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         int ret;
 349
 350         if (fs_info->readonly)
 351                 return 0;
 352
 353         eb = fs_info->tree_root->node;
 354         extent_buffer_get(eb);
 355         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 356         free_extent_buffer(eb);
 357         if (ret)
 358                 return ret;
 359
 360         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 361                 next = fs_info->dirty_cowonly_roots.next;
 362                 list_del_init(next);
 363                 root = list_entry(next, struct btrfs_root, dirty_list);
 364                 update_cowonly_root(trans, root);
 365         }
 366         return 0;
 367 }
 368
 369 static int __commit_transaction(struct btrfs_trans_handle *trans,
 370                                 struct btrfs_root *root)
 371 {
 372         u64 start;
 373         u64 end;
 374         struct extent_buffer *eb;
 375         struct extent_io_tree *tree = &root->fs_info->extent_cache;
 376         int ret;
 377
 378         while(1) {
 379                 ret = find_first_extent_bit(tree, 0, &start, &end,
 380                                             EXTENT_DIRTY);
 381                 if (ret)
 382                         break;
 383                 while(start <= end) {
 384                         eb = find_first_extent_buffer(tree, start);
 385                         BUG_ON(!eb || eb->start != start);
 386                         ret = write_tree_block(trans, root, eb);
 387                         BUG_ON(ret);
 388                         start += eb->len;
 389                         clear_extent_buffer_dirty(eb);
 390                         free_extent_buffer(eb);
 391                 }
 392         }
 393         return 0;
 394 }
 395
 396 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 397                              struct btrfs_root *root)
 398 {
 399         u64 transid = trans->transid;
 400         int ret = 0;
 401         struct btrfs_fs_info *fs_info = root->fs_info;
 402
 403         if (root->commit_root == root->node)
 404                 goto commit_tree;
 405
 406         free_extent_buffer(root->commit_root);
 407         root->commit_root = NULL;
 408
 409         btrfs_set_root_bytenr(&root->root_item, root->node->start);
 410         btrfs_set_root_generation(&root->root_item, trans->transid);
 411         root->root_item.level = btrfs_header_level(root->node);
 412         ret = btrfs_update_root(trans, root->fs_info->tree_root,
 413                                 &root->root_key, &root->root_item);
 414         BUG_ON(ret);
 415 commit_tree:
 416         ret = commit_tree_roots(trans, fs_info);
 417         BUG_ON(ret);
 418         ret = __commit_transaction(trans, root);
 419         BUG_ON(ret);
 420         write_ctree_super(trans, root);
 421         btrfs_finish_extent_commit(trans, fs_info->extent_root,
 422                                    &fs_info->pinned_extents);
 423         btrfs_free_transaction(root, trans);
 424         free_extent_buffer(root->commit_root);
 425         root->commit_root = NULL;
 426         fs_info->running_transaction = NULL;
 427         fs_info->last_trans_committed = transid;
 428         return 0;
 429 }
 430
 431 static int find_and_setup_root(struct btrfs_root *tree_root,
 432                                struct btrfs_fs_info *fs_info,
 433                                u64 objectid, struct btrfs_root *root)
 434 {
 435         int ret;
 436         u32 blocksize;
 437         u64 generation;
 438
 439         __setup_root(tree_root->nodesize, tree_root->leafsize,
 440                      tree_root->sectorsize, tree_root->stripesize,
 441                      root, fs_info, objectid);
 442         ret = btrfs_find_last_root(tree_root, objectid,
 443                                    &root->root_item, &root->root_key);
 444         BUG_ON(ret);
 445
 446         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 447         generation = btrfs_root_generation(&root->root_item);
 448         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 449                                      blocksize, generation);
 450         if (!extent_buffer_uptodate(root->node))
