disk-io.c

   1 /*
   2  * Copyright (C) 2007 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 #define _XOPEN_SOURCE 600
  20 #define __USE_XOPEN2K
  21 #define _GNU_SOURCE 1
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <sys/types.h>
  25 #include <sys/stat.h>
  26 #include <fcntl.h>
  27 #include <unistd.h>
  28 #include "kerncompat.h"
  29 #include "radix-tree.h"
  30 #include "ctree.h"
  31 #include "disk-io.h"
  32 #include "volumes.h"
  33 #include "transaction.h"
  34 #include "crc32c.h"
  35 #include "utils.h"
  36 #include "print-tree.h"
  37
  38 static int close_all_devices(struct btrfs_fs_info *fs_info);
  39
  40 static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
  41 {
  42
  43         struct btrfs_fs_devices *fs_devices;
  44         int ret = 1;
  45
  46         if (buf->start != btrfs_header_bytenr(buf)) {
  47                 printk("Check tree block failed, want=%Lu, have=%Lu\n",
  48                        buf->start, btrfs_header_bytenr(buf));
  49                 return ret;
  50         }
  51
  52         fs_devices = root->fs_info->fs_devices;
  53         while (fs_devices) {
  54                 if (!memcmp_extent_buffer(buf, fs_devices->fsid,
  55                                           (unsigned long)btrfs_header_fsid(buf),
  56                                           BTRFS_FSID_SIZE)) {
  57                         ret = 0;
  58                         break;
  59                 }
  60                 fs_devices = fs_devices->seed;
  61         }
  62         return ret;
  63 }
  64
  65 u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len)
  66 {
  67         return crc32c(seed, data, len);
  68 }
  69
  70 void btrfs_csum_final(u32 crc, char *result)
  71 {
  72         *(__le32 *)result = ~cpu_to_le32(crc);
  73 }
  74
  75 int csum_tree_block_size(struct extent_buffer *buf, u16 csum_size,
  76                          int verify)
  77 {
  78         char *result;
  79         u32 len;
  80         u32 crc = ~(u32)0;
  81
  82         result = malloc(csum_size * sizeof(char));
  83         if (!result)
  84                 return 1;
  85
  86         len = buf->len - BTRFS_CSUM_SIZE;
  87         crc = crc32c(crc, buf->data + BTRFS_CSUM_SIZE, len);
  88         btrfs_csum_final(crc, result);
  89
  90         if (verify) {
  91                 if (memcmp_extent_buffer(buf, result, 0, csum_size)) {
  92                         printk("checksum verify failed on %llu wanted %X "
  93                                "found %X\n", (unsigned long long)buf->start,
  94                                *((int *)result), *((char *)buf->data));
  95                         free(result);
  96                         return 1;
  97                 }
  98         } else {
  99                 write_extent_buffer(buf, result, 0, csum_size);
 100         }
 101         free(result);
 102         return 0;
 103 }
 104
 105 int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf,
 106                     int verify)
 107 {
 108         u16 csum_size =
 109                 btrfs_super_csum_size(&root->fs_info->super_copy);
 110         return csum_tree_block_size(buf, csum_size, verify);
 111 }
 112
 113 struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
 114                                             u64 bytenr, u32 blocksize)
 115 {
 116         return find_extent_buffer(&root->fs_info->extent_cache,
 117                                   bytenr, blocksize);
 118 }
 119
 120 struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
 121                                                  u64 bytenr, u32 blocksize)
 122 {
 123         return alloc_extent_buffer(&root->fs_info->extent_cache, bytenr,
 124                                    blocksize);
 125 }
 126
 127 int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 128                          u64 parent_transid)
 129 {
 130         int ret;
 131         struct extent_buffer *eb;
 132         u64 length;
 133         struct btrfs_multi_bio *multi = NULL;
 134         struct btrfs_device *device;
 135
 136         eb = btrfs_find_tree_block(root, bytenr, blocksize);
 137         if (eb && btrfs_buffer_uptodate(eb, parent_transid)) {
 138                 free_extent_buffer(eb);
 139                 return 0;
 140         }
 141
 142         length = blocksize;
 143         ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
 144                               bytenr, &length, &multi, 0);
 145         BUG_ON(ret);
 146         device = multi->stripes[0].dev;
 147         device->total_ios++;
 148         blocksize = min(blocksize, (u32)(64 * 1024));
 149         readahead(device->fd, multi->stripes[0].physical, blocksize);
 150         kfree(multi);
 151         return 0;
 152 }
 153
 154 static int verify_parent_transid(struct extent_io_tree *io_tree,
 155                                  struct extent_buffer *eb, u64 parent_transid,
 156                                  int ignore)
 157 {
 158         int ret;
 159
 160         if (!parent_transid || btrfs_header_generation(eb) == parent_transid)
 161                 return 0;
 162
 163         if (extent_buffer_uptodate(eb) &&
 164             btrfs_header_generation(eb) == parent_transid) {
 165                 ret = 0;
 166                 goto out;
 167         }
 168         printk("parent transid verify failed on %llu wanted %llu found %llu\n",
 169                (unsigned long long)eb->start,
 170                (unsigned long long)parent_transid,
 171                (unsigned long long)btrfs_header_generation(eb));
 172         if (ignore) {
 173                 printk("Ignoring transid failure\n");
 174                 return 0;
 175         }
 176
 177         ret = 1;
 178 out:
 179         clear_extent_buffer_uptodate(io_tree, eb);
 180         return ret;
 181
 182 }
 183
 184
 185 struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 186                                      u32 blocksize, u64 parent_transid)
 187 {
 188         int ret;
 189         struct extent_buffer *eb;
 190         u64 length;
 191         u64 best_transid = 0;
 192         struct btrfs_multi_bio *multi = NULL;
 193         struct btrfs_device *device;
 194         int mirror_num = 0;
 195         int good_mirror = 0;
 196         int num_copies;
 197         int ignore = 0;
 198
 199         eb = btrfs_find_create_tree_block(root, bytenr, blocksize);
 200         if (!eb)
 201                 return NULL;
 202
 203         if (btrfs_buffer_uptodate(eb, parent_transid))
 204                 return eb;
 205
 206         length = blocksize;
 207         while (1) {
 208                 ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
 209                                       eb->start, &length, &multi, mirror_num);
 210                 if (ret) {
 211                         printk("Couldn't map the block %Lu\n", bytenr);
 212                         break;
 213                 }
 214                 device = multi->stripes[0].dev;
 215                 eb->fd = device->fd;
 216                 device->total_ios++;
 217                 eb->dev_bytenr = multi->stripes[0].physical;
 218                 kfree(multi);
 219                 ret = read_extent_from_disk(eb);
 220
 221                 if (ret == 0 && check_tree_block(root, eb) == 0 &&
 222                     csum_tree_block(root, eb, 1) == 0 &&
 223                     verify_parent_transid(eb->tree, eb, parent_transid, ignore)
 224                     == 0) {
 225                         btrfs_set_buffer_uptodate(eb);
 226                         return eb;
 227                 }
 228                 if (ignore) {
 229                         if (check_tree_block(root, eb))
 230                                 printk("read block failed check_tree_block\n");
 231                         else
 232                                 printk("Csum didn't match\n");
 233                         break;
 234                 }
 235                 num_copies = btrfs_num_copies(&root->fs_info->mapping_tree,
 236                                               eb->start, eb->len);
 237                 if (num_copies == 1) {
 238                         ignore = 1;
 239                         continue;
 240                 }
 241                 if (btrfs_header_generation(eb) > best_transid) {
 242                         best_transid = btrfs_header_generation(eb);
 243                         good_mirror = mirror_num;
 244                 }
 245                 mirror_num++;
 246                 if (mirror_num > num_copies) {
 247                         mirror_num = good_mirror;
 248                         ignore = 1;
 249                         continue;
 250                 }
 251         }
 252         free_extent_buffer(eb);
 253         return NULL;
 254 }
 255
 256 int write_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 257                      struct extent_buffer *eb)
 258 {
 259         int ret;
 260         int dev_nr;
 261         u64 length;
 262         struct btrfs_multi_bio *multi = NULL;
 263
 264         if (check_tree_block(root, eb))
 265                 BUG();
 266         if (!btrfs_buffer_uptodate(eb, trans->transid))
 267                 BUG();
 268
 269         btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
 270         csum_tree_block(root, eb, 0);
 271
 272         dev_nr = 0;
 273         length = eb->len;
 274         ret = btrfs_map_block(&root->fs_info->mapping_tree, WRITE,
 275                               eb->start, &length, &multi, 0);
 276
 277         while(dev_nr < multi->num_stripes) {
 278                 BUG_ON(ret);
 279                 eb->fd = multi->stripes[dev_nr].dev->fd;
 280                 eb->dev_bytenr = multi->stripes[dev_nr].physical;
 281                 multi->stripes[dev_nr].dev->total_ios++;
 282                 dev_nr++;
 283                 ret = write_extent_to_disk(eb);
 284                 BUG_ON(ret);
 285         }
 286         kfree(multi);
 287         return 0;
 288 }
 289
 290 static int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
 291                         u32 stripesize, struct btrfs_root *root,
 292                         struct btrfs_fs_info *fs_info, u64 objectid)
 293 {
 294         root->node = NULL;
 295         root->commit_root = NULL;
 296         root->sectorsize = sectorsize;
 297         root->nodesize = nodesize;
 298         root->leafsize = leafsize;
 299         root->stripesize = stripesize;
 300         root->ref_cows = 0;
 301         root->track_dirty = 0;
 302
 303         root->fs_info = fs_info;
 304         root->objectid = objectid;
 305         root->last_trans = 0;
 306         root->highest_inode = 0;
 307         root->last_inode_alloc = 0;
 308
 309         INIT_LIST_HEAD(&root->dirty_list);
 310         memset(&root->root_key, 0, sizeof(root->root_key));
 311         memset(&root->root_item, 0, sizeof(root->root_item));
 312         root->root_key.objectid = objectid;
 313         return 0;
 314 }
 315
 316 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 317                                struct btrfs_root *root)
 318 {
 319         int ret;
 320         u64 old_root_bytenr;
 321         struct btrfs_root *tree_root = root->fs_info->tree_root;
 322
 323         btrfs_write_dirty_block_groups(trans, root);
 324         while(1) {
 325                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 326                 if (old_root_bytenr == root->node->start)
 327                         break;
