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 loop:
  59         if (root->fs_info->trans_no_join) {
  60                 if (!nofail) {
  61                         spin_unlock(&root->fs_info->trans_lock);
  62                         return -EBUSY;
  63                 }
  64         }
  65
  66         cur_trans = root->fs_info->running_transaction;
  67         if (cur_trans) {
  68                 atomic_inc(&cur_trans->use_count);
  69                 atomic_inc(&cur_trans->num_writers);
  70                 cur_trans->num_joined++;
  71                 spin_unlock(&root->fs_info->trans_lock);
  72                 return 0;
  73         }
  74         spin_unlock(&root->fs_info->trans_lock);
  75
  76         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  77         if (!cur_trans)
  78                 return -ENOMEM;
  79
  80         spin_lock(&root->fs_info->trans_lock);
  81         if (root->fs_info->running_transaction) {
  82                 /*
  83                  * someone started a transaction after we unlocked.  Make sure
  84                  * to redo the trans_no_join checks above
  85                  */
  86                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  87                 cur_trans = root->fs_info->running_transaction;
  88                 goto loop;
  89         }
  90
  91         atomic_set(&cur_trans->num_writers, 1);
  92         cur_trans->num_joined = 0;
  93         init_waitqueue_head(&cur_trans->writer_wait);
  94         init_waitqueue_head(&cur_trans->commit_wait);
  95         cur_trans->in_commit = 0;
  96         cur_trans->blocked = 0;
  97         /*
  98          * One for this trans handle, one so it will live on until we
  99          * commit the transaction.
 100          */
 101         atomic_set(&cur_trans->use_count, 2);
 102         cur_trans->commit_done = 0;
 103         cur_trans->start_time = get_seconds();
 104
 105         cur_trans->delayed_refs.root = RB_ROOT;
 106         cur_trans->delayed_refs.num_entries = 0;
 107         cur_trans->delayed_refs.num_heads_ready = 0;
 108         cur_trans->delayed_refs.num_heads = 0;
 109         cur_trans->delayed_refs.flushing = 0;
 110         cur_trans->delayed_refs.run_delayed_start = 0;
 111         spin_lock_init(&cur_trans->commit_lock);
 112         spin_lock_init(&cur_trans->delayed_refs.lock);
 113
 114         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 115         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 116         extent_io_tree_init(&cur_trans->dirty_pages,
 117                              root->fs_info->btree_inode->i_mapping);
 118         root->fs_info->generation++;
 119         cur_trans->transid = root->fs_info->generation;
 120         root->fs_info->running_transaction = cur_trans;
 121         spin_unlock(&root->fs_info->trans_lock);
 122
 123         return 0;
 124 }
 125
 126 /*
 127  * this does all the record keeping required to make sure that a reference
 128  * counted root is properly recorded in a given transaction.  This is required
 129  * to make sure the old root from before we joined the transaction is deleted
 130  * when the transaction commits
 131  */
 132 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 133                                struct btrfs_root *root)
 134 {
 135         if (root->ref_cows && root->last_trans < trans->transid) {
 136                 WARN_ON(root == root->fs_info->extent_root);
 137                 WARN_ON(root->commit_root != root->node);
 138
 139                 /*
 140                  * see below for in_trans_setup usage rules
 141                  * we have the reloc mutex held now, so there
 142                  * is only one writer in this function
 143                  */
 144                 root->in_trans_setup = 1;
 145
 146                 /* make sure readers find in_trans_setup before
 147                  * they find our root->last_trans update
 148                  */
 149                 smp_wmb();
 150
 151                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 152                 if (root->last_trans == trans->transid) {
 153                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 154                         return 0;
 155                 }
 156                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 157                            (unsigned long)root->root_key.objectid,
 158                            BTRFS_ROOT_TRANS_TAG);
 159                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 160                 root->last_trans = trans->transid;
 161
 162                 /* this is pretty tricky.  We don't want to
 163                  * take the relocation lock in btrfs_record_root_in_trans
 164                  * unless we're really doing the first setup for this root in
 165                  * this transaction.
 166                  *
 167                  * Normally we'd use root->last_trans as a flag to decide
 168                  * if we want to take the expensive mutex.
 169                  *
 170                  * But, we have to set root->last_trans before we
 171                  * init the relocation root, otherwise, we trip over warnings
 172                  * in ctree.c.  The solution used here is to flag ourselves
 173                  * with root->in_trans_setup.  When this is 1, we're still
 174                  * fixing up the reloc trees and everyone must wait.
 175                  *
 176                  * When this is zero, they can trust root->last_trans and fly
 177                  * through btrfs_record_root_in_trans without having to take the
 178                  * lock.  smp_wmb() makes sure that all the writes above are
 179                  * done before we pop in the zero below
 180                  */
 181                 btrfs_init_reloc_root(trans, root);
 182                 smp_wmb();
 183                 root->in_trans_setup = 0;
 184         }
 185         return 0;
 186 }
 187
 188
 189 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 190                                struct btrfs_root *root)
 191 {
 192         if (!root->ref_cows)
 193                 return 0;
 194
 195         /*
 196          * see record_root_in_trans for comments about in_trans_setup usage
 197          * and barriers
 198          */
 199         smp_rmb();
 200         if (root->last_trans == trans->transid &&
 201             !root->in_trans_setup)
 202                 return 0;
 203
 204         mutex_lock(&root->fs_info->reloc_mutex);
 205         record_root_in_trans(trans, root);
 206         mutex_unlock(&root->fs_info->reloc_mutex);
 207
 208         return 0;
 209 }
 210
 211 /* wait for commit against the current transaction to become unblocked
 212  * when this is done, it is safe to start a new transaction, but the current
 213  * transaction might not be fully on disk.
