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> | |
11 | #include <linux/highmem.h> | |
12 | #include <linux/time.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/string.h> | |
c8b97818 | 15 | #include <linux/backing-dev.h> |
c8b97818 | 16 | #include <linux/writeback.h> |
5a0e3ad6 | 17 | #include <linux/slab.h> |
fe308533 | 18 | #include <linux/sched/mm.h> |
19562430 | 19 | #include <linux/log2.h> |
c8b97818 CM |
20 | #include "ctree.h" |
21 | #include "disk-io.h" | |
22 | #include "transaction.h" | |
23 | #include "btrfs_inode.h" | |
24 | #include "volumes.h" | |
25 | #include "ordered-data.h" | |
c8b97818 CM |
26 | #include "compression.h" |
27 | #include "extent_io.h" | |
28 | #include "extent_map.h" | |
29 | ||
e128f9c3 DS |
30 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
31 | ||
32 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
33 | { | |
34 | switch (type) { | |
35 | case BTRFS_COMPRESS_ZLIB: | |
36 | case BTRFS_COMPRESS_LZO: | |
37 | case BTRFS_COMPRESS_ZSTD: | |
38 | case BTRFS_COMPRESS_NONE: | |
39 | return btrfs_compress_types[type]; | |
40 | } | |
41 | ||
42 | return NULL; | |
43 | } | |
44 | ||
8140dc30 | 45 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 46 | |
2ff7e61e | 47 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
48 | unsigned long disk_size) |
49 | { | |
0b246afa | 50 | u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); |
6c41761f | 51 | |
d20f7043 | 52 | return sizeof(struct compressed_bio) + |
0b246afa | 53 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size; |
d20f7043 CM |
54 | } |
55 | ||
f898ac6a | 56 | static int check_compressed_csum(struct btrfs_inode *inode, |
d20f7043 CM |
57 | struct compressed_bio *cb, |
58 | u64 disk_start) | |
59 | { | |
60 | int ret; | |
d20f7043 CM |
61 | struct page *page; |
62 | unsigned long i; | |
63 | char *kaddr; | |
64 | u32 csum; | |
65 | u32 *cb_sum = &cb->sums; | |
66 | ||
f898ac6a | 67 | if (inode->flags & BTRFS_INODE_NODATASUM) |
d20f7043 CM |
68 | return 0; |
69 | ||
70 | for (i = 0; i < cb->nr_pages; i++) { | |
71 | page = cb->compressed_pages[i]; | |
72 | csum = ~(u32)0; | |
73 | ||
7ac687d9 | 74 | kaddr = kmap_atomic(page); |
09cbfeaf | 75 | csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE); |
0b5e3daf | 76 | btrfs_csum_final(csum, (u8 *)&csum); |
7ac687d9 | 77 | kunmap_atomic(kaddr); |
d20f7043 CM |
78 | |
79 | if (csum != *cb_sum) { | |
f898ac6a | 80 | btrfs_print_data_csum_error(inode, disk_start, csum, |
0970a22e | 81 | *cb_sum, cb->mirror_num); |
d20f7043 CM |
82 | ret = -EIO; |
83 | goto fail; | |
84 | } | |
85 | cb_sum++; | |
86 | ||
87 | } | |
88 | ret = 0; | |
89 | fail: | |
90 | return ret; | |
91 | } | |
92 | ||
c8b97818 CM |
93 | /* when we finish reading compressed pages from the disk, we |
94 | * decompress them and then run the bio end_io routines on the | |
95 | * decompressed pages (in the inode address space). | |
96 | * | |
97 | * This allows the checksumming and other IO error handling routines | |
98 | * to work normally | |
99 | * | |
100 | * The compressed pages are freed here, and it must be run | |
101 | * in process context | |
102 | */ | |
4246a0b6 | 103 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 104 | { |
c8b97818 CM |
105 | struct compressed_bio *cb = bio->bi_private; |
106 | struct inode *inode; | |
107 | struct page *page; | |
108 | unsigned long index; | |
cf1167d5 | 109 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 110 | int ret = 0; |
c8b97818 | 111 | |
4e4cbee9 | 112 | if (bio->bi_status) |
c8b97818 CM |
113 | cb->errors = 1; |
114 | ||
115 | /* if there are more bios still pending for this compressed | |
116 | * extent, just exit | |
117 | */ | |
a50299ae | 118 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
119 | goto out; |
120 | ||
cf1167d5 LB |
121 | /* |
122 | * Record the correct mirror_num in cb->orig_bio so that | |
123 | * read-repair can work properly. | |
124 | */ | |
125 | ASSERT(btrfs_io_bio(cb->orig_bio)); | |
126 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; | |
127 | cb->mirror_num = mirror; | |
128 | ||
e6311f24 LB |
129 | /* |
130 | * Some IO in this cb have failed, just skip checksum as there | |
131 | * is no way it could be correct. | |
132 | */ | |
133 | if (cb->errors == 1) | |
134 | goto csum_failed; | |
135 | ||
d20f7043 | 136 | inode = cb->inode; |
f898ac6a | 137 | ret = check_compressed_csum(BTRFS_I(inode), cb, |
4f024f37 | 138 | (u64)bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
139 | if (ret) |
140 | goto csum_failed; | |
141 | ||
c8b97818 CM |
142 | /* ok, we're the last bio for this extent, lets start |
143 | * the decompression. | |
144 | */ | |
8140dc30 AJ |
145 | ret = btrfs_decompress_bio(cb); |
146 | ||
d20f7043 | 147 | csum_failed: |
c8b97818 CM |
148 | if (ret) |
149 | cb->errors = 1; | |
150 | ||
151 | /* release the compressed pages */ | |
152 | index = 0; | |
153 | for (index = 0; index < cb->nr_pages; index++) { | |
154 | page = cb->compressed_pages[index]; | |
155 | page->mapping = NULL; | |
09cbfeaf | 156 | put_page(page); |
c8b97818 CM |
157 | } |
158 | ||
159 | /* do io completion on the original bio */ | |
771ed689 | 160 | if (cb->errors) { |
c8b97818 | 161 | bio_io_error(cb->orig_bio); |
d20f7043 | 162 | } else { |
2c30c71b KO |
163 | int i; |
164 | struct bio_vec *bvec; | |
d20f7043 CM |
165 | |
166 | /* | |
167 | * we have verified the checksum already, set page | |
168 | * checked so the end_io handlers know about it | |
169 | */ | |
c09abff8 | 170 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2c30c71b | 171 | bio_for_each_segment_all(bvec, cb->orig_bio, i) |
d20f7043 | 172 | SetPageChecked(bvec->bv_page); |
2c30c71b | 173 | |
4246a0b6 | 174 | bio_endio(cb->orig_bio); |
d20f7043 | 175 | } |
c8b97818 CM |
176 | |
177 | /* finally free the cb struct */ | |
178 | kfree(cb->compressed_pages); | |
179 | kfree(cb); | |
180 | out: | |
181 | bio_put(bio); | |
182 | } | |
183 | ||
184 | /* | |
185 | * Clear the writeback bits on all of the file | |
186 | * pages for a compressed write | |
187 | */ | |
7bdcefc1 FM |
188 | static noinline void end_compressed_writeback(struct inode *inode, |
189 | const struct compressed_bio *cb) | |
c8b97818 | 190 | { |
09cbfeaf KS |
191 | unsigned long index = cb->start >> PAGE_SHIFT; |
192 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
193 | struct page *pages[16]; |
194 | unsigned long nr_pages = end_index - index + 1; | |
195 | int i; | |
196 | int ret; | |
197 | ||
7bdcefc1 FM |
198 | if (cb->errors) |
199 | mapping_set_error(inode->i_mapping, -EIO); | |
200 | ||
d397712b | 201 | while (nr_pages > 0) { |
c8b97818 | 202 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
203 | min_t(unsigned long, |
204 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
205 | if (ret == 0) { |
206 | nr_pages -= 1; | |
207 | index += 1; | |
208 | continue; | |
209 | } | |
210 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
211 | if (cb->errors) |
212 | SetPageError(pages[i]); | |
c8b97818 | 213 | end_page_writeback(pages[i]); |
09cbfeaf | 214 | put_page(pages[i]); |
c8b97818 CM |
215 | } |
216 | nr_pages -= ret; | |
217 | index += ret; | |
218 | } | |
219 | /* the inode may be gone now */ | |
c8b97818 CM |
220 | } |
221 | ||
222 | /* | |
223 | * do the cleanup once all the compressed pages hit the disk. | |
