent(tree, cur, page_end, NULL);
456 read_lock(&em_tree->lock);
457 em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
458 read_unlock(&em_tree->lock);
461 * At this point, we have a locked page in the page cache for
462 * these bytes in the file. But, we have to make sure they map
463 * to this compressed extent on disk.
465 if (!em || cur < em->start ||
466 (cur + fs_info->sectorsize > extent_map_end(em)) ||
467 (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
469 unlock_extent(tree, cur, page_end, NULL);
476 if (page->index == end_index) {
477 size_t zero_offset = offset_in_page(isize);
481 zeros = PAGE_SIZE - zero_offset;
482 memzero_page(page, zero_offset, zeros);
486 add_size = min(em->start + em->len, page_end + 1) - cur;
487 ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur));
488 if (ret != add_size) {
489 unlock_extent(tree, cur, page_end, NULL);
495 * If it's subpage, we also need to increase its
496 * subpage::readers number, as at endio we will decrease
497 * subpage::readers and to unlock the page.
499 if (fs_info->sectorsize < PAGE_SIZE)
500 btrfs_subpage_start_reader(fs_info, page, cur, add_size);
508 * for a compressed read, the bio we get passed has all the inode pages
509 * in it. We don't actually do IO on those pages but allocate new ones
510 * to hold the compressed pages on disk.
512 * bio->bi_iter.bi_sector points to the compressed extent on disk
513 * bio->bi_io_vec points to all of the inode pages
515 * After the compressed pages are read, we copy the bytes into the
516 * bio we were passed and then call the bio end_io calls
518 void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
521 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
522 struct extent_map_tree *em_tree;
523 struct compressed_bio *cb;
524 unsigned int compressed_len;
525 struct bio *comp_bio;
526 const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT;
527 u64 cur_disk_byte = disk_bytenr;
531 struct extent_map *em;
532 unsigned long pflags;
538 em_tree = &BTRFS_I(inode)->extent_tree;
540 file_offset = bio_first_bvec_all(bio)->bv_offset +
541 page_offset(bio_first_page_all(bio));
543 /* we need the actual starting offset of this extent in the file */
544 read_lock(&em_tree->lock);
545 em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
546 read_unlock(&em_tree->lock);
552 ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
553 compressed_len = em->block_len;
554 cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS);
556 ret = BLK_STS_RESOURCE;
560 cb->status = BLK_STS_OK;
563 cb->start = em->orig_start;
565 em_start = em->start;
567 cb->len = bio->bi_iter.bi_size;
568 cb->compressed_len = compressed_len;
569 cb->compress_type = em->compress_type;
575 cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
576 cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
577 if (!cb->compressed_pages) {
578 ret = BLK_STS_RESOURCE;
582 ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages);
584 ret = BLK_STS_RESOURCE;
588 add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags);
590 /* include any pages we added in add_ra-bio_pages */
591 cb->len = bio->bi_iter.bi_size;
593 comp_bio = btrfs_bio_alloc(BIO_MAX_VECS, REQ_OP_READ, BTRFS_I(cb->inode),
594 end_compressed_bio_read, cb);
595 comp_bio->bi_iter.bi_sector = (cur_disk_byte >> SECTOR_SHIFT);
597 while (cur_disk_byte < disk_bytenr + compressed_len) {
598 u64 offset = cur_disk_byte - disk_bytenr;
599 unsigned int index = offset >> PAGE_SHIFT;
600 unsigned int real_size;
602 struct page *page = cb->compressed_pages[index];
605 * We have various limit on the real read size:
607 * - compressed length boundary
609 real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset));
610 real_size = min_t(u64, real_size, compressed_len - offset);
611 ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize));
613 added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset));
615 * Maximum compressed extent is smaller than bio size limit,
616 * thus bio_add_page() should always success.
618 ASSERT(added == real_size);
619 cur_disk_byte += added;
623 psi_memstall_leave(&pflags);
626 * Stash the initial offset of this chunk, as there is no direct
627 * correlation between compressed pages and the original file offset.
628 * The field is only used for printing error messages anyway.
