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" | |
764c7c9a | 31 | #include "zoned.h" |
c8b97818 | 32 | |
e128f9c3 DS |
33 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
34 | ||
35 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
36 | { | |
37 | switch (type) { | |
38 | case BTRFS_COMPRESS_ZLIB: | |
39 | case BTRFS_COMPRESS_LZO: | |
40 | case BTRFS_COMPRESS_ZSTD: | |
41 | case BTRFS_COMPRESS_NONE: | |
42 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
43 | default: |
44 | break; | |
e128f9c3 DS |
45 | } |
46 | ||
47 | return NULL; | |
48 | } | |
49 | ||
aa53e3bf JT |
50 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
51 | { | |
52 | int i; | |
53 | ||
54 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
55 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
56 | ||
57 | if (len < comp_len) | |
58 | continue; | |
59 | ||
60 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
61 | return true; | |
62 | } | |
63 | return false; | |
64 | } | |
65 | ||
1e4eb746 DS |
66 | static int compression_compress_pages(int type, struct list_head *ws, |
67 | struct address_space *mapping, u64 start, struct page **pages, | |
68 | unsigned long *out_pages, unsigned long *total_in, | |
69 | unsigned long *total_out) | |
70 | { | |
71 | switch (type) { | |
72 | case BTRFS_COMPRESS_ZLIB: | |
73 | return zlib_compress_pages(ws, mapping, start, pages, | |
74 | out_pages, total_in, total_out); | |
75 | case BTRFS_COMPRESS_LZO: | |
76 | return lzo_compress_pages(ws, mapping, start, pages, | |
77 | out_pages, total_in, total_out); | |
78 | case BTRFS_COMPRESS_ZSTD: | |
79 | return zstd_compress_pages(ws, mapping, start, pages, | |
80 | out_pages, total_in, total_out); | |
81 | case BTRFS_COMPRESS_NONE: | |
82 | default: | |
83 | /* | |
1d8ba9e7 QW |
84 | * This can happen when compression races with remount setting |
85 | * it to 'no compress', while caller doesn't call | |
86 | * inode_need_compress() to check if we really need to | |
87 | * compress. | |
88 | * | |
89 | * Not a big deal, just need to inform caller that we | |
90 | * haven't allocated any pages yet. | |
1e4eb746 | 91 | */ |
1d8ba9e7 | 92 | *out_pages = 0; |
1e4eb746 DS |
93 | return -E2BIG; |
94 | } | |
95 | } | |
96 | ||
97 | static int compression_decompress_bio(int type, struct list_head *ws, | |
98 | struct compressed_bio *cb) | |
99 | { | |
100 | switch (type) { | |
101 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); | |
102 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
103 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
104 | case BTRFS_COMPRESS_NONE: | |
105 | default: | |
106 | /* | |
107 | * This can't happen, the type is validated several times | |
108 | * before we get here. | |
109 | */ | |
110 | BUG(); | |
111 | } | |
112 | } | |
113 | ||
114 | static int compression_decompress(int type, struct list_head *ws, | |
115 | unsigned char *data_in, struct page *dest_page, | |
116 | unsigned long start_byte, size_t srclen, size_t destlen) | |
117 | { | |
118 | switch (type) { | |
119 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
120 | start_byte, srclen, destlen); | |
121 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
122 | start_byte, srclen, destlen); | |
123 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
124 | start_byte, srclen, destlen); | |
125 | case BTRFS_COMPRESS_NONE: | |
126 | default: | |
127 | /* | |
128 | * This can't happen, the type is validated several times | |
129 | * before we get here. | |
130 | */ | |
131 | BUG(); | |
132 | } | |
133 | } | |
134 | ||
8140dc30 | 135 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 136 | |
2ff7e61e | 137 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
138 | unsigned long disk_size) |
139 | { | |
d20f7043 | 140 | return sizeof(struct compressed_bio) + |
713cebfb | 141 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size; |
d20f7043 CM |
142 | } |
143 | ||
5a9472fe | 144 | static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio, |
d20f7043 CM |
145 | u64 disk_start) |
146 | { | |
10fe6ca8 | 147 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
d5178578 | 148 | SHASH_DESC_ON_STACK(shash, fs_info->csum_shash); |
223486c2 | 149 | const u32 csum_size = fs_info->csum_size; |
04d4ba4c | 150 | const u32 sectorsize = fs_info->sectorsize; |
d20f7043 | 151 | struct page *page; |
1d08ce58 | 152 | unsigned int i; |
d20f7043 | 153 | char *kaddr; |
d5178578 | 154 | u8 csum[BTRFS_CSUM_SIZE]; |
5a9472fe | 155 | struct compressed_bio *cb = bio->bi_private; |
10fe6ca8 | 156 | u8 *cb_sum = cb->sums; |
d20f7043 | 157 | |
42437a63 | 158 | if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM)) |
d20f7043 CM |
159 | return 0; |
160 | ||
d5178578 JT |
161 | shash->tfm = fs_info->csum_shash; |
162 | ||
d20f7043 | 163 | for (i = 0; i < cb->nr_pages; i++) { |
04d4ba4c QW |
164 | u32 pg_offset; |
165 | u32 bytes_left = PAGE_SIZE; | |
d20f7043 | 166 | page = cb->compressed_pages[i]; |
d20f7043 | 167 | |
04d4ba4c QW |
168 | /* Determine the remaining bytes inside the page first */ |
169 | if (i == cb->nr_pages - 1) | |
170 | bytes_left = cb->compressed_len - i * PAGE_SIZE; | |
171 | ||
172 | /* Hash through the page sector by sector */ | |
173 | for (pg_offset = 0; pg_offset < bytes_left; | |
174 | pg_offset += sectorsize) { | |
4c2bf276 | 175 | kaddr = page_address(page); |
04d4ba4c QW |
176 | crypto_shash_digest(shash, kaddr + pg_offset, |
177 | sectorsize, csum); | |
04d4ba4c QW |
178 | |
179 | if (memcmp(&csum, cb_sum, csum_size) != 0) { | |
180 | btrfs_print_data_csum_error(inode, disk_start, | |
181 | csum, cb_sum, cb->mirror_num); | |
182 | if (btrfs_io_bio(bio)->device) | |
183 | btrfs_dev_stat_inc_and_print( | |
184 | btrfs_io_bio(bio)->device, | |
185 | BTRFS_DEV_STAT_CORRUPTION_ERRS); | |
186 | return -EIO; | |
187 | } | |
188 | cb_sum += csum_size; | |
189 | disk_start += sectorsize; | |
d20f7043 | 190 | } |
d20f7043 | 191 | } |
93c4c033 | 192 | return 0; |
d20f7043 CM |
193 | } |
194 | ||
c8b97818 CM |
195 | /* when we finish reading compressed pages from the disk, we |
196 | * decompress them and then run the bio end_io routines on the | |
197 | * decompressed pages (in the inode address space). | |
198 | * | |
199 | * This allows the checksumming and other IO error handling routines | |
200 | * to work normally | |
201 | * | |
202 | * The compressed pages are freed here, and it must be run | |
203 | * in process context | |
204 | */ | |
4246a0b6 | 205 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 206 | { |
c8b97818 CM |
207 | struct compressed_bio *cb = bio->bi_private; |
208 | struct inode *inode; | |
209 | struct page *page; | |
1d08ce58 | 210 | unsigned int index; |
cf1167d5 | 211 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 212 | int ret = 0; |
c8b97818 | 213 | |
4e4cbee9 | 214 | if (bio->bi_status) |
c8b97818 CM |
215 | cb->errors = 1; |
216 | ||
217 | /* if there are more bios still pending for this compressed | |
218 | * extent, just exit | |
219 | */ | |
a50299ae | 220 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
221 | goto out; |
222 | ||
cf1167d5 LB |
223 | /* |
224 | * Record the correct mirror_num in cb->orig_bio so that | |
225 | * read-repair can work properly. | |
226 | */ | |
cf1167d5 LB |
227 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; |
228 | cb->mirror_num = mirror; | |
229 | ||
e6311f24 LB |
230 | /* |
231 | * Some IO in this cb have failed, just skip checksum as there | |
232 | * is no way it could be correct. | |
233 | */ | |
234 | if (cb->errors == 1) | |
235 | goto csum_failed; | |
236 | ||
d20f7043 | 237 | inode = cb->inode; |
5a9472fe | 238 | ret = check_compressed_csum(BTRFS_I(inode), bio, |
1201b58b | 239 | bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
240 | if (ret) |
241 | goto csum_failed; | |
242 | ||
c8b97818 CM |
243 | /* ok, we're the last bio for this extent, lets start |
244 | * the decompression. | |
245 | */ | |
8140dc30 AJ |
246 | ret = btrfs_decompress_bio(cb); |
247 | ||
