kernel/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,
 228                                               const struct compressed_bio *cb)
 229 {
 230         unsigned long index = cb->start >> PAGE_CACHE_SHIFT;
 231         unsigned long end_index = (cb->start + cb->len - 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         if (cb->errors)
 238                 mapping_set_error(inode->i_mapping, -EIO);
 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                         if (cb->errors)
 251                                 SetPageError(pages[i]);
 252                         end_page_writeback(pages[i]);
 253                         page_cache_release(pages[i]);
 254                 }
 255                 nr_pages -= ret;
 256                 index += ret;
 257         }
 258         /* the inode may be gone now */
 259 }
 260
 261 /*
 262  * do the cleanup once all the compressed pages hit the disk.
 263  * This will clear writeback on the file pages and free the compressed
 264  * pages.
 265  *
 266  * This also calls the writeback end hooks for the file pages so that
 267  * metadata and checksums can be updated in the file.
 268  */
 269 static void end_compressed_bio_write(struct bio *bio, int err)
 270 {
 271         struct extent_io_tree *tree;
 272         struct compressed_bio *cb = bio->bi_private;
 273         struct inode *inode;
 274         struct page *page;
 275         unsigned long index;
 276
 277         if (err)
 278                 cb->errors = 1;
 279
 280         /* if there are more bios still pending for this compressed
 281          * extent, just exit
 282          */
 283         if (!atomic_dec_and_test(&cb->pending_bios))
 284                 goto out;
 285
 286         /* ok, we're the last bio for this extent, step one is to
 287          * call back into the FS and do all the end_io operations
 288          */
 289         inode = cb->inode;
 290         tree = &BTRFS_I(inode)->io_tree;
 291         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 292         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 293                                          cb->start,
 294                                          cb->start + cb->len - 1,
 295                                          NULL,
 296                                          err ? 0 : 1);
 297         cb->compressed_pages[0]->mapping = NULL;
 298
 299         end_compressed_writeback(inode, cb);
 300         /* note, our inode could be gone now */
 301
 302         /*
 303          * release the compressed pages, these came from alloc_page and
 304          * are not attached to the inode at all
 305          */
 306         index = 0;
 307         for (index = 0; index < cb->nr_pages; index++) {
 308                 page = cb->compressed_pages[index];
 309                 page->mapping = NULL;
 310                 page_cache_release(page);
 311         }
 312
 313         /* finally free the cb struct */
 314         kfree(cb->compressed_pages);
 315         kfree(cb);
 316 out:
 317         bio_put(bio);
 318 }
 319
 320 /*
 321  * worker function to build and submit bios for previously compressed pages.
 322  * The corresponding pages in the inode should be marked for writeback
 323  * and the compressed pages should have a reference on them for dropping
 324  * when the IO is complete.
 325  *
 326  * This also checksums the file bytes and gets things ready for
 327  * the end io hooks.
 328  */
 329 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 330                                  unsigned long len, u64 disk_start,
 331                                  unsigned long compressed_len,
 332                                  struct page **compressed_pages,
 333                                  unsigned long nr_pages)
 334 {
 335         struct bio *bio = NULL;
 336         struct btrfs_root *root = BTRFS_I(inode)->root;
 337         struct compressed_bio *cb;
 338         unsigned long bytes_left;
 339         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 340         int pg_index = 0;
 341         struct page *page;
 342         u64 first_byte = disk_start;
 343         struct block_device *bdev;
 344         int ret;
 345         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 346
 347         WARN_ON(start & ((u64)PAGE_CACHE_SIZE - 1));
 348         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 349         if (!cb)
 350                 return -ENOMEM;
 351         atomic_set(&cb->pending_bios, 0);
 352         cb->errors = 0;
 353         cb->inode = inode;
 354         cb->start = start;
 355         cb->len = len;
 356         cb->mirror_num = 0;
 357         cb->compressed_pages = compressed_pages;
 358         cb->compressed_len = compressed_len;
 359         cb->orig_bio = NULL;
 360         cb->nr_pages = nr_pages;
 361
 362         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 363
 364         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 365         if (!bio) {
 366                 kfree(cb);
 367                 return -ENOMEM;
 368         }
 369         bio->bi_private = cb;
 370         bio->bi_end_io = end_compressed_bio_write;
 371         atomic_inc(&cb->pending_bios);
 372
 373         /* create and submit bios for the compressed pages */
 374         bytes_left = compressed_len;
 375         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 376                 page = compressed_pages[pg_index];
 377                 page->mapping = inode->i_mapping;