224 | * This will clear writeback on the file pages and free the compressed | |
225 | * pages. | |
226 | * | |
227 | * This also calls the writeback end hooks for the file pages so that | |
228 | * metadata and checksums can be updated in the file. | |
229 | */ | |
4246a0b6 | 230 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 | 231 | { |
c8b97818 CM |
232 | struct compressed_bio *cb = bio->bi_private; |
233 | struct inode *inode; | |
234 | struct page *page; | |
235 | unsigned long index; | |
236 | ||
4e4cbee9 | 237 | if (bio->bi_status) |
c8b97818 CM |
238 | cb->errors = 1; |
239 | ||
240 | /* if there are more bios still pending for this compressed | |
241 | * extent, just exit | |
242 | */ | |
a50299ae | 243 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
244 | goto out; |
245 | ||
246 | /* ok, we're the last bio for this extent, step one is to | |
247 | * call back into the FS and do all the end_io operations | |
248 | */ | |
249 | inode = cb->inode; | |
70b99e69 | 250 | cb->compressed_pages[0]->mapping = cb->inode->i_mapping; |
7087a9d8 | 251 | btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0], |
c629732d | 252 | cb->start, cb->start + cb->len - 1, |
7087a9d8 | 253 | bio->bi_status ? BLK_STS_OK : BLK_STS_NOTSUPP); |
70b99e69 | 254 | cb->compressed_pages[0]->mapping = NULL; |
c8b97818 | 255 | |
7bdcefc1 | 256 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
257 | /* note, our inode could be gone now */ |
258 | ||
259 | /* | |
260 | * release the compressed pages, these came from alloc_page and | |
261 | * are not attached to the inode at all | |
262 | */ | |
263 | index = 0; | |
264 | for (index = 0; index < cb->nr_pages; index++) { | |
265 | page = cb->compressed_pages[index]; | |
266 | page->mapping = NULL; | |
09cbfeaf | 267 | put_page(page); |
c8b97818 CM |
268 | } |
269 | ||
270 | /* finally free the cb struct */ | |
271 | kfree(cb->compressed_pages); | |
272 | kfree(cb); | |
273 | out: | |
274 | bio_put(bio); | |
275 | } | |
276 | ||
277 | /* | |
278 | * worker function to build and submit bios for previously compressed pages. | |
279 | * The corresponding pages in the inode should be marked for writeback | |
280 | * and the compressed pages should have a reference on them for dropping | |
281 | * when the IO is complete. | |
282 | * | |
283 | * This also checksums the file bytes and gets things ready for | |
284 | * the end io hooks. | |
285 | */ | |
4e4cbee9 | 286 | blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start, |
c8b97818 CM |
287 | unsigned long len, u64 disk_start, |
288 | unsigned long compressed_len, | |
289 | struct page **compressed_pages, | |
f82b7359 LB |
290 | unsigned long nr_pages, |
291 | unsigned int write_flags) | |
c8b97818 | 292 | { |
0b246afa | 293 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 | 294 | struct bio *bio = NULL; |
c8b97818 CM |
295 | struct compressed_bio *cb; |
296 | unsigned long bytes_left; | |
306e16ce | 297 | int pg_index = 0; |
c8b97818 CM |
298 | struct page *page; |
299 | u64 first_byte = disk_start; | |
300 | struct block_device *bdev; | |
4e4cbee9 | 301 | blk_status_t ret; |
e55179b3 | 302 | int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM; |
c8b97818 | 303 | |
fdb1e121 | 304 | WARN_ON(!PAGE_ALIGNED(start)); |
2ff7e61e | 305 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 306 | if (!cb) |
4e4cbee9 | 307 | return BLK_STS_RESOURCE; |
a50299ae | 308 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
309 | cb->errors = 0; |
310 | cb->inode = inode; | |
311 | cb->start = start; | |
312 | cb->len = len; | |
d20f7043 | 313 | cb->mirror_num = 0; |
c8b97818 CM |
314 | cb->compressed_pages = compressed_pages; |
315 | cb->compressed_len = compressed_len; | |
316 | cb->orig_bio = NULL; | |
317 | cb->nr_pages = nr_pages; | |
318 | ||
0b246afa | 319 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 320 | |
c821e7f3 | 321 | bio = btrfs_bio_alloc(bdev, first_byte); |
f82b7359 | 322 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
323 | bio->bi_private = cb; |
324 | bio->bi_end_io = end_compressed_bio_write; | |
a50299ae | 325 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
326 | |
327 | /* create and submit bios for the compressed pages */ | |
328 | bytes_left = compressed_len; | |
306e16ce | 329 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 CH |
330 | int submit = 0; |
331 | ||
306e16ce | 332 | page = compressed_pages[pg_index]; |
c8b97818 | 333 | page->mapping = inode->i_mapping; |
4f024f37 | 334 | if (bio->bi_iter.bi_size) |
da12fe54 NB |
335 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, |
336 | 0); | |
c8b97818 | 337 | |
70b99e69 | 338 | page->mapping = NULL; |
4e4cbee9 | 339 | if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 340 | PAGE_SIZE) { |
af09abfe CM |
341 | /* |
342 | * inc the count before we submit the bio so | |
343 | * we know the end IO handler won't happen before | |
344 | * we inc the count. Otherwise, the cb might get | |
345 | * freed before we're done setting it up | |
346 | */ | |
a50299ae | 347 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
348 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
349 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 350 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 351 | |
e55179b3 | 352 | if (!skip_sum) { |
2ff7e61e | 353 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 354 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 355 | } |
d20f7043 | 356 | |
2ff7e61e | 357 | ret = btrfs_map_bio(fs_info, bio, 0, 1); |
f5daf2c7 | 358 | if (ret) { |
4e4cbee9 | 359 | bio->bi_status = ret; |
f5daf2c7 LB |
360 | bio_endio(bio); |
361 | } | |
c8b97818 | 362 | |
c821e7f3 | 363 | bio = btrfs_bio_alloc(bdev, first_byte); |
f82b7359 | 364 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
365 | bio->bi_private = cb; |
366 | bio->bi_end_io = end_compressed_bio_write; | |
09cbfeaf | 367 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 368 | } |
09cbfeaf | 369 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 370 | btrfs_info(fs_info, |
efe120a0 | 371 | "bytes left %lu compress len %lu nr %lu", |
cfbc246e CM |
372 | bytes_left, cb->compressed_len, cb->nr_pages); |
373 | } | |
09cbfeaf KS |
374 | bytes_left -= PAGE_SIZE; |
375 | first_byte += PAGE_SIZE; | |
771ed689 | 376 | cond_resched(); |
c8b97818 | 377 | } |
c8b97818 | 378 | |
0b246afa | 379 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 380 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 381 | |
e55179b3 | 382 | if (!skip_sum) { |
2ff7e61e | 383 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 384 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 385 | } |
d20f7043 | 386 | |
2ff7e61e | 387 | ret = btrfs_map_bio(fs_info, bio, 0, 1); |
f5daf2c7 | 388 | if (ret) { |
4e4cbee9 | 389 | bio->bi_status = ret; |
f5daf2c7 LB |
390 | bio_endio(bio); |
391 | } | |
c8b97818 | 392 | |
c8b97818 CM |
393 | return 0; |
394 | } | |
395 | ||
2a4d0c90 CH |
396 | static u64 bio_end_offset(struct bio *bio) |
397 | { | |
c45a8f2d | 398 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
399 | |
400 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
401 | } | |
402 | ||
771ed689 CM |
403 | static noinline int add_ra_bio_pages(struct inode *inode, |
404 | u64 compressed_end, | |
405 | struct compressed_bio *cb) | |
406 | { | |
407 | unsigned long end_index; | |