 214  */
 215 static void wait_current_trans(struct btrfs_root *root)
 216 {
 217         struct btrfs_transaction *cur_trans;
 218
 219         spin_lock(&root->fs_info->trans_lock);
 220         cur_trans = root->fs_info->running_transaction;
 221         if (cur_trans && cur_trans->blocked) {
 222                 atomic_inc(&cur_trans->use_count);
 223                 spin_unlock(&root->fs_info->trans_lock);
 224
 225                 wait_event(root->fs_info->transaction_wait,
 226                            !cur_trans->blocked);
 227                 put_transaction(cur_trans);
 228         } else {
 229                 spin_unlock(&root->fs_info->trans_lock);
 230         }
 231 }
 232
 233 enum btrfs_trans_type {
 234         TRANS_START,
 235         TRANS_JOIN,
 236         TRANS_USERSPACE,
 237         TRANS_JOIN_NOLOCK,
 238 };
 239
 240 static int may_wait_transaction(struct btrfs_root *root, int type)
 241 {
 242         if (root->fs_info->log_root_recovering)
 243                 return 0;
 244
 245         if (type == TRANS_USERSPACE)
 246                 return 1;
 247
 248         if (type == TRANS_START &&
 249             !atomic_read(&root->fs_info->open_ioctl_trans))
 250                 return 1;
 251
 252         return 0;
 253 }
 254
 255 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 256                                                     u64 num_items, int type)
 257 {
 258         struct btrfs_trans_handle *h;
 259         struct btrfs_transaction *cur_trans;
 260         u64 num_bytes = 0;
 261         int ret;
 262
 263         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 264                 return ERR_PTR(-EROFS);
 265
 266         if (current->journal_info) {
 267                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 268                 h = current->journal_info;
 269                 h->use_count++;
 270                 h->orig_rsv = h->block_rsv;
 271                 h->block_rsv = NULL;
 272                 goto got_it;
 273         }
 274
 275         /*
 276          * Do the reservation before we join the transaction so we can do all
 277          * the appropriate flushing if need be.
 278          */
 279         if (num_items > 0 && root != root->fs_info->chunk_root) {
 280                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 281                 ret = btrfs_block_rsv_add(root,
 282                                           &root->fs_info->trans_block_rsv,
 283                                           num_bytes);
 284                 if (ret)
 285                         return ERR_PTR(ret);
 286         }
 287 again:
 288         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 289         if (!h)
 290                 return ERR_PTR(-ENOMEM);
 291
 292         if (may_wait_transaction(root, type))
 293                 wait_current_trans(root);
 294
 295         do {
 296                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 297                 if (ret == -EBUSY)
 298                         wait_current_trans(root);
 299         } while (ret == -EBUSY);
 300
 301         if (ret < 0) {
 302                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 303                 return ERR_PTR(ret);
 304         }
 305
 306         cur_trans = root->fs_info->running_transaction;
 307
 308         h->transid = cur_trans->transid;
 309         h->transaction = cur_trans;
 310         h->blocks_used = 0;
 311         h->bytes_reserved = 0;
 312         h->delayed_ref_updates = 0;
 313         h->use_count = 1;
 314         h->block_rsv = NULL;
 315         h->orig_rsv = NULL;
 316
 317         smp_mb();
 318         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 319                 btrfs_commit_transaction(h, root);
 320                 goto again;
 321         }
 322
 323         if (num_bytes) {
 324                 h->block_rsv = &root->fs_info->trans_block_rsv;
 325                 h->bytes_reserved = num_bytes;
 326         }
 327
 328 got_it:
 329         btrfs_record_root_in_trans(h, root);
 330
 331         if (!current->journal_info && type != TRANS_USERSPACE)
 332                 current->journal_info = h;
 333         return h;
 334 }
 335
 336 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 337                                                    int num_items)
 338 {
 339         return start_transaction(root, num_items, TRANS_START);
 340 }
 341 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 342 {
 343         return start_transaction(root, 0, TRANS_JOIN);
 344 }
 345
 346 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 347 {
 348         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 349 }
 350
 351 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 352 {
 353         return start_transaction(root, 0, TRANS_USERSPACE);
 354 }
 355
 356 /* wait for a transaction commit to be fully complete */
 357 static noinline void wait_for_commit(struct btrfs_root *root,
 358                                     struct btrfs_transaction *commit)
 359 {
 360         wait_event(commit->commit_wait, commit->commit_done);
 361 }
 362
 363 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 364 {
 365         struct btrfs_transaction *cur_trans = NULL, *t;
 366         int ret;
 367
 368         ret = 0;
 369         if (transid) {
 370                 if (transid <= root->fs_info->last_trans_committed)
 371                         goto out;
 372
 373                 /* find specified transaction */
 374                 spin_lock(&root->fs_info->trans_lock);
 375                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 376                         if (t->transid == transid) {
 377                                 cur_trans = t;
 378                                 atomic_inc(&cur_trans->use_count);
 379                                 break;
 380                         }
 381                         if (t->transid > transid)
 382                                 break;
 383                 }
 384                 spin_unlock(&root->fs_info->trans_lock);
 385                 ret = -EINVAL;
 386                 if (!cur_trans)
 387                         goto out;  /* bad transid */
 388         } else {
 389                 /* find newest transaction that is committing | committed */
