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         return sizeof(struct compressed_bio) +
  90                 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
  91                 csum_size;
  92 }
  93
  94 static struct bio *compressed_bio_alloc(struct block_device *bdev,
  95                                         u64 first_byte, gfp_t gfp_flags)
  96 {
  97         int nr_vecs;
  98
  99         nr_vecs = bio_get_nr_vecs(bdev);
 100         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
 101 }
 102
 103 static int check_compressed_csum(struct inode *inode,
 104                                  struct compressed_bio *cb,
 105                                  u64 disk_start)
 106 {
 107         int ret;
 108         struct btrfs_root *root = BTRFS_I(inode)->root;
 109         struct page *page;
 110         unsigned long i;
 111         char *kaddr;
 112         u32 csum;
 113         u32 *cb_sum = &cb->sums;
 114
 115         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 116                 return 0;
 117
 118         for (i = 0; i < cb->nr_pages; i++) {
 119                 page = cb->compressed_pages[i];
 120                 csum = ~(u32)0;
 121
 122                 kaddr = kmap_atomic(page, KM_USER0);
 123                 csum = btrfs_csum_data(root, kaddr, csum, PAGE_CACHE_SIZE);
 124                 btrfs_csum_final(csum, (char *)&csum);
 125                 kunmap_atomic(kaddr, KM_USER0);
 126
 127                 if (csum != *cb_sum) {
 128                         printk(KERN_INFO "btrfs csum failed ino %lu "
 129                                "extent %llu csum %u "
 130                                "wanted %u mirror %d\n", inode->i_ino,
 131                                (unsigned long long)disk_start,
 132                                csum, *cb_sum, cb->mirror_num);
 133                         ret = -EIO;
 134                         goto fail;
 135                 }
 136                 cb_sum++;
 137
 138         }
 139         ret = 0;
 140 fail:
 141         return ret;
 142 }
 143
 144 /* when we finish reading compressed pages from the disk, we
 145  * decompress them and then run the bio end_io routines on the
 146  * decompressed pages (in the inode address space).
 147  *
 148  * This allows the checksumming and other IO error handling routines
 149  * to work normally
 150  *
 151  * The compressed pages are freed here, and it must be run
 152  * in process context
 153  */
 154 static void end_compressed_bio_read(struct bio *bio, int err)
 155 {
 156         struct compressed_bio *cb = bio->bi_private;
 157         struct inode *inode;
 158         struct page *page;
 159         unsigned long index;
 160         int ret;
 161
 162         if (err)
 163                 cb->errors = 1;
 164
 165         /* if there are more bios still pending for this compressed
 166          * extent, just exit
 167          */
 168         if (!atomic_dec_and_test(&cb->pending_bios))
 169                 goto out;
 170
 171         inode = cb->inode;
 172         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
 173         if (ret)
 174                 goto csum_failed;
 175
 176         /* ok, we're the last bio for this extent, lets start
 177          * the decompression.
 178          */
 179         ret = btrfs_decompress_biovec(cb->compress_type,
 180                                       cb->compressed_pages,
 181                                       cb->start,
 182                                       cb->orig_bio->bi_io_vec,
 183                                       cb->orig_bio->bi_vcnt,
 184                                       cb->compressed_len);
 185 csum_failed:
 186         if (ret)
 187                 cb->errors = 1;
 188
 189         /* release the compressed pages */
 190         index = 0;
 191         for (index = 0; index < cb->nr_pages; index++) {
 192                 page = cb->compressed_pages[index];
 193                 page->mapping = NULL;
 194                 page_cache_release(page);
 195         }
 196
 197         /* do io completion on the original bio */
 198         if (cb->errors) {
 199                 bio_io_error(cb->orig_bio);
 200         } else {
 201                 int bio_index = 0;
 202                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
 203
 204                 /*
 205                  * we have verified the checksum already, set page
 206                  * checked so the end_io handlers know about it
 207                  */
 208                 while (bio_index < cb->orig_bio->bi_vcnt) {
 209                         SetPageChecked(bvec->bv_page);
 210                         bvec++;
 211                         bio_index++;
 212                 }
 213                 bio_endio(cb->orig_bio, 0);
 214         }
 215
 216         /* finally free the cb struct */
 217         kfree(cb->compressed_pages);
 218         kfree(cb);
 219 out:
 220         bio_put(bio);
 221 }
 222
 223 /*
 224  * Clear the writeback bits on all of the file
 225  * pages for a compressed write
 226  */
 227 static noinline int end_compressed_writeback(struct inode *inode, u64 start,
 228                                              unsigned long ram_size)
 229 {
 230         unsigned long index = start >> PAGE_CACHE_SHIFT;
 231         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
 232         struct page *pages[16];
 233         unsigned long nr_pages = end_index - index + 1;
 234         int i;
 235         int ret;
 236
 237         while (nr_pages > 0) {
 238                 ret = find_get_pages_contig(inode->i_mapping, index,
 239                                      min_t(unsigned long,
 240                                      nr_pages, ARRAY_SIZE(pages)), pages);
 241                 if (ret == 0) {
 242                         nr_pages -= 1;
 243                         index += 1;
 244                         continue;
 245                 }
 246                 for (i = 0; i < ret; i++) {
 247                         end_page_writeback(pages[i]);
 248                         page_cache_release(pages[i]);
 249                 }
 250                 nr_pages -= ret;
 251                 index += ret;
 252         }
 253         /* the inode may be gone now */
 254         return 0;
 255 }
 256
 257 /*
 258  * do the cleanup once all the compressed pages hit the disk.