d20f7043 | 248 | csum_failed: |
c8b97818 CM |
249 | if (ret) |
250 | cb->errors = 1; | |
251 | ||
252 | /* release the compressed pages */ | |
253 | index = 0; | |
254 | for (index = 0; index < cb->nr_pages; index++) { | |
255 | page = cb->compressed_pages[index]; | |
256 | page->mapping = NULL; | |
09cbfeaf | 257 | put_page(page); |
c8b97818 CM |
258 | } |
259 | ||
260 | /* do io completion on the original bio */ | |
771ed689 | 261 | if (cb->errors) { |
c8b97818 | 262 | bio_io_error(cb->orig_bio); |
d20f7043 | 263 | } else { |
2c30c71b | 264 | struct bio_vec *bvec; |
6dc4f100 | 265 | struct bvec_iter_all iter_all; |
d20f7043 CM |
266 | |
267 | /* | |
268 | * we have verified the checksum already, set page | |
269 | * checked so the end_io handlers know about it | |
270 | */ | |
c09abff8 | 271 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2b070cfe | 272 | bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) |
d20f7043 | 273 | SetPageChecked(bvec->bv_page); |
2c30c71b | 274 | |
4246a0b6 | 275 | bio_endio(cb->orig_bio); |
d20f7043 | 276 | } |
c8b97818 CM |
277 | |
278 | /* finally free the cb struct */ | |
279 | kfree(cb->compressed_pages); | |
280 | kfree(cb); | |
281 | out: | |
282 | bio_put(bio); | |
283 | } | |
284 | ||
285 | /* | |
286 | * Clear the writeback bits on all of the file | |
287 | * pages for a compressed write | |
288 | */ | |
7bdcefc1 FM |
289 | static noinline void end_compressed_writeback(struct inode *inode, |
290 | const struct compressed_bio *cb) | |
c8b97818 | 291 | { |
09cbfeaf KS |
292 | unsigned long index = cb->start >> PAGE_SHIFT; |
293 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
294 | struct page *pages[16]; |
295 | unsigned long nr_pages = end_index - index + 1; | |
296 | int i; | |
297 | int ret; | |
298 | ||
7bdcefc1 FM |
299 | if (cb->errors) |
300 | mapping_set_error(inode->i_mapping, -EIO); | |
301 | ||
d397712b | 302 | while (nr_pages > 0) { |
c8b97818 | 303 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
304 | min_t(unsigned long, |
305 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
306 | if (ret == 0) { |
307 | nr_pages -= 1; | |
308 | index += 1; | |
309 | continue; | |
310 | } | |
311 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
312 | if (cb->errors) |
313 | SetPageError(pages[i]); | |
c8b97818 | 314 | end_page_writeback(pages[i]); |
09cbfeaf | 315 | put_page(pages[i]); |
c8b97818 CM |
316 | } |
317 | nr_pages -= ret; | |
318 | index += ret; | |
319 | } | |
320 | /* the inode may be gone now */ | |
c8b97818 CM |
321 | } |
322 | ||
323 | /* | |
324 | * do the cleanup once all the compressed pages hit the disk. | |
325 | * This will clear writeback on the file pages and free the compressed | |
326 | * pages. | |
327 | * | |
328 | * This also calls the writeback end hooks for the file pages so that | |
329 | * metadata and checksums can be updated in the file. | |
330 | */ | |
4246a0b6 | 331 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 | 332 | { |
c8b97818 CM |
333 | struct compressed_bio *cb = bio->bi_private; |
334 | struct inode *inode; | |
335 | struct page *page; | |
1d08ce58 | 336 | unsigned int index; |
c8b97818 | 337 | |
4e4cbee9 | 338 | if (bio->bi_status) |
c8b97818 CM |
339 | cb->errors = 1; |
340 | ||
341 | /* if there are more bios still pending for this compressed | |
342 | * extent, just exit | |
343 | */ | |
a50299ae | 344 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
345 | goto out; |
346 | ||
347 | /* ok, we're the last bio for this extent, step one is to | |
348 | * call back into the FS and do all the end_io operations | |
349 | */ | |
350 | inode = cb->inode; | |
764c7c9a | 351 | btrfs_record_physical_zoned(inode, cb->start, bio); |
38a39ac7 | 352 | btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, |
c629732d | 353 | cb->start, cb->start + cb->len - 1, |
240246f6 | 354 | !cb->errors); |
c8b97818 | 355 | |
7bdcefc1 | 356 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
357 | /* note, our inode could be gone now */ |
358 | ||
359 | /* | |
360 | * release the compressed pages, these came from alloc_page and | |
361 | * are not attached to the inode at all | |
362 | */ | |
363 | index = 0; | |
364 | for (index = 0; index < cb->nr_pages; index++) { | |
365 | page = cb->compressed_pages[index]; | |
366 | page->mapping = NULL; | |
09cbfeaf | 367 | put_page(page); |
c8b97818 CM |
368 | } |
369 | ||
370 | /* finally free the cb struct */ | |
371 | kfree(cb->compressed_pages); | |
372 | kfree(cb); | |
373 | out: | |
374 | bio_put(bio); | |
375 | } | |
376 | ||
377 | /* | |
378 | * worker function to build and submit bios for previously compressed pages. | |
379 | * The corresponding pages in the inode should be marked for writeback | |
380 | * and the compressed pages should have a reference on them for dropping | |
381 | * when the IO is complete. | |
382 | * | |
383 | * This also checksums the file bytes and gets things ready for | |
384 | * the end io hooks. | |
385 | */ | |
c7ee1819 | 386 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
65b5355f AJ |
387 | unsigned int len, u64 disk_start, |
388 | unsigned int compressed_len, | |
c8b97818 | 389 | struct page **compressed_pages, |
65b5355f | 390 | unsigned int nr_pages, |
ec39f769 CM |
391 | unsigned int write_flags, |
392 | struct cgroup_subsys_state *blkcg_css) | |
c8b97818 | 393 | { |
c7ee1819 | 394 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 395 | struct bio *bio = NULL; |
c8b97818 CM |
396 | struct compressed_bio *cb; |
397 | unsigned long bytes_left; | |
306e16ce | 398 | int pg_index = 0; |
c8b97818 CM |
399 | struct page *page; |
400 | u64 first_byte = disk_start; | |
4e4cbee9 | 401 | blk_status_t ret; |
c7ee1819 | 402 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
764c7c9a JT |
403 | const bool use_append = btrfs_use_zone_append(inode, disk_start); |
404 | const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; | |
c8b97818 | 405 | |
fdb1e121 | 406 | WARN_ON(!PAGE_ALIGNED(start)); |
2ff7e61e | 407 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 408 | if (!cb) |
4e4cbee9 | 409 | return BLK_STS_RESOURCE; |
a50299ae | 410 | refcount_set(&cb->pending_bios, 0); |
c8b97818 | 411 | cb->errors = 0; |
c7ee1819 | 412 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
413 | cb->start = start; |
414 | cb->len = len; | |
d20f7043 | 415 | cb->mirror_num = 0; |
c8b97818 CM |
416 | cb->compressed_pages = compressed_pages; |
417 | cb->compressed_len = compressed_len; | |
418 | cb->orig_bio = NULL; | |
419 | cb->nr_pages = nr_pages; | |
420 | ||
e749af44 | 421 | bio = btrfs_bio_alloc(first_byte); |
764c7c9a | 422 | bio->bi_opf = bio_op | write_flags; |
c8b97818 CM |
423 | bio->bi_private = cb; |
424 | bio->bi_end_io = end_compressed_bio_write; | |
ec39f769 | 425 | |
764c7c9a | 426 | if (use_append) { |
e7ff9e6b | 427 | struct btrfs_device *device; |
764c7c9a | 428 | |
e7ff9e6b JT |
429 | device = btrfs_zoned_get_device(fs_info, disk_start, PAGE_SIZE); |
430 | if (IS_ERR(device)) { | |
764c7c9a JT |
431 | kfree(cb); |
432 | bio_put(bio); | |
433 | return BLK_STS_NOTSUPP; | |
434 | } | |
435 | ||
e7ff9e6b | 436 | bio_set_dev(bio, device->bdev); |
764c7c9a JT |
437 | } |
438 | ||
ec39f769 CM |
439 | if (blkcg_css) { |
440 | bio->bi_opf |= REQ_CGROUP_PUNT; | |
46bcff2b | 441 | kthread_associate_blkcg(blkcg_css); |
ec39f769 | 442 | } |
a50299ae | 443 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
444 | |
445 | /* create and submit bios for the compressed pages */ | |
446 | bytes_left = compressed_len; | |
306e16ce | 447 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 | 448 | int submit = 0; |
4c80a97d | 449 | int len = 0; |
4e4cbee9 | 450 | |
306e16ce | 451 | page = compressed_pages[pg_index]; |
c7ee1819 | 452 | page->mapping = inode->vfs_inode.i_mapping; |
4f024f37 | 453 | if (bio->bi_iter.bi_size) |
da12fe54 NB |
454 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, |
455 | 0); | |
c8b97818 | 456 | |
4c80a97d QW |
457 | /* |
