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 int allocated_blocks = 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_blocknr = logical;
  42         return 0;
  43 }
  44
  45 static int check_tree_block(struct btrfs_root *root, struct btrfs_buffer *buf)
  46 {
  47         if (buf->blocknr != btrfs_header_blocknr(&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 blocknr)
  75 {
  76         struct btrfs_buffer *buf;
  77         int ret;
  78
  79         buf = malloc(sizeof(struct btrfs_buffer) + root->sectorsize);
  80         if (!buf)
  81                 return buf;
  82         allocated_blocks++;
  83         buf->blocknr = blocknr;
  84         buf->count = 2;
  85         INIT_LIST_HEAD(&buf->dirty);
  86         free_some_buffers(root);
  87         radix_tree_preload(GFP_KERNEL);
  88         ret = radix_tree_insert(&root->fs_info->cache_radix, blocknr, buf);
  89         radix_tree_preload_end();
  90         list_add_tail(&buf->cache, &root->fs_info->cache);
  91         root->fs_info->cache_size++;
  92         if (ret) {
  93                 free(buf);
  94                 return NULL;
  95         }
  96         return buf;
  97 }
  98
  99 struct btrfs_buffer *find_tree_block(struct btrfs_root *root, u64 blocknr)
 100 {
 101         struct btrfs_buffer *buf;
 102         buf = radix_tree_lookup(&root->fs_info->cache_radix, blocknr);
 103         if (buf) {
 104                 buf->count++;
 105         } else {
 106                 buf = alloc_tree_block(root, blocknr);
 107                 if (!buf) {
 108                         BUG();
 109                         return NULL;
 110                 }
 111         }
 112         return buf;
 113 }
 114
 115 struct btrfs_buffer *read_tree_block(struct btrfs_root *root, u64 blocknr)
 116 {
 117         struct btrfs_buffer *buf;
 118         int ret;
 119         buf = radix_tree_lookup(&root->fs_info->cache_radix, blocknr);
 120         if (buf) {
 121                 buf->count++;
 122                 if (check_tree_block(root, buf))
 123                         BUG();
 124         } else {
 125                 buf = alloc_tree_block(root, blocknr);
 126                 if (!buf)
 127                         return NULL;
 128                 btrfs_map_bh_to_logical(root, buf, blocknr);
 129                 ret = pread(buf->fd, &buf->node, root->sectorsize,
 130                             buf->dev_blocknr * root->sectorsize);
 131                 if (ret != root->sectorsize) {
 132                         free(buf);
 133                         return NULL;
 134                 }
 135                 if (check_tree_block(root, buf))
 136                         BUG();
 137         }
 138         return buf;
 139 }
 140
 141 int dirty_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 142                      struct btrfs_buffer *buf)
 143 {
 144         if (!list_empty(&buf->dirty))
 145                 return 0;
 146         list_add_tail(&buf->dirty, &root->fs_info->trans);
 147         buf->count++;
 148         if (check_tree_block(root, buf))
 149                 BUG();
 150         return 0;
 151 }
 152
 153 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 154                      struct btrfs_buffer *buf)
 155 {
 156         if (!list_empty(&buf->dirty)) {
 157                 list_del_init(&buf->dirty);
 158                 btrfs_block_release(root, buf);
 159         }
 160         return 0;
 161 }
 162
 163 int btrfs_csum_node(struct btrfs_root *root, struct btrfs_node *node)
 164 {
 165         u32 crc;
 166         size_t len = root->sectorsize - BTRFS_CSUM_SIZE;
 167
 168         crc = crc32c(0, (char *)(node) + BTRFS_CSUM_SIZE, len);
 169         memcpy(node->header.csum, &crc, BTRFS_CRC32_SIZE);
 170         return 0;
 171 }
 172
 173 int btrfs_csum_super(struct btrfs_root *root, struct btrfs_super_block *super)
 174 {
 175         u32 crc;
 176         char block[root->sectorsize];