 259  * This will clear writeback on the file pages and free the compressed
 260  * pages.
 261  *
 262  * This also calls the writeback end hooks for the file pages so that
 263  * metadata and checksums can be updated in the file.
 264  */
 265 static void end_compressed_bio_write(struct bio *bio, int err)
 266 {
 267         struct extent_io_tree *tree;
 268         struct compressed_bio *cb = bio->bi_private;
 269         struct inode *inode;
 270         struct page *page;
 271         unsigned long index;
 272
 273         if (err)
 274                 cb->errors = 1;
 275
 276         /* if there are more bios still pending for this compressed
 277          * extent, just exit
 278          */
 279         if (!atomic_dec_and_test(&cb->pending_bios))
 280                 goto out;
 281
 282         /* ok, we're the last bio for this extent, step one is to
 283          * call back into the FS and do all the end_io operations
 284          */
 285         inode = cb->inode;
 286         tree = &BTRFS_I(inode)->io_tree;
 287         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 288         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 289                                          cb->start,
 290                                          cb->start + cb->len - 1,
 291                                          NULL, 1);
 292         cb->compressed_pages[0]->mapping = NULL;
 293
 294         end_compressed_writeback(inode, cb->start, cb->len);
 295         /* note, our inode could be gone now */
 296
 297         /*
 298          * release the compressed pages, these came from alloc_page and
 299          * are not attached to the inode at all
 300          */
 301         index = 0;
 302         for (index = 0; index < cb->nr_pages; index++) {
 303                 page = cb->compressed_pages[index];
 304                 page->mapping = NULL;
 305                 page_cache_release(page);
 306         }
 307
 308         /* finally free the cb struct */
 309         kfree(cb->compressed_pages);
 310         kfree(cb);
 311 out:
 312         bio_put(bio);
 313 }
 314
 315 /*
 316  * worker function to build and submit bios for previously compressed pages.
 317  * The corresponding pages in the inode should be marked for writeback
 318  * and the compressed pages should have a reference on them for dropping
 319  * when the IO is complete.
 320  *
 321  * This also checksums the file bytes and gets things ready for
 322  * the end io hooks.
