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" | |
6a404910 | 31 | #include "subpage.h" |
764c7c9a | 32 | #include "zoned.h" |
c8b97818 | 33 | |
e128f9c3 DS |
34 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
35 | ||
36 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
37 | { | |
38 | switch (type) { | |
39 | case BTRFS_COMPRESS_ZLIB: | |
40 | case BTRFS_COMPRESS_LZO: | |
41 | case BTRFS_COMPRESS_ZSTD: | |
42 | case BTRFS_COMPRESS_NONE: | |
43 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
44 | default: |
45 | break; | |
e128f9c3 DS |
46 | } |
47 | ||
48 | return NULL; | |
49 | } | |
50 | ||
aa53e3bf JT |
51 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
52 | { | |
53 | int i; | |
54 | ||
55 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
56 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
57 | ||
58 | if (len < comp_len) | |
59 | continue; | |
60 | ||
61 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
62 | return true; | |
63 | } | |
64 | return false; | |
65 | } | |
66 | ||
1e4eb746 DS |
67 | static int compression_compress_pages(int type, struct list_head *ws, |
68 | struct address_space *mapping, u64 start, struct page **pages, | |
69 | unsigned long *out_pages, unsigned long *total_in, | |
70 | unsigned long *total_out) | |
71 | { | |
72 | switch (type) { | |
73 | case BTRFS_COMPRESS_ZLIB: | |
74 | return zlib_compress_pages(ws, mapping, start, pages, | |
75 | out_pages, total_in, total_out); | |
76 | case BTRFS_COMPRESS_LZO: | |
77 | return lzo_compress_pages(ws, mapping, start, pages, | |
78 | out_pages, total_in, total_out); | |
79 | case BTRFS_COMPRESS_ZSTD: | |
80 | return zstd_compress_pages(ws, mapping, start, pages, | |
81 | out_pages, total_in, total_out); | |
82 | case BTRFS_COMPRESS_NONE: | |
83 | default: | |
84 | /* | |
1d8ba9e7 QW |
85 | * This can happen when compression races with remount setting |
86 | * it to 'no compress', while caller doesn't call | |
87 | * inode_need_compress() to check if we really need to | |
88 | * compress. | |
89 | * | |
90 | * Not a big deal, just need to inform caller that we | |
91 | * haven't allocated any pages yet. | |
1e4eb746 | 92 | */ |
1d8ba9e7 | 93 | *out_pages = 0; |
1e4eb746 DS |
94 | return -E2BIG; |
95 | } | |
96 | } | |
97 | ||
98 | static int compression_decompress_bio(int type, struct list_head *ws, | |
99 | struct compressed_bio *cb) | |
100 | { | |
101 | switch (type) { | |
102 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); | |
103 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
104 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
105 | case BTRFS_COMPRESS_NONE: | |
106 | default: | |
107 | /* | |
108 | * This can't happen, the type is validated several times | |
109 | * before we get here. | |
110 | */ | |
111 | BUG(); | |
112 | } | |
113 | } | |
114 | ||
115 | static int compression_decompress(int type, struct list_head *ws, | |
116 | unsigned char *data_in, struct page *dest_page, | |
117 | unsigned long start_byte, size_t srclen, size_t destlen) | |
118 | { | |
119 | switch (type) { | |
120 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
121 | start_byte, srclen, destlen); | |
122 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
123 | start_byte, srclen, destlen); | |
124 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
125 | start_byte, srclen, destlen); | |
126 | case BTRFS_COMPRESS_NONE: | |
127 | default: | |
128 | /* | |
129 | * This can't happen, the type is validated several times | |
130 | * before we get here. | |
131 | */ | |
132 | BUG(); | |
133 | } | |
134 | } | |
135 | ||
8140dc30 | 136 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 137 | |
2ff7e61e | 138 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
139 | unsigned long disk_size) |
140 | { | |
d20f7043 | 141 | return sizeof(struct compressed_bio) + |
713cebfb | 142 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * fs_info->csum_size; |
d20f7043 CM |
143 | } |
144 | ||
5a9472fe | 145 | static int check_compressed_csum(struct btrfs_inode *inode, struct bio *bio, |
d20f7043 CM |
146 | u64 disk_start) |
147 | { | |
10fe6ca8 | 148 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
d5178578 | 149 | SHASH_DESC_ON_STACK(shash, fs_info->csum_shash); |
223486c2 | 150 | const u32 csum_size = fs_info->csum_size; |
04d4ba4c | 151 | const u32 sectorsize = fs_info->sectorsize; |
d20f7043 | 152 | struct page *page; |
1d08ce58 | 153 | unsigned int i; |
d20f7043 | 154 | char *kaddr; |
d5178578 | 155 | u8 csum[BTRFS_CSUM_SIZE]; |
5a9472fe | 156 | struct compressed_bio *cb = bio->bi_private; |
10fe6ca8 | 157 | u8 *cb_sum = cb->sums; |
d20f7043 | 158 | |
42437a63 | 159 | if (!fs_info->csum_root || (inode->flags & BTRFS_INODE_NODATASUM)) |
d20f7043 CM |
160 | return 0; |
161 | ||
d5178578 JT |
162 | shash->tfm = fs_info->csum_shash; |
163 | ||
d20f7043 | 164 | for (i = 0; i < cb->nr_pages; i++) { |
04d4ba4c QW |
165 | u32 pg_offset; |
166 | u32 bytes_left = PAGE_SIZE; | |
d20f7043 | 167 | page = cb->compressed_pages[i]; |
d20f7043 | 168 | |
04d4ba4c QW |
169 | /* Determine the remaining bytes inside the page first */ |
170 | if (i == cb->nr_pages - 1) | |
171 | bytes_left = cb->compressed_len - i * PAGE_SIZE; | |
172 | ||
173 | /* Hash through the page sector by sector */ | |
174 | for (pg_offset = 0; pg_offset < bytes_left; | |
175 | pg_offset += sectorsize) { | |
4c2bf276 | 176 | kaddr = page_address(page); |
04d4ba4c QW |
177 | crypto_shash_digest(shash, kaddr + pg_offset, |
178 | sectorsize, csum); | |
04d4ba4c QW |
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); | |
c3a3b19b | 183 | if (btrfs_bio(bio)->device) |
04d4ba4c | 184 | btrfs_dev_stat_inc_and_print( |
c3a3b19b | 185 | btrfs_bio(bio)->device, |
04d4ba4c QW |
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 | ||
6ec9765d QW |
196 | /* |
197 | * Reduce bio and io accounting for a compressed_bio with its corresponding bio. | |
198 | * | |
199 | * Return true if there is no pending bio nor io. | |
200 | * Return false otherwise. | |
201 | */ | |
202 | static bool dec_and_test_compressed_bio(struct compressed_bio *cb, struct bio *bio) | |
203 | { | |
204 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); | |
205 | unsigned int bi_size = 0; | |
206 | bool last_io = false; | |
207 | struct bio_vec *bvec; | |
208 | struct bvec_iter_all iter_all; | |
209 | ||
210 | /* | |
211 | * At endio time, bi_iter.bi_size doesn't represent the real bio size. | |
212 | * Thus here we have to iterate through all segments to grab correct | |
213 | * bio size. | |
214 | */ | |
215 | bio_for_each_segment_all(bvec, bio, iter_all) | |
216 | bi_size += bvec->bv_len; | |
217 | ||
218 | if (bio->bi_status) | |
219 | cb->errors = 1; | |
220 | ||
221 | ASSERT(bi_size && bi_size <= cb->compressed_len); | |
222 | last_io = refcount_sub_and_test(bi_size >> fs_info->sectorsize_bits, | |
223 | &cb->pending_sectors); | |
86ccbb4d QW |
224 | /* |
225 | * Here we must wake up the possible error handler after all other | |
226 | * operations on @cb finished, or we can race with | |
227 | * finish_compressed_bio_*() which may free @cb. | |
228 | */ | |
229 | wake_up_var(cb); | |
230 | ||
6ec9765d QW |
231 | return last_io; |
232 | } | |
233 | ||
86ccbb4d QW |
234 | static void finish_compressed_bio_read(struct compressed_bio *cb, struct bio *bio) |
235 | { | |
236 | unsigned int index; | |
237 | struct page *page; | |
238 | ||
239 | /* Release the compressed pages */ | |
240 | for (index = 0; index < cb->nr_pages; index++) { | |
241 | page = cb->compressed_pages[index]; | |
242 | page->mapping = NULL; | |
243 | put_page(page); | |
244 | } | |
245 | ||
246 | /* Do io completion on the original bio */ | |
247 | if (cb->errors) { | |
248 | bio_io_error(cb->orig_bio); | |
249 | } else { | |
250 | struct bio_vec *bvec; | |
251 | struct bvec_iter_all iter_all; | |
252 | ||
253 | ASSERT(bio); | |
254 | ASSERT(!bio->bi_status); | |
255 | /* | |
256 | * We have verified the checksum already, set page checked so | |
257 | * the end_io handlers know about it | |
258 | */ | |
259 | ASSERT(!bio_flagged(bio, BIO_CLONED)); | |
260 | bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) { | |
261 | u64 bvec_start = page_offset(bvec->bv_page) + | |
262 | bvec->bv_offset; | |
263 | ||
264 | btrfs_page_set_checked(btrfs_sb(cb->inode->i_sb), | |
265 | bvec->bv_page, bvec_start, | |
266 | bvec->bv_len); | |
267 | } | |
268 | ||
269 | bio_endio(cb->orig_bio); | |
270 | } | |
271 | ||
272 | /* Finally free the cb struct */ | |
273 | kfree(cb->compressed_pages); | |
274 | kfree(cb); | |
275 | } | |
276 | ||
c8b97818 CM |
277 | /* when we finish reading compressed pages from the disk, we |
