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 int btrfs_decompress_biovec(int type, struct page **pages_in,
  86                                    u64 disk_start, struct bio_vec *bvec,
  87                                    int vcnt, size_t srclen);
  88
  89 static inline int compressed_bio_size(struct btrfs_root *root,
  90                                       unsigned long disk_size)
  91 {
  92         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
  93
  94         return sizeof(struct compressed_bio) +
  95                 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
  96                 csum_size;
  97 }
  98
  99 static struct bio *compressed_bio_alloc(struct block_device *bdev,
 100                                         u64 first_byte, gfp_t gfp_flags)
 101 {
 102         int nr_vecs;
 103
 104         nr_vecs = bio_get_nr_vecs(bdev);
 105         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
 106 }
 107
 108 static int check_compressed_csum(struct inode *inode,
 109                                  struct compressed_bio *cb,
 110                                  u64 disk_start)
 111 {
 112         int ret;
 113         struct page *page;
 114         unsigned long i;
 115         char *kaddr;
 116         u32 csum;
 117         u32 *cb_sum = &cb->sums;
 118
 119         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 120                 return 0;
 121
 122         for (i = 0; i < cb->nr_pages; i++) {
 123                 page = cb->compressed_pages[i];
 124                 csum = ~(u32)0;
 125
 126                 kaddr = kmap_atomic(page);
 127                 csum = btrfs_csum_data(kaddr, csum, PAGE_CACHE_SIZE);
 128                 btrfs_csum_final(csum, (char *)&csum);
 129                 kunmap_atomic(kaddr);
 130
 131                 if (csum != *cb_sum) {
 132                         printk(KERN_INFO "btrfs csum failed ino %llu "
 133                                "extent %llu csum %u "
 134                                "wanted %u mirror %d\n",
 135                                btrfs_ino(inode), disk_start, csum, *cb_sum,
 136                                cb->mirror_num);
 137                         ret = -EIO;
 138                         goto fail;
 139                 }
 140                 cb_sum++;
 141
 142         }
 143         ret = 0;
 144 fail:
 145         return ret;
 146 }
 147
 148 /* when we finish reading compressed pages from the disk, we
 149  * decompress them and then run the bio end_io routines on the
 150  * decompressed pages (in the inode address space).
 151  *
 152  * This allows the checksumming and other IO error handling routines
 153  * to work normally
 154  *
 155  * The compressed pages are freed here, and it must be run
 156  * in process context
 157  */
 158 static void end_compressed_bio_read(struct bio *bio, int err)
 159 {
 160         struct compressed_bio *cb = bio->bi_private;
 161         struct inode *inode;
 162         struct page *page;
 163         unsigned long index;
 164         int ret;
 165
 166         if (err)
 167                 cb->errors = 1;
 168
 169         /* if there are more bios still pending for this compressed
 170          * extent, just exit
 171          */
 172         if (!atomic_dec_and_test(&cb->pending_bios))
 173                 goto out;
 174
 175         inode = cb->inode;
 176         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
 177         if (ret)
 178                 goto csum_failed;
 179
 180         /* ok, we're the last bio for this extent, lets start
 181          * the decompression.
