SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/btrfs/transaction.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 <linux/fs.h>
  20 #include <linux/sched.h>
  21 #include "ctree.h"
  22 #include "disk-io.h"
  23 #include "transaction.h"
  24
  25 static int total_trans = 0;
  26 extern struct kmem_cache *btrfs_trans_handle_cachep;
  27 extern struct kmem_cache *btrfs_transaction_cachep;
  28
  29 static struct workqueue_struct *trans_wq;
  30
  31 #define BTRFS_ROOT_TRANS_TAG 0
  32
  33 static void put_transaction(struct btrfs_transaction *transaction)
  34 {
  35         WARN_ON(transaction->use_count == 0);
  36         transaction->use_count--;
  37         if (transaction->use_count == 0) {
  38                 WARN_ON(total_trans == 0);
  39                 total_trans--;
  40                 list_del_init(&transaction->list);
  41                 memset(transaction, 0, sizeof(*transaction));
  42                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  43         }
  44 }
  45
  46 static int join_transaction(struct btrfs_root *root)
  47 {
  48         struct btrfs_transaction *cur_trans;
  49         cur_trans = root->fs_info->running_transaction;
  50         if (!cur_trans) {
  51                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  52                                              GFP_NOFS);
  53                 total_trans++;
  54                 BUG_ON(!cur_trans);
  55                 root->fs_info->generation++;
  56                 root->fs_info->running_transaction = cur_trans;
  57                 cur_trans->num_writers = 0;
  58                 cur_trans->transid = root->fs_info->generation;
  59                 init_waitqueue_head(&cur_trans->writer_wait);
  60                 init_waitqueue_head(&cur_trans->commit_wait);
  61                 cur_trans->in_commit = 0;
  62                 cur_trans->use_count = 1;
  63                 cur_trans->commit_done = 0;
  64                 cur_trans->start_time = get_seconds();
  65                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  66                 init_bit_radix(&cur_trans->dirty_pages);
  67         }
  68         cur_trans->num_writers++;
  69         return 0;
  70 }
  71
  72 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
  73                                                    int num_blocks)
  74 {
  75         struct btrfs_trans_handle *h =
  76                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
  77         int ret;
  78         u64 running_trans_id;
  79
  80         mutex_lock(&root->fs_info->trans_mutex);
  81         ret = join_transaction(root);
  82         BUG_ON(ret);
  83         running_trans_id = root->fs_info->running_transaction->transid;
  84
  85         if (root != root->fs_info->tree_root && root->last_trans <
  86             running_trans_id) {
  87                 WARN_ON(root == root->fs_info->extent_root);
  88                 WARN_ON(root->ref_cows != 1);
  89                 if (root->root_item.refs != 0) {
  90                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
  91                                            (unsigned long)root->root_key.objectid,
  92                                            BTRFS_ROOT_TRANS_TAG);
  93                         root->commit_root = root->node;
  94                         get_bh(root->node);
  95                 } else {
  96                         WARN_ON(1);
  97                 }
  98         }
  99         root->last_trans = running_trans_id;
 100         h->transid = running_trans_id;
 101         h->transaction = root->fs_info->running_transaction;
 102         h->blocks_reserved = num_blocks;
 103         h->blocks_used = 0;
 104         h->block_group = NULL;
 105         root->fs_info->running_transaction->use_count++;
 106         mutex_unlock(&root->fs_info->trans_mutex);
 107         return h;
 108 }
 109
 110 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 111                           struct btrfs_root *root)
 112 {
 113         struct btrfs_transaction *cur_trans;
 114
 115         mutex_lock(&root->fs_info->trans_mutex);
 116         cur_trans = root->fs_info->running_transaction;
 117         WARN_ON(cur_trans != trans->transaction);
 118         WARN_ON(cur_trans->num_writers < 1);
 119         cur_trans->num_writers--;
 120         if (waitqueue_active(&cur_trans->writer_wait))
 121                 wake_up(&cur_trans->writer_wait);
 122         put_transaction(cur_trans);
 123         mutex_unlock(&root->fs_info->trans_mutex);
 124         memset(trans, 0, sizeof(*trans));
 125         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 126         return 0;
 127 }
 128
 129
 130 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 131                                      struct btrfs_root *root)
 132 {
 133         unsigned long gang[16];
 134         int ret;
 135         int i;
 136         int err;
 137         int werr = 0;
 138         struct page *page;
 139         struct radix_tree_root *dirty_pages;
 140         struct inode *btree_inode = root->fs_info->btree_inode;
 141
 142         if (!trans || !trans->transaction) {