278 | * decompress them and then run the bio end_io routines on the | |
279 | * decompressed pages (in the inode address space). | |
280 | * | |
281 | * This allows the checksumming and other IO error handling routines | |
282 | * to work normally | |
283 | * | |
284 | * The compressed pages are freed here, and it must be run | |
285 | * in process context | |
286 | */ | |
4246a0b6 | 287 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 288 | { |
c8b97818 CM |
289 | struct compressed_bio *cb = bio->bi_private; |
290 | struct inode *inode; | |
c3a3b19b | 291 | unsigned int mirror = btrfs_bio(bio)->mirror_num; |
e6311f24 | 292 | int ret = 0; |
c8b97818 | 293 | |
6ec9765d | 294 | if (!dec_and_test_compressed_bio(cb, bio)) |
c8b97818 CM |
295 | goto out; |
296 | ||
cf1167d5 LB |
297 | /* |
298 | * Record the correct mirror_num in cb->orig_bio so that | |
299 | * read-repair can work properly. | |
300 | */ | |
c3a3b19b | 301 | btrfs_bio(cb->orig_bio)->mirror_num = mirror; |
cf1167d5 LB |
302 | cb->mirror_num = mirror; |
303 | ||
e6311f24 LB |
304 | /* |
305 | * Some IO in this cb have failed, just skip checksum as there | |
306 | * is no way it could be correct. | |
307 | */ | |
308 | if (cb->errors == 1) | |
309 | goto csum_failed; | |
310 | ||
d20f7043 | 311 | inode = cb->inode; |
5a9472fe | 312 | ret = check_compressed_csum(BTRFS_I(inode), bio, |
1201b58b | 313 | bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
314 | if (ret) |
315 | goto csum_failed; | |
316 | ||
c8b97818 CM |
317 | /* ok, we're the last bio for this extent, lets start |
318 | * the decompression. | |
319 | */ | |
8140dc30 AJ |
320 | ret = btrfs_decompress_bio(cb); |
321 | ||
d20f7043 | 322 | csum_failed: |
c8b97818 CM |
323 | if (ret) |
324 | cb->errors = 1; | |
86ccbb4d | 325 | finish_compressed_bio_read(cb, bio); |
c8b97818 CM |
326 | out: |
327 | bio_put(bio); | |
328 | } | |
329 | ||
330 | /* | |
331 | * Clear the writeback bits on all of the file | |
332 | * pages for a compressed write | |
333 | */ | |
7bdcefc1 FM |
334 | static noinline void end_compressed_writeback(struct inode *inode, |
335 | const struct compressed_bio *cb) | |
c8b97818 | 336 | { |
09cbfeaf KS |
337 | unsigned long index = cb->start >> PAGE_SHIFT; |
338 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
339 | struct page *pages[16]; |
340 | unsigned long nr_pages = end_index - index + 1; | |
341 | int i; | |
342 | int ret; | |
343 | ||
7bdcefc1 FM |
344 | if (cb->errors) |
345 | mapping_set_error(inode->i_mapping, -EIO); | |
346 | ||
d397712b | 347 | while (nr_pages > 0) { |
c8b97818 | 348 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
349 | min_t(unsigned long, |
350 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
351 | if (ret == 0) { |
352 | nr_pages -= 1; | |
353 | index += 1; | |
354 | continue; | |
355 | } | |
356 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
357 | if (cb->errors) |
358 | SetPageError(pages[i]); | |
c8b97818 | 359 | end_page_writeback(pages[i]); |
09cbfeaf | 360 | put_page(pages[i]); |
c8b97818 CM |
361 | } |
362 | nr_pages -= ret; | |
363 | index += ret; | |
364 | } | |
365 | /* the inode may be gone now */ | |
c8b97818 CM |
366 | } |
367 | ||
6853c64a | 368 | static void finish_compressed_bio_write(struct compressed_bio *cb) |
c8b97818 | 369 | { |
6853c64a | 370 | struct inode *inode = cb->inode; |
1d08ce58 | 371 | unsigned int index; |
c8b97818 | 372 | |
6853c64a QW |
373 | /* |
374 | * Ok, we're the last bio for this extent, step one is to call back | |
375 | * into the FS and do all the end_io operations. | |
c8b97818 | 376 | */ |
38a39ac7 | 377 | btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, |
c629732d | 378 | cb->start, cb->start + cb->len - 1, |
240246f6 | 379 | !cb->errors); |
c8b97818 | 380 | |
7bdcefc1 | 381 | end_compressed_writeback(inode, cb); |
6853c64a | 382 | /* Note, our inode could be gone now */ |
c8b97818 CM |
383 | |
384 | /* | |
6853c64a | 385 | * Release the compressed pages, these came from alloc_page and |
c8b97818 CM |
386 | * are not attached to the inode at all |
387 | */ | |
c8b97818 | 388 | for (index = 0; index < cb->nr_pages; index++) { |
6853c64a QW |
389 | struct page *page = cb->compressed_pages[index]; |
390 | ||
c8b97818 | 391 | page->mapping = NULL; |
09cbfeaf | 392 | put_page(page); |
c8b97818 CM |
393 | } |
394 | ||
6853c64a | 395 | /* Finally free the cb struct */ |
c8b97818 CM |
396 | kfree(cb->compressed_pages); |
397 | kfree(cb); | |
6853c64a QW |
398 | } |
399 | ||
400 | /* | |
401 | * Do the cleanup once all the compressed pages hit the disk. This will clear | |
402 | * writeback on the file pages and free the compressed pages. | |
403 | * | |
404 | * This also calls the writeback end hooks for the file pages so that metadata | |
405 | * and checksums can be updated in the file. | |
406 | */ | |
407 | static void end_compressed_bio_write(struct bio *bio) | |
408 | { | |
409 | struct compressed_bio *cb = bio->bi_private; | |
410 | ||
411 | if (!dec_and_test_compressed_bio(cb, bio)) | |
412 | goto out; | |
413 | ||
414 | btrfs_record_physical_zoned(cb->inode, cb->start, bio); | |
415 | ||
416 | finish_compressed_bio_write(cb); | |
c8b97818 CM |
417 | out: |
418 | bio_put(bio); | |
419 | } | |
420 | ||
2d4e0b84 QW |
421 | static blk_status_t submit_compressed_bio(struct btrfs_fs_info *fs_info, |
422 | struct compressed_bio *cb, | |
423 | struct bio *bio, int mirror_num) | |
424 | { | |
425 | blk_status_t ret; | |
426 | ||
427 | ASSERT(bio->bi_iter.bi_size); | |
2d4e0b84 QW |
428 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
429 | if (ret) | |
430 | return ret; | |
431 | ret = btrfs_map_bio(fs_info, bio, mirror_num); | |
432 | return ret; | |
433 | } | |
434 | ||
22c306fe | 435 | /* |
f472c28f QW |
436 | * Allocate a compressed_bio, which will be used to read/write on-disk |
437 | * (aka, compressed) * data. | |
438 | * | |
439 | * @cb: The compressed_bio structure, which records all the needed | |
440 | * information to bind the compressed data to the uncompressed | |
441 | * page cache. | |
442 | * @disk_byten: The logical bytenr where the compressed data will be read | |
443 | * from or written to. | |
444 | * @endio_func: The endio function to call after the IO for compressed data | |
445 | * is finished. | |
446 | * @next_stripe_start: Return value of logical bytenr of where next stripe starts. | |
447 | * Let the caller know to only fill the bio up to the stripe | |
448 | * boundary. | |
22c306fe | 449 | */ |
f472c28f QW |
450 | |
451 | ||
22c306fe | 452 | static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr, |
f472c28f QW |
453 | unsigned int opf, bio_end_io_t endio_func, |
454 | u64 *next_stripe_start) | |
22c306fe | 455 | { |
f472c28f QW |
456 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
457 | struct btrfs_io_geometry geom; | |
458 | struct extent_map *em; | |
22c306fe | 459 | struct bio *bio; |
f472c28f | 460 | int ret; |
22c306fe QW |
461 | |
462 | bio = btrfs_bio_alloc(BIO_MAX_VECS); | |
463 | ||
464 | bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT; | |
465 | bio->bi_opf = opf; | |
466 | bio->bi_private = cb; | |
467 | bio->bi_end_io = endio_func; | |
468 | ||
f472c28f QW |
469 | em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize); |
470 | if (IS_ERR(em)) { | |
471 | bio_put(bio); | |
472 | return ERR_CAST(em); | |
473 | } | |
22c306fe | 474 | |
f472c28f QW |
475 | if (bio_op(bio) == REQ_OP_ZONE_APPEND) |
476 | bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev); | |
477 | ||
478 | ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom); | |
479 | free_extent_map(em); | |
480 | if (ret < 0) { | |
481 | bio_put(bio); | |
482 | return ERR_PTR(ret); | |
22c306fe | 483 | } |
f472c28f QW |
484 | *next_stripe_start = disk_bytenr + geom.len; |
485 | ||
22c306fe QW |
486 | return bio; |
487 | } | |
488 | ||
c8b97818 CM |
489 | /* |
490 | * worker function to build and submit bios for previously compressed pages. | |
491 | * The corresponding pages in the inode should be marked for writeback | |
492 | * and the compressed pages should have a reference on them for dropping | |
493 | * when the IO is complete. | |
494 | * | |
495 | * This also checksums the file bytes and gets things ready for | |
496 | * the end io hooks. | |
497 | */ | |
c7ee1819 | 498 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
65b5355f AJ |
499 | unsigned int len, u64 disk_start, |
500 | unsigned int compressed_len, | |
c8b97818 | 501 | struct page **compressed_pages, |
65b5355f | 502 | unsigned int nr_pages, |
ec39f769 CM |
503 | unsigned int write_flags, |
504 | struct cgroup_subsys_state *blkcg_css) | |
c8b97818 | 505 | { |
c7ee1819 | 506 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 507 | struct bio *bio = NULL; |
