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         return btrfs_bio_alloc(bdev, first_byte >> 9, BIO_MAX_PAGES, 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 page *page;
 109         unsigned long i;
 110         char *kaddr;
 111         u32 csum;
 112         u32 *cb_sum = &cb->sums;
 113
 114         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 115                 return 0;
 116
 117         for (i = 0; i < cb->nr_pages; i++) {
 118                 page = cb->compressed_pages[i];
 119                 csum = ~(u32)0;
 120
 121                 kaddr = kmap_atomic(page);
 122                 csum = btrfs_csum_data(kaddr, csum, PAGE_SIZE);
 123                 btrfs_csum_final(csum, (char *)&csum);
 124                 kunmap_atomic(kaddr);
 125
 126                 if (csum != *cb_sum) {
 127                         btrfs_info(BTRFS_I(inode)->root->fs_info,
 128                            "csum failed ino %llu extent %llu csum %u wanted %u mirror %d",
 129                            btrfs_ino(inode), disk_start, csum, *cb_sum,
 130                            cb->mirror_num);
 131                         ret = -EIO;
 132                         goto fail;
 133                 }
 134                 cb_sum++;
 135
 136         }
 137         ret = 0;
 138 fail:
 139         return ret;
 140 }
 141
 142 /* when we finish reading compressed pages from the disk, we
 143  * decompress them and then run the bio end_io routines on the
 144  * decompressed pages (in the inode address space).
 145  *
 146  * This allows the checksumming and other IO error handling routines
 147  * to work normally
 148  *
 149  * The compressed pages are freed here, and it must be run
 150  * in process context
 151  */
 152 static void end_compressed_bio_read(struct bio *bio)
 153 {
 154         struct compressed_bio *cb = bio->bi_private;
 155         struct inode *inode;
 156         struct page *page;
 157         unsigned long index;
 158         int ret;
 159
 160         if (bio->bi_error)
 161                 cb->errors = 1;
 162
 163         /* if there are more bios still pending for this compressed
 164          * extent, just exit
 165          */
 166         if (!atomic_dec_and_test(&cb->pending_bios))
 167                 goto out;
 168
 169         inode = cb->inode;
 170         ret = check_compressed_csum(inode, cb,
 171                                     (u64)bio->bi_iter.bi_sector << 9);
 172         if (ret)
 173                 goto csum_failed;
 174
 175         /* ok, we're the last bio for this extent, lets start
 176          * the decompression.
 177          */
 178         ret = btrfs_decompress_biovec(cb->compress_type,
 179                                       cb->compressed_pages,
 180                                       cb->start,
 181                                       cb->orig_bio->bi_io_vec,
 182                                       cb->orig_bio->bi_vcnt,
 183                                       cb->compressed_len);
 184 csum_failed:
 185         if (ret)
 186                 cb->errors = 1;
 187
 188         /* release the compressed pages */
 189         index = 0;
 190         for (index = 0; index < cb->nr_pages; index++) {
 191                 page = cb->compressed_pages[index];
 192                 page->mapping = NULL;
 193                 put_page(page);
 194         }
 195
 196         /* do io completion on the original bio */
 197         if (cb->errors) {
 198                 bio_io_error(cb->orig_bio);
 199         } else {
 200                 int i;
 201                 struct bio_vec *bvec;
 202
 203                 /*
 204                  * we have verified the checksum already, set page
 205                  * checked so the end_io handlers know about it
 206                  */
 207                 bio_for_each_segment_all(bvec, cb->orig_bio, i)
 208                         SetPageChecked(bvec->bv_page);
 209
 210                 bio_endio(cb->orig_bio);
 211         }
 212
 213         /* finally free the cb struct */
 214         kfree(cb->compressed_pages);
 215         kfree(cb);
 216 out:
 217         bio_put(bio);
 218 }
 219
 220 /*
 221  * Clear the writeback bits on all of the file
 222  * pages for a compressed write
 223  */
 224 static noinline void end_compressed_writeback(struct inode *inode,
 225                                               const struct compressed_bio *cb)
 226 {
 227         unsigned long index = cb->start >> PAGE_SHIFT;
 228         unsigned long end_index = (cb->start + cb->len - 1) >> PAGE_SHIFT;
 229         struct page *pages[16];
 230         unsigned long nr_pages = end_index - index + 1;
 231         int i;
 232         int ret;
 233
 234         if (cb->errors)
 235                 mapping_set_error(inode->i_mapping, -EIO);
 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                         if (cb->errors)
 248                                 SetPageError(pages[i]);
 249                         end_page_writeback(pages[i]);
 250                         put_page(pages[i]);
 251                 }
 252                 nr_pages -= ret;
 253                 index += ret;
 254         }
 255         /* the inode may be gone now */
 256 }
 257
 258 /*
 259  * do the cleanup once all the compressed pages hit the disk.
