fs/btrfs/compression.c

   1 /*
   2  * Copyright (C) 2008 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/bit_spinlock.h>
  34 #include <linux/slab.h>
  35 #include "compat.h"
  36 #include "ctree.h"
  37 #include "disk-io.h"
  38 #include "transaction.h"
  39 #include "btrfs_inode.h"
  40 #include "volumes.h"
  41 #include "ordered-data.h"
  42 #include "compression.h"
  43 #include "extent_io.h"
  44 #include "extent_map.h"
  45
  46 struct compressed_bio {
  47         /* number of bios pending for this compressed extent */
  48         atomic_t pending_bios;
  49
  50         /* the pages with the compressed data on them */
  51         struct page **compressed_pages;
  52
  53         /* inode that owns this data */
  54         struct inode *inode;
  55
  56         /* starting offset in the inode for our pages */
  57         u64 start;
  58
  59         /* number of bytes in the inode we're working on */
  60         unsigned long len;
  61
  62         /* number of bytes on disk */
  63         unsigned long compressed_len;
  64
  65         /* the compression algorithm for this bio */
  66         int compress_type;
  67
  68         /* number of compressed pages in the array */
  69         unsigned long nr_pages;
  70
  71         /* IO errors */
  72         int errors;
  73         int mirror_num;
  74
  75         /* for reads, this is the bio we are copying the data into */
  76         struct bio *orig_bio;
  77
  78         /*
  79          * the start of a variable length array of checksums only
  80          * used by reads
  81          */
  82         u32 sums;
  83 };
  84
  85 static inline int compressed_bio_size(struct btrfs_root *root,
  86                                       unsigned long disk_size)
  87 {
  88         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
  89
  90         return sizeof(struct compressed_bio) +
  91                 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
  92                 csum_size;
  93 }
  94
  95 static struct bio *compressed_bio_alloc(struct block_device *bdev,
  96                                         u64 first_byte, gfp_t gfp_flags)
  97 {
  98         int nr_vecs;
  99
 100         nr_vecs = bio_get_nr_vecs(bdev);
 101         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
 102 }
 103
 104 static int check_compressed_csum(struct inode *inode,
 105                                  struct compressed_bio *cb,
 106                                  u64 disk_start)
 107 {
 108         int ret;
 109         struct btrfs_root *root = BTRFS_I(inode)->root;
 110         struct page *page;
 111         unsigned long i;
 112         char *kaddr;
 113         u32 csum;
 114         u32 *cb_sum = &cb->sums;
 115
 116         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 117                 return 0;
 118
 119         for (i = 0; i < cb->nr_pages; i++) {
 120                 page = cb->compressed_pages[i];
 121                 csum = ~(u32)0;
 122
 123                 kaddr = kmap_atomic(page, KM_USER0);
 124                 csum = btrfs_csum_data(root, kaddr, csum, PAGE_CACHE_SIZE);
 125                 btrfs_csum_final(csum, (char *)&csum);
 126                 kunmap_atomic(kaddr, KM_USER0);
 127
 128                 if (csum != *cb_sum) {
 129                         printk(KERN_INFO "btrfs csum failed ino %llu "
 130                                "extent %llu csum %u "
 131                                "wanted %u mirror %d\n",
 132                                (unsigned long long)btrfs_ino(inode),
 133                                (unsigned long long)disk_start,
 134                                csum, *cb_sum, cb->mirror_num);
 135                         ret = -EIO;
 136                         goto fail;
 137                 }
 138                 cb_sum++;
 139
 140         }
 141         ret = 0;
 142 fail:
 143         return ret;
 144 }
 145
 146 /* when we finish reading compressed pages from the disk, we
 147  * decompress them and then run the bio end_io routines on the
 148  * decompressed pages (in the inode address space).
