Commit | Line | Data |
---|---|---|
c1d7c514 | 1 | // SPDX-License-Identifier: GPL-2.0 |
c8b97818 CM |
2 | /* |
3 | * Copyright (C) 2008 Oracle. All rights reserved. | |
c8b97818 CM |
4 | */ |
5 | ||
6 | #include <linux/kernel.h> | |
7 | #include <linux/bio.h> | |
c8b97818 CM |
8 | #include <linux/file.h> |
9 | #include <linux/fs.h> | |
10 | #include <linux/pagemap.h> | |
a75b81c3 | 11 | #include <linux/pagevec.h> |
c8b97818 | 12 | #include <linux/highmem.h> |
e41d12f5 | 13 | #include <linux/kthread.h> |
c8b97818 CM |
14 | #include <linux/time.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/string.h> | |
c8b97818 | 17 | #include <linux/backing-dev.h> |
c8b97818 | 18 | #include <linux/writeback.h> |
4088a47e | 19 | #include <linux/psi.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
fe308533 | 21 | #include <linux/sched/mm.h> |
19562430 | 22 | #include <linux/log2.h> |
d5178578 | 23 | #include <crypto/hash.h> |
602cbe91 | 24 | #include "misc.h" |
c8b97818 | 25 | #include "ctree.h" |
ec8eb376 | 26 | #include "fs.h" |
c8b97818 CM |
27 | #include "disk-io.h" |
28 | #include "transaction.h" | |
29 | #include "btrfs_inode.h" | |
103c1972 | 30 | #include "bio.h" |
c8b97818 | 31 | #include "ordered-data.h" |
c8b97818 CM |
32 | #include "compression.h" |
33 | #include "extent_io.h" | |
34 | #include "extent_map.h" | |
6a404910 | 35 | #include "subpage.h" |
764c7c9a | 36 | #include "zoned.h" |
7c8ede16 | 37 | #include "file-item.h" |
7f0add25 | 38 | #include "super.h" |
c8b97818 | 39 | |
e128f9c3 DS |
40 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
41 | ||
42 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
43 | { | |
44 | switch (type) { | |
45 | case BTRFS_COMPRESS_ZLIB: | |
46 | case BTRFS_COMPRESS_LZO: | |
47 | case BTRFS_COMPRESS_ZSTD: | |
48 | case BTRFS_COMPRESS_NONE: | |
49 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
50 | default: |
51 | break; | |
e128f9c3 DS |
52 | } |
53 | ||
54 | return NULL; | |
55 | } | |
56 | ||
aa53e3bf JT |
57 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
58 | { | |
59 | int i; | |
60 | ||
61 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
62 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
63 | ||
64 | if (len < comp_len) | |
65 | continue; | |
66 | ||
67 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
68 | return true; | |
69 | } | |
70 | return false; | |
71 | } | |
72 | ||
1e4eb746 DS |
73 | static int compression_compress_pages(int type, struct list_head *ws, |
74 | struct address_space *mapping, u64 start, struct page **pages, | |
75 | unsigned long *out_pages, unsigned long *total_in, | |
76 | unsigned long *total_out) | |
77 | { | |
78 | switch (type) { | |
79 | case BTRFS_COMPRESS_ZLIB: | |
80 | return zlib_compress_pages(ws, mapping, start, pages, | |
81 | out_pages, total_in, total_out); | |
82 | case BTRFS_COMPRESS_LZO: | |
83 | return lzo_compress_pages(ws, mapping, start, pages, | |
84 | out_pages, total_in, total_out); | |
85 | case BTRFS_COMPRESS_ZSTD: | |
86 | return zstd_compress_pages(ws, mapping, start, pages, | |
87 | out_pages, total_in, total_out); | |
88 | case BTRFS_COMPRESS_NONE: | |
89 | default: | |
90 | /* | |
1d8ba9e7 QW |
91 | * This can happen when compression races with remount setting |
92 | * it to 'no compress', while caller doesn't call | |
93 | * inode_need_compress() to check if we really need to | |
94 | * compress. | |
95 | * | |
96 | * Not a big deal, just need to inform caller that we | |
97 | * haven't allocated any pages yet. | |
1e4eb746 | 98 | */ |
1d8ba9e7 | 99 | *out_pages = 0; |
1e4eb746 DS |
100 | return -E2BIG; |
101 | } | |
102 | } | |
103 | ||
4a9e803e SY |
104 | static int compression_decompress_bio(struct list_head *ws, |
105 | struct compressed_bio *cb) | |
1e4eb746 | 106 | { |
4a9e803e | 107 | switch (cb->compress_type) { |
1e4eb746 DS |
108 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); |
109 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
110 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
111 | case BTRFS_COMPRESS_NONE: | |
112 | default: | |
113 | /* | |
114 | * This can't happen, the type is validated several times | |
115 | * before we get here. | |
116 | */ | |
117 | BUG(); | |
118 | } | |
119 | } | |
120 | ||
121 | static int compression_decompress(int type, struct list_head *ws, | |
3e09b5b2 | 122 | const u8 *data_in, struct page *dest_page, |
1e4eb746 DS |
123 | unsigned long start_byte, size_t srclen, size_t destlen) |
124 | { | |
125 | switch (type) { | |
126 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
127 | start_byte, srclen, destlen); | |
128 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
129 | start_byte, srclen, destlen); | |
130 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
131 | start_byte, srclen, destlen); | |
132 | case BTRFS_COMPRESS_NONE: | |
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 | ||
8140dc30 | 142 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 143 | |
30493ff4 | 144 | static void end_compressed_bio_read(struct btrfs_bio *bbio) |
86ccbb4d | 145 | { |
30493ff4 | 146 | struct compressed_bio *cb = bbio->private; |
86ccbb4d QW |
147 | unsigned int index; |
148 | struct page *page; | |
149 | ||
30493ff4 QW |
150 | if (bbio->bio.bi_status) |
151 | cb->status = bbio->bio.bi_status; | |
152 | else | |
81bd9328 CH |
153 | cb->status = errno_to_blk_status(btrfs_decompress_bio(cb)); |
154 | ||
86ccbb4d QW |
155 | /* Release the compressed pages */ |
156 | for (index = 0; index < cb->nr_pages; index++) { | |
157 | page = cb->compressed_pages[index]; | |
158 | page->mapping = NULL; | |
159 | put_page(page); | |
160 | } | |
161 | ||
162 | /* Do io completion on the original bio */ | |
917f32a2 | 163 | btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status); |
86ccbb4d QW |
164 | |
165 | /* Finally free the cb struct */ | |
166 | kfree(cb->compressed_pages); | |
167 | kfree(cb); | |
917f32a2 | 168 | bio_put(&bbio->bio); |
c8b97818 CM |
169 | } |
170 | ||
171 | /* | |
172 | * Clear the writeback bits on all of the file | |
173 | * pages for a compressed write | |
174 | */ | |
7bdcefc1 FM |
175 | static noinline void end_compressed_writeback(struct inode *inode, |
176 | const struct compressed_bio *cb) | |
c8b97818 | 177 | { |
741ec653 | 178 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
09cbfeaf KS |
179 | unsigned long index = cb->start >> PAGE_SHIFT; |
180 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
a75b81c3 | 181 | struct folio_batch fbatch; |
606f82e7 | 182 | const int errno = blk_status_to_errno(cb->status); |
c8b97818 CM |
183 | int i; |
184 | int ret; | |
185 | ||
606f82e7 JB |
186 | if (errno) |
187 | mapping_set_error(inode->i_mapping, errno); | |
7bdcefc1 | 188 | |
a75b81c3 VMO |
189 | folio_batch_init(&fbatch); |
190 | while (index <= end_index) { | |
191 | ret = filemap_get_folios(inode->i_mapping, &index, end_index, | |
192 | &fbatch); | |
193 | ||
194 | if (ret == 0) | |
195 | return; | |
196 | ||
c8b97818 | 197 | for (i = 0; i < ret; i++) { |
a75b81c3 VMO |
198 | struct folio *folio = fbatch.folios[i]; |
199 | ||
606f82e7 | 200 | if (errno) |
a75b81c3 VMO |
201 | folio_set_error(folio); |
202 | btrfs_page_clamp_clear_writeback(fs_info, &folio->page, | |
741ec653 | 203 | cb->start, cb->len); |
c8b97818 | 204 | } |
a75b81c3 | 205 | folio_batch_release(&fbatch); |
c8b97818 CM |
206 | } |
207 | /* the inode may be gone now */ | |