 328                 btrfs_set_root_bytenr(&root->root_item,
 329                                        root->node->start);
 330                 btrfs_set_root_generation(&root->root_item,
 331                                           trans->transid);
 332                 root->root_item.level = btrfs_header_level(root->node);
 333                 ret = btrfs_update_root(trans, tree_root,
 334                                         &root->root_key,
 335                                         &root->root_item);
 336                 BUG_ON(ret);
 337                 btrfs_write_dirty_block_groups(trans, root);
 338         }
 339         return 0;
 340 }
 341
 342 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 343                              struct btrfs_fs_info *fs_info)
 344 {
 345         struct btrfs_root *root;
 346         struct list_head *next;
 347         struct extent_buffer *eb;
 348
 349         if (fs_info->readonly)
 350                 return 0;
 351
 352         eb = fs_info->tree_root->node;
 353         extent_buffer_get(eb);
 354         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 355         free_extent_buffer(eb);
 356
 357         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 358                 next = fs_info->dirty_cowonly_roots.next;
 359                 list_del_init(next);
 360                 root = list_entry(next, struct btrfs_root, dirty_list);
 361                 update_cowonly_root(trans, root);
 362         }
 363         return 0;
 364 }
 365
 366 static int __commit_transaction(struct btrfs_trans_handle *trans,
 367                                 struct btrfs_root *root)
 368 {
 369         u64 start;
 370         u64 end;
 371         struct extent_buffer *eb;
 372         struct extent_io_tree *tree = &root->fs_info->extent_cache;
 373         int ret;
 374
 375         while(1) {
 376                 ret = find_first_extent_bit(tree, 0, &start, &end,
 377                                             EXTENT_DIRTY);
 378                 if (ret)
 379                         break;
 380                 while(start <= end) {
 381                         eb = find_first_extent_buffer(tree, start);
 382                         BUG_ON(!eb || eb->start != start);
 383                         ret = write_tree_block(trans, root, eb);
 384                         BUG_ON(ret);
 385                         start += eb->len;
 386                         clear_extent_buffer_dirty(eb);
 387                         free_extent_buffer(eb);
 388                 }
 389         }
 390         return 0;
 391 }
 392
 393 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 394                              struct btrfs_root *root)
 395 {
 396         u64 transid = trans->transid;
 397         int ret = 0;
 398         struct btrfs_fs_info *fs_info = root->fs_info;
 399
 400         if (root->commit_root == root->node)
 401                 goto commit_tree;
 402
 403         free_extent_buffer(root->commit_root);
 404         root->commit_root = NULL;
 405
 406         btrfs_set_root_bytenr(&root->root_item, root->node->start);
 407         btrfs_set_root_generation(&root->root_item, trans->transid);
 408         root->root_item.level = btrfs_header_level(root->node);
 409         ret = btrfs_update_root(trans, root->fs_info->tree_root,
 410                                 &root->root_key, &root->root_item);
 411         BUG_ON(ret);
 412 commit_tree:
 413         ret = commit_tree_roots(trans, fs_info);
 414         BUG_ON(ret);
 415         ret = __commit_transaction(trans, root);
 416         BUG_ON(ret);
 417         write_ctree_super(trans, root);
 418         btrfs_finish_extent_commit(trans, fs_info->extent_root,
 419                                    &fs_info->pinned_extents);
 420         btrfs_free_transaction(root, trans);
 421         free_extent_buffer(root->commit_root);
 422         root->commit_root = NULL;
 423         fs_info->running_transaction = NULL;
 424         fs_info->last_trans_committed = transid;
 425         return 0;
 426 }
 427
 428 static int find_and_setup_root(struct btrfs_root *tree_root,
 429                                struct btrfs_fs_info *fs_info,
 430                                u64 objectid, struct btrfs_root *root)
 431 {
 432         int ret;
 433         u32 blocksize;
 434         u64 generation;
 435
 436         __setup_root(tree_root->nodesize, tree_root->leafsize,
 437                      tree_root->sectorsize, tree_root->stripesize,
 438                      root, fs_info, objectid);
 439         ret = btrfs_find_last_root(tree_root, objectid,
 440                                    &root->root_item, &root->root_key);
 441         BUG_ON(ret);
 442
 443         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 444         generation = btrfs_root_generation(&root->root_item);
 445         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 446                                      blocksize, generation);
 447         if (!extent_buffer_uptodate(root->node))
 448                 return -EIO;
 449
 450         return 0;
 451 }
 452
 453 static int find_and_setup_log_root(struct btrfs_root *tree_root,
 454                                struct btrfs_fs_info *fs_info,
 455                                struct btrfs_super_block *disk_super)
 456 {
 457         u32 blocksize;
 458         u64 blocknr = btrfs_super_log_root(disk_super);
 459         struct btrfs_root *log_root = malloc(sizeof(struct btrfs_root));
 460
 461         if (!log_root)
 462                 return -ENOMEM;
 463
 464         if (blocknr == 0) {
 465                 free(log_root);
 466                 return 0;
 467         }
 468
 469         blocksize = btrfs_level_size(tree_root,
