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 %llu "
 129                                "extent %llu csum %u "
 130                                "wanted %u mirror %d\n",
 131                                (unsigned long long)btrfs_ino(inode),
 132                                (unsigned long long)disk_start,
 133                                csum, *cb_sum, cb->mirror_num);
 134                         ret = -EIO;
 135                         goto fail;
 136                 }
 137                 cb_sum++;
 138
 139         }
 140         ret = 0;
 141 fail:
 142         return ret;
 143 }
 144
 145 /* when we finish reading compressed pages from the disk, we
 146  * decompress them and then run the bio end_io routines on the
 147  * decompressed pages (in the inode address space).
 148  *
 149  * This allows the checksumming and other IO error handling routines
 150  * to work normally
 151  *
 152  * The compressed pages are freed here, and it must be run
 153  * in process context
 154  */
 155 static void end_compressed_bio_read(struct bio *bio, int err)
 156 {
 157         struct compressed_bio *cb = bio->bi_private;
 158         struct inode *inode;
 159         struct page *page;
 160         unsigned long index;
 161         int ret;
 162
 163         if (err)
 164                 cb->errors = 1;
 165
 166         /* if there are more bios still pending for this compressed
 167          * extent, just exit
 168          */
 169         if (!atomic_dec_and_test(&cb->pending_bios))
 170                 goto out;
 171
 172         inode = cb->inode;
 173         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
 174         if (ret)
 175                 goto csum_failed;
 176
 177         /* ok, we're the last bio for this extent, lets start
 178          * the decompression.
 179          */
 180         ret = btrfs_decompress_biovec(cb->compress_type,
 181                                       cb->compressed_pages,
 182                                       cb->start,
 183                                       cb->orig_bio->bi_io_vec,
 184                                       cb->orig_bio->bi_vcnt,
 185                                       cb->compressed_len);
 186 csum_failed:
 187         if (ret)
 188                 cb->errors = 1;
 189
 190         /* release the compressed pages */
 191         index = 0;
 192         for (index = 0; index < cb->nr_pages; index++) {
 193                 page = cb->compressed_pages[index];
 194                 page->mapping = NULL;
 195                 page_cache_release(page);
 196         }
 197
 198         /* do io completion on the original bio */
 199         if (cb->errors) {
 200                 bio_io_error(cb->orig_bio);
 201         } else {
 202                 int bio_index = 0;
 203                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
 204
 205                 /*
 206                  * we have verified the checksum already, set page
 207                  * checked so the end_io handlers know about it
 208                  */
 209                 while (bio_index < cb->orig_bio->bi_vcnt) {
 210                         SetPageChecked(bvec->bv_page);
 211                         bvec++;
 212                         bio_index++;
 213                 }
 214                 bio_endio(cb->orig_bio, 0);
 215         }
 216
 217         /* finally free the cb struct */
 218         kfree(cb->compressed_pages);
 219         kfree(cb);
 220 out:
 221         bio_put(bio);
 222 }
 223
 224 /*
 225  * Clear the writeback bits on all of the file
 226  * pages for a compressed write
 227  */
 228 static noinline int end_compressed_writeback(struct inode *inode, u64 start,
 229                                              unsigned long ram_size)
 230 {
 231         unsigned long index = start >> PAGE_CACHE_SHIFT;
 232         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
 233         struct page *pages[16];
 234         unsigned long nr_pages = end_index - index + 1;
 235         int i;
 236         int ret;
 237
 238         while (nr_pages > 0) {
 239                 ret = find_get_pages_contig(inode->i_mapping, index,
 240                                      min_t(unsigned long,
 241                                      nr_pages, ARRAY_SIZE(pages)), pages);
 242                 if (ret == 0) {
 243                         nr_pages -= 1;
 244                         index += 1;
 245                         continue;
 246                 }
 247                 for (i = 0; i < ret; i++) {
 248                         end_page_writeback(pages[i]);
 249                         page_cache_release(pages[i]);
 250                 }
 251                 nr_pages -= ret;
 252                 index += ret;
 253         }
 254         /* the inode may be gone now */
 255         return 0;
 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
 342         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 343         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 344         if (!cb)
 345                 return -ENOMEM;
 346         atomic_set(&cb->pending_bios, 0);
 347         cb->errors = 0;
 348         cb->inode = inode;
 349         cb->start = start;
 350         cb->len = len;
 351         cb->mirror_num = 0;
 352         cb->compressed_pages = compressed_pages;
 353         cb->compressed_len = compressed_len;
 354         cb->orig_bio = NULL;
 355         cb->nr_pages = nr_pages;
 356
 357         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 358
