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 <linux/writeback.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/blkdev.h>
  24 #include "ctree.h"
  25 #include "disk-io.h"
  26 #include "transaction.h"
  27 #include "locking.h"
  28 #include "tree-log.h"
  29
  30 #define BTRFS_ROOT_TRANS_TAG 0
  31
  32 static noinline void put_transaction(struct btrfs_transaction *transaction)
  33 {
  34         WARN_ON(transaction->use_count == 0);
  35         transaction->use_count--;
  36         if (transaction->use_count == 0) {
  37                 list_del_init(&transaction->list);
  38                 memset(transaction, 0, sizeof(*transaction));
  39                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  40         }
  41 }
  42
  43 static noinline void switch_commit_root(struct btrfs_root *root)
  44 {
  45         free_extent_buffer(root->commit_root);
  46         root->commit_root = btrfs_root_node(root);
  47 }
  48
  49 /*
  50  * either allocate a new transaction or hop into the existing one
  51  */
  52 static noinline int join_transaction(struct btrfs_root *root)
  53 {
  54         struct btrfs_transaction *cur_trans;
  55         cur_trans = root->fs_info->running_transaction;
  56         if (!cur_trans) {
  57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  58                                              GFP_NOFS);
  59                 BUG_ON(!cur_trans);
  60                 root->fs_info->generation++;
  61                 cur_trans->num_writers = 1;
  62                 cur_trans->num_joined = 0;
  63                 cur_trans->transid = root->fs_info->generation;
  64                 init_waitqueue_head(&cur_trans->writer_wait);
  65                 init_waitqueue_head(&cur_trans->commit_wait);
  66                 cur_trans->in_commit = 0;
  67                 cur_trans->blocked = 0;
  68                 cur_trans->use_count = 1;
  69                 cur_trans->commit_done = 0;
  70                 cur_trans->start_time = get_seconds();
  71
  72                 cur_trans->delayed_refs.root.rb_node = NULL;
  73                 cur_trans->delayed_refs.num_entries = 0;
  74                 cur_trans->delayed_refs.num_heads_ready = 0;
  75                 cur_trans->delayed_refs.num_heads = 0;
  76                 cur_trans->delayed_refs.flushing = 0;
  77                 cur_trans->delayed_refs.run_delayed_start = 0;
  78                 spin_lock_init(&cur_trans->delayed_refs.lock);
  79
  80                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  81                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  82                 extent_io_tree_init(&cur_trans->dirty_pages,
  83                                      root->fs_info->btree_inode->i_mapping,
  84                                      GFP_NOFS);
  85                 spin_lock(&root->fs_info->new_trans_lock);
  86                 root->fs_info->running_transaction = cur_trans;
  87                 spin_unlock(&root->fs_info->new_trans_lock);
  88         } else {
  89                 cur_trans->num_writers++;
  90                 cur_trans->num_joined++;
  91         }
  92
  93         return 0;
  94 }
  95
  96 /*
  97  * this does all the record keeping required to make sure that a reference
  98  * counted root is properly recorded in a given transaction.  This is required
  99  * to make sure the old root from before we joined the transaction is deleted
 100  * when the transaction commits
 101  */
 102 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 103                                          struct btrfs_root *root)
 104 {
 105         if (root->ref_cows && root->last_trans < trans->transid) {
 106                 WARN_ON(root == root->fs_info->extent_root);
 107                 WARN_ON(root->commit_root != root->node);
 108
 109                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 110                            (unsigned long)root->root_key.objectid,
 111                            BTRFS_ROOT_TRANS_TAG);
 112                 root->last_trans = trans->transid;
 113                 btrfs_init_reloc_root(trans, root);
 114         }
 115         return 0;
 116 }
 117
 118 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 119                                struct btrfs_root *root)
 120 {
 121         if (!root->ref_cows)
 122                 return 0;
 123
 124         mutex_lock(&root->fs_info->trans_mutex);
 125         if (root->last_trans == trans->transid) {
 126                 mutex_unlock(&root->fs_info->trans_mutex);
 127                 return 0;
 128         }
 129
 130         record_root_in_trans(trans, root);
 131         mutex_unlock(&root->fs_info->trans_mutex);
 132         return 0;
 133 }
 134
 135 /* wait for commit against the current transaction to become unblocked
 136  * when this is done, it is safe to start a new transaction, but the current
 137  * transaction might not be fully on disk.
 138  */
 139 static void wait_current_trans(struct btrfs_root *root)
 140 {
 141         struct btrfs_transaction *cur_trans;
 142
 143         cur_trans = root->fs_info->running_transaction;
 144         if (cur_trans && cur_trans->blocked) {
 145                 DEFINE_WAIT(wait);
 146                 cur_trans->use_count++;
 147                 while (1) {
 148                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 149                                         TASK_UNINTERRUPTIBLE);
 150                         if (cur_trans->blocked) {
 151                                 mutex_unlock(&root->fs_info->trans_mutex);
 152                                 schedule();
 153                                 mutex_lock(&root->fs_info->trans_mutex);
 154                                 finish_wait(&root->fs_info->transaction_wait,
 155                                             &wait);
 156                         } else {
 157                                 finish_wait(&root->fs_info->transaction_wait,
 158                                             &wait);
 159                                 break;
 160                         }
 161                 }
 162                 put_transaction(cur_trans);
 163         }
 164 }
 165
 166 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 167                                              int num_blocks, int wait)
 168 {
 169         struct btrfs_trans_handle *h =
 170                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 171         int ret;
 172
 173         mutex_lock(&root->fs_info->trans_mutex);
 174         if (!root->fs_info->log_root_recovering &&
 175             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
 176                 wait_current_trans(root);
 177         ret = join_transaction(root);
 178         BUG_ON(ret);
 179
 180         h->transid = root->fs_info->running_transaction->transid;
 181         h->transaction = root->fs_info->running_transaction;
 182         h->blocks_reserved = num_blocks;
 183         h->blocks_used = 0;
 184         h->block_group = 0;
 185         h->alloc_exclude_nr = 0;
 186         h->alloc_exclude_start = 0;
 187         h->delayed_ref_updates = 0;
 188
 189         if (!current->journal_info)
 190                 current->journal_info = h;
 191
 192         root->fs_info->running_transaction->use_count++;
 193         record_root_in_trans(h, root);
 194         mutex_unlock(&root->fs_info->trans_mutex);
 195         return h;
 196 }
 197
 198 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 199                                                    int num_blocks)
 200 {
 201         return start_transaction(root, num_blocks, 1);
 202 }
 203 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 204                                                    int num_blocks)
 205 {
 206         return start_transaction(root, num_blocks, 0);
 207 }
 208
 209 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 210                                                          int num_blocks)
 211 {
 212         return start_transaction(r, num_blocks, 2);
 213 }
 214
 215 /* wait for a transaction commit to be fully complete */
 216 static noinline int wait_for_commit(struct btrfs_root *root,
 217                                     struct btrfs_transaction *commit)
 218 {
 219         DEFINE_WAIT(wait);
 220         mutex_lock(&root->fs_info->trans_mutex);
 221         while (!commit->commit_done) {
 222                 prepare_to_wait(&commit->commit_wait, &wait,
 223                                 TASK_UNINTERRUPTIBLE);
 224                 if (commit->commit_done)
 225                         break;
 226                 mutex_unlock(&root->fs_info->trans_mutex);
 227                 schedule();
 228                 mutex_lock(&root->fs_info->trans_mutex);
 229         }
 230         mutex_unlock(&root->fs_info->trans_mutex);
 231         finish_wait(&commit->commit_wait, &wait);
 232         return 0;
 233 }
 234
 235 #if 0
 236 /*
 237  * rate limit against the drop_snapshot code.  This helps to slow down new
 238  * operations if the drop_snapshot code isn't able to keep up.
 239  */
 240 static void throttle_on_drops(struct btrfs_root *root)
 241 {
 242         struct btrfs_fs_info *info = root->fs_info;
 243         int harder_count = 0;
 244
 245 harder:
 246         if (atomic_read(&info->throttles)) {
 247                 DEFINE_WAIT(wait);
 248                 int thr;
 249                 thr = atomic_read(&info->throttle_gen);
 250
 251                 do {
 252                         prepare_to_wait(&info->transaction_throttle,
 253                                         &wait, TASK_UNINTERRUPTIBLE);
 254                         if (!atomic_read(&info->throttles)) {
 255                                 finish_wait(&info->transaction_throttle, &wait);
 256                                 break;
 257                         }
 258                         schedule();
 259                         finish_wait(&info->transaction_throttle, &wait);
 260                 } while (thr == atomic_read(&info->throttle_gen));
 261                 harder_count++;
 262
 263                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 264                     harder_count < 2)
 265                         goto harder;
 266
 267                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 268                     harder_count < 10)
 269                         goto harder;
 270
 271                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 272                     harder_count < 20)
 273                         goto harder;
 274         }
 275 }
 276 #endif
 277
 278 void btrfs_throttle(struct btrfs_root *root)
 279 {
 280         mutex_lock(&root->fs_info->trans_mutex);
 281         if (!root->fs_info->open_ioctl_trans)
 282                 wait_current_trans(root);
 283         mutex_unlock(&root->fs_info->trans_mutex);
 284 }
 285
 286 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 287                           struct btrfs_root *root, int throttle)
 288 {
 289         struct btrfs_transaction *cur_trans;
 290         struct btrfs_fs_info *info = root->fs_info;
 291         int count = 0;
 292
 293         while (count < 4) {
 294                 unsigned long cur = trans->delayed_ref_updates;
 295                 trans->delayed_ref_updates = 0;
