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> | |
8 | #include <linux/buffer_head.h> | |
9 | #include <linux/file.h> | |
10 | #include <linux/fs.h> | |
11 | #include <linux/pagemap.h> | |
12 | #include <linux/highmem.h> | |
13 | #include <linux/time.h> | |
14 | #include <linux/init.h> | |
15 | #include <linux/string.h> | |
c8b97818 CM |
16 | #include <linux/backing-dev.h> |
17 | #include <linux/mpage.h> | |
18 | #include <linux/swap.h> | |
19 | #include <linux/writeback.h> | |
20 | #include <linux/bit_spinlock.h> | |
5a0e3ad6 | 21 | #include <linux/slab.h> |
fe308533 | 22 | #include <linux/sched/mm.h> |
19562430 | 23 | #include <linux/log2.h> |
c8b97818 CM |
24 | #include "ctree.h" |
25 | #include "disk-io.h" | |
26 | #include "transaction.h" | |
27 | #include "btrfs_inode.h" | |
28 | #include "volumes.h" | |
29 | #include "ordered-data.h" | |
c8b97818 CM |
30 | #include "compression.h" |
31 | #include "extent_io.h" | |
32 | #include "extent_map.h" | |
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]; | |
44 | } | |
45 | ||
46 | return NULL; | |
47 | } | |
48 | ||
8140dc30 | 49 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 50 | |
2ff7e61e | 51 | static inline int compressed_bio_size(struct btrfs_fs_info *fs_info, |
d20f7043 CM |
52 | unsigned long disk_size) |
53 | { | |
0b246afa | 54 | u16 csum_size = btrfs_super_csum_size(fs_info->super_copy); |
6c41761f | 55 | |
d20f7043 | 56 | return sizeof(struct compressed_bio) + |
0b246afa | 57 | (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size; |
d20f7043 CM |
58 | } |
59 | ||
f898ac6a | 60 | static int check_compressed_csum(struct btrfs_inode *inode, |
d20f7043 CM |
61 | struct compressed_bio *cb, |
62 | u64 disk_start) | |
63 | { | |
64 | int ret; | |
d20f7043 CM |
65 | struct page *page; |
66 | unsigned long i; | |
67 | char *kaddr; | |
68 | u32 csum; | |
69 | u32 *cb_sum = &cb->sums; | |
70 | ||
f898ac6a | 71 | if (inode->flags & BTRFS_INODE_NODATASUM) |
d20f7043 CM |
72 | return 0; |
73 | ||
74 | for (i = 0; i < cb->nr_pages; i++) { | |
75 | page = cb->compressed_pages[i]; | |
76 | csum = ~(u32)0; | |
77 | ||
7ac687d9 | 78 | kaddr = kmap_atomic(page); |
09cbfeaf | 79 | csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE); |
0b5e3daf | 80 | btrfs_csum_final(csum, (u8 *)&csum); |
7ac687d9 | 81 | kunmap_atomic(kaddr); |
d20f7043 CM |
82 | |
83 | if (csum != *cb_sum) { | |
f898ac6a | 84 | btrfs_print_data_csum_error(inode, disk_start, csum, |
0970a22e | 85 | *cb_sum, cb->mirror_num); |
d20f7043 CM |
86 | ret = -EIO; |
87 | goto fail; | |
88 | } | |
89 | cb_sum++; | |
90 | ||
91 | } | |
92 | ret = 0; | |
93 | fail: | |
94 | return ret; | |
95 | } | |
96 | ||
c8b97818 CM |
97 | /* when we finish reading compressed pages from the disk, we |
98 | * decompress them and then run the bio end_io routines on the | |
99 | * decompressed pages (in the inode address space). | |
100 | * | |
101 | * This allows the checksumming and other IO error handling routines | |
102 | * to work normally | |
103 | * | |
104 | * The compressed pages are freed here, and it must be run | |
105 | * in process context | |
106 | */ | |
4246a0b6 | 107 | static void end_compressed_bio_read(struct bio *bio) |
c8b97818 | 108 | { |
c8b97818 CM |
109 | struct compressed_bio *cb = bio->bi_private; |
110 | struct inode *inode; | |
111 | struct page *page; | |
112 | unsigned long index; | |
cf1167d5 | 113 | unsigned int mirror = btrfs_io_bio(bio)->mirror_num; |
e6311f24 | 114 | int ret = 0; |
c8b97818 | 115 | |
4e4cbee9 | 116 | if (bio->bi_status) |
c8b97818 CM |
117 | cb->errors = 1; |
118 | ||
119 | /* if there are more bios still pending for this compressed | |
120 | * extent, just exit | |
121 | */ | |
a50299ae | 122 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
123 | goto out; |
124 | ||
cf1167d5 LB |
125 | /* |
126 | * Record the correct mirror_num in cb->orig_bio so that | |
127 | * read-repair can work properly. | |
128 | */ | |
129 | ASSERT(btrfs_io_bio(cb->orig_bio)); | |
130 | btrfs_io_bio(cb->orig_bio)->mirror_num = mirror; | |
131 | cb->mirror_num = mirror; | |
132 | ||
e6311f24 LB |
133 | /* |
134 | * Some IO in this cb have failed, just skip checksum as there | |
135 | * is no way it could be correct. | |
136 | */ | |
137 | if (cb->errors == 1) | |
138 | goto csum_failed; | |
139 | ||
d20f7043 | 140 | inode = cb->inode; |
f898ac6a | 141 | ret = check_compressed_csum(BTRFS_I(inode), cb, |
4f024f37 | 142 | (u64)bio->bi_iter.bi_sector << 9); |
d20f7043 CM |
143 | if (ret) |
144 | goto csum_failed; | |
145 | ||
c8b97818 CM |
146 | /* ok, we're the last bio for this extent, lets start |
147 | * the decompression. | |
148 | */ | |
8140dc30 AJ |
149 | ret = btrfs_decompress_bio(cb); |
150 | ||
d20f7043 | 151 | csum_failed: |
c8b97818 CM |
152 | if (ret) |
153 | cb->errors = 1; | |
154 | ||
155 | /* release the compressed pages */ | |
156 | index = 0; | |
157 | for (index = 0; index < cb->nr_pages; index++) { | |
158 | page = cb->compressed_pages[index]; | |
159 | page->mapping = NULL; | |
09cbfeaf | 160 | put_page(page); |
c8b97818 CM |
161 | } |
162 | ||
163 | /* do io completion on the original bio */ | |
771ed689 | 164 | if (cb->errors) { |
c8b97818 | 165 | bio_io_error(cb->orig_bio); |
d20f7043 | 166 | } else { |
2c30c71b KO |
167 | int i; |
168 | struct bio_vec *bvec; | |
d20f7043 CM |
169 | |
170 | /* | |
171 | * we have verified the checksum already, set page | |
172 | * checked so the end_io handlers know about it | |
173 | */ | |
c09abff8 | 174 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
2c30c71b | 175 | bio_for_each_segment_all(bvec, cb->orig_bio, i) |
d20f7043 | 176 | SetPageChecked(bvec->bv_page); |
2c30c71b | 177 | |
4246a0b6 | 178 | bio_endio(cb->orig_bio); |
d20f7043 | 179 | } |
c8b97818 CM |
180 | |
181 | /* finally free the cb struct */ | |
182 | kfree(cb->compressed_pages); | |
183 | kfree(cb); | |
184 | out: | |
185 | bio_put(bio); | |
186 | } | |
187 | ||
188 | /* | |
189 | * Clear the writeback bits on all of the file | |
190 | * pages for a compressed write | |
191 | */ | |
7bdcefc1 FM |
192 | static noinline void end_compressed_writeback(struct inode *inode, |
193 | const struct compressed_bio *cb) | |
c8b97818 | 194 | { |
09cbfeaf KS |
195 | unsigned long index = cb->start >> PAGE_SHIFT; |
196 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
c8b97818 CM |
197 | struct page *pages[16]; |
198 | unsigned long nr_pages = end_index - index + 1; | |
199 | int i; | |
200 | int ret; | |
201 | ||
7bdcefc1 FM |
202 | if (cb->errors) |
203 | mapping_set_error(inode->i_mapping, -EIO); | |
204 | ||
d397712b | 205 | while (nr_pages > 0) { |
c8b97818 | 206 | ret = find_get_pages_contig(inode->i_mapping, index, |
5b050f04 CM |
207 | min_t(unsigned long, |
208 | nr_pages, ARRAY_SIZE(pages)), pages); | |
c8b97818 CM |
209 | if (ret == 0) { |
210 | nr_pages -= 1; | |
211 | index += 1; | |
212 | continue; | |
213 | } | |
214 | for (i = 0; i < ret; i++) { | |
7bdcefc1 FM |
215 | if (cb->errors) |
216 | SetPageError(pages[i]); | |
c8b97818 | 217 | end_page_writeback(pages[i]); |
09cbfeaf | 218 | put_page(pages[i]); |
c8b97818 CM |
219 | } |
220 | nr_pages -= ret; | |
221 | index += ret; | |
222 | } | |
223 | /* the inode may be gone now */ | |
c8b97818 CM |
224 | } |
225 | ||
226 | /* | |
227 | * do the cleanup once all the compressed pages hit the disk. | |
228 | * This will clear writeback on the file pages and free the compressed | |
229 | * pages. | |
230 | * | |
