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 <crypto/hash.h> |
24 | #include "misc.h" |
25 | #include "ctree.h" |
26 | #include "fs.h" |
27 | #include "disk-io.h" |
28 | #include "transaction.h" |
29 | #include "btrfs_inode.h" |
30 | #include "bio.h" |
31 | #include "ordered-data.h" |
32 | #include "compression.h" |
33 | #include "extent_io.h" |
34 | #include "extent_map.h" |
35 | #include "subpage.h" |
36 | #include "zoned.h" |
37 | #include "file-item.h" |
38 | #include "super.h" |
39 | |
40 | static struct bio_set btrfs_compressed_bioset; |
41 | |
42 | static const char* const btrfs_compress_types[] = { "" , "zlib" , "lzo" , "zstd" }; |
43 | |
44 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) |
45 | { |
46 | switch (type) { |
47 | case BTRFS_COMPRESS_ZLIB: |
48 | case BTRFS_COMPRESS_LZO: |
49 | case BTRFS_COMPRESS_ZSTD: |
50 | case BTRFS_COMPRESS_NONE: |
51 | return btrfs_compress_types[type]; |
52 | default: |
53 | break; |
54 | } |
55 | |
56 | return NULL; |
57 | } |
58 | |
59 | static inline struct compressed_bio *to_compressed_bio(struct btrfs_bio *bbio) |
60 | { |
61 | return container_of(bbio, struct compressed_bio, bbio); |
62 | } |
63 | |
64 | static struct compressed_bio *alloc_compressed_bio(struct btrfs_inode *inode, |
65 | u64 start, blk_opf_t op, |
66 | btrfs_bio_end_io_t end_io) |
67 | { |
68 | struct btrfs_bio *bbio; |
69 | |
70 | bbio = btrfs_bio(bio: bio_alloc_bioset(NULL, BTRFS_MAX_COMPRESSED_PAGES, opf: op, |
71 | GFP_NOFS, bs: &btrfs_compressed_bioset)); |
72 | btrfs_bio_init(bbio, fs_info: inode->root->fs_info, end_io, NULL); |
73 | bbio->inode = inode; |
74 | bbio->file_offset = start; |
75 | return to_compressed_bio(bbio); |
76 | } |
77 | |
78 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
79 | { |
80 | int i; |
81 | |
82 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { |
83 | size_t comp_len = strlen(btrfs_compress_types[i]); |
84 | |
85 | if (len < comp_len) |
86 | continue; |
87 | |
88 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) |
89 | return true; |
90 | } |
91 | return false; |
92 | } |
93 | |
94 | static int compression_compress_pages(int type, struct list_head *ws, |
95 | struct address_space *mapping, u64 start, struct page **pages, |
96 | unsigned long *out_pages, unsigned long *total_in, |
97 | unsigned long *total_out) |
98 | { |
99 | switch (type) { |
100 | case BTRFS_COMPRESS_ZLIB: |
101 | return zlib_compress_pages(ws, mapping, start, pages, |
102 | out_pages, total_in, total_out); |
103 | case BTRFS_COMPRESS_LZO: |
104 | return lzo_compress_pages(ws, mapping, start, pages, |
105 | out_pages, total_in, total_out); |
106 | case BTRFS_COMPRESS_ZSTD: |
107 | return zstd_compress_pages(ws, mapping, start, pages, |
108 | out_pages, total_in, total_out); |
109 | case BTRFS_COMPRESS_NONE: |
110 | default: |
111 | /* |
112 | * This can happen when compression races with remount setting |
113 | * it to 'no compress', while caller doesn't call |
114 | * inode_need_compress() to check if we really need to |
115 | * compress. |
116 | * |
117 | * Not a big deal, just need to inform caller that we |
118 | * haven't allocated any pages yet. |
119 | */ |
120 | *out_pages = 0; |
121 | return -E2BIG; |
122 | } |
123 | } |
124 | |
125 | static int compression_decompress_bio(struct list_head *ws, |
126 | struct compressed_bio *cb) |
127 | { |
128 | switch (cb->compress_type) { |
129 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); |
130 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); |
131 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); |
132 | case BTRFS_COMPRESS_NONE: |
133 | default: |
134 | /* |
135 | * This can't happen, the type is validated several times |
136 | * before we get here. |
137 | */ |
138 | BUG(); |
139 | } |
140 | } |
141 | |
142 | static int compression_decompress(int type, struct list_head *ws, |
143 | const u8 *data_in, struct page *dest_page, |
144 | unsigned long start_byte, size_t srclen, size_t destlen) |
145 | { |
146 | switch (type) { |
147 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, |
148 | start_byte, srclen, destlen); |
149 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, |
150 | start_byte, srclen, destlen); |
151 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, |
152 | start_byte, srclen, destlen); |
153 | case BTRFS_COMPRESS_NONE: |
154 | default: |
155 | /* |
156 | * This can't happen, the type is validated several times |
157 | * before we get here. |
158 | */ |
159 | BUG(); |
160 | } |
161 | } |
162 | |
163 | static void btrfs_free_compressed_pages(struct compressed_bio *cb) |
164 | { |
165 | for (unsigned int i = 0; i < cb->nr_pages; i++) |
166 | put_page(page: cb->compressed_pages[i]); |
167 | kfree(objp: cb->compressed_pages); |
168 | } |
169 | |
170 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
171 | |
172 | static void end_compressed_bio_read(struct btrfs_bio *bbio) |
173 | { |
174 | struct compressed_bio *cb = to_compressed_bio(bbio); |
175 | blk_status_t status = bbio->bio.bi_status; |
176 | |
177 | if (!status) |
178 | status = errno_to_blk_status(errno: btrfs_decompress_bio(cb)); |
179 | |
180 | btrfs_free_compressed_pages(cb); |
181 | btrfs_bio_end_io(bbio: cb->orig_bbio, status); |
182 | bio_put(&bbio->bio); |
183 | } |
184 | |
185 | /* |
186 | * Clear the writeback bits on all of the file |
187 | * pages for a compressed write |
188 | */ |
189 | static noinline void end_compressed_writeback(const struct compressed_bio *cb) |
190 | { |
191 | struct inode *inode = &cb->bbio.inode->vfs_inode; |
192 | struct btrfs_fs_info *fs_info = btrfs_sb(sb: inode->i_sb); |
193 | unsigned long index = cb->start >> PAGE_SHIFT; |
194 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; |
195 | struct folio_batch fbatch; |
196 | const int error = blk_status_to_errno(status: cb->bbio.bio.bi_status); |
197 | int i; |
198 | int ret; |
199 | |
200 | if (error) |
201 | mapping_set_error(mapping: inode->i_mapping, error); |
202 | |
203 | folio_batch_init(fbatch: &fbatch); |
204 | while (index <= end_index) { |
205 | ret = filemap_get_folios(mapping: inode->i_mapping, start: &index, end: end_index, |