 182          */
 183         ret = btrfs_decompress_biovec(cb->compress_type,
 184                                       cb->compressed_pages,
 185                                       cb->start,
 186                                       cb->orig_bio->bi_io_vec,
 187                                       cb->orig_bio->bi_vcnt,
 188                                       cb->compressed_len);
 189 csum_failed:
 190         if (ret)
 191                 cb->errors = 1;
 192
 193         /* release the compressed pages */
 194         index = 0;
 195         for (index = 0; index < cb->nr_pages; index++) {
 196                 page = cb->compressed_pages[index];
 197                 page->mapping = NULL;
 198                 page_cache_release(page);
 199         }
 200
 201         /* do io completion on the original bio */
 202         if (cb->errors) {
 203                 bio_io_error(cb->orig_bio);
 204         } else {
 205                 int bio_index = 0;
 206                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
 207
 208                 /*
 209                  * we have verified the checksum already, set page
 210                  * checked so the end_io handlers know about it
 211                  */
 212                 while (bio_index < cb->orig_bio->bi_vcnt) {
 213                         SetPageChecked(bvec->bv_page);
 214                         bvec++;
 215                         bio_index++;
 216                 }
 217                 bio_endio(cb->orig_bio, 0);
 218         }
 219
 220         /* finally free the cb struct */
 221         kfree(cb->compressed_pages);
 222         kfree(cb);
 223 out:
 224         bio_put(bio);
 225 }
 226
 227 /*
 228  * Clear the writeback bits on all of the file
 229  * pages for a compressed write
 230  */
 231 static noinline void end_compressed_writeback(struct inode *inode, u64 start,
 232                                               unsigned long ram_size)
 233 {
 234         unsigned long index = start >> PAGE_CACHE_SHIFT;
 235         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
 236         struct page *pages[16];
 237         unsigned long nr_pages = end_index - index + 1;
 238         int i;
 239         int ret;
 240
 241         while (nr_pages > 0) {
 242                 ret = find_get_pages_contig(inode->i_mapping, index,
 243                                      min_t(unsigned long,
 244                                      nr_pages, ARRAY_SIZE(pages)), pages);
 245                 if (ret == 0) {
 246                         nr_pages -= 1;
 247                         index += 1;
 248                         continue;
 249                 }
 250                 for (i = 0; i < ret; i++) {
 251                         end_page_writeback(pages[i]);
 252                         page_cache_release(pages[i]);
 253                 }
 254                 nr_pages -= ret;
 255                 index += ret;
 256         }
 257         /* the inode may be gone now */
 258 }
 259
 260 /*
 261  * do the cleanup once all the compressed pages hit the disk.
 262  * This will clear writeback on the file pages and free the compressed
 263  * pages.
 264  *
 265  * This also calls the writeback end hooks for the file pages so that
 266  * metadata and checksums can be updated in the file.
 267  */
 268 static void end_compressed_bio_write(struct bio *bio, int err)
 269 {
 270         struct extent_io_tree *tree;
 271         struct compressed_bio *cb = bio->bi_private;
 272         struct inode *inode;
 273         struct page *page;
 274         unsigned long index;
 275
 276         if (err)
 277                 cb->errors = 1;
 278
 279         /* if there are more bios still pending for this compressed
 280          * extent, just exit
 281          */
 282         if (!atomic_dec_and_test(&cb->pending_bios))
 283                 goto out;
 284
 285         /* ok, we're the last bio for this extent, step one is to
 286          * call back into the FS and do all the end_io operations
 287          */
 288         inode = cb->inode;
 289         tree = &BTRFS_I(inode)->io_tree;
 290         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 291         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 292                                          cb->start,
 293                                          cb->start + cb->len - 1,
 294                                          NULL, 1);
 295         cb->compressed_pages[0]->mapping = NULL;
 296
 297         end_compressed_writeback(inode, cb->start, cb->len);
 298         /* note, our inode could be gone now */
 299
 300         /*
 301          * release the compressed pages, these came from alloc_page and
 302          * are not attached to the inode at all
 303          */
 304         index = 0;
 305         for (index = 0; index < cb->nr_pages; index++) {
 306                 page = cb->compressed_pages[index];
 307                 page->mapping = NULL;
 308                 page_cache_release(page);
 309         }
 310
 311         /* finally free the cb struct */
 312         kfree(cb->compressed_pages);
 313         kfree(cb);
 314 out:
 315         bio_put(bio);
 316 }
 317
 318 /*
 319  * worker function to build and submit bios for previously compressed pages.
 320  * The corresponding pages in the inode should be marked for writeback
 321  * and the compressed pages should have a reference on them for dropping
 322  * when the IO is complete.
