]> git.karo-electronics.de Git - karo-tx-linux.git/blob - fs/btrfs/compression.c
Merge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[karo-tx-linux.git] / fs / btrfs / compression.c
1 /*
2  * Copyright (C) 2008 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/kernel.h>
20 #include <linux/bio.h>
21 #include <linux/buffer_head.h>
22 #include <linux/file.h>
23 #include <linux/fs.h>
24 #include <linux/pagemap.h>
25 #include <linux/highmem.h>
26 #include <linux/time.h>
27 #include <linux/init.h>
28 #include <linux/string.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mpage.h>
31 #include <linux/swap.h>
32 #include <linux/writeback.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/slab.h>
35 #include <linux/sched/mm.h>
36 #include "ctree.h"
37 #include "disk-io.h"
38 #include "transaction.h"
39 #include "btrfs_inode.h"
40 #include "volumes.h"
41 #include "ordered-data.h"
42 #include "compression.h"
43 #include "extent_io.h"
44 #include "extent_map.h"
45
46 static int btrfs_decompress_bio(struct compressed_bio *cb);
47
48 static inline int compressed_bio_size(struct btrfs_fs_info *fs_info,
49                                       unsigned long disk_size)
50 {
51         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
52
53         return sizeof(struct compressed_bio) +
54                 (DIV_ROUND_UP(disk_size, fs_info->sectorsize)) * csum_size;
55 }
56
57 static int check_compressed_csum(struct btrfs_inode *inode,
58                                  struct compressed_bio *cb,
59                                  u64 disk_start)
60 {
61         int ret;
62         struct page *page;
63         unsigned long i;
64         char *kaddr;
65         u32 csum;
66         u32 *cb_sum = &cb->sums;
67
68         if (inode->flags & BTRFS_INODE_NODATASUM)
69                 return 0;
70
71         for (i = 0; i < cb->nr_pages; i++) {
72                 page = cb->compressed_pages[i];
73                 csum = ~(u32)0;
74
75                 kaddr = kmap_atomic(page);
76                 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
77                 btrfs_csum_final(csum, (u8 *)&csum);
78                 kunmap_atomic(kaddr);
79
80                 if (csum != *cb_sum) {
81                         btrfs_print_data_csum_error(inode, disk_start, csum,
82                                         *cb_sum, cb->mirror_num);
83                         ret = -EIO;
84                         goto fail;
85                 }
86                 cb_sum++;
87
88         }
89         ret = 0;
90 fail:
91         return ret;
92 }
93
94 /* when we finish reading compressed pages from the disk, we
95  * decompress them and then run the bio end_io routines on the
96  * decompressed pages (in the inode address space).
97  *
98  * This allows the checksumming and other IO error handling routines
99  * to work normally
100  *
101  * The compressed pages are freed here, and it must be run
102  * in process context
103  */
104 static void end_compressed_bio_read(struct bio *bio)
105 {
106         struct compressed_bio *cb = bio->bi_private;
107         struct inode *inode;
108         struct page *page;
109         unsigned long index;
110         int ret;
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         inode = cb->inode;
122         ret = check_compressed_csum(BTRFS_I(inode), cb,
123                                     (u64)bio->bi_iter.bi_sector << 9);
124         if (ret)
125                 goto csum_failed;
126
127         /* ok, we're the last bio for this extent, lets start
128          * the decompression.
