xref: /linux/fs/btrfs/compression.c (revision e04e2b760ddbe3d7b283a05898c3a029085cd8cd)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2008 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/kernel.h>
7 #include <linux/bio.h>
8 #include <linux/file.h>
9 #include <linux/fs.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 <linux/shrinker.h>
24 #include <crypto/hash.h>
25 #include "misc.h"
26 #include "ctree.h"
27 #include "fs.h"
28 #include "btrfs_inode.h"
29 #include "bio.h"
30 #include "ordered-data.h"
31 #include "compression.h"
32 #include "extent_io.h"
33 #include "extent_map.h"
34 #include "subpage.h"
35 #include "messages.h"
36 #include "super.h"
37 
38 static struct bio_set btrfs_compressed_bioset;
39 
40 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
41 
42 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
43 {
44 	switch (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];
50 	default:
51 		break;
52 	}
53 
54 	return NULL;
55 }
56 
57 static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio)
58 {
59 	return container_of(bbio, struct compressed_bio, bbio);
60 }
61 
62 static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode,
63 						   u64 start, blk_opf_t op,
64 						   btrfs_bio_end_io_t end_io)
65 {
66 	struct btrfs_bio *bbio;
67 
68 	bbio = btrfs_bio(bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, op,
69 					  GFP_NOFS, &btrfs_compressed_bioset));
70 	btrfs_bio_init(bbio, inode->root->fs_info, end_io, NULL);
71 	bbio->inode = inode;
72 	bbio->file_offset = start;
73 	return to_compressed_bio(bbio);
74 }
75 
76 bool btrfs_compress_is_valid_type(const char *str, size_t len)
77 {
78 	int i;
79 
80 	for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
81 		size_t comp_len = strlen(btrfs_compress_types[i]);
82 
83 		if (len < comp_len)
84 			continue;
85 
86 		if (!strncmp(btrfs_compress_types[i], str, comp_len))
87 			return true;
88 	}
89 	return false;
90 }
91 
92 static int compression_compress_pages(int type, struct list_head *ws,
93 				      struct address_space *mapping, u64 start,
94 				      struct folio **folios, unsigned long *out_folios,
95 				      unsigned long *total_in, unsigned long *total_out)
96 {
97 	switch (type) {
98 	case BTRFS_COMPRESS_ZLIB:
99 		return zlib_compress_folios(ws, mapping, start, folios,
100 					    out_folios, total_in, total_out);
101 	case BTRFS_COMPRESS_LZO:
102 		return lzo_compress_folios(ws, mapping, start, folios,
103 					   out_folios, total_in, total_out);
104 	case BTRFS_COMPRESS_ZSTD:
105 		return zstd_compress_folios(ws, mapping, start, folios,
106 					    out_folios, total_in, total_out);
107 	case BTRFS_COMPRESS_NONE:
108 	default:
109 		/*
110 		 * This can happen when compression races with remount setting
111 		 * it to 'no compress', while caller doesn't call
112 		 * inode_need_compress() to check if we really need to
113 		 * compress.
114 		 *
115 		 * Not a big deal, just need to inform caller that we
116 		 * haven't allocated any pages yet.
117 		 */
118 		*out_folios = 0;
119 		return -E2BIG;
120 	}
121 }
122 
123 static int compression_decompress_bio(struct list_head *ws,
124 				      struct compressed_bio *cb)
125 {
126 	switch (cb->compress_type) {
127 	case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
128 	case BTRFS_COMPRESS_LZO:  return lzo_decompress_bio(ws, cb);
129 	case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
130 	case BTRFS_COMPRESS_NONE:
131 	default:
132 		/*
133 		 * This can't happen, the type is validated several times
134 		 * before we get here.
135 		 */
136 		BUG();
137 	}
138 }
139 
140 static int compression_decompress(int type, struct list_head *ws,
141 		const u8 *data_in, struct page *dest_page,
142 		unsigned long dest_pgoff, size_t srclen, size_t destlen)
143 {
144 	switch (type) {
145 	case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
146 						dest_pgoff, srclen, destlen);
147 	case BTRFS_COMPRESS_LZO:  return lzo_decompress(ws, data_in, dest_page,
148 						dest_pgoff, srclen, destlen);
149 	case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
150 						dest_pgoff, srclen, destlen);
151 	case BTRFS_COMPRESS_NONE:
152 	default:
153 		/*
154 		 * This can't happen, the type is validated several times
155 		 * before we get here.
156 		 */
157 		BUG();
158 	}
159 }
160 
161 static void btrfs_free_compressed_folios(struct compressed_bio *cb)
162 {
163 	for (unsigned int i = 0; i < cb->nr_folios; i++)
164 		btrfs_free_compr_folio(cb->compressed_folios[i]);
165 	kfree(cb->compressed_folios);
166 }
167 
168 static int btrfs_decompress_bio(struct compressed_bio *cb);
169 
170 /*
171  * Global cache of last unused pages for compression/decompression.
172  */
173 static struct btrfs_compr_pool {
174 	struct shrinker *shrinker;
175 	spinlock_t lock;
176 	struct list_head list;
177 	int count;
178 	int thresh;
179 } compr_pool;
180 
181 static unsigned long btrfs_compr_pool_count(struct shrinker *sh, struct shrink_control *sc)
182 {
183 	int ret;
184 
185 	/*
186 	 * We must not read the values more than once if 'ret' gets expanded in
187 	 * the return statement so we don't accidentally return a negative
188 	 * number, even if the first condition finds it positive.