 143                 return filemap_write_and_wait(btree_inode->i_mapping);
 144         }
 145         dirty_pages = &trans->transaction->dirty_pages;
 146         while(1) {
 147                 ret = find_first_radix_bit(dirty_pages, gang,
 148                                            0, ARRAY_SIZE(gang));
 149                 if (!ret)
 150                         break;
 151                 for (i = 0; i < ret; i++) {
 152                         /* FIXME EIO */
 153                         clear_radix_bit(dirty_pages, gang[i]);
 154                         page = find_lock_page(btree_inode->i_mapping,
 155                                               gang[i]);
 156                         if (!page)
 157                                 continue;
 158                         err = write_one_page(page, 0);
 159                         if (err)
 160                                 werr = err;
 161                         page_cache_release(page);
 162                 }
 163         }
 164         err = filemap_fdatawait(btree_inode->i_mapping);
 165         if (err)
 166                 werr = err;
 167         return werr;
 168 }
 169
 170 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 171                             struct btrfs_root *root)
 172 {
 173         int ret;
 174         u64 old_extent_block;
 175         struct btrfs_fs_info *fs_info = root->fs_info;
 176         struct btrfs_root *tree_root = fs_info->tree_root;
 177         struct btrfs_root *extent_root = fs_info->extent_root;
 178
 179         btrfs_write_dirty_block_groups(trans, extent_root);
 180         while(1) {
 181                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
 182                 if (old_extent_block == bh_blocknr(extent_root->node))
 183                         break;
 184                 btrfs_set_root_blocknr(&extent_root->root_item,
 185                                        bh_blocknr(extent_root->node));
 186                 ret = btrfs_update_root(trans, tree_root,
 187                                         &extent_root->root_key,
 188                                         &extent_root->root_item);
 189                 BUG_ON(ret);
 190                 btrfs_write_dirty_block_groups(trans, extent_root);
 191         }
 192         return 0;
 193 }
 194
 195 static int wait_for_commit(struct btrfs_root *root,
 196                            struct btrfs_transaction *commit)
 197 {
 198         DEFINE_WAIT(wait);
 199         mutex_lock(&root->fs_info->trans_mutex);
 200         while(!commit->commit_done) {
 201                 prepare_to_wait(&commit->commit_wait, &wait,
 202                                 TASK_UNINTERRUPTIBLE);
 203                 if (commit->commit_done)
 204                         break;
 205                 mutex_unlock(&root->fs_info->trans_mutex);
 206                 schedule();
 207                 mutex_lock(&root->fs_info->trans_mutex);
 208         }
 209         mutex_unlock(&root->fs_info->trans_mutex);
 210         finish_wait(&commit->commit_wait, &wait);
 211         return 0;
 212 }
 213
 214 struct dirty_root {
 215         struct list_head list;
 216         struct btrfs_key snap_key;
 217         struct buffer_head *commit_root;
 218         struct btrfs_root *root;
 219         int free_on_drop;
 220 };
 221
 222 int btrfs_add_dead_root(struct btrfs_root *root, struct list_head *dead_list)
 223 {
 224         struct dirty_root *dirty;
 225
 226         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 227         if (!dirty)
 228                 return -ENOMEM;
 229         memcpy(&dirty->snap_key, &root->root_key, sizeof(root->root_key));
 230         dirty->commit_root = root->node;
 231         dirty->root = root;
 232         dirty->free_on_drop = 1;
 233         list_add(&dirty->list, dead_list);
 234         return 0;
 235 }
 236
 237 static int add_dirty_roots(struct btrfs_trans_handle *trans,
 238                            struct radix_tree_root *radix,
 239                            struct list_head *list)
 240 {
 241         struct dirty_root *dirty;
 242         struct btrfs_root *gang[8];
 243         struct btrfs_root *root;
 244         struct btrfs_root_item tmp_item;
 245         int i;
 246         int ret;
 247         int err = 0;
 248         u32 refs;
 249
 250         while(1) {
 251                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 252                                                  ARRAY_SIZE(gang),
 253                                                  BTRFS_ROOT_TRANS_TAG);
 254                 if (ret == 0)
 255                         break;
 256                 for (i = 0; i < ret; i++) {
 257                         root = gang[i];
 258                         radix_tree_tag_clear(radix,
 259                                      (unsigned long)root->root_key.objectid,
 260                                      BTRFS_ROOT_TRANS_TAG);
 261                         if (root->commit_root == root->node) {
 262                                 WARN_ON(bh_blocknr(root->node) !=
 263                                         btrfs_root_blocknr(&root->root_item));
 264                                 brelse(root->commit_root);