 470                              btrfs_super_log_root_level(disk_super));
 471
 472         __setup_root(tree_root->nodesize, tree_root->leafsize,
 473                      tree_root->sectorsize, tree_root->stripesize,
 474                      log_root, fs_info, BTRFS_TREE_LOG_OBJECTID);
 475
 476         log_root->node = read_tree_block(tree_root, blocknr,
 477                                      blocksize,
 478                                      btrfs_super_generation(disk_super) + 1);
 479
 480         fs_info->log_root_tree = log_root;
 481
 482         if (!extent_buffer_uptodate(log_root->node)) {
 483                 free(log_root);
 484                 return -EIO;
 485         }
 486
 487         free(log_root);
 488         return 0;
 489 }
 490
 491
 492 int btrfs_free_fs_root(struct btrfs_fs_info *fs_info,
 493                        struct btrfs_root *root)
 494 {
 495         if (root->node)
 496                 free_extent_buffer(root->node);
 497         if (root->commit_root)
 498                 free_extent_buffer(root->commit_root);
 499         kfree(root);
 500         return 0;
 501 }
 502
 503 static int free_fs_roots(struct btrfs_fs_info *fs_info)
 504 {
 505         struct cache_extent *cache;
 506         struct btrfs_root *root;
 507
 508         while (1) {
 509                 cache = find_first_cache_extent(&fs_info->fs_root_cache, 0);
 510                 if (!cache)
 511                         break;
 512                 root = container_of(cache, struct btrfs_root, cache);
 513                 remove_cache_extent(&fs_info->fs_root_cache, cache);
 514                 btrfs_free_fs_root(fs_info, root);
 515         }
 516         return 0;
 517 }
 518
 519 struct btrfs_root *btrfs_read_fs_root_no_cache(struct btrfs_fs_info *fs_info,
 520                                                struct btrfs_key *location)
 521 {
 522         struct btrfs_root *root;
 523         struct btrfs_root *tree_root = fs_info->tree_root;
 524         struct btrfs_path *path;
 525         struct extent_buffer *l;
 526         u64 generation;
 527         u32 blocksize;
 528         int ret = 0;
 529
 530         root = malloc(sizeof(*root));
 531         if (!root)
 532                 return ERR_PTR(-ENOMEM);
 533         memset(root, 0, sizeof(*root));
 534         if (location->offset == (u64)-1) {
 535                 ret = find_and_setup_root(tree_root, fs_info,
 536                                           location->objectid, root);
 537                 if (ret) {
 538                         free(root);
 539                         return ERR_PTR(ret);
 540                 }
 541                 goto insert;
 542         }
 543
 544         __setup_root(tree_root->nodesize, tree_root->leafsize,
 545                      tree_root->sectorsize, tree_root->stripesize,
 546                      root, fs_info, location->objectid);
 547
 548         path = btrfs_alloc_path();
 549         BUG_ON(!path);
 550         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 551         if (ret != 0) {
 552                 if (ret > 0)
 553                         ret = -ENOENT;
 554                 goto out;
 555         }
 556         l = path->nodes[0];
 557         read_extent_buffer(l, &root->root_item,
 558                btrfs_item_ptr_offset(l, path->slots[0]),
 559                sizeof(root->root_item));
 560         memcpy(&root->root_key, location, sizeof(*location));
 561         ret = 0;
 562 out:
 563         btrfs_release_path(root, path);
 564         btrfs_free_path(path);
 565         if (ret) {
 566                 free(root);
 567                 return ERR_PTR(ret);
 568         }
 569         generation = btrfs_root_generation(&root->root_item);
 570         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 571         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 572                                      blocksize, generation);
 573         BUG_ON(!root->node);
 574 insert:
 575         root->ref_cows = 1;
 576         return root;
 577 }
 578
 579 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 580                                       struct btrfs_key *location)
 581 {
 582         struct btrfs_root *root;
 583         struct cache_extent *cache;
 584         int ret;
 585
 586         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 587                 return fs_info->tree_root;
 588         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 589                 return fs_info->extent_root;
 590         if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID)
 591                 return fs_info->chunk_root;
 592         if (location->objectid == BTRFS_DEV_TREE_OBJECTID)
 593                 return fs_info->dev_root;
 594         if (location->objectid == BTRFS_CSUM_TREE_OBJECTID)
 595                 return fs_info->csum_root;
 596
 597         BUG_ON(location->objectid == BTRFS_TREE_RELOC_OBJECTID ||
 598                location->offset != (u64)-1);
 599
 600         cache = find_cache_extent(&fs_info->fs_root_cache,
 601                                   location->objectid, 1);
 602         if (cache)
 603                 return container_of(cache, struct btrfs_root, cache);
 604
 605         root = btrfs_read_fs_root_no_cache(fs_info, location);
 606         if (IS_ERR(root))
 607                 return root;
 608
 609         root->cache.start = location->objectid;
 610         root->cache.size = 1;
 611         ret = insert_existing_cache_extent(&fs_info->fs_root_cache,
 612                                            &root->cache);
 613         BUG_ON(ret);
 614         return root;
 615 }
 616
 617 static struct btrfs_fs_info *__open_ctree_fd(int fp, const char *path,
 618                                              u64 sb_bytenr,