 323  *
 324  * This also checksums the file bytes and gets things ready for
 325  * the end io hooks.
 326  */
 327 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 328                                  unsigned long len, u64 disk_start,
 329                                  unsigned long compressed_len,
 330                                  struct page **compressed_pages,
 331                                  unsigned long nr_pages)
 332 {
 333         struct bio *bio = NULL;
 334         struct btrfs_root *root = BTRFS_I(inode)->root;
 335         struct compressed_bio *cb;
 336         unsigned long bytes_left;
 337         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 338         int pg_index = 0;
 339         struct page *page;
 340         u64 first_byte = disk_start;
 341         struct block_device *bdev;
 342         int ret;
 343         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 344
 345         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 346         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 347         if (!cb)
 348                 return -ENOMEM;
 349         atomic_set(&cb->pending_bios, 0);
 350         cb->errors = 0;
 351         cb->inode = inode;
 352         cb->start = start;
 353         cb->len = len;
 354         cb->mirror_num = 0;
 355         cb->compressed_pages = compressed_pages;
 356         cb->compressed_len = compressed_len;
 357         cb->orig_bio = NULL;
 358         cb->nr_pages = nr_pages;
 359
 360         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 361
 362         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 363         if(!bio) {
 364                 kfree(cb);
 365                 return -ENOMEM;
 366         }
 367         bio->bi_private = cb;
 368         bio->bi_end_io = end_compressed_bio_write;
 369         atomic_inc(&cb->pending_bios);
 370
 371         /* create and submit bios for the compressed pages */
 372         bytes_left = compressed_len;
 373         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 374                 page = compressed_pages[pg_index];
 375                 page->mapping = inode->i_mapping;
 376                 if (bio->bi_size)
 377                         ret = io_tree->ops->merge_bio_hook(WRITE, page, 0,
 378                                                            PAGE_CACHE_SIZE,
 379                                                            bio, 0);
 380                 else
 381                         ret = 0;
 382
 383                 page->mapping = NULL;
 384                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 385                     PAGE_CACHE_SIZE) {
 386                         bio_get(bio);
 387
 388                         /*
 389                          * inc the count before we submit the bio so
 390                          * we know the end IO handler won't happen before
 391                          * we inc the count.  Otherwise, the cb might get
 392                          * freed before we're done setting it up
 393                          */
 394                         atomic_inc(&cb->pending_bios);
 395                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 396                         BUG_ON(ret); /* -ENOMEM */
 397
 398                         if (!skip_sum) {
 399                                 ret = btrfs_csum_one_bio(root, inode, bio,
 400                                                          start, 1);
 401                                 BUG_ON(ret); /* -ENOMEM */
 402                         }
 403
 404                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 405                         BUG_ON(ret); /* -ENOMEM */
 406
 407                         bio_put(bio);
 408
 409                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 410                         BUG_ON(!bio);
 411                         bio->bi_private = cb;
 412                         bio->bi_end_io = end_compressed_bio_write;
 413                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 414                 }
 415                 if (bytes_left < PAGE_CACHE_SIZE) {
 416                         printk("bytes left %lu compress len %lu nr %lu\n",
 417                                bytes_left, cb->compressed_len, cb->nr_pages);
 418                 }
 419                 bytes_left -= PAGE_CACHE_SIZE;
 420                 first_byte += PAGE_CACHE_SIZE;
 421                 cond_resched();
 422         }
 423         bio_get(bio);
 424
 425         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 426         BUG_ON(ret); /* -ENOMEM */
 427
 428         if (!skip_sum) {
 429                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 430                 BUG_ON(ret); /* -ENOMEM */
 431         }
 432
 433         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 434         BUG_ON(ret); /* -ENOMEM */
 435
 436         bio_put(bio);
 437         return 0;
 438 }
 439
 440 static noinline int add_ra_bio_pages(struct inode *inode,
 441                                      u64 compressed_end,
 442                                      struct compressed_bio *cb)
 443 {
 444         unsigned long end_index;
 445         unsigned long pg_index;
 446         u64 last_offset;
 447         u64 isize = i_size_read(inode);
 448         int ret;
 449         struct page *page;
 450         unsigned long nr_pages = 0;
 451         struct extent_map *em;
 452         struct address_space *mapping = inode->i_mapping;
 453         struct extent_map_tree *em_tree;
 454         struct extent_io_tree *tree;
 455         u64 end;
 456         int misses = 0;
 457
 458         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 459         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 460         em_tree = &BTRFS_I(inode)->extent_tree;
 461         tree = &BTRFS_I(inode)->io_tree;
 462
 463         if (isize == 0)
 464                 return 0;
 465
 466         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 467
 468         while (last_offset < compressed_end) {
 469                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 470
 471                 if (pg_index > end_index)
 472                         break;
 473
 474                 rcu_read_lock();
 475                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 476                 rcu_read_unlock();
 477                 if (page) {
 478                         misses++;
 479                         if (misses > 4)
 480                                 break;
 481                         goto next;
 482                 }
 483
 484                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 485                                                                 ~__GFP_FS);
 486                 if (!page)
 487                         break;
 488
 489                 if (add_to_page_cache_lru(page, mapping, pg_index,
 490                                                                 GFP_NOFS)) {
 491                         page_cache_release(page);
 492                         goto next;
 493                 }
 494
 495                 end = last_offset + PAGE_CACHE_SIZE - 1;
 496                 /*
 497                  * at this point, we have a locked page in the page cache
 498                  * for these bytes in the file.  But, we have to make
 499                  * sure they map to this compressed extent on disk.
 500                  */
 501                 set_page_extent_mapped(page);
 502                 lock_extent(tree, last_offset, end);
 503                 read_lock(&em_tree->lock);
 504                 em = lookup_extent_mapping(em_tree, last_offset,
 505                                            PAGE_CACHE_SIZE);
 506                 read_unlock(&em_tree->lock);
 507
 508                 if (!em || last_offset < em->start ||
 509                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 510                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 511                         free_extent_map(em);
 512                         unlock_extent(tree, last_offset, end);
 513                         unlock_page(page);
 514                         page_cache_release(page);
 515                         break;
 516                 }
 517                 free_extent_map(em);
 518
 519                 if (page->index == end_index) {
 520                         char *userpage;
 521                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 522
 523                         if (zero_offset) {
 524                                 int zeros;
 525                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 526                                 userpage = kmap_atomic(page);
 527                                 memset(userpage + zero_offset, 0, zeros);
 528                                 flush_dcache_page(page);
 529                                 kunmap_atomic(userpage);
 530                         }
 531                 }
 532
 533                 ret = bio_add_page(cb->orig_bio, page,
 534                                    PAGE_CACHE_SIZE, 0);
 535
 536                 if (ret == PAGE_CACHE_SIZE) {
 537                         nr_pages++;
 538                         page_cache_release(page);
 539                 } else {
 540                         unlock_extent(tree, last_offset, end);
 541                         unlock_page(page);
 542                         page_cache_release(page);
 543                         break;
 544                 }
 545 next:
 546                 last_offset += PAGE_CACHE_SIZE;
 547         }
 548         return 0;
 549 }
 550
 551 /*
 552  * for a compressed read, the bio we get passed has all the inode pages
 553  * in it.  We don't actually do IO on those pages but allocate new ones
 554  * to hold the compressed pages on disk.
