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) {
 464                 trans->block_rsv = trans->orig_rsv;
 465                 return 0;
 466         }
 467
 468         btrfs_trans_release_metadata(trans, root);
 469         trans->block_rsv = NULL;
 470         while (count < 4) {
 471                 unsigned long cur = trans->delayed_ref_updates;
 472                 trans->delayed_ref_updates = 0;
 473                 if (cur &&
 474                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 475                         trans->delayed_ref_updates = 0;
 476
 477                         /*
 478                          * do a full flush if the transaction is trying
 479                          * to close
 480                          */
 481                         if (trans->transaction->delayed_refs.flushing)
 482                                 cur = 0;
 483                         btrfs_run_delayed_refs(trans, root, cur);
 484                 } else {
 485                         break;
 486                 }
 487                 count++;
 488         }
 489
 490         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 491             should_end_transaction(trans, root)) {
 492                 trans->transaction->blocked = 1;
 493                 smp_wmb();
 494         }
 495
 496         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 497                 if (throttle) {
 498                         /*
 499                          * We may race with somebody else here so end up having
 500                          * to call end_transaction on ourselves again, so inc
 501                          * our use_count.
 502                          */
 503                         trans->use_count++;
 504                         return btrfs_commit_transaction(trans, root);
 505                 } else {
 506                         wake_up_process(info->transaction_kthread);
 507                 }
 508         }
 509
 510         WARN_ON(cur_trans != info->running_transaction);
 511         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 512         atomic_dec(&cur_trans->num_writers);
 513
 514         smp_mb();
 515         if (waitqueue_active(&cur_trans->writer_wait))
 516                 wake_up(&cur_trans->writer_wait);
 517         put_transaction(cur_trans);
 518
 519         if (current->journal_info == trans)
 520                 current->journal_info = NULL;
 521         memset(trans, 0, sizeof(*trans));
 522         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 523
 524         if (throttle)
 525                 btrfs_run_delayed_iputs(root);
 526
 527         return 0;
 528 }
 529
 530 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 531                           struct btrfs_root *root)
 532 {
 533         int ret;
 534
 535         ret = __btrfs_end_transaction(trans, root, 0, 1);
 536         if (ret)
 537                 return ret;
 538         return 0;
 539 }
 540
 541 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 542                                    struct btrfs_root *root)
 543 {
 544         int ret;
 545
 546         ret = __btrfs_end_transaction(trans, root, 1, 1);
 547         if (ret)
 548                 return ret;
 549         return 0;
 550 }
 551
 552 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 553                                  struct btrfs_root *root)
 554 {
 555         int ret;
 556
 557         ret = __btrfs_end_transaction(trans, root, 0, 0);
 558         if (ret)
 559                 return ret;
 560         return 0;
 561 }
 562
 563 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 564                                 struct btrfs_root *root)
 565 {
 566         return __btrfs_end_transaction(trans, root, 1, 1);
 567 }
 568
 569 /*
 570  * when btree blocks are allocated, they have some corresponding bits set for
 571  * them in one of two extent_io trees.  This is used to make sure all of
 572  * those extents are sent to disk but does not wait on them
 573  */
 574 int btrfs_write_marked_extents(struct btrfs_root *root,
 575                                struct extent_io_tree *dirty_pages, int mark)
 576 {
 577         int err = 0;
 578         int werr = 0;
 579         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 580         u64 start = 0;
 581         u64 end;
 582
 583         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 584                                       mark)) {
 585                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 586                                    GFP_NOFS);
 587                 err = filemap_fdatawrite_range(mapping, start, end);
 588                 if (err)
 589                         werr = err;
 590                 cond_resched();
 591                 start = end + 1;
 592         }
 593         if (err)
 594                 werr = err;
 595         return werr;
 596 }
 597
 598 /*
 599  * when btree blocks are allocated, they have some corresponding bits set for
 600  * them in one of two extent_io trees.  This is used to make sure all of
 601  * those extents are on disk for transaction or log commit.  We wait
 602  * on all the pages and clear them from the dirty pages state tree
 603  */
 604 int btrfs_wait_marked_extents(struct btrfs_root *root,
 605                               struct extent_io_tree *dirty_pages, int mark)
 606 {
 607         int err = 0;
 608         int werr = 0;
 609         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 610         u64 start = 0;
 611         u64 end;
 612
