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 #define IS_ALIGNED(x, a)                (((x) & ((typeof(x))(a) - 1)) == 0)
  44
  45 /* Calculate max possible nritems for a leaf/node */
  46 static u32 max_nritems(u8 level, u32 nodesize)
  47 {
  48
  49         if (level == 0)
  50                 return ((nodesize - sizeof(struct btrfs_header)) /
  51                         sizeof(struct btrfs_item));
  52         return ((nodesize - sizeof(struct btrfs_header)) /
  53                 sizeof(struct btrfs_key_ptr));
  54 }
  55
  56 static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
  57 {
  58
  59         struct btrfs_fs_devices *fs_devices;
  60         int ret = BTRFS_BAD_FSID;
  61
  62         if (buf->start != btrfs_header_bytenr(buf))
  63                 return BTRFS_BAD_BYTENR;
  64         if (btrfs_header_level(buf) >= BTRFS_MAX_LEVEL)
  65                 return BTRFS_BAD_LEVEL;
  66         if (btrfs_header_nritems(buf) > max_nritems(btrfs_header_level(buf),
  67                                                     root->nodesize))
  68                 return BTRFS_BAD_NRITEMS;
  69
  70         fs_devices = root->fs_info->fs_devices;
  71         while (fs_devices) {
  72                 if (root->fs_info->ignore_fsid_mismatch ||
  73                     !memcmp_extent_buffer(buf, fs_devices->fsid,
  74                                           btrfs_header_fsid(),
  75                                           BTRFS_FSID_SIZE)) {
  76                         ret = 0;
  77                         break;
  78                 }
  79                 fs_devices = fs_devices->seed;
  80         }
  81         return ret;
  82 }
  83
  84 static void print_tree_block_error(struct btrfs_root *root,
  85                                 struct extent_buffer *eb,
  86                                 int err)
  87 {
  88         char fs_uuid[BTRFS_UUID_UNPARSED_SIZE] = {'\0'};
  89         char found_uuid[BTRFS_UUID_UNPARSED_SIZE] = {'\0'};
  90         u8 buf[BTRFS_UUID_SIZE];
  91
  92         switch (err) {
  93         case BTRFS_BAD_FSID:
  94                 read_extent_buffer(eb, buf, btrfs_header_fsid(),
  95                                    BTRFS_UUID_SIZE);
  96                 uuid_unparse(buf, found_uuid);
  97                 uuid_unparse(root->fs_info->fsid, fs_uuid);
  98                 fprintf(stderr, "fsid mismatch, want=%s, have=%s\n",
  99                         fs_uuid, found_uuid);
 100                 break;
 101         case BTRFS_BAD_BYTENR:
 102                 fprintf(stderr, "bytenr mismatch, want=%llu, have=%llu\n",
 103                         eb->start, btrfs_header_bytenr(eb));
 104                 break;
 105         case BTRFS_BAD_LEVEL:
 106                 fprintf(stderr, "bad level, %u > %u\n",
 107                         btrfs_header_level(eb), BTRFS_MAX_LEVEL);
 108                 break;
 109         case BTRFS_BAD_NRITEMS:
 110                 fprintf(stderr, "invalid nr_items: %u\n",
 111                         btrfs_header_nritems(eb));
 112                 break;
 113         }
 114 }
 115
 116 u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len)
 117 {
 118         return crc32c(seed, data, len);
 119 }
 120
 121 void btrfs_csum_final(u32 crc, char *result)
 122 {
 123         *(__le32 *)result = ~cpu_to_le32(crc);
 124 }
 125
 126 static int __csum_tree_block_size(struct extent_buffer *buf, u16 csum_size,
 127                                   int verify, int silent)
 128 {
 129         char *result;
 130         u32 len;
 131         u32 crc = ~(u32)0;
 132
 133         result = malloc(csum_size * sizeof(char));
 134         if (!result)
 135                 return 1;
 136
 137         len = buf->len - BTRFS_CSUM_SIZE;
 138         crc = crc32c(crc, buf->data + BTRFS_CSUM_SIZE, len);
 139         btrfs_csum_final(crc, result);
 140
 141         if (verify) {
 142                 if (memcmp_extent_buffer(buf, result, 0, csum_size)) {
 143                         if (!silent)
 144                                 printk("checksum verify failed on %llu found %08X wanted %08X\n",
 145                                        (unsigned long long)buf->start,
 146                                        *((u32 *)result),
 147                                        *((u32*)(char *)buf->data));
 148                         free(result);
 149                         return 1;
 150                 }
 151         } else {
 152                 write_extent_buffer(buf, result, 0, csum_size);
 153         }
 154         free(result);
 155         return 0;
 156 }
 157
 158 int csum_tree_block_size(struct extent_buffer *buf, u16 csum_size, int verify)
 159 {
 160         return __csum_tree_block_size(buf, csum_size, verify, 0);
 161 }
 162
 163 int verify_tree_block_csum_silent(struct extent_buffer *buf, u16 csum_size)
 164 {
 165         return __csum_tree_block_size(buf, csum_size, 1, 1);
 166 }
 167
 168 int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf,
 169                            int verify)
 170 {
 171         u16 csum_size =
 172                 btrfs_super_csum_size(root->fs_info->super_copy);
 173         if (verify && root->fs_info->suppress_check_block_errors)
 174                 return verify_tree_block_csum_silent(buf, csum_size);
 175         return csum_tree_block_size(buf, csum_size, verify);
 176 }
 177
 178 struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
 179                                             u64 bytenr, u32 blocksize)
 180 {
 181         return find_extent_buffer(&root->fs_info->extent_cache,
 182                                   bytenr, blocksize);
 183 }
 184
 185 struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
 186                                                  u64 bytenr, u32 blocksize)
 187 {
 188         return alloc_extent_buffer(&root->fs_info->extent_cache, bytenr,
 189                                    blocksize);
 190 }
 191
 192 void readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
 193                           u64 parent_transid)
 194 {
 195         struct extent_buffer *eb;
 196         u64 length;
 197         struct btrfs_multi_bio *multi = NULL;
 198         struct btrfs_device *device;
 199
 200         eb = btrfs_find_tree_block(root, bytenr, blocksize);
 201         if (!(eb && btrfs_buffer_uptodate(eb, parent_transid)) &&
 202             !btrfs_map_block(&root->fs_info->mapping_tree, READ,
 203                              bytenr, &length, &multi, 0, NULL)) {
 204                 device = multi->stripes[0].dev;
 205                 device->total_ios++;
 206                 blocksize = min(blocksize, (u32)(64 * 1024));
 207                 readahead(device->fd, multi->stripes[0].physical, blocksize);
 208         }
 209
 210         free_extent_buffer(eb);
 211         kfree(multi);
 212 }
 213
 214 static int verify_parent_transid(struct extent_io_tree *io_tree,
 215                                  struct extent_buffer *eb, u64 parent_transid,
 216                                  int ignore)
 217 {
 218         int ret;
 219
 220         if (!parent_transid || btrfs_header_generation(eb) == parent_transid)
 221                 return 0;
 222
 223         if (extent_buffer_uptodate(eb) &&
 224             btrfs_header_generation(eb) == parent_transid) {
 225                 ret = 0;
 226                 goto out;
 227         }
 228         printk("parent transid verify failed on %llu wanted %llu found %llu\n",
 229                (unsigned long long)eb->start,
 230                (unsigned long long)parent_transid,
 231                (unsigned long long)btrfs_header_generation(eb));
 232         if (ignore) {
 233                 eb->flags |= EXTENT_BAD_TRANSID;
 234                 printk("Ignoring transid failure\n");
 235                 return 0;
 236         }
 237
 238         ret = 1;
 239 out:
 240         clear_extent_buffer_uptodate(io_tree, eb);
 241         return ret;
 242
 243 }
 244
 245
 246 int read_whole_eb(struct btrfs_fs_info *info, struct extent_buffer *eb, int mirror)
 247 {
 248         unsigned long offset = 0;
 249         struct btrfs_multi_bio *multi = NULL;
 250         struct btrfs_device *device;
 251         int ret = 0;
 252         u64 read_len;
 253         unsigned long bytes_left = eb->len;
 254
