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 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <sys/types.h>
  22 #include <sys/stat.h>
  23 #include <fcntl.h>
  24 #include <unistd.h>
  25 #include <uuid/uuid.h>
  26 #include "kerncompat.h"
  27 #include "radix-tree.h"
  28 #include "ctree.h"
  29 #include "disk-io.h"
  30 #include "volumes.h"
  31 #include "transaction.h"
  32 #include "crc32c.h"
  33 #include "utils.h"
  34 #include "print-tree.h"
  35 #include "rbtree-utils.h"
  36
  37 /* specified errno for check_tree_block */
  38 #define BTRFS_BAD_BYTENR                (-1)
  39 #define BTRFS_BAD_FSID                  (-2)
  40 #define BTRFS_BAD_LEVEL                 (-3)
  41 #define BTRFS_BAD_NRITEMS               (-4)
  42
  43 /* Calculate max possible nritems for a leaf/node */
  44 static u32 max_nritems(u8 level, u32 nodesize)
  45 {
  46
  47         if (level == 0)
  48                 return ((nodesize - sizeof(struct btrfs_header)) /
  49                         sizeof(struct btrfs_item));
  50         return ((nodesize - sizeof(struct btrfs_header)) /
  51                 sizeof(struct btrfs_key_ptr));
  52 }
  53
  54 static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
  55 {
  56
  57         struct btrfs_fs_devices *fs_devices;
  58         int ret = BTRFS_BAD_FSID;
  59
  60         if (buf->start != btrfs_header_bytenr(buf))
  61                 return BTRFS_BAD_BYTENR;
  62         if (btrfs_header_level(buf) >= BTRFS_MAX_LEVEL)
  63                 return BTRFS_BAD_LEVEL;
  64         if (btrfs_header_nritems(buf) > max_nritems(btrfs_header_level(buf),
  65                                                     root->nodesize))
  66                 return BTRFS_BAD_NRITEMS;
  67
  68         fs_devices = root->fs_info->fs_devices;
  69         while (fs_devices) {
  70                 if (root->fs_info->ignore_fsid_mismatch ||
  71                     !memcmp_extent_buffer(buf, fs_devices->fsid,
  72                                           btrfs_header_fsid(),
  73                                           BTRFS_FSID_SIZE)) {
  74                         ret = 0;
  75                         break;
  76                 }
  77                 fs_devices = fs_devices->seed;
  78         }
  79         return ret;
  80 }
  81
  82 static void print_tree_block_error(struct btrfs_root *root,
  83                                 struct extent_buffer *eb,
  84                                 int err)
  85 {
  86         char fs_uuid[BTRFS_UUID_UNPARSED_SIZE] = {'\0'};
  87         char found_uuid[BTRFS_UUID_UNPARSED_SIZE] = {'\0'};
  88         u8 buf[BTRFS_UUID_SIZE];
  89
  90         switch (err) {
  91         case BTRFS_BAD_FSID:
  92                 read_extent_buffer(eb, buf, btrfs_header_fsid(),
  93                                    BTRFS_UUID_SIZE);
  94                 uuid_unparse(buf, found_uuid);
  95                 uuid_unparse(root->fs_info->fsid, fs_uuid);
  96                 fprintf(stderr, "fsid mismatch, want=%s, have=%s\n",
  97                         fs_uuid, found_uuid);
  98                 break;
  99         case BTRFS_BAD_BYTENR:
 100                 fprintf(stderr, "bytenr mismatch, want=%llu, have=%llu\n",
 101                         eb->start, btrfs_header_bytenr(eb));
 102                 break;
 103         case BTRFS_BAD_LEVEL:
 104                 fprintf(stderr, "bad level, %u > %u\n",
 105                         btrfs_header_level(eb), BTRFS_MAX_LEVEL);
 106                 break;
 107         case BTRFS_BAD_NRITEMS:
 108                 fprintf(stderr, "invalid nr_items: %u\n",
 109                         btrfs_header_nritems(eb));
 110                 break;
 111         }
 112 }
 113
 114 u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len)
 115 {
 116         return crc32c(seed, data, len);
 117 }
 118
 119 void btrfs_csum_final(u32 crc, char *result)
 120 {
 121         *(__le32 *)result = ~cpu_to_le32(crc);
 122 }
 123
 124 static int __csum_tree_block_size(struct extent_buffer *buf, u16 csum_size,
 125                                   int verify, int silent)
 126 {
 127         char *result;
 128         u32 len;
 129         u32 crc = ~(u32)0;
 130
 131         result = malloc(csum_size * sizeof(char));
 132         if (!result)
 133                 return 1;
 134
 135         len = buf->len - BTRFS_CSUM_SIZE;
 136         crc = crc32c(crc, buf->data + BTRFS_CSUM_SIZE, len);
 137         btrfs_csum_final(crc, result);
 138
 139         if (verify) {
 140                 if (memcmp_extent_buffer(buf, result, 0, csum_size)) {
 141                         if (!silent)
 142                                 printk("checksum verify failed on %llu found %08X wanted %08X\n",
 143                                        (unsigned long long)buf->start,
 144                                        *((u32 *)result),
 145                                        *((u32*)(char *)buf->data));
 146                         free(result);
 147                         return 1;
 148                 }
 149         } else {
 150                 write_extent_buffer(buf, result, 0, csum_size);
 151         }
 152         free(result);
 153         return 0;
 154 }
 155
 156 int csum_tree_block_size(struct extent_buffer *buf, u16 csum_size, int verify)
 157 {
 158         return __csum_tree_block_size(buf, csum_size, verify, 0);
 159 }
 160
 161 int verify_tree_block_csum_silent(struct extent_buffer *buf, u16 csum_size)
 162 {
 163         return __csum_tree_block_size(buf, csum_size, 1, 1);
 164 }
 165
 166 int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf,
 167                            int verify)
 168 {
 169         u16 csum_size =
 170                 btrfs_super_csum_size(root->fs_info->super_copy);
 171         if (verify && root->fs_info->suppress_check_block_errors)
 172                 return verify_tree_block_csum_silent(buf, csum_size);
 173         return csum_tree_block_size(buf, csum_size, verify);
 174 }
 175
 176 struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
 177                                             u64 bytenr, u32 blocksize)
 178 {
 179         return find_extent_buffer(&root->fs_info->extent_cache,
 180                                   bytenr, blocksize);
 181 }
 182
 183 struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
 184                                                  u64 bytenr, u32 blocksize)
 185 {
 186         return alloc_extent_buffer(&root->fs_info->extent_cache, bytenr,
 187                                    blocksize);
 188 }
 189
 190 void readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 191                           u64 parent_transid)
 192 {
 193         struct extent_buffer *eb;
 194         u64 length;
 195         struct btrfs_multi_bio *multi = NULL;
 196         struct btrfs_device *device;
 197
 198         eb = btrfs_find_tree_block(root, bytenr, blocksize);
 199         if (!(eb && btrfs_buffer_uptodate(eb, parent_transid)) &&
 200             !btrfs_map_block(&root->fs_info->mapping_tree, READ,
 201                              bytenr, &length, &multi, 0, NULL)) {
 202                 device = multi->stripes[0].dev;
 203                 device->total_ios++;
 204                 blocksize = min(blocksize, (u32)(64 * 1024));
 205                 readahead(device->fd, multi->stripes[0].physical, blocksize);
 206         }
 207
 208         free_extent_buffer(eb);
 209         kfree(multi);
 210 }
 211
 212 static int verify_parent_transid(struct extent_io_tree *io_tree,
 213                                  struct extent_buffer *eb, u64 parent_transid,
 214                                  int ignore)
 215 {
 216         int ret;
 217
 218         if (!parent_transid || btrfs_header_generation(eb) == parent_transid)
 219                 return 0;
 220
 221         if (extent_buffer_uptodate(eb) &&
 222             btrfs_header_generation(eb) == parent_transid) {
 223                 ret = 0;
 224                 goto out;
 225         }
 226         printk("parent transid verify failed on %llu wanted %llu found %llu\n",
 227                (unsigned long long)eb->start,
 228                (unsigned long long)parent_transid,
 229                (unsigned long long)btrfs_header_generation(eb));
 230         if (ignore) {
 231                 eb->flags |= EXTENT_BAD_TRANSID;
 232                 printk("Ignoring transid failure\n");
 233                 return 0;
 234         }
 235
 236         ret = 1;
 237 out:
 238         clear_extent_buffer_uptodate(io_tree, eb);
 239         return ret;
 240
 241 }
 242
 243
 244 int read_whole_eb(struct btrfs_fs_info *info, struct extent_buffer *eb, int mirror)
 245 {
 246         unsigned long offset = 0;
 247         struct btrfs_multi_bio *multi = NULL;
 248         struct btrfs_device *device;
 249         int ret = 0;
 250         u64 read_len;
 251         unsigned long bytes_left = eb->len;
 252
 253         while (bytes_left) {
 254                 read_len = bytes_left;
 255                 device = NULL;
 256
 257                 if (!info->on_restoring &&
 258                     eb->start != BTRFS_SUPER_INFO_OFFSET) {
 259                         ret = btrfs_map_block(&info->mapping_tree, READ,
 260                                               eb->start + offset, &read_len, &multi,
 261                                               mirror, NULL);
 262                         if (ret) {
 263                                 printk("Couldn't map the block %Lu\n", eb->start + offset);
 264                                 kfree(multi);
 265                                 return -EIO;
 266                         }
 267                         device = multi->stripes[0].dev;
 268
 269                         if (device->fd == 0) {
 270                                 kfree(multi);
 271                                 return -EIO;
 272                         }
 273
 274                         eb->fd = device->fd;
 275                         device->total_ios++;
 276                         eb->dev_bytenr = multi->stripes[0].physical;
 277                         kfree(multi);
 278                         multi = NULL;
 279                 } else {
 280                         /* special case for restore metadump */
 281                         list_for_each_entry(device, &info->fs_devices->devices, dev_list) {
 282                                 if (device->devid == 1)
 283                                         break;
 284                         }
 285
 286                         eb->fd = device->fd;
 287                         eb->dev_bytenr = eb->start;
 288                         device->total_ios++;
 289                 }
 290
 291                 if (read_len > bytes_left)
 292                         read_len = bytes_left;
 293
 294                 ret = read_extent_from_disk(eb, offset, read_len);
 295                 if (ret)
 296                         return -EIO;
 297                 offset += read_len;
 298                 bytes_left -= read_len;
 299         }
 300         return 0;
 301 }
 302
 303 struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 304                                      u32 blocksize, u64 parent_transid)
 305 {
 306         int ret;
 307         struct extent_buffer *eb;
 308         u64 best_transid = 0;
 309         int mirror_num = 0;
 310         int good_mirror = 0;
 311         int num_copies;
 312         int ignore = 0;
 313
 314         eb = btrfs_find_create_tree_block(root, bytenr, blocksize);
 315         if (!eb)
 316                 return ERR_PTR(-ENOMEM);
 317
 318         if (btrfs_buffer_uptodate(eb, parent_transid))
 319                 return eb;
 320
 321         while (1) {
 322                 ret = read_whole_eb(root->fs_info, eb, mirror_num);
 323                 if (ret == 0 && csum_tree_block(root, eb, 1) == 0 &&
 324                     check_tree_block(root, eb) == 0 &&
 325                     verify_parent_transid(eb->tree, eb, parent_transid, ignore)
 326                     == 0) {
 327                         if (eb->flags & EXTENT_BAD_TRANSID &&
 328                             list_empty(&eb->recow)) {
 329                                 list_add_tail(&eb->recow,
 330                                               &root->fs_info->recow_ebs);
 331                                 eb->refs++;
 332                         }
 333                         btrfs_set_buffer_uptodate(eb);
 334                         return eb;
 335                 }
 336                 if (ignore) {
 337                         if (check_tree_block(root, eb)) {
 338                                 if (!root->fs_info->suppress_check_block_errors)
 339                                         print_tree_block_error(root, eb,
 340                                                 check_tree_block(root, eb));
 341                         } else {
 342                                 if (!root->fs_info->suppress_check_block_errors)
 343                                         fprintf(stderr, "Csum didn't match\n");
 344                         }
 345                         ret = -EIO;
 346                         break;
 347                 }
 348                 num_copies = btrfs_num_copies(&root->fs_info->mapping_tree,
 349                                               eb->start, eb->len);
 350                 if (num_copies == 1) {
 351                         ignore = 1;
 352                         continue;
 353                 }
 354                 if (btrfs_header_generation(eb) > best_transid && mirror_num) {
 355                         best_transid = btrfs_header_generation(eb);
 356                         good_mirror = mirror_num;
 357                 }
 358                 mirror_num++;
 359                 if (mirror_num > num_copies) {
 360                         mirror_num = good_mirror;
 361                         ignore = 1;
 362                         continue;
 363                 }
 364         }
 365         free_extent_buffer(eb);
 366         return ERR_PTR(ret);
 367 }
 368
 369 int write_and_map_eb(struct btrfs_trans_handle *trans,
 370                      struct btrfs_root *root,
 371                      struct extent_buffer *eb)
 372 {
 373         int ret;
 374         int dev_nr;
 375         u64 length;
 376         u64 *raid_map = NULL;
 377         struct btrfs_multi_bio *multi = NULL;
 378
 379         dev_nr = 0;
 380         length = eb->len;
 381         ret = btrfs_map_block(&root->fs_info->mapping_tree, WRITE,
 382                               eb->start, &length, &multi, 0, &raid_map);
 383
 384         if (raid_map) {
 385                 ret = write_raid56_with_parity(root->fs_info, eb, multi,
 386                                                length, raid_map);
 387                 BUG_ON(ret);
 388         } else while (dev_nr < multi->num_stripes) {
 389                 BUG_ON(ret);
 390                 eb->fd = multi->stripes[dev_nr].dev->fd;