231 | * This also calls the writeback end hooks for the file pages so that | |
232 | * metadata and checksums can be updated in the file. | |
233 | */ | |
4246a0b6 | 234 | static void end_compressed_bio_write(struct bio *bio) |
c8b97818 CM |
235 | { |
236 | struct extent_io_tree *tree; | |
237 | struct compressed_bio *cb = bio->bi_private; | |
238 | struct inode *inode; | |
239 | struct page *page; | |
240 | unsigned long index; | |
241 | ||
4e4cbee9 | 242 | if (bio->bi_status) |
c8b97818 CM |
243 | cb->errors = 1; |
244 | ||
245 | /* if there are more bios still pending for this compressed | |
246 | * extent, just exit | |
247 | */ | |
a50299ae | 248 | if (!refcount_dec_and_test(&cb->pending_bios)) |
c8b97818 CM |
249 | goto out; |
250 | ||
251 | /* ok, we're the last bio for this extent, step one is to | |
252 | * call back into the FS and do all the end_io operations | |
253 | */ | |
254 | inode = cb->inode; | |
255 | tree = &BTRFS_I(inode)->io_tree; | |
70b99e69 | 256 | cb->compressed_pages[0]->mapping = cb->inode->i_mapping; |
c8b97818 CM |
257 | tree->ops->writepage_end_io_hook(cb->compressed_pages[0], |
258 | cb->start, | |
259 | cb->start + cb->len - 1, | |
7bdcefc1 | 260 | NULL, |
2dbe0c77 AJ |
261 | bio->bi_status ? |
262 | BLK_STS_OK : BLK_STS_NOTSUPP); | |
70b99e69 | 263 | cb->compressed_pages[0]->mapping = NULL; |
c8b97818 | 264 | |
7bdcefc1 | 265 | end_compressed_writeback(inode, cb); |
c8b97818 CM |
266 | /* note, our inode could be gone now */ |
267 | ||
268 | /* | |
269 | * release the compressed pages, these came from alloc_page and | |
270 | * are not attached to the inode at all | |
271 | */ | |
272 | index = 0; | |
273 | for (index = 0; index < cb->nr_pages; index++) { | |
274 | page = cb->compressed_pages[index]; | |
275 | page->mapping = NULL; | |
09cbfeaf | 276 | put_page(page); |
c8b97818 CM |
277 | } |
278 | ||
279 | /* finally free the cb struct */ | |
280 | kfree(cb->compressed_pages); | |
281 | kfree(cb); | |
282 | out: | |
283 | bio_put(bio); | |
284 | } | |
285 | ||
286 | /* | |
287 | * worker function to build and submit bios for previously compressed pages. | |
288 | * The corresponding pages in the inode should be marked for writeback | |
289 | * and the compressed pages should have a reference on them for dropping | |
290 | * when the IO is complete. | |
291 | * | |
292 | * This also checksums the file bytes and gets things ready for | |
293 | * the end io hooks. | |
294 | */ | |
4e4cbee9 | 295 | blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start, |
c8b97818 CM |
296 | unsigned long len, u64 disk_start, |
297 | unsigned long compressed_len, | |
298 | struct page **compressed_pages, | |
f82b7359 LB |
299 | unsigned long nr_pages, |
300 | unsigned int write_flags) | |
c8b97818 | 301 | { |
0b246afa | 302 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 | 303 | struct bio *bio = NULL; |
c8b97818 CM |
304 | struct compressed_bio *cb; |
305 | unsigned long bytes_left; | |
306 | struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree; | |
306e16ce | 307 | int pg_index = 0; |
c8b97818 CM |
308 | struct page *page; |
309 | u64 first_byte = disk_start; | |
310 | struct block_device *bdev; | |
4e4cbee9 | 311 | blk_status_t ret; |
e55179b3 | 312 | int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM; |
c8b97818 | 313 | |
09cbfeaf | 314 | WARN_ON(start & ((u64)PAGE_SIZE - 1)); |
2ff7e61e | 315 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
dac97e51 | 316 | if (!cb) |
4e4cbee9 | 317 | return BLK_STS_RESOURCE; |
a50299ae | 318 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
319 | cb->errors = 0; |
320 | cb->inode = inode; | |
321 | cb->start = start; | |
322 | cb->len = len; | |
d20f7043 | 323 | cb->mirror_num = 0; |
c8b97818 CM |
324 | cb->compressed_pages = compressed_pages; |
325 | cb->compressed_len = compressed_len; | |
326 | cb->orig_bio = NULL; | |
327 | cb->nr_pages = nr_pages; | |
328 | ||
0b246afa | 329 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 330 | |
c821e7f3 | 331 | bio = btrfs_bio_alloc(bdev, first_byte); |
f82b7359 | 332 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
333 | bio->bi_private = cb; |
334 | bio->bi_end_io = end_compressed_bio_write; | |
a50299ae | 335 | refcount_set(&cb->pending_bios, 1); |
c8b97818 CM |
336 | |
337 | /* create and submit bios for the compressed pages */ | |
338 | bytes_left = compressed_len; | |
306e16ce | 339 | for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) { |
4e4cbee9 CH |
340 | int submit = 0; |
341 | ||
306e16ce | 342 | page = compressed_pages[pg_index]; |
c8b97818 | 343 | page->mapping = inode->i_mapping; |
4f024f37 | 344 | if (bio->bi_iter.bi_size) |
4e4cbee9 | 345 | submit = io_tree->ops->merge_bio_hook(page, 0, |
09cbfeaf | 346 | PAGE_SIZE, |
c8b97818 | 347 | bio, 0); |
c8b97818 | 348 | |
70b99e69 | 349 | page->mapping = NULL; |
4e4cbee9 | 350 | if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 351 | PAGE_SIZE) { |
af09abfe CM |
352 | /* |
353 | * inc the count before we submit the bio so | |
354 | * we know the end IO handler won't happen before | |
355 | * we inc the count. Otherwise, the cb might get | |
356 | * freed before we're done setting it up | |
357 | */ | |
a50299ae | 358 | refcount_inc(&cb->pending_bios); |
0b246afa JM |
359 | ret = btrfs_bio_wq_end_io(fs_info, bio, |
360 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 361 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 362 | |
e55179b3 | 363 | if (!skip_sum) { |
2ff7e61e | 364 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 365 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 366 | } |
d20f7043 | 367 | |
2ff7e61e | 368 | ret = btrfs_map_bio(fs_info, bio, 0, 1); |
f5daf2c7 | 369 | if (ret) { |
4e4cbee9 | 370 | bio->bi_status = ret; |
f5daf2c7 LB |
371 | bio_endio(bio); |
372 | } | |
c8b97818 | 373 | |
c821e7f3 | 374 | bio = btrfs_bio_alloc(bdev, first_byte); |
f82b7359 | 375 | bio->bi_opf = REQ_OP_WRITE | write_flags; |
c8b97818 CM |
376 | bio->bi_private = cb; |
377 | bio->bi_end_io = end_compressed_bio_write; | |
09cbfeaf | 378 | bio_add_page(bio, page, PAGE_SIZE, 0); |
c8b97818 | 379 | } |
09cbfeaf | 380 | if (bytes_left < PAGE_SIZE) { |
0b246afa | 381 | btrfs_info(fs_info, |
efe120a0 | 382 | "bytes left %lu compress len %lu nr %lu", |
cfbc246e CM |
383 | bytes_left, cb->compressed_len, cb->nr_pages); |
384 | } | |
09cbfeaf KS |
385 | bytes_left -= PAGE_SIZE; |
386 | first_byte += PAGE_SIZE; | |
771ed689 | 387 | cond_resched(); |
c8b97818 | 388 | } |
c8b97818 | 389 | |
0b246afa | 390 | ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 391 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 392 | |
e55179b3 | 393 | if (!skip_sum) { |
2ff7e61e | 394 | ret = btrfs_csum_one_bio(inode, bio, start, 1); |
79787eaa | 395 | BUG_ON(ret); /* -ENOMEM */ |
e55179b3 | 396 | } |
d20f7043 | 397 | |
2ff7e61e | 398 | ret = btrfs_map_bio(fs_info, bio, 0, 1); |
f5daf2c7 | 399 | if (ret) { |
4e4cbee9 | 400 | bio->bi_status = ret; |
f5daf2c7 LB |
401 | bio_endio(bio); |
402 | } | |
c8b97818 | 403 | |
c8b97818 CM |
404 | return 0; |
405 | } | |
406 | ||
2a4d0c90 CH |
407 | static u64 bio_end_offset(struct bio *bio) |
408 | { | |
c45a8f2d | 409 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
410 | |
411 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
412 | } | |
413 | ||
771ed689 CM |
414 | static noinline int add_ra_bio_pages(struct inode *inode, |
415 | u64 compressed_end, | |
416 | struct compressed_bio *cb) | |
417 | { | |
418 | unsigned long end_index; | |
306e16ce | 419 | unsigned long pg_index; |
