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 read_extent_data(struct btrfs_root *root, char *data,
 370                            u64 logical, u64 *len, int mirror)
 371 {
 372         u64 offset = 0;
 373         struct btrfs_multi_bio *multi = NULL;
 374         struct btrfs_fs_info *info = root->fs_info;
 375         struct btrfs_device *device;
 376         int ret = 0;
 377         u64 max_len = *len;
 378
 379         ret = btrfs_map_block(&info->mapping_tree, READ, logical, len,
 380                               &multi, mirror, NULL);
 381         if (ret) {
 382                 fprintf(stderr, "Couldn't map the block %llu\n",
 383                                 logical + offset);
 384                 goto err;
 385         }
 386         device = multi->stripes[0].dev;
 387
 388         if (device->fd == 0)
 389                 goto err;
 390         if (*len > max_len)
 391                 *len = max_len;
 392
 393         ret = pread64(device->fd, data, *len, multi->stripes[0].physical);
 394         if (ret != *len)
 395                 ret = -EIO;
 396         else
 397                 ret = 0;
 398 err:
 399         kfree(multi);
 400         return ret;
 401 }
 402
 403 int write_and_map_eb(struct btrfs_trans_handle *trans,
 404                      struct btrfs_root *root,
 405                      struct extent_buffer *eb)
 406 {
 407         int ret;
 408         int dev_nr;
 409         u64 length;
 410         u64 *raid_map = NULL;
 411         struct btrfs_multi_bio *multi = NULL;
 412
 413         dev_nr = 0;
 414         length = eb->len;
 415         ret = btrfs_map_block(&root->fs_info->mapping_tree, WRITE,
 416                               eb->start, &length, &multi, 0, &raid_map);
 417
 418         if (raid_map) {
 419                 ret = write_raid56_with_parity(root->fs_info, eb, multi,
 420                                                length, raid_map);
 421                 BUG_ON(ret);
 422         } else while (dev_nr < multi->num_stripes) {
 423                 BUG_ON(ret);
 424                 eb->fd = multi->stripes[dev_nr].dev->fd;
 425                 eb->dev_bytenr = multi->stripes[dev_nr].physical;
 426                 multi->stripes[dev_nr].dev->total_ios++;
 427                 dev_nr++;
 428                 ret = write_extent_to_disk(eb);
 429                 BUG_ON(ret);
 430         }
 431         kfree(multi);
 432         return 0;
 433 }
 434
 435 int write_tree_block(struct btrfs_trans_handle *trans,
 436                      struct btrfs_root *root,
 437                      struct extent_buffer *eb)
 438 {
 439         if (check_tree_block(root, eb)) {
 440                 print_tree_block_error(root, eb, check_tree_block(root, eb));
 441                 BUG();
 442         }
 443
 444         if (trans && !btrfs_buffer_uptodate(eb, trans->transid))
 445                 BUG();
 446
 447         btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
 448         csum_tree_block(root, eb, 0);
 449
 450         return write_and_map_eb(trans, root, eb);
 451 }
 452
 453 int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
 454                         u32 stripesize, struct btrfs_root *root,
 455                         struct btrfs_fs_info *fs_info, u64 objectid)
 456 {
 457         root->node = NULL;
 458         root->commit_root = NULL;
 459         root->sectorsize = sectorsize;
 460         root->nodesize = nodesize;
 461         root->leafsize = leafsize;
 462         root->stripesize = stripesize;
 463         root->ref_cows = 0;
 464         root->track_dirty = 0;
 465
 466         root->fs_info = fs_info;
 467         root->objectid = objectid;
 468         root->last_trans = 0;
 469         root->highest_inode = 0;
 470         root->last_inode_alloc = 0;
 471
 472         INIT_LIST_HEAD(&root->dirty_list);
 473         INIT_LIST_HEAD(&root->orphan_data_extents);
 474         memset(&root->root_key, 0, sizeof(root->root_key));
 475         memset(&root->root_item, 0, sizeof(root->root_item));
 476         root->root_key.objectid = objectid;
 477         return 0;
 478 }
 479
 480 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 481                                struct btrfs_root *root)
 482 {
 483         int ret;
 484         u64 old_root_bytenr;
 485         struct btrfs_root *tree_root = root->fs_info->tree_root;
 486
 487         btrfs_write_dirty_block_groups(trans, root);
 488         while(1) {
 489                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 490                 if (old_root_bytenr == root->node->start)
 491                         break;
 492                 btrfs_set_root_bytenr(&root->root_item,
 493                                        root->node->start);
 494                 btrfs_set_root_generation(&root->root_item,
 495                                           trans->transid);
 496                 root->root_item.level = btrfs_header_level(root->node);
 497                 ret = btrfs_update_root(trans, tree_root,
 498                                         &root->root_key,
 499                                         &root->root_item);
 500                 BUG_ON(ret);
 501                 btrfs_write_dirty_block_groups(trans, root);
 502         }
 503         return 0;
 504 }
 505
 506 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 507                              struct btrfs_fs_info *fs_info)
 508 {
 509         struct btrfs_root *root;
 510         struct list_head *next;
 511         struct extent_buffer *eb;
 512         int ret;
 513
 514         if (fs_info->readonly)
 515                 return 0;
 516
 517         eb = fs_info->tree_root->node;
 518         extent_buffer_get(eb);
 519         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 520         free_extent_buffer(eb);
 521         if (ret)
 522                 return ret;
 523
 524         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 525                 next = fs_info->dirty_cowonly_roots.next;
 526                 list_del_init(next);
