fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30 #include "inode-map.h"
  31
  32 #define BTRFS_ROOT_TRANS_TAG 0
  33
  34 static noinline void put_transaction(struct btrfs_transaction *transaction)
  35 {
  36         WARN_ON(atomic_read(&transaction->use_count) == 0);
  37         if (atomic_dec_and_test(&transaction->use_count)) {
  38                 BUG_ON(!list_empty(&transaction->list));
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56
  57         spin_lock(&root->fs_info->trans_lock);
  58         if (root->fs_info->trans_no_join) {
  59                 if (!nofail) {
  60                         spin_unlock(&root->fs_info->trans_lock);
  61                         return -EBUSY;
  62                 }
  63         }
  64
  65         cur_trans = root->fs_info->running_transaction;
  66         if (cur_trans) {
  67                 atomic_inc(&cur_trans->use_count);
  68                 atomic_inc(&cur_trans->num_writers);
  69                 cur_trans->num_joined++;
  70                 spin_unlock(&root->fs_info->trans_lock);
  71                 return 0;
  72         }
  73         spin_unlock(&root->fs_info->trans_lock);
  74
  75         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  76         if (!cur_trans)
  77                 return -ENOMEM;
  78         spin_lock(&root->fs_info->trans_lock);
  79         if (root->fs_info->running_transaction) {
  80                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  81                 cur_trans = root->fs_info->running_transaction;
  82                 atomic_inc(&cur_trans->use_count);
  83                 atomic_inc(&cur_trans->num_writers);
  84                 cur_trans->num_joined++;
  85                 spin_unlock(&root->fs_info->trans_lock);
  86                 return 0;
  87         }
  88         atomic_set(&cur_trans->num_writers, 1);
  89         cur_trans->num_joined = 0;
  90         init_waitqueue_head(&cur_trans->writer_wait);
  91         init_waitqueue_head(&cur_trans->commit_wait);
  92         cur_trans->in_commit = 0;
  93         cur_trans->blocked = 0;
  94         /*
  95          * One for this trans handle, one so it will live on until we
  96          * commit the transaction.
  97          */
  98         atomic_set(&cur_trans->use_count, 2);
  99         cur_trans->commit_done = 0;
 100         cur_trans->start_time = get_seconds();
 101
 102         cur_trans->delayed_refs.root = RB_ROOT;
 103         cur_trans->delayed_refs.num_entries = 0;
 104         cur_trans->delayed_refs.num_heads_ready = 0;
 105         cur_trans->delayed_refs.num_heads = 0;
 106         cur_trans->delayed_refs.flushing = 0;
 107         cur_trans->delayed_refs.run_delayed_start = 0;
 108         spin_lock_init(&cur_trans->commit_lock);
 109         spin_lock_init(&cur_trans->delayed_refs.lock);
 110
 111         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 112         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 113         extent_io_tree_init(&cur_trans->dirty_pages,
 114                              root->fs_info->btree_inode->i_mapping);
 115         root->fs_info->generation++;
 116         cur_trans->transid = root->fs_info->generation;
 117         root->fs_info->running_transaction = cur_trans;
 118         spin_unlock(&root->fs_info->trans_lock);
 119
 120         return 0;
 121 }
 122
 123 /*
 124  * this does all the record keeping required to make sure that a reference
 125  * counted root is properly recorded in a given transaction.  This is required
 126  * to make sure the old root from before we joined the transaction is deleted
 127  * when the transaction commits
 128  */
 129 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 130                                struct btrfs_root *root)
 131 {
 132         if (root->ref_cows && root->last_trans < trans->transid) {
 133                 WARN_ON(root == root->fs_info->extent_root);
 134                 WARN_ON(root->commit_root != root->node);
 135
 136                 /*
 137                  * see below for in_trans_setup usage rules
 138                  * we have the reloc mutex held now, so there
 139                  * is only one writer in this function
 140                  */
 141                 root->in_trans_setup = 1;
 142
 143                 /* make sure readers find in_trans_setup before
 144                  * they find our root->last_trans update
 145                  */
 146                 smp_wmb();
 147
 148                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 149                 if (root->last_trans == trans->transid) {
 150                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 151                         return 0;
 152                 }
 153                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 154                            (unsigned long)root->root_key.objectid,
 155                            BTRFS_ROOT_TRANS_TAG);
 156                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 157                 root->last_trans = trans->transid;
 158
 159                 /* this is pretty tricky.  We don't want to
 160                  * take the relocation lock in btrfs_record_root_in_trans
 161                  * unless we're really doing the first setup for this root in
 162                  * this transaction.
 163                  *
 164                  * Normally we'd use root->last_trans as a flag to decide
 165                  * if we want to take the expensive mutex.
 166                  *
 167                  * But, we have to set root->last_trans before we
 168                  * init the relocation root, otherwise, we trip over warnings
 169                  * in ctree.c.  The solution used here is to flag ourselves
 170                  * with root->in_trans_setup.  When this is 1, we're still
 171                  * fixing up the reloc trees and everyone must wait.
