Commit | Line | Data |
---|---|---|
c1d7c514 | 1 | // SPDX-License-Identifier: GPL-2.0 |
c8b97818 CM |
2 | /* |
3 | * Copyright (C) 2008 Oracle. All rights reserved. | |
c8b97818 CM |
4 | */ |
5 | ||
6 | #include <linux/kernel.h> | |
7 | #include <linux/bio.h> | |
c8b97818 CM |
8 | #include <linux/file.h> |
9 | #include <linux/fs.h> | |
10 | #include <linux/pagemap.h> | |
11 | #include <linux/highmem.h> | |
12 | #include <linux/time.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/string.h> | |
c8b97818 | 15 | #include <linux/backing-dev.h> |
c8b97818 | 16 | #include <linux/writeback.h> |
5a0e3ad6 | 17 | #include <linux/slab.h> |
fe308533 | 18 | #include <linux/sched/mm.h> |
19562430 | 19 | #include <linux/log2.h> |
d5178578 | 20 | #include <crypto/hash.h> |
602cbe91 | 21 | #include "misc.h" |
c8b97818 CM |
22 | #include "ctree.h" |
23 | #include "disk-io.h" | |
24 | #include "transaction.h" | |
25 | #include "btrfs_inode.h" | |
26 | #include "volumes.h" | |
27 | #include "ordered-data.h" | |
c8b97818 CM |
28 | #include "compression.h" |
29 | #include "extent_io.h" | |
30 | #include "extent_map.h" | |
31 | ||
c4bf665a DS |
32 | int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, |
33 | u64 start, struct page **pages, unsigned long *out_pages, | |
34 | unsigned long *total_in, unsigned long *total_out); | |
35 | int zlib_decompress_bio(struct list_head *ws, struct compressed_bio *cb); | |
36 | int zlib_decompress(struct list_head *ws, unsigned char *data_in, | |
37 | struct page *dest_page, unsigned long start_byte, size_t srclen, | |
38 | size_t destlen); | |
d20f395f DS |
39 | struct list_head *zlib_alloc_workspace(unsigned int level); |
40 | void zlib_free_workspace(struct list_head *ws); | |
41 | struct list_head *zlib_get_workspace(unsigned int level); | |
c4bf665a DS |
42 | |
43 | int lzo_compress_pages(struct list_head *ws, struct address_space *mapping, | |
44 | u64 start, struct page **pages, unsigned long *out_pages, | |
45 | unsigned long *total_in, unsigned long *total_out); | |
46 | int lzo_decompress_bio(struct list_head *ws, struct compressed_bio *cb); | |
47 | int lzo_decompress(struct list_head *ws, unsigned char *data_in, | |
48 | struct page *dest_page, unsigned long start_byte, size_t srclen, | |
49 | size_t destlen); | |
d20f395f DS |
50 | struct list_head *lzo_alloc_workspace(unsigned int level); |
51 | void lzo_free_workspace(struct list_head *ws); | |
c4bf665a DS |
52 | |
53 | int zstd_compress_pages(struct list_head *ws, struct address_space *mapping, | |
54 | u64 start, struct page **pages, unsigned long *out_pages, | |
55 | unsigned long *total_in, unsigned long *total_out); | |
56 | int zstd_decompress_bio(struct list_head *ws, struct compressed_bio *cb); | |
57 | int zstd_decompress(struct list_head *ws, unsigned char *data_in, | |
58 | struct page *dest_page, unsigned long start_byte, size_t srclen, | |
59 | size_t destlen); | |
d5517033 | 60 | void zstd_init_workspace_manager(void); |
2510307e | 61 | void zstd_cleanup_workspace_manager(void); |
d20f395f DS |
62 | struct list_head *zstd_alloc_workspace(unsigned int level); |
63 | void zstd_free_workspace(struct list_head *ws); | |
64 | struct list_head *zstd_get_workspace(unsigned int level); | |
65 | void zstd_put_workspace(struct list_head *ws); | |
c4bf665a | 66 | |
e128f9c3 DS |
67 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
68 | ||
69 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
70 | { | |
71 | switch (type) { | |
72 | case BTRFS_COMPRESS_ZLIB: | |
73 | case BTRFS_COMPRESS_LZO: | |
74 | case BTRFS_COMPRESS_ZSTD: | |
75 | case BTRFS_COMPRESS_NONE: | |
76 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
77 | default: |
78 | break; | |
e128f9c3 DS |
79 | } |
80 | ||
81 | return NULL; | |
82 | } | |
83 | ||
aa53e3bf JT |
84 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
85 | { | |
86 | int i; | |
87 | ||
88 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
89 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
90 | ||
91 | if (len < comp_len) | |
92 | continue; | |
93 | ||
94 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
95 | return true; | |
96 | } | |
97 | return false; | |
98 | } | |
99 | ||
1e4eb746 DS |
100 | static int compression_compress_pages(int type, struct list_head *ws, |
101 | struct address_space *mapping, u64 start, struct page **pages, | |
102 | unsigned long *out_pages, unsigned long *total_in, | |
103 | unsigned long *total_out) | |
104 | { | |
105 | switch (type) { | |
106 | case BTRFS_COMPRESS_ZLIB: | |
107 | return zlib_compress_pages(ws, mapping, start, pages, | |
108 | out_pages, total_in, total_out); | |
109 | case BTRFS_COMPRESS_LZO: | |
110 | return lzo_compress_pages(ws, mapping, start, pages, | |
111 | out_pages, total_in, total_out); | |
112 | case BTRFS_COMPRESS_ZSTD: | |
113 | return zstd_compress_pages(ws, mapping, start, pages, | |
114 | out_pages, total_in, total_out); | |
115 | case BTRFS_COMPRESS_NONE: | |
116 | default: | |
117 | /* | |
118 | * This can't happen, the type is validated several times | |
119 | * before we get here. As a sane fallback, return what the | |
120 | * callers will understand as 'no compression happened'. | |
121 | */ | |
122 | return -E2BIG; | |
123 | } | |
124 | } | |
125 | ||
126 | static int compression_decompress_bio(int type, struct list_head *ws, | |
127 | struct compressed_bio *cb) | |
128 | { | |
129 | switch (type) { | |
130 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); | |
131 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
132 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
133 | case BTRFS_COMPRESS_NONE: | |
134 | default: | |
135 | /* | |
136 | * This can't happen, the type is validated several times | |
137 | * before we get here. | |
138 | */ | |
139 | BUG(); | |
140 | } | |
141 | } | |
142 | ||
143 | static int compression_decompress(int type, struct list_head *ws, | |
144 | unsigned char *data_in, struct page *dest_page, | |
145 | unsigned long start_byte, size_t srclen, size_t destlen) | |
146 | { | |
147 | switch (type) { | |
148 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
149 | start_byte, srclen, destlen); | |
150 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
151 | start_byte, srclen, destlen); | |
152 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
153 | start_byte, srclen, destlen); | |
154 | case BTRFS_COMPRESS_NONE: | |
155 | default: | |
156 | /* | |
157 | * This can't happen, the type is validated several times | |
158 | * before we get here. | |
159 | */ | |
160 | BUG(); | |
161 | } | |
162 | } | |
163 | ||
8140dc30 | 164 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 165 | |
2ff7e61e | 166 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
167 | unsigned long disk_size) |
168 | { | |
0b246afa | 169 | u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); |
6c41761f | 170 | |
d20f7043 | 171 | return sizeof(struct compressed_bio) + |
0b246afa | 172 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size; |
d20f7043 CM |
173 | } |
174 | ||
f898ac6a | 175 | static int check_compressed_csum(struct btrfs_inode *inode, |
d20f7043 CM |
176 | struct compressed_bio *cb, |
177 | u64 disk_start) | |
178 | { | |
10fe6ca8 | 179 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
d5178578 | 180 | SHASH_DESC_ON_STACK(shash, fs_info->csum_shash); |
10fe6ca8 | 181 | const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); |
d20f7043 | 182 | int ret; |
d20f7043 CM |
183 | struct page *page; |
184 | unsigned long i; | |
185 | char *kaddr; | |
d5178578 | 186 | u8 csum[BTRFS_CSUM_SIZE]; |
10fe6ca8 | 187 | u8 *cb_sum = cb->sums; |
d20f7043 | 188 | |
f898ac6a | 189 | if (inode->flags & BTRFS_INODE_NODATASUM) |
d20f7043 CM |
190 | return 0; |
191 | ||
d5178578 JT |
192 | shash->tfm = fs_info->csum_shash; |
193 | ||
d20f7043 CM |
194 | for (i = 0; i < cb->nr_pages; i++) { |
195 | page = cb->compressed_pages[i]; | |
d20f7043 | 196 | |
d5178578 | 197 | crypto_shash_init(shash); |
7ac687d9 | 198 | kaddr = kmap_atomic(page); |
d5178578 | 199 | crypto_shash_update(shash, kaddr, PAGE_SIZE); |
7ac687d9 | 200 | kunmap_atomic(kaddr); |
d5178578 | 201 | crypto_shash_final(shash, (u8 *)&csum); |
d20f7043 | 202 | |
10fe6ca8 | 203 | if (memcmp(&csum, cb_sum, csum_size)) { |
d5178578 | 204 | btrfs_print_data_csum_error(inode, disk_start, |
