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         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 342
 343         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 344         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 345         if (!cb)
 346                 return -ENOMEM;
 347         atomic_set(&cb->pending_bios, 0);
 348         cb->errors = 0;
 349         cb->inode = inode;
 350         cb->start = start;
 351         cb->len = len;
 352         cb->mirror_num = 0;
 353         cb->compressed_pages = compressed_pages;
 354         cb->compressed_len = compressed_len;
 355         cb->orig_bio = NULL;
 356         cb->nr_pages = nr_pages;
 357
 358         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 359
 360         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 361         if(!bio) {
 362                 kfree(cb);
 363                 return -ENOMEM;
 364         }
 365         bio->bi_private = cb;
 366         bio->bi_end_io = end_compressed_bio_write;
 367         atomic_inc(&cb->pending_bios);
 368
 369         /* create and submit bios for the compressed pages */
 370         bytes_left = compressed_len;
 371         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 372                 page = compressed_pages[pg_index];
 373                 page->mapping = inode->i_mapping;
 374                 if (bio->bi_size)
 375                         ret = io_tree->ops->merge_bio_hook(page, 0,
 376                                                            PAGE_CACHE_SIZE,
 377                                                            bio, 0);
 378                 else
 379                         ret = 0;
 380
 381                 page->mapping = NULL;
 382                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 383                     PAGE_CACHE_SIZE) {
 384                         bio_get(bio);
 385
 386                         /*
 387                          * inc the count before we submit the bio so
 388                          * we know the end IO handler won't happen before
 389                          * we inc the count.  Otherwise, the cb might get
 390                          * freed before we're done setting it up
 391                          */
 392                         atomic_inc(&cb->pending_bios);
 393                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 394                         BUG_ON(ret);
 395
 396                         if (!skip_sum) {
 397                                 ret = btrfs_csum_one_bio(root, inode, bio,
 398                                                          start, 1);
 399                                 BUG_ON(ret);
 400                         }
 401
 402                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 403                         BUG_ON(ret);
 404
 405                         bio_put(bio);
 406
 407                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 408                         bio->bi_private = cb;
 409                         bio->bi_end_io = end_compressed_bio_write;
 410                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 411                 }
 412                 if (bytes_left < PAGE_CACHE_SIZE) {
 413                         printk("bytes left %lu compress len %lu nr %lu\n",
 414                                bytes_left, cb->compressed_len, cb->nr_pages);
 415                 }
 416                 bytes_left -= PAGE_CACHE_SIZE;
 417                 first_byte += PAGE_CACHE_SIZE;
 418                 cond_resched();
 419         }
 420         bio_get(bio);
 421
 422         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 423         BUG_ON(ret);
 424
 425         if (!skip_sum) {
 426                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 427                 BUG_ON(ret);
 428         }
 429
 430         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 431         BUG_ON(ret);
 432
 433         bio_put(bio);
 434         return 0;
 435 }
 436
 437 static noinline int add_ra_bio_pages(struct inode *inode,
 438                                      u64 compressed_end,
 439                                      struct compressed_bio *cb)
 440 {
 441         unsigned long end_index;
 442         unsigned long pg_index;
 443         u64 last_offset;
 444         u64 isize = i_size_read(inode);
 445         int ret;
 446         struct page *page;
 447         unsigned long nr_pages = 0;
 448         struct extent_map *em;
 449         struct address_space *mapping = inode->i_mapping;
 450         struct extent_map_tree *em_tree;
 451         struct extent_io_tree *tree;
 452         u64 end;
 453         int misses = 0;
 454
 455         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 456         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 457         em_tree = &BTRFS_I(inode)->extent_tree;
 458         tree = &BTRFS_I(inode)->io_tree;
 459
 460         if (isize == 0)
 461                 return 0;
 462
 463         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 464
 465         while (last_offset < compressed_end) {
 466                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 467
 468                 if (pg_index > end_index)
 469                         break;
 470
 471                 rcu_read_lock();
 472                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 473                 rcu_read_unlock();
 474                 if (page) {
 475                         misses++;
 476                         if (misses > 4)
 477                                 break;
 478                         goto next;
 479                 }
 480
 481                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 482                                                                 ~__GFP_FS);
 483                 if (!page)
 484                         break;
 485
 486                 if (add_to_page_cache_lru(page, mapping, pg_index,
 487                                                                 GFP_NOFS)) {
 488                         page_cache_release(page);
 489                         goto next;
 490                 }
 491
 492                 end = last_offset + PAGE_CACHE_SIZE - 1;
 493                 /*
 494                  * at this point, we have a locked page in the page cache
 495                  * for these bytes in the file.  But, we have to make
 496                  * sure they map to this compressed extent on disk.
