xref: /linux/fs/btrfs/compression.c (revision ae22a94997b8a03dcb3c922857c203246711f9d4)
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, struct page **pages,
94                unsigned long *out_pages, unsigned long *total_in,
95                unsigned long *total_out)
96 {
97 	switch (type) {
98 	case BTRFS_COMPRESS_ZLIB:
99 		return zlib_compress_pages(ws, mapping, start, pages,
100 				out_pages, total_in, total_out);
101 	case BTRFS_COMPRESS_LZO:
102 		return lzo_compress_pages(ws, mapping, start, pages,
103 				out_pages, total_in, total_out);
104 	case BTRFS_COMPRESS_ZSTD:
105 		return zstd_compress_pages(ws, mapping, start, pages,
106 				out_pages, 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_pages = 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_pages(struct compressed_bio *cb)
162 {
163 	for (unsigned int i = 0; i < cb->nr_pages; i++)
164 		btrfs_free_compr_page(cb->compressed_pages[i]);
165 	kfree(cb->compressed_pages);
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 page *btrfs_alloc_compr_page(void)
227 {
228 	struct page *page = NULL;
229 
230 	spin_lock(&compr_pool.lock);
231 	if (compr_pool.count > 0) {
232 		page = list_first_entry(&compr_pool.list, struct page, lru);
233 		list_del_init(&page->lru);
234 		compr_pool.count--;
235 	}
236 	spin_unlock(&compr_pool.lock);
237 
238 	if (page)
239 		return page;
240 
241 	return alloc_page(GFP_NOFS);
242 }
243 
244 void btrfs_free_compr_page(struct page *page)
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(&page->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(page_ref_count(page) == 1);
261 	put_page(page);
262 }
263 
264 static void end_bbio_comprssed_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_pages(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_pages(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_comprssed_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_pages(struct compressed_bio *cb)
346 {
347 	struct bio *bio = &cb->bbio.bio;
348 	u32 offset = 0;
349 
350 	while (offset < cb->compressed_len) {
351 		u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE);
352 
353 		/* Maximum compressed extent is smaller than bio size limit. */
354 		__bio_add_page(bio, cb->compressed_pages[offset >> PAGE_SHIFT],
355 			       len, 0);
356 		offset += len;
357 	}
358 }
359 
360 /*
361  * worker function to build and submit bios for previously compressed pages.
362  * The corresponding pages in the inode should be marked for writeback
363  * and the compressed pages should have a reference on them for dropping
364  * when the IO is complete.
365  *
366  * This also checksums the file bytes and gets things ready for
367  * the end io hooks.
368  */
369 void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered,
370 				   struct page **compressed_pages,
371 				   unsigned int nr_pages,
372 				   blk_opf_t write_flags,
373 				   bool writeback)
374 {
375 	struct btrfs_inode *inode = BTRFS_I(ordered->inode);
376 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
377 	struct compressed_bio *cb;
378 
379 	ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize));
380 	ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize));
381 
382 	cb = alloc_compressed_bio(inode, ordered->file_offset,
383 				  REQ_OP_WRITE | write_flags,
384 				  end_bbio_comprssed_write);
385 	cb->start = ordered->file_offset;
386 	cb->len = ordered->num_bytes;
387 	cb->compressed_pages = compressed_pages;
388 	cb->compressed_len = ordered->disk_num_bytes;
389 	cb->writeback = writeback;
390 	INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work);
391 	cb->nr_pages = nr_pages;
392 	cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT;
393 	cb->bbio.ordered = ordered;
394 	btrfs_add_compressed_bio_pages(cb);
395 
396 	btrfs_submit_bio(&cb->bbio, 0);
397 }
398 
399 /*
400  * Add extra pages in the same compressed file extent so that we don't need to
401  * re-read the same extent again and again.
402  *
403  * NOTE: this won't work well for subpage, as for subpage read, we lock the
404  * full page then submit bio for each compressed/regular extents.
405  *
406  * This means, if we have several sectors in the same page points to the same
407  * on-disk compressed data, we will re-read the same extent many times and
408  * this function can only help for the next page.
409  */
410 static noinline int add_ra_bio_pages(struct inode *inode,
411 				     u64 compressed_end,
412 				     struct compressed_bio *cb,
413 				     int *memstall, unsigned long *pflags)
414 {
415 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
416 	unsigned long end_index;
417 	struct bio *orig_bio = &cb->orig_bbio->bio;
418 	u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size;
419 	u64 isize = i_size_read(inode);
420 	int ret;
421 	struct page *page;
422 	struct extent_map *em;
423 	struct address_space *mapping = inode->i_mapping;
424 	struct extent_map_tree *em_tree;
425 	struct extent_io_tree *tree;
426 	int sectors_missed = 0;
427 
428 	em_tree = &BTRFS_I(inode)->extent_tree;
429 	tree = &BTRFS_I(inode)->io_tree;
430 
431 	if (isize == 0)
432 		return 0;
433 
434 	/*
435 	 * For current subpage support, we only support 64K page size,
436 	 * which means maximum compressed extent size (128K) is just 2x page
437 	 * size.
438 	 * This makes readahead less effective, so here disable readahead for
439 	 * subpage for now, until full compressed write is supported.
