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         h->block_group = NULL;
  81         root->fs_info->running_transaction->use_count++;
  82         mutex_unlock(&root->fs_info->trans_mutex);
  83         h->magic = h->magic2 = TRANS_MAGIC;
  84         return h;
  85 }
  86
  87 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
  88                           struct btrfs_root *root)
  89 {
  90         struct btrfs_transaction *cur_trans;
  91
  92         WARN_ON(trans->magic != TRANS_MAGIC);
  93         WARN_ON(trans->magic2 != TRANS_MAGIC);
  94         mutex_lock(&root->fs_info->trans_mutex);
  95         cur_trans = root->fs_info->running_transaction;
  96         WARN_ON(cur_trans->num_writers < 1);
  97         if (waitqueue_active(&cur_trans->writer_wait))
  98                 wake_up(&cur_trans->writer_wait);
  99         cur_trans->num_writers--;
 100         put_transaction(cur_trans);
 101         mutex_unlock(&root->fs_info->trans_mutex);
 102         memset(trans, 0, sizeof(*trans));
 103         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 104         return 0;
 105 }
 106
 107
 108 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 109                                      struct btrfs_root *root)
 110 {
 111         unsigned long gang[16];
 112         int ret;
 113         int i;
 114         int err;
 115         int werr = 0;
 116         struct page *page;
 117         struct radix_tree_root *dirty_pages;
 118         struct inode *btree_inode = root->fs_info->btree_inode;
 119
 120         if (!trans || !trans->transaction) {
 121                 return filemap_write_and_wait(btree_inode->i_mapping);
 122         }
 123         dirty_pages = &trans->transaction->dirty_pages;
 124         while(1) {
 125                 ret = find_first_radix_bit(dirty_pages, gang, ARRAY_SIZE(gang));
 126                 if (!ret)
 127                         break;
 128                 for (i = 0; i < ret; i++) {
 129                         /* FIXME EIO */
 130                         clear_radix_bit(dirty_pages, gang[i]);
 131                         page = find_lock_page(btree_inode->i_mapping,
 132                                               gang[i]);
 133                         if (!page)
 134                                 continue;
 135                         err = write_one_page(page, 0);
 136                         if (err)
 137                                 werr = err;
 138                         page_cache_release(page);
 139                 }
 140         }
 141         err = filemap_fdatawait(btree_inode->i_mapping);
 142         if (err)
 143                 werr = err;
 144         return werr;
 145 }
 146
 147 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 148                             struct btrfs_root *root)
 149 {
 150         int ret;
 151         u64 old_extent_block;
 152         struct btrfs_fs_info *fs_info = root->fs_info;
 153         struct btrfs_root *tree_root = fs_info->tree_root;
 154         struct btrfs_root *extent_root = fs_info->extent_root;
 155         struct btrfs_root *dev_root = fs_info->dev_root;
 156
 157         if (btrfs_super_device_root(fs_info->disk_super) !=
 158             bh_blocknr(dev_root->node)) {
 159                 btrfs_set_super_device_root(fs_info->disk_super,
 160                                             bh_blocknr(dev_root->node));
 161         }
 162         btrfs_write_dirty_block_groups(trans, extent_root);
 163         while(1) {
 164                 old_extent_block = btrfs_root_blocknr(&extent_root->root_item);
 165                 if (old_extent_block == bh_blocknr(extent_root->node))
 166                         break;
 167                 btrfs_set_root_blocknr(&extent_root->root_item,
 168                                        bh_blocknr(extent_root->node));
 169                 ret = btrfs_update_root(trans, tree_root,
 170                                         &extent_root->root_key,
 171                                         &extent_root->root_item);
 172                 BUG_ON(ret);
 173                 btrfs_write_dirty_block_groups(trans, extent_root);
 174         }
 175         return 0;
 176 }
 177
 178 static int wait_for_commit(struct btrfs_root *root,
 179                            struct btrfs_transaction *commit)
 180 {
 181         DEFINE_WAIT(wait);
 182         while(!commit->commit_done) {
 183                 prepare_to_wait(&commit->commit_wait, &wait,
 184                                 TASK_UNINTERRUPTIBLE);
 185                 if (commit->commit_done)
 186                         break;
 187                 mutex_unlock(&root->fs_info->trans_mutex);
 188                 schedule();
 189                 mutex_lock(&root->fs_info->trans_mutex);
 190         }
 191         finish_wait(&commit->commit_wait, &wait);
 192         return 0;
 193 }
 194
 195 struct dirty_root {
 196         struct list_head list;
 197         struct btrfs_key snap_key;
 198         struct buffer_head *commit_root;
 199         struct btrfs_root *root;
 200 };
 201
 202 int add_dirty_roots(struct btrfs_trans_handle *trans,
 203                     struct radix_tree_root *radix, struct list_head *list)
 204 {
 205         struct dirty_root *dirty;
 206         struct btrfs_root *gang[8];
 207         struct btrfs_root *root;
 208         int i;
 209         int ret;
 210         int err;
 211         while(1) {
 212                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 213                                                  ARRAY_SIZE(gang),
 214                                                  BTRFS_ROOT_TRANS_TAG);
 215                 if (ret == 0)
 216                         break;
 217                 for (i = 0; i < ret; i++) {
 218                         root = gang[i];
 219                         radix_tree_tag_clear(radix,
 220                                      (unsigned long)root->root_key.objectid,
 221                                      BTRFS_ROOT_TRANS_TAG);
