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> | |
a75b81c3 | 11 | #include <linux/pagevec.h> |
c8b97818 | 12 | #include <linux/highmem.h> |
e41d12f5 | 13 | #include <linux/kthread.h> |
c8b97818 CM |
14 | #include <linux/time.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/string.h> | |
c8b97818 | 17 | #include <linux/backing-dev.h> |
c8b97818 | 18 | #include <linux/writeback.h> |
4088a47e | 19 | #include <linux/psi.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
fe308533 | 21 | #include <linux/sched/mm.h> |
19562430 | 22 | #include <linux/log2.h> |
d5178578 | 23 | #include <crypto/hash.h> |
602cbe91 | 24 | #include "misc.h" |
c8b97818 | 25 | #include "ctree.h" |
ec8eb376 | 26 | #include "fs.h" |
c8b97818 CM |
27 | #include "disk-io.h" |
28 | #include "transaction.h" | |
29 | #include "btrfs_inode.h" | |
30 | #include "volumes.h" | |
31 | #include "ordered-data.h" | |
c8b97818 CM |
32 | #include "compression.h" |
33 | #include "extent_io.h" | |
34 | #include "extent_map.h" | |
6a404910 | 35 | #include "subpage.h" |
764c7c9a | 36 | #include "zoned.h" |
c8b97818 | 37 | |
e128f9c3 DS |
38 | static const char* const btrfs_compress_types[] = { "", "zlib", "lzo", "zstd" }; |
39 | ||
40 | const char* btrfs_compress_type2str(enum btrfs_compression_type type) | |
41 | { | |
42 | switch (type) { | |
43 | case BTRFS_COMPRESS_ZLIB: | |
44 | case BTRFS_COMPRESS_LZO: | |
45 | case BTRFS_COMPRESS_ZSTD: | |
46 | case BTRFS_COMPRESS_NONE: | |
47 | return btrfs_compress_types[type]; | |
ce96b7ff CX |
48 | default: |
49 | break; | |
e128f9c3 DS |
50 | } |
51 | ||
52 | return NULL; | |
53 | } | |
54 | ||
aa53e3bf JT |
55 | bool btrfs_compress_is_valid_type(const char *str, size_t len) |
56 | { | |
57 | int i; | |
58 | ||
59 | for (i = 1; i < ARRAY_SIZE(btrfs_compress_types); i++) { | |
60 | size_t comp_len = strlen(btrfs_compress_types[i]); | |
61 | ||
62 | if (len < comp_len) | |
63 | continue; | |
64 | ||
65 | if (!strncmp(btrfs_compress_types[i], str, comp_len)) | |
66 | return true; | |
67 | } | |
68 | return false; | |
69 | } | |
70 | ||
1e4eb746 DS |
71 | static int compression_compress_pages(int type, struct list_head *ws, |
72 | struct address_space *mapping, u64 start, struct page **pages, | |
73 | unsigned long *out_pages, unsigned long *total_in, | |
74 | unsigned long *total_out) | |
75 | { | |
76 | switch (type) { | |
77 | case BTRFS_COMPRESS_ZLIB: | |
78 | return zlib_compress_pages(ws, mapping, start, pages, | |
79 | out_pages, total_in, total_out); | |
80 | case BTRFS_COMPRESS_LZO: | |
81 | return lzo_compress_pages(ws, mapping, start, pages, | |
82 | out_pages, total_in, total_out); | |
83 | case BTRFS_COMPRESS_ZSTD: | |
84 | return zstd_compress_pages(ws, mapping, start, pages, | |
85 | out_pages, total_in, total_out); | |
86 | case BTRFS_COMPRESS_NONE: | |
87 | default: | |
88 | /* | |
1d8ba9e7 QW |
89 | * This can happen when compression races with remount setting |
90 | * it to 'no compress', while caller doesn't call | |
91 | * inode_need_compress() to check if we really need to | |
92 | * compress. | |
93 | * | |
94 | * Not a big deal, just need to inform caller that we | |
95 | * haven't allocated any pages yet. | |
1e4eb746 | 96 | */ |
1d8ba9e7 | 97 | *out_pages = 0; |
1e4eb746 DS |
98 | return -E2BIG; |
99 | } | |
100 | } | |
101 | ||
4a9e803e SY |
102 | static int compression_decompress_bio(struct list_head *ws, |
103 | struct compressed_bio *cb) | |
1e4eb746 | 104 | { |
4a9e803e | 105 | switch (cb->compress_type) { |
1e4eb746 DS |
106 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress_bio(ws, cb); |
107 | case BTRFS_COMPRESS_LZO: return lzo_decompress_bio(ws, cb); | |
108 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress_bio(ws, cb); | |
109 | case BTRFS_COMPRESS_NONE: | |
110 | default: | |
111 | /* | |
112 | * This can't happen, the type is validated several times | |
113 | * before we get here. | |
114 | */ | |
115 | BUG(); | |
116 | } | |
117 | } | |
118 | ||
119 | static int compression_decompress(int type, struct list_head *ws, | |
120 | unsigned char *data_in, struct page *dest_page, | |
121 | unsigned long start_byte, size_t srclen, size_t destlen) | |
122 | { | |
123 | switch (type) { | |
124 | case BTRFS_COMPRESS_ZLIB: return zlib_decompress(ws, data_in, dest_page, | |
125 | start_byte, srclen, destlen); | |
126 | case BTRFS_COMPRESS_LZO: return lzo_decompress(ws, data_in, dest_page, | |
127 | start_byte, srclen, destlen); | |
128 | case BTRFS_COMPRESS_ZSTD: return zstd_decompress(ws, data_in, dest_page, | |
129 | start_byte, srclen, destlen); | |
130 | case BTRFS_COMPRESS_NONE: | |
131 | default: | |
132 | /* | |
133 | * This can't happen, the type is validated several times | |
134 | * before we get here. | |
135 | */ | |
136 | BUG(); | |
137 | } | |
138 | } | |
139 | ||
8140dc30 | 140 | static int btrfs_decompress_bio(struct compressed_bio *cb); |
48a3b636 | 141 | |
e14bfdb5 | 142 | static void finish_compressed_bio_read(struct compressed_bio *cb) |
86ccbb4d QW |
143 | { |
144 | unsigned int index; | |
145 | struct page *page; | |
146 | ||
81bd9328 CH |
147 | if (cb->status == BLK_STS_OK) |
148 | cb->status = errno_to_blk_status(btrfs_decompress_bio(cb)); | |
149 | ||
86ccbb4d QW |
150 | /* Release the compressed pages */ |
151 | for (index = 0; index < cb->nr_pages; index++) { | |
152 | page = cb->compressed_pages[index]; | |
153 | page->mapping = NULL; | |
154 | put_page(page); | |
155 | } | |
156 | ||
157 | /* Do io completion on the original bio */ | |
917f32a2 | 158 | btrfs_bio_end_io(btrfs_bio(cb->orig_bio), cb->status); |
86ccbb4d QW |
159 | |
160 | /* Finally free the cb struct */ | |
161 | kfree(cb->compressed_pages); | |
162 | kfree(cb); | |
163 | } | |
164 | ||
81bd9328 CH |
165 | /* |
166 | * Verify the checksums and kick off repair if needed on the uncompressed data | |
167 | * before decompressing it into the original bio and freeing the uncompressed | |
168 | * pages. | |
c8b97818 | 169 | */ |
917f32a2 | 170 | static void end_compressed_bio_read(struct btrfs_bio *bbio) |
c8b97818 | 171 | { |
917f32a2 | 172 | struct compressed_bio *cb = bbio->private; |
81bd9328 CH |
173 | struct inode *inode = cb->inode; |
174 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); | |
175 | struct btrfs_inode *bi = BTRFS_I(inode); | |
176 | bool csum = !(bi->flags & BTRFS_INODE_NODATASUM) && | |
177 | !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state); | |
917f32a2 | 178 | blk_status_t status = bbio->bio.bi_status; |
81bd9328 CH |
179 | struct bvec_iter iter; |
180 | struct bio_vec bv; | |
181 | u32 offset; | |
182 | ||
183 | btrfs_bio_for_each_sector(fs_info, bv, bbio, iter, offset) { | |
184 | u64 start = bbio->file_offset + offset; | |
185 | ||
186 | if (!status && | |
187 | (!csum || !btrfs_check_data_csum(inode, bbio, offset, | |
188 | bv.bv_page, bv.bv_offset))) { | |
0d0a762c JB |
189 | btrfs_clean_io_failure(bi, start, bv.bv_page, |
190 | bv.bv_offset); | |
81bd9328 CH |
191 | } else { |
192 | int ret; | |
e6311f24 | 193 | |
81bd9328 CH |
194 | refcount_inc(&cb->pending_ios); |
195 | ret = btrfs_repair_one_sector(inode, bbio, offset, | |
196 | bv.bv_page, bv.bv_offset, | |
197 | btrfs_submit_data_read_bio); | |
198 | if (ret) { | |
199 | refcount_dec(&cb->pending_ios); | |
200 | status = errno_to_blk_status(ret); | |
201 | } | |
202 | } | |
203 | } | |
d20f7043 | 204 | |
81bd9328 CH |
205 | if (status) |
206 | cb->status = status; | |
8140dc30 | 207 | |
81bd9328 CH |
208 | if (refcount_dec_and_test(&cb->pending_ios)) |
209 | finish_compressed_bio_read(cb); | |
210 | btrfs_bio_free_csum(bbio); | |
917f32a2 | 211 | bio_put(&bbio->bio); |
c8b97818 CM |
212 | } |
213 | ||
214 | /* | |
215 | * Clear the writeback bits on all of the file | |
216 | * pages for a compressed write | |
217 | */ | |
7bdcefc1 FM |
218 | static noinline void end_compressed_writeback(struct inode *inode, |
219 | const struct compressed_bio *cb) | |
c8b97818 | 220 | { |
741ec653 | 221 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
09cbfeaf KS |
222 | unsigned long index = cb->start >> PAGE_SHIFT; |
223 | unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT; | |
a75b81c3 | 224 | struct folio_batch fbatch; |
606f82e7 | 225 | const int errno = blk_status_to_errno(cb->status); |
c8b97818 CM |
226 | int i; |
227 | int ret; | |
228 | ||
606f82e7 JB |
229 | if (errno) |
230 | mapping_set_error(inode->i_mapping, errno); | |
7bdcefc1 | 231 | |
a75b81c3 VMO |
232 | folio_batch_init(&fbatch); |
233 | while (index <= end_index) { | |
234 | ret = filemap_get_folios(inode->i_mapping, &index, end_index, | |
