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                 WARN_ON(transaction->delayed_refs.root.rb_node);
  40                 WARN_ON(!list_empty(&transaction->delayed_refs.seq_head));
  41                 memset(transaction, 0, sizeof(*transaction));
  42                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  43         }
  44 }
  45
  46 static noinline void switch_commit_root(struct btrfs_root *root)
  47 {
  48         free_extent_buffer(root->commit_root);
  49         root->commit_root = btrfs_root_node(root);
  50 }
  51
  52 /*
  53  * either allocate a new transaction or hop into the existing one
  54  */
  55 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  56 {
  57         struct btrfs_transaction *cur_trans;
  58
  59         spin_lock(&root->fs_info->trans_lock);
  60 loop:
  61         if (root->fs_info->trans_no_join) {
  62                 if (!nofail) {
  63                         spin_unlock(&root->fs_info->trans_lock);
  64                         return -EBUSY;
  65                 }
  66         }
  67
  68         cur_trans = root->fs_info->running_transaction;
  69         if (cur_trans) {
  70                 atomic_inc(&cur_trans->use_count);
  71                 atomic_inc(&cur_trans->num_writers);
  72                 cur_trans->num_joined++;
  73                 spin_unlock(&root->fs_info->trans_lock);
  74                 return 0;
  75         }
  76         spin_unlock(&root->fs_info->trans_lock);
  77
  78         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  79         if (!cur_trans)
  80                 return -ENOMEM;
  81
  82         spin_lock(&root->fs_info->trans_lock);
  83         if (root->fs_info->running_transaction) {
  84                 /*
  85                  * someone started a transaction after we unlocked.  Make sure
  86                  * to redo the trans_no_join checks above
  87                  */
  88                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  89                 cur_trans = root->fs_info->running_transaction;
  90                 goto loop;
  91         }
  92
  93         atomic_set(&cur_trans->num_writers, 1);
  94         cur_trans->num_joined = 0;
  95         init_waitqueue_head(&cur_trans->writer_wait);
  96         init_waitqueue_head(&cur_trans->commit_wait);
  97         cur_trans->in_commit = 0;
  98         cur_trans->blocked = 0;
  99         /*
 100          * One for this trans handle, one so it will live on until we
 101          * commit the transaction.
 102          */
 103         atomic_set(&cur_trans->use_count, 2);
 104         cur_trans->commit_done = 0;
 105         cur_trans->start_time = get_seconds();
 106
 107         cur_trans->delayed_refs.root = RB_ROOT;
 108         cur_trans->delayed_refs.num_entries = 0;
 109         cur_trans->delayed_refs.num_heads_ready = 0;
 110         cur_trans->delayed_refs.num_heads = 0;
 111         cur_trans->delayed_refs.flushing = 0;
 112         cur_trans->delayed_refs.run_delayed_start = 0;
 113         cur_trans->delayed_refs.seq = 1;
 114         init_waitqueue_head(&cur_trans->delayed_refs.seq_wait);
 115         spin_lock_init(&cur_trans->commit_lock);
 116         spin_lock_init(&cur_trans->delayed_refs.lock);
 117         INIT_LIST_HEAD(&cur_trans->delayed_refs.seq_head);
 118
 119         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 120         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 121         extent_io_tree_init(&cur_trans->dirty_pages,
 122                              root->fs_info->btree_inode->i_mapping);
 123         root->fs_info->generation++;
 124         cur_trans->transid = root->fs_info->generation;
 125         root->fs_info->running_transaction = cur_trans;
 126         spin_unlock(&root->fs_info->trans_lock);
 127
 128         return 0;
 129 }
 130
 131 /*
 132  * this does all the record keeping required to make sure that a reference
 133  * counted root is properly recorded in a given transaction.  This is required
 134  * to make sure the old root from before we joined the transaction is deleted
 135  * when the transaction commits
 136  */
 137 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 138                                struct btrfs_root *root)
 139 {
 140         if (root->ref_cows && root->last_trans < trans->transid) {
 141                 WARN_ON(root == root->fs_info->extent_root);
 142                 WARN_ON(root->commit_root != root->node);
 143
 144                 /*
 145                  * see below for in_trans_setup usage rules
 146                  * we have the reloc mutex held now, so there
 147                  * is only one writer in this function
 148                  */
 149                 root->in_trans_setup = 1;
 150
 151                 /* make sure readers find in_trans_setup before
 152                  * they find our root->last_trans update
 153                  */
 154                 smp_wmb();
 155
 156                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 157                 if (root->last_trans == trans->transid) {
 158                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 159                         return 0;
 160                 }
 161                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 162                            (unsigned long)root->root_key.objectid,
 163                            BTRFS_ROOT_TRANS_TAG);
 164                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 165                 root->last_trans = trans->transid;
 166
 167                 /* this is pretty tricky.  We don't want to
 168                  * take the relocation lock in btrfs_record_root_in_trans
 169                  * unless we're really doing the first setup for this root in
 170                  * this transaction.