 378                 if (bio->bi_iter.bi_size)
 379                         ret = io_tree->ops->merge_bio_hook(WRITE, page, 0,
 380                                                            PAGE_CACHE_SIZE,
 381                                                            bio, 0);
 382                 else
 383                         ret = 0;
 384
 385                 page->mapping = NULL;
 386                 if (ret || bio_add_page(bio, page, PAGE_CACHE_SIZE, 0) <
 387                     PAGE_CACHE_SIZE) {
 388                         bio_get(bio);
 389
 390                         /*
 391                          * inc the count before we submit the bio so
 392                          * we know the end IO handler won't happen before
 393                          * we inc the count.  Otherwise, the cb might get
 394                          * freed before we're done setting it up
 395                          */
 396                         atomic_inc(&cb->pending_bios);
 397                         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
 398                                         BTRFS_WQ_ENDIO_DATA);
 399                         BUG_ON(ret); /* -ENOMEM */
 400
 401                         if (!skip_sum) {
 402                                 ret = btrfs_csum_one_bio(root, inode, bio,
 403                                                          start, 1);
 404                                 BUG_ON(ret); /* -ENOMEM */
 405                         }
 406
 407                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 408                         BUG_ON(ret); /* -ENOMEM */
 409
 410                         bio_put(bio);
 411
 412                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 413                         BUG_ON(!bio);
 414                         bio->bi_private = cb;
 415                         bio->bi_end_io = end_compressed_bio_write;
 416                         bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
 417                 }
 418                 if (bytes_left < PAGE_CACHE_SIZE) {
 419                         btrfs_info(BTRFS_I(inode)->root->fs_info,
 420                                         "bytes left %lu compress len %lu nr %lu",
 421                                bytes_left, cb->compressed_len, cb->nr_pages);
 422                 }
 423                 bytes_left -= PAGE_CACHE_SIZE;
 424                 first_byte += PAGE_CACHE_SIZE;
 425                 cond_resched();
 426         }
 427         bio_get(bio);
 428
 429         ret = btrfs_bio_wq_end_io(root->fs_info, bio, BTRFS_WQ_ENDIO_DATA);
 430         BUG_ON(ret); /* -ENOMEM */
 431
 432         if (!skip_sum) {
 433                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 434                 BUG_ON(ret); /* -ENOMEM */
 435         }
 436
 437         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 438         BUG_ON(ret); /* -ENOMEM */
 439
 440         bio_put(bio);
 441         return 0;
 442 }
 443
 444 static noinline int add_ra_bio_pages(struct inode *inode,
 445                                      u64 compressed_end,
 446                                      struct compressed_bio *cb)
 447 {
 448         unsigned long end_index;
 449         unsigned long pg_index;
 450         u64 last_offset;
 451         u64 isize = i_size_read(inode);
 452         int ret;
 453         struct page *page;
 454         unsigned long nr_pages = 0;
 455         struct extent_map *em;
 456         struct address_space *mapping = inode->i_mapping;
 457         struct extent_map_tree *em_tree;
 458         struct extent_io_tree *tree;
 459         u64 end;
 460         int misses = 0;
 461
 462         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 463         last_offset = (page_offset(page) + PAGE_CACHE_SIZE);
 464         em_tree = &BTRFS_I(inode)->extent_tree;
 465         tree = &BTRFS_I(inode)->io_tree;
 466
 467         if (isize == 0)
 468                 return 0;
 469
 470         end_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 471
 472         while (last_offset < compressed_end) {
 473                 pg_index = last_offset >> PAGE_CACHE_SHIFT;
 474
 475                 if (pg_index > end_index)
 476                         break;
 477
 478                 rcu_read_lock();
 479                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 480                 rcu_read_unlock();
 481                 if (page && !radix_tree_exceptional_entry(page)) {
 482                         misses++;
 483                         if (misses > 4)
 484                                 break;
 485                         goto next;
 486                 }
 487
 488                 page = __page_cache_alloc(mapping_gfp_mask(mapping) &
 489                                                                 ~__GFP_FS);
 490                 if (!page)
 491                         break;
 492
 493                 if (add_to_page_cache_lru(page, mapping, pg_index,
 494                                                                 GFP_NOFS)) {
 495                         page_cache_release(page);
 496                         goto next;
 497                 }
 498
 499                 end = last_offset + PAGE_CACHE_SIZE - 1;
 500                 /*
 501                  * at this point, we have a locked page in the page cache
 502                  * for these bytes in the file.  But, we have to make
 503                  * sure they map to this compressed extent on disk.