189 	 */
190 	ret = READ_ONCE(compr_pool.count) - READ_ONCE(compr_pool.thresh);
191 
192 	return ret > 0 ? ret : 0;
193 }
194 
195 static unsigned long btrfs_compr_pool_scan(struct shrinker *sh, struct shrink_control *sc)
196 {
197 	struct list_head remove;
198 	struct list_head *tmp, *next;
199 	int freed;
200 
201 	if (compr_pool.count == 0)
202 		return SHRINK_STOP;
203 
204 	INIT_LIST_HEAD(&remove);
205 
206 	/* For now, just simply drain the whole list. */
207 	spin_lock(&compr_pool.lock);
208 	list_splice_init(&compr_pool.list, &remove);
209 	freed = compr_pool.count;
210 	compr_pool.count = 0;
211 	spin_unlock(&compr_pool.lock);
212 
213 	list_for_each_safe(tmp, next, &remove) {
214 		struct page *page = list_entry(tmp, struct page, lru);
215 
216 		ASSERT(page_ref_count(page) == 1);
217 		put_page(page);
218 	}
219 
220 	return freed;
221 }
222 
223 /*
224  * Common wrappers for page allocation from compression wrappers
225  */
226 struct folio *btrfs_alloc_compr_folio(void)
227 {
228 	struct folio *folio = NULL;
229 
230 	spin_lock(&compr_pool.lock);
231 	if (compr_pool.count > 0) {
232 		folio = list_first_entry(&compr_pool.list, struct folio, lru);
233 		list_del_init(&folio->lru);
234 		compr_pool.count--;
235 	}
236 	spin_unlock(&compr_pool.lock);
237 
238 	if (folio)
239 		return folio;
240 
241 	return folio_alloc(GFP_NOFS, 0);
242 }
243 
244 void btrfs_free_compr_folio(struct folio *folio)
245 {
246 	bool do_free = false;
247 
248 	spin_lock(&compr_pool.lock);
249 	if (compr_pool.count > compr_pool.thresh) {
250 		do_free = true;
251 	} else {
252 		list_add(&folio->lru, &compr_pool.list);
253 		compr_pool.count++;
254 	}
255 	spin_unlock(&compr_pool.lock);
256 
257 	if (!do_free)
258 		return;
259 
260 	ASSERT(folio_ref_count(folio) == 1);
261 	folio_put(folio);
262 }
263 
264 static void end_bbio_compressed_read(struct btrfs_bio *bbio)
265 {
266 	struct compressed_bio *cb = to_compressed_bio(bbio);
267 	blk_status_t status = bbio->bio.bi_status;
268 
269 	if (!status)
270 		status = errno_to_blk_status(btrfs_decompress_bio(cb));
271 
272 	btrfs_free_compressed_folios(cb);
273 	btrfs_bio_end_io(cb->orig_bbio, status);
274 	bio_put(&bbio->bio);
275 }
276 
277 /*
278  * Clear the writeback bits on all of the file
279  * pages for a compressed write
280  */
281 static noinline void end_compressed_writeback(const struct compressed_bio *cb)
282 {
283 	struct inode *inode = &cb->bbio.inode->vfs_inode;
284 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
285 	unsigned long index = cb->start >> PAGE_SHIFT;
286 	unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
287 	struct folio_batch fbatch;
288 	const int error = blk_status_to_errno(cb->bbio.bio.bi_status);
289 	int i;
290 	int ret;
291 
292 	if (error)
293 		mapping_set_error(inode->i_mapping, error);
294 
295 	folio_batch_init(&fbatch);
296 	while (index <= end_index) {
297 		ret = filemap_get_folios(inode->i_mapping, &index, end_index,
298 				&fbatch);
299 
300 		if (ret == 0)
301 			return;
302 
303 		for (i = 0; i < ret; i++) {
304 			struct folio *folio = fbatch.folios[i];
305 
306 			btrfs_folio_clamp_clear_writeback(fs_info, folio,
307 							  cb->start, cb->len);
308 		}
309 		folio_batch_release(&fbatch);
310 	}
311 	/* the inode may be gone now */
312 }
313 
314 static void btrfs_finish_compressed_write_work(struct work_struct *work)
315 {
316 	struct compressed_bio *cb =
317 		container_of(work, struct compressed_bio, write_end_work);
318 
319 	btrfs_finish_ordered_extent(cb->bbio.ordered, NULL, cb->start, cb->len,
320 				    cb->bbio.bio.bi_status == BLK_STS_OK);
321 
322 	if (cb->writeback)
323 		end_compressed_writeback(cb);
324 	/* Note, our inode could be gone now */
325 
326 	btrfs_free_compressed_folios(cb);
327 	bio_put(&cb->bbio.bio);
328 }
329 
330 /*
331  * Do the cleanup once all the compressed pages hit the disk.  This will clear
332  * writeback on the file pages and free the compressed pages.
333  *
334  * This also calls the writeback end hooks for the file pages so that metadata
335  * and checksums can be updated in the file.
336  */
337 static void end_bbio_compressed_write(struct btrfs_bio *bbio)
338 {
339 	struct compressed_bio *cb = to_compressed_bio(bbio);
340 	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
341 
342 	queue_work(fs_info->compressed_write_workers, &cb->write_end_work);
343 }
344 
345 static void btrfs_add_compressed_bio_folios(struct compressed_bio *cb)
346 {
347 	struct bio *bio = &cb->bbio.bio;
348 	u32 offset = 0;
349 
350 	while (offset < cb->compressed_len) {
351 		int ret;
352 		u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
353 
354 		/* Maximum compressed extent is smaller than bio size limit. */
355 		ret = bio_add_folio(bio, cb->compressed_folios[offset >> PAGE_SHIFT],
356 				    len, 0);
357 		ASSERT(ret);
358 		offset += len;
359 	}
360 }
361 
362 /*
363  * worker function to build and submit bios for previously compressed pages.
364  * The corresponding pages in the inode should be marked for writeback
365  * and the compressed pages should have a reference on them for dropping
366  * when the IO is complete.
367  *
368  * This also checksums the file bytes and gets things ready for
369  * the end io hooks.
370  */
371 void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
372 				   struct folio **compressed_folios,
373 				   unsigned int nr_folios,
374 				   blk_opf_t write_flags,
375 				   bool writeback)
376 {
377 	struct btrfs_inode *inode = ordered->inode;
378 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
379 	struct compressed_bio *cb;
380 
381 	ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
382 	ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
383 
384 	cb = alloc_compressed_bio(inode, ordered->file_offset,
385 				  REQ_OP_WRITE | write_flags,
386 				  end_bbio_compressed_write);
387 	cb->start = ordered->file_offset;
388 	cb->len = ordered->num_bytes;
389 	cb->compressed_folios = compressed_folios;
390 	cb->compressed_len = ordered->disk_num_bytes;
391 	cb->writeback = writeback;
392 	INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
393 	cb->nr_folios = nr_folios;
394 	cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
395 	cb->bbio.ordered = ordered;
396 	btrfs_add_compressed_bio_folios(cb);
397 
398 	btrfs_submit_bio(&cb->bbio, 0);
399 }
400 
401 /*
402  * Add extra pages in the same compressed file extent so that we don't need to
403  * re-read the same extent again and again.
404  *
405  * NOTE: this won't work well for subpage, as for subpage read, we lock the
406  * full page then submit bio for each compressed/regular extents.
407  *
408  * This means, if we have several sectors in the same page points to the same
409  * on-disk compressed data, we will re-read the same extent many times and
410  * this function can only help for the next page.