 451                 return -EIO;
 452
 453         return 0;
 454 }
 455
 456 static int find_and_setup_log_root(struct btrfs_root *tree_root,
 457                                struct btrfs_fs_info *fs_info,
 458                                struct btrfs_super_block *disk_super)
 459 {
 460         u32 blocksize;
 461         u64 blocknr = btrfs_super_log_root(disk_super);
 462         struct btrfs_root *log_root = malloc(sizeof(struct btrfs_root));
 463
 464         if (!log_root)
 465                 return -ENOMEM;
 466
 467         if (blocknr == 0) {
 468                 free(log_root);
 469                 return 0;
 470         }
 471
 472         blocksize = btrfs_level_size(tree_root,
 473                              btrfs_super_log_root_level(disk_super));
 474
 475         __setup_root(tree_root->nodesize, tree_root->leafsize,
 476                      tree_root->sectorsize, tree_root->stripesize,
 477                      log_root, fs_info, BTRFS_TREE_LOG_OBJECTID);
 478
 479         log_root->node = read_tree_block(tree_root, blocknr,
 480                                      blocksize,
 481                                      btrfs_super_generation(disk_super) + 1);
 482
 483         fs_info->log_root_tree = log_root;
 484
 485         if (!extent_buffer_uptodate(log_root->node)) {
 486                 free(log_root);
 487                 return -EIO;
 488         }
 489
 490         free(log_root);
 491         return 0;
 492 }
 493
 494
 495 int btrfs_free_fs_root(struct btrfs_fs_info *fs_info,
 496                        struct btrfs_root *root)
 497 {
 498         if (root->node)
 499                 free_extent_buffer(root->node);
 500         if (root->commit_root)
 501                 free_extent_buffer(root->commit_root);
 502         kfree(root);
 503         return 0;
 504 }
 505
 506 static int free_fs_roots(struct btrfs_fs_info *fs_info)
 507 {
 508         struct cache_extent *cache;
 509         struct btrfs_root *root;
 510
 511         while (1) {
 512                 cache = find_first_cache_extent(&fs_info->fs_root_cache, 0);
 513                 if (!cache)
 514                         break;
 515                 root = container_of(cache, struct btrfs_root, cache);
 516                 remove_cache_extent(&fs_info->fs_root_cache, cache);
 517                 btrfs_free_fs_root(fs_info, root);
 518         }
 519         return 0;
 520 }
 521
 522 struct btrfs_root *btrfs_read_fs_root_no_cache(struct btrfs_fs_info *fs_info,
 523                                                struct btrfs_key *location)
 524 {
 525         struct btrfs_root *root;
 526         struct btrfs_root *tree_root = fs_info->tree_root;
 527         struct btrfs_path *path;
 528         struct extent_buffer *l;
 529         u64 generation;
 530         u32 blocksize;
 531         int ret = 0;
 532
 533         root = malloc(sizeof(*root));
 534         if (!root)
 535                 return ERR_PTR(-ENOMEM);
 536         memset(root, 0, sizeof(*root));
 537         if (location->offset == (u64)-1) {
 538                 ret = find_and_setup_root(tree_root, fs_info,
 539                                           location->objectid, root);
 540                 if (ret) {
 541                         free(root);
 542                         return ERR_PTR(ret);
 543                 }
 544                 goto insert;
 545         }
 546
 547         __setup_root(tree_root->nodesize, tree_root->leafsize,
 548                      tree_root->sectorsize, tree_root->stripesize,
 549                      root, fs_info, location->objectid);
 550
 551         path = btrfs_alloc_path();
 552         BUG_ON(!path);
 553         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 554         if (ret != 0) {
 555                 if (ret > 0)
 556                         ret = -ENOENT;
 557                 goto out;
 558         }
 559         l = path->nodes[0];
 560         read_extent_buffer(l, &root->root_item,
 561                btrfs_item_ptr_offset(l, path->slots[0]),
 562                sizeof(root->root_item));
 563         memcpy(&root->root_key, location, sizeof(*location));
 564         ret = 0;
 565 out:
 566         btrfs_release_path(root, path);