458 | * Page can only be added to bio if the current bio fits in | |
459 | * stripe. | |
460 | */ | |
461 | if (!submit) { | |
462 | if (pg_index == 0 && use_append) | |
463 | len = bio_add_zone_append_page(bio, page, | |
464 | PAGE_SIZE, 0); | |
465 | else | |
466 | len = bio_add_page(bio, page, PAGE_SIZE, 0); | |
467 | } | |
764c7c9a | 468 | |
70b99e69 | 469 | page->mapping = NULL; |
764c7c9a | 470 | if (submit || len < PAGE_SIZE) { |
af09abfe CM |
471 | /* |
472 | * inc the count before we submit the bio so | |
473 | * we know the end IO handler won't happen before | |
474 | * we inc the count. Otherwise, the cb might get | |
475 | * freed before we're done setting it up | |
476 | */ | |
a50299ae | 477 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
478 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
479 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 480 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 481 | |
e55179b3 | 482 | if (!skip_sum) { |
c7ee1819 | 483 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 484 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 485 | } |
d20f7043 | 486 | |
08635bae | 487 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 488 | if (ret) { |
4e4cbee9 | 489 | bio->bi_status = ret; |
f5daf2c7 LB |
490 | bio_endio(bio); |
491 | } | |
c8b97818 | 492 | |
e749af44 | 493 | bio = btrfs_bio_alloc(first_byte); |
764c7c9a | 494 | bio->bi_opf = bio_op | write_flags; |
c8b97818 CM |
495 | bio->bi_private = cb; |
496 | bio->bi_end_io = end_compressed_bio_write; | |
46bcff2b | 497 | if (blkcg_css) |
7b62e66c | 498 | bio->bi_opf |= REQ_CGROUP_PUNT; |
764c7c9a JT |
499 | /* |
500 | * Use bio_add_page() to ensure the bio has at least one | |
501 | * page. | |
502 | */ | |
09cbfeaf | 503 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 504 | } |
09cbfeaf | 505 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 506 | btrfs_info(fs_info, |
282ab3ff | 507 | "bytes left %lu compress len %u nr %u", |
cfbc246e CM |
508 | bytes_left, cb->compressed_len, cb->nr_pages); |
509 | } | |
09cbfeaf KS |
510 | bytes_left -= PAGE_SIZE; |
511 | first_byte += PAGE_SIZE; | |
771ed689 | 512 | cond_resched(); |
c8b97818 | 513 | } |
c8b97818 | 514 | |
0b246afa | 515 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 516 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 517 | |
e55179b3 | 518 | if (!skip_sum) { |
c7ee1819 | 519 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 520 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 521 | } |
d20f7043 | 522 | |
08635bae | 523 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 524 | if (ret) { |
4e4cbee9 | 525 | bio->bi_status = ret; |
f5daf2c7 LB |
526 | bio_endio(bio); |
527 | } | |
c8b97818 | 528 | |
46bcff2b DZ |
529 | if (blkcg_css) |
530 | kthread_associate_blkcg(NULL); | |
531 | ||
c8b97818 CM |
532 | return 0; |
533 | } | |
534 | ||
2a4d0c90 CH |
535 | static u64 bio_end_offset(struct bio *bio) |
536 | { | |
c45a8f2d | 537 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
538 | |
539 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
540 | } | |
541 | ||
771ed689 CM |
542 | static noinline int add_ra_bio_pages(struct inode *inode, |
543 | u64 compressed_end, | |
544 | struct compressed_bio *cb) | |
545 | { | |
546 | unsigned long end_index; | |
306e16ce | 547 | unsigned long pg_index; |
771ed689 CM |
548 | u64 last_offset; |
549 | u64 isize = i_size_read(inode); | |
550 | int ret; | |
551 | struct page *page; | |
552 | unsigned long nr_pages = 0; | |
553 | struct extent_map *em; | |
554 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
555 | struct extent_map_tree *em_tree; |
556 | struct extent_io_tree *tree; | |
557 | u64 end; | |
558 | int misses = 0; | |
559 | ||
2a4d0c90 | 560 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
561 | em_tree = &BTRFS_I(inode)->extent_tree; |
562 | tree = &BTRFS_I(inode)->io_tree; | |
563 | ||
564 | if (isize == 0) | |
565 | return 0; | |
566 | ||
09cbfeaf | 567 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 568 | |
d397712b | 569 | while (last_offset < compressed_end) { |
09cbfeaf | 570 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 571 | |
306e16ce | 572 | if (pg_index > end_index) |
771ed689 CM |
573 | break; |
574 | ||
0a943c65 | 575 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 576 | if (page && !xa_is_value(page)) { |
771ed689 CM |
577 | misses++; |
578 | if (misses > 4) | |
579 | break; | |
580 | goto next; | |
581 | } | |
582 | ||
c62d2555 MH |
583 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
584 | ~__GFP_FS)); | |
771ed689 CM |
585 | if (!page) |
586 | break; | |
587 | ||
c62d2555 | 588 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 589 | put_page(page); |
771ed689 CM |
590 | goto next; |
591 | } | |
592 | ||
771ed689 CM |
593 | /* |
594 | * at this point, we have a locked page in the page cache | |
595 | * for these bytes in the file. But, we have to make | |
596 | * sure they map to this compressed extent on disk. | |
597 | */ | |
32443de3 QW |
598 | ret = set_page_extent_mapped(page); |
599 | if (ret < 0) { | |
600 | unlock_page(page); | |
601 | put_page(page); | |
602 | break; | |
603 | } | |
604 | ||
605 | end = last_offset + PAGE_SIZE - 1; | |
d0082371 | 606 | lock_extent(tree, last_offset, end); |
890871be | 607 | read_lock(&em_tree->lock); |
771ed689 | 608 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 609 | PAGE_SIZE); |
890871be | 610 | read_unlock(&em_tree->lock); |
771ed689 CM |
611 | |
612 | if (!em || last_offset < em->start || | |
09cbfeaf | 613 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 614 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 615 | free_extent_map(em); |
d0082371 | 616 | unlock_extent(tree, last_offset, end); |
771ed689 | 617 | unlock_page(page); |
09cbfeaf | 618 | put_page(page); |
771ed689 CM |
619 | break; |
620 | } | |
621 | free_extent_map(em); | |
622 | ||
623 | if (page->index == end_index) { | |
7073017a | 624 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
625 | |
626 | if (zero_offset) { | |
627 | int zeros; | |
09cbfeaf | 628 | zeros = PAGE_SIZE - zero_offset; |
d048b9c2 | 629 | memzero_page(page, zero_offset, zeros); |
771ed689 | 630 | flush_dcache_page(page); |
771ed689 CM |
631 | } |
632 | } | |
633 | ||
634 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 635 | PAGE_SIZE, 0); |
771ed689 | 636 | |
09cbfeaf | 637 | if (ret == PAGE_SIZE) { |
771ed689 | 638 | nr_pages++; |
09cbfeaf | 639 | put_page(page); |
771ed689 | 640 | } else { |
d0082371 | 641 | unlock_extent(tree, last_offset, end); |
771ed689 | 642 | unlock_page(page); |
09cbfeaf | 643 | put_page(page); |
771ed689 CM |
644 | break; |
645 | } | |
646 | next: | |
09cbfeaf | 647 | last_offset += PAGE_SIZE; |
771ed689 | 648 | } |
771ed689 CM |
649 | return 0; |
650 | } | |
651 | ||
c8b97818 CM |
652 | /* |
653 | * for a compressed read, the bio we get passed has all the inode pages | |
654 | * in it. We don't actually do IO on those pages but allocate new ones | |
655 | * to hold the compressed pages on disk. | |
656 | * | |
4f024f37 | 657 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 658 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
659 | * |
660 | * After the compressed pages are read, we copy the bytes into the | |
661 | * bio we were passed and then call the bio end_io calls | |
662 | */ | |
4e4cbee9 | 663 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
664 | int mirror_num, unsigned long bio_flags) |
665 | { | |
0b246afa | 666 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
667 | struct extent_map_tree *em_tree; |
668 | struct compressed_bio *cb; | |
356b4a2d AJ |
669 | unsigned int compressed_len; |
670 | unsigned int nr_pages; | |
671 | unsigned int pg_index; | |
c8b97818 | 672 | struct page *page; |
c8b97818 | 673 | struct bio *comp_bio; |
1201b58b | 674 | u64 cur_disk_byte = bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
675 | u64 em_len; |
676 | u64 em_start; | |
c8b97818 | 677 | struct extent_map *em; |