 296                 if (cur &&
 297                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 298                         trans->delayed_ref_updates = 0;
 299
 300                         /*
 301                          * do a full flush if the transaction is trying
 302                          * to close
 303                          */
 304                         if (trans->transaction->delayed_refs.flushing)
 305                                 cur = 0;
 306                         btrfs_run_delayed_refs(trans, root, cur);
 307                 } else {
 308                         break;
 309                 }
 310                 count++;
 311         }
 312
 313         mutex_lock(&info->trans_mutex);
 314         cur_trans = info->running_transaction;
 315         WARN_ON(cur_trans != trans->transaction);
 316         WARN_ON(cur_trans->num_writers < 1);
 317         cur_trans->num_writers--;
 318
 319         if (waitqueue_active(&cur_trans->writer_wait))
 320                 wake_up(&cur_trans->writer_wait);
 321         put_transaction(cur_trans);
 322         mutex_unlock(&info->trans_mutex);
 323
 324         if (current->journal_info == trans)
 325                 current->journal_info = NULL;
 326         memset(trans, 0, sizeof(*trans));
 327         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 328
 329         return 0;
 330 }
 331
 332 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 333                           struct btrfs_root *root)
 334 {
 335         return __btrfs_end_transaction(trans, root, 0);
 336 }
 337
 338 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 339                                    struct btrfs_root *root)
 340 {
 341         return __btrfs_end_transaction(trans, root, 1);
 342 }
 343
 344 /*
 345  * when btree blocks are allocated, they have some corresponding bits set for
 346  * them in one of two extent_io trees.  This is used to make sure all of
 347  * those extents are sent to disk but does not wait on them
 348  */
 349 int btrfs_write_marked_extents(struct btrfs_root *root,
 350                                struct extent_io_tree *dirty_pages)
 351 {
 352         int ret;
 353         int err = 0;
 354         int werr = 0;
 355         struct page *page;
 356         struct inode *btree_inode = root->fs_info->btree_inode;
 357         u64 start = 0;
 358         u64 end;
 359         unsigned long index;
 360
 361         while (1) {
 362                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 363                                             EXTENT_DIRTY);
 364                 if (ret)
 365                         break;
 366                 while (start <= end) {
 367                         cond_resched();
 368
 369                         index = start >> PAGE_CACHE_SHIFT;
 370                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 371                         page = find_get_page(btree_inode->i_mapping, index);
 372                         if (!page)
 373                                 continue;
 374
 375                         btree_lock_page_hook(page);
 376                         if (!page->mapping) {
 377                                 unlock_page(page);
 378                                 page_cache_release(page);
 379                                 continue;
 380                         }
 381
 382                         if (PageWriteback(page)) {
 383                                 if (PageDirty(page))
 384                                         wait_on_page_writeback(page);
 385                                 else {
 386                                         unlock_page(page);
 387                                         page_cache_release(page);
 388                                         continue;
 389                                 }
 390                         }
 391                         err = write_one_page(page, 0);
 392                         if (err)
 393                                 werr = err;
 394                         page_cache_release(page);
 395                 }
 396         }
 397         if (err)
 398                 werr = err;
 399         return werr;
 400 }
 401
 402 /*
 403  * when btree blocks are allocated, they have some corresponding bits set for
 404  * them in one of two extent_io trees.  This is used to make sure all of
 405  * those extents are on disk for transaction or log commit.  We wait
 406  * on all the pages and clear them from the dirty pages state tree
 407  */
 408 int btrfs_wait_marked_extents(struct btrfs_root *root,
 409                               struct extent_io_tree *dirty_pages)
 410 {
 411         int ret;
 412         int err = 0;
 413         int werr = 0;
 414         struct page *page;
 415         struct inode *btree_inode = root->fs_info->btree_inode;
 416         u64 start = 0;
 417         u64 end;
 418         unsigned long index;
 419
 420         while (1) {
 421                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 422                                             EXTENT_DIRTY);
 423                 if (ret)
 424                         break;
 425
 426                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 427                 while (start <= end) {
 428                         index = start >> PAGE_CACHE_SHIFT;
 429                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 430                         page = find_get_page(btree_inode->i_mapping, index);
 431                         if (!page)
 432                                 continue;
 433                         if (PageDirty(page)) {
 434                                 btree_lock_page_hook(page);
 435                                 wait_on_page_writeback(page);
 436                                 err = write_one_page(page, 0);
 437                                 if (err)
 438                                         werr = err;
 439                         }
 440                         wait_on_page_writeback(page);
 441                         page_cache_release(page);
 442                         cond_resched();
 443                 }
 444         }
 445         if (err)
 446                 werr = err;
 447         return werr;
 448 }
 449
 450 /*
 451  * when btree blocks are allocated, they have some corresponding bits set for
 452  * them in one of two extent_io trees.  This is used to make sure all of
 453  * those extents are on disk for transaction or log commit
 454  */
 455 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 456                                         struct extent_io_tree *dirty_pages)
 457 {
 458         int ret;
 459         int ret2;
 460
 461         ret = btrfs_write_marked_extents(root, dirty_pages);
 462         ret2 = btrfs_wait_marked_extents(root, dirty_pages);
 463         return ret || ret2;
 464 }
 465
 466 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 467                                      struct btrfs_root *root)
 468 {
 469         if (!trans || !trans->transaction) {
 470                 struct inode *btree_inode;
 471                 btree_inode = root->fs_info->btree_inode;
 472                 return filemap_write_and_wait(btree_inode->i_mapping);
 473         }
 474         return btrfs_write_and_wait_marked_extents(root,
 475                                            &trans->transaction->dirty_pages);
 476 }
 477
 478 /*
 479  * this is used to update the root pointer in the tree of tree roots.
 480  *
 481  * But, in the case of the extent allocation tree, updating the root
 482  * pointer may allocate blocks which may change the root of the extent
 483  * allocation tree.
 484  *
 485  * So, this loops and repeats and makes sure the cowonly root didn't
 486  * change while the root pointer was being updated in the metadata.
 487  */
 488 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 489                                struct btrfs_root *root)
 490 {
 491         int ret;
 492         u64 old_root_bytenr;
 493         struct btrfs_root *tree_root = root->fs_info->tree_root;
 494
 495         btrfs_write_dirty_block_groups(trans, root);
 496
 497         while (1) {
 498                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 499                 if (old_root_bytenr == root->node->start)
 500                         break;
 501
 502                 btrfs_set_root_node(&root->root_item, root->node);
 503                 ret = btrfs_update_root(trans, tree_root,
 504                                         &root->root_key,
 505                                         &root->root_item);
 506                 BUG_ON(ret);
 507
 508                 ret = btrfs_write_dirty_block_groups(trans, root);
 509                 BUG_ON(ret);
 510         }
 511
 512         if (root != root->fs_info->extent_root)
 513                 switch_commit_root(root);
 514
 515         return 0;
 516 }
 517
 518 /*
 519  * update all the cowonly tree roots on disk
 520  */
 521 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 522                                          struct btrfs_root *root)
 523 {
 524         struct btrfs_fs_info *fs_info = root->fs_info;
 525         struct list_head *next;
 526         struct extent_buffer *eb;
 527         int ret;
 528
 529         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 530         BUG_ON(ret);
 531
 532         eb = btrfs_lock_root_node(fs_info->tree_root);
 533         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 534         btrfs_tree_unlock(eb);
 535         free_extent_buffer(eb);
 536
 537         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 538         BUG_ON(ret);
 539
 540         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 541                 next = fs_info->dirty_cowonly_roots.next;
 542                 list_del_init(next);
 543                 root = list_entry(next, struct btrfs_root, dirty_list);
 544
 545                 update_cowonly_root(trans, root);
 546         }
 547
 548         down_write(&fs_info->extent_commit_sem);
 549         switch_commit_root(fs_info->extent_root);
 550         up_write(&fs_info->extent_commit_sem);
 551
 552         return 0;
 553 }
 554
 555 /*
 556  * dead roots are old snapshots that need to be deleted.  This allocates
 557  * a dirty root struct and adds it into the list of dead roots that need to
 558  * be deleted
 559  */
 560 int btrfs_add_dead_root(struct btrfs_root *root)
 561 {
 562         mutex_lock(&root->fs_info->trans_mutex);
 563         list_add(&root->root_list, &root->fs_info->dead_roots);
 564         mutex_unlock(&root->fs_info->trans_mutex);
 565         return 0;
 566 }
 567
 568 /*
 569  * update all the cowonly tree roots on disk
 570  */
 571 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 572                                     struct btrfs_root *root)
 573 {
 574         struct btrfs_root *gang[8];
 575         struct btrfs_fs_info *fs_info = root->fs_info;
 576         int i;
 577         int ret;
 578         int err = 0;
 579
 580         while (1) {
 581                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 582                                                  (void **)gang, 0,
 583                                                  ARRAY_SIZE(gang),
 584                                                  BTRFS_ROOT_TRANS_TAG);
 585                 if (ret == 0)
 586                         break;
 587                 for (i = 0; i < ret; i++) {
 588                         root = gang[i];
 589                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 590                                         (unsigned long)root->root_key.objectid,
 591                                         BTRFS_ROOT_TRANS_TAG);
 592
 593                         btrfs_free_log(trans, root);
 594                         btrfs_update_reloc_root(trans, root);
 595
 596                         if (root->commit_root != root->node) {
 597                                 switch_commit_root(root);
 598                                 btrfs_set_root_node(&root->root_item,
 599                                                     root->node);
 600                         }
 601
 602                         err = btrfs_update_root(trans, fs_info->tree_root,
 603                                                 &root->root_key,
 604                                                 &root->root_item);
 605                         if (err)
 606                                 break;
 607                 }
 608         }
 609         return err;
 610 }
 611
 612 /*
 613  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 614  * otherwise every leaf in the btree is read and defragged.