 149  *
 150  * This allows the checksumming and other IO error handling routines
 151  * to work normally
 152  *
 153  * The compressed pages are freed here, and it must be run
 154  * in process context
 155  */
 156 static void end_compressed_bio_read(struct bio *bio, int err)
 157 {
 158         struct compressed_bio *cb = bio->bi_private;
 159         struct inode *inode;
 160         struct page *page;
 161         unsigned long index;
 162         int ret;
 163
 164         if (err)
 165                 cb->errors = 1;
 166
 167         /* if there are more bios still pending for this compressed
 168          * extent, just exit
 169          */
 170         if (!atomic_dec_and_test(&cb->pending_bios))
 171                 goto out;
 172
 173         inode = cb->inode;
 174         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
 175         if (ret)
 176                 goto csum_failed;
 177
 178         /* ok, we're the last bio for this extent, lets start
 179          * the decompression.
 180          */
 181         ret = btrfs_decompress_biovec(cb->compress_type,
 182                                       cb->compressed_pages,
 183                                       cb->start,
 184                                       cb->orig_bio->bi_io_vec,
 185                                       cb->orig_bio->bi_vcnt,
 186                                       cb->compressed_len);
 187 csum_failed:
 188         if (ret)
 189                 cb->errors = 1;
 190
 191         /* release the compressed pages */
 192         index = 0;
 193         for (index = 0; index < cb->nr_pages; index++) {
 194                 page = cb->compressed_pages[index];
 195                 page->mapping = NULL;
 196                 page_cache_release(page);
 197         }
 198
 199         /* do io completion on the original bio */
 200         if (cb->errors) {
 201                 bio_io_error(cb->orig_bio);
 202         } else {
 203                 int bio_index = 0;
 204                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
 205
 206                 /*
 207                  * we have verified the checksum already, set page
 208                  * checked so the end_io handlers know about it
 209                  */
 210                 while (bio_index < cb->orig_bio->bi_vcnt) {
 211                         SetPageChecked(bvec->bv_page);
 212                         bvec++;
 213                         bio_index++;
 214                 }
 215                 bio_endio(cb->orig_bio, 0);
 216         }
 217
 218         /* finally free the cb struct */
 219         kfree(cb->compressed_pages);
 220         kfree(cb);
 221 out:
 222         bio_put(bio);
 223 }
 224
 225 /*
 226  * Clear the writeback bits on all of the file
 227  * pages for a compressed write
 228  */
 229 static noinline int end_compressed_writeback(struct inode *inode, u64 start,
 230                                              unsigned long ram_size)
 231 {
 232         unsigned long index = start >> PAGE_CACHE_SHIFT;
 233         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
 234         struct page *pages[16];
 235         unsigned long nr_pages = end_index - index + 1;
 236         int i;
 237         int ret;
 238
 239         while (nr_pages > 0) {
 240                 ret = find_get_pages_contig(inode->i_mapping, index,
 241                                      min_t(unsigned long,
 242                                      nr_pages, ARRAY_SIZE(pages)), pages);
 243                 if (ret == 0) {
 244                         nr_pages -= 1;
 245                         index += 1;
 246                         continue;
 247                 }
 248                 for (i = 0; i < ret; i++) {
 249                         end_page_writeback(pages[i]);
 250                         page_cache_release(pages[i]);
 251                 }
 252                 nr_pages -= ret;
 253                 index += ret;
 254         }
 255         /* the inode may be gone now */
 256         return 0;
 257 }
 258
 259 /*
 260  * do the cleanup once all the compressed pages hit the disk.
 261  * This will clear writeback on the file pages and free the compressed
 262  * pages.
 263  *
 264  * This also calls the writeback end hooks for the file pages so that
 265  * metadata and checksums can be updated in the file.