ea41d6b2 | 205 | csum, cb_sum, cb->mirror_num); |
d20f7043 CM |
206 | ret = -EIO; |
207 | goto fail; | |
208 | } | |
10fe6ca8 | 209 | cb_sum += csum_size; |
d20f7043 CM |
210 | |
211 | } | |
212 | ret = 0; | |
213 | fail: | |
214 | return ret; | |
215 | } | |
216 | ||
c8b97818 CM |
217 | /* when we finish reading compressed pages from the disk, we |
218 | * decompress them and then run the bio end_io routines on the | |
219 | * decompressed pages (in the inode address space). | |
220 | * | |
221 | * This allows the checksumming and other IO error handling routines | |
222 | * to work normally | |
223 | * | |
224 | * The compressed pages are freed here, and it must be run | |
225 | * in process context | |
226 | */ | |
4246a0b6 | 227 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 228 | { |
c8b97818 CM |
229 | struct compressed_bio *cb = bio->bi_private; |
230 | struct inode *inode; | |
231 | struct page *page; | |
232 | unsigned long index; | |
cf1167d5 | 233 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 234 | int ret = 0; |
c8b97818 | 235 | |
4e4cbee9 | 236 | if (bio->bi_status) |
c8b97818 CM |
237 | cb->errors = 1; |
238 | ||
239 | /* if there are more bios still pending for this compressed | |
240 | * extent, just exit | |
241 | */ | |
a50299ae | 242 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
243 | goto out; |
244 | ||
cf1167d5 LB |
245 | /* |
246 | * Record the correct mirror_num in cb->orig_bio so that | |
247 | * read-repair can work properly. | |
248 | */ | |
249 | ASSERT(btrfs_io_bio(cb->orig_bio)); | |
250 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; | |
251 | cb->mirror_num = mirror; | |
252 | ||
e6311f24 LB |
253 | /* |
254 | * Some IO in this cb have failed, just skip checksum as there | |
255 | * is no way it could be correct. | |
256 | */ | |
257 | if (cb->errors == 1) | |
258 | goto csum_failed; | |
259 | ||
d20f7043 | 260 | inode = cb->inode; |
f898ac6a | 261 | ret = check_compressed_csum(BTRFS_I(inode), cb, |
4f024f37 | 262 | (u64)bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
263 | if (ret) |
264 | goto csum_failed; | |
265 | ||
c8b97818 CM |
266 | /* ok, we're the last bio for this extent, lets start |
267 | * the decompression. | |
268 | */ | |
8140dc30 AJ |
269 | ret = btrfs_decompress_bio(cb); |
270 | ||
d20f7043 | 271 | csum_failed: |
c8b97818 CM |
272 | if (ret) |
273 | cb->errors = 1; | |
274 | ||
275 | /* release the compressed pages */ | |
276 | index = 0; | |
277 | for (index = 0; index < cb->nr_pages; index++) { | |
278 | page = cb->compressed_pages[index]; | |
279 | page->mapping = NULL; | |
09cbfeaf | 280 | put_page(page); |
c8b97818 CM |
281 | } |
282 | ||
283 | /* do io completion on the original bio */ | |
771ed689 | 284 | if (cb->errors) { |
c8b97818 | 285 | bio_io_error(cb->orig_bio); |
d20f7043 | 286 | } else { |
2c30c71b | 287 | struct bio_vec *bvec; |
6dc4f100 | 288 | struct bvec_iter_all iter_all; |
d20f7043 CM |
289 | |
290 | /* | |
291 | * we have verified the checksum already, set page | |
292 | * checked so the end_io handlers know about it | |
293 | */ | |
c09abff8 | 294 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2b070cfe | 295 | bio_for_each_segment_all(bvec, cb->orig_bio, iter_all) |
d20f7043 | 296 | SetPageChecked(bvec->bv_page); |
2c30c71b | 297 | |
4246a0b6 | 298 | bio_endio(cb->orig_bio); |
d20f7043 | 299 | } |
c8b97818 CM |
300 | |
301 | /* finally free the cb struct */ | |
302 | kfree(cb->compressed_pages); | |
303 | kfree(cb); | |
304 | out: | |
305 | bio_put(bio); | |
306 | } | |
307 | ||
308 | /* | |
309 | * Clear the writeback bits on all of the file | |
310 | * pages for a compressed write | |
311 | */ | |
7bdcefc1 FM |
312 | static noinline void end_compressed_writeback(struct inode *inode, |
313 | const struct compressed_bio *cb) | |
c8b97818 | 314 | { |
09cbfeaf KS |
315 | unsigned long index = cb->start >> PAGE_SHIFT; |
316 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
317 | struct page *pages[16]; |
318 | unsigned long nr_pages = end_index - index + 1; | |
319 | int i; | |
320 | int ret; | |
321 | ||
7bdcefc1 FM |
322 | if (cb->errors) |
323 | mapping_set_error(inode->i_mapping, -EIO); | |
324 | ||
d397712b | 325 | while (nr_pages > 0) { |
c8b97818 | 326 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
327 | min_t(unsigned long, |
328 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
329 | if (ret == 0) { |
330 | nr_pages -= 1; | |
331 | index += 1; | |
332 | continue; | |
333 | } | |
334 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
335 | if (cb->errors) |
336 | SetPageError(pages[i]); | |
c8b97818 | 337 | end_page_writeback(pages[i]); |
09cbfeaf | 338 | put_page(pages[i]); |
c8b97818 CM |
339 | } |
340 | nr_pages -= ret; | |
341 | index += ret; | |
342 | } | |
343 | /* the inode may be gone now */ | |
c8b97818 CM |
344 | } |
345 | ||
346 | /* | |
347 | * do the cleanup once all the compressed pages hit the disk. | |
348 | * This will clear writeback on the file pages and free the compressed | |
349 | * pages. | |
350 | * | |
351 | * This also calls the writeback end hooks for the file pages so that | |
352 | * metadata and checksums can be updated in the file. | |
353 | */ | |
4246a0b6 | 354 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 | 355 | { |
c8b97818 CM |
356 | struct compressed_bio *cb = bio->bi_private; |
357 | struct inode *inode; | |
358 | struct page *page; | |
359 | unsigned long index; | |
360 | ||
4e4cbee9 | 361 | if (bio->bi_status) |
c8b97818 CM |
362 | cb->errors = 1; |
363 | ||
364 | /* if there are more bios still pending for this compressed | |
365 | * extent, just exit | |
366 | */ | |
a50299ae | 367 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
368 | goto out; |
369 | ||
370 | /* ok, we're the last bio for this extent, step one is to | |
371 | * call back into the FS and do all the end_io operations | |
372 | */ | |
373 | inode = cb->inode; | |
70b99e69 | 374 | cb->compressed_pages[0]->mapping = cb->inode->i_mapping; |
7087a9d8 | 375 | btrfs_writepage_endio_finish_ordered(cb->compressed_pages[0], |
c629732d | 376 | cb->start, cb->start + cb->len - 1, |
6a8d2136 | 377 | bio->bi_status == BLK_STS_OK); |
70b99e69 | 378 | cb->compressed_pages[0]->mapping = NULL; |
c8b97818 | 379 | |
7bdcefc1 | 380 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
381 | /* note, our inode could be gone now */ |
382 | ||
383 | /* | |
384 | * release the compressed pages, these came from alloc_page and | |
385 | * are not attached to the inode at all | |
386 | */ | |
387 | index = 0; | |
388 | for (index = 0; index < cb->nr_pages; index++) { | |
389 | page = cb->compressed_pages[index]; | |
390 | page->mapping = NULL; | |
09cbfeaf | 391 | put_page(page); |
c8b97818 CM |
392 | } |
393 | ||
394 | /* finally free the cb struct */ | |
395 | kfree(cb->compressed_pages); | |
396 | kfree(cb); | |
397 | out: | |
398 | bio_put(bio); | |
399 | } | |
400 | ||
401 | /* | |
402 | * worker function to build and submit bios for previously compressed pages. | |
403 | * The corresponding pages in the inode should be marked for writeback | |
404 | * and the compressed pages should have a reference on them for dropping | |
405 | * when the IO is complete. | |
406 | * | |
407 | * This also checksums the file bytes and gets things ready for | |
408 | * the end io hooks. | |
409 | */ | |
4e4cbee9 | 410 | blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start, |
c8b97818 CM |
411 | unsigned long len, u64 disk_start, |
412 | unsigned long compressed_len, | |
413 | struct page **compressed_pages, | |
f82b7359 | 414 | unsigned long nr_pages, |
ec39f769 CM |
415 | unsigned int write_flags, |
416 | struct cgroup_subsys_state *blkcg_css) | |
c8b97818 | 417 | { |
0b246afa | 418 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 | 419 | struct bio *bio = NULL; |
c8b97818 CM |
420 | struct compressed_bio *cb; |
421 | unsigned long bytes_left; | |
306e16ce | 422 | int pg_index = 0; |
c8b97818 CM |
423 | struct page *page; |
424 | u64 first_byte = disk_start; | |
425 | struct block_device *bdev; | |
4e4cbee9 | 426 | blk_status_t ret; |
e55179b3 | 427 | int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM; |
c8b97818 | 428 | |
fdb1e121 | 429 | WARN_ON(!PAGE_ALIGNED(start)); |