235 | &fbatch); | |
236 | ||
237 | if (ret == 0) | |
238 | return; | |
239 | ||
c8b97818 | 240 | for (i = 0; i < ret; i++) { |
a75b81c3 VMO |
241 | struct folio *folio = fbatch.folios[i]; |
242 | ||
606f82e7 | 243 | if (errno) |
a75b81c3 VMO |
244 | folio_set_error(folio); |
245 | btrfs_page_clamp_clear_writeback(fs_info, &folio->page, | |
741ec653 | 246 | cb->start, cb->len); |
c8b97818 | 247 | } |
a75b81c3 | 248 | folio_batch_release(&fbatch); |
c8b97818 CM |
249 | } |
250 | /* the inode may be gone now */ | |
c8b97818 CM |
251 | } |
252 | ||
6853c64a | 253 | static void finish_compressed_bio_write(struct compressed_bio *cb) |
c8b97818 | 254 | { |
6853c64a | 255 | struct inode *inode = cb->inode; |
1d08ce58 | 256 | unsigned int index; |
c8b97818 | 257 | |
6853c64a QW |
258 | /* |
259 | * Ok, we're the last bio for this extent, step one is to call back | |
260 | * into the FS and do all the end_io operations. | |
c8b97818 | 261 | */ |
38a39ac7 | 262 | btrfs_writepage_endio_finish_ordered(BTRFS_I(inode), NULL, |
c629732d | 263 | cb->start, cb->start + cb->len - 1, |
606f82e7 | 264 | cb->status == BLK_STS_OK); |
c8b97818 | 265 | |
7c0c7269 OS |
266 | if (cb->writeback) |
267 | end_compressed_writeback(inode, cb); | |
6853c64a | 268 | /* Note, our inode could be gone now */ |
c8b97818 CM |
269 | |
270 | /* | |
6853c64a | 271 | * Release the compressed pages, these came from alloc_page and |
c8b97818 CM |
272 | * are not attached to the inode at all |
273 | */ | |
c8b97818 | 274 | for (index = 0; index < cb->nr_pages; index++) { |
6853c64a QW |
275 | struct page *page = cb->compressed_pages[index]; |
276 | ||
c8b97818 | 277 | page->mapping = NULL; |
09cbfeaf | 278 | put_page(page); |
c8b97818 CM |
279 | } |
280 | ||
6853c64a | 281 | /* Finally free the cb struct */ |
c8b97818 CM |
282 | kfree(cb->compressed_pages); |
283 | kfree(cb); | |
6853c64a QW |
284 | } |
285 | ||
fed8a72d CH |
286 | static void btrfs_finish_compressed_write_work(struct work_struct *work) |
287 | { | |
288 | struct compressed_bio *cb = | |
289 | container_of(work, struct compressed_bio, write_end_work); | |
290 | ||
291 | finish_compressed_bio_write(cb); | |
292 | } | |
293 | ||
6853c64a QW |
294 | /* |
295 | * Do the cleanup once all the compressed pages hit the disk. This will clear | |
296 | * writeback on the file pages and free the compressed pages. | |
297 | * | |
298 | * This also calls the writeback end hooks for the file pages so that metadata | |
299 | * and checksums can be updated in the file. | |
300 | */ | |
917f32a2 | 301 | static void end_compressed_bio_write(struct btrfs_bio *bbio) |
6853c64a | 302 | { |
917f32a2 | 303 | struct compressed_bio *cb = bbio->private; |
6853c64a | 304 | |
917f32a2 CH |
305 | if (bbio->bio.bi_status) |
306 | cb->status = bbio->bio.bi_status; | |
6853c64a | 307 | |
524bcd1e | 308 | if (refcount_dec_and_test(&cb->pending_ios)) { |
fed8a72d | 309 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
6853c64a | 310 | |
917f32a2 | 311 | btrfs_record_physical_zoned(cb->inode, cb->start, &bbio->bio); |
fed8a72d CH |
312 | queue_work(fs_info->compressed_write_workers, &cb->write_end_work); |
313 | } | |
917f32a2 | 314 | bio_put(&bbio->bio); |
c8b97818 CM |
315 | } |
316 | ||
22c306fe | 317 | /* |
f472c28f QW |
318 | * Allocate a compressed_bio, which will be used to read/write on-disk |
319 | * (aka, compressed) * data. | |
320 | * | |
321 | * @cb: The compressed_bio structure, which records all the needed | |
322 | * information to bind the compressed data to the uncompressed | |
323 | * page cache. | |
324 | * @disk_byten: The logical bytenr where the compressed data will be read | |
325 | * from or written to. | |
326 | * @endio_func: The endio function to call after the IO for compressed data | |
327 | * is finished. | |
328 | * @next_stripe_start: Return value of logical bytenr of where next stripe starts. | |
329 | * Let the caller know to only fill the bio up to the stripe | |
330 | * boundary. | |
22c306fe | 331 | */ |
f472c28f QW |
332 | |
333 | ||
22c306fe | 334 | static struct bio *alloc_compressed_bio(struct compressed_bio *cb, u64 disk_bytenr, |
917f32a2 CH |
335 | blk_opf_t opf, |
336 | btrfs_bio_end_io_t endio_func, | |
f472c28f | 337 | u64 *next_stripe_start) |
22c306fe | 338 | { |
f472c28f QW |
339 | struct btrfs_fs_info *fs_info = btrfs_sb(cb->inode->i_sb); |
340 | struct btrfs_io_geometry geom; | |
341 | struct extent_map *em; | |
22c306fe | 342 | struct bio *bio; |
f472c28f | 343 | int ret; |
22c306fe | 344 | |
917f32a2 | 345 | bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, endio_func, cb); |
22c306fe | 346 | bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT; |
22c306fe | 347 | |
f472c28f QW |
348 | em = btrfs_get_chunk_map(fs_info, disk_bytenr, fs_info->sectorsize); |
349 | if (IS_ERR(em)) { | |
350 | bio_put(bio); | |
351 | return ERR_CAST(em); | |
352 | } | |
22c306fe | 353 | |
f472c28f QW |
354 | if (bio_op(bio) == REQ_OP_ZONE_APPEND) |
355 | bio_set_dev(bio, em->map_lookup->stripes[0].dev->bdev); | |
356 | ||
357 | ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio), disk_bytenr, &geom); | |
358 | free_extent_map(em); | |
359 | if (ret < 0) { | |
360 | bio_put(bio); | |
361 | return ERR_PTR(ret); | |
22c306fe | 362 | } |
f472c28f | 363 | *next_stripe_start = disk_bytenr + geom.len; |
524bcd1e | 364 | refcount_inc(&cb->pending_ios); |
22c306fe QW |
365 | return bio; |
366 | } | |
367 | ||
c8b97818 CM |
368 | /* |
369 | * worker function to build and submit bios for previously compressed pages. | |
370 | * The corresponding pages in the inode should be marked for writeback | |
371 | * and the compressed pages should have a reference on them for dropping | |
372 | * when the IO is complete. | |
373 | * | |
374 | * This also checksums the file bytes and gets things ready for | |
375 | * the end io hooks. | |
376 | */ | |
c7ee1819 | 377 | blk_status_t btrfs_submit_compressed_write(struct btrfs_inode *inode, u64 start, |
65b5355f AJ |
378 | unsigned int len, u64 disk_start, |
379 | unsigned int compressed_len, | |
c8b97818 | 380 | struct page **compressed_pages, |
65b5355f | 381 | unsigned int nr_pages, |
bf9486d6 | 382 | blk_opf_t write_flags, |
7c0c7269 OS |
383 | struct cgroup_subsys_state *blkcg_css, |
384 | bool writeback) | |
c8b97818 | 385 | { |
c7ee1819 | 386 | struct btrfs_fs_info *fs_info = inode->root->fs_info; |
c8b97818 | 387 | struct bio *bio = NULL; |
c8b97818 | 388 | struct compressed_bio *cb; |
91507240 | 389 | u64 cur_disk_bytenr = disk_start; |
f472c28f | 390 | u64 next_stripe_start; |
524bcd1e | 391 | blk_status_t ret = BLK_STS_OK; |
c7ee1819 | 392 | int skip_sum = inode->flags & BTRFS_INODE_NODATASUM; |
764c7c9a | 393 | const bool use_append = btrfs_use_zone_append(inode, disk_start); |
bf9486d6 | 394 | const enum req_op bio_op = use_append ? REQ_OP_ZONE_APPEND : REQ_OP_WRITE; |
c8b97818 | 395 | |
bbbff01a QW |
396 | ASSERT(IS_ALIGNED(start, fs_info->sectorsize) && |
397 | IS_ALIGNED(len, fs_info->sectorsize)); | |
81bd9328 | 398 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
dac97e51 | 399 | if (!cb) |
4e4cbee9 | 400 | return BLK_STS_RESOURCE; |
524bcd1e | 401 | refcount_set(&cb->pending_ios, 1); |
606f82e7 | 402 | cb->status = BLK_STS_OK; |
c7ee1819 | 403 | cb->inode = &inode->vfs_inode; |
c8b97818 CM |
404 | cb->start = start; |
405 | cb->len = len; | |
406 | cb->compressed_pages = compressed_pages; | |
407 | cb->compressed_len = compressed_len; | |
7c0c7269 | 408 | cb->writeback = writeback; |
fed8a72d | 409 | INIT_WORK(&cb->write_end_work, btrfs_finish_compressed_write_work); |
c8b97818 CM |
410 | cb->nr_pages = nr_pages; |
411 | ||
acee08aa DZ |
412 | if (blkcg_css) |
413 | kthread_associate_blkcg(blkcg_css); | |
414 | ||
91507240 QW |
415 | while (cur_disk_bytenr < disk_start + compressed_len) { |
416 | u64 offset = cur_disk_bytenr - disk_start; | |
417 | unsigned int index = offset >> PAGE_SHIFT; | |
418 | unsigned int real_size; | |
419 | unsigned int added; | |
420 | struct page *page = compressed_pages[index]; | |
421 | bool submit = false; | |
422 | ||
423 | /* Allocate new bio if submitted or not yet allocated */ | |
424 | if (!bio) { | |
425 | bio = alloc_compressed_bio(cb, cur_disk_bytenr, | |
426 | bio_op | write_flags, end_compressed_bio_write, | |
427 | &next_stripe_start); | |
428 | if (IS_ERR(bio)) { | |
429 | ret = errno_to_blk_status(PTR_ERR(bio)); | |
524bcd1e | 430 | break; |
91507240 | 431 | } |
acee08aa DZ |
432 | if (blkcg_css) |
433 | bio->bi_opf |= REQ_CGROUP_PUNT; | |