 390                 spin_lock(&root->fs_info->trans_lock);
 391                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 392                                             list) {
 393                         if (t->in_commit) {
 394                                 if (t->commit_done)
 395                                         break;
 396                                 cur_trans = t;
 397                                 atomic_inc(&cur_trans->use_count);
 398                                 break;
 399                         }
 400                 }
 401                 spin_unlock(&root->fs_info->trans_lock);
 402                 if (!cur_trans)
 403                         goto out;  /* nothing committing|committed */
 404         }
 405
 406         wait_for_commit(root, cur_trans);
 407
 408         put_transaction(cur_trans);
 409         ret = 0;
 410 out:
 411         return ret;
 412 }
 413
 414 void btrfs_throttle(struct btrfs_root *root)
 415 {
 416         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 417                 wait_current_trans(root);
 418 }
 419
 420 static int should_end_transaction(struct btrfs_trans_handle *trans,
 421                                   struct btrfs_root *root)
 422 {
 423         int ret;
 424
 425         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 426         return ret ? 1 : 0;
 427 }
 428
 429 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 430                                  struct btrfs_root *root)
 431 {
 432         struct btrfs_transaction *cur_trans = trans->transaction;
 433         struct btrfs_block_rsv *rsv = trans->block_rsv;
 434         int updates;
 435
 436         smp_mb();
 437         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 438                 return 1;
 439
 440         /*
 441          * We need to do this in case we're deleting csums so the global block
 442          * rsv get's used instead of the csum block rsv.
 443          */
 444         trans->block_rsv = NULL;
 445
 446         updates = trans->delayed_ref_updates;
 447         trans->delayed_ref_updates = 0;
 448         if (updates)
 449                 btrfs_run_delayed_refs(trans, root, updates);
 450
 451         trans->block_rsv = rsv;
 452
 453         return should_end_transaction(trans, root);
 454 }
 455
 456 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 457                           struct btrfs_root *root, int throttle, int lock)
 458 {
 459         struct btrfs_transaction *cur_trans = trans->transaction;
 460         struct btrfs_fs_info *info = root->fs_info;
 461         int count = 0;
 462
 463         if (trans->use_count > 1) {
 464                 trans->use_count--;
 465                 trans->block_rsv = trans->orig_rsv;
 466                 return 0;
 467         }
 468
 469         btrfs_trans_release_metadata(trans, root);
 470         trans->block_rsv = NULL;
 471         while (count < 4) {
 472                 unsigned long cur = trans->delayed_ref_updates;
 473                 trans->delayed_ref_updates = 0;
 474                 if (cur &&
 475                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 476                         trans->delayed_ref_updates = 0;
 477
 478                         /*
 479                          * do a full flush if the transaction is trying
 480                          * to close
 481                          */
 482                         if (trans->transaction->delayed_refs.flushing)
 483                                 cur = 0;
 484                         btrfs_run_delayed_refs(trans, root, cur);
 485                 } else {
 486                         break;
 487                 }
 488                 count++;
 489         }
 490
 491         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 492             should_end_transaction(trans, root)) {
 493                 trans->transaction->blocked = 1;
 494                 smp_wmb();
 495         }
 496
 497         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 498                 if (throttle)
 499                         return btrfs_commit_transaction(trans, root);
 500                 else
 501                         wake_up_process(info->transaction_kthread);
 502         }
 503
 504         WARN_ON(cur_trans != info->running_transaction);
 505         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 506         atomic_dec(&cur_trans->num_writers);
 507
 508         smp_mb();
 509         if (waitqueue_active(&cur_trans->writer_wait))
 510                 wake_up(&cur_trans->writer_wait);
 511         put_transaction(cur_trans);
 512
 513         if (current->journal_info == trans)
 514                 current->journal_info = NULL;
 515         memset(trans, 0, sizeof(*trans));
 516         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 517
 518         if (throttle)
 519                 btrfs_run_delayed_iputs(root);
 520
 521         return 0;
 522 }
 523
 524 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 525                           struct btrfs_root *root)
 526 {
 527         int ret;
 528
 529         ret = __btrfs_end_transaction(trans, root, 0, 1);
 530         if (ret)
 531                 return ret;
 532         return 0;
 533 }
 534
 535 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 536                                    struct btrfs_root *root)
 537 {
 538         int ret;
 539
 540         ret = __btrfs_end_transaction(trans, root, 1, 1);
 541         if (ret)
 542                 return ret;
 543         return 0;
 544 }
 545
 546 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 547                                  struct btrfs_root *root)
 548 {
 549         int ret;
 550
 551         ret = __btrfs_end_transaction(trans, root, 0, 0);
 552         if (ret)
 553                 return ret;
 554         return 0;
 555 }
 556
 557 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 558                                 struct btrfs_root *root)
 559 {
 560         return __btrfs_end_transaction(trans, root, 1, 1);
 561 }
 562
 563 /*
 564  * when btree blocks are allocated, they have some corresponding bits set for
 565  * them in one of two extent_io trees.  This is used to make sure all of
 566  * those extents are sent to disk but does not wait on them
 567  */
 568 int btrfs_write_marked_extents(struct btrfs_root *root,