 359         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 360         if(!bio) {
 361                 kfree(cb);
 362                 return -ENOMEM;
 363         }
 364         bio->bi_private = cb;
 365         bio->bi_end_io = end_compressed_bio_write;
 366         atomic_inc(&cb->pending_bios);
 367
 368         /* create and submit bios for the compressed pages */
 369         bytes_left = compressed_len;
 370         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 371                 page = compressed_pages[pg_index];
 372                 page->mapping = inode->i_mapping;
 373                 if (bio->bi_size)
 374                         ret = io_tree->ops->merge_bio_hook(page, 0,
 375                                                            PAGE_CACHE_SIZE,
 376                                                            bio, 0);
 377                 else
 378                         ret = 0;
 379
 380                 page->mapping = NULL;
 381                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 382                     PAGE_CACHE_SIZE) {
 383                         bio_get(bio);
 384
 385                         /*
 386                          * inc the count before we submit the bio so
 387                          * we know the end IO handler won't happen before
 388                          * we inc the count.  Otherwise, the cb might get
 389                          * freed before we're done setting it up
 390                          */
 391                         atomic_inc(&cb->pending_bios);
 392                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 393                         BUG_ON(ret);
 394
 395                         ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 396                         BUG_ON(ret);
 397
 398                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 399                         BUG_ON(ret);
 400
 401                         bio_put(bio);
 402
 403                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 404                         bio->bi_private = cb;
 405                         bio->bi_end_io = end_compressed_bio_write;
 406                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 407                 }
 408                 if (bytes_left < PAGE_CACHE_SIZE) {
 409                         printk("bytes left %lu compress len %lu nr %lu\n",
 410                                bytes_left, cb->compressed_len, cb->nr_pages);
 411                 }
 412                 bytes_left -= PAGE_CACHE_SIZE;
 413                 first_byte += PAGE_CACHE_SIZE;
 414                 cond_resched();
 415         }
 416         bio_get(bio);
 417
 418         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 419         BUG_ON(ret);
 420
 421         ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 422         BUG_ON(ret);
 423
 424         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 425         BUG_ON(ret);
 426
 427         bio_put(bio);
 428         return 0;
 429 }
 430
 431 static noinline int add_ra_bio_pages(struct inode *inode,
 432                                      u64 compressed_end,
 433                                      struct compressed_bio *cb)
 434 {
 435         unsigned long end_index;
 436         unsigned long pg_index;
 437         u64 last_offset;
 438         u64 isize = i_size_read(inode);
 439         int ret;
 440         struct page *page;
 441         unsigned long nr_pages = 0;
 442         struct extent_map *em;
 443         struct address_space *mapping = inode->i_mapping;
 444         struct extent_map_tree *em_tree;
 445         struct extent_io_tree *tree;
 446         u64 end;
 447         int misses = 0;
 448
 449         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 450         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 451         em_tree = &BTRFS_I(inode)->extent_tree;
 452         tree = &BTRFS_I(inode)->io_tree;
 453
 454         if (isize == 0)
 455                 return 0;
 456
 457         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 458
 459         while (last_offset < compressed_end) {
 460                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 461
 462                 if (pg_index > end_index)
 463                         break;
 464
 465                 rcu_read_lock();
 466                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 467                 rcu_read_unlock();
 468                 if (page) {
 469                         misses++;
 470                         if (misses > 4)
 471                                 break;
 472                         goto next;
 473                 }
 474
 475                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 476                                                                 ~__GFP_FS);
 477                 if (!page)
 478                         break;
 479
 480                 if (add_to_page_cache_lru(page, mapping, pg_index,
 481                                                                 GFP_NOFS)) {
 482                         page_cache_release(page);
 483                         goto next;
 484                 }
 485
 486                 end = last_offset + PAGE_CACHE_SIZE - 1;
 487                 /*
 488                  * at this point, we have a locked page in the page cache
 489                  * for these bytes in the file.  But, we have to make
 490                  * sure they map to this compressed extent on disk.
