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 #define BTRFS_ROOT_TRANS_TAG 0
  12
  13 #define TRANS_MAGIC 0xE1E10E
  14 static void put_transaction(struct btrfs_transaction *transaction)
  15 {
  16         WARN_ON(transaction->use_count == 0);
  17         transaction->use_count--;
  18         WARN_ON(transaction->magic != TRANS_MAGIC);
  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->magic = TRANS_MAGIC;
  44                 cur_trans->in_commit = 0;
  45                 cur_trans->use_count = 1;
  46                 cur_trans->commit_done = 0;
  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         root->fs_info->running_transaction->use_count++;
  81         mutex_unlock(&root->fs_info->trans_mutex);
  82         h->magic = h->magic2 = TRANS_MAGIC;
  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         WARN_ON(trans->magic != TRANS_MAGIC);
  92         WARN_ON(trans->magic2 != TRANS_MAGIC);
  93         mutex_lock(&root->fs_info->trans_mutex);
  94         cur_trans = root->fs_info->running_transaction;
  95         WARN_ON(cur_trans->num_writers < 1);
  96         if (waitqueue_active(&cur_trans->writer_wait))
  97                 wake_up(&cur_trans->writer_wait);
  98         cur_trans->num_writers--;
  99         put_transaction(cur_trans);
 100         mutex_unlock(&root->fs_info->trans_mutex);
 101         memset(trans, 0, sizeof(*trans));
 102         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 103         return 0;
 104 }
 105
 106
 107 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 108                                      struct btrfs_root *root)
 109 {
 110         unsigned long gang[16];
 111         int ret;
 112         int i;
 113         int err;
 114         int werr = 0;
 115         struct page *page;
 116         struct radix_tree_root *dirty_pages;
 117         struct inode *btree_inode = root->fs_info->btree_inode;
 118
 119         if (!trans || !trans->transaction) {
 120                 return filemap_write_and_wait(btree_inode->i_mapping);
 121         }
 122         dirty_pages = &trans->transaction->dirty_pages;
 123         while(1) {
 124                 ret = find_first_radix_bit(dirty_pages, gang, ARRAY_SIZE(gang));
 125                 if (!ret)
 126                         break;
 127                 for (i = 0; i < ret; i++) {
 128                         /* FIXME EIO */
 129                         clear_radix_bit(dirty_pages, gang[i]);
 130                         page = find_lock_page(btree_inode->i_mapping,
 131                                               gang[i]);
 132                         if (!page)
 133                                 continue;
 134                         err = write_one_page(page, 0);
 135                         if (err)
 136                                 werr = err;
 137                         page_cache_release(page);
 138                 }
 139         }
 140         err = filemap_fdatawait(btree_inode->i_mapping);
 141         if (err)
 142                 werr = err;
 143         return werr;
 144 }
 145
 146 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 147                             struct btrfs_root *root)
 148 {
 149         int ret;
 150         u64 old_extent_block;
 151         struct btrfs_fs_info *fs_info = root->fs_info;
 152         struct btrfs_root *tree_root = fs_info->tree_root;
 153         struct btrfs_root *extent_root = fs_info->extent_root;
 154         struct btrfs_root *dev_root = fs_info->dev_root;
 155
 156         if (btrfs_super_device_root(fs_info->disk_super) !=
 157             bh_blocknr(dev_root->node)) {
 158                 btrfs_set_super_device_root(fs_info->disk_super,
 159                                             bh_blocknr(dev_root->node));
 160         }
 161         btrfs_write_dirty_block_groups(trans, extent_root);
 162         while(1) {
 163                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
 164                 if (old_extent_block == bh_blocknr(extent_root->node))
 165                         break;
 166                 btrfs_set_root_blocknr(&extent_root->root_item,
 167                                        bh_blocknr(extent_root->node));
 168                 ret = btrfs_update_root(trans, tree_root,
 169                                         &extent_root->root_key,
 170                                         &extent_root->root_item);
 171                 BUG_ON(ret);
 172                 btrfs_write_dirty_block_groups(trans, extent_root);
 173         }
 174         return 0;
 175 }
 176
 177 static int wait_for_commit(struct btrfs_root *root,
 178                            struct btrfs_transaction *commit)
 179 {
 180         DEFINE_WAIT(wait);
 181         while(!commit->commit_done) {
 182                 prepare_to_wait(&commit->commit_wait, &wait,
 183                                 TASK_UNINTERRUPTIBLE);
 184                 if (commit->commit_done)
 185                         break;
 186                 mutex_unlock(&root->fs_info->trans_mutex);
 187                 schedule();
 188                 mutex_lock(&root->fs_info->trans_mutex);
 189         }
 190         finish_wait(&commit->commit_wait, &wait);
 191         return 0;
 192 }
 193
 194 struct dirty_root {
 195         struct list_head list;
 196         struct btrfs_key snap_key;
 197         struct buffer_head *commit_root;
 198         struct btrfs_root *root;
 199 };
 200
 201 int add_dirty_roots(struct btrfs_trans_handle *trans,
 202                     struct radix_tree_root *radix, struct list_head *list)
 203 {
 204         struct dirty_root *dirty;
 205         struct btrfs_root *gang[8];
 206         struct btrfs_root *root;
 207         int i;
 208         int ret;
 209         int err;
 210         while(1) {
 211                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 212                                                  ARRAY_SIZE(gang),
 213                                                  BTRFS_ROOT_TRANS_TAG);
 214                 if (ret == 0)
 215                         break;
 216                 for (i = 0; i < ret; i++) {
 217                         root = gang[i];