4e4cbee9 | 678 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 679 | int faili = 0; |
10fe6ca8 | 680 | u8 *sums; |
c8b97818 | 681 | |
c8b97818 CM |
682 | em_tree = &BTRFS_I(inode)->extent_tree; |
683 | ||
684 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 685 | read_lock(&em_tree->lock); |
c8b97818 | 686 | em = lookup_extent_mapping(em_tree, |
263663cd | 687 | page_offset(bio_first_page_all(bio)), |
be6a1361 | 688 | fs_info->sectorsize); |
890871be | 689 | read_unlock(&em_tree->lock); |
285190d9 | 690 | if (!em) |
4e4cbee9 | 691 | return BLK_STS_IOERR; |
c8b97818 | 692 | |
d20f7043 | 693 | compressed_len = em->block_len; |
2ff7e61e | 694 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 695 | if (!cb) |
696 | goto out; | |
697 | ||
a50299ae | 698 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
699 | cb->errors = 0; |
700 | cb->inode = inode; | |
d20f7043 | 701 | cb->mirror_num = mirror_num; |
10fe6ca8 | 702 | sums = cb->sums; |
c8b97818 | 703 | |
ff5b7ee3 | 704 | cb->start = em->orig_start; |
e04ca626 CM |
705 | em_len = em->len; |
706 | em_start = em->start; | |
d20f7043 | 707 | |
c8b97818 | 708 | free_extent_map(em); |
e04ca626 | 709 | em = NULL; |
c8b97818 | 710 | |
81381053 | 711 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 712 | cb->compressed_len = compressed_len; |
261507a0 | 713 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
714 | cb->orig_bio = bio; |
715 | ||
09cbfeaf | 716 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 717 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 718 | GFP_NOFS); |
6b82ce8d | 719 | if (!cb->compressed_pages) |
720 | goto fail1; | |
721 | ||
306e16ce | 722 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
b0ee5e1e | 723 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS); |
15e3004a JB |
724 | if (!cb->compressed_pages[pg_index]) { |
725 | faili = pg_index - 1; | |
0e9350de | 726 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 727 | goto fail2; |
15e3004a | 728 | } |
c8b97818 | 729 | } |
15e3004a | 730 | faili = nr_pages - 1; |
c8b97818 CM |
731 | cb->nr_pages = nr_pages; |
732 | ||
7f042a83 | 733 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 734 | |
771ed689 | 735 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 736 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 737 | |
e749af44 | 738 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 739 | comp_bio->bi_opf = REQ_OP_READ; |
c8b97818 CM |
740 | comp_bio->bi_private = cb; |
741 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 742 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 743 | |
306e16ce | 744 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
be6a1361 | 745 | u32 pg_len = PAGE_SIZE; |
4e4cbee9 CH |
746 | int submit = 0; |
747 | ||
be6a1361 QW |
748 | /* |
749 | * To handle subpage case, we need to make sure the bio only | |
750 | * covers the range we need. | |
751 | * | |
752 | * If we're at the last page, truncate the length to only cover | |
753 | * the remaining part. | |
754 | */ | |
755 | if (pg_index == nr_pages - 1) | |
756 | pg_len = min_t(u32, PAGE_SIZE, | |
757 | compressed_len - pg_index * PAGE_SIZE); | |
758 | ||
306e16ce | 759 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 760 | page->mapping = inode->i_mapping; |
09cbfeaf | 761 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 762 | |
4f024f37 | 763 | if (comp_bio->bi_iter.bi_size) |
be6a1361 | 764 | submit = btrfs_bio_fits_in_stripe(page, pg_len, |
da12fe54 | 765 | comp_bio, 0); |
c8b97818 | 766 | |
70b99e69 | 767 | page->mapping = NULL; |
be6a1361 | 768 | if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) { |
10fe6ca8 JT |
769 | unsigned int nr_sectors; |
770 | ||
0b246afa JM |
771 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
772 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 773 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 774 | |
af09abfe CM |
775 | /* |
776 | * inc the count before we submit the bio so | |
777 | * we know the end IO handler won't happen before | |
778 | * we inc the count. Otherwise, the cb might get | |
779 | * freed before we're done setting it up | |
780 | */ | |
a50299ae | 781 | refcount_inc(&cb->pending_bios); |
af09abfe | 782 | |
6275193e | 783 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
334c16d8 | 784 | BUG_ON(ret); /* -ENOMEM */ |
10fe6ca8 JT |
785 | |
786 | nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, | |
787 | fs_info->sectorsize); | |
713cebfb | 788 | sums += fs_info->csum_size * nr_sectors; |
d20f7043 | 789 | |
08635bae | 790 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 791 | if (ret) { |
4e4cbee9 | 792 | comp_bio->bi_status = ret; |
4246a0b6 CH |
793 | bio_endio(comp_bio); |
794 | } | |
c8b97818 | 795 | |
e749af44 | 796 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 797 | comp_bio->bi_opf = REQ_OP_READ; |
771ed689 CM |
798 | comp_bio->bi_private = cb; |
799 | comp_bio->bi_end_io = end_compressed_bio_read; | |
800 | ||
be6a1361 | 801 | bio_add_page(comp_bio, page, pg_len, 0); |
c8b97818 | 802 | } |
be6a1361 | 803 | cur_disk_byte += pg_len; |
c8b97818 | 804 | } |
c8b97818 | 805 | |
0b246afa | 806 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 807 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 808 | |
6275193e | 809 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
334c16d8 | 810 | BUG_ON(ret); /* -ENOMEM */ |
d20f7043 | 811 | |
08635bae | 812 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 813 | if (ret) { |
4e4cbee9 | 814 | comp_bio->bi_status = ret; |
4246a0b6 CH |
815 | bio_endio(comp_bio); |
816 | } | |
c8b97818 | 817 | |
c8b97818 | 818 | return 0; |
6b82ce8d | 819 | |
820 | fail2: | |
15e3004a JB |
821 | while (faili >= 0) { |
822 | __free_page(cb->compressed_pages[faili]); | |
823 | faili--; | |
824 | } | |
6b82ce8d | 825 | |
826 | kfree(cb->compressed_pages); | |
827 | fail1: | |
828 | kfree(cb); | |
829 | out: | |
830 | free_extent_map(em); | |
831 | return ret; | |
c8b97818 | 832 | } |
261507a0 | 833 | |
17b5a6c1 TT |
834 | /* |
835 | * Heuristic uses systematic sampling to collect data from the input data | |
836 | * range, the logic can be tuned by the following constants: | |
837 | * | |
838 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
839 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
840 | */ | |
841 | #define SAMPLING_READ_SIZE (16) | |
842 | #define SAMPLING_INTERVAL (256) | |
843 | ||
844 | /* | |
845 | * For statistical analysis of the input data we consider bytes that form a | |
846 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
847 | * many times the object appeared in the sample. | |
848 | */ | |
849 | #define BUCKET_SIZE (256) | |
850 | ||
851 | /* | |
852 | * The size of the sample is based on a statistical sampling rule of thumb. | |
853 | * The common way is to perform sampling tests as long as the number of | |
854 | * elements in each cell is at least 5. | |
855 | * | |
856 | * Instead of 5, we choose 32 to obtain more accurate results. | |
857 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
858 | * sample size bound by 8192. | |
859 | * | |
860 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
861 | * from up to 512 locations. | |
862 | */ | |
863 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
864 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
865 | ||
866 | struct bucket_item { | |
867 | u32 count; | |
868 | }; | |
4e439a0b TT |
869 | |
870 | struct heuristic_ws { | |
17b5a6c1 TT |
871 | /* Partial copy of input data */ |
872 | u8 *sample; | |
a440d48c | 873 | u32 sample_size; |
17b5a6c1 TT |
874 | /* Buckets store counters for each byte value */ |
875 | struct bucket_item *bucket; | |
440c840c TT |
876 | /* Sorting buffer */ |
877 | struct bucket_item *bucket_b; | |
4e439a0b TT |
878 | struct list_head list; |
879 | }; | |
880 | ||
92ee5530 DZ |
881 | static struct workspace_manager heuristic_wsm; |
882 | ||
4e439a0b TT |
883 | static void free_heuristic_ws(struct list_head *ws) |
884 | { | |