 555  *
 556  * bio->bi_sector points to the compressed extent on disk
 557  * bio->bi_io_vec points to all of the inode pages
 558  * bio->bi_vcnt is a count of pages
 559  *
 560  * After the compressed pages are read, we copy the bytes into the
 561  * bio we were passed and then call the bio end_io calls
 562  */
 563 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 564                                  int mirror_num, unsigned long bio_flags)
 565 {
 566         struct extent_io_tree *tree;
 567         struct extent_map_tree *em_tree;
 568         struct compressed_bio *cb;
 569         struct btrfs_root *root = BTRFS_I(inode)->root;
 570         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 571         unsigned long compressed_len;
 572         unsigned long nr_pages;
 573         unsigned long pg_index;
 574         struct page *page;
 575         struct block_device *bdev;
 576         struct bio *comp_bio;
 577         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 578         u64 em_len;
 579         u64 em_start;
 580         struct extent_map *em;
 581         int ret = -ENOMEM;
 582         int faili = 0;
 583         u32 *sums;
 584
 585         tree = &BTRFS_I(inode)->io_tree;
 586         em_tree = &BTRFS_I(inode)->extent_tree;
 587
 588         /* we need the actual starting offset of this extent in the file */
 589         read_lock(&em_tree->lock);
 590         em = lookup_extent_mapping(em_tree,
 591                                    page_offset(bio->bi_io_vec->bv_page),
 592                                    PAGE_CACHE_SIZE);
 593         read_unlock(&em_tree->lock);
 594         if (!em)
 595                 return -EIO;
 596
 597         compressed_len = em->block_len;
 598         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 599         if (!cb)
 600                 goto out;
 601
 602         atomic_set(&cb->pending_bios, 0);
 603         cb->errors = 0;
 604         cb->inode = inode;
 605         cb->mirror_num = mirror_num;
 606         sums = &cb->sums;
 607
 608         cb->start = em->orig_start;
 609         em_len = em->len;
 610         em_start = em->start;
 611
 612         free_extent_map(em);
 613         em = NULL;
 614
 615         cb->len = uncompressed_len;
 616         cb->compressed_len = compressed_len;
 617         cb->compress_type = extent_compress_type(bio_flags);
 618         cb->orig_bio = bio;
 619
 620         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 621                                  PAGE_CACHE_SIZE;
 622         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 623                                        GFP_NOFS);
 624         if (!cb->compressed_pages)
 625                 goto fail1;
 626
 627         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 628
 629         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 630                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 631                                                               __GFP_HIGHMEM);
 632                 if (!cb->compressed_pages[pg_index]) {
 633                         faili = pg_index - 1;
 634                         ret = -ENOMEM;
 635                         goto fail2;
 636                 }
 637         }
 638         faili = nr_pages - 1;
 639         cb->nr_pages = nr_pages;
 640
 641         /* In the parent-locked case, we only locked the range we are
 642          * interested in.  In all other cases, we can opportunistically
 643          * cache decompressed data that goes beyond the requested range. */
 644         if (!(bio_flags & EXTENT_BIO_PARENT_LOCKED))
 645                 add_ra_bio_pages(inode, em_start + em_len, cb);
 646
 647         /* include any pages we added in add_ra-bio_pages */
 648         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 649         cb->len = uncompressed_len;
 650
 651         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 652         if (!comp_bio)
 653                 goto fail2;
 654         comp_bio->bi_private = cb;
 655         comp_bio->bi_end_io = end_compressed_bio_read;
 656         atomic_inc(&cb->pending_bios);
 657
 658         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 659                 page = cb->compressed_pages[pg_index];
 660                 page->mapping = inode->i_mapping;
 661                 page->index = em_start >> PAGE_CACHE_SHIFT;
 662
 663                 if (comp_bio->bi_size)
 664                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 665                                                         PAGE_CACHE_SIZE,
 666                                                         comp_bio, 0);
 667                 else
 668                         ret = 0;
 669
 670                 page->mapping = NULL;
 671                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 672                     PAGE_CACHE_SIZE) {
 673                         bio_get(comp_bio);
 674
 675                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 676                         BUG_ON(ret); /* -ENOMEM */
 677
 678                         /*