 323  */
 324 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 325                                  unsigned long len, u64 disk_start,
 326                                  unsigned long compressed_len,
 327                                  struct page **compressed_pages,
 328                                  unsigned long nr_pages)
 329 {
 330         struct bio *bio = NULL;
 331         struct btrfs_root *root = BTRFS_I(inode)->root;
 332         struct compressed_bio *cb;
 333         unsigned long bytes_left;
 334         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 335         int page_index = 0;
 336         struct page *page;
 337         u64 first_byte = disk_start;
 338         struct block_device *bdev;
 339         int ret;
 340
 341         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 342         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 343         atomic_set(&cb->pending_bios, 0);
 344         cb->errors = 0;
 345         cb->inode = inode;
 346         cb->start = start;
 347         cb->len = len;
 348         cb->mirror_num = 0;
 349         cb->compressed_pages = compressed_pages;
 350         cb->compressed_len = compressed_len;
 351         cb->orig_bio = NULL;
 352         cb->nr_pages = nr_pages;
 353
 354         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 355
 356         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 357         bio->bi_private = cb;
 358         bio->bi_end_io = end_compressed_bio_write;
 359         atomic_inc(&cb->pending_bios);
 360
 361         /* create and submit bios for the compressed pages */
 362         bytes_left = compressed_len;
 363         for (page_index = 0; page_index < cb->nr_pages; page_index++) {
 364                 page = compressed_pages[page_index];
 365                 page->mapping = inode->i_mapping;
 366                 if (bio->bi_size)
 367                         ret = io_tree->ops->merge_bio_hook(page, 0,
 368                                                            PAGE_CACHE_SIZE,
 369                                                            bio, 0);
 370                 else
 371                         ret = 0;
 372
 373                 page->mapping = NULL;
 374                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 375                     PAGE_CACHE_SIZE) {
 376                         bio_get(bio);
 377
 378                         /*
 379                          * inc the count before we submit the bio so
 380                          * we know the end IO handler won't happen before
 381                          * we inc the count.  Otherwise, the cb might get
 382                          * freed before we're done setting it up
 383                          */
 384                         atomic_inc(&cb->pending_bios);
 385                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 386                         BUG_ON(ret);
 387
 388                         ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 389                         BUG_ON(ret);
 390
 391                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 392                         BUG_ON(ret);
 393
 394                         bio_put(bio);
 395
 396                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 397                         bio->bi_private = cb;
 398                         bio->bi_end_io = end_compressed_bio_write;
 399                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 400                 }
 401                 if (bytes_left < PAGE_CACHE_SIZE) {
 402                         printk("bytes left %lu compress len %lu nr %lu\n",
 403                                bytes_left, cb->compressed_len, cb->nr_pages);
 404                 }
 405                 bytes_left -= PAGE_CACHE_SIZE;
 406                 first_byte += PAGE_CACHE_SIZE;
 407                 cond_resched();
 408         }
 409         bio_get(bio);
 410
 411         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 412         BUG_ON(ret);
 413
 414         ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 415         BUG_ON(ret);
 416
 417         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 418         BUG_ON(ret);
 419
 420         bio_put(bio);
 421         return 0;
 422 }
 423
 424 static noinline int add_ra_bio_pages(struct inode *inode,
 425                                      u64 compressed_end,
 426                                      struct compressed_bio *cb)
 427 {
 428         unsigned long end_index;
 429         unsigned long page_index;
 430         u64 last_offset;
 431         u64 isize = i_size_read(inode);
 432         int ret;
 433         struct page *page;
 434         unsigned long nr_pages = 0;
 435         struct extent_map *em;
 436         struct address_space *mapping = inode->i_mapping;
 437         struct extent_map_tree *em_tree;
 438         struct extent_io_tree *tree;
 439         u64 end;
 440         int misses = 0;
 441
 442         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 443         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 444         em_tree = &BTRFS_I(inode)->extent_tree;
 445         tree = &BTRFS_I(inode)->io_tree;
 446
 447         if (isize == 0)
 448                 return 0;
 449
 450         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 451
 452         while (last_offset < compressed_end) {
 453                 page_index = last_offset >> PAGE_CACHE_SHIFT;
 454
 455                 if (page_index > end_index)
 456                         break;
 457
 458                 rcu_read_lock();
 459                 page = radix_tree_lookup(&mapping->page_tree, page_index);
 460                 rcu_read_unlock();
 461                 if (page) {
 462                         misses++;
 463                         if (misses > 4)
 464                                 break;
 465                         goto next;
 466                 }
 467
 468                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 469                                                                 ~__GFP_FS);
 470                 if (!page)
 471                         break;
 472
 473                 if (add_to_page_cache_lru(page, mapping, page_index,
 474                                                                 GFP_NOFS)) {
 475                         page_cache_release(page);
 476                         goto next;
 477                 }
 478
 479                 end = last_offset + PAGE_CACHE_SIZE - 1;
 480                 /*
 481                  * at this point, we have a locked page in the page cache
 482                  * for these bytes in the file.  But, we have to make
 483                  * sure they map to this compressed extent on disk.
