fs/btrfs/transaction.c

   1 #include <linux/module.h>
   2 #include <linux/fs.h>
   3 #include "ctree.h"
   4 #include "disk-io.h"
   5 #include "transaction.h"
   6
   7 static int total_trans = 0;
   8 extern struct kmem_cache *btrfs_trans_handle_cachep;
   9 extern struct kmem_cache *btrfs_transaction_cachep;
  10
  11 static struct workqueue_struct *trans_wq;
  12
  13 #define BTRFS_ROOT_TRANS_TAG 0
  14
  15 static void put_transaction(struct btrfs_transaction *transaction)
  16 {
  17         WARN_ON(transaction->use_count == 0);
  18         transaction->use_count--;
  19         if (transaction->use_count == 0) {
  20                 WARN_ON(total_trans == 0);
  21                 total_trans--;
  22                 list_del_init(&transaction->list);
  23                 memset(transaction, 0, sizeof(*transaction));
  24                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  25         }
  26 }
  27
  28 static int join_transaction(struct btrfs_root *root)
  29 {
  30         struct btrfs_transaction *cur_trans;
  31         cur_trans = root->fs_info->running_transaction;
  32         if (!cur_trans) {
  33                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  34                                              GFP_NOFS);
  35                 total_trans++;
  36                 BUG_ON(!cur_trans);
  37                 root->fs_info->generation++;
  38                 root->fs_info->running_transaction = cur_trans;
  39                 cur_trans->num_writers = 0;
  40                 cur_trans->transid = root->fs_info->generation;
  41                 init_waitqueue_head(&cur_trans->writer_wait);
  42                 init_waitqueue_head(&cur_trans->commit_wait);
  43                 cur_trans->in_commit = 0;
  44                 cur_trans->use_count = 1;
  45                 cur_trans->commit_done = 0;
  46                 cur_trans->start_time = get_seconds();
  47                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  48                 init_bit_radix(&cur_trans->dirty_pages);
  49         }
  50         cur_trans->num_writers++;
  51         return 0;
  52 }
  53
  54 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
  55                                                    int num_blocks)
  56 {
  57         struct btrfs_trans_handle *h =
  58                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
  59         int ret;
  60         u64 running_trans_id;
  61
  62         mutex_lock(&root->fs_info->trans_mutex);
  63         ret = join_transaction(root);
  64         BUG_ON(ret);
  65         running_trans_id = root->fs_info->running_transaction->transid;
  66
  67         if (root != root->fs_info->tree_root && root->last_trans <
  68             running_trans_id) {
  69                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
  70                                    (unsigned long)root->root_key.objectid,
  71                                    BTRFS_ROOT_TRANS_TAG);
  72                 root->commit_root = root->node;
  73                 get_bh(root->node);
  74         }
  75         root->last_trans = running_trans_id;
  76         h->transid = running_trans_id;
  77         h->transaction = root->fs_info->running_transaction;
  78         h->blocks_reserved = num_blocks;
  79         h->blocks_used = 0;
  80         h->block_group = NULL;
  81         root->fs_info->running_transaction->use_count++;
  82         mutex_unlock(&root->fs_info->trans_mutex);
  83         return h;
  84 }
  85
  86 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
  87                           struct btrfs_root *root)
  88 {
  89         struct btrfs_transaction *cur_trans;
  90
  91         mutex_lock(&root->fs_info->trans_mutex);
  92         cur_trans = root->fs_info->running_transaction;
  93         WARN_ON(cur_trans->num_writers < 1);
  94         if (waitqueue_active(&cur_trans->writer_wait))
  95                 wake_up(&cur_trans->writer_wait);
  96         cur_trans->num_writers--;
  97         put_transaction(cur_trans);
  98         mutex_unlock(&root->fs_info->trans_mutex);
  99         memset(trans, 0, sizeof(*trans));
 100         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 101         return 0;
 102 }
 103
 104
 105 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 106                                      struct btrfs_root *root)
 107 {
 108         unsigned long gang[16];
 109         int ret;