440 	 */
441 	if (fs_info->sectorsize < PAGE_SIZE)
442 		return 0;
443 
444 	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
445 
446 	while (cur < compressed_end) {
447 		u64 page_end;
448 		u64 pg_index = cur >> PAGE_SHIFT;
449 		u32 add_size;
450 
451 		if (pg_index > end_index)
452 			break;
453 
454 		page = xa_load(&mapping->i_pages, pg_index);
455 		if (page && !xa_is_value(page)) {
456 			sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >>
457 					  fs_info->sectorsize_bits;
458 
459 			/* Beyond threshold, no need to continue */
460 			if (sectors_missed > 4)
461 				break;
462 
463 			/*
464 			 * Jump to next page start as we already have page for
465 			 * current offset.
466 			 */
467 			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
468 			continue;
469 		}
470 
471 		page = __page_cache_alloc(mapping_gfp_constraint(mapping,
472 								 ~__GFP_FS));
473 		if (!page)
474 			break;
475 
476 		if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
477 			put_page(page);
478 			/* There is already a page, skip to page end */
479 			cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE;
480 			continue;
481 		}
482 
483 		if (!*memstall && PageWorkingset(page)) {
484 			psi_memstall_enter(pflags);
485 			*memstall = 1;
486 		}
487 
488 		ret = set_page_extent_mapped(page);
489 		if (ret < 0) {
490 			unlock_page(page);
491 			put_page(page);
492 			break;
493 		}
494 
495 		page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1;
496 		lock_extent(tree, cur, page_end, NULL);
497 		read_lock(&em_tree->lock);
498 		em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur);
499 		read_unlock(&em_tree->lock);
500 
501 		/*
502 		 * At this point, we have a locked page in the page cache for
503 		 * these bytes in the file.  But, we have to make sure they map
504 		 * to this compressed extent on disk.
505 		 */
506 		if (!em || cur < em->start ||
507 		    (cur + fs_info->sectorsize > extent_map_end(em)) ||
508 		    (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) {
509 			free_extent_map(em);
510 			unlock_extent(tree, cur, page_end, NULL);
511 			unlock_page(page);
512 			put_page(page);
513 			break;
514 		}
515 		free_extent_map(em);
516 
517 		if (page->index == end_index) {
518 			size_t zero_offset = offset_in_page(isize);
519 
520 			if (zero_offset) {
521 				int zeros;
522 				zeros = PAGE_SIZE - zero_offset;
523 				memzero_page(page, zero_offset, zeros);
524 			}
525 		}
526 
527 		add_size = min(em->start + em->len, page_end + 1) - cur;
528 		ret = bio_add_page(orig_bio, page, add_size, offset_in_page(cur));
529 		if (ret != add_size) {
530 			unlock_extent(tree, cur, page_end, NULL);
531 			unlock_page(page);
532 			put_page(page);
533 			break;
534 		}
535 		/*
536 		 * If it's subpage, we also need to increase its
537 		 * subpage::readers number, as at endio we will decrease
538 		 * subpage::readers and to unlock the page.
539 		 */
540 		if (fs_info->sectorsize < PAGE_SIZE)
541 			btrfs_subpage_start_reader(fs_info, page_folio(page),
542 						   cur, add_size);
543 		put_page(page);
544 		cur += add_size;
545 	}
546 	return 0;
547 }
548 
549 /*
550  * for a compressed read, the bio we get passed has all the inode pages
551  * in it.  We don't actually do IO on those pages but allocate new ones
552  * to hold the compressed pages on disk.
553  *
554  * bio->bi_iter.bi_sector points to the compressed extent on disk
555  * bio->bi_io_vec points to all of the inode pages
556  *
557  * After the compressed pages are read, we copy the bytes into the
558  * bio we were passed and then call the bio end_io calls
559  */
560 void btrfs_submit_compressed_read(struct btrfs_bio *bbio)
561 {
562 	struct btrfs_inode *inode = bbio->inode;
563 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
564 	struct extent_map_tree *em_tree = &inode->extent_tree;
565 	struct compressed_bio *cb;
566 	unsigned int compressed_len;
567 	u64 file_offset = bbio->file_offset;
568 	u64 em_len;
569 	u64 em_start;
570 	struct extent_map *em;
571 	unsigned long pflags;
572 	int memstall = 0;
573 	blk_status_t ret;
574 	int ret2;
575 
576 	/* we need the actual starting offset of this extent in the file */
577 	read_lock(&em_tree->lock);
578 	em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
579 	read_unlock(&em_tree->lock);
580 	if (!em) {
581 		ret = BLK_STS_IOERR;
582 		goto out;
583 	}
584 
585 	ASSERT(extent_map_is_compressed(em));
586 	compressed_len = em->block_len;
587 
588 	cb = alloc_compressed_bio(inode, file_offset, REQ_OP_READ,
589 				  end_bbio_comprssed_read);
590 
591 	cb->start = em->orig_start;
592 	em_len = em->len;
593 	em_start = em->start;
594 
595 	cb->len = bbio->bio.bi_iter.bi_size;
596 	cb->compressed_len = compressed_len;
597 	cb->compress_type = extent_map_compression(em);
598 	cb->orig_bbio = bbio;
599 
600 	free_extent_map(em);
601 
602 	cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
603 	cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS);
604 	if (!cb->compressed_pages) {
605 		ret = BLK_STS_RESOURCE;
606 		goto out_free_bio;
607 	}
608 
609 	ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages, 0);
610 	if (ret2) {
611 		ret = BLK_STS_RESOURCE;
612 		goto out_free_compressed_pages;
613 	}
614 
615 	add_ra_bio_pages(&inode->vfs_inode, em_start + em_len, cb, &memstall,
616 			 &pflags);
617 
618 	/* include any pages we added in add_ra-bio_pages */
619 	cb->len = bbio->bio.bi_iter.bi_size;
620 	cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector;
621 	btrfs_add_compressed_bio_pages(cb);
622 
623 	if (memstall)
624 		psi_memstall_leave(&pflags);
625 
626 	btrfs_submit_bio(&cb->bbio, 0);
627 	return;
628 
629 out_free_compressed_pages:
630 	kfree(cb->compressed_pages);
631 out_free_bio:
632 	bio_put(&cb->bbio.bio);
633 out:
634 	btrfs_bio_end_io(bbio, ret);
635 }
636 
637 /*
638  * Heuristic uses systematic sampling to collect data from the input data
639  * range, the logic can be tuned by the following constants:
640  *
641  * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
642  * @SAMPLING_INTERVAL  - range from which the sampled data can be collected
643  */
644 #define SAMPLING_READ_SIZE	(16)
645 #define SAMPLING_INTERVAL	(256)
646 
647 /*
648  * For statistical analysis of the input data we consider bytes that form a
649  * Galois Field of 256 objects. Each object has an attribute count, ie. how
650  * many times the object appeared in the sample.
