xref: /openbmc/linux/fs/btrfs/compression.c (revision 64f45351)
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/highmem.h>
12 #include <linux/time.h>
13 #include <linux/init.h>
14 #include <linux/string.h>
15 #include <linux/backing-dev.h>
16 #include <linux/writeback.h>
17 #include <linux/slab.h>
18 #include <linux/sched/mm.h>
19 #include <linux/log2.h>
20 #include <crypto/hash.h>
21 #include "misc.h"
22 #include "ctree.h"
23 #include "disk-io.h"
24 #include "transaction.h"
25 #include "btrfs_inode.h"
26 #include "volumes.h"
27 #include "ordered-data.h"
28 #include "compression.h"
29 #include "extent_io.h"
30 #include "extent_map.h"
31 #include "zoned.h"
32 
33 static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" };
34 
35 const char* btrfs_compress_type2str(enum btrfs_compression_type type)
36 {
37 	switch (type) {
38 	case BTRFS_COMPRESS_ZLIB:
39 	case BTRFS_COMPRESS_LZO:
40 	case BTRFS_COMPRESS_ZSTD:
41 	case BTRFS_COMPRESS_NONE:
42 		return btrfs_compress_types[type];
43 	default:
44 		break;
45 	}
46 
47 	return NULL;
48 }
49 
50 bool btrfs_compress_is_valid_type(const char *str, size_t len)
51 {
52 	int i;
53 
54 	for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) {
55 		size_t comp_len = strlen(btrfs_compress_types[i]);
56 
57 		if (len < comp_len)
58 			continue;
59 
60 		if (!strncmp(btrfs_compress_types[i], str, comp_len))
61 			return true;
62 	}
63 	return false;
64 }
65 
66 static int compression_compress_pages(int type, struct list_head *ws,
67                struct address_space *mapping, u64 start, struct page **pages,
68                unsigned long *out_pages, unsigned long *total_in,
69                unsigned long *total_out)
70 {
71 	switch (type) {
72 	case BTRFS_COMPRESS_ZLIB:
73 		return zlib_compress_pages(ws, mapping, start, pages,
74 				out_pages, total_in, total_out);
75 	case BTRFS_COMPRESS_LZO:
76 		return lzo_compress_pages(ws, mapping, start, pages,
77 				out_pages, total_in, total_out);
78 	case BTRFS_COMPRESS_ZSTD:
79 		return zstd_compress_pages(ws, mapping, start, pages,
80 				out_pages, total_in, total_out);
81 	case BTRFS_COMPRESS_NONE:
82 	default:
83 		/*
84 		 * This can happen when compression races with remount setting
85 		 * it to 'no compress', while caller doesn't call
86 		 * inode_need_compress() to check if we really need to
87 		 * compress.
88 		 *
89 		 * Not a big deal, just need to inform caller that we
90 		 * haven't allocated any pages yet.
91 		 */
92 		*out_pages = 0;
93 		return -E2BIG;
94 	}
95 }
96 
97 static int compression_decompress_bio(int type, struct list_head *ws,
98 		struct compressed_bio *cb)
99 {
100 	switch (type) {
101 	case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb);
102 	case BTRFS_COMPRESS_LZO:  return lzo_decompress_bio(ws, cb);
103 	case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb);
104 	case BTRFS_COMPRESS_NONE:
105 	default:
106 		/*
107 		 * This can't happen, the type is validated several times
108 		 * before we get here.
109 		 */
110 		BUG();
111 	}
112 }
113 
114 static int compression_decompress(int type, struct list_head *ws,
115                unsigned char *data_in, struct page *dest_page,
116                unsigned long start_byte, size_t srclen, size_t destlen)
117 {
118 	switch (type) {
119 	case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page,
120 						start_byte, srclen, destlen);
121 	case BTRFS_COMPRESS_LZO:  return lzo_decompress(ws, data_in, dest_page,
122 						start_byte, srclen, destlen);
123 	case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page,
124 						start_byte, srclen, destlen);
125 	case BTRFS_COMPRESS_NONE:
126 	default:
127 		/*
128 		 * This can't happen, the type is validated several times
129 		 * before we get here.
130 		 */
131 		BUG();
132 	}
133 }
134 
135 static int btrfs_decompress_bio(struct compressed_bio *cb);
136 
137 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
138 				      unsigned long disk_size)
139 {
140 	return sizeof(struct compressed_bio) +
141 		(DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size;
142 }
143 
144 static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio,
145 				 u64 disk_start)
146 {
147 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
148 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
149 	const u32 csum_size = fs_info->csum_size;
150 	const u32 sectorsize = fs_info->sectorsize;
151 	struct page *page;
152 	unsigned int i;
153 	char *kaddr;
154 	u8 csum[BTRFS_CSUM_SIZE];
155 	struct compressed_bio *cb = bio->bi_private;
156 	u8 *cb_sum = cb->sums;
157 
158 	if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM))
159 		return 0;
160 
161 	shash->tfm = fs_info->csum_shash;
162 
163 	for (i = 0; i < cb->nr_pages; i++) {
164 		u32 pg_offset;
165 		u32 bytes_left = PAGE_SIZE;
166 		page = cb->compressed_pages[i];
167 
168 		/* Determine the remaining bytes inside the page first */
169 		if (i == cb->nr_pages - 1)
170 			bytes_left = cb->compressed_len - i * PAGE_SIZE;
171 
172 		/* Hash through the page sector by sector */
173 		for (pg_offset = 0; pg_offset < bytes_left;
174 		     pg_offset += sectorsize) {
175 			kaddr = page_address(page);
176 			crypto_shash_digest(shash, kaddr + pg_offset,
177 					    sectorsize, csum);
178 
179 			if (memcmp(&csum, cb_sum, csum_size) != 0) {
180 				btrfs_print_data_csum_error(inode, disk_start,
181 						csum, cb_sum, cb->mirror_num);
182 				if (btrfs_io_bio(bio)->device)
183 					btrfs_dev_stat_inc_and_print(
184 						btrfs_io_bio(bio)->device,
185 						BTRFS_DEV_STAT_CORRUPTION_ERRS);
186 				return -EIO;
187 			}
188 			cb_sum += csum_size;
189 			disk_start += sectorsize;
190 		}
191 	}
192 	return 0;
193 }
194 
195 /* when we finish reading compressed pages from the disk, we
196  * decompress them and then run the bio end_io routines on the
197  * decompressed pages (in the inode address space).
198  *
199  * This allows the checksumming and other IO error handling routines
200  * to work normally
201  *
202  * The compressed pages are freed here, and it must be run
203  * in process context
204  */
205 static void end_compressed_bio_read(struct bio *bio)
206 {
207 	struct compressed_bio *cb = bio->bi_private;
208 	struct inode *inode;
209 	struct page *page;
210 	unsigned int index;
211 	unsigned int mirror = btrfs_io_bio(bio)->mirror_num;
212 	int ret = 0;
213 
214 	if (bio->bi_status)
215 		cb->errors = 1;
216 
217 	/* if there are more bios still pending for this compressed
218 	 * extent, just exit
219 	 */
220 	if (!refcount_dec_and_test(&cb->pending_bios))
221 		goto out;
222 
223 	/*
224 	 * Record the correct mirror_num in cb->orig_bio so that
225 	 * read-repair can work properly.
226 	 */
227 	btrfs_io_bio(cb->orig_bio)->mirror_num = mirror;
228 	cb->mirror_num = mirror;
229 
230 	/*
231 	 * Some IO in this cb have failed, just skip checksum as there
232 	 * is no way it could be correct.