 260  * This will clear writeback on the file pages and free the compressed
 261  * pages.
 262  *
 263  * This also calls the writeback end hooks for the file pages so that
 264  * metadata and checksums can be updated in the file.
 265  */
 266 static void end_compressed_bio_write(struct bio *bio)
 267 {
 268         struct extent_io_tree *tree;
 269         struct compressed_bio *cb = bio->bi_private;
 270         struct inode *inode;
 271         struct page *page;
 272         unsigned long index;
 273
 274         if (bio->bi_error)
 275                 cb->errors = 1;
 276
 277         /* if there are more bios still pending for this compressed
 278          * extent, just exit
 279          */
 280         if (!atomic_dec_and_test(&cb->pending_bios))
 281                 goto out;
 282
 283         /* ok, we're the last bio for this extent, step one is to
 284          * call back into the FS and do all the end_io operations
 285          */
 286         inode = cb->inode;
 287         tree = &BTRFS_I(inode)->io_tree;
 288         cb->compressed_pages[0]->mapping = cb->inode->i_mapping;
 289         tree->ops->writepage_end_io_hook(cb->compressed_pages[0],
 290                                          cb->start,
 291                                          cb->start + cb->len - 1,
 292                                          NULL,
 293                                          bio->bi_error ? 0 : 1);
 294         cb->compressed_pages[0]->mapping = NULL;
 295
 296         end_compressed_writeback(inode, cb);
 297         /* note, our inode could be gone now */
 298
 299         /*
 300          * release the compressed pages, these came from alloc_page and
 301          * are not attached to the inode at all
 302          */
 303         index = 0;
 304         for (index = 0; index < cb->nr_pages; index++) {
 305                 page = cb->compressed_pages[index];
 306                 page->mapping = NULL;
 307                 put_page(page);
 308         }
 309
 310         /* finally free the cb struct */
 311         kfree(cb->compressed_pages);
 312         kfree(cb);
 313 out:
 314         bio_put(bio);
 315 }
 316
 317 /*
 318  * worker function to build and submit bios for previously compressed pages.
 319  * The corresponding pages in the inode should be marked for writeback
 320  * and the compressed pages should have a reference on them for dropping
 321  * when the IO is complete.
 322  *
 323  * This also checksums the file bytes and gets things ready for
 324  * the end io hooks.