 504                  */
 505                 set_page_extent_mapped(page);
 506                 lock_extent(tree, last_offset, end);
 507                 read_lock(&em_tree->lock);
 508                 em = lookup_extent_mapping(em_tree, last_offset,
 509                                            PAGE_CACHE_SIZE);
 510                 read_unlock(&em_tree->lock);
 511
 512                 if (!em || last_offset < em->start ||
 513                     (last_offset + PAGE_CACHE_SIZE > extent_map_end(em)) ||
 514                     (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
 515                         free_extent_map(em);
 516                         unlock_extent(tree, last_offset, end);
 517                         unlock_page(page);
 518                         page_cache_release(page);
 519                         break;
 520                 }
 521                 free_extent_map(em);
 522
 523                 if (page->index == end_index) {
 524                         char *userpage;
 525                         size_t zero_offset = isize & (PAGE_CACHE_SIZE - 1);
 526
 527                         if (zero_offset) {
 528                                 int zeros;
 529                                 zeros = PAGE_CACHE_SIZE - zero_offset;
 530                                 userpage = kmap_atomic(page);
 531                                 memset(userpage + zero_offset, 0, zeros);
 532                                 flush_dcache_page(page);
 533                                 kunmap_atomic(userpage);
 534                         }
 535                 }
 536
 537                 ret = bio_add_page(cb->orig_bio, page,
 538                                    PAGE_CACHE_SIZE, 0);
 539
 540                 if (ret == PAGE_CACHE_SIZE) {
 541                         nr_pages++;
 542                         page_cache_release(page);
 543                 } else {
 544                         unlock_extent(tree, last_offset, end);
 545                         unlock_page(page);
 546                         page_cache_release(page);
 547                         break;
 548                 }
 549 next:
 550                 last_offset += PAGE_CACHE_SIZE;
 551         }
 552         return 0;
 553 }
 554
 555 /*
 556  * for a compressed read, the bio we get passed has all the inode pages
 557  * in it.  We don't actually do IO on those pages but allocate new ones
 558  * to hold the compressed pages on disk.
 559  *
 560  * bio->bi_iter.bi_sector points to the compressed extent on disk
 561  * bio->bi_io_vec points to all of the inode pages
 562  * bio->bi_vcnt is a count of pages
 563  *
 564  * After the compressed pages are read, we copy the bytes into the
 565  * bio we were passed and then call the bio end_io calls
 566  */
 567 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 568                                  int mirror_num, unsigned long bio_flags)
 569 {
 570         struct extent_io_tree *tree;
 571         struct extent_map_tree *em_tree;
 572         struct compressed_bio *cb;
 573         struct btrfs_root *root = BTRFS_I(inode)->root;
 574         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 575         unsigned long compressed_len;
 576         unsigned long nr_pages;
 577         unsigned long pg_index;
 578         struct page *page;
 579         struct block_device *bdev;
 580         struct bio *comp_bio;
 581         u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
 582         u64 em_len;
 583         u64 em_start;
 584         struct extent_map *em;
 585         int ret = -ENOMEM;
 586         int faili = 0;
 587         u32 *sums;
 588
 589         tree = &BTRFS_I(inode)->io_tree;
 590         em_tree = &BTRFS_I(inode)->extent_tree;
 591
 592         /* we need the actual starting offset of this extent in the file */
 593         read_lock(&em_tree->lock);
 594         em = lookup_extent_mapping(em_tree,
 595                                    page_offset(bio->bi_io_vec->bv_page),
 596                                    PAGE_CACHE_SIZE);
 597         read_unlock(&em_tree->lock);
 598         if (!em)
 599                 return -EIO;
 600
 601         compressed_len = em->block_len;
 602         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 603         if (!cb)
 604                 goto out;
 605
 606         atomic_set(&cb->pending_bios, 0);
 607         cb->errors = 0;
 608         cb->inode = inode;
 609         cb->mirror_num = mirror_num;
 610         sums = &cb->sums;
 611
 612         cb->start = em->orig_start;
 613         em_len = em->len;
 614         em_start = em->start;
 615
 616         free_extent_map(em);
 617         em = NULL;
 618
 619         cb->len = uncompressed_len;
 620         cb->compressed_len = compressed_len;
 621         cb->compress_type = extent_compress_type(bio_flags);
 622         cb->orig_bio = bio;
 623
 624         nr_pages = DIV_ROUND_UP(compressed_len, PAGE_CACHE_SIZE);