 527                 root = list_entry(next, struct btrfs_root, dirty_list);
 528                 update_cowonly_root(trans, root);
 529                 free_extent_buffer(root->commit_root);
 530                 root->commit_root = NULL;
 531         }
 532
 533         return 0;
 534 }
 535
 536 static int __commit_transaction(struct btrfs_trans_handle *trans,
 537                                 struct btrfs_root *root)
 538 {
 539         u64 start;
 540         u64 end;
 541         struct extent_buffer *eb;
 542         struct extent_io_tree *tree = &root->fs_info->extent_cache;
 543         int ret;
 544
 545         while(1) {
 546                 ret = find_first_extent_bit(tree, 0, &start, &end,
 547                                             EXTENT_DIRTY);
 548                 if (ret)
 549                         break;
 550                 while(start <= end) {
 551                         eb = find_first_extent_buffer(tree, start);
 552                         BUG_ON(!eb || eb->start != start);
 553                         ret = write_tree_block(trans, root, eb);
 554                         BUG_ON(ret);
 555                         start += eb->len;
 556                         clear_extent_buffer_dirty(eb);
 557                         free_extent_buffer(eb);
 558                 }
 559         }
 560         return 0;
 561 }
 562
 563 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 564                              struct btrfs_root *root)
 565 {
 566         u64 transid = trans->transid;
 567         int ret = 0;
 568         struct btrfs_fs_info *fs_info = root->fs_info;
 569
 570         if (root->commit_root == root->node)
 571                 goto commit_tree;
 572         if (root == root->fs_info->tree_root)
 573                 goto commit_tree;
 574         if (root == root->fs_info->chunk_root)
 575                 goto commit_tree;
 576
 577         free_extent_buffer(root->commit_root);
 578         root->commit_root = NULL;
 579
 580         btrfs_set_root_bytenr(&root->root_item, root->node->start);
 581         btrfs_set_root_generation(&root->root_item, trans->transid);
 582         root->root_item.level = btrfs_header_level(root->node);
 583         ret = btrfs_update_root(trans, root->fs_info->tree_root,
 584                                 &root->root_key, &root->root_item);
 585         BUG_ON(ret);
 586 commit_tree:
 587         ret = commit_tree_roots(trans, fs_info);
 588         BUG_ON(ret);
 589         ret = __commit_transaction(trans, root);
 590         BUG_ON(ret);
 591         write_ctree_super(trans, root);
 592         btrfs_finish_extent_commit(trans, fs_info->extent_root,
 593                                    &fs_info->pinned_extents);
 594         btrfs_free_transaction(root, trans);
 595         free_extent_buffer(root->commit_root);
 596         root->commit_root = NULL;
 597         fs_info->running_transaction = NULL;
 598         fs_info->last_trans_committed = transid;
 599         return 0;
 600 }
 601
 602 static int find_and_setup_root(struct btrfs_root *tree_root,
 603                                struct btrfs_fs_info *fs_info,
 604                                u64 objectid, struct btrfs_root *root)
 605 {
 606         int ret;
 607         u32 blocksize;
 608         u64 generation;
 609
 610         __setup_root(tree_root->nodesize, tree_root->leafsize,
 611                      tree_root->sectorsize, tree_root->stripesize,
 612                      root, fs_info, objectid);
 613         ret = btrfs_find_last_root(tree_root, objectid,
 614                                    &root->root_item, &root->root_key);
 615         if (ret)
 616                 return ret;
 617
 618         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 619         generation = btrfs_root_generation(&root->root_item);
 620         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 621                                      blocksize, generation);
 622         if (!extent_buffer_uptodate(root->node))
 623                 return -EIO;
 624
 625         return 0;
 626 }
 627
 628 static int find_and_setup_log_root(struct btrfs_root *tree_root,
 629                                struct btrfs_fs_info *fs_info,
 630                                struct btrfs_super_block *disk_super)
 631 {
 632         u32 blocksize;
 633         u64 blocknr = btrfs_super_log_root(disk_super);
 634         struct btrfs_root *log_root = malloc(sizeof(struct btrfs_root));
 635
 636         if (!log_root)
 637                 return -ENOMEM;
 638
 639         if (blocknr == 0) {
 640                 free(log_root);
 641                 return 0;
 642         }
 643
 644         blocksize = btrfs_level_size(tree_root,
 645                              btrfs_super_log_root_level(disk_super));
 646
 647         __setup_root(tree_root->nodesize, tree_root->leafsize,
 648                      tree_root->sectorsize, tree_root->stripesize,
 649                      log_root, fs_info, BTRFS_TREE_LOG_OBJECTID);
 650
 651         log_root->node = read_tree_block(tree_root, blocknr,
 652                                      blocksize,
 653                                      btrfs_super_generation(disk_super) + 1);
 654
 655         fs_info->log_root_tree = log_root;
 656
 657         if (!extent_buffer_uptodate(log_root->node)) {
 658                 free_extent_buffer(log_root->node);
 659                 free(log_root);
 660                 fs_info->log_root_tree = NULL;
 661                 return -EIO;
 662         }
 663
 664         return 0;
 665 }
 666
 667 int btrfs_free_fs_root(struct btrfs_root *root)
 668 {
 669         if (root->node)
 670                 free_extent_buffer(root->node);
 671         if (root->commit_root)
 672                 free_extent_buffer(root->commit_root);
 673         kfree(root);
 674         return 0;
 675 }
 676
 677 static void __free_fs_root(struct rb_node *node)
 678 {
 679         struct btrfs_root *root;
 680
 681         root = container_of(node, struct btrfs_root, rb_node);
 682         btrfs_free_fs_root(root);
 683 }
 684