411  */
412 static noinline int add_ra_bio_pages(struct inode *inode,
413 				     u64 compressed_end,
414 				     struct compressed_bio *cb,
415 				     int *memstall, unsigned long *pflags)
416 {
417 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
418 	unsigned long end_index;
419 	struct bio *orig_bio = &cb->orig_bbio->bio;
420 	u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
421 	u64 isize = i_size_read(inode);
422 	int ret;
423 	struct page *page;
424 	struct extent_map *em;
425 	struct address_space *mapping = inode->i_mapping;
426 	struct extent_map_tree *em_tree;
427 	struct extent_io_tree *tree;
428 	int sectors_missed = 0;
429 
430 	em_tree = &BTRFS_I(inode)->extent_tree;
431 	tree = &BTRFS_I(inode)->io_tree;
432 
433 	if (isize == 0)
434 		return 0;
435 
436 	/*
437 	 * For current subpage support, we only support 64K page size,
438 	 * which means maximum compressed extent size (128K) is just 2x page
439 	 * size.
440 	 * This makes readahead less effective, so here disable readahead for
441 	 * subpage for now, until full compressed write is supported.
442 	 */
443 	if (fs_info->sectorsize < PAGE_SIZE)
444 		return 0;
445 
446 	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
447 
448 	while (cur < compressed_end) {
449 		u64 page_end;
450 		u64 pg_index = cur >> PAGE_SHIFT;
451 		u32 add_size;
452 
453 		if (pg_index > end_index)
454 			break;
455 
456 		page = xa_load(&mapping->i_pages, pg_index);
457 		if (page && !xa_is_value(page)) {
458 			sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
459 					  fs_info->sectorsize_bits;
460 
461 			/* Beyond threshold, no need to continue */
462 			if (sectors_missed > 4)
463 				break;
464 
465 			/*
466 			 * Jump to next page start as we already have page for
467 			 * current offset.
468 			 */
469 			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
470 			continue;
471 		}
472 
473 		page = __page_cache_alloc(mapping_gfp_constraint(mapping,
474 								 ~__GFP_FS));
475 		if (!page)
476 			break;
477 
478 		if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
479 			put_page(page);
480 			/* There is already a page, skip to page end */
481 			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
482 			continue;
483 		}
484 
485 		if (!*memstall && PageWorkingset(page)) {
486 			psi_memstall_enter(pflags);
487 			*memstall = 1;
488 		}
489 
490 		ret = set_page_extent_mapped(page);
491 		if (ret < 0) {
492 			unlock_page(page);
493 			put_page(page);
494 			break;
495 		}
496 
497 		page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
498 		lock_extent(tree, cur, page_end, NULL);
499 		read_lock(&em_tree->lock);
500 		em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
501 		read_unlock(&em_tree->lock);
502 
503 		/*
504 		 * At this point, we have a locked page in the page cache for
505 		 * these bytes in the file.  But, we have to make sure they map
506 		 * to this compressed extent on disk.
507 		 */
508 		if (!em || cur < em->start ||
509 		    (cur + fs_info->sectorsize > extent_map_end(em)) ||
510 		    (extent_map_block_start(em) >> SECTOR_SHIFT) !=
511 		    orig_bio->bi_iter.bi_sector) {
512 			free_extent_map(em);
513 			unlock_extent(tree, cur, page_end, NULL);
514 			unlock_page(page);
515 			put_page(page);
516 			break;
517 		}
518 		add_size = min(em->start + em->len, page_end + 1) - cur;
519 		free_extent_map(em);
520 
521 		if (page->index == end_index) {
522 			size_t zero_offset = offset_in_page(isize);
523 
524 			if (zero_offset) {
525 				int zeros;
526 				zeros = PAGE_SIZE - zero_offset;
527 				memzero_page(page, zero_offset, zeros);
528 			}
529 		}
530 
531 		ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
532 		if (ret != add_size) {
533 			unlock_extent(tree, cur, page_end, NULL);
534 			unlock_page(page);
535 			put_page(page);
536 			break;
537 		}
538 		/*
539 		 * If it's subpage, we also need to increase its
540 		 * subpage::readers number, as at endio we will decrease
541 		 * subpage::readers and to unlock the page.
542 		 */
543 		if (fs_info->sectorsize < PAGE_SIZE)
544 			btrfs_subpage_start_reader(fs_info, page_folio(page),
545 						   cur, add_size);
546 		put_page(page);
547 		cur += add_size;
548 	}
549 	return 0;
550 }
551 
552 /*
553  * for a compressed read, the bio we get passed has all the inode pages
554  * in it.  We don't actually do IO on those pages but allocate new ones
555  * to hold the compressed pages on disk.
556  *
557  * bio->bi_iter.bi_sector points to the compressed extent on disk
558  * bio->bi_io_vec points to all of the inode pages
559  *
560  * After the compressed pages are read, we copy the bytes into the
561  * bio we were passed and then call the bio end_io calls
562  */
563 void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
564 {
565 	struct btrfs_inode *inode = bbio->inode;
566 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
567 	struct extent_map_tree *em_tree = &inode->extent_tree;
568 	struct compressed_bio *cb;
569 	unsigned int compressed_len;
570 	u64 file_offset = bbio->file_offset;
571 	u64 em_len;
572 	u64 em_start;
573 	struct extent_map *em;
574 	unsigned long pflags;
575 	int memstall = 0;
576 	blk_status_t ret;
577 	int ret2;
578 
579 	/* we need the actual starting offset of this extent in the file */
580 	read_lock(&em_tree->lock);
581 	em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
582 	read_unlock(&em_tree->lock);
583 	if (!em) {
584 		ret = BLK_STS_IOERR;
585 		goto out;
586 	}
587 
588 	ASSERT(extent_map_is_compressed(em));
589 	compressed_len = em->disk_num_bytes;
590 
591 	cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
592 				  end_bbio_compressed_read);
593 
594 	cb->start = em->start - em->offset;
595 	em_len = em->len;
596 	em_start = em->start;
597 
598 	cb->len = bbio->bio.bi_iter.bi_size;
599 	cb->compressed_len = compressed_len;
600 	cb->compress_type = extent_map_compression(em);
601 	cb->orig_bbio = bbio;
602 
603 	free_extent_map(em);
604 
605 	cb->nr_folios = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
606 	cb->compressed_folios = kcalloc(cb->nr_folios, sizeof(struct page *), GFP_NOFS);
607 	if (!cb->compressed_folios) {
608 		ret = BLK_STS_RESOURCE;
609 		goto out_free_bio;
610 	}
611 
612 	ret2 = btrfs_alloc_folio_array(cb->nr_folios, cb->compressed_folios);
613 	if (ret2) {
614 		ret = BLK_STS_RESOURCE;
615 		goto out_free_compressed_pages;
616 	}
617 
618 	add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
619 			 &pflags);
620 
621 	/* include any pages we added in add_ra-bio_pages */
622 	cb->len = bbio->bio.bi_iter.bi_size;
623 	cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
624 	btrfs_add_compressed_bio_folios(cb);
625 
626 	if (memstall)
627 		psi_memstall_leave(&pflags);
628 
629 	btrfs_submit_bio(&cb->bbio, 0);
630 	return;
631 
632 out_free_compressed_pages:
633 	kfree(cb->compressed_folios);
634 out_free_bio:
635 	bio_put(&cb->bbio.bio);
636 out:
637 	btrfs_bio_end_io(bbio, ret);
638 }
639 
640 /*
641  * Heuristic uses systematic sampling to collect data from the input data
642  * range, the logic can be tuned by the following constants:
643  *
644  * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
645  * @SAMPLING_INTERVAL  - range from which the sampled data can be collected
646  */
647 #define SAMPLING_READ_SIZE	(16)
648 #define SAMPLING_INTERVAL	(256)
649 
650 /*
651  * For statistical analysis of the input data we consider bytes that form a
652  * Galois Field of 256 objects. Each object has an attribute count, ie. how
653  * many times the object appeared in the sample.