 325  */
 326 int btrfs_submit_compressed_write(struct inode *inode, u64 start,
 327                                  unsigned long len, u64 disk_start,
 328                                  unsigned long compressed_len,
 329                                  struct page **compressed_pages,
 330                                  unsigned long nr_pages)
 331 {
 332         struct bio *bio = NULL;
 333         struct btrfs_root *root = BTRFS_I(inode)->root;
 334         struct compressed_bio *cb;
 335         unsigned long bytes_left;
 336         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 337         int pg_index = 0;
 338         struct page *page;
 339         u64 first_byte = disk_start;
 340         struct block_device *bdev;
 341         int ret;
 342         int skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 343
 344         WARN_ON(start & ((u64)PAGE_SIZE - 1));
 345         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 346         if (!cb)
 347                 return -ENOMEM;
 348         atomic_set(&cb->pending_bios, 0);
 349         cb->errors = 0;
 350         cb->inode = inode;
 351         cb->start = start;
 352         cb->len = len;
 353         cb->mirror_num = 0;
 354         cb->compressed_pages = compressed_pages;
 355         cb->compressed_len = compressed_len;
 356         cb->orig_bio = NULL;
 357         cb->nr_pages = nr_pages;
 358
 359         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 360
 361         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 362         if (!bio) {
 363                 kfree(cb);
 364                 return -ENOMEM;
 365         }
 366         bio->bi_private = cb;
 367         bio->bi_end_io = end_compressed_bio_write;
 368         atomic_inc(&cb->pending_bios);
 369
 370         /* create and submit bios for the compressed pages */
 371         bytes_left = compressed_len;
 372         for (pg_index = 0; pg_index < cb->nr_pages; pg_index++) {
 373                 page = compressed_pages[pg_index];
 374                 page->mapping = inode->i_mapping;
 375                 if (bio->bi_iter.bi_size)
 376                         ret = io_tree->ops->merge_bio_hook(WRITE, page, 0,
 377                                                            PAGE_SIZE,
 378                                                            bio, 0);
 379                 else
 380                         ret = 0;
 381
 382                 page->mapping = NULL;
 383                 if (ret || bio_add_page(bio, page, PAGE_SIZE, 0) <
 384                     PAGE_SIZE) {
 385                         bio_get(bio);
 386
 387                         /*
 388                          * inc the count before we submit the bio so
 389                          * we know the end IO handler won't happen before
 390                          * we inc the count.  Otherwise, the cb might get
 391                          * freed before we're done setting it up
 392                          */
 393                         atomic_inc(&cb->pending_bios);
 394                         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
 395                                         BTRFS_WQ_ENDIO_DATA);
 396                         BUG_ON(ret); /* -ENOMEM */
 397
 398                         if (!skip_sum) {
 399                                 ret = btrfs_csum_one_bio(root, inode, bio,
 400                                                          start, 1);
 401                                 BUG_ON(ret); /* -ENOMEM */
 402                         }
 403
 404                         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 405                         BUG_ON(ret); /* -ENOMEM */
 406
 407                         bio_put(bio);
 408
 409                         bio = compressed_bio_alloc(bdev, first_byte, GFP_NOFS);
 410                         BUG_ON(!bio);
 411                         bio->bi_private = cb;
 412                         bio->bi_end_io = end_compressed_bio_write;
 413                         bio_add_page(bio, page, PAGE_SIZE, 0);
 414                 }
 415                 if (bytes_left < PAGE_SIZE) {
 416                         btrfs_info(BTRFS_I(inode)->root->fs_info,
 417                                         "bytes left %lu compress len %lu nr %lu",
 418                                bytes_left, cb->compressed_len, cb->nr_pages);
 419                 }
 420                 bytes_left -= PAGE_SIZE;
 421                 first_byte += PAGE_SIZE;
 422                 cond_resched();
 423         }
 424         bio_get(bio);
 425
 426         ret = btrfs_bio_wq_end_io(root->fs_info, bio, BTRFS_WQ_ENDIO_DATA);
 427         BUG_ON(ret); /* -ENOMEM */
 428
 429         if (!skip_sum) {
 430                 ret = btrfs_csum_one_bio(root, inode, bio, start, 1);
 431                 BUG_ON(ret); /* -ENOMEM */
 432         }
 433
 434         ret = btrfs_map_bio(root, WRITE, bio, 0, 1);
 435         BUG_ON(ret); /* -ENOMEM */
 436
 437         bio_put(bio);
 438         return 0;
 439 }
 440
 441 static noinline int add_ra_bio_pages(struct inode *inode,
 442                                      u64 compressed_end,
 443                                      struct compressed_bio *cb)
 444 {
 445         unsigned long end_index;
 446         unsigned long pg_index;
 447         u64 last_offset;
 448         u64 isize = i_size_read(inode);
 449         int ret;
 450         struct page *page;
 451         unsigned long nr_pages = 0;
 452         struct extent_map *em;
 453         struct address_space *mapping = inode->i_mapping;
 454         struct extent_map_tree *em_tree;
 455         struct extent_io_tree *tree;
 456         u64 end;
 457         int misses = 0;
 458
 459         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 460         last_offset = (page_offset(page) + PAGE_SIZE);
 461         em_tree = &BTRFS_I(inode)->extent_tree;
 462         tree = &BTRFS_I(inode)->io_tree;
 463
 464         if (isize == 0)
 465                 return 0;
 466
 467         end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 468
 469         while (last_offset < compressed_end) {
 470                 pg_index = last_offset >> PAGE_SHIFT;
 471
 472                 if (pg_index > end_index)
 473                         break;
 474
 475                 rcu_read_lock();
 476                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 477                 rcu_read_unlock();
 478                 if (page && !radix_tree_exceptional_entry(page)) {
 479                         misses++;
 480                         if (misses > 4)
 481                                 break;
 482                         goto next;
 483                 }
 484
 485                 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
 486                                                                  ~__GFP_FS));
 487                 if (!page)
 488                         break;
 489
 490                 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
 491                         put_page(page);
 492                         goto next;
 493                 }
 494
 495                 end = last_offset + PAGE_SIZE - 1;
 496                 /*
 497                  * at this point, we have a locked page in the page cache
 498                  * for these bytes in the file.  But, we have to make
 499                  * sure they map to this compressed extent on disk.