 255         while (bytes_left) {
 256                 read_len = bytes_left;
 257                 device = NULL;
 258
 259                 if (!info->on_restoring &&
 260                     eb->start != BTRFS_SUPER_INFO_OFFSET) {
 261                         ret = btrfs_map_block(&info->mapping_tree, READ,
 262                                               eb->start + offset, &read_len, &multi,
 263                                               mirror, NULL);
 264                         if (ret) {
 265                                 printk("Couldn't map the block %Lu\n", eb->start + offset);
 266                                 kfree(multi);
 267                                 return -EIO;
 268                         }
 269                         device = multi->stripes[0].dev;
 270
 271                         if (device->fd <= 0) {
 272                                 kfree(multi);
 273                                 return -EIO;
 274                         }
 275
 276                         eb->fd = device->fd;
 277                         device->total_ios++;
 278                         eb->dev_bytenr = multi->stripes[0].physical;
 279                         kfree(multi);
 280                         multi = NULL;
 281                 } else {
 282                         /* special case for restore metadump */
 283                         list_for_each_entry(device, &info->fs_devices->devices, dev_list) {
 284                                 if (device->devid == 1)
 285                                         break;
 286                         }
 287
 288                         eb->fd = device->fd;
 289                         eb->dev_bytenr = eb->start;
 290                         device->total_ios++;
 291                 }
 292
 293                 if (read_len > bytes_left)
 294                         read_len = bytes_left;
 295
 296                 ret = read_extent_from_disk(eb, offset, read_len);
 297                 if (ret)
 298                         return -EIO;
 299                 offset += read_len;
 300                 bytes_left -= read_len;
 301         }
 302         return 0;
 303 }
 304
 305 struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 306                                      u32 blocksize, u64 parent_transid)
 307 {
 308         int ret;
 309         struct extent_buffer *eb;
 310         u64 best_transid = 0;
 311         int mirror_num = 0;
 312         int good_mirror = 0;
 313         int num_copies;
 314         int ignore = 0;
 315
 316         eb = btrfs_find_create_tree_block(root, bytenr, blocksize);
 317         if (!eb)
 318                 return ERR_PTR(-ENOMEM);
 319
 320         if (btrfs_buffer_uptodate(eb, parent_transid))
 321                 return eb;
 322
 323         while (1) {
 324                 ret = read_whole_eb(root->fs_info, eb, mirror_num);
 325                 if (ret == 0 && csum_tree_block(root, eb, 1) == 0 &&
 326                     check_tree_block(root, eb) == 0 &&
 327                     verify_parent_transid(eb->tree, eb, parent_transid, ignore)
 328                     == 0) {
 329                         if (eb->flags & EXTENT_BAD_TRANSID &&
 330                             list_empty(&eb->recow)) {
 331                                 list_add_tail(&eb->recow,
 332                                               &root->fs_info->recow_ebs);
 333                                 eb->refs++;
 334                         }
 335                         btrfs_set_buffer_uptodate(eb);
 336                         return eb;
 337                 }
 338                 if (ignore) {
 339                         if (check_tree_block(root, eb)) {
 340                                 if (!root->fs_info->suppress_check_block_errors)
 341                                         print_tree_block_error(root, eb,
 342                                                 check_tree_block(root, eb));
 343                         } else {
 344                                 if (!root->fs_info->suppress_check_block_errors)
 345                                         fprintf(stderr, "Csum didn't match\n");
 346                         }
 347                         ret = -EIO;
 348                         break;
 349                 }
 350                 num_copies = btrfs_num_copies(&root->fs_info->mapping_tree,
 351                                               eb->start, eb->len);
 352                 if (num_copies == 1) {
 353                         ignore = 1;
 354                         continue;
 355                 }
 356                 if (btrfs_header_generation(eb) > best_transid && mirror_num) {
 357                         best_transid = btrfs_header_generation(eb);
 358                         good_mirror = mirror_num;
 359                 }
 360                 mirror_num++;
 361                 if (mirror_num > num_copies) {
 362                         mirror_num = good_mirror;
 363                         ignore = 1;
 364                         continue;
 365                 }
 366         }
 367         free_extent_buffer(eb);
 368         return ERR_PTR(ret);
 369 }
 370
 371 int read_extent_data(struct btrfs_root *root, char *data,
 372                            u64 logical, u64 *len, int mirror)
 373 {
 374         u64 offset = 0;
 375         struct btrfs_multi_bio *multi = NULL;
 376         struct btrfs_fs_info *info = root->fs_info;
 377         struct btrfs_device *device;
 378         int ret = 0;
 379         u64 max_len = *len;
 380
 381         ret = btrfs_map_block(&info->mapping_tree, READ, logical, len,
 382                               &multi, mirror, NULL);
 383         if (ret) {
 384                 fprintf(stderr, "Couldn't map the block %llu\n",
 385                                 logical + offset);
 386                 goto err;
 387         }
 388         device = multi->stripes[0].dev;
 389
 390         if (device->fd <= 0)
 391                 goto err;
 392         if (*len > max_len)
 393                 *len = max_len;
 394
 395         ret = pread64(device->fd, data, *len, multi->stripes[0].physical);
 396         if (ret != *len)
 397                 ret = -EIO;
 398         else
 399                 ret = 0;
 400 err:
 401         kfree(multi);
 402         return ret;
 403 }
 404
 405 int write_and_map_eb(struct btrfs_trans_handle *trans,
 406                      struct btrfs_root *root,
 407                      struct extent_buffer *eb)
 408 {
 409         int ret;
 410         int dev_nr;
 411         u64 length;
 412         u64 *raid_map = NULL;
 413         struct btrfs_multi_bio *multi = NULL;
 414
 415         dev_nr = 0;
 416         length = eb->len;
 417         ret = btrfs_map_block(&root->fs_info->mapping_tree, WRITE,
 418                               eb->start, &length, &multi, 0, &raid_map);
 419
 420         if (raid_map) {
 421                 ret = write_raid56_with_parity(root->fs_info, eb, multi,
 422                                                length, raid_map);
 423                 BUG_ON(ret);
 424         } else while (dev_nr < multi->num_stripes) {
 425                 BUG_ON(ret);
 426                 eb->fd = multi->stripes[dev_nr].dev->fd;
 427                 eb->dev_bytenr = multi->stripes[dev_nr].physical;
 428                 multi->stripes[dev_nr].dev->total_ios++;
 429                 dev_nr++;
 430                 ret = write_extent_to_disk(eb);
 431                 BUG_ON(ret);
 432         }
 433         kfree(raid_map);
 434         kfree(multi);
 435         return 0;
 436 }
 437
 438 int write_tree_block(struct btrfs_trans_handle *trans,
 439                      struct btrfs_root *root,
 440                      struct extent_buffer *eb)
 441 {
 442         if (check_tree_block(root, eb)) {
 443                 print_tree_block_error(root, eb, check_tree_block(root, eb));
 444                 BUG();
 445         }
 446
 447         if (trans && !btrfs_buffer_uptodate(eb, trans->transid))
 448                 BUG();
 449
 450         btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
 451         csum_tree_block(root, eb, 0);
 452
 453         return write_and_map_eb(trans, root, eb);
 454 }
 455
 456 int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
 457                         u32 stripesize, struct btrfs_root *root,