 265                                 root->commit_root = NULL;
 266                                 continue;
 267                         }
 268                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 269                         BUG_ON(!dirty);
 270                         memcpy(&dirty->snap_key, &root->root_key,
 271                                sizeof(root->root_key));
 272                         dirty->commit_root = root->commit_root;
 273                         root->commit_root = NULL;
 274                         dirty->root = root;
 275                         dirty->free_on_drop = 0;
 276                         memcpy(&tmp_item, &root->root_item, sizeof(tmp_item));
 277
 278                         root->root_key.offset = root->fs_info->generation;
 279                         btrfs_set_root_blocknr(&root->root_item,
 280                                                bh_blocknr(root->node));
 281                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 282                                                 &root->root_key,
 283                                                 &root->root_item);
 284                         if (err)
 285                                 break;
 286                         refs = btrfs_root_refs(&tmp_item);
 287                         btrfs_set_root_refs(&tmp_item, refs - 1);
 288                         err = btrfs_update_root(trans, root->fs_info->tree_root,
 289                                                 &dirty->snap_key,
 290                                                 &tmp_item);
 291
 292                         BUG_ON(err);
 293                         if (refs == 1)
 294                                 list_add(&dirty->list, list);
 295                         else
 296                                 kfree(dirty);
 297                 }
 298         }
 299         return err;
 300 }
 301
 302 static int drop_dirty_roots(struct btrfs_root *tree_root,
 303                             struct list_head *list)
 304 {
 305         struct dirty_root *dirty;
 306         struct btrfs_trans_handle *trans;
 307         int ret = 0;
 308         while(!list_empty(list)) {
 309                 mutex_lock(&tree_root->fs_info->fs_mutex);
 310                 dirty = list_entry(list->next, struct dirty_root, list);
 311                 list_del_init(&dirty->list);
 312
 313                 trans = btrfs_start_transaction(tree_root, 1);
 314                 ret = btrfs_drop_snapshot(trans, dirty->root,
 315                                           dirty->commit_root);
 316                 BUG_ON(ret);
 317                 ret = btrfs_del_root(trans, tree_root, &dirty->snap_key);
 318                 if (ret)
 319                         break;
 320                 ret = btrfs_end_transaction(trans, tree_root);
 321                 BUG_ON(ret);
 322
 323                 if (dirty->free_on_drop)
 324                         kfree(dirty->root);
 325                 kfree(dirty);
 326                 mutex_unlock(&tree_root->fs_info->fs_mutex);
 327                 btrfs_btree_balance_dirty(tree_root);
 328         }
 329         return ret;
 330 }
 331
 332 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 333                              struct btrfs_root *root)
 334 {
 335         int ret = 0;
 336         struct btrfs_transaction *cur_trans;
 337         struct btrfs_transaction *prev_trans = NULL;
 338         struct list_head dirty_fs_roots;
 339         struct radix_tree_root pinned_copy;
 340         DEFINE_WAIT(wait);
 341
 342         init_bit_radix(&pinned_copy);
 343         INIT_LIST_HEAD(&dirty_fs_roots);
 344
 345         mutex_lock(&root->fs_info->trans_mutex);
 346         if (trans->transaction->in_commit) {
 347                 cur_trans = trans->transaction;
 348                 trans->transaction->use_count++;
 349                 mutex_unlock(&root->fs_info->trans_mutex);
 350                 btrfs_end_transaction(trans, root);
 351
 352                 mutex_unlock(&root->fs_info->fs_mutex);
 353                 ret = wait_for_commit(root, cur_trans);
 354                 BUG_ON(ret);
 355                 put_transaction(cur_trans);
 356                 mutex_lock(&root->fs_info->fs_mutex);
 357                 return 0;
 358         }
 359         trans->transaction->in_commit = 1;
 360         cur_trans = trans->transaction;
 361         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 362                 prev_trans = list_entry(cur_trans->list.prev,
 363                                         struct btrfs_transaction, list);
 364                 if (!prev_trans->commit_done) {
 365                         prev_trans->use_count++;
 366                         mutex_unlock(&root->fs_info->fs_mutex);
 367                         mutex_unlock(&root->fs_info->trans_mutex);
 368
 369                         wait_for_commit(root, prev_trans);
 370                         put_transaction(prev_trans);
 371
 372                         mutex_lock(&root->fs_info->fs_mutex);
 373                         mutex_lock(&root->fs_info->trans_mutex);
 374                 }
 375         }
 376         while (trans->transaction->num_writers > 1) {