233 	 */
234 	if (cb->errors == 1)
235 		goto csum_failed;
236 
237 	inode = cb->inode;
238 	ret = check_compressed_csum(BTRFS_I(inode), bio,
239 				    bio->bi_iter.bi_sector << 9);
240 	if (ret)
241 		goto csum_failed;
242 
243 	/* ok, we're the last bio for this extent, lets start
244 	 * the decompression.
245 	 */
246 	ret = btrfs_decompress_bio(cb);
247 
248 csum_failed:
249 	if (ret)
250 		cb->errors = 1;
251 
252 	/* release the compressed pages */
253 	index = 0;
254 	for (index = 0; index < cb->nr_pages; index++) {
255 		page = cb->compressed_pages[index];
256 		page->mapping = NULL;
257 		put_page(page);
258 	}
259 
260 	/* do io completion on the original bio */
261 	if (cb->errors) {
262 		bio_io_error(cb->orig_bio);
263 	} else {
264 		struct bio_vec *bvec;
265 		struct bvec_iter_all iter_all;
266 
267 		/*
268 		 * we have verified the checksum already, set page
269 		 * checked so the end_io handlers know about it
270 		 */
271 		ASSERT(!bio_flagged(bio, BIO_CLONED));
272 		bio_for_each_segment_all(bvec, cb->orig_bio, iter_all)
273 			SetPageChecked(bvec->bv_page);
274 
275 		bio_endio(cb->orig_bio);
276 	}
277 
278 	/* finally free the cb struct */
279 	kfree(cb->compressed_pages);
280 	kfree(cb);
281 out:
282 	bio_put(bio);
283 }
284 
285 /*
286  * Clear the writeback bits on all of the file
287  * pages for a compressed write
288  */
289 static noinline void end_compressed_writeback(struct inode *inode,
290 					      const struct compressed_bio *cb)
291 {
292 	unsigned long index = cb->start >> PAGE_SHIFT;
293 	unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
294 	struct page *pages[16];
295 	unsigned long nr_pages = end_index - index + 1;
296 	int i;
297 	int ret;
298 
299 	if (cb->errors)
300 		mapping_set_error(inode->i_mapping, -EIO);
301 
302 	while (nr_pages > 0) {
303 		ret = find_get_pages_contig(inode->i_mapping, index,
304 				     min_t(unsigned long,
305 				     nr_pages, ARRAY_SIZE(pages)), pages);
306 		if (ret == 0) {
307 			nr_pages -= 1;
308 			index += 1;
309 			continue;
310 		}
311 		for (i = 0; i < ret; i++) {
312 			if (cb->errors)
313 				SetPageError(pages[i]);
314 			end_page_writeback(pages[i]);
315 			put_page(pages[i]);
316 		}
317 		nr_pages -= ret;
318 		index += ret;
319 	}
320 	/* the inode may be gone now */
321 }
322 
323 /*
324  * do the cleanup once all the compressed pages hit the disk.
325  * This will clear writeback on the file pages and free the compressed
326  * pages.
327  *
328  * This also calls the writeback end hooks for the file pages so that
329  * metadata and checksums can be updated in the file.
330  */
331 static void end_compressed_bio_write(struct bio *bio)
332 {
333 	struct compressed_bio *cb = bio->bi_private;
334 	struct inode *inode;
335 	struct page *page;
336 	unsigned int index;
337 
338 	if (bio->bi_status)
339 		cb->errors = 1;
340 
341 	/* if there are more bios still pending for this compressed
342 	 * extent, just exit
343 	 */
344 	if (!refcount_dec_and_test(&cb->pending_bios))
345 		goto out;
346 
347 	/* ok, we're the last bio for this extent, step one is to
348 	 * call back into the FS and do all the end_io operations
349 	 */
350 	inode = cb->inode;
351 	btrfs_record_physical_zoned(inode, cb->start, bio);
352 	btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL,
353 			cb->start, cb->start + cb->len - 1,
354 			!cb->errors);
355 
356 	end_compressed_writeback(inode, cb);
357 	/* note, our inode could be gone now */
358 
359 	/*
360 	 * release the compressed pages, these came from alloc_page and
361 	 * are not attached to the inode at all
362 	 */
363 	index = 0;
364 	for (index = 0; index < cb->nr_pages; index++) {
365 		page = cb->compressed_pages[index];
366 		page->mapping = NULL;
367 		put_page(page);
368 	}
369 
370 	/* finally free the cb struct */
371 	kfree(cb->compressed_pages);
372 	kfree(cb);
373 out:
374 	bio_put(bio);
375 }
376 
377 /*
378  * worker function to build and submit bios for previously compressed pages.
379  * The corresponding pages in the inode should be marked for writeback
380  * and the compressed pages should have a reference on them for dropping
381  * when the IO is complete.
382  *
383  * This also checksums the file bytes and gets things ready for
384  * the end io hooks.
385  */
386 blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start,
387 				 unsigned int len, u64 disk_start,
388 				 unsigned int compressed_len,
389 				 struct page **compressed_pages,
390 				 unsigned int nr_pages,
391 				 unsigned int write_flags,
392 				 struct cgroup_subsys_state *blkcg_css)
393 {
394 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
395 	struct bio *bio = NULL;
396 	struct compressed_bio *cb;
397 	unsigned long bytes_left;
398 	int pg_index = 0;
399 	struct page *page;
400 	u64 first_byte = disk_start;
401 	blk_status_t ret;
402 	int skip_sum = inode->flags & BTRFS_INODE_NODATASUM;
403 	const bool use_append = btrfs_use_zone_append(inode, disk_start);
404 	const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE;
405 
406 	WARN_ON(!PAGE_ALIGNED(start));
407 	cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
408 	if (!cb)
409 		return BLK_STS_RESOURCE;
410 	refcount_set(&cb->pending_bios, 0);
411 	cb->errors = 0;
412 	cb->inode = &inode->vfs_inode;
413 	cb->start = start;
414 	cb->len = len;
415 	cb->mirror_num = 0;
416 	cb->compressed_pages = compressed_pages;
417 	cb->compressed_len = compressed_len;
418 	cb->orig_bio = NULL;
419 	cb->nr_pages = nr_pages;
420 
421 	bio = btrfs_bio_alloc(first_byte);
422 	bio->bi_opf = bio_op | write_flags;
423 	bio->bi_private = cb;
424 	bio->bi_end_io = end_compressed_bio_write;
425 
426 	if (use_append) {
427 		struct btrfs_device *device;
428 
429 		device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE);
430 		if (IS_ERR(device)) {
431 			kfree(cb);
432 			bio_put(bio);
433 			return BLK_STS_NOTSUPP;
434 		}
435 
436 		bio_set_dev(bio, device->bdev);
437 	}
438 
439 	if (blkcg_css) {
440 		bio->bi_opf |= REQ_CGROUP_PUNT;
441 		kthread_associate_blkcg(blkcg_css);
442 	}
443 	refcount_set(&cb->pending_bios, 1);
444 
445 	/* create and submit bios for the compressed pages */
446 	bytes_left = compressed_len;
447 	for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
448 		int submit = 0;
449 		int len = 0;
450 
451 		page = compressed_pages[pg_index];
452 		page->mapping = inode->vfs_inode.i_mapping;
453 		if (bio->bi_iter.bi_size)
454 			submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio,
455 							  0);
456 
457 		/*
458 		 * Page can only be added to bio if the current bio fits in
459 		 * stripe.