771ed689 CM |
420 | u64 last_offset; |
421 | u64 isize = i_size_read(inode); | |
422 | int ret; | |
423 | struct page *page; | |
424 | unsigned long nr_pages = 0; | |
425 | struct extent_map *em; | |
426 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
427 | struct extent_map_tree *em_tree; |
428 | struct extent_io_tree *tree; | |
429 | u64 end; | |
430 | int misses = 0; | |
431 | ||
2a4d0c90 | 432 | last_offset = bio_end_offset(cb->orig_bio); |
771ed689 CM |
433 | em_tree = &BTRFS_I(inode)->extent_tree; |
434 | tree = &BTRFS_I(inode)->io_tree; | |
435 | ||
436 | if (isize == 0) | |
437 | return 0; | |
438 | ||
09cbfeaf | 439 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 440 | |
d397712b | 441 | while (last_offset < compressed_end) { |
09cbfeaf | 442 | pg_index = last_offset >> PAGE_SHIFT; |
771ed689 | 443 | |
306e16ce | 444 | if (pg_index > end_index) |
771ed689 CM |
445 | break; |
446 | ||
447 | rcu_read_lock(); | |
b93b0163 | 448 | page = radix_tree_lookup(&mapping->i_pages, pg_index); |
771ed689 | 449 | rcu_read_unlock(); |
0cd6144a | 450 | if (page && !radix_tree_exceptional_entry(page)) { |
771ed689 CM |
451 | misses++; |
452 | if (misses > 4) | |
453 | break; | |
454 | goto next; | |
455 | } | |
456 | ||
c62d2555 MH |
457 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
458 | ~__GFP_FS)); | |
771ed689 CM |
459 | if (!page) |
460 | break; | |
461 | ||
c62d2555 | 462 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 463 | put_page(page); |
771ed689 CM |
464 | goto next; |
465 | } | |
466 | ||
09cbfeaf | 467 | end = last_offset + PAGE_SIZE - 1; |
771ed689 CM |
468 | /* |
469 | * at this point, we have a locked page in the page cache | |
470 | * for these bytes in the file. But, we have to make | |
471 | * sure they map to this compressed extent on disk. | |
472 | */ | |
473 | set_page_extent_mapped(page); | |
d0082371 | 474 | lock_extent(tree, last_offset, end); |
890871be | 475 | read_lock(&em_tree->lock); |
771ed689 | 476 | em = lookup_extent_mapping(em_tree, last_offset, |
09cbfeaf | 477 | PAGE_SIZE); |
890871be | 478 | read_unlock(&em_tree->lock); |
771ed689 CM |
479 | |
480 | if (!em || last_offset < em->start || | |
09cbfeaf | 481 | (last_offset + PAGE_SIZE > extent_map_end(em)) || |
4f024f37 | 482 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 483 | free_extent_map(em); |
d0082371 | 484 | unlock_extent(tree, last_offset, end); |
771ed689 | 485 | unlock_page(page); |
09cbfeaf | 486 | put_page(page); |
771ed689 CM |
487 | break; |
488 | } | |
489 | free_extent_map(em); | |
490 | ||
491 | if (page->index == end_index) { | |
492 | char *userpage; | |
09cbfeaf | 493 | size_t zero_offset = isize & (PAGE_SIZE - 1); |
771ed689 CM |
494 | |
495 | if (zero_offset) { | |
496 | int zeros; | |
09cbfeaf | 497 | zeros = PAGE_SIZE - zero_offset; |
7ac687d9 | 498 | userpage = kmap_atomic(page); |
771ed689 CM |
499 | memset(userpage + zero_offset, 0, zeros); |
500 | flush_dcache_page(page); | |
7ac687d9 | 501 | kunmap_atomic(userpage); |
771ed689 CM |
502 | } |
503 | } | |
504 | ||
505 | ret = bio_add_page(cb->orig_bio, page, | |
09cbfeaf | 506 | PAGE_SIZE, 0); |
771ed689 | 507 | |
09cbfeaf | 508 | if (ret == PAGE_SIZE) { |
771ed689 | 509 | nr_pages++; |
09cbfeaf | 510 | put_page(page); |
771ed689 | 511 | } else { |
d0082371 | 512 | unlock_extent(tree, last_offset, end); |
771ed689 | 513 | unlock_page(page); |
09cbfeaf | 514 | put_page(page); |
771ed689 CM |
515 | break; |
516 | } | |
517 | next: | |
09cbfeaf | 518 | last_offset += PAGE_SIZE; |
771ed689 | 519 | } |
771ed689 CM |
520 | return 0; |
521 | } | |
522 | ||
c8b97818 CM |
523 | /* |
524 | * for a compressed read, the bio we get passed has all the inode pages | |
525 | * in it. We don't actually do IO on those pages but allocate new ones | |
526 | * to hold the compressed pages on disk. | |
527 | * | |
4f024f37 | 528 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 529 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
530 | * |
531 | * After the compressed pages are read, we copy the bytes into the | |
532 | * bio we were passed and then call the bio end_io calls | |
533 | */ | |
4e4cbee9 | 534 | blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
c8b97818 CM |
535 | int mirror_num, unsigned long bio_flags) |
536 | { | |
0b246afa | 537 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
538 | struct extent_io_tree *tree; |
539 | struct extent_map_tree *em_tree; | |
540 | struct compressed_bio *cb; | |
c8b97818 CM |
541 | unsigned long compressed_len; |
542 | unsigned long nr_pages; | |
306e16ce | 543 | unsigned long pg_index; |
c8b97818 CM |
544 | struct page *page; |
545 | struct block_device *bdev; | |
546 | struct bio *comp_bio; | |
4f024f37 | 547 | u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9; |
e04ca626 CM |
548 | u64 em_len; |
549 | u64 em_start; | |
c8b97818 | 550 | struct extent_map *em; |
4e4cbee9 | 551 | blk_status_t ret = BLK_STS_RESOURCE; |
15e3004a | 552 | int faili = 0; |
d20f7043 | 553 | u32 *sums; |
c8b97818 CM |
554 | |
555 | tree = &BTRFS_I(inode)->io_tree; | |
556 | em_tree = &BTRFS_I(inode)->extent_tree; | |
557 | ||
558 | /* we need the actual starting offset of this extent in the file */ | |
890871be | 559 | read_lock(&em_tree->lock); |
c8b97818 | 560 | em = lookup_extent_mapping(em_tree, |
263663cd | 561 | page_offset(bio_first_page_all(bio)), |
09cbfeaf | 562 | PAGE_SIZE); |
890871be | 563 | read_unlock(&em_tree->lock); |
285190d9 | 564 | if (!em) |
4e4cbee9 | 565 | return BLK_STS_IOERR; |
c8b97818 | 566 | |
d20f7043 | 567 | compressed_len = em->block_len; |
2ff7e61e | 568 | cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS); |
6b82ce8d | 569 | if (!cb) |
570 | goto out; | |
571 | ||
a50299ae | 572 | refcount_set(&cb->pending_bios, 0); |
c8b97818 CM |
573 | cb->errors = 0; |
574 | cb->inode = inode; | |
d20f7043 CM |
575 | cb->mirror_num = mirror_num; |
576 | sums = &cb->sums; | |
c8b97818 | 577 | |
ff5b7ee3 | 578 | cb->start = em->orig_start; |
e04ca626 CM |
579 | em_len = em->len; |
580 | em_start = em->start; | |
d20f7043 | 581 | |
c8b97818 | 582 | free_extent_map(em); |
e04ca626 | 583 | em = NULL; |
c8b97818 | 584 | |
81381053 | 585 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 586 | cb->compressed_len = compressed_len; |
261507a0 | 587 | cb->compress_type = extent_compress_type(bio_flags); |
c8b97818 CM |
588 | cb->orig_bio = bio; |
589 | ||
09cbfeaf | 590 | nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
31e818fe | 591 | cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *), |
c8b97818 | 592 | GFP_NOFS); |
6b82ce8d | 593 | if (!cb->compressed_pages) |
594 | goto fail1; | |
595 | ||
0b246afa | 596 | bdev = fs_info->fs_devices->latest_bdev; |
c8b97818 | 597 | |
306e16ce DS |
598 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
599 | cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS | | |
c8b97818 | 600 | __GFP_HIGHMEM); |
15e3004a JB |
601 | if (!cb->compressed_pages[pg_index]) { |
602 | faili = pg_index - 1; | |
0e9350de | 603 | ret = BLK_STS_RESOURCE; |
6b82ce8d | 604 | goto fail2; |
15e3004a | 605 | } |
c8b97818 | 606 | } |
15e3004a | 607 | faili = nr_pages - 1; |
c8b97818 CM |
608 | cb->nr_pages = nr_pages; |
609 | ||
7f042a83 | 610 | add_ra_bio_pages(inode, em_start + em_len, cb); |
771ed689 | 611 | |
771ed689 | 612 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 613 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 614 | |
c821e7f3 | 615 | comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte); |
