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                         if (ret) {
 406                                 bio->bi_error = ret;
 407                                 bio_endio(bio);
 408                         }
 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_SIZE, 0);
 417                 }
 418                 if (bytes_left < PAGE_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_SIZE;
 424                 first_byte += PAGE_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         if (ret) {
 439                 bio->bi_error = ret;
 440                 bio_endio(bio);
 441         }
 442
 443         bio_put(bio);
 444         return 0;
 445 }
 446
 447 static noinline int add_ra_bio_pages(struct inode *inode,
 448                                      u64 compressed_end,
 449                                      struct compressed_bio *cb)
 450 {
 451         unsigned long end_index;
 452         unsigned long pg_index;
 453         u64 last_offset;
 454         u64 isize = i_size_read(inode);
 455         int ret;
 456         struct page *page;
 457         unsigned long nr_pages = 0;
 458         struct extent_map *em;
 459         struct address_space *mapping = inode->i_mapping;
 460         struct extent_map_tree *em_tree;
 461         struct extent_io_tree *tree;
 462         u64 end;
 463         int misses = 0;
 464
 465         page = cb->orig_bio->bi_io_vec[cb->orig_bio->bi_vcnt - 1].bv_page;
 466         last_offset = (page_offset(page) + PAGE_SIZE);
 467         em_tree = &BTRFS_I(inode)->extent_tree;
 468         tree = &BTRFS_I(inode)->io_tree;
 469
 470         if (isize == 0)
 471                 return 0;
 472
 473         end_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 474
 475         while (last_offset < compressed_end) {
 476                 pg_index = last_offset >> PAGE_SHIFT;
 477
 478                 if (pg_index > end_index)
 479                         break;
 480
 481                 rcu_read_lock();
 482                 page = radix_tree_lookup(&mapping->page_tree, pg_index);
 483                 rcu_read_unlock();
 484                 if (page && !radix_tree_exceptional_entry(page)) {
 485                         misses++;
 486                         if (misses > 4)
 487                                 break;
 488                         goto next;
 489                 }
 490
 491                 page = __page_cache_alloc(mapping_gfp_constraint(mapping,
 492                                                                  ~__GFP_FS));
 493                 if (!page)
 494                         break;
 495
 496                 if (add_to_page_cache_lru(page, mapping, pg_index, GFP_NOFS)) {
 497                         put_page(page);
 498                         goto next;
 499                 }
 500
 501                 end = last_offset + PAGE_SIZE - 1;
 502                 /*
 503                  * at this point, we have a locked page in the page cache
 504                  * for these bytes in the file.  But, we have to make
 505                  * sure they map to this compressed extent on disk.
 506                  */
 507                 set_page_extent_mapped(page);
 508                 lock_extent(tree, last_offset, end);
 509                 read_lock(&em_tree->lock);
 510                 em = lookup_extent_mapping(em_tree, last_offset,
 511                                            PAGE_SIZE);
 512                 read_unlock(&em_tree->lock);
 513
 514                 if (!em || last_offset < em->start ||
 515                     (last_offset + PAGE_SIZE > extent_map_end(em)) ||
 516                     (em->block_start >> 9) != cb->orig_bio->bi_iter.bi_sector) {
 517                         free_extent_map(em);
 518                         unlock_extent(tree, last_offset, end);
 519                         unlock_page(page);
 520                         put_page(page);
 521                         break;
 522                 }
 523                 free_extent_map(em);
 524
 525                 if (page->index == end_index) {
 526                         char *userpage;
 527                         size_t zero_offset = isize & (PAGE_SIZE - 1);
 528
 529                         if (zero_offset) {
 530                                 int zeros;
 531                                 zeros = PAGE_SIZE - zero_offset;
 532                                 userpage = kmap_atomic(page);
 533                                 memset(userpage + zero_offset, 0, zeros);
 534                                 flush_dcache_page(page);
 535                                 kunmap_atomic(userpage);
 536                         }
 537                 }
 538
 539                 ret = bio_add_page(cb->orig_bio, page,
 540                                    PAGE_SIZE, 0);
 541
 542                 if (ret == PAGE_SIZE) {
 543                         nr_pages++;
 544                         put_page(page);
 545                 } else {
 546                         unlock_extent(tree, last_offset, end);
 547                         unlock_page(page);
 548                         put_page(page);
 549                         break;
 550                 }
 551 next:
 552                 last_offset += PAGE_SIZE;
 553         }
 554         return 0;
 555 }
 556
 557 /*
 558  * for a compressed read, the bio we get passed has all the inode pages
 559  * in it.  We don't actually do IO on those pages but allocate new ones
 560  * to hold the compressed pages on disk.