 685 FREE_RB_BASED_TREE(fs_roots, __free_fs_root);
 686
 687 struct btrfs_root *btrfs_read_fs_root_no_cache(struct btrfs_fs_info *fs_info,
 688                                                struct btrfs_key *location)
 689 {
 690         struct btrfs_root *root;
 691         struct btrfs_root *tree_root = fs_info->tree_root;
 692         struct btrfs_path *path;
 693         struct extent_buffer *l;
 694         u64 generation;
 695         u32 blocksize;
 696         int ret = 0;
 697
 698         root = malloc(sizeof(*root));
 699         if (!root)
 700                 return ERR_PTR(-ENOMEM);
 701         memset(root, 0, sizeof(*root));
 702         if (location->offset == (u64)-1) {
 703                 ret = find_and_setup_root(tree_root, fs_info,
 704                                           location->objectid, root);
 705                 if (ret) {
 706                         free(root);
 707                         return ERR_PTR(ret);
 708                 }
 709                 goto insert;
 710         }
 711
 712         __setup_root(tree_root->nodesize, tree_root->leafsize,
 713                      tree_root->sectorsize, tree_root->stripesize,
 714                      root, fs_info, location->objectid);
 715
 716         path = btrfs_alloc_path();
 717         BUG_ON(!path);
 718         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 719         if (ret != 0) {
 720                 if (ret > 0)
 721                         ret = -ENOENT;
 722                 goto out;
 723         }
 724         l = path->nodes[0];
 725         read_extent_buffer(l, &root->root_item,
 726                btrfs_item_ptr_offset(l, path->slots[0]),
 727                sizeof(root->root_item));
 728         memcpy(&root->root_key, location, sizeof(*location));
 729         ret = 0;
 730 out:
 731         btrfs_free_path(path);
 732         if (ret) {
 733                 free(root);
 734                 return ERR_PTR(ret);
 735         }
 736         generation = btrfs_root_generation(&root->root_item);
 737         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 738         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 739                                      blocksize, generation);
 740         if (!extent_buffer_uptodate(root->node)) {
 741                 free(root);
 742                 return ERR_PTR(-EIO);
 743         }
 744 insert:
 745         root->ref_cows = 1;
 746         return root;
 747 }
 748
 749 static int btrfs_fs_roots_compare_objectids(struct rb_node *node,
 750                                             void *data)
 751 {
 752         u64 objectid = *((u64 *)data);
 753         struct btrfs_root *root;
 754
 755         root = rb_entry(node, struct btrfs_root, rb_node);
 756         if (objectid > root->objectid)
 757                 return 1;
 758         else if (objectid < root->objectid)
 759                 return -1;
 760         else
 761                 return 0;
 762 }
 763
 764 static int btrfs_fs_roots_compare_roots(struct rb_node *node1,
 765                                         struct rb_node *node2)
 766 {
 767         struct btrfs_root *root;
 768
 769         root = rb_entry(node2, struct btrfs_root, rb_node);
 770         return btrfs_fs_roots_compare_objectids(node1, (void *)&root->objectid);
 771 }
 772
 773 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 774                                       struct btrfs_key *location)
 775 {
 776         struct btrfs_root *root;
 777         struct rb_node *node;
 778         int ret;
 779         u64 objectid = location->objectid;
 780
 781         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 782                 return fs_info->tree_root;
 783         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 784                 return fs_info->extent_root;
 785         if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID)
 786                 return fs_info->chunk_root;
 787         if (location->objectid == BTRFS_DEV_TREE_OBJECTID)
 788                 return fs_info->dev_root;
 789         if (location->objectid == BTRFS_CSUM_TREE_OBJECTID)
 790                 return fs_info->csum_root;
 791         if (location->objectid == BTRFS_QUOTA_TREE_OBJECTID)
 792                 return fs_info->quota_root;
 793
 794         BUG_ON(location->objectid == BTRFS_TREE_RELOC_OBJECTID ||
 795                location->offset != (u64)-1);
 796
 797         node = rb_search(&fs_info->fs_root_tree, (void *)&objectid,
 798                          btrfs_fs_roots_compare_objectids, NULL);
 799         if (node)
 800                 return container_of(node, struct btrfs_root, rb_node);
 801
 802         root = btrfs_read_fs_root_no_cache(fs_info, location);
 803         if (IS_ERR(root))
 804                 return root;
 805
 806         ret = rb_insert(&fs_info->fs_root_tree, &root->rb_node,
 807                         btrfs_fs_roots_compare_roots);
 808         BUG_ON(ret);
 809         return root;
 810 }
 811
 812 void btrfs_free_fs_info(struct btrfs_fs_info *fs_info)
 813 {
 814         free(fs_info->tree_root);
 815         free(fs_info->extent_root);
 816         free(fs_info->chunk_root);
 817         free(fs_info->dev_root);
 818         free(fs_info->csum_root);
 819         free(fs_info->quota_root);
 820         free(fs_info->super_copy);
 821         free(fs_info->log_root_tree);
 822         free(fs_info);
 823 }
 824
 825 struct btrfs_fs_info *btrfs_new_fs_info(int writable, u64 sb_bytenr)
 826 {
 827         struct btrfs_fs_info *fs_info;
 828
 829         fs_info = malloc(sizeof(struct btrfs_fs_info));
 830         if (!fs_info)
 831                 return NULL;
 832
 833         memset(fs_info, 0, sizeof(struct btrfs_fs_info));
 834
 835         fs_info->tree_root = malloc(sizeof(struct btrfs_root));
 836         fs_info->extent_root = malloc(sizeof(struct btrfs_root));
 837         fs_info->chunk_root = malloc(sizeof(struct btrfs_root));
 838         fs_info->dev_root = malloc(sizeof(struct btrfs_root));
 839         fs_info->csum_root = malloc(sizeof(struct btrfs_root));