 613         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 614                                       EXTENT_NEED_WAIT)) {
 615                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 616                 err = filemap_fdatawait_range(mapping, start, end);
 617                 if (err)
 618                         werr = err;
 619                 cond_resched();
 620                 start = end + 1;
 621         }
 622         if (err)
 623                 werr = err;
 624         return werr;
 625 }
 626
 627 /*
 628  * when btree blocks are allocated, they have some corresponding bits set for
 629  * them in one of two extent_io trees.  This is used to make sure all of
 630  * those extents are on disk for transaction or log commit
 631  */
 632 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 633                                 struct extent_io_tree *dirty_pages, int mark)
 634 {
 635         int ret;
 636         int ret2;
 637
 638         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 639         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 640
 641         if (ret)
 642                 return ret;
 643         if (ret2)
 644                 return ret2;
 645         return 0;
 646 }
 647
 648 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 649                                      struct btrfs_root *root)
 650 {
 651         if (!trans || !trans->transaction) {
 652                 struct inode *btree_inode;
 653                 btree_inode = root->fs_info->btree_inode;
 654                 return filemap_write_and_wait(btree_inode->i_mapping);
 655         }
 656         return btrfs_write_and_wait_marked_extents(root,
 657                                            &trans->transaction->dirty_pages,
 658                                            EXTENT_DIRTY);
 659 }
 660
 661 /*
 662  * this is used to update the root pointer in the tree of tree roots.
 663  *
 664  * But, in the case of the extent allocation tree, updating the root
 665  * pointer may allocate blocks which may change the root of the extent
 666  * allocation tree.
 667  *
 668  * So, this loops and repeats and makes sure the cowonly root didn't
 669  * change while the root pointer was being updated in the metadata.
 670  */
 671 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 672                                struct btrfs_root *root)
 673 {
 674         int ret;
 675         u64 old_root_bytenr;
 676         u64 old_root_used;
 677         struct btrfs_root *tree_root = root->fs_info->tree_root;
 678
 679         old_root_used = btrfs_root_used(&root->root_item);
 680         btrfs_write_dirty_block_groups(trans, root);
 681
 682         while (1) {
 683                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 684                 if (old_root_bytenr == root->node->start &&
 685                     old_root_used == btrfs_root_used(&root->root_item))
 686                         break;
 687
 688                 btrfs_set_root_node(&root->root_item, root->node);
 689                 ret = btrfs_update_root(trans, tree_root,
 690                                         &root->root_key,
 691                                         &root->root_item);
 692                 BUG_ON(ret);
 693
 694                 old_root_used = btrfs_root_used(&root->root_item);
 695                 ret = btrfs_write_dirty_block_groups(trans, root);
 696                 BUG_ON(ret);
 697         }
 698
 699         if (root != root->fs_info->extent_root)
 700                 switch_commit_root(root);
 701
 702         return 0;
 703 }
 704
 705 /*
 706  * update all the cowonly tree roots on disk
 707  */
 708 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 709                                          struct btrfs_root *root)
 710 {
 711         struct btrfs_fs_info *fs_info = root->fs_info;
 712         struct list_head *next;
 713         struct extent_buffer *eb;
 714         int ret;
 715
 716         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 717         BUG_ON(ret);
 718
 719         eb = btrfs_lock_root_node(fs_info->tree_root);
 720         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 721         btrfs_tree_unlock(eb);
 722         free_extent_buffer(eb);
 723
 724         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 725         BUG_ON(ret);
 726
 727         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 728                 next = fs_info->dirty_cowonly_roots.next;
 729                 list_del_init(next);
 730                 root = list_entry(next, struct btrfs_root, dirty_list);
 731
 732                 update_cowonly_root(trans, root);
 733         }
 734
 735         down_write(&fs_info->extent_commit_sem);
 736         switch_commit_root(fs_info->extent_root);
 737         up_write(&fs_info->extent_commit_sem);
 738
 739         return 0;
 740 }
 741
 742 /*
 743  * dead roots are old snapshots that need to be deleted.  This allocates
 744  * a dirty root struct and adds it into the list of dead roots that need to
 745  * be deleted
 746  */
 747 int btrfs_add_dead_root(struct btrfs_root *root)
 748 {
 749         spin_lock(&root->fs_info->trans_lock);
 750         list_add(&root->root_list, &root->fs_info->dead_roots);
 751         spin_unlock(&root->fs_info->trans_lock);
 752         return 0;
 753 }
 754
 755 /*
 756  * update all the cowonly tree roots on disk
 757  */
 758 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 759                                     struct btrfs_root *root)
 760 {
 761         struct btrfs_root *gang[8];