2ff7e61e | 430 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 431 | if (!cb) |
4e4cbee9 | 432 | return BLK_STS_RESOURCE; |
a50299ae | 433 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
434 | cb->errors = 0; |
435 | cb->inode = inode; | |
436 | cb->start = start; | |
437 | cb->len = len; | |
d20f7043 | 438 | cb->mirror_num = 0; |
c8b97818 CM |
439 | cb->compressed_pages = compressed_pages; |
440 | cb->compressed_len = compressed_len; | |
441 | cb->orig_bio = NULL; | |
442 | cb->nr_pages = nr_pages; | |
443 | ||
0b246afa | 444 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 445 | |
e749af44 DS |
446 | bio = btrfs_bio_alloc(first_byte); |
447 | bio_set_dev(bio, bdev); | |
f82b7359 | 448 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
449 | bio->bi_private = cb; |
450 | bio->bi_end_io = end_compressed_bio_write; | |
ec39f769 CM |
451 | |
452 | if (blkcg_css) { | |
453 | bio->bi_opf |= REQ_CGROUP_PUNT; | |
454 | bio_associate_blkg_from_css(bio, blkcg_css); | |
455 | } | |
a50299ae | 456 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
457 | |
458 | /* create and submit bios for the compressed pages */ | |
459 | bytes_left = compressed_len; | |
306e16ce | 460 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 CH |
461 | int submit = 0; |
462 | ||
306e16ce | 463 | page = compressed_pages[pg_index]; |
c8b97818 | 464 | page->mapping = inode->i_mapping; |
4f024f37 | 465 | if (bio->bi_iter.bi_size) |
da12fe54 NB |
466 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, bio, |
467 | 0); | |
c8b97818 | 468 | |
70b99e69 | 469 | page->mapping = NULL; |
4e4cbee9 | 470 | if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 471 | PAGE_SIZE) { |
af09abfe CM |
472 | /* |
473 | * inc the count before we submit the bio so | |
474 | * we know the end IO handler won't happen before | |
475 | * we inc the count. Otherwise, the cb might get | |
476 | * freed before we're done setting it up | |
477 | */ | |
a50299ae | 478 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
479 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
480 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 481 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 482 | |
e55179b3 | 483 | if (!skip_sum) { |
2ff7e61e | 484 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 485 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 486 | } |
d20f7043 | 487 | |
08635bae | 488 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 489 | if (ret) { |
4e4cbee9 | 490 | bio->bi_status = ret; |
f5daf2c7 LB |
491 | bio_endio(bio); |
492 | } | |
c8b97818 | 493 | |
e749af44 DS |
494 | bio = btrfs_bio_alloc(first_byte); |
495 | bio_set_dev(bio, bdev); | |
f82b7359 | 496 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
497 | bio->bi_private = cb; |
498 | bio->bi_end_io = end_compressed_bio_write; | |
09cbfeaf | 499 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 500 | } |
09cbfeaf | 501 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 502 | btrfs_info(fs_info, |
efe120a0 | 503 | "bytes left %lu compress len %lu nr %lu", |
cfbc246e CM |
504 | bytes_left, cb->compressed_len, cb->nr_pages); |
505 | } | |
09cbfeaf KS |
506 | bytes_left -= PAGE_SIZE; |
507 | first_byte += PAGE_SIZE; | |
771ed689 | 508 | cond_resched(); |
c8b97818 | 509 | } |
c8b97818 | 510 | |
0b246afa | 511 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 512 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 513 | |
e55179b3 | 514 | if (!skip_sum) { |
2ff7e61e | 515 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 516 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 517 | } |
d20f7043 | 518 | |
08635bae | 519 | ret = btrfs_map_bio(fs_info, bio, 0); |
f5daf2c7 | 520 | if (ret) { |
4e4cbee9 | 521 | bio->bi_status = ret; |
f5daf2c7 LB |
522 | bio_endio(bio); |
523 | } | |
c8b97818 | 524 | |
c8b97818 CM |
525 | return 0; |
526 | } | |
527 | ||
2a4d0c90 CH |
528 | static u64 bio_end_offset(struct bio *bio) |
529 | { | |
c45a8f2d | 530 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
531 | |
532 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
533 | } | |
534 | ||
771ed689 CM |
535 | static noinline int add_ra_bio_pages(struct inode *inode, |
536 | u64 compressed_end, | |
537 | struct compressed_bio *cb) | |
538 | { | |
539 | unsigned long end_index; | |
306e16ce | 540 | unsigned long pg_index; |
771ed689 CM |
541 | u64 last_offset; |
542 | u64 isize = i_size_read(inode); | |
543 | int ret; | |
544 | struct page *page; | |
545 | unsigned long nr_pages = 0; | |
546 | struct extent_map *em; | |
547 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
548 | struct extent_map_tree *em_tree; |
549 | struct extent_io_tree *tree; | |
550 | u64 end; | |
551 | int misses = 0; | |
552 | ||
2a4d0c90 | 553 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
554 | em_tree = &BTRFS_I(inode)->extent_tree; |
555 | tree = &BTRFS_I(inode)->io_tree; | |
556 | ||
557 | if (isize == 0) | |
558 | return 0; | |
559 | ||
09cbfeaf | 560 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 561 | |
d397712b | 562 | while (last_offset < compressed_end) { |
09cbfeaf | 563 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 564 | |
306e16ce | 565 | if (pg_index > end_index) |
771ed689 CM |
566 | break; |
567 | ||
0a943c65 | 568 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 569 | if (page && !xa_is_value(page)) { |
771ed689 CM |
570 | misses++; |
571 | if (misses > 4) | |
572 | break; | |
573 | goto next; | |
574 | } | |
575 | ||
c62d2555 MH |
576 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
577 | ~__GFP_FS)); | |
771ed689 CM |
578 | if (!page) |
579 | break; | |
580 | ||
c62d2555 | 581 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 582 | put_page(page); |
771ed689 CM |
583 | goto next; |
584 | } | |
585 | ||
09cbfeaf | 586 | end = last_offset + PAGE_SIZE - 1; |
771ed689 CM |
587 | /* |
588 | * at this point, we have a locked page in the page cache | |
589 | * for these bytes in the file. But, we have to make | |
590 | * sure they map to this compressed extent on disk. | |
591 | */ | |
592 | set_page_extent_mapped(page); | |
d0082371 | 593 | lock_extent(tree, last_offset, end); |
890871be | 594 | read_lock(&em_tree->lock); |
771ed689 | 595 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 596 | PAGE_SIZE); |
890871be | 597 | read_unlock(&em_tree->lock); |
771ed689 CM |
598 | |
599 | if (!em || last_offset < em->start || | |
09cbfeaf | 600 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 601 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 602 | free_extent_map(em); |
d0082371 | 603 | unlock_extent(tree, last_offset, end); |
771ed689 | 604 | unlock_page(page); |
09cbfeaf | 605 | put_page(page); |
771ed689 CM |
606 | break; |
607 | } | |
608 | free_extent_map(em); | |
609 | ||
610 | if (page->index == end_index) { | |
611 | char *userpage; | |
7073017a | 612 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
613 | |
614 | if (zero_offset) { | |
615 | int zeros; | |
09cbfeaf | 616 | zeros = PAGE_SIZE - zero_offset; |
7ac687d9 | 617 | userpage = kmap_atomic(page); |
771ed689 CM |
618 | memset(userpage + zero_offset, 0, zeros); |
619 | flush_dcache_page(page); | |
7ac687d9 | 620 | kunmap_atomic(userpage); |
771ed689 CM |
621 | } |
622 | } | |
623 | ||
624 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 625 | PAGE_SIZE, 0); |
771ed689 | 626 | |
09cbfeaf | 627 | if (ret == PAGE_SIZE) { |
771ed689 | 628 | nr_pages++; |
09cbfeaf | 629 | put_page(page); |
771ed689 | 630 | } else { |
d0082371 | 631 | unlock_extent(tree, last_offset, end); |
771ed689 | 632 | unlock_page(page); |
09cbfeaf | 633 | put_page(page); |
771ed689 CM |
634 | break; |
635 | } | |
636 | next: | |
09cbfeaf | 637 | last_offset += PAGE_SIZE; |
771ed689 | 638 | } |
771ed689 CM |
639 | return 0; |
640 | } | |
641 | ||
c8b97818 CM |
642 | /* |
643 | * for a compressed read, the bio we get passed has all the inode pages | |
644 | * in it. We don't actually do IO on those pages but allocate new ones | |
645 | * to hold the compressed pages on disk. | |
646 | * | |
4f024f37 | 647 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 648 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
649 | * |
650 | * After the compressed pages are read, we copy the bytes into the | |
651 | * bio we were passed and then call the bio end_io calls | |
652 | */ | |