 266  */
 267 static void end_compressed_bio_write(struct bio *bio, int err)
 268 {
 269         struct extent_io_tree *tree;
 270         struct compressed_bio *cb = bio->bi_private;
 271         struct inode *inode;
 272         struct page *page;
 273         unsigned long index;
 274
 275         if (err)
 276                 cb->errors = 1;
 277
 278         /* if there are more bios still pending for this compressed
 279          * extent, just exit
 280          */
 281         if (!atomic_dec_and_test(&cb->pending_bios))
 282                 goto out;
 283
 284         /* ok, we're the last bio for this extent, step one is to
 285          * call back into the FS and do all the end_io operations
 286          */
 287         inode = cb->inode;
 288         tree = &BTRFS_I(inode)->io_tree;
 289         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 290         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 291                                          cb->start,
 292                                          cb->start + cb->len - 1,
 293                                          NULL, 1);
 294         cb->compressed_pages[0]->mapping = NULL;
 295
 296         end_compressed_writeback(inode, cb->start, cb->len);
 297         /* note, our inode could be gone now */
 298
 299         /*
 300          * release the compressed pages, these came from alloc_page and
 301          * are not attached to the inode at all
 302          */
 303         index = 0;
 304         for (index = 0; index < cb->nr_pages; index++) {
 305                 page = cb->compressed_pages[index];
 306                 page->mapping = NULL;
 307                 page_cache_release(page);
 308         }
 309
 310         /* finally free the cb struct */
 311         kfree(cb->compressed_pages);
 312         kfree(cb);
 313 out:
 314         bio_put(bio);
 315 }
 316
 317 /*
 318  * worker function to build and submit bios for previously compressed pages.
 319  * The corresponding pages in the inode should be marked for writeback
 320  * and the compressed pages should have a reference on them for dropping
 321  * when the IO is complete.
 322  *
 323  * This also checksums the file bytes and gets things ready for
 324  * the end io hooks.
 325  */
 326 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 327                                  unsigned long len, u64 disk_start,
 328                                  unsigned long compressed_len,
 329                                  struct page **compressed_pages,
 330                                  unsigned long nr_pages)
 331 {
 332         struct bio *bio = NULL;
 333         struct btrfs_root *root = BTRFS_I(inode)->root;
 334         struct compressed_bio *cb;
 335         unsigned long bytes_left;
 336         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 337         int pg_index = 0;
 338         struct page *page;
 339         u64 first_byte = disk_start;
 340         struct block_device *bdev;
 341         int ret;
 342         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 343
 344         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 345         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 346         if (!cb)
 347                 return -ENOMEM;
 348         atomic_set(&cb->pending_bios, 0);
 349         cb->errors = 0;
 350         cb->inode = inode;
 351         cb->start = start;
 352         cb->len = len;
 353         cb->mirror_num = 0;
 354         cb->compressed_pages = compressed_pages;
 355         cb->compressed_len = compressed_len;
 356         cb->orig_bio = NULL;
 357         cb->nr_pages = nr_pages;
 358
 359         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 360
 361         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 362         if(!bio) {
 363                 kfree(cb);
 364                 return -ENOMEM;
 365         }
 366         bio->bi_private = cb;
 367         bio->bi_end_io = end_compressed_bio_write;
 368         atomic_inc(&cb->pending_bios);
 369
 370         /* create and submit bios for the compressed pages */
 371         bytes_left = compressed_len;
 372         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 373                 page = compressed_pages[pg_index];
 374                 page->mapping = inode->i_mapping;
 375                 if (bio->bi_size)
 376                         ret = io_tree->ops->merge_bio_hook(page, 0,
 377                                                            PAGE_CACHE_SIZE,
 378                                                            bio, 0);
 379                 else
 380                         ret = 0;
 381
 382                 page->mapping = NULL;
 383                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 384                     PAGE_CACHE_SIZE) {
 385                         bio_get(bio);
 386
 387                         /*
 388                          * inc the count before we submit the bio so
 389                          * we know the end IO handler won't happen before
 390                          * we inc the count.  Otherwise, the cb might get
 391                          * freed before we're done setting it up
 392                          */
 393                         atomic_inc(&cb->pending_bios);
 394                         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 395                         BUG_ON(ret);
 396
 397                         if (!skip_sum) {
 398                                 ret = btrfs_csum_one_bio(root, inode, bio,
 399                                                          start, 1);
 400                                 BUG_ON(ret);
 401                         }
 402
 403                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 404                         BUG_ON(ret);
 405
 406                         bio_put(bio);
 407
 408                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 409                         bio->bi_private = cb;