 762         struct btrfs_fs_info *fs_info = root->fs_info;
 763         int i;
 764         int ret;
 765         int err = 0;
 766
 767         spin_lock(&fs_info->fs_roots_radix_lock);
 768         while (1) {
 769                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 770                                                  (void **)gang, 0,
 771                                                  ARRAY_SIZE(gang),
 772                                                  BTRFS_ROOT_TRANS_TAG);
 773                 if (ret == 0)
 774                         break;
 775                 for (i = 0; i < ret; i++) {
 776                         root = gang[i];
 777                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 778                                         (unsigned long)root->root_key.objectid,
 779                                         BTRFS_ROOT_TRANS_TAG);
 780                         spin_unlock(&fs_info->fs_roots_radix_lock);
 781
 782                         btrfs_free_log(trans, root);
 783                         btrfs_update_reloc_root(trans, root);
 784                         btrfs_orphan_commit_root(trans, root);
 785
 786                         btrfs_save_ino_cache(root, trans);
 787
 788                         if (root->commit_root != root->node) {
 789                                 mutex_lock(&root->fs_commit_mutex);
 790                                 switch_commit_root(root);
 791                                 btrfs_unpin_free_ino(root);
 792                                 mutex_unlock(&root->fs_commit_mutex);
 793
 794                                 btrfs_set_root_node(&root->root_item,
 795                                                     root->node);
 796                         }
 797
 798                         err = btrfs_update_root(trans, fs_info->tree_root,
 799                                                 &root->root_key,
 800                                                 &root->root_item);
 801                         spin_lock(&fs_info->fs_roots_radix_lock);
 802                         if (err)
 803                                 break;
 804                 }
 805         }
 806         spin_unlock(&fs_info->fs_roots_radix_lock);
 807         return err;
 808 }
 809
 810 /*
 811  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 812  * otherwise every leaf in the btree is read and defragged.
 813  */
 814 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 815 {
 816         struct btrfs_fs_info *info = root->fs_info;
 817         struct btrfs_trans_handle *trans;
 818         int ret;
 819         unsigned long nr;
 820
 821         if (xchg(&root->defrag_running, 1))
 822                 return 0;
 823
 824         while (1) {
 825                 trans = btrfs_start_transaction(root, 0);
 826                 if (IS_ERR(trans))
 827                         return PTR_ERR(trans);
 828
 829                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 830
 831                 nr = trans->blocks_used;
 832                 btrfs_end_transaction(trans, root);
 833                 btrfs_btree_balance_dirty(info->tree_root, nr);
 834                 cond_resched();
 835
 836                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 837                         break;
 838         }
 839         root->defrag_running = 0;
 840         return ret;
 841 }
 842
 843 /*
 844  * new snapshots need to be created at a very specific time in the
 845  * transaction commit.  This does the actual creation
 846  */
 847 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 848                                    struct btrfs_fs_info *fs_info,
 849                                    struct btrfs_pending_snapshot *pending)
 850 {
 851         struct btrfs_key key;
 852         struct btrfs_root_item *new_root_item;
 853         struct btrfs_root *tree_root = fs_info->tree_root;
 854         struct btrfs_root *root = pending->root;
 855         struct btrfs_root *parent_root;
 856         struct btrfs_block_rsv *rsv;
 857         struct inode *parent_inode;
 858         struct dentry *parent;
 859         struct dentry *dentry;
 860         struct extent_buffer *tmp;
 861         struct extent_buffer *old;
 862         int ret;
 863         u64 to_reserve = 0;
 864         u64 index = 0;
 865         u64 objectid;
 866         u64 root_flags;
 867
 868         rsv = trans->block_rsv;
 869
 870         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 871         if (!new_root_item) {
 872                 pending->error = -ENOMEM;
 873                 goto fail;
 874         }
 875
 876         ret = btrfs_find_free_objectid(tree_root, &objectid);
 877         if (ret) {
 878                 pending->error = ret;
 879                 goto fail;
 880         }
 881
 882         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 883
 884         if (to_reserve > 0) {
 885                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 886                                                   to_reserve);
 887                 if (ret) {
 888                         pending->error = ret;
 889                         goto fail;
 890                 }
 891         }
 892
 893         key.objectid = objectid;
 894         key.offset = (u64)-1;
 895         key.type = BTRFS_ROOT_ITEM_KEY;
 896
 897         trans->block_rsv = &pending->block_rsv;
 898
 899         dentry = pending->dentry;
 900         parent = dget_parent(dentry);
 901         parent_inode = parent->d_inode;
 902         parent_root = BTRFS_I(parent_inode)->root;