 171                  *
 172                  * Normally we'd use root->last_trans as a flag to decide
 173                  * if we want to take the expensive mutex.
 174                  *
 175                  * But, we have to set root->last_trans before we
 176                  * init the relocation root, otherwise, we trip over warnings
 177                  * in ctree.c.  The solution used here is to flag ourselves
 178                  * with root->in_trans_setup.  When this is 1, we're still
 179                  * fixing up the reloc trees and everyone must wait.
 180                  *
 181                  * When this is zero, they can trust root->last_trans and fly
 182                  * through btrfs_record_root_in_trans without having to take the
 183                  * lock.  smp_wmb() makes sure that all the writes above are
 184                  * done before we pop in the zero below
 185                  */
 186                 btrfs_init_reloc_root(trans, root);
 187                 smp_wmb();
 188                 root->in_trans_setup = 0;
 189         }
 190         return 0;
 191 }
 192
 193
 194 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 195                                struct btrfs_root *root)
 196 {
 197         if (!root->ref_cows)
 198                 return 0;
 199
 200         /*
 201          * see record_root_in_trans for comments about in_trans_setup usage
 202          * and barriers
 203          */
 204         smp_rmb();
 205         if (root->last_trans == trans->transid &&
 206             !root->in_trans_setup)
 207                 return 0;
 208
 209         mutex_lock(&root->fs_info->reloc_mutex);
 210         record_root_in_trans(trans, root);
 211         mutex_unlock(&root->fs_info->reloc_mutex);
 212
 213         return 0;
 214 }
 215
 216 /* wait for commit against the current transaction to become unblocked
 217  * when this is done, it is safe to start a new transaction, but the current
 218  * transaction might not be fully on disk.
 219  */
 220 static void wait_current_trans(struct btrfs_root *root)
 221 {
 222         struct btrfs_transaction *cur_trans;
 223
 224         spin_lock(&root->fs_info->trans_lock);
 225         cur_trans = root->fs_info->running_transaction;
 226         if (cur_trans && cur_trans->blocked) {
 227                 atomic_inc(&cur_trans->use_count);
 228                 spin_unlock(&root->fs_info->trans_lock);
 229
 230                 wait_event(root->fs_info->transaction_wait,
 231                            !cur_trans->blocked);
 232                 put_transaction(cur_trans);
 233         } else {
 234                 spin_unlock(&root->fs_info->trans_lock);
 235         }
 236 }
 237
 238 enum btrfs_trans_type {
 239         TRANS_START,
 240         TRANS_JOIN,
 241         TRANS_USERSPACE,
 242         TRANS_JOIN_NOLOCK,
 243 };
 244
 245 static int may_wait_transaction(struct btrfs_root *root, int type)
 246 {
 247         if (root->fs_info->log_root_recovering)
 248                 return 0;
 249
 250         if (type == TRANS_USERSPACE)
 251                 return 1;
 252
 253         if (type == TRANS_START &&
 254             !atomic_read(&root->fs_info->open_ioctl_trans))
 255                 return 1;
 256
 257         return 0;
 258 }
 259
 260 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 261                                                     u64 num_items, int type)
 262 {
 263         struct btrfs_trans_handle *h;
 264         struct btrfs_transaction *cur_trans;
 265         u64 num_bytes = 0;
 266         int ret;
 267
 268         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 269                 return ERR_PTR(-EROFS);
 270
 271         if (current->journal_info) {
 272                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 273                 h = current->journal_info;
 274                 h->use_count++;
 275                 h->orig_rsv = h->block_rsv;
 276                 h->block_rsv = NULL;
 277                 goto got_it;
 278         }
 279
 280         /*
 281          * Do the reservation before we join the transaction so we can do all
 282          * the appropriate flushing if need be.