764c7c9a | 434 | } |
4c80a97d | 435 | /* |
91507240 QW |
436 | * We should never reach next_stripe_start start as we will |
437 | * submit comp_bio when reach the boundary immediately. | |
4c80a97d | 438 | */ |
91507240 | 439 | ASSERT(cur_disk_bytenr != next_stripe_start); |
c8b97818 | 440 | |
91507240 QW |
441 | /* |
442 | * We have various limits on the real read size: | |
443 | * - stripe boundary | |
444 | * - page boundary | |
445 | * - compressed length boundary | |
446 | */ | |
447 | real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_bytenr); | |
448 | real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset)); | |
449 | real_size = min_t(u64, real_size, compressed_len - offset); | |
450 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
451 | ||
452 | if (use_append) | |
453 | added = bio_add_zone_append_page(bio, page, real_size, | |
454 | offset_in_page(offset)); | |
455 | else | |
456 | added = bio_add_page(bio, page, real_size, | |
457 | offset_in_page(offset)); | |
458 | /* Reached zoned boundary */ | |
459 | if (added == 0) | |
460 | submit = true; | |
461 | ||
462 | cur_disk_bytenr += added; | |
463 | /* Reached stripe boundary */ | |
464 | if (cur_disk_bytenr == next_stripe_start) | |
465 | submit = true; | |
466 | ||
467 | /* Finished the range */ | |
468 | if (cur_disk_bytenr == disk_start + compressed_len) | |
469 | submit = true; | |
470 | ||
471 | if (submit) { | |
e55179b3 | 472 | if (!skip_sum) { |
e331f6b1 | 473 | ret = btrfs_csum_one_bio(inode, bio, start, true); |
524bcd1e | 474 | if (ret) { |
917f32a2 | 475 | btrfs_bio_end_io(btrfs_bio(bio), ret); |
524bcd1e CH |
476 | break; |
477 | } | |
f5daf2c7 | 478 | } |
c8b97818 | 479 | |
fed8a72d | 480 | ASSERT(bio->bi_iter.bi_size); |
1a722d8f | 481 | btrfs_submit_bio(fs_info, bio, 0); |
91507240 | 482 | bio = NULL; |
c8b97818 | 483 | } |
771ed689 | 484 | cond_resched(); |
c8b97818 | 485 | } |
d20f7043 | 486 | |
acee08aa DZ |
487 | if (blkcg_css) |
488 | kthread_associate_blkcg(NULL); | |
489 | ||
524bcd1e CH |
490 | if (refcount_dec_and_test(&cb->pending_ios)) |
491 | finish_compressed_bio_write(cb); | |
6853c64a | 492 | return ret; |
c8b97818 CM |
493 | } |
494 | ||
2a4d0c90 CH |
495 | static u64 bio_end_offset(struct bio *bio) |
496 | { | |
c45a8f2d | 497 | struct bio_vec *last = bio_last_bvec_all(bio); |
2a4d0c90 CH |
498 | |
499 | return page_offset(last->bv_page) + last->bv_len + last->bv_offset; | |
500 | } | |
501 | ||
6a404910 QW |
502 | /* |
503 | * Add extra pages in the same compressed file extent so that we don't need to | |
504 | * re-read the same extent again and again. | |
505 | * | |
506 | * NOTE: this won't work well for subpage, as for subpage read, we lock the | |
507 | * full page then submit bio for each compressed/regular extents. | |
508 | * | |
509 | * This means, if we have several sectors in the same page points to the same | |
510 | * on-disk compressed data, we will re-read the same extent many times and | |
511 | * this function can only help for the next page. | |
512 | */ | |
771ed689 CM |
513 | static noinline int add_ra_bio_pages(struct inode *inode, |
514 | u64 compressed_end, | |
4088a47e | 515 | struct compressed_bio *cb, |
82e60d00 | 516 | int *memstall, unsigned long *pflags) |
771ed689 | 517 | { |
6a404910 | 518 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
771ed689 | 519 | unsigned long end_index; |
6a404910 | 520 | u64 cur = bio_end_offset(cb->orig_bio); |
771ed689 CM |
521 | u64 isize = i_size_read(inode); |
522 | int ret; | |
523 | struct page *page; | |
771ed689 CM |
524 | struct extent_map *em; |
525 | struct address_space *mapping = inode->i_mapping; | |
771ed689 CM |
526 | struct extent_map_tree *em_tree; |
527 | struct extent_io_tree *tree; | |
6a404910 | 528 | int sectors_missed = 0; |
771ed689 | 529 | |
771ed689 CM |
530 | em_tree = &BTRFS_I(inode)->extent_tree; |
531 | tree = &BTRFS_I(inode)->io_tree; | |
532 | ||
533 | if (isize == 0) | |
534 | return 0; | |
535 | ||
ca62e85d QW |
536 | /* |
537 | * For current subpage support, we only support 64K page size, | |
538 | * which means maximum compressed extent size (128K) is just 2x page | |
539 | * size. | |
540 | * This makes readahead less effective, so here disable readahead for | |
541 | * subpage for now, until full compressed write is supported. | |
542 | */ | |
543 | if (btrfs_sb(inode->i_sb)->sectorsize < PAGE_SIZE) | |
544 | return 0; | |
545 | ||
09cbfeaf | 546 | end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT; |
771ed689 | 547 | |
6a404910 QW |
548 | while (cur < compressed_end) { |
549 | u64 page_end; | |
550 | u64 pg_index = cur >> PAGE_SHIFT; | |
551 | u32 add_size; | |
771ed689 | 552 | |
306e16ce | 553 | if (pg_index > end_index) |
771ed689 CM |
554 | break; |
555 | ||
0a943c65 | 556 | page = xa_load(&mapping->i_pages, pg_index); |
3159f943 | 557 | if (page && !xa_is_value(page)) { |
6a404910 QW |
558 | sectors_missed += (PAGE_SIZE - offset_in_page(cur)) >> |
559 | fs_info->sectorsize_bits; | |
560 | ||
561 | /* Beyond threshold, no need to continue */ | |
562 | if (sectors_missed > 4) | |
771ed689 | 563 | break; |
6a404910 QW |
564 | |
565 | /* | |
566 | * Jump to next page start as we already have page for | |
567 | * current offset. | |
568 | */ | |
569 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
570 | continue; | |
771ed689 CM |
571 | } |
572 | ||
c62d2555 MH |
573 | page = __page_cache_alloc(mapping_gfp_constraint(mapping, |
574 | ~__GFP_FS)); | |
771ed689 CM |
575 | if (!page) |
576 | break; | |
577 | ||
c62d2555 | 578 | if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) { |
09cbfeaf | 579 | put_page(page); |
6a404910 QW |
580 | /* There is already a page, skip to page end */ |
581 | cur = (pg_index << PAGE_SHIFT) + PAGE_SIZE; | |
582 | continue; | |
771ed689 CM |
583 | } |
584 | ||
82e60d00 | 585 | if (!*memstall && PageWorkingset(page)) { |
4088a47e | 586 | psi_memstall_enter(pflags); |
82e60d00 JW |
587 | *memstall = 1; |
588 | } | |
4088a47e | 589 | |
32443de3 QW |
590 | ret = set_page_extent_mapped(page); |
591 | if (ret < 0) { | |
592 | unlock_page(page); | |
593 | put_page(page); | |
594 | break; | |
595 | } | |
596 | ||
6a404910 | 597 | page_end = (pg_index << PAGE_SHIFT) + PAGE_SIZE - 1; |
570eb97b | 598 | lock_extent(tree, cur, page_end, NULL); |
890871be | 599 | read_lock(&em_tree->lock); |
6a404910 | 600 | em = lookup_extent_mapping(em_tree, cur, page_end + 1 - cur); |
890871be | 601 | read_unlock(&em_tree->lock); |
771ed689 | 602 | |
6a404910 QW |
603 | /* |
604 | * At this point, we have a locked page in the page cache for | |
605 | * these bytes in the file. But, we have to make sure they map | |
606 | * to this compressed extent on disk. | |
607 | */ | |
608 | if (!em || cur < em->start || | |
609 | (cur + fs_info->sectorsize > extent_map_end(em)) || | |
4f024f37 | 610 | (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) { |
771ed689 | 611 | free_extent_map(em); |
570eb97b | 612 | unlock_extent(tree, cur, page_end, NULL); |
771ed689 | 613 | unlock_page(page); |
09cbfeaf | 614 | put_page(page); |
771ed689 CM |
615 | break; |
616 | } | |
617 | free_extent_map(em); | |
618 | ||
619 | if (page->index == end_index) { | |
7073017a | 620 | size_t zero_offset = offset_in_page(isize); |
771ed689 CM |
621 | |
622 | if (zero_offset) { | |
623 | int zeros; | |
09cbfeaf | 624 | zeros = PAGE_SIZE - zero_offset; |
d048b9c2 | 625 | memzero_page(page, zero_offset, zeros); |
771ed689 CM |
626 | } |
627 | } | |
628 | ||
6a404910 QW |
629 | add_size = min(em->start + em->len, page_end + 1) - cur; |
630 | ret = bio_add_page(cb->orig_bio, page, add_size, offset_in_page(cur)); | |
631 | if (ret != add_size) { | |
570eb97b | 632 | unlock_extent(tree, cur, page_end, NULL); |
771ed689 | 633 | unlock_page(page); |
09cbfeaf | 634 | put_page(page); |
771ed689 CM |
635 | break; |
636 | } | |
6a404910 QW |
637 | /* |
638 | * If it's subpage, we also need to increase its | |
639 | * subpage::readers number, as at endio we will decrease | |
640 | * subpage::readers and to unlock the page. | |
641 | */ | |
642 | if (fs_info->sectorsize < PAGE_SIZE) | |
643 | btrfs_subpage_start_reader(fs_info, page, cur, add_size); | |
644 | put_page(page); | |
645 | cur += add_size; | |
771ed689 | 646 | } |
771ed689 CM |
647 | return 0; |
648 | } | |
649 | ||
c8b97818 CM |
650 | /* |
651 | * for a compressed read, the bio we get passed has all the inode pages | |
652 | * in it. We don't actually do IO on those pages but allocate new ones | |
653 | * to hold the compressed pages on disk. | |
654 | * | |