206 | fbatch: &fbatch); |
207 | |
208 | if (ret == 0) |
209 | return; |
210 | |
211 | for (i = 0; i < ret; i++) { |
212 | struct folio *folio = fbatch.folios[i]; |
213 | |
214 | btrfs_page_clamp_clear_writeback(fs_info, page: &folio->page, |
215 | start: cb->start, len: cb->len); |
216 | } |
217 | folio_batch_release(fbatch: &fbatch); |
218 | } |
219 | /* the inode may be gone now */ |
220 | } |
221 | |
222 | static void btrfs_finish_compressed_write_work(struct work_struct *work) |
223 | { |
224 | struct compressed_bio *cb = |
225 | container_of(work, struct compressed_bio, write_end_work); |
226 | |
227 | btrfs_finish_ordered_extent(ordered: cb->bbio.ordered, NULL, file_offset: cb->start, len: cb->len, |
228 | uptodate: cb->bbio.bio.bi_status == BLK_STS_OK); |
229 | |
230 | if (cb->writeback) |
231 | end_compressed_writeback(cb); |
232 | /* Note, our inode could be gone now */ |
233 | |
234 | btrfs_free_compressed_pages(cb); |
235 | bio_put(&cb->bbio.bio); |
236 | } |
237 | |
238 | /* |
239 | * Do the cleanup once all the compressed pages hit the disk. This will clear |
240 | * writeback on the file pages and free the compressed pages. |
241 | * |
242 | * This also calls the writeback end hooks for the file pages so that metadata |
243 | * and checksums can be updated in the file. |
244 | */ |
245 | static void end_compressed_bio_write(struct btrfs_bio *bbio) |
246 | { |
247 | struct compressed_bio *cb = to_compressed_bio(bbio); |
248 | struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info; |
249 | |
250 | queue_work(wq: fs_info->compressed_write_workers, work: &cb->write_end_work); |
251 | } |
252 | |
253 | static void btrfs_add_compressed_bio_pages(struct compressed_bio *cb) |
254 | { |
255 | struct bio *bio = &cb->bbio.bio; |
256 | u32 offset = 0; |
257 | |
258 | while (offset < cb->compressed_len) { |
259 | u32 len = min_t(u32, cb->compressed_len - offset, PAGE_SIZE); |
260 | |
261 | /* Maximum compressed extent is smaller than bio size limit. */ |
262 | __bio_add_page(bio, page: cb->compressed_pages[offset >> PAGE_SHIFT], |
263 | len, off: 0); |
264 | offset += len; |
265 | } |
266 | } |
267 | |
268 | /* |
269 | * worker function to build and submit bios for previously compressed pages. |
270 | * The corresponding pages in the inode should be marked for writeback |
271 | * and the compressed pages should have a reference on them for dropping |
272 | * when the IO is complete. |
273 | * |
274 | * This also checksums the file bytes and gets things ready for |
275 | * the end io hooks. |
276 | */ |
277 | void btrfs_submit_compressed_write(struct btrfs_ordered_extent *ordered, |
278 | struct page **compressed_pages, |
279 | unsigned int nr_pages, |
280 | blk_opf_t write_flags, |
281 | bool writeback) |
282 | { |
283 | struct btrfs_inode *inode = BTRFS_I(inode: ordered->inode); |
284 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
285 | struct compressed_bio *cb; |
286 | |
287 | ASSERT(IS_ALIGNED(ordered->file_offset, fs_info->sectorsize)); |
288 | ASSERT(IS_ALIGNED(ordered->num_bytes, fs_info->sectorsize)); |
289 | |
290 | cb = alloc_compressed_bio(inode, start: ordered->file_offset, |
291 | op: REQ_OP_WRITE | write_flags, |
292 | end_io: end_compressed_bio_write); |
293 | cb->start = ordered->file_offset; |
294 | cb->len = ordered->num_bytes; |
295 | cb->compressed_pages = compressed_pages; |
296 | cb->compressed_len = ordered->disk_num_bytes; |
297 | cb->writeback = writeback; |
298 | INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); |
299 | cb->nr_pages = nr_pages; |
300 | cb->bbio.bio.bi_iter.bi_sector = ordered->disk_bytenr >> SECTOR_SHIFT; |
301 | cb->bbio.ordered = ordered; |
302 | btrfs_add_compressed_bio_pages(cb); |
303 | |
304 | btrfs_submit_bio(bbio: &cb->bbio, mirror_num: 0); |
305 | } |
306 | |
307 | /* |
308 | * Add extra pages in the same compressed file extent so that we don't need to |
309 | * re-read the same extent again and again. |
310 | * |
311 | * NOTE: this won't work well for subpage, as for subpage read, we lock the |
312 | * full page then submit bio for each compressed/regular extents. |
313 | * |
314 | * This means, if we have several sectors in the same page points to the same |
315 | * on-disk compressed data, we will re-read the same extent many times and |
316 | * this function can only help for the next page. |
317 | */ |
318 | static noinline int add_ra_bio_pages(struct inode *inode, |
319 | u64 compressed_end, |
320 | struct compressed_bio *cb, |
321 | int *memstall, unsigned long *pflags) |
322 | { |
323 | struct btrfs_fs_info *fs_info = btrfs_sb(sb: inode->i_sb); |
324 | unsigned long end_index; |
325 | struct bio *orig_bio = &cb->orig_bbio->bio; |
326 | u64 cur = cb->orig_bbio->file_offset + orig_bio->bi_iter.bi_size; |
327 | u64 isize = i_size_read(inode); |
328 | int ret; |
329 | struct page *page; |
330 | struct extent_map *em; |
331 | struct address_space *mapping = inode->i_mapping; |
332 | struct extent_map_tree *em_tree; |
333 | struct extent_io_tree *tree; |
334 | int sectors_missed = 0; |
335 | |
336 | em_tree = &BTRFS_I(inode)->extent_tree; |
337 | tree = &BTRFS_I(inode)->io_tree; |
338 | |
339 | if (isize == 0) |
340 | return 0; |
341 | |
342 | /* |
343 | * For current subpage support, we only support 64K page size, |
344 | * which means maximum compressed extent size (128K) is just 2x page |
345 | * size. |
346 | * This makes readahead less effective, so here disable readahead for |
347 | * subpage for now, until full compressed write is supported. |
348 | */ |
349 | if (btrfs_sb(sb: inode->i_sb)->sectorsize < PAGE_SIZE) |
350 | return 0; |
351 | |
352 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
353 | |
354 | while (cur < compressed_end) { |
355 | u64 page_end; |
356 | u64 pg_index = cur >> PAGE_SHIFT; |
357 | u32 add_size; |
358 | |
359 | if (pg_index > end_index) |
360 | break; |
361 | |
362 | page = xa_load(&mapping->i_pages, index: pg_index); |
363 | if (page && !xa_is_value(entry: page)) { |
364 | sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> |
365 | fs_info->sectorsize_bits; |
366 | |
367 | /* Beyond threshold, no need to continue */ |