 625         cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
 626                                        GFP_NOFS);
 627         if (!cb->compressed_pages)
 628                 goto fail1;
 629
 630         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 631
 632         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 633                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 634                                                               __GFP_HIGHMEM);
 635                 if (!cb->compressed_pages[pg_index]) {
 636                         faili = pg_index - 1;
 637                         ret = -ENOMEM;
 638                         goto fail2;
 639                 }
 640         }
 641         faili = nr_pages - 1;
 642         cb->nr_pages = nr_pages;
 643
 644         /* In the parent-locked case, we only locked the range we are
 645          * interested in.  In all other cases, we can opportunistically
 646          * cache decompressed data that goes beyond the requested range. */
 647         if (!(bio_flags & EXTENT_BIO_PARENT_LOCKED))
 648                 add_ra_bio_pages(inode, em_start + em_len, cb);
 649
 650         /* include any pages we added in add_ra-bio_pages */
 651         uncompressed_len = bio->bi_vcnt * PAGE_CACHE_SIZE;
 652         cb->len = uncompressed_len;
 653
 654         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 655         if (!comp_bio)
 656                 goto fail2;
 657         comp_bio->bi_private = cb;
 658         comp_bio->bi_end_io = end_compressed_bio_read;
 659         atomic_inc(&cb->pending_bios);
 660
 661         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 662                 page = cb->compressed_pages[pg_index];
 663                 page->mapping = inode->i_mapping;
 664                 page->index = em_start >> PAGE_CACHE_SHIFT;
 665
 666                 if (comp_bio->bi_iter.bi_size)
 667                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 668                                                         PAGE_CACHE_SIZE,
 669                                                         comp_bio, 0);
 670                 else
 671                         ret = 0;
 672
 673                 page->mapping = NULL;
 674                 if (ret || bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0) <
 675                     PAGE_CACHE_SIZE) {
 676                         bio_get(comp_bio);
 677
 678                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 679                                         BTRFS_WQ_ENDIO_DATA);
 680                         BUG_ON(ret); /* -ENOMEM */
 681
 682                         /*
 683                          * inc the count before we submit the bio so
 684                          * we know the end IO handler won't happen before
 685                          * we inc the count.  Otherwise, the cb might get
 686                          * freed before we're done setting it up
 687                          */
 688                         atomic_inc(&cb->pending_bios);
 689
 690                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 691                                 ret = btrfs_lookup_bio_sums(root, inode,
 692                                                         comp_bio, sums);
 693                                 BUG_ON(ret); /* -ENOMEM */
 694                         }
 695                         sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
 696                                              root->sectorsize);
 697
 698                         ret = btrfs_map_bio(root, READ, comp_bio,
 699                                             mirror_num, 0);
 700                         if (ret)
 701                                 bio_endio(comp_bio, ret);
 702
 703                         bio_put(comp_bio);
 704
 705                         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte,
 706                                                         GFP_NOFS);
 707                         BUG_ON(!comp_bio);
 708                         comp_bio->bi_private = cb;
 709                         comp_bio->bi_end_io = end_compressed_bio_read;
 710
 711                         bio_add_page(comp_bio, page, PAGE_CACHE_SIZE, 0);
 712                 }
 713                 cur_disk_byte += PAGE_CACHE_SIZE;
 714         }
 715         bio_get(comp_bio);
 716
 717         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 718                         BTRFS_WQ_ENDIO_DATA);
 719         BUG_ON(ret); /* -ENOMEM */
 720
 721         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 722                 ret = btrfs_lookup_bio_sums(root, inode, comp_bio, sums);
 723                 BUG_ON(ret); /* -ENOMEM */
 724         }
 725
 726         ret = btrfs_map_bio(root, READ, comp_bio, mirror_num, 0);
 727         if (ret)
 728                 bio_endio(comp_bio, ret);
 729
 730         bio_put(comp_bio);
 731         return 0;