 283          */
 284         if (num_items > 0 && root != root->fs_info->chunk_root) {
 285                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 286                 ret = btrfs_block_rsv_add(root,
 287                                           &root->fs_info->trans_block_rsv,
 288                                           num_bytes);
 289                 if (ret)
 290                         return ERR_PTR(ret);
 291         }
 292 again:
 293         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 294         if (!h)
 295                 return ERR_PTR(-ENOMEM);
 296
 297         if (may_wait_transaction(root, type))
 298                 wait_current_trans(root);
 299
 300         do {
 301                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 302                 if (ret == -EBUSY)
 303                         wait_current_trans(root);
 304         } while (ret == -EBUSY);
 305
 306         if (ret < 0) {
 307                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 308                 return ERR_PTR(ret);
 309         }
 310
 311         cur_trans = root->fs_info->running_transaction;
 312
 313         h->transid = cur_trans->transid;
 314         h->transaction = cur_trans;
 315         h->blocks_used = 0;
 316         h->bytes_reserved = 0;
 317         h->delayed_ref_updates = 0;
 318         h->use_count = 1;
 319         h->block_rsv = NULL;
 320         h->orig_rsv = NULL;
 321
 322         smp_mb();
 323         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 324                 btrfs_commit_transaction(h, root);
 325                 goto again;
 326         }
 327
 328         if (num_bytes) {
 329                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 330                                               (u64)h, num_bytes, 1);
 331                 h->block_rsv = &root->fs_info->trans_block_rsv;
 332                 h->bytes_reserved = num_bytes;
 333         }
 334
 335 got_it:
 336         btrfs_record_root_in_trans(h, root);
 337
 338         if (!current->journal_info && type != TRANS_USERSPACE)
 339                 current->journal_info = h;
 340         return h;
 341 }
 342
 343 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 344                                                    int num_items)
 345 {
 346         return start_transaction(root, num_items, TRANS_START);
 347 }
 348 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 349 {
 350         return start_transaction(root, 0, TRANS_JOIN);
 351 }
 352
 353 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 354 {
 355         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 356 }
 357
 358 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 359 {
 360         return start_transaction(root, 0, TRANS_USERSPACE);
 361 }
 362
 363 /* wait for a transaction commit to be fully complete */
 364 static noinline void wait_for_commit(struct btrfs_root *root,
 365                                     struct btrfs_transaction *commit)
 366 {
 367         wait_event(commit->commit_wait, commit->commit_done);
 368 }
 369
 370 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 371 {
 372         struct btrfs_transaction *cur_trans = NULL, *t;
 373         int ret;
 374
 375         ret = 0;
 376         if (transid) {
 377                 if (transid <= root->fs_info->last_trans_committed)
 378                         goto out;
 379
 380                 /* find specified transaction */
 381                 spin_lock(&root->fs_info->trans_lock);
 382                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 383                         if (t->transid == transid) {
 384                                 cur_trans = t;
 385                                 atomic_inc(&cur_trans->use_count);
 386                                 break;
 387                         }
 388                         if (t->transid > transid)
 389                                 break;
 390                 }
 391                 spin_unlock(&root->fs_info->trans_lock);
 392                 ret = -EINVAL;
 393                 if (!cur_trans)
 394                         goto out;  /* bad transid */
 395         } else {
 396                 /* find newest transaction that is committing | committed */
 397                 spin_lock(&root->fs_info->trans_lock);
 398                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 399                                             list) {
 400                         if (t->in_commit) {
 401                                 if (t->commit_done)
 402                                         break;
 403                                 cur_trans = t;
 404                                 atomic_inc(&cur_trans->use_count);
 405                                 break;
 406                         }
 407                 }
 408                 spin_unlock(&root->fs_info->trans_lock);
 409                 if (!cur_trans)
 410                         goto out;  /* nothing committing|committed */
 411         }
 412
 413         wait_for_commit(root, cur_trans);
 414
 415         put_transaction(cur_trans);
 416         ret = 0;
 417 out:
 418         return ret;
 419 }
 420
 421 void btrfs_throttle(struct btrfs_root *root)
 422 {
 423         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 424                 wait_current_trans(root);
 425 }
 426
 427 static int should_end_transaction(struct btrfs_trans_handle *trans,
 428                                   struct btrfs_root *root)
 429 {
 430         int ret;
 431
 432         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 433         return ret ? 1 : 0;
 434 }
 435
 436 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 437                                  struct btrfs_root *root)
 438 {
 439         struct btrfs_transaction *cur_trans = trans->transaction;
 440         struct btrfs_block_rsv *rsv = trans->block_rsv;
 441         int updates;
 442
 443         smp_mb();
 444         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 445                 return 1;
 446
 447         /*
 448          * We need to do this in case we're deleting csums so the global block
 449          * rsv get's used instead of the csum block rsv.