 110         int i;
 111         int err;
 112         int werr = 0;
 113         struct page *page;
 114         struct radix_tree_root *dirty_pages;
 115         struct inode *btree_inode = root->fs_info->btree_inode;
 116
 117         if (!trans || !trans->transaction) {
 118                 return filemap_write_and_wait(btree_inode->i_mapping);
 119         }
 120         dirty_pages = &trans->transaction->dirty_pages;
 121         while(1) {
 122                 ret = find_first_radix_bit(dirty_pages, gang,
 123                                            0, ARRAY_SIZE(gang));
 124                 if (!ret)
 125                         break;
 126                 for (i = 0; i < ret; i++) {
 127                         /* FIXME EIO */
 128                         clear_radix_bit(dirty_pages, gang[i]);
 129                         page = find_lock_page(btree_inode->i_mapping,
 130                                               gang[i]);
 131                         if (!page)
 132                                 continue;
 133                         err = write_one_page(page, 0);
 134                         if (err)
 135                                 werr = err;
 136                         page_cache_release(page);
 137                 }
 138         }
 139         err = filemap_fdatawait(btree_inode->i_mapping);
 140         if (err)
 141                 werr = err;
 142         return werr;
 143 }
 144
 145 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 146                             struct btrfs_root *root)
 147 {
 148         int ret;
 149         u64 old_extent_block;
 150         struct btrfs_fs_info *fs_info = root->fs_info;
 151         struct btrfs_root *tree_root = fs_info->tree_root;
 152         struct btrfs_root *extent_root = fs_info->extent_root;
 153
 154         btrfs_write_dirty_block_groups(trans, extent_root);
 155         while(1) {
 156                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
 157                 if (old_extent_block == bh_blocknr(extent_root->node))
 158                         break;
 159                 btrfs_set_root_blocknr(&extent_root->root_item,
 160                                        bh_blocknr(extent_root->node));
 161                 ret = btrfs_update_root(trans, tree_root,
 162                                         &extent_root->root_key,
 163                                         &extent_root->root_item);
 164                 BUG_ON(ret);
 165                 btrfs_write_dirty_block_groups(trans, extent_root);
 166         }
 167         return 0;
 168 }
 169
 170 static int wait_for_commit(struct btrfs_root *root,
 171                            struct btrfs_transaction *commit)
 172 {
 173         DEFINE_WAIT(wait);
 174         while(!commit->commit_done) {
 175                 prepare_to_wait(&commit->commit_wait, &wait,
 176                                 TASK_UNINTERRUPTIBLE);
 177                 if (commit->commit_done)
 178                         break;
 179                 mutex_unlock(&root->fs_info->trans_mutex);
 180                 schedule();
 181                 mutex_lock(&root->fs_info->trans_mutex);
 182         }
 183         finish_wait(&commit->commit_wait, &wait);
 184         return 0;
 185 }
 186
 187 struct dirty_root {
 188         struct list_head list;
 189         struct btrfs_key snap_key;
 190         struct buffer_head *commit_root;
 191         struct btrfs_root *root;
 192 };
 193
 194 static int add_dirty_roots(struct btrfs_trans_handle *trans,
 195                            struct radix_tree_root *radix,
 196                            struct list_head *list)
 197 {
 198         struct dirty_root *dirty;
 199         struct btrfs_root *gang[8];
 200         struct btrfs_root *root;
 201         int i;
 202         int ret;
 203         int err;
 204         while(1) {
 205                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 206                                                  ARRAY_SIZE(gang),
 207                                                  BTRFS_ROOT_TRANS_TAG);
 208                 if (ret == 0)
 209                         break;
 210                 for (i = 0; i < ret; i++) {
 211                         root = gang[i];
 212                         radix_tree_tag_clear(radix,
 213                                      (unsigned long)root->root_key.objectid,
 214                                      BTRFS_ROOT_TRANS_TAG);