 177         size_t len = root->sectorsize - BTRFS_CSUM_SIZE;
 178
 179         memset(block, 0, root->sectorsize);
 180         memcpy(block, super, sizeof(*super));
 181
 182         crc = crc32c(0, block + BTRFS_CSUM_SIZE, len);
 183         memcpy(super->csum, &crc, BTRFS_CRC32_SIZE);
 184         return 0;
 185 }
 186
 187 int write_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 188                      struct btrfs_buffer *buf)
 189 {
 190         int ret;
 191
 192         if (buf->blocknr != btrfs_header_blocknr(&buf->node.header))
 193                 BUG();
 194         btrfs_map_bh_to_logical(root, buf, buf->blocknr);
 195         if (check_tree_block(root, buf))
 196                 BUG();
 197
 198         btrfs_csum_node(root, &buf->node);
 199
 200         ret = pwrite(buf->fd, &buf->node, root->sectorsize,
 201                      buf->dev_blocknr * root->sectorsize);
 202         if (ret != root->sectorsize)
 203                 return ret;
 204         return 0;
 205 }
 206
 207 static int __commit_transaction(struct btrfs_trans_handle *trans, struct
 208                                 btrfs_root *root)
 209 {
 210         struct btrfs_buffer *b;
 211         int ret = 0;
 212         int wret;
 213         while(!list_empty(&root->fs_info->trans)) {
 214                 b = list_entry(root->fs_info->trans.next, struct btrfs_buffer,
 215                                dirty);
 216                 list_del_init(&b->dirty);
 217                 wret = write_tree_block(trans, root, b);
 218                 if (wret)
 219                         ret = wret;
 220                 btrfs_block_release(root, b);
 221         }
 222         return ret;
 223 }
 224
 225 static int commit_tree_roots(struct btrfs_trans_handle *trans,
 226                              struct btrfs_fs_info *fs_info)
 227 {
 228         int ret;
 229         u64 old_extent_block;
 230         struct btrfs_root *tree_root = fs_info->tree_root;
 231         struct btrfs_root *extent_root = fs_info->extent_root;
 232
 233         btrfs_write_dirty_block_groups(trans, fs_info->extent_root);
 234         while(1) {
 235                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
 236                 if (old_extent_block == extent_root->node->blocknr)
 237                         break;
 238                 btrfs_set_root_blocknr(&extent_root->root_item,
 239                                        extent_root->node->blocknr);
 240                 ret = btrfs_update_root(trans, tree_root,
 241                                         &extent_root->root_key,
 242                                         &extent_root->root_item);
 243                 BUG_ON(ret);
 244                 btrfs_write_dirty_block_groups(trans, fs_info->extent_root);
 245         }
 246         return 0;
 247 }
 248
 249 int btrfs_commit_transaction(struct btrfs_trans_handle *trans, struct
 250                              btrfs_root *root, struct btrfs_super_block *s)
 251 {
 252         int ret = 0;
 253         struct btrfs_buffer *snap = root->commit_root;
 254         struct btrfs_key snap_key;
 255
 256         if (root->commit_root == root->node)
 257                 return 0;
 258
 259         memcpy(&snap_key, &root->root_key, sizeof(snap_key));
 260         root->root_key.offset++;
 261
 262         btrfs_set_root_blocknr(&root->root_item, root->node->blocknr);
 263         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
 264                                 &root->root_key, &root->root_item);
 265         BUG_ON(ret);
 266
 267         ret = commit_tree_roots(trans, root->fs_info);
 268         BUG_ON(ret);
 269
 270         ret = __commit_transaction(trans, root);
 271         BUG_ON(ret);
 272
 273         write_ctree_super(trans, root, s);
 274         btrfs_finish_extent_commit(trans, root->fs_info->extent_root);
 275         btrfs_finish_extent_commit(trans, root->fs_info->tree_root);
 276
 277         root->commit_root = root->node;
