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> |
d5178578 | 20 | #include <crypto/hash.h> |
602cbe91 | 21 | #include "misc.h" |
c8b97818 CM |
22 | #include "ctree.h" |
23 | #include "disk-io.h" | |
24 | #include "transaction.h" | |
25 | #include "btrfs_inode.h" | |
26 | #include "volumes.h" | |
27 | #include "ordered-data.h" | |
c8b97818 CM |
28 | #include "compression.h" |
29 | #include "extent_io.h" | |
30 | #include "extent_map.h" | |
31 | ||
e128f9c3 DS |
32 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
33 | ||
34 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
35 | { | |
36 | switch (type) { | |
37 | case BTRFS_COMPRESS_ZLIB: | |
38 | case BTRFS_COMPRESS_LZO: | |
39 | case BTRFS_COMPRESS_ZSTD: | |
40 | case BTRFS_COMPRESS_NONE: | |
41 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
42 | default: |
43 | break; | |
e128f9c3 DS |
44 | } |
45 | ||
46 | return NULL; | |
47 | } | |
48 | ||
aa53e3bf JT |
49 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
50 | { | |
51 | int i; | |
52 | ||
53 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
54 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
55 | ||
56 | if (len < comp_len) | |
57 | continue; | |
58 | ||
59 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
60 | return true; | |
61 | } | |
62 | return false; | |
63 | } | |
64 | ||
1e4eb746 DS |
65 | static int compression_compress_pages(int type, struct list_head *ws, |
66 | struct address_space *mapping, u64 start, struct page **pages, | |
67 | unsigned long *out_pages, unsigned long *total_in, | |
68 | unsigned long *total_out) | |
69 | { | |
70 | switch (type) { | |
71 | case BTRFS_COMPRESS_ZLIB: | |
72 | return zlib_compress_pages(ws, mapping, start, pages, | |
73 | out_pages, total_in, total_out); | |
74 | case BTRFS_COMPRESS_LZO: | |
75 | return lzo_compress_pages(ws, mapping, start, pages, | |
76 | out_pages, total_in, total_out); | |
77 | case BTRFS_COMPRESS_ZSTD: | |
78 | return zstd_compress_pages(ws, mapping, start, pages, | |
79 | out_pages, total_in, total_out); | |
80 | case BTRFS_COMPRESS_NONE: | |
81 | default: | |
82 | /* | |
83 | * This can't happen, the type is validated several times | |
84 | * before we get here. As a sane fallback, return what the | |
85 | * callers will understand as 'no compression happened'. | |
86 | */ | |
87 | return -E2BIG; | |
88 | } | |
89 | } | |
90 | ||
91 | static int compression_decompress_bio(int type, struct list_head *ws, | |
92 | struct compressed_bio *cb) | |
93 | { | |
94 | switch (type) { | |
95 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); | |
96 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
97 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
98 | case BTRFS_COMPRESS_NONE: | |
99 | default: | |
100 | /* | |
101 | * This can't happen, the type is validated several times | |
102 | * before we get here. | |
103 | */ | |
104 | BUG(); | |
105 | } | |
106 | } | |
107 | ||
108 | static int compression_decompress(int type, struct list_head *ws, | |
109 | unsigned char *data_in, struct page *dest_page, | |
110 | unsigned long start_byte, size_t srclen, size_t destlen) | |
111 | { | |
112 | switch (type) { | |
113 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
114 | start_byte, srclen, destlen); | |
115 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
116 | start_byte, srclen, destlen); | |
117 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
118 | start_byte, srclen, destlen); | |
119 | case BTRFS_COMPRESS_NONE: | |
120 | default: | |
121 | /* | |
122 | * This can't happen, the type is validated several times | |
123 | * before we get here. | |
124 | */ | |
125 | BUG(); | |
126 | } | |
127 | } | |
128 | ||
8140dc30 | 129 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 130 | |
2ff7e61e | 131 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
132 | unsigned long disk_size) |
133 | { | |
d20f7043 | 134 | return sizeof(struct compressed_bio) + |
713cebfb | 135 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size; |
d20f7043 CM |
136 | } |
137 | ||
5a9472fe | 138 | static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio, |
d20f7043 CM |
139 | u64 disk_start) |
140 | { | |
10fe6ca8 | 141 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
d5178578 | 142 | SHASH_DESC_ON_STACK(shash, fs_info->csum_shash); |
223486c2 | 143 | const u32 csum_size = fs_info->csum_size; |
d20f7043 CM |
144 | struct page *page; |
145 | unsigned long i; | |
146 | char *kaddr; | |
d5178578 | 147 | u8 csum[BTRFS_CSUM_SIZE]; |
5a9472fe | 148 | struct compressed_bio *cb = bio->bi_private; |
10fe6ca8 | 149 | u8 *cb_sum = cb->sums; |
d20f7043 | 150 | |
42437a63 | 151 | if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM)) |
d20f7043 CM |
152 | return 0; |
153 | ||
d5178578 JT |
154 | shash->tfm = fs_info->csum_shash; |
155 | ||
d20f7043 CM |
156 | for (i = 0; i < cb->nr_pages; i++) { |
157 | page = cb->compressed_pages[i]; | |
d20f7043 | 158 | |
7ac687d9 | 159 | kaddr = kmap_atomic(page); |
fd08001f | 160 | crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum); |
7ac687d9 | 161 | kunmap_atomic(kaddr); |
d20f7043 | 162 | |
10fe6ca8 | 163 | if (memcmp(&csum, cb_sum, csum_size)) { |
d5178578 | 164 | btrfs_print_data_csum_error(inode, disk_start, |
ea41d6b2 | 165 | csum, cb_sum, cb->mirror_num); |
5a9472fe NB |
166 | if (btrfs_io_bio(bio)->device) |
167 | btrfs_dev_stat_inc_and_print( | |
168 | btrfs_io_bio(bio)->device, | |
169 | BTRFS_DEV_STAT_CORRUPTION_ERRS); | |
93c4c033 | 170 | return -EIO; |
d20f7043 | 171 | } |
10fe6ca8 | 172 | cb_sum += csum_size; |
d20f7043 | 173 | } |
93c4c033 | 174 | return 0; |
d20f7043 CM |
175 | } |
176 | ||
c8b97818 CM |
177 | /* when we finish reading compressed pages from the disk, we |
178 | * decompress them and then run the bio end_io routines on the | |
179 | * decompressed pages (in the inode address space). | |
180 | * | |
181 | * This allows the checksumming and other IO error handling routines | |
182 | * to work normally | |
183 | * | |
184 | * The compressed pages are freed here, and it must be run | |
185 | * in process context | |
186 | */ | |
4246a0b6 | 187 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 188 | { |
c8b97818 CM |
189 | struct compressed_bio *cb = bio->bi_private; |
190 | struct inode *inode; | |
191 | struct page *page; | |
192 | unsigned long index; | |
cf1167d5 | 193 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 194 | int ret = 0; |
c8b97818 | 195 | |
4e4cbee9 | 196 | if (bio->bi_status) |
c8b97818 CM |
197 | cb->errors = 1; |
198 | ||
199 | /* if there are more bios still pending for this compressed | |
200 | * extent, just exit | |
201 | */ | |
a50299ae | 202 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
203 | goto out; |
204 | ||
cf1167d5 LB |
205 | /* |
206 | * Record the correct mirror_num in cb->orig_bio so that | |
207 | * read-repair can work properly. | |
208 | */ | |
cf1167d5 LB |
209 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; |
210 | cb->mirror_num = mirror; | |
211 | ||
e6311f24 LB |
212 | /* |
213 | * Some IO in this cb have failed, just skip checksum as there | |
214 | * is no way it could be correct. | |
215 | */ | |
216 | if (cb->errors == 1) | |
217 | goto csum_failed; | |
218 | ||
d20f7043 | 219 | inode = cb->inode; |
5a9472fe | 220 | ret = check_compressed_csum(BTRFS_I(inode), bio, |
1201b58b | 221 | bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
222 | if (ret) |
223 | goto csum_failed; | |
224 | ||
c8b97818 CM |
225 | /* ok, we're the last bio for this extent, lets start |
226 | * the decompression. | |
227 | */ | |
8140dc30 AJ |
228 | ret = btrfs_decompress_bio(cb); |
229 | ||
d20f7043 | 230 | csum_failed: |
c8b97818 CM |
231 | if (ret) |
232 | cb->errors = 1; | |
233 | ||
234 | /* release the compressed pages */ | |
235 | index = 0; | |
236 | for (index = 0; index < cb->nr_pages; index++) { | |