 615  */
 616 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 617 {
 618         struct btrfs_fs_info *info = root->fs_info;
 619         int ret;
 620         struct btrfs_trans_handle *trans;
 621         unsigned long nr;
 622
 623         smp_mb();
 624         if (root->defrag_running)
 625                 return 0;
 626         trans = btrfs_start_transaction(root, 1);
 627         while (1) {
 628                 root->defrag_running = 1;
 629                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 630                 nr = trans->blocks_used;
 631                 btrfs_end_transaction(trans, root);
 632                 btrfs_btree_balance_dirty(info->tree_root, nr);
 633                 cond_resched();
 634
 635                 trans = btrfs_start_transaction(root, 1);
 636                 if (root->fs_info->closing || ret != -EAGAIN)
 637                         break;
 638         }
 639         root->defrag_running = 0;
 640         smp_mb();
 641         btrfs_end_transaction(trans, root);
 642         return 0;
 643 }
 644
 645 #if 0
 646 /*
 647  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 648  * sense not to join the transaction while it is trying to flush the current
 649  * queue of delayed refs out.
 650  *
 651  * This is used by the drop snapshot code only
 652  */
 653 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 654 {
 655         DEFINE_WAIT(wait);
 656
 657         mutex_lock(&info->trans_mutex);
 658         while (info->running_transaction &&
 659                info->running_transaction->delayed_refs.flushing) {
 660                 prepare_to_wait(&info->transaction_wait, &wait,
 661                                 TASK_UNINTERRUPTIBLE);
 662                 mutex_unlock(&info->trans_mutex);
 663
 664                 schedule();
 665
 666                 mutex_lock(&info->trans_mutex);
 667                 finish_wait(&info->transaction_wait, &wait);
 668         }
 669         mutex_unlock(&info->trans_mutex);
 670         return 0;
 671 }
 672
 673 /*
 674  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 675  * all of them
 676  */
 677 int btrfs_drop_dead_root(struct btrfs_root *root)
 678 {
 679         struct btrfs_trans_handle *trans;
 680         struct btrfs_root *tree_root = root->fs_info->tree_root;
 681         unsigned long nr;
 682         int ret;
 683
 684         while (1) {
 685                 /*
 686                  * we don't want to jump in and create a bunch of
 687                  * delayed refs if the transaction is starting to close
 688                  */
 689                 wait_transaction_pre_flush(tree_root->fs_info);
 690                 trans = btrfs_start_transaction(tree_root, 1);
 691
 692                 /*
 693                  * we've joined a transaction, make sure it isn't
 694                  * closing right now
 695                  */
 696                 if (trans->transaction->delayed_refs.flushing) {
 697                         btrfs_end_transaction(trans, tree_root);
 698                         continue;
 699                 }
 700
 701                 ret = btrfs_drop_snapshot(trans, root);
 702                 if (ret != -EAGAIN)
 703                         break;
 704
 705                 ret = btrfs_update_root(trans, tree_root,
 706                                         &root->root_key,
 707                                         &root->root_item);
 708                 if (ret)
 709                         break;
 710
 711                 nr = trans->blocks_used;
 712                 ret = btrfs_end_transaction(trans, tree_root);
 713                 BUG_ON(ret);
 714
 715                 btrfs_btree_balance_dirty(tree_root, nr);
 716                 cond_resched();
 717         }
 718         BUG_ON(ret);
 719
 720         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 721         BUG_ON(ret);
 722
 723         nr = trans->blocks_used;
 724         ret = btrfs_end_transaction(trans, tree_root);
 725         BUG_ON(ret);
 726
 727         free_extent_buffer(root->node);
 728         free_extent_buffer(root->commit_root);
 729         kfree(root);
 730
 731         btrfs_btree_balance_dirty(tree_root, nr);
 732         return ret;
 733 }
 734 #endif
 735
 736 /*
 737  * new snapshots need to be created at a very specific time in the
 738  * transaction commit.  This does the actual creation
 739  */
 740 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 741                                    struct btrfs_fs_info *fs_info,
 742                                    struct btrfs_pending_snapshot *pending)
 743 {
 744         struct btrfs_key key;
 745         struct btrfs_root_item *new_root_item;
 746         struct btrfs_root *tree_root = fs_info->tree_root;
 747         struct btrfs_root *root = pending->root;
 748         struct extent_buffer *tmp;
 749         struct extent_buffer *old;
 750         int ret;
 751         u64 objectid;
 752
 753         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 754         if (!new_root_item) {
 755                 ret = -ENOMEM;
 756                 goto fail;
 757         }
 758         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 759         if (ret)
 760                 goto fail;
 761
 762         record_root_in_trans(trans, root);
