disk-io.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #define _XOPEN_SOURCE 600
  20 #define __USE_XOPEN2K
  21 #include <stdio.h>
  22 #include <stdlib.h>
  23 #include <sys/types.h>
  24 #include <sys/stat.h>
  25 #include <fcntl.h>
  26 #include <unistd.h>
  27 #include "kerncompat.h"
  28 #include "radix-tree.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "transaction.h"
  32 #include "crc32c.h"
  33
  34 static u64 allocated_bytes = 0;
  35 int cache_max = 10000;
  36
  37 int btrfs_map_bh_to_logical(struct btrfs_root *root, struct btrfs_buffer *bh,
  38                              u64 logical)
  39 {
  40         bh->fd = root->fs_info->fp;
  41         bh->dev_bytenr = logical;
  42         return 0;
  43 }
  44
  45 static int check_tree_block(struct btrfs_root *root, struct btrfs_buffer *buf)
  46 {
  47         if (buf->bytenr != btrfs_header_bytenr(&buf->node.header))
  48                 BUG();
  49         if (memcmp(root->fs_info->disk_super->fsid, buf->node.header.fsid,
  50                    sizeof(buf->node.header.fsid)))
  51                 BUG();
  52         return 0;
  53 }
  54
  55 static int free_some_buffers(struct btrfs_root *root)
  56 {
  57         struct list_head *node, *next;
  58         struct btrfs_buffer *b;
  59         if (root->fs_info->cache_size < cache_max)
  60                 return 0;
  61         list_for_each_safe(node, next, &root->fs_info->cache) {
  62                 b = list_entry(node, struct btrfs_buffer, cache);
  63                 if (b->count == 1) {
  64                         BUG_ON(!list_empty(&b->dirty));
  65                         list_del_init(&b->cache);
  66                         btrfs_block_release(root, b);
  67                         if (root->fs_info->cache_size < cache_max)
  68                                 break;
  69                 }
  70         }
  71         return 0;
  72 }
  73
  74 struct btrfs_buffer *alloc_tree_block(struct btrfs_root *root, u64 bytenr,
  75                                       u32 blocksize)
  76 {
  77         struct btrfs_buffer *buf;
  78         int ret;
  79
  80         buf = malloc(sizeof(struct btrfs_buffer) + blocksize);
  81         if (!buf)
  82                 return buf;
  83         allocated_bytes += blocksize;
  84
  85         buf->bytenr = bytenr;
  86         buf->count = 2;
  87         buf->size = blocksize;
  88         buf->cache_node.start = bytenr;
  89         buf->cache_node.size = blocksize;
  90
  91         INIT_LIST_HEAD(&buf->dirty);
  92         free_some_buffers(root);
  93
  94         ret = insert_existing_cache_extent(&root->fs_info->extent_cache,
  95                                            &buf->cache_node);
  96
  97         list_add_tail(&buf->cache, &root->fs_info->cache);
  98         root->fs_info->cache_size += blocksize;
  99         if (ret) {
 100                 free(buf);
 101                 return NULL;
 102         }
 103         return buf;
 104 }
 105
 106 struct btrfs_buffer *find_tree_block(struct btrfs_root *root, u64 bytenr,
 107                                      u32 blocksize)
 108 {
 109         struct btrfs_buffer *buf;
 110         struct cache_extent *cache;
 111
 112         cache = find_cache_extent(&root->fs_info->extent_cache,
 113                                   bytenr, blocksize);
 114         if (cache) {
 115                 buf = container_of(cache, struct btrfs_buffer, cache_node);
 116                 buf->count++;
 117         } else {
 118                 buf = alloc_tree_block(root, bytenr, blocksize);
 119                 if (!buf) {
 120                         BUG();
 121                         return NULL;
 122                 }
 123         }
 124         return buf;
 125 }
 126
 127 struct btrfs_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 128                                      u32 blocksize)