 903         record_root_in_trans(trans, parent_root);
 904
 905         /*
 906          * insert the directory item
 907          */
 908         ret = btrfs_set_inode_index(parent_inode, &index);
 909         BUG_ON(ret);
 910         ret = btrfs_insert_dir_item(trans, parent_root,
 911                                 dentry->d_name.name, dentry->d_name.len,
 912                                 parent_inode, &key,
 913                                 BTRFS_FT_DIR, index);
 914         BUG_ON(ret);
 915
 916         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 917                                          dentry->d_name.len * 2);
 918         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 919         BUG_ON(ret);
 920
 921         /*
 922          * pull in the delayed directory update
 923          * and the delayed inode item
 924          * otherwise we corrupt the FS during
 925          * snapshot
 926          */
 927         ret = btrfs_run_delayed_items(trans, root);
 928         BUG_ON(ret);
 929
 930         record_root_in_trans(trans, root);
 931         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 932         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 933         btrfs_check_and_init_root_item(new_root_item);
 934
 935         root_flags = btrfs_root_flags(new_root_item);
 936         if (pending->readonly)
 937                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 938         else
 939                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 940         btrfs_set_root_flags(new_root_item, root_flags);
 941
 942         old = btrfs_lock_root_node(root);
 943         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 944         btrfs_set_lock_blocking(old);
 945
 946         btrfs_copy_root(trans, root, old, &tmp, objectid);
 947         btrfs_tree_unlock(old);
 948         free_extent_buffer(old);
 949
 950         btrfs_set_root_node(new_root_item, tmp);
 951         /* record when the snapshot was created in key.offset */
 952         key.offset = trans->transid;
 953         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 954         btrfs_tree_unlock(tmp);
 955         free_extent_buffer(tmp);
 956         BUG_ON(ret);
 957
 958         /*
 959          * insert root back/forward references
 960          */
 961         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 962                                  parent_root->root_key.objectid,
 963                                  btrfs_ino(parent_inode), index,
 964                                  dentry->d_name.name, dentry->d_name.len);
 965         BUG_ON(ret);
 966         dput(parent);
 967
 968         key.offset = (u64)-1;
 969         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 970         BUG_ON(IS_ERR(pending->snap));
 971
 972         btrfs_reloc_post_snapshot(trans, pending);
 973 fail:
 974         kfree(new_root_item);
 975         trans->block_rsv = rsv;
 976         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 977         return 0;
 978 }
 979
 980 /*
 981  * create all the snapshots we've scheduled for creation
 982  */
 983 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 984                                              struct btrfs_fs_info *fs_info)
 985 {
 986         struct btrfs_pending_snapshot *pending;
 987         struct list_head *head = &trans->transaction->pending_snapshots;
 988         int ret;
 989
 990         list_for_each_entry(pending, head, list) {
 991                 ret = create_pending_snapshot(trans, fs_info, pending);
 992                 BUG_ON(ret);
 993         }
 994         return 0;
 995 }
 996
 997 static void update_super_roots(struct btrfs_root *root)
 998 {
 999         struct btrfs_root_item *root_item;
1000         struct btrfs_super_block *super;
1001
1002         super = root->fs_info->super_copy;
1003
1004         root_item = &root->fs_info->chunk_root->root_item;
1005         super->chunk_root = root_item->bytenr;
1006         super->chunk_root_generation = root_item->generation;
1007         super->chunk_root_level = root_item->level;
1008
1009         root_item = &root->fs_info->tree_root->root_item;
1010         super->root = root_item->bytenr;
1011         super->generation = root_item->generation;
1012         super->root_level = root_item->level;
1013         if (btrfs_test_opt(root, SPACE_CACHE))
1014                 super->cache_generation = root_item->generation;
1015 }
1016
1017 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1018 {
1019         int ret = 0;
1020         spin_lock(&info->trans_lock);
1021         if (info->running_transaction)
1022                 ret = info->running_transaction->in_commit;
1023         spin_unlock(&info->trans_lock);
1024         return ret;
1025 }
1026
1027 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1028 {
1029         int ret = 0;
1030         spin_lock(&info->trans_lock);
1031         if (info->running_transaction)
1032                 ret = info->running_transaction->blocked;
1033         spin_unlock(&info->trans_lock);
1034         return ret;
1035 }
1036
1037 /*
1038  * wait for the current transaction commit to start and block subsequent
1039  * transaction joins
1040  */
1041 static void wait_current_trans_commit_start(struct btrfs_root *root,
1042                                             struct btrfs_transaction *trans)
1043 {
1044         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1045 }
1046
1047 /*
1048  * wait for the current transaction to start and then become unblocked.