 840         fs_info->quota_root = malloc(sizeof(struct btrfs_root));
 841         fs_info->super_copy = malloc(BTRFS_SUPER_INFO_SIZE);
 842
 843         if (!fs_info->tree_root || !fs_info->extent_root ||
 844             !fs_info->chunk_root || !fs_info->dev_root ||
 845             !fs_info->csum_root || !fs_info->quota_root ||
 846             !fs_info->super_copy)
 847                 goto free_all;
 848
 849         memset(fs_info->super_copy, 0, BTRFS_SUPER_INFO_SIZE);
 850         memset(fs_info->tree_root, 0, sizeof(struct btrfs_root));
 851         memset(fs_info->extent_root, 0, sizeof(struct btrfs_root));
 852         memset(fs_info->chunk_root, 0, sizeof(struct btrfs_root));
 853         memset(fs_info->dev_root, 0, sizeof(struct btrfs_root));
 854         memset(fs_info->csum_root, 0, sizeof(struct btrfs_root));
 855         memset(fs_info->quota_root, 0, sizeof(struct btrfs_root));
 856
 857         extent_io_tree_init(&fs_info->extent_cache);
 858         extent_io_tree_init(&fs_info->free_space_cache);
 859         extent_io_tree_init(&fs_info->block_group_cache);
 860         extent_io_tree_init(&fs_info->pinned_extents);
 861         extent_io_tree_init(&fs_info->pending_del);
 862         extent_io_tree_init(&fs_info->extent_ins);
 863         fs_info->excluded_extents = NULL;
 864
 865         fs_info->fs_root_tree = RB_ROOT;
 866         cache_tree_init(&fs_info->mapping_tree.cache_tree);
 867
 868         mutex_init(&fs_info->fs_mutex);
 869         INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
 870         INIT_LIST_HEAD(&fs_info->space_info);
 871         INIT_LIST_HEAD(&fs_info->recow_ebs);
 872
 873         if (!writable)
 874                 fs_info->readonly = 1;
 875
 876         fs_info->super_bytenr = sb_bytenr;
 877         fs_info->data_alloc_profile = (u64)-1;
 878         fs_info->metadata_alloc_profile = (u64)-1;
 879         fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
 880         return fs_info;
 881 free_all:
 882         btrfs_free_fs_info(fs_info);
 883         return NULL;
 884 }
 885
 886 int btrfs_check_fs_compatibility(struct btrfs_super_block *sb, int writable)
 887 {
 888         u64 features;
 889
 890         features = btrfs_super_incompat_flags(sb) &
 891                    ~BTRFS_FEATURE_INCOMPAT_SUPP;
 892         if (features) {
 893                 printk("couldn't open because of unsupported "
 894                        "option features (%Lx).\n",
 895                        (unsigned long long)features);
 896                 return -ENOTSUP;
 897         }
 898
 899         features = btrfs_super_incompat_flags(sb);
 900         if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) {
 901                 features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF;
 902                 btrfs_set_super_incompat_flags(sb, features);
 903         }
 904
 905         features = btrfs_super_compat_ro_flags(sb) &
 906                 ~BTRFS_FEATURE_COMPAT_RO_SUPP;
 907         if (writable && features) {
 908                 printk("couldn't open RDWR because of unsupported "
 909                        "option features (%Lx).\n",
 910                        (unsigned long long)features);
 911                 return -ENOTSUP;
 912         }
 913         return 0;
 914 }
 915
 916 static int find_best_backup_root(struct btrfs_super_block *super)
 917 {
 918         struct btrfs_root_backup *backup;
 919         u64 orig_gen = btrfs_super_generation(super);
 920         u64 gen = 0;
 921         int best_index = 0;
 922         int i;
 923
 924         for (i = 0; i < BTRFS_NUM_BACKUP_ROOTS; i++) {
 925                 backup = super->super_roots + i;
 926                 if (btrfs_backup_tree_root_gen(backup) != orig_gen &&
 927                     btrfs_backup_tree_root_gen(backup) > gen) {
 928                         best_index = i;
 929                         gen = btrfs_backup_tree_root_gen(backup);
 930                 }
 931         }
 932         return best_index;
 933 }
 934
 935 static int setup_root_or_create_block(struct btrfs_fs_info *fs_info,
 936                                       enum btrfs_open_ctree_flags flags,
 937                                       struct btrfs_root *info_root,
 938                                       u64 objectid, char *str)
 939 {
 940         struct btrfs_super_block *sb = fs_info->super_copy;
 941         struct btrfs_root *root = fs_info->tree_root;
 942         u32 leafsize = btrfs_super_leafsize(sb);
 943         int ret;
 944
 945         ret = find_and_setup_root(root, fs_info, objectid, info_root);
 946         if (ret) {
 947                 printk("Couldn't setup %s tree\n", str);
 948                 if (!(flags & OPEN_CTREE_PARTIAL))
 949                         return -EIO;
 950                 /*
 951                  * Need a blank node here just so we don't screw up in the
 952                  * million of places that assume a root has a valid ->node
 953                  */
 954                 info_root->node =
 955                         btrfs_find_create_tree_block(info_root, 0, leafsize);
 956                 if (!info_root->node)
 957                         return -ENOMEM;
 958                 clear_extent_buffer_uptodate(NULL, info_root->node);
 959         }
 960
 961         return 0;
 962 }
 963
 964 int btrfs_setup_all_roots(struct btrfs_fs_info *fs_info, u64 root_tree_bytenr,
 965                           enum btrfs_open_ctree_flags flags)
 966 {
 967         struct btrfs_super_block *sb = fs_info->super_copy;
 968         struct btrfs_root *root;
 969         struct btrfs_key key;
 970         u32 sectorsize;
 971         u32 nodesize;
 972         u32 leafsize;
 973         u32 stripesize;
 974         u64 generation;
 975         u32 blocksize;
 976         int ret;
 977
 978         nodesize = btrfs_super_nodesize(sb);
 979         leafsize = btrfs_super_leafsize(sb);
 980         sectorsize = btrfs_super_sectorsize(sb);
 981         stripesize = btrfs_super_stripesize(sb);
 982
 983         root = fs_info->tree_root;