4f024f37 | 655 | * bio->bi_iter.bi_sector points to the compressed extent on disk |
c8b97818 | 656 | * bio->bi_io_vec points to all of the inode pages |
c8b97818 CM |
657 | * |
658 | * After the compressed pages are read, we copy the bytes into the | |
659 | * bio we were passed and then call the bio end_io calls | |
660 | */ | |
cb4411dd | 661 | void btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, |
1d8fa2e2 | 662 | int mirror_num) |
c8b97818 | 663 | { |
0b246afa | 664 | struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb); |
c8b97818 CM |
665 | struct extent_map_tree *em_tree; |
666 | struct compressed_bio *cb; | |
356b4a2d | 667 | unsigned int compressed_len; |
f472c28f QW |
668 | struct bio *comp_bio = NULL; |
669 | const u64 disk_bytenr = bio->bi_iter.bi_sector << SECTOR_SHIFT; | |
670 | u64 cur_disk_byte = disk_bytenr; | |
671 | u64 next_stripe_start; | |
557023ea | 672 | u64 file_offset; |
e04ca626 CM |
673 | u64 em_len; |
674 | u64 em_start; | |
c8b97818 | 675 | struct extent_map *em; |
82e60d00 JW |
676 | unsigned long pflags; |
677 | int memstall = 0; | |
f9f15de8 | 678 | blk_status_t ret; |
dd137dd1 STD |
679 | int ret2; |
680 | int i; | |
c8b97818 | 681 | |
c8b97818 CM |
682 | em_tree = &BTRFS_I(inode)->extent_tree; |
683 | ||
557023ea QW |
684 | file_offset = bio_first_bvec_all(bio)->bv_offset + |
685 | page_offset(bio_first_page_all(bio)); | |
686 | ||
c8b97818 | 687 | /* we need the actual starting offset of this extent in the file */ |
890871be | 688 | read_lock(&em_tree->lock); |
557023ea | 689 | em = lookup_extent_mapping(em_tree, file_offset, fs_info->sectorsize); |
890871be | 690 | read_unlock(&em_tree->lock); |
f9f15de8 JB |
691 | if (!em) { |
692 | ret = BLK_STS_IOERR; | |
693 | goto out; | |
694 | } | |
c8b97818 | 695 | |
557023ea | 696 | ASSERT(em->compress_type != BTRFS_COMPRESS_NONE); |
d20f7043 | 697 | compressed_len = em->block_len; |
81bd9328 | 698 | cb = kmalloc(sizeof(struct compressed_bio), GFP_NOFS); |
f9f15de8 JB |
699 | if (!cb) { |
700 | ret = BLK_STS_RESOURCE; | |
6b82ce8d | 701 | goto out; |
f9f15de8 | 702 | } |
6b82ce8d | 703 | |
524bcd1e | 704 | refcount_set(&cb->pending_ios, 1); |
606f82e7 | 705 | cb->status = BLK_STS_OK; |
c8b97818 CM |
706 | cb->inode = inode; |
707 | ||
ff5b7ee3 | 708 | cb->start = em->orig_start; |
e04ca626 CM |
709 | em_len = em->len; |
710 | em_start = em->start; | |
d20f7043 | 711 | |
81381053 | 712 | cb->len = bio->bi_iter.bi_size; |
c8b97818 | 713 | cb->compressed_len = compressed_len; |
1d8fa2e2 | 714 | cb->compress_type = em->compress_type; |
c8b97818 CM |
715 | cb->orig_bio = bio; |
716 | ||
1d8fa2e2 GR |
717 | free_extent_map(em); |
718 | em = NULL; | |
719 | ||
dd137dd1 STD |
720 | cb->nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE); |
721 | cb->compressed_pages = kcalloc(cb->nr_pages, sizeof(struct page *), GFP_NOFS); | |
f9f15de8 JB |
722 | if (!cb->compressed_pages) { |
723 | ret = BLK_STS_RESOURCE; | |
dd137dd1 | 724 | goto fail; |
f9f15de8 | 725 | } |
6b82ce8d | 726 | |
dd137dd1 STD |
727 | ret2 = btrfs_alloc_page_array(cb->nr_pages, cb->compressed_pages); |
728 | if (ret2) { | |
729 | ret = BLK_STS_RESOURCE; | |
730 | goto fail; | |
c8b97818 | 731 | } |
c8b97818 | 732 | |
82e60d00 | 733 | add_ra_bio_pages(inode, em_start + em_len, cb, &memstall, &pflags); |
771ed689 | 734 | |
771ed689 | 735 | /* include any pages we added in add_ra-bio_pages */ |
81381053 | 736 | cb->len = bio->bi_iter.bi_size; |
771ed689 | 737 | |
f472c28f QW |
738 | while (cur_disk_byte < disk_bytenr + compressed_len) { |
739 | u64 offset = cur_disk_byte - disk_bytenr; | |
740 | unsigned int index = offset >> PAGE_SHIFT; | |
741 | unsigned int real_size; | |
742 | unsigned int added; | |
743 | struct page *page = cb->compressed_pages[index]; | |
744 | bool submit = false; | |
c8b97818 | 745 | |
f472c28f QW |
746 | /* Allocate new bio if submitted or not yet allocated */ |
747 | if (!comp_bio) { | |
748 | comp_bio = alloc_compressed_bio(cb, cur_disk_byte, | |
749 | REQ_OP_READ, end_compressed_bio_read, | |
750 | &next_stripe_start); | |
751 | if (IS_ERR(comp_bio)) { | |
524bcd1e CH |
752 | cb->status = errno_to_blk_status(PTR_ERR(comp_bio)); |
753 | break; | |
f472c28f QW |
754 | } |
755 | } | |
756 | /* | |
757 | * We should never reach next_stripe_start start as we will | |
758 | * submit comp_bio when reach the boundary immediately. | |
759 | */ | |
760 | ASSERT(cur_disk_byte != next_stripe_start); | |
761 | /* | |
762 | * We have various limit on the real read size: | |
763 | * - stripe boundary | |
764 | * - page boundary | |
765 | * - compressed length boundary | |
766 | */ | |
767 | real_size = min_t(u64, U32_MAX, next_stripe_start - cur_disk_byte); | |
768 | real_size = min_t(u64, real_size, PAGE_SIZE - offset_in_page(offset)); | |
769 | real_size = min_t(u64, real_size, compressed_len - offset); | |
770 | ASSERT(IS_ALIGNED(real_size, fs_info->sectorsize)); | |
4e4cbee9 | 771 | |
f472c28f | 772 | added = bio_add_page(comp_bio, page, real_size, offset_in_page(offset)); |
be6a1361 | 773 | /* |
f472c28f QW |
774 | * Maximum compressed extent is smaller than bio size limit, |
775 | * thus bio_add_page() should always success. | |
be6a1361 | 776 | */ |
f472c28f QW |
777 | ASSERT(added == real_size); |
778 | cur_disk_byte += added; | |
be6a1361 | 779 | |
f472c28f QW |
780 | /* Reached stripe boundary, need to submit */ |
781 | if (cur_disk_byte == next_stripe_start) | |
782 | submit = true; | |
d20f7043 | 783 | |
f472c28f QW |
784 | /* Has finished the range, need to submit */ |
785 | if (cur_disk_byte == disk_bytenr + compressed_len) | |
786 | submit = true; | |
c8b97818 | 787 | |
f472c28f | 788 | if (submit) { |
81bd9328 CH |
789 | /* Save the original iter for read repair */ |
790 | if (bio_op(comp_bio) == REQ_OP_READ) | |
791 | btrfs_bio(comp_bio)->iter = comp_bio->bi_iter; | |
10fe6ca8 | 792 | |
81bd9328 CH |
793 | /* |
794 | * Save the initial offset of this chunk, as there | |
795 | * is no direct correlation between compressed pages and | |
796 | * the original file offset. The field is only used for | |
797 | * priting error messages. | |
798 | */ | |
799 | btrfs_bio(comp_bio)->file_offset = file_offset; | |
10fe6ca8 | 800 | |
81bd9328 | 801 | ret = btrfs_lookup_bio_sums(inode, comp_bio, NULL); |
524bcd1e | 802 | if (ret) { |
917f32a2 | 803 | btrfs_bio_end_io(btrfs_bio(comp_bio), ret); |
524bcd1e CH |
804 | break; |
805 | } | |
d20f7043 | 806 | |
fed8a72d | 807 | ASSERT(comp_bio->bi_iter.bi_size); |
1a722d8f | 808 | btrfs_submit_bio(fs_info, comp_bio, mirror_num); |
f472c28f | 809 | comp_bio = NULL; |
c8b97818 | 810 | } |
c8b97818 | 811 | } |
524bcd1e | 812 | |
82e60d00 | 813 | if (memstall) |
4088a47e CH |
814 | psi_memstall_leave(&pflags); |
815 | ||
524bcd1e CH |
816 | if (refcount_dec_and_test(&cb->pending_ios)) |
817 | finish_compressed_bio_read(cb); | |
cb4411dd | 818 | return; |
6b82ce8d | 819 | |
dd137dd1 STD |
820 | fail: |
821 | if (cb->compressed_pages) { | |
822 | for (i = 0; i < cb->nr_pages; i++) { | |
823 | if (cb->compressed_pages[i]) | |
824 | __free_page(cb->compressed_pages[i]); | |
825 | } | |
15e3004a | 826 | } |
6b82ce8d | 827 | |
828 | kfree(cb->compressed_pages); | |
6b82ce8d | 829 | kfree(cb); |
830 | out: | |
831 | free_extent_map(em); | |
917f32a2 | 832 | btrfs_bio_end_io(btrfs_bio(bio), ret); |
cb4411dd | 833 | return; |
c8b97818 | 834 | } |
261507a0 | 835 | |
17b5a6c1 TT |
836 | /* |
837 | * Heuristic uses systematic sampling to collect data from the input data | |
838 | * range, the logic can be tuned by the following constants: | |
839 | * | |
840 | * @SAMPLING_READ_SIZE - how many bytes will be copied from for each sample | |
841 | * @SAMPLING_INTERVAL - range from which the sampled data can be collected | |
842 | */ | |
843 | #define SAMPLING_READ_SIZE (16) | |
844 | #define SAMPLING_INTERVAL (256) | |
845 | ||
846 | /* | |
847 | * For statistical analysis of the input data we consider bytes that form a | |
848 | * Galois Field of 256 objects. Each object has an attribute count, ie. how | |
849 | * many times the object appeared in the sample. | |
850 | */ | |
851 | #define BUCKET_SIZE (256) | |
852 | ||
853 | /* | |
854 | * The size of the sample is based on a statistical sampling rule of thumb. | |
855 | * The common way is to perform sampling tests as long as the number of | |
856 | * elements in each cell is at least 5. | |
857 | * | |
858 | * Instead of 5, we choose 32 to obtain more accurate results. | |
859 | * If the data contain the maximum number of symbols, which is 256, we obtain a | |