 491                  */
 492                 set_page_extent_mapped(page);
 493                 lock_extent(tree, last_offset, end, GFP_NOFS);
 494                 read_lock(&em_tree->lock);
 495                 em = lookup_extent_mapping(em_tree, last_offset,
 496                                            PAGE_CACHE_SIZE);
 497                 read_unlock(&em_tree->lock);
 498
 499                 if (!em || last_offset < em->start ||
 500                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 501                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 502                         free_extent_map(em);
 503                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 504                         unlock_page(page);
 505                         page_cache_release(page);
 506                         break;
 507                 }
 508                 free_extent_map(em);
 509
 510                 if (page->index == end_index) {
 511                         char *userpage;
 512                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 513
 514                         if (zero_offset) {
 515                                 int zeros;
 516                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 517                                 userpage = kmap_atomic(page, KM_USER0);
 518                                 memset(userpage + zero_offset, 0, zeros);
 519                                 flush_dcache_page(page);
 520                                 kunmap_atomic(userpage, KM_USER0);
 521                         }
 522                 }
 523
 524                 ret = bio_add_page(cb->orig_bio, page,
 525                                    PAGE_CACHE_SIZE, 0);
 526
 527                 if (ret == PAGE_CACHE_SIZE) {
 528                         nr_pages++;
 529                         page_cache_release(page);
 530                 } else {
 531                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 532                         unlock_page(page);
 533                         page_cache_release(page);
 534                         break;
 535                 }
 536 next:
 537                 last_offset += PAGE_CACHE_SIZE;
 538         }
 539         return 0;
 540 }
 541
 542 /*
 543  * for a compressed read, the bio we get passed has all the inode pages
 544  * in it.  We don't actually do IO on those pages but allocate new ones
 545  * to hold the compressed pages on disk.
 546  *
 547  * bio->bi_sector points to the compressed extent on disk
 548  * bio->bi_io_vec points to all of the inode pages
 549  * bio->bi_vcnt is a count of pages
 550  *
 551  * After the compressed pages are read, we copy the bytes into the
 552  * bio we were passed and then call the bio end_io calls
 553  */
 554 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 555                                  int mirror_num, unsigned long bio_flags)
 556 {
 557         struct extent_io_tree *tree;
 558         struct extent_map_tree *em_tree;
 559         struct compressed_bio *cb;
 560         struct btrfs_root *root = BTRFS_I(inode)->root;
 561         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 562         unsigned long compressed_len;
 563         unsigned long nr_pages;
 564         unsigned long pg_index;
 565         struct page *page;
 566         struct block_device *bdev;
 567         struct bio *comp_bio;
 568         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 569         u64 em_len;
 570         u64 em_start;
 571         struct extent_map *em;
 572         int ret = -ENOMEM;
 573         u32 *sums;
 574
 575         tree = &BTRFS_I(inode)->io_tree;
 576         em_tree = &BTRFS_I(inode)->extent_tree;
 577
 578         /* we need the actual starting offset of this extent in the file */
 579         read_lock(&em_tree->lock);
 580         em = lookup_extent_mapping(em_tree,
 581                                    page_offset(bio->bi_io_vec->bv_page),
 582                                    PAGE_CACHE_SIZE);
 583         read_unlock(&em_tree->lock);
 584
 585         compressed_len = em->block_len;
 586         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 587         if (!cb)
 588                 goto out;
 589
 590         atomic_set(&cb->pending_bios, 0);
 591         cb->errors = 0;
 592         cb->inode = inode;
 593         cb->mirror_num = mirror_num;
 594         sums = &cb->sums;
 595
 596         cb->start = em->orig_start;
 597         em_len = em->len;
 598         em_start = em->start;
 599
 600         free_extent_map(em);
 601         em = NULL;
 602
 603         cb->len = uncompressed_len;
 604         cb->compressed_len = compressed_len;
 605         cb->compress_type = extent_compress_type(bio_flags);
 606         cb->orig_bio = bio;
 607
 608         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 609                                  PAGE_CACHE_SIZE;
 610         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 611                                        GFP_NOFS);
 612         if (!cb->compressed_pages)
 613                 goto fail1;
 614
 615         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 616