 497                  */
 498                 set_page_extent_mapped(page);
 499                 lock_extent(tree, last_offset, end, GFP_NOFS);
 500                 read_lock(&em_tree->lock);
 501                 em = lookup_extent_mapping(em_tree, last_offset,
 502                                            PAGE_CACHE_SIZE);
 503                 read_unlock(&em_tree->lock);
 504
 505                 if (!em || last_offset < em->start ||
 506                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 507                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 508                         free_extent_map(em);
 509                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 510                         unlock_page(page);
 511                         page_cache_release(page);
 512                         break;
 513                 }
 514                 free_extent_map(em);
 515
 516                 if (page->index == end_index) {
 517                         char *userpage;
 518                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 519
 520                         if (zero_offset) {
 521                                 int zeros;
 522                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 523                                 userpage = kmap_atomic(page, KM_USER0);
 524                                 memset(userpage + zero_offset, 0, zeros);
 525                                 flush_dcache_page(page);
 526                                 kunmap_atomic(userpage, KM_USER0);
 527                         }
 528                 }
 529
 530                 ret = bio_add_page(cb->orig_bio, page,
 531                                    PAGE_CACHE_SIZE, 0);
 532
 533                 if (ret == PAGE_CACHE_SIZE) {
 534                         nr_pages++;
 535                         page_cache_release(page);
 536                 } else {
 537                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 538                         unlock_page(page);
 539                         page_cache_release(page);
 540                         break;
 541                 }
 542 next:
 543                 last_offset += PAGE_CACHE_SIZE;
 544         }
 545         return 0;
 546 }
 547
 548 /*
 549  * for a compressed read, the bio we get passed has all the inode pages
 550  * in it.  We don't actually do IO on those pages but allocate new ones
 551  * to hold the compressed pages on disk.
 552  *
 553  * bio->bi_sector points to the compressed extent on disk
 554  * bio->bi_io_vec points to all of the inode pages
 555  * bio->bi_vcnt is a count of pages
 556  *
 557  * After the compressed pages are read, we copy the bytes into the
 558  * bio we were passed and then call the bio end_io calls
 559  */
 560 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 561                                  int mirror_num, unsigned long bio_flags)
 562 {
 563         struct extent_io_tree *tree;
 564         struct extent_map_tree *em_tree;
 565         struct compressed_bio *cb;
 566         struct btrfs_root *root = BTRFS_I(inode)->root;
 567         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 568         unsigned long compressed_len;
 569         unsigned long nr_pages;
 570         unsigned long pg_index;
 571         struct page *page;
 572         struct block_device *bdev;
 573         struct bio *comp_bio;
 574         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 575         u64 em_len;
 576         u64 em_start;
 577         struct extent_map *em;
 578         int ret = -ENOMEM;
 579         u32 *sums;
 580
 581         tree = &BTRFS_I(inode)->io_tree;
 582         em_tree = &BTRFS_I(inode)->extent_tree;
 583
 584         /* we need the actual starting offset of this extent in the file */
 585         read_lock(&em_tree->lock);
 586         em = lookup_extent_mapping(em_tree,
 587                                    page_offset(bio->bi_io_vec->bv_page),
 588                                    PAGE_CACHE_SIZE);
 589         read_unlock(&em_tree->lock);
 590
 591         compressed_len = em->block_len;
 592         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 593         if (!cb)
 594                 goto out;
 595
 596         atomic_set(&cb->pending_bios, 0);
 597         cb->errors = 0;
 598         cb->inode = inode;
 599         cb->mirror_num = mirror_num;
 600         sums = &cb->sums;
 601
 602         cb->start = em->orig_start;
 603         em_len = em->len;
 604         em_start = em->start;
 605
 606         free_extent_map(em);
 607         em = NULL;
 608
 609         cb->len = uncompressed_len;
 610         cb->compressed_len = compressed_len;
 611         cb->compress_type = extent_compress_type(bio_flags);
 612         cb->orig_bio = bio;
 613
 614         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 615                                  PAGE_CACHE_SIZE;
 616         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 617                                        GFP_NOFS);
 618         if (!cb->compressed_pages)
 619                 goto fail1;
 620
 621         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 622