368 | if (sectors_missed > 4) |
369 | break; |
370 | |
371 | /* |
372 | * Jump to next page start as we already have page for |
373 | * current offset. |
374 | */ |
375 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; |
376 | continue; |
377 | } |
378 | |
379 | page = __page_cache_alloc(gfp: mapping_gfp_constraint(mapping, |
380 | gfp_mask: ~__GFP_FS)); |
381 | if (!page) |
382 | break; |
383 | |
384 | if (add_to_page_cache_lru(page, mapping, index: pg_index, GFP_NOFS)) { |
385 | put_page(page); |
386 | /* There is already a page, skip to page end */ |
387 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; |
388 | continue; |
389 | } |
390 | |
391 | if (!*memstall && PageWorkingset(page)) { |
392 | psi_memstall_enter(flags: pflags); |
393 | *memstall = 1; |
394 | } |
395 | |
396 | ret = set_page_extent_mapped(page); |
397 | if (ret < 0) { |
398 | unlock_page(page); |
399 | put_page(page); |
400 | break; |
401 | } |
402 | |
403 | page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; |
404 | lock_extent(tree, start: cur, end: page_end, NULL); |
405 | read_lock(&em_tree->lock); |
406 | em = lookup_extent_mapping(tree: em_tree, start: cur, len: page_end + 1 - cur); |
407 | read_unlock(&em_tree->lock); |
408 | |
409 | /* |
410 | * At this point, we have a locked page in the page cache for |
411 | * these bytes in the file. But, we have to make sure they map |
412 | * to this compressed extent on disk. |
413 | */ |
414 | if (!em || cur < em->start || |
415 | (cur + fs_info->sectorsize > extent_map_end(em)) || |
416 | (em->block_start >> SECTOR_SHIFT) != orig_bio->bi_iter.bi_sector) { |
417 | free_extent_map(em); |
418 | unlock_extent(tree, start: cur, end: page_end, NULL); |
419 | unlock_page(page); |
420 | put_page(page); |
421 | break; |
422 | } |
423 | free_extent_map(em); |
424 | |
425 | if (page->index == end_index) { |
426 | size_t zero_offset = offset_in_page(isize); |
427 | |
428 | if (zero_offset) { |
429 | int zeros; |
430 | zeros = PAGE_SIZE - zero_offset; |
431 | memzero_page(page, offset: zero_offset, len: zeros); |
432 | } |
433 | } |
434 | |
435 | add_size = min(em->start + em->len, page_end + 1) - cur; |
436 | ret = bio_add_page(bio: orig_bio, page, len: add_size, offset_in_page(cur)); |
437 | if (ret != add_size) { |
438 | unlock_extent(tree, start: cur, end: page_end, NULL); |
439 | unlock_page(page); |
440 | put_page(page); |
441 | break; |
442 | } |
443 | /* |
444 | * If it's subpage, we also need to increase its |
445 | * subpage::readers number, as at endio we will decrease |
446 | * subpage::readers and to unlock the page. |
447 | */ |
448 | if (fs_info->sectorsize < PAGE_SIZE) |
449 | btrfs_subpage_start_reader(fs_info, page, start: cur, len: add_size); |
450 | put_page(page); |
451 | cur += add_size; |
452 | } |
453 | return 0; |
454 | } |
455 | |
456 | /* |
457 | * for a compressed read, the bio we get passed has all the inode pages |
458 | * in it. We don't actually do IO on those pages but allocate new ones |
459 | * to hold the compressed pages on disk. |
460 | * |
461 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
462 | * bio->bi_io_vec points to all of the inode pages |
463 | * |
464 | * After the compressed pages are read, we copy the bytes into the |
465 | * bio we were passed and then call the bio end_io calls |
466 | */ |
467 | void btrfs_submit_compressed_read(struct btrfs_bio *bbio) |
468 | { |
469 | struct btrfs_inode *inode = bbio->inode; |
470 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
471 | struct extent_map_tree *em_tree = &inode->extent_tree; |
472 | struct compressed_bio *cb; |
473 | unsigned int compressed_len; |
474 | u64 file_offset = bbio->file_offset; |
475 | u64 em_len; |
476 | u64 em_start; |
477 | struct extent_map *em; |
478 | unsigned long pflags; |
479 | int memstall = 0; |
480 | blk_status_t ret; |
481 | int ret2; |
482 | |
483 | /* we need the actual starting offset of this extent in the file */ |
484 | read_lock(&em_tree->lock); |
485 | em = lookup_extent_mapping(tree: em_tree, start: file_offset, len: fs_info->sectorsize); |
486 | read_unlock(&em_tree->lock); |
487 | if (!em) { |
488 | ret = BLK_STS_IOERR; |
489 | goto out; |
490 | } |
491 | |
492 | ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); |
493 | compressed_len = em->block_len; |
494 | |
495 | cb = alloc_compressed_bio(inode, start: file_offset, op: REQ_OP_READ, |
496 | end_io: end_compressed_bio_read); |
497 | |
498 | cb->start = em->orig_start; |
499 | em_len = em->len; |
500 | em_start = em->start; |
501 | |
502 | cb->len = bbio->bio.bi_iter.bi_size; |
503 | cb->compressed_len = compressed_len; |
504 | cb->compress_type = em->compress_type; |
505 | cb->orig_bbio = bbio; |
506 | |
507 | free_extent_map(em); |
508 | |
509 | cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
510 | cb->compressed_pages = kcalloc(n: cb->nr_pages, size: sizeof(struct page *), GFP_NOFS); |
511 | if (!cb->compressed_pages) { |
512 | ret = BLK_STS_RESOURCE; |
513 | goto out_free_bio; |
514 | } |
515 | |
516 | ret2 = btrfs_alloc_page_array(nr_pages: cb->nr_pages, page_array: cb->compressed_pages); |
517 | if (ret2) { |
518 | ret = BLK_STS_RESOURCE; |
519 | goto out_free_compressed_pages; |
520 | } |
521 | |
522 | add_ra_bio_pages(inode: &inode->vfs_inode, compressed_end: em_start + em_len, cb, memstall: &memstall, |
523 | pflags: &pflags); |
524 | |
525 | /* include any pages we added in add_ra-bio_pages */ |
526 | cb->len = bbio->bio.bi_iter.bi_size; |
527 | cb->bbio.bio.bi_iter.bi_sector = bbio->bio.bi_iter.bi_sector; |
528 | btrfs_add_compressed_bio_pages(cb); |
529 | |
530 | if (memstall) |
531 | psi_memstall_leave(flags: &pflags); |
532 | |
533 | btrfs_submit_bio(bbio: &cb->bbio, mirror_num: 0); |
534 | return; |
535 | |
536 | out_free_compressed_pages: |
537 | kfree(objp: cb->compressed_pages); |
538 | out_free_bio: |
539 | bio_put(&cb->bbio.bio); |
540 | out: |
541 | btrfs_bio_end_io(bbio, status: ret); |
542 | } |
543 | |
544 | /* |
545 | * Heuristic uses systematic sampling to collect data from the input data |
546 | * range, the logic can be tuned by the following constants: |
547 | * |
