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 CM |
151 | struct page *page; |
152 | unsigned long i; | |
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) { | |
175 | kaddr = kmap_atomic(page); | |
176 | crypto_shash_digest(shash, kaddr + pg_offset, | |
177 | sectorsize, csum); | |
178 | kunmap_atomic(kaddr); | |
179 | ||
180 | if (memcmp(&csum, cb_sum, csum_size) != 0) { | |
181 | btrfs_print_data_csum_error(inode, disk_start, | |
182 | csum, cb_sum, cb->mirror_num); | |
183 | if (btrfs_io_bio(bio)->device) | |
184 | btrfs_dev_stat_inc_and_print( | |
185 | btrfs_io_bio(bio)->device, | |
186 | BTRFS_DEV_STAT_CORRUPTION_ERRS); | |
187 | return -EIO; | |
188 | } | |
189 | cb_sum += csum_size; | |
190 | disk_start += sectorsize; | |
d20f7043 | 191 | } |
d20f7043 | 192 | } |
93c4c033 | 193 | return 0; |
d20f7043 CM |
194 | } |
195 | ||
c8b97818 CM |
196 | /* when we finish reading compressed pages from the disk, we |
197 | * decompress them and then run the bio end_io routines on the | |
198 | * decompressed pages (in the inode address space). | |
199 | * | |
200 | * This allows the checksumming and other IO error handling routines | |
201 | * to work normally | |
202 | * | |
203 | * The compressed pages are freed here, and it must be run | |
204 | * in process context | |
205 | */ | |
4246a0b6 | 206 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 207 | { |
c8b97818 CM |
208 | struct compressed_bio *cb = bio->bi_private; |
209 | struct inode *inode; | |
210 | struct page *page; | |
211 | unsigned long index; | |
cf1167d5 | 212 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 213 | int ret = 0; |
c8b97818 | 214 | |
4e4cbee9 | 215 | if (bio->bi_status) |
c8b97818 CM |
216 | cb->errors = 1; |
217 | ||
218 | /* if there are more bios still pending for this compressed | |
219 | * extent, just exit | |
220 | */ | |
a50299ae | 221 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
222 | goto out; |
223 | ||
cf1167d5 LB |
224 | /* |
225 | * Record the correct mirror_num in cb->orig_bio so that | |
226 | * read-repair can work properly. | |
227 | */ | |
cf1167d5 LB |
228 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; |
229 | cb->mirror_num = mirror; | |
230 | ||
e6311f24 LB |
231 | /* |
232 | * Some IO in this cb have failed, just skip checksum as there | |
233 | * is no way it could be correct. | |
234 | */ | |
235 | if (cb->errors == 1) | |
236 | goto csum_failed; | |
237 | ||
d20f7043 | 238 | inode = cb->inode; |
5a9472fe | 239 | ret = check_compressed_csum(BTRFS_I(inode), bio, |
1201b58b | 240 | bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
241 | if (ret) |
242 | goto csum_failed; | |
243 | ||
c8b97818 CM |
244 | /* ok, we're the last bio for this extent, lets start |
245 | * the decompression. | |
246 | */ | |
8140dc30 AJ |
247 | ret = btrfs_decompress_bio(cb); |
248 | ||
d20f7043 | 249 | csum_failed: |
c8b97818 CM |
250 | if (ret) |
251 | cb->errors = 1; | |
252 | ||
253 | /* release the compressed pages */ | |
254 | index = 0; | |
255 | for (index = 0; index < cb->nr_pages; index++) { | |
256 | page = cb->compressed_pages[index]; | |
257 | page->mapping = NULL; | |
09cbfeaf | 258 | put_page(page); |
c8b97818 CM |
259 | } |
260 | ||
261 | /* do io completion on the original bio */ | |
771ed689 | 262 | if (cb->errors) { |
c8b97818 | 263 | bio_io_error(cb->orig_bio); |
d20f7043 | 264 | } else { |
2c30c71b | 265 | struct bio_vec *bvec; |
6dc4f100 | 266 | struct bvec_iter_all iter_all; |
d20f7043 CM |
267 | |
268 | /* | |
269 | * we have verified the checksum already, set page | |
270 | * checked so the end_io handlers know about it | |
271 | */ | |
c09abff8 | 272 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2b070cfe | 273 | bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) |
d20f7043 | 274 | SetPageChecked(bvec->bv_page); |
2c30c71b | 275 | |
4246a0b6 | 276 | bio_endio(cb->orig_bio); |
d20f7043 | 277 | } |
c8b97818 CM |
278 | |
279 | /* finally free the cb struct */ | |
280 | kfree(cb->compressed_pages); | |
281 | kfree(cb); | |
282 | out: | |
283 | bio_put(bio); | |
284 | } | |
285 | ||
286 | /* | |
287 | * Clear the writeback bits on all of the file | |
288 | * pages for a compressed write | |
289 | */ | |
7bdcefc1 FM |
290 | static noinline void end_compressed_writeback(struct inode *inode, |
291 | const struct compressed_bio *cb) | |
c8b97818 | 292 | { |
09cbfeaf KS |
293 | unsigned long index = cb->start >> PAGE_SHIFT; |
294 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
295 | struct page *pages[16]; |
296 | unsigned long nr_pages = end_index - index + 1; | |
297 | int i; | |
298 | int ret; | |
299 | ||
7bdcefc1 FM |
300 | if (cb->errors) |
301 | mapping_set_error(inode->i_mapping, -EIO); | |
302 | ||
d397712b | 303 | while (nr_pages > 0) { |
c8b97818 | 304 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
305 | min_t(unsigned long, |
306 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
307 | if (ret == 0) { |
308 | nr_pages -= 1; | |
309 | index += 1; | |
310 | continue; | |
311 | } | |
312 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
313 | if (cb->errors) |
314 | SetPageError(pages[i]); | |
c8b97818 | 315 | end_page_writeback(pages[i]); |
09cbfeaf | 316 | put_page(pages[i]); |
c8b97818 CM |
317 | } |
318 | nr_pages -= ret; | |
319 | index += ret; | |
320 | } | |
321 | /* the inode may be gone now */ | |
c8b97818 CM |
322 | } |
323 | ||
324 | /* | |
325 | * do the cleanup once all the compressed pages hit the disk. | |
326 | * This will clear writeback on the file pages and free the compressed | |
327 | * pages. | |
328 | * | |
329 | * This also calls the writeback end hooks for the file pages so that | |
330 | * metadata and checksums can be updated in the file. | |
331 | */ | |
4246a0b6 | 332 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 | 333 | { |
c8b97818 CM |
334 | struct compressed_bio *cb = bio->bi_private; |
335 | struct inode *inode; | |
336 | struct page *page; | |
337 | unsigned long index; | |
338 | ||
4e4cbee9 | 339 | if (bio->bi_status) |
c8b97818 CM |
340 | cb->errors = 1; |
341 | ||
342 | /* if there are more bios still pending for this compressed | |
343 | * extent, just exit | |
344 | */ | |
a50299ae | 345 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
346 | goto out; |
347 | ||
348 | /* ok, we're the last bio for this extent, step one is to | |
349 | * call back into the FS and do all the end_io operations | |
350 | */ | |
351 | inode = cb->inode; | |
70b99e69 | 352 | cb->compressed_pages[0]->mapping = cb->inode->i_mapping; |
764c7c9a | 353 | btrfs_record_physical_zoned(inode, cb->start, bio); |
7087a9d8 | 354 | btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0], |
c629732d | 355 | cb->start, cb->start + cb->len - 1, |
6a8d2136 | 356 | bio->bi_status == BLK_STS_OK); |
70b99e69 | 357 | cb->compressed_pages[0]->mapping = NULL; |
c8b97818 | 358 | |
7bdcefc1 | 359 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
360 | /* note, our inode could be gone now */ |
361 | ||
362 | /* | |
363 | * release the compressed pages, these came from alloc_page and | |
364 | * are not attached to the inode at all | |
365 | */ | |
366 | index = 0; | |
367 | for (index = 0; index < cb->nr_pages; index++) { | |
368 | page = cb->compressed_pages[index]; | |
369 | page->mapping = NULL; | |
09cbfeaf | 370 | put_page(page); |
c8b97818 CM |
371 | } |
372 | ||
373 | /* finally free the cb struct */ | |
374 | kfree(cb->compressed_pages); | |
375 | kfree(cb); | |
376 | out: | |