 500                  */
 501                 set_page_extent_mapped(page);
 502                 lock_extent(tree, last_offset, end);
 503                 read_lock(&em_tree->lock);
 504                 em = lookup_extent_mapping(em_tree, last_offset,
 505                                            PAGE_SIZE);
 506                 read_unlock(&em_tree->lock);
 507
 508                 if (!em || last_offset < em->start ||
 509                     (last_offset + PAGE_SIZE > extent_map_end(em)) ||
 510                     (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
 511                         free_extent_map(em);
 512                         unlock_extent(tree, last_offset, end);
 513                         unlock_page(page);
 514                         put_page(page);
 515                         break;
 516                 }
 517                 free_extent_map(em);
 518
 519                 if (page->index == end_index) {
 520                         char *userpage;
 521                         size_t zero_offset = isize & (PAGE_SIZE - 1);
 522
 523                         if (zero_offset) {
 524                                 int zeros;
 525                                 zeros = PAGE_SIZE - zero_offset;
 526                                 userpage = kmap_atomic(page);
 527                                 memset(userpage + zero_offset, 0, zeros);
 528                                 flush_dcache_page(page);
 529                                 kunmap_atomic(userpage);
 530                         }
 531                 }
 532
 533                 ret = bio_add_page(cb->orig_bio, page,
 534                                    PAGE_SIZE, 0);
 535
 536                 if (ret == PAGE_SIZE) {
 537                         nr_pages++;
 538                         put_page(page);
 539                 } else {
 540                         unlock_extent(tree, last_offset, end);
 541                         unlock_page(page);
 542                         put_page(page);
 543                         break;
 544                 }
 545 next:
 546                 last_offset += PAGE_SIZE;
 547         }
 548         return 0;
 549 }
 550
 551 /*
 552  * for a compressed read, the bio we get passed has all the inode pages
 553  * in it.  We don't actually do IO on those pages but allocate new ones
 554  * to hold the compressed pages on disk.
 555  *
 556  * bio->bi_iter.bi_sector points to the compressed extent on disk
 557  * bio->bi_io_vec points to all of the inode pages
 558  * bio->bi_vcnt is a count of pages
 559  *
 560  * After the compressed pages are read, we copy the bytes into the
 561  * bio we were passed and then call the bio end_io calls
 562  */
 563 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 564                                  int mirror_num, unsigned long bio_flags)
 565 {
 566         struct extent_io_tree *tree;
 567         struct extent_map_tree *em_tree;
 568         struct compressed_bio *cb;
 569         struct btrfs_root *root = BTRFS_I(inode)->root;
 570         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_SIZE;
 571         unsigned long compressed_len;
 572         unsigned long nr_pages;
 573         unsigned long pg_index;
 574         struct page *page;
 575         struct block_device *bdev;
 576         struct bio *comp_bio;
 577         u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
 578         u64 em_len;
 579         u64 em_start;
 580         struct extent_map *em;
 581         int ret = -ENOMEM;
 582         int faili = 0;
 583         u32 *sums;
 584
 585         tree = &BTRFS_I(inode)->io_tree;
 586         em_tree = &BTRFS_I(inode)->extent_tree;
 587
 588         /* we need the actual starting offset of this extent in the file */
 589         read_lock(&em_tree->lock);
 590         em = lookup_extent_mapping(em_tree,
 591                                    page_offset(bio->bi_io_vec->bv_page),
 592                                    PAGE_SIZE);
 593         read_unlock(&em_tree->lock);
 594         if (!em)
 595                 return -EIO;
 596
 597         compressed_len = em->block_len;
 598         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 599         if (!cb)
 600                 goto out;
 601
 602         atomic_set(&cb->pending_bios, 0);
 603         cb->errors = 0;
 604         cb->inode = inode;
 605         cb->mirror_num = mirror_num;
 606         sums = &cb->sums;
 607
 608         cb->start = em->orig_start;
 609         em_len = em->len;
 610         em_start = em->start;
 611
 612         free_extent_map(em);
 613         em = NULL;
 614
 615         cb->len = uncompressed_len;
 616         cb->compressed_len = compressed_len;
 617         cb->compress_type = extent_compress_type(bio_flags);
 618         cb->orig_bio = bio;
 619
 620         nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
 621         cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