4e4cbee9 | 653 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
654 | int mirror_num, unsigned long bio_flags) |
655 | { | |
0b246afa | 656 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
657 | struct extent_map_tree *em_tree; |
658 | struct compressed_bio *cb; | |
c8b97818 CM |
659 | unsigned long compressed_len; |
660 | unsigned long nr_pages; | |
306e16ce | 661 | unsigned long pg_index; |
c8b97818 CM |
662 | struct page *page; |
663 | struct block_device *bdev; | |
664 | struct bio *comp_bio; | |
4f024f37 | 665 | u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
666 | u64 em_len; |
667 | u64 em_start; | |
c8b97818 | 668 | struct extent_map *em; |
4e4cbee9 | 669 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 670 | int faili = 0; |
10fe6ca8 JT |
671 | const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); |
672 | u8 *sums; | |
c8b97818 | 673 | |
c8b97818 CM |
674 | em_tree = &BTRFS_I(inode)->extent_tree; |
675 | ||
676 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 677 | read_lock(&em_tree->lock); |
c8b97818 | 678 | em = lookup_extent_mapping(em_tree, |
263663cd | 679 | page_offset(bio_first_page_all(bio)), |
09cbfeaf | 680 | PAGE_SIZE); |
890871be | 681 | read_unlock(&em_tree->lock); |
285190d9 | 682 | if (!em) |
4e4cbee9 | 683 | return BLK_STS_IOERR; |
c8b97818 | 684 | |
d20f7043 | 685 | compressed_len = em->block_len; |
2ff7e61e | 686 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 687 | if (!cb) |
688 | goto out; | |
689 | ||
a50299ae | 690 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
691 | cb->errors = 0; |
692 | cb->inode = inode; | |
d20f7043 | 693 | cb->mirror_num = mirror_num; |
10fe6ca8 | 694 | sums = cb->sums; |
c8b97818 | 695 | |
ff5b7ee3 | 696 | cb->start = em->orig_start; |
e04ca626 CM |
697 | em_len = em->len; |
698 | em_start = em->start; | |
d20f7043 | 699 | |
c8b97818 | 700 | free_extent_map(em); |
e04ca626 | 701 | em = NULL; |
c8b97818 | 702 | |
81381053 | 703 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 704 | cb->compressed_len = compressed_len; |
261507a0 | 705 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
706 | cb->orig_bio = bio; |
707 | ||
09cbfeaf | 708 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 709 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 710 | GFP_NOFS); |
6b82ce8d | 711 | if (!cb->compressed_pages) |
712 | goto fail1; | |
713 | ||
0b246afa | 714 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 715 | |
306e16ce DS |
716 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
717 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS | | |
c8b97818 | 718 | __GFP_HIGHMEM); |
15e3004a JB |
719 | if (!cb->compressed_pages[pg_index]) { |
720 | faili = pg_index - 1; | |
0e9350de | 721 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 722 | goto fail2; |
15e3004a | 723 | } |
c8b97818 | 724 | } |
15e3004a | 725 | faili = nr_pages - 1; |
c8b97818 CM |
726 | cb->nr_pages = nr_pages; |
727 | ||
7f042a83 | 728 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 729 | |
771ed689 | 730 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 731 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 732 | |
e749af44 DS |
733 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
734 | bio_set_dev(comp_bio, bdev); | |
ebcc3263 | 735 | comp_bio->bi_opf = REQ_OP_READ; |
c8b97818 CM |
736 | comp_bio->bi_private = cb; |
737 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 738 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 739 | |
306e16ce | 740 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
4e4cbee9 CH |
741 | int submit = 0; |
742 | ||
306e16ce | 743 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 744 | page->mapping = inode->i_mapping; |
09cbfeaf | 745 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 746 | |
4f024f37 | 747 | if (comp_bio->bi_iter.bi_size) |
da12fe54 NB |
748 | submit = btrfs_bio_fits_in_stripe(page, PAGE_SIZE, |
749 | comp_bio, 0); | |
c8b97818 | 750 | |
70b99e69 | 751 | page->mapping = NULL; |
4e4cbee9 | 752 | if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 753 | PAGE_SIZE) { |
10fe6ca8 JT |
754 | unsigned int nr_sectors; |
755 | ||
0b246afa JM |
756 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
757 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 758 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 759 | |
af09abfe CM |
760 | /* |
761 | * inc the count before we submit the bio so | |
762 | * we know the end IO handler won't happen before | |
763 | * we inc the count. Otherwise, the cb might get | |
764 | * freed before we're done setting it up | |
765 | */ | |
a50299ae | 766 | refcount_inc(&cb->pending_bios); |
af09abfe | 767 | |
6cbff00f | 768 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
2ff7e61e | 769 | ret = btrfs_lookup_bio_sums(inode, comp_bio, |
10fe6ca8 | 770 | sums); |
79787eaa | 771 | BUG_ON(ret); /* -ENOMEM */ |
d20f7043 | 772 | } |
10fe6ca8 JT |
773 | |
774 | nr_sectors = DIV_ROUND_UP(comp_bio->bi_iter.bi_size, | |
775 | fs_info->sectorsize); | |
776 | sums += csum_size * nr_sectors; | |
d20f7043 | 777 | |
08635bae | 778 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 779 | if (ret) { |
4e4cbee9 | 780 | comp_bio->bi_status = ret; |
4246a0b6 CH |
781 | bio_endio(comp_bio); |
782 | } | |
c8b97818 | 783 | |
e749af44 DS |
784 | comp_bio = btrfs_bio_alloc(cur_disk_byte); |
785 | bio_set_dev(comp_bio, bdev); | |
ebcc3263 | 786 | comp_bio->bi_opf = REQ_OP_READ; |
771ed689 CM |
787 | comp_bio->bi_private = cb; |
788 | comp_bio->bi_end_io = end_compressed_bio_read; | |
789 | ||
09cbfeaf | 790 | bio_add_page(comp_bio, page, PAGE_SIZE, 0); |
c8b97818 | 791 | } |
09cbfeaf | 792 | cur_disk_byte += PAGE_SIZE; |
c8b97818 | 793 | } |
c8b97818 | 794 | |
0b246afa | 795 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 796 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 797 | |
c2db1073 | 798 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
10fe6ca8 | 799 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
79787eaa | 800 | BUG_ON(ret); /* -ENOMEM */ |
c2db1073 | 801 | } |
d20f7043 | 802 | |
08635bae | 803 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num); |
4246a0b6 | 804 | if (ret) { |
4e4cbee9 | 805 | comp_bio->bi_status = ret; |
4246a0b6 CH |
806 | bio_endio(comp_bio); |
807 | } | |
c8b97818 | 808 | |
c8b97818 | 809 | return 0; |
6b82ce8d | 810 | |
811 | fail2: | |
15e3004a JB |
812 | while (faili >= 0) { |
813 | __free_page(cb->compressed_pages[faili]); | |
814 | faili--; | |
815 | } | |
6b82ce8d | 816 | |
817 | kfree(cb->compressed_pages); | |
818 | fail1: | |
819 | kfree(cb); | |
820 | out: | |
821 | free_extent_map(em); | |
822 | return ret; | |
c8b97818 | 823 | } |
261507a0 | 824 | |
17b5a6c1 TT |
825 | /* |
826 | * Heuristic uses systematic sampling to collect data from the input data | |
827 | * range, the logic can be tuned by the following constants: | |
828 | * | |
829 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
830 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
831 | */ | |
832 | #define SAMPLING_READ_SIZE (16) | |
833 | #define SAMPLING_INTERVAL (256) | |
834 | ||
835 | /* | |
836 | * For statistical analysis of the input data we consider bytes that form a | |
837 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
838 | * many times the object appeared in the sample. | |
839 | */ | |
840 | #define BUCKET_SIZE (256) | |
841 | ||
842 | /* | |
843 | * The size of the sample is based on a statistical sampling rule of thumb. | |
844 | * The common way is to perform sampling tests as long as the number of | |
845 | * elements in each cell is at least 5. | |
846 | * | |
847 | * Instead of 5, we choose 32 to obtain more accurate results. | |
848 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
849 | * sample size bound by 8192. | |
850 | * | |
851 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
852 | * from up to 512 locations. | |
853 | */ | |
854 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
855 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
856 | ||
857 | struct bucket_item { | |
858 | u32 count; | |
859 | }; | |
4e439a0b TT |
860 | |
861 | struct heuristic_ws { | |
17b5a6c1 TT |
862 | /* Partial copy of input data */ |
863 | u8 *sample; | |
a440d48c | 864 | u32 sample_size; |
17b5a6c1 TT |