237 | page = cb->compressed_pages[index]; | |
238 | page->mapping = NULL; | |
09cbfeaf | 239 | put_page(page); |
c8b97818 CM |
240 | } |
241 | ||
242 | /* do io completion on the original bio */ | |
771ed689 | 243 | if (cb->errors) { |
c8b97818 | 244 | bio_io_error(cb->orig_bio); |
d20f7043 | 245 | } else { |
2c30c71b | 246 | struct bio_vec *bvec; |
6dc4f100 | 247 | struct bvec_iter_all iter_all; |
d20f7043 CM |
248 | |
249 | /* | |
250 | * we have verified the checksum already, set page | |
251 | * checked so the end_io handlers know about it | |
252 | */ | |
c09abff8 | 253 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2b070cfe | 254 | bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) |
d20f7043 | 255 | SetPageChecked(bvec->bv_page); |
2c30c71b | 256 | |
4246a0b6 | 257 | bio_endio(cb->orig_bio); |
d20f7043 | 258 | } |
c8b97818 CM |
259 | |
260 | /* finally free the cb struct */ | |
261 | kfree(cb->compressed_pages); | |
262 | kfree(cb); | |
263 | out: | |
264 | bio_put(bio); | |
265 | } | |
266 | ||
267 | /* | |
268 | * Clear the writeback bits on all of the file | |
269 | * pages for a compressed write | |
270 | */ | |
7bdcefc1 FM |
271 | static noinline void end_compressed_writeback(struct inode *inode, |
272 | const struct compressed_bio *cb) | |
c8b97818 | 273 | { |
09cbfeaf KS |
274 | unsigned long index = cb->start >> PAGE_SHIFT; |
275 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
276 | struct page *pages[16]; |
277 | unsigned long nr_pages = end_index - index + 1; | |
278 | int i; | |
279 | int ret; | |
280 | ||
7bdcefc1 FM |
281 | if (cb->errors) |
282 | mapping_set_error(inode->i_mapping, -EIO); | |
283 | ||
d397712b | 284 | while (nr_pages > 0) { |
c8b97818 | 285 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
286 | min_t(unsigned long, |
287 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
288 | if (ret == 0) { |
289 | nr_pages -= 1; | |
290 | index += 1; | |
291 | continue; | |
292 | } | |
293 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
294 | if (cb->errors) |
295 | SetPageError(pages[i]); | |
c8b97818 | 296 | end_page_writeback(pages[i]); |
09cbfeaf | 297 | put_page(pages[i]); |
c8b97818 CM |
298 | } |
299 | nr_pages -= ret; | |
300 | index += ret; | |
301 | } | |
302 | /* the inode may be gone now */ | |
c8b97818 CM |
303 | } |
304 | ||
305 | /* | |
306 | * do the cleanup once all the compressed pages hit the disk. | |
307 | * This will clear writeback on the file pages and free the compressed | |
308 | * pages. | |
309 | * | |
310 | * This also calls the writeback end hooks for the file pages so that | |
311 | * metadata and checksums can be updated in the file. | |
312 | */ | |
4246a0b6 | 313 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 | 314 | { |
c8b97818 CM |
315 | struct compressed_bio *cb = bio->bi_private; |
316 | struct inode *inode; | |
317 | struct page *page; | |
318 | unsigned long index; | |
319 | ||
4e4cbee9 | 320 | if (bio->bi_status) |
c8b97818 CM |
321 | cb->errors = 1; |
322 | ||
323 | /* if there are more bios still pending for this compressed | |
324 | * extent, just exit | |
325 | */ | |
a50299ae | 326 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
327 | goto out; |
328 | ||
329 | /* ok, we're the last bio for this extent, step one is to | |
330 | * call back into the FS and do all the end_io operations | |
331 | */ | |
332 | inode = cb->inode; | |
70b99e69 | 333 | cb->compressed_pages[0]->mapping = cb->inode->i_mapping; |
7087a9d8 | 334 | btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0], |
c629732d | 335 | cb->start, cb->start + cb->len - 1, |
6a8d2136 | 336 | bio->bi_status == BLK_STS_OK); |
70b99e69 | 337 | cb->compressed_pages[0]->mapping = NULL; |
c8b97818 | 338 | |
7bdcefc1 | 339 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
340 | /* note, our inode could be gone now */ |
341 | ||
342 | /* | |
343 | * release the compressed pages, these came from alloc_page and | |
344 | * are not attached to the inode at all | |
345 | */ | |
346 | index = 0; | |
347 | for (index = 0; index < cb->nr_pages; index++) { | |
348 | page = cb->compressed_pages[index]; | |
349 | page->mapping = NULL; | |
09cbfeaf | 350 | put_page(page); |
c8b97818 CM |
351 | } |
352 | ||
353 | /* finally free the cb struct */ | |
354 | kfree(cb->compressed_pages); | |
355 | kfree(cb); | |
356 | out: | |
357 | bio_put(bio); | |
358 | } | |
359 | ||
360 | /* | |
361 | * worker function to build and submit bios for previously compressed pages. | |
362 | * The corresponding pages in the inode should be marked for writeback | |
363 | * and the compressed pages should have a reference on them for dropping | |
364 | * when the IO is complete. | |
365 | * | |
366 | * This also checksums the file bytes and gets things ready for | |
367 | * the end io hooks. | |
368 | */ | |
c7ee1819 | 369 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
c8b97818 CM |
370 | unsigned long len, u64 disk_start, |
371 | unsigned long compressed_len, | |
372 | struct page **compressed_pages, | |
f82b7359 | 373 | unsigned long nr_pages, |
ec39f769 CM |
374 | unsigned int write_flags, |
375 | struct cgroup_subsys_state *blkcg_css) | |
c8b97818 | 376 | { |
c7ee1819 | 377 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 378 | struct bio *bio = NULL; |
c8b97818 CM |
379 | struct compressed_bio *cb; |
380 | unsigned long bytes_left; | |
306e16ce | 381 | int pg_index = 0; |
c8b97818 CM |
382 | struct page *page; |
383 | u64 first_byte = disk_start; | |
4e4cbee9 | 384 | blk_status_t ret; |
c7ee1819 | 385 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
c8b97818 | 386 | |
fdb1e121 | 387 | WARN_ON(!PAGE_ALIGNED(start)); |
2ff7e61e | 388 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 389 | if (!cb) |
4e4cbee9 | 390 | return BLK_STS_RESOURCE; |
a50299ae | 391 | refcount_set(&cb->pending_bios, 0); |
c8b97818 | 392 | cb->errors = 0; |
c7ee1819 | 393 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
394 | cb->start = start; |
395 | cb->len = len; | |
d20f7043 | 396 | cb->mirror_num = 0; |
c8b97818 CM |
397 | cb->compressed_pages = compressed_pages; |
398 | cb->compressed_len = compressed_len; | |
399 | cb->orig_bio = NULL; | |
400 | cb->nr_pages = nr_pages; | |
401 | ||
e749af44 | 402 | bio = btrfs_bio_alloc(first_byte); |
f82b7359 | 403 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
404 | bio->bi_private = cb; |
405 | bio->bi_end_io = end_compressed_bio_write; | |
ec39f769 CM |
406 | |
407 | if (blkcg_css) { | |
408 | bio->bi_opf |= REQ_CGROUP_PUNT; | |
46bcff2b | 409 | kthread_associate_blkcg(blkcg_css); |
ec39f769 | 410 | } |
a50299ae | 411 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
412 | |
413 | /* create and submit bios for the compressed pages */ | |
414 | bytes_left = compressed_len; | |
306e16ce | 415 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 CH |
416 | int submit = 0; |
417 | ||
306e16ce | 418 | page = compressed_pages[pg_index]; |
c7ee1819 | 419 | page->mapping = inode->vfs_inode.i_mapping; |
4f024f37 | 420 | if (bio->bi_iter.bi_size) |
da12fe54 NB |
421 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, |
422 | 0); | |
c8b97818 | 423 | |
70b99e69 | 424 | page->mapping = NULL; |
4e4cbee9 | 425 | if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 426 | PAGE_SIZE) { |
af09abfe CM |
427 | /* |
428 | * inc the count before we submit the bio so | |
429 | * we know the end IO handler won't happen before | |
430 | * we inc the count. Otherwise, the cb might get | |
431 | * freed before we're done setting it up | |
432 | */ | |
a50299ae | 433 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
434 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
435 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 436 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 437 | |
e55179b3 | 438 | if (!skip_sum) { |