 567         btrfs_free_path(path);
 568         if (ret) {
 569                 free(root);
 570                 return ERR_PTR(ret);
 571         }
 572         generation = btrfs_root_generation(&root->root_item);
 573         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 574         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 575                                      blocksize, generation);
 576         BUG_ON(!root->node);
 577 insert:
 578         root->ref_cows = 1;
 579         return root;
 580 }
 581
 582 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 583                                       struct btrfs_key *location)
 584 {
 585         struct btrfs_root *root;
 586         struct cache_extent *cache;
 587         int ret;
 588
 589         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 590                 return fs_info->tree_root;
 591         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 592                 return fs_info->extent_root;
 593         if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID)
 594                 return fs_info->chunk_root;
 595         if (location->objectid == BTRFS_DEV_TREE_OBJECTID)
 596                 return fs_info->dev_root;
 597         if (location->objectid == BTRFS_CSUM_TREE_OBJECTID)
 598                 return fs_info->csum_root;
 599
 600         BUG_ON(location->objectid == BTRFS_TREE_RELOC_OBJECTID ||
 601                location->offset != (u64)-1);
 602
 603         cache = find_cache_extent(&fs_info->fs_root_cache,
 604                                   location->objectid, 1);
 605         if (cache)
 606                 return container_of(cache, struct btrfs_root, cache);
 607
 608         root = btrfs_read_fs_root_no_cache(fs_info, location);
 609         if (IS_ERR(root))
 610                 return root;
 611
 612         root->cache.start = location->objectid;
 613         root->cache.size = 1;
 614         ret = insert_existing_cache_extent(&fs_info->fs_root_cache,
 615                                            &root->cache);
 616         BUG_ON(ret);
 617         return root;
 618 }
 619
 620 static struct btrfs_fs_info *__open_ctree_fd(int fp, const char *path,
 621                                              u64 sb_bytenr,
 622                                              u64 root_tree_bytenr, int writes,
 623                                              int partial)
 624 {
 625         u32 sectorsize;
 626         u32 nodesize;
 627         u32 leafsize;
 628         u32 blocksize;
 629         u32 stripesize;
 630         u64 generation;
 631         struct btrfs_key key;
 632         struct btrfs_root *tree_root = malloc(sizeof(struct btrfs_root));
 633         struct btrfs_root *extent_root = malloc(sizeof(struct btrfs_root));
 634         struct btrfs_root *chunk_root = malloc(sizeof(struct btrfs_root));
 635         struct btrfs_root *dev_root = malloc(sizeof(struct btrfs_root));
 636         struct btrfs_root *csum_root = malloc(sizeof(struct btrfs_root));
 637         struct btrfs_fs_info *fs_info = malloc(sizeof(*fs_info));
 638         int ret;
 639         struct btrfs_super_block *disk_super;
 640         struct btrfs_fs_devices *fs_devices = NULL;
 641         u64 total_devs;
 642         u64 features;
 643
 644         if (sb_bytenr == 0)
 645                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
 646
 647         ret = btrfs_scan_one_device(fp, path, &fs_devices,
 648                                     &total_devs, sb_bytenr);
 649
 650         if (ret) {
 651                 fprintf(stderr, "No valid Btrfs found on %s\n", path);
 652                 goto out;
 653         }
 654
 655         if (total_devs != 1) {
 656                 ret = btrfs_scan_for_fsid(fs_devices, total_devs, 1);
 657                 if (ret)
 658                         goto out;
 659         }
 660
 661         memset(fs_info, 0, sizeof(*fs_info));
 662         fs_info->tree_root = tree_root;
 663         fs_info->extent_root = extent_root;
 664         fs_info->chunk_root = chunk_root;
 665         fs_info->dev_root = dev_root;
 666         fs_info->csum_root = csum_root;
 667
 668         if (!writes)
 669                 fs_info->readonly = 1;