 569                                struct extent_io_tree *dirty_pages, int mark)
 570 {
 571         int err = 0;
 572         int werr = 0;
 573         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 574         u64 start = 0;
 575         u64 end;
 576
 577         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 578                                       mark)) {
 579                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 580                                    GFP_NOFS);
 581                 err = filemap_fdatawrite_range(mapping, start, end);
 582                 if (err)
 583                         werr = err;
 584                 cond_resched();
 585                 start = end + 1;
 586         }
 587         if (err)
 588                 werr = err;
 589         return werr;
 590 }
 591
 592 /*
 593  * when btree blocks are allocated, they have some corresponding bits set for
 594  * them in one of two extent_io trees.  This is used to make sure all of
 595  * those extents are on disk for transaction or log commit.  We wait
 596  * on all the pages and clear them from the dirty pages state tree
 597  */
 598 int btrfs_wait_marked_extents(struct btrfs_root *root,
 599                               struct extent_io_tree *dirty_pages, int mark)
 600 {
 601         int err = 0;
 602         int werr = 0;
 603         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 604         u64 start = 0;
 605         u64 end;
 606
 607         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 608                                       EXTENT_NEED_WAIT)) {
 609                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 610                 err = filemap_fdatawait_range(mapping, start, end);
 611                 if (err)
 612                         werr = err;
 613                 cond_resched();
 614                 start = end + 1;
 615         }
 616         if (err)
 617                 werr = err;
 618         return werr;
 619 }
 620
 621 /*
 622  * when btree blocks are allocated, they have some corresponding bits set for
 623  * them in one of two extent_io trees.  This is used to make sure all of
 624  * those extents are on disk for transaction or log commit
 625  */
 626 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 627                                 struct extent_io_tree *dirty_pages, int mark)
 628 {
 629         int ret;
 630         int ret2;
 631
 632         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 633         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 634
 635         if (ret)
 636                 return ret;
 637         if (ret2)
 638                 return ret2;
 639         return 0;
 640 }
 641
 642 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 643                                      struct btrfs_root *root)
 644 {
 645         if (!trans || !trans->transaction) {
 646                 struct inode *btree_inode;
 647                 btree_inode = root->fs_info->btree_inode;
 648                 return filemap_write_and_wait(btree_inode->i_mapping);
 649         }
 650         return btrfs_write_and_wait_marked_extents(root,
 651                                            &trans->transaction->dirty_pages,
 652                                            EXTENT_DIRTY);
 653 }
 654
 655 /*
 656  * this is used to update the root pointer in the tree of tree roots.
 657  *
 658  * But, in the case of the extent allocation tree, updating the root
 659  * pointer may allocate blocks which may change the root of the extent
 660  * allocation tree.
 661  *
 662  * So, this loops and repeats and makes sure the cowonly root didn't
 663  * change while the root pointer was being updated in the metadata.
 664  */
 665 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 666                                struct btrfs_root *root)
 667 {
 668         int ret;
 669         u64 old_root_bytenr;
 670         u64 old_root_used;
 671         struct btrfs_root *tree_root = root->fs_info->tree_root;
 672
 673         old_root_used = btrfs_root_used(&root->root_item);
 674         btrfs_write_dirty_block_groups(trans, root);
 675
 676         while (1) {
 677                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 678                 if (old_root_bytenr == root->node->start &&
 679                     old_root_used == btrfs_root_used(&root->root_item))
 680                         break;
 681
 682                 btrfs_set_root_node(&root->root_item, root->node);
 683                 ret = btrfs_update_root(trans, tree_root,
 684                                         &root->root_key,
 685                                         &root->root_item);
 686                 BUG_ON(ret);
 687
 688                 old_root_used = btrfs_root_used(&root->root_item);
 689                 ret = btrfs_write_dirty_block_groups(trans, root);
 690                 BUG_ON(ret);
 691         }
 692
 693         if (root != root->fs_info->extent_root)
 694                 switch_commit_root(root);
 695
 696         return 0;
 697 }
 698
 699 /*
 700  * update all the cowonly tree roots on disk
 701  */
 702 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 703                                          struct btrfs_root *root)
 704 {
 705         struct btrfs_fs_info *fs_info = root->fs_info;
 706         struct list_head *next;
 707         struct extent_buffer *eb;
 708         int ret;
 709
 710         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 711         BUG_ON(ret);
 712
 713         eb = btrfs_lock_root_node(fs_info->tree_root);
 714         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 715         btrfs_tree_unlock(eb);
 716         free_extent_buffer(eb);
 717
 718         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 719         BUG_ON(ret);
 720
 721         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 722                 next = fs_info->dirty_cowonly_roots.next;
 723                 list_del_init(next);
 724                 root = list_entry(next, struct btrfs_root, dirty_list);
 725
 726                 update_cowonly_root(trans, root);
 727         }
 728
 729         down_write(&fs_info->extent_commit_sem);
 730         switch_commit_root(fs_info->extent_root);
 731         up_write(&fs_info->extent_commit_sem);
 732
 733         return 0;
 734 }
 735
 736 /*
 737  * dead roots are old snapshots that need to be deleted.  This allocates
 738  * a dirty root struct and adds it into the list of dead roots that need to
 739  * be deleted
 740  */
 741 int btrfs_add_dead_root(struct btrfs_root *root)
 742 {
 743         spin_lock(&root->fs_info->trans_lock);
 744         list_add(&root->root_list, &root->fs_info->dead_roots);
 745         spin_unlock(&root->fs_info->trans_lock);
 746         return 0;
 747 }
 748
 749 /*
 750  * update all the cowonly tree roots on disk
 751  */
 752 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 753                                     struct btrfs_root *root)
 754 {
 755         struct btrfs_root *gang[8];
 756         struct btrfs_fs_info *fs_info = root->fs_info;
 757         int i;
 758         int ret;
 759         int err = 0;
 760
 761         spin_lock(&fs_info->fs_roots_radix_lock);
 762         while (1) {
 763                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 764                                                  (void **)gang, 0,
 765                                                  ARRAY_SIZE(gang),
 766                                                  BTRFS_ROOT_TRANS_TAG);
 767                 if (ret == 0)
 768                         break;
 769                 for (i = 0; i < ret; i++) {
 770                         root = gang[i];
 771                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 772                                         (unsigned long)root->root_key.objectid,
 773                                         BTRFS_ROOT_TRANS_TAG);
 774                         spin_unlock(&fs_info->fs_roots_radix_lock);
 775
 776                         btrfs_free_log(trans, root);
 777                         btrfs_update_reloc_root(trans, root);
 778                         btrfs_orphan_commit_root(trans, root);
 779
 780                         btrfs_save_ino_cache(root, trans);
 781
 782                         /* see comments in should_cow_block() */
 783                         root->force_cow = 0;
 784                         smp_wmb();
 785
 786                         if (root->commit_root != root->node) {
 787                                 mutex_lock(&root->fs_commit_mutex);
 788                                 switch_commit_root(root);
 789                                 btrfs_unpin_free_ino(root);
 790                                 mutex_unlock(&root->fs_commit_mutex);
 791
 792                                 btrfs_set_root_node(&root->root_item,
 793                                                     root->node);
 794                         }
 795
 796                         err = btrfs_update_root(trans, fs_info->tree_root,
 797                                                 &root->root_key,
 798                                                 &root->root_item);
 799                         spin_lock(&fs_info->fs_roots_radix_lock);
 800                         if (err)
 801                                 break;
 802                 }
 803         }
 804         spin_unlock(&fs_info->fs_roots_radix_lock);
 805         return err;
 806 }
 807
 808 /*
 809  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 810  * otherwise every leaf in the btree is read and defragged.
 811  */
 812 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 813 {
 814         struct btrfs_fs_info *info = root->fs_info;
 815         struct btrfs_trans_handle *trans;
 816         int ret;
 817         unsigned long nr;
 818
 819         if (xchg(&root->defrag_running, 1))
 820                 return 0;
 821
 822         while (1) {
 823                 trans = btrfs_start_transaction(root, 0);
 824                 if (IS_ERR(trans))
 825                         return PTR_ERR(trans);
 826
 827                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 828
 829                 nr = trans->blocks_used;
 830                 btrfs_end_transaction(trans, root);
 831                 btrfs_btree_balance_dirty(info->tree_root, nr);
 832                 cond_resched();
 833
 834                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 835                         break;
 836         }
 837         root->defrag_running = 0;
 838         return ret;
 839 }
 840
 841 /*
 842  * new snapshots need to be created at a very specific time in the
 843  * transaction commit.  This does the actual creation
 844  */
 845 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 846                                    struct btrfs_fs_info *fs_info,
 847                                    struct btrfs_pending_snapshot *pending)
 848 {
 849         struct btrfs_key key;
 850         struct btrfs_root_item *new_root_item;
 851         struct btrfs_root *tree_root = fs_info->tree_root;
 852         struct btrfs_root *root = pending->root;
 853         struct btrfs_root *parent_root;
 854         struct btrfs_block_rsv *rsv;
 855         struct inode *parent_inode;
 856         struct dentry *parent;
 857         struct dentry *dentry;
 858         struct extent_buffer *tmp;
 859         struct extent_buffer *old;
 860         int ret;
 861         u64 to_reserve = 0;
 862         u64 index = 0;
 863         u64 objectid;
 864         u64 root_flags;
 865
 866         rsv = trans->block_rsv;
 867
 868         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 869         if (!new_root_item) {
 870                 pending->error = -ENOMEM;
 871                 goto fail;
 872         }
 873
 874         ret = btrfs_find_free_objectid(tree_root, &objectid);
 875         if (ret) {
 876                 pending->error = ret;
 877                 goto fail;
 878         }
 879
 880         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 881
 882         if (to_reserve > 0) {
 883                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 884                                                   to_reserve);
 885                 if (ret) {
 886                         pending->error = ret;
 887                         goto fail;
 888                 }
 889         }
 890
 891         key.objectid = objectid;
 892         key.offset = (u64)-1;
 893         key.type = BTRFS_ROOT_ITEM_KEY;
 894
 895         trans->block_rsv = &pending->block_rsv;
 896