860 | * sample size bound by 8192. | |
861 | * | |
862 | * For a sample of at most 8KB of data per data range: 16 consecutive bytes | |
863 | * from up to 512 locations. | |
864 | */ | |
865 | #define MAX_SAMPLE_SIZE (BTRFS_MAX_UNCOMPRESSED * \ | |
866 | SAMPLING_READ_SIZE / SAMPLING_INTERVAL) | |
867 | ||
868 | struct bucket_item { | |
869 | u32 count; | |
870 | }; | |
4e439a0b TT |
871 | |
872 | struct heuristic_ws { | |
17b5a6c1 TT |
873 | /* Partial copy of input data */ |
874 | u8 *sample; | |
a440d48c | 875 | u32 sample_size; |
17b5a6c1 TT |
876 | /* Buckets store counters for each byte value */ |
877 | struct bucket_item *bucket; | |
440c840c TT |
878 | /* Sorting buffer */ |
879 | struct bucket_item *bucket_b; | |
4e439a0b TT |
880 | struct list_head list; |
881 | }; | |
882 | ||
92ee5530 DZ |
883 | static struct workspace_manager heuristic_wsm; |
884 | ||
4e439a0b TT |
885 | static void free_heuristic_ws(struct list_head *ws) |
886 | { | |
887 | struct heuristic_ws *workspace; | |
888 | ||
889 | workspace = list_entry(ws, struct heuristic_ws, list); | |
890 | ||
17b5a6c1 TT |
891 | kvfree(workspace->sample); |
892 | kfree(workspace->bucket); | |
440c840c | 893 | kfree(workspace->bucket_b); |
4e439a0b TT |
894 | kfree(workspace); |
895 | } | |
896 | ||
7bf49943 | 897 | static struct list_head *alloc_heuristic_ws(unsigned int level) |
4e439a0b TT |
898 | { |
899 | struct heuristic_ws *ws; | |
900 | ||
901 | ws = kzalloc(sizeof(*ws), GFP_KERNEL); | |
902 | if (!ws) | |
903 | return ERR_PTR(-ENOMEM); | |
904 | ||
17b5a6c1 TT |
905 | ws->sample = kvmalloc(MAX_SAMPLE_SIZE, GFP_KERNEL); |
906 | if (!ws->sample) | |
907 | goto fail; | |
908 | ||
909 | ws->bucket = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket), GFP_KERNEL); | |
910 | if (!ws->bucket) | |
911 | goto fail; | |
4e439a0b | 912 | |
440c840c TT |
913 | ws->bucket_b = kcalloc(BUCKET_SIZE, sizeof(*ws->bucket_b), GFP_KERNEL); |
914 | if (!ws->bucket_b) | |
915 | goto fail; | |
916 | ||
17b5a6c1 | 917 | INIT_LIST_HEAD(&ws->list); |
4e439a0b | 918 | return &ws->list; |
17b5a6c1 TT |
919 | fail: |
920 | free_heuristic_ws(&ws->list); | |
921 | return ERR_PTR(-ENOMEM); | |
4e439a0b TT |
922 | } |
923 | ||
ca4ac360 | 924 | const struct btrfs_compress_op btrfs_heuristic_compress = { |
be951045 | 925 | .workspace_manager = &heuristic_wsm, |
ca4ac360 DZ |
926 | }; |
927 | ||
e8c9f186 | 928 | static const struct btrfs_compress_op * const btrfs_compress_op[] = { |
ca4ac360 DZ |
929 | /* The heuristic is represented as compression type 0 */ |
930 | &btrfs_heuristic_compress, | |
261507a0 | 931 | &btrfs_zlib_compress, |
a6fa6fae | 932 | &btrfs_lzo_compress, |
5c1aab1d | 933 | &btrfs_zstd_compress, |
261507a0 LZ |
934 | }; |
935 | ||
c778df14 DS |
936 | static struct list_head *alloc_workspace(int type, unsigned int level) |
937 | { | |
938 | switch (type) { | |
939 | case BTRFS_COMPRESS_NONE: return alloc_heuristic_ws(level); | |
940 | case BTRFS_COMPRESS_ZLIB: return zlib_alloc_workspace(level); | |
941 | case BTRFS_COMPRESS_LZO: return lzo_alloc_workspace(level); | |
942 | case BTRFS_COMPRESS_ZSTD: return zstd_alloc_workspace(level); | |
943 | default: | |
944 | /* | |
945 | * This can't happen, the type is validated several times | |
946 | * before we get here. | |
947 | */ | |
948 | BUG(); | |
949 | } | |
950 | } | |
951 | ||
1e002351 DS |
952 | static void free_workspace(int type, struct list_head *ws) |
953 | { | |
954 | switch (type) { | |
955 | case BTRFS_COMPRESS_NONE: return free_heuristic_ws(ws); | |
956 | case BTRFS_COMPRESS_ZLIB: return zlib_free_workspace(ws); | |
957 | case BTRFS_COMPRESS_LZO: return lzo_free_workspace(ws); | |
958 | case BTRFS_COMPRESS_ZSTD: return zstd_free_workspace(ws); | |
959 | default: | |
960 | /* | |
961 | * This can't happen, the type is validated several times | |
962 | * before we get here. | |
963 | */ | |
964 | BUG(); | |
965 | } | |
966 | } | |
967 | ||
d5517033 | 968 | static void btrfs_init_workspace_manager(int type) |
261507a0 | 969 | { |
0cf25213 | 970 | struct workspace_manager *wsm; |
4e439a0b | 971 | struct list_head *workspace; |
261507a0 | 972 | |
0cf25213 | 973 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
974 | INIT_LIST_HEAD(&wsm->idle_ws); |
975 | spin_lock_init(&wsm->ws_lock); | |
976 | atomic_set(&wsm->total_ws, 0); | |
977 | init_waitqueue_head(&wsm->ws_wait); | |
f77dd0d6 | 978 | |
1666edab DZ |
979 | /* |
980 | * Preallocate one workspace for each compression type so we can | |
981 | * guarantee forward progress in the worst case | |
982 | */ | |
c778df14 | 983 | workspace = alloc_workspace(type, 0); |
1666edab DZ |
984 | if (IS_ERR(workspace)) { |
985 | pr_warn( | |
986 | "BTRFS: cannot preallocate compression workspace, will try later\n"); | |
987 | } else { | |
92ee5530 DZ |
988 | atomic_set(&wsm->total_ws, 1); |
989 | wsm->free_ws = 1; | |
990 | list_add(workspace, &wsm->idle_ws); | |
1666edab DZ |
991 | } |
992 | } | |
993 | ||
2510307e | 994 | static void btrfs_cleanup_workspace_manager(int type) |
1666edab | 995 | { |
2dba7143 | 996 | struct workspace_manager *wsman; |
1666edab DZ |
997 | struct list_head *ws; |
998 | ||
2dba7143 | 999 | wsman = btrfs_compress_op[type]->workspace_manager; |
1666edab DZ |
1000 | while (!list_empty(&wsman->idle_ws)) { |
1001 | ws = wsman->idle_ws.next; | |
1002 | list_del(ws); | |
1e002351 | 1003 | free_workspace(type, ws); |
1666edab | 1004 | atomic_dec(&wsman->total_ws); |
261507a0 | 1005 | } |
261507a0 LZ |
1006 | } |
1007 | ||
1008 | /* | |
e721e49d DS |
1009 | * This finds an available workspace or allocates a new one. |
1010 | * If it's not possible to allocate a new one, waits until there's one. | |
1011 | * Preallocation makes a forward progress guarantees and we do not return | |
1012 | * errors. | |
261507a0 | 1013 | */ |
5907a9bb | 1014 | struct list_head *btrfs_get_workspace(int type, unsigned int level) |
261507a0 | 1015 | { |
5907a9bb | 1016 | struct workspace_manager *wsm; |
261507a0 LZ |
1017 | struct list_head *workspace; |
1018 | int cpus = num_online_cpus(); | |
fe308533 | 1019 | unsigned nofs_flag; |
4e439a0b TT |
1020 | struct list_head *idle_ws; |
1021 | spinlock_t *ws_lock; | |
1022 | atomic_t *total_ws; | |
1023 | wait_queue_head_t *ws_wait; | |
1024 | int *free_ws; | |
1025 | ||
5907a9bb | 1026 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1027 | idle_ws = &wsm->idle_ws; |
1028 | ws_lock = &wsm->ws_lock; | |
1029 | total_ws = &wsm->total_ws; | |
1030 | ws_wait = &wsm->ws_wait; | |
1031 | free_ws = &wsm->free_ws; | |
261507a0 | 1032 | |
261507a0 | 1033 | again: |
d9187649 BL |
1034 | spin_lock(ws_lock); |
1035 | if (!list_empty(idle_ws)) { | |
1036 | workspace = idle_ws->next; | |
261507a0 | 1037 | list_del(workspace); |
6ac10a6a | 1038 | (*free_ws)--; |
d9187649 | 1039 | spin_unlock(ws_lock); |
261507a0 LZ |
1040 | return workspace; |
1041 | ||
1042 | } | |
6ac10a6a | 1043 | if (atomic_read(total_ws) > cpus) { |
261507a0 LZ |
1044 | DEFINE_WAIT(wait); |
1045 | ||
d9187649 BL |
1046 | spin_unlock(ws_lock); |
1047 | prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6ac10a6a | 1048 | if (atomic_read(total_ws) > cpus && !*free_ws) |
261507a0 | 1049 | schedule(); |
d9187649 | 1050 | finish_wait(ws_wait, &wait); |
261507a0 LZ |
1051 | goto again; |
1052 | } | |
6ac10a6a | 1053 | atomic_inc(total_ws); |
d9187649 | 1054 | spin_unlock(ws_lock); |
261507a0 | 1055 | |
fe308533 DS |
1056 | /* |
1057 | * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have | |
1058 | * to turn it off here because we might get called from the restricted | |
1059 | * context of btrfs_compress_bio/btrfs_compress_pages | |
1060 | */ | |
1061 | nofs_flag = memalloc_nofs_save(); | |
c778df14 | 1062 | workspace = alloc_workspace(type, level); |
fe308533 DS |
1063 | memalloc_nofs_restore(nofs_flag); |
1064 | ||
261507a0 | 1065 | if (IS_ERR(workspace)) { |
6ac10a6a | 1066 | atomic_dec(total_ws); |
d9187649 | 1067 | wake_up(ws_wait); |
e721e49d DS |
1068 | |
1069 | /* | |
1070 | * Do not return the error but go back to waiting. There's a | |
1071 | * workspace preallocated for each type and the compression | |
1072 | * time is bounded so we get to a workspace eventually. This | |
1073 | * makes our caller's life easier. | |
52356716 DS |
1074 | * |
1075 | * To prevent silent and low-probability deadlocks (when the | |
1076 | * initial preallocation fails), check if there are any | |
1077 | * workspaces at all. | |
e721e49d | 1078 | */ |
52356716 DS |
1079 | if (atomic_read(total_ws) == 0) { |
1080 | static DEFINE_RATELIMIT_STATE(_rs, | |