654  */
655 #define BUCKET_SIZE		(256)
656 
657 /*
658  * The size of the sample is based on a statistical sampling rule of thumb.
659  * The common way is to perform sampling tests as long as the number of
660  * elements in each cell is at least 5.
661  *
662  * Instead of 5, we choose 32 to obtain more accurate results.
663  * If the data contain the maximum number of symbols, which is 256, we obtain a
664  * sample size bound by 8192.
665  *
666  * For a sample of at most 8KB of data per data range: 16 consecutive bytes
667  * from up to 512 locations.
668  */
669 #define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\
670 				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
671 
672 struct bucket_item {
673 	u32 count;
674 };
675 
676 struct heuristic_ws {
677 	/* Partial copy of input data */
678 	u8 *sample;
679 	u32 sample_size;
680 	/* Buckets store counters for each byte value */
681 	struct bucket_item *bucket;
682 	/* Sorting buffer */
683 	struct bucket_item *bucket_b;
684 	struct list_head list;
685 };
686 
687 static struct workspace_manager heuristic_wsm;
688 
689 static void free_heuristic_ws(struct list_head *ws)
690 {
691 	struct heuristic_ws *workspace;
692 
693 	workspace = list_entry(ws, struct heuristic_ws, list);
694 
695 	kvfree(workspace->sample);
696 	kfree(workspace->bucket);
697 	kfree(workspace->bucket_b);
698 	kfree(workspace);
699 }
700 
701 static struct list_head *alloc_heuristic_ws(unsigned int level)
702 {
703 	struct heuristic_ws *ws;
704 
705 	ws = kzalloc(sizeof(*ws), GFP_KERNEL);
706 	if (!ws)
707 		return ERR_PTR(-ENOMEM);
708 
709 	ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
710 	if (!ws->sample)
711 		goto fail;
712 
713 	ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
714 	if (!ws->bucket)
715 		goto fail;
716 
717 	ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
718 	if (!ws->bucket_b)
719 		goto fail;
720 
721 	INIT_LIST_HEAD(&ws->list);
722 	return &ws->list;
723 fail:
724 	free_heuristic_ws(&ws->list);
725 	return ERR_PTR(-ENOMEM);
726 }
727 
728 const struct btrfs_compress_op btrfs_heuristic_compress = {
729 	.workspace_manager = &heuristic_wsm,
730 };
731 
732 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
733 	/* The heuristic is represented as compression type 0 */
734 	&btrfs_heuristic_compress,
735 	&btrfs_zlib_compress,
736 	&btrfs_lzo_compress,
737 	&btrfs_zstd_compress,
738 };
739 
740 static struct list_head *alloc_workspace(int type, unsigned int level)
741 {
742 	switch (type) {
743 	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
744 	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
745 	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(level);
746 	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
747 	default:
748 		/*
749 		 * This can't happen, the type is validated several times
750 		 * before we get here.
751 		 */
752 		BUG();
753 	}
754 }
755 
756 static void free_workspace(int type, struct list_head *ws)
757 {
758 	switch (type) {
759 	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
760 	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
761 	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);
762 	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
763 	default:
764 		/*
765 		 * This can't happen, the type is validated several times
766 		 * before we get here.
767 		 */
768 		BUG();
769 	}
770 }
771 
772 static void btrfs_init_workspace_manager(int type)
773 {
774 	struct workspace_manager *wsm;
775 	struct list_head *workspace;
776 
777 	wsm = btrfs_compress_op[type]->workspace_manager;
778 	INIT_LIST_HEAD(&wsm->idle_ws);
779 	spin_lock_init(&wsm->ws_lock);
780 	atomic_set(&wsm->total_ws, 0);
781 	init_waitqueue_head(&wsm->ws_wait);
782 
783 	/*
784 	 * Preallocate one workspace for each compression type so we can
785 	 * guarantee forward progress in the worst case
786 	 */
787 	workspace = alloc_workspace(type, 0);
788 	if (IS_ERR(workspace)) {
789 		pr_warn(
790 	"BTRFS: cannot preallocate compression workspace, will try later\n");
791 	} else {
792 		atomic_set(&wsm->total_ws, 1);
793 		wsm->free_ws = 1;
794 		list_add(workspace, &wsm->idle_ws);
795 	}
796 }
797 
798 static void btrfs_cleanup_workspace_manager(int type)
799 {
800 	struct workspace_manager *wsman;
801 	struct list_head *ws;
802 
803 	wsman = btrfs_compress_op[type]->workspace_manager;
804 	while (!list_empty(&wsman->idle_ws)) {
805 		ws = wsman->idle_ws.next;
806 		list_del(ws);
807 		free_workspace(type, ws);
808 		atomic_dec(&wsman->total_ws);
809 	}
810 }
811 
812 /*
813  * This finds an available workspace or allocates a new one.
814  * If it's not possible to allocate a new one, waits until there's one.
815  * Preallocation makes a forward progress guarantees and we do not return
816  * errors.
817  */
818 struct list_head *btrfs_get_workspace(int type, unsigned int level)
819 {
820 	struct workspace_manager *wsm;
821 	struct list_head *workspace;
822 	int cpus = num_online_cpus();
823 	unsigned nofs_flag;
824 	struct list_head *idle_ws;
825 	spinlock_t *ws_lock;
826 	atomic_t *total_ws;
827 	wait_queue_head_t *ws_wait;
828 	int *free_ws;
829 
830 	wsm = btrfs_compress_op[type]->workspace_manager;
831 	idle_ws	 = &wsm->idle_ws;
832 	ws_lock	 = &wsm->ws_lock;
833 	total_ws = &wsm->total_ws;
834 	ws_wait	 = &wsm->ws_wait;
835 	free_ws	 = &wsm->free_ws;
836 
837 again:
838 	spin_lock(ws_lock);
839 	if (!list_empty(idle_ws)) {
840 		workspace = idle_ws->next;
841 		list_del(workspace);
842 		(*free_ws)--;
843 		spin_unlock(ws_lock);
844 		return workspace;
845 
846 	}
847 	if (atomic_read(total_ws) > cpus) {
848 		DEFINE_WAIT(wait);
849 
850 		spin_unlock(ws_lock);
851 		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
852 		if (atomic_read(total_ws) > cpus && !*free_ws)
853 			schedule();
854 		finish_wait(ws_wait, &wait);
855 		goto again;
856 	}
857 	atomic_inc(total_ws);
858 	spin_unlock(ws_lock);
859 
860 	/*
861 	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
862 	 * to turn it off here because we might get called from the restricted
863 	 * context of btrfs_compress_bio/btrfs_compress_pages
864 	 */
865 	nofs_flag = memalloc_nofs_save();
866 	workspace = alloc_workspace(type, level);
867 	memalloc_nofs_restore(nofs_flag);
868 
869 	if (IS_ERR(workspace)) {
870 		atomic_dec(total_ws);
871 		wake_up(ws_wait);
872 
873 		/*
874 		 * Do not return the error but go back to waiting. There's a
875 		 * workspace preallocated for each type and the compression
876 		 * time is bounded so we get to a workspace eventually. This
877 		 * makes our caller's life easier.
878 		 *
879 		 * To prevent silent and low-probability deadlocks (when the
880 		 * initial preallocation fails), check if there are any
881 		 * workspaces at all.
882 		 */
883 		if (atomic_read(total_ws) == 0) {
884 			static DEFINE_RATELIMIT_STATE(_rs,
885 					/* once per minute */ 60 * HZ,
886 					/* no burst */ 1);
887 
888 			if (__ratelimit(&_rs)) {
889 				pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
890 			}
891 		}
892 		goto again;
893 	}
894 	return workspace;
895 }
896 
897 static struct list_head *get_workspace(int type, int level)
898 {
899 	switch (type) {
900 	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
901 	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
902 	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(type, level);
903 	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
904 	default:
905 		/*
906 		 * This can't happen, the type is validated several times
907 		 * before we get here.
908 		 */
909 		BUG();
910 	}
911 }
912 
913 /*
914  * put a workspace struct back on the list or free it if we have enough
915  * idle ones sitting around
916  */
917 void btrfs_put_workspace(int type, struct list_head *ws)
918 {
919 	struct workspace_manager *wsm;
920 	struct list_head *idle_ws;
921 	spinlock_t *ws_lock;
922 	atomic_t *total_ws;
923 	wait_queue_head_t *ws_wait;
924 	int *free_ws;
925 
926 	wsm = btrfs_compress_op[type]->workspace_manager;
927 	idle_ws	 = &wsm->idle_ws;
928 	ws_lock	 = &wsm->ws_lock;
929 	total_ws = &wsm->total_ws;
930 	ws_wait	 = &wsm->ws_wait;
931 	free_ws	 = &wsm->free_ws;
932 
933 	spin_lock(ws_lock);
934 	if (*free_ws <= num_online_cpus()) {
935 		list_add(ws, idle_ws);
936 		(*free_ws)++;
937 		spin_unlock(ws_lock);
938 		goto wake;
939 	}
940 	spin_unlock(ws_lock);
941 
942 	free_workspace(type, ws);
943 	atomic_dec(total_ws);
944 wake:
945 	cond_wake_up(ws_wait);
946 }
947 
948 static void put_workspace(int type, struct list_head *ws)
949 {
950 	switch (type) {
951 	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
952 	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
953 	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(type, ws);
954 	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
955 	default:
956 		/*
957 		 * This can't happen, the type is validated several times
958 		 * before we get here.
959 		 */
960 		BUG();
961 	}
962 }
963 
964 /*
965  * Adjust @level according to the limits of the compression algorithm or
966  * fallback to default
967  */
968 static unsigned int btrfs_compress_set_level(int type, unsigned level)
969 {
970 	const struct btrfs_compress_op *ops = btrfs_compress_op[type];
971 
972 	if (level == 0)
973 		level = ops->default_level;
974 	else
975 		level = min(level, ops->max_level);
976 
977 	return level;
978 }
979 
980 /* Wrapper around find_get_page(), with extra error message. */
981 int btrfs_compress_filemap_get_folio(struct address_space *mapping, u64 start,
982 				     struct folio **in_folio_ret)
983 {
984 	struct folio *in_folio;
985 
986 	/*
987 	 * The compressed write path should have the folio locked already, thus
988 	 * we only need to grab one reference.
989 	 */
990 	in_folio = filemap_get_folio(mapping, start >> PAGE_SHIFT);
991 	if (IS_ERR(in_folio)) {
992 		struct btrfs_inode *inode = BTRFS_I(mapping->host);
993 
994 		btrfs_crit(inode->root->fs_info,
995 		"failed to get page cache, root %lld ino %llu file offset %llu",
996 			   btrfs_root_id(inode->root), btrfs_ino(inode), start);
997 		return -ENOENT;
998 	}
999 	*in_folio_ret = in_folio;
1000 	return 0;
1001 }
1002 
1003 /*
1004  * Given an address space and start and length, compress the bytes into @pages
1005  * that are allocated on demand.
1006  *
1007  * @type_level is encoded algorithm and level, where level 0 means whatever
1008  * default the algorithm chooses and is opaque here;
1009  * - compression algo are 0-3
1010  * - the level are bits 4-7
1011  *
1012  * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1013  * and returns number of actually allocated pages
1014  *
1015  * @total_in is used to return the number of bytes actually read.  It
1016  * may be smaller than the input length if we had to exit early because we
1017  * ran out of room in the pages array or because we cross the
1018  * max_out threshold.
