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                 (DIV_ROUND_UP(disk_size, root->sectorsize)) * csum_size;
  95 }
  96
  97 static struct bio *compressed_bio_alloc(struct block_device *bdev,
  98                                         u64 first_byte, gfp_t gfp_flags)
  99 {
 100         int nr_vecs;
 101
 102         nr_vecs = bio_get_nr_vecs(bdev);
 103         return btrfs_bio_alloc(bdev, first_byte >> 9, nr_vecs, gfp_flags);
 104 }
 105
 106 static int check_compressed_csum(struct inode *inode,
 107                                  struct compressed_bio *cb,
 108                                  u64 disk_start)
 109 {
 110         int ret;
 111         struct page *page;
 112         unsigned long i;
 113         char *kaddr;
 114         u32 csum;
 115         u32 *cb_sum = &cb->sums;
 116
 117         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 118                 return 0;
 119
 120         for (i = 0; i < cb->nr_pages; i++) {
 121                 page = cb->compressed_pages[i];
 122                 csum = ~(u32)0;
 123
 124                 kaddr = kmap_atomic(page);
 125                 csum = btrfs_csum_data(kaddr, csum, PAGE_CACHE_SIZE);
 126                 btrfs_csum_final(csum, (char *)&csum);
 127                 kunmap_atomic(kaddr);
 128
 129                 if (csum != *cb_sum) {
 130                         btrfs_info(BTRFS_I(inode)->root->fs_info,
 131                            "csum failed ino %llu extent %llu csum %u wanted %u mirror %d",
 132                            btrfs_ino(inode), disk_start, csum, *cb_sum,
 133                            cb->mirror_num);
 134                         ret = -EIO;
 135                         goto fail;
 136                 }
 137                 cb_sum++;
 138
 139         }
 140         ret = 0;
 141 fail:
 142         return ret;
 143 }
 144
 145 /* when we finish reading compressed pages from the disk, we
 146  * decompress them and then run the bio end_io routines on the
 147  * decompressed pages (in the inode address space).
 148  *
 149  * This allows the checksumming and other IO error handling routines
 150  * to work normally
 151  *
 152  * The compressed pages are freed here, and it must be run
 153  * in process context
 154  */
 155 static void end_compressed_bio_read(struct bio *bio, int err)
 156 {
 157         struct compressed_bio *cb = bio->bi_private;
 158         struct inode *inode;
 159         struct page *page;
 160         unsigned long index;
 161         int ret;
 162
 163         if (err)
 164                 cb->errors = 1;
 165
 166         /* if there are more bios still pending for this compressed
 167          * extent, just exit
 168          */
 169         if (!atomic_dec_and_test(&cb->pending_bios))
 170                 goto out;
 171
 172         inode = cb->inode;
 173         ret = check_compressed_csum(inode, cb,
 174                                     (u64)bio->bi_iter.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 i;
 204                 struct bio_vec *bvec;
 205
 206                 /*
 207                  * we have verified the checksum already, set page
 208                  * checked so the end_io handlers know about it
 209                  */
 210                 bio_for_each_segment_all(bvec, cb->orig_bio, i)
 211                         SetPageChecked(bvec->bv_page);
 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 void 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 }
 255
 256 /*
 257  * do the cleanup once all the compressed pages hit the disk.
 258  * This will clear writeback on the file pages and free the compressed
 259  * pages.
 260  *
 261  * This also calls the writeback end hooks for the file pages so that
 262  * metadata and checksums can be updated in the file.
 263  */
 264 static void end_compressed_bio_write(struct bio *bio, int err)
 265 {
 266         struct extent_io_tree *tree;
 267         struct compressed_bio *cb = bio->bi_private;
 268         struct inode *inode;
 269         struct page *page;
 270         unsigned long index;
 271
 272         if (err)
 273                 cb->errors = 1;
 274
 275         /* if there are more bios still pending for this compressed
 276          * extent, just exit
 277          */
 278         if (!atomic_dec_and_test(&cb->pending_bios))
 279                 goto out;
 280
 281         /* ok, we're the last bio for this extent, step one is to
 282          * call back into the FS and do all the end_io operations
 283          */
 284         inode = cb->inode;
 285         tree = &BTRFS_I(inode)->io_tree;
 286         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 287         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 288                                          cb->start,
 289                                          cb->start + cb->len - 1,
 290                                          NULL, 1);
 291         cb->compressed_pages[0]->mapping = NULL;
 292
 293         end_compressed_writeback(inode, cb->start, cb->len);
 294         /* note, our inode could be gone now */
 295
 296         /*
 297          * release the compressed pages, these came from alloc_page and
 298          * are not attached to the inode at all
 299          */
 300         index = 0;
 301         for (index = 0; index < cb->nr_pages; index++) {
 302                 page = cb->compressed_pages[index];
 303                 page->mapping = NULL;
 304                 page_cache_release(page);
 305         }
 306
 307         /* finally free the cb struct */
 308         kfree(cb->compressed_pages);
 309         kfree(cb);
 310 out:
 311         bio_put(bio);
 312 }
 313
 314 /*
 315  * worker function to build and submit bios for previously compressed pages.