 670
 671         extent_io_tree_init(&fs_info->extent_cache);
 672         extent_io_tree_init(&fs_info->free_space_cache);
 673         extent_io_tree_init(&fs_info->block_group_cache);
 674         extent_io_tree_init(&fs_info->pinned_extents);
 675         extent_io_tree_init(&fs_info->pending_del);
 676         extent_io_tree_init(&fs_info->extent_ins);
 677         cache_tree_init(&fs_info->fs_root_cache);
 678
 679         cache_tree_init(&fs_info->mapping_tree.cache_tree);
 680
 681         mutex_init(&fs_info->fs_mutex);
 682         fs_info->fs_devices = fs_devices;
 683         INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
 684         INIT_LIST_HEAD(&fs_info->space_info);
 685
 686         __setup_root(4096, 4096, 4096, 4096, tree_root,
 687                      fs_info, BTRFS_ROOT_TREE_OBJECTID);
 688
 689         if (writes)
 690                 ret = btrfs_open_devices(fs_devices, O_RDWR);
 691         else
 692                 ret = btrfs_open_devices(fs_devices, O_RDONLY);
 693         if (ret)
 694                 goto out_cleanup;
 695
 696         fs_info->super_bytenr = sb_bytenr;
 697         disk_super = &fs_info->super_copy;
 698         ret = btrfs_read_dev_super(fs_devices->latest_bdev,
 699                                    disk_super, sb_bytenr);
 700         if (ret) {
 701                 printk("No valid btrfs found\n");
 702                 goto out_devices;
 703         }
 704
 705         memcpy(fs_info->fsid, &disk_super->fsid, BTRFS_FSID_SIZE);
 706
 707
 708         features = btrfs_super_incompat_flags(disk_super) &
 709                    ~BTRFS_FEATURE_INCOMPAT_SUPP;
 710         if (features) {
 711                 printk("couldn't open because of unsupported "
 712                        "option features (%Lx).\n",
 713                        (unsigned long long)features);
 714                 goto out_devices;
 715         }
 716
 717         features = btrfs_super_incompat_flags(disk_super);
 718         if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) {
 719                 features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF;
 720                 btrfs_set_super_incompat_flags(disk_super, features);
 721         }
 722
 723         features = btrfs_super_compat_ro_flags(disk_super) &
 724                 ~BTRFS_FEATURE_COMPAT_RO_SUPP;
 725         if (writes && features) {
 726                 printk("couldn't open RDWR because of unsupported "
 727                        "option features (%Lx).\n",
 728                        (unsigned long long)features);
 729                 goto out_devices;
 730         }
 731
 732         nodesize = btrfs_super_nodesize(disk_super);
 733         leafsize = btrfs_super_leafsize(disk_super);
 734         sectorsize = btrfs_super_sectorsize(disk_super);
 735         stripesize = btrfs_super_stripesize(disk_super);
 736         tree_root->nodesize = nodesize;
 737         tree_root->leafsize = leafsize;
 738         tree_root->sectorsize = sectorsize;
 739         tree_root->stripesize = stripesize;
 740
 741         ret = btrfs_read_sys_array(tree_root);
 742         if (ret)
 743                 goto out_devices;
 744         blocksize = btrfs_level_size(tree_root,
 745                                      btrfs_super_chunk_root_level(disk_super));
 746         generation = btrfs_super_chunk_root_generation(disk_super);
 747
 748         __setup_root(nodesize, leafsize, sectorsize, stripesize,
 749                      chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID);
 750
 751         chunk_root->node = read_tree_block(chunk_root,
 752                                            btrfs_super_chunk_root(disk_super),
 753                                            blocksize, generation);
 754         if (!extent_buffer_uptodate(chunk_root->node)) {
 755                 printk("Couldn't read chunk root\n");
 756                 goto out_devices;
 757         }
 758
 759         read_extent_buffer(chunk_root->node, fs_info->chunk_tree_uuid,
 760                  (unsigned long)btrfs_header_chunk_tree_uuid(chunk_root->node),
 761                  BTRFS_UUID_SIZE);
 762
 763         if (!(btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_METADUMP)) {