460 		 */
461 		if (!submit) {
462 			if (pg_index == 0 && use_append)
463 				len = bio_add_zone_append_page(bio, page,
464 							       PAGE_SIZE, 0);
465 			else
466 				len = bio_add_page(bio, page, PAGE_SIZE, 0);
467 		}
468 
469 		page->mapping = NULL;
470 		if (submit || len < PAGE_SIZE) {
471 			/*
472 			 * inc the count before we submit the bio so
473 			 * we know the end IO handler won't happen before
474 			 * we inc the count.  Otherwise, the cb might get
475 			 * freed before we're done setting it up
476 			 */
477 			refcount_inc(&cb->pending_bios);
478 			ret = btrfs_bio_wq_end_io(fs_info, bio,
479 						  BTRFS_WQ_ENDIO_DATA);
480 			BUG_ON(ret); /* -ENOMEM */
481 
482 			if (!skip_sum) {
483 				ret = btrfs_csum_one_bio(inode, bio, start, 1);
484 				BUG_ON(ret); /* -ENOMEM */
485 			}
486 
487 			ret = btrfs_map_bio(fs_info, bio, 0);
488 			if (ret) {
489 				bio->bi_status = ret;
490 				bio_endio(bio);
491 			}
492 
493 			bio = btrfs_bio_alloc(first_byte);
494 			bio->bi_opf = bio_op | write_flags;
495 			bio->bi_private = cb;
496 			bio->bi_end_io = end_compressed_bio_write;
497 			if (blkcg_css)
498 				bio->bi_opf |= REQ_CGROUP_PUNT;
499 			/*
500 			 * Use bio_add_page() to ensure the bio has at least one
501 			 * page.
502 			 */
503 			bio_add_page(bio, page, PAGE_SIZE, 0);
504 		}
505 		if (bytes_left < PAGE_SIZE) {
506 			btrfs_info(fs_info,
507 					"bytes left %lu compress len %u nr %u",
508 			       bytes_left, cb->compressed_len, cb->nr_pages);
509 		}
510 		bytes_left -= PAGE_SIZE;
511 		first_byte += PAGE_SIZE;
512 		cond_resched();
513 	}
514 
515 	ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
516 	BUG_ON(ret); /* -ENOMEM */
517 
518 	if (!skip_sum) {
519 		ret = btrfs_csum_one_bio(inode, bio, start, 1);
520 		BUG_ON(ret); /* -ENOMEM */
521 	}
522 
523 	ret = btrfs_map_bio(fs_info, bio, 0);
524 	if (ret) {
525 		bio->bi_status = ret;
526 		bio_endio(bio);
527 	}
528 
529 	if (blkcg_css)
530 		kthread_associate_blkcg(NULL);
531 
532 	return 0;
533 }
534 
535 static u64 bio_end_offset(struct bio *bio)
536 {
537 	struct bio_vec *last = bio_last_bvec_all(bio);
538 
539 	return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
540 }
541 
542 static noinline int add_ra_bio_pages(struct inode *inode,
543 				     u64 compressed_end,
544 				     struct compressed_bio *cb)
545 {
546 	unsigned long end_index;
547 	unsigned long pg_index;
548 	u64 last_offset;
549 	u64 isize = i_size_read(inode);
550 	int ret;
551 	struct page *page;
552 	unsigned long nr_pages = 0;
553 	struct extent_map *em;
554 	struct address_space *mapping = inode->i_mapping;
555 	struct extent_map_tree *em_tree;
556 	struct extent_io_tree *tree;
557 	u64 end;
558 	int misses = 0;
559 
560 	last_offset = bio_end_offset(cb->orig_bio);
561 	em_tree = &BTRFS_I(inode)->extent_tree;
562 	tree = &BTRFS_I(inode)->io_tree;
563 
564 	if (isize == 0)
565 		return 0;
566 
567 	/*
568 	 * For current subpage support, we only support 64K page size,
569 	 * which means maximum compressed extent size (128K) is just 2x page
570 	 * size.
571 	 * This makes readahead less effective, so here disable readahead for
572 	 * subpage for now, until full compressed write is supported.
573 	 */
574 	if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE)
575 		return 0;
576 
577 	end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
578 
579 	while (last_offset < compressed_end) {
580 		pg_index = last_offset >> PAGE_SHIFT;
581 
582 		if (pg_index > end_index)
583 			break;
584 
585 		page = xa_load(&mapping->i_pages, pg_index);
586 		if (page && !xa_is_value(page)) {
587 			misses++;
588 			if (misses > 4)
589 				break;
590 			goto next;
591 		}
592 
593 		page = __page_cache_alloc(mapping_gfp_constraint(mapping,
594 								 ~__GFP_FS));
595 		if (!page)
596 			break;
597 
598 		if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
599 			put_page(page);
600 			goto next;
601 		}
602 
603 		/*
604 		 * at this point, we have a locked page in the page cache
605 		 * for these bytes in the file.  But, we have to make
606 		 * sure they map to this compressed extent on disk.
607 		 */
608 		ret = set_page_extent_mapped(page);
609 		if (ret < 0) {
610 			unlock_page(page);
611 			put_page(page);
612 			break;
613 		}
614 
615 		end = last_offset + PAGE_SIZE - 1;
616 		lock_extent(tree, last_offset, end);
617 		read_lock(&em_tree->lock);
618 		em = lookup_extent_mapping(em_tree, last_offset,
619 					   PAGE_SIZE);
620 		read_unlock(&em_tree->lock);
621 
622 		if (!em || last_offset < em->start ||
623 		    (last_offset + PAGE_SIZE > extent_map_end(em)) ||
624 		    (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
625 			free_extent_map(em);
626 			unlock_extent(tree, last_offset, end);
627 			unlock_page(page);
628 			put_page(page);
629 			break;
630 		}
631 		free_extent_map(em);
632 
633 		if (page->index == end_index) {
634 			size_t zero_offset = offset_in_page(isize);
635 
636 			if (zero_offset) {
637 				int zeros;
638 				zeros = PAGE_SIZE - zero_offset;
639 				memzero_page(page, zero_offset, zeros);
640 				flush_dcache_page(page);
641 			}
642 		}
643 
644 		ret = bio_add_page(cb->orig_bio, page,
645 				   PAGE_SIZE, 0);
646 
647 		if (ret == PAGE_SIZE) {
648 			nr_pages++;
649 			put_page(page);
650 		} else {
651 			unlock_extent(tree, last_offset, end);
652 			unlock_page(page);
653 			put_page(page);
654 			break;
655 		}
656 next:
657 		last_offset += PAGE_SIZE;
658 	}
659 	return 0;
660 }
661 
662 /*
663  * for a compressed read, the bio we get passed has all the inode pages
664  * in it.  We don't actually do IO on those pages but allocate new ones
665  * to hold the compressed pages on disk.