c8b97818 CM |
208 | } |
209 | ||
6853c64a | 210 | static void finish_compressed_bio_write(struct compressed_bio *cb) |
c8b97818 | 211 | { |
6853c64a | 212 | struct inode *inode = cb->inode; |
1d08ce58 | 213 | unsigned int index; |
c8b97818 | 214 | |
6853c64a QW |
215 | /* |
216 | * Ok, we're the last bio for this extent, step one is to call back | |
217 | * into the FS and do all the end_io operations. | |
c8b97818 | 218 | */ |
38a39ac7 | 219 | btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, |
c629732d | 220 | cb->start, cb->start + cb->len - 1, |
606f82e7 | 221 | cb->status == BLK_STS_OK); |
c8b97818 | 222 | |
7c0c7269 OS |
223 | if (cb->writeback) |
224 | end_compressed_writeback(inode, cb); | |
6853c64a | 225 | /* Note, our inode could be gone now */ |
c8b97818 CM |
226 | |
227 | /* | |
6853c64a | 228 | * Release the compressed pages, these came from alloc_page and |
c8b97818 CM |
229 | * are not attached to the inode at all |
230 | */ | |
c8b97818 | 231 | for (index = 0; index < cb->nr_pages; index++) { |
6853c64a QW |
232 | struct page *page = cb->compressed_pages[index]; |
233 | ||
c8b97818 | 234 | page->mapping = NULL; |
09cbfeaf | 235 | put_page(page); |
c8b97818 CM |
236 | } |
237 | ||
6853c64a | 238 | /* Finally free the cb struct */ |
c8b97818 CM |
239 | kfree(cb->compressed_pages); |
240 | kfree(cb); | |
6853c64a QW |
241 | } |
242 | ||
fed8a72d CH |
243 | static void btrfs_finish_compressed_write_work(struct work_struct *work) |
244 | { | |
245 | struct compressed_bio *cb = | |
246 | container_of(work, struct compressed_bio, write_end_work); | |
247 | ||
248 | finish_compressed_bio_write(cb); | |
249 | } | |
250 | ||
6853c64a QW |
251 | /* |
252 | * Do the cleanup once all the compressed pages hit the disk. This will clear | |
253 | * writeback on the file pages and free the compressed pages. | |
254 | * | |
255 | * This also calls the writeback end hooks for the file pages so that metadata | |
256 | * and checksums can be updated in the file. | |
257 | */ | |
917f32a2 | 258 | static void end_compressed_bio_write(struct btrfs_bio *bbio) |
6853c64a | 259 | { |
917f32a2 | 260 | struct compressed_bio *cb = bbio->private; |
d5e4377d | 261 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
6853c64a | 262 | |
d5e4377d CH |
263 | cb->status = bbio->bio.bi_status; |
264 | queue_work(fs_info->compressed_write_workers, &cb->write_end_work); | |
6853c64a | 265 | |
917f32a2 | 266 | bio_put(&bbio->bio); |
c8b97818 CM |
267 | } |
268 | ||
269 | /* | |
270 | * worker function to build and submit bios for previously compressed pages. | |
271 | * The corresponding pages in the inode should be marked for writeback | |
272 | * and the compressed pages should have a reference on them for dropping | |
273 | * when the IO is complete. | |
274 | * | |
275 | * This also checksums the file bytes and gets things ready for | |
276 | * the end io hooks. | |
277 | */ | |
c7ee1819 | 278 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
65b5355f AJ |
279 | unsigned int len, u64 disk_start, |
280 | unsigned int compressed_len, | |
c8b97818 | 281 | struct page **compressed_pages, |
65b5355f | 282 | unsigned int nr_pages, |
bf9486d6 | 283 | blk_opf_t write_flags, |
7c0c7269 OS |
284 | struct cgroup_subsys_state *blkcg_css, |
285 | bool writeback) | |
c8b97818 | 286 | { |
c7ee1819 | 287 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 288 | struct bio *bio = NULL; |
c8b97818 | 289 | struct compressed_bio *cb; |
91507240 | 290 | u64 cur_disk_bytenr = disk_start; |
524bcd1e | 291 | blk_status_t ret = BLK_STS_OK; |
c8b97818 | 292 | |
bbbff01a QW |
293 | ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && |
294 | IS_ALIGNED(len, fs_info->sectorsize)); | |
81bd9328 | 295 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
dac97e51 | 296 | if (!cb) |
4e4cbee9 | 297 | return BLK_STS_RESOURCE; |
606f82e7 | 298 | cb->status = BLK_STS_OK; |
c7ee1819 | 299 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
300 | cb->start = start; |
301 | cb->len = len; | |
302 | cb->compressed_pages = compressed_pages; | |
303 | cb->compressed_len = compressed_len; | |
7c0c7269 | 304 | cb->writeback = writeback; |
fed8a72d | 305 | INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); |
c8b97818 CM |
306 | cb->nr_pages = nr_pages; |
307 | ||
d5e4377d | 308 | if (blkcg_css) { |
acee08aa | 309 | kthread_associate_blkcg(blkcg_css); |
d5e4377d CH |
310 | write_flags |= REQ_CGROUP_PUNT; |
311 | } | |
312 | ||
313 | write_flags |= REQ_BTRFS_ONE_ORDERED; | |
314 | bio = btrfs_bio_alloc(BIO_MAX_VECS, REQ_OP_WRITE | write_flags, | |
315 | BTRFS_I(cb->inode), end_compressed_bio_write, cb); | |
316 | bio->bi_iter.bi_sector = cur_disk_bytenr >> SECTOR_SHIFT; | |
317 | btrfs_bio(bio)->file_offset = start; | |
acee08aa | 318 | |
91507240 QW |
319 | while (cur_disk_bytenr < disk_start + compressed_len) { |
320 | u64 offset = cur_disk_bytenr - disk_start; | |
321 | unsigned int index = offset >> PAGE_SHIFT; | |
322 | unsigned int real_size; | |
323 | unsigned int added; | |
324 | struct page *page = compressed_pages[index]; | |
d5e4377d | 325 | |
91507240 QW |
326 | /* |
327 | * We have various limits on the real read size: | |
91507240 QW |
328 | * - page boundary |
329 | * - compressed length boundary | |
330 | */ | |
30493ff4 | 331 | real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset)); |
91507240 QW |
332 | real_size = min_t(u64, real_size, compressed_len - offset); |
333 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
334 | ||
d5e4377d CH |
335 | added = bio_add_page(bio, page, real_size, offset_in_page(offset)); |
336 | /* | |
337 | * Maximum compressed extent is smaller than bio size limit, | |
338 | * thus bio_add_page() should always success. | |
339 | */ | |
340 | ASSERT(added == real_size); | |
91507240 | 341 | cur_disk_bytenr += added; |
c8b97818 | 342 | } |
d20f7043 | 343 | |
d5e4377d CH |
344 | /* Finished the range. */ |
345 | ASSERT(bio->bi_iter.bi_size); | |
346 | btrfs_submit_bio(bio, 0); | |
acee08aa DZ |
347 | if (blkcg_css) |
348 | kthread_associate_blkcg(NULL); | |
6853c64a | 349 | return ret; |
c8b97818 CM |
350 | } |
351 | ||
2a4d0c90 CH |
352 | static u64 bio_end_offset(struct bio *bio) |
353 | { | |
c45a8f2d | 354 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
355 | |
356 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
357 | } | |
358 | ||
6a404910 QW |
359 | /* |
360 | * Add extra pages in the same compressed file extent so that we don't need to | |
361 | * re-read the same extent again and again. | |
362 | * | |
363 | * NOTE: this won't work well for subpage, as for subpage read, we lock the | |
364 | * full page then submit bio for each compressed/regular extents. | |
365 | * | |
366 | * This means, if we have several sectors in the same page points to the same | |
367 | * on-disk compressed data, we will re-read the same extent many times and | |
368 | * this function can only help for the next page. | |
369 | */ | |
771ed689 CM |
370 | static noinline int add_ra_bio_pages(struct inode *inode, |
371 | u64 compressed_end, | |
4088a47e | 372 | struct compressed_bio *cb, |
82e60d00 | 373 | int *memstall, unsigned long *pflags) |
771ed689 | 374 | { |
6a404910 | 375 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
771ed689 | 376 | unsigned long end_index; |
6a404910 | 377 | u64 cur = bio_end_offset(cb->orig_bio); |
771ed689 CM |