865 | /* Buckets store counters for each byte value */ |
866 | struct bucket_item *bucket; | |
440c840c TT |
867 | /* Sorting buffer */ |
868 | struct bucket_item *bucket_b; | |
4e439a0b TT |
869 | struct list_head list; |
870 | }; | |
871 | ||
92ee5530 DZ |
872 | static struct workspace_manager heuristic_wsm; |
873 | ||
4e439a0b TT |
874 | static void free_heuristic_ws(struct list_head *ws) |
875 | { | |
876 | struct heuristic_ws *workspace; | |
877 | ||
878 | workspace = list_entry(ws, struct heuristic_ws, list); | |
879 | ||
17b5a6c1 TT |
880 | kvfree(workspace->sample); |
881 | kfree(workspace->bucket); | |
440c840c | 882 | kfree(workspace->bucket_b); |
4e439a0b TT |
883 | kfree(workspace); |
884 | } | |
885 | ||
7bf49943 | 886 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
887 | { |
888 | struct heuristic_ws *ws; | |
889 | ||
890 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
891 | if (!ws) | |
892 | return ERR_PTR(-ENOMEM); | |
893 | ||
17b5a6c1 TT |
894 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
895 | if (!ws->sample) | |
896 | goto fail; | |
897 | ||
898 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
899 | if (!ws->bucket) | |
900 | goto fail; | |
4e439a0b | 901 | |
440c840c TT |
902 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
903 | if (!ws->bucket_b) | |
904 | goto fail; | |
905 | ||
17b5a6c1 | 906 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 907 | return &ws->list; |
17b5a6c1 TT |
908 | fail: |
909 | free_heuristic_ws(&ws->list); | |
910 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
911 | } |
912 | ||
ca4ac360 | 913 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 914 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
915 | }; |
916 | ||
e8c9f186 | 917 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
918 | /* The heuristic is represented as compression type 0 */ |
919 | &btrfs_heuristic_compress, | |
261507a0 | 920 | &btrfs_zlib_compress, |
a6fa6fae | 921 | &btrfs_lzo_compress, |
5c1aab1d | 922 | &btrfs_zstd_compress, |
261507a0 LZ |
923 | }; |
924 | ||
c778df14 DS |
925 | static struct list_head *alloc_workspace(int type, unsigned int level) |
926 | { | |
927 | switch (type) { | |
928 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
929 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
930 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
931 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
932 | default: | |
933 | /* | |
934 | * This can't happen, the type is validated several times | |
935 | * before we get here. | |
936 | */ | |
937 | BUG(); | |
938 | } | |
939 | } | |
940 | ||
1e002351 DS |
941 | static void free_workspace(int type, struct list_head *ws) |
942 | { | |
943 | switch (type) { | |
944 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
945 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
946 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
947 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
948 | default: | |
949 | /* | |
950 | * This can't happen, the type is validated several times | |
951 | * before we get here. | |
952 | */ | |
953 | BUG(); | |
954 | } | |
955 | } | |
956 | ||
d5517033 | 957 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 958 | { |
975db483 DS |
959 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; |
960 | struct workspace_manager *wsm = ops->workspace_manager; | |
4e439a0b | 961 | struct list_head *workspace; |
261507a0 | 962 | |
92ee5530 | 963 | wsm->ops = ops; |
10b94a51 | 964 | |
92ee5530 DZ |
965 | INIT_LIST_HEAD(&wsm->idle_ws); |
966 | spin_lock_init(&wsm->ws_lock); | |
967 | atomic_set(&wsm->total_ws, 0); | |
968 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 969 | |
1666edab DZ |
970 | /* |
971 | * Preallocate one workspace for each compression type so we can | |
972 | * guarantee forward progress in the worst case | |
973 | */ | |
c778df14 | 974 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
975 | if (IS_ERR(workspace)) { |
976 | pr_warn( | |
977 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
978 | } else { | |
92ee5530 DZ |
979 | atomic_set(&wsm->total_ws, 1); |
980 | wsm->free_ws = 1; | |
981 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
982 | } |
983 | } | |
984 | ||
2510307e | 985 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 986 | { |
2dba7143 | 987 | struct workspace_manager *wsman; |
1666edab DZ |
988 | struct list_head *ws; |
989 | ||
2dba7143 | 990 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
991 | while (!list_empty(&wsman->idle_ws)) { |
992 | ws = wsman->idle_ws.next; | |
993 | list_del(ws); | |
1e002351 | 994 | free_workspace(type, ws); |
1666edab | 995 | atomic_dec(&wsman->total_ws); |
261507a0 | 996 | } |
261507a0 LZ |
997 | } |
998 | ||
999 | /* | |
e721e49d DS |
1000 | * This finds an available workspace or allocates a new one. |
1001 | * If it's not possible to allocate a new one, waits until there's one. | |
1002 | * Preallocation makes a forward progress guarantees and we do not return | |
1003 | * errors. | |
261507a0 | 1004 | */ |
5907a9bb | 1005 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 1006 | { |
5907a9bb | 1007 | struct workspace_manager *wsm; |
261507a0 LZ |
1008 | struct list_head *workspace; |
1009 | int cpus = num_online_cpus(); | |
fe308533 | 1010 | unsigned nofs_flag; |
4e439a0b TT |
1011 | struct list_head *idle_ws; |
1012 | spinlock_t *ws_lock; | |
1013 | atomic_t *total_ws; | |
1014 | wait_queue_head_t *ws_wait; | |
1015 | int *free_ws; | |
1016 | ||
5907a9bb | 1017 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1018 | idle_ws = &wsm->idle_ws; |
1019 | ws_lock = &wsm->ws_lock; | |
1020 | total_ws = &wsm->total_ws; | |
1021 | ws_wait = &wsm->ws_wait; | |
1022 | free_ws = &wsm->free_ws; | |
261507a0 | 1023 | |
261507a0 | 1024 | again: |
d9187649 BL |
1025 | spin_lock(ws_lock); |
1026 | if (!list_empty(idle_ws)) { | |
1027 | workspace = idle_ws->next; | |
261507a0 | 1028 | list_del(workspace); |
6ac10a6a | 1029 | (*free_ws)--; |
d9187649 | 1030 | spin_unlock(ws_lock); |
261507a0 LZ |
1031 | return workspace; |
1032 | ||
1033 | } | |
6ac10a6a | 1034 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
1035 | DEFINE_WAIT(wait); |
1036 | ||
d9187649 BL |
1037 | spin_unlock(ws_lock); |
1038 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 1039 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 1040 | schedule(); |
d9187649 | 1041 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
1042 | goto again; |
1043 | } | |
6ac10a6a | 1044 | atomic_inc(total_ws); |
d9187649 | 1045 | spin_unlock(ws_lock); |
261507a0 | 1046 | |
fe308533 DS |
1047 | /* |
1048 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
1049 | * to turn it off here because we might get called from the restricted | |
1050 | * context of btrfs_compress_bio/btrfs_compress_pages | |
1051 | */ | |
1052 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1053 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1054 | memalloc_nofs_restore(nofs_flag); |
1055 | ||
261507a0 | 1056 | if (IS_ERR(workspace)) { |
6ac10a6a | 1057 | atomic_dec(total_ws); |
d9187649 | 1058 | wake_up(ws_wait); |
e721e49d DS |
1059 | |
1060 | /* | |
1061 | * Do not return the error but go back to waiting. There's a | |
1062 | * workspace preallocated for each type and the compression | |
1063 | * time is bounded so we get to a workspace eventually. This | |
1064 | * makes our caller's life easier. | |
52356716 DS |
1065 | * |
1066 | * To prevent silent and low-probability deadlocks (when the | |
1067 | * initial preallocation fails), check if there are any | |
1068 | * workspaces at all. | |
e721e49d | 1069 | */ |
52356716 DS |
1070 | if (atomic_read(total_ws) == 0) { |
1071 | static DEFINE_RATELIMIT_STATE(_rs, | |
1072 | /* once per minute */ 60 * HZ, | |
1073 | /* no burst */ 1); | |
1074 | ||
1075 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1076 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1077 | } |
1078 | } | |
e721e49d | 1079 | goto again; |
261507a0 LZ |
1080 | } |
1081 | return workspace; | |
1082 | } | |
1083 | ||
7bf49943 | 1084 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1085 | { |
6a0d1272 | 1086 | switch (type) { |
5907a9bb | 1087 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1088 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1089 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1090 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1091 | default: | |