 391                 eb->dev_bytenr = multi->stripes[dev_nr].physical;
 392                 multi->stripes[dev_nr].dev->total_ios++;
 393                 dev_nr++;
 394                 ret = write_extent_to_disk(eb);
 395                 BUG_ON(ret);
 396         }
 397         kfree(multi);
 398         return 0;
 399 }
 400
 401 int write_tree_block(struct btrfs_trans_handle *trans,
 402                      struct btrfs_root *root,
 403                      struct extent_buffer *eb)
 404 {
 405         if (check_tree_block(root, eb)) {
 406                 print_tree_block_error(root, eb, check_tree_block(root, eb));
 407                 BUG();
 408         }
 409
 410         if (trans && !btrfs_buffer_uptodate(eb, trans->transid))
 411                 BUG();
 412
 413         btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
 414         csum_tree_block(root, eb, 0);
 415
 416         return write_and_map_eb(trans, root, eb);
 417 }
 418
 419 int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
 420                         u32 stripesize, struct btrfs_root *root,
 421                         struct btrfs_fs_info *fs_info, u64 objectid)
 422 {
 423         root->node = NULL;
 424         root->commit_root = NULL;
 425         root->sectorsize = sectorsize;
 426         root->nodesize = nodesize;
 427         root->leafsize = leafsize;
 428         root->stripesize = stripesize;
 429         root->ref_cows = 0;
 430         root->track_dirty = 0;
 431
 432         root->fs_info = fs_info;
 433         root->objectid = objectid;
 434         root->last_trans = 0;
 435         root->highest_inode = 0;
 436         root->last_inode_alloc = 0;
 437
 438         INIT_LIST_HEAD(&root->dirty_list);
 439         INIT_LIST_HEAD(&root->orphan_data_extents);
 440         memset(&root->root_key, 0, sizeof(root->root_key));
 441         memset(&root->root_item, 0, sizeof(root->root_item));
 442         root->root_key.objectid = objectid;
 443         return 0;
 444 }
 445
 446 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 447                                struct btrfs_root *root)
 448 {
 449         int ret;
 450         u64 old_root_bytenr;
 451         struct btrfs_root *tree_root = root->fs_info->tree_root;
 452
 453         btrfs_write_dirty_block_groups(trans, root);
 454         while(1) {
 455                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 456                 if (old_root_bytenr == root->node->start)
 457                         break;
 458                 btrfs_set_root_bytenr(&root->root_item,
 459                                        root->node->start);
 460                 btrfs_set_root_generation(&root->root_item,
 461                                           trans->transid);
 462                 root->root_item.level = btrfs_header_level(root->node);
 463                 ret = btrfs_update_root(trans, tree_root,
 464                                         &root->root_key,
 465                                         &root->root_item);
 466                 BUG_ON(ret);
 467                 btrfs_write_dirty_block_groups(trans, root);
 468         }
 469         return 0;
 470 }
 471
 472 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 473                              struct btrfs_fs_info *fs_info)
 474 {
 475         struct btrfs_root *root;
 476         struct list_head *next;
 477         struct extent_buffer *eb;
 478         int ret;
 479
 480         if (fs_info->readonly)
 481                 return 0;
 482
 483         eb = fs_info->tree_root->node;
 484         extent_buffer_get(eb);
 485         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 486         free_extent_buffer(eb);
 487         if (ret)
 488                 return ret;
 489
 490         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 491                 next = fs_info->dirty_cowonly_roots.next;
 492                 list_del_init(next);
 493                 root = list_entry(next, struct btrfs_root, dirty_list);
 494                 update_cowonly_root(trans, root);
 495                 free_extent_buffer(root->commit_root);
 496                 root->commit_root = NULL;
 497         }
 498
 499         return 0;
 500 }
 501
 502 static int __commit_transaction(struct btrfs_trans_handle *trans,
 503                                 struct btrfs_root *root)
 504 {
 505         u64 start;
 506         u64 end;
 507         struct extent_buffer *eb;
 508         struct extent_io_tree *tree = &root->fs_info->extent_cache;
 509         int ret;
 510
 511         while(1) {
 512                 ret = find_first_extent_bit(tree, 0, &start, &end,
 513                                             EXTENT_DIRTY);
 514                 if (ret)
 515                         break;
 516                 while(start <= end) {
 517                         eb = find_first_extent_buffer(tree, start);
 518                         BUG_ON(!eb || eb->start != start);
 519                         ret = write_tree_block(trans, root, eb);
 520                         BUG_ON(ret);
 521                         start += eb->len;
 522                         clear_extent_buffer_dirty(eb);
 523                         free_extent_buffer(eb);
 524                 }
 525         }
 526         return 0;
 527 }
 528
 529 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 530                              struct btrfs_root *root)
 531 {
 532         u64 transid = trans->transid;
 533         int ret = 0;
 534         struct btrfs_fs_info *fs_info = root->fs_info;
 535
 536         if (root->commit_root == root->node)
 537                 goto commit_tree;
 538         if (root == root->fs_info->tree_root)
 539                 goto commit_tree;
 540
 541         free_extent_buffer(root->commit_root);
 542         root->commit_root = NULL;
 543
 544         btrfs_set_root_bytenr(&root->root_item, root->node->start);
 545         btrfs_set_root_generation(&root->root_item, trans->transid);
 546         root->root_item.level = btrfs_header_level(root->node);
 547         ret = btrfs_update_root(trans, root->fs_info->tree_root,
 548                                 &root->root_key, &root->root_item);
 549         BUG_ON(ret);
 550 commit_tree:
 551         ret = commit_tree_roots(trans, fs_info);
 552         BUG_ON(ret);
 553         ret = __commit_transaction(trans, root);
 554         BUG_ON(ret);
 555         write_ctree_super(trans, root);
 556         btrfs_finish_extent_commit(trans, fs_info->extent_root,
 557                                    &fs_info->pinned_extents);
 558         btrfs_free_transaction(root, trans);
 559         free_extent_buffer(root->commit_root);
 560         root->commit_root = NULL;
 561         fs_info->running_transaction = NULL;
 562         fs_info->last_trans_committed = transid;
 563         return 0;
 564 }
 565
 566 static int find_and_setup_root(struct btrfs_root *tree_root,
 567                                struct btrfs_fs_info *fs_info,
 568                                u64 objectid, struct btrfs_root *root)
 569 {
 570         int ret;
 571         u32 blocksize;
 572         u64 generation;
 573
 574         __setup_root(tree_root->nodesize, tree_root->leafsize,
 575                      tree_root->sectorsize, tree_root->stripesize,
 576                      root, fs_info, objectid);
 577         ret = btrfs_find_last_root(tree_root, objectid,
 578                                    &root->root_item, &root->root_key);
 579         if (ret)