c7ee1819 | 439 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 440 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 441 | } |
d20f7043 | 442 | |
08635bae | 443 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 444 | if (ret) { |
4e4cbee9 | 445 | bio->bi_status = ret; |
f5daf2c7 LB |
446 | bio_endio(bio); |
447 | } | |
c8b97818 | 448 | |
e749af44 | 449 | bio = btrfs_bio_alloc(first_byte); |
f82b7359 | 450 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
451 | bio->bi_private = cb; |
452 | bio->bi_end_io = end_compressed_bio_write; | |
46bcff2b | 453 | if (blkcg_css) |
7b62e66c | 454 | bio->bi_opf |= REQ_CGROUP_PUNT; |
09cbfeaf | 455 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 456 | } |
09cbfeaf | 457 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 458 | btrfs_info(fs_info, |
efe120a0 | 459 | "bytes left %lu compress len %lu nr %lu", |
cfbc246e CM |
460 | bytes_left, cb->compressed_len, cb->nr_pages); |
461 | } | |
09cbfeaf KS |
462 | bytes_left -= PAGE_SIZE; |
463 | first_byte += PAGE_SIZE; | |
771ed689 | 464 | cond_resched(); |
c8b97818 | 465 | } |
c8b97818 | 466 | |
0b246afa | 467 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 468 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 469 | |
e55179b3 | 470 | if (!skip_sum) { |
c7ee1819 | 471 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 472 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 473 | } |
d20f7043 | 474 | |
08635bae | 475 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 476 | if (ret) { |
4e4cbee9 | 477 | bio->bi_status = ret; |
f5daf2c7 LB |
478 | bio_endio(bio); |
479 | } | |
c8b97818 | 480 | |
46bcff2b DZ |
481 | if (blkcg_css) |
482 | kthread_associate_blkcg(NULL); | |
483 | ||
c8b97818 CM |
484 | return 0; |
485 | } | |
486 | ||
2a4d0c90 CH |
487 | static u64 bio_end_offset(struct bio *bio) |
488 | { | |
c45a8f2d | 489 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
490 | |
491 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
492 | } | |
493 | ||
771ed689 CM |
494 | static noinline int add_ra_bio_pages(struct inode *inode, |
495 | u64 compressed_end, | |
496 | struct compressed_bio *cb) | |
497 | { | |
498 | unsigned long end_index; | |
306e16ce | 499 | unsigned long pg_index; |
771ed689 CM |
500 | u64 last_offset; |
501 | u64 isize = i_size_read(inode); | |
502 | int ret; | |
503 | struct page *page; | |
504 | unsigned long nr_pages = 0; | |
505 | struct extent_map *em; | |
506 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
507 | struct extent_map_tree *em_tree; |
508 | struct extent_io_tree *tree; | |
509 | u64 end; | |
510 | int misses = 0; | |
511 | ||
2a4d0c90 | 512 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
513 | em_tree = &BTRFS_I(inode)->extent_tree; |
514 | tree = &BTRFS_I(inode)->io_tree; | |
515 | ||
516 | if (isize == 0) | |
517 | return 0; | |
518 | ||
09cbfeaf | 519 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 520 | |
d397712b | 521 | while (last_offset < compressed_end) { |
09cbfeaf | 522 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 523 | |
306e16ce | 524 | if (pg_index > end_index) |
771ed689 CM |
525 | break; |
526 | ||
0a943c65 | 527 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 528 | if (page && !xa_is_value(page)) { |
771ed689 CM |
529 | misses++; |
530 | if (misses > 4) | |
531 | break; | |
532 | goto next; | |
533 | } | |
534 | ||
c62d2555 MH |
535 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
536 | ~__GFP_FS)); | |
771ed689 CM |
537 | if (!page) |
538 | break; | |
539 | ||
c62d2555 | 540 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 541 | put_page(page); |
771ed689 CM |
542 | goto next; |
543 | } | |
544 | ||
09cbfeaf | 545 | end = last_offset + PAGE_SIZE - 1; |
771ed689 CM |
546 | /* |
547 | * at this point, we have a locked page in the page cache | |
548 | * for these bytes in the file. But, we have to make | |
549 | * sure they map to this compressed extent on disk. | |
550 | */ | |
551 | set_page_extent_mapped(page); | |
d0082371 | 552 | lock_extent(tree, last_offset, end); |
890871be | 553 | read_lock(&em_tree->lock); |
771ed689 | 554 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 555 | PAGE_SIZE); |
890871be | 556 | read_unlock(&em_tree->lock); |
771ed689 CM |
557 | |
558 | if (!em || last_offset < em->start || | |
09cbfeaf | 559 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 560 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 561 | free_extent_map(em); |
d0082371 | 562 | unlock_extent(tree, last_offset, end); |
771ed689 | 563 | unlock_page(page); |
09cbfeaf | 564 | put_page(page); |
771ed689 CM |
565 | break; |
566 | } | |
567 | free_extent_map(em); | |
568 | ||
569 | if (page->index == end_index) { | |
570 | char *userpage; | |
7073017a | 571 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
572 | |
573 | if (zero_offset) { | |
574 | int zeros; | |
09cbfeaf | 575 | zeros = PAGE_SIZE - zero_offset; |
7ac687d9 | 576 | userpage = kmap_atomic(page); |
771ed689 CM |
577 | memset(userpage + zero_offset, 0, zeros); |
578 | flush_dcache_page(page); | |
7ac687d9 | 579 | kunmap_atomic(userpage); |
771ed689 CM |
580 | } |
581 | } | |
582 | ||
583 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 584 | PAGE_SIZE, 0); |
771ed689 | 585 | |
09cbfeaf | 586 | if (ret == PAGE_SIZE) { |
771ed689 | 587 | nr_pages++; |
09cbfeaf | 588 | put_page(page); |
771ed689 | 589 | } else { |
d0082371 | 590 | unlock_extent(tree, last_offset, end); |
771ed689 | 591 | unlock_page(page); |
09cbfeaf | 592 | put_page(page); |
771ed689 CM |
593 | break; |
594 | } | |
595 | next: | |
09cbfeaf | 596 | last_offset += PAGE_SIZE; |
771ed689 | 597 | } |
771ed689 CM |
598 | return 0; |
599 | } | |
600 | ||
c8b97818 CM |
601 | /* |
602 | * for a compressed read, the bio we get passed has all the inode pages | |
603 | * in it. We don't actually do IO on those pages but allocate new ones | |
604 | * to hold the compressed pages on disk. | |
605 | * | |
4f024f37 | 606 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 607 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
608 | * |
609 | * After the compressed pages are read, we copy the bytes into the | |
610 | * bio we were passed and then call the bio end_io calls | |
611 | */ | |
4e4cbee9 | 612 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
613 | int mirror_num, unsigned long bio_flags) |
614 | { | |
0b246afa | 615 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
616 | struct extent_map_tree *em_tree; |
617 | struct compressed_bio *cb; | |
c8b97818 CM |
618 | unsigned long compressed_len; |
619 | unsigned long nr_pages; | |
306e16ce | 620 | unsigned long pg_index; |
c8b97818 | 621 | struct page *page; |
c8b97818 | 622 | struct bio *comp_bio; |
1201b58b | 623 | u64 cur_disk_byte = bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
624 | u64 em_len; |
625 | u64 em_start; | |
c8b97818 | 626 | struct extent_map *em; |
4e4cbee9 | 627 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 628 | int faili = 0; |
10fe6ca8 | 629 | u8 *sums; |
c8b97818 | 630 | |
c8b97818 CM |
631 | em_tree = &BTRFS_I(inode)->extent_tree; |
632 | ||
633 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 634 | read_lock(&em_tree->lock); |
c8b97818 | 635 | em = lookup_extent_mapping(em_tree, |
263663cd | 636 | page_offset(bio_first_page_all(bio)), |
09cbfeaf | 637 | PAGE_SIZE); |
890871be | 638 | read_unlock(&em_tree->lock); |
285190d9 | 639 | if (!em) |
4e4cbee9 | 640 | return BLK_STS_IOERR; |
c8b97818 | 641 | |
d20f7043 | 642 | compressed_len = em->block_len; |
2ff7e61e | 643 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 644 | if (!cb) |
645 | goto out; | |
646 | ||
a50299ae | 647 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
648 | cb->errors = 0; |
649 | cb->inode = inode; | |
d20f7043 | 650 | cb->mirror_num = mirror_num; |
10fe6ca8 | 651 | sums = cb->sums; |
c8b97818 | 652 | |