 984         __setup_root(nodesize, leafsize, sectorsize, stripesize,
 985                      root, fs_info, BTRFS_ROOT_TREE_OBJECTID);
 986         blocksize = btrfs_level_size(root, btrfs_super_root_level(sb));
 987         generation = btrfs_super_generation(sb);
 988
 989         if (!root_tree_bytenr && !(flags & OPEN_CTREE_BACKUP_ROOT)) {
 990                 root_tree_bytenr = btrfs_super_root(sb);
 991         } else if (flags & OPEN_CTREE_BACKUP_ROOT) {
 992                 struct btrfs_root_backup *backup;
 993                 int index = find_best_backup_root(sb);
 994                 if (index >= BTRFS_NUM_BACKUP_ROOTS) {
 995                         fprintf(stderr, "Invalid backup root number\n");
 996                         return -EIO;
 997                 }
 998                 backup = fs_info->super_copy->super_roots + index;
 999                 root_tree_bytenr = btrfs_backup_tree_root(backup);
1000                 generation = btrfs_backup_tree_root_gen(backup);
1001         }
1002
1003         root->node = read_tree_block(root, root_tree_bytenr, blocksize,
1004                                      generation);
1005         if (!extent_buffer_uptodate(root->node)) {
1006                 fprintf(stderr, "Couldn't read tree root\n");
1007                 return -EIO;
1008         }
1009
1010         ret = setup_root_or_create_block(fs_info, flags, fs_info->extent_root,
1011                                          BTRFS_EXTENT_TREE_OBJECTID, "extent");
1012         if (ret)
1013                 return ret;
1014         fs_info->extent_root->track_dirty = 1;
1015
1016         ret = find_and_setup_root(root, fs_info, BTRFS_DEV_TREE_OBJECTID,
1017                                   fs_info->dev_root);
1018         if (ret) {
1019                 printk("Couldn't setup device tree\n");
1020                 return -EIO;
1021         }
1022         fs_info->dev_root->track_dirty = 1;
1023
1024         ret = setup_root_or_create_block(fs_info, flags, fs_info->csum_root,
1025                                          BTRFS_CSUM_TREE_OBJECTID, "csum");
1026         if (ret)
1027                 return ret;
1028         fs_info->csum_root->track_dirty = 1;
1029
1030         ret = find_and_setup_root(root, fs_info, BTRFS_QUOTA_TREE_OBJECTID,
1031                                   fs_info->quota_root);
1032         if (ret == 0)
1033                 fs_info->quota_enabled = 1;
1034
1035         ret = find_and_setup_log_root(root, fs_info, sb);
1036         if (ret) {
1037                 printk("Couldn't setup log root tree\n");
1038                 if (!(flags & OPEN_CTREE_PARTIAL))
1039                         return -EIO;
1040         }
1041
1042         fs_info->generation = generation;
1043         fs_info->last_trans_committed = generation;
1044         if (extent_buffer_uptodate(fs_info->extent_root->node) &&
1045             !(flags & OPEN_CTREE_NO_BLOCK_GROUPS))
1046                 btrfs_read_block_groups(fs_info->tree_root);
1047
1048         key.objectid = BTRFS_FS_TREE_OBJECTID;
1049         key.type = BTRFS_ROOT_ITEM_KEY;
1050         key.offset = (u64)-1;
1051         fs_info->fs_root = btrfs_read_fs_root(fs_info, &key);
1052
1053         if (IS_ERR(fs_info->fs_root))
1054                 return -EIO;
1055         return 0;
1056 }
1057
1058 void btrfs_release_all_roots(struct btrfs_fs_info *fs_info)
1059 {
1060         if (fs_info->quota_root)
1061                 free_extent_buffer(fs_info->quota_root->node);
1062         if (fs_info->csum_root)
1063                 free_extent_buffer(fs_info->csum_root->node);
1064         if (fs_info->dev_root)
1065                 free_extent_buffer(fs_info->dev_root->node);
1066         if (fs_info->extent_root)
1067                 free_extent_buffer(fs_info->extent_root->node);
1068         if (fs_info->tree_root)
1069                 free_extent_buffer(fs_info->tree_root->node);
1070         if (fs_info->log_root_tree)
1071                 free_extent_buffer(fs_info->log_root_tree->node);
1072         if (fs_info->chunk_root)
1073                 free_extent_buffer(fs_info->chunk_root->node);
1074 }
1075
1076 static void free_map_lookup(struct cache_extent *ce)
1077 {
1078         struct map_lookup *map;
1079
1080         map = container_of(ce, struct map_lookup, ce);
1081         kfree(map);
1082 }
1083
1084 FREE_EXTENT_CACHE_BASED_TREE(mapping_cache, free_map_lookup);
1085
1086 void btrfs_cleanup_all_caches(struct btrfs_fs_info *fs_info)
1087 {
1088         while (!list_empty(&fs_info->recow_ebs)) {
1089                 struct extent_buffer *eb;
1090                 eb = list_first_entry(&fs_info->recow_ebs,
1091                                       struct extent_buffer, recow);
1092                 list_del_init(&eb->recow);
1093                 free_extent_buffer(eb);
1094         }
1095         free_mapping_cache_tree(&fs_info->mapping_tree.cache_tree);
1096         extent_io_tree_cleanup(&fs_info->extent_cache);
1097         extent_io_tree_cleanup(&fs_info->free_space_cache);
1098         extent_io_tree_cleanup(&fs_info->block_group_cache);
1099         extent_io_tree_cleanup(&fs_info->pinned_extents);
1100         extent_io_tree_cleanup(&fs_info->pending_del);
1101         extent_io_tree_cleanup(&fs_info->extent_ins);
1102 }
1103
1104 int btrfs_scan_fs_devices(int fd, const char *path,
1105                           struct btrfs_fs_devices **fs_devices,
1106                           u64 sb_bytenr, int super_recover,
1107                           int skip_devices)
1108 {
1109         u64 total_devs;
1110         u64 dev_size;
1111         off_t seek_ret;
1112         int ret;
1113         if (!sb_bytenr)
1114                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1115
1116         seek_ret = lseek(fd, 0, SEEK_END);
1117         if (seek_ret < 0)
1118                 return -errno;
1119
1120         dev_size = seek_ret;
1121         lseek(fd, 0, SEEK_SET);
1122         if (sb_bytenr > dev_size) {
1123                 fprintf(stderr, "Superblock bytenr is larger than device size\n");
1124                 return -EINVAL;
