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 "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "volumes.h"
  40 #include "ordered-data.h"
  41 #include "compression.h"
  42 #include "extent_io.h"
  43 #include "extent_map.h"
  44
  45 struct compressed_bio {
  46         /* number of bios pending for this compressed extent */
  47         atomic_t pending_bios;
  48
  49         /* the pages with the compressed data on them */
  50         struct page **compressed_pages;
  51
  52         /* inode that owns this data */
  53         struct inode *inode;
  54
  55         /* starting offset in the inode for our pages */
  56         u64 start;
  57
  58         /* number of bytes in the inode we're working on */
  59         unsigned long len;
  60
  61         /* number of bytes on disk */
  62         unsigned long compressed_len;
  63
  64         /* the compression algorithm for this bio */
  65         int compress_type;
  66
  67         /* number of compressed pages in the array */
  68         unsigned long nr_pages;
  69
  70         /* IO errors */
  71         int errors;
  72         int mirror_num;
  73
  74         /* for reads, this is the bio we are copying the data into */
  75         struct bio *orig_bio;
  76
  77         /*
  78          * the start of a variable length array of checksums only
  79          * used by reads
  80          */
  81         u32 sums;
  82 };
  83
  84 static int btrfs_decompress_biovec(int type, struct page **pages_in,
  85                                    u64 disk_start, struct bio_vec *bvec,
  86                                    int vcnt, size_t srclen);
  87
  88 static inline int compressed_bio_size(struct btrfs_root *root,
  89                                       unsigned long disk_size)
  90 {
  91         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
  92
  93         return sizeof(struct compressed_bio) +
  94                 ((disk_size + root->sectorsize - 1) / root->sectorsize) *
  95                 csum_size;
  96 }
  97
  98 static struct bio *compressed_bio_alloc(struct block_device *bdev,
  99                                         u64 first_byte, gfp_t gfp_flags)
 100 {
 101         int nr_vecs;
 102
 103         nr_vecs = bio_get_nr_vecs(bdev);
 104         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
 105 }
 106
 107 static int check_compressed_csum(struct inode *inode,
 108                                  struct compressed_bio *cb,
 109                                  u64 disk_start)
 110 {
 111         int ret;
 112         struct page *page;
 113         unsigned long i;
 114         char *kaddr;
 115         u32 csum;
 116         u32 *cb_sum = &cb->sums;
 117
 118         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 119                 return 0;
 120
 121         for (i = 0; i < cb->nr_pages; i++) {
 122                 page = cb->compressed_pages[i];
 123                 csum = ~(u32)0;
 124
 125                 kaddr = kmap_atomic(page);
 126                 csum = btrfs_csum_data(kaddr, csum, PAGE_CACHE_SIZE);
 127                 btrfs_csum_final(csum, (char *)&csum);
 128                 kunmap_atomic(kaddr);
 129
 130                 if (csum != *cb_sum) {
 131                         printk(KERN_INFO "btrfs csum failed ino %llu "
 132                                "extent %llu csum %u "
 133                                "wanted %u mirror %d\n",
 134                                btrfs_ino(inode), disk_start, csum, *cb_sum,
 135                                cb->mirror_num);
 136                         ret = -EIO;
 137                         goto fail;
 138                 }
 139                 cb_sum++;
 140
 141         }
 142         ret = 0;
 143 fail:
 144         return ret;
 145 }
 146
 147 /* when we finish reading compressed pages from the disk, we
 148  * decompress them and then run the bio end_io routines on the
 149  * decompressed pages (in the inode address space).
 150  *
 151  * This allows the checksumming and other IO error handling routines
 152  * to work normally
 153  *
 154  * The compressed pages are freed here, and it must be run
 155  * in process context
 156  */
 157 static void end_compressed_bio_read(struct bio *bio, int err)
 158 {
 159         struct compressed_bio *cb = bio->bi_private;
 160         struct inode *inode;
 161         struct page *page;
 162         unsigned long index;
 163         int ret;
 164
 165         if (err)
 166                 cb->errors = 1;
 167
 168         /* if there are more bios still pending for this compressed
 169          * extent, just exit
 170          */
 171         if (!atomic_dec_and_test(&cb->pending_bios))
 172                 goto out;
 173
 174         inode = cb->inode;
 175         ret = check_compressed_csum(inode, cb, (u64)bio->bi_sector << 9);
 176         if (ret)
 177                 goto csum_failed;
 178
 179         /* ok, we're the last bio for this extent, lets start
 180          * the decompression.