1049  * caller holds ref.
1050  */
1051 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1052                                          struct btrfs_transaction *trans)
1053 {
1054         wait_event(root->fs_info->transaction_wait,
1055                    trans->commit_done || (trans->in_commit && !trans->blocked));
1056 }
1057
1058 /*
1059  * commit transactions asynchronously. once btrfs_commit_transaction_async
1060  * returns, any subsequent transaction will not be allowed to join.
1061  */
1062 struct btrfs_async_commit {
1063         struct btrfs_trans_handle *newtrans;
1064         struct btrfs_root *root;
1065         struct delayed_work work;
1066 };
1067
1068 static void do_async_commit(struct work_struct *work)
1069 {
1070         struct btrfs_async_commit *ac =
1071                 container_of(work, struct btrfs_async_commit, work.work);
1072
1073         btrfs_commit_transaction(ac->newtrans, ac->root);
1074         kfree(ac);
1075 }
1076
1077 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1078                                    struct btrfs_root *root,
1079                                    int wait_for_unblock)
1080 {
1081         struct btrfs_async_commit *ac;
1082         struct btrfs_transaction *cur_trans;
1083
1084         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1085         if (!ac)
1086                 return -ENOMEM;
1087
1088         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1089         ac->root = root;
1090         ac->newtrans = btrfs_join_transaction(root);
1091         if (IS_ERR(ac->newtrans)) {
1092                 int err = PTR_ERR(ac->newtrans);
1093                 kfree(ac);
1094                 return err;
1095         }
1096
1097         /* take transaction reference */
1098         cur_trans = trans->transaction;
1099         atomic_inc(&cur_trans->use_count);
1100
1101         btrfs_end_transaction(trans, root);
1102         schedule_delayed_work(&ac->work, 0);
1103
1104         /* wait for transaction to start and unblock */
1105         if (wait_for_unblock)
1106                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1107         else
1108                 wait_current_trans_commit_start(root, cur_trans);
1109
1110         if (current->journal_info == trans)
1111                 current->journal_info = NULL;
1112
1113         put_transaction(cur_trans);
1114         return 0;
1115 }
1116
1117 /*
1118  * btrfs_transaction state sequence:
1119  *    in_commit = 0, blocked = 0  (initial)
1120  *    in_commit = 1, blocked = 1
1121  *    blocked = 0
1122  *    commit_done = 1
1123  */
1124 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1125                              struct btrfs_root *root)
1126 {
1127         unsigned long joined = 0;
1128         struct btrfs_transaction *cur_trans;
1129         struct btrfs_transaction *prev_trans = NULL;
1130         DEFINE_WAIT(wait);
1131         int ret;
1132         int should_grow = 0;
1133         unsigned long now = get_seconds();
1134         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1135
1136         btrfs_run_ordered_operations(root, 0);
1137
1138         btrfs_trans_release_metadata(trans, root);
1139         trans->block_rsv = NULL;
1140
1141         /* make a pass through all the delayed refs we have so far
1142          * any runnings procs may add more while we are here
1143          */
1144         ret = btrfs_run_delayed_refs(trans, root, 0);
1145         BUG_ON(ret);
1146
1147         cur_trans = trans->transaction;
1148         /*
1149          * set the flushing flag so procs in this transaction have to
1150          * start sending their work down.