377 | bio_put(bio); | |
378 | } | |
379 | ||
380 | /* | |
381 | * worker function to build and submit bios for previously compressed pages. | |
382 | * The corresponding pages in the inode should be marked for writeback | |
383 | * and the compressed pages should have a reference on them for dropping | |
384 | * when the IO is complete. | |
385 | * | |
386 | * This also checksums the file bytes and gets things ready for | |
387 | * the end io hooks. | |
388 | */ | |
c7ee1819 | 389 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
c8b97818 CM |
390 | unsigned long len, u64 disk_start, |
391 | unsigned long compressed_len, | |
392 | struct page **compressed_pages, | |
f82b7359 | 393 | unsigned long nr_pages, |
ec39f769 CM |
394 | unsigned int write_flags, |
395 | struct cgroup_subsys_state *blkcg_css) | |
c8b97818 | 396 | { |
c7ee1819 | 397 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 398 | struct bio *bio = NULL; |
c8b97818 CM |
399 | struct compressed_bio *cb; |
400 | unsigned long bytes_left; | |
306e16ce | 401 | int pg_index = 0; |
c8b97818 CM |
402 | struct page *page; |
403 | u64 first_byte = disk_start; | |
4e4cbee9 | 404 | blk_status_t ret; |
c7ee1819 | 405 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
764c7c9a JT |
406 | const bool use_append = btrfs_use_zone_append(inode, disk_start); |
407 | const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; | |
c8b97818 | 408 | |
fdb1e121 | 409 | WARN_ON(!PAGE_ALIGNED(start)); |
2ff7e61e | 410 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 411 | if (!cb) |
4e4cbee9 | 412 | return BLK_STS_RESOURCE; |
a50299ae | 413 | refcount_set(&cb->pending_bios, 0); |
c8b97818 | 414 | cb->errors = 0; |
c7ee1819 | 415 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
416 | cb->start = start; |
417 | cb->len = len; | |
d20f7043 | 418 | cb->mirror_num = 0; |
c8b97818 CM |
419 | cb->compressed_pages = compressed_pages; |
420 | cb->compressed_len = compressed_len; | |
421 | cb->orig_bio = NULL; | |
422 | cb->nr_pages = nr_pages; | |
423 | ||
e749af44 | 424 | bio = btrfs_bio_alloc(first_byte); |
764c7c9a | 425 | bio->bi_opf = bio_op | write_flags; |
c8b97818 CM |
426 | bio->bi_private = cb; |
427 | bio->bi_end_io = end_compressed_bio_write; | |
ec39f769 | 428 | |
764c7c9a JT |
429 | if (use_append) { |
430 | struct extent_map *em; | |
431 | struct map_lookup *map; | |
432 | struct block_device *bdev; | |
433 | ||
434 | em = btrfs_get_chunk_map(fs_info, disk_start, PAGE_SIZE); | |
435 | if (IS_ERR(em)) { | |
436 | kfree(cb); | |
437 | bio_put(bio); | |
438 | return BLK_STS_NOTSUPP; | |
439 | } | |
440 | ||
441 | map = em->map_lookup; | |
442 | /* We only support single profile for now */ | |
443 | ASSERT(map->num_stripes == 1); | |
444 | bdev = map->stripes[0].dev->bdev; | |
445 | ||
446 | bio_set_dev(bio, bdev); | |
447 | free_extent_map(em); | |
448 | } | |
449 | ||
ec39f769 CM |
450 | if (blkcg_css) { |
451 | bio->bi_opf |= REQ_CGROUP_PUNT; | |
46bcff2b | 452 | kthread_associate_blkcg(blkcg_css); |
ec39f769 | 453 | } |
a50299ae | 454 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
455 | |
456 | /* create and submit bios for the compressed pages */ | |
457 | bytes_left = compressed_len; | |
306e16ce | 458 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 | 459 | int submit = 0; |
4c80a97d | 460 | int len = 0; |
4e4cbee9 | 461 | |
306e16ce | 462 | page = compressed_pages[pg_index]; |
c7ee1819 | 463 | page->mapping = inode->vfs_inode.i_mapping; |
4f024f37 | 464 | if (bio->bi_iter.bi_size) |
da12fe54 NB |
465 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, |
466 | 0); | |
c8b97818 | 467 | |
4c80a97d QW |
468 | /* |
469 | * Page can only be added to bio if the current bio fits in | |
470 | * stripe. | |
471 | */ | |
472 | if (!submit) { | |
473 | if (pg_index == 0 && use_append) | |
474 | len = bio_add_zone_append_page(bio, page, | |
475 | PAGE_SIZE, 0); | |
476 | else | |
477 | len = bio_add_page(bio, page, PAGE_SIZE, 0); | |
478 | } | |
764c7c9a | 479 | |
70b99e69 | 480 | page->mapping = NULL; |
764c7c9a | 481 | if (submit || len < PAGE_SIZE) { |
af09abfe CM |
482 | /* |
483 | * inc the count before we submit the bio so | |
484 | * we know the end IO handler won't happen before | |
485 | * we inc the count. Otherwise, the cb might get | |
486 | * freed before we're done setting it up | |
487 | */ | |
a50299ae | 488 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
489 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
490 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 491 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 492 | |
e55179b3 | 493 | if (!skip_sum) { |
c7ee1819 | 494 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 495 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 496 | } |
d20f7043 | 497 | |
08635bae | 498 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 499 | if (ret) { |
4e4cbee9 | 500 | bio->bi_status = ret; |
f5daf2c7 LB |
501 | bio_endio(bio); |
502 | } | |
c8b97818 | 503 | |
e749af44 | 504 | bio = btrfs_bio_alloc(first_byte); |
764c7c9a | 505 | bio->bi_opf = bio_op | write_flags; |
c8b97818 CM |
506 | bio->bi_private = cb; |
507 | bio->bi_end_io = end_compressed_bio_write; | |
46bcff2b | 508 | if (blkcg_css) |
7b62e66c | 509 | bio->bi_opf |= REQ_CGROUP_PUNT; |
764c7c9a JT |
510 | /* |
511 | * Use bio_add_page() to ensure the bio has at least one | |
512 | * page. | |
513 | */ | |
09cbfeaf | 514 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 515 | } |
09cbfeaf | 516 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 517 | btrfs_info(fs_info, |
efe120a0 | 518 | "bytes left %lu compress len %lu nr %lu", |
cfbc246e CM |
519 | bytes_left, cb->compressed_len, cb->nr_pages); |
520 | } | |
09cbfeaf KS |
521 | bytes_left -= PAGE_SIZE; |
522 | first_byte += PAGE_SIZE; | |
771ed689 | 523 | cond_resched(); |
c8b97818 | 524 | } |
c8b97818 | 525 | |
0b246afa | 526 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 527 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 528 | |
e55179b3 | 529 | if (!skip_sum) { |
c7ee1819 | 530 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 531 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 532 | } |
d20f7043 | 533 | |
08635bae | 534 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 535 | if (ret) { |
4e4cbee9 | 536 | bio->bi_status = ret; |
f5daf2c7 LB |
537 | bio_endio(bio); |
538 | } | |
c8b97818 | 539 | |
46bcff2b DZ |
540 | if (blkcg_css) |
541 | kthread_associate_blkcg(NULL); | |
542 | ||
c8b97818 CM |
543 | return 0; |
544 | } | |
545 | ||
2a4d0c90 CH |
546 | static u64 bio_end_offset(struct bio *bio) |
547 | { | |
c45a8f2d | 548 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
549 | |
550 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
551 | } | |
552 | ||
771ed689 CM |
553 | static noinline int add_ra_bio_pages(struct inode *inode, |
554 | u64 compressed_end, | |
555 | struct compressed_bio *cb) | |
556 | { | |
557 | unsigned long end_index; | |
306e16ce | 558 | unsigned long pg_index; |
771ed689 CM |
559 | u64 last_offset; |
560 | u64 isize = i_size_read(inode); | |
561 | int ret; | |
562 | struct page *page; | |
563 | unsigned long nr_pages = 0; | |
564 | struct extent_map *em; | |
565 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
566 | struct extent_map_tree *em_tree; |
567 | struct extent_io_tree *tree; | |
568 | u64 end; | |
569 | int misses = 0; | |
570 | ||
2a4d0c90 | 571 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
572 | em_tree = &BTRFS_I(inode)->extent_tree; |