129          */
130         ret = btrfs_decompress_bio(cb);
131
132 csum_failed:
133         if (ret)
134                 cb->errors = 1;
135
136         /* release the compressed pages */
137         index = 0;
138         for (index = 0; index < cb->nr_pages; index++) {
139                 page = cb->compressed_pages[index];
140                 page->mapping = NULL;
141                 put_page(page);
142         }
143
144         /* do io completion on the original bio */
145         if (cb->errors) {
146                 bio_io_error(cb->orig_bio);
147         } else {
148                 int i;
149                 struct bio_vec *bvec;
150
151                 /*
152                  * we have verified the checksum already, set page
153                  * checked so the end_io handlers know about it
154                  */
155                 ASSERT(!bio_flagged(bio, BIO_CLONED));
156                 bio_for_each_segment_all(bvec, cb->orig_bio, i)
157                         SetPageChecked(bvec->bv_page);
158
159                 bio_endio(cb->orig_bio);
160         }
161
162         /* finally free the cb struct */
163         kfree(cb->compressed_pages);
164         kfree(cb);
165 out:
166         bio_put(bio);
167 }
168
169 /*
170  * Clear the writeback bits on all of the file
171  * pages for a compressed write
172  */
173 static noinline void end_compressed_writeback(struct inode *inode,
174                                               const struct compressed_bio *cb)
175 {
176         unsigned long index = cb->start >> PAGE_SHIFT;
177         unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
178         struct page *pages[16];
179         unsigned long nr_pages = end_index - index + 1;
180         int i;
181         int ret;
182
183         if (cb->errors)
184                 mapping_set_error(inode->i_mapping, -EIO);
185
186         while (nr_pages > 0) {
187                 ret = find_get_pages_contig(inode->i_mapping, index,
188                                      min_t(unsigned long,
189                                      nr_pages, ARRAY_SIZE(pages)), pages);
190                 if (ret == 0) {
191                         nr_pages -= 1;
192                         index += 1;
193                         continue;
194                 }
195                 for (i = 0; i < ret; i++) {
196                         if (cb->errors)
197                                 SetPageError(pages[i]);
198                         end_page_writeback(pages[i]);
199                         put_page(pages[i]);
200                 }
201                 nr_pages -= ret;
202                 index += ret;
203         }
204         /* the inode may be gone now */
205 }
206
207 /*
208  * do the cleanup once all the compressed pages hit the disk.
209  * This will clear writeback on the file pages and free the compressed
210  * pages.
211  *
212  * This also calls the writeback end hooks for the file pages so that
213  * metadata and checksums can be updated in the file.
214  */
215 static void end_compressed_bio_write(struct bio *bio)
216 {
217         struct extent_io_tree *tree;
218         struct compressed_bio *cb = bio->bi_private;
219         struct inode *inode;
220         struct page *page;
221         unsigned long index;
222
223         if (bio->bi_status)
224                 cb->errors = 1;
225
226         /* if there are more bios still pending for this compressed
227          * extent, just exit
228          */
229         if (!refcount_dec_and_test(&cb->pending_bios))
230                 goto out;
231
232         /* ok, we're the last bio for this extent, step one is to
233          * call back into the FS and do all the end_io operations
234          */
235         inode = cb->inode;
236         tree = &BTRFS_I(inode)->io_tree;
237         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
238         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
239                                          cb->start,
240                                          cb->start + cb->len - 1,
241                                          NULL,
242                                          bio->bi_status ? 0 : 1);
243         cb->compressed_pages[0]->mapping = NULL;
244
245         end_compressed_writeback(inode, cb);
246         /* note, our inode could be gone now */
247
248         /*
249          * release the compressed pages, these came from alloc_page and
250          * are not attached to the inode at all
251          */
252         index = 0;
253         for (index = 0; index < cb->nr_pages; index++) {
254                 page = cb->compressed_pages[index];
255                 page->mapping = NULL;
256                 put_page(page);
257         }
258
259         /* finally free the cb struct */
260         kfree(cb->compressed_pages);
261         kfree(cb);
262 out:
263         bio_put(bio);
264 }
265
266 /*
267  * worker function to build and submit bios for previously compressed pages.
268  * The corresponding pages in the inode should be marked for writeback
269  * and the compressed pages should have a reference on them for dropping
270  * when the IO is complete.
271  *
272  * This also checksums the file bytes and gets things ready for
273  * the end io hooks.