 278         root->node->count++;
 279         ret = btrfs_drop_snapshot(trans, root, snap);
 280         BUG_ON(ret);
 281
 282         ret = btrfs_del_root(trans, root->fs_info->tree_root, &snap_key);
 283         BUG_ON(ret);
 284         root->fs_info->generation = root->root_key.offset + 1;
 285
 286         return ret;
 287 }
 288
 289 static int __setup_root(struct btrfs_super_block *super,
 290                         struct btrfs_root *root,
 291                         struct btrfs_fs_info *fs_info,
 292                         u64 objectid, int fp)
 293 {
 294         root->node = NULL;
 295         root->commit_root = NULL;
 296         root->sectorsize = btrfs_super_sectorsize(super);
 297         root->nodesize = btrfs_super_nodesize(super);
 298         root->leafsize = btrfs_super_leafsize(super);
 299         root->ref_cows = 0;
 300         root->fs_info = fs_info;
 301         memset(&root->root_key, 0, sizeof(root->root_key));
 302         memset(&root->root_item, 0, sizeof(root->root_item));
 303         root->root_key.objectid = objectid;
 304         return 0;
 305 }
 306
 307 static int find_and_setup_root(struct btrfs_super_block *super,
 308                                struct btrfs_root *tree_root,
 309                                struct btrfs_fs_info *fs_info,
 310                                u64 objectid,
 311                                struct btrfs_root *root, int fp)
 312 {
 313         int ret;
 314
 315         __setup_root(super, root, fs_info, objectid, fp);
 316         ret = btrfs_find_last_root(tree_root, objectid,
 317                                    &root->root_item, &root->root_key);
 318         BUG_ON(ret);
 319
 320         root->node = read_tree_block(root,
 321                                      btrfs_root_blocknr(&root->root_item));
 322         BUG_ON(!root->node);
 323         return 0;
 324 }
 325
 326 struct btrfs_root *open_ctree(char *filename, struct btrfs_super_block *super)
 327 {
 328         int fp;
 329
 330         fp = open(filename, O_CREAT | O_RDWR, 0600);
 331         if (fp < 0) {
 332                 return NULL;
 333         }
 334         return open_ctree_fd(fp, super);
 335 }
 336
 337 struct btrfs_root *open_ctree_fd(int fp, struct btrfs_super_block *super)
 338 {
 339         struct btrfs_root *root = malloc(sizeof(struct btrfs_root));
 340         struct btrfs_root *extent_root = malloc(sizeof(struct btrfs_root));
 341         struct btrfs_root *tree_root = malloc(sizeof(struct btrfs_root));
 342         struct btrfs_fs_info *fs_info = malloc(sizeof(*fs_info));
 343         int ret;
 344
 345         INIT_RADIX_TREE(&fs_info->cache_radix, GFP_KERNEL);
 346         INIT_RADIX_TREE(&fs_info->block_group_radix, GFP_KERNEL);
 347         INIT_LIST_HEAD(&fs_info->trans);
 348         INIT_LIST_HEAD(&fs_info->cache);
 349         pending_tree_init(&fs_info->pending_tree);
 350         pending_tree_init(&fs_info->pinned_tree);
 351         pending_tree_init(&fs_info->del_pending);
 352         fs_info->cache_size = 0;
 353         fs_info->fp = fp;
 354         fs_info->running_transaction = NULL;
 355         fs_info->fs_root = root;
 356         fs_info->tree_root = tree_root;
 357         fs_info->extent_root = extent_root;
 358         fs_info->last_inode_alloc = 0;
 359         fs_info->last_inode_alloc_dirid = 0;
 360         fs_info->disk_super = super;
 361         memset(&fs_info->last_insert, 0, sizeof(fs_info->last_insert));
 362
 363         ret = pread(fp, super, sizeof(struct btrfs_super_block),
 364                      BTRFS_SUPER_INFO_OFFSET);
 365         if (ret == 0 || btrfs_super_root(super) == 0) {
 366                 BUG();
 367                 return NULL;
 368         }
 369         BUG_ON(ret < 0);
 370
 371         __setup_root(super, tree_root, fs_info, BTRFS_ROOT_TREE_OBJECTID, fp);
 372         tree_root->node = read_tree_block(tree_root, btrfs_super_root(super));