378 | u64 isize = i_size_read(inode); |
379 | int ret; | |
380 | struct page *page; | |
771ed689 CM |
381 | struct extent_map *em; |
382 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
383 | struct extent_map_tree *em_tree; |
384 | struct extent_io_tree *tree; | |
6a404910 | 385 | int sectors_missed = 0; |
771ed689 | 386 | |
771ed689 CM |
387 | em_tree = &BTRFS_I(inode)->extent_tree; |
388 | tree = &BTRFS_I(inode)->io_tree; | |
389 | ||
390 | if (isize == 0) | |
391 | return 0; | |
392 | ||
ca62e85d QW |
393 | /* |
394 | * For current subpage support, we only support 64K page size, | |
395 | * which means maximum compressed extent size (128K) is just 2x page | |
396 | * size. | |
397 | * This makes readahead less effective, so here disable readahead for | |
398 | * subpage for now, until full compressed write is supported. | |
399 | */ | |
400 | if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) | |
401 | return 0; | |
402 | ||
09cbfeaf | 403 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 404 | |
6a404910 QW |
405 | while (cur < compressed_end) { |
406 | u64 page_end; | |
407 | u64 pg_index = cur >> PAGE_SHIFT; | |
408 | u32 add_size; | |
771ed689 | 409 | |
306e16ce | 410 | if (pg_index > end_index) |
771ed689 CM |
411 | break; |
412 | ||
0a943c65 | 413 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 414 | if (page && !xa_is_value(page)) { |
6a404910 QW |
415 | sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> |
416 | fs_info->sectorsize_bits; | |
417 | ||
418 | /* Beyond threshold, no need to continue */ | |
419 | if (sectors_missed > 4) | |
771ed689 | 420 | break; |
6a404910 QW |
421 | |
422 | /* | |
423 | * Jump to next page start as we already have page for | |
424 | * current offset. | |
425 | */ | |
426 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
427 | continue; | |
771ed689 CM |
428 | } |
429 | ||
c62d2555 MH |
430 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
431 | ~__GFP_FS)); | |
771ed689 CM |
432 | if (!page) |
433 | break; | |
434 | ||
c62d2555 | 435 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 436 | put_page(page); |
6a404910 QW |
437 | /* There is already a page, skip to page end */ |
438 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
439 | continue; | |
771ed689 CM |
440 | } |
441 | ||
82e60d00 | 442 | if (!*memstall && PageWorkingset(page)) { |
4088a47e | 443 | psi_memstall_enter(pflags); |
82e60d00 JW |
444 | *memstall = 1; |
445 | } | |
4088a47e | 446 | |
32443de3 QW |
447 | ret = set_page_extent_mapped(page); |
448 | if (ret < 0) { | |
449 | unlock_page(page); | |
450 | put_page(page); | |
451 | break; | |
452 | } | |
453 | ||
6a404910 | 454 | page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; |
570eb97b | 455 | lock_extent(tree, cur, page_end, NULL); |
890871be | 456 | read_lock(&em_tree->lock); |
6a404910 | 457 | em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); |
890871be | 458 | read_unlock(&em_tree->lock); |
771ed689 | 459 | |
6a404910 QW |
460 | /* |
461 | * At this point, we have a locked page in the page cache for | |
462 | * these bytes in the file. But, we have to make sure they map | |
463 | * to this compressed extent on disk. | |
464 | */ | |
465 | if (!em || cur < em->start || | |
466 | (cur + fs_info->sectorsize > extent_map_end(em)) || | |
4f024f37 | 467 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 468 | free_extent_map(em); |
570eb97b | 469 | unlock_extent(tree, cur, page_end, NULL); |
771ed689 | 470 | unlock_page(page); |
09cbfeaf | 471 | put_page(page); |
771ed689 CM |
472 | break; |
473 | } | |
474 | free_extent_map(em); | |
475 | ||
476 | if (page->index == end_index) { | |
7073017a | 477 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
478 | |
479 | if (zero_offset) { | |
480 | int zeros; | |
09cbfeaf | 481 | zeros = PAGE_SIZE - zero_offset; |
d048b9c2 | 482 | memzero_page(page, zero_offset, zeros); |
771ed689 CM |
483 | } |
484 | } | |
485 | ||
6a404910 QW |
486 | add_size = min(em->start + em->len, page_end + 1) - cur; |
487 | ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur)); | |
488 | if (ret != add_size) { | |
570eb97b | 489 | unlock_extent(tree, cur, page_end, NULL); |
771ed689 | 490 | unlock_page(page); |
09cbfeaf | 491 | put_page(page); |
771ed689 CM |
492 | break; |
493 | } | |
6a404910 QW |
494 | /* |
495 | * If it's subpage, we also need to increase its | |
496 | * subpage::readers number, as at endio we will decrease | |
497 | * subpage::readers and to unlock the page. | |
498 | */ | |
499 | if (fs_info->sectorsize < PAGE_SIZE) | |
500 | btrfs_subpage_start_reader(fs_info, page, cur, add_size); | |
501 | put_page(page); | |
502 | cur += add_size; | |
771ed689 | 503 | } |
771ed689 CM |
504 | return 0; |
505 | } | |
506 | ||
c8b97818 CM |
507 | /* |
508 | * for a compressed read, the bio we get passed has all the inode pages | |
509 | * in it. We don't actually do IO on those pages but allocate new ones | |
510 | * to hold the compressed pages on disk. | |
511 | * | |
4f024f37 | 512 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 513 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
514 | * |
515 | * After the compressed pages are read, we copy the bytes into the | |
516 | * bio we were passed and then call the bio end_io calls | |
517 | */ | |
cb4411dd | 518 | void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
1d8fa2e2 | 519 | int mirror_num) |
c8b97818 | 520 | { |
0b246afa | 521 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
522 | struct extent_map_tree *em_tree; |
523 | struct compressed_bio *cb; | |
356b4a2d | 524 | unsigned int compressed_len; |
30493ff4 | 525 | struct bio *comp_bio; |
f472c28f QW |
526 | const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT; |
527 | u64 cur_disk_byte = disk_bytenr; | |
557023ea | 528 | u64 file_offset; |
e04ca626 CM |
529 | u64 em_len; |
530 | u64 em_start; | |
c8b97818 | 531 | struct extent_map *em; |
82e60d00 JW |
532 | unsigned long pflags; |
533 | int memstall = 0; | |
f9f15de8 | 534 | blk_status_t ret; |
dd137dd1 STD |
535 | int ret2; |
536 | int i; | |
c8b97818 | 537 | |
c8b97818 CM |
538 | em_tree = &BTRFS_I(inode)->extent_tree; |
539 | ||
557023ea QW |
540 | file_offset = bio_first_bvec_all(bio)->bv_offset + |
541 | page_offset(bio_first_page_all(bio)); | |
542 | ||
c8b97818 | 543 | /* we need the actual starting offset of this extent in the file */ |
890871be | 544 | read_lock(&em_tree->lock); |
557023ea | 545 | em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); |
890871be | 546 | read_unlock(&em_tree->lock); |
f9f15de8 JB |
547 | if (!em) { |
548 | ret = BLK_STS_IOERR; | |
549 | goto out; | |
550 | } | |
c8b97818 | 551 | |
557023ea | 552 | ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); |
d20f7043 | 553 | compressed_len = em->block_len; |
81bd9328 | 554 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
f9f15de8 JB |
555 | if (!cb) { |
556 | ret = BLK_STS_RESOURCE; | |
6b82ce8d | 557 | goto out; |
f9f15de8 | 558 | } |
6b82ce8d | 559 | |
606f82e7 | 560 | cb->status = BLK_STS_OK; |
c8b97818 CM |
561 | cb->inode = inode; |
562 | ||
ff5b7ee3 | 563 | cb->start = em->orig_start; |
e04ca626 CM |
564 | em_len = em->len; |
565 | em_start = em->start; | |
d20f7043 | 566 | |
81381053 | 567 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 568 | cb->compressed_len = compressed_len; |
1d8fa2e2 | 569 | cb->compress_type = em->compress_type; |