274  */
275 blk_status_t btrfs_submit_compressed_write(struct inode *inode, u64 start,
276                                  unsigned long len, u64 disk_start,
277                                  unsigned long compressed_len,
278                                  struct page **compressed_pages,
279                                  unsigned long nr_pages)
280 {
281         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
282         struct bio *bio = NULL;
283         struct compressed_bio *cb;
284         unsigned long bytes_left;
285         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
286         int pg_index = 0;
287         struct page *page;
288         u64 first_byte = disk_start;
289         struct block_device *bdev;
290         blk_status_t ret;
291         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
292
293         WARN_ON(start & ((u64)PAGE_SIZE - 1));
294         cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
295         if (!cb)
296                 return BLK_STS_RESOURCE;
297         refcount_set(&cb->pending_bios, 0);
298         cb->errors = 0;
299         cb->inode = inode;
300         cb->start = start;
301         cb->len = len;
302         cb->mirror_num = 0;
303         cb->compressed_pages = compressed_pages;
304         cb->compressed_len = compressed_len;
305         cb->orig_bio = NULL;
306         cb->nr_pages = nr_pages;
307
308         bdev = fs_info->fs_devices->latest_bdev;
309
310         bio = btrfs_bio_alloc(bdev, first_byte);
311         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
312         bio->bi_private = cb;
313         bio->bi_end_io = end_compressed_bio_write;
314         refcount_set(&cb->pending_bios, 1);
315
316         /* create and submit bios for the compressed pages */
317         bytes_left = compressed_len;
318         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
319                 int submit = 0;
320
321                 page = compressed_pages[pg_index];
322                 page->mapping = inode->i_mapping;
323                 if (bio->bi_iter.bi_size)
324                         submit = io_tree->ops->merge_bio_hook(page, 0,
325                                                            PAGE_SIZE,
326                                                            bio, 0);
327
328                 page->mapping = NULL;
329                 if (submit || bio_add_page(bio, page, PAGE_SIZE, 0) <
330                     PAGE_SIZE) {
331                         bio_get(bio);
332
333                         /*
334                          * inc the count before we submit the bio so
335                          * we know the end IO handler won't happen before
336                          * we inc the count.  Otherwise, the cb might get
337                          * freed before we're done setting it up
338                          */
339                         refcount_inc(&cb->pending_bios);
340                         ret = btrfs_bio_wq_end_io(fs_info, bio,
341                                                   BTRFS_WQ_ENDIO_DATA);
342                         BUG_ON(ret); /* -ENOMEM */
343
344                         if (!skip_sum) {
345                                 ret = btrfs_csum_one_bio(inode, bio, start, 1);
346                                 BUG_ON(ret); /* -ENOMEM */
347                         }
348
349                         ret = btrfs_map_bio(fs_info, bio, 0, 1);
350                         if (ret) {
351                                 bio->bi_status = ret;
352                                 bio_endio(bio);
353                         }
354
355                         bio_put(bio);
356
357                         bio = btrfs_bio_alloc(bdev, first_byte);
358                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
359                         bio->bi_private = cb;
360                         bio->bi_end_io = end_compressed_bio_write;
361                         bio_add_page(bio, page, PAGE_SIZE, 0);
362                 }
363                 if (bytes_left < PAGE_SIZE) {
364                         btrfs_info(fs_info,
365                                         "bytes left %lu compress len %lu nr %lu",
366                                bytes_left, cb->compressed_len, cb->nr_pages);
367                 }
368                 bytes_left -= PAGE_SIZE;
369                 first_byte += PAGE_SIZE;
370                 cond_resched();
371         }
372         bio_get(bio);
373
374         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
375         BUG_ON(ret); /* -ENOMEM */
376
377         if (!skip_sum) {
378                 ret = btrfs_csum_one_bio(inode, bio, start, 1);
379                 BUG_ON(ret); /* -ENOMEM */
380         }
381
382         ret = btrfs_map_bio(fs_info, bio, 0, 1);
383         if (ret) {
384                 bio->bi_status = ret;
385                 bio_endio(bio);