c8b97818 | 508 | struct compressed_bio *cb; |
91507240 | 509 | u64 cur_disk_bytenr = disk_start; |
f472c28f | 510 | u64 next_stripe_start; |
4e4cbee9 | 511 | blk_status_t ret; |
c7ee1819 | 512 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
764c7c9a JT |
513 | const bool use_append = btrfs_use_zone_append(inode, disk_start); |
514 | const unsigned int bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; | |
c8b97818 | 515 | |
bbbff01a QW |
516 | ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && |
517 | IS_ALIGNED(len, fs_info->sectorsize)); | |
2ff7e61e | 518 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 519 | if (!cb) |
4e4cbee9 | 520 | return BLK_STS_RESOURCE; |
6ec9765d | 521 | refcount_set(&cb->pending_sectors, compressed_len >> fs_info->sectorsize_bits); |
c8b97818 | 522 | cb->errors = 0; |
c7ee1819 | 523 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
524 | cb->start = start; |
525 | cb->len = len; | |
d20f7043 | 526 | cb->mirror_num = 0; |
c8b97818 CM |
527 | cb->compressed_pages = compressed_pages; |
528 | cb->compressed_len = compressed_len; | |
529 | cb->orig_bio = NULL; | |
530 | cb->nr_pages = nr_pages; | |
531 | ||
91507240 QW |
532 | while (cur_disk_bytenr < disk_start + compressed_len) { |
533 | u64 offset = cur_disk_bytenr - disk_start; | |
534 | unsigned int index = offset >> PAGE_SHIFT; | |
535 | unsigned int real_size; | |
536 | unsigned int added; | |
537 | struct page *page = compressed_pages[index]; | |
538 | bool submit = false; | |
539 | ||
540 | /* Allocate new bio if submitted or not yet allocated */ | |
541 | if (!bio) { | |
542 | bio = alloc_compressed_bio(cb, cur_disk_bytenr, | |
543 | bio_op | write_flags, end_compressed_bio_write, | |
544 | &next_stripe_start); | |
545 | if (IS_ERR(bio)) { | |
546 | ret = errno_to_blk_status(PTR_ERR(bio)); | |
547 | bio = NULL; | |
548 | goto finish_cb; | |
549 | } | |
550 | } | |
4c80a97d | 551 | /* |
91507240 QW |
552 | * We should never reach next_stripe_start start as we will |
553 | * submit comp_bio when reach the boundary immediately. | |
4c80a97d | 554 | */ |
91507240 | 555 | ASSERT(cur_disk_bytenr != next_stripe_start); |
764c7c9a | 556 | |
91507240 QW |
557 | /* |
558 | * We have various limits on the real read size: | |
559 | * - stripe boundary | |
560 | * - page boundary | |
561 | * - compressed length boundary | |
562 | */ | |
563 | real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr); | |
564 | real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset)); | |
565 | real_size = min_t(u64, real_size, compressed_len - offset); | |
566 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
567 | ||
568 | if (use_append) | |
569 | added = bio_add_zone_append_page(bio, page, real_size, | |
570 | offset_in_page(offset)); | |
571 | else | |
572 | added = bio_add_page(bio, page, real_size, | |
573 | offset_in_page(offset)); | |
574 | /* Reached zoned boundary */ | |
575 | if (added == 0) | |
576 | submit = true; | |
577 | ||
578 | cur_disk_bytenr += added; | |
579 | /* Reached stripe boundary */ | |
580 | if (cur_disk_bytenr == next_stripe_start) | |
581 | submit = true; | |
582 | ||
583 | /* Finished the range */ | |
584 | if (cur_disk_bytenr == disk_start + compressed_len) | |
585 | submit = true; | |
586 | ||
587 | if (submit) { | |
e55179b3 | 588 | if (!skip_sum) { |
c7ee1819 | 589 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
6853c64a QW |
590 | if (ret) |
591 | goto finish_cb; | |
e55179b3 | 592 | } |
d20f7043 | 593 | |
2d4e0b84 | 594 | ret = submit_compressed_bio(fs_info, cb, bio, 0); |
6853c64a QW |
595 | if (ret) |
596 | goto finish_cb; | |
91507240 | 597 | bio = NULL; |
c8b97818 | 598 | } |
771ed689 | 599 | cond_resched(); |
c8b97818 | 600 | } |
46bcff2b DZ |
601 | if (blkcg_css) |
602 | kthread_associate_blkcg(NULL); | |
603 | ||
c8b97818 | 604 | return 0; |
6853c64a QW |
605 | |
606 | finish_cb: | |
607 | if (bio) { | |
608 | bio->bi_status = ret; | |
609 | bio_endio(bio); | |
610 | } | |
91507240 QW |
611 | /* Last byte of @cb is submitted, endio will free @cb */ |
612 | if (cur_disk_bytenr == disk_start + compressed_len) | |
613 | return ret; | |
6853c64a | 614 | |
91507240 QW |
615 | wait_var_event(cb, refcount_read(&cb->pending_sectors) == |
616 | (disk_start + compressed_len - cur_disk_bytenr) >> | |
617 | fs_info->sectorsize_bits); | |
6853c64a QW |
618 | /* |
619 | * Even with previous bio ended, we should still have io not yet | |
620 | * submitted, thus need to finish manually. | |
621 | */ | |
622 | ASSERT(refcount_read(&cb->pending_sectors)); | |
623 | /* Now we are the only one referring @cb, can finish it safely. */ | |
624 | finish_compressed_bio_write(cb); | |
625 | return ret; | |
c8b97818 CM |
626 | } |
627 | ||
2a4d0c90 CH |
628 | static u64 bio_end_offset(struct bio *bio) |
629 | { | |
c45a8f2d | 630 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
631 | |
632 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
633 | } | |
634 | ||
6a404910 QW |
635 | /* |
636 | * Add extra pages in the same compressed file extent so that we don't need to | |
637 | * re-read the same extent again and again. | |
638 | * | |
639 | * NOTE: this won't work well for subpage, as for subpage read, we lock the | |
640 | * full page then submit bio for each compressed/regular extents. | |
641 | * | |
642 | * This means, if we have several sectors in the same page points to the same | |
643 | * on-disk compressed data, we will re-read the same extent many times and | |
644 | * this function can only help for the next page. | |
645 | */ | |
771ed689 CM |
646 | static noinline int add_ra_bio_pages(struct inode *inode, |
647 | u64 compressed_end, | |
648 | struct compressed_bio *cb) | |
649 | { | |
6a404910 | 650 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
771ed689 | 651 | unsigned long end_index; |
6a404910 | 652 | u64 cur = bio_end_offset(cb->orig_bio); |
771ed689 CM |
653 | u64 isize = i_size_read(inode); |
654 | int ret; | |
655 | struct page *page; | |
771ed689 CM |
656 | struct extent_map *em; |
657 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
658 | struct extent_map_tree *em_tree; |
659 | struct extent_io_tree *tree; | |
6a404910 | 660 | int sectors_missed = 0; |
771ed689 | 661 | |
771ed689 CM |
662 | em_tree = &BTRFS_I(inode)->extent_tree; |
663 | tree = &BTRFS_I(inode)->io_tree; | |
664 | ||
665 | if (isize == 0) | |
666 | return 0; | |
667 | ||
ca62e85d QW |
668 | /* |
669 | * For current subpage support, we only support 64K page size, | |
670 | * which means maximum compressed extent size (128K) is just 2x page | |
671 | * size. | |
672 | * This makes readahead less effective, so here disable readahead for | |
673 | * subpage for now, until full compressed write is supported. | |
674 | */ | |
675 | if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) | |
676 | return 0; | |
677 | ||
09cbfeaf | 678 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 679 | |
6a404910 QW |
680 | while (cur < compressed_end) { |
681 | u64 page_end; | |
682 | u64 pg_index = cur >> PAGE_SHIFT; | |
683 | u32 add_size; | |
771ed689 | 684 | |
306e16ce | 685 | if (pg_index > end_index) |
771ed689 CM |
686 | break; |
687 | ||
0a943c65 | 688 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 689 | if (page && !xa_is_value(page)) { |
6a404910 QW |
690 | sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> |
691 | fs_info->sectorsize_bits; | |
692 | ||
693 | /* Beyond threshold, no need to continue */ | |
694 | if (sectors_missed > 4) | |
771ed689 | 695 | break; |
6a404910 QW |
696 | |
697 | /* | |
698 | * Jump to next page start as we already have page for | |
699 | * current offset. | |
700 | */ | |
701 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
702 | continue; | |
771ed689 CM |
703 | } |
704 | ||
c62d2555 MH |
705 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
706 | ~__GFP_FS)); | |
771ed689 CM |
707 | if (!page) |
708 | break; | |
709 | ||
c62d2555 | 710 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 711 | put_page(page); |
6a404910 QW |
712 | /* There is already a page, skip to page end */ |
713 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
714 | continue; | |
771ed689 CM |
715 | } |
716 | ||
32443de3 QW |
717 | ret = set_page_extent_mapped(page); |
718 | if (ret < 0) { | |
719 | unlock_page(page); | |
720 | put_page(page); | |
721 | break; | |
722 | } | |
723 | ||
6a404910 QW |
724 | page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; |
725 | lock_extent(tree, cur, page_end); | |
890871be | 726 | read_lock(&em_tree->lock); |
6a404910 | 727 | em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); |
890871be | 728 | read_unlock(&em_tree->lock); |
771ed689 | 729 | |