 561  *
 562  * bio->bi_iter.bi_sector points to the compressed extent on disk
 563  * bio->bi_io_vec points to all of the inode pages
 564  * bio->bi_vcnt is a count of pages
 565  *
 566  * After the compressed pages are read, we copy the bytes into the
 567  * bio we were passed and then call the bio end_io calls
 568  */
 569 int btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,
 570                                  int mirror_num, unsigned long bio_flags)
 571 {
 572         struct extent_io_tree *tree;
 573         struct extent_map_tree *em_tree;
 574         struct compressed_bio *cb;
 575         struct btrfs_root *root = BTRFS_I(inode)->root;
 576         unsigned long uncompressed_len = bio->bi_vcnt * PAGE_SIZE;
 577         unsigned long compressed_len;
 578         unsigned long nr_pages;
 579         unsigned long pg_index;
 580         struct page *page;
 581         struct block_device *bdev;
 582         struct bio *comp_bio;
 583         u64 cur_disk_byte = (u64)bio->bi_iter.bi_sector << 9;
 584         u64 em_len;
 585         u64 em_start;
 586         struct extent_map *em;
 587         int ret = -ENOMEM;
 588         int faili = 0;
 589         u32 *sums;
 590
 591         tree = &BTRFS_I(inode)->io_tree;
 592         em_tree = &BTRFS_I(inode)->extent_tree;
 593
 594         /* we need the actual starting offset of this extent in the file */
 595         read_lock(&em_tree->lock);
 596         em = lookup_extent_mapping(em_tree,
 597                                    page_offset(bio->bi_io_vec->bv_page),
 598                                    PAGE_SIZE);
 599         read_unlock(&em_tree->lock);
 600         if (!em)
 601                 return -EIO;
 602
 603         compressed_len = em->block_len;
 604         cb = kmalloc(compressed_bio_size(root, compressed_len), GFP_NOFS);
 605         if (!cb)
 606                 goto out;
 607
 608         atomic_set(&cb->pending_bios, 0);
 609         cb->errors = 0;
 610         cb->inode = inode;
 611         cb->mirror_num = mirror_num;
 612         sums = &cb->sums;
 613
 614         cb->start = em->orig_start;
 615         em_len = em->len;
 616         em_start = em->start;
 617
 618         free_extent_map(em);
 619         em = NULL;
 620
 621         cb->len = uncompressed_len;
 622         cb->compressed_len = compressed_len;
 623         cb->compress_type = extent_compress_type(bio_flags);
 624         cb->orig_bio = bio;
 625
 626         nr_pages = DIV_ROUND_UP(compressed_len, PAGE_SIZE);
 627         cb->compressed_pages = kcalloc(nr_pages, sizeof(struct page *),
 628                                        GFP_NOFS);
 629         if (!cb->compressed_pages)
 630                 goto fail1;
 631
 632         bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
 633
 634         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 635                 cb->compressed_pages[pg_index] = alloc_page(GFP_NOFS |
 636                                                               __GFP_HIGHMEM);
 637                 if (!cb->compressed_pages[pg_index]) {
 638                         faili = pg_index - 1;
 639                         ret = -ENOMEM;
 640                         goto fail2;
 641                 }
 642         }
 643         faili = nr_pages - 1;
 644         cb->nr_pages = nr_pages;
 645
 646         add_ra_bio_pages(inode, em_start + em_len, cb);
 647
 648         /* include any pages we added in add_ra-bio_pages */
 649         uncompressed_len = bio->bi_vcnt * PAGE_SIZE;
 650         cb->len = uncompressed_len;
 651
 652         comp_bio = compressed_bio_alloc(bdev, cur_disk_byte, GFP_NOFS);
 653         if (!comp_bio)
 654                 goto fail2;
 655         comp_bio->bi_private = cb;
 656         comp_bio->bi_end_io = end_compressed_bio_read;
 657         atomic_inc(&cb->pending_bios);
 658
 659         for (pg_index = 0; pg_index < nr_pages; pg_index++) {
 660                 page = cb->compressed_pages[pg_index];
 661                 page->mapping = inode->i_mapping;
 662                 page->index = em_start >> PAGE_SHIFT;
 663
 664                 if (comp_bio->bi_iter.bi_size)
 665                         ret = tree->ops->merge_bio_hook(READ, page, 0,
 666                                                         PAGE_SIZE,