 458                         struct btrfs_fs_info *fs_info, u64 objectid)
 459 {
 460         root->node = NULL;
 461         root->commit_root = NULL;
 462         root->sectorsize = sectorsize;
 463         root->nodesize = nodesize;
 464         root->leafsize = leafsize;
 465         root->stripesize = stripesize;
 466         root->ref_cows = 0;
 467         root->track_dirty = 0;
 468
 469         root->fs_info = fs_info;
 470         root->objectid = objectid;
 471         root->last_trans = 0;
 472         root->highest_inode = 0;
 473         root->last_inode_alloc = 0;
 474
 475         INIT_LIST_HEAD(&root->dirty_list);
 476         INIT_LIST_HEAD(&root->orphan_data_extents);
 477         memset(&root->root_key, 0, sizeof(root->root_key));
 478         memset(&root->root_item, 0, sizeof(root->root_item));
 479         root->root_key.objectid = objectid;
 480         return 0;
 481 }
 482
 483 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 484                                struct btrfs_root *root)
 485 {
 486         int ret;
 487         u64 old_root_bytenr;
 488         struct btrfs_root *tree_root = root->fs_info->tree_root;
 489
 490         btrfs_write_dirty_block_groups(trans, root);
 491         while(1) {
 492                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 493                 if (old_root_bytenr == root->node->start)
 494                         break;
 495                 btrfs_set_root_bytenr(&root->root_item,
 496                                        root->node->start);
 497                 btrfs_set_root_generation(&root->root_item,
 498                                           trans->transid);
 499                 root->root_item.level = btrfs_header_level(root->node);
 500                 ret = btrfs_update_root(trans, tree_root,
 501                                         &root->root_key,
 502                                         &root->root_item);
 503                 BUG_ON(ret);
 504                 btrfs_write_dirty_block_groups(trans, root);
 505         }
 506         return 0;
 507 }
 508
 509 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 510                              struct btrfs_fs_info *fs_info)
 511 {
 512         struct btrfs_root *root;
 513         struct list_head *next;
 514         struct extent_buffer *eb;
 515         int ret;
 516
 517         if (fs_info->readonly)
 518                 return 0;
 519
 520         eb = fs_info->tree_root->node;
 521         extent_buffer_get(eb);
 522         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 523         free_extent_buffer(eb);
 524         if (ret)
 525                 return ret;
 526
 527         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 528                 next = fs_info->dirty_cowonly_roots.next;
 529                 list_del_init(next);
 530                 root = list_entry(next, struct btrfs_root, dirty_list);
 531                 update_cowonly_root(trans, root);
 532                 free_extent_buffer(root->commit_root);
 533                 root->commit_root = NULL;
 534         }
 535
 536         return 0;
 537 }
 538
 539 static int __commit_transaction(struct btrfs_trans_handle *trans,
 540                                 struct btrfs_root *root)
 541 {
 542         u64 start;
 543         u64 end;
 544         struct extent_buffer *eb;
 545         struct extent_io_tree *tree = &root->fs_info->extent_cache;
 546         int ret;
 547
 548         while(1) {
 549                 ret = find_first_extent_bit(tree, 0, &start, &end,
 550                                             EXTENT_DIRTY);
 551                 if (ret)
 552                         break;
 553                 while(start <= end) {
 554                         eb = find_first_extent_buffer(tree, start);
 555                         BUG_ON(!eb || eb->start != start);
 556                         ret = write_tree_block(trans, root, eb);
 557                         BUG_ON(ret);
 558                         start += eb->len;
 559                         clear_extent_buffer_dirty(eb);
 560                         free_extent_buffer(eb);
 561                 }
 562         }
 563         return 0;
 564 }
 565
 566 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 567                              struct btrfs_root *root)
 568 {
 569         u64 transid = trans->transid;
 570         int ret = 0;
 571         struct btrfs_fs_info *fs_info = root->fs_info;
 572
 573         if (root->commit_root == root->node)
 574                 goto commit_tree;
 575         if (root == root->fs_info->tree_root)
 576                 goto commit_tree;
 577         if (root == root->fs_info->chunk_root)
 578                 goto commit_tree;
 579
 580         free_extent_buffer(root->commit_root);
 581         root->commit_root = NULL;
 582
 583         btrfs_set_root_bytenr(&root->root_item, root->node->start);
 584         btrfs_set_root_generation(&root->root_item, trans->transid);
 585         root->root_item.level = btrfs_header_level(root->node);
 586         ret = btrfs_update_root(trans, root->fs_info->tree_root,
 587                                 &root->root_key, &root->root_item);
 588         BUG_ON(ret);
 589 commit_tree:
 590         ret = commit_tree_roots(trans, fs_info);
 591         BUG_ON(ret);
 592         ret = __commit_transaction(trans, root);
 593         BUG_ON(ret);
 594         write_ctree_super(trans, root);
 595         btrfs_finish_extent_commit(trans, fs_info->extent_root,
 596                                    &fs_info->pinned_extents);
 597         btrfs_free_transaction(root, trans);
 598         free_extent_buffer(root->commit_root);
 599         root->commit_root = NULL;
 600         fs_info->running_transaction = NULL;
 601         fs_info->last_trans_committed = transid;
 602         return 0;
 603 }
 604
 605 static int find_and_setup_root(struct btrfs_root *tree_root,
 606                                struct btrfs_fs_info *fs_info,
 607                                u64 objectid, struct btrfs_root *root)
 608 {
 609         int ret;
 610         u32 blocksize;
 611         u64 generation;
 612
 613         __setup_root(tree_root->nodesize, tree_root->leafsize,
 614                      tree_root->sectorsize, tree_root->stripesize,
 615                      root, fs_info, objectid);
 616         ret = btrfs_find_last_root(tree_root, objectid,
 617                                    &root->root_item, &root->root_key);
 618         if (ret)
 619                 return ret;
 620
 621         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 622         generation = btrfs_root_generation(&root->root_item);
 623         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 624                                      blocksize, generation);
 625         if (!extent_buffer_uptodate(root->node))
 626                 return -EIO;
 627
 628         return 0;
 629 }
 630
 631 static int find_and_setup_log_root(struct btrfs_root *tree_root,
 632                                struct btrfs_fs_info *fs_info,
 633                                struct btrfs_super_block *disk_super)
 634 {
 635         u32 blocksize;
 636         u64 blocknr = btrfs_super_log_root(disk_super);
 637         struct btrfs_root *log_root = malloc(sizeof(struct btrfs_root));
 638
 639         if (!log_root)
 640                 return -ENOMEM;
 641
 642         if (blocknr == 0) {
 643                 free(log_root);
 644                 return 0;
 645         }
 646
 647         blocksize = btrfs_level_size(tree_root,
 648                              btrfs_super_log_root_level(disk_super));
 649
 650         __setup_root(tree_root->nodesize, tree_root->leafsize,
 651                      tree_root->sectorsize, tree_root->stripesize,
 652                      log_root, fs_info, BTRFS_TREE_LOG_OBJECTID);
 653
 654         log_root->node = read_tree_block(tree_root, blocknr,
 655                                      blocksize,
 656                                      btrfs_super_generation(disk_super) + 1);
 657
 658         fs_info->log_root_tree = log_root;
 659
 660         if (!extent_buffer_uptodate(log_root->node)) {
 661                 free_extent_buffer(log_root->node);
 662                 free(log_root);
 663                 fs_info->log_root_tree = NULL;
 664                 return -EIO;