 410                         bio->bi_end_io = end_compressed_bio_write;
 411                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 412                 }
 413                 if (bytes_left < PAGE_CACHE_SIZE) {
 414                         printk("bytes left %lu compress len %lu nr %lu\n",
 415                                bytes_left, cb->compressed_len, cb->nr_pages);
 416                 }
 417                 bytes_left -= PAGE_CACHE_SIZE;
 418                 first_byte += PAGE_CACHE_SIZE;
 419                 cond_resched();
 420         }
 421         bio_get(bio);
 422
 423         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 424         BUG_ON(ret);
 425
 426         if (!skip_sum) {
 427                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 428                 BUG_ON(ret);
 429         }
 430
 431         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 432         BUG_ON(ret);
 433
 434         bio_put(bio);
 435         return 0;
 436 }
 437
 438 static noinline int add_ra_bio_pages(struct inode *inode,
 439                                      u64 compressed_end,
 440                                      struct compressed_bio *cb)
 441 {
 442         unsigned long end_index;
 443         unsigned long pg_index;
 444         u64 last_offset;
 445         u64 isize = i_size_read(inode);
 446         int ret;
 447         struct page *page;
 448         unsigned long nr_pages = 0;
 449         struct extent_map *em;
 450         struct address_space *mapping = inode->i_mapping;
 451         struct extent_map_tree *em_tree;
 452         struct extent_io_tree *tree;
 453         u64 end;
 454         int misses = 0;
 455
 456         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 457         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 458         em_tree = &BTRFS_I(inode)->extent_tree;
 459         tree = &BTRFS_I(inode)->io_tree;
 460
 461         if (isize == 0)
 462                 return 0;
 463
 464         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 465
 466         while (last_offset < compressed_end) {
 467                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 468
 469                 if (pg_index > end_index)
 470                         break;
 471
 472                 rcu_read_lock();
 473                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 474                 rcu_read_unlock();
 475                 if (page) {
 476                         misses++;
 477                         if (misses > 4)
 478                                 break;
 479                         goto next;
 480                 }
 481
 482                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 483                                                                 ~__GFP_FS);
 484                 if (!page)
 485                         break;
 486
 487                 if (add_to_page_cache_lru(page, mapping, pg_index,
 488                                                                 GFP_NOFS)) {
 489                         page_cache_release(page);
 490                         goto next;
 491                 }
 492
 493                 end = last_offset + PAGE_CACHE_SIZE - 1;
 494                 /*
 495                  * at this point, we have a locked page in the page cache
 496                  * for these bytes in the file.  But, we have to make
 497                  * sure they map to this compressed extent on disk.
 498                  */
 499                 set_page_extent_mapped(page);
 500                 lock_extent(tree, last_offset, end, GFP_NOFS);
 501                 read_lock(&em_tree->lock);
 502                 em = lookup_extent_mapping(em_tree, last_offset,
 503                                            PAGE_CACHE_SIZE);
 504                 read_unlock(&em_tree->lock);
 505
 506                 if (!em || last_offset < em->start ||
 507                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 508                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 509                         free_extent_map(em);
 510                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 511                         unlock_page(page);
 512                         page_cache_release(page);
 513                         break;
 514                 }
 515                 free_extent_map(em);
 516
 517                 if (page->index == end_index) {
 518                         char *userpage;
 519                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 520
 521                         if (zero_offset) {
 522                                 int zeros;
 523                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 524                                 userpage = kmap_atomic(page, KM_USER0);
 525                                 memset(userpage + zero_offset, 0, zeros);
 526                                 flush_dcache_page(page);
 527                                 kunmap_atomic(userpage, KM_USER0);
 528                         }
 529                 }
 530
 531                 ret = bio_add_page(cb->orig_bio, page,
 532                                    PAGE_CACHE_SIZE, 0);
 533
 534                 if (ret == PAGE_CACHE_SIZE) {
 535                         nr_pages++;
 536                         page_cache_release(page);
 537                 } else {
 538                         unlock_extent(tree, last_offset, end, GFP_NOFS);
 539                         unlock_page(page);
 540                         page_cache_release(page);
 541                         break;
 542                 }
 543 next:
 544                 last_offset += PAGE_CACHE_SIZE;
 545         }
 546         return 0;
 547 }
 548
 549 /*
 550  * for a compressed read, the bio we get passed has all the inode pages
 551  * in it.  We don't actually do IO on those pages but allocate new ones
 552  * to hold the compressed pages on disk.