 181          */
 182         ret = btrfs_decompress_biovec(cb->compress_type,
 183                                       cb->compressed_pages,
 184                                       cb->start,
 185                                       cb->orig_bio->bi_io_vec,
 186                                       cb->orig_bio->bi_vcnt,
 187                                       cb->compressed_len);
 188 csum_failed:
 189         if (ret)
 190                 cb->errors = 1;
 191
 192         /* release the compressed pages */
 193         index = 0;
 194         for (index = 0; index < cb->nr_pages; index++) {
 195                 page = cb->compressed_pages[index];
 196                 page->mapping = NULL;
 197                 page_cache_release(page);
 198         }
 199
 200         /* do io completion on the original bio */
 201         if (cb->errors) {
 202                 bio_io_error(cb->orig_bio);
 203         } else {
 204                 int bio_index = 0;
 205                 struct bio_vec *bvec = cb->orig_bio->bi_io_vec;
 206
 207                 /*
 208                  * we have verified the checksum already, set page
 209                  * checked so the end_io handlers know about it
 210                  */
 211                 while (bio_index < cb->orig_bio->bi_vcnt) {
 212                         SetPageChecked(bvec->bv_page);
 213                         bvec++;
 214                         bio_index++;
 215                 }
 216                 bio_endio(cb->orig_bio, 0);
 217         }
 218
 219         /* finally free the cb struct */
 220         kfree(cb->compressed_pages);
 221         kfree(cb);
 222 out:
 223         bio_put(bio);
 224 }
 225
 226 /*
 227  * Clear the writeback bits on all of the file
 228  * pages for a compressed write
 229  */
 230 static noinline void end_compressed_writeback(struct inode *inode, u64 start,
 231                                               unsigned long ram_size)
 232 {
 233         unsigned long index = start >> PAGE_CACHE_SHIFT;
 234         unsigned long end_index = (start + ram_size - 1) >> PAGE_CACHE_SHIFT;
 235         struct page *pages[16];
 236         unsigned long nr_pages = end_index - index + 1;
 237         int i;
 238         int ret;
 239
 240         while (nr_pages > 0) {
 241                 ret = find_get_pages_contig(inode->i_mapping, index,
 242                                      min_t(unsigned long,
 243                                      nr_pages, ARRAY_SIZE(pages)), pages);
 244                 if (ret == 0) {
 245                         nr_pages -= 1;
 246                         index += 1;
 247                         continue;
 248                 }
 249                 for (i = 0; i < ret; i++) {
 250                         end_page_writeback(pages[i]);
 251                         page_cache_release(pages[i]);
 252                 }
 253                 nr_pages -= ret;
 254                 index += ret;
 255         }
 256         /* the inode may be gone now */
 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(WRITE, 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); /* -ENOMEM */
 396
 397                         if (!skip_sum) {
 398                                 ret = btrfs_csum_one_bio(root, inode, bio,
 399                                                          start, 1);
 400                                 BUG_ON(ret); /* -ENOMEM */
 401                         }
 402
 403                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 404                         BUG_ON(ret); /* -ENOMEM */
 405
 406                         bio_put(bio);
 407
 408                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 409                         BUG_ON(!bio);
 410                         bio->bi_private = cb;
 411                         bio->bi_end_io = end_compressed_bio_write;
 412                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 413                 }
 414                 if (bytes_left < PAGE_CACHE_SIZE) {
 415                         printk("bytes left %lu compress len %lu nr %lu\n",
 416                                bytes_left, cb->compressed_len, cb->nr_pages);
 417                 }
 418                 bytes_left -= PAGE_CACHE_SIZE;
 419                 first_byte += PAGE_CACHE_SIZE;
 420                 cond_resched();
 421         }
 422         bio_get(bio);
 423
 424         ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
 425         BUG_ON(ret); /* -ENOMEM */
 426
 427         if (!skip_sum) {
 428                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 429                 BUG_ON(ret); /* -ENOMEM */
 430         }
 431
 432         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 433         BUG_ON(ret); /* -ENOMEM */
 434
 435         bio_put(bio);
 436         return 0;
 437 }
 438