ff5b7ee3 | 653 | cb->start = em->orig_start; |
e04ca626 CM |
654 | em_len = em->len; |
655 | em_start = em->start; | |
d20f7043 | 656 | |
c8b97818 | 657 | free_extent_map(em); |
e04ca626 | 658 | em = NULL; |
c8b97818 | 659 | |
81381053 | 660 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 661 | cb->compressed_len = compressed_len; |
261507a0 | 662 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
663 | cb->orig_bio = bio; |
664 | ||
09cbfeaf | 665 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 666 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 667 | GFP_NOFS); |
6b82ce8d | 668 | if (!cb->compressed_pages) |
669 | goto fail1; | |
670 | ||
306e16ce DS |
671 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
672 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS | | |
c8b97818 | 673 | __GFP_HIGHMEM); |
15e3004a JB |
674 | if (!cb->compressed_pages[pg_index]) { |
675 | faili = pg_index - 1; | |
0e9350de | 676 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 677 | goto fail2; |
15e3004a | 678 | } |
c8b97818 | 679 | } |
15e3004a | 680 | faili = nr_pages - 1; |
c8b97818 CM |
681 | cb->nr_pages = nr_pages; |
682 | ||
7f042a83 | 683 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 684 | |
771ed689 | 685 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 686 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 687 | |
e749af44 | 688 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 689 | comp_bio->bi_opf = REQ_OP_READ; |
c8b97818 CM |
690 | comp_bio->bi_private = cb; |
691 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 692 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 693 | |
306e16ce | 694 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
4e4cbee9 CH |
695 | int submit = 0; |
696 | ||
306e16ce | 697 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 698 | page->mapping = inode->i_mapping; |
09cbfeaf | 699 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 700 | |
4f024f37 | 701 | if (comp_bio->bi_iter.bi_size) |
da12fe54 NB |
702 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, |
703 | comp_bio, 0); | |
c8b97818 | 704 | |
70b99e69 | 705 | page->mapping = NULL; |
4e4cbee9 | 706 | if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 707 | PAGE_SIZE) { |
10fe6ca8 JT |
708 | unsigned int nr_sectors; |
709 | ||
0b246afa JM |
710 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
711 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 712 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 713 | |
af09abfe CM |
714 | /* |
715 | * inc the count before we submit the bio so | |
716 | * we know the end IO handler won't happen before | |
717 | * we inc the count. Otherwise, the cb might get | |
718 | * freed before we're done setting it up | |
719 | */ | |
a50299ae | 720 | refcount_inc(&cb->pending_bios); |
af09abfe | 721 | |
334c16d8 JB |
722 | ret = btrfs_lookup_bio_sums(inode, comp_bio, (u64)-1, |
723 | sums); | |
724 | BUG_ON(ret); /* -ENOMEM */ | |
10fe6ca8 JT |
725 | |
726 | nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, | |
727 | fs_info->sectorsize); | |
713cebfb | 728 | sums += fs_info->csum_size * nr_sectors; |
d20f7043 | 729 | |
08635bae | 730 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 731 | if (ret) { |
4e4cbee9 | 732 | comp_bio->bi_status = ret; |
4246a0b6 CH |
733 | bio_endio(comp_bio); |
734 | } | |
c8b97818 | 735 | |
e749af44 | 736 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 737 | comp_bio->bi_opf = REQ_OP_READ; |
771ed689 CM |
738 | comp_bio->bi_private = cb; |
739 | comp_bio->bi_end_io = end_compressed_bio_read; | |
740 | ||
09cbfeaf | 741 | bio_add_page(comp_bio, page, PAGE_SIZE, 0); |
c8b97818 | 742 | } |
09cbfeaf | 743 | cur_disk_byte += PAGE_SIZE; |
c8b97818 | 744 | } |
c8b97818 | 745 | |
0b246afa | 746 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 747 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 748 | |
334c16d8 JB |
749 | ret = btrfs_lookup_bio_sums(inode, comp_bio, (u64)-1, sums); |
750 | BUG_ON(ret); /* -ENOMEM */ | |
d20f7043 | 751 | |
08635bae | 752 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 753 | if (ret) { |
4e4cbee9 | 754 | comp_bio->bi_status = ret; |
4246a0b6 CH |
755 | bio_endio(comp_bio); |
756 | } | |
c8b97818 | 757 | |
c8b97818 | 758 | return 0; |
6b82ce8d | 759 | |
760 | fail2: | |
15e3004a JB |
761 | while (faili >= 0) { |
762 | __free_page(cb->compressed_pages[faili]); | |
763 | faili--; | |
764 | } | |
6b82ce8d | 765 | |
766 | kfree(cb->compressed_pages); | |
767 | fail1: | |
768 | kfree(cb); | |
769 | out: | |
770 | free_extent_map(em); | |
771 | return ret; | |
c8b97818 | 772 | } |
261507a0 | 773 | |
17b5a6c1 TT |
774 | /* |
775 | * Heuristic uses systematic sampling to collect data from the input data | |
776 | * range, the logic can be tuned by the following constants: | |
777 | * | |
778 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
779 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
780 | */ | |
781 | #define SAMPLING_READ_SIZE (16) | |
782 | #define SAMPLING_INTERVAL (256) | |
783 | ||
784 | /* | |
785 | * For statistical analysis of the input data we consider bytes that form a | |
786 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
787 | * many times the object appeared in the sample. | |
788 | */ | |
789 | #define BUCKET_SIZE (256) | |
790 | ||
791 | /* | |
792 | * The size of the sample is based on a statistical sampling rule of thumb. | |
793 | * The common way is to perform sampling tests as long as the number of | |
794 | * elements in each cell is at least 5. | |
795 | * | |
796 | * Instead of 5, we choose 32 to obtain more accurate results. | |
797 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
798 | * sample size bound by 8192. | |
799 | * | |
800 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
801 | * from up to 512 locations. | |
802 | */ | |
803 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
804 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
805 | ||
806 | struct bucket_item { | |
807 | u32 count; | |
808 | }; | |
4e439a0b TT |
809 | |
810 | struct heuristic_ws { | |
17b5a6c1 TT |
811 | /* Partial copy of input data */ |
812 | u8 *sample; | |
a440d48c | 813 | u32 sample_size; |
17b5a6c1 TT |
814 | /* Buckets store counters for each byte value */ |
815 | struct bucket_item *bucket; | |
440c840c TT |
816 | /* Sorting buffer */ |
817 | struct bucket_item *bucket_b; | |
4e439a0b TT |
818 | struct list_head list; |
819 | }; | |
820 | ||
92ee5530 DZ |
821 | static struct workspace_manager heuristic_wsm; |
822 | ||
4e439a0b TT |
823 | static void free_heuristic_ws(struct list_head *ws) |
824 | { | |
825 | struct heuristic_ws *workspace; | |
826 | ||
827 | workspace = list_entry(ws, struct heuristic_ws, list); | |
828 | ||
17b5a6c1 TT |
829 | kvfree(workspace->sample); |
830 | kfree(workspace->bucket); | |
440c840c | 831 | kfree(workspace->bucket_b); |
4e439a0b TT |
832 | kfree(workspace); |
833 | } | |
834 | ||
7bf49943 | 835 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
836 | { |
837 | struct heuristic_ws *ws; | |
838 | ||
839 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
840 | if (!ws) | |
841 | return ERR_PTR(-ENOMEM); | |
842 | ||
17b5a6c1 TT |
843 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
844 | if (!ws->sample) | |
845 | goto fail; | |
846 | ||
847 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
848 | if (!ws->bucket) | |
849 | goto fail; | |
4e439a0b | 850 | |
440c840c TT |
851 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
852 | if (!ws->bucket_b) | |
853 | goto fail; | |
854 | ||
17b5a6c1 | 855 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 856 | return &ws->list; |
17b5a6c1 TT |
857 | fail: |
858 | free_heuristic_ws(&ws->list); | |
859 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
860 | } |
861 | ||
ca4ac360 | 862 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 863 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