 665         }
 666
 667         return 0;
 668 }
 669
 670 int btrfs_free_fs_root(struct btrfs_root *root)
 671 {
 672         if (root->node)
 673                 free_extent_buffer(root->node);
 674         if (root->commit_root)
 675                 free_extent_buffer(root->commit_root);
 676         kfree(root);
 677         return 0;
 678 }
 679
 680 static void __free_fs_root(struct rb_node *node)
 681 {
 682         struct btrfs_root *root;
 683
 684         root = container_of(node, struct btrfs_root, rb_node);
 685         btrfs_free_fs_root(root);
 686 }
 687
 688 FREE_RB_BASED_TREE(fs_roots, __free_fs_root);
 689
 690 struct btrfs_root *btrfs_read_fs_root_no_cache(struct btrfs_fs_info *fs_info,
 691                                                struct btrfs_key *location)
 692 {
 693         struct btrfs_root *root;
 694         struct btrfs_root *tree_root = fs_info->tree_root;
 695         struct btrfs_path *path;
 696         struct extent_buffer *l;
 697         u64 generation;
 698         u32 blocksize;
 699         int ret = 0;
 700
 701         root = calloc(1, sizeof(*root));
 702         if (!root)
 703                 return ERR_PTR(-ENOMEM);
 704         if (location->offset == (u64)-1) {
 705                 ret = find_and_setup_root(tree_root, fs_info,
 706                                           location->objectid, root);
 707                 if (ret) {
 708                         free(root);
 709                         return ERR_PTR(ret);
 710                 }
 711                 goto insert;
 712         }
 713
 714         __setup_root(tree_root->nodesize, tree_root->leafsize,
 715                      tree_root->sectorsize, tree_root->stripesize,
 716                      root, fs_info, location->objectid);
 717
 718         path = btrfs_alloc_path();
 719         BUG_ON(!path);
 720         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 721         if (ret != 0) {
 722                 if (ret > 0)
 723                         ret = -ENOENT;
 724                 goto out;
 725         }
 726         l = path->nodes[0];
 727         read_extent_buffer(l, &root->root_item,
 728                btrfs_item_ptr_offset(l, path->slots[0]),
 729                sizeof(root->root_item));
 730         memcpy(&root->root_key, location, sizeof(*location));
 731         ret = 0;
 732 out:
 733         btrfs_free_path(path);
 734         if (ret) {
 735                 free(root);
 736                 return ERR_PTR(ret);
 737         }
 738         generation = btrfs_root_generation(&root->root_item);
 739         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 740         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 741                                      blocksize, generation);
 742         if (!extent_buffer_uptodate(root->node)) {
 743                 free(root);
 744                 return ERR_PTR(-EIO);
 745         }
 746 insert:
 747         root->ref_cows = 1;
 748         return root;
 749 }
 750
 751 static int btrfs_fs_roots_compare_objectids(struct rb_node *node,
 752                                             void *data)
 753 {
 754         u64 objectid = *((u64 *)data);
 755         struct btrfs_root *root;
 756
 757         root = rb_entry(node, struct btrfs_root, rb_node);
 758         if (objectid > root->objectid)
 759                 return 1;
 760         else if (objectid < root->objectid)
 761                 return -1;
 762         else
 763                 return 0;
 764 }
 765
 766 static int btrfs_fs_roots_compare_roots(struct rb_node *node1,
 767                                         struct rb_node *node2)
 768 {
 769         struct btrfs_root *root;
 770
 771         root = rb_entry(node2, struct btrfs_root, rb_node);
 772         return btrfs_fs_roots_compare_objectids(node1, (void *)&root->objectid);
 773 }
 774
 775 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 776                                       struct btrfs_key *location)
 777 {
 778         struct btrfs_root *root;
 779         struct rb_node *node;
 780         int ret;
 781         u64 objectid = location->objectid;
 782
 783         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 784                 return fs_info->tree_root;
 785         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 786                 return fs_info->extent_root;
 787         if (location->objectid == BTRFS_CHUNK_TREE_OBJECTID)
 788                 return fs_info->chunk_root;
 789         if (location->objectid == BTRFS_DEV_TREE_OBJECTID)
 790                 return fs_info->dev_root;
 791         if (location->objectid == BTRFS_CSUM_TREE_OBJECTID)
 792                 return fs_info->csum_root;
 793         if (location->objectid == BTRFS_QUOTA_TREE_OBJECTID)
 794                 return fs_info->quota_root;
 795
 796         BUG_ON(location->objectid == BTRFS_TREE_RELOC_OBJECTID ||
 797                location->offset != (u64)-1);
 798
 799         node = rb_search(&fs_info->fs_root_tree, (void *)&objectid,
 800                          btrfs_fs_roots_compare_objectids, NULL);
 801         if (node)
 802                 return container_of(node, struct btrfs_root, rb_node);
 803
 804         root = btrfs_read_fs_root_no_cache(fs_info, location);
 805         if (IS_ERR(root))
 806                 return root;
 807
 808         ret = rb_insert(&fs_info->fs_root_tree, &root->rb_node,
 809                         btrfs_fs_roots_compare_roots);
 810         BUG_ON(ret);
 811         return root;
 812 }
 813
 814 void btrfs_free_fs_info(struct btrfs_fs_info *fs_info)
 815 {
 816         free(fs_info->tree_root);
 817         free(fs_info->extent_root);
 818         free(fs_info->chunk_root);
 819         free(fs_info->dev_root);
 820         free(fs_info->csum_root);
 821         free(fs_info->quota_root);
 822         free(fs_info->super_copy);
 823         free(fs_info->log_root_tree);
 824         free(fs_info);
 825 }
 826
 827 struct btrfs_fs_info *btrfs_new_fs_info(int writable, u64 sb_bytenr)
 828 {
 829         struct btrfs_fs_info *fs_info;
 830
 831         fs_info = calloc(1, sizeof(struct btrfs_fs_info));
 832         if (!fs_info)
 833                 return NULL;
 834
 835         fs_info->tree_root = calloc(1, sizeof(struct btrfs_root));
 836         fs_info->extent_root = calloc(1, sizeof(struct btrfs_root));
 837         fs_info->chunk_root = calloc(1, sizeof(struct btrfs_root));
 838         fs_info->dev_root = calloc(1, sizeof(struct btrfs_root));
 839         fs_info->csum_root = calloc(1, sizeof(struct btrfs_root));
 840         fs_info->quota_root = calloc(1, sizeof(struct btrfs_root));
 841         fs_info->super_copy = calloc(1, 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         extent_io_tree_init(&fs_info->extent_cache);
 850         extent_io_tree_init(&fs_info->free_space_cache);
 851         extent_io_tree_init(&fs_info->block_group_cache);
 852         extent_io_tree_init(&fs_info->pinned_extents);
 853         extent_io_tree_init(&fs_info->pending_del);
 854         extent_io_tree_init(&fs_info->extent_ins);
 855         fs_info->excluded_extents = NULL;
 856
 857         fs_info->fs_root_tree = RB_ROOT;
 858         cache_tree_init(&fs_info->mapping_tree.cache_tree);
 859
 860         mutex_init(&fs_info->fs_mutex);
 861         INIT_LIST_HEAD(&fs_info->dirty_cowonly_roots);
 862         INIT_LIST_HEAD(&fs_info->space_info);
 863         INIT_LIST_HEAD(&fs_info->recow_ebs);
 864
 865         if (!writable)
 866                 fs_info->readonly = 1;
 867
 868         fs_info->super_bytenr = sb_bytenr;
 869         fs_info->data_alloc_profile = (u64)-1;
 870         fs_info->metadata_alloc_profile = (u64)-1;
 871         fs_info->system_alloc_profile = fs_info->metadata_alloc_profile;
 872         return fs_info;
 873 free_all:
 874         btrfs_free_fs_info(fs_info);
 875         return NULL;
 876 }
 877
 878 int btrfs_check_fs_compatibility(struct btrfs_super_block *sb, int writable)
 879 {
 880         u64 features;
 881
 882         features = btrfs_super_incompat_flags(sb) &
 883                    ~BTRFS_FEATURE_INCOMPAT_SUPP;
 884         if (features) {
 885                 printk("couldn't open because of unsupported "
 886                        "option features (%Lx).\n",
 887                        (unsigned long long)features);