 172                  *
 173                  * When this is zero, they can trust root->last_trans and fly
 174                  * through btrfs_record_root_in_trans without having to take the
 175                  * lock.  smp_wmb() makes sure that all the writes above are
 176                  * done before we pop in the zero below
 177                  */
 178                 btrfs_init_reloc_root(trans, root);
 179                 smp_wmb();
 180                 root->in_trans_setup = 0;
 181         }
 182         return 0;
 183 }
 184
 185
 186 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 187                                struct btrfs_root *root)
 188 {
 189         if (!root->ref_cows)
 190                 return 0;
 191
 192         /*
 193          * see record_root_in_trans for comments about in_trans_setup usage
 194          * and barriers
 195          */
 196         smp_rmb();
 197         if (root->last_trans == trans->transid &&
 198             !root->in_trans_setup)
 199                 return 0;
 200
 201         mutex_lock(&root->fs_info->reloc_mutex);
 202         record_root_in_trans(trans, root);
 203         mutex_unlock(&root->fs_info->reloc_mutex);
 204
 205         return 0;
 206 }
 207
 208 /* wait for commit against the current transaction to become unblocked
 209  * when this is done, it is safe to start a new transaction, but the current
 210  * transaction might not be fully on disk.
 211  */
 212 static void wait_current_trans(struct btrfs_root *root)
 213 {
 214         struct btrfs_transaction *cur_trans;
 215
 216         spin_lock(&root->fs_info->trans_lock);
 217         cur_trans = root->fs_info->running_transaction;
 218         if (cur_trans && cur_trans->blocked) {
 219                 atomic_inc(&cur_trans->use_count);
 220                 spin_unlock(&root->fs_info->trans_lock);
 221
 222                 wait_event(root->fs_info->transaction_wait,
 223                            !cur_trans->blocked);
 224                 put_transaction(cur_trans);
 225         } else {
 226                 spin_unlock(&root->fs_info->trans_lock);
 227         }
 228 }
 229
 230 enum btrfs_trans_type {
 231         TRANS_START,
 232         TRANS_JOIN,
 233         TRANS_USERSPACE,
 234         TRANS_JOIN_NOLOCK,
 235 };
 236
 237 static int may_wait_transaction(struct btrfs_root *root, int type)
 238 {
 239         if (root->fs_info->log_root_recovering)
 240                 return 0;
 241
 242         if (type == TRANS_USERSPACE)
 243                 return 1;
 244
 245         if (type == TRANS_START &&
 246             !atomic_read(&root->fs_info->open_ioctl_trans))
 247                 return 1;
 248
 249         return 0;
 250 }
 251
 252 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 253                                                     u64 num_items, int type)
 254 {
 255         struct btrfs_trans_handle *h;
 256         struct btrfs_transaction *cur_trans;
 257         u64 num_bytes = 0;
 258         int ret;
 259
 260         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 261                 return ERR_PTR(-EROFS);
 262
 263         if (current->journal_info) {
 264                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 265                 h = current->journal_info;
 266                 h->use_count++;
 267                 h->orig_rsv = h->block_rsv;
 268                 h->block_rsv = NULL;
 269                 goto got_it;
 270         }
 271
 272         /*
 273          * Do the reservation before we join the transaction so we can do all
 274          * the appropriate flushing if need be.
 275          */
 276         if (num_items > 0 && root != root->fs_info->chunk_root) {
 277                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 278                 ret = btrfs_block_rsv_add(root,
 279                                           &root->fs_info->trans_block_rsv,
 280                                           num_bytes);
 281                 if (ret)
 282                         return ERR_PTR(ret);
 283         }
 284 again:
 285         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 286         if (!h)
 287                 return ERR_PTR(-ENOMEM);
 288
 289         if (may_wait_transaction(root, type))
 290                 wait_current_trans(root);
 291
 292         do {
 293                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 294                 if (ret == -EBUSY)
 295                         wait_current_trans(root);
 296         } while (ret == -EBUSY);
 297
 298         if (ret < 0) {
 299                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 300                 return ERR_PTR(ret);
 301         }
 302
 303         cur_trans = root->fs_info->running_transaction;
 304
 305         h->transid = cur_trans->transid;
 306         h->transaction = cur_trans;
 307         h->blocks_used = 0;
 308         h->bytes_reserved = 0;
 309         h->delayed_ref_updates = 0;
 310         h->use_count = 1;
 311         h->block_rsv = NULL;
 312         h->orig_rsv = NULL;
 313
 314         smp_mb();
 315         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 316                 btrfs_commit_transaction(h, root);
 317                 goto again;
 318         }
 319
 320         if (num_bytes) {
 321                 h->block_rsv = &root->fs_info->trans_block_rsv;
 322                 h->bytes_reserved = num_bytes;
 323         }
 324
 325 got_it:
 326         btrfs_record_root_in_trans(h, root);
 327
 328         if (!current->journal_info && type != TRANS_USERSPACE)
 329                 current->journal_info = h;
 330         return h;
 331 }
 332
 333 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 334                                                    int num_items)
 335 {
 336         return start_transaction(root, num_items, TRANS_START);
 337 }
 338 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 339 {
 340         return start_transaction(root, 0, TRANS_JOIN);
 341 }
 342