386         }
387
388         bio_put(bio);
389         return 0;
390 }
391
392 static u64 bio_end_offset(struct bio *bio)
393 {
394         struct bio_vec *last = &bio->bi_io_vec[bio->bi_vcnt - 1];
395
396         return page_offset(last->bv_page) + last->bv_len + last->bv_offset;
397 }
398
399 static noinline int add_ra_bio_pages(struct inode *inode,
400                                      u64 compressed_end,
401                                      struct compressed_bio *cb)
402 {
403         unsigned long end_index;
404         unsigned long pg_index;
405         u64 last_offset;
406         u64 isize = i_size_read(inode);
407         int ret;
408         struct page *page;
409         unsigned long nr_pages = 0;
410         struct extent_map *em;
411         struct address_space *mapping = inode->i_mapping;
412         struct extent_map_tree *em_tree;
413         struct extent_io_tree *tree;
414         u64 end;
415         int misses = 0;
416
417         last_offset = bio_end_offset(cb->orig_bio);
418         em_tree = &BTRFS_I(inode)->extent_tree;
419         tree = &BTRFS_I(inode)->io_tree;
420
421         if (isize == 0)
422                 return 0;
423
424         end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
425
426         while (last_offset < compressed_end) {
427                 pg_index = last_offset >> PAGE_SHIFT;
428
429                 if (pg_index > end_index)
430                         break;
431
432                 rcu_read_lock();
433                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
434                 rcu_read_unlock();
435                 if (page && !radix_tree_exceptional_entry(page)) {
436                         misses++;
437                         if (misses > 4)
438                                 break;
439                         goto next;
440                 }
441
442                 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
443                                                                  ~__GFP_FS));
444                 if (!page)
445                         break;
446
447                 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
448                         put_page(page);
449                         goto next;
450                 }
451
452                 end = last_offset + PAGE_SIZE - 1;
453                 /*
454                  * at this point, we have a locked page in the page cache
455                  * for these bytes in the file.  But, we have to make
456                  * sure they map to this compressed extent on disk.
457                  */
458                 set_page_extent_mapped(page);
459                 lock_extent(tree, last_offset, end);
460                 read_lock(&em_tree->lock);
461                 em = lookup_extent_mapping(em_tree, last_offset,
462                                            PAGE_SIZE);
463                 read_unlock(&em_tree->lock);
464
465                 if (!em || last_offset < em->start ||
466                     (last_offset + PAGE_SIZE > extent_map_end(em)) ||
467                     (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
468                         free_extent_map(em);
469                         unlock_extent(tree, last_offset, end);
470                         unlock_page(page);
471                         put_page(page);
472                         break;
473                 }
474                 free_extent_map(em);
475
476                 if (page->index == end_index) {
477                         char *userpage;
478                         size_t zero_offset = isize & (PAGE_SIZE - 1);
479
480                         if (zero_offset) {
481                                 int zeros;
482                                 zeros = PAGE_SIZE - zero_offset;
483                                 userpage = kmap_atomic(page);
484                                 memset(userpage + zero_offset, 0, zeros);
485                                 flush_dcache_page(page);
486                                 kunmap_atomic(userpage);
487                         }
488                 }
489
490                 ret = bio_add_page(cb->orig_bio, page,
491                                    PAGE_SIZE, 0);
492
493                 if (ret == PAGE_SIZE) {
494                         nr_pages++;
495                         put_page(page);
496                 } else {
497                         unlock_extent(tree, last_offset, end);
498                         unlock_page(page);
499                         put_page(page);
500                         break;
501                 }
502 next:
503                 last_offset += PAGE_SIZE;
504         }
505         return 0;
506 }
507
508 /*
509  * for a compressed read, the bio we get passed has all the inode pages
510  * in it.  We don't actually do IO on those pages but allocate new ones
511  * to hold the compressed pages on disk.