666  *
667  * bio->bi_iter.bi_sector points to the compressed extent on disk
668  * bio->bi_io_vec points to all of the inode pages
669  *
670  * After the compressed pages are read, we copy the bytes into the
671  * bio we were passed and then call the bio end_io calls
672  */
673 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
674 				 int mirror_num, unsigned long bio_flags)
675 {
676 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
677 	struct extent_map_tree *em_tree;
678 	struct compressed_bio *cb;
679 	unsigned int compressed_len;
680 	unsigned int nr_pages;
681 	unsigned int pg_index;
682 	struct page *page;
683 	struct bio *comp_bio;
684 	u64 cur_disk_byte = bio->bi_iter.bi_sector << 9;
685 	u64 file_offset;
686 	u64 em_len;
687 	u64 em_start;
688 	struct extent_map *em;
689 	blk_status_t ret = BLK_STS_RESOURCE;
690 	int faili = 0;
691 	u8 *sums;
692 
693 	em_tree = &BTRFS_I(inode)->extent_tree;
694 
695 	file_offset = bio_first_bvec_all(bio)->bv_offset +
696 		      page_offset(bio_first_page_all(bio));
697 
698 	/* we need the actual starting offset of this extent in the file */
699 	read_lock(&em_tree->lock);
700 	em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize);
701 	read_unlock(&em_tree->lock);
702 	if (!em)
703 		return BLK_STS_IOERR;
704 
705 	ASSERT(em->compress_type != BTRFS_COMPRESS_NONE);
706 	compressed_len = em->block_len;
707 	cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
708 	if (!cb)
709 		goto out;
710 
711 	refcount_set(&cb->pending_bios, 0);
712 	cb->errors = 0;
713 	cb->inode = inode;
714 	cb->mirror_num = mirror_num;
715 	sums = cb->sums;
716 
717 	cb->start = em->orig_start;
718 	em_len = em->len;
719 	em_start = em->start;
720 
721 	free_extent_map(em);
722 	em = NULL;
723 
724 	cb->len = bio->bi_iter.bi_size;
725 	cb->compressed_len = compressed_len;
726 	cb->compress_type = extent_compress_type(bio_flags);
727 	cb->orig_bio = bio;
728 
729 	nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
730 	cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
731 				       GFP_NOFS);
732 	if (!cb->compressed_pages)
733 		goto fail1;
734 
735 	for (pg_index = 0; pg_index < nr_pages; pg_index++) {
736 		cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS);
737 		if (!cb->compressed_pages[pg_index]) {
738 			faili = pg_index - 1;
739 			ret = BLK_STS_RESOURCE;
740 			goto fail2;
741 		}
742 	}
743 	faili = nr_pages - 1;
744 	cb->nr_pages = nr_pages;
745 
746 	add_ra_bio_pages(inode, em_start + em_len, cb);
747 
748 	/* include any pages we added in add_ra-bio_pages */
749 	cb->len = bio->bi_iter.bi_size;
750 
751 	comp_bio = btrfs_bio_alloc(cur_disk_byte);
752 	comp_bio->bi_opf = REQ_OP_READ;
753 	comp_bio->bi_private = cb;
754 	comp_bio->bi_end_io = end_compressed_bio_read;
755 	refcount_set(&cb->pending_bios, 1);
756 
757 	for (pg_index = 0; pg_index < nr_pages; pg_index++) {
758 		u32 pg_len = PAGE_SIZE;
759 		int submit = 0;
760 
761 		/*
762 		 * To handle subpage case, we need to make sure the bio only
763 		 * covers the range we need.
764 		 *
765 		 * If we're at the last page, truncate the length to only cover
766 		 * the remaining part.
767 		 */
768 		if (pg_index == nr_pages - 1)
769 			pg_len = min_t(u32, PAGE_SIZE,
770 					compressed_len - pg_index * PAGE_SIZE);
771 
772 		page = cb->compressed_pages[pg_index];
773 		page->mapping = inode->i_mapping;
774 		page->index = em_start >> PAGE_SHIFT;
775 
776 		if (comp_bio->bi_iter.bi_size)
777 			submit = btrfs_bio_fits_in_stripe(page, pg_len,
778 							  comp_bio, 0);
779 
780 		page->mapping = NULL;
781 		if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) {
782 			unsigned int nr_sectors;
783 
784 			ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
785 						  BTRFS_WQ_ENDIO_DATA);
786 			BUG_ON(ret); /* -ENOMEM */
787 
788 			/*
789 			 * inc the count before we submit the bio so
790 			 * we know the end IO handler won't happen before
791 			 * we inc the count.  Otherwise, the cb might get
792 			 * freed before we're done setting it up
793 			 */
794 			refcount_inc(&cb->pending_bios);
795 
796 			ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
797 			BUG_ON(ret); /* -ENOMEM */
798 
799 			nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
800 						  fs_info->sectorsize);
801 			sums += fs_info->csum_size * nr_sectors;
802 
803 			ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
804 			if (ret) {
805 				comp_bio->bi_status = ret;
806 				bio_endio(comp_bio);
807 			}
808 
809 			comp_bio = btrfs_bio_alloc(cur_disk_byte);
810 			comp_bio->bi_opf = REQ_OP_READ;
811 			comp_bio->bi_private = cb;
812 			comp_bio->bi_end_io = end_compressed_bio_read;
813 
814 			bio_add_page(comp_bio, page, pg_len, 0);
815 		}
816 		cur_disk_byte += pg_len;
817 	}
818 
819 	ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
820 	BUG_ON(ret); /* -ENOMEM */
821 
822 	ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
823 	BUG_ON(ret); /* -ENOMEM */
824 
825 	ret = btrfs_map_bio(fs_info, comp_bio, mirror_num);
826 	if (ret) {
827 		comp_bio->bi_status = ret;
828 		bio_endio(comp_bio);
829 	}
830 
831 	return 0;
832 
833 fail2:
834 	while (faili >= 0) {
835 		__free_page(cb->compressed_pages[faili]);
836 		faili--;
837 	}
838 
839 	kfree(cb->compressed_pages);
840 fail1:
841 	kfree(cb);
842 out:
843 	free_extent_map(em);
844 	return ret;
845 }
846 
847 /*
848  * Heuristic uses systematic sampling to collect data from the input data
849  * range, the logic can be tuned by the following constants:
850  *
851  * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample
852  * @SAMPLING_INTERVAL  - range from which the sampled data can be collected
853  */
854 #define SAMPLING_READ_SIZE	(16)
855 #define SAMPLING_INTERVAL	(256)
856 
857 /*
858  * For statistical analysis of the input data we consider bytes that form a
859  * Galois Field of 256 objects. Each object has an attribute count, ie. how
860  * many times the object appeared in the sample.
861  */
862 #define BUCKET_SIZE		(256)
863 
864 /*
865  * The size of the sample is based on a statistical sampling rule of thumb.
866  * The common way is to perform sampling tests as long as the number of
867  * elements in each cell is at least 5.
868  *
869  * Instead of 5, we choose 32 to obtain more accurate results.
870  * If the data contain the maximum number of symbols, which is 256, we obtain a
871  * sample size bound by 8192.
872  *
873  * For a sample of at most 8KB of data per data range: 16 consecutive bytes
874  * from up to 512 locations.