 897         dentry = pending->dentry;
 898         parent = dget_parent(dentry);
 899         parent_inode = parent->d_inode;
 900         parent_root = BTRFS_I(parent_inode)->root;
 901         record_root_in_trans(trans, parent_root);
 902
 903         /*
 904          * insert the directory item
 905          */
 906         ret = btrfs_set_inode_index(parent_inode, &index);
 907         BUG_ON(ret);
 908         ret = btrfs_insert_dir_item(trans, parent_root,
 909                                 dentry->d_name.name, dentry->d_name.len,
 910                                 parent_inode, &key,
 911                                 BTRFS_FT_DIR, index);
 912         BUG_ON(ret);
 913
 914         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 915                                          dentry->d_name.len * 2);
 916         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 917         BUG_ON(ret);
 918
 919         /*
 920          * pull in the delayed directory update
 921          * and the delayed inode item
 922          * otherwise we corrupt the FS during
 923          * snapshot
 924          */
 925         ret = btrfs_run_delayed_items(trans, root);
 926         BUG_ON(ret);
 927
 928         record_root_in_trans(trans, root);
 929         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 930         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 931         btrfs_check_and_init_root_item(new_root_item);
 932
 933         root_flags = btrfs_root_flags(new_root_item);
 934         if (pending->readonly)
 935                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 936         else
 937                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 938         btrfs_set_root_flags(new_root_item, root_flags);
 939
 940         old = btrfs_lock_root_node(root);
 941         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 942         btrfs_set_lock_blocking(old);
 943
 944         btrfs_copy_root(trans, root, old, &tmp, objectid);
 945         btrfs_tree_unlock(old);
 946         free_extent_buffer(old);
 947
 948         /* see comments in should_cow_block() */
 949         root->force_cow = 1;
 950         smp_wmb();
 951
 952         btrfs_set_root_node(new_root_item, tmp);
 953         /* record when the snapshot was created in key.offset */
 954         key.offset = trans->transid;
 955         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 956         btrfs_tree_unlock(tmp);
 957         free_extent_buffer(tmp);
 958         BUG_ON(ret);
 959
 960         /*
 961          * insert root back/forward references
 962          */
 963         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 964                                  parent_root->root_key.objectid,
 965                                  btrfs_ino(parent_inode), index,
 966                                  dentry->d_name.name, dentry->d_name.len);
 967         BUG_ON(ret);
 968         dput(parent);
 969
 970         key.offset = (u64)-1;
 971         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 972         BUG_ON(IS_ERR(pending->snap));
 973
 974         btrfs_reloc_post_snapshot(trans, pending);
 975 fail:
 976         kfree(new_root_item);
 977         trans->block_rsv = rsv;
 978         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 979         return 0;
 980 }
 981
 982 /*
 983  * create all the snapshots we've scheduled for creation
 984  */
 985 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 986                                              struct btrfs_fs_info *fs_info)
 987 {
 988         struct btrfs_pending_snapshot *pending;
 989         struct list_head *head = &trans->transaction->pending_snapshots;
 990         int ret;
 991
 992         list_for_each_entry(pending, head, list) {
 993                 ret = create_pending_snapshot(trans, fs_info, pending);
 994                 BUG_ON(ret);
 995         }
 996         return 0;
 997 }
 998
 999 static void update_super_roots(struct btrfs_root *root)
1000 {
1001         struct btrfs_root_item *root_item;
1002         struct btrfs_super_block *super;
1003
1004         super = root->fs_info->super_copy;
1005
1006         root_item = &root->fs_info->chunk_root->root_item;
1007         super->chunk_root = root_item->bytenr;
1008         super->chunk_root_generation = root_item->generation;
1009         super->chunk_root_level = root_item->level;
1010
1011         root_item = &root->fs_info->tree_root->root_item;
1012         super->root = root_item->bytenr;
1013         super->generation = root_item->generation;
1014         super->root_level = root_item->level;
1015         if (btrfs_test_opt(root, SPACE_CACHE))
1016                 super->cache_generation = root_item->generation;
1017 }
1018
1019 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1020 {
1021         int ret = 0;
1022         spin_lock(&info->trans_lock);
1023         if (info->running_transaction)
1024                 ret = info->running_transaction->in_commit;
1025         spin_unlock(&info->trans_lock);
1026         return ret;
1027 }
1028
1029 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1030 {
1031         int ret = 0;
1032         spin_lock(&info->trans_lock);
1033         if (info->running_transaction)
1034                 ret = info->running_transaction->blocked;
1035         spin_unlock(&info->trans_lock);
1036         return ret;
1037 }
1038
1039 /*
1040  * wait for the current transaction commit to start and block subsequent
1041  * transaction joins
1042  */
1043 static void wait_current_trans_commit_start(struct btrfs_root *root,
1044                                             struct btrfs_transaction *trans)
1045 {
1046         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1047 }
1048
1049 /*
1050  * wait for the current transaction to start and then become unblocked.
1051  * caller holds ref.
1052  */
1053 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1054                                          struct btrfs_transaction *trans)
1055 {
1056         wait_event(root->fs_info->transaction_wait,
1057                    trans->commit_done || (trans->in_commit && !trans->blocked));
1058 }
1059
1060 /*
1061  * commit transactions asynchronously. once btrfs_commit_transaction_async
1062  * returns, any subsequent transaction will not be allowed to join.