1092 | /* | |
1093 | * This can't happen, the type is validated several times | |
1094 | * before we get here. | |
1095 | */ | |
1096 | BUG(); | |
1097 | } | |
929f4baf DZ |
1098 | } |
1099 | ||
261507a0 LZ |
1100 | /* |
1101 | * put a workspace struct back on the list or free it if we have enough | |
1102 | * idle ones sitting around | |
1103 | */ | |
a3bbd2a9 | 1104 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1105 | { |
a3bbd2a9 | 1106 | struct workspace_manager *wsm; |
4e439a0b TT |
1107 | struct list_head *idle_ws; |
1108 | spinlock_t *ws_lock; | |
1109 | atomic_t *total_ws; | |
1110 | wait_queue_head_t *ws_wait; | |
1111 | int *free_ws; | |
1112 | ||
a3bbd2a9 | 1113 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1114 | idle_ws = &wsm->idle_ws; |
1115 | ws_lock = &wsm->ws_lock; | |
1116 | total_ws = &wsm->total_ws; | |
1117 | ws_wait = &wsm->ws_wait; | |
1118 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1119 | |
1120 | spin_lock(ws_lock); | |
26b28dce | 1121 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1122 | list_add(ws, idle_ws); |
6ac10a6a | 1123 | (*free_ws)++; |
d9187649 | 1124 | spin_unlock(ws_lock); |
261507a0 LZ |
1125 | goto wake; |
1126 | } | |
d9187649 | 1127 | spin_unlock(ws_lock); |
261507a0 | 1128 | |
1e002351 | 1129 | free_workspace(type, ws); |
6ac10a6a | 1130 | atomic_dec(total_ws); |
261507a0 | 1131 | wake: |
093258e6 | 1132 | cond_wake_up(ws_wait); |
261507a0 LZ |
1133 | } |
1134 | ||
929f4baf DZ |
1135 | static void put_workspace(int type, struct list_head *ws) |
1136 | { | |
bd3a5287 | 1137 | switch (type) { |
a3bbd2a9 DS |
1138 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1139 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1140 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1141 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1142 | default: | |
1143 | /* | |
1144 | * This can't happen, the type is validated several times | |
1145 | * before we get here. | |
1146 | */ | |
1147 | BUG(); | |
1148 | } | |
929f4baf DZ |
1149 | } |
1150 | ||
261507a0 | 1151 | /* |
38c31464 DS |
1152 | * Given an address space and start and length, compress the bytes into @pages |
1153 | * that are allocated on demand. | |
261507a0 | 1154 | * |
f51d2b59 DS |
1155 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1156 | * default the algorithm chooses and is opaque here; | |
1157 | * - compression algo are 0-3 | |
1158 | * - the level are bits 4-7 | |
1159 | * | |
4d3a800e DS |
1160 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1161 | * and returns number of actually allocated pages | |
261507a0 | 1162 | * |
38c31464 DS |
1163 | * @total_in is used to return the number of bytes actually read. It |
1164 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1165 | * ran out of room in the pages array or because we cross the |
1166 | * max_out threshold. | |
1167 | * | |
38c31464 DS |
1168 | * @total_out is an in/out parameter, must be set to the input length and will |
1169 | * be also used to return the total number of compressed bytes | |
261507a0 | 1170 | * |
38c31464 | 1171 | * @max_out tells us the max number of bytes that we're allowed to |
261507a0 LZ |
1172 | * stuff into pages |
1173 | */ | |
f51d2b59 | 1174 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1175 | u64 start, struct page **pages, |
261507a0 LZ |
1176 | unsigned long *out_pages, |
1177 | unsigned long *total_in, | |
e5d74902 | 1178 | unsigned long *total_out) |
261507a0 | 1179 | { |
1972708a | 1180 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1181 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1182 | struct list_head *workspace; |
1183 | int ret; | |
1184 | ||
b0c1fe1e | 1185 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1186 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1187 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1188 | out_pages, total_in, total_out); | |
929f4baf | 1189 | put_workspace(type, workspace); |
261507a0 LZ |
1190 | return ret; |
1191 | } | |
1192 | ||
1193 | /* | |
1194 | * pages_in is an array of pages with compressed data. | |
1195 | * | |
1196 | * disk_start is the starting logical offset of this array in the file | |
1197 | * | |
974b1adc | 1198 | * orig_bio contains the pages from the file that we want to decompress into |
261507a0 LZ |
1199 | * |
1200 | * srclen is the number of bytes in pages_in | |
1201 | * | |
1202 | * The basic idea is that we have a bio that was created by readpages. | |
1203 | * The pages in the bio are for the uncompressed data, and they may not | |
1204 | * be contiguous. They all correspond to the range of bytes covered by | |
1205 | * the compressed extent. | |
1206 | */ | |
8140dc30 | 1207 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1208 | { |
1209 | struct list_head *workspace; | |
1210 | int ret; | |
8140dc30 | 1211 | int type = cb->compress_type; |
261507a0 | 1212 | |
7bf49943 | 1213 | workspace = get_workspace(type, 0); |
1e4eb746 | 1214 | ret = compression_decompress_bio(type, workspace, cb); |
929f4baf | 1215 | put_workspace(type, workspace); |
e1ddce71 | 1216 | |
261507a0 LZ |
1217 | return ret; |
1218 | } | |
1219 | ||
1220 | /* | |
1221 | * a less complex decompression routine. Our compressed data fits in a | |
1222 | * single page, and we want to read a single page out of it. | |
1223 | * start_byte tells us the offset into the compressed data we're interested in | |
1224 | */ | |
1225 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1226 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1227 | { | |
1228 | struct list_head *workspace; | |
1229 | int ret; | |
1230 | ||
7bf49943 | 1231 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1232 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1233 | start_byte, srclen, destlen); | |
929f4baf | 1234 | put_workspace(type, workspace); |
7bf49943 | 1235 | |
261507a0 LZ |
1236 | return ret; |
1237 | } | |
1238 | ||
1666edab DZ |
1239 | void __init btrfs_init_compress(void) |
1240 | { | |
d5517033 DS |
1241 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1242 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1243 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1244 | zstd_init_workspace_manager(); | |
1666edab DZ |
1245 | } |
1246 | ||
e67c718b | 1247 | void __cold btrfs_exit_compress(void) |
261507a0 | 1248 | { |
2510307e DS |
1249 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1250 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1251 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1252 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1253 | } |
3a39c18d LZ |
1254 | |
1255 | /* | |
1256 | * Copy uncompressed data from working buffer to pages. | |
1257 | * | |
1258 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1259 | * | |
1260 | * total_out is the last byte of the buffer | |
1261 | */ | |
14a3357b | 1262 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1263 | unsigned long total_out, u64 disk_start, |
974b1adc | 1264 | struct bio *bio) |
3a39c18d LZ |
1265 | { |
1266 | unsigned long buf_offset; | |
1267 | unsigned long current_buf_start; | |
1268 | unsigned long start_byte; | |
6e78b3f7 | 1269 | unsigned long prev_start_byte; |
3a39c18d LZ |
1270 | unsigned long working_bytes = total_out - buf_start; |
1271 | unsigned long bytes; | |
1272 | char *kaddr; | |
974b1adc | 1273 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1274 | |
1275 | /* | |
1276 | * start byte is the first byte of the page we're currently | |
1277 | * copying into relative to the start of the compressed data. | |
1278 | */ | |
974b1adc | 1279 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1280 | |
1281 | /* we haven't yet hit data corresponding to this page */ | |
1282 | if (total_out <= start_byte) | |
1283 | return 1; | |
1284 | ||
1285 | /* | |
1286 | * the start of the data we care about is offset into | |
1287 | * the middle of our working buffer | |
1288 | */ | |
1289 | if (total_out > start_byte && buf_start < start_byte) { | |
1290 | buf_offset = start_byte - buf_start; | |
1291 | working_bytes -= buf_offset; | |
1292 | } else { | |
1293 | buf_offset = 0; | |
1294 | } | |
1295 | current_buf_start = buf_start; | |
1296 | ||