6a404910 QW |
730 | /* |
731 | * At this point, we have a locked page in the page cache for | |
732 | * these bytes in the file. But, we have to make sure they map | |
733 | * to this compressed extent on disk. | |
734 | */ | |
735 | if (!em || cur < em->start || | |
736 | (cur + fs_info->sectorsize > extent_map_end(em)) || | |
4f024f37 | 737 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 738 | free_extent_map(em); |
6a404910 | 739 | unlock_extent(tree, cur, page_end); |
771ed689 | 740 | unlock_page(page); |
09cbfeaf | 741 | put_page(page); |
771ed689 CM |
742 | break; |
743 | } | |
744 | free_extent_map(em); | |
745 | ||
746 | if (page->index == end_index) { | |
7073017a | 747 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
748 | |
749 | if (zero_offset) { | |
750 | int zeros; | |
09cbfeaf | 751 | zeros = PAGE_SIZE - zero_offset; |
d048b9c2 | 752 | memzero_page(page, zero_offset, zeros); |
771ed689 | 753 | flush_dcache_page(page); |
771ed689 CM |
754 | } |
755 | } | |
756 | ||
6a404910 QW |
757 | add_size = min(em->start + em->len, page_end + 1) - cur; |
758 | ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur)); | |
759 | if (ret != add_size) { | |
760 | unlock_extent(tree, cur, page_end); | |
771ed689 | 761 | unlock_page(page); |
09cbfeaf | 762 | put_page(page); |
771ed689 CM |
763 | break; |
764 | } | |
6a404910 QW |
765 | /* |
766 | * If it's subpage, we also need to increase its | |
767 | * subpage::readers number, as at endio we will decrease | |
768 | * subpage::readers and to unlock the page. | |
769 | */ | |
770 | if (fs_info->sectorsize < PAGE_SIZE) | |
771 | btrfs_subpage_start_reader(fs_info, page, cur, add_size); | |
772 | put_page(page); | |
773 | cur += add_size; | |
771ed689 | 774 | } |
771ed689 CM |
775 | return 0; |
776 | } | |
777 | ||
c8b97818 CM |
778 | /* |
779 | * for a compressed read, the bio we get passed has all the inode pages | |
780 | * in it. We don't actually do IO on those pages but allocate new ones | |
781 | * to hold the compressed pages on disk. | |
782 | * | |
4f024f37 | 783 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 784 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
785 | * |
786 | * After the compressed pages are read, we copy the bytes into the | |
787 | * bio we were passed and then call the bio end_io calls | |
788 | */ | |
4e4cbee9 | 789 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
790 | int mirror_num, unsigned long bio_flags) |
791 | { | |
0b246afa | 792 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
793 | struct extent_map_tree *em_tree; |
794 | struct compressed_bio *cb; | |
356b4a2d AJ |
795 | unsigned int compressed_len; |
796 | unsigned int nr_pages; | |
797 | unsigned int pg_index; | |
f472c28f QW |
798 | struct bio *comp_bio = NULL; |
799 | const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT; | |
800 | u64 cur_disk_byte = disk_bytenr; | |
801 | u64 next_stripe_start; | |
557023ea | 802 | u64 file_offset; |
e04ca626 CM |
803 | u64 em_len; |
804 | u64 em_start; | |
c8b97818 | 805 | struct extent_map *em; |
4e4cbee9 | 806 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 807 | int faili = 0; |
10fe6ca8 | 808 | u8 *sums; |
c8b97818 | 809 | |
c8b97818 CM |
810 | em_tree = &BTRFS_I(inode)->extent_tree; |
811 | ||
557023ea QW |
812 | file_offset = bio_first_bvec_all(bio)->bv_offset + |
813 | page_offset(bio_first_page_all(bio)); | |
814 | ||
c8b97818 | 815 | /* we need the actual starting offset of this extent in the file */ |
890871be | 816 | read_lock(&em_tree->lock); |
557023ea | 817 | em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); |
890871be | 818 | read_unlock(&em_tree->lock); |
285190d9 | 819 | if (!em) |
4e4cbee9 | 820 | return BLK_STS_IOERR; |
c8b97818 | 821 | |
557023ea | 822 | ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); |
d20f7043 | 823 | compressed_len = em->block_len; |
2ff7e61e | 824 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 825 | if (!cb) |
826 | goto out; | |
827 | ||
6ec9765d | 828 | refcount_set(&cb->pending_sectors, compressed_len >> fs_info->sectorsize_bits); |
c8b97818 CM |
829 | cb->errors = 0; |
830 | cb->inode = inode; | |
d20f7043 | 831 | cb->mirror_num = mirror_num; |
10fe6ca8 | 832 | sums = cb->sums; |
c8b97818 | 833 | |
ff5b7ee3 | 834 | cb->start = em->orig_start; |
e04ca626 CM |
835 | em_len = em->len; |
836 | em_start = em->start; | |
d20f7043 | 837 | |
c8b97818 | 838 | free_extent_map(em); |
e04ca626 | 839 | em = NULL; |
c8b97818 | 840 | |
81381053 | 841 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 842 | cb->compressed_len = compressed_len; |
261507a0 | 843 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
844 | cb->orig_bio = bio; |
845 | ||
09cbfeaf | 846 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 847 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 848 | GFP_NOFS); |
6b82ce8d | 849 | if (!cb->compressed_pages) |
850 | goto fail1; | |
851 | ||
306e16ce | 852 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
b0ee5e1e | 853 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS); |
15e3004a JB |
854 | if (!cb->compressed_pages[pg_index]) { |
855 | faili = pg_index - 1; | |
0e9350de | 856 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 857 | goto fail2; |
15e3004a | 858 | } |
c8b97818 | 859 | } |
15e3004a | 860 | faili = nr_pages - 1; |
c8b97818 CM |
861 | cb->nr_pages = nr_pages; |
862 | ||
7f042a83 | 863 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 864 | |
771ed689 | 865 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 866 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 867 | |
f472c28f QW |
868 | while (cur_disk_byte < disk_bytenr + compressed_len) { |
869 | u64 offset = cur_disk_byte - disk_bytenr; | |
870 | unsigned int index = offset >> PAGE_SHIFT; | |
871 | unsigned int real_size; | |
872 | unsigned int added; | |
873 | struct page *page = cb->compressed_pages[index]; | |
874 | bool submit = false; | |
c8b97818 | 875 | |
f472c28f QW |
876 | /* Allocate new bio if submitted or not yet allocated */ |
877 | if (!comp_bio) { | |
878 | comp_bio = alloc_compressed_bio(cb, cur_disk_byte, | |
879 | REQ_OP_READ, end_compressed_bio_read, | |
880 | &next_stripe_start); | |
881 | if (IS_ERR(comp_bio)) { | |
882 | ret = errno_to_blk_status(PTR_ERR(comp_bio)); | |
883 | comp_bio = NULL; | |
884 | goto finish_cb; | |
885 | } | |
886 | } | |
887 | /* | |
888 | * We should never reach next_stripe_start start as we will | |
889 | * submit comp_bio when reach the boundary immediately. | |
890 | */ | |
891 | ASSERT(cur_disk_byte != next_stripe_start); | |
892 | /* | |
893 | * We have various limit on the real read size: | |
894 | * - stripe boundary | |
895 | * - page boundary | |
896 | * - compressed length boundary | |
897 | */ | |
898 | real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte); | |
899 | real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset)); | |
900 | real_size = min_t(u64, real_size, compressed_len - offset); | |
901 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
4e4cbee9 | 902 | |
f472c28f | 903 | added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset)); |
be6a1361 | 904 | /* |
f472c28f QW |
905 | * Maximum compressed extent is smaller than bio size limit, |
906 | * thus bio_add_page() should always success. | |
be6a1361 | 907 | */ |
f472c28f QW |
908 | ASSERT(added == real_size); |
909 | cur_disk_byte += added; | |
be6a1361 | 910 | |
f472c28f QW |
911 | /* Reached stripe boundary, need to submit */ |
912 | if (cur_disk_byte == next_stripe_start) | |
913 | submit = true; | |
d20f7043 | 914 | |
f472c28f QW |
915 | /* Has finished the range, need to submit */ |
916 | if (cur_disk_byte == disk_bytenr + compressed_len) | |
917 | submit = true; | |
c8b97818 | 918 | |
f472c28f | 919 | if (submit) { |
10fe6ca8 JT |
920 | unsigned int nr_sectors; |
921 | ||
6275193e | 922 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
86ccbb4d QW |
923 | if (ret) |
924 | goto finish_cb; | |
10fe6ca8 JT |
925 | |
926 | nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, | |
927 | fs_info->sectorsize); | |
713cebfb | 928 | sums += fs_info->csum_size * nr_sectors; |
d20f7043 | 929 | |
2d4e0b84 | 930 | ret = submit_compressed_bio(fs_info, cb, comp_bio, mirror_num); |
86ccbb4d QW |
931 | if (ret) |
932 | goto finish_cb; | |
f472c28f | 933 | comp_bio = NULL; |
c8b97818 | 934 | } |
c8b97818 | 935 | } |
c8b97818 | 936 | return 0; |
6b82ce8d | 937 | |
938 | fail2: | |
15e3004a JB |
939 | while (faili >= 0) { |
940 | __free_page(cb->compressed_pages[faili]); | |
941 | faili--; | |
942 | } | |
6b82ce8d | 943 | |
944 | kfree(cb->compressed_pages); | |
945 | fail1: | |
946 | kfree(cb); | |
947 | out: | |
948 | free_extent_map(em); | |
