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