 484                  */
 485                 set_page_extent_mapped(page);
 486                 lock_extent(tree, last_offset, end, GFP_NOFS);
 487                 read_lock(&em_tree->lock);
 488                 em = lookup_extent_mapping(em_tree, last_offset,
 489                                            PAGE_CACHE_SIZE);
 490                 read_unlock(&em_tree->lock);
 491
 492                 if (!em || last_offset < em->start ||
 493                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 494                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 495                         free_extent_map(em);
 496                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 497                         unlock_page(page);
 498                         page_cache_release(page);
 499                         break;
 500                 }
 501                 free_extent_map(em);
 502
 503                 if (page->index == end_index) {
 504                         char *userpage;
 505                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 506
 507                         if (zero_offset) {
 508                                 int zeros;
 509                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 510                                 userpage = kmap_atomic(page, KM_USER0);
 511                                 memset(userpage + zero_offset, 0, zeros);
 512                                 flush_dcache_page(page);
 513                                 kunmap_atomic(userpage, KM_USER0);
 514                         }
 515                 }
 516
 517                 ret = bio_add_page(cb->orig_bio, page,
 518                                    PAGE_CACHE_SIZE, 0);
 519
 520                 if (ret == PAGE_CACHE_SIZE) {
 521                         nr_pages++;
 522                         page_cache_release(page);
 523                 } else {
 524                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 525                         unlock_page(page);
 526                         page_cache_release(page);
 527                         break;
 528                 }
 529 next:
 530                 last_offset += PAGE_CACHE_SIZE;
 531         }
 532         return 0;
 533 }
 534
 535 /*
 536  * for a compressed read, the bio we get passed has all the inode pages
 537  * in it.  We don't actually do IO on those pages but allocate new ones
 538  * to hold the compressed pages on disk.
 539  *
 540  * bio->bi_sector points to the compressed extent on disk
 541  * bio->bi_io_vec points to all of the inode pages
 542  * bio->bi_vcnt is a count of pages
 543  *
 544  * After the compressed pages are read, we copy the bytes into the
 545  * bio we were passed and then call the bio end_io calls
 546  */
 547 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 548                                  int mirror_num, unsigned long bio_flags)
 549 {
 550         struct extent_io_tree *tree;
 551         struct extent_map_tree *em_tree;
 552         struct compressed_bio *cb;
 553         struct btrfs_root *root = BTRFS_I(inode)->root;
 554         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 555         unsigned long compressed_len;
 556         unsigned long nr_pages;
 557         unsigned long page_index;
 558         struct page *page;
 559         struct block_device *bdev;
 560         struct bio *comp_bio;
 561         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 562         u64 em_len;
 563         u64 em_start;
 564         struct extent_map *em;
 565         int ret = -ENOMEM;
 566         u32 *sums;
 567
 568         tree = &BTRFS_I(inode)->io_tree;
 569         em_tree = &BTRFS_I(inode)->extent_tree;
 570
 571         /* we need the actual starting offset of this extent in the file */
 572         read_lock(&em_tree->lock);
 573         em = lookup_extent_mapping(em_tree,
 574                                    page_offset(bio->bi_io_vec->bv_page),
 575                                    PAGE_CACHE_SIZE);
 576         read_unlock(&em_tree->lock);
 577
 578         compressed_len = em->block_len;
 579         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 580         if (!cb)
 581                 goto out;
 582
 583         atomic_set(&cb->pending_bios, 0);
 584         cb->errors = 0;
 585         cb->inode = inode;
 586         cb->mirror_num = mirror_num;
 587         sums = &cb->sums;
 588
 589         cb->start = em->orig_start;
 590         em_len = em->len;
 591         em_start = em->start;
 592
 593         free_extent_map(em);
 594         em = NULL;
 595
 596         cb->len = uncompressed_len;
 597         cb->compressed_len = compressed_len;
 598         cb->compress_type = extent_compress_type(bio_flags);
 599         cb->orig_bio = bio;
 600
 601         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 602                                  PAGE_CACHE_SIZE;
 603         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 604                                        GFP_NOFS);
 605         if (!cb->compressed_pages)
 606                 goto fail1;
 607