 377                 WARN_ON(cur_trans != trans->transaction);
 378                 prepare_to_wait(&trans->transaction->writer_wait, &wait,
 379                                 TASK_UNINTERRUPTIBLE);
 380                 if (trans->transaction->num_writers <= 1)
 381                         break;
 382                 mutex_unlock(&root->fs_info->fs_mutex);
 383                 mutex_unlock(&root->fs_info->trans_mutex);
 384                 schedule();
 385                 mutex_lock(&root->fs_info->fs_mutex);
 386                 mutex_lock(&root->fs_info->trans_mutex);
 387                 finish_wait(&trans->transaction->writer_wait, &wait);
 388         }
 389         finish_wait(&trans->transaction->writer_wait, &wait);
 390         WARN_ON(cur_trans != trans->transaction);
 391         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
 392                               &dirty_fs_roots);
 393         BUG_ON(ret);
 394
 395         ret = btrfs_commit_tree_roots(trans, root);
 396         BUG_ON(ret);
 397
 398         cur_trans = root->fs_info->running_transaction;
 399         root->fs_info->running_transaction = NULL;
 400         btrfs_set_super_generation(&root->fs_info->super_copy,
 401                                    cur_trans->transid);
 402         btrfs_set_super_root(&root->fs_info->super_copy,
 403                              bh_blocknr(root->fs_info->tree_root->node));
 404         memcpy(root->fs_info->disk_super, &root->fs_info->super_copy,
 405                sizeof(root->fs_info->super_copy));
 406
 407         btrfs_copy_pinned(root, &pinned_copy);
 408
 409         mutex_unlock(&root->fs_info->trans_mutex);
 410         mutex_unlock(&root->fs_info->fs_mutex);
 411         ret = btrfs_write_and_wait_transaction(trans, root);
 412         BUG_ON(ret);
 413         write_ctree_super(trans, root);
 414         mutex_lock(&root->fs_info->fs_mutex);
 415         btrfs_finish_extent_commit(trans, root, &pinned_copy);
 416         mutex_lock(&root->fs_info->trans_mutex);
 417         cur_trans->commit_done = 1;
 418         wake_up(&cur_trans->commit_wait);
 419         put_transaction(cur_trans);
 420         put_transaction(cur_trans);
 421         if (root->fs_info->closing)
 422                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
 423         else
 424                 list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
 425         mutex_unlock(&root->fs_info->trans_mutex);
 426         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 427
 428         if (root->fs_info->closing) {
 429                 mutex_unlock(&root->fs_info->fs_mutex);
 430                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
 431                 mutex_lock(&root->fs_info->fs_mutex);
 432         }
 433         return ret;
 434 }
 435
 436 void btrfs_transaction_cleaner(struct work_struct *work)
 437 {
 438         struct btrfs_fs_info *fs_info = container_of(work,
 439                                                      struct btrfs_fs_info,
 440                                                      trans_work.work);
 441
 442         struct btrfs_root *root = fs_info->tree_root;
 443         struct btrfs_transaction *cur;
 444         struct btrfs_trans_handle *trans;
 445         struct list_head dirty_roots;
 446         unsigned long now;
 447         unsigned long delay = HZ * 30;
 448         int ret;
 449
 450         INIT_LIST_HEAD(&dirty_roots);
 451         mutex_lock(&root->fs_info->fs_mutex);
 452         mutex_lock(&root->fs_info->trans_mutex);
 453         cur = root->fs_info->running_transaction;
 454         if (!cur) {
 455                 mutex_unlock(&root->fs_info->trans_mutex);
 456                 goto out;
 457         }
 458         now = get_seconds();
 459         if (now < cur->start_time || now - cur->start_time < 30) {
 460                 mutex_unlock(&root->fs_info->trans_mutex);
 461                 delay = HZ * 5;
 462                 goto out;
 463         }
 464         mutex_unlock(&root->fs_info->trans_mutex);
 465         trans = btrfs_start_transaction(root, 1);
 466         ret = btrfs_commit_transaction(trans, root);
 467 out:
 468         mutex_unlock(&root->fs_info->fs_mutex);
 469
 470         mutex_lock(&root->fs_info->trans_mutex);
 471         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
 472         mutex_unlock(&root->fs_info->trans_mutex);
 473
 474         if (!list_empty(&dirty_roots)) {
 475                 drop_dirty_roots(root, &dirty_roots);
 476         }
 477         btrfs_transaction_queue_work(root, delay);
 478 }
 479
 480 void btrfs_transaction_queue_work(struct btrfs_root *root, int delay)
 481 {
 482         queue_delayed_work(trans_wq, &root->fs_info->trans_work, delay);
 483 }
 484
 485 void btrfs_transaction_flush_work(struct btrfs_root *root)
 486 {
 487         cancel_rearming_delayed_workqueue(trans_wq, &root->fs_info->trans_work);
 488         flush_workqueue(trans_wq);
 489 }
 490
 491 void __init btrfs_init_transaction_sys(void)
 492 {
 493         trans_wq = create_workqueue("btrfs");
 494 }
 495
 496 void __exit btrfs_exit_transaction_sys(void)
 497 {
 498         destroy_workqueue(trans_wq);
 499 }
 500