 888                 return -ENOTSUP;
 889         }
 890
 891         features = btrfs_super_incompat_flags(sb);
 892         if (!(features & BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF)) {
 893                 features |= BTRFS_FEATURE_INCOMPAT_MIXED_BACKREF;
 894                 btrfs_set_super_incompat_flags(sb, features);
 895         }
 896
 897         features = btrfs_super_compat_ro_flags(sb) &
 898                 ~BTRFS_FEATURE_COMPAT_RO_SUPP;
 899         if (writable && features) {
 900                 printk("couldn't open RDWR because of unsupported "
 901                        "option features (%Lx).\n",
 902                        (unsigned long long)features);
 903                 return -ENOTSUP;
 904         }
 905         return 0;
 906 }
 907
 908 static int find_best_backup_root(struct btrfs_super_block *super)
 909 {
 910         struct btrfs_root_backup *backup;
 911         u64 orig_gen = btrfs_super_generation(super);
 912         u64 gen = 0;
 913         int best_index = 0;
 914         int i;
 915
 916         for (i = 0; i < BTRFS_NUM_BACKUP_ROOTS; i++) {
 917                 backup = super->super_roots + i;
 918                 if (btrfs_backup_tree_root_gen(backup) != orig_gen &&
 919                     btrfs_backup_tree_root_gen(backup) > gen) {
 920                         best_index = i;
 921                         gen = btrfs_backup_tree_root_gen(backup);
 922                 }
 923         }
 924         return best_index;
 925 }
 926
 927 static int setup_root_or_create_block(struct btrfs_fs_info *fs_info,
 928                                       enum btrfs_open_ctree_flags flags,
 929                                       struct btrfs_root *info_root,
 930                                       u64 objectid, char *str)
 931 {
 932         struct btrfs_super_block *sb = fs_info->super_copy;
 933         struct btrfs_root *root = fs_info->tree_root;
 934         u32 leafsize = btrfs_super_leafsize(sb);
 935         int ret;
 936
 937         ret = find_and_setup_root(root, fs_info, objectid, info_root);
 938         if (ret) {
 939                 printk("Couldn't setup %s tree\n", str);
 940                 if (!(flags & OPEN_CTREE_PARTIAL))
 941                         return -EIO;
 942                 /*
 943                  * Need a blank node here just so we don't screw up in the
 944                  * million of places that assume a root has a valid ->node
 945                  */
 946                 info_root->node =
 947                         btrfs_find_create_tree_block(info_root, 0, leafsize);
 948                 if (!info_root->node)
 949                         return -ENOMEM;
 950                 clear_extent_buffer_uptodate(NULL, info_root->node);
 951         }
 952
 953         return 0;
 954 }
 955
 956 int btrfs_setup_all_roots(struct btrfs_fs_info *fs_info, u64 root_tree_bytenr,
 957                           enum btrfs_open_ctree_flags flags)
 958 {
 959         struct btrfs_super_block *sb = fs_info->super_copy;
 960         struct btrfs_root *root;
 961         struct btrfs_key key;
 962         u32 sectorsize;
 963         u32 nodesize;
 964         u32 leafsize;
 965         u32 stripesize;
 966         u64 generation;
 967         u32 blocksize;
 968         int ret;
 969
 970         nodesize = btrfs_super_nodesize(sb);
 971         leafsize = btrfs_super_leafsize(sb);
 972         sectorsize = btrfs_super_sectorsize(sb);
 973         stripesize = btrfs_super_stripesize(sb);
 974
 975         root = fs_info->tree_root;
 976         __setup_root(nodesize, leafsize, sectorsize, stripesize,
 977                      root, fs_info, BTRFS_ROOT_TREE_OBJECTID);
 978         blocksize = btrfs_level_size(root, btrfs_super_root_level(sb));
 979         generation = btrfs_super_generation(sb);
 980
 981         if (!root_tree_bytenr && !(flags & OPEN_CTREE_BACKUP_ROOT)) {
 982                 root_tree_bytenr = btrfs_super_root(sb);
 983         } else if (flags & OPEN_CTREE_BACKUP_ROOT) {
 984                 struct btrfs_root_backup *backup;
 985                 int index = find_best_backup_root(sb);
 986                 if (index >= BTRFS_NUM_BACKUP_ROOTS) {
 987                         fprintf(stderr, "Invalid backup root number\n");
 988                         return -EIO;
 989                 }
 990                 backup = fs_info->super_copy->super_roots + index;
 991                 root_tree_bytenr = btrfs_backup_tree_root(backup);
 992                 generation = btrfs_backup_tree_root_gen(backup);
 993         }
 994
 995         root->node = read_tree_block(root, root_tree_bytenr, blocksize,
 996                                      generation);
 997         if (!extent_buffer_uptodate(root->node)) {
 998                 fprintf(stderr, "Couldn't read tree root\n");
 999                 return -EIO;
1000         }
1001
1002         ret = setup_root_or_create_block(fs_info, flags, fs_info->extent_root,
1003                                          BTRFS_EXTENT_TREE_OBJECTID, "extent");
1004         if (ret)
1005                 return ret;
1006         fs_info->extent_root->track_dirty = 1;
1007
1008         ret = find_and_setup_root(root, fs_info, BTRFS_DEV_TREE_OBJECTID,
1009                                   fs_info->dev_root);
1010         if (ret) {
1011                 printk("Couldn't setup device tree\n");
1012                 return -EIO;
1013         }
1014         fs_info->dev_root->track_dirty = 1;
1015
1016         ret = setup_root_or_create_block(fs_info, flags, fs_info->csum_root,
1017                                          BTRFS_CSUM_TREE_OBJECTID, "csum");
1018         if (ret)
1019                 return ret;
1020         fs_info->csum_root->track_dirty = 1;
1021
1022         ret = find_and_setup_root(root, fs_info, BTRFS_QUOTA_TREE_OBJECTID,
1023                                   fs_info->quota_root);
1024         if (ret == 0)
1025                 fs_info->quota_enabled = 1;
1026
1027         ret = find_and_setup_log_root(root, fs_info, sb);
1028         if (ret) {
1029                 printk("Couldn't setup log root tree\n");
1030                 if (!(flags & OPEN_CTREE_PARTIAL))
1031                         return -EIO;
1032         }
1033
1034         fs_info->generation = generation;
1035         fs_info->last_trans_committed = generation;
1036         if (extent_buffer_uptodate(fs_info->extent_root->node) &&
1037             !(flags & OPEN_CTREE_NO_BLOCK_GROUPS))
1038                 btrfs_read_block_groups(fs_info->tree_root);
1039
1040         key.objectid = BTRFS_FS_TREE_OBJECTID;
1041         key.type = BTRFS_ROOT_ITEM_KEY;
1042         key.offset = (u64)-1;
1043         fs_info->fs_root = btrfs_read_fs_root(fs_info, &key);
1044
1045         if (IS_ERR(fs_info->fs_root))
1046                 return -EIO;
1047         return 0;
1048 }
1049
1050 void btrfs_release_all_roots(struct btrfs_fs_info *fs_info)
1051 {
1052         if (fs_info->quota_root)
1053                 free_extent_buffer(fs_info->quota_root->node);
1054         if (fs_info->csum_root)
1055                 free_extent_buffer(fs_info->csum_root->node);
1056         if (fs_info->dev_root)
1057                 free_extent_buffer(fs_info->dev_root->node);
1058         if (fs_info->extent_root)
1059                 free_extent_buffer(fs_info->extent_root->node);
1060         if (fs_info->tree_root)
1061                 free_extent_buffer(fs_info->tree_root->node);
1062         if (fs_info->log_root_tree)
1063                 free_extent_buffer(fs_info->log_root_tree->node);
1064         if (fs_info->chunk_root)
1065                 free_extent_buffer(fs_info->chunk_root->node);
1066 }
1067
1068 static void free_map_lookup(struct cache_extent *ce)
1069 {
1070         struct map_lookup *map;
1071
1072         map = container_of(ce, struct map_lookup, ce);
1073         kfree(map);
1074 }
1075
1076 FREE_EXTENT_CACHE_BASED_TREE(mapping_cache, free_map_lookup);
1077
1078 void btrfs_cleanup_all_caches(struct btrfs_fs_info *fs_info)
1079 {
1080         while (!list_empty(&fs_info->recow_ebs)) {
1081                 struct extent_buffer *eb;
1082                 eb = list_first_entry(&fs_info->recow_ebs,
1083                                       struct extent_buffer, recow);
1084                 list_del_init(&eb->recow);