1151          */
1152         cur_trans->delayed_refs.flushing = 1;
1153
1154         ret = btrfs_run_delayed_refs(trans, root, 0);
1155         BUG_ON(ret);
1156
1157         spin_lock(&cur_trans->commit_lock);
1158         if (cur_trans->in_commit) {
1159                 spin_unlock(&cur_trans->commit_lock);
1160                 atomic_inc(&cur_trans->use_count);
1161                 btrfs_end_transaction(trans, root);
1162
1163                 wait_for_commit(root, cur_trans);
1164
1165                 put_transaction(cur_trans);
1166
1167                 return 0;
1168         }
1169
1170         trans->transaction->in_commit = 1;
1171         trans->transaction->blocked = 1;
1172         spin_unlock(&cur_trans->commit_lock);
1173         wake_up(&root->fs_info->transaction_blocked_wait);
1174
1175         spin_lock(&root->fs_info->trans_lock);
1176         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1177                 prev_trans = list_entry(cur_trans->list.prev,
1178                                         struct btrfs_transaction, list);
1179                 if (!prev_trans->commit_done) {
1180                         atomic_inc(&prev_trans->use_count);
1181                         spin_unlock(&root->fs_info->trans_lock);
1182
1183                         wait_for_commit(root, prev_trans);
1184
1185                         put_transaction(prev_trans);
1186                 } else {
1187                         spin_unlock(&root->fs_info->trans_lock);
1188                 }
1189         } else {
1190                 spin_unlock(&root->fs_info->trans_lock);
1191         }
1192
1193         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1194                 should_grow = 1;
1195
1196         do {
1197                 int snap_pending = 0;
1198
1199                 joined = cur_trans->num_joined;
1200                 if (!list_empty(&trans->transaction->pending_snapshots))
1201                         snap_pending = 1;
1202
1203                 WARN_ON(cur_trans != trans->transaction);
1204
1205                 if (flush_on_commit || snap_pending) {
1206                         btrfs_start_delalloc_inodes(root, 1);
1207                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1208                         BUG_ON(ret);
1209                 }
1210
1211                 ret = btrfs_run_delayed_items(trans, root);
1212                 BUG_ON(ret);
1213
1214                 /*
1215                  * rename don't use btrfs_join_transaction, so, once we
1216                  * set the transaction to blocked above, we aren't going
1217                  * to get any new ordered operations.  We can safely run
1218                  * it here and no for sure that nothing new will be added
1219                  * to the list
1220                  */
1221                 btrfs_run_ordered_operations(root, 1);
1222
1223                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1224                                 TASK_UNINTERRUPTIBLE);
1225
1226                 if (atomic_read(&cur_trans->num_writers) > 1)
1227                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1228                 else if (should_grow)
1229                         schedule_timeout(1);
1230
1231                 finish_wait(&cur_trans->writer_wait, &wait);
1232         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1233                  (should_grow && cur_trans->num_joined != joined));
1234
1235         /*
1236          * Ok now we need to make sure to block out any other joins while we
1237          * commit the transaction.  We could have started a join before setting
1238          * no_join so make sure to wait for num_writers to == 1 again.
1239          */
1240         spin_lock(&root->fs_info->trans_lock);
1241         root->fs_info->trans_no_join = 1;
1242         spin_unlock(&root->fs_info->trans_lock);
1243         wait_event(cur_trans->writer_wait,
1244                    atomic_read(&cur_trans->num_writers) == 1);
1245
1246         /*
1247          * the reloc mutex makes sure that we stop
1248          * the balancing code from coming in and moving
1249          * extents around in the middle of the commit
1250          */
1251         mutex_lock(&root->fs_info->reloc_mutex);
1252
1253         ret = btrfs_run_delayed_items(trans, root);
1254         BUG_ON(ret);
1255
1256         ret = create_pending_snapshots(trans, root->fs_info);
1257         BUG_ON(ret);
1258
1259         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1260         BUG_ON(ret);
1261
1262         /*
1263          * make sure none of the code above managed to slip in a
1264          * delayed item
1265          */
1266         btrfs_assert_delayed_root_empty(root);
1267
1268         WARN_ON(cur_trans != trans->transaction);
1269
1270         btrfs_scrub_pause(root);
1271         /* btrfs_commit_tree_roots is responsible for getting the
1272          * various roots consistent with each other.  Every pointer
1273          * in the tree of tree roots has to point to the most up to date
1274          * root for every subvolume and other tree.  So, we have to keep
1275          * the tree logging code from jumping in and changing any
1276          * of the trees.