885 | struct heuristic_ws *workspace; | |
886 | ||
887 | workspace = list_entry(ws, struct heuristic_ws, list); | |
888 | ||
17b5a6c1 TT |
889 | kvfree(workspace->sample); |
890 | kfree(workspace->bucket); | |
440c840c | 891 | kfree(workspace->bucket_b); |
4e439a0b TT |
892 | kfree(workspace); |
893 | } | |
894 | ||
7bf49943 | 895 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
896 | { |
897 | struct heuristic_ws *ws; | |
898 | ||
899 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
900 | if (!ws) | |
901 | return ERR_PTR(-ENOMEM); | |
902 | ||
17b5a6c1 TT |
903 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
904 | if (!ws->sample) | |
905 | goto fail; | |
906 | ||
907 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
908 | if (!ws->bucket) | |
909 | goto fail; | |
4e439a0b | 910 | |
440c840c TT |
911 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
912 | if (!ws->bucket_b) | |
913 | goto fail; | |
914 | ||
17b5a6c1 | 915 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 916 | return &ws->list; |
17b5a6c1 TT |
917 | fail: |
918 | free_heuristic_ws(&ws->list); | |
919 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
920 | } |
921 | ||
ca4ac360 | 922 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 923 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
924 | }; |
925 | ||
e8c9f186 | 926 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
927 | /* The heuristic is represented as compression type 0 */ |
928 | &btrfs_heuristic_compress, | |
261507a0 | 929 | &btrfs_zlib_compress, |
a6fa6fae | 930 | &btrfs_lzo_compress, |
5c1aab1d | 931 | &btrfs_zstd_compress, |
261507a0 LZ |
932 | }; |
933 | ||
c778df14 DS |
934 | static struct list_head *alloc_workspace(int type, unsigned int level) |
935 | { | |
936 | switch (type) { | |
937 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
938 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
939 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
940 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
941 | default: | |
942 | /* | |
943 | * This can't happen, the type is validated several times | |
944 | * before we get here. | |
945 | */ | |
946 | BUG(); | |
947 | } | |
948 | } | |
949 | ||
1e002351 DS |
950 | static void free_workspace(int type, struct list_head *ws) |
951 | { | |
952 | switch (type) { | |
953 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
954 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
955 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
956 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
957 | default: | |
958 | /* | |
959 | * This can't happen, the type is validated several times | |
960 | * before we get here. | |
961 | */ | |
962 | BUG(); | |
963 | } | |
964 | } | |
965 | ||
d5517033 | 966 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 967 | { |
0cf25213 | 968 | struct workspace_manager *wsm; |
4e439a0b | 969 | struct list_head *workspace; |
261507a0 | 970 | |
0cf25213 | 971 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
972 | INIT_LIST_HEAD(&wsm->idle_ws); |
973 | spin_lock_init(&wsm->ws_lock); | |
974 | atomic_set(&wsm->total_ws, 0); | |
975 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 976 | |
1666edab DZ |
977 | /* |
978 | * Preallocate one workspace for each compression type so we can | |
979 | * guarantee forward progress in the worst case | |
980 | */ | |
c778df14 | 981 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
982 | if (IS_ERR(workspace)) { |
983 | pr_warn( | |
984 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
985 | } else { | |
92ee5530 DZ |
986 | atomic_set(&wsm->total_ws, 1); |
987 | wsm->free_ws = 1; | |
988 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
989 | } |
990 | } | |
991 | ||
2510307e | 992 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 993 | { |
2dba7143 | 994 | struct workspace_manager *wsman; |
1666edab DZ |
995 | struct list_head *ws; |
996 | ||
2dba7143 | 997 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
998 | while (!list_empty(&wsman->idle_ws)) { |
999 | ws = wsman->idle_ws.next; | |
1000 | list_del(ws); | |
1e002351 | 1001 | free_workspace(type, ws); |
1666edab | 1002 | atomic_dec(&wsman->total_ws); |
261507a0 | 1003 | } |
261507a0 LZ |
1004 | } |
1005 | ||
1006 | /* | |
e721e49d DS |
1007 | * This finds an available workspace or allocates a new one. |
1008 | * If it's not possible to allocate a new one, waits until there's one. | |
1009 | * Preallocation makes a forward progress guarantees and we do not return | |
1010 | * errors. | |
261507a0 | 1011 | */ |
5907a9bb | 1012 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 1013 | { |
5907a9bb | 1014 | struct workspace_manager *wsm; |
261507a0 LZ |
1015 | struct list_head *workspace; |
1016 | int cpus = num_online_cpus(); | |
fe308533 | 1017 | unsigned nofs_flag; |
4e439a0b TT |
1018 | struct list_head *idle_ws; |
1019 | spinlock_t *ws_lock; | |
1020 | atomic_t *total_ws; | |
1021 | wait_queue_head_t *ws_wait; | |
1022 | int *free_ws; | |
1023 | ||
5907a9bb | 1024 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1025 | idle_ws = &wsm->idle_ws; |
1026 | ws_lock = &wsm->ws_lock; | |
1027 | total_ws = &wsm->total_ws; | |
1028 | ws_wait = &wsm->ws_wait; | |
1029 | free_ws = &wsm->free_ws; | |
261507a0 | 1030 | |
261507a0 | 1031 | again: |
d9187649 BL |
1032 | spin_lock(ws_lock); |
1033 | if (!list_empty(idle_ws)) { | |
1034 | workspace = idle_ws->next; | |
261507a0 | 1035 | list_del(workspace); |
6ac10a6a | 1036 | (*free_ws)--; |
d9187649 | 1037 | spin_unlock(ws_lock); |
261507a0 LZ |
1038 | return workspace; |
1039 | ||
1040 | } | |
6ac10a6a | 1041 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
1042 | DEFINE_WAIT(wait); |
1043 | ||
d9187649 BL |
1044 | spin_unlock(ws_lock); |
1045 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 1046 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 1047 | schedule(); |
d9187649 | 1048 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
1049 | goto again; |
1050 | } | |
6ac10a6a | 1051 | atomic_inc(total_ws); |
d9187649 | 1052 | spin_unlock(ws_lock); |
261507a0 | 1053 | |
fe308533 DS |
1054 | /* |
1055 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
1056 | * to turn it off here because we might get called from the restricted | |
1057 | * context of btrfs_compress_bio/btrfs_compress_pages | |
1058 | */ | |
1059 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1060 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1061 | memalloc_nofs_restore(nofs_flag); |
1062 | ||
261507a0 | 1063 | if (IS_ERR(workspace)) { |
6ac10a6a | 1064 | atomic_dec(total_ws); |
d9187649 | 1065 | wake_up(ws_wait); |
e721e49d DS |
1066 | |
1067 | /* | |
1068 | * Do not return the error but go back to waiting. There's a | |
1069 | * workspace preallocated for each type and the compression | |
1070 | * time is bounded so we get to a workspace eventually. This | |
1071 | * makes our caller's life easier. | |
52356716 DS |
1072 | * |
1073 | * To prevent silent and low-probability deadlocks (when the | |
1074 | * initial preallocation fails), check if there are any | |
1075 | * workspaces at all. | |
e721e49d | 1076 | */ |
52356716 DS |
1077 | if (atomic_read(total_ws) == 0) { |
1078 | static DEFINE_RATELIMIT_STATE(_rs, | |
1079 | /* once per minute */ 60 * HZ, | |
1080 | /* no burst */ 1); | |
1081 | ||
1082 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1083 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1084 | } |
1085 | } | |
e721e49d | 1086 | goto again; |
261507a0 LZ |
1087 | } |
1088 | return workspace; | |
1089 | } | |
1090 | ||
7bf49943 | 1091 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1092 | { |
6a0d1272 | 1093 | switch (type) { |
5907a9bb | 1094 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1095 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1096 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1097 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1098 | default: | |
1099 | /* | |