573 | tree = &BTRFS_I(inode)->io_tree; | |
574 | ||
575 | if (isize == 0) | |
576 | return 0; | |
577 | ||
09cbfeaf | 578 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 579 | |
d397712b | 580 | while (last_offset < compressed_end) { |
09cbfeaf | 581 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 582 | |
306e16ce | 583 | if (pg_index > end_index) |
771ed689 CM |
584 | break; |
585 | ||
0a943c65 | 586 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 587 | if (page && !xa_is_value(page)) { |
771ed689 CM |
588 | misses++; |
589 | if (misses > 4) | |
590 | break; | |
591 | goto next; | |
592 | } | |
593 | ||
c62d2555 MH |
594 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
595 | ~__GFP_FS)); | |
771ed689 CM |
596 | if (!page) |
597 | break; | |
598 | ||
c62d2555 | 599 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 600 | put_page(page); |
771ed689 CM |
601 | goto next; |
602 | } | |
603 | ||
771ed689 CM |
604 | /* |
605 | * at this point, we have a locked page in the page cache | |
606 | * for these bytes in the file. But, we have to make | |
607 | * sure they map to this compressed extent on disk. | |
608 | */ | |
32443de3 QW |
609 | ret = set_page_extent_mapped(page); |
610 | if (ret < 0) { | |
611 | unlock_page(page); | |
612 | put_page(page); | |
613 | break; | |
614 | } | |
615 | ||
616 | end = last_offset + PAGE_SIZE - 1; | |
d0082371 | 617 | lock_extent(tree, last_offset, end); |
890871be | 618 | read_lock(&em_tree->lock); |
771ed689 | 619 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 620 | PAGE_SIZE); |
890871be | 621 | read_unlock(&em_tree->lock); |
771ed689 CM |
622 | |
623 | if (!em || last_offset < em->start || | |
09cbfeaf | 624 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 625 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 626 | free_extent_map(em); |
d0082371 | 627 | unlock_extent(tree, last_offset, end); |
771ed689 | 628 | unlock_page(page); |
09cbfeaf | 629 | put_page(page); |
771ed689 CM |
630 | break; |
631 | } | |
632 | free_extent_map(em); | |
633 | ||
634 | if (page->index == end_index) { | |
7073017a | 635 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
636 | |
637 | if (zero_offset) { | |
638 | int zeros; | |
09cbfeaf | 639 | zeros = PAGE_SIZE - zero_offset; |
d048b9c2 | 640 | memzero_page(page, zero_offset, zeros); |
771ed689 | 641 | flush_dcache_page(page); |
771ed689 CM |
642 | } |
643 | } | |
644 | ||
645 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 646 | PAGE_SIZE, 0); |
771ed689 | 647 | |
09cbfeaf | 648 | if (ret == PAGE_SIZE) { |
771ed689 | 649 | nr_pages++; |
09cbfeaf | 650 | put_page(page); |
771ed689 | 651 | } else { |
d0082371 | 652 | unlock_extent(tree, last_offset, end); |
771ed689 | 653 | unlock_page(page); |
09cbfeaf | 654 | put_page(page); |
771ed689 CM |
655 | break; |
656 | } | |
657 | next: | |
09cbfeaf | 658 | last_offset += PAGE_SIZE; |
771ed689 | 659 | } |
771ed689 CM |
660 | return 0; |
661 | } | |
662 | ||
c8b97818 CM |
663 | /* |
664 | * for a compressed read, the bio we get passed has all the inode pages | |
665 | * in it. We don't actually do IO on those pages but allocate new ones | |
666 | * to hold the compressed pages on disk. | |
667 | * | |
4f024f37 | 668 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 669 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
670 | * |
671 | * After the compressed pages are read, we copy the bytes into the | |
672 | * bio we were passed and then call the bio end_io calls | |
673 | */ | |
4e4cbee9 | 674 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
675 | int mirror_num, unsigned long bio_flags) |
676 | { | |
0b246afa | 677 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
678 | struct extent_map_tree *em_tree; |
679 | struct compressed_bio *cb; | |
c8b97818 CM |
680 | unsigned long compressed_len; |
681 | unsigned long nr_pages; | |
306e16ce | 682 | unsigned long pg_index; |
c8b97818 | 683 | struct page *page; |
c8b97818 | 684 | struct bio *comp_bio; |
1201b58b | 685 | u64 cur_disk_byte = bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
686 | u64 em_len; |
687 | u64 em_start; | |
c8b97818 | 688 | struct extent_map *em; |
4e4cbee9 | 689 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 690 | int faili = 0; |
10fe6ca8 | 691 | u8 *sums; |
c8b97818 | 692 | |
c8b97818 CM |
693 | em_tree = &BTRFS_I(inode)->extent_tree; |
694 | ||
695 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 696 | read_lock(&em_tree->lock); |
c8b97818 | 697 | em = lookup_extent_mapping(em_tree, |
263663cd | 698 | page_offset(bio_first_page_all(bio)), |
be6a1361 | 699 | fs_info->sectorsize); |
890871be | 700 | read_unlock(&em_tree->lock); |
285190d9 | 701 | if (!em) |
4e4cbee9 | 702 | return BLK_STS_IOERR; |
c8b97818 | 703 | |
d20f7043 | 704 | compressed_len = em->block_len; |
2ff7e61e | 705 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 706 | if (!cb) |
707 | goto out; | |
708 | ||
a50299ae | 709 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
710 | cb->errors = 0; |
711 | cb->inode = inode; | |
d20f7043 | 712 | cb->mirror_num = mirror_num; |
10fe6ca8 | 713 | sums = cb->sums; |
c8b97818 | 714 | |
ff5b7ee3 | 715 | cb->start = em->orig_start; |
e04ca626 CM |
716 | em_len = em->len; |
717 | em_start = em->start; | |
d20f7043 | 718 | |
c8b97818 | 719 | free_extent_map(em); |
e04ca626 | 720 | em = NULL; |
c8b97818 | 721 | |
81381053 | 722 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 723 | cb->compressed_len = compressed_len; |
261507a0 | 724 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
725 | cb->orig_bio = bio; |
726 | ||
09cbfeaf | 727 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 728 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 729 | GFP_NOFS); |
6b82ce8d | 730 | if (!cb->compressed_pages) |
731 | goto fail1; | |
732 | ||
306e16ce DS |
733 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
734 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS | | |
c8b97818 | 735 | __GFP_HIGHMEM); |
15e3004a JB |
736 | if (!cb->compressed_pages[pg_index]) { |
737 | faili = pg_index - 1; | |
0e9350de | 738 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 739 | goto fail2; |
15e3004a | 740 | } |
c8b97818 | 741 | } |
15e3004a | 742 | faili = nr_pages - 1; |
c8b97818 CM |
743 | cb->nr_pages = nr_pages; |
744 | ||
7f042a83 | 745 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 746 | |
771ed689 | 747 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 748 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 749 | |
e749af44 | 750 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 751 | comp_bio->bi_opf = REQ_OP_READ; |
c8b97818 CM |
752 | comp_bio->bi_private = cb; |
753 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 754 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 755 | |
306e16ce | 756 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
be6a1361 | 757 | u32 pg_len = PAGE_SIZE; |
4e4cbee9 CH |
758 | int submit = 0; |
759 | ||
be6a1361 QW |
760 | /* |
761 | * To handle subpage case, we need to make sure the bio only | |
762 | * covers the range we need. | |
763 | * | |
764 | * If we're at the last page, truncate the length to only cover | |
765 | * the remaining part. | |
766 | */ | |
767 | if (pg_index == nr_pages - 1) | |
768 | pg_len = min_t(u32, PAGE_SIZE, | |
769 | compressed_len - pg_index * PAGE_SIZE); | |
770 | ||
306e16ce | 771 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 772 | page->mapping = inode->i_mapping; |