 622                                        GFP_NOFS);
 623         if (!cb->compressed_pages)
 624                 goto fail1;
 625
 626         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 627
 628         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 629                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 630                                                               __GFP_HIGHMEM);
 631                 if (!cb->compressed_pages[pg_index]) {
 632                         faili = pg_index - 1;
 633                         ret = -ENOMEM;
 634                         goto fail2;
 635                 }
 636         }
 637         faili = nr_pages - 1;
 638         cb->nr_pages = nr_pages;
 639
 640         add_ra_bio_pages(inode, em_start + em_len, cb);
 641
 642         /* include any pages we added in add_ra-bio_pages */
 643         uncompressed_len = bio->bi_vcnt * PAGE_SIZE;
 644         cb->len = uncompressed_len;
 645
 646         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 647         if (!comp_bio)
 648                 goto fail2;
 649         comp_bio->bi_private = cb;
 650         comp_bio->bi_end_io = end_compressed_bio_read;
 651         atomic_inc(&cb->pending_bios);
 652
 653         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 654                 page = cb->compressed_pages[pg_index];
 655                 page->mapping = inode->i_mapping;
 656                 page->index = em_start >> PAGE_SHIFT;
 657
 658                 if (comp_bio->bi_iter.bi_size)
 659                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 660                                                         PAGE_SIZE,
 661                                                         comp_bio, 0);
 662                 else
 663                         ret = 0;
 664
 665                 page->mapping = NULL;
 666                 if (ret || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
 667                     PAGE_SIZE) {
 668                         bio_get(comp_bio);
 669
 670                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 671                                         BTRFS_WQ_ENDIO_DATA);
 672                         BUG_ON(ret); /* -ENOMEM */
 673
 674                         /*
 675                          * inc the count before we submit the bio so
 676                          * we know the end IO handler won't happen before
 677                          * we inc the count.  Otherwise, the cb might get
 678                          * freed before we're done setting it up
 679                          */
 680                         atomic_inc(&cb->pending_bios);
 681
 682                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 683                                 ret = btrfs_lookup_bio_sums(root, inode,
 684                                                         comp_bio, sums);
 685                                 BUG_ON(ret); /* -ENOMEM */
 686                         }
 687                         sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
 688                                              root->sectorsize);
 689
 690                         ret = btrfs_map_bio(root, READ, comp_bio,
 691                                             mirror_num, 0);
 692                         if (ret) {
 693                                 bio->bi_error = ret;
 694                                 bio_endio(comp_bio);
 695                         }
 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_SIZE, 0);
 706                 }
 707                 cur_disk_byte += PAGE_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->bi_error = ret;
 723                 bio_endio(comp_bio);
 724         }
 725
 726         bio_put(comp_bio);
 727         return 0;
 728
 729 fail2:
 730         while (faili >= 0) {
 731                 __free_page(cb->compressed_pages[faili]);
 732                 faili--;
 733         }
 734
 735         kfree(cb->compressed_pages);
 736 fail1:
 737         kfree(cb);
 738 out:
 739         free_extent_map(em);
 740         return ret;
 741 }
 742
 743 static struct {
 744         struct list_head idle_ws;
 745         spinlock_t ws_lock;
 746         /* Number of free workspaces */
 747         int free_ws;
 748         /* Total number of allocated workspaces */
 749         atomic_t total_ws;
 750         /* Waiters for a free workspace */
 751         wait_queue_head_t ws_wait;
 752 } btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
 753
 754 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
 755         &btrfs_zlib_compress,
 756         &btrfs_lzo_compress,
 757 };
 758
 759 void __init btrfs_init_compress(void)
 760 {
 761         int i;
 762
 763         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 764                 struct list_head *workspace;
 765
 766                 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