 679                          * inc the count before we submit the bio so
 680                          * we know the end IO handler won't happen before
 681                          * we inc the count.  Otherwise, the cb might get
 682                          * freed before we're done setting it up
 683                          */
 684                         atomic_inc(&cb->pending_bios);
 685
 686                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 687                                 ret = btrfs_lookup_bio_sums(root, inode,
 688                                                         comp_bio, sums);
 689                                 BUG_ON(ret); /* -ENOMEM */
 690                         }
 691                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 692                                 root->sectorsize;
 693
 694                         ret = btrfs_map_bio(root, READ, comp_bio,
 695                                             mirror_num, 0);
 696                         if (ret)
 697                                 bio_endio(comp_bio, ret);
 698
 699                         bio_put(comp_bio);
 700
 701                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 702                                                         GFP_NOFS);
 703                         BUG_ON(!comp_bio);
 704                         comp_bio->bi_private = cb;
 705                         comp_bio->bi_end_io = end_compressed_bio_read;
 706
 707                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 708                 }
 709                 cur_disk_byte += PAGE_CACHE_SIZE;
 710         }
 711         bio_get(comp_bio);
 712
 713         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 714         BUG_ON(ret); /* -ENOMEM */
 715
 716         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 717                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 718                 BUG_ON(ret); /* -ENOMEM */
 719         }
 720
 721         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 722         if (ret)
 723                 bio_endio(comp_bio, ret);
 724
 725         bio_put(comp_bio);
 726         return 0;
 727
 728 fail2:
 729         while (faili >= 0) {
 730                 __free_page(cb->compressed_pages[faili]);
 731                 faili--;
 732         }
 733
 734         kfree(cb->compressed_pages);
 735 fail1:
 736         kfree(cb);
 737 out:
 738         free_extent_map(em);
 739         return ret;
 740 }
 741
 742 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 743 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 744 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 745 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 746 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 747
 748 static struct btrfs_compress_op *btrfs_compress_op[] = {
 749         &btrfs_zlib_compress,
 750         &btrfs_lzo_compress,
 751 };
 752
 753 void __init btrfs_init_compress(void)
 754 {
 755         int i;
 756
 757         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 758                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 759                 spin_lock_init(&comp_workspace_lock[i]);
 760                 atomic_set(&comp_alloc_workspace[i], 0);
 761                 init_waitqueue_head(&comp_workspace_wait[i]);
 762         }
 763 }
 764
 765 /*
 766  * this finds an available workspace or allocates a new one
 767  * ERR_PTR is returned if things go bad.
 768  */
 769 static struct list_head *find_workspace(int type)
 770 {
 771         struct list_head *workspace;
 772         int cpus = num_online_cpus();
 773         int idx = type - 1;
 774
 775         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 776         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 777         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 778         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 779         int *num_workspace                      = &comp_num_workspace[idx];
 780 again:
 781         spin_lock(workspace_lock);
 782         if (!list_empty(idle_workspace)) {
 783                 workspace = idle_workspace->next;
 784                 list_del(workspace);
 785                 (*num_workspace)--;
 786                 spin_unlock(workspace_lock);
 787                 return workspace;
 788
 789         }
 790         if (atomic_read(alloc_workspace) > cpus) {
 791                 DEFINE_WAIT(wait);
 792
 793                 spin_unlock(workspace_lock);
 794                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 795                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 796                         schedule();
 797                 finish_wait(workspace_wait, &wait);
 798                 goto again;
 799         }
 800         atomic_inc(alloc_workspace);
 801         spin_unlock(workspace_lock);
 802
 803         workspace = btrfs_compress_op[idx]->alloc_workspace();
 804         if (IS_ERR(workspace)) {
 805                 atomic_dec(alloc_workspace);
 806                 wake_up(workspace_wait);
 807         }
 808         return workspace;
 809 }
 810
 811 /*