 343 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 344 {
 345         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 346 }
 347
 348 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 349 {
 350         return start_transaction(root, 0, TRANS_USERSPACE);
 351 }
 352
 353 /* wait for a transaction commit to be fully complete */
 354 static noinline void wait_for_commit(struct btrfs_root *root,
 355                                     struct btrfs_transaction *commit)
 356 {
 357         wait_event(commit->commit_wait, commit->commit_done);
 358 }
 359
 360 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 361 {
 362         struct btrfs_transaction *cur_trans = NULL, *t;
 363         int ret;
 364
 365         ret = 0;
 366         if (transid) {
 367                 if (transid <= root->fs_info->last_trans_committed)
 368                         goto out;
 369
 370                 /* find specified transaction */
 371                 spin_lock(&root->fs_info->trans_lock);
 372                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 373                         if (t->transid == transid) {
 374                                 cur_trans = t;
 375                                 atomic_inc(&cur_trans->use_count);
 376                                 break;
 377                         }
 378                         if (t->transid > transid)
 379                                 break;
 380                 }
 381                 spin_unlock(&root->fs_info->trans_lock);
 382                 ret = -EINVAL;
 383                 if (!cur_trans)
 384                         goto out;  /* bad transid */
 385         } else {
 386                 /* find newest transaction that is committing | committed */
 387                 spin_lock(&root->fs_info->trans_lock);
 388                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 389                                             list) {
 390                         if (t->in_commit) {
 391                                 if (t->commit_done)
 392                                         break;
 393                                 cur_trans = t;
 394                                 atomic_inc(&cur_trans->use_count);
 395                                 break;
 396                         }
 397                 }
 398                 spin_unlock(&root->fs_info->trans_lock);
 399                 if (!cur_trans)
 400                         goto out;  /* nothing committing|committed */
 401         }
 402
 403         wait_for_commit(root, cur_trans);
 404
 405         put_transaction(cur_trans);
 406         ret = 0;
 407 out:
 408         return ret;
 409 }
 410
 411 void btrfs_throttle(struct btrfs_root *root)
 412 {
 413         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 414                 wait_current_trans(root);
 415 }
 416
 417 static int should_end_transaction(struct btrfs_trans_handle *trans,
 418                                   struct btrfs_root *root)
 419 {
 420         int ret;
 421
 422         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 423         return ret ? 1 : 0;
 424 }
 425
 426 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 427                                  struct btrfs_root *root)
 428 {
 429         struct btrfs_transaction *cur_trans = trans->transaction;
 430         struct btrfs_block_rsv *rsv = trans->block_rsv;
 431         int updates;
 432
 433         smp_mb();
 434         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 435                 return 1;
 436
 437         /*
 438          * We need to do this in case we're deleting csums so the global block
 439          * rsv get's used instead of the csum block rsv.
 440          */
 441         trans->block_rsv = NULL;
 442
 443         updates = trans->delayed_ref_updates;
 444         trans->delayed_ref_updates = 0;
 445         if (updates)
 446                 btrfs_run_delayed_refs(trans, root, updates);
 447
 448         trans->block_rsv = rsv;
 449
 450         return should_end_transaction(trans, root);
 451 }
 452
 453 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 454                           struct btrfs_root *root, int throttle, int lock)
 455 {
 456         struct btrfs_transaction *cur_trans = trans->transaction;
 457         struct btrfs_fs_info *info = root->fs_info;
 458         int count = 0;
 459
 460         if (--trans->use_count) {
 461                 trans->block_rsv = trans->orig_rsv;
 462                 return 0;
 463         }
 464
 465         btrfs_trans_release_metadata(trans, root);
 466         trans->block_rsv = NULL;
 467         while (count < 4) {
 468                 unsigned long cur = trans->delayed_ref_updates;
 469                 trans->delayed_ref_updates = 0;
 470                 if (cur &&
 471                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 472                         trans->delayed_ref_updates = 0;
 473
 474                         /*
 475                          * do a full flush if the transaction is trying
 476                          * to close
 477                          */
 478                         if (trans->transaction->delayed_refs.flushing)
 479                                 cur = 0;
 480                         btrfs_run_delayed_refs(trans, root, cur);
 481                 } else {
 482                         break;
 483                 }
 484                 count++;
 485         }
 486
 487         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 488             should_end_transaction(trans, root)) {
 489                 trans->transaction->blocked = 1;
 490                 smp_wmb();
 491         }
 492
 493         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 494                 if (throttle) {
 495                         /*
 496                          * We may race with somebody else here so end up having
 497                          * to call end_transaction on ourselves again, so inc
 498                          * our use_count.