512  *
513  * bio->bi_iter.bi_sector points to the compressed extent on disk
514  * bio->bi_io_vec points to all of the inode pages
515  *
516  * After the compressed pages are read, we copy the bytes into the
517  * bio we were passed and then call the bio end_io calls
518  */
519 blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
520                                  int mirror_num, unsigned long bio_flags)
521 {
522         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
523         struct extent_io_tree *tree;
524         struct extent_map_tree *em_tree;
525         struct compressed_bio *cb;
526         unsigned long compressed_len;
527         unsigned long nr_pages;
528         unsigned long pg_index;
529         struct page *page;
530         struct block_device *bdev;
531         struct bio *comp_bio;
532         u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
533         u64 em_len;
534         u64 em_start;
535         struct extent_map *em;
536         blk_status_t ret = BLK_STS_RESOURCE;
537         int faili = 0;
538         u32 *sums;
539
540         tree = &BTRFS_I(inode)->io_tree;
541         em_tree = &BTRFS_I(inode)->extent_tree;
542
543         /* we need the actual starting offset of this extent in the file */
544         read_lock(&em_tree->lock);
545         em = lookup_extent_mapping(em_tree,
546                                    page_offset(bio->bi_io_vec->bv_page),
547                                    PAGE_SIZE);
548         read_unlock(&em_tree->lock);
549         if (!em)
550                 return BLK_STS_IOERR;
551
552         compressed_len = em->block_len;
553         cb = kmalloc(compressed_bio_size(fs_info, compressed_len), GFP_NOFS);
554         if (!cb)
555                 goto out;
556
557         refcount_set(&cb->pending_bios, 0);
558         cb->errors = 0;
559         cb->inode = inode;
560         cb->mirror_num = mirror_num;
561         sums = &cb->sums;
562
563         cb->start = em->orig_start;
564         em_len = em->len;
565         em_start = em->start;
566
567         free_extent_map(em);
568         em = NULL;
569
570         cb->len = bio->bi_iter.bi_size;
571         cb->compressed_len = compressed_len;
572         cb->compress_type = extent_compress_type(bio_flags);
573         cb->orig_bio = bio;
574
575         nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
576         cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
577                                        GFP_NOFS);
578         if (!cb->compressed_pages)
579                 goto fail1;
580
581         bdev = fs_info->fs_devices->latest_bdev;
582
583         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
584                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
585                                                               __GFP_HIGHMEM);
586                 if (!cb->compressed_pages[pg_index]) {
587                         faili = pg_index - 1;
588                         ret = BLK_STS_RESOURCE;
589                         goto fail2;
590                 }
591         }
592         faili = nr_pages - 1;
593         cb->nr_pages = nr_pages;
594
595         add_ra_bio_pages(inode, em_start + em_len, cb);
596
597         /* include any pages we added in add_ra-bio_pages */
598         cb->len = bio->bi_iter.bi_size;
599
600         comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
601         bio_set_op_attrs (comp_bio, REQ_OP_READ, 0);
602         comp_bio->bi_private = cb;
603         comp_bio->bi_end_io = end_compressed_bio_read;
604         refcount_set(&cb->pending_bios, 1);
605
606         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
607                 int submit = 0;
608
609                 page = cb->compressed_pages[pg_index];
610                 page->mapping = inode->i_mapping;
611                 page->index = em_start >> PAGE_SHIFT;
612
613                 if (comp_bio->bi_iter.bi_size)
614                         submit = tree->ops->merge_bio_hook(page, 0,
615                                                         PAGE_SIZE,
616                                                         comp_bio, 0);
617
618                 page->mapping = NULL;
619                 if (submit || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
620                     PAGE_SIZE) {
621                         bio_get(comp_bio);
622
623                         ret = btrfs_bio_wq_end_io(fs_info, comp_bio,
624                                                   BTRFS_WQ_ENDIO_DATA);
625                         BUG_ON(ret); /* -ENOMEM */
626
627                         /*
628                          * inc the count before we submit the bio so
629                          * we know the end IO handler won't happen before
630                          * we inc the count.  Otherwise, the cb might get
631                          * freed before we're done setting it up
632                          */
633                         refcount_inc(&cb->pending_bios);
634
635                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
636                                 ret = btrfs_lookup_bio_sums(inode, comp_bio,
637                                                             sums);
638                                 BUG_ON(ret); /* -ENOMEM */
639                         }
640                         sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
641                                              fs_info->sectorsize);
642
643                         ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
644                         if (ret) {
645                                 comp_bio->bi_status = ret;
646                                 bio_endio(comp_bio);
647                         }
648
649                         bio_put(comp_bio);
650
651                         comp_bio = btrfs_bio_alloc(bdev, cur_disk_byte);
652                         bio_set_op_attrs(comp_bio, REQ_OP_READ, 0);
653                         comp_bio->bi_private = cb;
654                         comp_bio->bi_end_io = end_compressed_bio_read;
655
656                         bio_add_page(comp_bio, page, PAGE_SIZE, 0);
657                 }
658                 cur_disk_byte += PAGE_SIZE;
659         }
660         bio_get(comp_bio);
661
662         ret = btrfs_bio_wq_end_io(fs_info, comp_bio, BTRFS_WQ_ENDIO_DATA);
663         BUG_ON(ret); /* -ENOMEM */
664
665         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
666                 ret = btrfs_lookup_bio_sums(inode, comp_bio, sums);
667                 BUG_ON(ret); /* -ENOMEM */
668         }
669
670         ret = btrfs_map_bio(fs_info, comp_bio, mirror_num, 0);
671         if (ret) {
672                 comp_bio->bi_status = ret;
673                 bio_endio(comp_bio);
674         }
675
676         bio_put(comp_bio);
677         return 0;
678
679 fail2:
680         while (faili >= 0) {
681                 __free_page(cb->compressed_pages[faili]);
682                 faili--;
683         }
684
685         kfree(cb->compressed_pages);
686 fail1:
687         kfree(cb);
688 out:
689         free_extent_map(em);
690         return ret;
691 }
692
693 static struct {
694         struct list_head idle_ws;
695         spinlock_t ws_lock;
696         /* Number of free workspaces */
697         int free_ws;
698         /* Total number of allocated workspaces */
699         atomic_t total_ws;
700         /* Waiters for a free workspace */
701         wait_queue_head_t ws_wait;
702 } btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
703
704 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
705         &btrfs_zlib_compress,
706         &btrfs_lzo_compress,
707 };
708
709 void __init btrfs_init_compress(void)
710 {
711         int i;
712
713         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
714                 struct list_head *workspace;
715
716                 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
717                 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
718                 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
719                 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
720
721                 /*
722                  * Preallocate one workspace for each compression type so
723                  * we can guarantee forward progress in the worst case
724                  */
725                 workspace = btrfs_compress_op[i]->alloc_workspace();
726                 if (IS_ERR(workspace)) {
727                         pr_warn("BTRFS: cannot preallocate compression workspace, will try later\n");
728                 } else {
729                         atomic_set(&btrfs_comp_ws[i].total_ws, 1);
730                         btrfs_comp_ws[i].free_ws = 1;
731                         list_add(workspace, &btrfs_comp_ws[i].idle_ws);
732                 }
733         }
734 }
735
736 /*
737  * This finds an available workspace or allocates a new one.
738  * If it's not possible to allocate a new one, waits until there's one.
739  * Preallocation makes a forward progress guarantees and we do not return
740  * errors.
741  */
742 static struct list_head *find_workspace(int type)
743 {
744         struct list_head *workspace;
745         int cpus = num_online_cpus();
746         int idx = type - 1;
747         unsigned nofs_flag;
748
749         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
750         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
751         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
752         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
753         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
754 again:
755         spin_lock(ws_lock);
756         if (!list_empty(idle_ws)) {
757                 workspace = idle_ws->next;
758                 list_del(workspace);
759                 (*free_ws)--;
760                 spin_unlock(ws_lock);
761                 return workspace;
762
763         }
764         if (atomic_read(total_ws) > cpus) {
765                 DEFINE_WAIT(wait);
766
767                 spin_unlock(ws_lock);
768                 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
769                 if (atomic_read(total_ws) > cpus && !*free_ws)
770                         schedule();
771                 finish_wait(ws_wait, &wait);
772                 goto again;
773         }
774         atomic_inc(total_ws);
775         spin_unlock(ws_lock);
776
777         /*
778          * Allocation helpers call vmalloc that can't use GFP_NOFS, so we have
779          * to turn it off here because we might get called from the restricted
780          * context of btrfs_compress_bio/btrfs_compress_pages
781          */
782         nofs_flag = memalloc_nofs_save();
783         workspace = btrfs_compress_op[idx]->alloc_workspace();
784         memalloc_nofs_restore(nofs_flag);
785
786         if (IS_ERR(workspace)) {
787                 atomic_dec(total_ws);
788                 wake_up(ws_wait);
789
790                 /*
791                  * Do not return the error but go back to waiting. There's a
792                  * workspace preallocated for each type and the compression
793                  * time is bounded so we get to a workspace eventually. This
794                  * makes our caller's life easier.
795                  *
796                  * To prevent silent and low-probability deadlocks (when the
797                  * initial preallocation fails), check if there are any
798                  * workspaces at all.
799                  */
800                 if (atomic_read(total_ws) == 0) {
801                         static DEFINE_RATELIMIT_STATE(_rs,
802                                         /* once per minute */ 60 * HZ,
803                                         /* no burst */ 1);
804
805                         if (__ratelimit(&_rs)) {
806                                 pr_warn("BTRFS: no compression workspaces, low memory, retrying\n");
807                         }
808                 }
809                 goto again;
810         }
811         return workspace;
812 }
813
814 /*
815  * put a workspace struct back on the list or free it if we have enough
816  * idle ones sitting around
817  */
818 static void free_workspace(int type, struct list_head *workspace)
819 {
820         int idx = type - 1;
821         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
822         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
823         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
824         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
825         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
826
827         spin_lock(ws_lock);
828         if (*free_ws < num_online_cpus()) {
829                 list_add(workspace, idle_ws);
830                 (*free_ws)++;
831                 spin_unlock(ws_lock);
832                 goto wake;
833         }
834         spin_unlock(ws_lock);
835
836         btrfs_compress_op[idx]->free_workspace(workspace);
837         atomic_dec(total_ws);
838 wake:
839         /*
840          * Make sure counter is updated before we wake up waiters.
841          */
842         smp_mb();
843         if (waitqueue_active(ws_wait))
844                 wake_up(ws_wait);
845 }
846
847 /*
848  * cleanup function for module exit
849  */
850 static void free_workspaces(void)
851 {
852         struct list_head *workspace;
853         int i;
854
855         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
856                 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
857                         workspace = btrfs_comp_ws[i].idle_ws.next;
858                         list_del(workspace);
859                         btrfs_compress_op[i]->free_workspace(workspace);
860                         atomic_dec(&btrfs_comp_ws[i].total_ws);
861                 }
862         }
863 }
864
865 /*
866  * Given an address space and start and length, compress the bytes into @pages
867  * that are allocated on demand.
868  *
869  * @out_pages is an in/out parameter, holds maximum number of pages to allocate
870  * and returns number of actually allocated pages
871  *
872  * @total_in is used to return the number of bytes actually read.  It
873  * may be smaller than the input length if we had to exit early because we
874  * ran out of room in the pages array or because we cross the
875  * max_out threshold.
876  *
877  * @total_out is an in/out parameter, must be set to the input length and will
878  * be also used to return the total number of compressed bytes
879  *
880  * @max_out tells us the max number of bytes that we're allowed to
881  * stuff into pages
882  */
883 int btrfs_compress_pages(int type, struct address_space *mapping,
884                          u64 start, struct page **pages,
885                          unsigned long *out_pages,
886                          unsigned long *total_in,
887                          unsigned long *total_out)
888 {
889         struct list_head *workspace;
890         int ret;
891
892         workspace = find_workspace(type);
893
894         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
895                                                       start, pages,
896                                                       out_pages,
897                                                       total_in, total_out);
898         free_workspace(type, workspace);
899         return ret;
900 }
901
902 /*
903  * pages_in is an array of pages with compressed data.
904  *
905  * disk_start is the starting logical offset of this array in the file
906  *
907  * orig_bio contains the pages from the file that we want to decompress into
908  *
909  * srclen is the number of bytes in pages_in
910  *
911  * The basic idea is that we have a bio that was created by readpages.
912  * The pages in the bio are for the uncompressed data, and they may not
913  * be contiguous.  They all correspond to the range of bytes covered by
914  * the compressed extent.
915  */
916 static int btrfs_decompress_bio(struct compressed_bio *cb)
917 {
918         struct list_head *workspace;
919         int ret;
920         int type = cb->compress_type;
921
922         workspace = find_workspace(type);
923         ret = btrfs_compress_op[type - 1]->decompress_bio(workspace, cb);
924         free_workspace(type, workspace);
925
926         return ret;
927 }
928
929 /*
930  * a less complex decompression routine.  Our compressed data fits in a
931  * single page, and we want to read a single page out of it.