 619                                              u64 root_tree_bytenr, int writes,
 620                                              int partial)
 621 {
 622         u32 sectorsize;
 623         u32 nodesize;
 624         u32 leafsize;
 625         u32 blocksize;
 626         u32 stripesize;
 627         u64 generation;
 628         struct btrfs_key key;
 629         struct btrfs_root *tree_root = malloc(sizeof(struct btrfs_root));
 630         struct btrfs_root *extent_root = malloc(sizeof(struct btrfs_root));
 631         struct btrfs_root *chunk_root = malloc(sizeof(struct btrfs_root));
 632         struct btrfs_root *dev_root = malloc(sizeof(struct btrfs_root));
 633         struct btrfs_root *csum_root = malloc(sizeof(struct btrfs_root));
 634         struct btrfs_fs_info *fs_info = malloc(sizeof(*fs_info));
 635         int ret;
 636         struct btrfs_super_block *disk_super;
 637         struct btrfs_fs_devices *fs_devices = NULL;
 638         u64 total_devs;
 639         u64 features;
 640
 641         if (sb_bytenr == 0)
 642                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
 643
 644         ret = btrfs_scan_one_device(fp, path, &fs_devices,
 645                                     &total_devs, sb_bytenr);
 646
 647         if (ret) {
 648                 fprintf(stderr, "No valid Btrfs found on %s\n", path);
 649                 goto out;
 650         }
 651
 652         if (total_devs != 1) {
 653                 ret = btrfs_scan_for_fsid(fs_devices, total_devs, 1);
 654                 if (ret)
 655                         goto out;
 656         }
 657
 658         memset(fs_info, 0, sizeof(*fs_info));
 659         fs_info->tree_root = tree_root;
 660         fs_info->extent_root = extent_root;
 661         fs_info->chunk_root = chunk_root;
 662         fs_info->dev_root = dev_root;
 663         fs_info->csum_root = csum_root;
 664
 665         if (!writes)
 666                 fs_info->readonly = 1;
 667
 668         extent_io_tree_init(&fs_info->extent_cache);
 669         extent_io_tree_init(&fs_info->free_space_cache);
 670         extent_io_tree_init(&fs_info->block_group_cache);
 671         extent_io_tree_init(&fs_info->pinned_extents);
 672         extent_io_tree_init(&fs_info->pending_del);
 673         extent_io_tree_init(&fs_info->extent_ins);
 674         cache_tree_init(&fs_info->fs_root_cache);
 675
 676         cache_tree_init(&fs_info->mapping_tree.cache_tree);
 677
 678         mutex_init(&fs_info->fs_mutex);
 679         fs_info->fs_devices = fs_devices;
 680         INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
 681         INIT_LIST_HEAD(&fs_info->space_info);
 682
 683         __setup_root(4096, 4096, 4096, 4096, tree_root,
 684                      fs_info, BTRFS_ROOT_TREE_OBJECTID);
 685
 686         if (writes)
 687                 ret = btrfs_open_devices(fs_devices, O_RDWR);
 688         else
 689                 ret = btrfs_open_devices(fs_devices, O_RDONLY);
 690         if (ret)
 691                 goto out_cleanup;
 692
 693         fs_info->super_bytenr = sb_bytenr;
 694         disk_super = &fs_info->super_copy;
 695         ret = btrfs_read_dev_super(fs_devices->latest_bdev,
 696                                    disk_super, sb_bytenr);
 697         if (ret) {
 698                 printk("No valid btrfs found\n");
 699                 goto out_devices;
 700         }
 701
 702         memcpy(fs_info->fsid, &disk_super->fsid, BTRFS_FSID_SIZE);
 703
 704
 705         features = btrfs_super_incompat_flags(disk_super) &
 706                    ~BTRFS_FEATURE_INCOMPAT_SUPP;
 707         if (features) {
 708                 printk("couldn't open because of unsupported "
 709                        "option features (%Lx).\n",
 710                        (unsigned long long)features);
 711                 goto out_devices;
 712         }
 713
 714         features = btrfs_super_incompat_flags(disk_super);
 715         if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) {
 716                 features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF;
 717                 btrfs_set_super_incompat_flags(disk_super, features);
 718         }
 719
 720         features = btrfs_super_compat_ro_flags(disk_super) &
 721                 ~BTRFS_FEATURE_COMPAT_RO_SUPP;
 722         if (writes && features) {
 723                 printk("couldn't open RDWR because of unsupported "
 724                        "option features (%Lx).\n",
 725                        (unsigned long long)features);
 726                 goto out_devices;
 727         }
 728
 729         nodesize = btrfs_super_nodesize(disk_super);
 730         leafsize = btrfs_super_leafsize(disk_super);
 731         sectorsize = btrfs_super_sectorsize(disk_super);
 732         stripesize = btrfs_super_stripesize(disk_super);
 733         tree_root->nodesize = nodesize;
 734         tree_root->leafsize = leafsize;
 735         tree_root->sectorsize = sectorsize;
 736         tree_root->stripesize = stripesize;
 737
 738         ret = btrfs_read_sys_array(tree_root);
 739         if (ret)
 740                 goto out_devices;
 741         blocksize = btrfs_level_size(tree_root,
 742                                      btrfs_super_chunk_root_level(disk_super));
 743         generation = btrfs_super_chunk_root_generation(disk_super);
 744
 745         __setup_root(nodesize, leafsize, sectorsize, stripesize,
 746                      chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID);
 747
 748         chunk_root->node = read_tree_block(chunk_root,
 749                                            btrfs_super_chunk_root(disk_super),
 750                                            blocksize, generation);