 764                 ret = btrfs_read_chunk_tree(chunk_root);
 765                 if (ret)
 766                         goto out_failed;
 767         }
 768
 769         blocksize = btrfs_level_size(tree_root,
 770                                      btrfs_super_root_level(disk_super));
 771         generation = btrfs_super_generation(disk_super);
 772
 773         if (!root_tree_bytenr)
 774                 root_tree_bytenr = btrfs_super_root(disk_super);
 775         tree_root->node = read_tree_block(tree_root,
 776                                           root_tree_bytenr,
 777                                           blocksize, generation);
 778         if (!extent_buffer_uptodate(tree_root->node)) {
 779                 printk("Couldn't read tree root\n");
 780                 goto out_failed;
 781         }
 782         ret = find_and_setup_root(tree_root, fs_info,
 783                                   BTRFS_EXTENT_TREE_OBJECTID, extent_root);
 784         if (ret) {
 785                 printk("Couldn't setup extent tree\n");
 786                 goto out_failed;
 787         }
 788         extent_root->track_dirty = 1;
 789
 790         ret = find_and_setup_root(tree_root, fs_info,
 791                                   BTRFS_DEV_TREE_OBJECTID, dev_root);
 792         if (ret) {
 793                 printk("Couldn't setup device tree\n");
 794                 goto out_failed;
 795         }
 796         dev_root->track_dirty = 1;
 797
 798         ret = find_and_setup_root(tree_root, fs_info,
 799                                   BTRFS_CSUM_TREE_OBJECTID, csum_root);
 800         if (ret) {
 801                 printk("Couldn't setup csum tree\n");
 802                 if (!partial)
 803                         goto out_failed;
 804         }
 805         csum_root->track_dirty = 1;
 806
 807         find_and_setup_log_root(tree_root, fs_info, disk_super);
 808
 809         fs_info->generation = generation;
 810         fs_info->last_trans_committed = generation;
 811         btrfs_read_block_groups(fs_info->tree_root);
 812
 813         key.objectid = BTRFS_FS_TREE_OBJECTID;
 814         key.type = BTRFS_ROOT_ITEM_KEY;
 815         key.offset = (u64)-1;
 816         fs_info->fs_root = btrfs_read_fs_root(fs_info, &key);
 817
 818         if (!fs_info->fs_root)
 819                 goto out_failed;
 820
 821         fs_info->data_alloc_profile = (u64)-1;
 822         fs_info->metadata_alloc_profile = (u64)-1;
 823         fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
 824
 825         return fs_info;
 826
 827 out_failed:
 828         if (partial)
 829                 return fs_info;
 830
 831         if (fs_info->csum_root)
 832                 free_extent_buffer(fs_info->csum_root->node);
 833         if (fs_info->dev_root)
 834                 free_extent_buffer(fs_info->dev_root->node);
 835         if (fs_info->extent_root)
 836                 free_extent_buffer(fs_info->extent_root->node);
 837         if (fs_info->tree_root)
 838                 free_extent_buffer(fs_info->tree_root->node);
 839         if (fs_info->chunk_root)
 840                 free_extent_buffer(fs_info->chunk_root->node);
 841 out_devices:
 842         close_all_devices(fs_info);
 843 out_cleanup:
 844         extent_io_tree_cleanup(&fs_info->extent_cache);
 845         extent_io_tree_cleanup(&fs_info->free_space_cache);
 846         extent_io_tree_cleanup(&fs_info->block_group_cache);
 847         extent_io_tree_cleanup(&fs_info->pinned_extents);
 848         extent_io_tree_cleanup(&fs_info->pending_del);
 849         extent_io_tree_cleanup(&fs_info->extent_ins);
 850 out:
 851         free(tree_root);
 852         free(extent_root);
 853         free(chunk_root);
 854         free(dev_root);
 855         free(csum_root);
 856         free(fs_info);
 857         return NULL;
 858 }
 859
 860 struct btrfs_fs_info *open_ctree_fs_info(const char *filename,
 861                                          u64 sb_bytenr, int writes,
 862                                          int partial)
 863 {
 864         int fp;
 865         struct btrfs_fs_info *info;
 866         int flags = O_CREAT | O_RDWR;
 867
 868         if (!writes)
 869                 flags = O_RDONLY;
 870