875  */
876 #define MAX_SAMPLE_SIZE		(BTRFS_MAX_UNCOMPRESSED *		\
877 				 SAMPLING_READ_SIZE / SAMPLING_INTERVAL)
878 
879 struct bucket_item {
880 	u32 count;
881 };
882 
883 struct heuristic_ws {
884 	/* Partial copy of input data */
885 	u8 *sample;
886 	u32 sample_size;
887 	/* Buckets store counters for each byte value */
888 	struct bucket_item *bucket;
889 	/* Sorting buffer */
890 	struct bucket_item *bucket_b;
891 	struct list_head list;
892 };
893 
894 static struct workspace_manager heuristic_wsm;
895 
896 static void free_heuristic_ws(struct list_head *ws)
897 {
898 	struct heuristic_ws *workspace;
899 
900 	workspace = list_entry(ws, struct heuristic_ws, list);
901 
902 	kvfree(workspace->sample);
903 	kfree(workspace->bucket);
904 	kfree(workspace->bucket_b);
905 	kfree(workspace);
906 }
907 
908 static struct list_head *alloc_heuristic_ws(unsigned int level)
909 {
910 	struct heuristic_ws *ws;
911 
912 	ws = kzalloc(sizeof(*ws), GFP_KERNEL);
913 	if (!ws)
914 		return ERR_PTR(-ENOMEM);
915 
916 	ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL);
917 	if (!ws->sample)
918 		goto fail;
919 
920 	ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL);
921 	if (!ws->bucket)
922 		goto fail;
923 
924 	ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL);
925 	if (!ws->bucket_b)
926 		goto fail;
927 
928 	INIT_LIST_HEAD(&ws->list);
929 	return &ws->list;
930 fail:
931 	free_heuristic_ws(&ws->list);
932 	return ERR_PTR(-ENOMEM);
933 }
934 
935 const struct btrfs_compress_op btrfs_heuristic_compress = {
936 	.workspace_manager = &heuristic_wsm,
937 };
938 
939 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
940 	/* The heuristic is represented as compression type 0 */
941 	&btrfs_heuristic_compress,
942 	&btrfs_zlib_compress,
943 	&btrfs_lzo_compress,
944 	&btrfs_zstd_compress,
945 };
946 
947 static struct list_head *alloc_workspace(int type, unsigned int level)
948 {
949 	switch (type) {
950 	case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level);
951 	case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level);
952 	case BTRFS_COMPRESS_LZO:  return lzo_alloc_workspace(level);
953 	case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level);
954 	default:
955 		/*
956 		 * This can't happen, the type is validated several times
957 		 * before we get here.
958 		 */
959 		BUG();
960 	}
961 }
962 
963 static void free_workspace(int type, struct list_head *ws)
964 {
965 	switch (type) {
966 	case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws);
967 	case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws);
968 	case BTRFS_COMPRESS_LZO:  return lzo_free_workspace(ws);
969 	case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws);
970 	default:
971 		/*
972 		 * This can't happen, the type is validated several times
973 		 * before we get here.
974 		 */
975 		BUG();
976 	}
977 }
978 
979 static void btrfs_init_workspace_manager(int type)
980 {
981 	struct workspace_manager *wsm;
982 	struct list_head *workspace;
983 
984 	wsm = btrfs_compress_op[type]->workspace_manager;
985 	INIT_LIST_HEAD(&wsm->idle_ws);
986 	spin_lock_init(&wsm->ws_lock);
987 	atomic_set(&wsm->total_ws, 0);
988 	init_waitqueue_head(&wsm->ws_wait);
989 
990 	/*
991 	 * Preallocate one workspace for each compression type so we can
992 	 * guarantee forward progress in the worst case
993 	 */
994 	workspace = alloc_workspace(type, 0);
995 	if (IS_ERR(workspace)) {
996 		pr_warn(
997 	"BTRFS: cannot preallocate compression workspace, will try later\n");
998 	} else {
999 		atomic_set(&wsm->total_ws, 1);
1000 		wsm->free_ws = 1;
1001 		list_add(workspace, &wsm->idle_ws);
1002 	}
1003 }
1004 
1005 static void btrfs_cleanup_workspace_manager(int type)
1006 {
1007 	struct workspace_manager *wsman;
1008 	struct list_head *ws;
1009 
1010 	wsman = btrfs_compress_op[type]->workspace_manager;
1011 	while (!list_empty(&wsman->idle_ws)) {
1012 		ws = wsman->idle_ws.next;
1013 		list_del(ws);
1014 		free_workspace(type, ws);
1015 		atomic_dec(&wsman->total_ws);
1016 	}
1017 }
1018 
1019 /*
1020  * This finds an available workspace or allocates a new one.
1021  * If it's not possible to allocate a new one, waits until there's one.
1022  * Preallocation makes a forward progress guarantees and we do not return
1023  * errors.
1024  */
1025 struct list_head *btrfs_get_workspace(int type, unsigned int level)
1026 {
1027 	struct workspace_manager *wsm;
1028 	struct list_head *workspace;
1029 	int cpus = num_online_cpus();
1030 	unsigned nofs_flag;
1031 	struct list_head *idle_ws;
1032 	spinlock_t *ws_lock;
1033 	atomic_t *total_ws;
1034 	wait_queue_head_t *ws_wait;
1035 	int *free_ws;
1036 
1037 	wsm = btrfs_compress_op[type]->workspace_manager;
1038 	idle_ws	 = &wsm->idle_ws;
1039 	ws_lock	 = &wsm->ws_lock;
1040 	total_ws = &wsm->total_ws;
1041 	ws_wait	 = &wsm->ws_wait;
1042 	free_ws	 = &wsm->free_ws;
1043 
1044 again:
1045 	spin_lock(ws_lock);
1046 	if (!list_empty(idle_ws)) {
1047 		workspace = idle_ws->next;
1048 		list_del(workspace);
1049 		(*free_ws)--;
1050 		spin_unlock(ws_lock);
1051 		return workspace;
1052 
1053 	}
1054 	if (atomic_read(total_ws) > cpus) {
1055 		DEFINE_WAIT(wait);
1056 
1057 		spin_unlock(ws_lock);
1058 		prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
1059 		if (atomic_read(total_ws) > cpus && !*free_ws)
1060 			schedule();
1061 		finish_wait(ws_wait, &wait);
1062 		goto again;
1063 	}
1064 	atomic_inc(total_ws);
1065 	spin_unlock(ws_lock);
1066 
1067 	/*
1068 	 * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
1069 	 * to turn it off here because we might get called from the restricted
1070 	 * context of btrfs_compress_bio/btrfs_compress_pages
1071 	 */
1072 	nofs_flag = memalloc_nofs_save();
1073 	workspace = alloc_workspace(type, level);
1074 	memalloc_nofs_restore(nofs_flag);
1075 
1076 	if (IS_ERR(workspace)) {
1077 		atomic_dec(total_ws);
1078 		wake_up(ws_wait);
1079 
1080 		/*
1081 		 * Do not return the error but go back to waiting. There's a
1082 		 * workspace preallocated for each type and the compression
1083 		 * time is bounded so we get to a workspace eventually. This
1084 		 * makes our caller's life easier.
1085 		 *
1086 		 * To prevent silent and low-probability deadlocks (when the
1087 		 * initial preallocation fails), check if there are any
1088 		 * workspaces at all.
1089 		 */
1090 		if (atomic_read(total_ws) == 0) {
1091 			static DEFINE_RATELIMIT_STATE(_rs,
1092 					/* once per minute */ 60 * HZ,
1093 					/* no burst */ 1);
1094 
1095 			if (__ratelimit(&_rs)) {
1096 				pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
1097 			}
1098 		}
1099 		goto again;
1100 	}
1101 	return workspace;
1102 }
1103 
1104 static struct list_head *get_workspace(int type, int level)
1105 {
1106 	switch (type) {
1107 	case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level);
1108 	case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level);
1109 	case BTRFS_COMPRESS_LZO:  return btrfs_get_workspace(type, level);
1110 	case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level);
1111 	default:
1112 		/*
1113 		 * This can't happen, the type is validated several times
1114 		 * before we get here.