 450          */
 451         trans->block_rsv = NULL;
 452
 453         updates = trans->delayed_ref_updates;
 454         trans->delayed_ref_updates = 0;
 455         if (updates)
 456                 btrfs_run_delayed_refs(trans, root, updates);
 457
 458         trans->block_rsv = rsv;
 459
 460         return should_end_transaction(trans, root);
 461 }
 462
 463 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 464                           struct btrfs_root *root, int throttle, int lock)
 465 {
 466         struct btrfs_transaction *cur_trans = trans->transaction;
 467         struct btrfs_fs_info *info = root->fs_info;
 468         int count = 0;
 469
 470         if (--trans->use_count) {
 471                 trans->block_rsv = trans->orig_rsv;
 472                 return 0;
 473         }
 474
 475         btrfs_trans_release_metadata(trans, root);
 476         trans->block_rsv = NULL;
 477         while (count < 2) {
 478                 unsigned long cur = trans->delayed_ref_updates;
 479                 trans->delayed_ref_updates = 0;
 480                 if (cur &&
 481                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 482                         trans->delayed_ref_updates = 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                         /* see comments in should_cow_block() */
 789                         root->force_cow = 0;
 790                         smp_wmb();
 791
 792                         if (root->commit_root != root->node) {
 793                                 mutex_lock(&root->fs_commit_mutex);
 794                                 switch_commit_root(root);
 795                                 btrfs_unpin_free_ino(root);
 796                                 mutex_unlock(&root->fs_commit_mutex);
 797
 798                                 btrfs_set_root_node(&root->root_item,
 799                                                     root->node);
 800                         }
 801
 802                         err = btrfs_update_root(trans, fs_info->tree_root,
 803                                                 &root->root_key,
 804                                                 &root->root_item);
 805                         spin_lock(&fs_info->fs_roots_radix_lock);
 806                         if (err)
 807                                 break;
 808                 }
 809         }
 810         spin_unlock(&fs_info->fs_roots_radix_lock);
 811         return err;
 812 }
 813
 814 /*
 815  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 816  * otherwise every leaf in the btree is read and defragged.
 817  */
 818 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 819 {
 820         struct btrfs_fs_info *info = root->fs_info;
 821         struct btrfs_trans_handle *trans;
 822         int ret;
 823         unsigned long nr;
 824
 825         if (xchg(&root->defrag_running, 1))
 826                 return 0;
 827
 828         while (1) {
 829                 trans = btrfs_start_transaction(root, 0);
 830                 if (IS_ERR(trans))
 831                         return PTR_ERR(trans);
 832
 833                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 834
 835                 nr = trans->blocks_used;
 836                 btrfs_end_transaction(trans, root);
 837                 btrfs_btree_balance_dirty(info->tree_root, nr);
 838                 cond_resched();
 839
 840                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 841                         break;
 842         }
 843         root->defrag_running = 0;
 844         return ret;
 845 }
 846
 847 /*
 848  * new snapshots need to be created at a very specific time in the
 849  * transaction commit.  This does the actual creation
 850  */
 851 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 852                                    struct btrfs_fs_info *fs_info,
 853                                    struct btrfs_pending_snapshot *pending)
 854 {
 855         struct btrfs_key key;
 856         struct btrfs_root_item *new_root_item;
 857         struct btrfs_root *tree_root = fs_info->tree_root;
 858         struct btrfs_root *root = pending->root;
 859         struct btrfs_root *parent_root;
 860         struct btrfs_block_rsv *rsv;
 861         struct inode *parent_inode;
 862         struct dentry *parent;
 863         struct dentry *dentry;
 864         struct extent_buffer *tmp;
 865         struct extent_buffer *old;
 866         int ret;
 867         u64 to_reserve = 0;
 868         u64 index = 0;
 869         u64 objectid;
 870         u64 root_flags;
 871
 872         rsv = trans->block_rsv;
 873
 874         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 875         if (!new_root_item) {
 876                 pending->error = -ENOMEM;
 877                 goto fail;
 878         }
 879
 880         ret = btrfs_find_free_objectid(tree_root, &objectid);
 881         if (ret) {
 882                 pending->error = ret;
 883                 goto fail;
 884         }
 885
 886         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 887
 888         if (to_reserve > 0) {
 889                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 890                                                   to_reserve);
 891                 if (ret) {
 892                         pending->error = ret;
 893                         goto fail;
 894                 }
 895         }
 896
 897         key.objectid = objectid;
 898         key.offset = (u64)-1;
 899         key.type = BTRFS_ROOT_ITEM_KEY;
 900
 901         trans->block_rsv = &pending->block_rsv;
 902
 903         dentry = pending->dentry;
 904         parent = dget_parent(dentry);
 905         parent_inode = parent->d_inode;