651  */
652 #define BUCKET_SIZE		(256)
653 
654 /*
655  * The size of the sample is based on a statistical sampling rule of thumb.
656  * The common way is to perform sampling tests as long as the number of
657  * elements in each cell is at least 5.
658  *
659  * Instead of 5, we choose 32 to obtain more accurate results.
660  * If the data contain the maximum number of symbols, which is 256, we obtain a
661  * sample size bound by 8192.
662  *
663  * For a sample of at most 8KB of data per data range: 16 consecutive bytes
664  * from up to 512 locations.
665  */
666 #define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\
667 				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
668 
669 struct bucket_item {
670 	u32 count;
671 };
672 
673 struct heuristic_ws {
674 	/* Partial copy of input data */
675 	u8 *sample;
676 	u32 sample_size;
677 	/* Buckets store counters for each byte value */
678 	struct bucket_item *bucket;
679 	/* Sorting buffer */
680 	struct bucket_item *bucket_b;
681 	struct list_head list;
682 };
683 
684 static struct workspace_manager heuristic_wsm;
685 
686 static void free_heuristic_ws(struct list_head *ws)
687 {
688 	struct heuristic_ws *workspace;
689 
690 	workspace = list_entry(ws, struct heuristic_ws, list);
691 
692 	kvfree(workspace->sample);
693 	kfree(workspace->bucket);
694 	kfree(workspace->bucket_b);
695 	kfree(workspace);
696 }
697 
698 static struct list_head *alloc_heuristic_ws(unsigned int level)
699 {
700 	struct heuristic_ws *ws;
701 
702 	ws = kzalloc(sizeof(*ws), GFP_KERNEL);
703 	if (!ws)
704 		return ERR_PTR(-ENOMEM);
705 
706 	ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
707 	if (!ws->sample)
708 		goto fail;
709 
710 	ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
711 	if (!ws->bucket)
712 		goto fail;
713 
714 	ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
715 	if (!ws->bucket_b)
716 		goto fail;
717 
718 	INIT_LIST_HEAD(&ws->list);
719 	return &ws->list;
720 fail:
721 	free_heuristic_ws(&ws->list);
722 	return ERR_PTR(-ENOMEM);
723 }
724 
725 const struct btrfs_compress_op btrfs_heuristic_compress = {
726 	.workspace_manager = &heuristic_wsm,
727 };
728 
729 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
730 	/* The heuristic is represented as compression type 0 */
731 	&btrfs_heuristic_compress,
732 	&btrfs_zlib_compress,
733 	&btrfs_lzo_compress,
734 	&btrfs_zstd_compress,
735 };
736 
737 static struct list_head *alloc_workspace(int type, unsigned int level)
738 {
739 	switch (type) {
740 	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
741 	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
742 	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(level);
743 	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
744 	default:
745 		/*
746 		 * This can't happen, the type is validated several times
747 		 * before we get here.
748 		 */
749 		BUG();
750 	}
751 }
752 
753 static void free_workspace(int type, struct list_head *ws)
754 {
755 	switch (type) {
756 	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
757 	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
758 	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);
759 	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
760 	default:
761 		/*
762 		 * This can't happen, the type is validated several times
763 		 * before we get here.
764 		 */
765 		BUG();
766 	}
767 }
768 
769 static void btrfs_init_workspace_manager(int type)
770 {
771 	struct workspace_manager *wsm;
772 	struct list_head *workspace;
773 
774 	wsm = btrfs_compress_op[type]->workspace_manager;
775 	INIT_LIST_HEAD(&wsm->idle_ws);
776 	spin_lock_init(&wsm->ws_lock);
777 	atomic_set(&wsm->total_ws, 0);
778 	init_waitqueue_head(&wsm->ws_wait);
779 
780 	/*
781 	 * Preallocate one workspace for each compression type so we can
782 	 * guarantee forward progress in the worst case
783 	 */
784 	workspace = alloc_workspace(type, 0);
785 	if (IS_ERR(workspace)) {
786 		pr_warn(
787 	"BTRFS: cannot preallocate compression workspace, will try later\n");
788 	} else {
789 		atomic_set(&wsm->total_ws, 1);
790 		wsm->free_ws = 1;
791 		list_add(workspace, &wsm->idle_ws);
792 	}
793 }
794 
795 static void btrfs_cleanup_workspace_manager(int type)
796 {
797 	struct workspace_manager *wsman;
798 	struct list_head *ws;
799 
800 	wsman = btrfs_compress_op[type]->workspace_manager;
801 	while (!list_empty(&wsman->idle_ws)) {
802 		ws = wsman->idle_ws.next;
803 		list_del(ws);
804 		free_workspace(type, ws);
805 		atomic_dec(&wsman->total_ws);
806 	}
807 }
808 
809 /*
810  * This finds an available workspace or allocates a new one.