 580                 return ret;
 581
 582         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 583         generation = btrfs_root_generation(&root->root_item);
 584         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 585                                      blocksize, generation);
 586         if (!extent_buffer_uptodate(root->node))
 587                 return -EIO;
 588
 589         return 0;
 590 }
 591
 592 static int find_and_setup_log_root(struct btrfs_root *tree_root,
 593                                struct btrfs_fs_info *fs_info,
 594                                struct btrfs_super_block *disk_super)
 595 {
 596         u32 blocksize;
 597         u64 blocknr = btrfs_super_log_root(disk_super);
 598         struct btrfs_root *log_root = malloc(sizeof(struct btrfs_root));
 599
 600         if (!log_root)
 601                 return -ENOMEM;
 602
 603         if (blocknr == 0) {
 604                 free(log_root);
 605                 return 0;
 606         }
 607
 608         blocksize = btrfs_level_size(tree_root,
 609                              btrfs_super_log_root_level(disk_super));
 610
 611         __setup_root(tree_root->nodesize, tree_root->leafsize,
 612                      tree_root->sectorsize, tree_root->stripesize,
 613                      log_root, fs_info, BTRFS_TREE_LOG_OBJECTID);
 614
 615         log_root->node = read_tree_block(tree_root, blocknr,
 616                                      blocksize,
 617                                      btrfs_super_generation(disk_super) + 1);
 618
 619         fs_info->log_root_tree = log_root;
 620
 621         if (!extent_buffer_uptodate(log_root->node)) {
 622                 free_extent_buffer(log_root->node);
 623                 free(log_root);
 624                 fs_info->log_root_tree = NULL;
 625                 return -EIO;
 626         }
 627
 628         return 0;
 629 }
 630
 631 int btrfs_free_fs_root(struct btrfs_root *root)
 632 {
 633         if (root->node)
 634                 free_extent_buffer(root->node);
 635         if (root->commit_root)
 636                 free_extent_buffer(root->commit_root);
 637         kfree(root);
 638         return 0;
 639 }
 640
 641 static void __free_fs_root(struct rb_node *node)
 642 {
 643         struct btrfs_root *root;
 644
 645         root = container_of(node, struct btrfs_root, rb_node);
 646         btrfs_free_fs_root(root);
 647 }
 648
 649 FREE_RB_BASED_TREE(fs_roots, __free_fs_root);
 650
 651 struct btrfs_root *btrfs_read_fs_root_no_cache(struct btrfs_fs_info *fs_info,
 652                                                struct btrfs_key *location)
 653 {
 654         struct btrfs_root *root;
 655         struct btrfs_root *tree_root = fs_info->tree_root;
 656         struct btrfs_path *path;
 657         struct extent_buffer *l;
 658         u64 generation;
 659         u32 blocksize;
 660         int ret = 0;
 661
 662         root = malloc(sizeof(*root));
 663         if (!root)
 664                 return ERR_PTR(-ENOMEM);
 665         memset(root, 0, sizeof(*root));
 666         if (location->offset == (u64)-1) {
 667                 ret = find_and_setup_root(tree_root, fs_info,
 668                                           location->objectid, root);
 669                 if (ret) {
 670                         free(root);
 671                         return ERR_PTR(ret);
 672                 }
 673                 goto insert;
 674         }
 675
 676         __setup_root(tree_root->nodesize, tree_root->leafsize,
 677                      tree_root->sectorsize, tree_root->stripesize,
 678                      root, fs_info, location->objectid);
 679
 680         path = btrfs_alloc_path();
 681         BUG_ON(!path);
 682         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 683         if (ret != 0) {
 684                 if (ret > 0)
 685                         ret = -ENOENT;
 686                 goto out;
 687         }
 688         l = path->nodes[0];
 689         read_extent_buffer(l, &root->root_item,
 690                btrfs_item_ptr_offset(l, path->slots[0]),
 691                sizeof(root->root_item));
 692         memcpy(&root->root_key, location, sizeof(*location));
 693         ret = 0;
 694 out:
 695         btrfs_free_path(path);
 696         if (ret) {
 697                 free(root);
 698                 return ERR_PTR(ret);
 699         }
 700         generation = btrfs_root_generation(&root->root_item);
 701         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 702         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 703                                      blocksize, generation);
 704         if (!extent_buffer_uptodate(root->node)) {
 705                 free(root);
 706                 return ERR_PTR(-EIO);
 707         }
 708 insert:
 709         root->ref_cows = 1;
 710         return root;
 711 }
 712
 713 static int btrfs_fs_roots_compare_objectids(struct rb_node *node,
 714                                             void *data)
 715 {
 716         u64 objectid = *((u64 *)data);
 717         struct btrfs_root *root;
 718
 719         root = rb_entry(node, struct btrfs_root, rb_node);
 720         if (objectid > root->objectid)
 721                 return 1;
 722         else if (objectid < root->objectid)
 723                 return -1;
 724         else
 725                 return 0;
 726 }
 727
 728 static int btrfs_fs_roots_compare_roots(struct rb_node *node1,
 729                                         struct rb_node *node2)
 730 {
 731         struct btrfs_root *root;
 732
 733         root = rb_entry(node2, struct btrfs_root, rb_node);
 734         return btrfs_fs_roots_compare_objectids(node1, (void *)&root->objectid);
 735 }
 736
 737 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 738                                       struct btrfs_key *location)
 739 {
 740         struct btrfs_root *root;
 741         struct rb_node *node;
 742         int ret;
 743         u64 objectid = location->objectid;
 744
 745         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 746                 return fs_info->tree_root;
 747         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 748                 return fs_info->extent_root;
 749         if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID)
 750                 return fs_info->chunk_root;
 751         if (location->objectid == BTRFS_DEV_TREE_OBJECTID)
 752                 return fs_info->dev_root;
 753         if (location->objectid == BTRFS_CSUM_TREE_OBJECTID)
 754                 return fs_info->csum_root;
 755         if (location->objectid == BTRFS_QUOTA_TREE_OBJECTID)
 756                 return fs_info->quota_root;
 757
 758         BUG_ON(location->objectid == BTRFS_TREE_RELOC_OBJECTID ||
 759                location->offset != (u64)-1);
 760
 761         node = rb_search(&fs_info->fs_root_tree, (void *)&objectid,
 762                          btrfs_fs_roots_compare_objectids, NULL);
 763         if (node)
 764                 return container_of(node, struct btrfs_root, rb_node);
 765
 766         root = btrfs_read_fs_root_no_cache(fs_info, location);
 767         if (IS_ERR(root))
 768                 return root;
 769
 770         ret = rb_insert(&fs_info->fs_root_tree, &root->rb_node,
 771                         btrfs_fs_roots_compare_roots);
 772         BUG_ON(ret);
 773         return root;
 774 }
 775
 776 void btrfs_free_fs_info(struct btrfs_fs_info *fs_info)