 906         parent_root = BTRFS_I(parent_inode)->root;
 907         record_root_in_trans(trans, parent_root);
 908
 909         /*
 910          * insert the directory item
 911          */
 912         ret = btrfs_set_inode_index(parent_inode, &index);
 913         BUG_ON(ret);
 914         ret = btrfs_insert_dir_item(trans, parent_root,
 915                                 dentry->d_name.name, dentry->d_name.len,
 916                                 parent_inode, &key,
 917                                 BTRFS_FT_DIR, index);
 918         BUG_ON(ret);
 919
 920         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 921                                          dentry->d_name.len * 2);
 922         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 923         BUG_ON(ret);
 924
 925         /*
 926          * pull in the delayed directory update
 927          * and the delayed inode item
 928          * otherwise we corrupt the FS during
 929          * snapshot
 930          */
 931         ret = btrfs_run_delayed_items(trans, root);
 932         BUG_ON(ret);
 933
 934         record_root_in_trans(trans, root);
 935         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 936         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 937         btrfs_check_and_init_root_item(new_root_item);
 938
 939         root_flags = btrfs_root_flags(new_root_item);
 940         if (pending->readonly)
 941                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 942         else
 943                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 944         btrfs_set_root_flags(new_root_item, root_flags);
 945
 946         old = btrfs_lock_root_node(root);
 947         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 948         btrfs_set_lock_blocking(old);
 949
 950         btrfs_copy_root(trans, root, old, &tmp, objectid);
 951         btrfs_tree_unlock(old);
 952         free_extent_buffer(old);
 953
 954         /* see comments in should_cow_block() */
 955         root->force_cow = 1;
 956         smp_wmb();
 957
 958         btrfs_set_root_node(new_root_item, tmp);
 959         /* record when the snapshot was created in key.offset */
 960         key.offset = trans->transid;
 961         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 962         btrfs_tree_unlock(tmp);
 963         free_extent_buffer(tmp);
 964         BUG_ON(ret);
 965
 966         /*
 967          * insert root back/forward references
 968          */
 969         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 970                                  parent_root->root_key.objectid,
 971                                  btrfs_ino(parent_inode), index,
 972                                  dentry->d_name.name, dentry->d_name.len);
 973         BUG_ON(ret);
 974         dput(parent);
 975
 976         key.offset = (u64)-1;
 977         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 978         BUG_ON(IS_ERR(pending->snap));
 979
 980         btrfs_reloc_post_snapshot(trans, pending);
 981 fail:
 982         kfree(new_root_item);
 983         trans->block_rsv = rsv;
 984         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 985         return 0;
 986 }
 987
 988 /*
 989  * create all the snapshots we've scheduled for creation
 990  */
 991 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 992                                              struct btrfs_fs_info *fs_info)
 993 {
 994         struct btrfs_pending_snapshot *pending;
 995         struct list_head *head = &trans->transaction->pending_snapshots;
 996         int ret;
 997
 998         list_for_each_entry(pending, head, list) {
 999                 ret = create_pending_snapshot(trans, fs_info, pending);
1000                 BUG_ON(ret);
1001         }
1002         return 0;
1003 }
1004
1005 static void update_super_roots(struct btrfs_root *root)
1006 {
1007         struct btrfs_root_item *root_item;
1008         struct btrfs_super_block *super;
1009
1010         super = root->fs_info->super_copy;
1011
1012         root_item = &root->fs_info->chunk_root->root_item;
1013         super->chunk_root = root_item->bytenr;
1014         super->chunk_root_generation = root_item->generation;
1015         super->chunk_root_level = root_item->level;
1016
1017         root_item = &root->fs_info->tree_root->root_item;
1018         super->root = root_item->bytenr;
1019         super->generation = root_item->generation;
1020         super->root_level = root_item->level;
1021         if (btrfs_test_opt(root, SPACE_CACHE))
1022                 super->cache_generation = root_item->generation;
1023 }
1024
1025 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1026 {
1027         int ret = 0;
1028         spin_lock(&info->trans_lock);
1029         if (info->running_transaction)
1030                 ret = info->running_transaction->in_commit;
1031         spin_unlock(&info->trans_lock);
1032         return ret;
1033 }
1034
1035 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1036 {
1037         int ret = 0;
1038         spin_lock(&info->trans_lock);
1039         if (info->running_transaction)
1040                 ret = info->running_transaction->blocked;
1041         spin_unlock(&info->trans_lock);
1042         return ret;
1043 }
1044
1045 /*
1046  * wait for the current transaction commit to start and block subsequent
1047  * transaction joins
1048  */
1049 static void wait_current_trans_commit_start(struct btrfs_root *root,
1050                                             struct btrfs_transaction *trans)
1051 {
1052         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1053 }
1054
1055 /*
1056  * wait for the current transaction to start and then become unblocked.