 751         if (!extent_buffer_uptodate(chunk_root->node)) {
 752                 printk("Couldn't read chunk root\n");
 753                 goto out_devices;
 754         }
 755
 756         read_extent_buffer(chunk_root->node, fs_info->chunk_tree_uuid,
 757                  (unsigned long)btrfs_header_chunk_tree_uuid(chunk_root->node),
 758                  BTRFS_UUID_SIZE);
 759
 760         if (!(btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_METADUMP)) {
 761                 ret = btrfs_read_chunk_tree(chunk_root);
 762                 if (ret)
 763                         goto out_failed;
 764         }
 765
 766         blocksize = btrfs_level_size(tree_root,
 767                                      btrfs_super_root_level(disk_super));
 768         generation = btrfs_super_generation(disk_super);
 769
 770         if (!root_tree_bytenr)
 771                 root_tree_bytenr = btrfs_super_root(disk_super);
 772         tree_root->node = read_tree_block(tree_root,
 773                                           root_tree_bytenr,
 774                                           blocksize, generation);
 775         if (!extent_buffer_uptodate(tree_root->node)) {
 776                 printk("Couldn't read tree root\n");
 777                 goto out_failed;
 778         }
 779         ret = find_and_setup_root(tree_root, fs_info,
 780                                   BTRFS_EXTENT_TREE_OBJECTID, extent_root);
 781         if (ret) {
 782                 printk("Couldn't setup extent tree\n");
 783                 goto out_failed;
 784         }
 785         extent_root->track_dirty = 1;
 786
 787         ret = find_and_setup_root(tree_root, fs_info,
 788                                   BTRFS_DEV_TREE_OBJECTID, dev_root);
 789         if (ret) {
 790                 printk("Couldn't setup device tree\n");
 791                 goto out_failed;
 792         }
 793         dev_root->track_dirty = 1;
 794
 795         ret = find_and_setup_root(tree_root, fs_info,
 796                                   BTRFS_CSUM_TREE_OBJECTID, csum_root);
 797         if (ret) {
 798                 printk("Couldn't setup csum tree\n");
 799                 if (!partial)
 800                         goto out_failed;
 801         }
 802         csum_root->track_dirty = 1;
 803
 804         find_and_setup_log_root(tree_root, fs_info, disk_super);
 805
 806         fs_info->generation = generation;
 807         fs_info->last_trans_committed = generation;
 808         btrfs_read_block_groups(fs_info->tree_root);
 809
 810         key.objectid = BTRFS_FS_TREE_OBJECTID;
 811         key.type = BTRFS_ROOT_ITEM_KEY;
 812         key.offset = (u64)-1;
 813         fs_info->fs_root = btrfs_read_fs_root(fs_info, &key);
 814
 815         if (!fs_info->fs_root)
 816                 goto out_failed;
 817
 818         fs_info->data_alloc_profile = (u64)-1;
 819         fs_info->metadata_alloc_profile = (u64)-1;
 820         fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
 821
 822         return fs_info;
 823
 824 out_failed:
 825         if (partial)
 826                 return fs_info;
 827
 828         if (fs_info->csum_root)
 829                 free_extent_buffer(fs_info->csum_root->node);
 830         if (fs_info->dev_root)
 831                 free_extent_buffer(fs_info->dev_root->node);
 832         if (fs_info->extent_root)
 833                 free_extent_buffer(fs_info->extent_root->node);
 834         if (fs_info->tree_root)
 835                 free_extent_buffer(fs_info->tree_root->node);
 836         if (fs_info->chunk_root)
 837                 free_extent_buffer(fs_info->chunk_root->node);
 838 out_devices:
 839         close_all_devices(fs_info);
 840 out_cleanup:
 841         extent_io_tree_cleanup(&fs_info->extent_cache);
 842         extent_io_tree_cleanup(&fs_info->free_space_cache);
 843         extent_io_tree_cleanup(&fs_info->block_group_cache);
 844         extent_io_tree_cleanup(&fs_info->pinned_extents);
 845         extent_io_tree_cleanup(&fs_info->pending_del);
 846         extent_io_tree_cleanup(&fs_info->extent_ins);
 847 out:
 848         free(tree_root);
 849         free(extent_root);
 850         free(chunk_root);
 851         free(dev_root);
 852         free(csum_root);
 853         free(fs_info);
 854         return NULL;
 855 }
 856
 857 struct btrfs_fs_info *open_ctree_fs_info(const char *filename,
 858                                          u64 sb_bytenr, int writes,
 859                                          int partial)
 860 {
 861         int fp;
 862         struct btrfs_fs_info *info;
 863         int flags = O_CREAT | O_RDWR;
 864
 865         if (!writes)
 866                 flags = O_RDONLY;
 867
 868         fp = open(filename, flags, 0600);
 869         if (fp < 0) {
 870                 fprintf (stderr, "Could not open %s\n", filename);
 871                 return NULL;
 872         }
 873         info = __open_ctree_fd(fp, filename, sb_bytenr, 0, writes, partial);
 874         close(fp);
 875         return info;
 876 }
 877
 878 struct btrfs_root *open_ctree(const char *filename, u64 sb_bytenr, int writes)
 879 {
 880         struct btrfs_fs_info *info;
 881
 882         info = open_ctree_fs_info(filename, sb_bytenr, writes, 0);
 883         if (!info)
 884                 return NULL;
 885         return info->fs_root;
 886 }
 887
 888 struct btrfs_root *open_ctree_recovery(const char *filename, u64 sb_bytenr,
 889                                        u64 root_tree_bytenr)
 890 {
 891         int fp;
 892         struct btrfs_fs_info *info;
 893
 894
 895         fp = open(filename, O_RDONLY);
 896         if (fp < 0) {
 897                 fprintf (stderr, "Could not open %s\n", filename);
 898                 return NULL;
 899         }