1100 | * This can't happen, the type is validated several times | |
1101 | * before we get here. | |
1102 | */ | |
1103 | BUG(); | |
1104 | } | |
929f4baf DZ |
1105 | } |
1106 | ||
261507a0 LZ |
1107 | /* |
1108 | * put a workspace struct back on the list or free it if we have enough | |
1109 | * idle ones sitting around | |
1110 | */ | |
a3bbd2a9 | 1111 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1112 | { |
a3bbd2a9 | 1113 | struct workspace_manager *wsm; |
4e439a0b TT |
1114 | struct list_head *idle_ws; |
1115 | spinlock_t *ws_lock; | |
1116 | atomic_t *total_ws; | |
1117 | wait_queue_head_t *ws_wait; | |
1118 | int *free_ws; | |
1119 | ||
a3bbd2a9 | 1120 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1121 | idle_ws = &wsm->idle_ws; |
1122 | ws_lock = &wsm->ws_lock; | |
1123 | total_ws = &wsm->total_ws; | |
1124 | ws_wait = &wsm->ws_wait; | |
1125 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1126 | |
1127 | spin_lock(ws_lock); | |
26b28dce | 1128 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1129 | list_add(ws, idle_ws); |
6ac10a6a | 1130 | (*free_ws)++; |
d9187649 | 1131 | spin_unlock(ws_lock); |
261507a0 LZ |
1132 | goto wake; |
1133 | } | |
d9187649 | 1134 | spin_unlock(ws_lock); |
261507a0 | 1135 | |
1e002351 | 1136 | free_workspace(type, ws); |
6ac10a6a | 1137 | atomic_dec(total_ws); |
261507a0 | 1138 | wake: |
093258e6 | 1139 | cond_wake_up(ws_wait); |
261507a0 LZ |
1140 | } |
1141 | ||
929f4baf DZ |
1142 | static void put_workspace(int type, struct list_head *ws) |
1143 | { | |
bd3a5287 | 1144 | switch (type) { |
a3bbd2a9 DS |
1145 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1146 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1147 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1148 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1149 | default: | |
1150 | /* | |
1151 | * This can't happen, the type is validated several times | |
1152 | * before we get here. | |
1153 | */ | |
1154 | BUG(); | |
1155 | } | |
929f4baf DZ |
1156 | } |
1157 | ||
adbab642 AJ |
1158 | /* |
1159 | * Adjust @level according to the limits of the compression algorithm or | |
1160 | * fallback to default | |
1161 | */ | |
1162 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1163 | { | |
1164 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1165 | ||
1166 | if (level == 0) | |
1167 | level = ops->default_level; | |
1168 | else | |
1169 | level = min(level, ops->max_level); | |
1170 | ||
1171 | return level; | |
1172 | } | |
1173 | ||
261507a0 | 1174 | /* |
38c31464 DS |
1175 | * Given an address space and start and length, compress the bytes into @pages |
1176 | * that are allocated on demand. | |
261507a0 | 1177 | * |
f51d2b59 DS |
1178 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1179 | * default the algorithm chooses and is opaque here; | |
1180 | * - compression algo are 0-3 | |
1181 | * - the level are bits 4-7 | |
1182 | * | |
4d3a800e DS |
1183 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1184 | * and returns number of actually allocated pages | |
261507a0 | 1185 | * |
38c31464 DS |
1186 | * @total_in is used to return the number of bytes actually read. It |
1187 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1188 | * ran out of room in the pages array or because we cross the |
1189 | * max_out threshold. | |
1190 | * | |
38c31464 DS |
1191 | * @total_out is an in/out parameter, must be set to the input length and will |
1192 | * be also used to return the total number of compressed bytes | |
261507a0 | 1193 | */ |
f51d2b59 | 1194 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1195 | u64 start, struct page **pages, |
261507a0 LZ |
1196 | unsigned long *out_pages, |
1197 | unsigned long *total_in, | |
e5d74902 | 1198 | unsigned long *total_out) |
261507a0 | 1199 | { |
1972708a | 1200 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1201 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1202 | struct list_head *workspace; |
1203 | int ret; | |
1204 | ||
b0c1fe1e | 1205 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1206 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1207 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1208 | out_pages, total_in, total_out); | |
929f4baf | 1209 | put_workspace(type, workspace); |
261507a0 LZ |
1210 | return ret; |
1211 | } | |
1212 | ||
8140dc30 | 1213 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1214 | { |
1215 | struct list_head *workspace; | |
1216 | int ret; | |
8140dc30 | 1217 | int type = cb->compress_type; |
261507a0 | 1218 | |
7bf49943 | 1219 | workspace = get_workspace(type, 0); |
1e4eb746 | 1220 | ret = compression_decompress_bio(type, workspace, cb); |
929f4baf | 1221 | put_workspace(type, workspace); |
e1ddce71 | 1222 | |
261507a0 LZ |
1223 | return ret; |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * a less complex decompression routine. Our compressed data fits in a | |
1228 | * single page, and we want to read a single page out of it. | |
1229 | * start_byte tells us the offset into the compressed data we're interested in | |
1230 | */ | |
1231 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1232 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1233 | { | |
1234 | struct list_head *workspace; | |
1235 | int ret; | |
1236 | ||
7bf49943 | 1237 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1238 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1239 | start_byte, srclen, destlen); | |
929f4baf | 1240 | put_workspace(type, workspace); |
7bf49943 | 1241 | |
261507a0 LZ |
1242 | return ret; |
1243 | } | |
1244 | ||
1666edab DZ |
1245 | void __init btrfs_init_compress(void) |
1246 | { | |
d5517033 DS |
1247 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1248 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1249 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1250 | zstd_init_workspace_manager(); | |
1666edab DZ |
1251 | } |
1252 | ||
e67c718b | 1253 | void __cold btrfs_exit_compress(void) |
261507a0 | 1254 | { |
2510307e DS |
1255 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1256 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1257 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1258 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1259 | } |
3a39c18d LZ |
1260 | |
1261 | /* | |
1262 | * Copy uncompressed data from working buffer to pages. | |
1263 | * | |
1264 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1265 | * | |
1266 | * total_out is the last byte of the buffer | |
1267 | */ | |
14a3357b | 1268 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1269 | unsigned long total_out, u64 disk_start, |
974b1adc | 1270 | struct bio *bio) |
3a39c18d LZ |
1271 | { |
1272 | unsigned long buf_offset; | |
1273 | unsigned long current_buf_start; | |
1274 | unsigned long start_byte; | |
6e78b3f7 | 1275 | unsigned long prev_start_byte; |
3a39c18d LZ |
1276 | unsigned long working_bytes = total_out - buf_start; |
1277 | unsigned long bytes; | |
974b1adc | 1278 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1279 | |
1280 | /* | |
1281 | * start byte is the first byte of the page we're currently | |
1282 | * copying into relative to the start of the compressed data. | |
1283 | */ | |
974b1adc | 1284 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1285 | |
1286 | /* we haven't yet hit data corresponding to this page */ | |
1287 | if (total_out <= start_byte) | |
1288 | return 1; | |
1289 | ||
1290 | /* | |
1291 | * the start of the data we care about is offset into | |
1292 | * the middle of our working buffer | |
1293 | */ | |
1294 | if (total_out > start_byte && buf_start < start_byte) { | |
1295 | buf_offset = start_byte - buf_start; | |
1296 | working_bytes -= buf_offset; | |
1297 | } else { | |
1298 | buf_offset = 0; | |
1299 | } | |
1300 | current_buf_start = buf_start; | |
1301 | ||
1302 | /* copy bytes from the working buffer into the pages */ | |
1303 | while (working_bytes > 0) { | |
974b1adc | 1304 | bytes = min_t(unsigned long, bvec.bv_len, |