 553  *
 554  * bio->bi_sector points to the compressed extent on disk
 555  * bio->bi_io_vec points to all of the inode pages
 556  * bio->bi_vcnt is a count of pages
 557  *
 558  * After the compressed pages are read, we copy the bytes into the
 559  * bio we were passed and then call the bio end_io calls
 560  */
 561 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 562                                  int mirror_num, unsigned long bio_flags)
 563 {
 564         struct extent_io_tree *tree;
 565         struct extent_map_tree *em_tree;
 566         struct compressed_bio *cb;
 567         struct btrfs_root *root = BTRFS_I(inode)->root;
 568         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 569         unsigned long compressed_len;
 570         unsigned long nr_pages;
 571         unsigned long pg_index;
 572         struct page *page;
 573         struct block_device *bdev;
 574         struct bio *comp_bio;
 575         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 576         u64 em_len;
 577         u64 em_start;
 578         struct extent_map *em;
 579         int ret = -ENOMEM;
 580         u32 *sums;
 581
 582         tree = &BTRFS_I(inode)->io_tree;
 583         em_tree = &BTRFS_I(inode)->extent_tree;
 584
 585         /* we need the actual starting offset of this extent in the file */
 586         read_lock(&em_tree->lock);
 587         em = lookup_extent_mapping(em_tree,
 588                                    page_offset(bio->bi_io_vec->bv_page),
 589                                    PAGE_CACHE_SIZE);
 590         read_unlock(&em_tree->lock);
 591
 592         compressed_len = em->block_len;
 593         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 594         if (!cb)
 595                 goto out;
 596
 597         atomic_set(&cb->pending_bios, 0);
 598         cb->errors = 0;
 599         cb->inode = inode;
 600         cb->mirror_num = mirror_num;
 601         sums = &cb->sums;
 602
 603         cb->start = em->orig_start;
 604         em_len = em->len;
 605         em_start = em->start;
 606
 607         free_extent_map(em);
 608         em = NULL;
 609
 610         cb->len = uncompressed_len;
 611         cb->compressed_len = compressed_len;
 612         cb->compress_type = extent_compress_type(bio_flags);
 613         cb->orig_bio = bio;
 614
 615         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 616                                  PAGE_CACHE_SIZE;
 617         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 618                                        GFP_NOFS);
 619         if (!cb->compressed_pages)
 620                 goto fail1;
 621
 622         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 623
 624         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 625                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 626                                                               __GFP_HIGHMEM);
 627                 if (!cb->compressed_pages[pg_index])
 628                         goto fail2;
 629         }
 630         cb->nr_pages = nr_pages;
 631
 632         add_ra_bio_pages(inode, em_start + em_len, cb);
 633
 634         /* include any pages we added in add_ra-bio_pages */
 635         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 636         cb->len = uncompressed_len;
 637
 638         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 639         if (!comp_bio)
 640                 goto fail2;
 641         comp_bio->bi_private = cb;
 642         comp_bio->bi_end_io = end_compressed_bio_read;
 643         atomic_inc(&cb->pending_bios);
 644
 645         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 646                 page = cb->compressed_pages[pg_index];
 647                 page->mapping = inode->i_mapping;
 648                 page->index = em_start >> PAGE_CACHE_SHIFT;
 649
 650                 if (comp_bio->bi_size)
 651                         ret = tree->ops->merge_bio_hook(page, 0,
 652                                                         PAGE_CACHE_SIZE,
 653                                                         comp_bio, 0);
 654                 else
 655                         ret = 0;
 656
 657                 page->mapping = NULL;
 658                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 659                     PAGE_CACHE_SIZE) {
 660                         bio_get(comp_bio);
 661
 662                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 663                         BUG_ON(ret);
 664
 665                         /*
 666                          * inc the count before we submit the bio so
 667                          * we know the end IO handler won't happen before
 668                          * we inc the count.  Otherwise, the cb might get
 669                          * freed before we're done setting it up