 732
 733 fail2:
 734         while (faili >= 0) {
 735                 __free_page(cb->compressed_pages[faili]);
 736                 faili--;
 737         }
 738
 739         kfree(cb->compressed_pages);
 740 fail1:
 741         kfree(cb);
 742 out:
 743         free_extent_map(em);
 744         return ret;
 745 }
 746
 747 static struct list_head comp_idle_workspace[BTRFS_COMPRESS_TYPES];
 748 static spinlock_t comp_workspace_lock[BTRFS_COMPRESS_TYPES];
 749 static int comp_num_workspace[BTRFS_COMPRESS_TYPES];
 750 static atomic_t comp_alloc_workspace[BTRFS_COMPRESS_TYPES];
 751 static wait_queue_head_t comp_workspace_wait[BTRFS_COMPRESS_TYPES];
 752
 753 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
 754         &btrfs_zlib_compress,
 755         &btrfs_lzo_compress,
 756 };
 757
 758 void __init btrfs_init_compress(void)
 759 {
 760         int i;
 761
 762         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 763                 INIT_LIST_HEAD(&comp_idle_workspace[i]);
 764                 spin_lock_init(&comp_workspace_lock[i]);
 765                 atomic_set(&comp_alloc_workspace[i], 0);
 766                 init_waitqueue_head(&comp_workspace_wait[i]);
 767         }
 768 }
 769
 770 /*
 771  * this finds an available workspace or allocates a new one
 772  * ERR_PTR is returned if things go bad.
 773  */
 774 static struct list_head *find_workspace(int type)
 775 {
 776         struct list_head *workspace;
 777         int cpus = num_online_cpus();
 778         int idx = type - 1;
 779
 780         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 781         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 782         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 783         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 784         int *num_workspace                      = &comp_num_workspace[idx];
 785 again:
 786         spin_lock(workspace_lock);
 787         if (!list_empty(idle_workspace)) {
 788                 workspace = idle_workspace->next;
 789                 list_del(workspace);
 790                 (*num_workspace)--;
 791                 spin_unlock(workspace_lock);
 792                 return workspace;
 793
 794         }
 795         if (atomic_read(alloc_workspace) > cpus) {
 796                 DEFINE_WAIT(wait);
 797
 798                 spin_unlock(workspace_lock);
 799                 prepare_to_wait(workspace_wait, &wait, TASK_UNINTERRUPTIBLE);
 800                 if (atomic_read(alloc_workspace) > cpus && !*num_workspace)
 801                         schedule();
 802                 finish_wait(workspace_wait, &wait);
 803                 goto again;
 804         }
 805         atomic_inc(alloc_workspace);
 806         spin_unlock(workspace_lock);
 807
 808         workspace = btrfs_compress_op[idx]->alloc_workspace();
 809         if (IS_ERR(workspace)) {
 810                 atomic_dec(alloc_workspace);
 811                 wake_up(workspace_wait);
 812         }
 813         return workspace;
 814 }
 815
 816 /*
 817  * put a workspace struct back on the list or free it if we have enough
 818  * idle ones sitting around
 819  */
 820 static void free_workspace(int type, struct list_head *workspace)
 821 {
 822         int idx = type - 1;
 823         struct list_head *idle_workspace        = &comp_idle_workspace[idx];
 824         spinlock_t *workspace_lock              = &comp_workspace_lock[idx];
 825         atomic_t *alloc_workspace               = &comp_alloc_workspace[idx];
 826         wait_queue_head_t *workspace_wait       = &comp_workspace_wait[idx];
 827         int *num_workspace                      = &comp_num_workspace[idx];
 828
 829         spin_lock(workspace_lock);
 830         if (*num_workspace < num_online_cpus()) {
 831                 list_add(workspace, idle_workspace);
 832                 (*num_workspace)++;
 833                 spin_unlock(workspace_lock);
 834                 goto wake;
 835         }
 836         spin_unlock(workspace_lock);
 837
 838         btrfs_compress_op[idx]->free_workspace(workspace);
 839         atomic_dec(alloc_workspace);
 840 wake:
 841         smp_mb();
 842         if (waitqueue_active(workspace_wait))
 843                 wake_up(workspace_wait);
 844 }
 845
 846 /*
 847  * cleanup function for module exit
 848  */
 849 static void free_workspaces(void)
 850 {
 851         struct list_head *workspace;
 852         int i;
 853
 854         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 855                 while (!list_empty(&comp_idle_workspace[i])) {
 856                         workspace = comp_idle_workspace[i].next;
 857                         list_del(workspace);
 858                         btrfs_compress_op[i]->free_workspace(workspace);
 859                         atomic_dec(&comp_alloc_workspace[i]);
 860                 }
 861         }
 862 }
 863
 864 /*
 865  * given an address space and start/len, compress the bytes.