09cbfeaf | 773 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 774 | |
4f024f37 | 775 | if (comp_bio->bi_iter.bi_size) |
be6a1361 | 776 | submit = btrfs_bio_fits_in_stripe(page, pg_len, |
da12fe54 | 777 | comp_bio, 0); |
c8b97818 | 778 | |
70b99e69 | 779 | page->mapping = NULL; |
be6a1361 | 780 | if (submit || bio_add_page(comp_bio, page, pg_len, 0) < pg_len) { |
10fe6ca8 JT |
781 | unsigned int nr_sectors; |
782 | ||
0b246afa JM |
783 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
784 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 785 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 786 | |
af09abfe CM |
787 | /* |
788 | * inc the count before we submit the bio so | |
789 | * we know the end IO handler won't happen before | |
790 | * we inc the count. Otherwise, the cb might get | |
791 | * freed before we're done setting it up | |
792 | */ | |
a50299ae | 793 | refcount_inc(&cb->pending_bios); |
af09abfe | 794 | |
6275193e | 795 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
334c16d8 | 796 | BUG_ON(ret); /* -ENOMEM */ |
10fe6ca8 JT |
797 | |
798 | nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, | |
799 | fs_info->sectorsize); | |
713cebfb | 800 | sums += fs_info->csum_size * nr_sectors; |
d20f7043 | 801 | |
08635bae | 802 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 803 | if (ret) { |
4e4cbee9 | 804 | comp_bio->bi_status = ret; |
4246a0b6 CH |
805 | bio_endio(comp_bio); |
806 | } | |
c8b97818 | 807 | |
e749af44 | 808 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
ebcc3263 | 809 | comp_bio->bi_opf = REQ_OP_READ; |
771ed689 CM |
810 | comp_bio->bi_private = cb; |
811 | comp_bio->bi_end_io = end_compressed_bio_read; | |
812 | ||
be6a1361 | 813 | bio_add_page(comp_bio, page, pg_len, 0); |
c8b97818 | 814 | } |
be6a1361 | 815 | cur_disk_byte += pg_len; |
c8b97818 | 816 | } |
c8b97818 | 817 | |
0b246afa | 818 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 819 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 820 | |
6275193e | 821 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
334c16d8 | 822 | BUG_ON(ret); /* -ENOMEM */ |
d20f7043 | 823 | |
08635bae | 824 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 825 | if (ret) { |
4e4cbee9 | 826 | comp_bio->bi_status = ret; |
4246a0b6 CH |
827 | bio_endio(comp_bio); |
828 | } | |
c8b97818 | 829 | |
c8b97818 | 830 | return 0; |
6b82ce8d | 831 | |
832 | fail2: | |
15e3004a JB |
833 | while (faili >= 0) { |
834 | __free_page(cb->compressed_pages[faili]); | |
835 | faili--; | |
836 | } | |
6b82ce8d | 837 | |
838 | kfree(cb->compressed_pages); | |
839 | fail1: | |
840 | kfree(cb); | |
841 | out: | |
842 | free_extent_map(em); | |
843 | return ret; | |
c8b97818 | 844 | } |
261507a0 | 845 | |
17b5a6c1 TT |
846 | /* |
847 | * Heuristic uses systematic sampling to collect data from the input data | |
848 | * range, the logic can be tuned by the following constants: | |
849 | * | |
850 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
851 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
852 | */ | |
853 | #define SAMPLING_READ_SIZE (16) | |
854 | #define SAMPLING_INTERVAL (256) | |
855 | ||
856 | /* | |
857 | * For statistical analysis of the input data we consider bytes that form a | |
858 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
859 | * many times the object appeared in the sample. | |
860 | */ | |
861 | #define BUCKET_SIZE (256) | |
862 | ||
863 | /* | |
864 | * The size of the sample is based on a statistical sampling rule of thumb. | |
865 | * The common way is to perform sampling tests as long as the number of | |
866 | * elements in each cell is at least 5. | |
867 | * | |
868 | * Instead of 5, we choose 32 to obtain more accurate results. | |
869 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
870 | * sample size bound by 8192. | |
871 | * | |
872 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
873 | * from up to 512 locations. | |
874 | */ | |
875 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
876 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
877 | ||
878 | struct bucket_item { | |
879 | u32 count; | |
880 | }; | |
4e439a0b TT |
881 | |
882 | struct heuristic_ws { | |
17b5a6c1 TT |
883 | /* Partial copy of input data */ |
884 | u8 *sample; | |
a440d48c | 885 | u32 sample_size; |
17b5a6c1 TT |
886 | /* Buckets store counters for each byte value */ |
887 | struct bucket_item *bucket; | |
440c840c TT |
888 | /* Sorting buffer */ |
889 | struct bucket_item *bucket_b; | |
4e439a0b TT |
890 | struct list_head list; |
891 | }; | |
892 | ||
92ee5530 DZ |
893 | static struct workspace_manager heuristic_wsm; |
894 | ||
4e439a0b TT |
895 | static void free_heuristic_ws(struct list_head *ws) |
896 | { | |
897 | struct heuristic_ws *workspace; | |
898 | ||
899 | workspace = list_entry(ws, struct heuristic_ws, list); | |
900 | ||
17b5a6c1 TT |
901 | kvfree(workspace->sample); |
902 | kfree(workspace->bucket); | |
440c840c | 903 | kfree(workspace->bucket_b); |
4e439a0b TT |
904 | kfree(workspace); |
905 | } | |
906 | ||
7bf49943 | 907 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
908 | { |
909 | struct heuristic_ws *ws; | |
910 | ||
911 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
912 | if (!ws) | |
913 | return ERR_PTR(-ENOMEM); | |
914 | ||
17b5a6c1 TT |
915 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
916 | if (!ws->sample) | |
917 | goto fail; | |
918 | ||
919 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
920 | if (!ws->bucket) | |
921 | goto fail; | |
4e439a0b | 922 | |
440c840c TT |
923 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
924 | if (!ws->bucket_b) | |
925 | goto fail; | |
926 | ||
17b5a6c1 | 927 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 928 | return &ws->list; |
17b5a6c1 TT |
929 | fail: |
930 | free_heuristic_ws(&ws->list); | |
931 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
932 | } |
933 | ||
ca4ac360 | 934 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 935 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
936 | }; |
937 | ||
e8c9f186 | 938 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
939 | /* The heuristic is represented as compression type 0 */ |
940 | &btrfs_heuristic_compress, | |
261507a0 | 941 | &btrfs_zlib_compress, |
a6fa6fae | 942 | &btrfs_lzo_compress, |
5c1aab1d | 943 | &btrfs_zstd_compress, |
261507a0 LZ |
944 | }; |
945 | ||
c778df14 DS |
946 | static struct list_head *alloc_workspace(int type, unsigned int level) |
947 | { | |
948 | switch (type) { | |
949 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
950 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
951 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
952 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
953 | default: | |
954 | /* | |
955 | * This can't happen, the type is validated several times | |
956 | * before we get here. | |
957 | */ | |
958 | BUG(); | |
959 | } | |
960 | } | |
961 | ||
1e002351 DS |
962 | static void free_workspace(int type, struct list_head *ws) |
963 | { | |
964 | switch (type) { | |
965 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
966 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
967 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
968 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
969 | default: | |
970 | /* | |
971 | * This can't happen, the type is validated several times | |
972 | * before we get here. | |
973 | */ | |
974 | BUG(); | |
975 | } | |
976 | } | |
977 | ||