 608         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 609
 610         for (page_index = 0; page_index < nr_pages; page_index++) {
 611                 cb->compressed_pages[page_index] = alloc_page(GFP_NOFS |
 612                                                               __GFP_HIGHMEM);
 613                 if (!cb->compressed_pages[page_index])
 614                         goto fail2;
 615         }
 616         cb->nr_pages = nr_pages;
 617
 618         add_ra_bio_pages(inode, em_start + em_len, cb);
 619
 620         /* include any pages we added in add_ra-bio_pages */
 621         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 622         cb->len = uncompressed_len;
 623
 624         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 625         if (!comp_bio)
 626                 goto fail2;
 627         comp_bio->bi_private = cb;
 628         comp_bio->bi_end_io = end_compressed_bio_read;
 629         atomic_inc(&cb->pending_bios);
 630
 631         for (page_index = 0; page_index < nr_pages; page_index++) {
 632                 page = cb->compressed_pages[page_index];
 633                 page->mapping = inode->i_mapping;
 634                 page->index = em_start >> PAGE_CACHE_SHIFT;
 635
 636                 if (comp_bio->bi_size)
 637                         ret = tree->ops->merge_bio_hook(page, 0,
 638                                                         PAGE_CACHE_SIZE,
 639                                                         comp_bio, 0);
 640                 else
 641                         ret = 0;
 642
 643                 page->mapping = NULL;
 644                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 645                     PAGE_CACHE_SIZE) {
 646                         bio_get(comp_bio);
 647
 648                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 649                         BUG_ON(ret);
 650
 651                         /*
 652                          * inc the count before we submit the bio so
 653                          * we know the end IO handler won't happen before
 654                          * we inc the count.  Otherwise, the cb might get
 655                          * freed before we're done setting it up
 656                          */
 657                         atomic_inc(&cb->pending_bios);
 658
 659                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 660                                 btrfs_lookup_bio_sums(root, inode, comp_bio,
 661                                                       sums);
 662                         }
 663                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 664                                 root->sectorsize;
 665
 666                         ret = btrfs_map_bio(root, READ, comp_bio,
 667                                             mirror_num, 0);
 668                         BUG_ON(ret);
 669
 670                         bio_put(comp_bio);
 671
 672                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 673                                                         GFP_NOFS);
 674                         comp_bio->bi_private = cb;
 675                         comp_bio->bi_end_io = end_compressed_bio_read;
 676
 677                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 678                 }
 679                 cur_disk_byte += PAGE_CACHE_SIZE;
 680         }
 681         bio_get(comp_bio);
 682
 683         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 684         BUG_ON(ret);
 685
 686         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM))
 687                 btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 688
 689         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 690         BUG_ON(ret);
 691
 692         bio_put(comp_bio);
 693         return 0;
 694
 695 fail2:
 696         for (page_index = 0; page_index < nr_pages; page_index++)
 697                 free_page((unsigned long)cb->compressed_pages[page_index]);
 698
 699         kfree(cb->compressed_pages);
 700 fail1:
 701         kfree(cb);
 702 out:
 703         free_extent_map(em);
 704         return ret;
 705 }
 706
 707 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 708 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 709 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 710 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 711 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 712
 713 struct btrfs_compress_op *btrfs_compress_op[] = {
 714         &btrfs_zlib_compress,
 715         &btrfs_lzo_compress,
 716 };
 717
 718 int __init btrfs_init_compress(void)
 719 {
 720         int i;
 721
 722         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 723                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 724                 spin_lock_init(&comp_workspace_lock[i]);
 725                 atomic_set(&comp_alloc_workspace[i], 0);
 726                 init_waitqueue_head(&comp_workspace_wait[i]);
 727         }
 728         return 0;
 729 }
 730
 731 /*
 732  * this finds an available workspace or allocates a new one
 733  * ERR_PTR is returned if things go bad.