548 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample |
549 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected |
550 | */ |
551 | #define SAMPLING_READ_SIZE (16) |
552 | #define SAMPLING_INTERVAL (256) |
553 | |
554 | /* |
555 | * For statistical analysis of the input data we consider bytes that form a |
556 | * Galois Field of 256 objects. Each object has an attribute count, ie. how |
557 | * many times the object appeared in the sample. |
558 | */ |
559 | #define BUCKET_SIZE (256) |
560 | |
561 | /* |
562 | * The size of the sample is based on a statistical sampling rule of thumb. |
563 | * The common way is to perform sampling tests as long as the number of |
564 | * elements in each cell is at least 5. |
565 | * |
566 | * Instead of 5, we choose 32 to obtain more accurate results. |
567 | * If the data contain the maximum number of symbols, which is 256, we obtain a |
568 | * sample size bound by 8192. |
569 | * |
570 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes |
571 | * from up to 512 locations. |
572 | */ |
573 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ |
574 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) |
575 | |
576 | struct bucket_item { |
577 | u32 count; |
578 | }; |
579 | |
580 | struct heuristic_ws { |
581 | /* Partial copy of input data */ |
582 | u8 *sample; |
583 | u32 sample_size; |
584 | /* Buckets store counters for each byte value */ |
585 | struct bucket_item *bucket; |
586 | /* Sorting buffer */ |
587 | struct bucket_item *bucket_b; |
588 | struct list_head list; |
589 | }; |
590 | |
591 | static struct workspace_manager heuristic_wsm; |
592 | |
593 | static void free_heuristic_ws(struct list_head *ws) |
594 | { |
595 | struct heuristic_ws *workspace; |
596 | |
597 | workspace = list_entry(ws, struct heuristic_ws, list); |
598 | |
599 | kvfree(addr: workspace->sample); |
600 | kfree(objp: workspace->bucket); |
601 | kfree(objp: workspace->bucket_b); |
602 | kfree(objp: workspace); |
603 | } |
604 | |
605 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
606 | { |
607 | struct heuristic_ws *ws; |
608 | |
609 | ws = kzalloc(size: sizeof(*ws), GFP_KERNEL); |
610 | if (!ws) |
611 | return ERR_PTR(error: -ENOMEM); |
612 | |
613 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
614 | if (!ws->sample) |
615 | goto fail; |
616 | |
617 | ws->bucket = kcalloc(BUCKET_SIZE, size: sizeof(*ws->bucket), GFP_KERNEL); |
618 | if (!ws->bucket) |
619 | goto fail; |
620 | |
621 | ws->bucket_b = kcalloc(BUCKET_SIZE, size: sizeof(*ws->bucket_b), GFP_KERNEL); |
622 | if (!ws->bucket_b) |
623 | goto fail; |
624 | |
625 | INIT_LIST_HEAD(list: &ws->list); |
626 | return &ws->list; |
627 | fail: |
628 | free_heuristic_ws(ws: &ws->list); |
629 | return ERR_PTR(error: -ENOMEM); |
630 | } |
631 | |
632 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
633 | .workspace_manager = &heuristic_wsm, |
634 | }; |
635 | |
636 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
637 | /* The heuristic is represented as compression type 0 */ |
638 | &btrfs_heuristic_compress, |
639 | &btrfs_zlib_compress, |
640 | &btrfs_lzo_compress, |
641 | &btrfs_zstd_compress, |
642 | }; |
643 | |
644 | static struct list_head *alloc_workspace(int type, unsigned int level) |
645 | { |
646 | switch (type) { |
647 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); |
648 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); |
649 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); |
650 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); |
651 | default: |
652 | /* |
653 | * This can't happen, the type is validated several times |
654 | * before we get here. |
655 | */ |
656 | BUG(); |
657 | } |
658 | } |
659 | |
660 | static void free_workspace(int type, struct list_head *ws) |
661 | { |
662 | switch (type) { |
663 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); |
664 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); |
665 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); |
666 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); |
667 | default: |
668 | /* |
669 | * This can't happen, the type is validated several times |
670 | * before we get here. |
671 | */ |
672 | BUG(); |
673 | } |
674 | } |
675 | |
676 | static void btrfs_init_workspace_manager(int type) |
677 | { |
678 | struct workspace_manager *wsm; |
679 | struct list_head *workspace; |
680 | |
681 | wsm = btrfs_compress_op[type]->workspace_manager; |
682 | INIT_LIST_HEAD(list: &wsm->idle_ws); |
683 | spin_lock_init(&wsm->ws_lock); |
684 | atomic_set(v: &wsm->total_ws, i: 0); |
685 | init_waitqueue_head(&wsm->ws_wait); |
686 | |
687 | /* |
688 | * Preallocate one workspace for each compression type so we can |
689 | * guarantee forward progress in the worst case |
690 | */ |
691 | workspace = alloc_workspace(type, level: 0); |
692 | if (IS_ERR(ptr: workspace)) { |
693 | pr_warn( |
694 | "BTRFS: cannot preallocate compression workspace, will try later\n" ); |
695 | } else { |
696 | atomic_set(v: &wsm->total_ws, i: 1); |
697 | wsm->free_ws = 1; |
698 | list_add(new: workspace, head: &wsm->idle_ws); |
699 | } |
700 | } |
701 | |
702 | static void btrfs_cleanup_workspace_manager(int type) |
703 | { |
704 | struct workspace_manager *wsman; |
705 | struct list_head *ws; |
706 | |
707 | wsman = btrfs_compress_op[type]->workspace_manager; |
708 | while (!list_empty(head: &wsman->idle_ws)) { |
709 | ws = wsman->idle_ws.next; |
710 | list_del(entry: ws); |
711 | free_workspace(type, ws); |
712 | atomic_dec(v: &wsman->total_ws); |
713 | } |
714 | } |
715 | |
716 | /* |
717 | * This finds an available workspace or allocates a new one. |
718 | * If it's not possible to allocate a new one, waits until there's one. |
719 | * Preallocation makes a forward progress guarantees and we do not return |
720 | * errors. |
721 | */ |
722 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
723 | { |
724 | struct workspace_manager *wsm; |
725 | struct list_head *workspace; |
726 | int cpus = num_online_cpus(); |
727 | unsigned nofs_flag; |
728 | struct list_head *idle_ws; |
729 | spinlock_t *ws_lock; |
730 | atomic_t *total_ws; |