 763         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 764         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 765
 766         key.objectid = objectid;
 767         /* record when the snapshot was created in key.offset */
 768         key.offset = trans->transid;
 769         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 770
 771         old = btrfs_lock_root_node(root);
 772         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 773         btrfs_set_lock_blocking(old);
 774
 775         btrfs_copy_root(trans, root, old, &tmp, objectid);
 776         btrfs_tree_unlock(old);
 777         free_extent_buffer(old);
 778
 779         btrfs_set_root_node(new_root_item, tmp);
 780         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 781                                 new_root_item);
 782         btrfs_tree_unlock(tmp);
 783         free_extent_buffer(tmp);
 784         if (ret)
 785                 goto fail;
 786
 787         key.offset = (u64)-1;
 788         memcpy(&pending->root_key, &key, sizeof(key));
 789 fail:
 790         kfree(new_root_item);
 791         btrfs_unreserve_metadata_space(root, 6);
 792         return ret;
 793 }
 794
 795 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 796                                    struct btrfs_pending_snapshot *pending)
 797 {
 798         int ret;
 799         int namelen;
 800         u64 index = 0;
 801         struct btrfs_trans_handle *trans;
 802         struct inode *parent_inode;
 803         struct inode *inode;
 804         struct btrfs_root *parent_root;
 805
 806         parent_inode = pending->dentry->d_parent->d_inode;
 807         parent_root = BTRFS_I(parent_inode)->root;
 808         trans = btrfs_join_transaction(parent_root, 1);
 809
 810         /*
 811          * insert the directory item
 812          */
 813         namelen = strlen(pending->name);
 814         ret = btrfs_set_inode_index(parent_inode, &index);
 815         ret = btrfs_insert_dir_item(trans, parent_root,
 816                             pending->name, namelen,
 817                             parent_inode->i_ino,
 818                             &pending->root_key, BTRFS_FT_DIR, index);
 819
 820         if (ret)
 821                 goto fail;
 822
 823         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 824         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 825         BUG_ON(ret);
 826
 827         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 828                                  pending->root_key.objectid,
 829                                  parent_root->root_key.objectid,
 830                                  parent_inode->i_ino, index, pending->name,
 831                                  namelen);
 832
 833         BUG_ON(ret);
 834
 835         inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
 836         d_instantiate(pending->dentry, inode);
 837 fail:
 838         btrfs_end_transaction(trans, fs_info->fs_root);
 839         return ret;
 840 }
 841
 842 /*
 843  * create all the snapshots we've scheduled for creation
 844  */
 845 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 846                                              struct btrfs_fs_info *fs_info)
 847 {
 848         struct btrfs_pending_snapshot *pending;
 849         struct list_head *head = &trans->transaction->pending_snapshots;
 850         int ret;
 851
 852         list_for_each_entry(pending, head, list) {
 853                 ret = create_pending_snapshot(trans, fs_info, pending);
 854                 BUG_ON(ret);
 855         }
 856         return 0;
 857 }
 858
 859 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 860                                              struct btrfs_fs_info *fs_info)
 861 {
 862         struct btrfs_pending_snapshot *pending;
 863         struct list_head *head = &trans->transaction->pending_snapshots;
 864         int ret;
 865
 866         while (!list_empty(head)) {
 867                 pending = list_entry(head->next,
 868                                      struct btrfs_pending_snapshot, list);
 869                 ret = finish_pending_snapshot(fs_info, pending);
 870                 BUG_ON(ret);
 871                 list_del(&pending->list);
 872                 kfree(pending->name);
 873                 kfree(pending);
 874         }
 875         return 0;
 876 }
 877
 878 static void update_super_roots(struct btrfs_root *root)
 879 {
 880         struct btrfs_root_item *root_item;
 881         struct btrfs_super_block *super;
 882
 883         super = &root->fs_info->super_copy;
 884
 885         root_item = &root->fs_info->chunk_root->root_item;
 886         super->chunk_root = root_item->bytenr;
 887         super->chunk_root_generation = root_item->generation;
 888         super->chunk_root_level = root_item->level;
 889
 890         root_item = &root->fs_info->tree_root->root_item;
 891         super->root = root_item->bytenr;
 892         super->generation = root_item->generation;
 893         super->root_level = root_item->level;
 894 }
 895
 896 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 897 {
 898         int ret = 0;
 899         spin_lock(&info->new_trans_lock);
 900         if (info->running_transaction)
 901                 ret = info->running_transaction->in_commit;