 215                         if (root->commit_root == root->node) {
 216                                 WARN_ON(bh_blocknr(root->node) !=
 217                                         btrfs_root_blocknr(&root->root_item));
 218                                 brelse(root->commit_root);
 219                                 root->commit_root = NULL;
 220                                 continue;
 221                         }
 222                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 223                         BUG_ON(!dirty);
 224                         memcpy(&dirty->snap_key, &root->root_key,
 225                                sizeof(root->root_key));
 226                         dirty->commit_root = root->commit_root;
 227                         root->commit_root = NULL;
 228                         dirty->root = root;
 229                         root->root_key.offset = root->fs_info->generation;
 230                         btrfs_set_root_blocknr(&root->root_item,
 231                                                bh_blocknr(root->node));
 232                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 233                                                 &root->root_key,
 234                                                 &root->root_item);
 235                         BUG_ON(err);
 236                         list_add(&dirty->list, list);
 237                 }
 238         }
 239         return 0;
 240 }
 241
 242 static int drop_dirty_roots(struct btrfs_root *tree_root,
 243                             struct list_head *list)
 244 {
 245         struct dirty_root *dirty;
 246         struct btrfs_trans_handle *trans;
 247         int ret;
 248         while(!list_empty(list)) {
 249                 mutex_lock(&tree_root->fs_info->fs_mutex);
 250                 dirty = list_entry(list->next, struct dirty_root, list);
 251                 list_del_init(&dirty->list);
 252                 trans = btrfs_start_transaction(tree_root, 1);
 253                 ret = btrfs_drop_snapshot(trans, dirty->root,
 254                                           dirty->commit_root);
 255                 BUG_ON(ret);
 256
 257                 ret = btrfs_del_root(trans, tree_root, &dirty->snap_key);
 258                 BUG_ON(ret);
 259                 ret = btrfs_end_transaction(trans, tree_root);
 260                 BUG_ON(ret);
 261                 kfree(dirty);
 262                 mutex_unlock(&tree_root->fs_info->fs_mutex);
 263         }
 264         return 0;
 265 }
 266
 267 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 268                              struct btrfs_root *root)
 269 {
 270         int ret = 0;
 271         struct btrfs_transaction *cur_trans;
 272         struct btrfs_transaction *prev_trans = NULL;
 273         struct list_head dirty_fs_roots;
 274         DEFINE_WAIT(wait);
 275
 276         INIT_LIST_HEAD(&dirty_fs_roots);
 277
 278         mutex_lock(&root->fs_info->trans_mutex);
 279         if (trans->transaction->in_commit) {
 280                 cur_trans = trans->transaction;
 281                 trans->transaction->use_count++;
 282                 btrfs_end_transaction(trans, root);
 283                 ret = wait_for_commit(root, cur_trans);
 284                 BUG_ON(ret);
 285                 put_transaction(cur_trans);
 286                 mutex_unlock(&root->fs_info->trans_mutex);
 287                 return 0;
 288         }
 289         cur_trans = trans->transaction;
 290         trans->transaction->in_commit = 1;
 291         while (trans->transaction->num_writers > 1) {
 292                 WARN_ON(cur_trans != trans->transaction);
 293                 prepare_to_wait(&trans->transaction->writer_wait, &wait,
 294                                 TASK_UNINTERRUPTIBLE);
 295                 if (trans->transaction->num_writers <= 1)
 296                         break;
 297                 mutex_unlock(&root->fs_info->trans_mutex);
 298                 schedule();
 299                 mutex_lock(&root->fs_info->trans_mutex);
 300                 finish_wait(&trans->transaction->writer_wait, &wait);
 301         }
 302         finish_wait(&trans->transaction->writer_wait, &wait);
 303         WARN_ON(cur_trans != trans->transaction);
 304         add_dirty_roots(trans, &root->fs_info->fs_roots_radix, &dirty_fs_roots);
 305         ret = btrfs_commit_tree_roots(trans, root);
 306         BUG_ON(ret);
 307         cur_trans = root->fs_info->running_transaction;