864 | }; |
865 | ||
e8c9f186 | 866 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
867 | /* The heuristic is represented as compression type 0 */ |
868 | &btrfs_heuristic_compress, | |
261507a0 | 869 | &btrfs_zlib_compress, |
a6fa6fae | 870 | &btrfs_lzo_compress, |
5c1aab1d | 871 | &btrfs_zstd_compress, |
261507a0 LZ |
872 | }; |
873 | ||
c778df14 DS |
874 | static struct list_head *alloc_workspace(int type, unsigned int level) |
875 | { | |
876 | switch (type) { | |
877 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
878 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
879 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
880 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
881 | default: | |
882 | /* | |
883 | * This can't happen, the type is validated several times | |
884 | * before we get here. | |
885 | */ | |
886 | BUG(); | |
887 | } | |
888 | } | |
889 | ||
1e002351 DS |
890 | static void free_workspace(int type, struct list_head *ws) |
891 | { | |
892 | switch (type) { | |
893 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
894 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
895 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
896 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
897 | default: | |
898 | /* | |
899 | * This can't happen, the type is validated several times | |
900 | * before we get here. | |
901 | */ | |
902 | BUG(); | |
903 | } | |
904 | } | |
905 | ||
d5517033 | 906 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 907 | { |
0cf25213 | 908 | struct workspace_manager *wsm; |
4e439a0b | 909 | struct list_head *workspace; |
261507a0 | 910 | |
0cf25213 | 911 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
912 | INIT_LIST_HEAD(&wsm->idle_ws); |
913 | spin_lock_init(&wsm->ws_lock); | |
914 | atomic_set(&wsm->total_ws, 0); | |
915 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 916 | |
1666edab DZ |
917 | /* |
918 | * Preallocate one workspace for each compression type so we can | |
919 | * guarantee forward progress in the worst case | |
920 | */ | |
c778df14 | 921 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
922 | if (IS_ERR(workspace)) { |
923 | pr_warn( | |
924 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
925 | } else { | |
92ee5530 DZ |
926 | atomic_set(&wsm->total_ws, 1); |
927 | wsm->free_ws = 1; | |
928 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
929 | } |
930 | } | |
931 | ||
2510307e | 932 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 933 | { |
2dba7143 | 934 | struct workspace_manager *wsman; |
1666edab DZ |
935 | struct list_head *ws; |
936 | ||
2dba7143 | 937 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
938 | while (!list_empty(&wsman->idle_ws)) { |
939 | ws = wsman->idle_ws.next; | |
940 | list_del(ws); | |
1e002351 | 941 | free_workspace(type, ws); |
1666edab | 942 | atomic_dec(&wsman->total_ws); |
261507a0 | 943 | } |
261507a0 LZ |
944 | } |
945 | ||
946 | /* | |
e721e49d DS |
947 | * This finds an available workspace or allocates a new one. |
948 | * If it's not possible to allocate a new one, waits until there's one. | |
949 | * Preallocation makes a forward progress guarantees and we do not return | |
950 | * errors. | |
261507a0 | 951 | */ |
5907a9bb | 952 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 953 | { |
5907a9bb | 954 | struct workspace_manager *wsm; |
261507a0 LZ |
955 | struct list_head *workspace; |
956 | int cpus = num_online_cpus(); | |
fe308533 | 957 | unsigned nofs_flag; |
4e439a0b TT |
958 | struct list_head *idle_ws; |
959 | spinlock_t *ws_lock; | |
960 | atomic_t *total_ws; | |
961 | wait_queue_head_t *ws_wait; | |
962 | int *free_ws; | |
963 | ||
5907a9bb | 964 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
965 | idle_ws = &wsm->idle_ws; |
966 | ws_lock = &wsm->ws_lock; | |
967 | total_ws = &wsm->total_ws; | |
968 | ws_wait = &wsm->ws_wait; | |
969 | free_ws = &wsm->free_ws; | |
261507a0 | 970 | |
261507a0 | 971 | again: |
d9187649 BL |
972 | spin_lock(ws_lock); |
973 | if (!list_empty(idle_ws)) { | |
974 | workspace = idle_ws->next; | |
261507a0 | 975 | list_del(workspace); |
6ac10a6a | 976 | (*free_ws)--; |
d9187649 | 977 | spin_unlock(ws_lock); |
261507a0 LZ |
978 | return workspace; |
979 | ||
980 | } | |
6ac10a6a | 981 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
982 | DEFINE_WAIT(wait); |
983 | ||
d9187649 BL |
984 | spin_unlock(ws_lock); |
985 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 986 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 987 | schedule(); |
d9187649 | 988 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
989 | goto again; |
990 | } | |
6ac10a6a | 991 | atomic_inc(total_ws); |
d9187649 | 992 | spin_unlock(ws_lock); |
261507a0 | 993 | |
fe308533 DS |
994 | /* |
995 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
996 | * to turn it off here because we might get called from the restricted | |
997 | * context of btrfs_compress_bio/btrfs_compress_pages | |
998 | */ | |
999 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1000 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1001 | memalloc_nofs_restore(nofs_flag); |
1002 | ||
261507a0 | 1003 | if (IS_ERR(workspace)) { |
6ac10a6a | 1004 | atomic_dec(total_ws); |
d9187649 | 1005 | wake_up(ws_wait); |
e721e49d DS |
1006 | |
1007 | /* | |
1008 | * Do not return the error but go back to waiting. There's a | |
1009 | * workspace preallocated for each type and the compression | |
1010 | * time is bounded so we get to a workspace eventually. This | |
1011 | * makes our caller's life easier. | |
52356716 DS |
1012 | * |
1013 | * To prevent silent and low-probability deadlocks (when the | |
1014 | * initial preallocation fails), check if there are any | |
1015 | * workspaces at all. | |
e721e49d | 1016 | */ |
52356716 DS |
1017 | if (atomic_read(total_ws) == 0) { |
1018 | static DEFINE_RATELIMIT_STATE(_rs, | |
1019 | /* once per minute */ 60 * HZ, | |
1020 | /* no burst */ 1); | |
1021 | ||
1022 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1023 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1024 | } |
1025 | } | |
e721e49d | 1026 | goto again; |
261507a0 LZ |
1027 | } |
1028 | return workspace; | |
1029 | } | |
1030 | ||
7bf49943 | 1031 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1032 | { |
6a0d1272 | 1033 | switch (type) { |
5907a9bb | 1034 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1035 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1036 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1037 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1038 | default: | |
1039 | /* | |
1040 | * This can't happen, the type is validated several times | |
1041 | * before we get here. | |
1042 | */ | |
1043 | BUG(); | |
1044 | } | |
929f4baf DZ |
1045 | } |
1046 | ||
261507a0 LZ |
1047 | /* |
1048 | * put a workspace struct back on the list or free it if we have enough | |
1049 | * idle ones sitting around | |
1050 | */ | |
a3bbd2a9 | 1051 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1052 | { |
a3bbd2a9 | 1053 | struct workspace_manager *wsm; |
4e439a0b TT |
1054 | struct list_head *idle_ws; |
1055 | spinlock_t *ws_lock; | |
1056 | atomic_t *total_ws; | |
1057 | wait_queue_head_t *ws_wait; | |
1058 | int *free_ws; | |
1059 | ||
a3bbd2a9 | 1060 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1061 | idle_ws = &wsm->idle_ws; |
1062 | ws_lock = &wsm->ws_lock; | |
1063 | total_ws = &wsm->total_ws; | |
1064 | ws_wait = &wsm->ws_wait; | |
1065 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1066 | |
1067 | spin_lock(ws_lock); | |
26b28dce | 1068 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1069 | list_add(ws, idle_ws); |
6ac10a6a | 1070 | (*free_ws)++; |
d9187649 | 1071 | spin_unlock(ws_lock); |
261507a0 LZ |
1072 | goto wake; |
1073 | } | |