1085                 free_extent_buffer(eb);
1086         }
1087         free_mapping_cache_tree(&fs_info->mapping_tree.cache_tree);
1088         extent_io_tree_cleanup(&fs_info->extent_cache);
1089         extent_io_tree_cleanup(&fs_info->free_space_cache);
1090         extent_io_tree_cleanup(&fs_info->block_group_cache);
1091         extent_io_tree_cleanup(&fs_info->pinned_extents);
1092         extent_io_tree_cleanup(&fs_info->pending_del);
1093         extent_io_tree_cleanup(&fs_info->extent_ins);
1094 }
1095
1096 int btrfs_scan_fs_devices(int fd, const char *path,
1097                           struct btrfs_fs_devices **fs_devices,
1098                           u64 sb_bytenr, int super_recover,
1099                           int skip_devices)
1100 {
1101         u64 total_devs;
1102         u64 dev_size;
1103         off_t seek_ret;
1104         int ret;
1105         if (!sb_bytenr)
1106                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1107
1108         seek_ret = lseek(fd, 0, SEEK_END);
1109         if (seek_ret < 0)
1110                 return -errno;
1111
1112         dev_size = seek_ret;
1113         lseek(fd, 0, SEEK_SET);
1114         if (sb_bytenr > dev_size) {
1115                 fprintf(stderr, "Superblock bytenr is larger than device size\n");
1116                 return -EINVAL;
1117         }
1118
1119         ret = btrfs_scan_one_device(fd, path, fs_devices,
1120                                     &total_devs, sb_bytenr, super_recover);
1121         if (ret) {
1122                 fprintf(stderr, "No valid Btrfs found on %s\n", path);
1123                 return ret;
1124         }
1125
1126         if (!skip_devices && total_devs != 1) {
1127                 ret = btrfs_scan_lblkid();
1128                 if (ret)
1129                         return ret;
1130         }
1131         return 0;
1132 }
1133
1134 int btrfs_setup_chunk_tree_and_device_map(struct btrfs_fs_info *fs_info)
1135 {
1136         struct btrfs_super_block *sb = fs_info->super_copy;
1137         u32 sectorsize;
1138         u32 nodesize;
1139         u32 leafsize;
1140         u32 blocksize;
1141         u32 stripesize;
1142         u64 generation;
1143         int ret;
1144
1145         nodesize = btrfs_super_nodesize(sb);
1146         leafsize = btrfs_super_leafsize(sb);
1147         sectorsize = btrfs_super_sectorsize(sb);
1148         stripesize = btrfs_super_stripesize(sb);
1149
1150         __setup_root(nodesize, leafsize, sectorsize, stripesize,
1151                      fs_info->chunk_root, fs_info, BTRFS_CHUNK_TREE_OBJECTID);
1152
1153         ret = btrfs_read_sys_array(fs_info->chunk_root);
1154         if (ret)
1155                 return ret;
1156
1157         blocksize = btrfs_level_size(fs_info->chunk_root,
1158                                      btrfs_super_chunk_root_level(sb));
1159         generation = btrfs_super_chunk_root_generation(sb);
1160
1161         fs_info->chunk_root->node = read_tree_block(fs_info->chunk_root,
1162                                                     btrfs_super_chunk_root(sb),
1163                                                     blocksize, generation);
1164         if (!extent_buffer_uptodate(fs_info->chunk_root->node)) {
1165                 fprintf(stderr, "Couldn't read chunk root\n");
1166                 return -EIO;
1167         }
1168
1169         if (!(btrfs_super_flags(sb) & BTRFS_SUPER_FLAG_METADUMP)) {
1170                 ret = btrfs_read_chunk_tree(fs_info->chunk_root);
1171                 if (ret) {
1172                         fprintf(stderr, "Couldn't read chunk tree\n");
1173                         return ret;
1174                 }
1175         }
1176         return 0;
1177 }
1178
1179 static struct btrfs_fs_info *__open_ctree_fd(int fp, const char *path,
1180                                              u64 sb_bytenr,
1181                                              u64 root_tree_bytenr,
1182                                              enum btrfs_open_ctree_flags flags)
1183 {
1184         struct btrfs_fs_info *fs_info;
1185         struct btrfs_super_block *disk_super;
1186         struct btrfs_fs_devices *fs_devices = NULL;
1187         struct extent_buffer *eb;
1188         int ret;
1189         int oflags;
1190
1191         if (sb_bytenr == 0)
1192                 sb_bytenr = BTRFS_SUPER_INFO_OFFSET;
1193
1194         /* try to drop all the caches */
1195         if (posix_fadvise(fp, 0, 0, POSIX_FADV_DONTNEED))
1196                 fprintf(stderr, "Warning, could not drop caches\n");
1197
1198         fs_info = btrfs_new_fs_info(flags & OPEN_CTREE_WRITES, sb_bytenr);
1199         if (!fs_info) {
1200                 fprintf(stderr, "Failed to allocate memory for fs_info\n");
1201                 return NULL;
1202         }
1203         if (flags & OPEN_CTREE_RESTORE)
1204                 fs_info->on_restoring = 1;
1205         if (flags & OPEN_CTREE_SUPPRESS_CHECK_BLOCK_ERRORS)
1206                 fs_info->suppress_check_block_errors = 1;
1207         if (flags & OPEN_CTREE_IGNORE_FSID_MISMATCH)
1208                 fs_info->ignore_fsid_mismatch = 1;
1209
1210         ret = btrfs_scan_fs_devices(fp, path, &fs_devices, sb_bytenr,
1211                                     (flags & OPEN_CTREE_RECOVER_SUPER),
1212                                     (flags & OPEN_CTREE_NO_DEVICES));
1213         if (ret)
1214                 goto out;
1215
1216         fs_info->fs_devices = fs_devices;
1217         if (flags & OPEN_CTREE_WRITES)
1218                 oflags = O_RDWR;
1219         else
1220                 oflags = O_RDONLY;
1221
1222         if (flags & OPEN_CTREE_EXCLUSIVE)
1223                 oflags |= O_EXCL;
1224
1225         ret = btrfs_open_devices(fs_devices, oflags);
1226         if (ret)
1227                 goto out;
1228
1229         disk_super = fs_info->super_copy;
1230         if (!(flags & OPEN_CTREE_RECOVER_SUPER))
1231                 ret = btrfs_read_dev_super(fs_devices->latest_bdev,
1232                                            disk_super, sb_bytenr, 1);
1233         else
1234                 ret = btrfs_read_dev_super(fp, disk_super, sb_bytenr, 0);
1235         if (ret) {
1236                 printk("No valid btrfs found\n");
1237                 goto out_devices;
1238         }
1239
1240         if (btrfs_super_flags(disk_super) & BTRFS_SUPER_FLAG_CHANGING_FSID &&
1241             !fs_info->ignore_fsid_mismatch) {
1242                 fprintf(stderr, "ERROR: Filesystem UUID change in progress\n");
1243                 goto out_devices;
1244         }
1245
1246         memcpy(fs_info->fsid, &disk_super->fsid, BTRFS_FSID_SIZE);
1247
1248         ret = btrfs_check_fs_compatibility(fs_info->super_copy,
1249                                            flags & OPEN_CTREE_WRITES);
1250         if (ret)
1251                 goto out_devices;
1252
1253         ret = btrfs_setup_chunk_tree_and_device_map(fs_info);
1254         if (ret)
1255                 goto out_chunk;
1256
1257         eb = fs_info->chunk_root->node;
1258         read_extent_buffer(eb, fs_info->chunk_tree_uuid,
1259                            btrfs_header_chunk_tree_uuid(eb),
1260                            BTRFS_UUID_SIZE);
1261
1262         ret = btrfs_setup_all_roots(fs_info, root_tree_bytenr, flags);
1263         if (ret && !(flags & __OPEN_CTREE_RETURN_CHUNK_ROOT))
1264                 goto out_chunk;
1265
1266         return fs_info;
1267
1268 out_chunk:
1269         btrfs_release_all_roots(fs_info);
1270         btrfs_cleanup_all_caches(fs_info);
1271 out_devices:
1272         btrfs_close_devices(fs_devices);
1273 out:
1274         btrfs_free_fs_info(fs_info);
1275         return NULL;
1276 }
1277
1278 struct btrfs_fs_info *open_ctree_fs_info(const char *filename,
1279                                          u64 sb_bytenr, u64 root_tree_bytenr,
1280                                          enum btrfs_open_ctree_flags flags)
1281 {
1282         int fp;
1283         struct btrfs_fs_info *info;
1284         int oflags = O_CREAT | O_RDWR;
1285
1286         if (!(flags & OPEN_CTREE_WRITES))
1287                 oflags = O_RDONLY;
1288
1289         fp = open(filename, oflags, 0600);
1290         if (fp < 0) {
1291                 fprintf (stderr, "Could not open %s\n", filename);
1292                 return NULL;
1293         }