306e16ce | 408 | unsigned long pg_index; |
771ed689 CM |
409 | u64 last_offset; |
410 | u64 isize = i_size_read(inode); | |
411 | int ret; | |
412 | struct page *page; | |
413 | unsigned long nr_pages = 0; | |
414 | struct extent_map *em; | |
415 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
416 | struct extent_map_tree *em_tree; |
417 | struct extent_io_tree *tree; | |
418 | u64 end; | |
419 | int misses = 0; | |
420 | ||
2a4d0c90 | 421 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
422 | em_tree = &BTRFS_I(inode)->extent_tree; |
423 | tree = &BTRFS_I(inode)->io_tree; | |
424 | ||
425 | if (isize == 0) | |
426 | return 0; | |
427 | ||
09cbfeaf | 428 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 429 | |
d397712b | 430 | while (last_offset < compressed_end) { |
09cbfeaf | 431 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 432 | |
306e16ce | 433 | if (pg_index > end_index) |
771ed689 CM |
434 | break; |
435 | ||
0a943c65 | 436 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 437 | if (page && !xa_is_value(page)) { |
771ed689 CM |
438 | misses++; |
439 | if (misses > 4) | |
440 | break; | |
441 | goto next; | |
442 | } | |
443 | ||
c62d2555 MH |
444 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
445 | ~__GFP_FS)); | |
771ed689 CM |
446 | if (!page) |
447 | break; | |
448 | ||
c62d2555 | 449 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 450 | put_page(page); |
771ed689 CM |
451 | goto next; |
452 | } | |
453 | ||
09cbfeaf | 454 | end = last_offset + PAGE_SIZE - 1; |
771ed689 CM |
455 | /* |
456 | * at this point, we have a locked page in the page cache | |
457 | * for these bytes in the file. But, we have to make | |
458 | * sure they map to this compressed extent on disk. | |
459 | */ | |
460 | set_page_extent_mapped(page); | |
d0082371 | 461 | lock_extent(tree, last_offset, end); |
890871be | 462 | read_lock(&em_tree->lock); |
771ed689 | 463 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 464 | PAGE_SIZE); |
890871be | 465 | read_unlock(&em_tree->lock); |
771ed689 CM |
466 | |
467 | if (!em || last_offset < em->start || | |
09cbfeaf | 468 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 469 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 470 | free_extent_map(em); |
d0082371 | 471 | unlock_extent(tree, last_offset, end); |
771ed689 | 472 | unlock_page(page); |
09cbfeaf | 473 | put_page(page); |
771ed689 CM |
474 | break; |
475 | } | |
476 | free_extent_map(em); | |
477 | ||
478 | if (page->index == end_index) { | |
479 | char *userpage; | |
7073017a | 480 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
481 | |
482 | if (zero_offset) { | |
483 | int zeros; | |
09cbfeaf | 484 | zeros = PAGE_SIZE - zero_offset; |
7ac687d9 | 485 | userpage = kmap_atomic(page); |
771ed689 CM |
486 | memset(userpage + zero_offset, 0, zeros); |
487 | flush_dcache_page(page); | |
7ac687d9 | 488 | kunmap_atomic(userpage); |
771ed689 CM |
489 | } |
490 | } | |
491 | ||
492 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 493 | PAGE_SIZE, 0); |
771ed689 | 494 | |
09cbfeaf | 495 | if (ret == PAGE_SIZE) { |
771ed689 | 496 | nr_pages++; |
09cbfeaf | 497 | put_page(page); |
771ed689 | 498 | } else { |
d0082371 | 499 | unlock_extent(tree, last_offset, end); |
771ed689 | 500 | unlock_page(page); |
09cbfeaf | 501 | put_page(page); |
771ed689 CM |
502 | break; |
503 | } | |
504 | next: | |
09cbfeaf | 505 | last_offset += PAGE_SIZE; |
771ed689 | 506 | } |
771ed689 CM |
507 | return 0; |
508 | } | |
509 | ||
c8b97818 CM |
510 | /* |
511 | * for a compressed read, the bio we get passed has all the inode pages | |
512 | * in it. We don't actually do IO on those pages but allocate new ones | |
513 | * to hold the compressed pages on disk. | |
514 | * | |
4f024f37 | 515 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 516 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
517 | * |
518 | * After the compressed pages are read, we copy the bytes into the | |
519 | * bio we were passed and then call the bio end_io calls | |
520 | */ | |
4e4cbee9 | 521 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
522 | int mirror_num, unsigned long bio_flags) |
523 | { | |
0b246afa | 524 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
525 | struct extent_map_tree *em_tree; |
526 | struct compressed_bio *cb; | |
c8b97818 CM |
527 | unsigned long compressed_len; |
528 | unsigned long nr_pages; | |
306e16ce | 529 | unsigned long pg_index; |
c8b97818 CM |
530 | struct page *page; |
531 | struct block_device *bdev; | |
532 | struct bio *comp_bio; | |
4f024f37 | 533 | u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
534 | u64 em_len; |
535 | u64 em_start; | |
c8b97818 | 536 | struct extent_map *em; |
4e4cbee9 | 537 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 538 | int faili = 0; |
d20f7043 | 539 | u32 *sums; |
c8b97818 | 540 | |
c8b97818 CM |
541 | em_tree = &BTRFS_I(inode)->extent_tree; |
542 | ||
543 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 544 | read_lock(&em_tree->lock); |
c8b97818 | 545 | em = lookup_extent_mapping(em_tree, |
263663cd | 546 | page_offset(bio_first_page_all(bio)), |
09cbfeaf | 547 | PAGE_SIZE); |
890871be | 548 | read_unlock(&em_tree->lock); |
285190d9 | 549 | if (!em) |
4e4cbee9 | 550 | return BLK_STS_IOERR; |
c8b97818 | 551 | |
d20f7043 | 552 | compressed_len = em->block_len; |
2ff7e61e | 553 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 554 | if (!cb) |
555 | goto out; | |
556 | ||
a50299ae | 557 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
558 | cb->errors = 0; |
559 | cb->inode = inode; | |
d20f7043 CM |
560 | cb->mirror_num = mirror_num; |
561 | sums = &cb->sums; | |
c8b97818 | 562 | |
ff5b7ee3 | 563 | cb->start = em->orig_start; |
e04ca626 CM |
564 | em_len = em->len; |
565 | em_start = em->start; | |
d20f7043 | 566 | |
c8b97818 | 567 | free_extent_map(em); |
e04ca626 | 568 | em = NULL; |
c8b97818 | 569 | |
81381053 | 570 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 571 | cb->compressed_len = compressed_len; |
261507a0 | 572 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
573 | cb->orig_bio = bio; |
574 | ||
09cbfeaf | 575 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 576 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 577 | GFP_NOFS); |
6b82ce8d | 578 | if (!cb->compressed_pages) |
579 | goto fail1; | |
580 | ||
0b246afa | 581 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 582 | |
306e16ce DS |
583 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
584 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS | | |
c8b97818 | 585 | __GFP_HIGHMEM); |
15e3004a JB |
586 | if (!cb->compressed_pages[pg_index]) { |
587 | faili = pg_index - 1; | |
0e9350de | 588 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 589 | goto fail2; |
15e3004a | 590 | } |
c8b97818 | 591 | } |
15e3004a | 592 | faili = nr_pages - 1; |
c8b97818 CM |
593 | cb->nr_pages = nr_pages; |
594 | ||
7f042a83 | 595 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 596 | |
771ed689 | 597 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 598 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 599 | |
c821e7f3 | 600 | comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte); |
ebcc3263 | 601 | comp_bio->bi_opf = REQ_OP_READ; |
c8b97818 CM |
602 | comp_bio->bi_private = cb; |