 439 static noinline int add_ra_bio_pages(struct inode *inode,
 440                                      u64 compressed_end,
 441                                      struct compressed_bio *cb)
 442 {
 443         unsigned long end_index;
 444         unsigned long pg_index;
 445         u64 last_offset;
 446         u64 isize = i_size_read(inode);
 447         int ret;
 448         struct page *page;
 449         unsigned long nr_pages = 0;
 450         struct extent_map *em;
 451         struct address_space *mapping = inode->i_mapping;
 452         struct extent_map_tree *em_tree;
 453         struct extent_io_tree *tree;
 454         u64 end;
 455         int misses = 0;
 456
 457         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 458         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 459         em_tree = &BTRFS_I(inode)->extent_tree;
 460         tree = &BTRFS_I(inode)->io_tree;
 461
 462         if (isize == 0)
 463                 return 0;
 464
 465         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 466
 467         while (last_offset < compressed_end) {
 468                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 469
 470                 if (pg_index > end_index)
 471                         break;
 472
 473                 rcu_read_lock();
 474                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 475                 rcu_read_unlock();
 476                 if (page) {
 477                         misses++;
 478                         if (misses > 4)
 479                                 break;
 480                         goto next;
 481                 }
 482
 483                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 484                                                                 ~__GFP_FS);
 485                 if (!page)
 486                         break;
 487
 488                 if (add_to_page_cache_lru(page, mapping, pg_index,
 489                                                                 GFP_NOFS)) {
 490                         page_cache_release(page);
 491                         goto next;
 492                 }
 493
 494                 end = last_offset + PAGE_CACHE_SIZE - 1;
 495                 /*
 496                  * at this point, we have a locked page in the page cache
 497                  * for these bytes in the file.  But, we have to make
 498                  * sure they map to this compressed extent on disk.
 499                  */
 500                 set_page_extent_mapped(page);
 501                 lock_extent(tree, last_offset, end);
 502                 read_lock(&em_tree->lock);
 503                 em = lookup_extent_mapping(em_tree, last_offset,
 504                                            PAGE_CACHE_SIZE);
 505                 read_unlock(&em_tree->lock);
 506
 507                 if (!em || last_offset < em->start ||
 508                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 509                     (em->block_start >> 9) != cb->orig_bio->bi_sector) {
 510                         free_extent_map(em);
 511                         unlock_extent(tree, last_offset, end);
 512                         unlock_page(page);
 513                         page_cache_release(page);
 514                         break;
 515                 }
 516                 free_extent_map(em);
 517
 518                 if (page->index == end_index) {
 519                         char *userpage;
 520                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 521
 522                         if (zero_offset) {
 523                                 int zeros;
 524                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 525                                 userpage = kmap_atomic(page);
 526                                 memset(userpage + zero_offset, 0, zeros);
 527                                 flush_dcache_page(page);
 528                                 kunmap_atomic(userpage);
 529                         }
 530                 }
 531
 532                 ret = bio_add_page(cb->orig_bio, page,
 533                                    PAGE_CACHE_SIZE, 0);
 534
 535                 if (ret == PAGE_CACHE_SIZE) {
 536                         nr_pages++;
 537                         page_cache_release(page);
 538                 } else {
 539                         unlock_extent(tree, last_offset, end);
 540                         unlock_page(page);
 541                         page_cache_release(page);
 542                         break;
 543                 }
 544 next:
 545                 last_offset += PAGE_CACHE_SIZE;
 546         }
 547         return 0;
 548 }
 549
 550 /*
 551  * for a compressed read, the bio we get passed has all the inode pages
 552  * in it.  We don't actually do IO on those pages but allocate new ones
 553  * to hold the compressed pages on disk.