 499                          */
 500                         trans->use_count++;
 501                         return btrfs_commit_transaction(trans, root);
 502                 } else {
 503                         wake_up_process(info->transaction_kthread);
 504                 }
 505         }
 506
 507         WARN_ON(cur_trans != info->running_transaction);
 508         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 509         atomic_dec(&cur_trans->num_writers);
 510
 511         smp_mb();
 512         if (waitqueue_active(&cur_trans->writer_wait))
 513                 wake_up(&cur_trans->writer_wait);
 514         put_transaction(cur_trans);
 515
 516         if (current->journal_info == trans)
 517                 current->journal_info = NULL;
 518         memset(trans, 0, sizeof(*trans));
 519         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 520
 521         if (throttle)
 522                 btrfs_run_delayed_iputs(root);
 523
 524         return 0;
 525 }
 526
 527 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 528                           struct btrfs_root *root)
 529 {
 530         int ret;
 531
 532         ret = __btrfs_end_transaction(trans, root, 0, 1);
 533         if (ret)
 534                 return ret;
 535         return 0;
 536 }
 537
 538 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 539                                    struct btrfs_root *root)
 540 {
 541         int ret;
 542
 543         ret = __btrfs_end_transaction(trans, root, 1, 1);
 544         if (ret)
 545                 return ret;
 546         return 0;
 547 }
 548
 549 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 550                                  struct btrfs_root *root)
 551 {
 552         int ret;
 553
 554         ret = __btrfs_end_transaction(trans, root, 0, 0);
 555         if (ret)
 556                 return ret;
 557         return 0;
 558 }
 559
 560 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 561                                 struct btrfs_root *root)
 562 {
 563         return __btrfs_end_transaction(trans, root, 1, 1);
 564 }
 565
 566 /*
 567  * when btree blocks are allocated, they have some corresponding bits set for
 568  * them in one of two extent_io trees.  This is used to make sure all of
 569  * those extents are sent to disk but does not wait on them
 570  */
 571 int btrfs_write_marked_extents(struct btrfs_root *root,
 572                                struct extent_io_tree *dirty_pages, int mark)
 573 {
 574         int err = 0;
 575         int werr = 0;
 576         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 577         u64 start = 0;
 578         u64 end;
 579
 580         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 581                                       mark)) {
 582                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 583                                    GFP_NOFS);
 584                 err = filemap_fdatawrite_range(mapping, start, end);
 585                 if (err)
 586                         werr = err;
 587                 cond_resched();
 588                 start = end + 1;
 589         }
 590         if (err)
 591                 werr = err;
 592         return werr;
 593 }
 594
 595 /*
 596  * when btree blocks are allocated, they have some corresponding bits set for
 597  * them in one of two extent_io trees.  This is used to make sure all of
 598  * those extents are on disk for transaction or log commit.  We wait
 599  * on all the pages and clear them from the dirty pages state tree
 600  */
 601 int btrfs_wait_marked_extents(struct btrfs_root *root,
 602                               struct extent_io_tree *dirty_pages, int mark)
 603 {
 604         int err = 0;
 605         int werr = 0;
 606         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 607         u64 start = 0;
 608         u64 end;
 609
 610         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 611                                       EXTENT_NEED_WAIT)) {
 612                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 613                 err = filemap_fdatawait_range(mapping, start, end);
 614                 if (err)
 615                         werr = err;
 616                 cond_resched();
 617                 start = end + 1;
 618         }
 619         if (err)
 620                 werr = err;
 621         return werr;
 622 }
 623
 624 /*
 625  * when btree blocks are allocated, they have some corresponding bits set for
 626  * them in one of two extent_io trees.  This is used to make sure all of
 627  * those extents are on disk for transaction or log commit
 628  */
 629 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 630                                 struct extent_io_tree *dirty_pages, int mark)
 631 {
 632         int ret;
 633         int ret2;
 634
 635         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 636         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 637
 638         if (ret)
 639                 return ret;
 640         if (ret2)
 641                 return ret2;
 642         return 0;
 643 }
 644
 645 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 646                                      struct btrfs_root *root)
 647 {
 648         if (!trans || !trans->transaction) {
 649                 struct inode *btree_inode;
 650                 btree_inode = root->fs_info->btree_inode;
 651                 return filemap_write_and_wait(btree_inode->i_mapping);
 652         }
 653         return btrfs_write_and_wait_marked_extents(root,
 654                                            &trans->transaction->dirty_pages,
 655                                            EXTENT_DIRTY);
 656 }
 657
 658 /*
 659  * this is used to update the root pointer in the tree of tree roots.
 660  *
 661  * But, in the case of the extent allocation tree, updating the root
 662  * pointer may allocate blocks which may change the root of the extent
 663  * allocation tree.
 664  *
 665  * So, this loops and repeats and makes sure the cowonly root didn't
 666  * change while the root pointer was being updated in the metadata.
 667  */
 668 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 669                                struct btrfs_root *root)
 670 {
 671         int ret;
 672         u64 old_root_bytenr;
 673         u64 old_root_used;
 674         struct btrfs_root *tree_root = root->fs_info->tree_root;
 675
 676         old_root_used = btrfs_root_used(&root->root_item);
 677         btrfs_write_dirty_block_groups(trans, root);
 678
 679         while (1) {
 680                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 681                 if (old_root_bytenr == root->node->start &&
 682                     old_root_used == btrfs_root_used(&root->root_item))
 683                         break;
 684
 685                 btrfs_set_root_node(&root->root_item, root->node);
 686                 ret = btrfs_update_root(trans, tree_root,
 687                                         &root->root_key,