603 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 604 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 605 | |
306e16ce | 606 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
4e4cbee9 CH |
607 | int submit = 0; |
608 | ||
306e16ce | 609 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 610 | page->mapping = inode->i_mapping; |
09cbfeaf | 611 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 612 | |
4f024f37 | 613 | if (comp_bio->bi_iter.bi_size) |
da12fe54 NB |
614 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, |
615 | comp_bio, 0); | |
c8b97818 | 616 | |
70b99e69 | 617 | page->mapping = NULL; |
4e4cbee9 | 618 | if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 619 | PAGE_SIZE) { |
0b246afa JM |
620 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
621 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 622 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 623 | |
af09abfe CM |
624 | /* |
625 | * inc the count before we submit the bio so | |
626 | * we know the end IO handler won't happen before | |
627 | * we inc the count. Otherwise, the cb might get | |
628 | * freed before we're done setting it up | |
629 | */ | |
a50299ae | 630 | refcount_inc(&cb->pending_bios); |
af09abfe | 631 | |
6cbff00f | 632 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
2ff7e61e JM |
633 | ret = btrfs_lookup_bio_sums(inode, comp_bio, |
634 | sums); | |
79787eaa | 635 | BUG_ON(ret); /* -ENOMEM */ |
d20f7043 | 636 | } |
ed6078f7 | 637 | sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size, |
0b246afa | 638 | fs_info->sectorsize); |
d20f7043 | 639 | |
2ff7e61e | 640 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0); |
4246a0b6 | 641 | if (ret) { |
4e4cbee9 | 642 | comp_bio->bi_status = ret; |
4246a0b6 CH |
643 | bio_endio(comp_bio); |
644 | } | |
c8b97818 | 645 | |
c821e7f3 | 646 | comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte); |
ebcc3263 | 647 | comp_bio->bi_opf = REQ_OP_READ; |
771ed689 CM |
648 | comp_bio->bi_private = cb; |
649 | comp_bio->bi_end_io = end_compressed_bio_read; | |
650 | ||
09cbfeaf | 651 | bio_add_page(comp_bio, page, PAGE_SIZE, 0); |
c8b97818 | 652 | } |
09cbfeaf | 653 | cur_disk_byte += PAGE_SIZE; |
c8b97818 | 654 | } |
c8b97818 | 655 | |
0b246afa | 656 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 657 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 658 | |
c2db1073 | 659 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
2ff7e61e | 660 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
79787eaa | 661 | BUG_ON(ret); /* -ENOMEM */ |
c2db1073 | 662 | } |
d20f7043 | 663 | |
2ff7e61e | 664 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0); |
4246a0b6 | 665 | if (ret) { |
4e4cbee9 | 666 | comp_bio->bi_status = ret; |
4246a0b6 CH |
667 | bio_endio(comp_bio); |
668 | } | |
c8b97818 | 669 | |
c8b97818 | 670 | return 0; |
6b82ce8d | 671 | |
672 | fail2: | |
15e3004a JB |
673 | while (faili >= 0) { |
674 | __free_page(cb->compressed_pages[faili]); | |
675 | faili--; | |
676 | } | |
6b82ce8d | 677 | |
678 | kfree(cb->compressed_pages); | |
679 | fail1: | |
680 | kfree(cb); | |
681 | out: | |
682 | free_extent_map(em); | |
683 | return ret; | |
c8b97818 | 684 | } |
261507a0 | 685 | |
17b5a6c1 TT |
686 | /* |
687 | * Heuristic uses systematic sampling to collect data from the input data | |
688 | * range, the logic can be tuned by the following constants: | |
689 | * | |
690 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
691 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
692 | */ | |
693 | #define SAMPLING_READ_SIZE (16) | |
694 | #define SAMPLING_INTERVAL (256) | |
695 | ||
696 | /* | |
697 | * For statistical analysis of the input data we consider bytes that form a | |
698 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
699 | * many times the object appeared in the sample. | |
700 | */ | |
701 | #define BUCKET_SIZE (256) | |
702 | ||
703 | /* | |
704 | * The size of the sample is based on a statistical sampling rule of thumb. | |
705 | * The common way is to perform sampling tests as long as the number of | |
706 | * elements in each cell is at least 5. | |
707 | * | |
708 | * Instead of 5, we choose 32 to obtain more accurate results. | |
709 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
710 | * sample size bound by 8192. | |
711 | * | |
712 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
713 | * from up to 512 locations. | |
714 | */ | |
715 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
716 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
717 | ||
718 | struct bucket_item { | |
719 | u32 count; | |
720 | }; | |
4e439a0b TT |
721 | |
722 | struct heuristic_ws { | |
17b5a6c1 TT |
723 | /* Partial copy of input data */ |
724 | u8 *sample; | |
a440d48c | 725 | u32 sample_size; |
17b5a6c1 TT |
726 | /* Buckets store counters for each byte value */ |
727 | struct bucket_item *bucket; | |
440c840c TT |
728 | /* Sorting buffer */ |
729 | struct bucket_item *bucket_b; | |
4e439a0b TT |
730 | struct list_head list; |
731 | }; | |
732 | ||
733 | static void free_heuristic_ws(struct list_head *ws) | |
734 | { | |
735 | struct heuristic_ws *workspace; | |
736 | ||
737 | workspace = list_entry(ws, struct heuristic_ws, list); | |
738 | ||
17b5a6c1 TT |
739 | kvfree(workspace->sample); |
740 | kfree(workspace->bucket); | |
440c840c | 741 | kfree(workspace->bucket_b); |
4e439a0b TT |
742 | kfree(workspace); |
743 | } | |
744 | ||
745 | static struct list_head *alloc_heuristic_ws(void) | |
746 | { | |
747 | struct heuristic_ws *ws; | |
748 | ||
749 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
750 | if (!ws) | |
751 | return ERR_PTR(-ENOMEM); | |
752 | ||
17b5a6c1 TT |
753 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
754 | if (!ws->sample) | |
755 | goto fail; | |
756 | ||
757 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
758 | if (!ws->bucket) | |
759 | goto fail; | |
4e439a0b | 760 | |
440c840c TT |
761 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
762 | if (!ws->bucket_b) | |
763 | goto fail; | |
764 | ||
17b5a6c1 | 765 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 766 | return &ws->list; |
17b5a6c1 TT |
767 | fail: |
768 | free_heuristic_ws(&ws->list); | |
769 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
770 | } |
771 | ||
acce85de | 772 | struct workspace_manager { |
d9187649 BL |
773 | struct list_head idle_ws; |
774 | spinlock_t ws_lock; | |
6ac10a6a DS |
775 | /* Number of free workspaces */ |
776 | int free_ws; | |
777 | /* Total number of allocated workspaces */ | |
778 | atomic_t total_ws; | |
779 | /* Waiters for a free workspace */ | |
d9187649 | 780 | wait_queue_head_t ws_wait; |
4e439a0b TT |
781 | }; |
782 | ||
acce85de | 783 | static struct workspace_manager wsm[BTRFS_COMPRESS_TYPES]; |
4e439a0b | 784 | |
acce85de | 785 | static struct workspace_manager btrfs_heuristic_ws; |
261507a0 | 786 | |
e8c9f186 | 787 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
261507a0 | 788 | &btrfs_zlib_compress, |
a6fa6fae | 789 | &btrfs_lzo_compress, |
5c1aab1d | 790 | &btrfs_zstd_compress, |
261507a0 LZ |
791 | }; |
792 | ||
143bede5 | 793 | void __init btrfs_init_compress(void) |
261507a0 | 794 | { |
4e439a0b | 795 | struct list_head *workspace; |
261507a0 LZ |
796 | int i; |
797 | ||
4e439a0b TT |
798 | INIT_LIST_HEAD(&btrfs_heuristic_ws.idle_ws); |
799 | spin_lock_init(&btrfs_heuristic_ws.ws_lock); | |