 316  * The corresponding pages in the inode should be marked for writeback
 317  * and the compressed pages should have a reference on them for dropping
 318  * when the IO is complete.
 319  *
 320  * This also checksums the file bytes and gets things ready for
 321  * the end io hooks.
 322  */
 323 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 324                                  unsigned long len, u64 disk_start,
 325                                  unsigned long compressed_len,
 326                                  struct page **compressed_pages,
 327                                  unsigned long nr_pages)
 328 {
 329         struct bio *bio = NULL;
 330         struct btrfs_root *root = BTRFS_I(inode)->root;
 331         struct compressed_bio *cb;
 332         unsigned long bytes_left;
 333         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 334         int pg_index = 0;
 335         struct page *page;
 336         u64 first_byte = disk_start;
 337         struct block_device *bdev;
 338         int ret;
 339         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 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 (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 370                 page = compressed_pages[pg_index];
 371                 page->mapping = inode->i_mapping;
 372                 if (bio->bi_iter.bi_size)
 373                         ret = io_tree->ops->merge_bio_hook(WRITE, 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,
 392                                         BTRFS_WQ_ENDIO_DATA);
 393                         BUG_ON(ret); /* -ENOMEM */
 394
 395                         if (!skip_sum) {
 396                                 ret = btrfs_csum_one_bio(root, inode, bio,
 397                                                          start, 1);
 398                                 BUG_ON(ret); /* -ENOMEM */
 399                         }
 400
 401                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 402                         BUG_ON(ret); /* -ENOMEM */
 403
 404                         bio_put(bio);
 405
 406                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 407                         BUG_ON(!bio);
 408                         bio->bi_private = cb;
 409                         bio->bi_end_io = end_compressed_bio_write;
 410                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 411                 }
 412                 if (bytes_left < PAGE_CACHE_SIZE) {
 413                         btrfs_info(BTRFS_I(inode)->root->fs_info,
 414                                         "bytes left %lu compress len %lu nr %lu",
 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, BTRFS_WQ_ENDIO_DATA);
 424         BUG_ON(ret); /* -ENOMEM */
 425
 426         if (!skip_sum) {
 427                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 428                 BUG_ON(ret); /* -ENOMEM */
 429         }
 430
 431         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 432         BUG_ON(ret); /* -ENOMEM */
 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 && !radix_tree_exceptional_entry(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);
 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_iter.bi_sector) {
 509                         free_extent_map(em);
 510                         unlock_extent(tree, last_offset, end);
 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);
 525                                 memset(userpage + zero_offset, 0, zeros);
 526                                 flush_dcache_page(page);
 527                                 kunmap_atomic(userpage);
 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);
 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_iter.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_iter.bi_sector << 9;
 576         u64 em_len;
 577         u64 em_start;
 578         struct extent_map *em;
 579         int ret = -ENOMEM;
 580         int faili = 0;
 581         u32 *sums;
 582
 583         tree = &BTRFS_I(inode)->io_tree;
 584         em_tree = &BTRFS_I(inode)->extent_tree;
 585
 586         /* we need the actual starting offset of this extent in the file */
 587         read_lock(&em_tree->lock);
 588         em = lookup_extent_mapping(em_tree,
 589                                    page_offset(bio->bi_io_vec->bv_page),
 590                                    PAGE_CACHE_SIZE);
 591         read_unlock(&em_tree->lock);
 592         if (!em)
 593                 return -EIO;
 594
 595         compressed_len = em->block_len;
 596         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 597         if (!cb)
 598                 goto out;
 599
 600         atomic_set(&cb->pending_bios, 0);
 601         cb->errors = 0;
 602         cb->inode = inode;
 603         cb->mirror_num = mirror_num;
 604         sums = &cb->sums;
 605
 606         cb->start = em->orig_start;
 607         em_len = em->len;
 608         em_start = em->start;
 609
 610         free_extent_map(em);
 611         em = NULL;
 612
 613         cb->len = uncompressed_len;
 614         cb->compressed_len = compressed_len;
 615         cb->compress_type = extent_compress_type(bio_flags);
 616         cb->orig_bio = bio;
 617
 618         nr_pages = DIV_ROUND_UP(compressed_len, PAGE_CACHE_SIZE);
 619         cb->compressed_pages = kzalloc(sizeof(struct page *) * nr_pages,
 620                                        GFP_NOFS);
 621         if (!cb->compressed_pages)
 622                 goto fail1;
 623
 624         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 625
 626         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 627                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 628                                                               __GFP_HIGHMEM);