d9187649 | 1074 | spin_unlock(ws_lock); |
261507a0 | 1075 | |
1e002351 | 1076 | free_workspace(type, ws); |
6ac10a6a | 1077 | atomic_dec(total_ws); |
261507a0 | 1078 | wake: |
093258e6 | 1079 | cond_wake_up(ws_wait); |
261507a0 LZ |
1080 | } |
1081 | ||
929f4baf DZ |
1082 | static void put_workspace(int type, struct list_head *ws) |
1083 | { | |
bd3a5287 | 1084 | switch (type) { |
a3bbd2a9 DS |
1085 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1086 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1087 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1088 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1089 | default: | |
1090 | /* | |
1091 | * This can't happen, the type is validated several times | |
1092 | * before we get here. | |
1093 | */ | |
1094 | BUG(); | |
1095 | } | |
929f4baf DZ |
1096 | } |
1097 | ||
adbab642 AJ |
1098 | /* |
1099 | * Adjust @level according to the limits of the compression algorithm or | |
1100 | * fallback to default | |
1101 | */ | |
1102 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1103 | { | |
1104 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1105 | ||
1106 | if (level == 0) | |
1107 | level = ops->default_level; | |
1108 | else | |
1109 | level = min(level, ops->max_level); | |
1110 | ||
1111 | return level; | |
1112 | } | |
1113 | ||
261507a0 | 1114 | /* |
38c31464 DS |
1115 | * Given an address space and start and length, compress the bytes into @pages |
1116 | * that are allocated on demand. | |
261507a0 | 1117 | * |
f51d2b59 DS |
1118 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1119 | * default the algorithm chooses and is opaque here; | |
1120 | * - compression algo are 0-3 | |
1121 | * - the level are bits 4-7 | |
1122 | * | |
4d3a800e DS |
1123 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1124 | * and returns number of actually allocated pages | |
261507a0 | 1125 | * |
38c31464 DS |
1126 | * @total_in is used to return the number of bytes actually read. It |
1127 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1128 | * ran out of room in the pages array or because we cross the |
1129 | * max_out threshold. | |
1130 | * | |
38c31464 DS |
1131 | * @total_out is an in/out parameter, must be set to the input length and will |
1132 | * be also used to return the total number of compressed bytes | |
261507a0 | 1133 | * |
38c31464 | 1134 | * @max_out tells us the max number of bytes that we're allowed to |
261507a0 LZ |
1135 | * stuff into pages |
1136 | */ | |
f51d2b59 | 1137 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1138 | u64 start, struct page **pages, |
261507a0 LZ |
1139 | unsigned long *out_pages, |
1140 | unsigned long *total_in, | |
e5d74902 | 1141 | unsigned long *total_out) |
261507a0 | 1142 | { |
1972708a | 1143 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1144 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1145 | struct list_head *workspace; |
1146 | int ret; | |
1147 | ||
b0c1fe1e | 1148 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1149 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1150 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1151 | out_pages, total_in, total_out); | |
929f4baf | 1152 | put_workspace(type, workspace); |
261507a0 LZ |
1153 | return ret; |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * pages_in is an array of pages with compressed data. | |
1158 | * | |
1159 | * disk_start is the starting logical offset of this array in the file | |
1160 | * | |
974b1adc | 1161 | * orig_bio contains the pages from the file that we want to decompress into |
261507a0 LZ |
1162 | * |
1163 | * srclen is the number of bytes in pages_in | |
1164 | * | |
1165 | * The basic idea is that we have a bio that was created by readpages. | |
1166 | * The pages in the bio are for the uncompressed data, and they may not | |
1167 | * be contiguous. They all correspond to the range of bytes covered by | |
1168 | * the compressed extent. | |
1169 | */ | |
8140dc30 | 1170 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1171 | { |
1172 | struct list_head *workspace; | |
1173 | int ret; | |
8140dc30 | 1174 | int type = cb->compress_type; |
261507a0 | 1175 | |
7bf49943 | 1176 | workspace = get_workspace(type, 0); |
1e4eb746 | 1177 | ret = compression_decompress_bio(type, workspace, cb); |
929f4baf | 1178 | put_workspace(type, workspace); |
e1ddce71 | 1179 | |
261507a0 LZ |
1180 | return ret; |
1181 | } | |
1182 | ||
1183 | /* | |
1184 | * a less complex decompression routine. Our compressed data fits in a | |
1185 | * single page, and we want to read a single page out of it. | |
1186 | * start_byte tells us the offset into the compressed data we're interested in | |
1187 | */ | |
1188 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1189 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1190 | { | |
1191 | struct list_head *workspace; | |
1192 | int ret; | |
1193 | ||
7bf49943 | 1194 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1195 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1196 | start_byte, srclen, destlen); | |
929f4baf | 1197 | put_workspace(type, workspace); |
7bf49943 | 1198 | |
261507a0 LZ |
1199 | return ret; |
1200 | } | |
1201 | ||
1666edab DZ |
1202 | void __init btrfs_init_compress(void) |
1203 | { | |
d5517033 DS |
1204 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1205 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1206 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1207 | zstd_init_workspace_manager(); | |
1666edab DZ |
1208 | } |
1209 | ||
e67c718b | 1210 | void __cold btrfs_exit_compress(void) |
261507a0 | 1211 | { |
2510307e DS |
1212 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1213 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1214 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1215 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1216 | } |
3a39c18d LZ |
1217 | |
1218 | /* | |
1219 | * Copy uncompressed data from working buffer to pages. | |
1220 | * | |
1221 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1222 | * | |
1223 | * total_out is the last byte of the buffer | |
1224 | */ | |
14a3357b | 1225 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1226 | unsigned long total_out, u64 disk_start, |
974b1adc | 1227 | struct bio *bio) |
3a39c18d LZ |
1228 | { |
1229 | unsigned long buf_offset; | |
1230 | unsigned long current_buf_start; | |
1231 | unsigned long start_byte; | |
6e78b3f7 | 1232 | unsigned long prev_start_byte; |
3a39c18d LZ |
1233 | unsigned long working_bytes = total_out - buf_start; |
1234 | unsigned long bytes; | |
1235 | char *kaddr; | |
974b1adc | 1236 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1237 | |
1238 | /* | |
1239 | * start byte is the first byte of the page we're currently | |
1240 | * copying into relative to the start of the compressed data. | |
1241 | */ | |
974b1adc | 1242 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1243 | |
1244 | /* we haven't yet hit data corresponding to this page */ | |
1245 | if (total_out <= start_byte) | |
1246 | return 1; | |
1247 | ||
1248 | /* | |
1249 | * the start of the data we care about is offset into | |
1250 | * the middle of our working buffer | |
1251 | */ | |
1252 | if (total_out > start_byte && buf_start < start_byte) { | |
1253 | buf_offset = start_byte - buf_start; | |
1254 | working_bytes -= buf_offset; | |
1255 | } else { | |
1256 | buf_offset = 0; | |
1257 | } | |
1258 | current_buf_start = buf_start; | |
1259 | ||
1260 | /* copy bytes from the working buffer into the pages */ | |
1261 | while (working_bytes > 0) { | |
974b1adc | 1262 | bytes = min_t(unsigned long, bvec.bv_len, |
3fd396af | 1263 | PAGE_SIZE - (buf_offset % PAGE_SIZE)); |
3a39c18d | 1264 | bytes = min(bytes, working_bytes); |
974b1adc CH |
1265 | |
1266 | kaddr = kmap_atomic(bvec.bv_page); | |
1267 | memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes); | |