 670                          */
 671                         atomic_inc(&cb->pending_bios);
 672
 673                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 674                                 ret = btrfs_lookup_bio_sums(root, inode,
 675                                                         comp_bio, sums);
 676                                 BUG_ON(ret);
 677                         }
 678                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 679                                 root->sectorsize;
 680
 681                         ret = btrfs_map_bio(root, READ, comp_bio,
 682                                             mirror_num, 0);
 683                         BUG_ON(ret);
 684
 685                         bio_put(comp_bio);
 686
 687                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 688                                                         GFP_NOFS);
 689                         comp_bio->bi_private = cb;
 690                         comp_bio->bi_end_io = end_compressed_bio_read;
 691
 692                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 693                 }
 694                 cur_disk_byte += PAGE_CACHE_SIZE;
 695         }
 696         bio_get(comp_bio);
 697
 698         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 699         BUG_ON(ret);
 700
 701         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 702                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 703                 BUG_ON(ret);
 704         }
 705
 706         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 707         BUG_ON(ret);
 708
 709         bio_put(comp_bio);
 710         return 0;
 711
 712 fail2:
 713         for (pg_index = 0; pg_index < nr_pages; pg_index++)
 714                 free_page((unsigned long)cb->compressed_pages[pg_index]);
 715
 716         kfree(cb->compressed_pages);
 717 fail1:
 718         kfree(cb);
 719 out:
 720         free_extent_map(em);
 721         return ret;
 722 }
 723
 724 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 725 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 726 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 727 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 728 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 729
 730 struct btrfs_compress_op *btrfs_compress_op[] = {
 731         &btrfs_zlib_compress,
 732         &btrfs_lzo_compress,
 733 };
 734
 735 int __init btrfs_init_compress(void)
 736 {
 737         int i;
 738
 739         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 740                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 741                 spin_lock_init(&comp_workspace_lock[i]);
 742                 atomic_set(&comp_alloc_workspace[i], 0);
 743                 init_waitqueue_head(&comp_workspace_wait[i]);
 744         }
 745         return 0;
 746 }
 747
 748 /*
 749  * this finds an available workspace or allocates a new one
 750  * ERR_PTR is returned if things go bad.
 751  */
 752 static struct list_head *find_workspace(int type)
 753 {
 754         struct list_head *workspace;
 755         int cpus = num_online_cpus();
 756         int idx = type - 1;
 757
 758         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 759         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 760         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 761         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 762         int *num_workspace                      = &comp_num_workspace[idx];
 763 again:
 764         spin_lock(workspace_lock);
 765         if (!list_empty(idle_workspace)) {
 766                 workspace = idle_workspace->next;
 767                 list_del(workspace);
 768                 (*num_workspace)--;
 769                 spin_unlock(workspace_lock);
 770                 return workspace;
 771
 772         }
 773         if (atomic_read(alloc_workspace) > cpus) {
 774                 DEFINE_WAIT(wait);
 775
 776                 spin_unlock(workspace_lock);
 777                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 778                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 779                         schedule();
 780                 finish_wait(workspace_wait, &wait);
 781                 goto again;
 782         }
 783         atomic_inc(alloc_workspace);
 784         spin_unlock(workspace_lock);
 785
 786         workspace = btrfs_compress_op[idx]->alloc_workspace();
 787         if (IS_ERR(workspace)) {
 788                 atomic_dec(alloc_workspace);
 789                 wake_up(workspace_wait);
 790         }
 791         return workspace;
 792 }
 793
 794 /*
 795  * put a workspace struct back on the list or free it if we have enough
 796  * idle ones sitting around
 797  */
 798 static void free_workspace(int type, struct list_head *workspace)