 767                 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
 768                 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
 769                 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
 770
 771                 /*
 772                  * Preallocate one workspace for each compression type so
 773                  * we can guarantee forward progress in the worst case
 774                  */
 775                 workspace = btrfs_compress_op[i]->alloc_workspace();
 776                 if (IS_ERR(workspace)) {
 777                         printk(KERN_WARNING
 778         "BTRFS: cannot preallocate compression workspace, will try later");
 779                 } else {
 780                         atomic_set(&btrfs_comp_ws[i].total_ws, 1);
 781                         btrfs_comp_ws[i].free_ws = 1;
 782                         list_add(workspace, &btrfs_comp_ws[i].idle_ws);
 783                 }
 784         }
 785 }
 786
 787 /*
 788  * This finds an available workspace or allocates a new one.
 789  * If it's not possible to allocate a new one, waits until there's one.
 790  * Preallocation makes a forward progress guarantees and we do not return
 791  * errors.
 792  */
 793 static struct list_head *find_workspace(int type)
 794 {
 795         struct list_head *workspace;
 796         int cpus = num_online_cpus();
 797         int idx = type - 1;
 798
 799         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
 800         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
 801         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
 802         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
 803         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
 804 again:
 805         spin_lock(ws_lock);
 806         if (!list_empty(idle_ws)) {
 807                 workspace = idle_ws->next;
 808                 list_del(workspace);
 809                 (*free_ws)--;
 810                 spin_unlock(ws_lock);
 811                 return workspace;
 812
 813         }
 814         if (atomic_read(total_ws) > cpus) {
 815                 DEFINE_WAIT(wait);
 816
 817                 spin_unlock(ws_lock);
 818                 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
 819                 if (atomic_read(total_ws) > cpus && !*free_ws)
 820                         schedule();
 821                 finish_wait(ws_wait, &wait);
 822                 goto again;
 823         }
 824         atomic_inc(total_ws);
 825         spin_unlock(ws_lock);
 826
 827         workspace = btrfs_compress_op[idx]->alloc_workspace();
 828         if (IS_ERR(workspace)) {
 829                 atomic_dec(total_ws);
 830                 wake_up(ws_wait);
 831
 832                 /*
 833                  * Do not return the error but go back to waiting. There's a
 834                  * workspace preallocated for each type and the compression
 835                  * time is bounded so we get to a workspace eventually. This
 836                  * makes our caller's life easier.
 837                  *
 838                  * To prevent silent and low-probability deadlocks (when the
 839                  * initial preallocation fails), check if there are any
 840                  * workspaces at all.
 841                  */
 842                 if (atomic_read(total_ws) == 0) {
 843                         static DEFINE_RATELIMIT_STATE(_rs,
 844                                         /* once per minute */ 60 * HZ,
 845                                         /* no burst */ 1);
 846
 847                         if (__ratelimit(&_rs)) {
 848                                 printk(KERN_WARNING
 849                             "no compression workspaces, low memory, retrying");
 850                         }
 851                 }
 852                 goto again;
 853         }
 854         return workspace;
 855 }
 856
 857 /*
 858  * put a workspace struct back on the list or free it if we have enough
 859  * idle ones sitting around
 860  */
 861 static void free_workspace(int type, struct list_head *workspace)
 862 {
 863         int idx = type - 1;
 864         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
 865         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
 866         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
 867         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
 868         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
 869
 870         spin_lock(ws_lock);
 871         if (*free_ws < num_online_cpus()) {
 872                 list_add(workspace, idle_ws);
 873                 (*free_ws)++;
 874                 spin_unlock(ws_lock);
 875                 goto wake;
 876         }
 877         spin_unlock(ws_lock);
 878
 879         btrfs_compress_op[idx]->free_workspace(workspace);
 880         atomic_dec(total_ws);
 881 wake:
 882         /*
 883          * Make sure counter is updated before we wake up waiters.