1297 | /* copy bytes from the working buffer into the pages */ | |
1298 | while (working_bytes > 0) { | |
974b1adc CH |
1299 | bytes = min_t(unsigned long, bvec.bv_len, |
1300 | PAGE_SIZE - buf_offset); | |
3a39c18d | 1301 | bytes = min(bytes, working_bytes); |
974b1adc CH |
1302 | |
1303 | kaddr = kmap_atomic(bvec.bv_page); | |
1304 | memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes); | |
7ac687d9 | 1305 | kunmap_atomic(kaddr); |
974b1adc | 1306 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1307 | |
3a39c18d LZ |
1308 | buf_offset += bytes; |
1309 | working_bytes -= bytes; | |
1310 | current_buf_start += bytes; | |
1311 | ||
1312 | /* check if we need to pick another page */ | |
974b1adc CH |
1313 | bio_advance(bio, bytes); |
1314 | if (!bio->bi_iter.bi_size) | |
1315 | return 0; | |
1316 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1317 | prev_start_byte = start_byte; |
974b1adc | 1318 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1319 | |
974b1adc | 1320 | /* |
6e78b3f7 OS |
1321 | * We need to make sure we're only adjusting |
1322 | * our offset into compression working buffer when | |
1323 | * we're switching pages. Otherwise we can incorrectly | |
1324 | * keep copying when we were actually done. | |
974b1adc | 1325 | */ |
6e78b3f7 OS |
1326 | if (start_byte != prev_start_byte) { |
1327 | /* | |
1328 | * make sure our new page is covered by this | |
1329 | * working buffer | |
1330 | */ | |
1331 | if (total_out <= start_byte) | |
1332 | return 1; | |
3a39c18d | 1333 | |
6e78b3f7 OS |
1334 | /* |
1335 | * the next page in the biovec might not be adjacent | |
1336 | * to the last page, but it might still be found | |
1337 | * inside this working buffer. bump our offset pointer | |
1338 | */ | |
1339 | if (total_out > start_byte && | |
1340 | current_buf_start < start_byte) { | |
1341 | buf_offset = start_byte - buf_start; | |
1342 | working_bytes = total_out - start_byte; | |
1343 | current_buf_start = buf_start + buf_offset; | |
1344 | } | |
3a39c18d LZ |
1345 | } |
1346 | } | |
1347 | ||
1348 | return 1; | |
1349 | } | |
c2fcdcdf | 1350 | |
19562430 TT |
1351 | /* |
1352 | * Shannon Entropy calculation | |
1353 | * | |
52042d8e | 1354 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1355 | * Try calculating entropy to estimate the average minimum number of bits |
1356 | * needed to encode the sampled data. | |
1357 | * | |
1358 | * For convenience, return the percentage of needed bits, instead of amount of | |
1359 | * bits directly. | |
1360 | * | |
1361 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1362 | * and can be compressible with high probability | |
1363 | * | |
1364 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1365 | * | |
1366 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1367 | */ | |
1368 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1369 | #define ENTROPY_LVL_HIGH (80) | |
1370 | ||
1371 | /* | |
1372 | * For increasead precision in shannon_entropy calculation, | |
1373 | * let's do pow(n, M) to save more digits after comma: | |
1374 | * | |
1375 | * - maximum int bit length is 64 | |
1376 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1377 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1378 | * | |
1379 | * So use pow(n, 4). | |
1380 | */ | |
1381 | static inline u32 ilog2_w(u64 n) | |
1382 | { | |
1383 | return ilog2(n * n * n * n); | |
1384 | } | |
1385 | ||
1386 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1387 | { | |
1388 | const u32 entropy_max = 8 * ilog2_w(2); | |
1389 | u32 entropy_sum = 0; | |
1390 | u32 p, p_base, sz_base; | |
1391 | u32 i; | |
1392 | ||
1393 | sz_base = ilog2_w(ws->sample_size); | |
1394 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1395 | p = ws->bucket[i].count; | |
1396 | p_base = ilog2_w(p); | |
1397 | entropy_sum += p * (sz_base - p_base); | |
1398 | } | |
1399 | ||
1400 | entropy_sum /= ws->sample_size; | |
1401 | return entropy_sum * 100 / entropy_max; | |
1402 | } | |
1403 | ||
440c840c TT |
1404 | #define RADIX_BASE 4U |
1405 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1406 | ||
1407 | static u8 get4bits(u64 num, int shift) { | |
1408 | u8 low4bits; | |
1409 | ||
1410 | num >>= shift; | |
1411 | /* Reverse order */ | |
1412 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1413 | return low4bits; | |
1414 | } | |
1415 | ||
440c840c TT |
1416 | /* |
1417 | * Use 4 bits as radix base | |
52042d8e | 1418 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1419 | * |
1420 | * @array - array that will be sorted | |
1421 | * @array_buf - buffer array to store sorting results | |
1422 | * must be equal in size to @array | |
1423 | * @num - array size | |
440c840c | 1424 | */ |
23ae8c63 | 1425 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1426 | int num) |
858177d3 | 1427 | { |
440c840c TT |
1428 | u64 max_num; |
1429 | u64 buf_num; | |
1430 | u32 counters[COUNTERS_SIZE]; | |
1431 | u32 new_addr; | |
1432 | u32 addr; | |
1433 | int bitlen; | |
1434 | int shift; | |
1435 | int i; | |
858177d3 | 1436 | |
440c840c TT |
1437 | /* |
1438 | * Try avoid useless loop iterations for small numbers stored in big | |
1439 | * counters. Example: 48 33 4 ... in 64bit array | |
1440 | */ | |
23ae8c63 | 1441 | max_num = array[0].count; |
440c840c | 1442 | for (i = 1; i < num; i++) { |
23ae8c63 | 1443 | buf_num = array[i].count; |
440c840c TT |
1444 | if (buf_num > max_num) |
1445 | max_num = buf_num; | |
1446 | } | |
1447 | ||
1448 | buf_num = ilog2(max_num); | |
1449 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1450 | ||
1451 | shift = 0; | |
1452 | while (shift < bitlen) { | |
1453 | memset(counters, 0, sizeof(counters)); | |
1454 | ||
1455 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1456 | buf_num = array[i].count; |
440c840c TT |
1457 | addr = get4bits(buf_num, shift); |
1458 | counters[addr]++; | |
1459 | } | |
1460 | ||
1461 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1462 | counters[i] += counters[i - 1]; | |
1463 | ||
1464 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1465 | buf_num = array[i].count; |
440c840c TT |
1466 | addr = get4bits(buf_num, shift); |
1467 | counters[addr]--; | |
1468 | new_addr = counters[addr]; | |
7add17be | 1469 | array_buf[new_addr] = array[i]; |
440c840c TT |
1470 | } |
1471 | ||
1472 | shift += RADIX_BASE; | |
1473 | ||
1474 | /* | |
1475 | * Normal radix expects to move data from a temporary array, to | |
1476 | * the main one. But that requires some CPU time. Avoid that | |
1477 | * by doing another sort iteration to original array instead of | |
1478 | * memcpy() | |
1479 | */ | |
1480 | memset(counters, 0, sizeof(counters)); | |
1481 | ||
1482 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1483 | buf_num = array_buf[i].count; |
440c840c TT |
1484 | addr = get4bits(buf_num, shift); |
1485 | counters[addr]++; | |
1486 | } | |
1487 | ||
1488 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1489 | counters[i] += counters[i - 1]; | |
1490 | ||
1491 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1492 | buf_num = array_buf[i].count; |
440c840c TT |
1493 | addr = get4bits(buf_num, shift); |
1494 | counters[addr]--; | |
1495 | new_addr = counters[addr]; | |
7add17be | 1496 | array[new_addr] = array_buf[i]; |
440c840c TT |
1497 | } |
1498 | ||
1499 | shift += RADIX_BASE; | |
1500 | } | |
858177d3 TT |
1501 | } |
1502 | ||
1503 | /* | |
1504 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1505 | * | |
1506 | * There are several types of structured binary data that use nearly all byte | |
1507 | * values. The distribution can be uniform and counts in all buckets will be | |
1508 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1509 | * | |
1510 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1511 | * be potentially compressible, but we have to take a few more steps to decide | |
1512 | * how much. | |
1513 | * | |
1514 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1515 | * compression algo can easy fix that | |
1516 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1517 | * probability is not compressible | |
1518 | */ | |
1519 | #define BYTE_CORE_SET_LOW (64) | |
1520 | #define BYTE_CORE_SET_HIGH (200) | |
1521 | ||
1522 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1523 | { | |
1524 | u32 i; | |
1525 | u32 coreset_sum = 0; | |