800 | atomic_set(&btrfs_heuristic_ws.total_ws, 0); | |
801 | init_waitqueue_head(&btrfs_heuristic_ws.ws_wait); | |
f77dd0d6 | 802 | |
4e439a0b TT |
803 | workspace = alloc_heuristic_ws(); |
804 | if (IS_ERR(workspace)) { | |
805 | pr_warn( | |
806 | "BTRFS: cannot preallocate heuristic workspace, will try later\n"); | |
807 | } else { | |
808 | atomic_set(&btrfs_heuristic_ws.total_ws, 1); | |
809 | btrfs_heuristic_ws.free_ws = 1; | |
810 | list_add(workspace, &btrfs_heuristic_ws.idle_ws); | |
811 | } | |
812 | ||
813 | for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) { | |
acce85de DZ |
814 | INIT_LIST_HEAD(&wsm[i].idle_ws); |
815 | spin_lock_init(&wsm[i].ws_lock); | |
816 | atomic_set(&wsm[i].total_ws, 0); | |
817 | init_waitqueue_head(&wsm[i].ws_wait); | |
f77dd0d6 DS |
818 | |
819 | /* | |
820 | * Preallocate one workspace for each compression type so | |
821 | * we can guarantee forward progress in the worst case | |
822 | */ | |
823 | workspace = btrfs_compress_op[i]->alloc_workspace(); | |
824 | if (IS_ERR(workspace)) { | |
62e85577 | 825 | pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n"); |
f77dd0d6 | 826 | } else { |
acce85de DZ |
827 | atomic_set(&wsm[i].total_ws, 1); |
828 | wsm[i].free_ws = 1; | |
829 | list_add(workspace, &wsm[i].idle_ws); | |
f77dd0d6 | 830 | } |
261507a0 | 831 | } |
261507a0 LZ |
832 | } |
833 | ||
834 | /* | |
e721e49d DS |
835 | * This finds an available workspace or allocates a new one. |
836 | * If it's not possible to allocate a new one, waits until there's one. | |
837 | * Preallocation makes a forward progress guarantees and we do not return | |
838 | * errors. | |
261507a0 | 839 | */ |
4e439a0b | 840 | static struct list_head *__find_workspace(int type, bool heuristic) |
261507a0 LZ |
841 | { |
842 | struct list_head *workspace; | |
843 | int cpus = num_online_cpus(); | |
844 | int idx = type - 1; | |
fe308533 | 845 | unsigned nofs_flag; |
4e439a0b TT |
846 | struct list_head *idle_ws; |
847 | spinlock_t *ws_lock; | |
848 | atomic_t *total_ws; | |
849 | wait_queue_head_t *ws_wait; | |
850 | int *free_ws; | |
851 | ||
852 | if (heuristic) { | |
853 | idle_ws = &btrfs_heuristic_ws.idle_ws; | |
854 | ws_lock = &btrfs_heuristic_ws.ws_lock; | |
855 | total_ws = &btrfs_heuristic_ws.total_ws; | |
856 | ws_wait = &btrfs_heuristic_ws.ws_wait; | |
857 | free_ws = &btrfs_heuristic_ws.free_ws; | |
858 | } else { | |
acce85de DZ |
859 | idle_ws = &wsm[idx].idle_ws; |
860 | ws_lock = &wsm[idx].ws_lock; | |
861 | total_ws = &wsm[idx].total_ws; | |
862 | ws_wait = &wsm[idx].ws_wait; | |
863 | free_ws = &wsm[idx].free_ws; | |
4e439a0b | 864 | } |
261507a0 | 865 | |
261507a0 | 866 | again: |
d9187649 BL |
867 | spin_lock(ws_lock); |
868 | if (!list_empty(idle_ws)) { | |
869 | workspace = idle_ws->next; | |
261507a0 | 870 | list_del(workspace); |
6ac10a6a | 871 | (*free_ws)--; |
d9187649 | 872 | spin_unlock(ws_lock); |
261507a0 LZ |
873 | return workspace; |
874 | ||
875 | } | |
6ac10a6a | 876 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
877 | DEFINE_WAIT(wait); |
878 | ||
d9187649 BL |
879 | spin_unlock(ws_lock); |
880 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 881 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 882 | schedule(); |
d9187649 | 883 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
884 | goto again; |
885 | } | |
6ac10a6a | 886 | atomic_inc(total_ws); |
d9187649 | 887 | spin_unlock(ws_lock); |
261507a0 | 888 | |
fe308533 DS |
889 | /* |
890 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
891 | * to turn it off here because we might get called from the restricted | |
892 | * context of btrfs_compress_bio/btrfs_compress_pages | |
893 | */ | |
894 | nofs_flag = memalloc_nofs_save(); | |
4e439a0b TT |
895 | if (heuristic) |
896 | workspace = alloc_heuristic_ws(); | |
897 | else | |
898 | workspace = btrfs_compress_op[idx]->alloc_workspace(); | |
fe308533 DS |
899 | memalloc_nofs_restore(nofs_flag); |
900 | ||
261507a0 | 901 | if (IS_ERR(workspace)) { |
6ac10a6a | 902 | atomic_dec(total_ws); |
d9187649 | 903 | wake_up(ws_wait); |
e721e49d DS |
904 | |
905 | /* | |
906 | * Do not return the error but go back to waiting. There's a | |
907 | * workspace preallocated for each type and the compression | |
908 | * time is bounded so we get to a workspace eventually. This | |
909 | * makes our caller's life easier. | |
52356716 DS |
910 | * |
911 | * To prevent silent and low-probability deadlocks (when the | |
912 | * initial preallocation fails), check if there are any | |
913 | * workspaces at all. | |
e721e49d | 914 | */ |
52356716 DS |
915 | if (atomic_read(total_ws) == 0) { |
916 | static DEFINE_RATELIMIT_STATE(_rs, | |
917 | /* once per minute */ 60 * HZ, | |
918 | /* no burst */ 1); | |
919 | ||
920 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 921 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
922 | } |
923 | } | |
e721e49d | 924 | goto again; |
261507a0 LZ |
925 | } |
926 | return workspace; | |
927 | } | |
928 | ||
4e439a0b TT |
929 | static struct list_head *find_workspace(int type) |
930 | { | |
931 | return __find_workspace(type, false); | |
932 | } | |
933 | ||
261507a0 LZ |
934 | /* |
935 | * put a workspace struct back on the list or free it if we have enough | |
936 | * idle ones sitting around | |
937 | */ | |
4e439a0b TT |
938 | static void __free_workspace(int type, struct list_head *workspace, |
939 | bool heuristic) | |
261507a0 LZ |
940 | { |
941 | int idx = type - 1; | |
4e439a0b TT |
942 | struct list_head *idle_ws; |
943 | spinlock_t *ws_lock; | |
944 | atomic_t *total_ws; | |
945 | wait_queue_head_t *ws_wait; | |
946 | int *free_ws; | |
947 | ||
948 | if (heuristic) { | |
949 | idle_ws = &btrfs_heuristic_ws.idle_ws; | |
950 | ws_lock = &btrfs_heuristic_ws.ws_lock; | |
951 | total_ws = &btrfs_heuristic_ws.total_ws; | |
952 | ws_wait = &btrfs_heuristic_ws.ws_wait; | |
953 | free_ws = &btrfs_heuristic_ws.free_ws; | |
954 | } else { | |
acce85de DZ |
955 | idle_ws = &wsm[idx].idle_ws; |
956 | ws_lock = &wsm[idx].ws_lock; | |
957 | total_ws = &wsm[idx].total_ws; | |
958 | ws_wait = &wsm[idx].ws_wait; | |
959 | free_ws = &wsm[idx].free_ws; | |
4e439a0b | 960 | } |
d9187649 BL |
961 | |
962 | spin_lock(ws_lock); | |
26b28dce | 963 | if (*free_ws <= num_online_cpus()) { |
d9187649 | 964 | list_add(workspace, idle_ws); |
6ac10a6a | 965 | (*free_ws)++; |
d9187649 | 966 | spin_unlock(ws_lock); |
261507a0 LZ |
967 | goto wake; |
968 | } | |
d9187649 | 969 | spin_unlock(ws_lock); |
261507a0 | 970 | |
4e439a0b TT |
971 | if (heuristic) |
972 | free_heuristic_ws(workspace); | |
973 | else | |
974 | btrfs_compress_op[idx]->free_workspace(workspace); | |
6ac10a6a | 975 | atomic_dec(total_ws); |
261507a0 | 976 | wake: |
093258e6 | 977 | cond_wake_up(ws_wait); |
261507a0 LZ |
978 | } |
979 | ||
4e439a0b TT |
980 | static void free_workspace(int type, struct list_head *ws) |
981 | { | |
982 | return __free_workspace(type, ws, false); | |
983 | } | |
984 | ||
261507a0 LZ |
985 | /* |
986 | * cleanup function for module exit | |
987 | */ | |
988 | static void free_workspaces(void) | |
989 | { | |
990 | struct list_head *workspace; | |
991 | int i; | |
992 | ||
4e439a0b TT |
993 | while (!list_empty(&btrfs_heuristic_ws.idle_ws)) { |
994 | workspace = btrfs_heuristic_ws.idle_ws.next; | |
995 | list_del(workspace); | |