1019  *
1020  * @total_out is an in/out parameter, must be set to the input length and will
1021  * be also used to return the total number of compressed bytes
1022  */
1023 int btrfs_compress_folios(unsigned int type_level, struct address_space *mapping,
1024 			 u64 start, struct folio **folios, unsigned long *out_folios,
1025 			 unsigned long *total_in, unsigned long *total_out)
1026 {
1027 	int type = btrfs_compress_type(type_level);
1028 	int level = btrfs_compress_level(type_level);
1029 	struct list_head *workspace;
1030 	int ret;
1031 
1032 	level = btrfs_compress_set_level(type, level);
1033 	workspace = get_workspace(type, level);
1034 	ret = compression_compress_pages(type, workspace, mapping, start, folios,
1035 					 out_folios, total_in, total_out);
1036 	put_workspace(type, workspace);
1037 	return ret;
1038 }
1039 
1040 static int btrfs_decompress_bio(struct compressed_bio *cb)
1041 {
1042 	struct list_head *workspace;
1043 	int ret;
1044 	int type = cb->compress_type;
1045 
1046 	workspace = get_workspace(type, 0);
1047 	ret = compression_decompress_bio(workspace, cb);
1048 	put_workspace(type, workspace);
1049 
1050 	if (!ret)
1051 		zero_fill_bio(&cb->orig_bbio->bio);
1052 	return ret;
1053 }
1054 
1055 /*
1056  * a less complex decompression routine.  Our compressed data fits in a
1057  * single page, and we want to read a single page out of it.
1058  * start_byte tells us the offset into the compressed data we're interested in
1059  */
1060 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1061 		     unsigned long dest_pgoff, size_t srclen, size_t destlen)
1062 {
1063 	struct btrfs_fs_info *fs_info = page_to_fs_info(dest_page);
1064 	struct list_head *workspace;
1065 	const u32 sectorsize = fs_info->sectorsize;
1066 	int ret;
1067 
1068 	/*
1069 	 * The full destination page range should not exceed the page size.
1070 	 * And the @destlen should not exceed sectorsize, as this is only called for
1071 	 * inline file extents, which should not exceed sectorsize.
1072 	 */
1073 	ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);
1074 
1075 	workspace = get_workspace(type, 0);
1076 	ret = compression_decompress(type, workspace, data_in, dest_page,
1077 				     dest_pgoff, srclen, destlen);
1078 	put_workspace(type, workspace);
1079 
1080 	return ret;
1081 }
1082 
1083 int __init btrfs_init_compress(void)
1084 {
1085 	if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1086 			offsetof(struct compressed_bio, bbio.bio),
1087 			BIOSET_NEED_BVECS))
1088 		return -ENOMEM;
1089 
1090 	compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1091 	if (!compr_pool.shrinker)
1092 		return -ENOMEM;
1093 
1094 	btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1095 	btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1096 	btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1097 	zstd_init_workspace_manager();
1098 
1099 	spin_lock_init(&compr_pool.lock);
1100 	INIT_LIST_HEAD(&compr_pool.list);
1101 	compr_pool.count = 0;
1102 	/* 128K / 4K = 32, for 8 threads is 256 pages. */
1103 	compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1104 	compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1105 	compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1106 	compr_pool.shrinker->batch = 32;
1107 	compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1108 	shrinker_register(compr_pool.shrinker);
1109 
1110 	return 0;
1111 }
1112 
1113 void __cold btrfs_exit_compress(void)
1114 {
1115 	/* For now scan drains all pages and does not touch the parameters. */
1116 	btrfs_compr_pool_scan(NULL, NULL);
1117 	shrinker_free(compr_pool.shrinker);
1118 
1119 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1120 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1121 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1122 	zstd_cleanup_workspace_manager();
1123 	bioset_exit(&btrfs_compressed_bioset);
1124 }
1125 
1126 /*
1127  * Copy decompressed data from working buffer to pages.
1128  *
1129  * @buf:		The decompressed data buffer
1130  * @buf_len:		The decompressed data length
1131  * @decompressed:	Number of bytes that are already decompressed inside the
1132  * 			compressed extent
1133  * @cb:			The compressed extent descriptor
1134  * @orig_bio:		The original bio that the caller wants to read for
1135  *
1136  * An easier to understand graph is like below:
1137  *
1138  * 		|<- orig_bio ->|     |<- orig_bio->|
1139  * 	|<-------      full decompressed extent      ----->|
1140  * 	|<-----------    @cb range   ---->|
1141  * 	|			|<-- @buf_len -->|
1142  * 	|<--- @decompressed --->|
1143  *
1144  * Note that, @cb can be a subpage of the full decompressed extent, but
1145  * @cb->start always has the same as the orig_file_offset value of the full
1146  * decompressed extent.
1147  *
1148  * When reading compressed extent, we have to read the full compressed extent,
1149  * while @orig_bio may only want part of the range.
1150  * Thus this function will ensure only data covered by @orig_bio will be copied
1151  * to.
1152  *
1153  * Return 0 if we have copied all needed contents for @orig_bio.
1154  * Return >0 if we need continue decompress.
1155  */
1156 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1157 			      struct compressed_bio *cb, u32 decompressed)
1158 {
1159 	struct bio *orig_bio = &cb->orig_bbio->bio;
1160 	/* Offset inside the full decompressed extent */
1161 	u32 cur_offset;
1162 
1163 	cur_offset = decompressed;
1164 	/* The main loop to do the copy */
1165 	while (cur_offset < decompressed + buf_len) {
1166 		struct bio_vec bvec;
1167 		size_t copy_len;
1168 		u32 copy_start;
1169 		/* Offset inside the full decompressed extent */
1170 		u32 bvec_offset;
1171 
1172 		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1173 		/*
1174 		 * cb->start may underflow, but subtracting that value can still
1175 		 * give us correct offset inside the full decompressed extent.
1176 		 */
1177 		bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1178 
1179 		/* Haven't reached the bvec range, exit */
1180 		if (decompressed + buf_len <= bvec_offset)
1181 			return 1;
1182 
1183 		copy_start = max(cur_offset, bvec_offset);
1184 		copy_len = min(bvec_offset + bvec.bv_len,
1185 			       decompressed + buf_len) - copy_start;
1186 		ASSERT(copy_len);
1187 
1188 		/*
1189 		 * Extra range check to ensure we didn't go beyond
1190 		 * @buf + @buf_len.
1191 		 */
1192 		ASSERT(copy_start - decompressed < buf_len);
1193 		memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1194 			       buf + copy_start - decompressed, copy_len);
1195 		cur_offset += copy_len;
1196 
1197 		bio_advance(orig_bio, copy_len);
1198 		/* Finished the bio */
1199 		if (!orig_bio->bi_iter.bi_size)
1200 			return 0;
1201 	}
1202 	return 1;
1203 }
1204 
1205 /*
1206  * Shannon Entropy calculation
1207  *
1208  * Pure byte distribution analysis fails to determine compressibility of data.