949 | return ret; | |
86ccbb4d QW |
950 | finish_cb: |
951 | if (comp_bio) { | |
952 | comp_bio->bi_status = ret; | |
953 | bio_endio(comp_bio); | |
954 | } | |
f472c28f QW |
955 | /* All bytes of @cb is submitted, endio will free @cb */ |
956 | if (cur_disk_byte == disk_bytenr + compressed_len) | |
957 | return ret; | |
958 | ||
959 | wait_var_event(cb, refcount_read(&cb->pending_sectors) == | |
960 | (disk_bytenr + compressed_len - cur_disk_byte) >> | |
961 | fs_info->sectorsize_bits); | |
86ccbb4d QW |
962 | /* |
963 | * Even with previous bio ended, we should still have io not yet | |
964 | * submitted, thus need to finish @cb manually. | |
965 | */ | |
966 | ASSERT(refcount_read(&cb->pending_sectors)); | |
967 | /* Now we are the only one referring @cb, can finish it safely. */ | |
968 | finish_compressed_bio_read(cb, NULL); | |
969 | return ret; | |
c8b97818 | 970 | } |
261507a0 | 971 | |
17b5a6c1 TT |
972 | /* |
973 | * Heuristic uses systematic sampling to collect data from the input data | |
974 | * range, the logic can be tuned by the following constants: | |
975 | * | |
976 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
977 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
978 | */ | |
979 | #define SAMPLING_READ_SIZE (16) | |
980 | #define SAMPLING_INTERVAL (256) | |
981 | ||
982 | /* | |
983 | * For statistical analysis of the input data we consider bytes that form a | |
984 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
985 | * many times the object appeared in the sample. | |
986 | */ | |
987 | #define BUCKET_SIZE (256) | |
988 | ||
989 | /* | |
990 | * The size of the sample is based on a statistical sampling rule of thumb. | |
991 | * The common way is to perform sampling tests as long as the number of | |
992 | * elements in each cell is at least 5. | |
993 | * | |
994 | * Instead of 5, we choose 32 to obtain more accurate results. | |
995 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
996 | * sample size bound by 8192. | |
997 | * | |
998 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
999 | * from up to 512 locations. | |
1000 | */ | |
1001 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
1002 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
1003 | ||
1004 | struct bucket_item { | |
1005 | u32 count; | |
1006 | }; | |
4e439a0b TT |
1007 | |
1008 | struct heuristic_ws { | |
17b5a6c1 TT |
1009 | /* Partial copy of input data */ |
1010 | u8 *sample; | |
a440d48c | 1011 | u32 sample_size; |
17b5a6c1 TT |
1012 | /* Buckets store counters for each byte value */ |
1013 | struct bucket_item *bucket; | |
440c840c TT |
1014 | /* Sorting buffer */ |
1015 | struct bucket_item *bucket_b; | |
4e439a0b TT |
1016 | struct list_head list; |
1017 | }; | |
1018 | ||
92ee5530 DZ |
1019 | static struct workspace_manager heuristic_wsm; |
1020 | ||
4e439a0b TT |
1021 | static void free_heuristic_ws(struct list_head *ws) |
1022 | { | |
1023 | struct heuristic_ws *workspace; | |
1024 | ||
1025 | workspace = list_entry(ws, struct heuristic_ws, list); | |
1026 | ||
17b5a6c1 TT |
1027 | kvfree(workspace->sample); |
1028 | kfree(workspace->bucket); | |
440c840c | 1029 | kfree(workspace->bucket_b); |
4e439a0b TT |
1030 | kfree(workspace); |
1031 | } | |
1032 | ||
7bf49943 | 1033 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
1034 | { |
1035 | struct heuristic_ws *ws; | |
1036 | ||
1037 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
1038 | if (!ws) | |
1039 | return ERR_PTR(-ENOMEM); | |
1040 | ||
17b5a6c1 TT |
1041 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
1042 | if (!ws->sample) | |
1043 | goto fail; | |
1044 | ||
1045 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
1046 | if (!ws->bucket) | |
1047 | goto fail; | |
4e439a0b | 1048 | |
440c840c TT |
1049 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
1050 | if (!ws->bucket_b) | |
1051 | goto fail; | |
1052 | ||
17b5a6c1 | 1053 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 1054 | return &ws->list; |
17b5a6c1 TT |
1055 | fail: |
1056 | free_heuristic_ws(&ws->list); | |
1057 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
1058 | } |
1059 | ||
ca4ac360 | 1060 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 1061 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
1062 | }; |
1063 | ||
e8c9f186 | 1064 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
1065 | /* The heuristic is represented as compression type 0 */ |
1066 | &btrfs_heuristic_compress, | |
261507a0 | 1067 | &btrfs_zlib_compress, |
a6fa6fae | 1068 | &btrfs_lzo_compress, |
5c1aab1d | 1069 | &btrfs_zstd_compress, |
261507a0 LZ |
1070 | }; |
1071 | ||
c778df14 DS |
1072 | static struct list_head *alloc_workspace(int type, unsigned int level) |
1073 | { | |
1074 | switch (type) { | |
1075 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
1076 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
1077 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
1078 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
1079 | default: | |
1080 | /* | |
1081 | * This can't happen, the type is validated several times | |
1082 | * before we get here. | |
1083 | */ | |
1084 | BUG(); | |
1085 | } | |
1086 | } | |
1087 | ||
1e002351 DS |
1088 | static void free_workspace(int type, struct list_head *ws) |
1089 | { | |
1090 | switch (type) { | |
1091 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
1092 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
1093 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
1094 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
1095 | default: | |
1096 | /* | |
1097 | * This can't happen, the type is validated several times | |
1098 | * before we get here. | |
1099 | */ | |
1100 | BUG(); | |
1101 | } | |
1102 | } | |
1103 | ||
d5517033 | 1104 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 1105 | { |
0cf25213 | 1106 | struct workspace_manager *wsm; |
4e439a0b | 1107 | struct list_head *workspace; |
261507a0 | 1108 | |
0cf25213 | 1109 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1110 | INIT_LIST_HEAD(&wsm->idle_ws); |
1111 | spin_lock_init(&wsm->ws_lock); | |
1112 | atomic_set(&wsm->total_ws, 0); | |
1113 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 1114 | |
1666edab DZ |
1115 | /* |
1116 | * Preallocate one workspace for each compression type so we can | |
1117 | * guarantee forward progress in the worst case | |
1118 | */ | |
c778df14 | 1119 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
1120 | if (IS_ERR(workspace)) { |
1121 | pr_warn( | |
1122 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
1123 | } else { | |
92ee5530 DZ |
1124 | atomic_set(&wsm->total_ws, 1); |
1125 | wsm->free_ws = 1; | |
1126 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
1127 | } |
1128 | } | |
1129 | ||
2510307e | 1130 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 1131 | { |
2dba7143 | 1132 | struct workspace_manager *wsman; |
1666edab DZ |
1133 | struct list_head *ws; |
1134 | ||
2dba7143 | 1135 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
1136 | while (!list_empty(&wsman->idle_ws)) { |
1137 | ws = wsman->idle_ws.next; | |
1138 | list_del(ws); | |
1e002351 | 1139 | free_workspace(type, ws); |
1666edab | 1140 | atomic_dec(&wsman->total_ws); |
261507a0 | 1141 | } |
261507a0 LZ |
1142 | } |
1143 | ||
1144 | /* | |
e721e49d DS |
1145 | * This finds an available workspace or allocates a new one. |
1146 | * If it's not possible to allocate a new one, waits until there's one. | |
1147 | * Preallocation makes a forward progress guarantees and we do not return | |
1148 | * errors. | |
261507a0 | 1149 | */ |
5907a9bb | 1150 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 1151 | { |
5907a9bb | 1152 | struct workspace_manager *wsm; |
261507a0 LZ |
1153 | struct list_head *workspace; |
1154 | int cpus = num_online_cpus(); | |
fe308533 | 1155 | unsigned nofs_flag; |
4e439a0b TT |
1156 | struct list_head *idle_ws; |
1157 | spinlock_t *ws_lock; | |
1158 | atomic_t *total_ws; | |
1159 | wait_queue_head_t *ws_wait; | |
1160 | int *free_ws; | |
1161 | ||
5907a9bb | 1162 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1163 | idle_ws = &wsm->idle_ws; |
1164 | ws_lock = &wsm->ws_lock; | |
1165 | total_ws = &wsm->total_ws; | |
1166 | ws_wait = &wsm->ws_wait; | |
1167 | free_ws = &wsm->free_ws; | |
261507a0 | 1168 | |
261507a0 | 1169 | again: |
d9187649 BL |
1170 | spin_lock(ws_lock); |
1171 | if (!list_empty(idle_ws)) { | |
1172 | workspace = idle_ws->next; | |
261507a0 | 1173 | list_del(workspace); |
6ac10a6a | 1174 | (*free_ws)--; |
d9187649 | 1175 | spin_unlock(ws_lock); |
261507a0 LZ |
1176 | return workspace; |
1177 | ||
1178 | } | |
6ac10a6a | 1179 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