c8b97818 CM |
570 | cb->orig_bio = bio; |
571 | ||
1d8fa2e2 GR |
572 | free_extent_map(em); |
573 | em = NULL; | |
574 | ||
dd137dd1 STD |
575 | cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
576 | cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS); | |
f9f15de8 JB |
577 | if (!cb->compressed_pages) { |
578 | ret = BLK_STS_RESOURCE; | |
dd137dd1 | 579 | goto fail; |
f9f15de8 | 580 | } |
6b82ce8d | 581 | |
dd137dd1 STD |
582 | ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages); |
583 | if (ret2) { | |
584 | ret = BLK_STS_RESOURCE; | |
585 | goto fail; | |
c8b97818 | 586 | } |
c8b97818 | 587 | |
82e60d00 | 588 | add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags); |
771ed689 | 589 | |
771ed689 | 590 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 591 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 592 | |
30493ff4 QW |
593 | comp_bio = btrfs_bio_alloc(BIO_MAX_VECS, REQ_OP_READ, BTRFS_I(cb->inode), |
594 | end_compressed_bio_read, cb); | |
595 | comp_bio->bi_iter.bi_sector = (cur_disk_byte >> SECTOR_SHIFT); | |
596 | ||
f472c28f QW |
597 | while (cur_disk_byte < disk_bytenr + compressed_len) { |
598 | u64 offset = cur_disk_byte - disk_bytenr; | |
599 | unsigned int index = offset >> PAGE_SHIFT; | |
600 | unsigned int real_size; | |
601 | unsigned int added; | |
602 | struct page *page = cb->compressed_pages[index]; | |
c8b97818 | 603 | |
f472c28f QW |
604 | /* |
605 | * We have various limit on the real read size: | |
f472c28f QW |
606 | * - page boundary |
607 | * - compressed length boundary | |
608 | */ | |
30493ff4 | 609 | real_size = min_t(u64, U32_MAX, PAGE_SIZE - offset_in_page(offset)); |
f472c28f QW |
610 | real_size = min_t(u64, real_size, compressed_len - offset); |
611 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
4e4cbee9 | 612 | |
f472c28f | 613 | added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset)); |
be6a1361 | 614 | /* |
f472c28f QW |
615 | * Maximum compressed extent is smaller than bio size limit, |
616 | * thus bio_add_page() should always success. | |
be6a1361 | 617 | */ |
f472c28f QW |
618 | ASSERT(added == real_size); |
619 | cur_disk_byte += added; | |
c8b97818 | 620 | } |
524bcd1e | 621 | |
82e60d00 | 622 | if (memstall) |
4088a47e CH |
623 | psi_memstall_leave(&pflags); |
624 | ||
30493ff4 QW |
625 | /* |
626 | * Stash the initial offset of this chunk, as there is no direct | |
627 | * correlation between compressed pages and the original file offset. | |
628 | * The field is only used for printing error messages anyway. | |
629 | */ | |
630 | btrfs_bio(comp_bio)->file_offset = file_offset; | |
631 | ||
632 | ASSERT(comp_bio->bi_iter.bi_size); | |
285599b6 | 633 | btrfs_submit_bio(comp_bio, mirror_num); |
cb4411dd | 634 | return; |
6b82ce8d | 635 | |
dd137dd1 STD |
636 | fail: |
637 | if (cb->compressed_pages) { | |
638 | for (i = 0; i < cb->nr_pages; i++) { | |
639 | if (cb->compressed_pages[i]) | |
640 | __free_page(cb->compressed_pages[i]); | |
641 | } | |
15e3004a | 642 | } |
6b82ce8d | 643 | |
644 | kfree(cb->compressed_pages); | |
6b82ce8d | 645 | kfree(cb); |
646 | out: | |
647 | free_extent_map(em); | |
917f32a2 | 648 | btrfs_bio_end_io(btrfs_bio(bio), ret); |
cb4411dd | 649 | return; |
c8b97818 | 650 | } |
261507a0 | 651 | |
17b5a6c1 TT |
652 | /* |
653 | * Heuristic uses systematic sampling to collect data from the input data | |
654 | * range, the logic can be tuned by the following constants: | |
655 | * | |
656 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
657 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
658 | */ | |
659 | #define SAMPLING_READ_SIZE (16) | |
660 | #define SAMPLING_INTERVAL (256) | |
661 | ||
662 | /* | |
663 | * For statistical analysis of the input data we consider bytes that form a | |
664 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
665 | * many times the object appeared in the sample. | |
666 | */ | |
667 | #define BUCKET_SIZE (256) | |
668 | ||
669 | /* | |
670 | * The size of the sample is based on a statistical sampling rule of thumb. | |
671 | * The common way is to perform sampling tests as long as the number of | |
672 | * elements in each cell is at least 5. | |
673 | * | |
674 | * Instead of 5, we choose 32 to obtain more accurate results. | |
675 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
676 | * sample size bound by 8192. | |
677 | * | |
678 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
679 | * from up to 512 locations. | |
680 | */ | |
681 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
682 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
683 | ||
684 | struct bucket_item { | |
685 | u32 count; | |
686 | }; | |
4e439a0b TT |
687 | |
688 | struct heuristic_ws { | |
17b5a6c1 TT |
689 | /* Partial copy of input data */ |
690 | u8 *sample; | |
a440d48c | 691 | u32 sample_size; |
17b5a6c1 TT |
692 | /* Buckets store counters for each byte value */ |
693 | struct bucket_item *bucket; | |
440c840c TT |
694 | /* Sorting buffer */ |
695 | struct bucket_item *bucket_b; | |
4e439a0b TT |
696 | struct list_head list; |
697 | }; | |
698 | ||
92ee5530 DZ |
699 | static struct workspace_manager heuristic_wsm; |
700 | ||
4e439a0b TT |
701 | static void free_heuristic_ws(struct list_head *ws) |
702 | { | |
703 | struct heuristic_ws *workspace; | |
704 | ||
705 | workspace = list_entry(ws, struct heuristic_ws, list); | |
706 | ||
17b5a6c1 TT |
707 | kvfree(workspace->sample); |
708 | kfree(workspace->bucket); | |
440c840c | 709 | kfree(workspace->bucket_b); |
4e439a0b TT |
710 | kfree(workspace); |
711 | } | |
712 | ||
7bf49943 | 713 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
714 | { |
715 | struct heuristic_ws *ws; | |
716 | ||
717 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
718 | if (!ws) | |
719 | return ERR_PTR(-ENOMEM); | |
720 | ||
17b5a6c1 TT |
721 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
722 | if (!ws->sample) | |
723 | goto fail; | |
724 | ||
725 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
726 | if (!ws->bucket) | |
727 | goto fail; | |
4e439a0b | 728 | |
440c840c TT |
729 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
730 | if (!ws->bucket_b) | |
731 | goto fail; | |
732 | ||
17b5a6c1 | 733 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 734 | return &ws->list; |
17b5a6c1 TT |
735 | fail: |
736 | free_heuristic_ws(&ws->list); | |
737 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
738 | } |
739 | ||
ca4ac360 | 740 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 741 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
742 | }; |
743 | ||
e8c9f186 | 744 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
745 | /* The heuristic is represented as compression type 0 */ |
746 | &btrfs_heuristic_compress, | |
261507a0 | 747 | &btrfs_zlib_compress, |
a6fa6fae | 748 | &btrfs_lzo_compress, |
5c1aab1d | 749 | &btrfs_zstd_compress, |
261507a0 LZ |
750 | }; |
751 | ||
c778df14 DS |
752 | static struct list_head *alloc_workspace(int type, unsigned int level) |
753 | { | |
754 | switch (type) { | |
755 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
756 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
757 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
758 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
759 | default: | |
760 | /* | |
761 | * This can't happen, the type is validated several times | |
762 | * before we get here. | |