 866  *
 867  * pages are allocated to hold the compressed result and stored
 868  * in 'pages'
 869  *
 870  * out_pages is used to return the number of pages allocated.  There
 871  * may be pages allocated even if we return an error
 872  *
 873  * total_in is used to return the number of bytes actually read.  It
 874  * may be smaller then len if we had to exit early because we
 875  * ran out of room in the pages array or because we cross the
 876  * max_out threshold.
 877  *
 878  * total_out is used to return the total number of compressed bytes
 879  *
 880  * max_out tells us the max number of bytes that we're allowed to
 881  * stuff into pages
 882  */
 883 int btrfs_compress_pages(int type, struct address_space *mapping,
 884                          u64 start, unsigned long len,
 885                          struct page **pages,
 886                          unsigned long nr_dest_pages,
 887                          unsigned long *out_pages,
 888                          unsigned long *total_in,
 889                          unsigned long *total_out,
 890                          unsigned long max_out)
 891 {
 892         struct list_head *workspace;
 893         int ret;
 894
 895         workspace = find_workspace(type);
 896         if (IS_ERR(workspace))
 897                 return PTR_ERR(workspace);
 898
 899         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 900                                                       start, len, pages,
 901                                                       nr_dest_pages, out_pages,
 902                                                       total_in, total_out,
 903                                                       max_out);
 904         free_workspace(type, workspace);
 905         return ret;
 906 }
 907
 908 /*
 909  * pages_in is an array of pages with compressed data.
 910  *
 911  * disk_start is the starting logical offset of this array in the file
 912  *
 913  * bvec is a bio_vec of pages from the file that we want to decompress into
 914  *
 915  * vcnt is the count of pages in the biovec
 916  *
 917  * srclen is the number of bytes in pages_in
 918  *
 919  * The basic idea is that we have a bio that was created by readpages.
 920  * The pages in the bio are for the uncompressed data, and they may not
 921  * be contiguous.  They all correspond to the range of bytes covered by
 922  * the compressed extent.
 923  */
 924 static int btrfs_decompress_biovec(int type, struct page **pages_in,
 925                                    u64 disk_start, struct bio_vec *bvec,
 926                                    int vcnt, size_t srclen)
 927 {
 928         struct list_head *workspace;
 929         int ret;
 930
 931         workspace = find_workspace(type);
 932         if (IS_ERR(workspace))
 933                 return PTR_ERR(workspace);
 934
 935         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 936                                                          disk_start,
 937                                                          bvec, vcnt, srclen);
 938         free_workspace(type, workspace);
 939         return ret;
 940 }
 941
 942 /*
 943  * a less complex decompression routine.  Our compressed data fits in a
 944  * single page, and we want to read a single page out of it.
 945  * start_byte tells us the offset into the compressed data we're interested in
 946  */
 947 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 948                      unsigned long start_byte, size_t srclen, size_t destlen)
 949 {
 950         struct list_head *workspace;
 951         int ret;
 952
 953         workspace = find_workspace(type);
 954         if (IS_ERR(workspace))
 955                 return PTR_ERR(workspace);
 956
 957         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 958                                                   dest_page, start_byte,
 959                                                   srclen, destlen);
 960
 961         free_workspace(type, workspace);
 962         return ret;
 963 }
 964
 965 void btrfs_exit_compress(void)
 966 {
 967         free_workspaces();
 968 }
 969
 970 /*
 971  * Copy uncompressed data from working buffer to pages.
 972  *
 973  * buf_start is the byte offset we're of the start of our workspace buffer.