 218                         radix_tree_tag_clear(radix,
 219                                      (unsigned long)root->root_key.objectid,
 220                                      BTRFS_ROOT_TRANS_TAG);
 221                         if (root->commit_root == root->node) {
 222                                 WARN_ON(bh_blocknr(root->node) !=
 223                                         btrfs_root_blocknr(&root->root_item));
 224                                 brelse(root->commit_root);
 225                                 root->commit_root = NULL;
 226                                 continue;
 227                         }
 228                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 229                         BUG_ON(!dirty);
 230                         memcpy(&dirty->snap_key, &root->root_key,
 231                                sizeof(root->root_key));
 232                         dirty->commit_root = root->commit_root;
 233                         root->commit_root = NULL;
 234                         dirty->root = root;
 235                         root->root_key.offset = root->fs_info->generation;
 236                         btrfs_set_root_blocknr(&root->root_item,
 237                                                bh_blocknr(root->node));
 238                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 239                                                 &root->root_key,
 240                                                 &root->root_item);
 241                         BUG_ON(err);
 242                         list_add(&dirty->list, list);
 243                 }
 244         }
 245         return 0;
 246 }
 247
 248 int drop_dirty_roots(struct btrfs_root *tree_root, struct list_head *list)
 249 {
 250         struct dirty_root *dirty;
 251         struct btrfs_trans_handle *trans;
 252         int ret;
 253
 254         while(!list_empty(list)) {
 255                 dirty = list_entry(list->next, struct dirty_root, list);
 256                 list_del_init(&dirty->list);
 257                 trans = btrfs_start_transaction(tree_root, 1);
 258                 ret = btrfs_drop_snapshot(trans, dirty->root,
 259                                           dirty->commit_root);
 260                 BUG_ON(ret);
 261
 262                 ret = btrfs_del_root(trans, tree_root, &dirty->snap_key);
 263                 BUG_ON(ret);
 264                 ret = btrfs_end_transaction(trans, tree_root);
 265                 BUG_ON(ret);
 266                 kfree(dirty);
 267         }
 268         return 0;
 269 }
 270
 271 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 272                              struct btrfs_root *root)
 273 {
 274         int ret = 0;
 275         struct btrfs_transaction *cur_trans;
 276         struct btrfs_transaction *prev_trans = NULL;
 277         struct list_head dirty_fs_roots;
 278         DEFINE_WAIT(wait);
 279
 280         INIT_LIST_HEAD(&dirty_fs_roots);
 281
 282         mutex_lock(&root->fs_info->trans_mutex);
 283         if (trans->transaction->in_commit) {
 284                 cur_trans = trans->transaction;
 285                 trans->transaction->use_count++;
 286                 btrfs_end_transaction(trans, root);
 287                 ret = wait_for_commit(root, cur_trans);
 288                 BUG_ON(ret);
 289                 put_transaction(cur_trans);
 290                 mutex_unlock(&root->fs_info->trans_mutex);
 291                 return 0;
 292         }
 293         cur_trans = trans->transaction;
 294         trans->transaction->in_commit = 1;
 295         while (trans->transaction->num_writers > 1) {
 296                 WARN_ON(cur_trans != trans->transaction);
 297                 prepare_to_wait(&trans->transaction->writer_wait, &wait,
 298                                 TASK_UNINTERRUPTIBLE);
 299                 if (trans->transaction->num_writers <= 1)
 300                         break;
 301                 mutex_unlock(&root->fs_info->trans_mutex);
 302                 schedule();
 303                 mutex_lock(&root->fs_info->trans_mutex);
 304                 finish_wait(&trans->transaction->writer_wait, &wait);
 305         }
 306         finish_wait(&trans->transaction->writer_wait, &wait);
 307         WARN_ON(cur_trans != trans->transaction);
 308         add_dirty_roots(trans, &root->fs_info->fs_roots_radix, &dirty_fs_roots);
 309         ret = btrfs_commit_tree_roots(trans, root);
 310         BUG_ON(ret);
 311         cur_trans = root->fs_info->running_transaction;
 312         root->fs_info->running_transaction = NULL;
 313         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 314                 prev_trans = list_entry(cur_trans->list.prev,
 315                                         struct btrfs_transaction, list);
 316                 if (prev_trans->commit_done)
 317                         prev_trans = NULL;
 318                 else
 319                         prev_trans->use_count++;
 320         }
 321         mutex_unlock(&root->fs_info->trans_mutex);
 322         mutex_unlock(&root->fs_info->fs_mutex);
 323         ret = btrfs_write_and_wait_transaction(trans, root);
 324         if (prev_trans) {
 325                 mutex_lock(&root->fs_info->trans_mutex);
 326                 wait_for_commit(root, prev_trans);
 327                 put_transaction(prev_trans);
 328                 mutex_unlock(&root->fs_info->trans_mutex);
 329         }
 330         btrfs_set_super_generation(root->fs_info->disk_super,
 331                                    cur_trans->transid);
 332         BUG_ON(ret);
 333         write_ctree_super(trans, root);
 334
 335         mutex_lock(&root->fs_info->fs_mutex);
 336         btrfs_finish_extent_commit(trans, root);
 337         mutex_lock(&root->fs_info->trans_mutex);
 338         cur_trans->commit_done = 1;
 339         wake_up(&cur_trans->commit_wait);
 340         put_transaction(cur_trans);
 341         put_transaction(cur_trans);
 342         mutex_unlock(&root->fs_info->trans_mutex);
 343         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 344
 345         drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
 346         return ret;
 347 }
 348