1180 | DEFINE_WAIT(wait); |
1181 | ||
d9187649 BL |
1182 | spin_unlock(ws_lock); |
1183 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 1184 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 1185 | schedule(); |
d9187649 | 1186 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
1187 | goto again; |
1188 | } | |
6ac10a6a | 1189 | atomic_inc(total_ws); |
d9187649 | 1190 | spin_unlock(ws_lock); |
261507a0 | 1191 | |
fe308533 DS |
1192 | /* |
1193 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
1194 | * to turn it off here because we might get called from the restricted | |
1195 | * context of btrfs_compress_bio/btrfs_compress_pages | |
1196 | */ | |
1197 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1198 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1199 | memalloc_nofs_restore(nofs_flag); |
1200 | ||
261507a0 | 1201 | if (IS_ERR(workspace)) { |
6ac10a6a | 1202 | atomic_dec(total_ws); |
d9187649 | 1203 | wake_up(ws_wait); |
e721e49d DS |
1204 | |
1205 | /* | |
1206 | * Do not return the error but go back to waiting. There's a | |
1207 | * workspace preallocated for each type and the compression | |
1208 | * time is bounded so we get to a workspace eventually. This | |
1209 | * makes our caller's life easier. | |
52356716 DS |
1210 | * |
1211 | * To prevent silent and low-probability deadlocks (when the | |
1212 | * initial preallocation fails), check if there are any | |
1213 | * workspaces at all. | |
e721e49d | 1214 | */ |
52356716 DS |
1215 | if (atomic_read(total_ws) == 0) { |
1216 | static DEFINE_RATELIMIT_STATE(_rs, | |
1217 | /* once per minute */ 60 * HZ, | |
1218 | /* no burst */ 1); | |
1219 | ||
1220 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1221 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1222 | } |
1223 | } | |
e721e49d | 1224 | goto again; |
261507a0 LZ |
1225 | } |
1226 | return workspace; | |
1227 | } | |
1228 | ||
7bf49943 | 1229 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1230 | { |
6a0d1272 | 1231 | switch (type) { |
5907a9bb | 1232 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1233 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1234 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1235 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1236 | default: | |
1237 | /* | |
1238 | * This can't happen, the type is validated several times | |
1239 | * before we get here. | |
1240 | */ | |
1241 | BUG(); | |
1242 | } | |
929f4baf DZ |
1243 | } |
1244 | ||
261507a0 LZ |
1245 | /* |
1246 | * put a workspace struct back on the list or free it if we have enough | |
1247 | * idle ones sitting around | |
1248 | */ | |
a3bbd2a9 | 1249 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1250 | { |
a3bbd2a9 | 1251 | struct workspace_manager *wsm; |
4e439a0b TT |
1252 | struct list_head *idle_ws; |
1253 | spinlock_t *ws_lock; | |
1254 | atomic_t *total_ws; | |
1255 | wait_queue_head_t *ws_wait; | |
1256 | int *free_ws; | |
1257 | ||
a3bbd2a9 | 1258 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1259 | idle_ws = &wsm->idle_ws; |
1260 | ws_lock = &wsm->ws_lock; | |
1261 | total_ws = &wsm->total_ws; | |
1262 | ws_wait = &wsm->ws_wait; | |
1263 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1264 | |
1265 | spin_lock(ws_lock); | |
26b28dce | 1266 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1267 | list_add(ws, idle_ws); |
6ac10a6a | 1268 | (*free_ws)++; |
d9187649 | 1269 | spin_unlock(ws_lock); |
261507a0 LZ |
1270 | goto wake; |
1271 | } | |
d9187649 | 1272 | spin_unlock(ws_lock); |
261507a0 | 1273 | |
1e002351 | 1274 | free_workspace(type, ws); |
6ac10a6a | 1275 | atomic_dec(total_ws); |
261507a0 | 1276 | wake: |
093258e6 | 1277 | cond_wake_up(ws_wait); |
261507a0 LZ |
1278 | } |
1279 | ||
929f4baf DZ |
1280 | static void put_workspace(int type, struct list_head *ws) |
1281 | { | |
bd3a5287 | 1282 | switch (type) { |
a3bbd2a9 DS |
1283 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1284 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1285 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1286 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1287 | default: | |
1288 | /* | |
1289 | * This can't happen, the type is validated several times | |
1290 | * before we get here. | |
1291 | */ | |
1292 | BUG(); | |
1293 | } | |
929f4baf DZ |
1294 | } |
1295 | ||
adbab642 AJ |
1296 | /* |
1297 | * Adjust @level according to the limits of the compression algorithm or | |
1298 | * fallback to default | |
1299 | */ | |
1300 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1301 | { | |
1302 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1303 | ||
1304 | if (level == 0) | |
1305 | level = ops->default_level; | |
1306 | else | |
1307 | level = min(level, ops->max_level); | |
1308 | ||
1309 | return level; | |
1310 | } | |
1311 | ||
261507a0 | 1312 | /* |
38c31464 DS |
1313 | * Given an address space and start and length, compress the bytes into @pages |
1314 | * that are allocated on demand. | |
261507a0 | 1315 | * |
f51d2b59 DS |
1316 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1317 | * default the algorithm chooses and is opaque here; | |
1318 | * - compression algo are 0-3 | |
1319 | * - the level are bits 4-7 | |
1320 | * | |
4d3a800e DS |
1321 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1322 | * and returns number of actually allocated pages | |
261507a0 | 1323 | * |
38c31464 DS |
1324 | * @total_in is used to return the number of bytes actually read. It |
1325 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1326 | * ran out of room in the pages array or because we cross the |
1327 | * max_out threshold. | |
1328 | * | |
38c31464 DS |
1329 | * @total_out is an in/out parameter, must be set to the input length and will |
1330 | * be also used to return the total number of compressed bytes | |
261507a0 | 1331 | */ |
f51d2b59 | 1332 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1333 | u64 start, struct page **pages, |
261507a0 LZ |
1334 | unsigned long *out_pages, |
1335 | unsigned long *total_in, | |
e5d74902 | 1336 | unsigned long *total_out) |
261507a0 | 1337 | { |
1972708a | 1338 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1339 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1340 | struct list_head *workspace; |
1341 | int ret; | |
1342 | ||
b0c1fe1e | 1343 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1344 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1345 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1346 | out_pages, total_in, total_out); | |
929f4baf | 1347 | put_workspace(type, workspace); |
261507a0 LZ |
1348 | return ret; |
1349 | } | |
1350 | ||
8140dc30 | 1351 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1352 | { |
1353 | struct list_head *workspace; | |
1354 | int ret; | |
8140dc30 | 1355 | int type = cb->compress_type; |
261507a0 | 1356 | |
7bf49943 | 1357 | workspace = get_workspace(type, 0); |
1e4eb746 | 1358 | ret = compression_decompress_bio(type, workspace, cb); |
929f4baf | 1359 | put_workspace(type, workspace); |
e1ddce71 | 1360 | |
261507a0 LZ |
1361 | return ret; |
1362 | } | |
1363 | ||
1364 | /* | |
1365 | * a less complex decompression routine. Our compressed data fits in a | |
1366 | * single page, and we want to read a single page out of it. | |
1367 | * start_byte tells us the offset into the compressed data we're interested in | |
1368 | */ | |
1369 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1370 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1371 | { | |
1372 | struct list_head *workspace; | |
1373 | int ret; | |
1374 | ||
7bf49943 | 1375 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1376 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1377 | start_byte, srclen, destlen); | |
929f4baf | 1378 | put_workspace(type, workspace); |
7bf49943 | 1379 | |
261507a0 LZ |
1380 | return ret; |
1381 | } | |
1382 | ||
1666edab DZ |
1383 | void __init btrfs_init_compress(void) |
1384 | { | |
d5517033 DS |
1385 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1386 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1387 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1388 | zstd_init_workspace_manager(); | |
1666edab DZ |
1389 | } |
1390 | ||
e67c718b | 1391 | void __cold btrfs_exit_compress(void) |
261507a0 | 1392 | { |
2510307e DS |
1393 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1394 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1395 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1396 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1397 | } |
3a39c18d LZ |
1398 | |
1399 | /* | |
1c3dc173 | 1400 | * Copy decompressed data from working buffer to pages. |
3a39c18d | 1401 | * |