 129 {
 130         struct btrfs_buffer *buf;
 131         int ret;
 132         struct cache_extent *cache;
 133
 134         cache = find_cache_extent(&root->fs_info->extent_cache,
 135                                   bytenr, blocksize);
 136         if (cache) {
 137                 buf = container_of(cache, struct btrfs_buffer, cache_node);
 138                 buf->count++;
 139                 if (check_tree_block(root, buf))
 140                         BUG();
 141         } else {
 142                 buf = alloc_tree_block(root, bytenr, blocksize);
 143                 if (!buf)
 144                         return NULL;
 145                 btrfs_map_bh_to_logical(root, buf, bytenr);
 146                 ret = pread(buf->fd, &buf->node, blocksize,
 147                             buf->dev_bytenr);
 148                 if (ret != blocksize) {
 149                         free(buf);
 150                         return NULL;
 151                 }
 152                 if (check_tree_block(root, buf))
 153                         BUG();
 154         }
 155         return buf;
 156 }
 157
 158 int dirty_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 159                      struct btrfs_buffer *buf)
 160 {
 161         if (!list_empty(&buf->dirty))
 162                 return 0;
 163         list_add_tail(&buf->dirty, &root->fs_info->trans);
 164         buf->count++;
 165         if (check_tree_block(root, buf))
 166                 BUG();
 167         return 0;
 168 }
 169
 170 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 171                      struct btrfs_buffer *buf)
 172 {
 173         if (!list_empty(&buf->dirty)) {
 174                 list_del_init(&buf->dirty);
 175                 btrfs_block_release(root, buf);
 176         }
 177         return 0;
 178 }
 179
 180 int btrfs_csum_node(struct btrfs_root *root, struct btrfs_node *node)
 181 {
 182         u32 crc = ~(u32)0;
 183         size_t len = btrfs_level_size(root, btrfs_header_level(&node->header)) -
 184                                       BTRFS_CSUM_SIZE;
 185
 186         crc = crc32c(crc, (char *)(node) + BTRFS_CSUM_SIZE, len);
 187         crc = ~cpu_to_le32(crc);
 188         memcpy(node->header.csum, &crc, BTRFS_CRC32_SIZE);
 189         return 0;
 190 }
 191
 192 int btrfs_csum_super(struct btrfs_root *root, struct btrfs_super_block *super)
 193 {
 194         u32 crc = ~(u32)0;
 195         char block[512];
 196         size_t len = 512 - BTRFS_CSUM_SIZE;
 197
 198         memset(block, 0, 512);
 199         memcpy(block, super, sizeof(*super));
 200
 201         crc = crc32c(crc, block + BTRFS_CSUM_SIZE, len);
 202         crc = ~cpu_to_le32(crc);
 203         memcpy(super->csum, &crc, BTRFS_CRC32_SIZE);
 204         return 0;
 205 }
 206
 207 int write_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 208                      struct btrfs_buffer *buf)
 209 {
 210         int ret;
 211
 212         if (buf->bytenr != btrfs_header_bytenr(&buf->node.header))
 213                 BUG();
 214         btrfs_map_bh_to_logical(root, buf, buf->bytenr);
 215         if (check_tree_block(root, buf))
 216                 BUG();
 217
 218         btrfs_csum_node(root, &buf->node);
 219
 220         ret = pwrite(buf->fd, &buf->node, buf->size,
 221                      buf->dev_bytenr);
 222         if (ret != buf->size)
 223                 return ret;
 224         return 0;
 225 }
 226
 227 static int __commit_transaction(struct btrfs_trans_handle *trans, struct
 228                                 btrfs_root *root)
 229 {
 230         struct btrfs_buffer *b;
 231         int ret = 0;
 232         int wret;
 233         while(!list_empty(&root->fs_info->trans)) {
 234                 b = list_entry(root->fs_info->trans.next, struct btrfs_buffer,