 812  * put a workspace struct back on the list or free it if we have enough
 813  * idle ones sitting around
 814  */
 815 static void free_workspace(int type, struct list_head *workspace)
 816 {
 817         int idx = type - 1;
 818         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 819         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 820         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 821         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 822         int *num_workspace                      = &comp_num_workspace[idx];
 823
 824         spin_lock(workspace_lock);
 825         if (*num_workspace < num_online_cpus()) {
 826                 list_add_tail(workspace, idle_workspace);
 827                 (*num_workspace)++;
 828                 spin_unlock(workspace_lock);
 829                 goto wake;
 830         }
 831         spin_unlock(workspace_lock);
 832
 833         btrfs_compress_op[idx]->free_workspace(workspace);
 834         atomic_dec(alloc_workspace);
 835 wake:
 836         smp_mb();
 837         if (waitqueue_active(workspace_wait))
 838                 wake_up(workspace_wait);
 839 }
 840
 841 /*
 842  * cleanup function for module exit
 843  */
 844 static void free_workspaces(void)
 845 {
 846         struct list_head *workspace;
 847         int i;
 848
 849         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 850                 while (!list_empty(&comp_idle_workspace[i])) {
 851                         workspace = comp_idle_workspace[i].next;
 852                         list_del(workspace);
 853                         btrfs_compress_op[i]->free_workspace(workspace);
 854                         atomic_dec(&comp_alloc_workspace[i]);
 855                 }
 856         }
 857 }
 858
 859 /*
 860  * given an address space and start/len, compress the bytes.
 861  *
 862  * pages are allocated to hold the compressed result and stored
 863  * in 'pages'
 864  *
 865  * out_pages is used to return the number of pages allocated.  There
 866  * may be pages allocated even if we return an error
 867  *
 868  * total_in is used to return the number of bytes actually read.  It
 869  * may be smaller then len if we had to exit early because we
 870  * ran out of room in the pages array or because we cross the
 871  * max_out threshold.
 872  *
 873  * total_out is used to return the total number of compressed bytes
 874  *
 875  * max_out tells us the max number of bytes that we're allowed to
 876  * stuff into pages
 877  */
 878 int btrfs_compress_pages(int type, struct address_space *mapping,
 879                          u64 start, unsigned long len,
 880                          struct page **pages,
 881                          unsigned long nr_dest_pages,
 882                          unsigned long *out_pages,
 883                          unsigned long *total_in,
 884                          unsigned long *total_out,
 885                          unsigned long max_out)
 886 {
 887         struct list_head *workspace;
 888         int ret;
 889
 890         workspace = find_workspace(type);
 891         if (IS_ERR(workspace))
 892                 return -1;
 893
 894         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 895                                                       start, len, pages,
 896                                                       nr_dest_pages, out_pages,
 897                                                       total_in, total_out,
 898                                                       max_out);
 899         free_workspace(type, workspace);
 900         return ret;
 901 }
 902
 903 /*
 904  * pages_in is an array of pages with compressed data.
 905  *
 906  * disk_start is the starting logical offset of this array in the file
 907  *
 908  * bvec is a bio_vec of pages from the file that we want to decompress into
 909  *
 910  * vcnt is the count of pages in the biovec
 911  *
 912  * srclen is the number of bytes in pages_in
 913  *
 914  * The basic idea is that we have a bio that was created by readpages.
 915  * The pages in the bio are for the uncompressed data, and they may not
 916  * be contiguous.  They all correspond to the range of bytes covered by
 917  * the compressed extent.
 918  */
 919 static int btrfs_decompress_biovec(int type, struct page **pages_in,
 920                                    u64 disk_start, struct bio_vec *bvec,
 921                                    int vcnt, size_t srclen)
 922 {
 923         struct list_head *workspace;
 924         int ret;
 925
 926         workspace = find_workspace(type);
 927         if (IS_ERR(workspace))
 928                 return -ENOMEM;
 929
 930         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 931                                                          disk_start,
 932                                                          bvec, vcnt, srclen);
 933         free_workspace(type, workspace);
 934         return ret;
 935 }
 936
 937 /*
 938  * a less complex decompression routine.  Our compressed data fits in a
 939  * single page, and we want to read a single page out of it.