630 btrfs_bio(comp_bio)->file_offset = file_offset;
632 ASSERT(comp_bio->bi_iter.bi_size);
633 btrfs_submit_bio(comp_bio, mirror_num);
637 if (cb->compressed_pages) {
638 for (i = 0; i < cb->nr_pages; i++) {
639 if (cb->compressed_pages[i])
640 __free_page(cb->compressed_pages[i]);
644 kfree(cb->compressed_pages);
648 btrfs_bio_end_io(btrfs_bio(bio), ret);
653 * Heuristic uses systematic sampling to collect data from the input data
654 * range, the logic can be tuned by the following constants:
656 * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
657 * @SAMPLING_INTERVAL - range from which the sampled data can be collected
659 #define SAMPLING_READ_SIZE (16)
660 #define SAMPLING_INTERVAL (256)
663 * For statistical analysis of the input data we consider bytes that form a
664 * Galois Field of 256 objects. Each object has an attribute count, ie. how
665 * many times the object appeared in the sample.
667 #define BUCKET_SIZE (256)
670 * The size of the sample is based on a statistical sampling rule of thumb.
671 * The common way is to perform sampling tests as long as the number of
672 * elements in each cell is at least 5.
674 * Instead of 5, we choose 32 to obtain more accurate results.
675 * If the data contain the maximum number of symbols, which is 256, we obtain a
676 * sample size bound by 8192.
678 * For a sample of at most 8KB of data per data range: 16 consecutive bytes
679 * from up to 512 locations.
681 #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \
682 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
688 struct heuristic_ws {
689 /* Partial copy of input data */
692 /* Buckets store counters for each byte value */
693 struct bucket_item *bucket;
695 struct bucket_item *bucket_b;
696 struct list_head list;
699 static struct workspace_manager heuristic_wsm;
701 static void free_heuristic_ws(struct list_head *ws)
703 struct heuristic_ws *workspace;
705 workspace = list_entry(ws, struct heuristic_ws, list);
707 kvfree(workspace->sample);
708 kfree(workspace->bucket);
709 kfree(workspace->bucket_b);
713 static struct list_head *alloc_heuristic_ws(unsigned int level)
715 struct heuristic_ws *ws;
717 ws = kzalloc(sizeof(*ws), GFP_KERNEL);
719 return ERR_PTR(-ENOMEM);
721 ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
725 ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
729 ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
733 INIT_LIST_HEAD(&ws->list);
736 free_heuristic_ws(&ws->list);
737 return ERR_PTR(-ENOMEM);
740 const struct btrfs_compress_op btrfs_heuristic_compress = {
741 .workspace_manager = &heuristic_wsm,
744 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
745 /* The heuristic is represented as compression type 0 */
746 &btrfs_heuristic_compress,
747 &btrfs_zlib_compress,
749 &btrfs_zstd_compress,
752 static struct list_head *alloc_workspace(int type, unsigned int level)
755 case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
756 case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
757 case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level);
758 case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
761 * This can't happen, the type is validated several times
762 * before we get here.
768 static void free_workspace(int type, struct list_head *ws)
771 case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
772 case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
773 case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws);
774 case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
777 * This can't happen, the type is validated several times
778 * before we get here.
784 static void btrfs_init_workspace_manager(int type)
786 struct workspace_manager *wsm;
787 struct list_head *workspace;
789 wsm = btrfs_compress_op[type]->workspace_manager;
790 INIT_LIST_HEAD(&wsm->idle_ws);
791 spin_lock_init(&wsm->ws_lock);
792 atomic_set(&wsm->total_ws, 0);
793 init_waitqueue_head(&wsm->ws_wait);
796 * Preallocate one workspace for each compression type so we can
797 * guarantee forward progress in the worst case
799 workspace = alloc_workspace(type, 0);
800 if (IS_ERR(workspace)) {
802 "BTRFS: cannot preallocate compression workspace, will try later\n");
804 atomic_set(&wsm->total_ws, 1);
806 list_add(workspace, &wsm->idle_ws);
810 static void btrfs_cleanup_workspace_manager(int type)
812 struct workspace_manager *wsman;
813 struct list_head *ws;
815 wsman = btrfs_compress_op[type]->workspace_manager;
816 while (!list_empty(&wsman->idle_ws)) {
817 ws = wsman->idle_ws.next;
819 free_workspace(type, ws);
820 atomic_dec(&wsman->total_ws);
825 * This finds an available workspace or allocates a new one.
826 * If it's not possible to allocate a new one, waits until there's one.