 617         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 618                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 619                                                               __GFP_HIGHMEM);
 620                 if (!cb->compressed_pages[pg_index])
 621                         goto fail2;
 622         }
 623         cb->nr_pages = nr_pages;
 624
 625         add_ra_bio_pages(inode, em_start + em_len, cb);
 626
 627         /* include any pages we added in add_ra-bio_pages */
 628         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 629         cb->len = uncompressed_len;
 630
 631         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 632         if (!comp_bio)
 633                 goto fail2;
 634         comp_bio->bi_private = cb;
 635         comp_bio->bi_end_io = end_compressed_bio_read;
 636         atomic_inc(&cb->pending_bios);
 637
 638         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 639                 page = cb->compressed_pages[pg_index];
 640                 page->mapping = inode->i_mapping;
 641                 page->index = em_start >> PAGE_CACHE_SHIFT;
 642
 643                 if (comp_bio->bi_size)
 644                         ret = tree->ops->merge_bio_hook(page, 0,
 645                                                         PAGE_CACHE_SIZE,
 646                                                         comp_bio, 0);
 647                 else
 648                         ret = 0;
 649
 650                 page->mapping = NULL;
 651                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 652                     PAGE_CACHE_SIZE) {
 653                         bio_get(comp_bio);
 654
 655                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 656                         BUG_ON(ret);
 657
 658                         /*
 659                          * inc the count before we submit the bio so
 660                          * we know the end IO handler won't happen before
 661                          * we inc the count.  Otherwise, the cb might get
 662                          * freed before we're done setting it up
 663                          */
 664                         atomic_inc(&cb->pending_bios);
 665
 666                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 667                                 ret = btrfs_lookup_bio_sums(root, inode,
 668                                                         comp_bio, sums);
 669                                 BUG_ON(ret);
 670                         }
 671                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 672                                 root->sectorsize;
 673
 674                         ret = btrfs_map_bio(root, READ, comp_bio,
 675                                             mirror_num, 0);
 676                         BUG_ON(ret);
 677
 678                         bio_put(comp_bio);
 679
 680                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 681                                                         GFP_NOFS);
 682                         comp_bio->bi_private = cb;
 683                         comp_bio->bi_end_io = end_compressed_bio_read;
 684
 685                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 686                 }
 687                 cur_disk_byte += PAGE_CACHE_SIZE;
 688         }
 689         bio_get(comp_bio);
 690
 691         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 692         BUG_ON(ret);
 693
 694         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 695                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 696                 BUG_ON(ret);
 697         }
 698
 699         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 700         BUG_ON(ret);
 701
 702         bio_put(comp_bio);
 703         return 0;
 704
 705 fail2:
 706         for (pg_index = 0; pg_index < nr_pages; pg_index++)
 707                 free_page((unsigned long)cb->compressed_pages[pg_index]);
 708
 709         kfree(cb->compressed_pages);
 710 fail1:
 711         kfree(cb);
 712 out:
 713         free_extent_map(em);
 714         return ret;
 715 }
 716
 717 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 718 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 719 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 720 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 721 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 722
 723 struct btrfs_compress_op *btrfs_compress_op[] = {
 724         &btrfs_zlib_compress,
 725         &btrfs_lzo_compress,
 726 };
 727
 728 int __init btrfs_init_compress(void)
 729 {
 730         int i;
 731
 732         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 733                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 734                 spin_lock_init(&comp_workspace_lock[i]);
 735                 atomic_set(&comp_alloc_workspace[i], 0);
 736                 init_waitqueue_head(&comp_workspace_wait[i]);
 737         }
 738         return 0;
 739 }
 740
 741 /*
 742  * this finds an available workspace or allocates a new one
 743  * ERR_PTR is returned if things go bad.