1277          *
1278          * At this point in the commit, there can't be any tree-log
1279          * writers, but a little lower down we drop the trans mutex
1280          * and let new people in.  By holding the tree_log_mutex
1281          * from now until after the super is written, we avoid races
1282          * with the tree-log code.
1283          */
1284         mutex_lock(&root->fs_info->tree_log_mutex);
1285
1286         ret = commit_fs_roots(trans, root);
1287         BUG_ON(ret);
1288
1289         /* commit_fs_roots gets rid of all the tree log roots, it is now
1290          * safe to free the root of tree log roots
1291          */
1292         btrfs_free_log_root_tree(trans, root->fs_info);
1293
1294         ret = commit_cowonly_roots(trans, root);
1295         BUG_ON(ret);
1296
1297         btrfs_prepare_extent_commit(trans, root);
1298
1299         cur_trans = root->fs_info->running_transaction;
1300
1301         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1302                             root->fs_info->tree_root->node);
1303         switch_commit_root(root->fs_info->tree_root);
1304
1305         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1306                             root->fs_info->chunk_root->node);
1307         switch_commit_root(root->fs_info->chunk_root);
1308
1309         update_super_roots(root);
1310
1311         if (!root->fs_info->log_root_recovering) {
1312                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1313                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1314         }
1315
1316         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1317                sizeof(*root->fs_info->super_copy));
1318
1319         trans->transaction->blocked = 0;
1320         spin_lock(&root->fs_info->trans_lock);
1321         root->fs_info->running_transaction = NULL;
1322         root->fs_info->trans_no_join = 0;
1323         spin_unlock(&root->fs_info->trans_lock);
1324         mutex_unlock(&root->fs_info->reloc_mutex);
1325
1326         wake_up(&root->fs_info->transaction_wait);
1327
1328         ret = btrfs_write_and_wait_transaction(trans, root);
1329         BUG_ON(ret);
1330         write_ctree_super(trans, root, 0);
1331
1332         /*
1333          * the super is written, we can safely allow the tree-loggers
1334          * to go about their business
1335          */
1336         mutex_unlock(&root->fs_info->tree_log_mutex);
1337
1338         btrfs_finish_extent_commit(trans, root);
1339
1340         cur_trans->commit_done = 1;
1341
1342         root->fs_info->last_trans_committed = cur_trans->transid;
1343
1344         wake_up(&cur_trans->commit_wait);
1345
1346         spin_lock(&root->fs_info->trans_lock);
1347         list_del_init(&cur_trans->list);
1348         spin_unlock(&root->fs_info->trans_lock);
1349
1350         put_transaction(cur_trans);
1351         put_transaction(cur_trans);
1352
1353         trace_btrfs_transaction_commit(root);
1354
1355         btrfs_scrub_continue(root);
1356
1357         if (current->journal_info == trans)
1358                 current->journal_info = NULL;
1359
1360         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1361
1362         if (current != root->fs_info->transaction_kthread)
1363                 btrfs_run_delayed_iputs(root);
1364
1365         return ret;
1366 }
1367
1368 /*
1369  * interface function to delete all the snapshots we have scheduled for deletion
1370  */
1371 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1372 {
1373         LIST_HEAD(list);
1374         struct btrfs_fs_info *fs_info = root->fs_info;
1375
1376         spin_lock(&fs_info->trans_lock);
1377         list_splice_init(&fs_info->dead_roots, &list);
1378         spin_unlock(&fs_info->trans_lock);
1379
1380         while (!list_empty(&list)) {
1381                 root = list_entry(list.next, struct btrfs_root, root_list);
1382                 list_del(&root->root_list);
1383
1384                 btrfs_kill_all_delayed_nodes(root);
1385
1386                 if (btrfs_header_backref_rev(root->node) <
1387                     BTRFS_MIXED_BACKREF_REV)
1388                         btrfs_drop_snapshot(root, NULL, 0);
1389                 else
1390                         btrfs_drop_snapshot(root, NULL, 1);
1391         }
1392         return 0;
1393 }