827 * Preallocation makes a forward progress guarantees and we do not return
830 struct list_head *btrfs_get_workspace(int type, unsigned int level)
832 struct workspace_manager *wsm;
833 struct list_head *workspace;
834 int cpus = num_online_cpus();
836 struct list_head *idle_ws;
839 wait_queue_head_t *ws_wait;
842 wsm = btrfs_compress_op[type]->workspace_manager;
843 idle_ws = &wsm->idle_ws;
844 ws_lock = &wsm->ws_lock;
845 total_ws = &wsm->total_ws;
846 ws_wait = &wsm->ws_wait;
847 free_ws = &wsm->free_ws;
851 if (!list_empty(idle_ws)) {
852 workspace = idle_ws->next;
855 spin_unlock(ws_lock);
859 if (atomic_read(total_ws) > cpus) {
862 spin_unlock(ws_lock);
863 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
864 if (atomic_read(total_ws) > cpus && !*free_ws)
866 finish_wait(ws_wait, &wait);
869 atomic_inc(total_ws);
870 spin_unlock(ws_lock);
873 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
874 * to turn it off here because we might get called from the restricted
875 * context of btrfs_compress_bio/btrfs_compress_pages
877 nofs_flag = memalloc_nofs_save();
878 workspace = alloc_workspace(type, level);
879 memalloc_nofs_restore(nofs_flag);
881 if (IS_ERR(workspace)) {
882 atomic_dec(total_ws);
886 * Do not return the error but go back to waiting. There's a
887 * workspace preallocated for each type and the compression
888 * time is bounded so we get to a workspace eventually. This
889 * makes our caller's life easier.
891 * To prevent silent and low-probability deadlocks (when the
892 * initial preallocation fails), check if there are any
895 if (atomic_read(total_ws) == 0) {
896 static DEFINE_RATELIMIT_STATE(_rs,
897 /* once per minute */ 60 * HZ,
900 if (__ratelimit(&_rs)) {
901 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
909 static struct list_head *get_workspace(int type, int level)
912 case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
913 case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
914 case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level);
915 case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
918 * This can't happen, the type is validated several times
919 * before we get here.
926 * put a workspace struct back on the list or free it if we have enough
927 * idle ones sitting around
929 void btrfs_put_workspace(int type, struct list_head *ws)
931 struct workspace_manager *wsm;
932 struct list_head *idle_ws;
935 wait_queue_head_t *ws_wait;
938 wsm = btrfs_compress_op[type]->workspace_manager;
939 idle_ws = &wsm->idle_ws;
940 ws_lock = &wsm->ws_lock;
941 total_ws = &wsm->total_ws;
942 ws_wait = &wsm->ws_wait;
943 free_ws = &wsm->free_ws;
946 if (*free_ws <= num_online_cpus()) {
947 list_add(ws, idle_ws);
949 spin_unlock(ws_lock);
952 spin_unlock(ws_lock);
954 free_workspace(type, ws);
955 atomic_dec(total_ws);
957 cond_wake_up(ws_wait);
960 static void put_workspace(int type, struct list_head *ws)
963 case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
964 case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
965 case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws);
966 case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
969 * This can't happen, the type is validated several times
970 * before we get here.
977 * Adjust @level according to the limits of the compression algorithm or
978 * fallback to default
980 static unsigned int btrfs_compress_set_level(int type, unsigned level)
982 const struct btrfs_compress_op *ops = btrfs_compress_op[type];
985 level = ops->default_level;
987 level = min(level, ops->max_level);
993 * Given an address space and start and length, compress the bytes into @pages
994 * that are allocated on demand.
996 * @type_level is encoded algorithm and level, where level 0 means whatever
997 * default the algorithm chooses and is opaque here;
998 * - compression algo are 0-3
999 * - the level are bits 4-7
1001 * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1002 * and returns number of actually allocated pages
1004 * @total_in is used to return the number of bytes actually read. It
1005 * may be smaller than the input length if we had to exit early because we
1006 * ran out of room in the pages array or because we cross the
1007 * max_out threshold.