7ac687d9 | 1268 | kunmap_atomic(kaddr); |
974b1adc | 1269 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1270 | |
3a39c18d LZ |
1271 | buf_offset += bytes; |
1272 | working_bytes -= bytes; | |
1273 | current_buf_start += bytes; | |
1274 | ||
1275 | /* check if we need to pick another page */ | |
974b1adc CH |
1276 | bio_advance(bio, bytes); |
1277 | if (!bio->bi_iter.bi_size) | |
1278 | return 0; | |
1279 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1280 | prev_start_byte = start_byte; |
974b1adc | 1281 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1282 | |
974b1adc | 1283 | /* |
6e78b3f7 OS |
1284 | * We need to make sure we're only adjusting |
1285 | * our offset into compression working buffer when | |
1286 | * we're switching pages. Otherwise we can incorrectly | |
1287 | * keep copying when we were actually done. | |
974b1adc | 1288 | */ |
6e78b3f7 OS |
1289 | if (start_byte != prev_start_byte) { |
1290 | /* | |
1291 | * make sure our new page is covered by this | |
1292 | * working buffer | |
1293 | */ | |
1294 | if (total_out <= start_byte) | |
1295 | return 1; | |
3a39c18d | 1296 | |
6e78b3f7 OS |
1297 | /* |
1298 | * the next page in the biovec might not be adjacent | |
1299 | * to the last page, but it might still be found | |
1300 | * inside this working buffer. bump our offset pointer | |
1301 | */ | |
1302 | if (total_out > start_byte && | |
1303 | current_buf_start < start_byte) { | |
1304 | buf_offset = start_byte - buf_start; | |
1305 | working_bytes = total_out - start_byte; | |
1306 | current_buf_start = buf_start + buf_offset; | |
1307 | } | |
3a39c18d LZ |
1308 | } |
1309 | } | |
1310 | ||
1311 | return 1; | |
1312 | } | |
c2fcdcdf | 1313 | |
19562430 TT |
1314 | /* |
1315 | * Shannon Entropy calculation | |
1316 | * | |
52042d8e | 1317 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1318 | * Try calculating entropy to estimate the average minimum number of bits |
1319 | * needed to encode the sampled data. | |
1320 | * | |
1321 | * For convenience, return the percentage of needed bits, instead of amount of | |
1322 | * bits directly. | |
1323 | * | |
1324 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1325 | * and can be compressible with high probability | |
1326 | * | |
1327 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1328 | * | |
1329 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1330 | */ | |
1331 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1332 | #define ENTROPY_LVL_HIGH (80) | |
1333 | ||
1334 | /* | |
1335 | * For increasead precision in shannon_entropy calculation, | |
1336 | * let's do pow(n, M) to save more digits after comma: | |
1337 | * | |
1338 | * - maximum int bit length is 64 | |
1339 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1340 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1341 | * | |
1342 | * So use pow(n, 4). | |
1343 | */ | |
1344 | static inline u32 ilog2_w(u64 n) | |
1345 | { | |
1346 | return ilog2(n * n * n * n); | |
1347 | } | |
1348 | ||
1349 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1350 | { | |
1351 | const u32 entropy_max = 8 * ilog2_w(2); | |
1352 | u32 entropy_sum = 0; | |
1353 | u32 p, p_base, sz_base; | |
1354 | u32 i; | |
1355 | ||
1356 | sz_base = ilog2_w(ws->sample_size); | |
1357 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1358 | p = ws->bucket[i].count; | |
1359 | p_base = ilog2_w(p); | |
1360 | entropy_sum += p * (sz_base - p_base); | |
1361 | } | |
1362 | ||
1363 | entropy_sum /= ws->sample_size; | |
1364 | return entropy_sum * 100 / entropy_max; | |
1365 | } | |
1366 | ||
440c840c TT |
1367 | #define RADIX_BASE 4U |
1368 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1369 | ||
1370 | static u8 get4bits(u64 num, int shift) { | |
1371 | u8 low4bits; | |
1372 | ||
1373 | num >>= shift; | |
1374 | /* Reverse order */ | |
1375 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1376 | return low4bits; | |
1377 | } | |
1378 | ||
440c840c TT |
1379 | /* |
1380 | * Use 4 bits as radix base | |
52042d8e | 1381 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1382 | * |
1383 | * @array - array that will be sorted | |
1384 | * @array_buf - buffer array to store sorting results | |
1385 | * must be equal in size to @array | |
1386 | * @num - array size | |
440c840c | 1387 | */ |
23ae8c63 | 1388 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1389 | int num) |
858177d3 | 1390 | { |
440c840c TT |
1391 | u64 max_num; |
1392 | u64 buf_num; | |
1393 | u32 counters[COUNTERS_SIZE]; | |
1394 | u32 new_addr; | |
1395 | u32 addr; | |
1396 | int bitlen; | |
1397 | int shift; | |
1398 | int i; | |
858177d3 | 1399 | |
440c840c TT |
1400 | /* |
1401 | * Try avoid useless loop iterations for small numbers stored in big | |
1402 | * counters. Example: 48 33 4 ... in 64bit array | |
1403 | */ | |
23ae8c63 | 1404 | max_num = array[0].count; |
440c840c | 1405 | for (i = 1; i < num; i++) { |
23ae8c63 | 1406 | buf_num = array[i].count; |
440c840c TT |
1407 | if (buf_num > max_num) |
1408 | max_num = buf_num; | |
1409 | } | |
1410 | ||
1411 | buf_num = ilog2(max_num); | |
1412 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1413 | ||
1414 | shift = 0; | |
1415 | while (shift < bitlen) { | |
1416 | memset(counters, 0, sizeof(counters)); | |
1417 | ||
1418 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1419 | buf_num = array[i].count; |
440c840c TT |
1420 | addr = get4bits(buf_num, shift); |
1421 | counters[addr]++; | |
1422 | } | |
1423 | ||
1424 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1425 | counters[i] += counters[i - 1]; | |
1426 | ||
1427 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1428 | buf_num = array[i].count; |
440c840c TT |
1429 | addr = get4bits(buf_num, shift); |
1430 | counters[addr]--; | |
1431 | new_addr = counters[addr]; | |
7add17be | 1432 | array_buf[new_addr] = array[i]; |
440c840c TT |
1433 | } |
1434 | ||
1435 | shift += RADIX_BASE; | |
1436 | ||
1437 | /* | |
1438 | * Normal radix expects to move data from a temporary array, to | |
1439 | * the main one. But that requires some CPU time. Avoid that | |
1440 | * by doing another sort iteration to original array instead of | |
1441 | * memcpy() | |
1442 | */ | |
1443 | memset(counters, 0, sizeof(counters)); | |
1444 | ||
1445 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1446 | buf_num = array_buf[i].count; |
440c840c TT |
1447 | addr = get4bits(buf_num, shift); |
1448 | counters[addr]++; | |
1449 | } | |
1450 | ||
1451 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1452 | counters[i] += counters[i - 1]; | |
1453 | ||
1454 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1455 | buf_num = array_buf[i].count; |
440c840c TT |
1456 | addr = get4bits(buf_num, shift); |
1457 | counters[addr]--; | |
1458 | new_addr = counters[addr]; | |
7add17be | 1459 | array[new_addr] = array_buf[i]; |
440c840c TT |
1460 | } |
1461 | ||
1462 | shift += RADIX_BASE; | |
1463 | } | |
858177d3 TT |
1464 | } |
1465 | ||
1466 | /* | |
1467 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1468 | * | |
1469 | * There are several types of structured binary data that use nearly all byte | |
1470 | * values. The distribution can be uniform and counts in all buckets will be | |
1471 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1472 | * | |
1473 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1474 | * be potentially compressible, but we have to take a few more steps to decide | |
1475 | * how much. | |
1476 | * | |
1477 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1478 | * compression algo can easy fix that | |
1479 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1480 | * probability is not compressible | |
1481 | */ | |
1482 | #define BYTE_CORE_SET_LOW (64) | |
1483 | #define BYTE_CORE_SET_HIGH (200) | |
1484 | ||
1485 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1486 | { | |