731 | wait_queue_head_t *ws_wait; |
732 | int *free_ws; |
733 | |
734 | wsm = btrfs_compress_op[type]->workspace_manager; |
735 | idle_ws = &wsm->idle_ws; |
736 | ws_lock = &wsm->ws_lock; |
737 | total_ws = &wsm->total_ws; |
738 | ws_wait = &wsm->ws_wait; |
739 | free_ws = &wsm->free_ws; |
740 | |
741 | again: |
742 | spin_lock(lock: ws_lock); |
743 | if (!list_empty(head: idle_ws)) { |
744 | workspace = idle_ws->next; |
745 | list_del(entry: workspace); |
746 | (*free_ws)--; |
747 | spin_unlock(lock: ws_lock); |
748 | return workspace; |
749 | |
750 | } |
751 | if (atomic_read(v: total_ws) > cpus) { |
752 | DEFINE_WAIT(wait); |
753 | |
754 | spin_unlock(lock: ws_lock); |
755 | prepare_to_wait(wq_head: ws_wait, wq_entry: &wait, TASK_UNINTERRUPTIBLE); |
756 | if (atomic_read(v: total_ws) > cpus && !*free_ws) |
757 | schedule(); |
758 | finish_wait(wq_head: ws_wait, wq_entry: &wait); |
759 | goto again; |
760 | } |
761 | atomic_inc(v: total_ws); |
762 | spin_unlock(lock: ws_lock); |
763 | |
764 | /* |
765 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have |
766 | * to turn it off here because we might get called from the restricted |
767 | * context of btrfs_compress_bio/btrfs_compress_pages |
768 | */ |
769 | nofs_flag = memalloc_nofs_save(); |
770 | workspace = alloc_workspace(type, level); |
771 | memalloc_nofs_restore(flags: nofs_flag); |
772 | |
773 | if (IS_ERR(ptr: workspace)) { |
774 | atomic_dec(v: total_ws); |
775 | wake_up(ws_wait); |
776 | |
777 | /* |
778 | * Do not return the error but go back to waiting. There's a |
779 | * workspace preallocated for each type and the compression |
780 | * time is bounded so we get to a workspace eventually. This |
781 | * makes our caller's life easier. |
782 | * |
783 | * To prevent silent and low-probability deadlocks (when the |
784 | * initial preallocation fails), check if there are any |
785 | * workspaces at all. |
786 | */ |
787 | if (atomic_read(v: total_ws) == 0) { |
788 | static DEFINE_RATELIMIT_STATE(_rs, |
789 | /* once per minute */ 60 * HZ, |
790 | /* no burst */ 1); |
791 | |
792 | if (__ratelimit(&_rs)) { |
793 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n" ); |
794 | } |
795 | } |
796 | goto again; |
797 | } |
798 | return workspace; |
799 | } |
800 | |
801 | static struct list_head *get_workspace(int type, int level) |
802 | { |
803 | switch (type) { |
804 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
805 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
806 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
807 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
808 | default: |
809 | /* |
810 | * This can't happen, the type is validated several times |
811 | * before we get here. |
812 | */ |
813 | BUG(); |
814 | } |
815 | } |
816 | |
817 | /* |
818 | * put a workspace struct back on the list or free it if we have enough |
819 | * idle ones sitting around |
820 | */ |
821 | void btrfs_put_workspace(int type, struct list_head *ws) |
822 | { |
823 | struct workspace_manager *wsm; |
824 | struct list_head *idle_ws; |
825 | spinlock_t *ws_lock; |
826 | atomic_t *total_ws; |
827 | wait_queue_head_t *ws_wait; |
828 | int *free_ws; |
829 | |
830 | wsm = btrfs_compress_op[type]->workspace_manager; |
831 | idle_ws = &wsm->idle_ws; |
832 | ws_lock = &wsm->ws_lock; |
833 | total_ws = &wsm->total_ws; |
834 | ws_wait = &wsm->ws_wait; |
835 | free_ws = &wsm->free_ws; |
836 | |
837 | spin_lock(lock: ws_lock); |
838 | if (*free_ws <= num_online_cpus()) { |
839 | list_add(new: ws, head: idle_ws); |
840 | (*free_ws)++; |
841 | spin_unlock(lock: ws_lock); |
842 | goto wake; |
843 | } |
844 | spin_unlock(lock: ws_lock); |
845 | |
846 | free_workspace(type, ws); |
847 | atomic_dec(v: total_ws); |
848 | wake: |
849 | cond_wake_up(wq: ws_wait); |
850 | } |
851 | |
852 | static void put_workspace(int type, struct list_head *ws) |
853 | { |
854 | switch (type) { |
855 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
856 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); |
857 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); |
858 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
859 | default: |
860 | /* |
861 | * This can't happen, the type is validated several times |
862 | * before we get here. |
863 | */ |
864 | BUG(); |
865 | } |
866 | } |
867 | |
868 | /* |
869 | * Adjust @level according to the limits of the compression algorithm or |
870 | * fallback to default |
871 | */ |
872 | static unsigned int btrfs_compress_set_level(int type, unsigned level) |
873 | { |
874 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; |
875 | |
876 | if (level == 0) |
877 | level = ops->default_level; |
878 | else |
879 | level = min(level, ops->max_level); |
880 | |
881 | return level; |
882 | } |
883 | |
884 | /* |
885 | * Given an address space and start and length, compress the bytes into @pages |
886 | * that are allocated on demand. |
887 | * |
888 | * @type_level is encoded algorithm and level, where level 0 means whatever |
889 | * default the algorithm chooses and is opaque here; |
890 | * - compression algo are 0-3 |
891 | * - the level are bits 4-7 |
892 | * |
893 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
894 | * and returns number of actually allocated pages |
895 | * |
896 | * @total_in is used to return the number of bytes actually read. It |
897 | * may be smaller than the input length if we had to exit early because we |
898 | * ran out of room in the pages array or because we cross the |
899 | * max_out threshold. |
900 | * |
901 | * @total_out is an in/out parameter, must be set to the input length and will |
902 | * be also used to return the total number of compressed bytes |
903 | */ |
904 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
905 | u64 start, struct page **pages, |
906 | unsigned long *out_pages, |
907 | unsigned long *total_in, |
908 | unsigned long *total_out) |
909 | { |
910 | int type = btrfs_compress_type(type_level); |
911 | int level = btrfs_compress_level(type_level); |
912 | struct list_head *workspace; |
913 | int ret; |
914 | |
915 | level = btrfs_compress_set_level(type, level); |
916 | workspace = get_workspace(type, level); |