 974  *
 975  * total_out is the last byte of the buffer
 976  */
 977 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
 978                               unsigned long total_out, u64 disk_start,
 979                               struct bio_vec *bvec, int vcnt,
 980                               unsigned long *pg_index,
 981                               unsigned long *pg_offset)
 982 {
 983         unsigned long buf_offset;
 984         unsigned long current_buf_start;
 985         unsigned long start_byte;
 986         unsigned long working_bytes = total_out - buf_start;
 987         unsigned long bytes;
 988         char *kaddr;
 989         struct page *page_out = bvec[*pg_index].bv_page;
 990
 991         /*
 992          * start byte is the first byte of the page we're currently
 993          * copying into relative to the start of the compressed data.
 994          */
 995         start_byte = page_offset(page_out) - disk_start;
 996
 997         /* we haven't yet hit data corresponding to this page */
 998         if (total_out <= start_byte)
 999                 return 1;
1000
1001         /*
1002          * the start of the data we care about is offset into
1003          * the middle of our working buffer
1004          */
1005         if (total_out > start_byte && buf_start < start_byte) {
1006                 buf_offset = start_byte - buf_start;
1007                 working_bytes -= buf_offset;
1008         } else {
1009                 buf_offset = 0;
1010         }
1011         current_buf_start = buf_start;
1012
1013         /* copy bytes from the working buffer into the pages */
1014         while (working_bytes > 0) {
1015                 bytes = min(PAGE_CACHE_SIZE - *pg_offset,
1016                             PAGE_CACHE_SIZE - buf_offset);
1017                 bytes = min(bytes, working_bytes);
1018                 kaddr = kmap_atomic(page_out);
1019                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1020                 kunmap_atomic(kaddr);
1021                 flush_dcache_page(page_out);
1022
1023                 *pg_offset += bytes;
1024                 buf_offset += bytes;
1025                 working_bytes -= bytes;
1026                 current_buf_start += bytes;
1027
1028                 /* check if we need to pick another page */
1029                 if (*pg_offset == PAGE_CACHE_SIZE) {
1030                         (*pg_index)++;
1031                         if (*pg_index >= vcnt)
1032                                 return 0;
1033
1034                         page_out = bvec[*pg_index].bv_page;
1035                         *pg_offset = 0;
1036                         start_byte = page_offset(page_out) - disk_start;
1037
1038                         /*
1039                          * make sure our new page is covered by this
1040                          * working buffer
1041                          */
1042                         if (total_out <= start_byte)
1043                                 return 1;
1044
1045                         /*
1046                          * the next page in the biovec might not be adjacent
1047                          * to the last page, but it might still be found
1048                          * inside this working buffer. bump our offset pointer
1049                          */
1050                         if (total_out > start_byte &&
1051                             current_buf_start < start_byte) {
1052                                 buf_offset = start_byte - buf_start;
1053                                 working_bytes = total_out - start_byte;
1054                                 current_buf_start = buf_start + buf_offset;
1055                         }
1056                 }
1057         }
1058
1059         return 1;
1060 }
1061
1062 /*
1063  * When uncompressing data, we need to make sure and zero any parts of
1064  * the biovec that were not filled in by the decompression code.  pg_index
1065  * and pg_offset indicate the last page and the last offset of that page
1066  * that have been filled in.  This will zero everything remaining in the
1067  * biovec.
1068  */
1069 void btrfs_clear_biovec_end(struct bio_vec *bvec, int vcnt,
1070                                    unsigned long pg_index,
1071                                    unsigned long pg_offset)
1072 {
1073         while (pg_index < vcnt) {
1074                 struct page *page = bvec[pg_index].bv_page;
1075                 unsigned long off = bvec[pg_index].bv_offset;
1076                 unsigned long len = bvec[pg_index].bv_len;
1077
1078                 if (pg_offset < off)
1079                         pg_offset = off;
1080                 if (pg_offset < off + len) {
1081                         unsigned long bytes = off + len - pg_offset;
1082                         char *kaddr;
1083
1084                         kaddr = kmap_atomic(page);
1085                         memset(kaddr + pg_offset, 0, bytes);
1086                         kunmap_atomic(kaddr);
1087                 }
1088                 pg_index++;
1089                 pg_offset = 0;
1090         }
1091 }