1081 | /* once per minute */ 60 * HZ, | |
1082 | /* no burst */ 1); | |
1083 | ||
1084 | if (__ratelimit(&_rs)) { | |
ab8d0fc4 | 1085 | pr_warn("BTRFS: no compression workspaces, low memory, retrying\n"); |
52356716 DS |
1086 | } |
1087 | } | |
e721e49d | 1088 | goto again; |
261507a0 LZ |
1089 | } |
1090 | return workspace; | |
1091 | } | |
1092 | ||
7bf49943 | 1093 | static struct list_head *get_workspace(int type, int level) |
929f4baf | 1094 | { |
6a0d1272 | 1095 | switch (type) { |
5907a9bb | 1096 | case BTRFS_COMPRESS_NONE: return btrfs_get_workspace(type, level); |
6a0d1272 | 1097 | case BTRFS_COMPRESS_ZLIB: return zlib_get_workspace(level); |
5907a9bb | 1098 | case BTRFS_COMPRESS_LZO: return btrfs_get_workspace(type, level); |
6a0d1272 DS |
1099 | case BTRFS_COMPRESS_ZSTD: return zstd_get_workspace(level); |
1100 | default: | |
1101 | /* | |
1102 | * This can't happen, the type is validated several times | |
1103 | * before we get here. | |
1104 | */ | |
1105 | BUG(); | |
1106 | } | |
929f4baf DZ |
1107 | } |
1108 | ||
261507a0 LZ |
1109 | /* |
1110 | * put a workspace struct back on the list or free it if we have enough | |
1111 | * idle ones sitting around | |
1112 | */ | |
a3bbd2a9 | 1113 | void btrfs_put_workspace(int type, struct list_head *ws) |
261507a0 | 1114 | { |
a3bbd2a9 | 1115 | struct workspace_manager *wsm; |
4e439a0b TT |
1116 | struct list_head *idle_ws; |
1117 | spinlock_t *ws_lock; | |
1118 | atomic_t *total_ws; | |
1119 | wait_queue_head_t *ws_wait; | |
1120 | int *free_ws; | |
1121 | ||
a3bbd2a9 | 1122 | wsm = btrfs_compress_op[type]->workspace_manager; |
92ee5530 DZ |
1123 | idle_ws = &wsm->idle_ws; |
1124 | ws_lock = &wsm->ws_lock; | |
1125 | total_ws = &wsm->total_ws; | |
1126 | ws_wait = &wsm->ws_wait; | |
1127 | free_ws = &wsm->free_ws; | |
d9187649 BL |
1128 | |
1129 | spin_lock(ws_lock); | |
26b28dce | 1130 | if (*free_ws <= num_online_cpus()) { |
929f4baf | 1131 | list_add(ws, idle_ws); |
6ac10a6a | 1132 | (*free_ws)++; |
d9187649 | 1133 | spin_unlock(ws_lock); |
261507a0 LZ |
1134 | goto wake; |
1135 | } | |
d9187649 | 1136 | spin_unlock(ws_lock); |
261507a0 | 1137 | |
1e002351 | 1138 | free_workspace(type, ws); |
6ac10a6a | 1139 | atomic_dec(total_ws); |
261507a0 | 1140 | wake: |
093258e6 | 1141 | cond_wake_up(ws_wait); |
261507a0 LZ |
1142 | } |
1143 | ||
929f4baf DZ |
1144 | static void put_workspace(int type, struct list_head *ws) |
1145 | { | |
bd3a5287 | 1146 | switch (type) { |
a3bbd2a9 DS |
1147 | case BTRFS_COMPRESS_NONE: return btrfs_put_workspace(type, ws); |
1148 | case BTRFS_COMPRESS_ZLIB: return btrfs_put_workspace(type, ws); | |
1149 | case BTRFS_COMPRESS_LZO: return btrfs_put_workspace(type, ws); | |
bd3a5287 DS |
1150 | case BTRFS_COMPRESS_ZSTD: return zstd_put_workspace(ws); |
1151 | default: | |
1152 | /* | |
1153 | * This can't happen, the type is validated several times | |
1154 | * before we get here. | |
1155 | */ | |
1156 | BUG(); | |
1157 | } | |
929f4baf DZ |
1158 | } |
1159 | ||
adbab642 AJ |
1160 | /* |
1161 | * Adjust @level according to the limits of the compression algorithm or | |
1162 | * fallback to default | |
1163 | */ | |
1164 | static unsigned int btrfs_compress_set_level(int type, unsigned level) | |
1165 | { | |
1166 | const struct btrfs_compress_op *ops = btrfs_compress_op[type]; | |
1167 | ||
1168 | if (level == 0) | |
1169 | level = ops->default_level; | |
1170 | else | |
1171 | level = min(level, ops->max_level); | |
1172 | ||
1173 | return level; | |
1174 | } | |
1175 | ||
261507a0 | 1176 | /* |
38c31464 DS |
1177 | * Given an address space and start and length, compress the bytes into @pages |
1178 | * that are allocated on demand. | |
261507a0 | 1179 | * |
f51d2b59 DS |
1180 | * @type_level is encoded algorithm and level, where level 0 means whatever |
1181 | * default the algorithm chooses and is opaque here; | |
1182 | * - compression algo are 0-3 | |
1183 | * - the level are bits 4-7 | |
1184 | * | |
4d3a800e DS |
1185 | * @out_pages is an in/out parameter, holds maximum number of pages to allocate |
1186 | * and returns number of actually allocated pages | |
261507a0 | 1187 | * |
38c31464 DS |
1188 | * @total_in is used to return the number of bytes actually read. It |
1189 | * may be smaller than the input length if we had to exit early because we | |
261507a0 LZ |
1190 | * ran out of room in the pages array or because we cross the |
1191 | * max_out threshold. | |
1192 | * | |
38c31464 DS |
1193 | * @total_out is an in/out parameter, must be set to the input length and will |
1194 | * be also used to return the total number of compressed bytes | |
261507a0 | 1195 | */ |
f51d2b59 | 1196 | int btrfs_compress_pages(unsigned int type_level, struct address_space *mapping, |
38c31464 | 1197 | u64 start, struct page **pages, |
261507a0 LZ |
1198 | unsigned long *out_pages, |
1199 | unsigned long *total_in, | |
e5d74902 | 1200 | unsigned long *total_out) |
261507a0 | 1201 | { |
1972708a | 1202 | int type = btrfs_compress_type(type_level); |
7bf49943 | 1203 | int level = btrfs_compress_level(type_level); |
261507a0 LZ |
1204 | struct list_head *workspace; |
1205 | int ret; | |
1206 | ||
b0c1fe1e | 1207 | level = btrfs_compress_set_level(type, level); |
7bf49943 | 1208 | workspace = get_workspace(type, level); |
1e4eb746 DS |
1209 | ret = compression_compress_pages(type, workspace, mapping, start, pages, |
1210 | out_pages, total_in, total_out); | |
929f4baf | 1211 | put_workspace(type, workspace); |
261507a0 LZ |
1212 | return ret; |
1213 | } | |
1214 | ||
8140dc30 | 1215 | static int btrfs_decompress_bio(struct compressed_bio *cb) |
261507a0 LZ |
1216 | { |
1217 | struct list_head *workspace; | |
1218 | int ret; | |
8140dc30 | 1219 | int type = cb->compress_type; |
261507a0 | 1220 | |
7bf49943 | 1221 | workspace = get_workspace(type, 0); |
4a9e803e | 1222 | ret = compression_decompress_bio(workspace, cb); |
929f4baf | 1223 | put_workspace(type, workspace); |
e1ddce71 | 1224 | |
261507a0 LZ |
1225 | return ret; |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * a less complex decompression routine. Our compressed data fits in a | |
1230 | * single page, and we want to read a single page out of it. | |
1231 | * start_byte tells us the offset into the compressed data we're interested in | |
1232 | */ | |
1233 | int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page, | |
1234 | unsigned long start_byte, size_t srclen, size_t destlen) | |
1235 | { | |
1236 | struct list_head *workspace; | |
1237 | int ret; | |
1238 | ||
7bf49943 | 1239 | workspace = get_workspace(type, 0); |
1e4eb746 DS |
1240 | ret = compression_decompress(type, workspace, data_in, dest_page, |
1241 | start_byte, srclen, destlen); | |
929f4baf | 1242 | put_workspace(type, workspace); |
7bf49943 | 1243 | |
261507a0 LZ |
1244 | return ret; |
1245 | } | |
1246 | ||
5565b8e0 | 1247 | int __init btrfs_init_compress(void) |
1666edab | 1248 | { |
d5517033 DS |
1249 | btrfs_init_workspace_manager(BTRFS_COMPRESS_NONE); |
1250 | btrfs_init_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1251 | btrfs_init_workspace_manager(BTRFS_COMPRESS_LZO); | |
1252 | zstd_init_workspace_manager(); | |
5565b8e0 | 1253 | return 0; |
1666edab DZ |
1254 | } |
1255 | ||
e67c718b | 1256 | void __cold btrfs_exit_compress(void) |
261507a0 | 1257 | { |
2510307e DS |
1258 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_NONE); |
1259 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_ZLIB); | |
1260 | btrfs_cleanup_workspace_manager(BTRFS_COMPRESS_LZO); | |
1261 | zstd_cleanup_workspace_manager(); | |
261507a0 | 1262 | } |
3a39c18d LZ |
1263 | |
1264 | /* | |
1c3dc173 | 1265 | * Copy decompressed data from working buffer to pages. |
3a39c18d | 1266 | * |
1c3dc173 QW |
1267 | * @buf: The decompressed data buffer |
1268 | * @buf_len: The decompressed data length | |
1269 | * @decompressed: Number of bytes that are already decompressed inside the | |
1270 | * compressed extent | |
1271 | * @cb: The compressed extent descriptor | |
1272 | * @orig_bio: The original bio that the caller wants to read for | |
3a39c18d | 1273 | * |
1c3dc173 QW |
1274 | * An easier to understand graph is like below: |
1275 | * | |
1276 | * |<- orig_bio ->| |<- orig_bio->| | |
1277 | * |<------- full decompressed extent ----->| | |
1278 | * |<----------- @cb range ---->| | |
1279 | * | |<-- @buf_len -->| | |
1280 | * |<--- @decompressed --->| | |
1281 | * | |
1282 | * Note that, @cb can be a subpage of the full decompressed extent, but | |
1283 | * @cb->start always has the same as the orig_file_offset value of the full | |