 884          */
 885         smp_mb();
 886         if (waitqueue_active(ws_wait))
 887                 wake_up(ws_wait);
 888 }
 889
 890 /*
 891  * cleanup function for module exit
 892  */
 893 static void free_workspaces(void)
 894 {
 895         struct list_head *workspace;
 896         int i;
 897
 898         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 899                 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
 900                         workspace = btrfs_comp_ws[i].idle_ws.next;
 901                         list_del(workspace);
 902                         btrfs_compress_op[i]->free_workspace(workspace);
 903                         atomic_dec(&btrfs_comp_ws[i].total_ws);
 904                 }
 905         }
 906 }
 907
 908 /*
 909  * given an address space and start/len, compress the bytes.
 910  *
 911  * pages are allocated to hold the compressed result and stored
 912  * in 'pages'
 913  *
 914  * out_pages is used to return the number of pages allocated.  There
 915  * may be pages allocated even if we return an error
 916  *
 917  * total_in is used to return the number of bytes actually read.  It
 918  * may be smaller then len if we had to exit early because we
 919  * ran out of room in the pages array or because we cross the
 920  * max_out threshold.
 921  *
 922  * total_out is used to return the total number of compressed bytes
 923  *
 924  * max_out tells us the max number of bytes that we're allowed to
 925  * stuff into pages
 926  */
 927 int btrfs_compress_pages(int type, struct address_space *mapping,
 928                          u64 start, unsigned long len,
 929                          struct page **pages,
 930                          unsigned long nr_dest_pages,
 931                          unsigned long *out_pages,
 932                          unsigned long *total_in,
 933                          unsigned long *total_out,
 934                          unsigned long max_out)
 935 {
 936         struct list_head *workspace;
 937         int ret;
 938
 939         workspace = find_workspace(type);
 940
 941         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 942                                                       start, len, pages,
 943                                                       nr_dest_pages, out_pages,
 944                                                       total_in, total_out,
 945                                                       max_out);
 946         free_workspace(type, workspace);
 947         return ret;
 948 }
 949
 950 /*
 951  * pages_in is an array of pages with compressed data.
 952  *
 953  * disk_start is the starting logical offset of this array in the file
 954  *
 955  * bvec is a bio_vec of pages from the file that we want to decompress into
 956  *
 957  * vcnt is the count of pages in the biovec
 958  *
 959  * srclen is the number of bytes in pages_in
 960  *
 961  * The basic idea is that we have a bio that was created by readpages.
 962  * The pages in the bio are for the uncompressed data, and they may not
 963  * be contiguous.  They all correspond to the range of bytes covered by
 964  * the compressed extent.
 965  */
 966 static int btrfs_decompress_biovec(int type, struct page **pages_in,
 967                                    u64 disk_start, struct bio_vec *bvec,
 968                                    int vcnt, size_t srclen)
 969 {
 970         struct list_head *workspace;
 971         int ret;
 972
 973         workspace = find_workspace(type);
 974
 975         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 976                                                          disk_start,
 977                                                          bvec, vcnt, srclen);
 978         free_workspace(type, workspace);
 979         return ret;
 980 }
 981
 982 /*
 983  * a less complex decompression routine.  Our compressed data fits in a
 984  * single page, and we want to read a single page out of it.
 985  * start_byte tells us the offset into the compressed data we're interested in
 986  */
 987 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 988                      unsigned long start_byte, size_t srclen, size_t destlen)
 989 {
 990         struct list_head *workspace;
 991         int ret;
 992
 993         workspace = find_workspace(type);
 994
 995         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
 996                                                   dest_page, start_byte,
 997                                                   srclen, destlen);
 998
 999         free_workspace(type, workspace);
1000         return ret;
1001 }
1002
1003 void btrfs_exit_compress(void)
1004 {
1005         free_workspaces();
1006 }
1007
1008 /*
1009  * Copy uncompressed data from working buffer to pages.
1010  *
1011  * buf_start is the byte offset we're of the start of our workspace buffer.