 777 {
 778         free(fs_info->tree_root);
 779         free(fs_info->extent_root);
 780         free(fs_info->chunk_root);
 781         free(fs_info->dev_root);
 782         free(fs_info->csum_root);
 783         free(fs_info->quota_root);
 784         free(fs_info->super_copy);
 785         free(fs_info->log_root_tree);
 786         free(fs_info);
 787 }
 788
 789 struct btrfs_fs_info *btrfs_new_fs_info(int writable, u64 sb_bytenr)
 790 {
 791         struct btrfs_fs_info *fs_info;
 792
 793         fs_info = malloc(sizeof(struct btrfs_fs_info));
 794         if (!fs_info)
 795                 return NULL;
 796
 797         memset(fs_info, 0, sizeof(struct btrfs_fs_info));
 798
 799         fs_info->tree_root = malloc(sizeof(struct btrfs_root));
 800         fs_info->extent_root = malloc(sizeof(struct btrfs_root));
 801         fs_info->chunk_root = malloc(sizeof(struct btrfs_root));
 802         fs_info->dev_root = malloc(sizeof(struct btrfs_root));
 803         fs_info->csum_root = malloc(sizeof(struct btrfs_root));
 804         fs_info->quota_root = malloc(sizeof(struct btrfs_root));
 805         fs_info->super_copy = malloc(BTRFS_SUPER_INFO_SIZE);
 806
 807         if (!fs_info->tree_root || !fs_info->extent_root ||
 808             !fs_info->chunk_root || !fs_info->dev_root ||
 809             !fs_info->csum_root || !fs_info->quota_root ||
 810             !fs_info->super_copy)
 811                 goto free_all;
 812
 813         memset(fs_info->super_copy, 0, BTRFS_SUPER_INFO_SIZE);
 814         memset(fs_info->tree_root, 0, sizeof(struct btrfs_root));
 815         memset(fs_info->extent_root, 0, sizeof(struct btrfs_root));
 816         memset(fs_info->chunk_root, 0, sizeof(struct btrfs_root));
 817         memset(fs_info->dev_root, 0, sizeof(struct btrfs_root));
 818         memset(fs_info->csum_root, 0, sizeof(struct btrfs_root));
 819         memset(fs_info->quota_root, 0, sizeof(struct btrfs_root));
 820
 821         extent_io_tree_init(&fs_info->extent_cache);
 822         extent_io_tree_init(&fs_info->free_space_cache);
 823         extent_io_tree_init(&fs_info->block_group_cache);
 824         extent_io_tree_init(&fs_info->pinned_extents);
 825         extent_io_tree_init(&fs_info->pending_del);
 826         extent_io_tree_init(&fs_info->extent_ins);
 827         fs_info->excluded_extents = NULL;
 828
 829         fs_info->fs_root_tree = RB_ROOT;
 830         cache_tree_init(&fs_info->mapping_tree.cache_tree);
 831
 832         mutex_init(&fs_info->fs_mutex);
 833         INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
 834         INIT_LIST_HEAD(&fs_info->space_info);
 835         INIT_LIST_HEAD(&fs_info->recow_ebs);
 836
 837         if (!writable)
 838                 fs_info->readonly = 1;
 839
 840         fs_info->super_bytenr = sb_bytenr;
 841         fs_info->data_alloc_profile = (u64)-1;
 842         fs_info->metadata_alloc_profile = (u64)-1;
 843         fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
 844         return fs_info;
 845 free_all:
 846         btrfs_free_fs_info(fs_info);
 847         return NULL;
 848 }
 849
 850 int btrfs_check_fs_compatibility(struct btrfs_super_block *sb, int writable)
 851 {
 852         u64 features;
 853
 854         features = btrfs_super_incompat_flags(sb) &
 855                    ~BTRFS_FEATURE_INCOMPAT_SUPP;
 856         if (features) {
 857                 printk("couldn't open because of unsupported "
 858                        "option features (%Lx).\n",
 859                        (unsigned long long)features);
 860                 return -ENOTSUP;
 861         }
 862
 863         features = btrfs_super_incompat_flags(sb);
 864         if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) {
 865                 features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF;
 866                 btrfs_set_super_incompat_flags(sb, features);
 867         }
 868
 869         features = btrfs_super_compat_ro_flags(sb) &
 870                 ~BTRFS_FEATURE_COMPAT_RO_SUPP;
 871         if (writable && features) {
 872                 printk("couldn't open RDWR because of unsupported "
 873                        "option features (%Lx).\n",
 874                        (unsigned long long)features);
 875                 return -ENOTSUP;
 876         }
 877         return 0;
 878 }
 879
 880 static int find_best_backup_root(struct btrfs_super_block *super)
 881 {
 882         struct btrfs_root_backup *backup;
 883         u64 orig_gen = btrfs_super_generation(super);
 884         u64 gen = 0;
 885         int best_index = 0;
 886         int i;
 887
 888         for (i = 0; i < BTRFS_NUM_BACKUP_ROOTS; i++) {
 889                 backup = super->super_roots + i;
 890                 if (btrfs_backup_tree_root_gen(backup) != orig_gen &&
 891                     btrfs_backup_tree_root_gen(backup) > gen) {
 892                         best_index = i;
 893                         gen = btrfs_backup_tree_root_gen(backup);
 894                 }
 895         }
 896         return best_index;
 897 }
 898
 899 static int setup_root_or_create_block(struct btrfs_fs_info *fs_info,
 900                                       enum btrfs_open_ctree_flags flags,
 901                                       struct btrfs_root *info_root,
 902                                       u64 objectid, char *str)
 903 {
 904         struct btrfs_super_block *sb = fs_info->super_copy;
 905         struct btrfs_root *root = fs_info->tree_root;
 906         u32 leafsize = btrfs_super_leafsize(sb);
 907         int ret;
 908
 909         ret = find_and_setup_root(root, fs_info, objectid, info_root);
 910         if (ret) {
 911                 printk("Couldn't setup %s tree\n", str);
 912                 if (!(flags & OPEN_CTREE_PARTIAL))
 913                         return -EIO;
 914                 /*
 915                  * Need a blank node here just so we don't screw up in the
 916                  * million of places that assume a root has a valid ->node
 917                  */
 918                 info_root->node =
 919                         btrfs_find_create_tree_block(info_root, 0, leafsize);
 920                 if (!info_root->node)
 921                         return -ENOMEM;
 922                 clear_extent_buffer_uptodate(NULL, info_root->node);
 923         }
 924
 925         return 0;
 926 }
 927
 928 int btrfs_setup_all_roots(struct btrfs_fs_info *fs_info, u64 root_tree_bytenr,
 929                           enum btrfs_open_ctree_flags flags)
 930 {
 931         struct btrfs_super_block *sb = fs_info->super_copy;
 932         struct btrfs_root *root;
 933         struct btrfs_key key;
 934         u32 sectorsize;
 935         u32 nodesize;
 936         u32 leafsize;
 937         u32 stripesize;
 938         u64 generation;
 939         u32 blocksize;
 940         int ret;
 941
 942         nodesize = btrfs_super_nodesize(sb);
 943         leafsize = btrfs_super_leafsize(sb);
 944         sectorsize = btrfs_super_sectorsize(sb);
 945         stripesize = btrfs_super_stripesize(sb);
 946
 947         root = fs_info->tree_root;
 948         __setup_root(nodesize, leafsize, sectorsize, stripesize,
 949                      root, fs_info, BTRFS_ROOT_TREE_OBJECTID);
 950         blocksize = btrfs_level_size(root, btrfs_super_root_level(sb));
 951         generation = btrfs_super_generation(sb);
 952
 953         if (!root_tree_bytenr && !(flags & OPEN_CTREE_BACKUP_ROOT)) {
 954                 root_tree_bytenr = btrfs_super_root(sb);