811  * If it's not possible to allocate a new one, waits until there's one.
812  * Preallocation makes a forward progress guarantees and we do not return
813  * errors.
814  */
815 struct list_head *btrfs_get_workspace(int type, unsigned int level)
816 {
817 	struct workspace_manager *wsm;
818 	struct list_head *workspace;
819 	int cpus = num_online_cpus();
820 	unsigned nofs_flag;
821 	struct list_head *idle_ws;
822 	spinlock_t *ws_lock;
823 	atomic_t *total_ws;
824 	wait_queue_head_t *ws_wait;
825 	int *free_ws;
826 
827 	wsm = btrfs_compress_op[type]->workspace_manager;
828 	idle_ws	 = &wsm->idle_ws;
829 	ws_lock	 = &wsm->ws_lock;
830 	total_ws = &wsm->total_ws;
831 	ws_wait	 = &wsm->ws_wait;
832 	free_ws	 = &wsm->free_ws;
833 
834 again:
835 	spin_lock(ws_lock);
836 	if (!list_empty(idle_ws)) {
837 		workspace = idle_ws->next;
838 		list_del(workspace);
839 		(*free_ws)--;
840 		spin_unlock(ws_lock);
841 		return workspace;
842 
843 	}
844 	if (atomic_read(total_ws) > cpus) {
845 		DEFINE_WAIT(wait);
846 
847 		spin_unlock(ws_lock);
848 		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
849 		if (atomic_read(total_ws) > cpus && !*free_ws)
850 			schedule();
851 		finish_wait(ws_wait, &wait);
852 		goto again;
853 	}
854 	atomic_inc(total_ws);
855 	spin_unlock(ws_lock);
856 
857 	/*
858 	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
859 	 * to turn it off here because we might get called from the restricted
860 	 * context of btrfs_compress_bio/btrfs_compress_pages
861 	 */
862 	nofs_flag = memalloc_nofs_save();
863 	workspace = alloc_workspace(type, level);
864 	memalloc_nofs_restore(nofs_flag);
865 
866 	if (IS_ERR(workspace)) {
867 		atomic_dec(total_ws);
868 		wake_up(ws_wait);
869 
870 		/*
871 		 * Do not return the error but go back to waiting. There's a
872 		 * workspace preallocated for each type and the compression
873 		 * time is bounded so we get to a workspace eventually. This
874 		 * makes our caller's life easier.
875 		 *
876 		 * To prevent silent and low-probability deadlocks (when the
877 		 * initial preallocation fails), check if there are any
878 		 * workspaces at all.
879 		 */
880 		if (atomic_read(total_ws) == 0) {
881 			static DEFINE_RATELIMIT_STATE(_rs,
882 					/* once per minute */ 60 * HZ,
883 					/* no burst */ 1);
884 
885 			if (__ratelimit(&_rs)) {
886 				pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
887 			}
888 		}
889 		goto again;
890 	}
891 	return workspace;
892 }
893 
894 static struct list_head *get_workspace(int type, int level)
895 {
896 	switch (type) {
897 	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
898 	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
899 	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(type, level);
900 	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
901 	default:
902 		/*
903 		 * This can't happen, the type is validated several times
904 		 * before we get here.
905 		 */
906 		BUG();
907 	}
908 }
909 
910 /*
911  * put a workspace struct back on the list or free it if we have enough
912  * idle ones sitting around
913  */
914 void btrfs_put_workspace(int type, struct list_head *ws)
915 {
916 	struct workspace_manager *wsm;
917 	struct list_head *idle_ws;
918 	spinlock_t *ws_lock;
919 	atomic_t *total_ws;
920 	wait_queue_head_t *ws_wait;
921 	int *free_ws;
922 
923 	wsm = btrfs_compress_op[type]->workspace_manager;
924 	idle_ws	 = &wsm->idle_ws;
925 	ws_lock	 = &wsm->ws_lock;
926 	total_ws = &wsm->total_ws;
927 	ws_wait	 = &wsm->ws_wait;
928 	free_ws	 = &wsm->free_ws;
929 
930 	spin_lock(ws_lock);
931 	if (*free_ws <= num_online_cpus()) {
932 		list_add(ws, idle_ws);
933 		(*free_ws)++;
934 		spin_unlock(ws_lock);
935 		goto wake;
936 	}
937 	spin_unlock(ws_lock);
938 
939 	free_workspace(type, ws);
940 	atomic_dec(total_ws);
941 wake:
942 	cond_wake_up(ws_wait);
943 }
944 
945 static void put_workspace(int type, struct list_head *ws)
946 {
947 	switch (type) {
948 	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
949 	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
950 	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(type, ws);
951 	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
952 	default:
953 		/*
954 		 * This can't happen, the type is validated several times
955 		 * before we get here.