763 | */ | |
764 | BUG(); | |
765 | } | |
766 | } | |
767 | ||
1e002351 DS |
768 | static void free_workspace(int type, struct list_head *ws) |
769 | { | |
770 | switch (type) { | |
771 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
772 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
773 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
774 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
775 | default: | |
776 | /* | |
777 | * This can't happen, the type is validated several times | |
778 | * before we get here. | |
779 | */ | |
780 | BUG(); | |
781 | } | |
782 | } | |
783 | ||
d5517033 | 784 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 785 | { |
0cf25213 | 786 | struct workspace_manager *wsm; |
4e439a0b | 787 | struct list_head *workspace; |
261507a0 | 788 | |
0cf25213 | 789 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
790 | INIT_LIST_HEAD(&wsm->idle_ws); |
791 | spin_lock_init(&wsm->ws_lock); | |
792 | atomic_set(&wsm->total_ws, 0); | |
793 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 794 | |
1666edab DZ |
795 | /* |
796 | * Preallocate one workspace for each compression type so we can | |
797 | * guarantee forward progress in the worst case | |
798 | */ | |
c778df14 | 799 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
800 | if (IS_ERR(workspace)) { |
801 | pr_warn( | |
802 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
803 | } else { | |
92ee5530 DZ |
804 | atomic_set(&wsm->total_ws, 1); |
805 | wsm->free_ws = 1; | |
806 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
807 | } |
808 | } | |
809 | ||
2510307e | 810 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 811 | { |
2dba7143 | 812 | struct workspace_manager *wsman; |
1666edab DZ |
813 | struct list_head *ws; |
814 | ||
2dba7143 | 815 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
816 | while (!list_empty(&wsman->idle_ws)) { |
817 | ws = wsman->idle_ws.next; | |
818 | list_del(ws); | |
1e002351 | 819 | free_workspace(type, ws); |
1666edab | 820 | atomic_dec(&wsman->total_ws); |
261507a0 | 821 | } |
261507a0 LZ |
822 | } |
823 | ||
824 | /* | |
e721e49d DS |
825 | * This finds an available workspace or allocates a new one. |
826 | * If it's not possible to allocate a new one, waits until there's one. | |
827 | * Preallocation makes a forward progress guarantees and we do not return | |
828 | * errors. | |
261507a0 | 829 | */ |
5907a9bb | 830 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 831 | { |
5907a9bb | 832 | struct workspace_manager *wsm; |
261507a0 LZ |
833 | struct list_head *workspace; |
834 | int cpus = num_online_cpus(); | |
fe308533 | 835 | unsigned nofs_flag; |
4e439a0b TT |
836 | struct list_head *idle_ws; |
837 | spinlock_t *ws_lock; | |
838 | atomic_t *total_ws; | |
839 | wait_queue_head_t *ws_wait; | |
840 | int *free_ws; | |
841 | ||
5907a9bb | 842 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
843 | idle_ws = &wsm->idle_ws; |
844 | ws_lock = &wsm->ws_lock; | |
845 | total_ws = &wsm->total_ws; | |
846 | ws_wait = &wsm->ws_wait; | |
847 | free_ws = &wsm->free_ws; | |
261507a0 | 848 | |
261507a0 | 849 | again: |
d9187649 BL |
850 | spin_lock(ws_lock); |
851 | if (!list_empty(idle_ws)) { | |
852 | workspace = idle_ws->next; | |
261507a0 | 853 | list_del(workspace); |
6ac10a6a | 854 | (*free_ws)--; |
d9187649 | 855 | spin_unlock(ws_lock); |
261507a0 LZ |
856 | return workspace; |
857 | ||
858 | } | |
6ac10a6a | 859 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
860 | DEFINE_WAIT(wait); |
861 | ||
d9187649 BL |
862 | spin_unlock(ws_lock); |
863 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 864 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 865 | schedule(); |
d9187649 | 866 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
867 | goto again; |
868 | } | |
6ac10a6a | 869 | atomic_inc(total_ws); |
d9187649 | 870 | spin_unlock(ws_lock); |
261507a0 | 871 | |
fe308533 DS |
872 | /* |
873 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
874 | * to turn it off here because we might get called from the restricted | |
875 | * context of btrfs_compress_bio/btrfs_compress_pages | |
876 | */ | |
877 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 878 | workspace = alloc_workspace(type, level); |
fe308533 DS |
879 | memalloc_nofs_restore(nofs_flag); |
880 | ||
261507a0 | 881 | if (IS_ERR(workspace)) { |
6ac10a6a | 882 | atomic_dec(total_ws); |
d9187649 | 883 | wake_up(ws_wait); |
e721e49d DS |
884 | |
885 | /* | |
886 | * Do not return the error but go back to waiting. There's a | |
887 | * workspace preallocated for each type and the compression | |
888 | * time is bounded so we get to a workspace eventually. This | |
889 | * makes our caller's life easier. | |
52356716 DS |
890 | * |
891 | * To prevent silent and low-probability deadlocks (when the | |
892 | * initial preallocation fails), check if there are any | |
893 | * workspaces at all. | |
e721e49d | 894 | */ |
52356716 DS |
895 | if (atomic_read(total_ws) == 0) { |
896 | static DEFINE_RATELIMIT_STATE(_rs, | |
897 | /* once per minute */ 60 * HZ, | |
898 | /* no burst */ 1); | |
899 | ||
900 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 901 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
902 | } |
903 | } | |
e721e49d | 904 | goto again; |
261507a0 LZ |
905 | } |
906 | return workspace; | |
907 | } | |
908 | ||
7bf49943 | 909 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 910 | { |
6a0d1272 | 911 | switch (type) { |
5907a9bb | 912 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 913 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 914 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
915 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
916 | default: | |
917 | /* | |
918 | * This can't happen, the type is validated several times | |
919 | * before we get here. | |
920 | */ | |
921 | BUG(); | |
922 | } | |
929f4baf DZ |
923 | } |
924 | ||
261507a0 LZ |
925 | /* |
926 | * put a workspace struct back on the list or free it if we have enough | |
927 | * idle ones sitting around | |
928 | */ | |
a3bbd2a9 | 929 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 930 | { |
a3bbd2a9 | 931 | struct workspace_manager *wsm; |
4e439a0b TT |
932 | struct list_head *idle_ws; |
933 | spinlock_t *ws_lock; | |
934 | atomic_t *total_ws; | |
935 | wait_queue_head_t *ws_wait; | |
936 | int *free_ws; | |
937 | ||
a3bbd2a9 | 938 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
939 | idle_ws = &wsm->idle_ws; |
940 | ws_lock = &wsm->ws_lock; | |
941 | total_ws = &wsm->total_ws; | |
942 | ws_wait = &wsm->ws_wait; | |
943 | free_ws = &wsm->free_ws; | |
d9187649 BL |
944 | |
945 | spin_lock(ws_lock); | |
26b28dce | 946 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 947 | list_add(ws, idle_ws); |
6ac10a6a | 948 | (*free_ws)++; |
d9187649 | 949 | spin_unlock(ws_lock); |
261507a0 LZ |
950 | goto wake; |
951 | } | |
d9187649 | 952 | spin_unlock(ws_lock); |
261507a0 | 953 | |
1e002351 | 954 | free_workspace(type, ws); |
6ac10a6a | 955 | atomic_dec(total_ws); |
261507a0 | 956 | wake: |
093258e6 | 957 | cond_wake_up(ws_wait); |
261507a0 LZ |
958 | } |
959 | ||
929f4baf DZ |
960 | static void put_workspace(int type, struct list_head *ws) |
961 | { | |
bd3a5287 | 962 | switch (type) { |
a3bbd2a9 DS |
963 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
964 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