 554  *
 555  * bio->bi_sector points to the compressed extent on disk
 556  * bio->bi_io_vec points to all of the inode pages
 557  * bio->bi_vcnt is a count of pages
 558  *
 559  * After the compressed pages are read, we copy the bytes into the
 560  * bio we were passed and then call the bio end_io calls
 561  */
 562 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 563                                  int mirror_num, unsigned long bio_flags)
 564 {
 565         struct extent_io_tree *tree;
 566         struct extent_map_tree *em_tree;
 567         struct compressed_bio *cb;
 568         struct btrfs_root *root = BTRFS_I(inode)->root;
 569         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 570         unsigned long compressed_len;
 571         unsigned long nr_pages;
 572         unsigned long pg_index;
 573         struct page *page;
 574         struct block_device *bdev;
 575         struct bio *comp_bio;
 576         u64 cur_disk_byte = (u64)bio->bi_sector << 9;
 577         u64 em_len;
 578         u64 em_start;
 579         struct extent_map *em;
 580         int ret = -ENOMEM;
 581         int faili = 0;
 582         u32 *sums;
 583
 584         tree = &BTRFS_I(inode)->io_tree;
 585         em_tree = &BTRFS_I(inode)->extent_tree;
 586
 587         /* we need the actual starting offset of this extent in the file */
 588         read_lock(&em_tree->lock);
 589         em = lookup_extent_mapping(em_tree,
 590                                    page_offset(bio->bi_io_vec->bv_page),
 591                                    PAGE_CACHE_SIZE);
 592         read_unlock(&em_tree->lock);
 593         if (!em)
 594                 return -EIO;
 595
 596         compressed_len = em->block_len;
 597         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 598         if (!cb)
 599                 goto out;
 600
 601         atomic_set(&cb->pending_bios, 0);
 602         cb->errors = 0;
 603         cb->inode = inode;
 604         cb->mirror_num = mirror_num;
 605         sums = &cb->sums;
 606
 607         cb->start = em->orig_start;
 608         em_len = em->len;
 609         em_start = em->start;
 610
 611         free_extent_map(em);
 612         em = NULL;
 613
 614         cb->len = uncompressed_len;
 615         cb->compressed_len = compressed_len;
 616         cb->compress_type = extent_compress_type(bio_flags);
 617         cb->orig_bio = bio;
 618
 619         nr_pages = (compressed_len + PAGE_CACHE_SIZE - 1) /
 620                                  PAGE_CACHE_SIZE;
 621         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 622                                        GFP_NOFS);
 623         if (!cb->compressed_pages)
 624                 goto fail1;
 625
 626         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 627
 628         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 629                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 630                                                               __GFP_HIGHMEM);
 631                 if (!cb->compressed_pages[pg_index]) {
 632                         faili = pg_index - 1;
 633                         ret = -ENOMEM;
 634                         goto fail2;
 635                 }
 636         }
 637         faili = nr_pages - 1;
 638         cb->nr_pages = nr_pages;
 639
 640         /* In the parent-locked case, we only locked the range we are
 641          * interested in.  In all other cases, we can opportunistically
 642          * cache decompressed data that goes beyond the requested range. */
 643         if (!(bio_flags & EXTENT_BIO_PARENT_LOCKED))
 644                 add_ra_bio_pages(inode, em_start + em_len, cb);
 645
 646         /* include any pages we added in add_ra-bio_pages */
 647         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 648         cb->len = uncompressed_len;
 649
 650         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 651         if (!comp_bio)
 652                 goto fail2;
 653         comp_bio->bi_private = cb;
 654         comp_bio->bi_end_io = end_compressed_bio_read;
 655         atomic_inc(&cb->pending_bios);
 656
 657         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 658                 page = cb->compressed_pages[pg_index];
 659                 page->mapping = inode->i_mapping;