917 | ret = compression_compress_pages(type, ws: workspace, mapping, start, pages, |
918 | out_pages, total_in, total_out); |
919 | put_workspace(type, ws: workspace); |
920 | return ret; |
921 | } |
922 | |
923 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
924 | { |
925 | struct list_head *workspace; |
926 | int ret; |
927 | int type = cb->compress_type; |
928 | |
929 | workspace = get_workspace(type, level: 0); |
930 | ret = compression_decompress_bio(ws: workspace, cb); |
931 | put_workspace(type, ws: workspace); |
932 | |
933 | if (!ret) |
934 | zero_fill_bio(bio: &cb->orig_bbio->bio); |
935 | return ret; |
936 | } |
937 | |
938 | /* |
939 | * a less complex decompression routine. Our compressed data fits in a |
940 | * single page, and we want to read a single page out of it. |
941 | * start_byte tells us the offset into the compressed data we're interested in |
942 | */ |
943 | int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page, |
944 | unsigned long start_byte, size_t srclen, size_t destlen) |
945 | { |
946 | struct list_head *workspace; |
947 | int ret; |
948 | |
949 | workspace = get_workspace(type, level: 0); |
950 | ret = compression_decompress(type, ws: workspace, data_in, dest_page, |
951 | start_byte, srclen, destlen); |
952 | put_workspace(type, ws: workspace); |
953 | |
954 | return ret; |
955 | } |
956 | |
957 | int __init btrfs_init_compress(void) |
958 | { |
959 | if (bioset_init(&btrfs_compressed_bioset, BIO_POOL_SIZE, |
960 | offsetof(struct compressed_bio, bbio.bio), |
961 | flags: BIOSET_NEED_BVECS)) |
962 | return -ENOMEM; |
963 | btrfs_init_workspace_manager(type: BTRFS_COMPRESS_NONE); |
964 | btrfs_init_workspace_manager(type: BTRFS_COMPRESS_ZLIB); |
965 | btrfs_init_workspace_manager(type: BTRFS_COMPRESS_LZO); |
966 | zstd_init_workspace_manager(); |
967 | return 0; |
968 | } |
969 | |
970 | void __cold btrfs_exit_compress(void) |
971 | { |
972 | btrfs_cleanup_workspace_manager(type: BTRFS_COMPRESS_NONE); |
973 | btrfs_cleanup_workspace_manager(type: BTRFS_COMPRESS_ZLIB); |
974 | btrfs_cleanup_workspace_manager(type: BTRFS_COMPRESS_LZO); |
975 | zstd_cleanup_workspace_manager(); |
976 | bioset_exit(&btrfs_compressed_bioset); |
977 | } |
978 | |
979 | /* |
980 | * Copy decompressed data from working buffer to pages. |
981 | * |
982 | * @buf: The decompressed data buffer |
983 | * @buf_len: The decompressed data length |
984 | * @decompressed: Number of bytes that are already decompressed inside the |
985 | * compressed extent |
986 | * @cb: The compressed extent descriptor |
987 | * @orig_bio: The original bio that the caller wants to read for |
988 | * |
989 | * An easier to understand graph is like below: |
990 | * |
991 | * |<- orig_bio ->| |<- orig_bio->| |
992 | * |<------- full decompressed extent ----->| |
993 | * |<----------- @cb range ---->| |
994 | * | |<-- @buf_len -->| |
995 | * |<--- @decompressed --->| |
996 | * |
997 | * Note that, @cb can be a subpage of the full decompressed extent, but |
998 | * @cb->start always has the same as the orig_file_offset value of the full |
999 | * decompressed extent. |
1000 | * |
1001 | * When reading compressed extent, we have to read the full compressed extent, |
1002 | * while @orig_bio may only want part of the range. |
1003 | * Thus this function will ensure only data covered by @orig_bio will be copied |
1004 | * to. |
1005 | * |
1006 | * Return 0 if we have copied all needed contents for @orig_bio. |
1007 | * Return >0 if we need continue decompress. |
1008 | */ |
1009 | int btrfs_decompress_buf2page(const char *buf, u32 buf_len, |
1010 | struct compressed_bio *cb, u32 decompressed) |
1011 | { |
1012 | struct bio *orig_bio = &cb->orig_bbio->bio; |
1013 | /* Offset inside the full decompressed extent */ |
1014 | u32 cur_offset; |
1015 | |
1016 | cur_offset = decompressed; |
1017 | /* The main loop to do the copy */ |
1018 | while (cur_offset < decompressed + buf_len) { |
1019 | struct bio_vec bvec; |
1020 | size_t copy_len; |
1021 | u32 copy_start; |
1022 | /* Offset inside the full decompressed extent */ |
1023 | u32 bvec_offset; |
1024 | |
1025 | bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); |
1026 | /* |
1027 | * cb->start may underflow, but subtracting that value can still |
1028 | * give us correct offset inside the full decompressed extent. |
1029 | */ |
1030 | bvec_offset = page_offset(page: bvec.bv_page) + bvec.bv_offset - cb->start; |
1031 | |
1032 | /* Haven't reached the bvec range, exit */ |
1033 | if (decompressed + buf_len <= bvec_offset) |
1034 | return 1; |
1035 | |
1036 | copy_start = max(cur_offset, bvec_offset); |
1037 | copy_len = min(bvec_offset + bvec.bv_len, |
1038 | decompressed + buf_len) - copy_start; |
1039 | ASSERT(copy_len); |
1040 | |
1041 | /* |
1042 | * Extra range check to ensure we didn't go beyond |
1043 | * @buf + @buf_len. |
1044 | */ |
1045 | ASSERT(copy_start - decompressed < buf_len); |
1046 | memcpy_to_page(page: bvec.bv_page, offset: bvec.bv_offset, |
1047 | from: buf + copy_start - decompressed, len: copy_len); |
1048 | cur_offset += copy_len; |
1049 | |
1050 | bio_advance(bio: orig_bio, nbytes: copy_len); |
1051 | /* Finished the bio */ |
1052 | if (!orig_bio->bi_iter.bi_size) |
1053 | return 0; |
1054 | } |
1055 | return 1; |
1056 | } |
1057 | |
1058 | /* |
1059 | * Shannon Entropy calculation |
1060 | * |
1061 | * Pure byte distribution analysis fails to determine compressibility of data. |
1062 | * Try calculating entropy to estimate the average minimum number of bits |
1063 | * needed to encode the sampled data. |
1064 | * |
1065 | * For convenience, return the percentage of needed bits, instead of amount of |
1066 | * bits directly. |
1067 | * |
1068 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy |
1069 | * and can be compressible with high probability |
1070 | * |
1071 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability |
1072 | * |
1073 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. |
1074 | */ |
1075 | #define ENTROPY_LVL_ACEPTABLE (65) |
1076 | #define ENTROPY_LVL_HIGH (80) |
1077 | |
1078 | /* |
1079 | * For increasead precision in shannon_entropy calculation, |