1115 		 */
1116 		BUG();
1117 	}
1118 }
1119 
1120 /*
1121  * put a workspace struct back on the list or free it if we have enough
1122  * idle ones sitting around
1123  */
1124 void btrfs_put_workspace(int type, struct list_head *ws)
1125 {
1126 	struct workspace_manager *wsm;
1127 	struct list_head *idle_ws;
1128 	spinlock_t *ws_lock;
1129 	atomic_t *total_ws;
1130 	wait_queue_head_t *ws_wait;
1131 	int *free_ws;
1132 
1133 	wsm = btrfs_compress_op[type]->workspace_manager;
1134 	idle_ws	 = &wsm->idle_ws;
1135 	ws_lock	 = &wsm->ws_lock;
1136 	total_ws = &wsm->total_ws;
1137 	ws_wait	 = &wsm->ws_wait;
1138 	free_ws	 = &wsm->free_ws;
1139 
1140 	spin_lock(ws_lock);
1141 	if (*free_ws <= num_online_cpus()) {
1142 		list_add(ws, idle_ws);
1143 		(*free_ws)++;
1144 		spin_unlock(ws_lock);
1145 		goto wake;
1146 	}
1147 	spin_unlock(ws_lock);
1148 
1149 	free_workspace(type, ws);
1150 	atomic_dec(total_ws);
1151 wake:
1152 	cond_wake_up(ws_wait);
1153 }
1154 
1155 static void put_workspace(int type, struct list_head *ws)
1156 {
1157 	switch (type) {
1158 	case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws);
1159 	case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws);
1160 	case BTRFS_COMPRESS_LZO:  return btrfs_put_workspace(type, ws);
1161 	case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws);
1162 	default:
1163 		/*
1164 		 * This can't happen, the type is validated several times
1165 		 * before we get here.
1166 		 */
1167 		BUG();
1168 	}
1169 }
1170 
1171 /*
1172  * Adjust @level according to the limits of the compression algorithm or
1173  * fallback to default
1174  */
1175 static unsigned int btrfs_compress_set_level(int type, unsigned level)
1176 {
1177 	const struct btrfs_compress_op *ops = btrfs_compress_op[type];
1178 
1179 	if (level == 0)
1180 		level = ops->default_level;
1181 	else
1182 		level = min(level, ops->max_level);
1183 
1184 	return level;
1185 }
1186 
1187 /*
1188  * Given an address space and start and length, compress the bytes into @pages
1189  * that are allocated on demand.
1190  *
1191  * @type_level is encoded algorithm and level, where level 0 means whatever
1192  * default the algorithm chooses and is opaque here;
1193  * - compression algo are 0-3
1194  * - the level are bits 4-7
1195  *
1196  * @out_pages is an in/out parameter, holds maximum number of pages to allocate
1197  * and returns number of actually allocated pages
1198  *
1199  * @total_in is used to return the number of bytes actually read.  It
1200  * may be smaller than the input length if we had to exit early because we
1201  * ran out of room in the pages array or because we cross the
1202  * max_out threshold.
1203  *
1204  * @total_out is an in/out parameter, must be set to the input length and will
1205  * be also used to return the total number of compressed bytes
1206  */
1207 int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping,
1208 			 u64 start, struct page **pages,
1209 			 unsigned long *out_pages,
1210 			 unsigned long *total_in,
1211 			 unsigned long *total_out)
1212 {
1213 	int type = btrfs_compress_type(type_level);
1214 	int level = btrfs_compress_level(type_level);
1215 	struct list_head *workspace;
1216 	int ret;
1217 
1218 	level = btrfs_compress_set_level(type, level);
1219 	workspace = get_workspace(type, level);
1220 	ret = compression_compress_pages(type, workspace, mapping, start, pages,
1221 					 out_pages, total_in, total_out);
1222 	put_workspace(type, workspace);
1223 	return ret;
1224 }
1225 
1226 static int btrfs_decompress_bio(struct compressed_bio *cb)
1227 {
1228 	struct list_head *workspace;
1229 	int ret;
1230 	int type = cb->compress_type;
1231 
1232 	workspace = get_workspace(type, 0);
1233 	ret = compression_decompress_bio(type, workspace, cb);
1234 	put_workspace(type, workspace);
1235 
1236 	return ret;
1237 }
1238 
1239 /*
1240  * a less complex decompression routine.  Our compressed data fits in a
1241  * single page, and we want to read a single page out of it.
1242  * start_byte tells us the offset into the compressed data we're interested in
1243  */
1244 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
1245 		     unsigned long start_byte, size_t srclen, size_t destlen)
1246 {
1247 	struct list_head *workspace;
1248 	int ret;
1249 
1250 	workspace = get_workspace(type, 0);
1251 	ret = compression_decompress(type, workspace, data_in, dest_page,
1252 				     start_byte, srclen, destlen);
1253 	put_workspace(type, workspace);
1254 
1255 	return ret;
1256 }
1257 
1258 void __init btrfs_init_compress(void)
1259 {
1260 	btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE);
1261 	btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB);
1262 	btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO);
1263 	zstd_init_workspace_manager();
1264 }
1265 
1266 void __cold btrfs_exit_compress(void)
1267 {
1268 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE);
1269 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB);
1270 	btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO);
1271 	zstd_cleanup_workspace_manager();
1272 }
1273 
1274 /*
1275  * Copy decompressed data from working buffer to pages.
1276  *
1277  * @buf:		The decompressed data buffer
1278  * @buf_len:		The decompressed data length
1279  * @decompressed:	Number of bytes that are already decompressed inside the
1280  * 			compressed extent
1281  * @cb:			The compressed extent descriptor
1282  * @orig_bio:		The original bio that the caller wants to read for
1283  *
1284  * An easier to understand graph is like below:
1285  *
1286  * 		|<- orig_bio ->|     |<- orig_bio->|
1287  * 	|<-------      full decompressed extent      ----->|
1288  * 	|<-----------    @cb range   ---->|
1289  * 	|			|<-- @buf_len -->|
1290  * 	|<--- @decompressed --->|
1291  *
1292  * Note that, @cb can be a subpage of the full decompressed extent, but
1293  * @cb->start always has the same as the orig_file_offset value of the full
1294  * decompressed extent.
1295  *
1296  * When reading compressed extent, we have to read the full compressed extent,
1297  * while @orig_bio may only want part of the range.
1298  * Thus this function will ensure only data covered by @orig_bio will be copied
1299  * to.
1300  *
1301  * Return 0 if we have copied all needed contents for @orig_bio.
1302  * Return >0 if we need continue decompress.
1303  */
1304 int btrfs_decompress_buf2page(const char *buf, u32 buf_len,
1305 			      struct compressed_bio *cb, u32 decompressed)
1306 {
1307 	struct bio *orig_bio = cb->orig_bio;
1308 	/* Offset inside the full decompressed extent */
1309 	u32 cur_offset;
1310 
1311 	cur_offset = decompressed;
1312 	/* The main loop to do the copy */
1313 	while (cur_offset < decompressed + buf_len) {
1314 		struct bio_vec bvec;
1315 		size_t copy_len;
1316 		u32 copy_start;
1317 		/* Offset inside the full decompressed extent */
1318 		u32 bvec_offset;
1319 
1320 		bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter);
1321 		/*
1322 		 * cb->start may underflow, but subtracting that value can still
1323 		 * give us correct offset inside the full decompressed extent.
1324 		 */
1325 		bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start;
1326 
1327 		/* Haven't reached the bvec range, exit */
1328 		if (decompressed + buf_len <= bvec_offset)
1329 			return 1;
1330 
1331 		copy_start = max(cur_offset, bvec_offset);
1332 		copy_len = min(bvec_offset + bvec.bv_len,
1333 			       decompressed + buf_len) - copy_start;
1334 		ASSERT(copy_len);
1335 
1336 		/*
1337 		 * Extra range check to ensure we didn't go beyond
1338 		 * @buf + @buf_len.