 660                 page->index = em_start >> PAGE_CACHE_SHIFT;
 661
 662                 if (comp_bio->bi_size)
 663                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 664                                                         PAGE_CACHE_SIZE,
 665                                                         comp_bio, 0);
 666                 else
 667                         ret = 0;
 668
 669                 page->mapping = NULL;
 670                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 671                     PAGE_CACHE_SIZE) {
 672                         bio_get(comp_bio);
 673
 674                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 675                         BUG_ON(ret); /* -ENOMEM */
 676
 677                         /*
 678                          * inc the count before we submit the bio so
 679                          * we know the end IO handler won't happen before
 680                          * we inc the count.  Otherwise, the cb might get
 681                          * freed before we're done setting it up
 682                          */
 683                         atomic_inc(&cb->pending_bios);
 684
 685                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 686                                 ret = btrfs_lookup_bio_sums(root, inode,
 687                                                         comp_bio, sums);
 688                                 BUG_ON(ret); /* -ENOMEM */
 689                         }
 690                         sums += (comp_bio->bi_size + root->sectorsize - 1) /
 691                                 root->sectorsize;
 692
 693                         ret = btrfs_map_bio(root, READ, comp_bio,
 694                                             mirror_num, 0);
 695                         if (ret)
 696                                 bio_endio(comp_bio, ret);
 697
 698                         bio_put(comp_bio);
 699
 700                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 701                                                         GFP_NOFS);
 702                         BUG_ON(!comp_bio);
 703                         comp_bio->bi_private = cb;
 704                         comp_bio->bi_end_io = end_compressed_bio_read;
 705
 706                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 707                 }
 708                 cur_disk_byte += PAGE_CACHE_SIZE;
 709         }
 710         bio_get(comp_bio);
 711
 712         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio, 0);
 713         BUG_ON(ret); /* -ENOMEM */
 714
 715         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 716                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 717                 BUG_ON(ret); /* -ENOMEM */
 718         }
 719
 720         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 721         if (ret)
 722                 bio_endio(comp_bio, ret);
 723
 724         bio_put(comp_bio);
 725         return 0;
 726
 727 fail2:
 728         while (faili >= 0) {
 729                 __free_page(cb->compressed_pages[faili]);
 730                 faili--;
 731         }
 732
 733         kfree(cb->compressed_pages);
 734 fail1:
 735         kfree(cb);
 736 out:
 737         free_extent_map(em);
 738         return ret;
 739 }
 740
 741 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 742 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 743 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 744 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 745 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 746
 747 static struct btrfs_compress_op *btrfs_compress_op[] = {
 748         &btrfs_zlib_compress,
 749         &btrfs_lzo_compress,
 750 };
 751
 752 void __init btrfs_init_compress(void)
 753 {
 754         int i;
 755
 756         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 757                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 758                 spin_lock_init(&comp_workspace_lock[i]);
 759                 atomic_set(&comp_alloc_workspace[i], 0);
 760                 init_waitqueue_head(&comp_workspace_wait[i]);
 761         }
 762 }
 763
 764 /*
 765  * this finds an available workspace or allocates a new one
 766  * ERR_PTR is returned if things go bad.