 688                                         &root->root_item);
 689                 BUG_ON(ret);
 690
 691                 old_root_used = btrfs_root_used(&root->root_item);
 692                 ret = btrfs_write_dirty_block_groups(trans, root);
 693                 BUG_ON(ret);
 694         }
 695
 696         if (root != root->fs_info->extent_root)
 697                 switch_commit_root(root);
 698
 699         return 0;
 700 }
 701
 702 /*
 703  * update all the cowonly tree roots on disk
 704  */
 705 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 706                                          struct btrfs_root *root)
 707 {
 708         struct btrfs_fs_info *fs_info = root->fs_info;
 709         struct list_head *next;
 710         struct extent_buffer *eb;
 711         int ret;
 712
 713         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 714         BUG_ON(ret);
 715
 716         eb = btrfs_lock_root_node(fs_info->tree_root);
 717         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 718         btrfs_tree_unlock(eb);
 719         free_extent_buffer(eb);
 720
 721         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 722         BUG_ON(ret);
 723
 724         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 725                 next = fs_info->dirty_cowonly_roots.next;
 726                 list_del_init(next);
 727                 root = list_entry(next, struct btrfs_root, dirty_list);
 728
 729                 update_cowonly_root(trans, root);
 730         }
 731
 732         down_write(&fs_info->extent_commit_sem);
 733         switch_commit_root(fs_info->extent_root);
 734         up_write(&fs_info->extent_commit_sem);
 735
 736         return 0;
 737 }
 738
 739 /*
 740  * dead roots are old snapshots that need to be deleted.  This allocates
 741  * a dirty root struct and adds it into the list of dead roots that need to
 742  * be deleted
 743  */
 744 int btrfs_add_dead_root(struct btrfs_root *root)
 745 {
 746         spin_lock(&root->fs_info->trans_lock);
 747         list_add(&root->root_list, &root->fs_info->dead_roots);
 748         spin_unlock(&root->fs_info->trans_lock);
 749         return 0;
 750 }
 751
 752 /*
 753  * update all the cowonly tree roots on disk
 754  */
 755 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 756                                     struct btrfs_root *root)
 757 {
 758         struct btrfs_root *gang[8];
 759         struct btrfs_fs_info *fs_info = root->fs_info;
 760         int i;
 761         int ret;
 762         int err = 0;
 763
 764         spin_lock(&fs_info->fs_roots_radix_lock);
 765         while (1) {
 766                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 767                                                  (void **)gang, 0,
 768                                                  ARRAY_SIZE(gang),
 769                                                  BTRFS_ROOT_TRANS_TAG);
 770                 if (ret == 0)
 771                         break;
 772                 for (i = 0; i < ret; i++) {
 773                         root = gang[i];
 774                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 775                                         (unsigned long)root->root_key.objectid,
 776                                         BTRFS_ROOT_TRANS_TAG);
 777                         spin_unlock(&fs_info->fs_roots_radix_lock);
 778
 779                         btrfs_free_log(trans, root);
 780                         btrfs_update_reloc_root(trans, root);
 781                         btrfs_orphan_commit_root(trans, root);
 782
 783                         btrfs_save_ino_cache(root, trans);
 784
 785                         if (root->commit_root != root->node) {
 786                                 mutex_lock(&root->fs_commit_mutex);
 787                                 switch_commit_root(root);
 788                                 btrfs_unpin_free_ino(root);
 789                                 mutex_unlock(&root->fs_commit_mutex);
 790
 791                                 btrfs_set_root_node(&root->root_item,
 792                                                     root->node);
 793                         }
 794
 795                         err = btrfs_update_root(trans, fs_info->tree_root,
 796                                                 &root->root_key,
 797                                                 &root->root_item);
 798                         spin_lock(&fs_info->fs_roots_radix_lock);
 799                         if (err)
 800                                 break;
 801                 }
 802         }
 803         spin_unlock(&fs_info->fs_roots_radix_lock);
 804         return err;
 805 }
 806
 807 /*
 808  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 809  * otherwise every leaf in the btree is read and defragged.
 810  */
 811 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 812 {
 813         struct btrfs_fs_info *info = root->fs_info;
 814         struct btrfs_trans_handle *trans;
 815         int ret;
 816         unsigned long nr;
 817
 818         if (xchg(&root->defrag_running, 1))
 819                 return 0;
 820
 821         while (1) {
 822                 trans = btrfs_start_transaction(root, 0);
 823                 if (IS_ERR(trans))
 824                         return PTR_ERR(trans);
 825
 826                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 827
 828                 nr = trans->blocks_used;
 829                 btrfs_end_transaction(trans, root);
 830                 btrfs_btree_balance_dirty(info->tree_root, nr);
 831                 cond_resched();
 832
 833                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 834                         break;
 835         }
 836         root->defrag_running = 0;
 837         return ret;
 838 }
 839
 840 /*
 841  * new snapshots need to be created at a very specific time in the
 842  * transaction commit.  This does the actual creation
 843  */
 844 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 845                                    struct btrfs_fs_info *fs_info,
 846                                    struct btrfs_pending_snapshot *pending)
 847 {
 848         struct btrfs_key key;
 849         struct btrfs_root_item *new_root_item;
 850         struct btrfs_root *tree_root = fs_info->tree_root;
 851         struct btrfs_root *root = pending->root;
 852         struct btrfs_root *parent_root;
 853         struct btrfs_block_rsv *rsv;
 854         struct inode *parent_inode;
 855         struct dentry *parent;
 856         struct dentry *dentry;
 857         struct extent_buffer *tmp;
 858         struct extent_buffer *old;
 859         int ret;
 860         u64 to_reserve = 0;
 861         u64 index = 0;
 862         u64 objectid;
 863         u64 root_flags;
 864
 865         rsv = trans->block_rsv;
 866
 867         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 868         if (!new_root_item) {
 869                 pending->error = -ENOMEM;