d5517033 | 978 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 979 | { |
0cf25213 | 980 | struct workspace_manager *wsm; |
4e439a0b | 981 | struct list_head *workspace; |
261507a0 | 982 | |
0cf25213 | 983 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
984 | INIT_LIST_HEAD(&wsm->idle_ws); |
985 | spin_lock_init(&wsm->ws_lock); | |
986 | atomic_set(&wsm->total_ws, 0); | |
987 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 988 | |
1666edab DZ |
989 | /* |
990 | * Preallocate one workspace for each compression type so we can | |
991 | * guarantee forward progress in the worst case | |
992 | */ | |
c778df14 | 993 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
994 | if (IS_ERR(workspace)) { |
995 | pr_warn( | |
996 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
997 | } else { | |
92ee5530 DZ |
998 | atomic_set(&wsm->total_ws, 1); |
999 | wsm->free_ws = 1; | |
1000 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
1001 | } |
1002 | } | |
1003 | ||
2510307e | 1004 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 1005 | { |
2dba7143 | 1006 | struct workspace_manager *wsman; |
1666edab DZ |
1007 | struct list_head *ws; |
1008 | ||
2dba7143 | 1009 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
1010 | while (!list_empty(&wsman->idle_ws)) { |
1011 | ws = wsman->idle_ws.next; | |
1012 | list_del(ws); | |
1e002351 | 1013 | free_workspace(type, ws); |
1666edab | 1014 | atomic_dec(&wsman->total_ws); |
261507a0 | 1015 | } |
261507a0 LZ |
1016 | } |
1017 | ||
1018 | /* | |
e721e49d DS |
1019 | * This finds an available workspace or allocates a new one. |
1020 | * If it's not possible to allocate a new one, waits until there's one. | |
1021 | * Preallocation makes a forward progress guarantees and we do not return | |
1022 | * errors. | |
261507a0 | 1023 | */ |
5907a9bb | 1024 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 1025 | { |
5907a9bb | 1026 | struct workspace_manager *wsm; |
261507a0 LZ |
1027 | struct list_head *workspace; |
1028 | int cpus = num_online_cpus(); | |
fe308533 | 1029 | unsigned nofs_flag; |
4e439a0b TT |
1030 | struct list_head *idle_ws; |
1031 | spinlock_t *ws_lock; | |
1032 | atomic_t *total_ws; | |
1033 | wait_queue_head_t *ws_wait; | |
1034 | int *free_ws; | |
1035 | ||
5907a9bb | 1036 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1037 | idle_ws = &wsm->idle_ws; |
1038 | ws_lock = &wsm->ws_lock; | |
1039 | total_ws = &wsm->total_ws; | |
1040 | ws_wait = &wsm->ws_wait; | |
1041 | free_ws = &wsm->free_ws; | |
261507a0 | 1042 | |
261507a0 | 1043 | again: |
d9187649 BL |
1044 | spin_lock(ws_lock); |
1045 | if (!list_empty(idle_ws)) { | |
1046 | workspace = idle_ws->next; | |
261507a0 | 1047 | list_del(workspace); |
6ac10a6a | 1048 | (*free_ws)--; |
d9187649 | 1049 | spin_unlock(ws_lock); |
261507a0 LZ |
1050 | return workspace; |
1051 | ||
1052 | } | |
6ac10a6a | 1053 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
1054 | DEFINE_WAIT(wait); |
1055 | ||
d9187649 BL |
1056 | spin_unlock(ws_lock); |
1057 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 1058 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 1059 | schedule(); |
d9187649 | 1060 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
1061 | goto again; |
1062 | } | |
6ac10a6a | 1063 | atomic_inc(total_ws); |
d9187649 | 1064 | spin_unlock(ws_lock); |
261507a0 | 1065 | |
fe308533 DS |
1066 | /* |
1067 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
1068 | * to turn it off here because we might get called from the restricted | |
1069 | * context of btrfs_compress_bio/btrfs_compress_pages | |
1070 | */ | |
1071 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1072 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1073 | memalloc_nofs_restore(nofs_flag); |
1074 | ||
261507a0 | 1075 | if (IS_ERR(workspace)) { |
6ac10a6a | 1076 | atomic_dec(total_ws); |
d9187649 | 1077 | wake_up(ws_wait); |
e721e49d DS |
1078 | |
1079 | /* | |
1080 | * Do not return the error but go back to waiting. There's a | |
1081 | * workspace preallocated for each type and the compression | |
1082 | * time is bounded so we get to a workspace eventually. This | |
1083 | * makes our caller's life easier. | |
52356716 DS |
1084 | * |
1085 | * To prevent silent and low-probability deadlocks (when the | |
1086 | * initial preallocation fails), check if there are any | |
1087 | * workspaces at all. | |
e721e49d | 1088 | */ |
52356716 DS |
1089 | if (atomic_read(total_ws) == 0) { |
1090 | static DEFINE_RATELIMIT_STATE(_rs, | |
1091 | /* once per minute */ 60 * HZ, | |
1092 | /* no burst */ 1); | |
1093 | ||
1094 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1095 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1096 | } |
1097 | } | |
e721e49d | 1098 | goto again; |
261507a0 LZ |
1099 | } |
1100 | return workspace; | |
1101 | } | |
1102 | ||
7bf49943 | 1103 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1104 | { |
6a0d1272 | 1105 | switch (type) { |
5907a9bb | 1106 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1107 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1108 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1109 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1110 | default: | |
1111 | /* | |
1112 | * This can't happen, the type is validated several times | |
1113 | * before we get here. | |
1114 | */ | |
1115 | BUG(); | |
1116 | } | |
929f4baf DZ |
1117 | } |
1118 | ||
261507a0 LZ |
1119 | /* |
1120 | * put a workspace struct back on the list or free it if we have enough | |
1121 | * idle ones sitting around | |
1122 | */ | |
a3bbd2a9 | 1123 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1124 | { |
a3bbd2a9 | 1125 | struct workspace_manager *wsm; |
4e439a0b TT |
1126 | struct list_head *idle_ws; |
1127 | spinlock_t *ws_lock; | |
1128 | atomic_t *total_ws; | |
1129 | wait_queue_head_t *ws_wait; | |
1130 | int *free_ws; | |
1131 | ||
a3bbd2a9 | 1132 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1133 | idle_ws = &wsm->idle_ws; |
1134 | ws_lock = &wsm->ws_lock; | |
1135 | total_ws = &wsm->total_ws; | |
1136 | ws_wait = &wsm->ws_wait; | |
1137 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1138 | |
1139 | spin_lock(ws_lock); | |
26b28dce | 1140 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1141 | list_add(ws, idle_ws); |
6ac10a6a | 1142 | (*free_ws)++; |
d9187649 | 1143 | spin_unlock(ws_lock); |
261507a0 LZ |
1144 | goto wake; |
1145 | } | |
d9187649 | 1146 | spin_unlock(ws_lock); |
261507a0 | 1147 | |
1e002351 | 1148 | free_workspace(type, ws); |
6ac10a6a | 1149 | atomic_dec(total_ws); |
261507a0 | 1150 | wake: |
093258e6 | 1151 | cond_wake_up(ws_wait); |
261507a0 LZ |
1152 | } |
1153 | ||
929f4baf DZ |
1154 | static void put_workspace(int type, struct list_head *ws) |
1155 | { | |
bd3a5287 | 1156 | switch (type) { |
a3bbd2a9 DS |
1157 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1158 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1159 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1160 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1161 | default: | |
1162 | /* | |
1163 | * This can't happen, the type is validated several times | |
1164 | * before we get here. | |
1165 | */ | |
1166 | BUG(); | |
1167 | } | |
929f4baf DZ |
1168 | } |
1169 | ||
adbab642 AJ |
1170 | /* |
1171 | * Adjust @level according to the limits of the compression algorithm or | |
1172 | * fallback to default | |
1173 | */ | |
1174 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1175 | { | |