 222                         if (root->commit_root == root->node) {
 223                                 WARN_ON(bh_blocknr(root->node) !=
 224                                         btrfs_root_blocknr(&root->root_item));
 225                                 brelse(root->commit_root);
 226                                 root->commit_root = NULL;
 227                                 continue;
 228                         }
 229                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 230                         BUG_ON(!dirty);
 231                         memcpy(&dirty->snap_key, &root->root_key,
 232                                sizeof(root->root_key));
 233                         dirty->commit_root = root->commit_root;
 234                         root->commit_root = NULL;
 235                         dirty->root = root;
 236                         root->root_key.offset = root->fs_info->generation;
 237                         btrfs_set_root_blocknr(&root->root_item,
 238                                                bh_blocknr(root->node));
 239                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 240                                                 &root->root_key,
 241                                                 &root->root_item);
 242                         BUG_ON(err);
 243                         list_add(&dirty->list, list);
 244                 }
 245         }
 246         return 0;
 247 }
 248
 249 int drop_dirty_roots(struct btrfs_root *tree_root, struct list_head *list)
 250 {
 251         struct dirty_root *dirty;
 252         struct btrfs_trans_handle *trans;
 253         int ret;
 254
 255         while(!list_empty(list)) {
 256                 dirty = list_entry(list->next, struct dirty_root, list);
 257                 list_del_init(&dirty->list);
 258                 trans = btrfs_start_transaction(tree_root, 1);
 259                 ret = btrfs_drop_snapshot(trans, dirty->root,
 260                                           dirty->commit_root);
 261                 BUG_ON(ret);
 262
 263                 ret = btrfs_del_root(trans, tree_root, &dirty->snap_key);
 264                 BUG_ON(ret);
 265                 ret = btrfs_end_transaction(trans, tree_root);
 266                 BUG_ON(ret);
 267                 kfree(dirty);
 268         }
 269         return 0;
 270 }
 271
 272 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 273                              struct btrfs_root *root)
 274 {
 275         int ret = 0;
 276         struct btrfs_transaction *cur_trans;
 277         struct btrfs_transaction *prev_trans = NULL;
 278         struct list_head dirty_fs_roots;
 279         DEFINE_WAIT(wait);
 280
 281         INIT_LIST_HEAD(&dirty_fs_roots);
 282
 283         mutex_lock(&root->fs_info->trans_mutex);
 284         if (trans->transaction->in_commit) {
 285                 cur_trans = trans->transaction;
 286                 trans->transaction->use_count++;
 287                 btrfs_end_transaction(trans, root);
 288                 ret = wait_for_commit(root, cur_trans);
 289                 BUG_ON(ret);
 290                 put_transaction(cur_trans);
 291                 mutex_unlock(&root->fs_info->trans_mutex);
 292                 return 0;
 293         }
 294         cur_trans = trans->transaction;
 295         trans->transaction->in_commit = 1;
 296         while (trans->transaction->num_writers > 1) {
 297                 WARN_ON(cur_trans != trans->transaction);
 298                 prepare_to_wait(&trans->transaction->writer_wait, &wait,
 299                                 TASK_UNINTERRUPTIBLE);
 300                 if (trans->transaction->num_writers <= 1)
 301                         break;
 302                 mutex_unlock(&root->fs_info->trans_mutex);
 303                 schedule();
 304                 mutex_lock(&root->fs_info->trans_mutex);
 305                 finish_wait(&trans->transaction->writer_wait, &wait);
 306         }
 307         finish_wait(&trans->transaction->writer_wait, &wait);
 308         WARN_ON(cur_trans != trans->transaction);
 309         add_dirty_roots(trans, &root->fs_info->fs_roots_radix, &dirty_fs_roots);
 310         ret = btrfs_commit_tree_roots(trans, root);
 311         BUG_ON(ret);
 312         cur_trans = root->fs_info->running_transaction;
 313         root->fs_info->running_transaction = NULL;
 314         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 315                 prev_trans = list_entry(cur_trans->list.prev,
 316                                         struct btrfs_transaction, list);
 317                 if (prev_trans->commit_done)
 318                         prev_trans = NULL;
 319                 else
 320                         prev_trans->use_count++;
 321         }
 322         mutex_unlock(&root->fs_info->trans_mutex);
 323         mutex_unlock(&root->fs_info->fs_mutex);
 324         ret = btrfs_write_and_wait_transaction(trans, root);
 325         if (prev_trans) {
 326                 mutex_lock(&root->fs_info->trans_mutex);
 327                 wait_for_commit(root, prev_trans);
 328                 put_transaction(prev_trans);
 329                 mutex_unlock(&root->fs_info->trans_mutex);
 330         }
 331         btrfs_set_super_generation(root->fs_info->disk_super,
 332                                    cur_trans->transid);
 333         BUG_ON(ret);
 334         write_ctree_super(trans, root);
 335
 336         mutex_lock(&root->fs_info->fs_mutex);
 337         btrfs_finish_extent_commit(trans, root);
 338         mutex_lock(&root->fs_info->trans_mutex);
 339         cur_trans->commit_done = 1;
 340         wake_up(&cur_trans->commit_wait);
 341         put_transaction(cur_trans);
 342         put_transaction(cur_trans);
 343         mutex_unlock(&root->fs_info->trans_mutex);
 344         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 345
 346         drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
 347         return ret;
 348 }
 349