 871         fp = open(filename, flags, 0600);
 872         if (fp < 0) {
 873                 fprintf (stderr, "Could not open %s\n", filename);
 874                 return NULL;
 875         }
 876         info = __open_ctree_fd(fp, filename, sb_bytenr, 0, writes, partial);
 877         close(fp);
 878         return info;
 879 }
 880
 881 struct btrfs_root *open_ctree(const char *filename, u64 sb_bytenr, int writes)
 882 {
 883         struct btrfs_fs_info *info;
 884
 885         info = open_ctree_fs_info(filename, sb_bytenr, writes, 0);
 886         if (!info)
 887                 return NULL;
 888         return info->fs_root;
 889 }
 890
 891 struct btrfs_root *open_ctree_recovery(const char *filename, u64 sb_bytenr,
 892                                        u64 root_tree_bytenr)
 893 {
 894         int fp;
 895         struct btrfs_fs_info *info;
 896
 897
 898         fp = open(filename, O_RDONLY);
 899         if (fp < 0) {
 900                 fprintf (stderr, "Could not open %s\n", filename);
 901                 return NULL;
 902         }
 903         info = __open_ctree_fd(fp, filename, sb_bytenr,
 904                                root_tree_bytenr, 0, 0);
 905         close(fp);
 906
 907         if (!info)
 908                 return NULL;
 909         return info->fs_root;
 910 }
 911
 912 struct btrfs_root *open_ctree_fd(int fp, const char *path, u64 sb_bytenr,
 913                                  int writes)
 914 {
 915         struct btrfs_fs_info *info;
 916         info = __open_ctree_fd(fp, path, sb_bytenr, 0, writes, 0);
 917         if (!info)
 918                 return NULL;
 919         return info->fs_root;
 920 }
 921
 922 int btrfs_read_dev_super(int fd, struct btrfs_super_block *sb, u64 sb_bytenr)
 923 {
 924         u8 fsid[BTRFS_FSID_SIZE];
 925         int fsid_is_initialized = 0;
 926         struct btrfs_super_block buf;
 927         int i;
 928         int ret;
 929         u64 transid = 0;
 930         u64 bytenr;
 931
 932         if (sb_bytenr != BTRFS_SUPER_INFO_OFFSET) {
 933                 ret = pread64(fd, &buf, sizeof(buf), sb_bytenr);
 934                 if (ret < sizeof(buf))
 935                         return -1;
 936
 937                 if (btrfs_super_bytenr(&buf) != sb_bytenr ||
 938                     buf.magic != cpu_to_le64(BTRFS_MAGIC))
 939                         return -1;
 940
 941                 memcpy(sb, &buf, sizeof(*sb));
 942                 return 0;
 943         }
 944
 945         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 946                 bytenr = btrfs_sb_offset(i);
 947                 ret = pread64(fd, &buf, sizeof(buf), bytenr);
 948                 if (ret < sizeof(buf))
 949                         break;
 950
 951                 if (btrfs_super_bytenr(&buf) != bytenr )
 952                         continue;
 953                 /* if magic is NULL, the device was removed */
 954                 if (buf.magic == 0 && i == 0)
 955                         return -1;
 956                 if (buf.magic != cpu_to_le64(BTRFS_MAGIC))
 957                         continue;
 958
 959                 if (!fsid_is_initialized) {
 960                         memcpy(fsid, buf.fsid, sizeof(fsid));
 961                         fsid_is_initialized = 1;
 962                 } else if (memcmp(fsid, buf.fsid, sizeof(fsid))) {
 963                         /*
 964                          * the superblocks (the original one and
 965                          * its backups) contain data of different
 966                          * filesystems -> the super cannot be trusted
 967                          */
 968                         continue;
 969                 }
 970
 971                 if (btrfs_super_generation(&buf) > transid) {
 972                         memcpy(sb, &buf, sizeof(*sb));
 973                         transid = btrfs_super_generation(&buf);
 974                 }
 975         }
 976
 977         return transid > 0 ? 0 : -1;
 978 }
 979
 980 int write_dev_supers(struct btrfs_root *root, struct btrfs_super_block *sb,
 981                      struct btrfs_device *device)
 982 {
 983         u64 bytenr;
 984         u32 crc;
 985         int i, ret;
 986         void *buf;
 987