956 		 */
957 		BUG();
958 	}
959 }
960 
961 /*
962  * Adjust @level according to the limits of the compression algorithm or
963  * fallback to default
964  */
965 static unsigned int btrfs_compress_set_level(int type, unsigned level)
966 {
967 	const struct btrfs_compress_op *ops = btrfs_compress_op[type];
968 
969 	if (level == 0)
970 		level = ops->default_level;
971 	else
972 		level = min(level, ops->max_level);
973 
974 	return level;
975 }
976 
977 /*
978  * Given an address space and start and length, compress the bytes into @pages
979  * that are allocated on demand.
980  *
981  * @type_level is encoded algorithm and level, where level 0 means whatever
982  * default the algorithm chooses and is opaque here;
983  * - compression algo are 0-3
984  * - the level are bits 4-7
985  *
986  * @out_pages is an in/out parameter, holds maximum number of pages to allocate
987  * and returns number of actually allocated pages
988  *
989  * @total_in is used to return the number of bytes actually read.  It
990  * may be smaller than the input length if we had to exit early because we
991  * ran out of room in the pages array or because we cross the
992  * max_out threshold.
993  *
994  * @total_out is an in/out parameter, must be set to the input length and will
995  * be also used to return the total number of compressed bytes
996  */
997 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
998 			 u64 start, struct page **pages,
999 			 unsigned long *out_pages,
1000 			 unsigned long *total_in,
1001 			 unsigned long *total_out)
1002 {
1003 	int type = btrfs_compress_type(type_level);
1004 	int level = btrfs_compress_level(type_level);
1005 	struct list_head *workspace;
1006 	int ret;
1007 
1008 	level = btrfs_compress_set_level(type, level);
1009 	workspace = get_workspace(type, level);
1010 	ret = compression_compress_pages(type, workspace, mapping, start, pages,
1011 					 out_pages, total_in, total_out);
1012 	put_workspace(type, workspace);
1013 	return ret;
1014 }
1015 
1016 static int btrfs_decompress_bio(struct compressed_bio *cb)
1017 {
1018 	struct list_head *workspace;
1019 	int ret;
1020 	int type = cb->compress_type;
1021 
1022 	workspace = get_workspace(type, 0);
1023 	ret = compression_decompress_bio(workspace, cb);
1024 	put_workspace(type, workspace);
1025 
1026 	if (!ret)
1027 		zero_fill_bio(&cb->orig_bbio->bio);
1028 	return ret;
1029 }
1030 
1031 /*
1032  * a less complex decompression routine.  Our compressed data fits in a
1033  * single page, and we want to read a single page out of it.
1034  * start_byte tells us the offset into the compressed data we're interested in
1035  */
1036 int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page,
1037 		     unsigned long dest_pgoff, size_t srclen, size_t destlen)
1038 {
1039 	struct btrfs_fs_info *fs_info = page_to_fs_info(dest_page);
1040 	struct list_head *workspace;
1041 	const u32 sectorsize = fs_info->sectorsize;
1042 	int ret;
1043 
1044 	/*
1045 	 * The full destination page range should not exceed the page size.
1046 	 * And the @destlen should not exceed sectorsize, as this is only called for
1047 	 * inline file extents, which should not exceed sectorsize.
1048 	 */
1049 	ASSERT(dest_pgoff + destlen <= PAGE_SIZE && destlen <= sectorsize);
1050 
1051 	workspace = get_workspace(type, 0);
1052 	ret = compression_decompress(type, workspace, data_in, dest_page,
1053 				     dest_pgoff, srclen, destlen);
1054 	put_workspace(type, workspace);
1055 
1056 	return ret;
1057 }
1058 
1059 int __init btrfs_init_compress(void)
1060 {
1061 	if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE,
1062 			offsetof(struct compressed_bio, bbio.bio),
1063 			BIOSET_NEED_BVECS))
1064 		return -ENOMEM;
1065 
1066 	compr_pool.shrinker = shrinker_alloc(SHRINKER_NONSLAB, "btrfs-compr-pages");
1067 	if (!compr_pool.shrinker)
1068 		return -ENOMEM;
1069 
1070 	btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1071 	btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1072 	btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1073 	zstd_init_workspace_manager();
1074 
1075 	spin_lock_init(&compr_pool.lock);
1076 	INIT_LIST_HEAD(&compr_pool.list);
1077 	compr_pool.count = 0;
1078 	/* 128K / 4K = 32, for 8 threads is 256 pages. */
1079 	compr_pool.thresh = BTRFS_MAX_COMPRESSED / PAGE_SIZE * 8;
1080 	compr_pool.shrinker->count_objects = btrfs_compr_pool_count;
1081 	compr_pool.shrinker->scan_objects = btrfs_compr_pool_scan;
1082 	compr_pool.shrinker->batch = 32;
1083 	compr_pool.shrinker->seeks = DEFAULT_SEEKS;
1084 	shrinker_register(compr_pool.shrinker);
1085 
1086 	return 0;
1087 }
1088 
1089 void __cold btrfs_exit_compress(void)
1090 {
1091 	/* For now scan drains all pages and does not touch the parameters. */
1092 	btrfs_compr_pool_scan(NULL, NULL);
1093 	shrinker_free(compr_pool.shrinker);
1094 
1095 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1096 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1097 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1098 	zstd_cleanup_workspace_manager();
1099 	bioset_exit(&btrfs_compressed_bioset);
1100 }
1101 
1102 /*
1103  * Copy decompressed data from working buffer to pages.