 235                                dirty);
 236                 list_del_init(&b->dirty);
 237                 wret = write_tree_block(trans, root, b);
 238                 if (wret)
 239                         ret = wret;
 240                 btrfs_block_release(root, b);
 241         }
 242         return ret;
 243 }
 244
 245 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 246                              struct btrfs_fs_info *fs_info)
 247 {
 248         int ret;
 249         u64 old_extent_bytenr;
 250         struct btrfs_root *tree_root = fs_info->tree_root;
 251         struct btrfs_root *extent_root = fs_info->extent_root;
 252
 253         btrfs_write_dirty_block_groups(trans, fs_info->extent_root);
 254         while(1) {
 255                 old_extent_bytenr = btrfs_root_bytenr(&extent_root->root_item);
 256                 if (old_extent_bytenr == extent_root->node->bytenr)
 257                         break;
 258                 btrfs_set_root_bytenr(&extent_root->root_item,
 259                                        extent_root->node->bytenr);
 260                 extent_root->root_item.level =
 261                         btrfs_header_level(&extent_root->node->node.header);
 262                 ret = btrfs_update_root(trans, tree_root,
 263                                         &extent_root->root_key,
 264                                         &extent_root->root_item);
 265                 BUG_ON(ret);
 266                 btrfs_write_dirty_block_groups(trans, fs_info->extent_root);
 267         }
 268         return 0;
 269 }
 270
 271 int btrfs_commit_transaction(struct btrfs_trans_handle *trans, struct
 272                              btrfs_root *root, struct btrfs_super_block *s)
 273 {
 274         int ret = 0;
 275         struct btrfs_buffer *snap = root->commit_root;
 276         struct btrfs_key snap_key;
 277
 278         if (root->commit_root == root->node)
 279                 return 0;
 280
 281         memcpy(&snap_key, &root->root_key, sizeof(snap_key));
 282         root->root_key.offset++;
 283
 284         btrfs_set_root_bytenr(&root->root_item, root->node->bytenr);
 285         root->root_item.level =
 286                         btrfs_header_level(&root->node->node.header);
 287         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
 288                                 &root->root_key, &root->root_item);
 289         BUG_ON(ret);
 290
 291         ret = commit_tree_roots(trans, root->fs_info);
 292         BUG_ON(ret);
 293
 294         ret = __commit_transaction(trans, root);
 295         BUG_ON(ret);
 296
 297         write_ctree_super(trans, root, s);
 298         btrfs_finish_extent_commit(trans, root->fs_info->extent_root);
 299         btrfs_finish_extent_commit(trans, root->fs_info->tree_root);
 300
 301         root->commit_root = root->node;
 302         root->node->count++;
 303         ret = btrfs_drop_snapshot(trans, root, snap);
 304         BUG_ON(ret);
 305         ret = btrfs_del_root(trans, root->fs_info->tree_root, &snap_key);
 306         BUG_ON(ret);
 307         btrfs_free_transaction(root, trans);
 308         return ret;
 309 }
 310
 311 static int __setup_root(struct btrfs_super_block *super,
 312                         struct btrfs_root *root,
 313                         struct btrfs_fs_info *fs_info,
 314                         u64 objectid, int fp)
 315 {
 316         root->node = NULL;
 317         root->commit_root = NULL;
 318         root->sectorsize = btrfs_super_sectorsize(super);
 319         root->nodesize = btrfs_super_nodesize(super);
 320         root->leafsize = btrfs_super_leafsize(super);
 321         root->stripesize = btrfs_super_stripesize(super);
 322         root->ref_cows = 0;
 323         root->fs_info = fs_info;
 324         memset(&root->root_key, 0, sizeof(root->root_key));