1284 | * decompressed extent. | |
1285 | * | |
1286 | * When reading compressed extent, we have to read the full compressed extent, | |
1287 | * while @orig_bio may only want part of the range. | |
1288 | * Thus this function will ensure only data covered by @orig_bio will be copied | |
1289 | * to. | |
1290 | * | |
1291 | * Return 0 if we have copied all needed contents for @orig_bio. | |
1292 | * Return >0 if we need continue decompress. | |
3a39c18d | 1293 | */ |
1c3dc173 QW |
1294 | int btrfs_decompress_buf2page(const char *buf, u32 buf_len, |
1295 | struct compressed_bio *cb, u32 decompressed) | |
3a39c18d | 1296 | { |
1c3dc173 QW |
1297 | struct bio *orig_bio = cb->orig_bio; |
1298 | /* Offset inside the full decompressed extent */ | |
1299 | u32 cur_offset; | |
1300 | ||
1301 | cur_offset = decompressed; | |
1302 | /* The main loop to do the copy */ | |
1303 | while (cur_offset < decompressed + buf_len) { | |
1304 | struct bio_vec bvec; | |
1305 | size_t copy_len; | |
1306 | u32 copy_start; | |
1307 | /* Offset inside the full decompressed extent */ | |
1308 | u32 bvec_offset; | |
1309 | ||
1310 | bvec = bio_iter_iovec(orig_bio, orig_bio->bi_iter); | |
1311 | /* | |
1312 | * cb->start may underflow, but subtracting that value can still | |
1313 | * give us correct offset inside the full decompressed extent. | |
1314 | */ | |
1315 | bvec_offset = page_offset(bvec.bv_page) + bvec.bv_offset - cb->start; | |
974b1adc | 1316 | |
1c3dc173 QW |
1317 | /* Haven't reached the bvec range, exit */ |
1318 | if (decompressed + buf_len <= bvec_offset) | |
1319 | return 1; | |
3a39c18d | 1320 | |
1c3dc173 QW |
1321 | copy_start = max(cur_offset, bvec_offset); |
1322 | copy_len = min(bvec_offset + bvec.bv_len, | |
1323 | decompressed + buf_len) - copy_start; | |
1324 | ASSERT(copy_len); | |
3a39c18d | 1325 | |
974b1adc | 1326 | /* |
1c3dc173 QW |
1327 | * Extra range check to ensure we didn't go beyond |
1328 | * @buf + @buf_len. | |
974b1adc | 1329 | */ |
1c3dc173 QW |
1330 | ASSERT(copy_start - decompressed < buf_len); |
1331 | memcpy_to_page(bvec.bv_page, bvec.bv_offset, | |
1332 | buf + copy_start - decompressed, copy_len); | |
1c3dc173 | 1333 | cur_offset += copy_len; |
3a39c18d | 1334 | |
1c3dc173 QW |
1335 | bio_advance(orig_bio, copy_len); |
1336 | /* Finished the bio */ | |
1337 | if (!orig_bio->bi_iter.bi_size) | |
1338 | return 0; | |
3a39c18d | 1339 | } |
3a39c18d LZ |
1340 | return 1; |
1341 | } | |
c2fcdcdf | 1342 | |
19562430 TT |
1343 | /* |
1344 | * Shannon Entropy calculation | |
1345 | * | |
52042d8e | 1346 | * Pure byte distribution analysis fails to determine compressibility of data. |
19562430 TT |
1347 | * Try calculating entropy to estimate the average minimum number of bits |
1348 | * needed to encode the sampled data. | |
1349 | * | |
1350 | * For convenience, return the percentage of needed bits, instead of amount of | |
1351 | * bits directly. | |
1352 | * | |
1353 | * @ENTROPY_LVL_ACEPTABLE - below that threshold, sample has low byte entropy | |
1354 | * and can be compressible with high probability | |
1355 | * | |
1356 | * @ENTROPY_LVL_HIGH - data are not compressible with high probability | |
1357 | * | |
1358 | * Use of ilog2() decreases precision, we lower the LVL to 5 to compensate. | |
1359 | */ | |
1360 | #define ENTROPY_LVL_ACEPTABLE (65) | |
1361 | #define ENTROPY_LVL_HIGH (80) | |
1362 | ||
1363 | /* | |
1364 | * For increasead precision in shannon_entropy calculation, | |
1365 | * let's do pow(n, M) to save more digits after comma: | |
1366 | * | |
1367 | * - maximum int bit length is 64 | |
1368 | * - ilog2(MAX_SAMPLE_SIZE) -> 13 | |
1369 | * - 13 * 4 = 52 < 64 -> M = 4 | |
1370 | * | |
1371 | * So use pow(n, 4). | |
1372 | */ | |
1373 | static inline u32 ilog2_w(u64 n) | |
1374 | { | |
1375 | return ilog2(n * n * n * n); | |
1376 | } | |
1377 | ||
1378 | static u32 shannon_entropy(struct heuristic_ws *ws) | |
1379 | { | |
1380 | const u32 entropy_max = 8 * ilog2_w(2); | |
1381 | u32 entropy_sum = 0; | |
1382 | u32 p, p_base, sz_base; | |
1383 | u32 i; | |
1384 | ||
1385 | sz_base = ilog2_w(ws->sample_size); | |
1386 | for (i = 0; i < BUCKET_SIZE && ws->bucket[i].count > 0; i++) { | |
1387 | p = ws->bucket[i].count; | |
1388 | p_base = ilog2_w(p); | |
1389 | entropy_sum += p * (sz_base - p_base); | |
1390 | } | |
1391 | ||
1392 | entropy_sum /= ws->sample_size; | |
1393 | return entropy_sum * 100 / entropy_max; | |
1394 | } | |
1395 | ||
440c840c TT |
1396 | #define RADIX_BASE 4U |
1397 | #define COUNTERS_SIZE (1U << RADIX_BASE) | |
1398 | ||
1399 | static u8 get4bits(u64 num, int shift) { | |
1400 | u8 low4bits; | |
1401 | ||
1402 | num >>= shift; | |
1403 | /* Reverse order */ | |
1404 | low4bits = (COUNTERS_SIZE - 1) - (num % COUNTERS_SIZE); | |
1405 | return low4bits; | |
1406 | } | |
1407 | ||
440c840c TT |
1408 | /* |
1409 | * Use 4 bits as radix base | |
52042d8e | 1410 | * Use 16 u32 counters for calculating new position in buf array |
440c840c TT |
1411 | * |
1412 | * @array - array that will be sorted | |
1413 | * @array_buf - buffer array to store sorting results | |
1414 | * must be equal in size to @array | |
1415 | * @num - array size | |
440c840c | 1416 | */ |
23ae8c63 | 1417 | static void radix_sort(struct bucket_item *array, struct bucket_item *array_buf, |
36243c91 | 1418 | int num) |
858177d3 | 1419 | { |
440c840c TT |
1420 | u64 max_num; |
1421 | u64 buf_num; | |
1422 | u32 counters[COUNTERS_SIZE]; | |
1423 | u32 new_addr; | |
1424 | u32 addr; | |
1425 | int bitlen; | |
1426 | int shift; | |
1427 | int i; | |
858177d3 | 1428 | |
440c840c TT |
1429 | /* |
1430 | * Try avoid useless loop iterations for small numbers stored in big | |
1431 | * counters. Example: 48 33 4 ... in 64bit array | |
1432 | */ | |
23ae8c63 | 1433 | max_num = array[0].count; |
440c840c | 1434 | for (i = 1; i < num; i++) { |
23ae8c63 | 1435 | buf_num = array[i].count; |
440c840c TT |
1436 | if (buf_num > max_num) |
1437 | max_num = buf_num; | |
1438 | } | |
1439 | ||
1440 | buf_num = ilog2(max_num); | |
1441 | bitlen = ALIGN(buf_num, RADIX_BASE * 2); | |
1442 | ||
1443 | shift = 0; | |
1444 | while (shift < bitlen) { | |
1445 | memset(counters, 0, sizeof(counters)); | |
1446 | ||
1447 | for (i = 0; i < num; i++) { | |
23ae8c63 | 1448 | buf_num = array[i].count; |
440c840c TT |
1449 | addr = get4bits(buf_num, shift); |
1450 | counters[addr]++; | |
1451 | } | |
1452 | ||
1453 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1454 | counters[i] += counters[i - 1]; | |
1455 | ||
1456 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1457 | buf_num = array[i].count; |
440c840c TT |
1458 | addr = get4bits(buf_num, shift); |
1459 | counters[addr]--; | |
1460 | new_addr = counters[addr]; | |
7add17be | 1461 | array_buf[new_addr] = array[i]; |
440c840c TT |
1462 | } |
1463 | ||
1464 | shift += RADIX_BASE; | |
1465 | ||
1466 | /* | |
1467 | * Normal radix expects to move data from a temporary array, to | |
1468 | * the main one. But that requires some CPU time. Avoid that | |
1469 | * by doing another sort iteration to original array instead of | |
1470 | * memcpy() | |
1471 | */ | |
1472 | memset(counters, 0, sizeof(counters)); | |
1473 | ||
1474 | for (i = 0; i < num; i ++) { | |
23ae8c63 | 1475 | buf_num = array_buf[i].count; |
440c840c TT |
1476 | addr = get4bits(buf_num, shift); |
1477 | counters[addr]++; | |
1478 | } | |
1479 | ||
1480 | for (i = 1; i < COUNTERS_SIZE; i++) | |
1481 | counters[i] += counters[i - 1]; | |
1482 | ||
1483 | for (i = num - 1; i >= 0; i--) { | |
23ae8c63 | 1484 | buf_num = array_buf[i].count; |
440c840c TT |
1485 | addr = get4bits(buf_num, shift); |
1486 | counters[addr]--; | |
1487 | new_addr = counters[addr]; | |
7add17be | 1488 | array[new_addr] = array_buf[i]; |
440c840c TT |
1489 | } |
1490 | ||
1491 | shift += RADIX_BASE; | |
1492 | } | |
858177d3 TT |
1493 | } |
1494 | ||
1495 | /* | |
1496 | * Size of the core byte set - how many bytes cover 90% of the sample | |
1497 | * | |
1498 | * There are several types of structured binary data that use nearly all byte | |
1499 | * values. The distribution can be uniform and counts in all buckets will be | |
1500 | * nearly the same (eg. encrypted data). Unlikely to be compressible. | |
1501 | * | |
1502 | * Other possibility is normal (Gaussian) distribution, where the data could | |
1503 | * be potentially compressible, but we have to take a few more steps to decide | |
1504 | * how much. | |
1505 | * | |
1506 | * @BYTE_CORE_SET_LOW - main part of byte values repeated frequently, | |
1507 | * compression algo can easy fix that | |