1104  *
1105  * @buf:		The decompressed data buffer
1106  * @buf_len:		The decompressed data length
1107  * @decompressed:	Number of bytes that are already decompressed inside the
1108  * 			compressed extent
1109  * @cb:			The compressed extent descriptor
1110  * @orig_bio:		The original bio that the caller wants to read for
1111  *
1112  * An easier to understand graph is like below:
1113  *
1114  * 		|<- orig_bio ->|     |<- orig_bio->|
1115  * 	|<-------      full decompressed extent      ----->|
1116  * 	|<-----------    @cb range   ---->|
1117  * 	|			|<-- @buf_len -->|
1118  * 	|<--- @decompressed --->|
1119  *
1120  * Note that, @cb can be a subpage of the full decompressed extent, but
1121  * @cb->start always has the same as the orig_file_offset value of the full
1122  * decompressed extent.
1123  *
1124  * When reading compressed extent, we have to read the full compressed extent,
1125  * while @orig_bio may only want part of the range.
1126  * Thus this function will ensure only data covered by @orig_bio will be copied
1127  * to.
1128  *
1129  * Return 0 if we have copied all needed contents for @orig_bio.
1130  * Return >0 if we need continue decompress.
1131  */
1132 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1133 			      struct compressed_bio *cb, u32 decompressed)
1134 {
1135 	struct bio *orig_bio = &cb->orig_bbio->bio;
1136 	/* Offset inside the full decompressed extent */
1137 	u32 cur_offset;
1138 
1139 	cur_offset = decompressed;
1140 	/* The main loop to do the copy */
1141 	while (cur_offset < decompressed + buf_len) {
1142 		struct bio_vec bvec;
1143 		size_t copy_len;
1144 		u32 copy_start;
1145 		/* Offset inside the full decompressed extent */
1146 		u32 bvec_offset;
1147 
1148 		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1149 		/*
1150 		 * cb->start may underflow, but subtracting that value can still
1151 		 * give us correct offset inside the full decompressed extent.
1152 		 */
1153 		bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1154 
1155 		/* Haven't reached the bvec range, exit */
1156 		if (decompressed + buf_len <= bvec_offset)
1157 			return 1;
1158 
1159 		copy_start = max(cur_offset, bvec_offset);
1160 		copy_len = min(bvec_offset + bvec.bv_len,
1161 			       decompressed + buf_len) - copy_start;
1162 		ASSERT(copy_len);
1163 
1164 		/*
1165 		 * Extra range check to ensure we didn't go beyond
1166 		 * @buf + @buf_len.
1167 		 */
1168 		ASSERT(copy_start - decompressed < buf_len);
1169 		memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1170 			       buf + copy_start - decompressed, copy_len);
1171 		cur_offset += copy_len;
1172 
1173 		bio_advance(orig_bio, copy_len);
1174 		/* Finished the bio */
1175 		if (!orig_bio->bi_iter.bi_size)
1176 			return 0;
1177 	}
1178 	return 1;
1179 }
1180 
1181 /*
1182  * Shannon Entropy calculation
1183  *
1184  * Pure byte distribution analysis fails to determine compressibility of data.
1185  * Try calculating entropy to estimate the average minimum number of bits
1186  * needed to encode the sampled data.
1187  *
1188  * For convenience, return the percentage of needed bits, instead of amount of
1189  * bits directly.
1190  *
1191  * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1192  *			    and can be compressible with high probability
1193  *
1194  * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1195  *
1196  * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1197  */
1198 #define ENTROPY_LVL_ACEPTABLE		(65)
1199 #define ENTROPY_LVL_HIGH		(80)
1200 
1201 /*
1202  * For increasead precision in shannon_entropy calculation,
1203  * let's do pow(n, M) to save more digits after comma:
1204  *
1205  * - maximum int bit length is 64
1206  * - ilog2(MAX_SAMPLE_SIZE)	-> 13
1207  * - 13 * 4 = 52 < 64		-> M = 4
1208  *
1209  * So use pow(n, 4).