 325         memset(&root->root_item, 0, sizeof(root->root_item));
 326         root->root_key.objectid = objectid;
 327         return 0;
 328 }
 329
 330 struct btrfs_buffer *read_root_block(struct btrfs_root *root, u64 bytenr,
 331                                             u8 level)
 332 {
 333         struct btrfs_buffer *node;
 334         u32 size = btrfs_level_size(root, level);
 335
 336         node = read_tree_block(root, bytenr, size);
 337         BUG_ON(!node);
 338         return node;
 339 }
 340
 341 static int find_and_setup_root(struct btrfs_super_block *super,
 342                                struct btrfs_root *tree_root,
 343                                struct btrfs_fs_info *fs_info,
 344                                u64 objectid,
 345                                struct btrfs_root *root, int fp)
 346 {
 347         int ret;
 348
 349         __setup_root(super, root, fs_info, objectid, fp);
 350         ret = btrfs_find_last_root(tree_root, objectid,
 351                                    &root->root_item, &root->root_key);
 352         BUG_ON(ret);
 353         root->node = read_root_block(root,
 354                                      btrfs_root_bytenr(&root->root_item),
 355                                      root->root_item.level);
 356         BUG_ON(!root->node);
 357         return 0;
 358 }
 359
 360 struct btrfs_root *open_ctree(char *filename, struct btrfs_super_block *super)
 361 {
 362         int fp;
 363
 364         fp = open(filename, O_CREAT | O_RDWR, 0600);
 365         if (fp < 0) {
 366                 return NULL;
 367         }
 368         return open_ctree_fd(fp, super);
 369 }
 370
 371 struct btrfs_root *open_ctree_fd(int fp, struct btrfs_super_block *super)
 372 {
 373         struct btrfs_root *root = malloc(sizeof(struct btrfs_root));
 374         struct btrfs_root *extent_root = malloc(sizeof(struct btrfs_root));
 375         struct btrfs_root *tree_root = malloc(sizeof(struct btrfs_root));
 376         struct btrfs_fs_info *fs_info = malloc(sizeof(*fs_info));
 377         int ret;
 378
 379         INIT_LIST_HEAD(&fs_info->trans);
 380         INIT_LIST_HEAD(&fs_info->cache);
 381         cache_tree_init(&fs_info->extent_cache);
 382         cache_tree_init(&fs_info->pending_tree);
 383         cache_tree_init(&fs_info->pinned_tree);
 384         cache_tree_init(&fs_info->del_pending);
 385         cache_tree_init(&fs_info->block_group_cache);
 386         fs_info->cache_size = 0;
 387         fs_info->fp = fp;
 388         fs_info->running_transaction = NULL;
 389         fs_info->fs_root = root;
 390         fs_info->tree_root = tree_root;
 391         fs_info->extent_root = extent_root;
 392         fs_info->last_inode_alloc = 0;
 393         fs_info->last_inode_alloc_dirid = 0;
 394         fs_info->disk_super = super;
 395         memset(&fs_info->last_insert, 0, sizeof(fs_info->last_insert));
 396
 397         ret = pread(fp, super, sizeof(struct btrfs_super_block),
 398                      BTRFS_SUPER_INFO_OFFSET);
 399         if (ret == 0 || btrfs_super_root(super) == 0) {
 400                 BUG();
 401                 return NULL;
 402         }
 403         BUG_ON(ret < 0);
 404
 405         __setup_root(super, tree_root, fs_info, BTRFS_ROOT_TREE_OBJECTID, fp);
 406         tree_root->node = read_root_block(tree_root, btrfs_super_root(super),
 407                                           btrfs_super_root_level(super));
 408         BUG_ON(!tree_root->node);
 409
 410         ret = find_and_setup_root(super, tree_root, fs_info,
 411                                   BTRFS_EXTENT_TREE_OBJECTID, extent_root, fp);
 412         BUG_ON(ret);
 413
 414         ret = find_and_setup_root(super, tree_root, fs_info,
 415                                   BTRFS_FS_TREE_OBJECTID, root, fp);