1209  * Try calculating entropy to estimate the average minimum number of bits
1210  * needed to encode the sampled data.
1211  *
1212  * For convenience, return the percentage of needed bits, instead of amount of
1213  * bits directly.
1214  *
1215  * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1216  *			    and can be compressible with high probability
1217  *
1218  * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1219  *
1220  * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1221  */
1222 #define ENTROPY_LVL_ACEPTABLE		(65)
1223 #define ENTROPY_LVL_HIGH		(80)
1224 
1225 /*
1226  * For increasead precision in shannon_entropy calculation,
1227  * let's do pow(n, M) to save more digits after comma:
1228  *
1229  * - maximum int bit length is 64
1230  * - ilog2(MAX_SAMPLE_SIZE)	-> 13
1231  * - 13 * 4 = 52 < 64		-> M = 4
1232  *
1233  * So use pow(n, 4).
1234  */
1235 static inline u32 ilog2_w(u64 n)
1236 {
1237 	return ilog2(n * n * n * n);
1238 }
1239 
1240 static u32 shannon_entropy(struct heuristic_ws *ws)
1241 {
1242 	const u32 entropy_max = 8 * ilog2_w(2);
1243 	u32 entropy_sum = 0;
1244 	u32 p, p_base, sz_base;
1245 	u32 i;
1246 
1247 	sz_base = ilog2_w(ws->sample_size);
1248 	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1249 		p = ws->bucket[i].count;
1250 		p_base = ilog2_w(p);
1251 		entropy_sum += p * (sz_base - p_base);
1252 	}
1253 
1254 	entropy_sum /= ws->sample_size;
1255 	return entropy_sum * 100 / entropy_max;
1256 }
1257 
1258 #define RADIX_BASE		4U
1259 #define COUNTERS_SIZE		(1U << RADIX_BASE)
1260 
1261 static u8 get4bits(u64 num, int shift) {
1262 	u8 low4bits;
1263 
1264 	num >>= shift;
1265 	/* Reverse order */
1266 	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1267 	return low4bits;
1268 }
1269 
1270 /*
1271  * Use 4 bits as radix base
1272  * Use 16 u32 counters for calculating new position in buf array
1273  *
1274  * @array     - array that will be sorted
1275  * @array_buf - buffer array to store sorting results
1276  *              must be equal in size to @array
1277  * @num       - array size
1278  */
1279 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1280 		       int num)
1281 {
1282 	u64 max_num;
1283 	u64 buf_num;
1284 	u32 counters[COUNTERS_SIZE];
1285 	u32 new_addr;
1286 	u32 addr;
1287 	int bitlen;
1288 	int shift;
1289 	int i;
1290 
1291 	/*
1292 	 * Try avoid useless loop iterations for small numbers stored in big
1293 	 * counters.  Example: 48 33 4 ... in 64bit array
1294 	 */
1295 	max_num = array[0].count;
1296 	for (i = 1; i < num; i++) {
1297 		buf_num = array[i].count;
1298 		if (buf_num > max_num)
1299 			max_num = buf_num;
1300 	}
1301 
1302 	buf_num = ilog2(max_num);
1303 	bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1304 
1305 	shift = 0;
1306 	while (shift < bitlen) {
1307 		memset(counters, 0, sizeof(counters));
1308 
1309 		for (i = 0; i < num; i++) {
1310 			buf_num = array[i].count;
1311 			addr = get4bits(buf_num, shift);
1312 			counters[addr]++;
1313 		}
1314 
1315 		for (i = 1; i < COUNTERS_SIZE; i++)
1316 			counters[i] += counters[i - 1];
1317 
1318 		for (i = num - 1; i >= 0; i--) {
1319 			buf_num = array[i].count;
1320 			addr = get4bits(buf_num, shift);
1321 			counters[addr]--;
1322 			new_addr = counters[addr];
1323 			array_buf[new_addr] = array[i];
1324 		}
1325 
1326 		shift += RADIX_BASE;
1327 
1328 		/*
1329 		 * Normal radix expects to move data from a temporary array, to
1330 		 * the main one.  But that requires some CPU time. Avoid that
1331 		 * by doing another sort iteration to original array instead of
1332 		 * memcpy()
1333 		 */
1334 		memset(counters, 0, sizeof(counters));
1335 
1336 		for (i = 0; i < num; i ++) {
1337 			buf_num = array_buf[i].count;
1338 			addr = get4bits(buf_num, shift);
1339 			counters[addr]++;
1340 		}
1341 
1342 		for (i = 1; i < COUNTERS_SIZE; i++)
1343 			counters[i] += counters[i - 1];
1344 
1345 		for (i = num - 1; i >= 0; i--) {
1346 			buf_num = array_buf[i].count;
1347 			addr = get4bits(buf_num, shift);
1348 			counters[addr]--;
1349 			new_addr = counters[addr];
1350 			array[new_addr] = array_buf[i];
1351 		}
1352 
1353 		shift += RADIX_BASE;
1354 	}
1355 }
1356 
1357 /*
1358  * Size of the core byte set - how many bytes cover 90% of the sample
1359  *
1360  * There are several types of structured binary data that use nearly all byte
1361  * values. The distribution can be uniform and counts in all buckets will be
1362  * nearly the same (eg. encrypted data). Unlikely to be compressible.
1363  *
1364  * Other possibility is normal (Gaussian) distribution, where the data could
1365  * be potentially compressible, but we have to take a few more steps to decide
1366  * how much.
1367  *
1368  * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,
1369  *                       compression algo can easy fix that
1370  * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1371  *                       probability is not compressible
1372  */
1373 #define BYTE_CORE_SET_LOW		(64)
1374 #define BYTE_CORE_SET_HIGH		(200)
1375 
1376 static int byte_core_set_size(struct heuristic_ws *ws)
1377 {
1378 	u32 i;
1379 	u32 coreset_sum = 0;
1380 	const u32 core_set_threshold = ws->sample_size * 90 / 100;
1381 	struct bucket_item *bucket = ws->bucket;
1382 
1383 	/* Sort in reverse order */
1384 	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1385 
1386 	for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1387 		coreset_sum += bucket[i].count;
1388 
1389 	if (coreset_sum > core_set_threshold)
1390 		return i;
1391 
1392 	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1393 		coreset_sum += bucket[i].count;
1394 		if (coreset_sum > core_set_threshold)
1395 			break;
1396 	}
1397 
1398 	return i;
1399 }
1400 
1401 /*
1402  * Count byte values in buckets.