996 | free_heuristic_ws(workspace); | |
997 | atomic_dec(&btrfs_heuristic_ws.total_ws); | |
998 | } | |
999 | ||
261507a0 | 1000 | for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) { |
acce85de DZ |
1001 | while (!list_empty(&wsm[i].idle_ws)) { |
1002 | workspace = wsm[i].idle_ws.next; | |
261507a0 LZ |
1003 | list_del(workspace); |
1004 | btrfs_compress_op[i]->free_workspace(workspace); | |
acce85de | 1005 | atomic_dec(&wsm[i].total_ws); |
261507a0 LZ |
1006 | } |
1007 | } | |
1008 | } | |
1009 | ||
1010 | /* | |
38c31464 DS |
1011 | * Given an address space and start and length, compress the bytes into @pages |
1012 | * that are allocated on demand. | |
261507a0 | 1013 | * |
f51d2b59 DS |
1014 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1015 | * default the algorithm chooses and is opaque here; | |
1016 | * - compression algo are 0-3 | |
1017 | * - the level are bits 4-7 | |
1018 | * | |
4d3a800e DS |
1019 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1020 | * and returns number of actually allocated pages | |
261507a0 | 1021 | * |
38c31464 DS |
1022 | * @total_in is used to return the number of bytes actually read. It |
1023 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1024 | * ran out of room in the pages array or because we cross the |
1025 | * max_out threshold. | |
1026 | * | |
38c31464 DS |
1027 | * @total_out is an in/out parameter, must be set to the input length and will |
1028 | * be also used to return the total number of compressed bytes | |
261507a0 | 1029 | * |
38c31464 | 1030 | * @max_out tells us the max number of bytes that we're allowed to |
261507a0 LZ |
1031 | * stuff into pages |
1032 | */ | |
f51d2b59 | 1033 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1034 | u64 start, struct page **pages, |
261507a0 LZ |
1035 | unsigned long *out_pages, |
1036 | unsigned long *total_in, | |
e5d74902 | 1037 | unsigned long *total_out) |
261507a0 | 1038 | { |
1972708a | 1039 | int type = btrfs_compress_type(type_level); |
261507a0 LZ |
1040 | struct list_head *workspace; |
1041 | int ret; | |
1042 | ||
1043 | workspace = find_workspace(type); | |
261507a0 | 1044 | |
f51d2b59 | 1045 | btrfs_compress_op[type - 1]->set_level(workspace, type_level); |
261507a0 | 1046 | ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping, |
38c31464 | 1047 | start, pages, |
4d3a800e | 1048 | out_pages, |
e5d74902 | 1049 | total_in, total_out); |
261507a0 LZ |
1050 | free_workspace(type, workspace); |
1051 | return ret; | |
1052 | } | |
1053 | ||
1054 | /* | |
1055 | * pages_in is an array of pages with compressed data. | |
1056 | * | |
1057 | * disk_start is the starting logical offset of this array in the file | |
1058 | * | |
974b1adc | 1059 | * orig_bio contains the pages from the file that we want to decompress into |
261507a0 LZ |
1060 | * |
1061 | * srclen is the number of bytes in pages_in | |
1062 | * | |
1063 | * The basic idea is that we have a bio that was created by readpages. | |
1064 | * The pages in the bio are for the uncompressed data, and they may not | |
1065 | * be contiguous. They all correspond to the range of bytes covered by | |
1066 | * the compressed extent. | |
1067 | */ | |
8140dc30 | 1068 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1069 | { |
1070 | struct list_head *workspace; | |
1071 | int ret; | |
8140dc30 | 1072 | int type = cb->compress_type; |
261507a0 LZ |
1073 | |
1074 | workspace = find_workspace(type); | |
e1ddce71 | 1075 | ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb); |
261507a0 | 1076 | free_workspace(type, workspace); |
e1ddce71 | 1077 | |
261507a0 LZ |
1078 | return ret; |
1079 | } | |
1080 | ||
1081 | /* | |
1082 | * a less complex decompression routine. Our compressed data fits in a | |
1083 | * single page, and we want to read a single page out of it. | |
1084 | * start_byte tells us the offset into the compressed data we're interested in | |
1085 | */ | |
1086 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1087 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1088 | { | |
1089 | struct list_head *workspace; | |
1090 | int ret; | |
1091 | ||
1092 | workspace = find_workspace(type); | |
261507a0 LZ |
1093 | |
1094 | ret = btrfs_compress_op[type-1]->decompress(workspace, data_in, | |
1095 | dest_page, start_byte, | |
1096 | srclen, destlen); | |
1097 | ||
1098 | free_workspace(type, workspace); | |
1099 | return ret; | |
1100 | } | |
1101 | ||
e67c718b | 1102 | void __cold btrfs_exit_compress(void) |
261507a0 LZ |
1103 | { |
1104 | free_workspaces(); | |
1105 | } | |
3a39c18d LZ |
1106 | |
1107 | /* | |
1108 | * Copy uncompressed data from working buffer to pages. | |
1109 | * | |
1110 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1111 | * | |
1112 | * total_out is the last byte of the buffer | |
1113 | */ | |
14a3357b | 1114 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1115 | unsigned long total_out, u64 disk_start, |
974b1adc | 1116 | struct bio *bio) |
3a39c18d LZ |
1117 | { |
1118 | unsigned long buf_offset; | |
1119 | unsigned long current_buf_start; | |
1120 | unsigned long start_byte; | |
6e78b3f7 | 1121 | unsigned long prev_start_byte; |
3a39c18d LZ |
1122 | unsigned long working_bytes = total_out - buf_start; |
1123 | unsigned long bytes; | |
1124 | char *kaddr; | |
974b1adc | 1125 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1126 | |
1127 | /* | |
1128 | * start byte is the first byte of the page we're currently | |
1129 | * copying into relative to the start of the compressed data. | |
1130 | */ | |
974b1adc | 1131 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1132 | |
1133 | /* we haven't yet hit data corresponding to this page */ | |
1134 | if (total_out <= start_byte) | |
1135 | return 1; | |
1136 | ||
1137 | /* | |
1138 | * the start of the data we care about is offset into | |
1139 | * the middle of our working buffer | |
1140 | */ | |
1141 | if (total_out > start_byte && buf_start < start_byte) { | |
1142 | buf_offset = start_byte - buf_start; | |
1143 | working_bytes -= buf_offset; | |
1144 | } else { | |
1145 | buf_offset = 0; | |
1146 | } | |
1147 | current_buf_start = buf_start; | |
1148 | ||
1149 | /* copy bytes from the working buffer into the pages */ | |
1150 | while (working_bytes > 0) { | |
974b1adc CH |
1151 | bytes = min_t(unsigned long, bvec.bv_len, |
1152 | PAGE_SIZE - buf_offset); | |
3a39c18d | 1153 | bytes = min(bytes, working_bytes); |
974b1adc CH |
1154 | |
1155 | kaddr = kmap_atomic(bvec.bv_page); | |
1156 | memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes); | |
7ac687d9 | 1157 | kunmap_atomic(kaddr); |
974b1adc | 1158 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1159 | |
3a39c18d LZ |
1160 | buf_offset += bytes; |
1161 | working_bytes -= bytes; | |
1162 | current_buf_start += bytes; | |
1163 | ||
1164 | /* check if we need to pick another page */ | |
974b1adc CH |
1165 | bio_advance(bio, bytes); |
1166 | if (!bio->bi_iter.bi_size) | |
1167 | return 0; | |
1168 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1169 | prev_start_byte = start_byte; |
974b1adc | 1170 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1171 | |
974b1adc | 1172 | /* |
6e78b3f7 OS |
1173 | * We need to make sure we're only adjusting |
1174 | * our offset into compression working buffer when | |
1175 | * we're switching pages. Otherwise we can incorrectly | |