 900         info = __open_ctree_fd(fp, filename, sb_bytenr,
 901                                root_tree_bytenr, 0, 0);
 902         close(fp);
 903
 904         if (!info)
 905                 return NULL;
 906         return info->fs_root;
 907 }
 908
 909 struct btrfs_root *open_ctree_fd(int fp, const char *path, u64 sb_bytenr,
 910                                  int writes)
 911 {
 912         struct btrfs_fs_info *info;
 913         info = __open_ctree_fd(fp, path, sb_bytenr, 0, writes, 0);
 914         if (!info)
 915                 return NULL;
 916         return info->fs_root;
 917 }
 918
 919 int btrfs_read_dev_super(int fd, struct btrfs_super_block *sb, u64 sb_bytenr)
 920 {
 921         u8 fsid[BTRFS_FSID_SIZE];
 922         int fsid_is_initialized = 0;
 923         struct btrfs_super_block buf;
 924         int i;
 925         int ret;
 926         u64 transid = 0;
 927         u64 bytenr;
 928
 929         if (sb_bytenr != BTRFS_SUPER_INFO_OFFSET) {
 930                 ret = pread64(fd, &buf, sizeof(buf), sb_bytenr);
 931                 if (ret < sizeof(buf))
 932                         return -1;
 933
 934                 if (btrfs_super_bytenr(&buf) != sb_bytenr ||
 935                     strncmp((char *)(&buf.magic), BTRFS_MAGIC,
 936                             sizeof(buf.magic)))
 937                         return -1;
 938
 939                 memcpy(sb, &buf, sizeof(*sb));
 940                 return 0;
 941         }
 942
 943         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 944                 bytenr = btrfs_sb_offset(i);
 945                 ret = pread64(fd, &buf, sizeof(buf), bytenr);
 946                 if (ret < sizeof(buf))
 947                         break;
 948
 949                 if (btrfs_super_bytenr(&buf) != bytenr )
 950                         continue;
 951                 /* if magic is NULL, the device was removed */
 952                 if (buf.magic == 0 && i == 0)
 953                         return -1;
 954                 if (strncmp((char *)(&buf.magic), BTRFS_MAGIC,
 955                             sizeof(buf.magic)))
 956                         continue;
 957
 958                 if (!fsid_is_initialized) {
 959                         memcpy(fsid, buf.fsid, sizeof(fsid));
 960                         fsid_is_initialized = 1;
 961                 } else if (memcmp(fsid, buf.fsid, sizeof(fsid))) {
 962                         /*
 963                          * the superblocks (the original one and
 964                          * its backups) contain data of different
 965                          * filesystems -> the super cannot be trusted
 966                          */
 967                         continue;
 968                 }
 969
 970                 if (btrfs_super_generation(&buf) > transid) {
 971                         memcpy(sb, &buf, sizeof(*sb));
 972                         transid = btrfs_super_generation(&buf);
 973                 }
 974         }
 975
 976         return transid > 0 ? 0 : -1;
 977 }
 978
 979 int write_dev_supers(struct btrfs_root *root, struct btrfs_super_block *sb,
 980                      struct btrfs_device *device)
 981 {
 982         u64 bytenr;
 983         u32 crc;
 984         int i, ret;
 985
 986         if (root->fs_info->super_bytenr != BTRFS_SUPER_INFO_OFFSET) {
 987                 btrfs_set_super_bytenr(sb, root->fs_info->super_bytenr);
 988                 crc = ~(u32)0;
 989                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
 990                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
 991                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
 992
 993                 ret = pwrite64(device->fd, sb, BTRFS_SUPER_INFO_SIZE,
 994                                root->fs_info->super_bytenr);
 995                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
 996                 return 0;
 997         }
 998
 999         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1000                 bytenr = btrfs_sb_offset(i);
1001                 if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes)
1002                         break;
1003
1004                 btrfs_set_super_bytenr(sb, bytenr);
1005
1006                 crc = ~(u32)0;
1007                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1008                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1009                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1010
1011                 ret = pwrite64(device->fd, sb, BTRFS_SUPER_INFO_SIZE, bytenr);
1012                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1013         }
1014         return 0;
1015 }
1016
1017 int write_all_supers(struct btrfs_root *root)
1018 {
1019         struct list_head *cur;
1020         struct list_head *head = &root->fs_info->fs_devices->devices;
1021         struct btrfs_device *dev;
1022         struct btrfs_super_block *sb;
1023         struct btrfs_dev_item *dev_item;
1024         int ret;
1025         u64 flags;
1026
1027         sb = &root->fs_info->super_copy;
1028         dev_item = &sb->dev_item;
1029         list_for_each(cur, head) {
1030                 dev = list_entry(cur, struct btrfs_device, dev_list);
1031                 if (!dev->writeable)
1032                         continue;
1033
1034                 btrfs_set_stack_device_generation(dev_item, 0);
1035                 btrfs_set_stack_device_type(dev_item, dev->type);
1036                 btrfs_set_stack_device_id(dev_item, dev->devid);
1037                 btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes);
1038                 btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used);
1039                 btrfs_set_stack_device_io_align(dev_item, dev->io_align);
