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/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30 #include "inode-map.h"
  31
  32 #define BTRFS_ROOT_TRANS_TAG 0
  33
  34 static noinline void put_transaction(struct btrfs_transaction *transaction)
  35 {
  36         WARN_ON(atomic_read(&transaction->use_count) == 0);
  37         if (atomic_dec_and_test(&transaction->use_count)) {
  38                 BUG_ON(!list_empty(&transaction->list));
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56
  57         spin_lock(&root->fs_info->trans_lock);
  58         if (root->fs_info->trans_no_join) {
  59                 if (!nofail) {
  60                         spin_unlock(&root->fs_info->trans_lock);
  61                         return -EBUSY;
  62                 }
  63         }
  64
  65         cur_trans = root->fs_info->running_transaction;
  66         if (cur_trans) {
  67                 atomic_inc(&cur_trans->use_count);
  68                 atomic_inc(&cur_trans->num_writers);
  69                 cur_trans->num_joined++;
  70                 spin_unlock(&root->fs_info->trans_lock);
  71                 return 0;
  72         }
  73         spin_unlock(&root->fs_info->trans_lock);
  74
  75         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  76         if (!cur_trans)
  77                 return -ENOMEM;
  78         spin_lock(&root->fs_info->trans_lock);
  79         if (root->fs_info->running_transaction) {
  80                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  81                 cur_trans = root->fs_info->running_transaction;
  82                 atomic_inc(&cur_trans->use_count);
  83                 atomic_inc(&cur_trans->num_writers);
  84                 cur_trans->num_joined++;
  85                 spin_unlock(&root->fs_info->trans_lock);
  86                 return 0;
  87         }
  88         atomic_set(&cur_trans->num_writers, 1);
  89         cur_trans->num_joined = 0;
  90         init_waitqueue_head(&cur_trans->writer_wait);
  91         init_waitqueue_head(&cur_trans->commit_wait);
  92         cur_trans->in_commit = 0;
  93         cur_trans->blocked = 0;
  94         /*
  95          * One for this trans handle, one so it will live on until we
  96          * commit the transaction.
  97          */
  98         atomic_set(&cur_trans->use_count, 2);
  99         cur_trans->commit_done = 0;
 100         cur_trans->start_time = get_seconds();
 101
 102         cur_trans->delayed_refs.root = RB_ROOT;
 103         cur_trans->delayed_refs.num_entries = 0;
 104         cur_trans->delayed_refs.num_heads_ready = 0;
 105         cur_trans->delayed_refs.num_heads = 0;
 106         cur_trans->delayed_refs.flushing = 0;
 107         cur_trans->delayed_refs.run_delayed_start = 0;
 108         spin_lock_init(&cur_trans->commit_lock);
 109         spin_lock_init(&cur_trans->delayed_refs.lock);
 110
 111         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 112         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 113         extent_io_tree_init(&cur_trans->dirty_pages,
 114                              root->fs_info->btree_inode->i_mapping);
 115         root->fs_info->generation++;
 116         cur_trans->transid = root->fs_info->generation;
 117         root->fs_info->running_transaction = cur_trans;
 118         spin_unlock(&root->fs_info->trans_lock);
 119
 120         return 0;
 121 }
 122
 123 /*
 124  * this does all the record keeping required to make sure that a reference
 125  * counted root is properly recorded in a given transaction.  This is required
 126  * to make sure the old root from before we joined the transaction is deleted
 127  * when the transaction commits
 128  */
 129 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 130                                struct btrfs_root *root)
 131 {
 132         if (root->ref_cows && root->last_trans < trans->transid) {
 133                 WARN_ON(root == root->fs_info->extent_root);
 134                 WARN_ON(root->commit_root != root->node);
 135
 136                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 137                 if (root->last_trans == trans->transid) {
 138                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 139                         return 0;
 140                 }
 141                 root->last_trans = trans->transid;
 142                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 143                            (unsigned long)root->root_key.objectid,
 144                            BTRFS_ROOT_TRANS_TAG);
 145                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 146                 btrfs_init_reloc_root(trans, root);
 147         }
 148         return 0;
 149 }
 150
 151 /* wait for commit against the current transaction to become unblocked
 152  * when this is done, it is safe to start a new transaction, but the current
 153  * transaction might not be fully on disk.
 154  */
 155 static void wait_current_trans(struct btrfs_root *root)
 156 {
 157         struct btrfs_transaction *cur_trans;
 158
 159         spin_lock(&root->fs_info->trans_lock);
 160         cur_trans = root->fs_info->running_transaction;
 161         if (cur_trans && cur_trans->blocked) {
 162                 DEFINE_WAIT(wait);
 163                 atomic_inc(&cur_trans->use_count);
 164                 spin_unlock(&root->fs_info->trans_lock);
 165                 while (1) {
 166                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 167                                         TASK_UNINTERRUPTIBLE);
 168                         if (!cur_trans->blocked)
 169                                 break;
 170                         schedule();
 171                 }
 172                 finish_wait(&root->fs_info->transaction_wait, &wait);
 173                 put_transaction(cur_trans);
 174         } else {
 175                 spin_unlock(&root->fs_info->trans_lock);
 176         }
 177 }
 178
 179 enum btrfs_trans_type {
 180         TRANS_START,
 181         TRANS_JOIN,
 182         TRANS_USERSPACE,
 183         TRANS_JOIN_NOLOCK,
 184 };
 185
 186 static int may_wait_transaction(struct btrfs_root *root, int type)
 187 {
 188         if (root->fs_info->log_root_recovering)
 189                 return 0;
 190
 191         if (type == TRANS_USERSPACE)
 192                 return 1;
 193
 194         if (type == TRANS_START &&
 195             !atomic_read(&root->fs_info->open_ioctl_trans))
 196                 return 1;
 197
 198         return 0;
 199 }
 200
 201 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 202                                                     u64 num_items, int type)
 203 {
 204         struct btrfs_trans_handle *h;
 205         struct btrfs_transaction *cur_trans;
 206         int retries = 0;
 207         int ret;
 208
 209         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 210                 return ERR_PTR(-EROFS);
 211
 212         if (current->journal_info) {
 213                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 214                 h = current->journal_info;
 215                 h->use_count++;
 216                 h->orig_rsv = h->block_rsv;
 217                 h->block_rsv = NULL;
 218                 goto got_it;
 219         }
 220 again:
 221         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 222         if (!h)
 223                 return ERR_PTR(-ENOMEM);
 224
 225         if (may_wait_transaction(root, type))
 226                 wait_current_trans(root);
 227
 228         do {
 229                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 230                 if (ret == -EBUSY)
 231                         wait_current_trans(root);
 232         } while (ret == -EBUSY);
 233
 234         if (ret < 0) {
 235                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 236                 return ERR_PTR(ret);
 237         }
 238
 239         cur_trans = root->fs_info->running_transaction;
 240
 241         h->transid = cur_trans->transid;
 242         h->transaction = cur_trans;
 243         h->blocks_used = 0;
 244         h->bytes_reserved = 0;
 245         h->delayed_ref_updates = 0;
 246         h->use_count = 1;
 247         h->block_rsv = NULL;
 248         h->orig_rsv = NULL;
 249
 250         smp_mb();
 251         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 252                 btrfs_commit_transaction(h, root);
 253                 goto again;
 254         }
 255
 256         if (num_items > 0) {
 257                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 258                 if (ret == -EAGAIN && !retries) {
 259                         retries++;
 260                         btrfs_commit_transaction(h, root);
 261                         goto again;
 262                 } else if (ret == -EAGAIN) {
 263                         /*
 264                          * We have already retried and got EAGAIN, so really we
 265                          * don't have space, so set ret to -ENOSPC.
 266                          */
 267                         ret = -ENOSPC;
 268                 }
 269
 270                 if (ret < 0) {
 271                         btrfs_end_transaction(h, root);
 272                         return ERR_PTR(ret);
 273                 }
 274         }
 275
 276 got_it:
 277         btrfs_record_root_in_trans(h, root);
 278
 279         if (!current->journal_info && type != TRANS_USERSPACE)
 280                 current->journal_info = h;
 281         return h;
 282 }
 283
 284 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 285                                                    int num_items)
 286 {
 287         return start_transaction(root, num_items, TRANS_START);
 288 }
 289 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 290 {
 291         return start_transaction(root, 0, TRANS_JOIN);
 292 }
 293
 294 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 295 {
 296         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 297 }
 298
 299 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 300 {
 301         return start_transaction(root, 0, TRANS_USERSPACE);
 302 }
 303
 304 /* wait for a transaction commit to be fully complete */
 305 static noinline int wait_for_commit(struct btrfs_root *root,
 306                                     struct btrfs_transaction *commit)
 307 {
 308         DEFINE_WAIT(wait);
 309         while (!commit->commit_done) {
 310                 prepare_to_wait(&commit->commit_wait, &wait,
 311                                 TASK_UNINTERRUPTIBLE);
 312                 if (commit->commit_done)
 313                         break;
 314                 schedule();
 315         }
 316         finish_wait(&commit->commit_wait, &wait);
 317         return 0;
 318 }
 319
 320 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 321 {
 322         struct btrfs_transaction *cur_trans = NULL, *t;
 323         int ret;
 324
 325         ret = 0;
 326         if (transid) {
 327                 if (transid <= root->fs_info->last_trans_committed)
 328                         goto out;
 329
 330                 /* find specified transaction */
 331                 spin_lock(&root->fs_info->trans_lock);
 332                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 333                         if (t->transid == transid) {
 334                                 cur_trans = t;
 335                                 atomic_inc(&cur_trans->use_count);
 336                                 break;
 337                         }
 338                         if (t->transid > transid)
 339                                 break;
 340                 }
 341                 spin_unlock(&root->fs_info->trans_lock);
 342                 ret = -EINVAL;
 343                 if (!cur_trans)
 344                         goto out;  /* bad transid */
 345         } else {
 346                 /* find newest transaction that is committing | committed */
 347                 spin_lock(&root->fs_info->trans_lock);
 348                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 349                                             list) {
 350                         if (t->in_commit) {
 351                                 if (t->commit_done)
 352                                         break;
 353                                 cur_trans = t;
 354                                 atomic_inc(&cur_trans->use_count);
 355                                 break;
 356                         }
 357                 }
 358                 spin_unlock(&root->fs_info->trans_lock);
 359                 if (!cur_trans)
 360                         goto out;  /* nothing committing|committed */
 361         }
 362
 363         wait_for_commit(root, cur_trans);
 364
 365         put_transaction(cur_trans);
 366         ret = 0;
 367 out:
 368         return ret;
 369 }
 370
 371 void btrfs_throttle(struct btrfs_root *root)
 372 {
 373         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 374                 wait_current_trans(root);
 375 }
 376
 377 static int should_end_transaction(struct btrfs_trans_handle *trans,
 378                                   struct btrfs_root *root)
 379 {
 380         int ret;
 381         ret = btrfs_block_rsv_check(trans, root,
 382                                     &root->fs_info->global_block_rsv, 0, 5);
 383         return ret ? 1 : 0;
 384 }
 385
 386 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 387                                  struct btrfs_root *root)
 388 {
 389         struct btrfs_transaction *cur_trans = trans->transaction;
 390         int updates;
 391
 392         smp_mb();
 393         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 394                 return 1;
 395
 396         updates = trans->delayed_ref_updates;
 397         trans->delayed_ref_updates = 0;
 398         if (updates)
 399                 btrfs_run_delayed_refs(trans, root, updates);
 400
 401         return should_end_transaction(trans, root);
 402 }
 403
 404 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 405                           struct btrfs_root *root, int throttle, int lock)
 406 {
 407         struct btrfs_transaction *cur_trans = trans->transaction;
 408         struct btrfs_fs_info *info = root->fs_info;
 409         int count = 0;
 410
 411         if (--trans->use_count) {
 412                 trans->block_rsv = trans->orig_rsv;
 413                 return 0;
 414         }
 415
 416         while (count < 4) {
 417                 unsigned long cur = trans->delayed_ref_updates;
 418                 trans->delayed_ref_updates = 0;
 419                 if (cur &&
 420                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 421                         trans->delayed_ref_updates = 0;
 422
 423                         /*
 424                          * do a full flush if the transaction is trying
 425                          * to close
 426                          */
 427                         if (trans->transaction->delayed_refs.flushing)
 428                                 cur = 0;
 429                         btrfs_run_delayed_refs(trans, root, cur);
 430                 } else {
 431                         break;
 432                 }
 433                 count++;
 434         }
 435
 436         btrfs_trans_release_metadata(trans, root);
 437
 438         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 439             should_end_transaction(trans, root)) {
 440                 trans->transaction->blocked = 1;
 441                 smp_wmb();
 442         }
 443
 444         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 445                 if (throttle)
 446                         return btrfs_commit_transaction(trans, root);
 447                 else
 448                         wake_up_process(info->transaction_kthread);
 449         }
 450
 451         WARN_ON(cur_trans != info->running_transaction);
 452         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 453         atomic_dec(&cur_trans->num_writers);
 454
 455         smp_mb();
 456         if (waitqueue_active(&cur_trans->writer_wait))
 457                 wake_up(&cur_trans->writer_wait);
 458         put_transaction(cur_trans);
 459
 460         if (current->journal_info == trans)
 461                 current->journal_info = NULL;
 462         memset(trans, 0, sizeof(*trans));
 463         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 464
 465         if (throttle)
 466                 btrfs_run_delayed_iputs(root);
 467
 468         return 0;
 469 }
 470
 471 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 472                           struct btrfs_root *root)
 473 {
 474         int ret;
 475
 476         ret = __btrfs_end_transaction(trans, root, 0, 1);
 477         if (ret)
 478                 return ret;
 479         return 0;
 480 }
 481
 482 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 483                                    struct btrfs_root *root)
 484 {
 485         int ret;
 486
 487         ret = __btrfs_end_transaction(trans, root, 1, 1);
 488         if (ret)
 489                 return ret;
 490         return 0;
 491 }
 492
 493 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 494                                  struct btrfs_root *root)
 495 {
 496         int ret;
 497
 498         ret = __btrfs_end_transaction(trans, root, 0, 0);
 499         if (ret)
 500                 return ret;
 501         return 0;
 502 }
 503
 504 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 505                                 struct btrfs_root *root)
 506 {
 507         return __btrfs_end_transaction(trans, root, 1, 1);
 508 }
 509
 510 /*
 511  * when btree blocks are allocated, they have some corresponding bits set for
 512  * them in one of two extent_io trees.  This is used to make sure all of
 513  * those extents are sent to disk but does not wait on them
 514  */
 515 int btrfs_write_marked_extents(struct btrfs_root *root,
 516                                struct extent_io_tree *dirty_pages, int mark)
 517 {
 518         int ret;
 519         int err = 0;
 520         int werr = 0;
 521         struct page *page;
 522         struct inode *btree_inode = root->fs_info->btree_inode;
 523         u64 start = 0;
 524         u64 end;
 525         unsigned long index;
 526
 527         while (1) {
 528                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 529                                             mark);
 530                 if (ret)
 531                         break;
 532                 while (start <= end) {
 533                         cond_resched();
 534
 535                         index = start >> PAGE_CACHE_SHIFT;
 536                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 537                         page = find_get_page(btree_inode->i_mapping, index);
 538                         if (!page)
 539                                 continue;
 540
 541                         btree_lock_page_hook(page);
 542                         if (!page->mapping) {
 543                                 unlock_page(page);
 544                                 page_cache_release(page);
 545                                 continue;
 546                         }
 547
 548                         if (PageWriteback(page)) {
 549                                 if (PageDirty(page))
 550                                         wait_on_page_writeback(page);
 551                                 else {
 552                                         unlock_page(page);
 553                                         page_cache_release(page);
 554                                         continue;
 555                                 }
 556                         }
 557                         err = write_one_page(page, 0);
 558                         if (err)
 559                                 werr = err;
 560                         page_cache_release(page);
 561                 }
 562         }
 563         if (err)
 564                 werr = err;
 565         return werr;
 566 }
 567
 568 /*
 569  * when btree blocks are allocated, they have some corresponding bits set for
 570  * them in one of two extent_io trees.  This is used to make sure all of
 571  * those extents are on disk for transaction or log commit.  We wait
 572  * on all the pages and clear them from the dirty pages state tree
 573  */
 574 int btrfs_wait_marked_extents(struct btrfs_root *root,
 575                               struct extent_io_tree *dirty_pages, int mark)
 576 {
 577         int ret;
 578         int err = 0;
 579         int werr = 0;
 580         struct page *page;
 581         struct inode *btree_inode = root->fs_info->btree_inode;
 582         u64 start = 0;
 583         u64 end;
 584         unsigned long index;
 585
 586         while (1) {
 587                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 588                                             mark);
 589                 if (ret)
 590                         break;
 591
 592                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 593                 while (start <= end) {
 594                         index = start >> PAGE_CACHE_SHIFT;
 595                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 596                         page = find_get_page(btree_inode->i_mapping, index);
 597                         if (!page)
 598                                 continue;
 599                         if (PageDirty(page)) {
 600                                 btree_lock_page_hook(page);
 601                                 wait_on_page_writeback(page);
 602                                 err = write_one_page(page, 0);
 603                                 if (err)
 604                                         werr = err;
 605                         }
 606                         wait_on_page_writeback(page);
 607                         page_cache_release(page);
 608                         cond_resched();
 609                 }
 610         }
 611         if (err)
 612                 werr = err;
 613         return werr;
 614 }
 615
 616 /*
 617  * when btree blocks are allocated, they have some corresponding bits set for
 618  * them in one of two extent_io trees.  This is used to make sure all of
 619  * those extents are on disk for transaction or log commit
 620  */
 621 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 622                                 struct extent_io_tree *dirty_pages, int mark)
 623 {
 624         int ret;
 625         int ret2;
 626
 627         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 628         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 629         return ret || ret2;
 630 }
 631
 632 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 633                                      struct btrfs_root *root)
 634 {
 635         if (!trans || !trans->transaction) {
 636                 struct inode *btree_inode;
 637                 btree_inode = root->fs_info->btree_inode;
 638                 return filemap_write_and_wait(btree_inode->i_mapping);
 639         }
 640         return btrfs_write_and_wait_marked_extents(root,
 641                                            &trans->transaction->dirty_pages,
 642                                            EXTENT_DIRTY);
 643 }
 644
 645 /*
 646  * this is used to update the root pointer in the tree of tree roots.
 647  *
 648  * But, in the case of the extent allocation tree, updating the root
 649  * pointer may allocate blocks which may change the root of the extent
 650  * allocation tree.
 651  *
 652  * So, this loops and repeats and makes sure the cowonly root didn't
 653  * change while the root pointer was being updated in the metadata.
 654  */
 655 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 656                                struct btrfs_root *root)
 657 {
 658         int ret;
 659         u64 old_root_bytenr;
 660         u64 old_root_used;
 661         struct btrfs_root *tree_root = root->fs_info->tree_root;
 662
 663         old_root_used = btrfs_root_used(&root->root_item);
 664         btrfs_write_dirty_block_groups(trans, root);
 665
 666         while (1) {
 667                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 668                 if (old_root_bytenr == root->node->start &&
 669                     old_root_used == btrfs_root_used(&root->root_item))
 670                         break;
 671
 672                 btrfs_set_root_node(&root->root_item, root->node);
 673                 ret = btrfs_update_root(trans, tree_root,
 674                                         &root->root_key,
 675                                         &root->root_item);
 676                 BUG_ON(ret);
 677
 678                 old_root_used = btrfs_root_used(&root->root_item);
 679                 ret = btrfs_write_dirty_block_groups(trans, root);
 680                 BUG_ON(ret);
 681         }
 682
 683         if (root != root->fs_info->extent_root)
 684                 switch_commit_root(root);
 685
 686         return 0;
 687 }
 688
 689 /*
 690  * update all the cowonly tree roots on disk
 691  */
 692 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 693                                          struct btrfs_root *root)
 694 {
 695         struct btrfs_fs_info *fs_info = root->fs_info;
 696         struct list_head *next;
 697         struct extent_buffer *eb;
 698         int ret;
 699
 700         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 701         BUG_ON(ret);
 702
 703         eb = btrfs_lock_root_node(fs_info->tree_root);
 704         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 705         btrfs_tree_unlock(eb);
 706         free_extent_buffer(eb);
 707
 708         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 709         BUG_ON(ret);
 710
 711         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 712                 next = fs_info->dirty_cowonly_roots.next;
 713                 list_del_init(next);
 714                 root = list_entry(next, struct btrfs_root, dirty_list);
 715
 716                 update_cowonly_root(trans, root);
 717         }
 718
 719         down_write(&fs_info->extent_commit_sem);
 720         switch_commit_root(fs_info->extent_root);
 721         up_write(&fs_info->extent_commit_sem);
 722
 723         return 0;
 724 }
 725
 726 /*
 727  * dead roots are old snapshots that need to be deleted.  This allocates
 728  * a dirty root struct and adds it into the list of dead roots that need to
 729  * be deleted
 730  */
 731 int btrfs_add_dead_root(struct btrfs_root *root)
 732 {
 733         spin_lock(&root->fs_info->trans_lock);
 734         list_add(&root->root_list, &root->fs_info->dead_roots);
 735         spin_unlock(&root->fs_info->trans_lock);
 736         return 0;
 737 }
 738
 739 /*
 740  * update all the cowonly tree roots on disk
 741  */
 742 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 743                                     struct btrfs_root *root)
 744 {
 745         struct btrfs_root *gang[8];
 746         struct btrfs_fs_info *fs_info = root->fs_info;
 747         int i;
 748         int ret;
 749         int err = 0;
 750
 751         spin_lock(&fs_info->fs_roots_radix_lock);
 752         while (1) {
 753                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 754                                                  (void **)gang, 0,
 755                                                  ARRAY_SIZE(gang),
 756                                                  BTRFS_ROOT_TRANS_TAG);
 757                 if (ret == 0)
 758                         break;
 759                 for (i = 0; i < ret; i++) {
 760                         root = gang[i];
 761                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 762                                         (unsigned long)root->root_key.objectid,
 763                                         BTRFS_ROOT_TRANS_TAG);
 764                         spin_unlock(&fs_info->fs_roots_radix_lock);
 765
 766                         btrfs_free_log(trans, root);
 767                         btrfs_update_reloc_root(trans, root);
 768                         btrfs_orphan_commit_root(trans, root);
 769
 770                         btrfs_save_ino_cache(root, trans);
 771
 772                         if (root->commit_root != root->node) {
 773                                 mutex_lock(&root->fs_commit_mutex);
 774                                 switch_commit_root(root);
 775                                 btrfs_unpin_free_ino(root);
 776                                 mutex_unlock(&root->fs_commit_mutex);
 777
 778                                 btrfs_set_root_node(&root->root_item,
 779                                                     root->node);
 780                         }
 781
 782                         err = btrfs_update_root(trans, fs_info->tree_root,
 783                                                 &root->root_key,
 784                                                 &root->root_item);
 785                         spin_lock(&fs_info->fs_roots_radix_lock);
 786                         if (err)
 787                                 break;
 788                 }
 789         }
 790         spin_unlock(&fs_info->fs_roots_radix_lock);
 791         return err;
 792 }
 793
 794 /*
 795  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 796  * otherwise every leaf in the btree is read and defragged.
 797  */
 798 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 799 {
 800         struct btrfs_fs_info *info = root->fs_info;
 801         struct btrfs_trans_handle *trans;
 802         int ret;
 803         unsigned long nr;
 804
 805         if (xchg(&root->defrag_running, 1))
 806                 return 0;
 807
 808         while (1) {
 809                 trans = btrfs_start_transaction(root, 0);
 810                 if (IS_ERR(trans))
 811                         return PTR_ERR(trans);
 812
 813                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 814
 815                 nr = trans->blocks_used;
 816                 btrfs_end_transaction(trans, root);
 817                 btrfs_btree_balance_dirty(info->tree_root, nr);
 818                 cond_resched();
 819
 820                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 821                         break;
 822         }
 823         root->defrag_running = 0;
 824         return ret;
 825 }
 826
 827 /*
 828  * new snapshots need to be created at a very specific time in the
 829  * transaction commit.  This does the actual creation
 830  */
 831 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 832                                    struct btrfs_fs_info *fs_info,
 833                                    struct btrfs_pending_snapshot *pending)
 834 {
 835         struct btrfs_key key;
 836         struct btrfs_root_item *new_root_item;
 837         struct btrfs_root *tree_root = fs_info->tree_root;
 838         struct btrfs_root *root = pending->root;
 839         struct btrfs_root *parent_root;
 840         struct inode *parent_inode;
 841         struct dentry *parent;
 842         struct dentry *dentry;
 843         struct extent_buffer *tmp;
 844         struct extent_buffer *old;
 845         int ret;
 846         u64 to_reserve = 0;
 847         u64 index = 0;
 848         u64 objectid;
 849         u64 root_flags;
 850
 851         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 852         if (!new_root_item) {
 853                 pending->error = -ENOMEM;
 854                 goto fail;
 855         }
 856
 857         ret = btrfs_find_free_objectid(tree_root, &objectid);
 858         if (ret) {
 859                 pending->error = ret;
 860                 goto fail;
 861         }
 862
 863         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 864         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 865
 866         if (to_reserve > 0) {
 867                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 868                                           to_reserve);
 869                 if (ret) {
 870                         pending->error = ret;
 871                         goto fail;
 872                 }
 873         }
 874
 875         key.objectid = objectid;
 876         key.offset = (u64)-1;
 877         key.type = BTRFS_ROOT_ITEM_KEY;
 878
 879         trans->block_rsv = &pending->block_rsv;
 880
 881         dentry = pending->dentry;
 882         parent = dget_parent(dentry);
 883         parent_inode = parent->d_inode;
 884         parent_root = BTRFS_I(parent_inode)->root;
 885         btrfs_record_root_in_trans(trans, parent_root);
 886
 887         /*
 888          * insert the directory item
 889          */
 890         ret = btrfs_set_inode_index(parent_inode, &index);
 891         BUG_ON(ret);
 892         ret = btrfs_insert_dir_item(trans, parent_root,
 893                                 dentry->d_name.name, dentry->d_name.len,
 894                                 parent_inode, &key,
 895                                 BTRFS_FT_DIR, index);
 896         BUG_ON(ret);
 897
 898         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 899                                          dentry->d_name.len * 2);
 900         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 901         BUG_ON(ret);
 902
 903         btrfs_record_root_in_trans(trans, root);
 904         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 905         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 906         btrfs_check_and_init_root_item(new_root_item);
 907
 908         root_flags = btrfs_root_flags(new_root_item);
 909         if (pending->readonly)
 910                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 911         else
 912                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 913         btrfs_set_root_flags(new_root_item, root_flags);
 914
 915         old = btrfs_lock_root_node(root);
 916         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 917         btrfs_set_lock_blocking(old);
 918
 919         btrfs_copy_root(trans, root, old, &tmp, objectid);
 920         btrfs_tree_unlock(old);
 921         free_extent_buffer(old);
 922
 923         btrfs_set_root_node(new_root_item, tmp);
 924         /* record when the snapshot was created in key.offset */
 925         key.offset = trans->transid;
 926         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 927         btrfs_tree_unlock(tmp);
 928         free_extent_buffer(tmp);
 929         BUG_ON(ret);
 930
 931         /*
 932          * insert root back/forward references
 933          */
 934         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 935                                  parent_root->root_key.objectid,
 936                                  btrfs_ino(parent_inode), index,
 937                                  dentry->d_name.name, dentry->d_name.len);
 938         BUG_ON(ret);
 939         dput(parent);
 940
 941         key.offset = (u64)-1;
 942         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 943         BUG_ON(IS_ERR(pending->snap));
 944
 945         btrfs_reloc_post_snapshot(trans, pending);
 946         btrfs_orphan_post_snapshot(trans, pending);
 947 fail:
 948         kfree(new_root_item);
 949         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 950         return 0;
 951 }
 952
 953 /*
 954  * create all the snapshots we've scheduled for creation
 955  */
 956 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 957                                              struct btrfs_fs_info *fs_info)
 958 {
 959         struct btrfs_pending_snapshot *pending;
 960         struct list_head *head = &trans->transaction->pending_snapshots;
 961         int ret;
 962
 963         list_for_each_entry(pending, head, list) {
 964                 /*
 965                  * We must deal with the delayed items before creating
 966                  * snapshots, or we will create a snapthot with inconsistent
 967                  * information.
 968                 */
 969                 ret = btrfs_run_delayed_items(trans, fs_info->fs_root);
 970                 BUG_ON(ret);
 971
 972                 ret = create_pending_snapshot(trans, fs_info, pending);
 973                 BUG_ON(ret);
 974         }
 975         return 0;
 976 }
 977
 978 static void update_super_roots(struct btrfs_root *root)
 979 {
 980         struct btrfs_root_item *root_item;
 981         struct btrfs_super_block *super;
 982
 983         super = &root->fs_info->super_copy;
 984
 985         root_item = &root->fs_info->chunk_root->root_item;
 986         super->chunk_root = root_item->bytenr;
 987         super->chunk_root_generation = root_item->generation;
 988         super->chunk_root_level = root_item->level;
 989
 990         root_item = &root->fs_info->tree_root->root_item;
 991         super->root = root_item->bytenr;
 992         super->generation = root_item->generation;
 993         super->root_level = root_item->level;
 994         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
 995                 super->cache_generation = root_item->generation;
 996 }
 997
 998 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 999 {
1000         int ret = 0;
1001         spin_lock(&info->trans_lock);
1002         if (info->running_transaction)
1003                 ret = info->running_transaction->in_commit;
1004         spin_unlock(&info->trans_lock);
1005         return ret;
1006 }
1007
1008 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1009 {
1010         int ret = 0;
1011         spin_lock(&info->trans_lock);
1012         if (info->running_transaction)
1013                 ret = info->running_transaction->blocked;
1014         spin_unlock(&info->trans_lock);
1015         return ret;
1016 }
1017
1018 /*
1019  * wait for the current transaction commit to start and block subsequent
1020  * transaction joins
1021  */
1022 static void wait_current_trans_commit_start(struct btrfs_root *root,
1023                                             struct btrfs_transaction *trans)
1024 {
1025         DEFINE_WAIT(wait);
1026
1027         if (trans->in_commit)
1028                 return;
1029
1030         while (1) {
1031                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1032                                 TASK_UNINTERRUPTIBLE);
1033                 if (trans->in_commit) {
1034                         finish_wait(&root->fs_info->transaction_blocked_wait,
1035                                     &wait);
1036                         break;
1037                 }
1038                 schedule();
1039                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1040         }
1041 }
1042
1043 /*
1044  * wait for the current transaction to start and then become unblocked.
1045  * caller holds ref.
1046  */
1047 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1048                                          struct btrfs_transaction *trans)
1049 {
1050         DEFINE_WAIT(wait);
1051
1052         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1053                 return;
1054
1055         while (1) {
1056                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1057                                 TASK_UNINTERRUPTIBLE);
1058                 if (trans->commit_done ||
1059                     (trans->in_commit && !trans->blocked)) {
1060                         finish_wait(&root->fs_info->transaction_wait,
1061                                     &wait);
1062                         break;
1063                 }
1064                 schedule();
1065                 finish_wait(&root->fs_info->transaction_wait,
1066                             &wait);
1067         }
1068 }
1069
1070 /*
1071  * commit transactions asynchronously. once btrfs_commit_transaction_async
1072  * returns, any subsequent transaction will not be allowed to join.
1073  */
1074 struct btrfs_async_commit {
1075         struct btrfs_trans_handle *newtrans;
1076         struct btrfs_root *root;
1077         struct delayed_work work;
1078 };
1079
1080 static void do_async_commit(struct work_struct *work)
1081 {
1082         struct btrfs_async_commit *ac =
1083                 container_of(work, struct btrfs_async_commit, work.work);
1084
1085         btrfs_commit_transaction(ac->newtrans, ac->root);
1086         kfree(ac);
1087 }
1088
1089 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1090                                    struct btrfs_root *root,
1091                                    int wait_for_unblock)
1092 {
1093         struct btrfs_async_commit *ac;
1094         struct btrfs_transaction *cur_trans;
1095
1096         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1097         if (!ac)
1098                 return -ENOMEM;
1099
1100         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1101         ac->root = root;
1102         ac->newtrans = btrfs_join_transaction(root);
1103         if (IS_ERR(ac->newtrans)) {
1104                 int err = PTR_ERR(ac->newtrans);
1105                 kfree(ac);
1106                 return err;
1107         }
1108
1109         /* take transaction reference */
1110         cur_trans = trans->transaction;
1111         atomic_inc(&cur_trans->use_count);
1112
1113         btrfs_end_transaction(trans, root);
1114         schedule_delayed_work(&ac->work, 0);
1115
1116         /* wait for transaction to start and unblock */
1117         if (wait_for_unblock)
1118                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1119         else
1120                 wait_current_trans_commit_start(root, cur_trans);
1121
1122         if (current->journal_info == trans)
1123                 current->journal_info = NULL;
1124
1125         put_transaction(cur_trans);
1126         return 0;
1127 }
1128
1129 /*
1130  * btrfs_transaction state sequence:
1131  *    in_commit = 0, blocked = 0  (initial)
1132  *    in_commit = 1, blocked = 1
1133  *    blocked = 0
1134  *    commit_done = 1
1135  */
1136 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1137                              struct btrfs_root *root)
1138 {
1139         unsigned long joined = 0;
1140         struct btrfs_transaction *cur_trans;
1141         struct btrfs_transaction *prev_trans = NULL;
1142         DEFINE_WAIT(wait);
1143         int ret;
1144         int should_grow = 0;
1145         unsigned long now = get_seconds();
1146         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1147
1148         btrfs_run_ordered_operations(root, 0);
1149
1150         /* make a pass through all the delayed refs we have so far
1151          * any runnings procs may add more while we are here
1152          */
1153         ret = btrfs_run_delayed_refs(trans, root, 0);
1154         BUG_ON(ret);
1155
1156         btrfs_trans_release_metadata(trans, root);
1157
1158         cur_trans = trans->transaction;
1159         /*
1160          * set the flushing flag so procs in this transaction have to
1161          * start sending their work down.
1162          */
1163         cur_trans->delayed_refs.flushing = 1;
1164
1165         ret = btrfs_run_delayed_refs(trans, root, 0);
1166         BUG_ON(ret);
1167
1168         spin_lock(&cur_trans->commit_lock);
1169         if (cur_trans->in_commit) {
1170                 spin_unlock(&cur_trans->commit_lock);
1171                 atomic_inc(&cur_trans->use_count);
1172                 btrfs_end_transaction(trans, root);
1173
1174                 ret = wait_for_commit(root, cur_trans);
1175                 BUG_ON(ret);
1176
1177                 put_transaction(cur_trans);
1178
1179                 return 0;
1180         }
1181
1182         trans->transaction->in_commit = 1;
1183         trans->transaction->blocked = 1;
1184         spin_unlock(&cur_trans->commit_lock);
1185         wake_up(&root->fs_info->transaction_blocked_wait);
1186
1187         spin_lock(&root->fs_info->trans_lock);
1188         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1189                 prev_trans = list_entry(cur_trans->list.prev,
1190                                         struct btrfs_transaction, list);
1191                 if (!prev_trans->commit_done) {
1192                         atomic_inc(&prev_trans->use_count);
1193                         spin_unlock(&root->fs_info->trans_lock);
1194
1195                         wait_for_commit(root, prev_trans);
1196
1197                         put_transaction(prev_trans);
1198                 } else {
1199                         spin_unlock(&root->fs_info->trans_lock);
1200                 }
1201         } else {
1202                 spin_unlock(&root->fs_info->trans_lock);
1203         }
1204
1205         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1206                 should_grow = 1;
1207
1208         do {
1209                 int snap_pending = 0;
1210
1211                 joined = cur_trans->num_joined;
1212                 if (!list_empty(&trans->transaction->pending_snapshots))
1213                         snap_pending = 1;
1214
1215                 WARN_ON(cur_trans != trans->transaction);
1216
1217                 if (flush_on_commit || snap_pending) {
1218                         btrfs_start_delalloc_inodes(root, 1);
1219                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1220                         BUG_ON(ret);
1221                 }
1222
1223                 ret = btrfs_run_delayed_items(trans, root);
1224                 BUG_ON(ret);
1225
1226                 /*
1227                  * rename don't use btrfs_join_transaction, so, once we
1228                  * set the transaction to blocked above, we aren't going
1229                  * to get any new ordered operations.  We can safely run
1230                  * it here and no for sure that nothing new will be added
1231                  * to the list
1232                  */
1233                 btrfs_run_ordered_operations(root, 1);
1234
1235                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1236                                 TASK_UNINTERRUPTIBLE);
1237
1238                 if (atomic_read(&cur_trans->num_writers) > 1)
1239                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1240                 else if (should_grow)
1241                         schedule_timeout(1);
1242
1243                 finish_wait(&cur_trans->writer_wait, &wait);
1244                 spin_lock(&root->fs_info->trans_lock);
1245                 root->fs_info->trans_no_join = 1;
1246                 spin_unlock(&root->fs_info->trans_lock);
1247         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1248                  (should_grow && cur_trans->num_joined != joined));
1249
1250         ret = create_pending_snapshots(trans, root->fs_info);
1251         BUG_ON(ret);
1252
1253         ret = btrfs_run_delayed_items(trans, root);
1254         BUG_ON(ret);
1255
1256         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1257         BUG_ON(ret);
1258
1259         WARN_ON(cur_trans != trans->transaction);
1260
1261         btrfs_scrub_pause(root);
1262         /* btrfs_commit_tree_roots is responsible for getting the
1263          * various roots consistent with each other.  Every pointer
1264          * in the tree of tree roots has to point to the most up to date
1265          * root for every subvolume and other tree.  So, we have to keep
1266          * the tree logging code from jumping in and changing any
1267          * of the trees.
1268          *
1269          * At this point in the commit, there can't be any tree-log
1270          * writers, but a little lower down we drop the trans mutex
1271          * and let new people in.  By holding the tree_log_mutex
1272          * from now until after the super is written, we avoid races
1273          * with the tree-log code.
1274          */
1275         mutex_lock(&root->fs_info->tree_log_mutex);
1276
1277         ret = commit_fs_roots(trans, root);
1278         BUG_ON(ret);
1279
1280         /* commit_fs_roots gets rid of all the tree log roots, it is now
1281          * safe to free the root of tree log roots
1282          */
1283         btrfs_free_log_root_tree(trans, root->fs_info);
1284
1285         ret = commit_cowonly_roots(trans, root);
1286         BUG_ON(ret);
1287
1288         btrfs_prepare_extent_commit(trans, root);
1289
1290         cur_trans = root->fs_info->running_transaction;
1291
1292         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1293                             root->fs_info->tree_root->node);
1294         switch_commit_root(root->fs_info->tree_root);
1295
1296         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1297                             root->fs_info->chunk_root->node);
1298         switch_commit_root(root->fs_info->chunk_root);
1299
1300         update_super_roots(root);
1301
1302         if (!root->fs_info->log_root_recovering) {
1303                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1304                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1305         }
1306
1307         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1308                sizeof(root->fs_info->super_copy));
1309
1310         trans->transaction->blocked = 0;
1311         spin_lock(&root->fs_info->trans_lock);
1312         root->fs_info->running_transaction = NULL;
1313         root->fs_info->trans_no_join = 0;
1314         spin_unlock(&root->fs_info->trans_lock);
1315
1316         wake_up(&root->fs_info->transaction_wait);
1317
1318         ret = btrfs_write_and_wait_transaction(trans, root);
1319         BUG_ON(ret);
1320         write_ctree_super(trans, root, 0);
1321
1322         /*
1323          * the super is written, we can safely allow the tree-loggers
1324          * to go about their business
1325          */
1326         mutex_unlock(&root->fs_info->tree_log_mutex);
1327
1328         btrfs_finish_extent_commit(trans, root);
1329
1330         cur_trans->commit_done = 1;
1331
1332         root->fs_info->last_trans_committed = cur_trans->transid;
1333
1334         wake_up(&cur_trans->commit_wait);
1335
1336         spin_lock(&root->fs_info->trans_lock);
1337         list_del_init(&cur_trans->list);
1338         spin_unlock(&root->fs_info->trans_lock);
1339
1340         put_transaction(cur_trans);
1341         put_transaction(cur_trans);
1342
1343         trace_btrfs_transaction_commit(root);
1344
1345         btrfs_scrub_continue(root);
1346
1347         if (current->journal_info == trans)
1348                 current->journal_info = NULL;
1349
1350         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1351
1352         if (current != root->fs_info->transaction_kthread)
1353                 btrfs_run_delayed_iputs(root);
1354
1355         return ret;
1356 }
1357
1358 /*
1359  * interface function to delete all the snapshots we have scheduled for deletion
1360  */
1361 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1362 {
1363         LIST_HEAD(list);
1364         struct btrfs_fs_info *fs_info = root->fs_info;
1365
1366         spin_lock(&fs_info->trans_lock);
1367         list_splice_init(&fs_info->dead_roots, &list);
1368         spin_unlock(&fs_info->trans_lock);
1369
1370         while (!list_empty(&list)) {
1371                 root = list_entry(list.next, struct btrfs_root, root_list);
1372                 list_del(&root->root_list);
1373
1374                 btrfs_kill_all_delayed_nodes(root);
1375
1376                 if (btrfs_header_backref_rev(root->node) <
1377                     BTRFS_MIXED_BACKREF_REV)
1378                         btrfs_drop_snapshot(root, NULL, 0);
1379                 else
1380                         btrfs_drop_snapshot(root, NULL, 1);
1381         }
1382         return 0;
1383 }