1210  */
1211 static inline u32 ilog2_w(u64 n)
1212 {
1213 	return ilog2(n * n * n * n);
1214 }
1215 
1216 static u32 shannon_entropy(struct heuristic_ws *ws)
1217 {
1218 	const u32 entropy_max = 8 * ilog2_w(2);
1219 	u32 entropy_sum = 0;
1220 	u32 p, p_base, sz_base;
1221 	u32 i;
1222 
1223 	sz_base = ilog2_w(ws->sample_size);
1224 	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1225 		p = ws->bucket[i].count;
1226 		p_base = ilog2_w(p);
1227 		entropy_sum += p * (sz_base - p_base);
1228 	}
1229 
1230 	entropy_sum /= ws->sample_size;
1231 	return entropy_sum * 100 / entropy_max;
1232 }
1233 
1234 #define RADIX_BASE		4U
1235 #define COUNTERS_SIZE		(1U << RADIX_BASE)
1236 
1237 static u8 get4bits(u64 num, int shift) {
1238 	u8 low4bits;
1239 
1240 	num >>= shift;
1241 	/* Reverse order */
1242 	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1243 	return low4bits;
1244 }
1245 
1246 /*
1247  * Use 4 bits as radix base
1248  * Use 16 u32 counters for calculating new position in buf array
1249  *
1250  * @array     - array that will be sorted
1251  * @array_buf - buffer array to store sorting results
1252  *              must be equal in size to @array
1253  * @num       - array size
1254  */
1255 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1256 		       int num)
1257 {
1258 	u64 max_num;
1259 	u64 buf_num;
1260 	u32 counters[COUNTERS_SIZE];
1261 	u32 new_addr;
1262 	u32 addr;
1263 	int bitlen;
1264 	int shift;
1265 	int i;
1266 
1267 	/*
1268 	 * Try avoid useless loop iterations for small numbers stored in big
1269 	 * counters.  Example: 48 33 4 ... in 64bit array
1270 	 */
1271 	max_num = array[0].count;
1272 	for (i = 1; i < num; i++) {
1273 		buf_num = array[i].count;
1274 		if (buf_num > max_num)
1275 			max_num = buf_num;
1276 	}
1277 
1278 	buf_num = ilog2(max_num);
1279 	bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1280 
1281 	shift = 0;
1282 	while (shift < bitlen) {
1283 		memset(counters, 0, sizeof(counters));
1284 
1285 		for (i = 0; i < num; i++) {
1286 			buf_num = array[i].count;
1287 			addr = get4bits(buf_num, shift);
1288 			counters[addr]++;
1289 		}
1290 
1291 		for (i = 1; i < COUNTERS_SIZE; i++)
1292 			counters[i] += counters[i - 1];
1293 
1294 		for (i = num - 1; i >= 0; i--) {
1295 			buf_num = array[i].count;
1296 			addr = get4bits(buf_num, shift);
1297 			counters[addr]--;
1298 			new_addr = counters[addr];
1299 			array_buf[new_addr] = array[i];
1300 		}
1301 
1302 		shift += RADIX_BASE;
1303 
1304 		/*
1305 		 * Normal radix expects to move data from a temporary array, to
1306 		 * the main one.  But that requires some CPU time. Avoid that
1307 		 * by doing another sort iteration to original array instead of
1308 		 * memcpy()
1309 		 */
1310 		memset(counters, 0, sizeof(counters));
1311 
1312 		for (i = 0; i < num; i ++) {
1313 			buf_num = array_buf[i].count;
1314 			addr = get4bits(buf_num, shift);
1315 			counters[addr]++;
1316 		}
1317 
1318 		for (i = 1; i < COUNTERS_SIZE; i++)
1319 			counters[i] += counters[i - 1];
1320 
1321 		for (i = num - 1; i >= 0; i--) {
1322 			buf_num = array_buf[i].count;
1323 			addr = get4bits(buf_num, shift);
1324 			counters[addr]--;
1325 			new_addr = counters[addr];
1326 			array[new_addr] = array_buf[i];
1327 		}
1328 
1329 		shift += RADIX_BASE;
1330 	}
1331 }
1332 
1333 /*
1334  * Size of the core byte set - how many bytes cover 90% of the sample
1335  *
1336  * There are several types of structured binary data that use nearly all byte
1337  * values. The distribution can be uniform and counts in all buckets will be
1338  * nearly the same (eg. encrypted data). Unlikely to be compressible.
1339  *
1340  * Other possibility is normal (Gaussian) distribution, where the data could
1341  * be potentially compressible, but we have to take a few more steps to decide
1342  * how much.
1343  *
1344  * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,
1345  *                       compression algo can easy fix that
1346  * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1347  *                       probability is not compressible
1348  */
1349 #define BYTE_CORE_SET_LOW		(64)
1350 #define BYTE_CORE_SET_HIGH		(200)
1351 
1352 static int byte_core_set_size(struct heuristic_ws *ws)
1353 {
1354 	u32 i;
1355 	u32 coreset_sum = 0;
1356 	const u32 core_set_threshold = ws->sample_size * 90 / 100;
1357 	struct bucket_item *bucket = ws->bucket;
1358 
1359 	/* Sort in reverse order */
1360 	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1361 
1362 	for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1363 		coreset_sum += bucket[i].count;
1364 
1365 	if (coreset_sum > core_set_threshold)
1366 		return i;
1367 
1368 	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1369 		coreset_sum += bucket[i].count;
1370 		if (coreset_sum > core_set_threshold)
1371 			break;
1372 	}
1373 
1374 	return i;
1375 }
1376 
1377 /*
1378  * Count byte values in buckets.
1379  * This heuristic can detect textual data (configs, xml, json, html, etc).
1380  * Because in most text-like data byte set is restricted to limited number of
1381  * possible characters, and that restriction in most cases makes data easy to
1382  * compress.
1383  *
1384  * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1385  *	less - compressible
1386  *	more - need additional analysis
1387  */
1388 #define BYTE_SET_THRESHOLD		(64)
1389 
1390 static u32 byte_set_size(const struct heuristic_ws *ws)
1391 {
1392 	u32 i;
1393 	u32 byte_set_size = 0;
1394 
1395 	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1396 		if (ws->bucket[i].count > 0)
1397 			byte_set_size++;
1398 	}
1399 
1400 	/*
1401 	 * Continue collecting count of byte values in buckets.  If the byte
1402 	 * set size is bigger then the threshold, it's pointless to continue,
1403 	 * the detection technique would fail for this type of data.