 955         } else if (flags & OPEN_CTREE_BACKUP_ROOT) {
 956                 struct btrfs_root_backup *backup;
 957                 int index = find_best_backup_root(sb);
 958                 if (index >= BTRFS_NUM_BACKUP_ROOTS) {
 959                         fprintf(stderr, "Invalid backup root number\n");
 960                         return -EIO;
 961                 }
 962                 backup = fs_info->super_copy->super_roots + index;
 963                 root_tree_bytenr = btrfs_backup_tree_root(backup);
 964                 generation = btrfs_backup_tree_root_gen(backup);
 965         }
 966
 967         root->node = read_tree_block(root, root_tree_bytenr, blocksize,
 968                                      generation);
 969         if (!extent_buffer_uptodate(root->node)) {
 970                 fprintf(stderr, "Couldn't read tree root\n");
 971                 return -EIO;
 972         }
 973
 974         ret = setup_root_or_create_block(fs_info, flags, fs_info->extent_root,
 975                                          BTRFS_EXTENT_TREE_OBJECTID, "extent");
 976         if (ret)
 977                 return ret;
 978         fs_info->extent_root->track_dirty = 1;
 979
 980         ret = find_and_setup_root(root, fs_info, BTRFS_DEV_TREE_OBJECTID,
 981                                   fs_info->dev_root);
 982         if (ret) {
 983                 printk("Couldn't setup device tree\n");
 984                 return -EIO;
 985         }
 986         fs_info->dev_root->track_dirty = 1;
 987
 988         ret = setup_root_or_create_block(fs_info, flags, fs_info->csum_root,
 989                                          BTRFS_CSUM_TREE_OBJECTID, "csum");
 990         if (ret)
 991                 return ret;
 992         fs_info->csum_root->track_dirty = 1;
 993
 994         ret = find_and_setup_root(root, fs_info, BTRFS_QUOTA_TREE_OBJECTID,
 995                                   fs_info->quota_root);
 996         if (ret == 0)
 997                 fs_info->quota_enabled = 1;
 998
 999         ret = find_and_setup_log_root(root, fs_info, sb);
1000         if (ret) {
1001                 printk("Couldn't setup log root tree\n");
1002                 if (!(flags & OPEN_CTREE_PARTIAL))
1003                         return -EIO;
1004         }
1005
1006         fs_info->generation = generation;
1007         fs_info->last_trans_committed = generation;
1008         if (extent_buffer_uptodate(fs_info->extent_root->node) &&
1009             !(flags & OPEN_CTREE_NO_BLOCK_GROUPS))
1010                 btrfs_read_block_groups(fs_info->tree_root);
1011
1012         key.objectid = BTRFS_FS_TREE_OBJECTID;
1013         key.type = BTRFS_ROOT_ITEM_KEY;
1014         key.offset = (u64)-1;
1015         fs_info->fs_root = btrfs_read_fs_root(fs_info, &key);
1016
1017         if (IS_ERR(fs_info->fs_root))
1018                 return -EIO;
1019         return 0;
1020 }
1021
1022 void btrfs_release_all_roots(struct btrfs_fs_info *fs_info)
1023 {
1024         if (fs_info->quota_root)
1025                 free_extent_buffer(fs_info->quota_root->node);
1026         if (fs_info->csum_root)
1027                 free_extent_buffer(fs_info->csum_root->node);
1028         if (fs_info->dev_root)
1029                 free_extent_buffer(fs_info->dev_root->node);
1030         if (fs_info->extent_root)
1031                 free_extent_buffer(fs_info->extent_root->node);
1032         if (fs_info->tree_root)
1033                 free_extent_buffer(fs_info->tree_root->node);
1034         if (fs_info->log_root_tree)
1035                 free_extent_buffer(fs_info->log_root_tree->node);
1036         if (fs_info->chunk_root)
1037                 free_extent_buffer(fs_info->chunk_root->node);
1038 }
1039
1040 static void free_map_lookup(struct cache_extent *ce)
1041 {
1042         struct map_lookup *map;
1043
1044         map = container_of(ce, struct map_lookup, ce);
1045         kfree(map);
1046 }
1047
1048 FREE_EXTENT_CACHE_BASED_TREE(mapping_cache, free_map_lookup);
1049
1050 void btrfs_cleanup_all_caches(struct btrfs_fs_info *fs_info)
1051 {
1052         while (!list_empty(&fs_info->recow_ebs)) {
1053                 struct extent_buffer *eb;
1054                 eb = list_first_entry(&fs_info->recow_ebs,
1055                                       struct extent_buffer, recow);
1056                 list_del_init(&eb->recow);
1057                 free_extent_buffer(eb);
1058         }
1059         free_mapping_cache_tree(&fs_info->mapping_tree.cache_tree);
1060         extent_io_tree_cleanup(&fs_info->extent_cache);
1061         extent_io_tree_cleanup(&fs_info->free_space_cache);
1062         extent_io_tree_cleanup(&fs_info->block_group_cache);
1063         extent_io_tree_cleanup(&fs_info->pinned_extents);
1064         extent_io_tree_cleanup(&fs_info->pending_del);
1065         extent_io_tree_cleanup(&fs_info->extent_ins);
1066 }
1067
1068 int btrfs_scan_fs_devices(int fd, const char *path,
1069                           struct btrfs_fs_devices **fs_devices,
1070                           u64 sb_bytenr, int super_recover,
1071                           int skip_devices)
1072 {
1073         u64 total_devs;
1074         u64 dev_size;
1075         off_t seek_ret;
1076         int ret;
1077         if (!sb_bytenr)
1078                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1079
1080         seek_ret = lseek(fd, 0, SEEK_END);
1081         if (seek_ret < 0)
1082                 return -errno;
1083
1084         dev_size = seek_ret;
1085         lseek(fd, 0, SEEK_SET);
1086         if (sb_bytenr > dev_size) {
1087                 fprintf(stderr, "Superblock bytenr is larger than device size\n");
1088                 return -EINVAL;
1089         }
1090
1091         ret = btrfs_scan_one_device(fd, path, fs_devices,
1092                                     &total_devs, sb_bytenr, super_recover);
1093         if (ret) {
1094                 fprintf(stderr, "No valid Btrfs found on %s\n", path);
1095                 return ret;
1096         }
1097
1098         if (!skip_devices && total_devs != 1) {
1099                 ret = btrfs_scan_lblkid();
1100                 if (ret)
1101                         return ret;
1102         }
1103         return 0;
1104 }
1105
1106 int btrfs_setup_chunk_tree_and_device_map(struct btrfs_fs_info *fs_info)
1107 {
1108         struct btrfs_super_block *sb = fs_info->super_copy;
1109         u32 sectorsize;
1110         u32 nodesize;
1111         u32 leafsize;
1112         u32 blocksize;
1113         u32 stripesize;
1114         u64 generation;
1115         int ret;
1116
1117         nodesize = btrfs_super_nodesize(sb);
1118         leafsize = btrfs_super_leafsize(sb);
1119         sectorsize = btrfs_super_sectorsize(sb);
1120         stripesize = btrfs_super_stripesize(sb);
1121
1122         __setup_root(nodesize, leafsize, sectorsize, stripesize,
1123                      fs_info->chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID);
1124
1125         ret = btrfs_read_sys_array(fs_info->chunk_root);
1126         if (ret)
1127                 return ret;
1128
1129         blocksize = btrfs_level_size(fs_info->chunk_root,
1130                                      btrfs_super_chunk_root_level(sb));
1131         generation = btrfs_super_chunk_root_generation(sb);
1132
1133         fs_info->chunk_root->node = read_tree_block(fs_info->chunk_root,
1134                                                     btrfs_super_chunk_root(sb),