965 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
966 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
967 | default: | |
968 | /* | |
969 | * This can't happen, the type is validated several times | |
970 | * before we get here. | |
971 | */ | |
972 | BUG(); | |
973 | } | |
929f4baf DZ |
974 | } |
975 | ||
adbab642 AJ |
976 | /* |
977 | * Adjust @level according to the limits of the compression algorithm or | |
978 | * fallback to default | |
979 | */ | |
980 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
981 | { | |
982 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
983 | ||
984 | if (level == 0) | |
985 | level = ops->default_level; | |
986 | else | |
987 | level = min(level, ops->max_level); | |
988 | ||
989 | return level; | |
990 | } | |
991 | ||
261507a0 | 992 | /* |
38c31464 DS |
993 | * Given an address space and start and length, compress the bytes into @pages |
994 | * that are allocated on demand. | |
261507a0 | 995 | * |
f51d2b59 DS |
996 | * @type_level is encoded algorithm and level, where level 0 means whatever |
997 | * default the algorithm chooses and is opaque here; | |
998 | * - compression algo are 0-3 | |
999 | * - the level are bits 4-7 | |
1000 | * | |
4d3a800e DS |
1001 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1002 | * and returns number of actually allocated pages | |
261507a0 | 1003 | * |
38c31464 DS |
1004 | * @total_in is used to return the number of bytes actually read. It |
1005 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1006 | * ran out of room in the pages array or because we cross the |
1007 | * max_out threshold. | |
1008 | * | |
38c31464 DS |
1009 | * @total_out is an in/out parameter, must be set to the input length and will |
1010 | * be also used to return the total number of compressed bytes | |
261507a0 | 1011 | */ |
f51d2b59 | 1012 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1013 | u64 start, struct page **pages, |
261507a0 LZ |
1014 | unsigned long *out_pages, |
1015 | unsigned long *total_in, | |
e5d74902 | 1016 | unsigned long *total_out) |
261507a0 | 1017 | { |
1972708a | 1018 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1019 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1020 | struct list_head *workspace; |
1021 | int ret; | |
1022 | ||
b0c1fe1e | 1023 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1024 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1025 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1026 | out_pages, total_in, total_out); | |
929f4baf | 1027 | put_workspace(type, workspace); |
261507a0 LZ |
1028 | return ret; |
1029 | } | |
1030 | ||
8140dc30 | 1031 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1032 | { |
1033 | struct list_head *workspace; | |
1034 | int ret; | |
8140dc30 | 1035 | int type = cb->compress_type; |
261507a0 | 1036 | |
7bf49943 | 1037 | workspace = get_workspace(type, 0); |
4a9e803e | 1038 | ret = compression_decompress_bio(workspace, cb); |
929f4baf | 1039 | put_workspace(type, workspace); |
e1ddce71 | 1040 | |
261507a0 LZ |
1041 | return ret; |
1042 | } | |
1043 | ||
1044 | /* | |
1045 | * a less complex decompression routine. Our compressed data fits in a | |
1046 | * single page, and we want to read a single page out of it. | |
1047 | * start_byte tells us the offset into the compressed data we're interested in | |
1048 | */ | |
3e09b5b2 | 1049 | int btrfs_decompress(int type, const u8 *data_in, struct page *dest_page, |
261507a0 LZ |
1050 | unsigned long start_byte, size_t srclen, size_t destlen) |
1051 | { | |
1052 | struct list_head *workspace; | |
1053 | int ret; | |
1054 | ||
7bf49943 | 1055 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1056 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1057 | start_byte, srclen, destlen); | |
929f4baf | 1058 | put_workspace(type, workspace); |
7bf49943 | 1059 | |
261507a0 LZ |
1060 | return ret; |
1061 | } | |
1062 | ||
5565b8e0 | 1063 | int __init btrfs_init_compress(void) |
1666edab | 1064 | { |
d5517033 DS |
1065 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1066 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1067 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1068 | zstd_init_workspace_manager(); | |
5565b8e0 | 1069 | return 0; |
1666edab DZ |
1070 | } |
1071 | ||
e67c718b | 1072 | void __cold btrfs_exit_compress(void) |
261507a0 | 1073 | { |
2510307e DS |
1074 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1075 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1076 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1077 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1078 | } |
3a39c18d LZ |
1079 | |
1080 | /* | |
1c3dc173 | 1081 | * Copy decompressed data from working buffer to pages. |
3a39c18d | 1082 | * |
1c3dc173 QW |
1083 | * @buf: The decompressed data buffer |
1084 | * @buf_len: The decompressed data length | |
1085 | * @decompressed: Number of bytes that are already decompressed inside the | |
1086 | * compressed extent | |
1087 | * @cb: The compressed extent descriptor | |
1088 | * @orig_bio: The original bio that the caller wants to read for | |
3a39c18d | 1089 | * |
1c3dc173 QW |
1090 | * An easier to understand graph is like below: |
1091 | * | |
1092 | * |<- orig_bio ->| |<- orig_bio->| | |
1093 | * |<------- full decompressed extent ----->| | |
1094 | * |<----------- @cb range ---->| | |
1095 | * | |<-- @buf_len -->| | |
1096 | * |<--- @decompressed --->| | |
1097 | * | |
1098 | * Note that, @cb can be a subpage of the full decompressed extent, but | |
1099 | * @cb->start always has the same as the orig_file_offset value of the full | |
1100 | * decompressed extent. | |
1101 | * | |
1102 | * When reading compressed extent, we have to read the full compressed extent, | |
1103 | * while @orig_bio may only want part of the range. | |
1104 | * Thus this function will ensure only data covered by @orig_bio will be copied | |
1105 | * to. | |
1106 | * | |
1107 | * Return 0 if we have copied all needed contents for @orig_bio. | |
1108 | * Return >0 if we need continue decompress. | |
3a39c18d | 1109 | */ |
1c3dc173 QW |
1110 | int btrfs_decompress_buf2page(const char *buf, u32 buf_len, |
1111 | struct compressed_bio *cb, u32 decompressed) | |
3a39c18d | 1112 | { |
1c3dc173 QW |
1113 | struct bio *orig_bio = cb->orig_bio; |
1114 | /* Offset inside the full decompressed extent */ | |
1115 | u32 cur_offset; | |
1116 | ||
1117 | cur_offset = decompressed; | |
1118 | /* The main loop to do the copy */ | |
1119 | while (cur_offset < decompressed + buf_len) { | |
1120 | struct bio_vec bvec; | |
1121 | size_t copy_len; | |
1122 | u32 copy_start; | |
1123 | /* Offset inside the full decompressed extent */ | |
1124 | u32 bvec_offset; | |
1125 | ||
1126 | bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); | |
1127 | /* | |
1128 | * cb->start may underflow, but subtracting that value can still | |
1129 | * give us correct offset inside the full decompressed extent. | |
1130 | */ | |
1131 | bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; | |
974b1adc | 1132 | |
1c3dc173 QW |
1133 | /* Haven't reached the bvec range, exit */ |
1134 | if (decompressed + buf_len <= bvec_offset) | |
1135 | return 1; | |
3a39c18d | 1136 | |
1c3dc173 QW |
1137 | copy_start = max(cur_offset, bvec_offset); |
1138 | copy_len = min(bvec_offset + bvec.bv_len, | |
1139 | decompressed + buf_len) - copy_start; | |