 667                                                         comp_bio, 0);
 668                 else
 669                         ret = 0;
 670
 671                 page->mapping = NULL;
 672                 if (ret || bio_add_page(comp_bio, page, PAGE_SIZE, 0) <
 673                     PAGE_SIZE) {
 674                         bio_get(comp_bio);
 675
 676                         ret = btrfs_bio_wq_end_io(root->fs_info, comp_bio,
 677                                         BTRFS_WQ_ENDIO_DATA);
 678                         BUG_ON(ret); /* -ENOMEM */
 679
 680                         /*
 681                          * inc the count before we submit the bio so
 682                          * we know the end IO handler won't happen before
 683                          * we inc the count.  Otherwise, the cb might get
 684                          * freed before we're done setting it up
 685                          */
 686                         atomic_inc(&cb->pending_bios);
 687
 688                         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
 689                                 ret = btrfs_lookup_bio_sums(root, inode,
 690                                                         comp_bio, sums);
 691                                 BUG_ON(ret); /* -ENOMEM */
 692                         }
 693                         sums += DIV_ROUND_UP(comp_bio->bi_iter.bi_size,
 694                                              root->sectorsize);
 695
 696                         ret = btrfs_map_bio(root, READ, comp_bio,
 697                                             mirror_num, 0);
 698                         if (ret) {
 699                                 bio->bi_error = ret;
 700                                 bio_endio(comp_bio);
 701                         }
 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_SIZE, 0);
 712                 }
 713                 cur_disk_byte += PAGE_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->bi_error = ret;
 729                 bio_endio(comp_bio);
 730         }
 731
 732         bio_put(comp_bio);
 733         return 0;
 734
 735 fail2:
 736         while (faili >= 0) {
 737                 __free_page(cb->compressed_pages[faili]);
 738                 faili--;
 739         }
 740
 741         kfree(cb->compressed_pages);
 742 fail1:
 743         kfree(cb);
 744 out:
 745         free_extent_map(em);
 746         return ret;
 747 }
 748
 749 static struct {
 750         struct list_head idle_ws;
 751         spinlock_t ws_lock;
 752         /* Number of free workspaces */
 753         int free_ws;
 754         /* Total number of allocated workspaces */
 755         atomic_t total_ws;
 756         /* Waiters for a free workspace */
 757         wait_queue_head_t ws_wait;
 758 } btrfs_comp_ws[BTRFS_COMPRESS_TYPES];
 759
 760 static const struct btrfs_compress_op * const btrfs_compress_op[] = {
 761         &btrfs_zlib_compress,
 762         &btrfs_lzo_compress,
 763 };
 764
 765 void __init btrfs_init_compress(void)
 766 {
 767         int i;
 768
 769         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 770                 struct list_head *workspace;
 771
 772                 INIT_LIST_HEAD(&btrfs_comp_ws[i].idle_ws);
 773                 spin_lock_init(&btrfs_comp_ws[i].ws_lock);
 774                 atomic_set(&btrfs_comp_ws[i].total_ws, 0);
 775                 init_waitqueue_head(&btrfs_comp_ws[i].ws_wait);
 776
 777                 /*
 778                  * Preallocate one workspace for each compression type so
 779                  * we can guarantee forward progress in the worst case
 780                  */
 781                 workspace = btrfs_compress_op[i]->alloc_workspace();
 782                 if (IS_ERR(workspace)) {
 783                         printk(KERN_WARNING
 784         "BTRFS: cannot preallocate compression workspace, will try later");
 785                 } else {
 786                         atomic_set(&btrfs_comp_ws[i].total_ws, 1);
 787                         btrfs_comp_ws[i].free_ws = 1;
 788                         list_add(workspace, &btrfs_comp_ws[i].idle_ws);
 789                 }
 790         }
 791 }
 792
 793 /*
 794  * This finds an available workspace or allocates a new one.