 902         spin_unlock(&info->new_trans_lock);
 903         return ret;
 904 }
 905
 906 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 907                              struct btrfs_root *root)
 908 {
 909         unsigned long joined = 0;
 910         unsigned long timeout = 1;
 911         struct btrfs_transaction *cur_trans;
 912         struct btrfs_transaction *prev_trans = NULL;
 913         DEFINE_WAIT(wait);
 914         int ret;
 915         int should_grow = 0;
 916         unsigned long now = get_seconds();
 917         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 918
 919         btrfs_run_ordered_operations(root, 0);
 920
 921         /* make a pass through all the delayed refs we have so far
 922          * any runnings procs may add more while we are here
 923          */
 924         ret = btrfs_run_delayed_refs(trans, root, 0);
 925         BUG_ON(ret);
 926
 927         cur_trans = trans->transaction;
 928         /*
 929          * set the flushing flag so procs in this transaction have to
 930          * start sending their work down.
 931          */
 932         cur_trans->delayed_refs.flushing = 1;
 933
 934         ret = btrfs_run_delayed_refs(trans, root, 0);
 935         BUG_ON(ret);
 936
 937         mutex_lock(&root->fs_info->trans_mutex);
 938         if (cur_trans->in_commit) {
 939                 cur_trans->use_count++;
 940                 mutex_unlock(&root->fs_info->trans_mutex);
 941                 btrfs_end_transaction(trans, root);
 942
 943                 ret = wait_for_commit(root, cur_trans);
 944                 BUG_ON(ret);
 945
 946                 mutex_lock(&root->fs_info->trans_mutex);
 947                 put_transaction(cur_trans);
 948                 mutex_unlock(&root->fs_info->trans_mutex);
 949
 950                 return 0;
 951         }
 952
 953         trans->transaction->in_commit = 1;
 954         trans->transaction->blocked = 1;
 955         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 956                 prev_trans = list_entry(cur_trans->list.prev,
 957                                         struct btrfs_transaction, list);
 958                 if (!prev_trans->commit_done) {
 959                         prev_trans->use_count++;
 960                         mutex_unlock(&root->fs_info->trans_mutex);
 961
 962                         wait_for_commit(root, prev_trans);
 963
 964                         mutex_lock(&root->fs_info->trans_mutex);
 965                         put_transaction(prev_trans);
 966                 }
 967         }
 968
 969         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 970                 should_grow = 1;
 971
 972         do {
 973                 int snap_pending = 0;
 974                 joined = cur_trans->num_joined;
 975                 if (!list_empty(&trans->transaction->pending_snapshots))
 976                         snap_pending = 1;
 977
 978                 WARN_ON(cur_trans != trans->transaction);
 979                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 980                                 TASK_UNINTERRUPTIBLE);
 981
 982                 if (cur_trans->num_writers > 1)
 983                         timeout = MAX_SCHEDULE_TIMEOUT;
 984                 else if (should_grow)
 985                         timeout = 1;
 986
 987                 mutex_unlock(&root->fs_info->trans_mutex);
 988
 989                 if (flush_on_commit) {
 990                         btrfs_start_delalloc_inodes(root);
 991                         ret = btrfs_wait_ordered_extents(root, 0);
 992                         BUG_ON(ret);
 993                 } else if (snap_pending) {
 994                         ret = btrfs_wait_ordered_extents(root, 1);
 995                         BUG_ON(ret);
 996                 }
 997
 998                 /*
 999                  * rename don't use btrfs_join_transaction, so, once we
1000                  * set the transaction to blocked above, we aren't going
1001                  * to get any new ordered operations.  We can safely run
1002                  * it here and no for sure that nothing new will be added
1003                  * to the list
1004                  */
1005                 btrfs_run_ordered_operations(root, 1);
1006
1007                 smp_mb();
1008                 if (cur_trans->num_writers > 1 || should_grow)
1009                         schedule_timeout(timeout);
1010
1011                 mutex_lock(&root->fs_info->trans_mutex);
1012                 finish_wait(&cur_trans->writer_wait, &wait);
1013         } while (cur_trans->num_writers > 1 ||
1014                  (should_grow && cur_trans->num_joined != joined));
1015
1016         ret = create_pending_snapshots(trans, root->fs_info);
1017         BUG_ON(ret);
1018
1019         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1020         BUG_ON(ret);
1021
1022         WARN_ON(cur_trans != trans->transaction);
1023
1024         /* btrfs_commit_tree_roots is responsible for getting the
1025          * various roots consistent with each other.  Every pointer
1026          * in the tree of tree roots has to point to the most up to date
1027          * root for every subvolume and other tree.  So, we have to keep
1028          * the tree logging code from jumping in and changing any
1029          * of the trees.