1125         }
1126
1127         ret = btrfs_scan_one_device(fd, path, fs_devices,
1128                                     &total_devs, sb_bytenr, super_recover);
1129         if (ret) {
1130                 fprintf(stderr, "No valid Btrfs found on %s\n", path);
1131                 return ret;
1132         }
1133
1134         if (!skip_devices && total_devs != 1) {
1135                 ret = btrfs_scan_lblkid();
1136                 if (ret)
1137                         return ret;
1138         }
1139         return 0;
1140 }
1141
1142 int btrfs_setup_chunk_tree_and_device_map(struct btrfs_fs_info *fs_info)
1143 {
1144         struct btrfs_super_block *sb = fs_info->super_copy;
1145         u32 sectorsize;
1146         u32 nodesize;
1147         u32 leafsize;
1148         u32 blocksize;
1149         u32 stripesize;
1150         u64 generation;
1151         int ret;
1152
1153         nodesize = btrfs_super_nodesize(sb);
1154         leafsize = btrfs_super_leafsize(sb);
1155         sectorsize = btrfs_super_sectorsize(sb);
1156         stripesize = btrfs_super_stripesize(sb);
1157
1158         __setup_root(nodesize, leafsize, sectorsize, stripesize,
1159                      fs_info->chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID);
1160
1161         ret = btrfs_read_sys_array(fs_info->chunk_root);
1162         if (ret)
1163                 return ret;
1164
1165         blocksize = btrfs_level_size(fs_info->chunk_root,
1166                                      btrfs_super_chunk_root_level(sb));
1167         generation = btrfs_super_chunk_root_generation(sb);
1168
1169         fs_info->chunk_root->node = read_tree_block(fs_info->chunk_root,
1170                                                     btrfs_super_chunk_root(sb),
1171                                                     blocksize, generation);
1172         if (!extent_buffer_uptodate(fs_info->chunk_root->node)) {
1173                 fprintf(stderr, "Couldn't read chunk root\n");
1174                 return -EIO;
1175         }
1176
1177         if (!(btrfs_super_flags(sb) & BTRFS_SUPER_FLAG_METADUMP)) {
1178                 ret = btrfs_read_chunk_tree(fs_info->chunk_root);
1179                 if (ret) {
1180                         fprintf(stderr, "Couldn't read chunk tree\n");
1181                         return ret;
1182                 }
1183         }
1184         return 0;
1185 }
1186
1187 static struct btrfs_fs_info *__open_ctree_fd(int fp, const char *path,
1188                                              u64 sb_bytenr,
1189                                              u64 root_tree_bytenr,
1190                                              enum btrfs_open_ctree_flags flags)
1191 {
1192         struct btrfs_fs_info *fs_info;
1193         struct btrfs_super_block *disk_super;
1194         struct btrfs_fs_devices *fs_devices = NULL;
1195         struct extent_buffer *eb;
1196         int ret;
1197         int oflags;
1198
1199         if (sb_bytenr == 0)
1200                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1201
1202         /* try to drop all the caches */
1203         if (posix_fadvise(fp, 0, 0, POSIX_FADV_DONTNEED))
1204                 fprintf(stderr, "Warning, could not drop caches\n");
1205
1206         fs_info = btrfs_new_fs_info(flags & OPEN_CTREE_WRITES, sb_bytenr);
1207         if (!fs_info) {
1208                 fprintf(stderr, "Failed to allocate memory for fs_info\n");
1209                 return NULL;
1210         }
1211         if (flags & OPEN_CTREE_RESTORE)
1212                 fs_info->on_restoring = 1;
1213         if (flags & OPEN_CTREE_SUPPRESS_CHECK_BLOCK_ERRORS)
1214                 fs_info->suppress_check_block_errors = 1;
1215         if (flags & OPEN_CTREE_IGNORE_FSID_MISMATCH)
1216                 fs_info->ignore_fsid_mismatch = 1;
1217
1218         ret = btrfs_scan_fs_devices(fp, path, &fs_devices, sb_bytenr,
1219                                     (flags & OPEN_CTREE_RECOVER_SUPER),
1220                                     (flags & OPEN_CTREE_NO_DEVICES));
1221         if (ret)
1222                 goto out;
1223
1224         fs_info->fs_devices = fs_devices;
1225         if (flags & OPEN_CTREE_WRITES)
1226                 oflags = O_RDWR;
1227         else
1228                 oflags = O_RDONLY;
1229
1230         if (flags & OPEN_CTREE_EXCLUSIVE)
1231                 oflags |= O_EXCL;
1232
1233         ret = btrfs_open_devices(fs_devices, oflags);
1234         if (ret)
1235                 goto out;
1236
1237         disk_super = fs_info->super_copy;
1238         if (!(flags & OPEN_CTREE_RECOVER_SUPER))
1239                 ret = btrfs_read_dev_super(fs_devices->latest_bdev,
1240                                            disk_super, sb_bytenr, 1);
1241         else
1242                 ret = btrfs_read_dev_super(fp, disk_super, sb_bytenr, 0);
1243         if (ret) {
1244                 printk("No valid btrfs found\n");
1245                 goto out_devices;
1246         }
1247
1248         if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_CHANGING_FSID &&
1249             !fs_info->ignore_fsid_mismatch) {
1250                 fprintf(stderr, "ERROR: Filesystem UUID change in progress\n");
1251                 goto out_devices;
1252         }
1253
1254         memcpy(fs_info->fsid, &disk_super->fsid, BTRFS_FSID_SIZE);
1255
1256         ret = btrfs_check_fs_compatibility(fs_info->super_copy,
1257                                            flags & OPEN_CTREE_WRITES);
1258         if (ret)
1259                 goto out_devices;
1260
1261         ret = btrfs_setup_chunk_tree_and_device_map(fs_info);
1262         if (ret)
1263                 goto out_chunk;
1264
1265         eb = fs_info->chunk_root->node;
1266         read_extent_buffer(eb, fs_info->chunk_tree_uuid,
1267                            btrfs_header_chunk_tree_uuid(eb),
1268                            BTRFS_UUID_SIZE);
1269
1270         ret = btrfs_setup_all_roots(fs_info, root_tree_bytenr, flags);
1271         if (ret && !(flags & __OPEN_CTREE_RETURN_CHUNK_ROOT))
1272                 goto out_chunk;