1135                                                     blocksize, generation);
1136         if (!extent_buffer_uptodate(fs_info->chunk_root->node)) {
1137                 fprintf(stderr, "Couldn't read chunk root\n");
1138                 return -EIO;
1139         }
1140
1141         if (!(btrfs_super_flags(sb) & BTRFS_SUPER_FLAG_METADUMP)) {
1142                 ret = btrfs_read_chunk_tree(fs_info->chunk_root);
1143                 if (ret) {
1144                         fprintf(stderr, "Couldn't read chunk tree\n");
1145                         return ret;
1146                 }
1147         }
1148         return 0;
1149 }
1150
1151 static struct btrfs_fs_info *__open_ctree_fd(int fp, const char *path,
1152                                              u64 sb_bytenr,
1153                                              u64 root_tree_bytenr,
1154                                              enum btrfs_open_ctree_flags flags)
1155 {
1156         struct btrfs_fs_info *fs_info;
1157         struct btrfs_super_block *disk_super;
1158         struct btrfs_fs_devices *fs_devices = NULL;
1159         struct extent_buffer *eb;
1160         int ret;
1161         int oflags;
1162
1163         if (sb_bytenr == 0)
1164                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1165
1166         /* try to drop all the caches */
1167         if (posix_fadvise(fp, 0, 0, POSIX_FADV_DONTNEED))
1168                 fprintf(stderr, "Warning, could not drop caches\n");
1169
1170         fs_info = btrfs_new_fs_info(flags & OPEN_CTREE_WRITES, sb_bytenr);
1171         if (!fs_info) {
1172                 fprintf(stderr, "Failed to allocate memory for fs_info\n");
1173                 return NULL;
1174         }
1175         if (flags & OPEN_CTREE_RESTORE)
1176                 fs_info->on_restoring = 1;
1177         if (flags & OPEN_CTREE_SUPPRESS_CHECK_BLOCK_ERRORS)
1178                 fs_info->suppress_check_block_errors = 1;
1179         if (flags & OPEN_CTREE_IGNORE_FSID_MISMATCH)
1180                 fs_info->ignore_fsid_mismatch = 1;
1181
1182         ret = btrfs_scan_fs_devices(fp, path, &fs_devices, sb_bytenr,
1183                                     (flags & OPEN_CTREE_RECOVER_SUPER),
1184                                     (flags & OPEN_CTREE_NO_DEVICES));
1185         if (ret)
1186                 goto out;
1187
1188         fs_info->fs_devices = fs_devices;
1189         if (flags & OPEN_CTREE_WRITES)
1190                 oflags = O_RDWR;
1191         else
1192                 oflags = O_RDONLY;
1193
1194         if (flags & OPEN_CTREE_EXCLUSIVE)
1195                 oflags |= O_EXCL;
1196
1197         ret = btrfs_open_devices(fs_devices, oflags);
1198         if (ret)
1199                 goto out;
1200
1201         disk_super = fs_info->super_copy;
1202         if (!(flags & OPEN_CTREE_RECOVER_SUPER))
1203                 ret = btrfs_read_dev_super(fs_devices->latest_bdev,
1204                                            disk_super, sb_bytenr, 1);
1205         else
1206                 ret = btrfs_read_dev_super(fp, disk_super, sb_bytenr, 0);
1207         if (ret) {
1208                 printk("No valid btrfs found\n");
1209                 goto out_devices;
1210         }
1211
1212         if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_CHANGING_FSID &&
1213             !fs_info->ignore_fsid_mismatch) {
1214                 fprintf(stderr, "ERROR: Filesystem UUID change in progress\n");
1215                 goto out_devices;
1216         }
1217
1218         memcpy(fs_info->fsid, &disk_super->fsid, BTRFS_FSID_SIZE);
1219
1220         ret = btrfs_check_fs_compatibility(fs_info->super_copy,
1221                                            flags & OPEN_CTREE_WRITES);
1222         if (ret)
1223                 goto out_devices;
1224
1225         ret = btrfs_setup_chunk_tree_and_device_map(fs_info);
1226         if (ret)
1227                 goto out_chunk;
1228
1229         eb = fs_info->chunk_root->node;
1230         read_extent_buffer(eb, fs_info->chunk_tree_uuid,
1231                            btrfs_header_chunk_tree_uuid(eb),
1232                            BTRFS_UUID_SIZE);
1233
1234         ret = btrfs_setup_all_roots(fs_info, root_tree_bytenr, flags);
1235         if (ret && !(flags & __OPEN_CTREE_RETURN_CHUNK_ROOT))
1236                 goto out_chunk;
1237
1238         return fs_info;
1239
1240 out_chunk:
1241         btrfs_release_all_roots(fs_info);
1242         btrfs_cleanup_all_caches(fs_info);
1243 out_devices:
1244         btrfs_close_devices(fs_devices);
1245 out:
1246         btrfs_free_fs_info(fs_info);
1247         return NULL;
1248 }
1249
1250 struct btrfs_fs_info *open_ctree_fs_info(const char *filename,
1251                                          u64 sb_bytenr, u64 root_tree_bytenr,
1252                                          enum btrfs_open_ctree_flags flags)
1253 {
1254         int fp;
1255         struct btrfs_fs_info *info;
1256         int oflags = O_CREAT | O_RDWR;
1257
1258         if (!(flags & OPEN_CTREE_WRITES))
1259                 oflags = O_RDONLY;
1260
1261         fp = open(filename, oflags, 0600);
1262         if (fp < 0) {
1263                 fprintf (stderr, "Could not open %s\n", filename);
1264                 return NULL;
1265         }
1266         info = __open_ctree_fd(fp, filename, sb_bytenr, root_tree_bytenr,
1267                                flags);
1268         close(fp);
1269         return info;
1270 }
1271
1272 struct btrfs_root *open_ctree(const char *filename, u64 sb_bytenr,
1273                               enum btrfs_open_ctree_flags flags)
1274 {
1275         struct btrfs_fs_info *info;
1276
1277         info = open_ctree_fs_info(filename, sb_bytenr, 0, flags);
1278         if (!info)
1279                 return NULL;
1280         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1281                 return info->chunk_root;
1282         return info->fs_root;
1283 }
1284
1285 struct btrfs_root *open_ctree_fd(int fp, const char *path, u64 sb_bytenr,
1286                                  enum btrfs_open_ctree_flags flags)
1287 {
1288         struct btrfs_fs_info *info;
1289         info = __open_ctree_fd(fp, path, sb_bytenr, 0, flags);
1290         if (!info)
1291                 return NULL;
1292         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1293                 return info->chunk_root;
1294         return info->fs_root;
1295 }
1296
1297 int btrfs_read_dev_super(int fd, struct btrfs_super_block *sb, u64 sb_bytenr,
1298                          int super_recover)
1299 {
1300         u8 fsid[BTRFS_FSID_SIZE];
1301         int fsid_is_initialized = 0;
1302         struct btrfs_super_block buf;
1303         int i;
1304         int ret;
1305         int max_super = super_recover ? BTRFS_SUPER_MIRROR_MAX : 1;
1306         u64 transid = 0;
1307         u64 bytenr;
1308
1309         if (sb_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1310                 ret = pread64(fd, &buf, sizeof(buf), sb_bytenr);
1311                 if (ret < sizeof(buf))
1312                         return -1;
1313
1314                 if (btrfs_super_bytenr(&buf) != sb_bytenr ||
1315                     btrfs_super_magic(&buf) != BTRFS_MAGIC)
1316                         return -1;
1317
1318                 memcpy(sb, &buf, sizeof(*sb));
1319                 return 0;
1320         }
1321
1322         /*
1323         * we would like to check all the supers, but that would make
1324         * a btrfs mount succeed after a mkfs from a different FS.