1080 | * let's do pow(n, M) to save more digits after comma: |
1081 | * |
1082 | * - maximum int bit length is 64 |
1083 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 |
1084 | * - 13 * 4 = 52 < 64 -> M = 4 |
1085 | * |
1086 | * So use pow(n, 4). |
1087 | */ |
1088 | static inline u32 ilog2_w(u64 n) |
1089 | { |
1090 | return ilog2(n * n * n * n); |
1091 | } |
1092 | |
1093 | static u32 shannon_entropy(struct heuristic_ws *ws) |
1094 | { |
1095 | const u32 entropy_max = 8 * ilog2_w(n: 2); |
1096 | u32 entropy_sum = 0; |
1097 | u32 p, p_base, sz_base; |
1098 | u32 i; |
1099 | |
1100 | sz_base = ilog2_w(n: ws->sample_size); |
1101 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { |
1102 | p = ws->bucket[i].count; |
1103 | p_base = ilog2_w(n: p); |
1104 | entropy_sum += p * (sz_base - p_base); |
1105 | } |
1106 | |
1107 | entropy_sum /= ws->sample_size; |
1108 | return entropy_sum * 100 / entropy_max; |
1109 | } |
1110 | |
1111 | #define RADIX_BASE 4U |
1112 | #define COUNTERS_SIZE (1U << RADIX_BASE) |
1113 | |
1114 | static u8 get4bits(u64 num, int shift) { |
1115 | u8 low4bits; |
1116 | |
1117 | num >>= shift; |
1118 | /* Reverse order */ |
1119 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); |
1120 | return low4bits; |
1121 | } |
1122 | |
1123 | /* |
1124 | * Use 4 bits as radix base |
1125 | * Use 16 u32 counters for calculating new position in buf array |
1126 | * |
1127 | * @array - array that will be sorted |
1128 | * @array_buf - buffer array to store sorting results |
1129 | * must be equal in size to @array |
1130 | * @num - array size |
1131 | */ |
1132 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
1133 | int num) |
1134 | { |
1135 | u64 max_num; |
1136 | u64 buf_num; |
1137 | u32 counters[COUNTERS_SIZE]; |
1138 | u32 new_addr; |
1139 | u32 addr; |
1140 | int bitlen; |
1141 | int shift; |
1142 | int i; |
1143 | |
1144 | /* |
1145 | * Try avoid useless loop iterations for small numbers stored in big |
1146 | * counters. Example: 48 33 4 ... in 64bit array |
1147 | */ |
1148 | max_num = array[0].count; |
1149 | for (i = 1; i < num; i++) { |
1150 | buf_num = array[i].count; |
1151 | if (buf_num > max_num) |
1152 | max_num = buf_num; |
1153 | } |
1154 | |
1155 | buf_num = ilog2(max_num); |
1156 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); |
1157 | |
1158 | shift = 0; |
1159 | while (shift < bitlen) { |
1160 | memset(counters, 0, sizeof(counters)); |
1161 | |
1162 | for (i = 0; i < num; i++) { |
1163 | buf_num = array[i].count; |
1164 | addr = get4bits(num: buf_num, shift); |
1165 | counters[addr]++; |
1166 | } |
1167 | |
1168 | for (i = 1; i < COUNTERS_SIZE; i++) |
1169 | counters[i] += counters[i - 1]; |
1170 | |
1171 | for (i = num - 1; i >= 0; i--) { |
1172 | buf_num = array[i].count; |
1173 | addr = get4bits(num: buf_num, shift); |
1174 | counters[addr]--; |
1175 | new_addr = counters[addr]; |
1176 | array_buf[new_addr] = array[i]; |
1177 | } |
1178 | |
1179 | shift += RADIX_BASE; |
1180 | |
1181 | /* |
1182 | * Normal radix expects to move data from a temporary array, to |
1183 | * the main one. But that requires some CPU time. Avoid that |
1184 | * by doing another sort iteration to original array instead of |
1185 | * memcpy() |
1186 | */ |
1187 | memset(counters, 0, sizeof(counters)); |
1188 | |
1189 | for (i = 0; i < num; i ++) { |
1190 | buf_num = array_buf[i].count; |
1191 | addr = get4bits(num: buf_num, shift); |
1192 | counters[addr]++; |
1193 | } |
1194 | |
1195 | for (i = 1; i < COUNTERS_SIZE; i++) |
1196 | counters[i] += counters[i - 1]; |
1197 | |
1198 | for (i = num - 1; i >= 0; i--) { |
1199 | buf_num = array_buf[i].count; |
1200 | addr = get4bits(num: buf_num, shift); |
1201 | counters[addr]--; |
1202 | new_addr = counters[addr]; |
1203 | array[new_addr] = array_buf[i]; |
1204 | } |
1205 | |
1206 | shift += RADIX_BASE; |
1207 | } |
1208 | } |
1209 | |
1210 | /* |
1211 | * Size of the core byte set - how many bytes cover 90% of the sample |
1212 | * |
1213 | * There are several types of structured binary data that use nearly all byte |
1214 | * values. The distribution can be uniform and counts in all buckets will be |
1215 | * nearly the same (eg. encrypted data). Unlikely to be compressible. |
1216 | * |
1217 | * Other possibility is normal (Gaussian) distribution, where the data could |
1218 | * be potentially compressible, but we have to take a few more steps to decide |
1219 | * how much. |
1220 | * |
1221 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, |
1222 | * compression algo can easy fix that |
1223 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high |
1224 | * probability is not compressible |
1225 | */ |
1226 | #define BYTE_CORE_SET_LOW (64) |
1227 | #define BYTE_CORE_SET_HIGH (200) |
1228 | |
1229 | static int byte_core_set_size(struct heuristic_ws *ws) |
1230 | { |
1231 | u32 i; |
1232 | u32 coreset_sum = 0; |
1233 | const u32 core_set_threshold = ws->sample_size * 90 / 100; |
1234 | struct bucket_item *bucket = ws->bucket; |
1235 | |
1236 | /* Sort in reverse order */ |
1237 | radix_sort(array: ws->bucket, array_buf: ws->bucket_b, BUCKET_SIZE); |
1238 | |
1239 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) |
1240 | coreset_sum += bucket[i].count; |
1241 | |
1242 | if (coreset_sum > core_set_threshold) |
1243 | return i; |
1244 | |
1245 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { |
1246 | coreset_sum += bucket[i].count; |
1247 | if (coreset_sum > core_set_threshold) |
1248 | break; |
1249 | } |
1250 | |
1251 | return i; |
1252 | } |
1253 | |
1254 | /* |
1255 | * Count byte values in buckets. |
1256 | * This heuristic can detect textual data (configs, xml, json, html, etc). |
1257 | * Because in most text-like data byte set is restricted to limited number of |
1258 | * possible characters, and that restriction in most cases makes data easy to |
1259 | * compress. |
1260 | * |
1261 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: |
1262 | * less - compressible |
1263 | * more - need additional analysis |
1264 | */ |
1265 | #define BYTE_SET_THRESHOLD (64) |
1266 | |