37226b21 | 616 | bio_set_op_attrs (comp_bio, REQ_OP_READ, 0); |
c8b97818 CM |
617 | comp_bio->bi_private = cb; |
618 | comp_bio->bi_end_io = end_compressed_bio_read; | |
a50299ae | 619 | refcount_set(&cb->pending_bios, 1); |
c8b97818 | 620 | |
306e16ce | 621 | for (pg_index = 0; pg_index < nr_pages; pg_index++) { |
4e4cbee9 CH |
622 | int submit = 0; |
623 | ||
306e16ce | 624 | page = cb->compressed_pages[pg_index]; |
c8b97818 | 625 | page->mapping = inode->i_mapping; |
09cbfeaf | 626 | page->index = em_start >> PAGE_SHIFT; |
d20f7043 | 627 | |
4f024f37 | 628 | if (comp_bio->bi_iter.bi_size) |
4e4cbee9 | 629 | submit = tree->ops->merge_bio_hook(page, 0, |
09cbfeaf | 630 | PAGE_SIZE, |
c8b97818 | 631 | comp_bio, 0); |
c8b97818 | 632 | |
70b99e69 | 633 | page->mapping = NULL; |
4e4cbee9 | 634 | if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) < |
09cbfeaf | 635 | PAGE_SIZE) { |
0b246afa JM |
636 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, |
637 | BTRFS_WQ_ENDIO_DATA); | |
79787eaa | 638 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 639 | |
af09abfe CM |
640 | /* |
641 | * inc the count before we submit the bio so | |
642 | * we know the end IO handler won't happen before | |
643 | * we inc the count. Otherwise, the cb might get | |
644 | * freed before we're done setting it up | |
645 | */ | |
a50299ae | 646 | refcount_inc(&cb->pending_bios); |
af09abfe | 647 | |
6cbff00f | 648 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
2ff7e61e JM |
649 | ret = btrfs_lookup_bio_sums(inode, comp_bio, |
650 | sums); | |
79787eaa | 651 | BUG_ON(ret); /* -ENOMEM */ |
d20f7043 | 652 | } |
ed6078f7 | 653 | sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size, |
0b246afa | 654 | fs_info->sectorsize); |
d20f7043 | 655 | |
2ff7e61e | 656 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0); |
4246a0b6 | 657 | if (ret) { |
4e4cbee9 | 658 | comp_bio->bi_status = ret; |
4246a0b6 CH |
659 | bio_endio(comp_bio); |
660 | } | |
c8b97818 | 661 | |
c821e7f3 | 662 | comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte); |
37226b21 | 663 | bio_set_op_attrs(comp_bio, REQ_OP_READ, 0); |
771ed689 CM |
664 | comp_bio->bi_private = cb; |
665 | comp_bio->bi_end_io = end_compressed_bio_read; | |
666 | ||
09cbfeaf | 667 | bio_add_page(comp_bio, page, PAGE_SIZE, 0); |
c8b97818 | 668 | } |
09cbfeaf | 669 | cur_disk_byte += PAGE_SIZE; |
c8b97818 | 670 | } |
c8b97818 | 671 | |
0b246afa | 672 | ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA); |
79787eaa | 673 | BUG_ON(ret); /* -ENOMEM */ |
c8b97818 | 674 | |
c2db1073 | 675 | if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) { |
2ff7e61e | 676 | ret = btrfs_lookup_bio_sums(inode, comp_bio, sums); |
79787eaa | 677 | BUG_ON(ret); /* -ENOMEM */ |
c2db1073 | 678 | } |
d20f7043 | 679 | |
2ff7e61e | 680 | ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0); |
4246a0b6 | 681 | if (ret) { |
4e4cbee9 | 682 | comp_bio->bi_status = ret; |
4246a0b6 CH |
683 | bio_endio(comp_bio); |
684 | } | |
c8b97818 | 685 | |
c8b97818 | 686 | return 0; |
6b82ce8d | 687 | |
688 | fail2: | |
15e3004a JB |
689 | while (faili >= 0) { |
690 | __free_page(cb->compressed_pages[faili]); | |
691 | faili--; | |
692 | } | |
6b82ce8d | 693 | |
694 | kfree(cb->compressed_pages); | |
695 | fail1: | |
696 | kfree(cb); | |
697 | out: | |
698 | free_extent_map(em); | |
699 | return ret; | |
c8b97818 | 700 | } |
261507a0 | 701 | |
17b5a6c1 TT |
702 | /* |
703 | * Heuristic uses systematic sampling to collect data from the input data | |
704 | * range, the logic can be tuned by the following constants: | |
705 | * | |
706 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
707 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
708 | */ | |
709 | #define SAMPLING_READ_SIZE (16) | |
710 | #define SAMPLING_INTERVAL (256) | |
711 | ||
712 | /* | |
713 | * For statistical analysis of the input data we consider bytes that form a | |
714 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
715 | * many times the object appeared in the sample. | |
716 | */ | |
717 | #define BUCKET_SIZE (256) | |
718 | ||
719 | /* | |
720 | * The size of the sample is based on a statistical sampling rule of thumb. | |
721 | * The common way is to perform sampling tests as long as the number of | |
722 | * elements in each cell is at least 5. | |
723 | * | |
724 | * Instead of 5, we choose 32 to obtain more accurate results. | |
725 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
726 | * sample size bound by 8192. | |
727 | * | |
728 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
729 | * from up to 512 locations. | |
730 | */ | |
731 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
732 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
733 | ||
734 | struct bucket_item { | |
735 | u32 count; | |
736 | }; | |
4e439a0b TT |
737 | |
738 | struct heuristic_ws { | |
17b5a6c1 TT |
739 | /* Partial copy of input data */ |
740 | u8 *sample; | |
a440d48c | 741 | u32 sample_size; |
17b5a6c1 TT |
742 | /* Buckets store counters for each byte value */ |
743 | struct bucket_item *bucket; | |
440c840c TT |
744 | /* Sorting buffer */ |
745 | struct bucket_item *bucket_b; | |
4e439a0b TT |
746 | struct list_head list; |
747 | }; | |
748 | ||
749 | static void free_heuristic_ws(struct list_head *ws) | |
750 | { | |
751 | struct heuristic_ws *workspace; | |
752 | ||
753 | workspace = list_entry(ws, struct heuristic_ws, list); | |
754 | ||
17b5a6c1 TT |
755 | kvfree(workspace->sample); |
756 | kfree(workspace->bucket); | |
440c840c | 757 | kfree(workspace->bucket_b); |
4e439a0b TT |
758 | kfree(workspace); |
759 | } | |
760 | ||
761 | static struct list_head *alloc_heuristic_ws(void) | |
762 | { | |
763 | struct heuristic_ws *ws; | |
764 | ||
765 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
766 | if (!ws) | |
767 | return ERR_PTR(-ENOMEM); | |
768 | ||
17b5a6c1 TT |
769 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
770 | if (!ws->sample) | |
771 | goto fail; | |
772 | ||
773 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
774 | if (!ws->bucket) | |
775 | goto fail; | |
4e439a0b | 776 | |
440c840c TT |
777 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
778 | if (!ws->bucket_b) | |
779 | goto fail; | |
780 | ||
17b5a6c1 | 781 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 782 | return &ws->list; |
17b5a6c1 TT |
783 | fail: |
784 | free_heuristic_ws(&ws->list); | |
785 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
786 | } |
787 | ||
788 | struct workspaces_list { | |
d9187649 BL |
789 | struct list_head idle_ws; |
790 | spinlock_t ws_lock; | |
6ac10a6a DS |
791 | /* Number of free workspaces */ |
792 | int free_ws; | |
793 | /* Total number of allocated workspaces */ | |
794 | atomic_t total_ws; | |
795 | /* Waiters for a free workspace */ | |
d9187649 | 796 | wait_queue_head_t ws_wait; |
4e439a0b TT |
797 | }; |
798 | ||
799 | static struct workspaces_list btrfs_comp_ws[BTRFS_COMPRESS_TYPES]; | |
800 | ||
801 | static struct workspaces_list btrfs_heuristic_ws; | |
261507a0 | 802 | |
e8c9f186 | 803 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
261507a0 | 804 | &btrfs_zlib_compress, |
a6fa6fae | 805 | &btrfs_lzo_compress, |
5c1aab1d | 806 | &btrfs_zstd_compress, |
261507a0 LZ |
807 | }; |
808 | ||
143bede5 | 809 | void __init btrfs_init_compress(void) |
261507a0 | 810 | { |
4e439a0b | 811 | struct list_head *workspace; |
261507a0 LZ |
812 | int i; |
813 | ||
4e439a0b TT |