 767  */
 768 static struct list_head *find_workspace(int type)
 769 {
 770         struct list_head *workspace;
 771         int cpus = num_online_cpus();
 772         int idx = type - 1;
 773
 774         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 775         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 776         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 777         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 778         int *num_workspace                      = &comp_num_workspace[idx];
 779 again:
 780         spin_lock(workspace_lock);
 781         if (!list_empty(idle_workspace)) {
 782                 workspace = idle_workspace->next;
 783                 list_del(workspace);
 784                 (*num_workspace)--;
 785                 spin_unlock(workspace_lock);
 786                 return workspace;
 787
 788         }
 789         if (atomic_read(alloc_workspace) > cpus) {
 790                 DEFINE_WAIT(wait);
 791
 792                 spin_unlock(workspace_lock);
 793                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 794                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 795                         schedule();
 796                 finish_wait(workspace_wait, &wait);
 797                 goto again;
 798         }
 799         atomic_inc(alloc_workspace);
 800         spin_unlock(workspace_lock);
 801
 802         workspace = btrfs_compress_op[idx]->alloc_workspace();
 803         if (IS_ERR(workspace)) {
 804                 atomic_dec(alloc_workspace);
 805                 wake_up(workspace_wait);
 806         }
 807         return workspace;
 808 }
 809
 810 /*
 811  * put a workspace struct back on the list or free it if we have enough
 812  * idle ones sitting around
 813  */
 814 static void free_workspace(int type, struct list_head *workspace)
 815 {
 816         int idx = type - 1;
 817         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 818         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 819         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 820         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 821         int *num_workspace                      = &comp_num_workspace[idx];
 822
 823         spin_lock(workspace_lock);
 824         if (*num_workspace < num_online_cpus()) {
 825                 list_add_tail(workspace, idle_workspace);
 826                 (*num_workspace)++;
 827                 spin_unlock(workspace_lock);
 828                 goto wake;
 829         }
 830         spin_unlock(workspace_lock);
 831
 832         btrfs_compress_op[idx]->free_workspace(workspace);
 833         atomic_dec(alloc_workspace);
 834 wake:
 835         smp_mb();
 836         if (waitqueue_active(workspace_wait))
 837                 wake_up(workspace_wait);
 838 }
 839
 840 /*
 841  * cleanup function for module exit
 842  */
 843 static void free_workspaces(void)
 844 {
 845         struct list_head *workspace;
 846         int i;
 847
 848         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 849                 while (!list_empty(&comp_idle_workspace[i])) {
 850                         workspace = comp_idle_workspace[i].next;
 851                         list_del(workspace);
 852                         btrfs_compress_op[i]->free_workspace(workspace);
 853                         atomic_dec(&comp_alloc_workspace[i]);
 854                 }
 855         }
 856 }
 857
 858 /*
 859  * given an address space and start/len, compress the bytes.
 860  *
 861  * pages are allocated to hold the compressed result and stored
 862  * in 'pages'
 863  *
 864  * out_pages is used to return the number of pages allocated.  There
 865  * may be pages allocated even if we return an error
 866  *
 867  * total_in is used to return the number of bytes actually read.  It
 868  * may be smaller then len if we had to exit early because we
 869  * ran out of room in the pages array or because we cross the
 870  * max_out threshold.
 871  *
 872  * total_out is used to return the total number of compressed bytes
 873  *
 874  * max_out tells us the max number of bytes that we're allowed to
 875  * stuff into pages
 876  */
 877 int btrfs_compress_pages(int type, struct address_space *mapping,
 878                          u64 start, unsigned long len,
 879                          struct page **pages,
 880                          unsigned long nr_dest_pages,
 881                          unsigned long *out_pages,
 882                          unsigned long *total_in,
 883                          unsigned long *total_out,
 884                          unsigned long max_out)
 885 {
 886         struct list_head *workspace;
 887         int ret;
 888
 889         workspace = find_workspace(type);
 890         if (IS_ERR(workspace))
 891                 return -1;
 892
 893         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 894                                                       start, len, pages,
 895                                                       nr_dest_pages, out_pages,
 896                                                       total_in, total_out,
 897                                                       max_out);
 898         free_workspace(type, workspace);
 899         return ret;
 900 }
 901
 902 /*
 903  * pages_in is an array of pages with compressed data.
 904  *
 905  * disk_start is the starting logical offset of this array in the file
 906  *
 907  * bvec is a bio_vec of pages from the file that we want to decompress into
 908  *
 909  * vcnt is the count of pages in the biovec
 910  *
 911  * srclen is the number of bytes in pages_in
 912  *
 913  * The basic idea is that we have a bio that was created by readpages.
 914  * The pages in the bio are for the uncompressed data, and they may not
 915  * be contiguous.  They all correspond to the range of bytes covered by
 916  * the compressed extent.
 917  */
 918 static int btrfs_decompress_biovec(int type, struct page **pages_in,
 919                                    u64 disk_start, struct bio_vec *bvec,
 920                                    int vcnt, size_t srclen)
 921 {
 922         struct list_head *workspace;
 923         int ret;
 924
 925         workspace = find_workspace(type);
 926         if (IS_ERR(workspace))
 927                 return -ENOMEM;
 928
 929         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 930                                                          disk_start,
 931                                                          bvec, vcnt, srclen);
 932         free_workspace(type, workspace);
 933         return ret;
 934 }
 935
 936 /*
 937  * a less complex decompression routine.  Our compressed data fits in a
 938  * single page, and we want to read a single page out of it.