 940  * start_byte tells us the offset into the compressed data we're interested in
 941  */
 942 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 943                      unsigned long start_byte, size_t srclen, size_t destlen)
 944 {
 945         struct list_head *workspace;
 946         int ret;
 947
 948         workspace = find_workspace(type);
 949         if (IS_ERR(workspace))
 950                 return -ENOMEM;
 951
 952         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 953                                                   dest_page, start_byte,
 954                                                   srclen, destlen);
 955
 956         free_workspace(type, workspace);
 957         return ret;
 958 }
 959
 960 void btrfs_exit_compress(void)
 961 {
 962         free_workspaces();
 963 }
 964
 965 /*
 966  * Copy uncompressed data from working buffer to pages.
 967  *
 968  * buf_start is the byte offset we're of the start of our workspace buffer.
 969  *
 970  * total_out is the last byte of the buffer
 971  */
 972 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 973                               unsigned long total_out, u64 disk_start,
 974                               struct bio_vec *bvec, int vcnt,
 975                               unsigned long *pg_index,
 976                               unsigned long *pg_offset)
 977 {
 978         unsigned long buf_offset;
 979         unsigned long current_buf_start;
 980         unsigned long start_byte;
 981         unsigned long working_bytes = total_out - buf_start;
 982         unsigned long bytes;
 983         char *kaddr;
 984         struct page *page_out = bvec[*pg_index].bv_page;
 985
 986         /*
 987          * start byte is the first byte of the page we're currently
 988          * copying into relative to the start of the compressed data.
 989          */
 990         start_byte = page_offset(page_out) - disk_start;
 991
 992         /* we haven't yet hit data corresponding to this page */
 993         if (total_out <= start_byte)
 994                 return 1;
 995
 996         /*
 997          * the start of the data we care about is offset into
 998          * the middle of our working buffer
 999          */
1000         if (total_out > start_byte && buf_start < start_byte) {
1001                 buf_offset = start_byte - buf_start;
1002                 working_bytes -= buf_offset;
1003         } else {
1004                 buf_offset = 0;
1005         }
1006         current_buf_start = buf_start;
1007
1008         /* copy bytes from the working buffer into the pages */
1009         while (working_bytes > 0) {
1010                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
1011                             PAGE_CACHE_SIZE - buf_offset);
1012                 bytes = min(bytes, working_bytes);
1013                 kaddr = kmap_atomic(page_out);
1014                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1015                 kunmap_atomic(kaddr);
1016                 flush_dcache_page(page_out);
1017
1018                 *pg_offset += bytes;
1019                 buf_offset += bytes;
1020                 working_bytes -= bytes;
1021                 current_buf_start += bytes;
1022
1023                 /* check if we need to pick another page */
1024                 if (*pg_offset == PAGE_CACHE_SIZE) {
1025                         (*pg_index)++;
1026                         if (*pg_index >= vcnt)
1027                                 return 0;
1028
1029                         page_out = bvec[*pg_index].bv_page;
1030                         *pg_offset = 0;
1031                         start_byte = page_offset(page_out) - disk_start;
1032
1033                         /*
1034                          * make sure our new page is covered by this
1035                          * working buffer
1036                          */
1037                         if (total_out <= start_byte)
1038                                 return 1;
1039
1040                         /*
1041                          * the next page in the biovec might not be adjacent
1042                          * to the last page, but it might still be found
1043                          * inside this working buffer. bump our offset pointer
1044                          */
1045                         if (total_out > start_byte &&
1046                             current_buf_start < start_byte) {
1047                                 buf_offset = start_byte - buf_start;
1048                                 working_bytes = total_out - start_byte;
1049                                 current_buf_start = buf_start + buf_offset;
1050                         }
1051                 }
1052         }
1053
1054         return 1;
1055 }