 870                 goto fail;
 871         }
 872
 873         ret = btrfs_find_free_objectid(tree_root, &objectid);
 874         if (ret) {
 875                 pending->error = ret;
 876                 goto fail;
 877         }
 878
 879         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 880
 881         if (to_reserve > 0) {
 882                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
 883                                           to_reserve);
 884                 if (ret) {
 885                         pending->error = ret;
 886                         goto fail;
 887                 }
 888         }
 889
 890         key.objectid = objectid;
 891         key.offset = (u64)-1;
 892         key.type = BTRFS_ROOT_ITEM_KEY;
 893
 894         trans->block_rsv = &pending->block_rsv;
 895
 896         dentry = pending->dentry;
 897         parent = dget_parent(dentry);
 898         parent_inode = parent->d_inode;
 899         parent_root = BTRFS_I(parent_inode)->root;
 900         record_root_in_trans(trans, parent_root);
 901
 902         /*
 903          * insert the directory item
 904          */
 905         ret = btrfs_set_inode_index(parent_inode, &index);
 906         BUG_ON(ret);
 907         ret = btrfs_insert_dir_item(trans, parent_root,
 908                                 dentry->d_name.name, dentry->d_name.len,
 909                                 parent_inode, &key,
 910                                 BTRFS_FT_DIR, index);
 911         BUG_ON(ret);
 912
 913         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 914                                          dentry->d_name.len * 2);
 915         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 916         BUG_ON(ret);
 917
 918         /*
 919          * pull in the delayed directory update
 920          * and the delayed inode item
 921          * otherwise we corrupt the FS during
 922          * snapshot
 923          */
 924         ret = btrfs_run_delayed_items(trans, root);
 925         BUG_ON(ret);
 926
 927         record_root_in_trans(trans, root);
 928         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 929         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 930         btrfs_check_and_init_root_item(new_root_item);
 931
 932         root_flags = btrfs_root_flags(new_root_item);
 933         if (pending->readonly)
 934                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 935         else
 936                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 937         btrfs_set_root_flags(new_root_item, root_flags);
 938
 939         old = btrfs_lock_root_node(root);
 940         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 941         btrfs_set_lock_blocking(old);
 942
 943         btrfs_copy_root(trans, root, old, &tmp, objectid);
 944         btrfs_tree_unlock(old);
 945         free_extent_buffer(old);
 946
 947         btrfs_set_root_node(new_root_item, tmp);
 948         /* record when the snapshot was created in key.offset */
 949         key.offset = trans->transid;
 950         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 951         btrfs_tree_unlock(tmp);
 952         free_extent_buffer(tmp);
 953         BUG_ON(ret);
 954
 955         /*
 956          * insert root back/forward references
 957          */
 958         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 959                                  parent_root->root_key.objectid,
 960                                  btrfs_ino(parent_inode), index,
 961                                  dentry->d_name.name, dentry->d_name.len);
 962         BUG_ON(ret);
 963         dput(parent);
 964
 965         key.offset = (u64)-1;
 966         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 967         BUG_ON(IS_ERR(pending->snap));
 968
 969         btrfs_reloc_post_snapshot(trans, pending);
 970 fail:
 971         kfree(new_root_item);
 972         trans->block_rsv = rsv;
 973         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 974         return 0;
 975 }
 976
 977 /*
 978  * create all the snapshots we've scheduled for creation
 979  */
 980 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 981                                              struct btrfs_fs_info *fs_info)
 982 {
 983         struct btrfs_pending_snapshot *pending;
 984         struct list_head *head = &trans->transaction->pending_snapshots;
 985         int ret;
 986
 987         list_for_each_entry(pending, head, list) {
 988                 ret = create_pending_snapshot(trans, fs_info, pending);
 989                 BUG_ON(ret);
 990         }
 991         return 0;
 992 }
 993
 994 static void update_super_roots(struct btrfs_root *root)
 995 {
 996         struct btrfs_root_item *root_item;
 997         struct btrfs_super_block *super;
 998
 999         super = root->fs_info->super_copy;
1000
1001         root_item = &root->fs_info->chunk_root->root_item;
1002         super->chunk_root = root_item->bytenr;
1003         super->chunk_root_generation = root_item->generation;
1004         super->chunk_root_level = root_item->level;
1005
1006         root_item = &root->fs_info->tree_root->root_item;
1007         super->root = root_item->bytenr;
1008         super->generation = root_item->generation;
1009         super->root_level = root_item->level;
1010         if (btrfs_test_opt(root, SPACE_CACHE))
1011                 super->cache_generation = root_item->generation;
1012 }
1013
1014 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1015 {
1016         int ret = 0;
1017         spin_lock(&info->trans_lock);
1018         if (info->running_transaction)
1019                 ret = info->running_transaction->in_commit;
1020         spin_unlock(&info->trans_lock);
1021         return ret;
1022 }
1023
1024 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1025 {
1026         int ret = 0;
1027         spin_lock(&info->trans_lock);
1028         if (info->running_transaction)
1029                 ret = info->running_transaction->blocked;
1030         spin_unlock(&info->trans_lock);
1031         return ret;
1032 }
1033
1034 /*
1035  * wait for the current transaction commit to start and block subsequent
1036  * transaction joins
1037  */
1038 static void wait_current_trans_commit_start(struct btrfs_root *root,
1039                                             struct btrfs_transaction *trans)
1040 {
1041         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1042 }
1043
1044 /*
1045  * wait for the current transaction to start and then become unblocked.
1046  * caller holds ref.
1047  */
1048 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1049                                          struct btrfs_transaction *trans)
1050 {
1051         wait_event(root->fs_info->transaction_wait,
1052                    trans->commit_done || (trans->in_commit && !trans->blocked));
1053 }
1054
1055 /*
1056  * commit transactions asynchronously. once btrfs_commit_transaction_async
1057  * returns, any subsequent transaction will not be allowed to join.