814 | INIT_LIST_HEAD(&btrfs_heuristic_ws.idle_ws); |
815 | spin_lock_init(&btrfs_heuristic_ws.ws_lock); | |
816 | atomic_set(&btrfs_heuristic_ws.total_ws, 0); | |
817 | init_waitqueue_head(&btrfs_heuristic_ws.ws_wait); | |
f77dd0d6 | 818 | |
4e439a0b TT |
819 | workspace = alloc_heuristic_ws(); |
820 | if (IS_ERR(workspace)) { | |
821 | pr_warn( | |
822 | "BTRFS: cannot preallocate heuristic workspace, will try later\n"); | |
823 | } else { | |
824 | atomic_set(&btrfs_heuristic_ws.total_ws, 1); | |
825 | btrfs_heuristic_ws.free_ws = 1; | |
826 | list_add(workspace, &btrfs_heuristic_ws.idle_ws); | |
827 | } | |
828 | ||
829 | for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) { | |
d9187649 BL |
830 | INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws); |
831 | spin_lock_init(&btrfs_comp_ws[i].ws_lock); | |
6ac10a6a | 832 | atomic_set(&btrfs_comp_ws[i].total_ws, 0); |
d9187649 | 833 | init_waitqueue_head(&btrfs_comp_ws[i].ws_wait); |
f77dd0d6 DS |
834 | |
835 | /* | |
836 | * Preallocate one workspace for each compression type so | |
837 | * we can guarantee forward progress in the worst case | |
838 | */ | |
839 | workspace = btrfs_compress_op[i]->alloc_workspace(); | |
840 | if (IS_ERR(workspace)) { | |
62e85577 | 841 | pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n"); |
f77dd0d6 DS |
842 | } else { |
843 | atomic_set(&btrfs_comp_ws[i].total_ws, 1); | |
844 | btrfs_comp_ws[i].free_ws = 1; | |
845 | list_add(workspace, &btrfs_comp_ws[i].idle_ws); | |
846 | } | |
261507a0 | 847 | } |
261507a0 LZ |
848 | } |
849 | ||
850 | /* | |
e721e49d DS |
851 | * This finds an available workspace or allocates a new one. |
852 | * If it's not possible to allocate a new one, waits until there's one. | |
853 | * Preallocation makes a forward progress guarantees and we do not return | |
854 | * errors. | |
261507a0 | 855 | */ |
4e439a0b | 856 | static struct list_head *__find_workspace(int type, bool heuristic) |
261507a0 LZ |
857 | { |
858 | struct list_head *workspace; | |
859 | int cpus = num_online_cpus(); | |
860 | int idx = type - 1; | |
fe308533 | 861 | unsigned nofs_flag; |
4e439a0b TT |
862 | struct list_head *idle_ws; |
863 | spinlock_t *ws_lock; | |
864 | atomic_t *total_ws; | |
865 | wait_queue_head_t *ws_wait; | |
866 | int *free_ws; | |
867 | ||
868 | if (heuristic) { | |
869 | idle_ws = &btrfs_heuristic_ws.idle_ws; | |
870 | ws_lock = &btrfs_heuristic_ws.ws_lock; | |
871 | total_ws = &btrfs_heuristic_ws.total_ws; | |
872 | ws_wait = &btrfs_heuristic_ws.ws_wait; | |
873 | free_ws = &btrfs_heuristic_ws.free_ws; | |
874 | } else { | |
875 | idle_ws = &btrfs_comp_ws[idx].idle_ws; | |
876 | ws_lock = &btrfs_comp_ws[idx].ws_lock; | |
877 | total_ws = &btrfs_comp_ws[idx].total_ws; | |
878 | ws_wait = &btrfs_comp_ws[idx].ws_wait; | |
879 | free_ws = &btrfs_comp_ws[idx].free_ws; | |
880 | } | |
261507a0 | 881 | |
261507a0 | 882 | again: |
d9187649 BL |
883 | spin_lock(ws_lock); |
884 | if (!list_empty(idle_ws)) { | |
885 | workspace = idle_ws->next; | |
261507a0 | 886 | list_del(workspace); |
6ac10a6a | 887 | (*free_ws)--; |
d9187649 | 888 | spin_unlock(ws_lock); |
261507a0 LZ |
889 | return workspace; |
890 | ||
891 | } | |
6ac10a6a | 892 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
893 | DEFINE_WAIT(wait); |
894 | ||
d9187649 BL |
895 | spin_unlock(ws_lock); |
896 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 897 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 898 | schedule(); |
d9187649 | 899 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
900 | goto again; |
901 | } | |
6ac10a6a | 902 | atomic_inc(total_ws); |
d9187649 | 903 | spin_unlock(ws_lock); |
261507a0 | 904 | |
fe308533 DS |
905 | /* |
906 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
907 | * to turn it off here because we might get called from the restricted | |
908 | * context of btrfs_compress_bio/btrfs_compress_pages | |
909 | */ | |
910 | nofs_flag = memalloc_nofs_save(); | |
4e439a0b TT |
911 | if (heuristic) |
912 | workspace = alloc_heuristic_ws(); | |
913 | else | |
914 | workspace = btrfs_compress_op[idx]->alloc_workspace(); | |
fe308533 DS |
915 | memalloc_nofs_restore(nofs_flag); |
916 | ||
261507a0 | 917 | if (IS_ERR(workspace)) { |
6ac10a6a | 918 | atomic_dec(total_ws); |
d9187649 | 919 | wake_up(ws_wait); |
e721e49d DS |
920 | |
921 | /* | |
922 | * Do not return the error but go back to waiting. There's a | |
923 | * workspace preallocated for each type and the compression | |
924 | * time is bounded so we get to a workspace eventually. This | |
925 | * makes our caller's life easier. | |
52356716 DS |
926 | * |
927 | * To prevent silent and low-probability deadlocks (when the | |
928 | * initial preallocation fails), check if there are any | |
929 | * workspaces at all. | |
e721e49d | 930 | */ |
52356716 DS |
931 | if (atomic_read(total_ws) == 0) { |
932 | static DEFINE_RATELIMIT_STATE(_rs, | |
933 | /* once per minute */ 60 * HZ, | |
934 | /* no burst */ 1); | |
935 | ||
936 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 937 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
938 | } |
939 | } | |
e721e49d | 940 | goto again; |
261507a0 LZ |
941 | } |
942 | return workspace; | |
943 | } | |
944 | ||
4e439a0b TT |
945 | static struct list_head *find_workspace(int type) |
946 | { | |
947 | return __find_workspace(type, false); | |
948 | } | |
949 | ||
261507a0 LZ |
950 | /* |
951 | * put a workspace struct back on the list or free it if we have enough | |
952 | * idle ones sitting around | |
953 | */ | |
4e439a0b TT |
954 | static void __free_workspace(int type, struct list_head *workspace, |
955 | bool heuristic) | |
261507a0 LZ |
956 | { |
957 | int idx = type - 1; | |
4e439a0b TT |
958 | struct list_head *idle_ws; |
959 | spinlock_t *ws_lock; | |
960 | atomic_t *total_ws; | |
961 | wait_queue_head_t *ws_wait; | |
962 | int *free_ws; | |
963 | ||
964 | if (heuristic) { | |
965 | idle_ws = &btrfs_heuristic_ws.idle_ws; | |
966 | ws_lock = &btrfs_heuristic_ws.ws_lock; | |
967 | total_ws = &btrfs_heuristic_ws.total_ws; | |
968 | ws_wait = &btrfs_heuristic_ws.ws_wait; | |
969 | free_ws = &btrfs_heuristic_ws.free_ws; | |
970 | } else { | |
971 | idle_ws = &btrfs_comp_ws[idx].idle_ws; | |
972 | ws_lock = &btrfs_comp_ws[idx].ws_lock; | |
973 | total_ws = &btrfs_comp_ws[idx].total_ws; | |
974 | ws_wait = &btrfs_comp_ws[idx].ws_wait; | |
975 | free_ws = &btrfs_comp_ws[idx].free_ws; | |
976 | } | |
d9187649 BL |
977 | |
978 | spin_lock(ws_lock); | |
26b28dce | 979 | if (*free_ws <= num_online_cpus()) { |
d9187649 | 980 | list_add(workspace, idle_ws); |
6ac10a6a | 981 | (*free_ws)++; |
d9187649 | 982 | spin_unlock(ws_lock); |
261507a0 LZ |
983 | goto wake; |
984 | } | |
d9187649 | 985 | spin_unlock(ws_lock); |
261507a0 | 986 | |
4e439a0b TT |
987 | if (heuristic) |
988 | free_heuristic_ws(workspace); | |
989 | else | |
990 | btrfs_compress_op[idx]->free_workspace(workspace); | |
6ac10a6a | 991 | atomic_dec(total_ws); |
261507a0 | 992 | wake: |
a83342aa DS |
993 | /* |
994 | * Make sure counter is updated before we wake up waiters. | |
995 | */ | |
66657b31 | 996 | smp_mb(); |
d9187649 BL |
997 | if (waitqueue_active(ws_wait)) |
998 | wake_up(ws_wait); | |
261507a0 LZ |
999 | } |
1000 | ||
4e439a0b TT |
1001 | static void free_workspace(int type, struct list_head *ws) |
1002 | { | |
1003 | return __free_workspace(type, ws, false); | |
1004 | } | |
1005 | ||
261507a0 LZ |
1006 | /* |
1007 | * cleanup function for module exit | |
1008 | */ | |
1009 | static void free_workspaces(void) | |