 373         BUG_ON(!tree_root->node);
 374
 375         ret = find_and_setup_root(super, tree_root, fs_info,
 376                                   BTRFS_EXTENT_TREE_OBJECTID, extent_root, fp);
 377         BUG_ON(ret);
 378
 379         ret = find_and_setup_root(super, tree_root, fs_info,
 380                                   BTRFS_FS_TREE_OBJECTID, root, fp);
 381         BUG_ON(ret);
 382
 383         root->commit_root = root->node;
 384         root->node->count++;
 385         root->ref_cows = 1;
 386         root->fs_info->generation = root->root_key.offset + 1;
 387         btrfs_read_block_groups(root);
 388         return root;
 389 }
 390
 391 int write_ctree_super(struct btrfs_trans_handle *trans, struct btrfs_root
 392                       *root, struct btrfs_super_block *s)
 393 {
 394         int ret;
 395
 396         btrfs_set_super_root(s, root->fs_info->tree_root->node->blocknr);
 397         btrfs_csum_super(root, s);
 398
 399         ret = pwrite(root->fs_info->fp, s, sizeof(*s),
 400                      BTRFS_SUPER_INFO_OFFSET);
 401         if (ret != sizeof(*s)) {
 402                 fprintf(stderr, "failed to write new super block err %d\n", ret);
 403                 return ret;
 404         }
 405         return 0;
 406 }
 407
 408 static int drop_cache(struct btrfs_root *root)
 409 {
 410         while(!list_empty(&root->fs_info->cache)) {
 411                 struct btrfs_buffer *b = list_entry(root->fs_info->cache.next,
 412                                                     struct btrfs_buffer,
 413                                                     cache);
 414                 list_del_init(&b->cache);
 415                 btrfs_block_release(root, b);
 416         }
 417         return 0;
 418 }
 419
 420 int close_ctree(struct btrfs_root *root, struct btrfs_super_block *s)
 421 {
 422         int ret;
 423         struct btrfs_trans_handle *trans;
 424
 425         trans = root->fs_info->running_transaction;
 426         btrfs_commit_transaction(trans, root, s);
 427         ret = commit_tree_roots(trans, root->fs_info);
 428         BUG_ON(ret);
 429         ret = __commit_transaction(trans, root);
 430         BUG_ON(ret);
 431         write_ctree_super(trans, root, s);
 432         drop_cache(root);
 433         BUG_ON(!list_empty(&root->fs_info->trans));
 434
 435         btrfs_free_block_groups(root->fs_info);
 436         close(root->fs_info->fp);
 437         if (root->node)
 438                 btrfs_block_release(root, root->node);
 439         if (root->fs_info->extent_root->node)
 440                 btrfs_block_release(root->fs_info->extent_root,
 441                                     root->fs_info->extent_root->node);
 442         if (root->fs_info->tree_root->node)
 443                 btrfs_block_release(root->fs_info->tree_root,
 444                                     root->fs_info->tree_root->node);
 445         btrfs_block_release(root, root->commit_root);
 446         free(root);
 447         printf("on close %d blocks are allocated\n", allocated_blocks);
 448         return 0;
 449 }
 450
 451 void btrfs_block_release(struct btrfs_root *root, struct btrfs_buffer *buf)
 452 {
 453         buf->count--;
 454         if (buf->count < 0)
 455                 BUG();
 456         if (buf->count == 0) {
 457                 BUG_ON(!list_empty(&buf->cache));
 458                 BUG_ON(!list_empty(&buf->dirty));
 459                 if (!radix_tree_lookup(&root->fs_info->cache_radix,
 460                                        buf->blocknr))
 461                         BUG();
 462                 radix_tree_delete(&root->fs_info->cache_radix, buf->blocknr);
 463                 memset(buf, 0, sizeof(*buf));
 464                 free(buf);
 465                 BUG_ON(allocated_blocks == 0);
 466                 allocated_blocks--;
 467                 BUG_ON(root->fs_info->cache_size == 0);
 468                 root->fs_info->cache_size--;
 469         }
 470 }
 471