1140 | ASSERT(copy_len); | |
3a39c18d | 1141 | |
974b1adc | 1142 | /* |
1c3dc173 QW |
1143 | * Extra range check to ensure we didn't go beyond |
1144 | * @buf + @buf_len. | |
974b1adc | 1145 | */ |
1c3dc173 QW |
1146 | ASSERT(copy_start - decompressed < buf_len); |
1147 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, | |
1148 | buf + copy_start - decompressed, copy_len); | |
1c3dc173 | 1149 | cur_offset += copy_len; |
3a39c18d | 1150 | |
1c3dc173 QW |
1151 | bio_advance(orig_bio, copy_len); |
1152 | /* Finished the bio */ | |
1153 | if (!orig_bio->bi_iter.bi_size) | |
1154 | return 0; | |
3a39c18d | 1155 | } |
3a39c18d LZ |
1156 | return 1; |
1157 | } | |
c2fcdcdf | 1158 | |
19562430 TT |
1159 | /* |
1160 | * Shannon Entropy calculation | |
1161 | * | |
52042d8e | 1162 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1163 | * Try calculating entropy to estimate the average minimum number of bits |
1164 | * needed to encode the sampled data. | |
1165 | * | |
1166 | * For convenience, return the percentage of needed bits, instead of amount of | |
1167 | * bits directly. | |
1168 | * | |
1169 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1170 | * and can be compressible with high probability | |
1171 | * | |
1172 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1173 | * | |
1174 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1175 | */ | |
1176 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1177 | #define ENTROPY_LVL_HIGH (80) | |
1178 | ||
1179 | /* | |
1180 | * For increasead precision in shannon_entropy calculation, | |
1181 | * let's do pow(n, M) to save more digits after comma: | |
1182 | * | |
1183 | * - maximum int bit length is 64 | |
1184 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1185 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1186 | * | |
1187 | * So use pow(n, 4). | |
1188 | */ | |
1189 | static inline u32 ilog2_w(u64 n) | |
1190 | { | |
1191 | return ilog2(n * n * n * n); | |
1192 | } | |
1193 | ||
1194 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1195 | { | |
1196 | const u32 entropy_max = 8 * ilog2_w(2); | |
1197 | u32 entropy_sum = 0; | |
1198 | u32 p, p_base, sz_base; | |
1199 | u32 i; | |
1200 | ||
1201 | sz_base = ilog2_w(ws->sample_size); | |
1202 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1203 | p = ws->bucket[i].count; | |
1204 | p_base = ilog2_w(p); | |
1205 | entropy_sum += p * (sz_base - p_base); | |
1206 | } | |
1207 | ||
1208 | entropy_sum /= ws->sample_size; | |
1209 | return entropy_sum * 100 / entropy_max; | |
1210 | } | |
1211 | ||
440c840c TT |
1212 | #define RADIX_BASE 4U |
1213 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1214 | ||
1215 | static u8 get4bits(u64 num, int shift) { | |
1216 | u8 low4bits; | |
1217 | ||
1218 | num >>= shift; | |
1219 | /* Reverse order */ | |
1220 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1221 | return low4bits; | |
1222 | } | |
1223 | ||
440c840c TT |
1224 | /* |
1225 | * Use 4 bits as radix base | |
52042d8e | 1226 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1227 | * |
1228 | * @array - array that will be sorted | |
1229 | * @array_buf - buffer array to store sorting results | |
1230 | * must be equal in size to @array | |
1231 | * @num - array size | |
440c840c | 1232 | */ |
23ae8c63 | 1233 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1234 | int num) |
858177d3 | 1235 | { |
440c840c TT |
1236 | u64 max_num; |
1237 | u64 buf_num; | |
1238 | u32 counters[COUNTERS_SIZE]; | |
1239 | u32 new_addr; | |
1240 | u32 addr; | |
1241 | int bitlen; | |
1242 | int shift; | |
1243 | int i; | |
858177d3 | 1244 | |
440c840c TT |
1245 | /* |
1246 | * Try avoid useless loop iterations for small numbers stored in big | |
1247 | * counters. Example: 48 33 4 ... in 64bit array | |
1248 | */ | |
23ae8c63 | 1249 | max_num = array[0].count; |
440c840c | 1250 | for (i = 1; i < num; i++) { |
23ae8c63 | 1251 | buf_num = array[i].count; |
440c840c TT |
1252 | if (buf_num > max_num) |
1253 | max_num = buf_num; | |
1254 | } | |
1255 | ||
1256 | buf_num = ilog2(max_num); | |
1257 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1258 | ||
1259 | shift = 0; | |
1260 | while (shift < bitlen) { | |
1261 | memset(counters, 0, sizeof(counters)); | |
1262 | ||
1263 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1264 | buf_num = array[i].count; |
440c840c TT |
1265 | addr = get4bits(buf_num, shift); |
1266 | counters[addr]++; | |
1267 | } | |
1268 | ||
1269 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1270 | counters[i] += counters[i - 1]; | |
1271 | ||
1272 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1273 | buf_num = array[i].count; |
440c840c TT |
1274 | addr = get4bits(buf_num, shift); |
1275 | counters[addr]--; | |
1276 | new_addr = counters[addr]; | |
7add17be | 1277 | array_buf[new_addr] = array[i]; |
440c840c TT |
1278 | } |
1279 | ||
1280 | shift += RADIX_BASE; | |
1281 | ||
1282 | /* | |
1283 | * Normal radix expects to move data from a temporary array, to | |
1284 | * the main one. But that requires some CPU time. Avoid that | |
1285 | * by doing another sort iteration to original array instead of | |
1286 | * memcpy() | |
1287 | */ | |
1288 | memset(counters, 0, sizeof(counters)); | |
1289 | ||
1290 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1291 | buf_num = array_buf[i].count; |
440c840c TT |
1292 | addr = get4bits(buf_num, shift); |
1293 | counters[addr]++; | |
1294 | } | |
1295 | ||
1296 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1297 | counters[i] += counters[i - 1]; | |
1298 | ||
1299 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1300 | buf_num = array_buf[i].count; |
440c840c TT |
1301 | addr = get4bits(buf_num, shift); |
1302 | counters[addr]--; | |
1303 | new_addr = counters[addr]; | |
7add17be | 1304 | array[new_addr] = array_buf[i]; |
440c840c TT |
1305 | } |
1306 | ||
1307 | shift += RADIX_BASE; | |
1308 | } | |
858177d3 TT |
1309 | } |
1310 | ||
1311 | /* | |
1312 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1313 | * | |
1314 | * There are several types of structured binary data that use nearly all byte | |
1315 | * values. The distribution can be uniform and counts in all buckets will be | |
1316 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1317 | * | |
1318 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1319 | * be potentially compressible, but we have to take a few more steps to decide | |
1320 | * how much. | |
1321 | * | |
1322 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1323 | * compression algo can easy fix that | |
1324 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1325 | * probability is not compressible | |
1326 | */ | |
1327 | #define BYTE_CORE_SET_LOW (64) | |
1328 | #define BYTE_CORE_SET_HIGH (200) | |
1329 | ||
1330 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1331 | { | |
1332 | u32 i; | |
1333 | u32 coreset_sum = 0; | |
1334 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1335 | struct bucket_item *bucket = ws->bucket; | |
1336 | ||
1337 | /* Sort in reverse order */ | |
36243c91 | 1338 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1339 | |
1340 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1341 | coreset_sum += bucket[i].count; | |
1342 | ||
1343 | if (coreset_sum > core_set_threshold) | |
1344 | return i; | |
1345 | ||
1346 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1347 | coreset_sum += bucket[i].count; | |