1058  */
1059 struct btrfs_async_commit {
1060         struct btrfs_trans_handle *newtrans;
1061         struct btrfs_root *root;
1062         struct delayed_work work;
1063 };
1064
1065 static void do_async_commit(struct work_struct *work)
1066 {
1067         struct btrfs_async_commit *ac =
1068                 container_of(work, struct btrfs_async_commit, work.work);
1069
1070         btrfs_commit_transaction(ac->newtrans, ac->root);
1071         kfree(ac);
1072 }
1073
1074 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1075                                    struct btrfs_root *root,
1076                                    int wait_for_unblock)
1077 {
1078         struct btrfs_async_commit *ac;
1079         struct btrfs_transaction *cur_trans;
1080
1081         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1082         if (!ac)
1083                 return -ENOMEM;
1084
1085         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1086         ac->root = root;
1087         ac->newtrans = btrfs_join_transaction(root);
1088         if (IS_ERR(ac->newtrans)) {
1089                 int err = PTR_ERR(ac->newtrans);
1090                 kfree(ac);
1091                 return err;
1092         }
1093
1094         /* take transaction reference */
1095         cur_trans = trans->transaction;
1096         atomic_inc(&cur_trans->use_count);
1097
1098         btrfs_end_transaction(trans, root);
1099         schedule_delayed_work(&ac->work, 0);
1100
1101         /* wait for transaction to start and unblock */
1102         if (wait_for_unblock)
1103                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1104         else
1105                 wait_current_trans_commit_start(root, cur_trans);
1106
1107         if (current->journal_info == trans)
1108                 current->journal_info = NULL;
1109
1110         put_transaction(cur_trans);
1111         return 0;
1112 }
1113
1114 /*
1115  * btrfs_transaction state sequence:
1116  *    in_commit = 0, blocked = 0  (initial)
1117  *    in_commit = 1, blocked = 1
1118  *    blocked = 0
1119  *    commit_done = 1
1120  */
1121 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1122                              struct btrfs_root *root)
1123 {
1124         unsigned long joined = 0;
1125         struct btrfs_transaction *cur_trans;
1126         struct btrfs_transaction *prev_trans = NULL;
1127         DEFINE_WAIT(wait);
1128         int ret;
1129         int should_grow = 0;
1130         unsigned long now = get_seconds();
1131         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1132
1133         btrfs_run_ordered_operations(root, 0);
1134
1135         btrfs_trans_release_metadata(trans, root);
1136         trans->block_rsv = NULL;
1137
1138         /* make a pass through all the delayed refs we have so far
1139          * any runnings procs may add more while we are here
1140          */
1141         ret = btrfs_run_delayed_refs(trans, root, 0);
1142         BUG_ON(ret);
1143
1144         cur_trans = trans->transaction;
1145         /*
1146          * set the flushing flag so procs in this transaction have to
1147          * start sending their work down.
1148          */
1149         cur_trans->delayed_refs.flushing = 1;
1150
1151         ret = btrfs_run_delayed_refs(trans, root, 0);
1152         BUG_ON(ret);
1153
1154         spin_lock(&cur_trans->commit_lock);
1155         if (cur_trans->in_commit) {
1156                 spin_unlock(&cur_trans->commit_lock);
1157                 atomic_inc(&cur_trans->use_count);
1158                 btrfs_end_transaction(trans, root);
1159
1160                 wait_for_commit(root, cur_trans);
1161
1162                 put_transaction(cur_trans);
1163
1164                 return 0;
1165         }
1166
1167         trans->transaction->in_commit = 1;
1168         trans->transaction->blocked = 1;
1169         spin_unlock(&cur_trans->commit_lock);
1170         wake_up(&root->fs_info->transaction_blocked_wait);
1171
1172         spin_lock(&root->fs_info->trans_lock);
1173         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1174                 prev_trans = list_entry(cur_trans->list.prev,
1175                                         struct btrfs_transaction, list);
1176                 if (!prev_trans->commit_done) {
1177                         atomic_inc(&prev_trans->use_count);
1178                         spin_unlock(&root->fs_info->trans_lock);
1179
1180                         wait_for_commit(root, prev_trans);
1181
1182                         put_transaction(prev_trans);
1183                 } else {
1184                         spin_unlock(&root->fs_info->trans_lock);
1185                 }
1186         } else {
1187                 spin_unlock(&root->fs_info->trans_lock);
1188         }
1189
1190         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1191                 should_grow = 1;
1192
1193         do {
1194                 int snap_pending = 0;
1195
1196                 joined = cur_trans->num_joined;
1197                 if (!list_empty(&trans->transaction->pending_snapshots))
1198                         snap_pending = 1;
1199
1200                 WARN_ON(cur_trans != trans->transaction);
1201
1202                 if (flush_on_commit || snap_pending) {
1203                         btrfs_start_delalloc_inodes(root, 1);
1204                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1205                         BUG_ON(ret);
1206                 }
1207
1208                 ret = btrfs_run_delayed_items(trans, root);
1209                 BUG_ON(ret);
1210
1211                 /*
1212                  * rename don't use btrfs_join_transaction, so, once we
1213                  * set the transaction to blocked above, we aren't going
1214                  * to get any new ordered operations.  We can safely run
1215                  * it here and no for sure that nothing new will be added
1216                  * to the list
1217                  */
1218                 btrfs_run_ordered_operations(root, 1);
1219
1220                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1221                                 TASK_UNINTERRUPTIBLE);
1222
1223                 if (atomic_read(&cur_trans->num_writers) > 1)
1224                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1225                 else if (should_grow)
1226                         schedule_timeout(1);
1227
1228                 finish_wait(&cur_trans->writer_wait, &wait);
1229         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1230                  (should_grow && cur_trans->num_joined != joined));
1231
1232         /*
1233          * Ok now we need to make sure to block out any other joins while we
1234          * commit the transaction.  We could have started a join before setting
1235          * no_join so make sure to wait for num_writers to == 1 again.