 734  */
 735 static struct list_head *find_workspace(int type)
 736 {
 737         struct list_head *workspace;
 738         int cpus = num_online_cpus();
 739         int idx = type - 1;
 740
 741         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 742         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 743         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 744         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 745         int *num_workspace                      = &comp_num_workspace[idx];
 746 again:
 747         spin_lock(workspace_lock);
 748         if (!list_empty(idle_workspace)) {
 749                 workspace = idle_workspace->next;
 750                 list_del(workspace);
 751                 (*num_workspace)--;
 752                 spin_unlock(workspace_lock);
 753                 return workspace;
 754
 755         }
 756         if (atomic_read(alloc_workspace) > cpus) {
 757                 DEFINE_WAIT(wait);
 758
 759                 spin_unlock(workspace_lock);
 760                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 761                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 762                         schedule();
 763                 finish_wait(workspace_wait, &wait);
 764                 goto again;
 765         }
 766         atomic_inc(alloc_workspace);
 767         spin_unlock(workspace_lock);
 768
 769         workspace = btrfs_compress_op[idx]->alloc_workspace();
 770         if (IS_ERR(workspace)) {
 771                 atomic_dec(alloc_workspace);
 772                 wake_up(workspace_wait);
 773         }
 774         return workspace;
 775 }
 776
 777 /*
 778  * put a workspace struct back on the list or free it if we have enough
 779  * idle ones sitting around
 780  */
 781 static void free_workspace(int type, struct list_head *workspace)
 782 {
 783         int idx = type - 1;
 784         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 785         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 786         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 787         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 788         int *num_workspace                      = &comp_num_workspace[idx];
 789
 790         spin_lock(workspace_lock);
 791         if (*num_workspace < num_online_cpus()) {
 792                 list_add_tail(workspace, idle_workspace);
 793                 (*num_workspace)++;
 794                 spin_unlock(workspace_lock);
 795                 goto wake;
 796         }
 797         spin_unlock(workspace_lock);
 798
 799         btrfs_compress_op[idx]->free_workspace(workspace);
 800         atomic_dec(alloc_workspace);
 801 wake:
 802         if (waitqueue_active(workspace_wait))
 803                 wake_up(workspace_wait);
 804 }
 805
 806 /*
 807  * cleanup function for module exit
 808  */
 809 static void free_workspaces(void)
 810 {
 811         struct list_head *workspace;
 812         int i;
 813
 814         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 815                 while (!list_empty(&comp_idle_workspace[i])) {
 816                         workspace = comp_idle_workspace[i].next;
 817                         list_del(workspace);
 818                         btrfs_compress_op[i]->free_workspace(workspace);
 819                         atomic_dec(&comp_alloc_workspace[i]);
 820                 }
 821         }
 822 }
 823
 824 /*
 825  * given an address space and start/len, compress the bytes.
 826  *
 827  * pages are allocated to hold the compressed result and stored
 828  * in 'pages'
 829  *
 830  * out_pages is used to return the number of pages allocated.  There
 831  * may be pages allocated even if we return an error
 832  *
 833  * total_in is used to return the number of bytes actually read.  It
 834  * may be smaller then len if we had to exit early because we
 835  * ran out of room in the pages array or because we cross the
 836  * max_out threshold.
 837  *
 838  * total_out is used to return the total number of compressed bytes
 839  *
 840  * max_out tells us the max number of bytes that we're allowed to
 841  * stuff into pages
 842  */
 843 int btrfs_compress_pages(int type, struct address_space *mapping,
 844                          u64 start, unsigned long len,
 845                          struct page **pages,
 846                          unsigned long nr_dest_pages,
 847                          unsigned long *out_pages,
 848                          unsigned long *total_in,
 849                          unsigned long *total_out,
 850                          unsigned long max_out)
 851 {
 852         struct list_head *workspace;
 853         int ret;
 854
 855         workspace = find_workspace(type);
 856         if (IS_ERR(workspace))
 857                 return -1;
 858
 859         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 860                                                       start, len, pages,
 861                                                       nr_dest_pages, out_pages,
 862                                                       total_in, total_out,
 863                                                       max_out);
 864         free_workspace(type, workspace);
 865         return ret;
 866 }
 867
 868 /*
 869  * pages_in is an array of pages with compressed data.
 870  *
 871  * disk_start is the starting logical offset of this array in the file
 872  *
 873  * bvec is a bio_vec of pages from the file that we want to decompress into
 874  *
 875  * vcnt is the count of pages in the biovec
 876  *
 877  * srclen is the number of bytes in pages_in
 878  *
 879  * The basic idea is that we have a bio that was created by readpages.
 880  * The pages in the bio are for the uncompressed data, and they may not
 881  * be contiguous.  They all correspond to the range of bytes covered by
 882  * the compressed extent.