 795  * If it's not possible to allocate a new one, waits until there's one.
 796  * Preallocation makes a forward progress guarantees and we do not return
 797  * errors.
 798  */
 799 static struct list_head *find_workspace(int type)
 800 {
 801         struct list_head *workspace;
 802         int cpus = num_online_cpus();
 803         int idx = type - 1;
 804
 805         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
 806         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
 807         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
 808         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
 809         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
 810 again:
 811         spin_lock(ws_lock);
 812         if (!list_empty(idle_ws)) {
 813                 workspace = idle_ws->next;
 814                 list_del(workspace);
 815                 (*free_ws)--;
 816                 spin_unlock(ws_lock);
 817                 return workspace;
 818
 819         }
 820         if (atomic_read(total_ws) > cpus) {
 821                 DEFINE_WAIT(wait);
 822
 823                 spin_unlock(ws_lock);
 824                 prepare_to_wait(ws_wait, &wait, TASK_UNINTERRUPTIBLE);
 825                 if (atomic_read(total_ws) > cpus && !*free_ws)
 826                         schedule();
 827                 finish_wait(ws_wait, &wait);
 828                 goto again;
 829         }
 830         atomic_inc(total_ws);
 831         spin_unlock(ws_lock);
 832
 833         workspace = btrfs_compress_op[idx]->alloc_workspace();
 834         if (IS_ERR(workspace)) {
 835                 atomic_dec(total_ws);
 836                 wake_up(ws_wait);
 837
 838                 /*
 839                  * Do not return the error but go back to waiting. There's a
 840                  * workspace preallocated for each type and the compression
 841                  * time is bounded so we get to a workspace eventually. This
 842                  * makes our caller's life easier.
 843                  *
 844                  * To prevent silent and low-probability deadlocks (when the
 845                  * initial preallocation fails), check if there are any
 846                  * workspaces at all.
 847                  */
 848                 if (atomic_read(total_ws) == 0) {
 849                         static DEFINE_RATELIMIT_STATE(_rs,
 850                                         /* once per minute */ 60 * HZ,
 851                                         /* no burst */ 1);
 852
 853                         if (__ratelimit(&_rs)) {
 854                                 printk(KERN_WARNING
 855                             "no compression workspaces, low memory, retrying");
 856                         }
 857                 }
 858                 goto again;
 859         }
 860         return workspace;
 861 }
 862
 863 /*
 864  * put a workspace struct back on the list or free it if we have enough
 865  * idle ones sitting around
 866  */
 867 static void free_workspace(int type, struct list_head *workspace)
 868 {
 869         int idx = type - 1;
 870         struct list_head *idle_ws       = &btrfs_comp_ws[idx].idle_ws;
 871         spinlock_t *ws_lock             = &btrfs_comp_ws[idx].ws_lock;
 872         atomic_t *total_ws              = &btrfs_comp_ws[idx].total_ws;
 873         wait_queue_head_t *ws_wait      = &btrfs_comp_ws[idx].ws_wait;
 874         int *free_ws                    = &btrfs_comp_ws[idx].free_ws;
 875
 876         spin_lock(ws_lock);
 877         if (*free_ws < num_online_cpus()) {
 878                 list_add(workspace, idle_ws);
 879                 (*free_ws)++;
 880                 spin_unlock(ws_lock);
 881                 goto wake;
 882         }
 883         spin_unlock(ws_lock);
 884
 885         btrfs_compress_op[idx]->free_workspace(workspace);
 886         atomic_dec(total_ws);
 887 wake:
 888         /*
 889          * Make sure counter is updated before we wake up waiters.