 988         buf = calloc(1, BTRFS_SUPER_INFO_SIZE);
 989         BUG_ON(!buf);
 990
 991         if (root->fs_info->super_bytenr != BTRFS_SUPER_INFO_OFFSET) {
 992                 btrfs_set_super_bytenr(sb, root->fs_info->super_bytenr);
 993                 crc = ~(u32)0;
 994                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
 995                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
 996                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
 997
 998                 memcpy(buf, sb, sizeof(*sb));
 999                 ret = pwrite64(device->fd, buf, BTRFS_SUPER_INFO_SIZE,
1000                                root->fs_info->super_bytenr);
1001                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1002                 goto out;
1003         }
1004
1005         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1006                 bytenr = btrfs_sb_offset(i);
1007                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
1008                         break;
1009
1010                 btrfs_set_super_bytenr(sb, bytenr);
1011
1012                 crc = ~(u32)0;
1013                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1014                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1015                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1016
1017                 memcpy(buf, sb, sizeof(*sb));
1018                 ret = pwrite64(device->fd, buf, BTRFS_SUPER_INFO_SIZE, bytenr);
1019                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1020         }
1021 out:
1022         free(buf);
1023         return 0;
1024 }
1025
1026 int write_all_supers(struct btrfs_root *root)
1027 {
1028         struct list_head *cur;
1029         struct list_head *head = &root->fs_info->fs_devices->devices;
1030         struct btrfs_device *dev;
1031         struct btrfs_super_block *sb;
1032         struct btrfs_dev_item *dev_item;
1033         int ret;
1034         u64 flags;
1035
1036         sb = &root->fs_info->super_copy;
1037         dev_item = &sb->dev_item;
1038         list_for_each(cur, head) {
1039                 dev = list_entry(cur, struct btrfs_device, dev_list);
1040                 if (!dev->writeable)
1041                         continue;
1042
1043                 btrfs_set_stack_device_generation(dev_item, 0);
1044                 btrfs_set_stack_device_type(dev_item, dev->type);
1045                 btrfs_set_stack_device_id(dev_item, dev->devid);
1046                 btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes);
1047                 btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used);
1048                 btrfs_set_stack_device_io_align(dev_item, dev->io_align);
1049                 btrfs_set_stack_device_io_width(dev_item, dev->io_width);
1050                 btrfs_set_stack_device_sector_size(dev_item, dev->sector_size);
1051                 memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE);
1052                 memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE);
1053
1054                 flags = btrfs_super_flags(sb);
1055                 btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN);
1056
1057                 ret = write_dev_supers(root, sb, dev);
1058                 BUG_ON(ret);
1059         }
1060         return 0;
1061 }
1062
1063 int write_ctree_super(struct btrfs_trans_handle *trans,
1064                       struct btrfs_root *root)
1065 {
1066         int ret;
1067         struct btrfs_root *tree_root = root->fs_info->tree_root;
1068         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
1069
1070         if (root->fs_info->readonly)
1071                 return 0;
1072
1073         btrfs_set_super_generation(&root->fs_info->super_copy,
1074                                    trans->transid);
1075         btrfs_set_super_root(&root->fs_info->super_copy,
1076                              tree_root->node->start);
1077         btrfs_set_super_root_level(&root->fs_info->super_copy,
1078                                    btrfs_header_level(tree_root->node));
1079         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
1080                                    chunk_root->node->start);