3fd396af | 1305 | PAGE_SIZE - (buf_offset % PAGE_SIZE)); |
3a39c18d | 1306 | bytes = min(bytes, working_bytes); |
974b1adc | 1307 | |
3590ec58 IW |
1308 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, buf + buf_offset, |
1309 | bytes); | |
974b1adc | 1310 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1311 | |
3a39c18d LZ |
1312 | buf_offset += bytes; |
1313 | working_bytes -= bytes; | |
1314 | current_buf_start += bytes; | |
1315 | ||
1316 | /* check if we need to pick another page */ | |
974b1adc CH |
1317 | bio_advance(bio, bytes); |
1318 | if (!bio->bi_iter.bi_size) | |
1319 | return 0; | |
1320 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1321 | prev_start_byte = start_byte; |
974b1adc | 1322 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1323 | |
974b1adc | 1324 | /* |
6e78b3f7 OS |
1325 | * We need to make sure we're only adjusting |
1326 | * our offset into compression working buffer when | |
1327 | * we're switching pages. Otherwise we can incorrectly | |
1328 | * keep copying when we were actually done. | |
974b1adc | 1329 | */ |
6e78b3f7 OS |
1330 | if (start_byte != prev_start_byte) { |
1331 | /* | |
1332 | * make sure our new page is covered by this | |
1333 | * working buffer | |
1334 | */ | |
1335 | if (total_out <= start_byte) | |
1336 | return 1; | |
3a39c18d | 1337 | |
6e78b3f7 OS |
1338 | /* |
1339 | * the next page in the biovec might not be adjacent | |
1340 | * to the last page, but it might still be found | |
1341 | * inside this working buffer. bump our offset pointer | |
1342 | */ | |
1343 | if (total_out > start_byte && | |
1344 | current_buf_start < start_byte) { | |
1345 | buf_offset = start_byte - buf_start; | |
1346 | working_bytes = total_out - start_byte; | |
1347 | current_buf_start = buf_start + buf_offset; | |
1348 | } | |
3a39c18d LZ |
1349 | } |
1350 | } | |
1351 | ||
1352 | return 1; | |
1353 | } | |
c2fcdcdf | 1354 | |
19562430 TT |
1355 | /* |
1356 | * Shannon Entropy calculation | |
1357 | * | |
52042d8e | 1358 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1359 | * Try calculating entropy to estimate the average minimum number of bits |
1360 | * needed to encode the sampled data. | |
1361 | * | |
1362 | * For convenience, return the percentage of needed bits, instead of amount of | |
1363 | * bits directly. | |
1364 | * | |
1365 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1366 | * and can be compressible with high probability | |
1367 | * | |
1368 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1369 | * | |
1370 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1371 | */ | |
1372 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1373 | #define ENTROPY_LVL_HIGH (80) | |
1374 | ||
1375 | /* | |
1376 | * For increasead precision in shannon_entropy calculation, | |
1377 | * let's do pow(n, M) to save more digits after comma: | |
1378 | * | |
1379 | * - maximum int bit length is 64 | |
1380 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1381 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1382 | * | |
1383 | * So use pow(n, 4). | |
1384 | */ | |
1385 | static inline u32 ilog2_w(u64 n) | |
1386 | { | |
1387 | return ilog2(n * n * n * n); | |
1388 | } | |
1389 | ||
1390 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1391 | { | |
1392 | const u32 entropy_max = 8 * ilog2_w(2); | |
1393 | u32 entropy_sum = 0; | |
1394 | u32 p, p_base, sz_base; | |
1395 | u32 i; | |
1396 | ||
1397 | sz_base = ilog2_w(ws->sample_size); | |
1398 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1399 | p = ws->bucket[i].count; | |
1400 | p_base = ilog2_w(p); | |
1401 | entropy_sum += p * (sz_base - p_base); | |
1402 | } | |
1403 | ||
1404 | entropy_sum /= ws->sample_size; | |
1405 | return entropy_sum * 100 / entropy_max; | |
1406 | } | |
1407 | ||
440c840c TT |
1408 | #define RADIX_BASE 4U |
1409 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1410 | ||
1411 | static u8 get4bits(u64 num, int shift) { | |
1412 | u8 low4bits; | |
1413 | ||
1414 | num >>= shift; | |
1415 | /* Reverse order */ | |
1416 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1417 | return low4bits; | |
1418 | } | |
1419 | ||
440c840c TT |
1420 | /* |
1421 | * Use 4 bits as radix base | |
52042d8e | 1422 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1423 | * |
1424 | * @array - array that will be sorted | |
1425 | * @array_buf - buffer array to store sorting results | |
1426 | * must be equal in size to @array | |
1427 | * @num - array size | |
440c840c | 1428 | */ |
23ae8c63 | 1429 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1430 | int num) |
858177d3 | 1431 | { |
440c840c TT |
1432 | u64 max_num; |
1433 | u64 buf_num; | |
1434 | u32 counters[COUNTERS_SIZE]; | |
1435 | u32 new_addr; | |
1436 | u32 addr; | |
1437 | int bitlen; | |
1438 | int shift; | |
1439 | int i; | |
858177d3 | 1440 | |
440c840c TT |
1441 | /* |
1442 | * Try avoid useless loop iterations for small numbers stored in big | |
1443 | * counters. Example: 48 33 4 ... in 64bit array | |
1444 | */ | |
23ae8c63 | 1445 | max_num = array[0].count; |
440c840c | 1446 | for (i = 1; i < num; i++) { |
23ae8c63 | 1447 | buf_num = array[i].count; |
440c840c TT |
1448 | if (buf_num > max_num) |
1449 | max_num = buf_num; | |
1450 | } | |
1451 | ||
1452 | buf_num = ilog2(max_num); | |
1453 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1454 | ||
1455 | shift = 0; | |
1456 | while (shift < bitlen) { | |
1457 | memset(counters, 0, sizeof(counters)); | |
1458 | ||
1459 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1460 | buf_num = array[i].count; |
440c840c TT |
1461 | addr = get4bits(buf_num, shift); |
1462 | counters[addr]++; | |
1463 | } | |
1464 | ||
1465 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1466 | counters[i] += counters[i - 1]; | |
1467 | ||
1468 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1469 | buf_num = array[i].count; |
440c840c TT |
1470 | addr = get4bits(buf_num, shift); |
1471 | counters[addr]--; | |
1472 | new_addr = counters[addr]; | |
7add17be | 1473 | array_buf[new_addr] = array[i]; |
440c840c TT |
1474 | } |
1475 | ||
1476 | shift += RADIX_BASE; | |
1477 | ||
1478 | /* | |
1479 | * Normal radix expects to move data from a temporary array, to | |
1480 | * the main one. But that requires some CPU time. Avoid that | |
1481 | * by doing another sort iteration to original array instead of | |
1482 | * memcpy() | |
1483 | */ | |
1484 | memset(counters, 0, sizeof(counters)); | |
1485 | ||
1486 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1487 | buf_num = array_buf[i].count; |
440c840c TT |
1488 | addr = get4bits(buf_num, shift); |
1489 | counters[addr]++; | |
1490 | } | |
1491 | ||
1492 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1493 | counters[i] += counters[i - 1]; | |
1494 | ||
1495 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1496 | buf_num = array_buf[i].count; |
440c840c TT |
1497 | addr = get4bits(buf_num, shift); |
1498 | counters[addr]--; | |
1499 | new_addr = counters[addr]; | |
7add17be | 1500 | array[new_addr] = array_buf[i]; |
440c840c TT |
1501 | } |
1502 | ||
1503 | shift += RADIX_BASE; | |
1504 | } | |
858177d3 TT |
1505 | } |
1506 | ||
1507 | /* | |
1508 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1509 | * | |
1510 | * There are several types of structured binary data that use nearly all byte | |
1511 | * values. The distribution can be uniform and counts in all buckets will be | |
1512 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1513 | * | |
1514 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1515 | * be potentially compressible, but we have to take a few more steps to decide | |
1516 | * how much. | |
1517 | * | |
1518 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1519 | * compression algo can easy fix that | |
1520 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1521 | * probability is not compressible | |
1522 | */ | |
1523 | #define BYTE_CORE_SET_LOW (64) | |
1524 | #define BYTE_CORE_SET_HIGH (200) | |
