~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/btrfs/compression.c

Version: ~ [ linux-5.5-rc7 ] ~ [ linux-5.4.13 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.97 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.166 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.210 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.210 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.81 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp