Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux...
[~shefty/rdma-dev.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 "compat.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 struct compressed_bio {
47         /* number of bios pending for this compressed extent */
48         atomic_t pending_bios;
49
50         /* the pages with the compressed data on them */
51         struct page **compressed_pages;
52
53         /* inode that owns this data */
54         struct inode *inode;
55
56         /* starting offset in the inode for our pages */
57         u64 start;
58
59         /* number of bytes in the inode we're working on */
60         unsigned long len;
61
62         /* number of bytes on disk */
63         unsigned long compressed_len;
64
65         /* the compression algorithm for this bio */
66         int compress_type;
67
68         /* number of compressed pages in the array */
69         unsigned long nr_pages;
70
71         /* IO errors */
72         int errors;
73         int mirror_num;
74
75         /* for reads, this is the bio we are copying the data into */
76         struct bio *orig_bio;
77
78         /*
79          * the start of a variable length array of checksums only
80          * used by reads
81          */
82         u32 sums;
83 };
84
85 static inline int compressed_bio_size(struct btrfs_root *root,
86                                       unsigned long disk_size)
87 {
88         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
89
90         return sizeof(struct compressed_bio) +
91                 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
92                 csum_size;
93 }
94
95 static struct bio *compressed_bio_alloc(struct block_device *bdev,
96                                         u64 first_byte, gfp_t gfp_flags)
97 {
98         int nr_vecs;
99
100         nr_vecs = bio_get_nr_vecs(bdev);
101         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
102 }
103
104 static int check_compressed_csum(struct inode *inode,
105                                  struct compressed_bio *cb,
106                                  u64 disk_start)
107 {
108         int ret;
109         struct btrfs_root *root = BTRFS_I(inode)->root;
110         struct page *page;
111         unsigned long i;
112         char *kaddr;
113         u32 csum;
114         u32 *cb_sum = &cb->sums;
115
116         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
117                 return 0;
118
119         for (i = 0; i < cb->nr_pages; i++) {
120                 page = cb->compressed_pages[i];
121                 csum = ~(u32)0;
122
123                 kaddr = kmap_atomic(page);
124                 csum = btrfs_csum_data(root, kaddr, csum, PAGE_CACHE_SIZE);
125                 btrfs_csum_final(csum, (char *)&csum);
126                 kunmap_atomic(kaddr);
127
128                 if (csum != *cb_sum) {
129                         printk(KERN_INFO "btrfs csum failed ino %llu "
130                                "extent %llu csum %u "
131                                "wanted %u mirror %d\n",
132                                (unsigned long long)btrfs_ino(inode),
133                                (unsigned long long)disk_start,
134                                csum, *cb_sum, cb->mirror_num);
135                         ret = -EIO;
136                         goto fail;
137                 }
138                 cb_sum++;
139
140         }
141         ret = 0;
142 fail:
143         return ret;
144 }
145
146 /* when we finish reading compressed pages from the disk, we
147  * decompress them and then run the bio end_io routines on the
148  * decompressed pages (in the inode address space).
149  *
150  * This allows the checksumming and other IO error handling routines
151  * to work normally
152  *
153  * The compressed pages are freed here, and it must be run
154  * in process context
155  */
156 static void end_compressed_bio_read(struct bio *bio, int err)
157 {
158         struct compressed_bio *cb = bio->bi_private;
159         struct inode *inode;
160         struct page *page;
161         unsigned long index;
162         int ret;
163
164         if (err)
165                 cb->errors = 1;
166
167         /* if there are more bios still pending for this compressed
168          * extent, just exit
169          */
170         if (!atomic_dec_and_test(&cb->pending_bios))
171                 goto out;
172
173         inode = cb->inode;
174         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
175         if (ret)
176                 goto csum_failed;
177
178         /* ok, we're the last bio for this extent, lets start
179          * the decompression.
180          */
181         ret = btrfs_decompress_biovec(cb->compress_type,
182                                       cb->compressed_pages,
183                                       cb->start,
184                                       cb->orig_bio->bi_io_vec,
185                                       cb->orig_bio->bi_vcnt,
186                                       cb->compressed_len);
187 csum_failed:
188         if (ret)
189                 cb->errors = 1;
190
191         /* release the compressed pages */
192         index = 0;
193         for (index = 0; index < cb->nr_pages; index++) {
194                 page = cb->compressed_pages[index];
195                 page->mapping = NULL;
196                 page_cache_release(page);
197         }
198
199         /* do io completion on the original bio */
200         if (cb->errors) {
201                 bio_io_error(cb->orig_bio);
202         } else {
203                 int bio_index = 0;
204                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
205
206                 /*
207                  * we have verified the checksum already, set page
208                  * checked so the end_io handlers know about it
209                  */
210                 while (bio_index < cb->orig_bio->bi_vcnt) {
211                         SetPageChecked(bvec->bv_page);
212                         bvec++;
213                         bio_index++;
214                 }
215                 bio_endio(cb->orig_bio, 0);
216         }
217
218         /* finally free the cb struct */
219         kfree(cb->compressed_pages);
220         kfree(cb);
221 out:
222         bio_put(bio);
223 }
224
225 /*
226  * Clear the writeback bits on all of the file
227  * pages for a compressed write
228  */
229 static noinline void end_compressed_writeback(struct inode *inode, u64 start,
230                                               unsigned long ram_size)
231 {
232         unsigned long index = start >> PAGE_CACHE_SHIFT;
233         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
234         struct page *pages[16];
235         unsigned long nr_pages = end_index - index + 1;
236         int i;
237         int ret;
238
239         while (nr_pages > 0) {
240                 ret = find_get_pages_contig(inode->i_mapping, index,
241                                      min_t(unsigned long,
242                                      nr_pages, ARRAY_SIZE(pages)), pages);
243                 if (ret == 0) {
244                         nr_pages -= 1;
245                         index += 1;
246                         continue;
247                 }
248                 for (i = 0; i < ret; i++) {
249                         end_page_writeback(pages[i]);
250                         page_cache_release(pages[i]);
251                 }
252                 nr_pages -= ret;
253                 index += ret;
254         }
255         /* the inode may be gone now */
256 }
257
258 /*
259  * do the cleanup once all the compressed pages hit the disk.
260  * This will clear writeback on the file pages and free the compressed
261  * pages.
262  *
263  * This also calls the writeback end hooks for the file pages so that
264  * metadata and checksums can be updated in the file.
265  */
266 static void end_compressed_bio_write(struct bio *bio, int err)
267 {
268         struct extent_io_tree *tree;
269         struct compressed_bio *cb = bio->bi_private;
270         struct inode *inode;
271         struct page *page;
272         unsigned long index;
273
274         if (err)
275                 cb->errors = 1;
276
277         /* if there are more bios still pending for this compressed
278          * extent, just exit
279          */
280         if (!atomic_dec_and_test(&cb->pending_bios))
281                 goto out;
282
283         /* ok, we're the last bio for this extent, step one is to
284          * call back into the FS and do all the end_io operations
285          */
286         inode = cb->inode;
287         tree = &BTRFS_I(inode)->io_tree;
288         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
289         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
290                                          cb->start,
291                                          cb->start + cb->len - 1,
292                                          NULL, 1);
293         cb->compressed_pages[0]->mapping = NULL;
294
295         end_compressed_writeback(inode, cb->start, cb->len);
296         /* note, our inode could be gone now */
297
298         /*
299          * release the compressed pages, these came from alloc_page and
300          * are not attached to the inode at all
301          */
302         index = 0;
303         for (index = 0; index < cb->nr_pages; index++) {
304                 page = cb->compressed_pages[index];
305                 page->mapping = NULL;
306                 page_cache_release(page);
307         }
308
309         /* finally free the cb struct */
310         kfree(cb->compressed_pages);
311         kfree(cb);
312 out:
313         bio_put(bio);
314 }
315
316 /*
317  * worker function to build and submit bios for previously compressed pages.
318  * The corresponding pages in the inode should be marked for writeback
319  * and the compressed pages should have a reference on them for dropping
320  * when the IO is complete.
321  *
322  * This also checksums the file bytes and gets things ready for
323  * the end io hooks.
324  */
325 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
326                                  unsigned long len, u64 disk_start,
327                                  unsigned long compressed_len,
328                                  struct page **compressed_pages,
329                                  unsigned long nr_pages)
330 {
331         struct bio *bio = NULL;
332         struct btrfs_root *root = BTRFS_I(inode)->root;
333         struct compressed_bio *cb;
334         unsigned long bytes_left;
335         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
336         int pg_index = 0;
337         struct page *page;
338         u64 first_byte = disk_start;
339         struct block_device *bdev;
340         int ret;
341         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
342
343         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
344         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
345         if (!cb)
346                 return -ENOMEM;
347         atomic_set(&cb->pending_bios, 0);
348         cb->errors = 0;
349         cb->inode = inode;
350         cb->start = start;
351         cb->len = len;
352         cb->mirror_num = 0;
353         cb->compressed_pages = compressed_pages;
354         cb->compressed_len = compressed_len;
355         cb->orig_bio = NULL;
356         cb->nr_pages = nr_pages;
357
358         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
359
360         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
361         if(!bio) {
362                 kfree(cb);
363                 return -ENOMEM;
364         }
365         bio->bi_private = cb;
366         bio->bi_end_io = end_compressed_bio_write;
367         atomic_inc(&cb->pending_bios);
368
369         /* create and submit bios for the compressed pages */
370         bytes_left = compressed_len;
371         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
372                 page = compressed_pages[pg_index];
373                 page->mapping = inode->i_mapping;
374                 if (bio->bi_size)
375                         ret = io_tree->ops->merge_bio_hook(page, 0,
376                                                            PAGE_CACHE_SIZE,
377                                                            bio, 0);
378                 else
379                         ret = 0;
380
381                 page->mapping = NULL;
382                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
383                     PAGE_CACHE_SIZE) {
384                         bio_get(bio);
385
386                         /*
387                          * inc the count before we submit the bio so
388                          * we know the end IO handler won't happen before
389                          * we inc the count.  Otherwise, the cb might get
390                          * freed before we're done setting it up
391                          */
392                         atomic_inc(&cb->pending_bios);
393                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
394                         BUG_ON(ret); /* -ENOMEM */
395
396                         if (!skip_sum) {
397                                 ret = btrfs_csum_one_bio(root, inode, bio,
398                                                          start, 1);
399                                 BUG_ON(ret); /* -ENOMEM */
400                         }
401
402                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
403                         BUG_ON(ret); /* -ENOMEM */
404
405                         bio_put(bio);
406
407                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
408                         bio->bi_private = cb;
409                         bio->bi_end_io = end_compressed_bio_write;
410                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
411                 }
412                 if (bytes_left < PAGE_CACHE_SIZE) {
413                         printk("bytes left %lu compress len %lu nr %lu\n",
414                                bytes_left, cb->compressed_len, cb->nr_pages);
415                 }
416                 bytes_left -= PAGE_CACHE_SIZE;
417                 first_byte += PAGE_CACHE_SIZE;
418                 cond_resched();
419         }
420         bio_get(bio);
421
422         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
423         BUG_ON(ret); /* -ENOMEM */
424
425         if (!skip_sum) {
426                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
427                 BUG_ON(ret); /* -ENOMEM */
428         }
429
430         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
431         BUG_ON(ret); /* -ENOMEM */
432
433         bio_put(bio);
434         return 0;
435 }
436
437 static noinline int add_ra_bio_pages(struct inode *inode,
438                                      u64 compressed_end,
439                                      struct compressed_bio *cb)
440 {
441         unsigned long end_index;
442         unsigned long pg_index;
443         u64 last_offset;
444         u64 isize = i_size_read(inode);
445         int ret;
446         struct page *page;
447         unsigned long nr_pages = 0;
448         struct extent_map *em;
449         struct address_space *mapping = inode->i_mapping;
450         struct extent_map_tree *em_tree;
451         struct extent_io_tree *tree;
452         u64 end;
453         int misses = 0;
454
455         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
456         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
457         em_tree = &BTRFS_I(inode)->extent_tree;
458         tree = &BTRFS_I(inode)->io_tree;
459
460         if (isize == 0)
461                 return 0;
462
463         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
464
465         while (last_offset < compressed_end) {
466                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
467
468                 if (pg_index > end_index)
469                         break;
470
471                 rcu_read_lock();
472                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
473                 rcu_read_unlock();
474                 if (page) {
475                         misses++;
476                         if (misses > 4)
477                                 break;
478                         goto next;
479                 }
480
481                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
482                                                                 ~__GFP_FS);
483                 if (!page)
484                         break;
485
486                 if (add_to_page_cache_lru(page, mapping, pg_index,
487                                                                 GFP_NOFS)) {
488                         page_cache_release(page);
489                         goto next;
490                 }
491
492                 end = last_offset + PAGE_CACHE_SIZE - 1;
493                 /*
494                  * at this point, we have a locked page in the page cache
495                  * for these bytes in the file.  But, we have to make
496                  * sure they map to this compressed extent on disk.
497                  */
498                 set_page_extent_mapped(page);
499                 lock_extent(tree, last_offset, end);
500                 read_lock(&em_tree->lock);
501                 em = lookup_extent_mapping(em_tree, last_offset,
502                                            PAGE_CACHE_SIZE);
503                 read_unlock(&em_tree->lock);
504
505                 if (!em || last_offset < em->start ||
506                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
507                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
508                         free_extent_map(em);
509                         unlock_extent(tree, last_offset, end);
510                         unlock_page(page);
511                         page_cache_release(page);
512                         break;
513                 }
514                 free_extent_map(em);
515
516                 if (page->index == end_index) {
517                         char *userpage;
518                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
519
520                         if (zero_offset) {
521                                 int zeros;
522                                 zeros = PAGE_CACHE_SIZE - zero_offset;
523                                 userpage = kmap_atomic(page);
524                                 memset(userpage + zero_offset, 0, zeros);
525                                 flush_dcache_page(page);
526                                 kunmap_atomic(userpage);
527                         }
528                 }
529
530                 ret = bio_add_page(cb->orig_bio, page,
531                                    PAGE_CACHE_SIZE, 0);
532
533                 if (ret == PAGE_CACHE_SIZE) {
534                         nr_pages++;
535                         page_cache_release(page);
536                 } else {
537                         unlock_extent(tree, last_offset, end);
538                         unlock_page(page);
539                         page_cache_release(page);
540                         break;
541                 }
542 next:
543                 last_offset += PAGE_CACHE_SIZE;
544         }
545         return 0;
546 }
547
548 /*
549  * for a compressed read, the bio we get passed has all the inode pages
550  * in it.  We don't actually do IO on those pages but allocate new ones
551  * to hold the compressed pages on disk.
552  *
553  * bio->bi_sector points to the compressed extent on disk
554  * bio->bi_io_vec points to all of the inode pages
555  * bio->bi_vcnt is a count of pages
556  *
557  * After the compressed pages are read, we copy the bytes into the
558  * bio we were passed and then call the bio end_io calls
559  */
560 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
561                                  int mirror_num, unsigned long bio_flags)
562 {
563         struct extent_io_tree *tree;
564         struct extent_map_tree *em_tree;
565         struct compressed_bio *cb;
566         struct btrfs_root *root = BTRFS_I(inode)->root;
567         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
568         unsigned long compressed_len;
569         unsigned long nr_pages;
570         unsigned long pg_index;
571         struct page *page;
572         struct block_device *bdev;
573         struct bio *comp_bio;
574         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
575         u64 em_len;
576         u64 em_start;
577         struct extent_map *em;
578         int ret = -ENOMEM;
579         u32 *sums;
580
581         tree = &BTRFS_I(inode)->io_tree;
582         em_tree = &BTRFS_I(inode)->extent_tree;
583
584         /* we need the actual starting offset of this extent in the file */
585         read_lock(&em_tree->lock);
586         em = lookup_extent_mapping(em_tree,
587                                    page_offset(bio->bi_io_vec->bv_page),
588                                    PAGE_CACHE_SIZE);
589         read_unlock(&em_tree->lock);
590         if (!em)
591                 return -EIO;
592
593         compressed_len = em->block_len;
594         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
595         if (!cb)
596                 goto out;
597
598         atomic_set(&cb->pending_bios, 0);
599         cb->errors = 0;
600         cb->inode = inode;
601         cb->mirror_num = mirror_num;
602         sums = &cb->sums;
603
604         cb->start = em->orig_start;
605         em_len = em->len;
606         em_start = em->start;
607
608         free_extent_map(em);
609         em = NULL;
610
611         cb->len = uncompressed_len;
612         cb->compressed_len = compressed_len;
613         cb->compress_type = extent_compress_type(bio_flags);
614         cb->orig_bio = bio;
615
616         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
617                                  PAGE_CACHE_SIZE;
618         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
619                                        GFP_NOFS);
620         if (!cb->compressed_pages)
621                 goto fail1;
622
623         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
624
625         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
626                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
627                                                               __GFP_HIGHMEM);
628                 if (!cb->compressed_pages[pg_index])
629                         goto fail2;
630         }
631         cb->nr_pages = nr_pages;
632
633         add_ra_bio_pages(inode, em_start + em_len, cb);
634
635         /* include any pages we added in add_ra-bio_pages */
636         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
637         cb->len = uncompressed_len;
638
639         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
640         if (!comp_bio)
641                 goto fail2;
642         comp_bio->bi_private = cb;
643         comp_bio->bi_end_io = end_compressed_bio_read;
644         atomic_inc(&cb->pending_bios);
645
646         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
647                 page = cb->compressed_pages[pg_index];
648                 page->mapping = inode->i_mapping;
649                 page->index = em_start >> PAGE_CACHE_SHIFT;
650
651                 if (comp_bio->bi_size)
652                         ret = tree->ops->merge_bio_hook(page, 0,
653                                                         PAGE_CACHE_SIZE,
654                                                         comp_bio, 0);
655                 else
656                         ret = 0;
657
658                 page->mapping = NULL;
659                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
660                     PAGE_CACHE_SIZE) {
661                         bio_get(comp_bio);
662
663                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
664                         BUG_ON(ret); /* -ENOMEM */
665
666                         /*
667                          * inc the count before we submit the bio so
668                          * we know the end IO handler won't happen before
669                          * we inc the count.  Otherwise, the cb might get
670                          * freed before we're done setting it up
671                          */
672                         atomic_inc(&cb->pending_bios);
673
674                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
675                                 ret = btrfs_lookup_bio_sums(root, inode,
676                                                         comp_bio, sums);
677                                 BUG_ON(ret); /* -ENOMEM */
678                         }
679                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
680                                 root->sectorsize;
681
682                         ret = btrfs_map_bio(root, READ, comp_bio,
683                                             mirror_num, 0);
684                         BUG_ON(ret); /* -ENOMEM */
685
686                         bio_put(comp_bio);
687
688                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
689                                                         GFP_NOFS);
690                         comp_bio->bi_private = cb;
691                         comp_bio->bi_end_io = end_compressed_bio_read;
692
693                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
694                 }
695                 cur_disk_byte += PAGE_CACHE_SIZE;
696         }
697         bio_get(comp_bio);
698
699         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
700         BUG_ON(ret); /* -ENOMEM */
701
702         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
703                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
704                 BUG_ON(ret); /* -ENOMEM */
705         }
706
707         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
708         BUG_ON(ret); /* -ENOMEM */
709
710         bio_put(comp_bio);
711         return 0;
712
713 fail2:
714         for (pg_index = 0; pg_index < nr_pages; pg_index++)
715                 free_page((unsigned long)cb->compressed_pages[pg_index]);
716
717         kfree(cb->compressed_pages);
718 fail1:
719         kfree(cb);
720 out:
721         free_extent_map(em);
722         return ret;
723 }
724
725 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
726 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
727 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
728 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
729 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
730
731 struct btrfs_compress_op *btrfs_compress_op[] = {
732         &btrfs_zlib_compress,
733         &btrfs_lzo_compress,
734 };
735
736 void __init btrfs_init_compress(void)
737 {
738         int i;
739
740         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
741                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
742                 spin_lock_init(&comp_workspace_lock[i]);
743                 atomic_set(&comp_alloc_workspace[i], 0);
744                 init_waitqueue_head(&comp_workspace_wait[i]);
745         }
746 }
747
748 /*
749  * this finds an available workspace or allocates a new one
750  * ERR_PTR is returned if things go bad.
751  */
752 static struct list_head *find_workspace(int type)
753 {
754         struct list_head *workspace;
755         int cpus = num_online_cpus();
756         int idx = type - 1;
757
758         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
759         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
760         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
761         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
762         int *num_workspace                      = &comp_num_workspace[idx];
763 again:
764         spin_lock(workspace_lock);
765         if (!list_empty(idle_workspace)) {
766                 workspace = idle_workspace->next;
767                 list_del(workspace);
768                 (*num_workspace)--;
769                 spin_unlock(workspace_lock);
770                 return workspace;
771
772         }
773         if (atomic_read(alloc_workspace) > cpus) {
774                 DEFINE_WAIT(wait);
775
776                 spin_unlock(workspace_lock);
777                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
778                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
779                         schedule();
780                 finish_wait(workspace_wait, &wait);
781                 goto again;
782         }
783         atomic_inc(alloc_workspace);
784         spin_unlock(workspace_lock);
785
786         workspace = btrfs_compress_op[idx]->alloc_workspace();
787         if (IS_ERR(workspace)) {
788                 atomic_dec(alloc_workspace);
789                 wake_up(workspace_wait);
790         }
791         return workspace;
792 }
793
794 /*
795  * put a workspace struct back on the list or free it if we have enough
796  * idle ones sitting around
797  */
798 static void free_workspace(int type, struct list_head *workspace)
799 {
800         int idx = type - 1;
801         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
802         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
803         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
804         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
805         int *num_workspace                      = &comp_num_workspace[idx];
806
807         spin_lock(workspace_lock);
808         if (*num_workspace < num_online_cpus()) {
809                 list_add_tail(workspace, idle_workspace);
810                 (*num_workspace)++;
811                 spin_unlock(workspace_lock);
812                 goto wake;
813         }
814         spin_unlock(workspace_lock);
815
816         btrfs_compress_op[idx]->free_workspace(workspace);
817         atomic_dec(alloc_workspace);
818 wake:
819         if (waitqueue_active(workspace_wait))
820                 wake_up(workspace_wait);
821 }
822
823 /*
824  * cleanup function for module exit
825  */
826 static void free_workspaces(void)
827 {
828         struct list_head *workspace;
829         int i;
830
831         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
832                 while (!list_empty(&comp_idle_workspace[i])) {
833                         workspace = comp_idle_workspace[i].next;
834                         list_del(workspace);
835                         btrfs_compress_op[i]->free_workspace(workspace);
836                         atomic_dec(&comp_alloc_workspace[i]);
837                 }
838         }
839 }
840
841 /*
842  * given an address space and start/len, compress the bytes.
843  *
844  * pages are allocated to hold the compressed result and stored
845  * in 'pages'
846  *
847  * out_pages is used to return the number of pages allocated.  There
848  * may be pages allocated even if we return an error
849  *
850  * total_in is used to return the number of bytes actually read.  It
851  * may be smaller then len if we had to exit early because we
852  * ran out of room in the pages array or because we cross the
853  * max_out threshold.
854  *
855  * total_out is used to return the total number of compressed bytes
856  *
857  * max_out tells us the max number of bytes that we're allowed to
858  * stuff into pages
859  */
860 int btrfs_compress_pages(int type, struct address_space *mapping,
861                          u64 start, unsigned long len,
862                          struct page **pages,
863                          unsigned long nr_dest_pages,
864                          unsigned long *out_pages,
865                          unsigned long *total_in,
866                          unsigned long *total_out,
867                          unsigned long max_out)
868 {
869         struct list_head *workspace;
870         int ret;
871
872         workspace = find_workspace(type);
873         if (IS_ERR(workspace))
874                 return -1;
875
876         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
877                                                       start, len, pages,
878                                                       nr_dest_pages, out_pages,
879                                                       total_in, total_out,
880                                                       max_out);
881         free_workspace(type, workspace);
882         return ret;
883 }
884
885 /*
886  * pages_in is an array of pages with compressed data.
887  *
888  * disk_start is the starting logical offset of this array in the file
889  *
890  * bvec is a bio_vec of pages from the file that we want to decompress into
891  *
892  * vcnt is the count of pages in the biovec
893  *
894  * srclen is the number of bytes in pages_in
895  *
896  * The basic idea is that we have a bio that was created by readpages.
897  * The pages in the bio are for the uncompressed data, and they may not
898  * be contiguous.  They all correspond to the range of bytes covered by
899  * the compressed extent.