1176 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1177 | ||
1178 | if (level == 0) | |
1179 | level = ops->default_level; | |
1180 | else | |
1181 | level = min(level, ops->max_level); | |
1182 | ||
1183 | return level; | |
1184 | } | |
1185 | ||
261507a0 | 1186 | /* |
38c31464 DS |
1187 | * Given an address space and start and length, compress the bytes into @pages |
1188 | * that are allocated on demand. | |
261507a0 | 1189 | * |
f51d2b59 DS |
1190 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1191 | * default the algorithm chooses and is opaque here; | |
1192 | * - compression algo are 0-3 | |
1193 | * - the level are bits 4-7 | |
1194 | * | |
4d3a800e DS |
1195 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1196 | * and returns number of actually allocated pages | |
261507a0 | 1197 | * |
38c31464 DS |
1198 | * @total_in is used to return the number of bytes actually read. It |
1199 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1200 | * ran out of room in the pages array or because we cross the |
1201 | * max_out threshold. | |
1202 | * | |
38c31464 DS |
1203 | * @total_out is an in/out parameter, must be set to the input length and will |
1204 | * be also used to return the total number of compressed bytes | |
261507a0 | 1205 | * |
38c31464 | 1206 | * @max_out tells us the max number of bytes that we're allowed to |
261507a0 LZ |
1207 | * stuff into pages |
1208 | */ | |
f51d2b59 | 1209 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1210 | u64 start, struct page **pages, |
261507a0 LZ |
1211 | unsigned long *out_pages, |
1212 | unsigned long *total_in, | |
e5d74902 | 1213 | unsigned long *total_out) |
261507a0 | 1214 | { |
1972708a | 1215 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1216 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1217 | struct list_head *workspace; |
1218 | int ret; | |
1219 | ||
b0c1fe1e | 1220 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1221 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1222 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1223 | out_pages, total_in, total_out); | |
929f4baf | 1224 | put_workspace(type, workspace); |
261507a0 LZ |
1225 | return ret; |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * pages_in is an array of pages with compressed data. | |
1230 | * | |
1231 | * disk_start is the starting logical offset of this array in the file | |
1232 | * | |
974b1adc | 1233 | * orig_bio contains the pages from the file that we want to decompress into |
261507a0 LZ |
1234 | * |
1235 | * srclen is the number of bytes in pages_in | |
1236 | * | |
1237 | * The basic idea is that we have a bio that was created by readpages. | |
1238 | * The pages in the bio are for the uncompressed data, and they may not | |
1239 | * be contiguous. They all correspond to the range of bytes covered by | |
1240 | * the compressed extent. | |
1241 | */ | |
8140dc30 | 1242 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1243 | { |
1244 | struct list_head *workspace; | |
1245 | int ret; | |
8140dc30 | 1246 | int type = cb->compress_type; |
261507a0 | 1247 | |
7bf49943 | 1248 | workspace = get_workspace(type, 0); |
1e4eb746 | 1249 | ret = compression_decompress_bio(type, workspace, cb); |
929f4baf | 1250 | put_workspace(type, workspace); |
e1ddce71 | 1251 | |
261507a0 LZ |
1252 | return ret; |
1253 | } | |
1254 | ||
1255 | /* | |
1256 | * a less complex decompression routine. Our compressed data fits in a | |
1257 | * single page, and we want to read a single page out of it. | |
1258 | * start_byte tells us the offset into the compressed data we're interested in | |
1259 | */ | |
1260 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1261 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1262 | { | |
1263 | struct list_head *workspace; | |
1264 | int ret; | |
1265 | ||
7bf49943 | 1266 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1267 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1268 | start_byte, srclen, destlen); | |
929f4baf | 1269 | put_workspace(type, workspace); |
7bf49943 | 1270 | |
261507a0 LZ |
1271 | return ret; |
1272 | } | |
1273 | ||
1666edab DZ |
1274 | void __init btrfs_init_compress(void) |
1275 | { | |
d5517033 DS |
1276 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1277 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1278 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1279 | zstd_init_workspace_manager(); | |
1666edab DZ |
1280 | } |
1281 | ||
e67c718b | 1282 | void __cold btrfs_exit_compress(void) |
261507a0 | 1283 | { |
2510307e DS |
1284 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1285 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1286 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1287 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1288 | } |
3a39c18d LZ |
1289 | |
1290 | /* | |
1291 | * Copy uncompressed data from working buffer to pages. | |
1292 | * | |
1293 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1294 | * | |
1295 | * total_out is the last byte of the buffer | |
1296 | */ | |
14a3357b | 1297 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1298 | unsigned long total_out, u64 disk_start, |
974b1adc | 1299 | struct bio *bio) |
3a39c18d LZ |
1300 | { |
1301 | unsigned long buf_offset; | |
1302 | unsigned long current_buf_start; | |
1303 | unsigned long start_byte; | |
6e78b3f7 | 1304 | unsigned long prev_start_byte; |
3a39c18d LZ |
1305 | unsigned long working_bytes = total_out - buf_start; |
1306 | unsigned long bytes; | |
974b1adc | 1307 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1308 | |
1309 | /* | |
1310 | * start byte is the first byte of the page we're currently | |
1311 | * copying into relative to the start of the compressed data. | |
1312 | */ | |
974b1adc | 1313 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1314 | |
1315 | /* we haven't yet hit data corresponding to this page */ | |
1316 | if (total_out <= start_byte) | |
1317 | return 1; | |
1318 | ||
1319 | /* | |
1320 | * the start of the data we care about is offset into | |
1321 | * the middle of our working buffer | |
1322 | */ | |
1323 | if (total_out > start_byte && buf_start < start_byte) { | |
1324 | buf_offset = start_byte - buf_start; | |
1325 | working_bytes -= buf_offset; | |
1326 | } else { | |
1327 | buf_offset = 0; | |
1328 | } | |
1329 | current_buf_start = buf_start; | |
1330 | ||
1331 | /* copy bytes from the working buffer into the pages */ | |
1332 | while (working_bytes > 0) { | |
974b1adc | 1333 | bytes = min_t(unsigned long, bvec.bv_len, |
3fd396af | 1334 | PAGE_SIZE - (buf_offset % PAGE_SIZE)); |
3a39c18d | 1335 | bytes = min(bytes, working_bytes); |
974b1adc | 1336 | |
3590ec58 IW |
1337 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, buf + buf_offset, |
1338 | bytes); | |
974b1adc | 1339 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1340 | |
3a39c18d LZ |
1341 | buf_offset += bytes; |
1342 | working_bytes -= bytes; | |
1343 | current_buf_start += bytes; | |
1344 | ||
1345 | /* check if we need to pick another page */ | |
974b1adc CH |
1346 | bio_advance(bio, bytes); |
1347 | if (!bio->bi_iter.bi_size) | |
1348 | return 0; | |
1349 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1350 | prev_start_byte = start_byte; |
974b1adc | 1351 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1352 | |
974b1adc | 1353 | /* |
6e78b3f7 OS |
1354 | * We need to make sure we're only adjusting |
1355 | * our offset into compression working buffer when | |
1356 | * we're switching pages. Otherwise we can incorrectly | |