1508 | * @BYTE_CORE_SET_HIGH - data have uniform distribution and with high | |
1509 | * probability is not compressible | |
1510 | */ | |
1511 | #define BYTE_CORE_SET_LOW (64) | |
1512 | #define BYTE_CORE_SET_HIGH (200) | |
1513 | ||
1514 | static int byte_core_set_size(struct heuristic_ws *ws) | |
1515 | { | |
1516 | u32 i; | |
1517 | u32 coreset_sum = 0; | |
1518 | const u32 core_set_threshold = ws->sample_size * 90 / 100; | |
1519 | struct bucket_item *bucket = ws->bucket; | |
1520 | ||
1521 | /* Sort in reverse order */ | |
36243c91 | 1522 | radix_sort(ws->bucket, ws->bucket_b, BUCKET_SIZE); |
858177d3 TT |
1523 | |
1524 | for (i = 0; i < BYTE_CORE_SET_LOW; i++) | |
1525 | coreset_sum += bucket[i].count; | |
1526 | ||
1527 | if (coreset_sum > core_set_threshold) | |
1528 | return i; | |
1529 | ||
1530 | for (; i < BYTE_CORE_SET_HIGH && bucket[i].count > 0; i++) { | |
1531 | coreset_sum += bucket[i].count; | |
1532 | if (coreset_sum > core_set_threshold) | |
1533 | break; | |
1534 | } | |
1535 | ||
1536 | return i; | |
1537 | } | |
1538 | ||
a288e92c TT |
1539 | /* |
1540 | * Count byte values in buckets. | |
1541 | * This heuristic can detect textual data (configs, xml, json, html, etc). | |
1542 | * Because in most text-like data byte set is restricted to limited number of | |
1543 | * possible characters, and that restriction in most cases makes data easy to | |
1544 | * compress. | |
1545 | * | |
1546 | * @BYTE_SET_THRESHOLD - consider all data within this byte set size: | |
1547 | * less - compressible | |
1548 | * more - need additional analysis | |
1549 | */ | |
1550 | #define BYTE_SET_THRESHOLD (64) | |
1551 | ||
1552 | static u32 byte_set_size(const struct heuristic_ws *ws) | |
1553 | { | |
1554 | u32 i; | |
1555 | u32 byte_set_size = 0; | |
1556 | ||
1557 | for (i = 0; i < BYTE_SET_THRESHOLD; i++) { | |
1558 | if (ws->bucket[i].count > 0) | |
1559 | byte_set_size++; | |
1560 | } | |
1561 | ||
1562 | /* | |
1563 | * Continue collecting count of byte values in buckets. If the byte | |
1564 | * set size is bigger then the threshold, it's pointless to continue, | |
1565 | * the detection technique would fail for this type of data. | |
1566 | */ | |
1567 | for (; i < BUCKET_SIZE; i++) { | |
1568 | if (ws->bucket[i].count > 0) { | |
1569 | byte_set_size++; | |
1570 | if (byte_set_size > BYTE_SET_THRESHOLD) | |
1571 | return byte_set_size; | |
1572 | } | |
1573 | } | |
1574 | ||
1575 | return byte_set_size; | |
1576 | } | |
1577 | ||
1fe4f6fa TT |
1578 | static bool sample_repeated_patterns(struct heuristic_ws *ws) |
1579 | { | |
1580 | const u32 half_of_sample = ws->sample_size / 2; | |
1581 | const u8 *data = ws->sample; | |
1582 | ||
1583 | return memcmp(&data[0], &data[half_of_sample], half_of_sample) == 0; | |
1584 | } | |
1585 | ||
a440d48c TT |
1586 | static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end, |
1587 | struct heuristic_ws *ws) | |
1588 | { | |
1589 | struct page *page; | |
1590 | u64 index, index_end; | |
1591 | u32 i, curr_sample_pos; | |
1592 | u8 *in_data; | |
1593 | ||
1594 | /* | |
1595 | * Compression handles the input data by chunks of 128KiB | |
1596 | * (defined by BTRFS_MAX_UNCOMPRESSED) | |
1597 | * | |
1598 | * We do the same for the heuristic and loop over the whole range. | |
1599 | * | |
1600 | * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will | |
1601 | * process no more than BTRFS_MAX_UNCOMPRESSED at a time. | |
1602 | */ | |
1603 | if (end - start > BTRFS_MAX_UNCOMPRESSED) | |
1604 | end = start + BTRFS_MAX_UNCOMPRESSED; | |
1605 | ||
1606 | index = start >> PAGE_SHIFT; | |
1607 | index_end = end >> PAGE_SHIFT; | |
1608 | ||
1609 | /* Don't miss unaligned end */ | |
1610 | if (!IS_ALIGNED(end, PAGE_SIZE)) | |
1611 | index_end++; | |
1612 | ||
1613 | curr_sample_pos = 0; | |
1614 | while (index < index_end) { | |
1615 | page = find_get_page(inode->i_mapping, index); | |
58c1a35c | 1616 | in_data = kmap_local_page(page); |
a440d48c TT |
1617 | /* Handle case where the start is not aligned to PAGE_SIZE */ |
1618 | i = start % PAGE_SIZE; | |
1619 | while (i < PAGE_SIZE - SAMPLING_READ_SIZE) { | |
1620 | /* Don't sample any garbage from the last page */ | |
1621 | if (start > end - SAMPLING_READ_SIZE) | |
1622 | break; | |
1623 | memcpy(&ws->sample[curr_sample_pos], &in_data[i], | |
1624 | SAMPLING_READ_SIZE); | |
1625 | i += SAMPLING_INTERVAL; | |
1626 | start += SAMPLING_INTERVAL; | |
1627 | curr_sample_pos += SAMPLING_READ_SIZE; | |
1628 | } | |
58c1a35c | 1629 | kunmap_local(in_data); |
a440d48c TT |
1630 | put_page(page); |
1631 | ||
1632 | index++; | |
1633 | } | |
1634 | ||
1635 | ws->sample_size = curr_sample_pos; | |
1636 | } | |
1637 | ||
c2fcdcdf TT |
1638 | /* |
1639 | * Compression heuristic. | |
1640 | * | |
1641 | * For now is's a naive and optimistic 'return true', we'll extend the logic to | |
1642 | * quickly (compared to direct compression) detect data characteristics | |
1643 | * (compressible/uncompressible) to avoid wasting CPU time on uncompressible | |
1644 | * data. | |
1645 | * | |
1646 | * The following types of analysis can be performed: | |
1647 | * - detect mostly zero data | |
1648 | * - detect data with low "byte set" size (text, etc) | |
1649 | * - detect data with low/high "core byte" set | |
1650 | * | |
1651 | * Return non-zero if the compression should be done, 0 otherwise. | |
1652 | */ | |
1653 | int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end) | |
1654 | { | |
7bf49943 | 1655 | struct list_head *ws_list = get_workspace(0, 0); |
4e439a0b | 1656 | struct heuristic_ws *ws; |
a440d48c TT |
1657 | u32 i; |
1658 | u8 byte; | |
19562430 | 1659 | int ret = 0; |
c2fcdcdf | 1660 | |
4e439a0b TT |
1661 | ws = list_entry(ws_list, struct heuristic_ws, list); |
1662 | ||
a440d48c TT |
1663 | heuristic_collect_sample(inode, start, end, ws); |
1664 | ||
1fe4f6fa TT |
1665 | if (sample_repeated_patterns(ws)) { |
1666 | ret = 1; | |
1667 | goto out; | |
1668 | } | |
1669 | ||
a440d48c TT |
1670 | memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE); |
1671 | ||
1672 | for (i = 0; i < ws->sample_size; i++) { | |
1673 | byte = ws->sample[i]; | |
1674 | ws->bucket[byte].count++; | |
c2fcdcdf TT |
1675 | } |
1676 | ||
a288e92c TT |
1677 | i = byte_set_size(ws); |
1678 | if (i < BYTE_SET_THRESHOLD) { | |
1679 | ret = 2; | |
1680 | goto out; | |
1681 | } | |
1682 | ||
858177d3 TT |
1683 | i = byte_core_set_size(ws); |
1684 | if (i <= BYTE_CORE_SET_LOW) { | |
1685 | ret = 3; | |
1686 | goto out; | |
1687 | } | |
1688 | ||
1689 | if (i >= BYTE_CORE_SET_HIGH) { | |
1690 | ret = 0; | |
1691 | goto out; | |
1692 | } | |
1693 | ||
19562430 TT |
1694 | i = shannon_entropy(ws); |
1695 | if (i <= ENTROPY_LVL_ACEPTABLE) { | |
1696 | ret = 4; | |
1697 | goto out; | |
1698 | } | |
1699 | ||
1700 | /* | |
1701 | * For the levels below ENTROPY_LVL_HIGH, additional analysis would be | |
1702 | * needed to give green light to compression. | |
1703 | * | |
1704 | * For now just assume that compression at that level is not worth the | |
1705 | * resources because: | |
1706 | * | |
1707 | * 1. it is possible to defrag the data later | |
1708 | * | |
1709 | * 2. the data would turn out to be hardly compressible, eg. 150 byte | |
1710 | * values, every bucket has counter at level ~54. The heuristic would | |
1711 | * be confused. This can happen when data have some internal repeated | |
1712 | * patterns like "abbacbbc...". This can be detected by analyzing | |
1713 | * pairs of bytes, which is too costly. | |
1714 | */ | |
1715 | if (i < ENTROPY_LVL_HIGH) { | |
1716 | ret = 5; | |
1717 | goto out; | |
1718 | } else { | |
1719 | ret = 0; | |
1720 | goto out; | |
1721 | } | |
1722 | ||
1fe4f6fa | 1723 | out: |
929f4baf | 1724 | put_workspace(0, ws_list); |
c2fcdcdf TT |
1725 | return ret; |
1726 | } | |
f51d2b59 | 1727 | |
d0ab62ce DZ |
1728 | /* |
1729 | * Convert the compression suffix (eg. after "zlib" starting with ":") to | |
1730 | * level, unrecognized string will set the default level | |
1731 | */ | |
1732 | unsigned int btrfs_compress_str2level(unsigned int type, const char *str) | |
f51d2b59 | 1733 | { |
d0ab62ce DZ |
1734 | unsigned int level = 0; |
1735 | int ret; | |
1736 | ||
1737 | if (!type) | |
f51d2b59 DS |
1738 | return 0; |
1739 | ||
d0ab62ce DZ |
1740 | if (str[0] == ':') { |
1741 | ret = kstrtouint(str + 1, 10, &level); | |
1742 | if (ret) | |
1743 | level = 0; | |
1744 | } | |
1745 | ||
b0c1fe1e DS |
1746 | level = btrfs_compress_set_level(type, level); |
1747 | ||
1748 | return level; | |
1749 | } |