 308         root->fs_info->running_transaction = NULL;
 309         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 310                 prev_trans = list_entry(cur_trans->list.prev,
 311                                         struct btrfs_transaction, list);
 312                 if (prev_trans->commit_done)
 313                         prev_trans = NULL;
 314                 else
 315                         prev_trans->use_count++;
 316         }
 317         mutex_unlock(&root->fs_info->trans_mutex);
 318         mutex_unlock(&root->fs_info->fs_mutex);
 319         ret = btrfs_write_and_wait_transaction(trans, root);
 320         if (prev_trans) {
 321                 mutex_lock(&root->fs_info->trans_mutex);
 322                 wait_for_commit(root, prev_trans);
 323                 put_transaction(prev_trans);
 324                 mutex_unlock(&root->fs_info->trans_mutex);
 325         }
 326         btrfs_set_super_generation(root->fs_info->disk_super,
 327                                    cur_trans->transid);
 328         BUG_ON(ret);
 329         write_ctree_super(trans, root);
 330
 331         mutex_lock(&root->fs_info->fs_mutex);
 332         btrfs_finish_extent_commit(trans, root);
 333         mutex_lock(&root->fs_info->trans_mutex);
 334         cur_trans->commit_done = 1;
 335         wake_up(&cur_trans->commit_wait);
 336         put_transaction(cur_trans);
 337         put_transaction(cur_trans);
 338         if (root->fs_info->closing)
 339                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
 340         else
 341                 list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
 342         mutex_unlock(&root->fs_info->trans_mutex);
 343         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 344
 345         if (root->fs_info->closing) {
 346                 mutex_unlock(&root->fs_info->fs_mutex);
 347                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
 348                 mutex_lock(&root->fs_info->fs_mutex);
 349         }
 350         return ret;
 351 }
 352
 353 void btrfs_transaction_cleaner(struct work_struct *work)
 354 {
 355         struct btrfs_fs_info *fs_info = container_of(work,
 356                                                      struct btrfs_fs_info,
 357                                                      trans_work.work);
 358
 359         struct btrfs_root *root = fs_info->tree_root;
 360         struct btrfs_transaction *cur;
 361         struct btrfs_trans_handle *trans;
 362         struct list_head dirty_roots;
 363         unsigned long now;
 364         unsigned long delay = HZ * 30;
 365         int ret;
 366
 367         INIT_LIST_HEAD(&dirty_roots);
 368         mutex_lock(&root->fs_info->fs_mutex);
 369         mutex_lock(&root->fs_info->trans_mutex);
 370         cur = root->fs_info->running_transaction;
 371         if (!cur) {
 372                 mutex_unlock(&root->fs_info->trans_mutex);
 373                 goto out;
 374         }
 375         now = get_seconds();
 376         if (now < cur->start_time || now - cur->start_time < 30) {
 377                 mutex_unlock(&root->fs_info->trans_mutex);
 378                 delay = HZ * 5;
 379                 goto out;
 380         }
 381         mutex_unlock(&root->fs_info->trans_mutex);
 382         trans = btrfs_start_transaction(root, 1);
 383         ret = btrfs_commit_transaction(trans, root);
 384 out:
 385         mutex_unlock(&root->fs_info->fs_mutex);
 386
 387         mutex_lock(&root->fs_info->trans_mutex);
 388         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
 389         mutex_unlock(&root->fs_info->trans_mutex);
 390
 391         if (!list_empty(&dirty_roots)) {
 392                 drop_dirty_roots(root, &dirty_roots);
 393         }
 394         btrfs_transaction_queue_work(root, delay);
 395 }
 396
 397 void btrfs_transaction_queue_work(struct btrfs_root *root, int delay)
 398 {
 399         queue_delayed_work(trans_wq, &root->fs_info->trans_work, delay);
 400 }
 401
 402 void btrfs_transaction_flush_work(struct btrfs_root *root)
 403 {
 404         cancel_rearming_delayed_workqueue(trans_wq, &root->fs_info->trans_work);
 405         flush_workqueue(trans_wq);
 406 }
 407
 408 void __init btrfs_init_transaction_sys(void)
 409 {
 410         trans_wq = create_workqueue("btrfs");
 411 }
 412
 413 void __exit btrfs_exit_transaction_sys(void)
 414 {
 415         destroy_workqueue(trans_wq);
 416 }
 417