1339 		 */
1340 		ASSERT(copy_start - decompressed < buf_len);
1341 		memcpy_to_page(bvec.bv_page, bvec.bv_offset,
1342 			       buf + copy_start - decompressed, copy_len);
1343 		flush_dcache_page(bvec.bv_page);
1344 		cur_offset += copy_len;
1345 
1346 		bio_advance(orig_bio, copy_len);
1347 		/* Finished the bio */
1348 		if (!orig_bio->bi_iter.bi_size)
1349 			return 0;
1350 	}
1351 	return 1;
1352 }
1353 
1354 /*
1355  * Shannon Entropy calculation
1356  *
1357  * Pure byte distribution analysis fails to determine compressibility of data.
1358  * Try calculating entropy to estimate the average minimum number of bits
1359  * needed to encode the sampled data.
1360  *
1361  * For convenience, return the percentage of needed bits, instead of amount of
1362  * bits directly.
1363  *
1364  * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy
1365  *			    and can be compressible with high probability
1366  *
1367  * @ENTROPY_LVL_HIGH - data are not compressible with high probability
1368  *
1369  * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate.
1370  */
1371 #define ENTROPY_LVL_ACEPTABLE		(65)
1372 #define ENTROPY_LVL_HIGH		(80)
1373 
1374 /*
1375  * For increasead precision in shannon_entropy calculation,
1376  * let's do pow(n, M) to save more digits after comma:
1377  *
1378  * - maximum int bit length is 64
1379  * - ilog2(MAX_SAMPLE_SIZE)	-> 13
1380  * - 13 * 4 = 52 < 64		-> M = 4
1381  *
1382  * So use pow(n, 4).
1383  */
1384 static inline u32 ilog2_w(u64 n)
1385 {
1386 	return ilog2(n * n * n * n);
1387 }
1388 
1389 static u32 shannon_entropy(struct heuristic_ws *ws)
1390 {
1391 	const u32 entropy_max = 8 * ilog2_w(2);
1392 	u32 entropy_sum = 0;
1393 	u32 p, p_base, sz_base;
1394 	u32 i;
1395 
1396 	sz_base = ilog2_w(ws->sample_size);
1397 	for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) {
1398 		p = ws->bucket[i].count;
1399 		p_base = ilog2_w(p);
1400 		entropy_sum += p * (sz_base - p_base);
1401 	}
1402 
1403 	entropy_sum /= ws->sample_size;
1404 	return entropy_sum * 100 / entropy_max;
1405 }
1406 
1407 #define RADIX_BASE		4U
1408 #define COUNTERS_SIZE		(1U << RADIX_BASE)
1409 
1410 static u8 get4bits(u64 num, int shift) {
1411 	u8 low4bits;
1412 
1413 	num >>= shift;
1414 	/* Reverse order */
1415 	low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE);
1416 	return low4bits;
1417 }
1418 
1419 /*
1420  * Use 4 bits as radix base
1421  * Use 16 u32 counters for calculating new position in buf array
1422  *
1423  * @array     - array that will be sorted
1424  * @array_buf - buffer array to store sorting results
1425  *              must be equal in size to @array
1426  * @num       - array size
1427  */
1428 static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf,
1429 		       int num)
1430 {
1431 	u64 max_num;
1432 	u64 buf_num;
1433 	u32 counters[COUNTERS_SIZE];
1434 	u32 new_addr;
1435 	u32 addr;
1436 	int bitlen;
1437 	int shift;
1438 	int i;
1439 
1440 	/*
1441 	 * Try avoid useless loop iterations for small numbers stored in big
1442 	 * counters.  Example: 48 33 4 ... in 64bit array
1443 	 */
1444 	max_num = array[0].count;
1445 	for (i = 1; i < num; i++) {
1446 		buf_num = array[i].count;
1447 		if (buf_num > max_num)
1448 			max_num = buf_num;
1449 	}
1450 
1451 	buf_num = ilog2(max_num);
1452 	bitlen = ALIGN(buf_num, RADIX_BASE * 2);
1453 
1454 	shift = 0;
1455 	while (shift < bitlen) {
1456 		memset(counters, 0, sizeof(counters));
1457 
1458 		for (i = 0; i < num; i++) {
1459 			buf_num = array[i].count;
1460 			addr = get4bits(buf_num, shift);
1461 			counters[addr]++;
1462 		}
1463 
1464 		for (i = 1; i < COUNTERS_SIZE; i++)
1465 			counters[i] += counters[i - 1];
1466 
1467 		for (i = num - 1; i >= 0; i--) {
1468 			buf_num = array[i].count;
1469 			addr = get4bits(buf_num, shift);
1470 			counters[addr]--;
1471 			new_addr = counters[addr];
1472 			array_buf[new_addr] = array[i];
1473 		}
1474 
1475 		shift += RADIX_BASE;
1476 
1477 		/*
1478 		 * Normal radix expects to move data from a temporary array, to
1479 		 * the main one.  But that requires some CPU time. Avoid that
1480 		 * by doing another sort iteration to original array instead of
1481 		 * memcpy()
1482 		 */
1483 		memset(counters, 0, sizeof(counters));
1484 
1485 		for (i = 0; i < num; i ++) {
1486 			buf_num = array_buf[i].count;
1487 			addr = get4bits(buf_num, shift);
1488 			counters[addr]++;
1489 		}
1490 
1491 		for (i = 1; i < COUNTERS_SIZE; i++)
1492 			counters[i] += counters[i - 1];
1493 
1494 		for (i = num - 1; i >= 0; i--) {
1495 			buf_num = array_buf[i].count;
1496 			addr = get4bits(buf_num, shift);
1497 			counters[addr]--;
1498 			new_addr = counters[addr];
1499 			array[new_addr] = array_buf[i];
1500 		}
1501 
1502 		shift += RADIX_BASE;
1503 	}
1504 }
1505 
1506 /*
1507  * Size of the core byte set - how many bytes cover 90% of the sample
1508  *
1509  * There are several types of structured binary data that use nearly all byte
1510  * values. The distribution can be uniform and counts in all buckets will be
1511  * nearly the same (eg. encrypted data). Unlikely to be compressible.
1512  *
1513  * Other possibility is normal (Gaussian) distribution, where the data could
1514  * be potentially compressible, but we have to take a few more steps to decide
1515  * how much.
1516  *
1517  * @BYTE_CORE_SET_LOW  - main part of byte values repeated frequently,
1518  *                       compression algo can easy fix that
1519  * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high
1520  *                       probability is not compressible
1521  */
1522 #define BYTE_CORE_SET_LOW		(64)
1523 #define BYTE_CORE_SET_HIGH		(200)
1524 
1525 static int byte_core_set_size(struct heuristic_ws *ws)
1526 {
1527 	u32 i;
1528 	u32 coreset_sum = 0;
1529 	const u32 core_set_threshold = ws->sample_size * 90 / 100;
1530 	struct bucket_item *bucket = ws->bucket;
1531 
1532 	/* Sort in reverse order */
1533 	radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE);
1534 
1535 	for (i = 0; i < BYTE_CORE_SET_LOW; i++)
1536 		coreset_sum += bucket[i].count;
1537 
1538 	if (coreset_sum > core_set_threshold)
1539 		return i;
1540 
1541 	for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) {
1542 		coreset_sum += bucket[i].count;
1543 		if (coreset_sum > core_set_threshold)
1544 			break;
1545 	}
1546 
1547 	return i;
1548 }
1549 
1550 /*
1551  * Count byte values in buckets.