1009 * @total_out is an in/out parameter, must be set to the input length and will
1010 * be also used to return the total number of compressed bytes
1012 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1013 u64 start, struct page **pages,
1014 unsigned long *out_pages,
1015 unsigned long *total_in,
1016 unsigned long *total_out)
1018 int type = btrfs_compress_type(type_level);
1019 int level = btrfs_compress_level(type_level);
1020 struct list_head *workspace;
1023 level = btrfs_compress_set_level(type, level);
1024 workspace = get_workspace(type, level);
1025 ret = compression_compress_pages(type, workspace, mapping, start, pages,
1026 out_pages, total_in, total_out);
1027 put_workspace(type, workspace);
1031 static int btrfs_decompress_bio(struct compressed_bio *cb)
1033 struct list_head *workspace;
1035 int type = cb->compress_type;
1037 workspace = get_workspace(type, 0);
1038 ret = compression_decompress_bio(workspace, cb);
1039 put_workspace(type, workspace);
1045 * a less complex decompression routine. Our compressed data fits in a
1046 * single page, and we want to read a single page out of it.
1047 * start_byte tells us the offset into the compressed data we're interested in
1049 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1050 unsigned long start_byte, size_t srclen, size_t destlen)
1052 struct list_head *workspace;
1055 workspace = get_workspace(type, 0);
1056 ret = compression_decompress(type, workspace, data_in, dest_page,
1057 start_byte, srclen, destlen);
1058 put_workspace(type, workspace);
1063 int __init btrfs_init_compress(void)
1065 btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1066 btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1067 btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1068 zstd_init_workspace_manager();
1072 void __cold btrfs_exit_compress(void)
1074 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1075 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1076 btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1077 zstd_cleanup_workspace_manager();
1081 * Copy decompressed data from working buffer to pages.
1083 * @buf: The decompressed data buffer
1084 * @buf_len: The decompressed data length
1085 * @decompressed: Number of bytes that are already decompressed inside the
1087 * @cb: The compressed extent descriptor
1088 * @orig_bio: The original bio that the caller wants to read for
1090 * An easier to understand graph is like below:
1092 * |<- orig_bio ->| |<- orig_bio->|
1093 * |<------- full decompressed extent ----->|
1094 * |<----------- @cb range ---->|
1095 * | |<-- @buf_len -->|
1096 * |<--- @decompressed --->|
1098 * Note that, @cb can be a subpage of the full decompressed extent, but
1099 * @cb->start always has the same as the orig_file_offset value of the full
1100 * decompressed extent.
1102 * When reading compressed extent, we have to read the full compressed extent,
1103 * while @orig_bio may only want part of the range.
1104 * Thus this function will ensure only data covered by @orig_bio will be copied
1107 * Return 0 if we have copied all needed contents for @orig_bio.
1108 * Return >0 if we need continue decompress.
1110 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1111 struct compressed_bio *cb, u32 decompressed)
1113 struct bio *orig_bio = cb->orig_bio;
1114 /* Offset inside the full decompressed extent */
1117 cur_offset = decompressed;
1118 /* The main loop to do the copy */
1119 while (cur_offset < decompressed + buf_len) {
1120 struct bio_vec bvec;
1123 /* Offset inside the full decompressed extent */
1126 bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1128 * cb->start may underflow, but subtracting that value can still
1129 * give us correct offset inside the full decompressed extent.
1131 bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1133 /* Haven't reached the bvec range, exit */
1134 if (decompressed + buf_len <= bvec_offset)
1137 copy_start = max(cur_offset, bvec_offset);
1138 copy_len = min(bvec_offset + bvec.bv_len,
1139 decompressed + buf_len) - copy_start;
1143 * Extra range check to ensure we didn't go beyond
1146 ASSERT(copy_start - decompressed < buf_len);
1147 memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1148 buf + copy_start - decompressed, copy_len);
1149 cur_offset += copy_len;
1151 bio_advance(orig_bio, copy_len);
1152 /* Finished the bio */
1153 if (!orig_bio->bi_iter.bi_size)
1160 * Shannon Entropy calculation
1162 * Pure byte distribution analysis fails to determine compressibility of data.
1163 * Try calculating entropy to estimate the average minimum number of bits
1164 * needed to encode the sampled data.