1010 | { | |
1011 | struct list_head *workspace; | |
1012 | int i; | |
1013 | ||
4e439a0b TT |
1014 | while (!list_empty(&btrfs_heuristic_ws.idle_ws)) { |
1015 | workspace = btrfs_heuristic_ws.idle_ws.next; | |
1016 | list_del(workspace); | |
1017 | free_heuristic_ws(workspace); | |
1018 | atomic_dec(&btrfs_heuristic_ws.total_ws); | |
1019 | } | |
1020 | ||
261507a0 | 1021 | for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) { |
d9187649 BL |
1022 | while (!list_empty(&btrfs_comp_ws[i].idle_ws)) { |
1023 | workspace = btrfs_comp_ws[i].idle_ws.next; | |
261507a0 LZ |
1024 | list_del(workspace); |
1025 | btrfs_compress_op[i]->free_workspace(workspace); | |
6ac10a6a | 1026 | atomic_dec(&btrfs_comp_ws[i].total_ws); |
261507a0 LZ |
1027 | } |
1028 | } | |
1029 | } | |
1030 | ||
1031 | /* | |
38c31464 DS |
1032 | * Given an address space and start and length, compress the bytes into @pages |
1033 | * that are allocated on demand. | |
261507a0 | 1034 | * |
f51d2b59 DS |
1035 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1036 | * default the algorithm chooses and is opaque here; | |
1037 | * - compression algo are 0-3 | |
1038 | * - the level are bits 4-7 | |
1039 | * | |
4d3a800e DS |
1040 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1041 | * and returns number of actually allocated pages | |
261507a0 | 1042 | * |
38c31464 DS |
1043 | * @total_in is used to return the number of bytes actually read. It |
1044 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1045 | * ran out of room in the pages array or because we cross the |
1046 | * max_out threshold. | |
1047 | * | |
38c31464 DS |
1048 | * @total_out is an in/out parameter, must be set to the input length and will |
1049 | * be also used to return the total number of compressed bytes | |
261507a0 | 1050 | * |
38c31464 | 1051 | * @max_out tells us the max number of bytes that we're allowed to |
261507a0 LZ |
1052 | * stuff into pages |
1053 | */ | |
f51d2b59 | 1054 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1055 | u64 start, struct page **pages, |
261507a0 LZ |
1056 | unsigned long *out_pages, |
1057 | unsigned long *total_in, | |
e5d74902 | 1058 | unsigned long *total_out) |
261507a0 LZ |
1059 | { |
1060 | struct list_head *workspace; | |
1061 | int ret; | |
f51d2b59 | 1062 | int type = type_level & 0xF; |
261507a0 LZ |
1063 | |
1064 | workspace = find_workspace(type); | |
261507a0 | 1065 | |
f51d2b59 | 1066 | btrfs_compress_op[type - 1]->set_level(workspace, type_level); |
261507a0 | 1067 | ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping, |
38c31464 | 1068 | start, pages, |
4d3a800e | 1069 | out_pages, |
e5d74902 | 1070 | total_in, total_out); |
261507a0 LZ |
1071 | free_workspace(type, workspace); |
1072 | return ret; | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * pages_in is an array of pages with compressed data. | |
1077 | * | |
1078 | * disk_start is the starting logical offset of this array in the file | |
1079 | * | |
974b1adc | 1080 | * orig_bio contains the pages from the file that we want to decompress into |
261507a0 LZ |
1081 | * |
1082 | * srclen is the number of bytes in pages_in | |
1083 | * | |
1084 | * The basic idea is that we have a bio that was created by readpages. | |
1085 | * The pages in the bio are for the uncompressed data, and they may not | |
1086 | * be contiguous. They all correspond to the range of bytes covered by | |
1087 | * the compressed extent. | |
1088 | */ | |
8140dc30 | 1089 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1090 | { |
1091 | struct list_head *workspace; | |
1092 | int ret; | |
8140dc30 | 1093 | int type = cb->compress_type; |
261507a0 LZ |
1094 | |
1095 | workspace = find_workspace(type); | |
e1ddce71 | 1096 | ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb); |
261507a0 | 1097 | free_workspace(type, workspace); |
e1ddce71 | 1098 | |
261507a0 LZ |
1099 | return ret; |
1100 | } | |
1101 | ||
1102 | /* | |
1103 | * a less complex decompression routine. Our compressed data fits in a | |
1104 | * single page, and we want to read a single page out of it. | |
1105 | * start_byte tells us the offset into the compressed data we're interested in | |
1106 | */ | |
1107 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1108 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1109 | { | |
1110 | struct list_head *workspace; | |
1111 | int ret; | |
1112 | ||
1113 | workspace = find_workspace(type); | |
261507a0 LZ |
1114 | |
1115 | ret = btrfs_compress_op[type-1]->decompress(workspace, data_in, | |
1116 | dest_page, start_byte, | |
1117 | srclen, destlen); | |
1118 | ||
1119 | free_workspace(type, workspace); | |
1120 | return ret; | |
1121 | } | |
1122 | ||
e67c718b | 1123 | void __cold btrfs_exit_compress(void) |
261507a0 LZ |
1124 | { |
1125 | free_workspaces(); | |
1126 | } | |
3a39c18d LZ |
1127 | |
1128 | /* | |
1129 | * Copy uncompressed data from working buffer to pages. | |
1130 | * | |
1131 | * buf_start is the byte offset we're of the start of our workspace buffer. | |
1132 | * | |
1133 | * total_out is the last byte of the buffer | |
1134 | */ | |
14a3357b | 1135 | int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, |
3a39c18d | 1136 | unsigned long total_out, u64 disk_start, |
974b1adc | 1137 | struct bio *bio) |
3a39c18d LZ |
1138 | { |
1139 | unsigned long buf_offset; | |
1140 | unsigned long current_buf_start; | |
1141 | unsigned long start_byte; | |
6e78b3f7 | 1142 | unsigned long prev_start_byte; |
3a39c18d LZ |
1143 | unsigned long working_bytes = total_out - buf_start; |
1144 | unsigned long bytes; | |
1145 | char *kaddr; | |
974b1adc | 1146 | struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter); |
3a39c18d LZ |
1147 | |
1148 | /* | |
1149 | * start byte is the first byte of the page we're currently | |
1150 | * copying into relative to the start of the compressed data. | |
1151 | */ | |
974b1adc | 1152 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d LZ |
1153 | |
1154 | /* we haven't yet hit data corresponding to this page */ | |
1155 | if (total_out <= start_byte) | |
1156 | return 1; | |
1157 | ||
1158 | /* | |
1159 | * the start of the data we care about is offset into | |
1160 | * the middle of our working buffer | |
1161 | */ | |
1162 | if (total_out > start_byte && buf_start < start_byte) { | |
1163 | buf_offset = start_byte - buf_start; | |
1164 | working_bytes -= buf_offset; | |
1165 | } else { | |
1166 | buf_offset = 0; | |
1167 | } | |
1168 | current_buf_start = buf_start; | |
1169 | ||
1170 | /* copy bytes from the working buffer into the pages */ | |
1171 | while (working_bytes > 0) { | |
974b1adc CH |
1172 | bytes = min_t(unsigned long, bvec.bv_len, |
1173 | PAGE_SIZE - buf_offset); | |
3a39c18d | 1174 | bytes = min(bytes, working_bytes); |
974b1adc CH |
1175 | |
1176 | kaddr = kmap_atomic(bvec.bv_page); | |
1177 | memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes); | |
7ac687d9 | 1178 | kunmap_atomic(kaddr); |
974b1adc | 1179 | flush_dcache_page(bvec.bv_page); |
3a39c18d | 1180 | |
3a39c18d LZ |
1181 | buf_offset += bytes; |
1182 | working_bytes -= bytes; | |
1183 | current_buf_start += bytes; | |
1184 | ||
1185 | /* check if we need to pick another page */ | |
974b1adc CH |
1186 | bio_advance(bio, bytes); |
1187 | if (!bio->bi_iter.bi_size) | |
1188 | return 0; | |
1189 | bvec = bio_iter_iovec(bio, bio->bi_iter); | |
6e78b3f7 | 1190 | prev_start_byte = start_byte; |
974b1adc | 1191 | start_byte = page_offset(bvec.bv_page) - disk_start; |