 799 {
 800         int idx = type - 1;
 801         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 802         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 803         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 804         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 805         int *num_workspace                      = &comp_num_workspace[idx];
 806
 807         spin_lock(workspace_lock);
 808         if (*num_workspace < num_online_cpus()) {
 809                 list_add_tail(workspace, idle_workspace);
 810                 (*num_workspace)++;
 811                 spin_unlock(workspace_lock);
 812                 goto wake;
 813         }
 814         spin_unlock(workspace_lock);
 815
 816         btrfs_compress_op[idx]->free_workspace(workspace);
 817         atomic_dec(alloc_workspace);
 818 wake:
 819         if (waitqueue_active(workspace_wait))
 820                 wake_up(workspace_wait);
 821 }
 822
 823 /*
 824  * cleanup function for module exit
 825  */
 826 static void free_workspaces(void)
 827 {
 828         struct list_head *workspace;
 829         int i;
 830
 831         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 832                 while (!list_empty(&comp_idle_workspace[i])) {
 833                         workspace = comp_idle_workspace[i].next;
 834                         list_del(workspace);
 835                         btrfs_compress_op[i]->free_workspace(workspace);
 836                         atomic_dec(&comp_alloc_workspace[i]);
 837                 }
 838         }
 839 }
 840
 841 /*
 842  * given an address space and start/len, compress the bytes.
 843  *
 844  * pages are allocated to hold the compressed result and stored
 845  * in 'pages'
 846  *
 847  * out_pages is used to return the number of pages allocated.  There
 848  * may be pages allocated even if we return an error
 849  *
 850  * total_in is used to return the number of bytes actually read.  It
 851  * may be smaller then len if we had to exit early because we
 852  * ran out of room in the pages array or because we cross the
 853  * max_out threshold.
 854  *
 855  * total_out is used to return the total number of compressed bytes
 856  *
 857  * max_out tells us the max number of bytes that we're allowed to
 858  * stuff into pages
 859  */
 860 int btrfs_compress_pages(int type, struct address_space *mapping,
 861                          u64 start, unsigned long len,
 862                          struct page **pages,
 863                          unsigned long nr_dest_pages,
 864                          unsigned long *out_pages,
 865                          unsigned long *total_in,
 866                          unsigned long *total_out,
 867                          unsigned long max_out)
 868 {
 869         struct list_head *workspace;
 870         int ret;
 871
 872         workspace = find_workspace(type);
 873         if (IS_ERR(workspace))
 874                 return -1;
 875
 876         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 877                                                       start, len, pages,
 878                                                       nr_dest_pages, out_pages,
 879                                                       total_in, total_out,
 880                                                       max_out);
 881         free_workspace(type, workspace);
 882         return ret;
 883 }
 884
 885 /*
 886  * pages_in is an array of pages with compressed data.
 887  *
 888  * disk_start is the starting logical offset of this array in the file
 889  *
 890  * bvec is a bio_vec of pages from the file that we want to decompress into
 891  *
 892  * vcnt is the count of pages in the biovec
 893  *
 894  * srclen is the number of bytes in pages_in
 895  *
 896  * The basic idea is that we have a bio that was created by readpages.
 897  * The pages in the bio are for the uncompressed data, and they may not
 898  * be contiguous.  They all correspond to the range of bytes covered by
 899  * the compressed extent.
 900  */
 901 int btrfs_decompress_biovec(int type, struct page **pages_in, u64 disk_start,
 902                             struct bio_vec *bvec, int vcnt, size_t srclen)
 903 {
 904         struct list_head *workspace;
 905         int ret;
 906
 907         workspace = find_workspace(type);
 908         if (IS_ERR(workspace))
 909                 return -ENOMEM;
 910
 911         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 912                                                          disk_start,
 913                                                          bvec, vcnt, srclen);
 914         free_workspace(type, workspace);
 915         return ret;
 916 }
 917
 918 /*
 919  * a less complex decompression routine.  Our compressed data fits in a
 920  * single page, and we want to read a single page out of it.