1166 * For convenience, return the percentage of needed bits, instead of amount of
1169 * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1170 * and can be compressible with high probability
1172 * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1174 * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1176 #define ENTROPY_LVL_ACEPTABLE (65)
1177 #define ENTROPY_LVL_HIGH (80)
1180 * For increasead precision in shannon_entropy calculation,
1181 * let's do pow(n, M) to save more digits after comma:
1183 * - maximum int bit length is 64
1184 * - ilog2(MAX_SAMPLE_SIZE) -> 13
1185 * - 13 * 4 = 52 < 64 -> M = 4
1189 static inline u32 ilog2_w(u64 n)
1191 return ilog2(n * n * n * n);
1194 static u32 shannon_entropy(struct heuristic_ws *ws)
1196 const u32 entropy_max = 8 * ilog2_w(2);
1197 u32 entropy_sum = 0;
1198 u32 p, p_base, sz_base;
1201 sz_base = ilog2_w(ws->sample_size);
1202 for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1203 p = ws->bucket[i].count;
1204 p_base = ilog2_w(p);
1205 entropy_sum += p * (sz_base - p_base);
1208 entropy_sum /= ws->sample_size;
1209 return entropy_sum * 100 / entropy_max;
1212 #define RADIX_BASE 4U
1213 #define COUNTERS_SIZE (1U << RADIX_BASE)
1215 static u8 get4bits(u64 num, int shift) {
1220 low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1225 * Use 4 bits as radix base
1226 * Use 16 u32 counters for calculating new position in buf array
1228 * @array - array that will be sorted
1229 * @array_buf - buffer array to store sorting results
1230 * must be equal in size to @array
1233 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1238 u32 counters[COUNTERS_SIZE];
1246 * Try avoid useless loop iterations for small numbers stored in big
1247 * counters. Example: 48 33 4 ... in 64bit array
1249 max_num = array[0].count;
1250 for (i = 1; i < num; i++) {
1251 buf_num = array[i].count;
1252 if (buf_num > max_num)
1256 buf_num = ilog2(max_num);
1257 bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1260 while (shift < bitlen) {
1261 memset(counters, 0, sizeof(counters));
1263 for (i = 0; i < num; i++) {
1264 buf_num = array[i].count;
1265 addr = get4bits(buf_num, shift);
1269 for (i = 1; i < COUNTERS_SIZE; i++)
1270 counters[i] += counters[i - 1];
1272 for (i = num - 1; i >= 0; i--) {
1273 buf_num = array[i].count;
1274 addr = get4bits(buf_num, shift);
1276 new_addr = counters[addr];
1277 array_buf[new_addr] = array[i];
1280 shift += RADIX_BASE;
1283 * Normal radix expects to move data from a temporary array, to
1284 * the main one. But that requires some CPU time. Avoid that
1285 * by doing another sort iteration to original array instead of
1288 memset(counters, 0, sizeof(counters));
1290 for (i = 0; i < num; i ++) {
1291 buf_num = array_buf[i].count;
1292 addr = get4bits(buf_num, shift);
1296 for (i = 1; i < COUNTERS_SIZE; i++)
1297 counters[i] += counters[i - 1];
1299 for (i = num - 1; i >= 0; i--) {
1300 buf_num = array_buf[i].count;
1301 addr = get4bits(buf_num, shift);
1303 new_addr = counters[addr];
1304 array[new_addr] = array_buf[i];
1307 shift += RADIX_BASE;
1312 * Size of the core byte set - how many bytes cover 90% of the sample
1314 * There are several types of structured binary data that use nearly all byte
1315 * values. The distribution can be uniform and counts in all buckets will be
1316 * nearly the same (eg. encrypted data). Unlikely to be compressible.
1318 * Other possibility is normal (Gaussian) distribution, where the data could
1319 * be potentially compressible, but we have to take a few more steps to decide
1322 * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently,
1323 * compression algo can easy fix that
1324 * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1325 * probability is not compressible
1327 #define BYTE_CORE_SET_LOW (64)
1328 #define BYTE_CORE_SET_HIGH (200)
1330 static int byte_core_set_size(struct heuristic_ws *ws)
1333 u32 coreset_sum = 0;
1334 const u32 core_set_threshold = ws->sample_size * 90 / 100;
1335 struct bucket_item *bucket = ws->bucket;
1337 /* Sort in reverse order */
1338 radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1340 for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1341 coreset_sum += bucket[i].count;
1343 if (coreset_sum > core_set_threshold)
1346 for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1347 coreset_sum += bucket[i].count;
1348 if (coreset_sum > core_set_threshold)
1356 * Count byte values in buckets.
1357 * This heuristic can detect textual data (configs, xml, json, html, etc).