1236          */
1237         spin_lock(&root->fs_info->trans_lock);
1238         root->fs_info->trans_no_join = 1;
1239         spin_unlock(&root->fs_info->trans_lock);
1240         wait_event(cur_trans->writer_wait,
1241                    atomic_read(&cur_trans->num_writers) == 1);
1242
1243         /*
1244          * the reloc mutex makes sure that we stop
1245          * the balancing code from coming in and moving
1246          * extents around in the middle of the commit
1247          */
1248         mutex_lock(&root->fs_info->reloc_mutex);
1249
1250         ret = btrfs_run_delayed_items(trans, root);
1251         BUG_ON(ret);
1252
1253         ret = create_pending_snapshots(trans, root->fs_info);
1254         BUG_ON(ret);
1255
1256         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1257         BUG_ON(ret);
1258
1259         /*
1260          * make sure none of the code above managed to slip in a
1261          * delayed item
1262          */
1263         btrfs_assert_delayed_root_empty(root);
1264
1265         WARN_ON(cur_trans != trans->transaction);
1266
1267         btrfs_scrub_pause(root);
1268         /* btrfs_commit_tree_roots is responsible for getting the
1269          * various roots consistent with each other.  Every pointer
1270          * in the tree of tree roots has to point to the most up to date
1271          * root for every subvolume and other tree.  So, we have to keep
1272          * the tree logging code from jumping in and changing any
1273          * of the trees.
1274          *
1275          * At this point in the commit, there can't be any tree-log
1276          * writers, but a little lower down we drop the trans mutex
1277          * and let new people in.  By holding the tree_log_mutex
1278          * from now until after the super is written, we avoid races
1279          * with the tree-log code.
1280          */
1281         mutex_lock(&root->fs_info->tree_log_mutex);
1282
1283         ret = commit_fs_roots(trans, root);
1284         BUG_ON(ret);
1285
1286         /* commit_fs_roots gets rid of all the tree log roots, it is now
1287          * safe to free the root of tree log roots
1288          */
1289         btrfs_free_log_root_tree(trans, root->fs_info);
1290
1291         ret = commit_cowonly_roots(trans, root);
1292         BUG_ON(ret);
1293
1294         btrfs_prepare_extent_commit(trans, root);
1295
1296         cur_trans = root->fs_info->running_transaction;
1297
1298         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1299                             root->fs_info->tree_root->node);
1300         switch_commit_root(root->fs_info->tree_root);
1301
1302         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1303                             root->fs_info->chunk_root->node);
1304         switch_commit_root(root->fs_info->chunk_root);
1305
1306         update_super_roots(root);
1307
1308         if (!root->fs_info->log_root_recovering) {
1309                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1310                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1311         }
1312
1313         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1314                sizeof(*root->fs_info->super_copy));
1315
1316         trans->transaction->blocked = 0;
1317         spin_lock(&root->fs_info->trans_lock);
1318         root->fs_info->running_transaction = NULL;
1319         root->fs_info->trans_no_join = 0;
1320         spin_unlock(&root->fs_info->trans_lock);
1321         mutex_unlock(&root->fs_info->reloc_mutex);
1322
1323         wake_up(&root->fs_info->transaction_wait);
1324
1325         ret = btrfs_write_and_wait_transaction(trans, root);
1326         BUG_ON(ret);
1327         write_ctree_super(trans, root, 0);
1328
1329         /*
1330          * the super is written, we can safely allow the tree-loggers
1331          * to go about their business
1332          */
1333         mutex_unlock(&root->fs_info->tree_log_mutex);
1334
1335         btrfs_finish_extent_commit(trans, root);
1336
1337         cur_trans->commit_done = 1;
1338
1339         root->fs_info->last_trans_committed = cur_trans->transid;
1340
1341         wake_up(&cur_trans->commit_wait);
1342
1343         spin_lock(&root->fs_info->trans_lock);
1344         list_del_init(&cur_trans->list);
1345         spin_unlock(&root->fs_info->trans_lock);
1346
1347         put_transaction(cur_trans);
1348         put_transaction(cur_trans);
1349
1350         trace_btrfs_transaction_commit(root);
1351
1352         btrfs_scrub_continue(root);
1353
1354         if (current->journal_info == trans)
1355                 current->journal_info = NULL;
1356
1357         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1358
1359         if (current != root->fs_info->transaction_kthread)
1360                 btrfs_run_delayed_iputs(root);
1361
1362         return ret;
1363 }
1364
1365 /*
1366  * interface function to delete all the snapshots we have scheduled for deletion
1367  */
1368 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1369 {
1370         LIST_HEAD(list);
1371         struct btrfs_fs_info *fs_info = root->fs_info;
1372
1373         spin_lock(&fs_info->trans_lock);
1374         list_splice_init(&fs_info->dead_roots, &list);
1375         spin_unlock(&fs_info->trans_lock);
1376
1377         while (!list_empty(&list)) {
1378                 root = list_entry(list.next, struct btrfs_root, root_list);
1379                 list_del(&root->root_list);
1380
1381                 btrfs_kill_all_delayed_nodes(root);
1382
1383                 if (btrfs_header_backref_rev(root->node) <
1384                     BTRFS_MIXED_BACKREF_REV)
1385                         btrfs_drop_snapshot(root, NULL, 0);
1386                 else
1387                         btrfs_drop_snapshot(root, NULL, 1);
1388         }
1389         return 0;
1390 }