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