 890          */
 891         smp_mb();
 892         if (waitqueue_active(ws_wait))
 893                 wake_up(ws_wait);
 894 }
 895
 896 /*
 897  * cleanup function for module exit
 898  */
 899 static void free_workspaces(void)
 900 {
 901         struct list_head *workspace;
 902         int i;
 903
 904         for (i = 0; i < BTRFS_COMPRESS_TYPES; i++) {
 905                 while (!list_empty(&btrfs_comp_ws[i].idle_ws)) {
 906                         workspace = btrfs_comp_ws[i].idle_ws.next;
 907                         list_del(workspace);
 908                         btrfs_compress_op[i]->free_workspace(workspace);
 909                         atomic_dec(&btrfs_comp_ws[i].total_ws);
 910                 }
 911         }
 912 }
 913
 914 /*
 915  * given an address space and start/len, compress the bytes.
 916  *
 917  * pages are allocated to hold the compressed result and stored
 918  * in 'pages'
 919  *
 920  * out_pages is used to return the number of pages allocated.  There
 921  * may be pages allocated even if we return an error
 922  *
 923  * total_in is used to return the number of bytes actually read.  It
 924  * may be smaller then len if we had to exit early because we
 925  * ran out of room in the pages array or because we cross the
 926  * max_out threshold.
 927  *
 928  * total_out is used to return the total number of compressed bytes
 929  *
 930  * max_out tells us the max number of bytes that we're allowed to
 931  * stuff into pages
 932  */
 933 int btrfs_compress_pages(int type, struct address_space *mapping,
 934                          u64 start, unsigned long len,
 935                          struct page **pages,
 936                          unsigned long nr_dest_pages,
 937                          unsigned long *out_pages,
 938                          unsigned long *total_in,
 939                          unsigned long *total_out,
 940                          unsigned long max_out)
 941 {
 942         struct list_head *workspace;
 943         int ret;
 944
 945         workspace = find_workspace(type);
 946
 947         ret = btrfs_compress_op[type-1]->compress_pages(workspace, mapping,
 948                                                       start, len, pages,
 949                                                       nr_dest_pages, out_pages,
 950                                                       total_in, total_out,
 951                                                       max_out);
 952         free_workspace(type, workspace);
 953         return ret;
 954 }
 955
 956 /*
 957  * pages_in is an array of pages with compressed data.
 958  *
 959  * disk_start is the starting logical offset of this array in the file
 960  *
 961  * bvec is a bio_vec of pages from the file that we want to decompress into
 962  *
 963  * vcnt is the count of pages in the biovec
 964  *
 965  * srclen is the number of bytes in pages_in
 966  *
 967  * The basic idea is that we have a bio that was created by readpages.
 968  * The pages in the bio are for the uncompressed data, and they may not
 969  * be contiguous.  They all correspond to the range of bytes covered by
 970  * the compressed extent.
 971  */
 972 static int btrfs_decompress_biovec(int type, struct page **pages_in,
 973                                    u64 disk_start, struct bio_vec *bvec,
 974                                    int vcnt, size_t srclen)
 975 {
 976         struct list_head *workspace;
 977         int ret;
 978
 979         workspace = find_workspace(type);
 980
 981         ret = btrfs_compress_op[type-1]->decompress_biovec(workspace, pages_in,
 982                                                          disk_start,
 983                                                          bvec, vcnt, srclen);
 984         free_workspace(type, workspace);
 985         return ret;
 986 }
 987
 988 /*
 989  * a less complex decompression routine.  Our compressed data fits in a
 990  * single page, and we want to read a single page out of it.
 991  * start_byte tells us the offset into the compressed data we're interested in
 992  */
 993 int btrfs_decompress(int type, unsigned char *data_in, struct page *dest_page,
 994                      unsigned long start_byte, size_t srclen, size_t destlen)
 995 {
 996         struct list_head *workspace;
 997         int ret;
 998
 999         workspace = find_workspace(type);
1000
1001         ret = btrfs_compress_op[type-1]->decompress(workspace, data_in,
1002                                                   dest_page, start_byte,
1003                                                   srclen, destlen);
1004
1005         free_workspace(type, workspace);
1006         return ret;
1007 }
1008
1009 void btrfs_exit_compress(void)
1010 {
1011         free_workspaces();
1012 }
1013
1014 /*
1015  * Copy uncompressed data from working buffer to pages.
1016  *
1017  * buf_start is the byte offset we're of the start of our workspace buffer.