1273
1274         return fs_info;
1275
1276 out_chunk:
1277         btrfs_release_all_roots(fs_info);
1278         btrfs_cleanup_all_caches(fs_info);
1279 out_devices:
1280         btrfs_close_devices(fs_devices);
1281 out:
1282         btrfs_free_fs_info(fs_info);
1283         return NULL;
1284 }
1285
1286 struct btrfs_fs_info *open_ctree_fs_info(const char *filename,
1287                                          u64 sb_bytenr, u64 root_tree_bytenr,
1288                                          enum btrfs_open_ctree_flags flags)
1289 {
1290         int fp;
1291         struct btrfs_fs_info *info;
1292         int oflags = O_CREAT | O_RDWR;
1293
1294         if (!(flags & OPEN_CTREE_WRITES))
1295                 oflags = O_RDONLY;
1296
1297         fp = open(filename, oflags, 0600);
1298         if (fp < 0) {
1299                 fprintf (stderr, "Could not open %s\n", filename);
1300                 return NULL;
1301         }
1302         info = __open_ctree_fd(fp, filename, sb_bytenr, root_tree_bytenr,
1303                                flags);
1304         close(fp);
1305         return info;
1306 }
1307
1308 struct btrfs_root *open_ctree(const char *filename, u64 sb_bytenr,
1309                               enum btrfs_open_ctree_flags flags)
1310 {
1311         struct btrfs_fs_info *info;
1312
1313         info = open_ctree_fs_info(filename, sb_bytenr, 0, flags);
1314         if (!info)
1315                 return NULL;
1316         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1317                 return info->chunk_root;
1318         return info->fs_root;
1319 }
1320
1321 struct btrfs_root *open_ctree_fd(int fp, const char *path, u64 sb_bytenr,
1322                                  enum btrfs_open_ctree_flags flags)
1323 {
1324         struct btrfs_fs_info *info;
1325         info = __open_ctree_fd(fp, path, sb_bytenr, 0, flags);
1326         if (!info)
1327                 return NULL;
1328         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1329                 return info->chunk_root;
1330         return info->fs_root;
1331 }
1332
1333 int btrfs_read_dev_super(int fd, struct btrfs_super_block *sb, u64 sb_bytenr,
1334                          int super_recover)
1335 {
1336         u8 fsid[BTRFS_FSID_SIZE];
1337         int fsid_is_initialized = 0;
1338         struct btrfs_super_block buf;
1339         int i;
1340         int ret;
1341         int max_super = super_recover ? BTRFS_SUPER_MIRROR_MAX : 1;
1342         u64 transid = 0;
1343         u64 bytenr;
1344
1345         if (sb_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1346                 ret = pread64(fd, &buf, sizeof(buf), sb_bytenr);
1347                 if (ret < sizeof(buf))
1348                         return -1;
1349
1350                 if (btrfs_super_bytenr(&buf) != sb_bytenr ||
1351                     btrfs_super_magic(&buf) != BTRFS_MAGIC)
1352                         return -1;
1353
1354                 memcpy(sb, &buf, sizeof(*sb));
1355                 return 0;
1356         }
1357
1358         /*
1359         * we would like to check all the supers, but that would make
1360         * a btrfs mount succeed after a mkfs from a different FS.
1361         * So, we need to add a special mount option to scan for
1362         * later supers, using BTRFS_SUPER_MIRROR_MAX instead
1363         */
1364
1365         for (i = 0; i < max_super; i++) {
1366                 bytenr = btrfs_sb_offset(i);
1367                 ret = pread64(fd, &buf, sizeof(buf), bytenr);
1368                 if (ret < sizeof(buf))
1369                         break;
1370
1371                 if (btrfs_super_bytenr(&buf) != bytenr )
1372                         continue;
1373                 /* if magic is NULL, the device was removed */
1374                 if (btrfs_super_magic(&buf) == 0 && i == 0)
1375                         return -1;
1376                 if (btrfs_super_magic(&buf) != BTRFS_MAGIC)
1377                         continue;
1378
1379                 if (!fsid_is_initialized) {
1380                         memcpy(fsid, buf.fsid, sizeof(fsid));
1381                         fsid_is_initialized = 1;
1382                 } else if (memcmp(fsid, buf.fsid, sizeof(fsid))) {
1383                         /*
1384                          * the superblocks (the original one and
1385                          * its backups) contain data of different
1386                          * filesystems -> the super cannot be trusted
1387                          */
1388                         continue;
1389                 }
1390
1391                 if (btrfs_super_generation(&buf) > transid) {
1392                         memcpy(sb, &buf, sizeof(*sb));
1393                         transid = btrfs_super_generation(&buf);
1394                 }
1395         }
1396
1397         return transid > 0 ? 0 : -1;
1398 }
1399
1400 static int write_dev_supers(struct btrfs_root *root,
1401                             struct btrfs_super_block *sb,
1402                             struct btrfs_device *device)
1403 {
1404         u64 bytenr;
1405         u32 crc;
1406         int i, ret;
1407
1408         if (root->fs_info->super_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1409                 btrfs_set_super_bytenr(sb, root->fs_info->super_bytenr);
1410                 crc = ~(u32)0;
1411                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1412                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1413                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1414
1415                 /*
1416                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1417                  * zero filled, we can use it directly
1418                  */
1419                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1420                                 BTRFS_SUPER_INFO_SIZE,
1421                                 root->fs_info->super_bytenr);
1422                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1423                 return 0;
1424         }
1425