1294         info = __open_ctree_fd(fp, filename, sb_bytenr, root_tree_bytenr,
1295                                flags);
1296         close(fp);
1297         return info;
1298 }
1299
1300 struct btrfs_root *open_ctree(const char *filename, u64 sb_bytenr,
1301                               enum btrfs_open_ctree_flags flags)
1302 {
1303         struct btrfs_fs_info *info;
1304
1305         info = open_ctree_fs_info(filename, sb_bytenr, 0, flags);
1306         if (!info)
1307                 return NULL;
1308         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1309                 return info->chunk_root;
1310         return info->fs_root;
1311 }
1312
1313 struct btrfs_root *open_ctree_fd(int fp, const char *path, u64 sb_bytenr,
1314                                  enum btrfs_open_ctree_flags flags)
1315 {
1316         struct btrfs_fs_info *info;
1317         info = __open_ctree_fd(fp, path, sb_bytenr, 0, flags);
1318         if (!info)
1319                 return NULL;
1320         if (flags & __OPEN_CTREE_RETURN_CHUNK_ROOT)
1321                 return info->chunk_root;
1322         return info->fs_root;
1323 }
1324
1325 /*
1326  * Check if the super is valid:
1327  * - nodesize/sectorsize - minimum, maximum, alignment
1328  * - tree block starts   - alignment
1329  * - number of devices   - something sane
1330  * - sys array size      - maximum
1331  */
1332 static int check_super(struct btrfs_super_block *sb)
1333 {
1334         char result[BTRFS_CSUM_SIZE];
1335         u32 crc;
1336         u16 csum_type;
1337         int csum_size;
1338
1339         if (btrfs_super_magic(sb) != BTRFS_MAGIC) {
1340                 fprintf(stderr, "ERROR: superblock magic doesn't match\n");
1341                 return -EIO;
1342         }
1343
1344         csum_type = btrfs_super_csum_type(sb);
1345         if (csum_type >= ARRAY_SIZE(btrfs_csum_sizes)) {
1346                 fprintf(stderr, "ERROR: unsupported checksum algorithm %u\n",
1347                         csum_type);
1348                 return -EIO;
1349         }
1350         csum_size = btrfs_csum_sizes[csum_type];
1351
1352         crc = ~(u32)0;
1353         crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1354                               BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1355         btrfs_csum_final(crc, result);
1356
1357         if (memcmp(result, sb->csum, csum_size)) {
1358                 fprintf(stderr, "ERROR: superblock checksum mismatch\n");
1359                 return -EIO;
1360         }
1361         if (btrfs_super_root_level(sb) >= BTRFS_MAX_LEVEL) {
1362                 fprintf(stderr, "ERROR: tree_root level too big: %d >= %d\n",
1363                         btrfs_super_root_level(sb), BTRFS_MAX_LEVEL);
1364                 return -EIO;
1365         }
1366         if (btrfs_super_chunk_root_level(sb) >= BTRFS_MAX_LEVEL) {
1367                 fprintf(stderr, "ERROR: chunk_root level too big: %d >= %d\n",
1368                         btrfs_super_chunk_root_level(sb), BTRFS_MAX_LEVEL);
1369                 return -EIO;
1370         }
1371         if (btrfs_super_log_root_level(sb) >= BTRFS_MAX_LEVEL) {
1372                 fprintf(stderr, "ERROR: log_root level too big: %d >= %d\n",
1373                         btrfs_super_log_root_level(sb), BTRFS_MAX_LEVEL);
1374                 return -EIO;
1375         }
1376
1377         if (!IS_ALIGNED(btrfs_super_root(sb), 4096)) {
1378                 fprintf(stderr, "ERROR: tree_root block unaligned: %llu\n",
1379                         btrfs_super_root(sb));
1380                 return -EIO;
1381         }
1382         if (!IS_ALIGNED(btrfs_super_chunk_root(sb), 4096)) {
1383                 fprintf(stderr, "ERROR: chunk_root block unaligned: %llu\n",
1384                         btrfs_super_chunk_root(sb));
1385                 return -EIO;
1386         }
1387         if (!IS_ALIGNED(btrfs_super_log_root(sb), 4096)) {
1388                 fprintf(stderr, "ERROR: log_root block unaligned: %llu\n",
1389                         btrfs_super_log_root(sb));
1390                 return -EIO;
1391         }
1392         if (btrfs_super_nodesize(sb) < 4096) {
1393                 fprintf(stderr, "ERROR: nodesize too small: %u < 4096\n",
1394                         btrfs_super_nodesize(sb));
1395                 return -EIO;
1396         }
1397         if (!IS_ALIGNED(btrfs_super_nodesize(sb), 4096)) {
1398                 fprintf(stderr, "ERROR: nodesize unaligned: %u\n",
1399                         btrfs_super_nodesize(sb));
1400                 return -EIO;
1401         }
1402         if (btrfs_super_sectorsize(sb) < 4096) {
1403                 fprintf(stderr, "ERROR: sectorsize too small: %u < 4096\n",
1404                         btrfs_super_sectorsize(sb));
1405                 return -EIO;
1406         }
1407         if (!IS_ALIGNED(btrfs_super_sectorsize(sb), 4096)) {
1408                 fprintf(stderr, "ERROR: sectorsize unaligned: %u\n",
1409                         btrfs_super_sectorsize(sb));
1410                 return -EIO;
1411         }
1412
1413         if (memcmp(sb->fsid, sb->dev_item.fsid, BTRFS_UUID_SIZE) != 0) {
1414                 char fsid[BTRFS_UUID_UNPARSED_SIZE];
1415                 char dev_fsid[BTRFS_UUID_UNPARSED_SIZE];
1416
1417                 uuid_unparse(sb->fsid, fsid);
1418                 uuid_unparse(sb->dev_item.fsid, dev_fsid);
1419                 printk(KERN_ERR
1420                         "ERROR: dev_item UUID does not match fsid: %s != %s\n",
1421                         dev_fsid, fsid);
1422                 return -EIO;
1423         }
1424
1425         /*
1426          * Hint to catch really bogus numbers, bitflips or so
1427          */
1428         if (btrfs_super_num_devices(sb) > (1UL << 31)) {
1429                 fprintf(stderr, "WARNING: suspicious number of devices: %llu\n",
1430                         btrfs_super_num_devices(sb));
1431         }
1432
1433         if (btrfs_super_num_devices(sb) == 0) {
1434                 fprintf(stderr, "ERROR: number of devices is 0\n");
1435                 return -EIO;
1436         }
1437
1438         /*
1439          * Obvious sys_chunk_array corruptions, it must hold at least one key
1440          * and one chunk
1441          */
1442         if (btrfs_super_sys_array_size(sb) > BTRFS_SYSTEM_CHUNK_ARRAY_SIZE) {
1443                 fprintf(stderr, "BTRFS: system chunk array too big %u > %u\n",
1444                         btrfs_super_sys_array_size(sb),
1445                         BTRFS_SYSTEM_CHUNK_ARRAY_SIZE);
1446                 return -EIO;
1447         }
1448         if (btrfs_super_sys_array_size(sb) < sizeof(struct btrfs_disk_key)
1449                         + sizeof(struct btrfs_chunk)) {
1450                 fprintf(stderr, "BTRFS: system chunk array too small %u < %lu\n",
1451                         btrfs_super_sys_array_size(sb),
1452                         sizeof(struct btrfs_disk_key) +
1453                         sizeof(struct btrfs_chunk));
1454                 return -EIO;
1455         }
1456
1457         return 0;
1458 }
1459
1460 int btrfs_read_dev_super(int fd, struct btrfs_super_block *sb, u64 sb_bytenr,
1461                          int super_recover)
1462 {
1463         u8 fsid[BTRFS_FSID_SIZE];
1464         int fsid_is_initialized = 0;
1465         char tmp[BTRFS_SUPER_INFO_SIZE];
1466         struct btrfs_super_block *buf = (struct btrfs_super_block *)tmp;
1467         int i;
1468         int ret;
1469         int max_super = super_recover ? BTRFS_SUPER_MIRROR_MAX : 1;
1470         u64 transid = 0;
1471         u64 bytenr;
1472
1473         if (sb_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1474                 ret = pread64(fd, buf, BTRFS_SUPER_INFO_SIZE, sb_bytenr);
1475                 if (ret < BTRFS_SUPER_INFO_SIZE)
1476                         return -1;
1477
1478                 if (btrfs_super_bytenr(buf) != sb_bytenr)
1479                         return -1;
1480
1481                 if (check_super(buf))
1482                         return -1;
1483                 memcpy(sb, buf, BTRFS_SUPER_INFO_SIZE);
1484                 return 0;
1485         }
1486
1487         /*
1488         * we would like to check all the supers, but that would make
1489         * a btrfs mount succeed after a mkfs from a different FS.