1081         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
1082                                          btrfs_header_level(chunk_root->node));
1083         btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
1084                                 btrfs_header_generation(chunk_root->node));
1085
1086         ret = write_all_supers(root);
1087         if (ret)
1088                 fprintf(stderr, "failed to write new super block err %d\n", ret);
1089         return ret;
1090 }
1091
1092 static int close_all_devices(struct btrfs_fs_info *fs_info)
1093 {
1094         struct list_head *list;
1095         struct list_head *next;
1096         struct btrfs_device *device;
1097
1098         return 0;
1099
1100         list = &fs_info->fs_devices->devices;
1101         list_for_each(next, list) {
1102                 device = list_entry(next, struct btrfs_device, dev_list);
1103                 close(device->fd);
1104         }
1105         return 0;
1106 }
1107
1108 int close_ctree(struct btrfs_root *root)
1109 {
1110         int ret;
1111         struct btrfs_trans_handle *trans;
1112         struct btrfs_fs_info *fs_info = root->fs_info;
1113
1114         if (fs_info->last_trans_committed !=
1115             fs_info->generation) {
1116                 trans = btrfs_start_transaction(root, 1);
1117                 btrfs_commit_transaction(trans, root);
1118                 trans = btrfs_start_transaction(root, 1);
1119                 ret = commit_tree_roots(trans, fs_info);
1120                 BUG_ON(ret);
1121                 ret = __commit_transaction(trans, root);
1122                 BUG_ON(ret);
1123                 write_ctree_super(trans, root);
1124                 btrfs_free_transaction(root, trans);
1125         }
1126         btrfs_free_block_groups(fs_info);
1127
1128         free_fs_roots(fs_info);
1129
1130         if (fs_info->extent_root->node)
1131                 free_extent_buffer(fs_info->extent_root->node);
1132         if (fs_info->tree_root->node)
1133                 free_extent_buffer(fs_info->tree_root->node);
1134         if (fs_info->chunk_root->node)
1135                 free_extent_buffer(fs_info->chunk_root->node);
1136         if (fs_info->dev_root->node)
1137                 free_extent_buffer(fs_info->dev_root->node);
1138         if (fs_info->csum_root->node)
1139                 free_extent_buffer(fs_info->csum_root->node);
1140
1141         if (fs_info->log_root_tree) {
1142                 if (fs_info->log_root_tree->node)
1143                         free_extent_buffer(fs_info->log_root_tree->node);
1144                 free(fs_info->log_root_tree);
1145         }
1146
1147         close_all_devices(fs_info);
1148         extent_io_tree_cleanup(&fs_info->extent_cache);
1149         extent_io_tree_cleanup(&fs_info->free_space_cache);
1150         extent_io_tree_cleanup(&fs_info->block_group_cache);
1151         extent_io_tree_cleanup(&fs_info->pinned_extents);
1152         extent_io_tree_cleanup(&fs_info->pending_del);
1153         extent_io_tree_cleanup(&fs_info->extent_ins);
1154
1155         free(fs_info->tree_root);
1156         free(fs_info->extent_root);
1157         free(fs_info->chunk_root);
1158         free(fs_info->dev_root);
1159         free(fs_info->csum_root);
1160         free(fs_info);
1161
1162         return 0;
1163 }
1164
1165 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1166                      struct extent_buffer *eb)
1167 {
1168         return clear_extent_buffer_dirty(eb);
1169 }
1170
1171 int wait_on_tree_block_writeback(struct btrfs_root *root,
1172                                  struct extent_buffer *eb)
1173 {
1174         return 0;
1175 }
1176
1177 void btrfs_mark_buffer_dirty(struct extent_buffer *eb)
1178 {
1179         set_extent_buffer_dirty(eb);
1180 }
1181
1182 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid)
1183 {
1184         int ret;
1185
1186         ret = extent_buffer_uptodate(buf);
1187         if (!ret)
1188                 return ret;
1189
1190         ret = verify_parent_transid(buf->tree, buf, parent_transid, 1);
1191         return !ret;
1192 }
1193
1194 int btrfs_set_buffer_uptodate(struct extent_buffer *eb)
1195 {
1196         return set_extent_buffer_uptodate(eb);
1197 }