 744  */
 745 static struct list_head *find_workspace(int type)
 746 {
 747         struct list_head *workspace;
 748         int cpus = num_online_cpus();
 749         int idx = type - 1;
 750
 751         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 752         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 753         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 754         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 755         int *num_workspace                      = &comp_num_workspace[idx];
 756 again:
 757         spin_lock(workspace_lock);
 758         if (!list_empty(idle_workspace)) {
 759                 workspace = idle_workspace->next;
 760                 list_del(workspace);
 761                 (*num_workspace)--;
 762                 spin_unlock(workspace_lock);
 763                 return workspace;
 764
 765         }
 766         if (atomic_read(alloc_workspace) > cpus) {
 767                 DEFINE_WAIT(wait);
 768
 769                 spin_unlock(workspace_lock);
 770                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 771                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 772                         schedule();
 773                 finish_wait(workspace_wait, &wait);
 774                 goto again;
 775         }
 776         atomic_inc(alloc_workspace);
 777         spin_unlock(workspace_lock);
 778
 779         workspace = btrfs_compress_op[idx]->alloc_workspace();
 780         if (IS_ERR(workspace)) {
 781                 atomic_dec(alloc_workspace);
 782                 wake_up(workspace_wait);
 783         }
 784         return workspace;
 785 }
 786
 787 /*
 788  * put a workspace struct back on the list or free it if we have enough
 789  * idle ones sitting around
 790  */
 791 static void free_workspace(int type, struct list_head *workspace)
 792 {
 793         int idx = type - 1;
 794         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 795         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 796         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 797         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 798         int *num_workspace                      = &comp_num_workspace[idx];
 799
 800         spin_lock(workspace_lock);
 801         if (*num_workspace < num_online_cpus()) {
 802                 list_add_tail(workspace, idle_workspace);
 803                 (*num_workspace)++;
 804                 spin_unlock(workspace_lock);
 805                 goto wake;
 806         }
 807         spin_unlock(workspace_lock);
 808
 809         btrfs_compress_op[idx]->free_workspace(workspace);
 810         atomic_dec(alloc_workspace);
 811 wake:
 812         if (waitqueue_active(workspace_wait))
 813                 wake_up(workspace_wait);
 814 }
 815
 816 /*
 817  * cleanup function for module exit
 818  */
 819 static void free_workspaces(void)
 820 {
 821         struct list_head *workspace;
 822         int i;
 823
 824         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 825                 while (!list_empty(&comp_idle_workspace[i])) {
 826                         workspace = comp_idle_workspace[i].next;
 827                         list_del(workspace);
 828                         btrfs_compress_op[i]->free_workspace(workspace);
 829                         atomic_dec(&comp_alloc_workspace[i]);
 830                 }
 831         }
 832 }
 833
 834 /*
 835  * given an address space and start/len, compress the bytes.
 836  *
 837  * pages are allocated to hold the compressed result and stored
 838  * in 'pages'
 839  *
 840  * out_pages is used to return the number of pages allocated.  There
 841  * may be pages allocated even if we return an error
 842  *
 843  * total_in is used to return the number of bytes actually read.  It
 844  * may be smaller then len if we had to exit early because we
 845  * ran out of room in the pages array or because we cross the
 846  * max_out threshold.
 847  *
 848  * total_out is used to return the total number of compressed bytes
 849  *
 850  * max_out tells us the max number of bytes that we're allowed to
 851  * stuff into pages
 852  */
 853 int btrfs_compress_pages(int type, struct address_space *mapping,
 854                          u64 start, unsigned long len,
 855                          struct page **pages,
 856                          unsigned long nr_dest_pages,
 857                          unsigned long *out_pages,
 858                          unsigned long *total_in,
 859                          unsigned long *total_out,
 860                          unsigned long max_out)
 861 {
 862         struct list_head *workspace;
 863         int ret;
 864
 865         workspace = find_workspace(type);
 866         if (IS_ERR(workspace))
 867                 return -1;
 868
 869         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 870                                                       start, len, pages,
 871                                                       nr_dest_pages, out_pages,
 872                                                       total_in, total_out,
 873                                                       max_out);
 874         free_workspace(type, workspace);
 875         return ret;
 876 }
 877
 878 /*
 879  * pages_in is an array of pages with compressed data.
 880  *
 881  * disk_start is the starting logical offset of this array in the file
 882  *
 883  * bvec is a bio_vec of pages from the file that we want to decompress into
 884  *
 885  * vcnt is the count of pages in the biovec
 886  *
 887  * srclen is the number of bytes in pages_in
 888  *
 889  * The basic idea is that we have a bio that was created by readpages.
 890  * The pages in the bio are for the uncompressed data, and they may not
 891  * be contiguous.  They all correspond to the range of bytes covered by
 892  * the compressed extent.