 623         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 624                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 625                                                               __GFP_HIGHMEM);
 626                 if (!cb->compressed_pages[pg_index])
 627                         goto fail2;
 628         }
 629         cb->nr_pages = nr_pages;
 630
 631         add_ra_bio_pages(inode, em_start + em_len, cb);
 632
 633         /* include any pages we added in add_ra-bio_pages */
 634         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 635         cb->len = uncompressed_len;
 636
 637         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 638         if (!comp_bio)
 639                 goto fail2;
 640         comp_bio->bi_private = cb;
 641         comp_bio->bi_end_io = end_compressed_bio_read;
 642         atomic_inc(&cb->pending_bios);
 643
 644         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 645                 page = cb->compressed_pages[pg_index];
 646                 page->mapping = inode->i_mapping;
 647                 page->index = em_start >> PAGE_CACHE_SHIFT;
 648
 649                 if (comp_bio->bi_size)
 650                         ret = tree->ops->merge_bio_hook(page, 0,
 651                                                         PAGE_CACHE_SIZE,
 652                                                         comp_bio, 0);
 653                 else
 654                         ret = 0;
 655
 656                 page->mapping = NULL;
 657                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 658                     PAGE_CACHE_SIZE) {
 659                         bio_get(comp_bio);
 660
 661                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 662                         BUG_ON(ret);
 663
 664                         /*
 665                          * inc the count before we submit the bio so
 666                          * we know the end IO handler won't happen before
 667                          * we inc the count.  Otherwise, the cb might get
 668                          * freed before we're done setting it up
 669                          */
 670                         atomic_inc(&cb->pending_bios);
 671
 672                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 673                                 ret = btrfs_lookup_bio_sums(root, inode,
 674                                                         comp_bio, sums);
 675                                 BUG_ON(ret);
 676                         }
 677                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 678                                 root->sectorsize;
 679
 680                         ret = btrfs_map_bio(root, READ, comp_bio,
 681                                             mirror_num, 0);
 682                         BUG_ON(ret);
 683
 684                         bio_put(comp_bio);
 685
 686                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 687                                                         GFP_NOFS);
 688                         comp_bio->bi_private = cb;
 689                         comp_bio->bi_end_io = end_compressed_bio_read;
 690
 691                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 692                 }
 693                 cur_disk_byte += PAGE_CACHE_SIZE;
 694         }
 695         bio_get(comp_bio);
 696
 697         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 698         BUG_ON(ret);
 699
 700         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 701                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 702                 BUG_ON(ret);
 703         }
 704
 705         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 706         BUG_ON(ret);
 707
 708         bio_put(comp_bio);
 709         return 0;
 710
 711 fail2:
 712         for (pg_index = 0; pg_index < nr_pages; pg_index++)
 713                 free_page((unsigned long)cb->compressed_pages[pg_index]);
 714
 715         kfree(cb->compressed_pages);
 716 fail1:
 717         kfree(cb);
 718 out:
 719         free_extent_map(em);
 720         return ret;
 721 }
 722
 723 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 724 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 725 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 726 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 727 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 728
 729 struct btrfs_compress_op *btrfs_compress_op[] = {
 730         &btrfs_zlib_compress,
 731         &btrfs_lzo_compress,
 732 };
 733
 734 int __init btrfs_init_compress(void)
 735 {
 736         int i;
 737
 738         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 739                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 740                 spin_lock_init(&comp_workspace_lock[i]);
 741                 atomic_set(&comp_alloc_workspace[i], 0);
 742                 init_waitqueue_head(&comp_workspace_wait[i]);
 743         }
 744         return 0;
 745 }
 746
 747 /*
 748  * this finds an available workspace or allocates a new one
 749  * ERR_PTR is returned if things go bad.