 921  * start_byte tells us the offset into the compressed data we're interested in
 922  */
 923 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 924                      unsigned long start_byte, size_t srclen, size_t destlen)
 925 {
 926         struct list_head *workspace;
 927         int ret;
 928
 929         workspace = find_workspace(type);
 930         if (IS_ERR(workspace))
 931                 return -ENOMEM;
 932
 933         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 934                                                   dest_page, start_byte,
 935                                                   srclen, destlen);
 936
 937         free_workspace(type, workspace);
 938         return ret;
 939 }
 940
 941 void btrfs_exit_compress(void)
 942 {
 943         free_workspaces();
 944 }
 945
 946 /*
 947  * Copy uncompressed data from working buffer to pages.
 948  *
 949  * buf_start is the byte offset we're of the start of our workspace buffer.
 950  *
 951  * total_out is the last byte of the buffer
 952  */
 953 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 954                               unsigned long total_out, u64 disk_start,
 955                               struct bio_vec *bvec, int vcnt,
 956                               unsigned long *pg_index,
 957                               unsigned long *pg_offset)
 958 {
 959         unsigned long buf_offset;
 960         unsigned long current_buf_start;
 961         unsigned long start_byte;
 962         unsigned long working_bytes = total_out - buf_start;
 963         unsigned long bytes;
 964         char *kaddr;
 965         struct page *page_out = bvec[*pg_index].bv_page;
 966
 967         /*
 968          * start byte is the first byte of the page we're currently
 969          * copying into relative to the start of the compressed data.
 970          */
 971         start_byte = page_offset(page_out) - disk_start;
 972
 973         /* we haven't yet hit data corresponding to this page */
 974         if (total_out <= start_byte)
 975                 return 1;
 976
 977         /*
 978          * the start of the data we care about is offset into
 979          * the middle of our working buffer
 980          */
 981         if (total_out > start_byte && buf_start < start_byte) {
 982                 buf_offset = start_byte - buf_start;
 983                 working_bytes -= buf_offset;
 984         } else {
 985                 buf_offset = 0;
 986         }
 987         current_buf_start = buf_start;
 988
 989         /* copy bytes from the working buffer into the pages */
 990         while (working_bytes > 0) {
 991                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
 992                             PAGE_CACHE_SIZE - buf_offset);
 993                 bytes = min(bytes, working_bytes);
 994                 kaddr = kmap_atomic(page_out, KM_USER0);
 995                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
 996                 kunmap_atomic(kaddr, KM_USER0);
 997                 flush_dcache_page(page_out);
 998
 999                 *pg_offset += bytes;
1000                 buf_offset += bytes;
1001                 working_bytes -= bytes;
1002                 current_buf_start += bytes;
1003
1004                 /* check if we need to pick another page */
1005                 if (*pg_offset == PAGE_CACHE_SIZE) {
1006                         (*pg_index)++;
1007                         if (*pg_index >= vcnt)
1008                                 return 0;
1009
1010                         page_out = bvec[*pg_index].bv_page;
1011                         *pg_offset = 0;
1012                         start_byte = page_offset(page_out) - disk_start;
1013
1014                         /*
1015                          * make sure our new page is covered by this
1016                          * working buffer
1017                          */
1018                         if (total_out <= start_byte)
1019                                 return 1;
1020
1021                         /*
1022                          * the next page in the biovec might not be adjacent
1023                          * to the last page, but it might still be found
1024                          * inside this working buffer. bump our offset pointer
1025                          */
1026                         if (total_out > start_byte &&
1027                             current_buf_start < start_byte) {
1028                                 buf_offset = start_byte - buf_start;
1029                                 working_bytes = total_out - start_byte;
1030                                 current_buf_start = buf_start + buf_offset;
1031                         }
1032                 }
1033         }
1034
1035         return 1;
1036 }