1525 | ||
1526 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1527 | { | |
1528 | u32 i; | |
1529 | u32 coreset_sum = 0; | |
1530 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1531 | struct bucket_item *bucket = ws->bucket; | |
1532 | ||
1533 | /* Sort in reverse order */ | |
36243c91 | 1534 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1535 | |
1536 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1537 | coreset_sum += bucket[i].count; | |
1538 | ||
1539 | if (coreset_sum > core_set_threshold) | |
1540 | return i; | |
1541 | ||
1542 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1543 | coreset_sum += bucket[i].count; | |
1544 | if (coreset_sum > core_set_threshold) | |
1545 | break; | |
1546 | } | |
1547 | ||
1548 | return i; | |
1549 | } | |
1550 | ||
a288e92c TT |
1551 | /* |
1552 | * Count byte values in buckets. | |
1553 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1554 | * Because in most text-like data byte set is restricted to limited number of | |
1555 | * possible characters, and that restriction in most cases makes data easy to | |
1556 | * compress. | |
1557 | * | |
1558 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1559 | * less - compressible | |
1560 | * more - need additional analysis | |
1561 | */ | |
1562 | #define BYTE_SET_THRESHOLD (64) | |
1563 | ||
1564 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1565 | { | |
1566 | u32 i; | |
1567 | u32 byte_set_size = 0; | |
1568 | ||
1569 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1570 | if (ws->bucket[i].count > 0) | |
1571 | byte_set_size++; | |
1572 | } | |
1573 | ||
1574 | /* | |
1575 | * Continue collecting count of byte values in buckets. If the byte | |
1576 | * set size is bigger then the threshold, it's pointless to continue, | |
1577 | * the detection technique would fail for this type of data. | |
1578 | */ | |
1579 | for (; i < BUCKET_SIZE; i++) { | |
1580 | if (ws->bucket[i].count > 0) { | |
1581 | byte_set_size++; | |
1582 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1583 | return byte_set_size; | |
1584 | } | |
1585 | } | |
1586 | ||
1587 | return byte_set_size; | |
1588 | } | |
1589 | ||
1fe4f6fa TT |
1590 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1591 | { | |
1592 | const u32 half_of_sample = ws->sample_size / 2; | |
1593 | const u8 *data = ws->sample; | |
1594 | ||
1595 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1596 | } | |
1597 | ||
a440d48c TT |
1598 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1599 | struct heuristic_ws *ws) | |
1600 | { | |
1601 | struct page *page; | |
1602 | u64 index, index_end; | |
1603 | u32 i, curr_sample_pos; | |
1604 | u8 *in_data; | |
1605 | ||
1606 | /* | |
1607 | * Compression handles the input data by chunks of 128KiB | |
1608 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1609 | * | |
1610 | * We do the same for the heuristic and loop over the whole range. | |
1611 | * | |
1612 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1613 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1614 | */ | |
1615 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1616 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1617 | ||
1618 | index = start >> PAGE_SHIFT; | |
1619 | index_end = end >> PAGE_SHIFT; | |
1620 | ||
1621 | /* Don't miss unaligned end */ | |
1622 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1623 | index_end++; | |
1624 | ||
1625 | curr_sample_pos = 0; | |
1626 | while (index < index_end) { | |
1627 | page = find_get_page(inode->i_mapping, index); | |
58c1a35c | 1628 | in_data = kmap_local_page(page); |
a440d48c TT |
1629 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1630 | i = start % PAGE_SIZE; | |
1631 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1632 | /* Don't sample any garbage from the last page */ | |
1633 | if (start > end - SAMPLING_READ_SIZE) | |
1634 | break; | |
1635 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1636 | SAMPLING_READ_SIZE); | |
1637 | i += SAMPLING_INTERVAL; | |
1638 | start += SAMPLING_INTERVAL; | |
1639 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1640 | } | |
58c1a35c | 1641 | kunmap_local(in_data); |
a440d48c TT |
1642 | put_page(page); |
1643 | ||
1644 | index++; | |
1645 | } | |
1646 | ||
1647 | ws->sample_size = curr_sample_pos; | |
1648 | } | |
1649 | ||
c2fcdcdf TT |
1650 | /* |
1651 | * Compression heuristic. | |
1652 | * | |
1653 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1654 | * quickly (compared to direct compression) detect data characteristics | |
1655 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1656 | * data. | |
1657 | * | |
1658 | * The following types of analysis can be performed: | |
1659 | * - detect mostly zero data | |
1660 | * - detect data with low "byte set" size (text, etc) | |
1661 | * - detect data with low/high "core byte" set | |
1662 | * | |
1663 | * Return non-zero if the compression should be done, 0 otherwise. | |
1664 | */ | |
1665 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1666 | { | |
7bf49943 | 1667 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1668 | struct heuristic_ws *ws; |
a440d48c TT |
1669 | u32 i; |
1670 | u8 byte; | |
19562430 | 1671 | int ret = 0; |
c2fcdcdf | 1672 | |
4e439a0b TT |
1673 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1674 | ||
a440d48c TT |
1675 | heuristic_collect_sample(inode, start, end, ws); |
1676 | ||
1fe4f6fa TT |
1677 | if (sample_repeated_patterns(ws)) { |
1678 | ret = 1; | |
1679 | goto out; | |
1680 | } | |
1681 | ||
a440d48c TT |
1682 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1683 | ||
1684 | for (i = 0; i < ws->sample_size; i++) { | |
1685 | byte = ws->sample[i]; | |
1686 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1687 | } |
1688 | ||
a288e92c TT |
1689 | i = byte_set_size(ws); |
1690 | if (i < BYTE_SET_THRESHOLD) { | |
1691 | ret = 2; | |
1692 | goto out; | |
1693 | } | |
1694 | ||
858177d3 TT |
1695 | i = byte_core_set_size(ws); |
1696 | if (i <= BYTE_CORE_SET_LOW) { | |
1697 | ret = 3; | |
1698 | goto out; | |
1699 | } | |
1700 | ||
1701 | if (i >= BYTE_CORE_SET_HIGH) { | |
1702 | ret = 0; | |
1703 | goto out; | |
1704 | } | |
1705 | ||
19562430 TT |
1706 | i = shannon_entropy(ws); |
1707 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1708 | ret = 4; | |
1709 | goto out; | |
1710 | } | |
1711 | ||
1712 | /* | |
1713 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1714 | * needed to give green light to compression. | |
1715 | * | |
1716 | * For now just assume that compression at that level is not worth the | |
1717 | * resources because: | |
1718 | * | |
1719 | * 1. it is possible to defrag the data later | |
1720 | * | |
1721 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1722 | * values, every bucket has counter at level ~54. The heuristic would | |
1723 | * be confused. This can happen when data have some internal repeated | |
1724 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1725 | * pairs of bytes, which is too costly. | |
1726 | */ | |
1727 | if (i < ENTROPY_LVL_HIGH) { | |
1728 | ret = 5; | |
1729 | goto out; | |
1730 | } else { | |
1731 | ret = 0; | |
1732 | goto out; | |
1733 | } | |
1734 | ||
1fe4f6fa | 1735 | out: |
929f4baf | 1736 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1737 | return ret; |
1738 | } | |
f51d2b59 | 1739 | |
d0ab62ce DZ |
1740 | /* |
1741 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1742 | * level, unrecognized string will set the default level | |
1743 | */ | |
1744 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1745 | { |
d0ab62ce DZ |
1746 | unsigned int level = 0; |
1747 | int ret; | |
1748 | ||
1749 | if (!type) | |
f51d2b59 DS |
1750 | return 0; |
1751 | ||
d0ab62ce DZ |
1752 | if (str[0] == ':') { |
1753 | ret = kstrtouint(str + 1, 10, &level); | |
1754 | if (ret) | |
1755 | level = 0; | |
1756 | } | |
1757 | ||
b0c1fe1e DS |
1758 | level = btrfs_compress_set_level(type, level); |
1759 | ||
1760 | return level; | |
1761 | } |