 750  */
 751 static struct list_head *find_workspace(int type)
 752 {
 753         struct list_head *workspace;
 754         int cpus = num_online_cpus();
 755         int idx = type - 1;
 756
 757         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 758         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 759         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 760         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 761         int *num_workspace                      = &comp_num_workspace[idx];
 762 again:
 763         spin_lock(workspace_lock);
 764         if (!list_empty(idle_workspace)) {
 765                 workspace = idle_workspace->next;
 766                 list_del(workspace);
 767                 (*num_workspace)--;
 768                 spin_unlock(workspace_lock);
 769                 return workspace;
 770
 771         }
 772         if (atomic_read(alloc_workspace) > cpus) {
 773                 DEFINE_WAIT(wait);
 774
 775                 spin_unlock(workspace_lock);
 776                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 777                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 778                         schedule();
 779                 finish_wait(workspace_wait, &wait);
 780                 goto again;
 781         }
 782         atomic_inc(alloc_workspace);
 783         spin_unlock(workspace_lock);
 784
 785         workspace = btrfs_compress_op[idx]->alloc_workspace();
 786         if (IS_ERR(workspace)) {
 787                 atomic_dec(alloc_workspace);
 788                 wake_up(workspace_wait);
 789         }
 790         return workspace;
 791 }
 792
 793 /*
 794  * put a workspace struct back on the list or free it if we have enough
 795  * idle ones sitting around
 796  */
 797 static void free_workspace(int type, struct list_head *workspace)
 798 {
 799         int idx = type - 1;
 800         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 801         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 802         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 803         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 804         int *num_workspace                      = &comp_num_workspace[idx];
 805
 806         spin_lock(workspace_lock);
 807         if (*num_workspace < num_online_cpus()) {
 808                 list_add_tail(workspace, idle_workspace);
 809                 (*num_workspace)++;
 810                 spin_unlock(workspace_lock);
 811                 goto wake;
 812         }
 813         spin_unlock(workspace_lock);
 814
 815         btrfs_compress_op[idx]->free_workspace(workspace);
 816         atomic_dec(alloc_workspace);
 817 wake:
 818         if (waitqueue_active(workspace_wait))
 819                 wake_up(workspace_wait);
 820 }
 821
 822 /*
 823  * cleanup function for module exit
 824  */
 825 static void free_workspaces(void)
 826 {
 827         struct list_head *workspace;
 828         int i;
 829
 830         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 831                 while (!list_empty(&comp_idle_workspace[i])) {
 832                         workspace = comp_idle_workspace[i].next;
 833                         list_del(workspace);
 834                         btrfs_compress_op[i]->free_workspace(workspace);
 835                         atomic_dec(&comp_alloc_workspace[i]);
 836                 }
 837         }
 838 }
 839
 840 /*
 841  * given an address space and start/len, compress the bytes.
 842  *
 843  * pages are allocated to hold the compressed result and stored
 844  * in 'pages'
 845  *
 846  * out_pages is used to return the number of pages allocated.  There
 847  * may be pages allocated even if we return an error
 848  *
 849  * total_in is used to return the number of bytes actually read.  It
 850  * may be smaller then len if we had to exit early because we
 851  * ran out of room in the pages array or because we cross the
 852  * max_out threshold.
 853  *
 854  * total_out is used to return the total number of compressed bytes
 855  *
 856  * max_out tells us the max number of bytes that we're allowed to
 857  * stuff into pages
 858  */
 859 int btrfs_compress_pages(int type, struct address_space *mapping,
 860                          u64 start, unsigned long len,
 861                          struct page **pages,
 862                          unsigned long nr_dest_pages,
 863                          unsigned long *out_pages,
 864                          unsigned long *total_in,
 865                          unsigned long *total_out,
 866                          unsigned long max_out)
 867 {
 868         struct list_head *workspace;
 869         int ret;
 870
 871         workspace = find_workspace(type);
 872         if (IS_ERR(workspace))
 873                 return -1;
 874
 875         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 876                                                       start, len, pages,
 877                                                       nr_dest_pages, out_pages,
 878                                                       total_in, total_out,
 879                                                       max_out);
 880         free_workspace(type, workspace);
 881         return ret;
 882 }
 883
 884 /*
 885  * pages_in is an array of pages with compressed data.
 886  *
 887  * disk_start is the starting logical offset of this array in the file
 888  *
 889  * bvec is a bio_vec of pages from the file that we want to decompress into
 890  *
 891  * vcnt is the count of pages in the biovec
 892  *
 893  * srclen is the number of bytes in pages_in
 894  *
 895  * The basic idea is that we have a bio that was created by readpages.
 896  * The pages in the bio are for the uncompressed data, and they may not
 897  * be contiguous.  They all correspond to the range of bytes covered by
 898  * the compressed extent.