1526 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1527 | struct bucket_item *bucket = ws->bucket; | |
1528 | ||
1529 | /* Sort in reverse order */ | |
36243c91 | 1530 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1531 | |
1532 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1533 | coreset_sum += bucket[i].count; | |
1534 | ||
1535 | if (coreset_sum > core_set_threshold) | |
1536 | return i; | |
1537 | ||
1538 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1539 | coreset_sum += bucket[i].count; | |
1540 | if (coreset_sum > core_set_threshold) | |
1541 | break; | |
1542 | } | |
1543 | ||
1544 | return i; | |
1545 | } | |
1546 | ||
a288e92c TT |
1547 | /* |
1548 | * Count byte values in buckets. | |
1549 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1550 | * Because in most text-like data byte set is restricted to limited number of | |
1551 | * possible characters, and that restriction in most cases makes data easy to | |
1552 | * compress. | |
1553 | * | |
1554 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1555 | * less - compressible | |
1556 | * more - need additional analysis | |
1557 | */ | |
1558 | #define BYTE_SET_THRESHOLD (64) | |
1559 | ||
1560 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1561 | { | |
1562 | u32 i; | |
1563 | u32 byte_set_size = 0; | |
1564 | ||
1565 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1566 | if (ws->bucket[i].count > 0) | |
1567 | byte_set_size++; | |
1568 | } | |
1569 | ||
1570 | /* | |
1571 | * Continue collecting count of byte values in buckets. If the byte | |
1572 | * set size is bigger then the threshold, it's pointless to continue, | |
1573 | * the detection technique would fail for this type of data. | |
1574 | */ | |
1575 | for (; i < BUCKET_SIZE; i++) { | |
1576 | if (ws->bucket[i].count > 0) { | |
1577 | byte_set_size++; | |
1578 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1579 | return byte_set_size; | |
1580 | } | |
1581 | } | |
1582 | ||
1583 | return byte_set_size; | |
1584 | } | |
1585 | ||
1fe4f6fa TT |
1586 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1587 | { | |
1588 | const u32 half_of_sample = ws->sample_size / 2; | |
1589 | const u8 *data = ws->sample; | |
1590 | ||
1591 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1592 | } | |
1593 | ||
a440d48c TT |
1594 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1595 | struct heuristic_ws *ws) | |
1596 | { | |
1597 | struct page *page; | |
1598 | u64 index, index_end; | |
1599 | u32 i, curr_sample_pos; | |
1600 | u8 *in_data; | |
1601 | ||
1602 | /* | |
1603 | * Compression handles the input data by chunks of 128KiB | |
1604 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1605 | * | |
1606 | * We do the same for the heuristic and loop over the whole range. | |
1607 | * | |
1608 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1609 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1610 | */ | |
1611 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1612 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1613 | ||
1614 | index = start >> PAGE_SHIFT; | |
1615 | index_end = end >> PAGE_SHIFT; | |
1616 | ||
1617 | /* Don't miss unaligned end */ | |
1618 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1619 | index_end++; | |
1620 | ||
1621 | curr_sample_pos = 0; | |
1622 | while (index < index_end) { | |
1623 | page = find_get_page(inode->i_mapping, index); | |
1624 | in_data = kmap(page); | |
1625 | /* Handle case where the start is not aligned to PAGE_SIZE */ | |
1626 | i = start % PAGE_SIZE; | |
1627 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1628 | /* Don't sample any garbage from the last page */ | |
1629 | if (start > end - SAMPLING_READ_SIZE) | |
1630 | break; | |
1631 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1632 | SAMPLING_READ_SIZE); | |
1633 | i += SAMPLING_INTERVAL; | |
1634 | start += SAMPLING_INTERVAL; | |
1635 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1636 | } | |
1637 | kunmap(page); | |
1638 | put_page(page); | |
1639 | ||
1640 | index++; | |
1641 | } | |
1642 | ||
1643 | ws->sample_size = curr_sample_pos; | |
1644 | } | |
1645 | ||
c2fcdcdf TT |
1646 | /* |
1647 | * Compression heuristic. | |
1648 | * | |
1649 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1650 | * quickly (compared to direct compression) detect data characteristics | |
1651 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1652 | * data. | |
1653 | * | |
1654 | * The following types of analysis can be performed: | |
1655 | * - detect mostly zero data | |
1656 | * - detect data with low "byte set" size (text, etc) | |
1657 | * - detect data with low/high "core byte" set | |
1658 | * | |
1659 | * Return non-zero if the compression should be done, 0 otherwise. | |
1660 | */ | |
1661 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1662 | { | |
7bf49943 | 1663 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1664 | struct heuristic_ws *ws; |
a440d48c TT |
1665 | u32 i; |
1666 | u8 byte; | |
19562430 | 1667 | int ret = 0; |
c2fcdcdf | 1668 | |
4e439a0b TT |
1669 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1670 | ||
a440d48c TT |
1671 | heuristic_collect_sample(inode, start, end, ws); |
1672 | ||
1fe4f6fa TT |
1673 | if (sample_repeated_patterns(ws)) { |
1674 | ret = 1; | |
1675 | goto out; | |
1676 | } | |
1677 | ||
a440d48c TT |
1678 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1679 | ||
1680 | for (i = 0; i < ws->sample_size; i++) { | |
1681 | byte = ws->sample[i]; | |
1682 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1683 | } |
1684 | ||
a288e92c TT |
1685 | i = byte_set_size(ws); |
1686 | if (i < BYTE_SET_THRESHOLD) { | |
1687 | ret = 2; | |
1688 | goto out; | |
1689 | } | |
1690 | ||
858177d3 TT |
1691 | i = byte_core_set_size(ws); |
1692 | if (i <= BYTE_CORE_SET_LOW) { | |
1693 | ret = 3; | |
1694 | goto out; | |
1695 | } | |
1696 | ||
1697 | if (i >= BYTE_CORE_SET_HIGH) { | |
1698 | ret = 0; | |
1699 | goto out; | |
1700 | } | |
1701 | ||
19562430 TT |
1702 | i = shannon_entropy(ws); |
1703 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1704 | ret = 4; | |
1705 | goto out; | |
1706 | } | |
1707 | ||
1708 | /* | |
1709 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1710 | * needed to give green light to compression. | |
1711 | * | |
1712 | * For now just assume that compression at that level is not worth the | |
1713 | * resources because: | |
1714 | * | |
1715 | * 1. it is possible to defrag the data later | |
1716 | * | |
1717 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1718 | * values, every bucket has counter at level ~54. The heuristic would | |
1719 | * be confused. This can happen when data have some internal repeated | |
1720 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1721 | * pairs of bytes, which is too costly. | |
1722 | */ | |
1723 | if (i < ENTROPY_LVL_HIGH) { | |
1724 | ret = 5; | |
1725 | goto out; | |
1726 | } else { | |
1727 | ret = 0; | |
1728 | goto out; | |
1729 | } | |
1730 | ||
1fe4f6fa | 1731 | out: |
929f4baf | 1732 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1733 | return ret; |
1734 | } | |
f51d2b59 | 1735 | |
d0ab62ce DZ |
1736 | /* |
1737 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1738 | * level, unrecognized string will set the default level | |
1739 | */ | |
1740 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1741 | { |
d0ab62ce DZ |
1742 | unsigned int level = 0; |
1743 | int ret; | |
1744 | ||
1745 | if (!type) | |
f51d2b59 DS |
1746 | return 0; |
1747 | ||
d0ab62ce DZ |
1748 | if (str[0] == ':') { |
1749 | ret = kstrtouint(str + 1, 10, &level); | |
1750 | if (ret) | |
1751 | level = 0; | |
1752 | } | |
1753 | ||
b0c1fe1e DS |
1754 | level = btrfs_compress_set_level(type, level); |
1755 | ||
1756 | return level; | |
1757 | } | |
1758 | ||
1759 | /* | |
1760 | * Adjust @level according to the limits of the compression algorithm or | |
1761 | * fallback to default | |
1762 | */ | |
1763 | unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1764 | { | |
1765 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1766 | ||
1767 | if (level == 0) | |
1768 | level = ops->default_level; | |
1769 | else | |
1770 | level = min(level, ops->max_level); | |
f51d2b59 | 1771 | |
d0ab62ce | 1772 | return level; |
f51d2b59 | 1773 | } |