 416         BUG_ON(ret);
 417
 418         root->commit_root = root->node;
 419         root->node->count++;
 420         root->ref_cows = 1;
 421         root->fs_info->generation = btrfs_super_generation(super) + 1;
 422         btrfs_read_block_groups(root);
 423         return root;
 424 }
 425
 426 int write_ctree_super(struct btrfs_trans_handle *trans, struct btrfs_root
 427                       *root, struct btrfs_super_block *s)
 428 {
 429         int ret;
 430         btrfs_set_super_root(s, root->fs_info->tree_root->node->bytenr);
 431         btrfs_set_super_generation(s, trans->transid);
 432         btrfs_set_super_root_level(s,
 433               btrfs_header_level(&root->fs_info->tree_root->node->node.header));
 434         btrfs_csum_super(root, s);
 435
 436         ret = pwrite(root->fs_info->fp, s, sizeof(*s),
 437                      BTRFS_SUPER_INFO_OFFSET);
 438         if (ret != sizeof(*s)) {
 439                 fprintf(stderr, "failed to write new super block err %d\n", ret);
 440                 return ret;
 441         }
 442         return 0;
 443 }
 444
 445 static int drop_cache(struct btrfs_root *root)
 446 {
 447         while(!list_empty(&root->fs_info->cache)) {
 448                 struct btrfs_buffer *b = list_entry(root->fs_info->cache.next,
 449                                                     struct btrfs_buffer,
 450                                                     cache);
 451                 list_del_init(&b->cache);
 452                 btrfs_block_release(root, b);
 453         }
 454         return 0;
 455 }
 456
 457 int close_ctree(struct btrfs_root *root, struct btrfs_super_block *s)
 458 {
 459         int ret;
 460         struct btrfs_trans_handle *trans;
 461         trans = btrfs_start_transaction(root, 1);
 462         btrfs_commit_transaction(trans, root, s);
 463         trans = btrfs_start_transaction(root, 1);
 464         ret = commit_tree_roots(trans, root->fs_info);
 465         BUG_ON(ret);
 466         ret = __commit_transaction(trans, root);
 467         BUG_ON(ret);
 468         write_ctree_super(trans, root, s);
 469         btrfs_free_transaction(root, trans);
 470         drop_cache(root);
 471         BUG_ON(!list_empty(&root->fs_info->trans));
 472         btrfs_free_block_groups(root->fs_info);
 473         close(root->fs_info->fp);
 474         if (root->node)
 475                 btrfs_block_release(root, root->node);
 476         if (root->fs_info->extent_root->node)
 477                 btrfs_block_release(root->fs_info->extent_root,
 478                                     root->fs_info->extent_root->node);
 479         if (root->fs_info->tree_root->node)
 480                 btrfs_block_release(root->fs_info->tree_root,
 481                                     root->fs_info->tree_root->node);
 482         btrfs_block_release(root, root->commit_root);
 483         free(root);
 484         printf("on close %llu blocks are allocated\n",
 485                (unsigned long long)allocated_bytes);
 486         return 0;
 487 }
 488
 489 void btrfs_block_release(struct btrfs_root *root, struct btrfs_buffer *buf)
 490 {
 491         buf->count--;
 492         if (buf->count < 0)
 493                 BUG();
 494         if (buf->count == 0) {
 495                 BUG_ON(!list_empty(&buf->cache));
 496                 BUG_ON(!list_empty(&buf->dirty));
 497
 498                 remove_cache_extent(&root->fs_info->extent_cache,
 499                                     &buf->cache_node);
 500                 BUG_ON(allocated_bytes == 0);
 501                 allocated_bytes -= buf->size;
 502                 BUG_ON(root->fs_info->cache_size == 0);
 503                 root->fs_info->cache_size -= buf->size;
 504
 505                 memset(buf, 0, sizeof(*buf));
 506                 free(buf);
 507         }
 508 }
 509