1487 | u32 i; | |
1488 | u32 coreset_sum = 0; | |
1489 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1490 | struct bucket_item *bucket = ws->bucket; | |
1491 | ||
1492 | /* Sort in reverse order */ | |
36243c91 | 1493 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1494 | |
1495 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1496 | coreset_sum += bucket[i].count; | |
1497 | ||
1498 | if (coreset_sum > core_set_threshold) | |
1499 | return i; | |
1500 | ||
1501 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1502 | coreset_sum += bucket[i].count; | |
1503 | if (coreset_sum > core_set_threshold) | |
1504 | break; | |
1505 | } | |
1506 | ||
1507 | return i; | |
1508 | } | |
1509 | ||
a288e92c TT |
1510 | /* |
1511 | * Count byte values in buckets. | |
1512 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1513 | * Because in most text-like data byte set is restricted to limited number of | |
1514 | * possible characters, and that restriction in most cases makes data easy to | |
1515 | * compress. | |
1516 | * | |
1517 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1518 | * less - compressible | |
1519 | * more - need additional analysis | |
1520 | */ | |
1521 | #define BYTE_SET_THRESHOLD (64) | |
1522 | ||
1523 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1524 | { | |
1525 | u32 i; | |
1526 | u32 byte_set_size = 0; | |
1527 | ||
1528 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1529 | if (ws->bucket[i].count > 0) | |
1530 | byte_set_size++; | |
1531 | } | |
1532 | ||
1533 | /* | |
1534 | * Continue collecting count of byte values in buckets. If the byte | |
1535 | * set size is bigger then the threshold, it's pointless to continue, | |
1536 | * the detection technique would fail for this type of data. | |
1537 | */ | |
1538 | for (; i < BUCKET_SIZE; i++) { | |
1539 | if (ws->bucket[i].count > 0) { | |
1540 | byte_set_size++; | |
1541 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1542 | return byte_set_size; | |
1543 | } | |
1544 | } | |
1545 | ||
1546 | return byte_set_size; | |
1547 | } | |
1548 | ||
1fe4f6fa TT |
1549 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1550 | { | |
1551 | const u32 half_of_sample = ws->sample_size / 2; | |
1552 | const u8 *data = ws->sample; | |
1553 | ||
1554 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1555 | } | |
1556 | ||
a440d48c TT |
1557 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1558 | struct heuristic_ws *ws) | |
1559 | { | |
1560 | struct page *page; | |
1561 | u64 index, index_end; | |
1562 | u32 i, curr_sample_pos; | |
1563 | u8 *in_data; | |
1564 | ||
1565 | /* | |
1566 | * Compression handles the input data by chunks of 128KiB | |
1567 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1568 | * | |
1569 | * We do the same for the heuristic and loop over the whole range. | |
1570 | * | |
1571 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1572 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1573 | */ | |
1574 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1575 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1576 | ||
1577 | index = start >> PAGE_SHIFT; | |
1578 | index_end = end >> PAGE_SHIFT; | |
1579 | ||
1580 | /* Don't miss unaligned end */ | |
1581 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1582 | index_end++; | |
1583 | ||
1584 | curr_sample_pos = 0; | |
1585 | while (index < index_end) { | |
1586 | page = find_get_page(inode->i_mapping, index); | |
1587 | in_data = kmap(page); | |
1588 | /* Handle case where the start is not aligned to PAGE_SIZE */ | |
1589 | i = start % PAGE_SIZE; | |
1590 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1591 | /* Don't sample any garbage from the last page */ | |
1592 | if (start > end - SAMPLING_READ_SIZE) | |
1593 | break; | |
1594 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1595 | SAMPLING_READ_SIZE); | |
1596 | i += SAMPLING_INTERVAL; | |
1597 | start += SAMPLING_INTERVAL; | |
1598 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1599 | } | |
1600 | kunmap(page); | |
1601 | put_page(page); | |
1602 | ||
1603 | index++; | |
1604 | } | |
1605 | ||
1606 | ws->sample_size = curr_sample_pos; | |
1607 | } | |
1608 | ||
c2fcdcdf TT |
1609 | /* |
1610 | * Compression heuristic. | |
1611 | * | |
1612 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1613 | * quickly (compared to direct compression) detect data characteristics | |
1614 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1615 | * data. | |
1616 | * | |
1617 | * The following types of analysis can be performed: | |
1618 | * - detect mostly zero data | |
1619 | * - detect data with low "byte set" size (text, etc) | |
1620 | * - detect data with low/high "core byte" set | |
1621 | * | |
1622 | * Return non-zero if the compression should be done, 0 otherwise. | |
1623 | */ | |
1624 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1625 | { | |
7bf49943 | 1626 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1627 | struct heuristic_ws *ws; |
a440d48c TT |
1628 | u32 i; |
1629 | u8 byte; | |
19562430 | 1630 | int ret = 0; |
c2fcdcdf | 1631 | |
4e439a0b TT |
1632 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1633 | ||
a440d48c TT |
1634 | heuristic_collect_sample(inode, start, end, ws); |
1635 | ||
1fe4f6fa TT |
1636 | if (sample_repeated_patterns(ws)) { |
1637 | ret = 1; | |
1638 | goto out; | |
1639 | } | |
1640 | ||
a440d48c TT |
1641 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1642 | ||
1643 | for (i = 0; i < ws->sample_size; i++) { | |
1644 | byte = ws->sample[i]; | |
1645 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1646 | } |
1647 | ||
a288e92c TT |
1648 | i = byte_set_size(ws); |
1649 | if (i < BYTE_SET_THRESHOLD) { | |
1650 | ret = 2; | |
1651 | goto out; | |
1652 | } | |
1653 | ||
858177d3 TT |
1654 | i = byte_core_set_size(ws); |
1655 | if (i <= BYTE_CORE_SET_LOW) { | |
1656 | ret = 3; | |
1657 | goto out; | |
1658 | } | |
1659 | ||
1660 | if (i >= BYTE_CORE_SET_HIGH) { | |
1661 | ret = 0; | |
1662 | goto out; | |
1663 | } | |
1664 | ||
19562430 TT |
1665 | i = shannon_entropy(ws); |
1666 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1667 | ret = 4; | |
1668 | goto out; | |
1669 | } | |
1670 | ||
1671 | /* | |
1672 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1673 | * needed to give green light to compression. | |
1674 | * | |
1675 | * For now just assume that compression at that level is not worth the | |
1676 | * resources because: | |
1677 | * | |
1678 | * 1. it is possible to defrag the data later | |
1679 | * | |
1680 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1681 | * values, every bucket has counter at level ~54. The heuristic would | |
1682 | * be confused. This can happen when data have some internal repeated | |
1683 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1684 | * pairs of bytes, which is too costly. | |
1685 | */ | |
1686 | if (i < ENTROPY_LVL_HIGH) { | |
1687 | ret = 5; | |
1688 | goto out; | |
1689 | } else { | |
1690 | ret = 0; | |
1691 | goto out; | |
1692 | } | |
1693 | ||
1fe4f6fa | 1694 | out: |
929f4baf | 1695 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1696 | return ret; |
1697 | } | |
f51d2b59 | 1698 | |
d0ab62ce DZ |
1699 | /* |
1700 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1701 | * level, unrecognized string will set the default level | |
1702 | */ | |
1703 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1704 | { |
d0ab62ce DZ |
1705 | unsigned int level = 0; |
1706 | int ret; | |
1707 | ||
1708 | if (!type) | |
f51d2b59 DS |
1709 | return 0; |
1710 | ||
d0ab62ce DZ |
1711 | if (str[0] == ':') { |
1712 | ret = kstrtouint(str + 1, 10, &level); | |
1713 | if (ret) | |
1714 | level = 0; | |
1715 | } | |
1716 | ||
b0c1fe1e DS |
1717 | level = btrfs_compress_set_level(type, level); |
1718 | ||
1719 | return level; | |
1720 | } |