1063  */
1064 struct btrfs_async_commit {
1065         struct btrfs_trans_handle *newtrans;
1066         struct btrfs_root *root;
1067         struct delayed_work work;
1068 };
1069
1070 static void do_async_commit(struct work_struct *work)
1071 {
1072         struct btrfs_async_commit *ac =
1073                 container_of(work, struct btrfs_async_commit, work.work);
1074
1075         btrfs_commit_transaction(ac->newtrans, ac->root);
1076         kfree(ac);
1077 }
1078
1079 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1080                                    struct btrfs_root *root,
1081                                    int wait_for_unblock)
1082 {
1083         struct btrfs_async_commit *ac;
1084         struct btrfs_transaction *cur_trans;
1085
1086         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1087         if (!ac)
1088                 return -ENOMEM;
1089
1090         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1091         ac->root = root;
1092         ac->newtrans = btrfs_join_transaction(root);
1093         if (IS_ERR(ac->newtrans)) {
1094                 int err = PTR_ERR(ac->newtrans);
1095                 kfree(ac);
1096                 return err;
1097         }
1098
1099         /* take transaction reference */
1100         cur_trans = trans->transaction;
1101         atomic_inc(&cur_trans->use_count);
1102
1103         btrfs_end_transaction(trans, root);
1104         schedule_delayed_work(&ac->work, 0);
1105
1106         /* wait for transaction to start and unblock */
1107         if (wait_for_unblock)
1108                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1109         else
1110                 wait_current_trans_commit_start(root, cur_trans);
1111
1112         if (current->journal_info == trans)
1113                 current->journal_info = NULL;
1114
1115         put_transaction(cur_trans);
1116         return 0;
1117 }
1118
1119 /*
1120  * btrfs_transaction state sequence:
1121  *    in_commit = 0, blocked = 0  (initial)
1122  *    in_commit = 1, blocked = 1
1123  *    blocked = 0
1124  *    commit_done = 1
1125  */
1126 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1127                              struct btrfs_root *root)
1128 {
1129         unsigned long joined = 0;
1130         struct btrfs_transaction *cur_trans;
1131         struct btrfs_transaction *prev_trans = NULL;
1132         DEFINE_WAIT(wait);
1133         int ret;
1134         int should_grow = 0;
1135         unsigned long now = get_seconds();
1136         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1137
1138         btrfs_run_ordered_operations(root, 0);
1139
1140         btrfs_trans_release_metadata(trans, root);
1141         trans->block_rsv = NULL;
1142
1143         /* make a pass through all the delayed refs we have so far
1144          * any runnings procs may add more while we are here
1145          */
1146         ret = btrfs_run_delayed_refs(trans, root, 0);
1147         BUG_ON(ret);
1148
1149         cur_trans = trans->transaction;
1150         /*
1151          * set the flushing flag so procs in this transaction have to
1152          * start sending their work down.
1153          */
1154         cur_trans->delayed_refs.flushing = 1;
1155
1156         ret = btrfs_run_delayed_refs(trans, root, 0);
1157         BUG_ON(ret);
1158
1159         spin_lock(&cur_trans->commit_lock);
1160         if (cur_trans->in_commit) {
1161                 spin_unlock(&cur_trans->commit_lock);
1162                 atomic_inc(&cur_trans->use_count);
1163                 btrfs_end_transaction(trans, root);
1164
1165                 wait_for_commit(root, cur_trans);
1166
1167                 put_transaction(cur_trans);
1168
1169                 return 0;
1170         }
1171
1172         trans->transaction->in_commit = 1;
1173         trans->transaction->blocked = 1;
1174         spin_unlock(&cur_trans->commit_lock);
1175         wake_up(&root->fs_info->transaction_blocked_wait);
1176
1177         spin_lock(&root->fs_info->trans_lock);
1178         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1179                 prev_trans = list_entry(cur_trans->list.prev,
1180                                         struct btrfs_transaction, list);
1181                 if (!prev_trans->commit_done) {
1182                         atomic_inc(&prev_trans->use_count);
1183                         spin_unlock(&root->fs_info->trans_lock);
1184
1185                         wait_for_commit(root, prev_trans);
1186
1187                         put_transaction(prev_trans);
1188                 } else {
1189                         spin_unlock(&root->fs_info->trans_lock);
1190                 }
1191         } else {
1192                 spin_unlock(&root->fs_info->trans_lock);
1193         }
1194
1195         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1196                 should_grow = 1;
1197
1198         do {
1199                 int snap_pending = 0;
1200
1201                 joined = cur_trans->num_joined;
1202                 if (!list_empty(&trans->transaction->pending_snapshots))
1203                         snap_pending = 1;
1204
1205                 WARN_ON(cur_trans != trans->transaction);
1206
1207                 if (flush_on_commit || snap_pending) {
1208                         btrfs_start_delalloc_inodes(root, 1);
1209                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1210                         BUG_ON(ret);
1211                 }
1212
1213                 ret = btrfs_run_delayed_items(trans, root);
1214                 BUG_ON(ret);
1215
1216                 /*
1217                  * rename don't use btrfs_join_transaction, so, once we
1218                  * set the transaction to blocked above, we aren't going
1219                  * to get any new ordered operations.  We can safely run
1220                  * it here and no for sure that nothing new will be added
1221                  * to the list
1222                  */
1223                 btrfs_run_ordered_operations(root, 1);
1224
1225                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1226                                 TASK_UNINTERRUPTIBLE);
1227
1228                 if (atomic_read(&cur_trans->num_writers) > 1)
1229                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1230                 else if (should_grow)
1231                         schedule_timeout(1);
1232
1233                 finish_wait(&cur_trans->writer_wait, &wait);
1234         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1235                  (should_grow && cur_trans->num_joined != joined));
1236
1237         /*
1238          * Ok now we need to make sure to block out any other joins while we
1239          * commit the transaction.  We could have started a join before setting
1240          * no_join so make sure to wait for num_writers to == 1 again.