900  */
901 int btrfs_decompress_biovec(int type, struct page **pages_in, u64 disk_start,
902                             struct bio_vec *bvec, int vcnt, size_t srclen)
903 {
904         struct list_head *workspace;
905         int ret;
906
907         workspace = find_workspace(type);
908         if (IS_ERR(workspace))
909                 return -ENOMEM;
910
911         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
912                                                          disk_start,
913                                                          bvec, vcnt, srclen);
914         free_workspace(type, workspace);
915         return ret;
916 }
917
918 /*
919  * a less complex decompression routine.  Our compressed data fits in a
920  * single page, and we want to read a single page out of it.
921  * start_byte tells us the offset into the compressed data we're interested in
922  */
923 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
924                      unsigned long start_byte, size_t srclen, size_t destlen)
925 {
926         struct list_head *workspace;
927         int ret;
928
929         workspace = find_workspace(type);
930         if (IS_ERR(workspace))
931                 return -ENOMEM;
932
933         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
934                                                   dest_page, start_byte,
935                                                   srclen, destlen);
936
937         free_workspace(type, workspace);
938         return ret;
939 }
940
941 void btrfs_exit_compress(void)
942 {
943         free_workspaces();
944 }
945
946 /*
947  * Copy uncompressed data from working buffer to pages.
948  *
949  * buf_start is the byte offset we're of the start of our workspace buffer.
950  *
951  * total_out is the last byte of the buffer
952  */
953 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
954                               unsigned long total_out, u64 disk_start,
955                               struct bio_vec *bvec, int vcnt,
956                               unsigned long *pg_index,
957                               unsigned long *pg_offset)
958 {
959         unsigned long buf_offset;
960         unsigned long current_buf_start;
961         unsigned long start_byte;
962         unsigned long working_bytes = total_out - buf_start;
963         unsigned long bytes;
964         char *kaddr;
965         struct page *page_out = bvec[*pg_index].bv_page;
966
967         /*
968          * start byte is the first byte of the page we're currently
969          * copying into relative to the start of the compressed data.
970          */
971         start_byte = page_offset(page_out) - disk_start;
972
973         /* we haven't yet hit data corresponding to this page */
974         if (total_out <= start_byte)
975                 return 1;
976
977         /*
978          * the start of the data we care about is offset into
979          * the middle of our working buffer
980          */
981         if (total_out > start_byte && buf_start < start_byte) {
982                 buf_offset = start_byte - buf_start;
983                 working_bytes -= buf_offset;
984         } else {
985                 buf_offset = 0;
986         }
987         current_buf_start = buf_start;
988
989         /* copy bytes from the working buffer into the pages */
990         while (working_bytes > 0) {
991                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
992                             PAGE_CACHE_SIZE - buf_offset);
993                 bytes = min(bytes, working_bytes);
994                 kaddr = kmap_atomic(page_out);
995                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
996                 kunmap_atomic(kaddr);
997                 flush_dcache_page(page_out);
998
999                 *pg_offset += bytes;
1000                 buf_offset += bytes;
1001                 working_bytes -= bytes;
1002                 current_buf_start += bytes;
1003
1004                 /* check if we need to pick another page */
1005                 if (*pg_offset == PAGE_CACHE_SIZE) {
1006                         (*pg_index)++;
1007                         if (*pg_index >= vcnt)
1008                                 return 0;
1009
1010                         page_out = bvec[*pg_index].bv_page;
1011                         *pg_offset = 0;
1012                         start_byte = page_offset(page_out) - disk_start;
1013
1014                         /*
1015                          * make sure our new page is covered by this
1016                          * working buffer
1017                          */
1018                         if (total_out <= start_byte)
1019                                 return 1;
1020
1021                         /*
1022                          * the next page in the biovec might not be adjacent
1023                          * to the last page, but it might still be found
1024                          * inside this working buffer. bump our offset pointer
1025                          */
1026                         if (total_out > start_byte &&
1027                             current_buf_start < start_byte) {
1028                                 buf_offset = start_byte - buf_start;
1029                                 working_bytes = total_out - start_byte;
1030                                 current_buf_start = buf_start + buf_offset;
1031                         }
1032                 }
1033         }
1034
1035         return 1;
1036 }