1267 | static u32 byte_set_size(const struct heuristic_ws *ws) |
1268 | { |
1269 | u32 i; |
1270 | u32 byte_set_size = 0; |
1271 | |
1272 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { |
1273 | if (ws->bucket[i].count > 0) |
1274 | byte_set_size++; |
1275 | } |
1276 | |
1277 | /* |
1278 | * Continue collecting count of byte values in buckets. If the byte |
1279 | * set size is bigger then the threshold, it's pointless to continue, |
1280 | * the detection technique would fail for this type of data. |
1281 | */ |
1282 | for (; i < BUCKET_SIZE; i++) { |
1283 | if (ws->bucket[i].count > 0) { |
1284 | byte_set_size++; |
1285 | if (byte_set_size > BYTE_SET_THRESHOLD) |
1286 | return byte_set_size; |
1287 | } |
1288 | } |
1289 | |
1290 | return byte_set_size; |
1291 | } |
1292 | |
1293 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1294 | { |
1295 | const u32 half_of_sample = ws->sample_size / 2; |
1296 | const u8 *data = ws->sample; |
1297 | |
1298 | return memcmp(p: &data[0], q: &data[half_of_sample], size: half_of_sample) == 0; |
1299 | } |
1300 | |
1301 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1302 | struct heuristic_ws *ws) |
1303 | { |
1304 | struct page *page; |
1305 | u64 index, index_end; |
1306 | u32 i, curr_sample_pos; |
1307 | u8 *in_data; |
1308 | |
1309 | /* |
1310 | * Compression handles the input data by chunks of 128KiB |
1311 | * (defined by BTRFS_MAX_UNCOMPRESSED) |
1312 | * |
1313 | * We do the same for the heuristic and loop over the whole range. |
1314 | * |
1315 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will |
1316 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. |
1317 | */ |
1318 | if (end - start > BTRFS_MAX_UNCOMPRESSED) |
1319 | end = start + BTRFS_MAX_UNCOMPRESSED; |
1320 | |
1321 | index = start >> PAGE_SHIFT; |
1322 | index_end = end >> PAGE_SHIFT; |
1323 | |
1324 | /* Don't miss unaligned end */ |
1325 | if (!PAGE_ALIGNED(end)) |
1326 | index_end++; |
1327 | |
1328 | curr_sample_pos = 0; |
1329 | while (index < index_end) { |
1330 | page = find_get_page(mapping: inode->i_mapping, offset: index); |
1331 | in_data = kmap_local_page(page); |
1332 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1333 | i = start % PAGE_SIZE; |
1334 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { |
1335 | /* Don't sample any garbage from the last page */ |
1336 | if (start > end - SAMPLING_READ_SIZE) |
1337 | break; |
1338 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], |
1339 | SAMPLING_READ_SIZE); |
1340 | i += SAMPLING_INTERVAL; |
1341 | start += SAMPLING_INTERVAL; |
1342 | curr_sample_pos += SAMPLING_READ_SIZE; |
1343 | } |
1344 | kunmap_local(in_data); |
1345 | put_page(page); |
1346 | |
1347 | index++; |
1348 | } |
1349 | |
1350 | ws->sample_size = curr_sample_pos; |
1351 | } |
1352 | |
1353 | /* |
1354 | * Compression heuristic. |
1355 | * |
1356 | * For now is's a naive and optimistic 'return true', we'll extend the logic to |
1357 | * quickly (compared to direct compression) detect data characteristics |
1358 | * (compressible/incompressible) to avoid wasting CPU time on incompressible |
1359 | * data. |
1360 | * |
1361 | * The following types of analysis can be performed: |
1362 | * - detect mostly zero data |
1363 | * - detect data with low "byte set" size (text, etc) |
1364 | * - detect data with low/high "core byte" set |
1365 | * |
1366 | * Return non-zero if the compression should be done, 0 otherwise. |
1367 | */ |
1368 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) |
1369 | { |
1370 | struct list_head *ws_list = get_workspace(type: 0, level: 0); |
1371 | struct heuristic_ws *ws; |
1372 | u32 i; |
1373 | u8 byte; |
1374 | int ret = 0; |
1375 | |
1376 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1377 | |
1378 | heuristic_collect_sample(inode, start, end, ws); |
1379 | |
1380 | if (sample_repeated_patterns(ws)) { |
1381 | ret = 1; |
1382 | goto out; |
1383 | } |
1384 | |
1385 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1386 | |
1387 | for (i = 0; i < ws->sample_size; i++) { |
1388 | byte = ws->sample[i]; |
1389 | ws->bucket[byte].count++; |
1390 | } |
1391 | |
1392 | i = byte_set_size(ws); |
1393 | if (i < BYTE_SET_THRESHOLD) { |
1394 | ret = 2; |
1395 | goto out; |
1396 | } |
1397 | |
1398 | i = byte_core_set_size(ws); |
1399 | if (i <= BYTE_CORE_SET_LOW) { |
1400 | ret = 3; |
1401 | goto out; |
1402 | } |
1403 | |
1404 | if (i >= BYTE_CORE_SET_HIGH) { |
1405 | ret = 0; |
1406 | goto out; |
1407 | } |
1408 | |
1409 | i = shannon_entropy(ws); |
1410 | if (i <= ENTROPY_LVL_ACEPTABLE) { |
1411 | ret = 4; |
1412 | goto out; |
1413 | } |
1414 | |
1415 | /* |
1416 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be |
1417 | * needed to give green light to compression. |
1418 | * |
1419 | * For now just assume that compression at that level is not worth the |
1420 | * resources because: |
1421 | * |
1422 | * 1. it is possible to defrag the data later |
1423 | * |
1424 | * 2. the data would turn out to be hardly compressible, eg. 150 byte |
1425 | * values, every bucket has counter at level ~54. The heuristic would |
1426 | * be confused. This can happen when data have some internal repeated |
1427 | * patterns like "abbacbbc...". This can be detected by analyzing |
1428 | * pairs of bytes, which is too costly. |
1429 | */ |
1430 | if (i < ENTROPY_LVL_HIGH) { |
1431 | ret = 5; |
1432 | goto out; |
1433 | } else { |
1434 | ret = 0; |
1435 | goto out; |
1436 | } |
1437 | |
1438 | out: |
1439 | put_workspace(type: 0, ws: ws_list); |
1440 | return ret; |
1441 | } |
1442 | |
1443 | /* |
1444 | * Convert the compression suffix (eg. after "zlib" starting with ":") to |
1445 | * level, unrecognized string will set the default level |
1446 | */ |
1447 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) |
1448 | { |
1449 | unsigned int level = 0; |
1450 | int ret; |
1451 | |
1452 | if (!type) |
1453 | return 0; |
1454 | |
1455 | if (str[0] == ':') { |
1456 | ret = kstrtouint(s: str + 1, base: 10, res: &level); |
1457 | if (ret) |
1458 | level = 0; |
1459 | } |
1460 | |
1461 | level = btrfs_compress_set_level(type, level); |
1462 | |
1463 | return level; |
1464 | } |
1465 | |