1040                 btrfs_set_stack_device_io_width(dev_item, dev->io_width);
1041                 btrfs_set_stack_device_sector_size(dev_item, dev->sector_size);
1042                 memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE);
1043                 memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE);
1044
1045                 flags = btrfs_super_flags(sb);
1046                 btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN);
1047
1048                 ret = write_dev_supers(root, sb, dev);
1049                 BUG_ON(ret);
1050         }
1051         return 0;
1052 }
1053
1054 int write_ctree_super(struct btrfs_trans_handle *trans,
1055                       struct btrfs_root *root)
1056 {
1057         int ret;
1058         struct btrfs_root *tree_root = root->fs_info->tree_root;
1059         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
1060
1061         if (root->fs_info->readonly)
1062                 return 0;
1063
1064         btrfs_set_super_generation(&root->fs_info->super_copy,
1065                                    trans->transid);
1066         btrfs_set_super_root(&root->fs_info->super_copy,
1067                              tree_root->node->start);
1068         btrfs_set_super_root_level(&root->fs_info->super_copy,
1069                                    btrfs_header_level(tree_root->node));
1070         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
1071                                    chunk_root->node->start);
1072         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
1073                                          btrfs_header_level(chunk_root->node));
1074         btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
1075                                 btrfs_header_generation(chunk_root->node));
1076
1077         ret = write_all_supers(root);
1078         if (ret)
1079                 fprintf(stderr, "failed to write new super block err %d\n", ret);
1080         return ret;
1081 }
1082
1083 static int close_all_devices(struct btrfs_fs_info *fs_info)
1084 {
1085         struct list_head *list;
1086         struct list_head *next;
1087         struct btrfs_device *device;
1088
1089         return 0;
1090
1091         list = &fs_info->fs_devices->devices;
1092         list_for_each(next, list) {
1093                 device = list_entry(next, struct btrfs_device, dev_list);
1094                 close(device->fd);
1095         }
1096         return 0;
1097 }
1098
1099 int close_ctree(struct btrfs_root *root)
1100 {
1101         int ret;
1102         struct btrfs_trans_handle *trans;
1103         struct btrfs_fs_info *fs_info = root->fs_info;
1104
1105         if (fs_info->last_trans_committed !=
1106             fs_info->generation) {
1107                 trans = btrfs_start_transaction(root, 1);
1108                 btrfs_commit_transaction(trans, root);
1109                 trans = btrfs_start_transaction(root, 1);
1110                 ret = commit_tree_roots(trans, fs_info);
1111                 BUG_ON(ret);
1112                 ret = __commit_transaction(trans, root);
1113                 BUG_ON(ret);
1114                 write_ctree_super(trans, root);
1115                 btrfs_free_transaction(root, trans);
1116         }
1117         btrfs_free_block_groups(fs_info);
1118
1119         free_fs_roots(fs_info);
1120
1121         if (fs_info->extent_root->node)
1122                 free_extent_buffer(fs_info->extent_root->node);
1123         if (fs_info->tree_root->node)
1124                 free_extent_buffer(fs_info->tree_root->node);
1125         if (fs_info->chunk_root->node)
1126                 free_extent_buffer(fs_info->chunk_root->node);
1127         if (fs_info->dev_root->node)
1128                 free_extent_buffer(fs_info->dev_root->node);
1129         if (fs_info->csum_root->node)
1130                 free_extent_buffer(fs_info->csum_root->node);
1131
1132         if (fs_info->log_root_tree) {
1133                 if (fs_info->log_root_tree->node)
1134                         free_extent_buffer(fs_info->log_root_tree->node);
1135                 free(fs_info->log_root_tree);
1136         }
1137
1138         close_all_devices(fs_info);
1139         extent_io_tree_cleanup(&fs_info->extent_cache);
1140         extent_io_tree_cleanup(&fs_info->free_space_cache);
1141         extent_io_tree_cleanup(&fs_info->block_group_cache);
1142         extent_io_tree_cleanup(&fs_info->pinned_extents);
1143         extent_io_tree_cleanup(&fs_info->pending_del);
1144         extent_io_tree_cleanup(&fs_info->extent_ins);
1145
1146         free(fs_info->tree_root);
1147         free(fs_info->extent_root);
1148         free(fs_info->chunk_root);
1149         free(fs_info->dev_root);
1150         free(fs_info->csum_root);
1151         free(fs_info);
1152
1153         return 0;
1154 }
1155
1156 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1157                      struct extent_buffer *eb)
1158 {
1159         return clear_extent_buffer_dirty(eb);
1160 }
1161
1162 int wait_on_tree_block_writeback(struct btrfs_root *root,
1163                                  struct extent_buffer *eb)
1164 {
1165         return 0;
1166 }
1167
1168 void btrfs_mark_buffer_dirty(struct extent_buffer *eb)
1169 {
1170         set_extent_buffer_dirty(eb);
1171 }
1172
1173 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid)
1174 {
1175         int ret;
1176
1177         ret = extent_buffer_uptodate(buf);
1178         if (!ret)
1179                 return ret;
1180
1181         ret = verify_parent_transid(buf->tree, buf, parent_transid, 1);
1182         return !ret;
1183 }
1184
1185 int btrfs_set_buffer_uptodate(struct extent_buffer *eb)
1186 {
1187         return set_extent_buffer_uptodate(eb);
1188 }