3a39c18d | 1192 | |
974b1adc | 1193 | /* |
6e78b3f7 OS |
1194 | * We need to make sure we're only adjusting |
1195 | * our offset into compression working buffer when | |
1196 | * we're switching pages. Otherwise we can incorrectly | |
1197 | * keep copying when we were actually done. | |
974b1adc | 1198 | */ |
6e78b3f7 OS |
1199 | if (start_byte != prev_start_byte) { |
1200 | /* | |
1201 | * make sure our new page is covered by this | |
1202 | * working buffer | |
1203 | */ | |
1204 | if (total_out <= start_byte) | |
1205 | return 1; | |
3a39c18d | 1206 | |
6e78b3f7 OS |
1207 | /* |
1208 | * the next page in the biovec might not be adjacent | |
1209 | * to the last page, but it might still be found | |
1210 | * inside this working buffer. bump our offset pointer | |
1211 | */ | |
1212 | if (total_out > start_byte && | |
1213 | current_buf_start < start_byte) { | |
1214 | buf_offset = start_byte - buf_start; | |
1215 | working_bytes = total_out - start_byte; | |
1216 | current_buf_start = buf_start + buf_offset; | |
1217 | } | |
3a39c18d LZ |
1218 | } |
1219 | } | |
1220 | ||
1221 | return 1; | |
1222 | } | |
c2fcdcdf | 1223 | |
19562430 TT |
1224 | /* |
1225 | * Shannon Entropy calculation | |
1226 | * | |
1227 | * Pure byte distribution analysis fails to determine compressiability of data. | |
1228 | * Try calculating entropy to estimate the average minimum number of bits | |
1229 | * needed to encode the sampled data. | |
1230 | * | |
1231 | * For convenience, return the percentage of needed bits, instead of amount of | |
1232 | * bits directly. | |
1233 | * | |
1234 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1235 | * and can be compressible with high probability | |
1236 | * | |
1237 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1238 | * | |
1239 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1240 | */ | |
1241 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1242 | #define ENTROPY_LVL_HIGH (80) | |
1243 | ||
1244 | /* | |
1245 | * For increasead precision in shannon_entropy calculation, | |
1246 | * let's do pow(n, M) to save more digits after comma: | |
1247 | * | |
1248 | * - maximum int bit length is 64 | |
1249 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1250 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1251 | * | |
1252 | * So use pow(n, 4). | |
1253 | */ | |
1254 | static inline u32 ilog2_w(u64 n) | |
1255 | { | |
1256 | return ilog2(n * n * n * n); | |
1257 | } | |
1258 | ||
1259 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1260 | { | |
1261 | const u32 entropy_max = 8 * ilog2_w(2); | |
1262 | u32 entropy_sum = 0; | |
1263 | u32 p, p_base, sz_base; | |
1264 | u32 i; | |
1265 | ||
1266 | sz_base = ilog2_w(ws->sample_size); | |
1267 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1268 | p = ws->bucket[i].count; | |
1269 | p_base = ilog2_w(p); | |
1270 | entropy_sum += p * (sz_base - p_base); | |
1271 | } | |
1272 | ||
1273 | entropy_sum /= ws->sample_size; | |
1274 | return entropy_sum * 100 / entropy_max; | |
1275 | } | |
1276 | ||
440c840c TT |
1277 | #define RADIX_BASE 4U |
1278 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1279 | ||
1280 | static u8 get4bits(u64 num, int shift) { | |
1281 | u8 low4bits; | |
1282 | ||
1283 | num >>= shift; | |
1284 | /* Reverse order */ | |
1285 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1286 | return low4bits; | |
1287 | } | |
1288 | ||
440c840c TT |
1289 | /* |
1290 | * Use 4 bits as radix base | |
1291 | * Use 16 u32 counters for calculating new possition in buf array | |
1292 | * | |
1293 | * @array - array that will be sorted | |
1294 | * @array_buf - buffer array to store sorting results | |
1295 | * must be equal in size to @array | |
1296 | * @num - array size | |
440c840c | 1297 | */ |
23ae8c63 | 1298 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1299 | int num) |
858177d3 | 1300 | { |
440c840c TT |
1301 | u64 max_num; |
1302 | u64 buf_num; | |
1303 | u32 counters[COUNTERS_SIZE]; | |
1304 | u32 new_addr; | |
1305 | u32 addr; | |
1306 | int bitlen; | |
1307 | int shift; | |
1308 | int i; | |
858177d3 | 1309 | |
440c840c TT |
1310 | /* |
1311 | * Try avoid useless loop iterations for small numbers stored in big | |
1312 | * counters. Example: 48 33 4 ... in 64bit array | |
1313 | */ | |
23ae8c63 | 1314 | max_num = array[0].count; |
440c840c | 1315 | for (i = 1; i < num; i++) { |
23ae8c63 | 1316 | buf_num = array[i].count; |
440c840c TT |
1317 | if (buf_num > max_num) |
1318 | max_num = buf_num; | |
1319 | } | |
1320 | ||
1321 | buf_num = ilog2(max_num); | |
1322 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1323 | ||
1324 | shift = 0; | |
1325 | while (shift < bitlen) { | |
1326 | memset(counters, 0, sizeof(counters)); | |
1327 | ||
1328 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1329 | buf_num = array[i].count; |
440c840c TT |
1330 | addr = get4bits(buf_num, shift); |
1331 | counters[addr]++; | |
1332 | } | |
1333 | ||
1334 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1335 | counters[i] += counters[i - 1]; | |
1336 | ||
1337 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1338 | buf_num = array[i].count; |
440c840c TT |
1339 | addr = get4bits(buf_num, shift); |
1340 | counters[addr]--; | |
1341 | new_addr = counters[addr]; | |
7add17be | 1342 | array_buf[new_addr] = array[i]; |
440c840c TT |
1343 | } |
1344 | ||
1345 | shift += RADIX_BASE; | |
1346 | ||
1347 | /* | |
1348 | * Normal radix expects to move data from a temporary array, to | |
1349 | * the main one. But that requires some CPU time. Avoid that | |
1350 | * by doing another sort iteration to original array instead of | |
1351 | * memcpy() | |
1352 | */ | |
1353 | memset(counters, 0, sizeof(counters)); | |
1354 | ||
1355 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1356 | buf_num = array_buf[i].count; |
440c840c TT |
1357 | addr = get4bits(buf_num, shift); |
1358 | counters[addr]++; | |
1359 | } | |
1360 | ||
1361 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1362 | counters[i] += counters[i - 1]; | |
1363 | ||
1364 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1365 | buf_num = array_buf[i].count; |
440c840c TT |
1366 | addr = get4bits(buf_num, shift); |
1367 | counters[addr]--; | |
1368 | new_addr = counters[addr]; | |
7add17be | 1369 | array[new_addr] = array_buf[i]; |
440c840c TT |
1370 | } |
1371 | ||
1372 | shift += RADIX_BASE; | |
1373 | } | |
858177d3 TT |
1374 | } |
1375 | ||
1376 | /* | |
1377 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1378 | * | |
1379 | * There are several types of structured binary data that use nearly all byte | |
1380 | * values. The distribution can be uniform and counts in all buckets will be | |
1381 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1382 | * | |
1383 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1384 | * be potentially compressible, but we have to take a few more steps to decide | |
1385 | * how much. | |
1386 | * | |
1387 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1388 | * compression algo can easy fix that | |
1389 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1390 | * probability is not compressible | |
1391 | */ | |
1392 | #define BYTE_CORE_SET_LOW (64) | |
1393 | #define BYTE_CORE_SET_HIGH (200) | |
1394 | ||
1395 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1396 | { | |
1397 | u32 i; | |
1398 | u32 coreset_sum = 0; | |
1399 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1400 | struct bucket_item *bucket = ws->bucket; | |
1401 | ||