 899  */
 900 int btrfs_decompress_biovec(int type, struct page **pages_in, u64 disk_start,
 901                             struct bio_vec *bvec, int vcnt, size_t srclen)
 902 {
 903         struct list_head *workspace;
 904         int ret;
 905
 906         workspace = find_workspace(type);
 907         if (IS_ERR(workspace))
 908                 return -ENOMEM;
 909
 910         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 911                                                          disk_start,
 912                                                          bvec, vcnt, srclen);
 913         free_workspace(type, workspace);
 914         return ret;
 915 }
 916
 917 /*
 918  * a less complex decompression routine.  Our compressed data fits in a
 919  * single page, and we want to read a single page out of it.
 920  * start_byte tells us the offset into the compressed data we're interested in
 921  */
 922 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 923                      unsigned long start_byte, size_t srclen, size_t destlen)
 924 {
 925         struct list_head *workspace;
 926         int ret;
 927
 928         workspace = find_workspace(type);
 929         if (IS_ERR(workspace))
 930                 return -ENOMEM;
 931
 932         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 933                                                   dest_page, start_byte,
 934                                                   srclen, destlen);
 935
 936         free_workspace(type, workspace);
 937         return ret;
 938 }
 939
 940 void btrfs_exit_compress(void)
 941 {
 942         free_workspaces();
 943 }
 944
 945 /*
 946  * Copy uncompressed data from working buffer to pages.
 947  *
 948  * buf_start is the byte offset we're of the start of our workspace buffer.
 949  *
 950  * total_out is the last byte of the buffer
 951  */
 952 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 953                               unsigned long total_out, u64 disk_start,
 954                               struct bio_vec *bvec, int vcnt,
 955                               unsigned long *pg_index,
 956                               unsigned long *pg_offset)
 957 {
 958         unsigned long buf_offset;
 959         unsigned long current_buf_start;
 960         unsigned long start_byte;
 961         unsigned long working_bytes = total_out - buf_start;
 962         unsigned long bytes;
 963         char *kaddr;
 964         struct page *page_out = bvec[*pg_index].bv_page;
 965
 966         /*
 967          * start byte is the first byte of the page we're currently
 968          * copying into relative to the start of the compressed data.
 969          */
 970         start_byte = page_offset(page_out) - disk_start;
 971
 972         /* we haven't yet hit data corresponding to this page */
 973         if (total_out <= start_byte)
 974                 return 1;
 975
 976         /*
 977          * the start of the data we care about is offset into
 978          * the middle of our working buffer
 979          */
 980         if (total_out > start_byte && buf_start < start_byte) {
 981                 buf_offset = start_byte - buf_start;
 982                 working_bytes -= buf_offset;
 983         } else {
 984                 buf_offset = 0;
 985         }
 986         current_buf_start = buf_start;
 987
 988         /* copy bytes from the working buffer into the pages */
 989         while (working_bytes > 0) {
 990                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
 991                             PAGE_CACHE_SIZE - buf_offset);
 992                 bytes = min(bytes, working_bytes);
 993                 kaddr = kmap_atomic(page_out, KM_USER0);
 994                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
 995                 kunmap_atomic(kaddr, KM_USER0);
 996                 flush_dcache_page(page_out);
 997
 998                 *pg_offset += bytes;
 999                 buf_offset += bytes;
1000                 working_bytes -= bytes;
1001                 current_buf_start += bytes;
1002
1003                 /* check if we need to pick another page */
1004                 if (*pg_offset == PAGE_CACHE_SIZE) {
1005                         (*pg_index)++;
1006                         if (*pg_index >= vcnt)
1007                                 return 0;
1008
1009                         page_out = bvec[*pg_index].bv_page;
1010                         *pg_offset = 0;
1011                         start_byte = page_offset(page_out) - disk_start;
1012
1013                         /*
1014                          * make sure our new page is covered by this
1015                          * working buffer
1016                          */
1017                         if (total_out <= start_byte)
1018                                 return 1;
1019
1020                         /*
1021                          * the next page in the biovec might not be adjacent
1022                          * to the last page, but it might still be found
1023                          * inside this working buffer. bump our offset pointer
1024                          */
1025                         if (total_out > start_byte &&
1026                             current_buf_start < start_byte) {
1027                                 buf_offset = start_byte - buf_start;
1028                                 working_bytes = total_out - start_byte;
1029                                 current_buf_start = buf_start + buf_offset;
1030                         }
1031                 }
1032         }
1033
1034         return 1;
1035 }