1018  *
1019  * total_out is the last byte of the buffer
1020  */
1021 int btrfs_decompress_buf2page(char *buf, unsigned long buf_start,
1022                               unsigned long total_out, u64 disk_start,
1023                               struct bio_vec *bvec, int vcnt,
1024                               unsigned long *pg_index,
1025                               unsigned long *pg_offset)
1026 {
1027         unsigned long buf_offset;
1028         unsigned long current_buf_start;
1029         unsigned long start_byte;
1030         unsigned long working_bytes = total_out - buf_start;
1031         unsigned long bytes;
1032         char *kaddr;
1033         struct page *page_out = bvec[*pg_index].bv_page;
1034
1035         /*
1036          * start byte is the first byte of the page we're currently
1037          * copying into relative to the start of the compressed data.
1038          */
1039         start_byte = page_offset(page_out) - disk_start;
1040
1041         /* we haven't yet hit data corresponding to this page */
1042         if (total_out <= start_byte)
1043                 return 1;
1044
1045         /*
1046          * the start of the data we care about is offset into
1047          * the middle of our working buffer
1048          */
1049         if (total_out > start_byte && buf_start < start_byte) {
1050                 buf_offset = start_byte - buf_start;
1051                 working_bytes -= buf_offset;
1052         } else {
1053                 buf_offset = 0;
1054         }
1055         current_buf_start = buf_start;
1056
1057         /* copy bytes from the working buffer into the pages */
1058         while (working_bytes > 0) {
1059                 bytes = min(PAGE_SIZE - *pg_offset,
1060                             PAGE_SIZE - buf_offset);
1061                 bytes = min(bytes, working_bytes);
1062                 kaddr = kmap_atomic(page_out);
1063                 memcpy(kaddr + *pg_offset, buf + buf_offset, bytes);
1064                 kunmap_atomic(kaddr);
1065                 flush_dcache_page(page_out);
1066
1067                 *pg_offset += bytes;
1068                 buf_offset += bytes;
1069                 working_bytes -= bytes;
1070                 current_buf_start += bytes;
1071
1072                 /* check if we need to pick another page */
1073                 if (*pg_offset == PAGE_SIZE) {
1074                         (*pg_index)++;
1075                         if (*pg_index >= vcnt)
1076                                 return 0;
1077
1078                         page_out = bvec[*pg_index].bv_page;
1079                         *pg_offset = 0;
1080                         start_byte = page_offset(page_out) - disk_start;
1081
1082                         /*
1083                          * make sure our new page is covered by this
1084                          * working buffer
1085                          */
1086                         if (total_out <= start_byte)
1087                                 return 1;
1088
1089                         /*
1090                          * the next page in the biovec might not be adjacent
1091                          * to the last page, but it might still be found
1092                          * inside this working buffer. bump our offset pointer
1093                          */
1094                         if (total_out > start_byte &&
1095                             current_buf_start < start_byte) {
1096                                 buf_offset = start_byte - buf_start;
1097                                 working_bytes = total_out - start_byte;
1098                                 current_buf_start = buf_start + buf_offset;
1099                         }
1100                 }
1101         }
1102
1103         return 1;
1104 }
1105
1106 /*
1107  * When uncompressing data, we need to make sure and zero any parts of
1108  * the biovec that were not filled in by the decompression code.  pg_index
1109  * and pg_offset indicate the last page and the last offset of that page
1110  * that have been filled in.  This will zero everything remaining in the
1111  * biovec.
1112  */
1113 void btrfs_clear_biovec_end(struct bio_vec *bvec, int vcnt,
1114                                    unsigned long pg_index,
1115                                    unsigned long pg_offset)
1116 {
1117         while (pg_index < vcnt) {
1118                 struct page *page = bvec[pg_index].bv_page;
1119                 unsigned long off = bvec[pg_index].bv_offset;
1120                 unsigned long len = bvec[pg_index].bv_len;
1121
1122                 if (pg_offset < off)
1123                         pg_offset = off;
1124                 if (pg_offset < off + len) {
1125                         unsigned long bytes = off + len - pg_offset;
1126                         char *kaddr;
1127
1128                         kaddr = kmap_atomic(page);
1129                         memset(kaddr + pg_offset, 0, bytes);
1130                         kunmap_atomic(kaddr);
1131                 }
1132                 pg_index++;
1133                 pg_offset = 0;
1134         }
1135 }