1402 | /* Sort in reverse order */ | |
36243c91 | 1403 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1404 | |
1405 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1406 | coreset_sum += bucket[i].count; | |
1407 | ||
1408 | if (coreset_sum > core_set_threshold) | |
1409 | return i; | |
1410 | ||
1411 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1412 | coreset_sum += bucket[i].count; | |
1413 | if (coreset_sum > core_set_threshold) | |
1414 | break; | |
1415 | } | |
1416 | ||
1417 | return i; | |
1418 | } | |
1419 | ||
a288e92c TT |
1420 | /* |
1421 | * Count byte values in buckets. | |
1422 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1423 | * Because in most text-like data byte set is restricted to limited number of | |
1424 | * possible characters, and that restriction in most cases makes data easy to | |
1425 | * compress. | |
1426 | * | |
1427 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1428 | * less - compressible | |
1429 | * more - need additional analysis | |
1430 | */ | |
1431 | #define BYTE_SET_THRESHOLD (64) | |
1432 | ||
1433 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1434 | { | |
1435 | u32 i; | |
1436 | u32 byte_set_size = 0; | |
1437 | ||
1438 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1439 | if (ws->bucket[i].count > 0) | |
1440 | byte_set_size++; | |
1441 | } | |
1442 | ||
1443 | /* | |
1444 | * Continue collecting count of byte values in buckets. If the byte | |
1445 | * set size is bigger then the threshold, it's pointless to continue, | |
1446 | * the detection technique would fail for this type of data. | |
1447 | */ | |
1448 | for (; i < BUCKET_SIZE; i++) { | |
1449 | if (ws->bucket[i].count > 0) { | |
1450 | byte_set_size++; | |
1451 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1452 | return byte_set_size; | |
1453 | } | |
1454 | } | |
1455 | ||
1456 | return byte_set_size; | |
1457 | } | |
1458 | ||
1fe4f6fa TT |
1459 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1460 | { | |
1461 | const u32 half_of_sample = ws->sample_size / 2; | |
1462 | const u8 *data = ws->sample; | |
1463 | ||
1464 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1465 | } | |
1466 | ||
a440d48c TT |
1467 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1468 | struct heuristic_ws *ws) | |
1469 | { | |
1470 | struct page *page; | |
1471 | u64 index, index_end; | |
1472 | u32 i, curr_sample_pos; | |
1473 | u8 *in_data; | |
1474 | ||
1475 | /* | |
1476 | * Compression handles the input data by chunks of 128KiB | |
1477 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1478 | * | |
1479 | * We do the same for the heuristic and loop over the whole range. | |
1480 | * | |
1481 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1482 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1483 | */ | |
1484 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1485 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1486 | ||
1487 | index = start >> PAGE_SHIFT; | |
1488 | index_end = end >> PAGE_SHIFT; | |
1489 | ||
1490 | /* Don't miss unaligned end */ | |
1491 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1492 | index_end++; | |
1493 | ||
1494 | curr_sample_pos = 0; | |
1495 | while (index < index_end) { | |
1496 | page = find_get_page(inode->i_mapping, index); | |
1497 | in_data = kmap(page); | |
1498 | /* Handle case where the start is not aligned to PAGE_SIZE */ | |
1499 | i = start % PAGE_SIZE; | |
1500 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1501 | /* Don't sample any garbage from the last page */ | |
1502 | if (start > end - SAMPLING_READ_SIZE) | |
1503 | break; | |
1504 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1505 | SAMPLING_READ_SIZE); | |
1506 | i += SAMPLING_INTERVAL; | |
1507 | start += SAMPLING_INTERVAL; | |
1508 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1509 | } | |
1510 | kunmap(page); | |
1511 | put_page(page); | |
1512 | ||
1513 | index++; | |
1514 | } | |
1515 | ||
1516 | ws->sample_size = curr_sample_pos; | |
1517 | } | |
1518 | ||
c2fcdcdf TT |
1519 | /* |
1520 | * Compression heuristic. | |
1521 | * | |
1522 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1523 | * quickly (compared to direct compression) detect data characteristics | |
1524 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1525 | * data. | |
1526 | * | |
1527 | * The following types of analysis can be performed: | |
1528 | * - detect mostly zero data | |
1529 | * - detect data with low "byte set" size (text, etc) | |
1530 | * - detect data with low/high "core byte" set | |
1531 | * | |
1532 | * Return non-zero if the compression should be done, 0 otherwise. | |
1533 | */ | |
1534 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1535 | { | |
4e439a0b TT |
1536 | struct list_head *ws_list = __find_workspace(0, true); |
1537 | struct heuristic_ws *ws; | |
a440d48c TT |
1538 | u32 i; |
1539 | u8 byte; | |
19562430 | 1540 | int ret = 0; |
c2fcdcdf | 1541 | |
4e439a0b TT |
1542 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1543 | ||
a440d48c TT |
1544 | heuristic_collect_sample(inode, start, end, ws); |
1545 | ||
1fe4f6fa TT |
1546 | if (sample_repeated_patterns(ws)) { |
1547 | ret = 1; | |
1548 | goto out; | |
1549 | } | |
1550 | ||
a440d48c TT |
1551 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1552 | ||
1553 | for (i = 0; i < ws->sample_size; i++) { | |
1554 | byte = ws->sample[i]; | |
1555 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1556 | } |
1557 | ||
a288e92c TT |
1558 | i = byte_set_size(ws); |
1559 | if (i < BYTE_SET_THRESHOLD) { | |
1560 | ret = 2; | |
1561 | goto out; | |
1562 | } | |
1563 | ||
858177d3 TT |
1564 | i = byte_core_set_size(ws); |
1565 | if (i <= BYTE_CORE_SET_LOW) { | |
1566 | ret = 3; | |
1567 | goto out; | |
1568 | } | |
1569 | ||
1570 | if (i >= BYTE_CORE_SET_HIGH) { | |
1571 | ret = 0; | |
1572 | goto out; | |
1573 | } | |
1574 | ||
19562430 TT |
1575 | i = shannon_entropy(ws); |
1576 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1577 | ret = 4; | |
1578 | goto out; | |
1579 | } | |
1580 | ||
1581 | /* | |
1582 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1583 | * needed to give green light to compression. | |
1584 | * | |
1585 | * For now just assume that compression at that level is not worth the | |
1586 | * resources because: | |
1587 | * | |
1588 | * 1. it is possible to defrag the data later | |
1589 | * | |
1590 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1591 | * values, every bucket has counter at level ~54. The heuristic would | |
1592 | * be confused. This can happen when data have some internal repeated | |
1593 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1594 | * pairs of bytes, which is too costly. | |
1595 | */ | |
1596 | if (i < ENTROPY_LVL_HIGH) { | |
1597 | ret = 5; | |
1598 | goto out; | |
1599 | } else { | |
1600 | ret = 0; | |
1601 | goto out; | |
1602 | } | |
1603 | ||
1fe4f6fa | 1604 | out: |
4e439a0b | 1605 | __free_workspace(0, ws_list, true); |
c2fcdcdf TT |
1606 | return ret; |
1607 | } | |
f51d2b59 DS |
1608 | |
1609 | unsigned int btrfs_compress_str2level(const char *str) | |
1610 | { | |
1611 | if (strncmp(str, "zlib", 4) != 0) | |
1612 | return 0; | |
1613 | ||
fa4d885a AB |
1614 | /* Accepted form: zlib:1 up to zlib:9 and nothing left after the number */ |
1615 | if (str[4] == ':' && '1' <= str[5] && str[5] <= '9' && str[6] == 0) | |
1616 | return str[5] - '0'; | |
f51d2b59 | 1617 | |
eae8d825 | 1618 | return BTRFS_ZLIB_DEFAULT_LEVEL; |
f51d2b59 | 1619 | } |