1c3dc173 QW |
1402 | * @buf: The decompressed data buffer |
1403 | * @buf_len: The decompressed data length | |
1404 | * @decompressed: Number of bytes that are already decompressed inside the | |
1405 | * compressed extent | |
1406 | * @cb: The compressed extent descriptor | |
1407 | * @orig_bio: The original bio that the caller wants to read for | |
3a39c18d | 1408 | * |
1c3dc173 QW |
1409 | * An easier to understand graph is like below: |
1410 | * | |
1411 | * |<- orig_bio ->| |<- orig_bio->| | |
1412 | * |<------- full decompressed extent ----->| | |
1413 | * |<----------- @cb range ---->| | |
1414 | * | |<-- @buf_len -->| | |
1415 | * |<--- @decompressed --->| | |
1416 | * | |
1417 | * Note that, @cb can be a subpage of the full decompressed extent, but | |
1418 | * @cb->start always has the same as the orig_file_offset value of the full | |
1419 | * decompressed extent. | |
1420 | * | |
1421 | * When reading compressed extent, we have to read the full compressed extent, | |
1422 | * while @orig_bio may only want part of the range. | |
1423 | * Thus this function will ensure only data covered by @orig_bio will be copied | |
1424 | * to. | |
1425 | * | |
1426 | * Return 0 if we have copied all needed contents for @orig_bio. | |
1427 | * Return >0 if we need continue decompress. | |
3a39c18d | 1428 | */ |
1c3dc173 QW |
1429 | int btrfs_decompress_buf2page(const char *buf, u32 buf_len, |
1430 | struct compressed_bio *cb, u32 decompressed) | |
3a39c18d | 1431 | { |
1c3dc173 QW |
1432 | struct bio *orig_bio = cb->orig_bio; |
1433 | /* Offset inside the full decompressed extent */ | |
1434 | u32 cur_offset; | |
1435 | ||
1436 | cur_offset = decompressed; | |
1437 | /* The main loop to do the copy */ | |
1438 | while (cur_offset < decompressed + buf_len) { | |
1439 | struct bio_vec bvec; | |
1440 | size_t copy_len; | |
1441 | u32 copy_start; | |
1442 | /* Offset inside the full decompressed extent */ | |
1443 | u32 bvec_offset; | |
1444 | ||
1445 | bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); | |
1446 | /* | |
1447 | * cb->start may underflow, but subtracting that value can still | |
1448 | * give us correct offset inside the full decompressed extent. | |
1449 | */ | |
1450 | bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; | |
974b1adc | 1451 | |
1c3dc173 QW |
1452 | /* Haven't reached the bvec range, exit */ |
1453 | if (decompressed + buf_len <= bvec_offset) | |
1454 | return 1; | |
3a39c18d | 1455 | |
1c3dc173 QW |
1456 | copy_start = max(cur_offset, bvec_offset); |
1457 | copy_len = min(bvec_offset + bvec.bv_len, | |
1458 | decompressed + buf_len) - copy_start; | |
1459 | ASSERT(copy_len); | |
3a39c18d | 1460 | |
974b1adc | 1461 | /* |
1c3dc173 QW |
1462 | * Extra range check to ensure we didn't go beyond |
1463 | * @buf + @buf_len. | |
974b1adc | 1464 | */ |
1c3dc173 QW |
1465 | ASSERT(copy_start - decompressed < buf_len); |
1466 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, | |
1467 | buf + copy_start - decompressed, copy_len); | |
1468 | flush_dcache_page(bvec.bv_page); | |
1469 | cur_offset += copy_len; | |
3a39c18d | 1470 | |
1c3dc173 QW |
1471 | bio_advance(orig_bio, copy_len); |
1472 | /* Finished the bio */ | |
1473 | if (!orig_bio->bi_iter.bi_size) | |
1474 | return 0; | |
3a39c18d | 1475 | } |
3a39c18d LZ |
1476 | return 1; |
1477 | } | |
c2fcdcdf | 1478 | |
19562430 TT |
1479 | /* |
1480 | * Shannon Entropy calculation | |
1481 | * | |
52042d8e | 1482 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1483 | * Try calculating entropy to estimate the average minimum number of bits |
1484 | * needed to encode the sampled data. | |
1485 | * | |
1486 | * For convenience, return the percentage of needed bits, instead of amount of | |
1487 | * bits directly. | |
1488 | * | |
1489 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1490 | * and can be compressible with high probability | |
1491 | * | |
1492 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1493 | * | |
1494 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1495 | */ | |
1496 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1497 | #define ENTROPY_LVL_HIGH (80) | |
1498 | ||
1499 | /* | |
1500 | * For increasead precision in shannon_entropy calculation, | |
1501 | * let's do pow(n, M) to save more digits after comma: | |
1502 | * | |
1503 | * - maximum int bit length is 64 | |
1504 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1505 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1506 | * | |
1507 | * So use pow(n, 4). | |
1508 | */ | |
1509 | static inline u32 ilog2_w(u64 n) | |
1510 | { | |
1511 | return ilog2(n * n * n * n); | |
1512 | } | |
1513 | ||
1514 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1515 | { | |
1516 | const u32 entropy_max = 8 * ilog2_w(2); | |
1517 | u32 entropy_sum = 0; | |
1518 | u32 p, p_base, sz_base; | |
1519 | u32 i; | |
1520 | ||
1521 | sz_base = ilog2_w(ws->sample_size); | |
1522 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1523 | p = ws->bucket[i].count; | |
1524 | p_base = ilog2_w(p); | |
1525 | entropy_sum += p * (sz_base - p_base); | |
1526 | } | |
1527 | ||
1528 | entropy_sum /= ws->sample_size; | |
1529 | return entropy_sum * 100 / entropy_max; | |
1530 | } | |
1531 | ||
440c840c TT |
1532 | #define RADIX_BASE 4U |
1533 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1534 | ||
1535 | static u8 get4bits(u64 num, int shift) { | |
1536 | u8 low4bits; | |
1537 | ||
1538 | num >>= shift; | |
1539 | /* Reverse order */ | |
1540 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1541 | return low4bits; | |
1542 | } | |
1543 | ||
440c840c TT |
1544 | /* |
1545 | * Use 4 bits as radix base | |
52042d8e | 1546 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1547 | * |
1548 | * @array - array that will be sorted | |
1549 | * @array_buf - buffer array to store sorting results | |
1550 | * must be equal in size to @array | |
1551 | * @num - array size | |
440c840c | 1552 | */ |
23ae8c63 | 1553 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1554 | int num) |
858177d3 | 1555 | { |
440c840c TT |
1556 | u64 max_num; |
1557 | u64 buf_num; | |
1558 | u32 counters[COUNTERS_SIZE]; | |
1559 | u32 new_addr; | |
1560 | u32 addr; | |
1561 | int bitlen; | |
1562 | int shift; | |
1563 | int i; | |
858177d3 | 1564 | |
440c840c TT |
1565 | /* |
1566 | * Try avoid useless loop iterations for small numbers stored in big | |
1567 | * counters. Example: 48 33 4 ... in 64bit array | |
1568 | */ | |
23ae8c63 | 1569 | max_num = array[0].count; |
440c840c | 1570 | for (i = 1; i < num; i++) { |
23ae8c63 | 1571 | buf_num = array[i].count; |
440c840c TT |
1572 | if (buf_num > max_num) |
1573 | max_num = buf_num; | |
1574 | } | |
1575 | ||
1576 | buf_num = ilog2(max_num); | |
1577 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1578 | ||
1579 | shift = 0; | |
1580 | while (shift < bitlen) { | |
1581 | memset(counters, 0, sizeof(counters)); | |
1582 | ||
1583 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1584 | buf_num = array[i].count; |
440c840c TT |
1585 | addr = get4bits(buf_num, shift); |
1586 | counters[addr]++; | |
1587 | } | |
1588 | ||
1589 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1590 | counters[i] += counters[i - 1]; | |
1591 | ||
1592 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1593 | buf_num = array[i].count; |
440c840c TT |
1594 | addr = get4bits(buf_num, shift); |
1595 | counters[addr]--; | |
1596 | new_addr = counters[addr]; | |
7add17be | 1597 | array_buf[new_addr] = array[i]; |
440c840c TT |
1598 | } |
1599 | ||
1600 | shift += RADIX_BASE; | |
1601 | ||
1602 | /* | |
1603 | * Normal radix expects to move data from a temporary array, to | |
1604 | * the main one. But that requires some CPU time. Avoid that | |
1605 | * by doing another sort iteration to original array instead of | |
1606 | * memcpy() | |
1607 | */ | |
1608 | memset(counters, 0, sizeof(counters)); | |
1609 | ||
1610 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1611 | buf_num = array_buf[i].count; |
440c840c TT |
1612 | addr = get4bits(buf_num, shift); |
1613 | counters[addr]++; | |
1614 | } | |
1615 | ||
1616 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1617 | counters[i] += counters[i - 1]; | |
1618 | ||
1619 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1620 | buf_num = array_buf[i].count; |
440c840c TT |
1621 | addr = get4bits(buf_num, shift); |
1622 | counters[addr]--; | |
1623 | new_addr = counters[addr]; | |
7add17be | 1624 | array[new_addr] = array_buf[i]; |
440c840c TT |
1625 | } |
1626 | ||
1627 | shift += RADIX_BASE; | |
1628 | } | |
858177d3 TT |
1629 | } |
1630 | ||
1631 | /* | |
1632 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1633 | * | |