1325         * So, we need to add a special mount option to scan for
1326         * later supers, using BTRFS_SUPER_MIRROR_MAX instead
1327         */
1328
1329         for (i = 0; i < max_super; i++) {
1330                 bytenr = btrfs_sb_offset(i);
1331                 ret = pread64(fd, &buf, sizeof(buf), bytenr);
1332                 if (ret < sizeof(buf))
1333                         break;
1334
1335                 if (btrfs_super_bytenr(&buf) != bytenr )
1336                         continue;
1337                 /* if magic is NULL, the device was removed */
1338                 if (btrfs_super_magic(&buf) == 0 && i == 0)
1339                         return -1;
1340                 if (btrfs_super_magic(&buf) != BTRFS_MAGIC)
1341                         continue;
1342
1343                 if (!fsid_is_initialized) {
1344                         memcpy(fsid, buf.fsid, sizeof(fsid));
1345                         fsid_is_initialized = 1;
1346                 } else if (memcmp(fsid, buf.fsid, sizeof(fsid))) {
1347                         /*
1348                          * the superblocks (the original one and
1349                          * its backups) contain data of different
1350                          * filesystems -> the super cannot be trusted
1351                          */
1352                         continue;
1353                 }
1354
1355                 if (btrfs_super_generation(&buf) > transid) {
1356                         memcpy(sb, &buf, sizeof(*sb));
1357                         transid = btrfs_super_generation(&buf);
1358                 }
1359         }
1360
1361         return transid > 0 ? 0 : -1;
1362 }
1363
1364 static int write_dev_supers(struct btrfs_root *root,
1365                             struct btrfs_super_block *sb,
1366                             struct btrfs_device *device)
1367 {
1368         u64 bytenr;
1369         u32 crc;
1370         int i, ret;
1371
1372         if (root->fs_info->super_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1373                 btrfs_set_super_bytenr(sb, root->fs_info->super_bytenr);
1374                 crc = ~(u32)0;
1375                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1376                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1377                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1378
1379                 /*
1380                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1381                  * zero filled, we can use it directly
1382                  */
1383                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1384                                 BTRFS_SUPER_INFO_SIZE,
1385                                 root->fs_info->super_bytenr);
1386                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1387                 return 0;
1388         }
1389
1390         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1391                 bytenr = btrfs_sb_offset(i);
1392                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
1393                         break;
1394
1395                 btrfs_set_super_bytenr(sb, bytenr);
1396
1397                 crc = ~(u32)0;
1398                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1399                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1400                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1401
1402                 /*
1403                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1404                  * zero filled, we can use it directly
1405                  */
1406                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1407                                 BTRFS_SUPER_INFO_SIZE, bytenr);
1408                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1409         }
1410
1411         return 0;
1412 }
1413
1414 int write_all_supers(struct btrfs_root *root)
1415 {
1416         struct list_head *cur;
1417         struct list_head *head = &root->fs_info->fs_devices->devices;
1418         struct btrfs_device *dev;
1419         struct btrfs_super_block *sb;
1420         struct btrfs_dev_item *dev_item;
1421         int ret;
1422         u64 flags;
1423
1424         sb = root->fs_info->super_copy;
1425         dev_item = &sb->dev_item;
1426         list_for_each(cur, head) {
1427                 dev = list_entry(cur, struct btrfs_device, dev_list);
1428                 if (!dev->writeable)
1429                         continue;
1430
1431                 btrfs_set_stack_device_generation(dev_item, 0);
1432                 btrfs_set_stack_device_type(dev_item, dev->type);
1433                 btrfs_set_stack_device_id(dev_item, dev->devid);
1434                 btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes);
1435                 btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used);
1436                 btrfs_set_stack_device_io_align(dev_item, dev->io_align);
1437                 btrfs_set_stack_device_io_width(dev_item, dev->io_width);
1438                 btrfs_set_stack_device_sector_size(dev_item, dev->sector_size);
1439                 memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE);
1440                 memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE);
1441
1442                 flags = btrfs_super_flags(sb);
1443                 btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN);
1444
1445                 ret = write_dev_supers(root, sb, dev);
1446                 BUG_ON(ret);
1447         }
1448         return 0;
1449 }
1450
1451 int write_ctree_super(struct btrfs_trans_handle *trans,
1452                       struct btrfs_root *root)
1453 {
1454         int ret;
1455         struct btrfs_root *tree_root = root->fs_info->tree_root;
1456         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
1457
1458         if (root->fs_info->readonly)
1459                 return 0;
1460
1461         btrfs_set_super_generation(root->fs_info->super_copy,
1462                                    trans->transid);
1463         btrfs_set_super_root(root->fs_info->super_copy,
1464                              tree_root->node->start);
1465         btrfs_set_super_root_level(root->fs_info->super_copy,
1466                                    btrfs_header_level(tree_root->node));
1467         btrfs_set_super_chunk_root(root->fs_info->super_copy,
1468                                    chunk_root->node->start);
1469         btrfs_set_super_chunk_root_level(root->fs_info->super_copy,
1470                                          btrfs_header_level(chunk_root->node));
1471         btrfs_set_super_chunk_root_generation(root->fs_info->super_copy,
1472                                 btrfs_header_generation(chunk_root->node));
1473
1474         ret = write_all_supers(root);
1475         if (ret)
1476                 fprintf(stderr, "failed to write new super block err %d\n", ret);
1477         return ret;
1478 }
1479
1480 int close_ctree(struct btrfs_root *root)
1481 {
1482         int ret;
1483         struct btrfs_trans_handle *trans;
1484         struct btrfs_fs_info *fs_info = root->fs_info;
1485
1486         if (fs_info->last_trans_committed !=
1487             fs_info->generation) {
1488                 trans = btrfs_start_transaction(root, 1);
1489                 btrfs_commit_transaction(trans, root);
1490                 trans = btrfs_start_transaction(root, 1);
1491                 ret = commit_tree_roots(trans, fs_info);
1492                 BUG_ON(ret);
1493                 ret = __commit_transaction(trans, root);
1494                 BUG_ON(ret);
1495                 write_ctree_super(trans, root);
1496                 btrfs_free_transaction(root, trans);
1497         }
1498         btrfs_free_block_groups(fs_info);
1499
1500         free_fs_roots_tree(&fs_info->fs_root_tree);
1501
1502         btrfs_release_all_roots(fs_info);
1503         btrfs_close_devices(fs_info->fs_devices);
1504         btrfs_cleanup_all_caches(fs_info);
1505         btrfs_free_fs_info(fs_info);
1506         return 0;
1507 }
1508
1509 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1510                      struct extent_buffer *eb)
1511 {
1512         return clear_extent_buffer_dirty(eb);
1513 }
1514
1515 int wait_on_tree_block_writeback(struct btrfs_root *root,
1516                                  struct extent_buffer *eb)
1517 {
1518         return 0;
1519 }
1520
1521 void btrfs_mark_buffer_dirty(struct extent_buffer *eb)
1522 {
1523         set_extent_buffer_dirty(eb);
1524 }
1525
1526 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid)
1527 {
1528         int ret;
1529
1530         ret = extent_buffer_uptodate(buf);
1531         if (!ret)
1532                 return ret;
1533
1534         ret = verify_parent_transid(buf->tree, buf, parent_transid, 1);
1535         return !ret;
1536 }
1537
1538 int btrfs_set_buffer_uptodate(struct extent_buffer *eb)
1539 {
1540         return set_extent_buffer_uptodate(eb);
1541 }