1241          */
1242         spin_lock(&root->fs_info->trans_lock);
1243         root->fs_info->trans_no_join = 1;
1244         spin_unlock(&root->fs_info->trans_lock);
1245         wait_event(cur_trans->writer_wait,
1246                    atomic_read(&cur_trans->num_writers) == 1);
1247
1248         /*
1249          * the reloc mutex makes sure that we stop
1250          * the balancing code from coming in and moving
1251          * extents around in the middle of the commit
1252          */
1253         mutex_lock(&root->fs_info->reloc_mutex);
1254
1255         ret = btrfs_run_delayed_items(trans, root);
1256         BUG_ON(ret);
1257
1258         ret = create_pending_snapshots(trans, root->fs_info);
1259         BUG_ON(ret);
1260
1261         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1262         BUG_ON(ret);
1263
1264         /*
1265          * make sure none of the code above managed to slip in a
1266          * delayed item
1267          */
1268         btrfs_assert_delayed_root_empty(root);
1269
1270         WARN_ON(cur_trans != trans->transaction);
1271
1272         btrfs_scrub_pause(root);
1273         /* btrfs_commit_tree_roots is responsible for getting the
1274          * various roots consistent with each other.  Every pointer
1275          * in the tree of tree roots has to point to the most up to date
1276          * root for every subvolume and other tree.  So, we have to keep
1277          * the tree logging code from jumping in and changing any
1278          * of the trees.
1279          *
1280          * At this point in the commit, there can't be any tree-log
1281          * writers, but a little lower down we drop the trans mutex
1282          * and let new people in.  By holding the tree_log_mutex
1283          * from now until after the super is written, we avoid races
1284          * with the tree-log code.
1285          */
1286         mutex_lock(&root->fs_info->tree_log_mutex);
1287
1288         ret = commit_fs_roots(trans, root);
1289         BUG_ON(ret);
1290
1291         /* commit_fs_roots gets rid of all the tree log roots, it is now
1292          * safe to free the root of tree log roots
1293          */
1294         btrfs_free_log_root_tree(trans, root->fs_info);
1295
1296         ret = commit_cowonly_roots(trans, root);
1297         BUG_ON(ret);
1298
1299         btrfs_prepare_extent_commit(trans, root);
1300
1301         cur_trans = root->fs_info->running_transaction;
1302
1303         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1304                             root->fs_info->tree_root->node);
1305         switch_commit_root(root->fs_info->tree_root);
1306
1307         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1308                             root->fs_info->chunk_root->node);
1309         switch_commit_root(root->fs_info->chunk_root);
1310
1311         update_super_roots(root);
1312
1313         if (!root->fs_info->log_root_recovering) {
1314                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1315                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1316         }
1317
1318         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1319                sizeof(*root->fs_info->super_copy));
1320
1321         trans->transaction->blocked = 0;
1322         spin_lock(&root->fs_info->trans_lock);
1323         root->fs_info->running_transaction = NULL;
1324         root->fs_info->trans_no_join = 0;
1325         spin_unlock(&root->fs_info->trans_lock);
1326         mutex_unlock(&root->fs_info->reloc_mutex);
1327
1328         wake_up(&root->fs_info->transaction_wait);
1329
1330         ret = btrfs_write_and_wait_transaction(trans, root);
1331         BUG_ON(ret);
1332         write_ctree_super(trans, root, 0);
1333
1334         /*
1335          * the super is written, we can safely allow the tree-loggers
1336          * to go about their business
1337          */
1338         mutex_unlock(&root->fs_info->tree_log_mutex);
1339
1340         btrfs_finish_extent_commit(trans, root);
1341
1342         cur_trans->commit_done = 1;
1343
1344         root->fs_info->last_trans_committed = cur_trans->transid;
1345
1346         wake_up(&cur_trans->commit_wait);
1347
1348         spin_lock(&root->fs_info->trans_lock);
1349         list_del_init(&cur_trans->list);
1350         spin_unlock(&root->fs_info->trans_lock);
1351
1352         put_transaction(cur_trans);
1353         put_transaction(cur_trans);
1354
1355         trace_btrfs_transaction_commit(root);
1356
1357         btrfs_scrub_continue(root);
1358
1359         if (current->journal_info == trans)
1360                 current->journal_info = NULL;
1361
1362         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1363
1364         if (current != root->fs_info->transaction_kthread)
1365                 btrfs_run_delayed_iputs(root);
1366
1367         return ret;
1368 }
1369
1370 /*
1371  * interface function to delete all the snapshots we have scheduled for deletion
1372  */
1373 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1374 {
1375         LIST_HEAD(list);
1376         struct btrfs_fs_info *fs_info = root->fs_info;
1377
1378         spin_lock(&fs_info->trans_lock);
1379         list_splice_init(&fs_info->dead_roots, &list);
1380         spin_unlock(&fs_info->trans_lock);
1381
1382         while (!list_empty(&list)) {
1383                 root = list_entry(list.next, struct btrfs_root, root_list);
1384                 list_del(&root->root_list);
1385
1386                 btrfs_kill_all_delayed_nodes(root);
1387
1388                 if (btrfs_header_backref_rev(root->node) <
1389                     BTRFS_MIXED_BACKREF_REV)
1390                         btrfs_drop_snapshot(root, NULL, 0);
1391                 else
1392                         btrfs_drop_snapshot(root, NULL, 1);
1393         }
1394         return 0;
1395 }