 883  */
 884 int btrfs_decompress_biovec(int type, struct page **pages_in, u64 disk_start,
 885                             struct bio_vec *bvec, int vcnt, size_t srclen)
 886 {
 887         struct list_head *workspace;
 888         int ret;
 889
 890         workspace = find_workspace(type);
 891         if (IS_ERR(workspace))
 892                 return -ENOMEM;
 893
 894         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 895                                                          disk_start,
 896                                                          bvec, vcnt, srclen);
 897         free_workspace(type, workspace);
 898         return ret;
 899 }
 900
 901 /*
 902  * a less complex decompression routine.  Our compressed data fits in a
 903  * single page, and we want to read a single page out of it.
 904  * start_byte tells us the offset into the compressed data we're interested in
 905  */
 906 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 907                      unsigned long start_byte, size_t srclen, size_t destlen)
 908 {
 909         struct list_head *workspace;
 910         int ret;
 911
 912         workspace = find_workspace(type);
 913         if (IS_ERR(workspace))
 914                 return -ENOMEM;
 915
 916         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 917                                                   dest_page, start_byte,
 918                                                   srclen, destlen);
 919
 920         free_workspace(type, workspace);
 921         return ret;
 922 }
 923
 924 void btrfs_exit_compress(void)
 925 {
 926         free_workspaces();
 927 }
 928
 929 /*
 930  * Copy uncompressed data from working buffer to pages.
 931  *
 932  * buf_start is the byte offset we're of the start of our workspace buffer.
 933  *
 934  * total_out is the last byte of the buffer
 935  */
 936 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 937                               unsigned long total_out, u64 disk_start,
 938                               struct bio_vec *bvec, int vcnt,
 939                               unsigned long *page_index,
 940                               unsigned long *pg_offset)
 941 {
 942         unsigned long buf_offset;
 943         unsigned long current_buf_start;
 944         unsigned long start_byte;
 945         unsigned long working_bytes = total_out - buf_start;
 946         unsigned long bytes;
 947         char *kaddr;
 948         struct page *page_out = bvec[*page_index].bv_page;
 949
 950         /*
 951          * start byte is the first byte of the page we're currently
 952          * copying into relative to the start of the compressed data.
 953          */
 954         start_byte = page_offset(page_out) - disk_start;
 955
 956         /* we haven't yet hit data corresponding to this page */
 957         if (total_out <= start_byte)
 958                 return 1;
 959
 960         /*
 961          * the start of the data we care about is offset into
 962          * the middle of our working buffer
 963          */
 964         if (total_out > start_byte && buf_start < start_byte) {
 965                 buf_offset = start_byte - buf_start;
 966                 working_bytes -= buf_offset;
 967         } else {
 968                 buf_offset = 0;
 969         }
 970         current_buf_start = buf_start;
 971
 972         /* copy bytes from the working buffer into the pages */
 973         while (working_bytes > 0) {
 974                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
 975                             PAGE_CACHE_SIZE - buf_offset);
 976                 bytes = min(bytes, working_bytes);
 977                 kaddr = kmap_atomic(page_out, KM_USER0);
 978                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
 979                 kunmap_atomic(kaddr, KM_USER0);
 980                 flush_dcache_page(page_out);
 981
 982                 *pg_offset += bytes;
 983                 buf_offset += bytes;
 984                 working_bytes -= bytes;
 985                 current_buf_start += bytes;
 986
 987                 /* check if we need to pick another page */
 988                 if (*pg_offset == PAGE_CACHE_SIZE) {
 989                         (*page_index)++;
 990                         if (*page_index >= vcnt)
 991                                 return 0;
 992
 993                         page_out = bvec[*page_index].bv_page;
 994                         *pg_offset = 0;
 995                         start_byte = page_offset(page_out) - disk_start;
 996
 997                         /*
 998                          * make sure our new page is covered by this
 999                          * working buffer
1000                          */
1001                         if (total_out <= start_byte)
1002                                 return 1;
1003
1004                         /*
1005                          * the next page in the biovec might not be adjacent
1006                          * to the last page, but it might still be found
1007                          * inside this working buffer. bump our offset pointer
1008                          */
1009                         if (total_out > start_byte &&
1010                             current_buf_start < start_byte) {
1011                                 buf_offset = start_byte - buf_start;
1012                                 working_bytes = total_out - start_byte;
1013                                 current_buf_start = buf_start + buf_offset;
1014                         }
1015                 }
1016         }
1017
1018         return 1;
1019 }