1634 | * There are several types of structured binary data that use nearly all byte | |
1635 | * values. The distribution can be uniform and counts in all buckets will be | |
1636 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1637 | * | |
1638 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1639 | * be potentially compressible, but we have to take a few more steps to decide | |
1640 | * how much. | |
1641 | * | |
1642 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1643 | * compression algo can easy fix that | |
1644 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1645 | * probability is not compressible | |
1646 | */ | |
1647 | #define BYTE_CORE_SET_LOW (64) | |
1648 | #define BYTE_CORE_SET_HIGH (200) | |
1649 | ||
1650 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1651 | { | |
1652 | u32 i; | |
1653 | u32 coreset_sum = 0; | |
1654 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1655 | struct bucket_item *bucket = ws->bucket; | |
1656 | ||
1657 | /* Sort in reverse order */ | |
36243c91 | 1658 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1659 | |
1660 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1661 | coreset_sum += bucket[i].count; | |
1662 | ||
1663 | if (coreset_sum > core_set_threshold) | |
1664 | return i; | |
1665 | ||
1666 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1667 | coreset_sum += bucket[i].count; | |
1668 | if (coreset_sum > core_set_threshold) | |
1669 | break; | |
1670 | } | |
1671 | ||
1672 | return i; | |
1673 | } | |
1674 | ||
a288e92c TT |
1675 | /* |
1676 | * Count byte values in buckets. | |
1677 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1678 | * Because in most text-like data byte set is restricted to limited number of | |
1679 | * possible characters, and that restriction in most cases makes data easy to | |
1680 | * compress. | |
1681 | * | |
1682 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1683 | * less - compressible | |
1684 | * more - need additional analysis | |
1685 | */ | |
1686 | #define BYTE_SET_THRESHOLD (64) | |
1687 | ||
1688 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1689 | { | |
1690 | u32 i; | |
1691 | u32 byte_set_size = 0; | |
1692 | ||
1693 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1694 | if (ws->bucket[i].count > 0) | |
1695 | byte_set_size++; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Continue collecting count of byte values in buckets. If the byte | |
1700 | * set size is bigger then the threshold, it's pointless to continue, | |
1701 | * the detection technique would fail for this type of data. | |
1702 | */ | |
1703 | for (; i < BUCKET_SIZE; i++) { | |
1704 | if (ws->bucket[i].count > 0) { | |
1705 | byte_set_size++; | |
1706 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1707 | return byte_set_size; | |
1708 | } | |
1709 | } | |
1710 | ||
1711 | return byte_set_size; | |
1712 | } | |
1713 | ||
1fe4f6fa TT |
1714 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1715 | { | |
1716 | const u32 half_of_sample = ws->sample_size / 2; | |
1717 | const u8 *data = ws->sample; | |
1718 | ||
1719 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1720 | } | |
1721 | ||
a440d48c TT |
1722 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1723 | struct heuristic_ws *ws) | |
1724 | { | |
1725 | struct page *page; | |
1726 | u64 index, index_end; | |
1727 | u32 i, curr_sample_pos; | |
1728 | u8 *in_data; | |
1729 | ||
1730 | /* | |
1731 | * Compression handles the input data by chunks of 128KiB | |
1732 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1733 | * | |
1734 | * We do the same for the heuristic and loop over the whole range. | |
1735 | * | |
1736 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1737 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1738 | */ | |
1739 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1740 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1741 | ||
1742 | index = start >> PAGE_SHIFT; | |
1743 | index_end = end >> PAGE_SHIFT; | |
1744 | ||
1745 | /* Don't miss unaligned end */ | |
1746 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1747 | index_end++; | |
1748 | ||
1749 | curr_sample_pos = 0; | |
1750 | while (index < index_end) { | |
1751 | page = find_get_page(inode->i_mapping, index); | |
58c1a35c | 1752 | in_data = kmap_local_page(page); |
a440d48c TT |
1753 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1754 | i = start % PAGE_SIZE; | |
1755 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1756 | /* Don't sample any garbage from the last page */ | |
1757 | if (start > end - SAMPLING_READ_SIZE) | |
1758 | break; | |
1759 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1760 | SAMPLING_READ_SIZE); | |
1761 | i += SAMPLING_INTERVAL; | |
1762 | start += SAMPLING_INTERVAL; | |
1763 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1764 | } | |
58c1a35c | 1765 | kunmap_local(in_data); |
a440d48c TT |
1766 | put_page(page); |
1767 | ||
1768 | index++; | |
1769 | } | |
1770 | ||
1771 | ws->sample_size = curr_sample_pos; | |
1772 | } | |
1773 | ||
c2fcdcdf TT |
1774 | /* |
1775 | * Compression heuristic. | |
1776 | * | |
1777 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1778 | * quickly (compared to direct compression) detect data characteristics | |
1779 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1780 | * data. | |
1781 | * | |
1782 | * The following types of analysis can be performed: | |
1783 | * - detect mostly zero data | |
1784 | * - detect data with low "byte set" size (text, etc) | |
1785 | * - detect data with low/high "core byte" set | |
1786 | * | |
1787 | * Return non-zero if the compression should be done, 0 otherwise. | |
1788 | */ | |
1789 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1790 | { | |
7bf49943 | 1791 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1792 | struct heuristic_ws *ws; |
a440d48c TT |
1793 | u32 i; |
1794 | u8 byte; | |
19562430 | 1795 | int ret = 0; |
c2fcdcdf | 1796 | |
4e439a0b TT |
1797 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1798 | ||
a440d48c TT |
1799 | heuristic_collect_sample(inode, start, end, ws); |
1800 | ||
1fe4f6fa TT |
1801 | if (sample_repeated_patterns(ws)) { |
1802 | ret = 1; | |
1803 | goto out; | |
1804 | } | |
1805 | ||
a440d48c TT |
1806 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1807 | ||
1808 | for (i = 0; i < ws->sample_size; i++) { | |
1809 | byte = ws->sample[i]; | |
1810 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1811 | } |
1812 | ||
a288e92c TT |
1813 | i = byte_set_size(ws); |
1814 | if (i < BYTE_SET_THRESHOLD) { | |
1815 | ret = 2; | |
1816 | goto out; | |
1817 | } | |
1818 | ||
858177d3 TT |
1819 | i = byte_core_set_size(ws); |
1820 | if (i <= BYTE_CORE_SET_LOW) { | |
1821 | ret = 3; | |
1822 | goto out; | |
1823 | } | |
1824 | ||
1825 | if (i >= BYTE_CORE_SET_HIGH) { | |
1826 | ret = 0; | |
1827 | goto out; | |
1828 | } | |
1829 | ||
19562430 TT |
1830 | i = shannon_entropy(ws); |
1831 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1832 | ret = 4; | |
1833 | goto out; | |
1834 | } | |
1835 | ||
1836 | /* | |
1837 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1838 | * needed to give green light to compression. | |
1839 | * | |
1840 | * For now just assume that compression at that level is not worth the | |
1841 | * resources because: | |
1842 | * | |
1843 | * 1. it is possible to defrag the data later | |
1844 | * | |
1845 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1846 | * values, every bucket has counter at level ~54. The heuristic would | |
1847 | * be confused. This can happen when data have some internal repeated | |
1848 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1849 | * pairs of bytes, which is too costly. | |
1850 | */ | |
1851 | if (i < ENTROPY_LVL_HIGH) { | |
1852 | ret = 5; | |
1853 | goto out; | |
1854 | } else { | |
1855 | ret = 0; | |
1856 | goto out; | |
1857 | } | |
1858 | ||
1fe4f6fa | 1859 | out: |
929f4baf | 1860 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1861 | return ret; |
1862 | } | |
f51d2b59 | 1863 | |
d0ab62ce DZ |
1864 | /* |
1865 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1866 | * level, unrecognized string will set the default level | |
1867 | */ | |
1868 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1869 | { |
d0ab62ce DZ |
1870 | unsigned int level = 0; |
1871 | int ret; | |
1872 | ||
1873 | if (!type) | |
f51d2b59 DS |
1874 | return 0; |
1875 | ||
d0ab62ce DZ |
1876 | if (str[0] == ':') { |
1877 | ret = kstrtouint(str + 1, 10, &level); | |
1878 | if (ret) | |
1879 | level = 0; | |
1880 | } | |
1881 | ||
b0c1fe1e DS |
1882 | level = btrfs_compress_set_level(type, level); |
1883 | ||
1884 | return level; | |
1885 | } |