 939  * start_byte tells us the offset into the compressed data we're interested in
 940  */
 941 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 942                      unsigned long start_byte, size_t srclen, size_t destlen)
 943 {
 944         struct list_head *workspace;
 945         int ret;
 946
 947         workspace = find_workspace(type);
 948         if (IS_ERR(workspace))
 949                 return -ENOMEM;
 950
 951         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 952                                                   dest_page, start_byte,
 953                                                   srclen, destlen);
 954
 955         free_workspace(type, workspace);
 956         return ret;
 957 }
 958
 959 void btrfs_exit_compress(void)
 960 {
 961         free_workspaces();
 962 }
 963
 964 /*
 965  * Copy uncompressed data from working buffer to pages.
 966  *
 967  * buf_start is the byte offset we're of the start of our workspace buffer.
 968  *
 969  * total_out is the last byte of the buffer
 970  */
 971 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 972                               unsigned long total_out, u64 disk_start,
 973                               struct bio_vec *bvec, int vcnt,
 974                               unsigned long *pg_index,
 975                               unsigned long *pg_offset)
 976 {
 977         unsigned long buf_offset;
 978         unsigned long current_buf_start;
 979         unsigned long start_byte;
 980         unsigned long working_bytes = total_out - buf_start;
 981         unsigned long bytes;
 982         char *kaddr;
 983         struct page *page_out = bvec[*pg_index].bv_page;
 984
 985         /*
 986          * start byte is the first byte of the page we're currently
 987          * copying into relative to the start of the compressed data.
 988          */
 989         start_byte = page_offset(page_out) - disk_start;
 990
 991         /* we haven't yet hit data corresponding to this page */
 992         if (total_out <= start_byte)
 993                 return 1;
 994
 995         /*
 996          * the start of the data we care about is offset into
 997          * the middle of our working buffer
 998          */
 999         if (total_out > start_byte && buf_start < start_byte) {
1000                 buf_offset = start_byte - buf_start;
1001                 working_bytes -= buf_offset;
1002         } else {
1003                 buf_offset = 0;
1004         }
1005         current_buf_start = buf_start;
1006
1007         /* copy bytes from the working buffer into the pages */
1008         while (working_bytes > 0) {
1009                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
1010                             PAGE_CACHE_SIZE - buf_offset);
1011                 bytes = min(bytes, working_bytes);
1012                 kaddr = kmap_atomic(page_out);
1013                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1014                 kunmap_atomic(kaddr);
1015                 flush_dcache_page(page_out);
1016
1017                 *pg_offset += bytes;
1018                 buf_offset += bytes;
1019                 working_bytes -= bytes;
1020                 current_buf_start += bytes;
1021
1022                 /* check if we need to pick another page */
1023                 if (*pg_offset == PAGE_CACHE_SIZE) {
1024                         (*pg_index)++;
1025                         if (*pg_index >= vcnt)
1026                                 return 0;
1027
1028                         page_out = bvec[*pg_index].bv_page;
1029                         *pg_offset = 0;
1030                         start_byte = page_offset(page_out) - disk_start;
1031
1032                         /*
1033                          * make sure our new page is covered by this
1034                          * working buffer
1035                          */
1036                         if (total_out <= start_byte)
1037                                 return 1;
1038
1039                         /*
1040                          * the next page in the biovec might not be adjacent
1041                          * to the last page, but it might still be found
1042                          * inside this working buffer. bump our offset pointer
1043                          */
1044                         if (total_out > start_byte &&
1045                             current_buf_start < start_byte) {
1046                                 buf_offset = start_byte - buf_start;
1047                                 working_bytes = total_out - start_byte;
1048                                 current_buf_start = buf_start + buf_offset;
1049                         }
1050                 }
1051         }
1052
1053         return 1;
1054 }