1490         * So, we need to add a special mount option to scan for
1491         * later supers, using BTRFS_SUPER_MIRROR_MAX instead
1492         */
1493
1494         for (i = 0; i < max_super; i++) {
1495                 bytenr = btrfs_sb_offset(i);
1496                 ret = pread64(fd, buf, BTRFS_SUPER_INFO_SIZE, bytenr);
1497                 if (ret < BTRFS_SUPER_INFO_SIZE)
1498                         break;
1499
1500                 if (btrfs_super_bytenr(buf) != bytenr )
1501                         continue;
1502                 /* if magic is NULL, the device was removed */
1503                 if (btrfs_super_magic(buf) == 0 && i == 0)
1504                         break;
1505                 if (check_super(buf))
1506                         continue;
1507
1508                 if (!fsid_is_initialized) {
1509                         memcpy(fsid, buf->fsid, sizeof(fsid));
1510                         fsid_is_initialized = 1;
1511                 } else if (memcmp(fsid, buf->fsid, sizeof(fsid))) {
1512                         /*
1513                          * the superblocks (the original one and
1514                          * its backups) contain data of different
1515                          * filesystems -> the super cannot be trusted
1516                          */
1517                         continue;
1518                 }
1519
1520                 if (btrfs_super_generation(buf) > transid) {
1521                         memcpy(sb, buf, BTRFS_SUPER_INFO_SIZE);
1522                         transid = btrfs_super_generation(buf);
1523                 }
1524         }
1525
1526         return transid > 0 ? 0 : -1;
1527 }
1528
1529 static int write_dev_supers(struct btrfs_root *root,
1530                             struct btrfs_super_block *sb,
1531                             struct btrfs_device *device)
1532 {
1533         u64 bytenr;
1534         u32 crc;
1535         int i, ret;
1536
1537         if (root->fs_info->super_bytenr != BTRFS_SUPER_INFO_OFFSET) {
1538                 btrfs_set_super_bytenr(sb, root->fs_info->super_bytenr);
1539                 crc = ~(u32)0;
1540                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1541                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1542                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1543
1544                 /*
1545                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1546                  * zero filled, we can use it directly
1547                  */
1548                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1549                                 BTRFS_SUPER_INFO_SIZE,
1550                                 root->fs_info->super_bytenr);
1551                 if (ret != BTRFS_SUPER_INFO_SIZE)
1552                         goto write_err;
1553                 return 0;
1554         }
1555
1556         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1557                 bytenr = btrfs_sb_offset(i);
1558                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
1559                         break;
1560
1561                 btrfs_set_super_bytenr(sb, bytenr);
1562
1563                 crc = ~(u32)0;
1564                 crc = btrfs_csum_data(NULL, (char *)sb + BTRFS_CSUM_SIZE, crc,
1565                                       BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE);
1566                 btrfs_csum_final(crc, (char *)&sb->csum[0]);
1567
1568                 /*
1569                  * super_copy is BTRFS_SUPER_INFO_SIZE bytes and is
1570                  * zero filled, we can use it directly
1571                  */
1572                 ret = pwrite64(device->fd, root->fs_info->super_copy,
1573                                 BTRFS_SUPER_INFO_SIZE, bytenr);
1574                 if (ret != BTRFS_SUPER_INFO_SIZE)
1575                         goto write_err;
1576         }
1577
1578         return 0;
1579
1580 write_err:
1581         if (ret > 0)
1582                 fprintf(stderr, "WARNING: failed to write all sb data\n");
1583         else
1584                 fprintf(stderr, "WARNING: failed to write sb: %s\n",
1585                         strerror(errno));
1586         return ret;
1587 }
1588
1589 int write_all_supers(struct btrfs_root *root)
1590 {
1591         struct list_head *cur;
1592         struct list_head *head = &root->fs_info->fs_devices->devices;
1593         struct btrfs_device *dev;
1594         struct btrfs_super_block *sb;
1595         struct btrfs_dev_item *dev_item;
1596         int ret;
1597         u64 flags;
1598
1599         sb = root->fs_info->super_copy;
1600         dev_item = &sb->dev_item;
1601         list_for_each(cur, head) {
1602                 dev = list_entry(cur, struct btrfs_device, dev_list);
1603                 if (!dev->writeable)
1604                         continue;
1605
1606                 btrfs_set_stack_device_generation(dev_item, 0);
1607                 btrfs_set_stack_device_type(dev_item, dev->type);
1608                 btrfs_set_stack_device_id(dev_item, dev->devid);
1609                 btrfs_set_stack_device_total_bytes(dev_item, dev->total_bytes);
1610                 btrfs_set_stack_device_bytes_used(dev_item, dev->bytes_used);
1611                 btrfs_set_stack_device_io_align(dev_item, dev->io_align);
1612                 btrfs_set_stack_device_io_width(dev_item, dev->io_width);
1613                 btrfs_set_stack_device_sector_size(dev_item, dev->sector_size);
1614                 memcpy(dev_item->uuid, dev->uuid, BTRFS_UUID_SIZE);
1615                 memcpy(dev_item->fsid, dev->fs_devices->fsid, BTRFS_UUID_SIZE);
1616
1617                 flags = btrfs_super_flags(sb);
1618                 btrfs_set_super_flags(sb, flags | BTRFS_HEADER_FLAG_WRITTEN);
1619
1620                 ret = write_dev_supers(root, sb, dev);
1621                 BUG_ON(ret);
1622         }
1623         return 0;
1624 }
1625
1626 int write_ctree_super(struct btrfs_trans_handle *trans,
1627                       struct btrfs_root *root)
1628 {
1629         int ret;
1630         struct btrfs_root *tree_root = root->fs_info->tree_root;
1631         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
1632
1633         if (root->fs_info->readonly)
1634                 return 0;
1635
1636         btrfs_set_super_generation(root->fs_info->super_copy,
1637                                    trans->transid);
1638         btrfs_set_super_root(root->fs_info->super_copy,
1639                              tree_root->node->start);
1640         btrfs_set_super_root_level(root->fs_info->super_copy,
1641                                    btrfs_header_level(tree_root->node));
1642         btrfs_set_super_chunk_root(root->fs_info->super_copy,
1643                                    chunk_root->node->start);
1644         btrfs_set_super_chunk_root_level(root->fs_info->super_copy,
1645                                          btrfs_header_level(chunk_root->node));
1646         btrfs_set_super_chunk_root_generation(root->fs_info->super_copy,
1647                                 btrfs_header_generation(chunk_root->node));
1648
1649         ret = write_all_supers(root);
1650         if (ret)
1651                 fprintf(stderr, "failed to write new super block err %d\n", ret);
1652         return ret;
1653 }
1654
1655 int close_ctree(struct btrfs_root *root)
1656 {
1657         int ret;
1658         struct btrfs_trans_handle *trans;
1659         struct btrfs_fs_info *fs_info = root->fs_info;
1660
1661         if (fs_info->last_trans_committed !=
1662             fs_info->generation) {
1663                 trans = btrfs_start_transaction(root, 1);
1664                 btrfs_commit_transaction(trans, root);
1665                 trans = btrfs_start_transaction(root, 1);
1666                 ret = commit_tree_roots(trans, fs_info);
1667                 BUG_ON(ret);
1668                 ret = __commit_transaction(trans, root);
1669                 BUG_ON(ret);
1670                 write_ctree_super(trans, root);
1671                 btrfs_free_transaction(root, trans);
1672         }
1673         btrfs_free_block_groups(fs_info);
1674
1675         free_fs_roots_tree(&fs_info->fs_root_tree);
1676
1677         btrfs_release_all_roots(fs_info);
1678         btrfs_close_devices(fs_info->fs_devices);
1679         btrfs_cleanup_all_caches(fs_info);
1680         btrfs_free_fs_info(fs_info);
1681         return 0;
1682 }
1683
1684 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1685                      struct extent_buffer *eb)
1686 {
1687         return clear_extent_buffer_dirty(eb);
1688 }
1689
1690 int wait_on_tree_block_writeback(struct btrfs_root *root,
1691                                  struct extent_buffer *eb)
1692 {
1693         return 0;
1694 }
1695
1696 void btrfs_mark_buffer_dirty(struct extent_buffer *eb)
1697 {
1698         set_extent_buffer_dirty(eb);
1699 }
1700
1701 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid)
1702 {
1703         int ret;
1704
1705         ret = extent_buffer_uptodate(buf);
1706         if (!ret)
1707                 return ret;
1708
1709         ret = verify_parent_transid(buf->tree, buf, parent_transid, 1);
1710         return !ret;
1711 }
1712
1713 int btrfs_set_buffer_uptodate(struct extent_buffer *eb)
1714 {
1715         return set_extent_buffer_uptodate(eb);
1716 }