 629                 if (!cb->compressed_pages[pg_index]) {
 630                         faili = pg_index - 1;
 631                         ret = -ENOMEM;
 632                         goto fail2;
 633                 }
 634         }
 635         faili = nr_pages - 1;
 636         cb->nr_pages = nr_pages;
 637
 638         /* In the parent-locked case, we only locked the range we are
 639          * interested in.  In all other cases, we can opportunistically
 640          * cache decompressed data that goes beyond the requested range. */
 641         if (!(bio_flags & EXTENT_BIO_PARENT_LOCKED))
 642                 add_ra_bio_pages(inode, em_start + em_len, cb);
 643
 644         /* include any pages we added in add_ra-bio_pages */
 645         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 646         cb->len = uncompressed_len;
 647
 648         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 649         if (!comp_bio)
 650                 goto fail2;
 651         comp_bio->bi_private = cb;
 652         comp_bio->bi_end_io = end_compressed_bio_read;
 653         atomic_inc(&cb->pending_bios);
 654
 655         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 656                 page = cb->compressed_pages[pg_index];
 657                 page->mapping = inode->i_mapping;
 658                 page->index = em_start >> PAGE_CACHE_SHIFT;
 659
 660                 if (comp_bio->bi_iter.bi_size)
 661                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 662                                                         PAGE_CACHE_SIZE,
 663                                                         comp_bio, 0);
 664                 else
 665                         ret = 0;
 666
 667                 page->mapping = NULL;
 668                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 669                     PAGE_CACHE_SIZE) {
 670                         bio_get(comp_bio);
 671
 672                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 673                                         BTRFS_WQ_ENDIO_DATA);
 674                         BUG_ON(ret); /* -ENOMEM */
 675
 676                         /*
 677                          * inc the count before we submit the bio so
 678                          * we know the end IO handler won't happen before
 679                          * we inc the count.  Otherwise, the cb might get
 680                          * freed before we're done setting it up
 681                          */
 682                         atomic_inc(&cb->pending_bios);
 683
 684                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 685                                 ret = btrfs_lookup_bio_sums(root, inode,
 686                                                         comp_bio, sums);
 687                                 BUG_ON(ret); /* -ENOMEM */
 688                         }
 689                         sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
 690                                              root->sectorsize);
 691
 692                         ret = btrfs_map_bio(root, READ, comp_bio,
 693                                             mirror_num, 0);
 694                         if (ret)
 695                                 bio_endio(comp_bio, ret);
 696
 697                         bio_put(comp_bio);
 698
 699                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 700                                                         GFP_NOFS);
 701                         BUG_ON(!comp_bio);
 702                         comp_bio->bi_private = cb;
 703                         comp_bio->bi_end_io = end_compressed_bio_read;
 704
 705                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 706                 }
 707                 cur_disk_byte += PAGE_CACHE_SIZE;
 708         }
 709         bio_get(comp_bio);
 710
 711         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 712                         BTRFS_WQ_ENDIO_DATA);
 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(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 PTR_ERR(workspace);
 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 PTR_ERR(workspace);
 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 PTR_ERR(workspace);
 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 }
1055
1056 /*
1057  * When uncompressing data, we need to make sure and zero any parts of
1058  * the biovec that were not filled in by the decompression code.  pg_index
1059  * and pg_offset indicate the last page and the last offset of that page
1060  * that have been filled in.  This will zero everything remaining in the
1061  * biovec.
1062  */
1063 void btrfs_clear_biovec_end(struct bio_vec *bvec, int vcnt,
1064                                    unsigned long pg_index,
1065                                    unsigned long pg_offset)
1066 {
1067         while (pg_index < vcnt) {
1068                 struct page *page = bvec[pg_index].bv_page;
1069                 unsigned long off = bvec[pg_index].bv_offset;
1070                 unsigned long len = bvec[pg_index].bv_len;
1071
1072                 if (pg_offset < off)
1073                         pg_offset = off;
1074                 if (pg_offset < off + len) {
1075                         unsigned long bytes = off + len - pg_offset;
1076                         char *kaddr;
1077
1078                         kaddr = kmap_atomic(page);
1079                         memset(kaddr + pg_offset, 0, bytes);
1080                         kunmap_atomic(kaddr);
1081                 }
1082                 pg_index++;
1083                 pg_offset = 0;
1084         }
1085 }