1348 | if (coreset_sum > core_set_threshold) | |
1349 | break; | |
1350 | } | |
1351 | ||
1352 | return i; | |
1353 | } | |
1354 | ||
a288e92c TT |
1355 | /* |
1356 | * Count byte values in buckets. | |
1357 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1358 | * Because in most text-like data byte set is restricted to limited number of | |
1359 | * possible characters, and that restriction in most cases makes data easy to | |
1360 | * compress. | |
1361 | * | |
1362 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1363 | * less - compressible | |
1364 | * more - need additional analysis | |
1365 | */ | |
1366 | #define BYTE_SET_THRESHOLD (64) | |
1367 | ||
1368 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1369 | { | |
1370 | u32 i; | |
1371 | u32 byte_set_size = 0; | |
1372 | ||
1373 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1374 | if (ws->bucket[i].count > 0) | |
1375 | byte_set_size++; | |
1376 | } | |
1377 | ||
1378 | /* | |
1379 | * Continue collecting count of byte values in buckets. If the byte | |
1380 | * set size is bigger then the threshold, it's pointless to continue, | |
1381 | * the detection technique would fail for this type of data. | |
1382 | */ | |
1383 | for (; i < BUCKET_SIZE; i++) { | |
1384 | if (ws->bucket[i].count > 0) { | |
1385 | byte_set_size++; | |
1386 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1387 | return byte_set_size; | |
1388 | } | |
1389 | } | |
1390 | ||
1391 | return byte_set_size; | |
1392 | } | |
1393 | ||
1fe4f6fa TT |
1394 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1395 | { | |
1396 | const u32 half_of_sample = ws->sample_size / 2; | |
1397 | const u8 *data = ws->sample; | |
1398 | ||
1399 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1400 | } | |
1401 | ||
a440d48c TT |
1402 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1403 | struct heuristic_ws *ws) | |
1404 | { | |
1405 | struct page *page; | |
1406 | u64 index, index_end; | |
1407 | u32 i, curr_sample_pos; | |
1408 | u8 *in_data; | |
1409 | ||
1410 | /* | |
1411 | * Compression handles the input data by chunks of 128KiB | |
1412 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1413 | * | |
1414 | * We do the same for the heuristic and loop over the whole range. | |
1415 | * | |
1416 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1417 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1418 | */ | |
1419 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1420 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1421 | ||
1422 | index = start >> PAGE_SHIFT; | |
1423 | index_end = end >> PAGE_SHIFT; | |
1424 | ||
1425 | /* Don't miss unaligned end */ | |
ce394a7f | 1426 | if (!PAGE_ALIGNED(end)) |
a440d48c TT |
1427 | index_end++; |
1428 | ||
1429 | curr_sample_pos = 0; | |
1430 | while (index < index_end) { | |
1431 | page = find_get_page(inode->i_mapping, index); | |
58c1a35c | 1432 | in_data = kmap_local_page(page); |
a440d48c TT |
1433 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1434 | i = start % PAGE_SIZE; | |
1435 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1436 | /* Don't sample any garbage from the last page */ | |
1437 | if (start > end - SAMPLING_READ_SIZE) | |
1438 | break; | |
1439 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1440 | SAMPLING_READ_SIZE); | |
1441 | i += SAMPLING_INTERVAL; | |
1442 | start += SAMPLING_INTERVAL; | |
1443 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1444 | } | |
58c1a35c | 1445 | kunmap_local(in_data); |
a440d48c TT |
1446 | put_page(page); |
1447 | ||
1448 | index++; | |
1449 | } | |
1450 | ||
1451 | ws->sample_size = curr_sample_pos; | |
1452 | } | |
1453 | ||
c2fcdcdf TT |
1454 | /* |
1455 | * Compression heuristic. | |
1456 | * | |
1457 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1458 | * quickly (compared to direct compression) detect data characteristics | |
67da05b3 | 1459 | * (compressible/incompressible) to avoid wasting CPU time on incompressible |
c2fcdcdf TT |
1460 | * data. |
1461 | * | |
1462 | * The following types of analysis can be performed: | |
1463 | * - detect mostly zero data | |
1464 | * - detect data with low "byte set" size (text, etc) | |
1465 | * - detect data with low/high "core byte" set | |
1466 | * | |
1467 | * Return non-zero if the compression should be done, 0 otherwise. | |
1468 | */ | |
1469 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1470 | { | |
7bf49943 | 1471 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1472 | struct heuristic_ws *ws; |
a440d48c TT |
1473 | u32 i; |
1474 | u8 byte; | |
19562430 | 1475 | int ret = 0; |
c2fcdcdf | 1476 | |
4e439a0b TT |
1477 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1478 | ||
a440d48c TT |
1479 | heuristic_collect_sample(inode, start, end, ws); |
1480 | ||
1fe4f6fa TT |
1481 | if (sample_repeated_patterns(ws)) { |
1482 | ret = 1; | |
1483 | goto out; | |
1484 | } | |
1485 | ||
a440d48c TT |
1486 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1487 | ||
1488 | for (i = 0; i < ws->sample_size; i++) { | |
1489 | byte = ws->sample[i]; | |
1490 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1491 | } |
1492 | ||
a288e92c TT |
1493 | i = byte_set_size(ws); |
1494 | if (i < BYTE_SET_THRESHOLD) { | |
1495 | ret = 2; | |
1496 | goto out; | |
1497 | } | |
1498 | ||
858177d3 TT |
1499 | i = byte_core_set_size(ws); |
1500 | if (i <= BYTE_CORE_SET_LOW) { | |
1501 | ret = 3; | |
1502 | goto out; | |
1503 | } | |
1504 | ||
1505 | if (i >= BYTE_CORE_SET_HIGH) { | |
1506 | ret = 0; | |
1507 | goto out; | |
1508 | } | |
1509 | ||
19562430 TT |
1510 | i = shannon_entropy(ws); |
1511 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1512 | ret = 4; | |
1513 | goto out; | |
1514 | } | |
1515 | ||
1516 | /* | |
1517 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1518 | * needed to give green light to compression. | |
1519 | * | |
1520 | * For now just assume that compression at that level is not worth the | |
1521 | * resources because: | |
1522 | * | |
1523 | * 1. it is possible to defrag the data later | |
1524 | * | |
1525 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1526 | * values, every bucket has counter at level ~54. The heuristic would | |
1527 | * be confused. This can happen when data have some internal repeated | |
1528 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1529 | * pairs of bytes, which is too costly. | |
1530 | */ | |
1531 | if (i < ENTROPY_LVL_HIGH) { | |
1532 | ret = 5; | |
1533 | goto out; | |
1534 | } else { | |
1535 | ret = 0; | |
1536 | goto out; | |
1537 | } | |
1538 | ||
1fe4f6fa | 1539 | out: |
929f4baf | 1540 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1541 | return ret; |
1542 | } | |
f51d2b59 | 1543 | |
d0ab62ce DZ |
1544 | /* |
1545 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1546 | * level, unrecognized string will set the default level | |
1547 | */ | |
1548 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1549 | { |
d0ab62ce DZ |
1550 | unsigned int level = 0; |
1551 | int ret; | |
1552 | ||
1553 | if (!type) | |
f51d2b59 DS |
1554 | return 0; |
1555 | ||
d0ab62ce DZ |
1556 | if (str[0] == ':') { |
1557 | ret = kstrtouint(str + 1, 10, &level); | |
1558 | if (ret) | |
1559 | level = 0; | |
1560 | } | |
1561 | ||
b0c1fe1e DS |
1562 | level = btrfs_compress_set_level(type, level); |
1563 | ||
1564 | return level; | |
1565 | } |