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