1358 * Because in most text-like data byte set is restricted to limited number of
1359 * possible characters, and that restriction in most cases makes data easy to
1362 * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1363 * less - compressible
1364 * more - need additional analysis
1366 #define BYTE_SET_THRESHOLD (64)
1368 static u32 byte_set_size(const struct heuristic_ws *ws)
1371 u32 byte_set_size = 0;
1373 for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1374 if (ws->bucket[i].count > 0)
1379 * Continue collecting count of byte values in buckets. If the byte
1380 * set size is bigger then the threshold, it's pointless to continue,
1381 * the detection technique would fail for this type of data.
1383 for (; i < BUCKET_SIZE; i++) {
1384 if (ws->bucket[i].count > 0) {
1386 if (byte_set_size > BYTE_SET_THRESHOLD)
1387 return byte_set_size;
1391 return byte_set_size;
1394 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1396 const u32 half_of_sample = ws->sample_size / 2;
1397 const u8 *data = ws->sample;
1399 return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1402 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1403 struct heuristic_ws *ws)
1406 u64 index, index_end;
1407 u32 i, curr_sample_pos;
1411 * Compression handles the input data by chunks of 128KiB
1412 * (defined by BTRFS_MAX_UNCOMPRESSED)
1414 * We do the same for the heuristic and loop over the whole range.
1416 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1417 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1419 if (end - start > BTRFS_MAX_UNCOMPRESSED)
1420 end = start + BTRFS_MAX_UNCOMPRESSED;
1422 index = start >> PAGE_SHIFT;
1423 index_end = end >> PAGE_SHIFT;
1425 /* Don't miss unaligned end */
1426 if (!PAGE_ALIGNED(end))
1429 curr_sample_pos = 0;
1430 while (index < index_end) {
1431 page = find_get_page(inode->i_mapping, index);
1432 in_data = kmap_local_page(page);
1433 /* Handle case where the start is not aligned to PAGE_SIZE */
1434 i = start % PAGE_SIZE;
1435 while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1436 /* Don't sample any garbage from the last page */
1437 if (start > end - SAMPLING_READ_SIZE)
1439 memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1440 SAMPLING_READ_SIZE);
1441 i += SAMPLING_INTERVAL;
1442 start += SAMPLING_INTERVAL;
1443 curr_sample_pos += SAMPLING_READ_SIZE;
1445 kunmap_local(in_data);
1451 ws->sample_size = curr_sample_pos;
1455 * Compression heuristic.
1457 * For now is's a naive and optimistic 'return true', we'll extend the logic to
1458 * quickly (compared to direct compression) detect data characteristics
1459 * (compressible/incompressible) to avoid wasting CPU time on incompressible
1462 * The following types of analysis can be performed:
1463 * - detect mostly zero data
1464 * - detect data with low "byte set" size (text, etc)
1465 * - detect data with low/high "core byte" set
1467 * Return non-zero if the compression should be done, 0 otherwise.
1469 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1471 struct list_head *ws_list = get_workspace(0, 0);
1472 struct heuristic_ws *ws;
1477 ws = list_entry(ws_list, struct heuristic_ws, list);
1479 heuristic_collect_sample(inode, start, end, ws);
1481 if (sample_repeated_patterns(ws)) {
1486 memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1488 for (i = 0; i < ws->sample_size; i++) {
1489 byte = ws->sample[i];
1490 ws->bucket[byte].count++;
1493 i = byte_set_size(ws);
1494 if (i < BYTE_SET_THRESHOLD) {
1499 i = byte_core_set_size(ws);
1500 if (i <= BYTE_CORE_SET_LOW) {
1505 if (i >= BYTE_CORE_SET_HIGH) {
1510 i = shannon_entropy(ws);
1511 if (i <= ENTROPY_LVL_ACEPTABLE) {
1517 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1518 * needed to give green light to compression.
1520 * For now just assume that compression at that level is not worth the
1521 * resources because:
1523 * 1. it is possible to defrag the data later
1525 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1526 * values, every bucket has counter at level ~54. The heuristic would
1527 * be confused. This can happen when data have some internal repeated
1528 * patterns like "abbacbbc...". This can be detected by analyzing
1529 * pairs of bytes, which is too costly.
1531 if (i < ENTROPY_LVL_HIGH) {
1540 put_workspace(0, ws_list);
1545 * Convert the compression suffix (eg. after "zlib" starting with ":") to
1546 * level, unrecognized string will set the default level
1548 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1550 unsigned int level = 0;
1556 if (str[0] == ':') {
1557 ret = kstrtouint(str + 1, 10, &level);
1562 level = btrfs_compress_set_level(type, level);