1552  * This heuristic can detect textual data (configs, xml, json, html, etc).
1553  * Because in most text-like data byte set is restricted to limited number of
1554  * possible characters, and that restriction in most cases makes data easy to
1555  * compress.
1556  *
1557  * @BYTE_SET_THRESHOLD - consider all data within this byte set size:
1558  *	less - compressible
1559  *	more - need additional analysis
1560  */
1561 #define BYTE_SET_THRESHOLD		(64)
1562 
1563 static u32 byte_set_size(const struct heuristic_ws *ws)
1564 {
1565 	u32 i;
1566 	u32 byte_set_size = 0;
1567 
1568 	for (i = 0; i < BYTE_SET_THRESHOLD; i++) {
1569 		if (ws->bucket[i].count > 0)
1570 			byte_set_size++;
1571 	}
1572 
1573 	/*
1574 	 * Continue collecting count of byte values in buckets.  If the byte
1575 	 * set size is bigger then the threshold, it's pointless to continue,
1576 	 * the detection technique would fail for this type of data.
1577 	 */
1578 	for (; i < BUCKET_SIZE; i++) {
1579 		if (ws->bucket[i].count > 0) {
1580 			byte_set_size++;
1581 			if (byte_set_size > BYTE_SET_THRESHOLD)
1582 				return byte_set_size;
1583 		}
1584 	}
1585 
1586 	return byte_set_size;
1587 }
1588 
1589 static bool sample_repeated_patterns(struct heuristic_ws *ws)
1590 {
1591 	const u32 half_of_sample = ws->sample_size / 2;
1592 	const u8 *data = ws->sample;
1593 
1594 	return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0;
1595 }
1596 
1597 static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
1598 				     struct heuristic_ws *ws)
1599 {
1600 	struct page *page;
1601 	u64 index, index_end;
1602 	u32 i, curr_sample_pos;
1603 	u8 *in_data;
1604 
1605 	/*
1606 	 * Compression handles the input data by chunks of 128KiB
1607 	 * (defined by BTRFS_MAX_UNCOMPRESSED)
1608 	 *
1609 	 * We do the same for the heuristic and loop over the whole range.
1610 	 *
1611 	 * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
1612 	 * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
1613 	 */
1614 	if (end - start > BTRFS_MAX_UNCOMPRESSED)
1615 		end = start + BTRFS_MAX_UNCOMPRESSED;
1616 
1617 	index = start >> PAGE_SHIFT;
1618 	index_end = end >> PAGE_SHIFT;
1619 
1620 	/* Don't miss unaligned end */
1621 	if (!IS_ALIGNED(end, PAGE_SIZE))
1622 		index_end++;
1623 
1624 	curr_sample_pos = 0;
1625 	while (index < index_end) {
1626 		page = find_get_page(inode->i_mapping, index);
1627 		in_data = kmap_local_page(page);
1628 		/* Handle case where the start is not aligned to PAGE_SIZE */
1629 		i = start % PAGE_SIZE;
1630 		while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
1631 			/* Don't sample any garbage from the last page */
1632 			if (start > end - SAMPLING_READ_SIZE)
1633 				break;
1634 			memcpy(&ws->sample[curr_sample_pos], &in_data[i],
1635 					SAMPLING_READ_SIZE);
1636 			i += SAMPLING_INTERVAL;
1637 			start += SAMPLING_INTERVAL;
1638 			curr_sample_pos += SAMPLING_READ_SIZE;
1639 		}
1640 		kunmap_local(in_data);
1641 		put_page(page);
1642 
1643 		index++;
1644 	}
1645 
1646 	ws->sample_size = curr_sample_pos;
1647 }
1648 
1649 /*
1650  * Compression heuristic.
1651  *
1652  * For now is's a naive and optimistic 'return true', we'll extend the logic to
1653  * quickly (compared to direct compression) detect data characteristics
1654  * (compressible/uncompressible) to avoid wasting CPU time on uncompressible
1655  * data.
1656  *
1657  * The following types of analysis can be performed:
1658  * - detect mostly zero data
1659  * - detect data with low "byte set" size (text, etc)
1660  * - detect data with low/high "core byte" set
1661  *
1662  * Return non-zero if the compression should be done, 0 otherwise.
1663  */
1664 int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
1665 {
1666 	struct list_head *ws_list = get_workspace(0, 0);
1667 	struct heuristic_ws *ws;
1668 	u32 i;
1669 	u8 byte;
1670 	int ret = 0;
1671 
1672 	ws = list_entry(ws_list, struct heuristic_ws, list);
1673 
1674 	heuristic_collect_sample(inode, start, end, ws);
1675 
1676 	if (sample_repeated_patterns(ws)) {
1677 		ret = 1;
1678 		goto out;
1679 	}
1680 
1681 	memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
1682 
1683 	for (i = 0; i < ws->sample_size; i++) {
1684 		byte = ws->sample[i];
1685 		ws->bucket[byte].count++;
1686 	}
1687 
1688 	i = byte_set_size(ws);
1689 	if (i < BYTE_SET_THRESHOLD) {
1690 		ret = 2;
1691 		goto out;
1692 	}
1693 
1694 	i = byte_core_set_size(ws);
1695 	if (i <= BYTE_CORE_SET_LOW) {
1696 		ret = 3;
1697 		goto out;
1698 	}
1699 
1700 	if (i >= BYTE_CORE_SET_HIGH) {
1701 		ret = 0;
1702 		goto out;
1703 	}
1704 
1705 	i = shannon_entropy(ws);
1706 	if (i <= ENTROPY_LVL_ACEPTABLE) {
1707 		ret = 4;
1708 		goto out;
1709 	}
1710 
1711 	/*
1712 	 * For the levels below ENTROPY_LVL_HIGH, additional analysis would be
1713 	 * needed to give green light to compression.
1714 	 *
1715 	 * For now just assume that compression at that level is not worth the
1716 	 * resources because:
1717 	 *
1718 	 * 1. it is possible to defrag the data later
1719 	 *
1720 	 * 2. the data would turn out to be hardly compressible, eg. 150 byte
1721 	 * values, every bucket has counter at level ~54. The heuristic would
1722 	 * be confused. This can happen when data have some internal repeated
1723 	 * patterns like "abbacbbc...". This can be detected by analyzing
1724 	 * pairs of bytes, which is too costly.
1725 	 */
1726 	if (i < ENTROPY_LVL_HIGH) {
1727 		ret = 5;
1728 		goto out;
1729 	} else {
1730 		ret = 0;
1731 		goto out;
1732 	}
1733 
1734 out:
1735 	put_workspace(0, ws_list);
1736 	return ret;
1737 }
1738 
1739 /*
1740  * Convert the compression suffix (eg. after "zlib" starting with ":") to
1741  * level, unrecognized string will set the default level
1742  */
1743 unsigned int btrfs_compress_str2level(unsigned int type, const char *str)
1744 {
1745 	unsigned int level = 0;
1746 	int ret;
1747 
1748 	if (!type)
1749 		return 0;
1750 
1751 	if (str[0] == ':') {
1752 		ret = kstrtouint(str + 1, 10, &level);
1753 		if (ret)
1754 			level = 0;
1755 	}
1756 
1757 	level = btrfs_compress_set_level(type, level);
1758 
1759 	return level;
1760 }
1761