 893  */
 894 int btrfs_decompress_biovec(int type, struct page **pages_in, u64 disk_start,
 895                             struct bio_vec *bvec, int vcnt, size_t srclen)
 896 {
 897         struct list_head *workspace;
 898         int ret;
 899
 900         workspace = find_workspace(type);
 901         if (IS_ERR(workspace))
 902                 return -ENOMEM;
 903
 904         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 905                                                          disk_start,
 906                                                          bvec, vcnt, srclen);
 907         free_workspace(type, workspace);
 908         return ret;
 909 }
 910
 911 /*
 912  * a less complex decompression routine.  Our compressed data fits in a
 913  * single page, and we want to read a single page out of it.
 914  * start_byte tells us the offset into the compressed data we're interested in
 915  */
 916 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 917                      unsigned long start_byte, size_t srclen, size_t destlen)
 918 {
 919         struct list_head *workspace;
 920         int ret;
 921
 922         workspace = find_workspace(type);
 923         if (IS_ERR(workspace))
 924                 return -ENOMEM;
 925
 926         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 927                                                   dest_page, start_byte,
 928                                                   srclen, destlen);
 929
 930         free_workspace(type, workspace);
 931         return ret;
 932 }
 933
 934 void btrfs_exit_compress(void)
 935 {
 936         free_workspaces();
 937 }
 938
 939 /*
 940  * Copy uncompressed data from working buffer to pages.
 941  *
 942  * buf_start is the byte offset we're of the start of our workspace buffer.
 943  *
 944  * total_out is the last byte of the buffer
 945  */
 946 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 947                               unsigned long total_out, u64 disk_start,
 948                               struct bio_vec *bvec, int vcnt,
 949                               unsigned long *pg_index,
 950                               unsigned long *pg_offset)
 951 {
 952         unsigned long buf_offset;
 953         unsigned long current_buf_start;
 954         unsigned long start_byte;
 955         unsigned long working_bytes = total_out - buf_start;
 956         unsigned long bytes;
 957         char *kaddr;
 958         struct page *page_out = bvec[*pg_index].bv_page;
 959
 960         /*
 961          * start byte is the first byte of the page we're currently
 962          * copying into relative to the start of the compressed data.
 963          */
 964         start_byte = page_offset(page_out) - disk_start;
 965
 966         /* we haven't yet hit data corresponding to this page */
 967         if (total_out <= start_byte)
 968                 return 1;
 969
 970         /*
 971          * the start of the data we care about is offset into
 972          * the middle of our working buffer
 973          */
 974         if (total_out > start_byte && buf_start < start_byte) {
 975                 buf_offset = start_byte - buf_start;
 976                 working_bytes -= buf_offset;
 977         } else {
 978                 buf_offset = 0;
 979         }
 980         current_buf_start = buf_start;
 981
 982         /* copy bytes from the working buffer into the pages */
 983         while (working_bytes > 0) {
 984                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
 985                             PAGE_CACHE_SIZE - buf_offset);
 986                 bytes = min(bytes, working_bytes);
 987                 kaddr = kmap_atomic(page_out, KM_USER0);
 988                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
 989                 kunmap_atomic(kaddr, KM_USER0);
 990                 flush_dcache_page(page_out);
 991
 992                 *pg_offset += bytes;
 993                 buf_offset += bytes;
 994                 working_bytes -= bytes;
 995                 current_buf_start += bytes;
 996
 997                 /* check if we need to pick another page */
 998                 if (*pg_offset == PAGE_CACHE_SIZE) {
 999                         (*pg_index)++;
1000                         if (*pg_index >= vcnt)
1001                                 return 0;
1002
1003                         page_out = bvec[*pg_index].bv_page;
1004                         *pg_offset = 0;
1005                         start_byte = page_offset(page_out) - disk_start;
1006
1007                         /*
1008                          * make sure our new page is covered by this
1009                          * working buffer
1010                          */
1011                         if (total_out <= start_byte)
1012                                 return 1;
1013
1014                         /*
1015                          * the next page in the biovec might not be adjacent
1016                          * to the last page, but it might still be found
1017                          * inside this working buffer. bump our offset pointer
1018                          */
1019                         if (total_out > start_byte &&
1020                             current_buf_start < start_byte) {
1021                                 buf_offset = start_byte - buf_start;
1022                                 working_bytes = total_out - start_byte;
1023                                 current_buf_start = buf_start + buf_offset;
1024                         }
1025                 }
1026         }
1027
1028         return 1;
1029 }