1357 | * keep copying when we were actually done. | |
974b1adc | 1358 | */ |
6e78b3f7 OS |
1359 | if (start_byte != prev_start_byte) { |
1360 | /* | |
1361 | * make sure our new page is covered by this | |
1362 | * working buffer | |
1363 | */ | |
1364 | if (total_out <= start_byte) | |
1365 | return 1; | |
3a39c18d | 1366 | |
6e78b3f7 OS |
1367 | /* |
1368 | * the next page in the biovec might not be adjacent | |
1369 | * to the last page, but it might still be found | |
1370 | * inside this working buffer. bump our offset pointer | |
1371 | */ | |
1372 | if (total_out > start_byte && | |
1373 | current_buf_start < start_byte) { | |
1374 | buf_offset = start_byte - buf_start; | |
1375 | working_bytes = total_out - start_byte; | |
1376 | current_buf_start = buf_start + buf_offset; | |
1377 | } | |
3a39c18d LZ |
1378 | } |
1379 | } | |
1380 | ||
1381 | return 1; | |
1382 | } | |
c2fcdcdf | 1383 | |
19562430 TT |
1384 | /* |
1385 | * Shannon Entropy calculation | |
1386 | * | |
52042d8e | 1387 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1388 | * Try calculating entropy to estimate the average minimum number of bits |
1389 | * needed to encode the sampled data. | |
1390 | * | |
1391 | * For convenience, return the percentage of needed bits, instead of amount of | |
1392 | * bits directly. | |
1393 | * | |
1394 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1395 | * and can be compressible with high probability | |
1396 | * | |
1397 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1398 | * | |
1399 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1400 | */ | |
1401 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1402 | #define ENTROPY_LVL_HIGH (80) | |
1403 | ||
1404 | /* | |
1405 | * For increasead precision in shannon_entropy calculation, | |
1406 | * let's do pow(n, M) to save more digits after comma: | |
1407 | * | |
1408 | * - maximum int bit length is 64 | |
1409 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1410 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1411 | * | |
1412 | * So use pow(n, 4). | |
1413 | */ | |
1414 | static inline u32 ilog2_w(u64 n) | |
1415 | { | |
1416 | return ilog2(n * n * n * n); | |
1417 | } | |
1418 | ||
1419 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1420 | { | |
1421 | const u32 entropy_max = 8 * ilog2_w(2); | |
1422 | u32 entropy_sum = 0; | |
1423 | u32 p, p_base, sz_base; | |
1424 | u32 i; | |
1425 | ||
1426 | sz_base = ilog2_w(ws->sample_size); | |
1427 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1428 | p = ws->bucket[i].count; | |
1429 | p_base = ilog2_w(p); | |
1430 | entropy_sum += p * (sz_base - p_base); | |
1431 | } | |
1432 | ||
1433 | entropy_sum /= ws->sample_size; | |
1434 | return entropy_sum * 100 / entropy_max; | |
1435 | } | |
1436 | ||
440c840c TT |
1437 | #define RADIX_BASE 4U |
1438 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1439 | ||
1440 | static u8 get4bits(u64 num, int shift) { | |
1441 | u8 low4bits; | |
1442 | ||
1443 | num >>= shift; | |
1444 | /* Reverse order */ | |
1445 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1446 | return low4bits; | |
1447 | } | |
1448 | ||
440c840c TT |
1449 | /* |
1450 | * Use 4 bits as radix base | |
52042d8e | 1451 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1452 | * |
1453 | * @array - array that will be sorted | |
1454 | * @array_buf - buffer array to store sorting results | |
1455 | * must be equal in size to @array | |
1456 | * @num - array size | |
440c840c | 1457 | */ |
23ae8c63 | 1458 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1459 | int num) |
858177d3 | 1460 | { |
440c840c TT |
1461 | u64 max_num; |
1462 | u64 buf_num; | |
1463 | u32 counters[COUNTERS_SIZE]; | |
1464 | u32 new_addr; | |
1465 | u32 addr; | |
1466 | int bitlen; | |
1467 | int shift; | |
1468 | int i; | |
858177d3 | 1469 | |
440c840c TT |
1470 | /* |
1471 | * Try avoid useless loop iterations for small numbers stored in big | |
1472 | * counters. Example: 48 33 4 ... in 64bit array | |
1473 | */ | |
23ae8c63 | 1474 | max_num = array[0].count; |
440c840c | 1475 | for (i = 1; i < num; i++) { |
23ae8c63 | 1476 | buf_num = array[i].count; |
440c840c TT |
1477 | if (buf_num > max_num) |
1478 | max_num = buf_num; | |
1479 | } | |
1480 | ||
1481 | buf_num = ilog2(max_num); | |
1482 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1483 | ||
1484 | shift = 0; | |
1485 | while (shift < bitlen) { | |
1486 | memset(counters, 0, sizeof(counters)); | |
1487 | ||
1488 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1489 | buf_num = array[i].count; |
440c840c TT |
1490 | addr = get4bits(buf_num, shift); |
1491 | counters[addr]++; | |
1492 | } | |
1493 | ||
1494 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1495 | counters[i] += counters[i - 1]; | |
1496 | ||
1497 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1498 | buf_num = array[i].count; |
440c840c TT |
1499 | addr = get4bits(buf_num, shift); |
1500 | counters[addr]--; | |
1501 | new_addr = counters[addr]; | |
7add17be | 1502 | array_buf[new_addr] = array[i]; |
440c840c TT |
1503 | } |
1504 | ||
1505 | shift += RADIX_BASE; | |
1506 | ||
1507 | /* | |
1508 | * Normal radix expects to move data from a temporary array, to | |
1509 | * the main one. But that requires some CPU time. Avoid that | |
1510 | * by doing another sort iteration to original array instead of | |
1511 | * memcpy() | |
1512 | */ | |
1513 | memset(counters, 0, sizeof(counters)); | |
1514 | ||
1515 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1516 | buf_num = array_buf[i].count; |
440c840c TT |
1517 | addr = get4bits(buf_num, shift); |
1518 | counters[addr]++; | |
1519 | } | |
1520 | ||
1521 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1522 | counters[i] += counters[i - 1]; | |
1523 | ||
1524 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1525 | buf_num = array_buf[i].count; |
440c840c TT |
1526 | addr = get4bits(buf_num, shift); |
1527 | counters[addr]--; | |
1528 | new_addr = counters[addr]; | |
7add17be | 1529 | array[new_addr] = array_buf[i]; |
440c840c TT |
1530 | } |
1531 | ||
1532 | shift += RADIX_BASE; | |
1533 | } | |
858177d3 TT |
1534 | } |
1535 | ||
1536 | /* | |
1537 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1538 | * | |
1539 | * There are several types of structured binary data that use nearly all byte | |
1540 | * values. The distribution can be uniform and counts in all buckets will be | |
1541 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1542 | * | |
1543 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1544 | * be potentially compressible, but we have to take a few more steps to decide | |
1545 | * how much. | |
1546 | * | |
1547 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1548 | * compression algo can easy fix that | |
1549 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1550 | * probability is not compressible | |
1551 | */ | |
1552 | #define BYTE_CORE_SET_LOW (64) | |
1553 | #define BYTE_CORE_SET_HIGH (200) | |
1554 | ||
1555 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1556 | { | |
1557 | u32 i; | |
1558 | u32 coreset_sum = 0; | |
1559 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1560 | struct bucket_item *bucket = ws->bucket; | |
1561 | ||
1562 | /* Sort in reverse order */ | |
36243c91 | 1563 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1564 | |
1565 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1566 | coreset_sum += bucket[i].count; | |
1567 | ||
