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