932  * start_byte tells us the offset into the compressed data we're interested in
933  */
934 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
935                      unsigned long start_byte, size_t srclen, size_t destlen)
936 {
937         struct list_head *workspace;
938         int ret;
939
940         workspace = find_workspace(type);
941
942         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
943                                                   dest_page, start_byte,
944                                                   srclen, destlen);
945
946         free_workspace(type, workspace);
947         return ret;
948 }
949
950 void btrfs_exit_compress(void)
951 {
952         free_workspaces();
953 }
954
955 /*
956  * Copy uncompressed data from working buffer to pages.
957  *
958  * buf_start is the byte offset we're of the start of our workspace buffer.
959  *
960  * total_out is the last byte of the buffer
961  */
962 int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
963                               unsigned long total_out, u64 disk_start,
964                               struct bio *bio)
965 {
966         unsigned long buf_offset;
967         unsigned long current_buf_start;
968         unsigned long start_byte;
969         unsigned long prev_start_byte;
970         unsigned long working_bytes = total_out - buf_start;
971         unsigned long bytes;
972         char *kaddr;
973         struct bio_vec bvec = bio_iter_iovec(bio, bio->bi_iter);
974
975         /*
976          * start byte is the first byte of the page we're currently
977          * copying into relative to the start of the compressed data.
978          */
979         start_byte = page_offset(bvec.bv_page) - disk_start;
980
981         /* we haven't yet hit data corresponding to this page */
982         if (total_out <= start_byte)
983                 return 1;
984
985         /*
986          * the start of the data we care about is offset into
987          * the middle of our working buffer
988          */
989         if (total_out > start_byte && buf_start < start_byte) {
990                 buf_offset = start_byte - buf_start;
991                 working_bytes -= buf_offset;
992         } else {
993                 buf_offset = 0;
994         }
995         current_buf_start = buf_start;
996
997         /* copy bytes from the working buffer into the pages */
998         while (working_bytes > 0) {
999                 bytes = min_t(unsigned long, bvec.bv_len,
1000                                 PAGE_SIZE - buf_offset);
1001                 bytes = min(bytes, working_bytes);
1002
1003                 kaddr = kmap_atomic(bvec.bv_page);
1004                 memcpy(kaddr + bvec.bv_offset, buf + buf_offset, bytes);
1005                 kunmap_atomic(kaddr);
1006                 flush_dcache_page(bvec.bv_page);
1007
1008                 buf_offset += bytes;
1009                 working_bytes -= bytes;
1010                 current_buf_start += bytes;
1011
1012                 /* check if we need to pick another page */
1013                 bio_advance(bio, bytes);
1014                 if (!bio->bi_iter.bi_size)
1015                         return 0;
1016                 bvec = bio_iter_iovec(bio, bio->bi_iter);
1017                 prev_start_byte = start_byte;
1018                 start_byte = page_offset(bvec.bv_page) - disk_start;
1019
1020                 /*
1021                  * We need to make sure we're only adjusting
1022                  * our offset into compression working buffer when
1023                  * we're switching pages.  Otherwise we can incorrectly
1024                  * keep copying when we were actually done.
1025                  */
1026                 if (start_byte != prev_start_byte) {
1027                         /*
1028                          * make sure our new page is covered by this
1029                          * working buffer
1030                          */
1031                         if (total_out <= start_byte)
1032                                 return 1;
1033
1034                         /*
1035                          * the next page in the biovec might not be adjacent
1036                          * to the last page, but it might still be found
1037                          * inside this working buffer. bump our offset pointer
1038                          */
1039                         if (total_out > start_byte &&
1040                             current_buf_start < start_byte) {
1041                                 buf_offset = start_byte - buf_start;
1042                                 working_bytes = total_out - start_byte;
1043                                 current_buf_start = buf_start + buf_offset;
1044                         }
1045                 }
1046         }
1047
1048         return 1;
1049 }