1568 | if (coreset_sum > core_set_threshold) | |
1569 | return i; | |
1570 | ||
1571 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1572 | coreset_sum += bucket[i].count; | |
1573 | if (coreset_sum > core_set_threshold) | |
1574 | break; | |
1575 | } | |
1576 | ||
1577 | return i; | |
1578 | } | |
1579 | ||
a288e92c TT |
1580 | /* |
1581 | * Count byte values in buckets. | |
1582 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1583 | * Because in most text-like data byte set is restricted to limited number of | |
1584 | * possible characters, and that restriction in most cases makes data easy to | |
1585 | * compress. | |
1586 | * | |
1587 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1588 | * less - compressible | |
1589 | * more - need additional analysis | |
1590 | */ | |
1591 | #define BYTE_SET_THRESHOLD (64) | |
1592 | ||
1593 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1594 | { | |
1595 | u32 i; | |
1596 | u32 byte_set_size = 0; | |
1597 | ||
1598 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1599 | if (ws->bucket[i].count > 0) | |
1600 | byte_set_size++; | |
1601 | } | |
1602 | ||
1603 | /* | |
1604 | * Continue collecting count of byte values in buckets. If the byte | |
1605 | * set size is bigger then the threshold, it's pointless to continue, | |
1606 | * the detection technique would fail for this type of data. | |
1607 | */ | |
1608 | for (; i < BUCKET_SIZE; i++) { | |
1609 | if (ws->bucket[i].count > 0) { | |
1610 | byte_set_size++; | |
1611 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1612 | return byte_set_size; | |
1613 | } | |
1614 | } | |
1615 | ||
1616 | return byte_set_size; | |
1617 | } | |
1618 | ||
1fe4f6fa TT |
1619 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1620 | { | |
1621 | const u32 half_of_sample = ws->sample_size / 2; | |
1622 | const u8 *data = ws->sample; | |
1623 | ||
1624 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1625 | } | |
1626 | ||
a440d48c TT |
1627 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1628 | struct heuristic_ws *ws) | |
1629 | { | |
1630 | struct page *page; | |
1631 | u64 index, index_end; | |
1632 | u32 i, curr_sample_pos; | |
1633 | u8 *in_data; | |
1634 | ||
1635 | /* | |
1636 | * Compression handles the input data by chunks of 128KiB | |
1637 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1638 | * | |
1639 | * We do the same for the heuristic and loop over the whole range. | |
1640 | * | |
1641 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1642 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1643 | */ | |
1644 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1645 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1646 | ||
1647 | index = start >> PAGE_SHIFT; | |
1648 | index_end = end >> PAGE_SHIFT; | |
1649 | ||
1650 | /* Don't miss unaligned end */ | |
1651 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1652 | index_end++; | |
1653 | ||
1654 | curr_sample_pos = 0; | |
1655 | while (index < index_end) { | |
1656 | page = find_get_page(inode->i_mapping, index); | |
58c1a35c | 1657 | in_data = kmap_local_page(page); |
a440d48c TT |
1658 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1659 | i = start % PAGE_SIZE; | |
1660 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1661 | /* Don't sample any garbage from the last page */ | |
1662 | if (start > end - SAMPLING_READ_SIZE) | |
1663 | break; | |
1664 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1665 | SAMPLING_READ_SIZE); | |
1666 | i += SAMPLING_INTERVAL; | |
1667 | start += SAMPLING_INTERVAL; | |
1668 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1669 | } | |
58c1a35c | 1670 | kunmap_local(in_data); |
a440d48c TT |
1671 | put_page(page); |
1672 | ||
1673 | index++; | |
1674 | } | |
1675 | ||
1676 | ws->sample_size = curr_sample_pos; | |
1677 | } | |
1678 | ||
c2fcdcdf TT |
1679 | /* |
1680 | * Compression heuristic. | |
1681 | * | |
1682 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1683 | * quickly (compared to direct compression) detect data characteristics | |
1684 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1685 | * data. | |
1686 | * | |
1687 | * The following types of analysis can be performed: | |
1688 | * - detect mostly zero data | |
1689 | * - detect data with low "byte set" size (text, etc) | |
1690 | * - detect data with low/high "core byte" set | |
1691 | * | |
1692 | * Return non-zero if the compression should be done, 0 otherwise. | |
1693 | */ | |
1694 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1695 | { | |
7bf49943 | 1696 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1697 | struct heuristic_ws *ws; |
a440d48c TT |
1698 | u32 i; |
1699 | u8 byte; | |
19562430 | 1700 | int ret = 0; |
c2fcdcdf | 1701 | |
4e439a0b TT |
1702 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1703 | ||
a440d48c TT |
1704 | heuristic_collect_sample(inode, start, end, ws); |
1705 | ||
1fe4f6fa TT |
1706 | if (sample_repeated_patterns(ws)) { |
1707 | ret = 1; | |
1708 | goto out; | |
1709 | } | |
1710 | ||
a440d48c TT |
1711 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1712 | ||
1713 | for (i = 0; i < ws->sample_size; i++) { | |
1714 | byte = ws->sample[i]; | |
1715 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1716 | } |
1717 | ||
a288e92c TT |
1718 | i = byte_set_size(ws); |
1719 | if (i < BYTE_SET_THRESHOLD) { | |
1720 | ret = 2; | |
1721 | goto out; | |
1722 | } | |
1723 | ||
858177d3 TT |
1724 | i = byte_core_set_size(ws); |
1725 | if (i <= BYTE_CORE_SET_LOW) { | |
1726 | ret = 3; | |
1727 | goto out; | |
1728 | } | |
1729 | ||
1730 | if (i >= BYTE_CORE_SET_HIGH) { | |
1731 | ret = 0; | |
1732 | goto out; | |
1733 | } | |
1734 | ||
19562430 TT |
1735 | i = shannon_entropy(ws); |
1736 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1737 | ret = 4; | |
1738 | goto out; | |
1739 | } | |
1740 | ||
1741 | /* | |
1742 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1743 | * needed to give green light to compression. | |
1744 | * | |
1745 | * For now just assume that compression at that level is not worth the | |
1746 | * resources because: | |
1747 | * | |
1748 | * 1. it is possible to defrag the data later | |
1749 | * | |
1750 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1751 | * values, every bucket has counter at level ~54. The heuristic would | |
1752 | * be confused. This can happen when data have some internal repeated | |
1753 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1754 | * pairs of bytes, which is too costly. | |
1755 | */ | |
1756 | if (i < ENTROPY_LVL_HIGH) { | |
1757 | ret = 5; | |
1758 | goto out; | |
1759 | } else { | |
1760 | ret = 0; | |
1761 | goto out; | |
1762 | } | |
1763 | ||
1fe4f6fa | 1764 | out: |
929f4baf | 1765 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1766 | return ret; |
1767 | } | |
f51d2b59 | 1768 | |
d0ab62ce DZ |
1769 | /* |
1770 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1771 | * level, unrecognized string will set the default level | |
1772 | */ | |
1773 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1774 | { |
d0ab62ce DZ |
1775 | unsigned int level = 0; |
1776 | int ret; | |
1777 | ||
1778 | if (!type) | |
f51d2b59 DS |
1779 | return 0; |
1780 | ||
d0ab62ce DZ |
1781 | if (str[0] == ':') { |
1782 | ret = kstrtouint(str + 1, 10, &level); | |
1783 | if (ret) | |
1784 | level = 0; | |
1785 | } | |
1786 | ||
b0c1fe1e DS |
1787 | level = btrfs_compress_set_level(type, level); |
1788 | ||
1789 | return level; | |
1790 | } |