1012  *
1013  * total_out is the last byte of the buffer
1014  */
1015 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
1016                               unsigned long total_out, u64 disk_start,
1017                               struct bio_vec *bvec, int vcnt,
1018                               unsigned long *pg_index,
1019                               unsigned long *pg_offset)
1020 {
1021         unsigned long buf_offset;
1022         unsigned long current_buf_start;
1023         unsigned long start_byte;
1024         unsigned long working_bytes = total_out - buf_start;
1025         unsigned long bytes;
1026         char *kaddr;
1027         struct page *page_out = bvec[*pg_index].bv_page;
1028
1029         /*
1030          * start byte is the first byte of the page we're currently
1031          * copying into relative to the start of the compressed data.
1032          */
1033         start_byte = page_offset(page_out) - disk_start;
1034
1035         /* we haven't yet hit data corresponding to this page */
1036         if (total_out <= start_byte)
1037                 return 1;
1038
1039         /*
1040          * the start of the data we care about is offset into
1041          * the middle of our working buffer
1042          */
1043         if (total_out > start_byte && buf_start < start_byte) {
1044                 buf_offset = start_byte - buf_start;
1045                 working_bytes -= buf_offset;
1046         } else {
1047                 buf_offset = 0;
1048         }
1049         current_buf_start = buf_start;
1050
1051         /* copy bytes from the working buffer into the pages */
1052         while (working_bytes > 0) {
1053                 bytes = min(PAGE_SIZE - *pg_offset,
1054                             PAGE_SIZE - buf_offset);
1055                 bytes = min(bytes, working_bytes);
1056                 kaddr = kmap_atomic(page_out);
1057                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1058                 kunmap_atomic(kaddr);
1059                 flush_dcache_page(page_out);
1060
1061                 *pg_offset += bytes;
1062                 buf_offset += bytes;
1063                 working_bytes -= bytes;
1064                 current_buf_start += bytes;
1065
1066                 /* check if we need to pick another page */
1067                 if (*pg_offset == PAGE_SIZE) {
1068                         (*pg_index)++;
1069                         if (*pg_index >= vcnt)
1070                                 return 0;
1071
1072                         page_out = bvec[*pg_index].bv_page;
1073                         *pg_offset = 0;
1074                         start_byte = page_offset(page_out) - disk_start;
1075
1076                         /*
1077                          * make sure our new page is covered by this
1078                          * working buffer
1079                          */
1080                         if (total_out <= start_byte)
1081                                 return 1;
1082
1083                         /*
1084                          * the next page in the biovec might not be adjacent
1085                          * to the last page, but it might still be found
1086                          * inside this working buffer. bump our offset pointer
1087                          */
1088                         if (total_out > start_byte &&
1089                             current_buf_start < start_byte) {
1090                                 buf_offset = start_byte - buf_start;
1091                                 working_bytes = total_out - start_byte;
1092                                 current_buf_start = buf_start + buf_offset;
1093                         }
1094                 }
1095         }
1096
1097         return 1;
1098 }
1099
1100 /*
1101  * When uncompressing data, we need to make sure and zero any parts of
1102  * the biovec that were not filled in by the decompression code.  pg_index
1103  * and pg_offset indicate the last page and the last offset of that page
1104  * that have been filled in.  This will zero everything remaining in the
1105  * biovec.
1106  */
1107 void btrfs_clear_biovec_end(struct bio_vec *bvec, int vcnt,
1108                                    unsigned long pg_index,
1109                                    unsigned long pg_offset)
1110 {
1111         while (pg_index < vcnt) {
1112                 struct page *page = bvec[pg_index].bv_page;
1113                 unsigned long off = bvec[pg_index].bv_offset;
1114                 unsigned long len = bvec[pg_index].bv_len;
1115
1116                 if (pg_offset < off)
1117                         pg_offset = off;
1118                 if (pg_offset < off + len) {
1119                         unsigned long bytes = off + len - pg_offset;
1120                         char *kaddr;
1121
1122                         kaddr = kmap_atomic(page);
1123                         memset(kaddr + pg_offset, 0, bytes);
1124                         kunmap_atomic(kaddr);
1125                 }
1126                 pg_index++;
1127                 pg_offset = 0;
1128         }
1129 }