1426         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1427                 bytenr = btrfs_sb_offset(i);
1428                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
1429                         break;
1430
1431                 btrfs_set_super_bytenr(sb, bytenr);
1432
1433                 crc = ~(u32)0;
1434                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1435                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1436                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1437
1438                 /*
1439                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1440                  * zero filled, we can use it directly
1441                  */
1442                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1443                                 BTRFS_SUPER_INFO_SIZE, bytenr);
1444                 BUG_ON(ret != BTRFS_SUPER_INFO_SIZE);
1445         }
1446
1447         return 0;
1448 }
1449
1450 int write_all_supers(struct btrfs_root *root)
1451 {
1452         struct list_head *cur;
1453         struct list_head *head = &root->fs_info->fs_devices->devices;
1454         struct btrfs_device *dev;
1455         struct btrfs_super_block *sb;
1456         struct btrfs_dev_item *dev_item;
1457         int ret;
1458         u64 flags;
1459
1460         sb = root->fs_info->super_copy;
1461         dev_item = &sb->dev_item;
1462         list_for_each(cur, head) {
1463                 dev = list_entry(cur, struct btrfs_device, dev_list);
1464                 if (!dev->writeable)
1465                         continue;
1466
1467                 btrfs_set_stack_device_generation(dev_item, 0);
1468                 btrfs_set_stack_device_type(dev_item, dev->type);
1469                 btrfs_set_stack_device_id(dev_item, dev->devid);
1470                 btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes);
1471                 btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used);
1472                 btrfs_set_stack_device_io_align(dev_item, dev->io_align);
1473                 btrfs_set_stack_device_io_width(dev_item, dev->io_width);
1474                 btrfs_set_stack_device_sector_size(dev_item, dev->sector_size);
1475                 memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE);
1476                 memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE);
1477
1478                 flags = btrfs_super_flags(sb);
1479                 btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN);
1480
1481                 ret = write_dev_supers(root, sb, dev);
1482                 BUG_ON(ret);
1483         }
1484         return 0;
1485 }
1486
1487 int write_ctree_super(struct btrfs_trans_handle *trans,
1488                       struct btrfs_root *root)
1489 {
1490         int ret;
1491         struct btrfs_root *tree_root = root->fs_info->tree_root;
1492         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
1493
1494         if (root->fs_info->readonly)
1495                 return 0;
1496
1497         btrfs_set_super_generation(root->fs_info->super_copy,
1498                                    trans->transid);
1499         btrfs_set_super_root(root->fs_info->super_copy,
1500                              tree_root->node->start);
1501         btrfs_set_super_root_level(root->fs_info->super_copy,
1502                                    btrfs_header_level(tree_root->node));
1503         btrfs_set_super_chunk_root(root->fs_info->super_copy,
1504                                    chunk_root->node->start);
1505         btrfs_set_super_chunk_root_level(root->fs_info->super_copy,
1506                                          btrfs_header_level(chunk_root->node));
1507         btrfs_set_super_chunk_root_generation(root->fs_info->super_copy,
1508                                 btrfs_header_generation(chunk_root->node));
1509
1510         ret = write_all_supers(root);
1511         if (ret)
1512                 fprintf(stderr, "failed to write new super block err %d\n", ret);
1513         return ret;
1514 }
1515
1516 int close_ctree(struct btrfs_root *root)
1517 {
1518         int ret;
1519         struct btrfs_trans_handle *trans;
1520         struct btrfs_fs_info *fs_info = root->fs_info;
1521
1522         if (fs_info->last_trans_committed !=
1523             fs_info->generation) {
1524                 trans = btrfs_start_transaction(root, 1);
1525                 btrfs_commit_transaction(trans, root);
1526                 trans = btrfs_start_transaction(root, 1);
1527                 ret = commit_tree_roots(trans, fs_info);
1528                 BUG_ON(ret);
1529                 ret = __commit_transaction(trans, root);
1530                 BUG_ON(ret);
1531                 write_ctree_super(trans, root);
1532                 btrfs_free_transaction(root, trans);
1533         }
1534         btrfs_free_block_groups(fs_info);
1535
1536         free_fs_roots_tree(&fs_info->fs_root_tree);
1537
1538         btrfs_release_all_roots(fs_info);
1539         btrfs_close_devices(fs_info->fs_devices);
1540         btrfs_cleanup_all_caches(fs_info);
1541         btrfs_free_fs_info(fs_info);
1542         return 0;
1543 }
1544
1545 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1546                      struct extent_buffer *eb)
1547 {
1548         return clear_extent_buffer_dirty(eb);
1549 }
1550
1551 int wait_on_tree_block_writeback(struct btrfs_root *root,
1552                                  struct extent_buffer *eb)
1553 {
1554         return 0;
1555 }
1556
1557 void btrfs_mark_buffer_dirty(struct extent_buffer *eb)
1558 {
1559         set_extent_buffer_dirty(eb);
1560 }
1561
1562 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid)
1563 {
1564         int ret;
1565
1566         ret = extent_buffer_uptodate(buf);
1567         if (!ret)
1568                 return ret;
1569
1570         ret = verify_parent_transid(buf->tree, buf, parent_transid, 1);
1571         return !ret;
1572 }
1573
1574 int btrfs_set_buffer_uptodate(struct extent_buffer *eb)
1575 {
1576         return set_extent_buffer_uptodate(eb);
1577 }