1404 	 */
1405 	for (; i < BUCKET_SIZE; i++) {
1406 		if (ws->bucket[i].count > 0) {
1407 			byte_set_size++;
1408 			if (byte_set_size > BYTE_SET_THRESHOLD)
1409 				return byte_set_size;
1410 		}
1411 	}
1412 
1413 	return byte_set_size;
1414 }
1415 
1416 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1417 {
1418 	const u32 half_of_sample = ws->sample_size / 2;
1419 	const u8 *data = ws->sample;
1420 
1421 	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1422 }
1423 
1424 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1425 				     struct heuristic_ws *ws)
1426 {
1427 	struct page *page;
1428 	u64 index, index_end;
1429 	u32 i, curr_sample_pos;
1430 	u8 *in_data;
1431 
1432 	/*
1433 	 * Compression handles the input data by chunks of 128KiB
1434 	 * (defined by BTRFS_MAX_UNCOMPRESSED)
1435 	 *
1436 	 * We do the same for the heuristic and loop over the whole range.
1437 	 *
1438 	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1439 	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1440 	 */
1441 	if (end - start > BTRFS_MAX_UNCOMPRESSED)
1442 		end = start + BTRFS_MAX_UNCOMPRESSED;
1443 
1444 	index = start >> PAGE_SHIFT;
1445 	index_end = end >> PAGE_SHIFT;
1446 
1447 	/* Don't miss unaligned end */
1448 	if (!PAGE_ALIGNED(end))
1449 		index_end++;
1450 
1451 	curr_sample_pos = 0;
1452 	while (index < index_end) {
1453 		page = find_get_page(inode->i_mapping, index);
1454 		in_data = kmap_local_page(page);
1455 		/* Handle case where the start is not aligned to PAGE_SIZE */
1456 		i = start % PAGE_SIZE;
1457 		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1458 			/* Don't sample any garbage from the last page */
1459 			if (start > end - SAMPLING_READ_SIZE)
1460 				break;
1461 			memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1462 					SAMPLING_READ_SIZE);
1463 			i += SAMPLING_INTERVAL;
1464 			start += SAMPLING_INTERVAL;
1465 			curr_sample_pos += SAMPLING_READ_SIZE;
1466 		}
1467 		kunmap_local(in_data);
1468 		put_page(page);
1469 
1470 		index++;
1471 	}
1472 
1473 	ws->sample_size = curr_sample_pos;
1474 }
1475 
1476 /*
1477  * Compression heuristic.
1478  *
1479  * The following types of analysis can be performed:
1480  * - detect mostly zero data
1481  * - detect data with low "byte set" size (text, etc)
1482  * - detect data with low/high "core byte" set
1483  *
1484  * Return non-zero if the compression should be done, 0 otherwise.
1485  */
1486 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1487 {
1488 	struct list_head *ws_list = get_workspace(0, 0);
1489 	struct heuristic_ws *ws;
1490 	u32 i;
1491 	u8 byte;
1492 	int ret = 0;
1493 
1494 	ws = list_entry(ws_list, struct heuristic_ws, list);
1495 
1496 	heuristic_collect_sample(inode, start, end, ws);
1497 
1498 	if (sample_repeated_patterns(ws)) {
1499 		ret = 1;
1500 		goto out;
1501 	}
1502 
1503 	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1504 
1505 	for (i = 0; i < ws->sample_size; i++) {
1506 		byte = ws->sample[i];
1507 		ws->bucket[byte].count++;
1508 	}
1509 
1510 	i = byte_set_size(ws);
1511 	if (i < BYTE_SET_THRESHOLD) {
1512 		ret = 2;
1513 		goto out;
1514 	}
1515 
1516 	i = byte_core_set_size(ws);
1517 	if (i <= BYTE_CORE_SET_LOW) {
1518 		ret = 3;
1519 		goto out;
1520 	}
1521 
1522 	if (i >= BYTE_CORE_SET_HIGH) {
1523 		ret = 0;
1524 		goto out;
1525 	}
1526 
1527 	i = shannon_entropy(ws);
1528 	if (i <= ENTROPY_LVL_ACEPTABLE) {
1529 		ret = 4;
1530 		goto out;
1531 	}
1532 
1533 	/*
1534 	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1535 	 * needed to give green light to compression.
1536 	 *
1537 	 * For now just assume that compression at that level is not worth the
1538 	 * resources because:
1539 	 *
1540 	 * 1. it is possible to defrag the data later
1541 	 *
1542 	 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1543 	 * values, every bucket has counter at level ~54. The heuristic would
1544 	 * be confused. This can happen when data have some internal repeated
1545 	 * patterns like "abbacbbc...". This can be detected by analyzing
1546 	 * pairs of bytes, which is too costly.
1547 	 */
1548 	if (i < ENTROPY_LVL_HIGH) {
1549 		ret = 5;
1550 		goto out;
1551 	} else {
1552 		ret = 0;
1553 		goto out;
1554 	}
1555 
1556 out:
1557 	put_workspace(0, ws_list);
1558 	return ret;
1559 }
1560 
1561 /*
1562  * Convert the compression suffix (eg. after "zlib" starting with ":") to
1563  * level, unrecognized string will set the default level
1564  */
1565 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1566 {
1567 	unsigned int level = 0;
1568 	int ret;
1569 
1570 	if (!type)
1571 		return 0;
1572 
1573 	if (str[0] == ':') {
1574 		ret = kstrtouint(str + 1, 10, &level);
1575 		if (ret)
1576 			level = 0;
1577 	}
1578 
1579 	level = btrfs_compress_set_level(type, level);
1580 
1581 	return level;
1582 }
1583