1176 | * keep copying when we were actually done. | |
974b1adc | 1177 | */ |
6e78b3f7 OS |
1178 | if (start_byte != prev_start_byte) { |
1179 | /* | |
1180 | * make sure our new page is covered by this | |
1181 | * working buffer | |
1182 | */ | |
1183 | if (total_out <= start_byte) | |
1184 | return 1; | |
3a39c18d | 1185 | |
6e78b3f7 OS |
1186 | /* |
1187 | * the next page in the biovec might not be adjacent | |
1188 | * to the last page, but it might still be found | |
1189 | * inside this working buffer. bump our offset pointer | |
1190 | */ | |
1191 | if (total_out > start_byte && | |
1192 | current_buf_start < start_byte) { | |
1193 | buf_offset = start_byte - buf_start; | |
1194 | working_bytes = total_out - start_byte; | |
1195 | current_buf_start = buf_start + buf_offset; | |
1196 | } | |
3a39c18d LZ |
1197 | } |
1198 | } | |
1199 | ||
1200 | return 1; | |
1201 | } | |
c2fcdcdf | 1202 | |
19562430 TT |
1203 | /* |
1204 | * Shannon Entropy calculation | |
1205 | * | |
52042d8e | 1206 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1207 | * Try calculating entropy to estimate the average minimum number of bits |
1208 | * needed to encode the sampled data. | |
1209 | * | |
1210 | * For convenience, return the percentage of needed bits, instead of amount of | |
1211 | * bits directly. | |
1212 | * | |
1213 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1214 | * and can be compressible with high probability | |
1215 | * | |
1216 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1217 | * | |
1218 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1219 | */ | |
1220 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1221 | #define ENTROPY_LVL_HIGH (80) | |
1222 | ||
1223 | /* | |
1224 | * For increasead precision in shannon_entropy calculation, | |
1225 | * let's do pow(n, M) to save more digits after comma: | |
1226 | * | |
1227 | * - maximum int bit length is 64 | |
1228 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1229 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1230 | * | |
1231 | * So use pow(n, 4). | |
1232 | */ | |
1233 | static inline u32 ilog2_w(u64 n) | |
1234 | { | |
1235 | return ilog2(n * n * n * n); | |
1236 | } | |
1237 | ||
1238 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1239 | { | |
1240 | const u32 entropy_max = 8 * ilog2_w(2); | |
1241 | u32 entropy_sum = 0; | |
1242 | u32 p, p_base, sz_base; | |
1243 | u32 i; | |
1244 | ||
1245 | sz_base = ilog2_w(ws->sample_size); | |
1246 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1247 | p = ws->bucket[i].count; | |
1248 | p_base = ilog2_w(p); | |
1249 | entropy_sum += p * (sz_base - p_base); | |
1250 | } | |
1251 | ||
1252 | entropy_sum /= ws->sample_size; | |
1253 | return entropy_sum * 100 / entropy_max; | |
1254 | } | |
1255 | ||
440c840c TT |
1256 | #define RADIX_BASE 4U |
1257 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1258 | ||
1259 | static u8 get4bits(u64 num, int shift) { | |
1260 | u8 low4bits; | |
1261 | ||
1262 | num >>= shift; | |
1263 | /* Reverse order */ | |
1264 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1265 | return low4bits; | |
1266 | } | |
1267 | ||
440c840c TT |
1268 | /* |
1269 | * Use 4 bits as radix base | |
52042d8e | 1270 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1271 | * |
1272 | * @array - array that will be sorted | |
1273 | * @array_buf - buffer array to store sorting results | |
1274 | * must be equal in size to @array | |
1275 | * @num - array size | |
440c840c | 1276 | */ |
23ae8c63 | 1277 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1278 | int num) |
858177d3 | 1279 | { |
440c840c TT |
1280 | u64 max_num; |
1281 | u64 buf_num; | |
1282 | u32 counters[COUNTERS_SIZE]; | |
1283 | u32 new_addr; | |
1284 | u32 addr; | |
1285 | int bitlen; | |
1286 | int shift; | |
1287 | int i; | |
858177d3 | 1288 | |
440c840c TT |
1289 | /* |
1290 | * Try avoid useless loop iterations for small numbers stored in big | |
1291 | * counters. Example: 48 33 4 ... in 64bit array | |
1292 | */ | |
23ae8c63 | 1293 | max_num = array[0].count; |
440c840c | 1294 | for (i = 1; i < num; i++) { |
23ae8c63 | 1295 | buf_num = array[i].count; |
440c840c TT |
1296 | if (buf_num > max_num) |
1297 | max_num = buf_num; | |
1298 | } | |
1299 | ||
1300 | buf_num = ilog2(max_num); | |
1301 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1302 | ||
1303 | shift = 0; | |
1304 | while (shift < bitlen) { | |
1305 | memset(counters, 0, sizeof(counters)); | |
1306 | ||
1307 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1308 | buf_num = array[i].count; |
440c840c TT |
1309 | addr = get4bits(buf_num, shift); |
1310 | counters[addr]++; | |
1311 | } | |
1312 | ||
1313 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1314 | counters[i] += counters[i - 1]; | |
1315 | ||
1316 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1317 | buf_num = array[i].count; |
440c840c TT |
1318 | addr = get4bits(buf_num, shift); |
1319 | counters[addr]--; | |
1320 | new_addr = counters[addr]; | |
7add17be | 1321 | array_buf[new_addr] = array[i]; |
440c840c TT |
1322 | } |
1323 | ||
1324 | shift += RADIX_BASE; | |
1325 | ||
1326 | /* | |
1327 | * Normal radix expects to move data from a temporary array, to | |
1328 | * the main one. But that requires some CPU time. Avoid that | |
1329 | * by doing another sort iteration to original array instead of | |
1330 | * memcpy() | |
1331 | */ | |
1332 | memset(counters, 0, sizeof(counters)); | |
1333 | ||
1334 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1335 | buf_num = array_buf[i].count; |
440c840c TT |
1336 | addr = get4bits(buf_num, shift); |
1337 | counters[addr]++; | |
1338 | } | |
1339 | ||
1340 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1341 | counters[i] += counters[i - 1]; | |
1342 | ||
1343 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1344 | buf_num = array_buf[i].count; |
440c840c TT |
1345 | addr = get4bits(buf_num, shift); |
1346 | counters[addr]--; | |
1347 | new_addr = counters[addr]; | |
7add17be | 1348 | array[new_addr] = array_buf[i]; |
440c840c TT |
1349 | } |
1350 | ||
1351 | shift += RADIX_BASE; | |
1352 | } | |
858177d3 TT |
1353 | } |
1354 | ||
1355 | /* | |
1356 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1357 | * | |
1358 | * There are several types of structured binary data that use nearly all byte | |
1359 | * values. The distribution can be uniform and counts in all buckets will be | |
1360 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1361 | * | |
1362 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1363 | * be potentially compressible, but we have to take a few more steps to decide | |
1364 | * how much. | |
1365 | * | |
1366 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1367 | * compression algo can easy fix that | |
1368 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1369 | * probability is not compressible | |
1370 | */ | |
1371 | #define BYTE_CORE_SET_LOW (64) | |
1372 | #define BYTE_CORE_SET_HIGH (200) | |
1373 | ||
1374 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1375 | { | |
1376 | u32 i; | |
1377 | u32 coreset_sum = 0; | |
1378 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1379 | struct bucket_item *bucket = ws->bucket; | |
1380 | ||
1381 | /* Sort in reverse order */ | |
36243c91 | 1382 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1383 | |