1030          *
1031          * At this point in the commit, there can't be any tree-log
1032          * writers, but a little lower down we drop the trans mutex
1033          * and let new people in.  By holding the tree_log_mutex
1034          * from now until after the super is written, we avoid races
1035          * with the tree-log code.
1036          */
1037         mutex_lock(&root->fs_info->tree_log_mutex);
1038
1039         ret = commit_fs_roots(trans, root);
1040         BUG_ON(ret);
1041
1042         /* commit_fs_roots gets rid of all the tree log roots, it is now
1043          * safe to free the root of tree log roots
1044          */
1045         btrfs_free_log_root_tree(trans, root->fs_info);
1046
1047         ret = commit_cowonly_roots(trans, root);
1048         BUG_ON(ret);
1049
1050         btrfs_prepare_extent_commit(trans, root);
1051
1052         cur_trans = root->fs_info->running_transaction;
1053         spin_lock(&root->fs_info->new_trans_lock);
1054         root->fs_info->running_transaction = NULL;
1055         spin_unlock(&root->fs_info->new_trans_lock);
1056
1057         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1058                             root->fs_info->tree_root->node);
1059         switch_commit_root(root->fs_info->tree_root);
1060
1061         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1062                             root->fs_info->chunk_root->node);
1063         switch_commit_root(root->fs_info->chunk_root);
1064
1065         update_super_roots(root);
1066
1067         if (!root->fs_info->log_root_recovering) {
1068                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1069                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1070         }
1071
1072         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1073                sizeof(root->fs_info->super_copy));
1074
1075         trans->transaction->blocked = 0;
1076
1077         wake_up(&root->fs_info->transaction_wait);
1078
1079         mutex_unlock(&root->fs_info->trans_mutex);
1080         ret = btrfs_write_and_wait_transaction(trans, root);
1081         BUG_ON(ret);
1082         write_ctree_super(trans, root, 0);
1083
1084         /*
1085          * the super is written, we can safely allow the tree-loggers
1086          * to go about their business
1087          */
1088         mutex_unlock(&root->fs_info->tree_log_mutex);
1089
1090         btrfs_finish_extent_commit(trans, root);
1091
1092         /* do the directory inserts of any pending snapshot creations */
1093         finish_pending_snapshots(trans, root->fs_info);
1094
1095         mutex_lock(&root->fs_info->trans_mutex);
1096
1097         cur_trans->commit_done = 1;
1098
1099         root->fs_info->last_trans_committed = cur_trans->transid;
1100
1101         wake_up(&cur_trans->commit_wait);
1102
1103         put_transaction(cur_trans);
1104         put_transaction(cur_trans);
1105
1106         mutex_unlock(&root->fs_info->trans_mutex);
1107
1108         if (current->journal_info == trans)
1109                 current->journal_info = NULL;
1110
1111         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1112         return ret;
1113 }
1114
1115 /*
1116  * interface function to delete all the snapshots we have scheduled for deletion
1117  */
1118 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1119 {
1120         LIST_HEAD(list);
1121         struct btrfs_fs_info *fs_info = root->fs_info;
1122
1123         mutex_lock(&fs_info->trans_mutex);
1124         list_splice_init(&fs_info->dead_roots, &list);
1125         mutex_unlock(&fs_info->trans_mutex);
1126
1127         while (!list_empty(&list)) {
1128                 root = list_entry(list.next, struct btrfs_root, root_list);
1129                 list_del(&root->root_list);
1130
1131                 if (btrfs_header_backref_rev(root->node) <
1132                     BTRFS_MIXED_BACKREF_REV)
1133                         btrfs_drop_snapshot(root, 0);
1134                 else
1135                         btrfs_drop_snapshot(root, 1);
1136         }
1137         return 0;
1138 }