1403  * This heuristic can detect textual data (configs, xml, json, html, etc).
1404  * Because in most text-like data byte set is restricted to limited number of
1405  * possible characters, and that restriction in most cases makes data easy to
1406  * compress.
1407  *
1408  * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1409  *	less - compressible
1410  *	more - need additional analysis
1411  */
1412 #define BYTE_SET_THRESHOLD		(64)
1413 
1414 static u32 byte_set_size(const struct heuristic_ws *ws)
1415 {
1416 	u32 i;
1417 	u32 byte_set_size = 0;
1418 
1419 	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1420 		if (ws->bucket[i].count > 0)
1421 			byte_set_size++;
1422 	}
1423 
1424 	/*
1425 	 * Continue collecting count of byte values in buckets.  If the byte
1426 	 * set size is bigger then the threshold, it's pointless to continue,
1427 	 * the detection technique would fail for this type of data.
1428 	 */
1429 	for (; i < BUCKET_SIZE; i++) {
1430 		if (ws->bucket[i].count > 0) {
1431 			byte_set_size++;
1432 			if (byte_set_size > BYTE_SET_THRESHOLD)
1433 				return byte_set_size;
1434 		}
1435 	}
1436 
1437 	return byte_set_size;
1438 }
1439 
1440 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1441 {
1442 	const u32 half_of_sample = ws->sample_size / 2;
1443 	const u8 *data = ws->sample;
1444 
1445 	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1446 }
1447 
1448 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1449 				     struct heuristic_ws *ws)
1450 {
1451 	struct page *page;
1452 	u64 index, index_end;
1453 	u32 i, curr_sample_pos;
1454 	u8 *in_data;
1455 
1456 	/*
1457 	 * Compression handles the input data by chunks of 128KiB
1458 	 * (defined by BTRFS_MAX_UNCOMPRESSED)
1459 	 *
1460 	 * We do the same for the heuristic and loop over the whole range.
1461 	 *
1462 	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1463 	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1464 	 */
1465 	if (end - start > BTRFS_MAX_UNCOMPRESSED)
1466 		end = start + BTRFS_MAX_UNCOMPRESSED;
1467 
1468 	index = start >> PAGE_SHIFT;
1469 	index_end = end >> PAGE_SHIFT;
1470 
1471 	/* Don't miss unaligned end */
1472 	if (!PAGE_ALIGNED(end))
1473 		index_end++;
1474 
1475 	curr_sample_pos = 0;
1476 	while (index < index_end) {
1477 		page = find_get_page(inode->i_mapping, index);
1478 		in_data = kmap_local_page(page);
1479 		/* Handle case where the start is not aligned to PAGE_SIZE */
1480 		i = start % PAGE_SIZE;
1481 		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1482 			/* Don't sample any garbage from the last page */
1483 			if (start > end - SAMPLING_READ_SIZE)
1484 				break;
1485 			memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1486 					SAMPLING_READ_SIZE);
1487 			i += SAMPLING_INTERVAL;
1488 			start += SAMPLING_INTERVAL;
1489 			curr_sample_pos += SAMPLING_READ_SIZE;
1490 		}
1491 		kunmap_local(in_data);
1492 		put_page(page);
1493 
1494 		index++;
1495 	}
1496 
1497 	ws->sample_size = curr_sample_pos;
1498 }
1499 
1500 /*
1501  * Compression heuristic.
1502  *
1503  * The following types of analysis can be performed:
1504  * - detect mostly zero data
1505  * - detect data with low "byte set" size (text, etc)
1506  * - detect data with low/high "core byte" set
1507  *
1508  * Return non-zero if the compression should be done, 0 otherwise.
1509  */
1510 int btrfs_compress_heuristic(struct btrfs_inode *inode, u64 start, u64 end)
1511 {
1512 	struct list_head *ws_list = get_workspace(0, 0);
1513 	struct heuristic_ws *ws;
1514 	u32 i;
1515 	u8 byte;
1516 	int ret = 0;
1517 
1518 	ws = list_entry(ws_list, struct heuristic_ws, list);
1519 
1520 	heuristic_collect_sample(&inode->vfs_inode, start, end, ws);
1521 
1522 	if (sample_repeated_patterns(ws)) {
1523 		ret = 1;
1524 		goto out;
1525 	}
1526 
1527 	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1528 
1529 	for (i = 0; i < ws->sample_size; i++) {
1530 		byte = ws->sample[i];
1531 		ws->bucket[byte].count++;
1532 	}
1533 
1534 	i = byte_set_size(ws);
1535 	if (i < BYTE_SET_THRESHOLD) {
1536 		ret = 2;
1537 		goto out;
1538 	}
1539 
1540 	i = byte_core_set_size(ws);
1541 	if (i <= BYTE_CORE_SET_LOW) {
1542 		ret = 3;
1543 		goto out;
1544 	}
1545 
1546 	if (i >= BYTE_CORE_SET_HIGH) {
1547 		ret = 0;
1548 		goto out;
1549 	}
1550 
1551 	i = shannon_entropy(ws);
1552 	if (i <= ENTROPY_LVL_ACEPTABLE) {
1553 		ret = 4;
1554 		goto out;
1555 	}
1556 
1557 	/*
1558 	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1559 	 * needed to give green light to compression.
1560 	 *
1561 	 * For now just assume that compression at that level is not worth the
1562 	 * resources because:
1563 	 *
1564 	 * 1. it is possible to defrag the data later
1565 	 *
1566 	 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1567 	 * values, every bucket has counter at level ~54. The heuristic would
1568 	 * be confused. This can happen when data have some internal repeated
1569 	 * patterns like "abbacbbc...". This can be detected by analyzing
1570 	 * pairs of bytes, which is too costly.
1571 	 */
1572 	if (i < ENTROPY_LVL_HIGH) {
1573 		ret = 5;
1574 		goto out;
1575 	} else {
1576 		ret = 0;
1577 		goto out;
1578 	}
1579 
1580 out:
1581 	put_workspace(0, ws_list);
1582 	return ret;
1583 }
1584 
1585 /*
1586  * Convert the compression suffix (eg. after "zlib" starting with ":") to
1587  * level, unrecognized string will set the default level
1588  */
1589 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1590 {
1591 	unsigned int level = 0;
1592 	int ret;
1593 
1594 	if (!type)
1595 		return 0;
1596 
1597 	if (str[0] == ':') {
1598 		ret = kstrtouint(str + 1, 10, &level);
1599 		if (ret)
1600 			level = 0;
1601 	}
1602 
1603 	level = btrfs_compress_set_level(type, level);
1604 
1605 	return level;
1606 }
1607