1384 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1385 | coreset_sum += bucket[i].count; | |
1386 | ||
1387 | if (coreset_sum > core_set_threshold) | |
1388 | return i; | |
1389 | ||
1390 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1391 | coreset_sum += bucket[i].count; | |
1392 | if (coreset_sum > core_set_threshold) | |
1393 | break; | |
1394 | } | |
1395 | ||
1396 | return i; | |
1397 | } | |
1398 | ||
a288e92c TT |
1399 | /* |
1400 | * Count byte values in buckets. | |
1401 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1402 | * Because in most text-like data byte set is restricted to limited number of | |
1403 | * possible characters, and that restriction in most cases makes data easy to | |
1404 | * compress. | |
1405 | * | |
1406 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1407 | * less - compressible | |
1408 | * more - need additional analysis | |
1409 | */ | |
1410 | #define BYTE_SET_THRESHOLD (64) | |
1411 | ||
1412 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1413 | { | |
1414 | u32 i; | |
1415 | u32 byte_set_size = 0; | |
1416 | ||
1417 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1418 | if (ws->bucket[i].count > 0) | |
1419 | byte_set_size++; | |
1420 | } | |
1421 | ||
1422 | /* | |
1423 | * Continue collecting count of byte values in buckets. If the byte | |
1424 | * set size is bigger then the threshold, it's pointless to continue, | |
1425 | * the detection technique would fail for this type of data. | |
1426 | */ | |
1427 | for (; i < BUCKET_SIZE; i++) { | |
1428 | if (ws->bucket[i].count > 0) { | |
1429 | byte_set_size++; | |
1430 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1431 | return byte_set_size; | |
1432 | } | |
1433 | } | |
1434 | ||
1435 | return byte_set_size; | |
1436 | } | |
1437 | ||
1fe4f6fa TT |
1438 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1439 | { | |
1440 | const u32 half_of_sample = ws->sample_size / 2; | |
1441 | const u8 *data = ws->sample; | |
1442 | ||
1443 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1444 | } | |
1445 | ||
a440d48c TT |
1446 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1447 | struct heuristic_ws *ws) | |
1448 | { | |
1449 | struct page *page; | |
1450 | u64 index, index_end; | |
1451 | u32 i, curr_sample_pos; | |
1452 | u8 *in_data; | |
1453 | ||
1454 | /* | |
1455 | * Compression handles the input data by chunks of 128KiB | |
1456 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1457 | * | |
1458 | * We do the same for the heuristic and loop over the whole range. | |
1459 | * | |
1460 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1461 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1462 | */ | |
1463 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1464 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1465 | ||
1466 | index = start >> PAGE_SHIFT; | |
1467 | index_end = end >> PAGE_SHIFT; | |
1468 | ||
1469 | /* Don't miss unaligned end */ | |
1470 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1471 | index_end++; | |
1472 | ||
1473 | curr_sample_pos = 0; | |
1474 | while (index < index_end) { | |
1475 | page = find_get_page(inode->i_mapping, index); | |
1476 | in_data = kmap(page); | |
1477 | /* Handle case where the start is not aligned to PAGE_SIZE */ | |
1478 | i = start % PAGE_SIZE; | |
1479 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1480 | /* Don't sample any garbage from the last page */ | |
1481 | if (start > end - SAMPLING_READ_SIZE) | |
1482 | break; | |
1483 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1484 | SAMPLING_READ_SIZE); | |
1485 | i += SAMPLING_INTERVAL; | |
1486 | start += SAMPLING_INTERVAL; | |
1487 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1488 | } | |
1489 | kunmap(page); | |
1490 | put_page(page); | |
1491 | ||
1492 | index++; | |
1493 | } | |
1494 | ||
1495 | ws->sample_size = curr_sample_pos; | |
1496 | } | |
1497 | ||
c2fcdcdf TT |
1498 | /* |
1499 | * Compression heuristic. | |
1500 | * | |
1501 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1502 | * quickly (compared to direct compression) detect data characteristics | |
1503 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1504 | * data. | |
1505 | * | |
1506 | * The following types of analysis can be performed: | |
1507 | * - detect mostly zero data | |
1508 | * - detect data with low "byte set" size (text, etc) | |
1509 | * - detect data with low/high "core byte" set | |
1510 | * | |
1511 | * Return non-zero if the compression should be done, 0 otherwise. | |
1512 | */ | |
1513 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1514 | { | |
4e439a0b TT |
1515 | struct list_head *ws_list = __find_workspace(0, true); |
1516 | struct heuristic_ws *ws; | |
a440d48c TT |
1517 | u32 i; |
1518 | u8 byte; | |
19562430 | 1519 | int ret = 0; |
c2fcdcdf | 1520 | |
4e439a0b TT |
1521 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1522 | ||
a440d48c TT |
1523 | heuristic_collect_sample(inode, start, end, ws); |
1524 | ||
1fe4f6fa TT |
1525 | if (sample_repeated_patterns(ws)) { |
1526 | ret = 1; | |
1527 | goto out; | |
1528 | } | |
1529 | ||
a440d48c TT |
1530 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1531 | ||
1532 | for (i = 0; i < ws->sample_size; i++) { | |
1533 | byte = ws->sample[i]; | |
1534 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1535 | } |
1536 | ||
a288e92c TT |
1537 | i = byte_set_size(ws); |
1538 | if (i < BYTE_SET_THRESHOLD) { | |
1539 | ret = 2; | |
1540 | goto out; | |
1541 | } | |
1542 | ||
858177d3 TT |
1543 | i = byte_core_set_size(ws); |
1544 | if (i <= BYTE_CORE_SET_LOW) { | |
1545 | ret = 3; | |
1546 | goto out; | |
1547 | } | |
1548 | ||
1549 | if (i >= BYTE_CORE_SET_HIGH) { | |
1550 | ret = 0; | |
1551 | goto out; | |
1552 | } | |
1553 | ||
19562430 TT |
1554 | i = shannon_entropy(ws); |
1555 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1556 | ret = 4; | |
1557 | goto out; | |
1558 | } | |
1559 | ||
1560 | /* | |
1561 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1562 | * needed to give green light to compression. | |
1563 | * | |
1564 | * For now just assume that compression at that level is not worth the | |
1565 | * resources because: | |
1566 | * | |
1567 | * 1. it is possible to defrag the data later | |
1568 | * | |
1569 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1570 | * values, every bucket has counter at level ~54. The heuristic would | |
1571 | * be confused. This can happen when data have some internal repeated | |
1572 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1573 | * pairs of bytes, which is too costly. | |
1574 | */ | |
1575 | if (i < ENTROPY_LVL_HIGH) { | |
1576 | ret = 5; | |
1577 | goto out; | |
1578 | } else { | |
1579 | ret = 0; | |
1580 | goto out; | |
1581 | } | |
1582 | ||
1fe4f6fa | 1583 | out: |
4e439a0b | 1584 | __free_workspace(0, ws_list, true); |
c2fcdcdf TT |
1585 | return ret; |
1586 | } | |
f51d2b59 DS |
1587 | |
1588 | unsigned int btrfs_compress_str2level(const char *str) | |
1589 | { | |
1590 | if (strncmp(str, "zlib", 4) != 0) | |
1591 | return 0; | |
1592 | ||
fa4d885a AB |
1593 | /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */ |
1594 | if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0) | |
1595 | return str[5] - '0'; | |
f51d2b59 | 1596 | |
eae8d825 | 1597 | return BTRFS_ZLIB_DEFAULT_LEVEL; |
f51d2b59 | 1598 | } |