1057  * caller holds ref.
1058  */
1059 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1060                                          struct btrfs_transaction *trans)
1061 {
1062         wait_event(root->fs_info->transaction_wait,
1063                    trans->commit_done || (trans->in_commit && !trans->blocked));
1064 }
1065
1066 /*
1067  * commit transactions asynchronously. once btrfs_commit_transaction_async
1068  * returns, any subsequent transaction will not be allowed to join.
1069  */
1070 struct btrfs_async_commit {
1071         struct btrfs_trans_handle *newtrans;
1072         struct btrfs_root *root;
1073         struct delayed_work work;
1074 };
1075
1076 static void do_async_commit(struct work_struct *work)
1077 {
1078         struct btrfs_async_commit *ac =
1079                 container_of(work, struct btrfs_async_commit, work.work);
1080
1081         btrfs_commit_transaction(ac->newtrans, ac->root);
1082         kfree(ac);
1083 }
1084
1085 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1086                                    struct btrfs_root *root,
1087                                    int wait_for_unblock)
1088 {
1089         struct btrfs_async_commit *ac;
1090         struct btrfs_transaction *cur_trans;
1091
1092         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1093         if (!ac)
1094                 return -ENOMEM;
1095
1096         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1097         ac->root = root;
1098         ac->newtrans = btrfs_join_transaction(root);
1099         if (IS_ERR(ac->newtrans)) {
1100                 int err = PTR_ERR(ac->newtrans);
1101                 kfree(ac);
1102                 return err;
1103         }
1104
1105         /* take transaction reference */
1106         cur_trans = trans->transaction;
1107         atomic_inc(&cur_trans->use_count);
1108
1109         btrfs_end_transaction(trans, root);
1110         schedule_delayed_work(&ac->work, 0);
1111
1112         /* wait for transaction to start and unblock */
1113         if (wait_for_unblock)
1114                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1115         else
1116                 wait_current_trans_commit_start(root, cur_trans);
1117
1118         if (current->journal_info == trans)
1119                 current->journal_info = NULL;
1120
1121         put_transaction(cur_trans);
1122         return 0;
1123 }
1124
1125 /*
1126  * btrfs_transaction state sequence:
1127  *    in_commit = 0, blocked = 0  (initial)
1128  *    in_commit = 1, blocked = 1
1129  *    blocked = 0
1130  *    commit_done = 1
1131  */
1132 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1133                              struct btrfs_root *root)
1134 {
1135         unsigned long joined = 0;
1136         struct btrfs_transaction *cur_trans;
1137         struct btrfs_transaction *prev_trans = NULL;
1138         DEFINE_WAIT(wait);
1139         int ret;
1140         int should_grow = 0;
1141         unsigned long now = get_seconds();
1142         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1143
1144         btrfs_run_ordered_operations(root, 0);
1145
1146         btrfs_trans_release_metadata(trans, root);
1147         trans->block_rsv = NULL;
1148
1149         /* make a pass through all the delayed refs we have so far
1150          * any runnings procs may add more while we are here
1151          */
1152         ret = btrfs_run_delayed_refs(trans, root, 0);
1153         BUG_ON(ret);
1154
1155         cur_trans = trans->transaction;
1156         /*
1157          * set the flushing flag so procs in this transaction have to
1158          * start sending their work down.
1159          */
1160         cur_trans->delayed_refs.flushing = 1;
1161
1162         ret = btrfs_run_delayed_refs(trans, root, 0);
1163         BUG_ON(ret);
1164
1165         spin_lock(&cur_trans->commit_lock);
1166         if (cur_trans->in_commit) {
1167                 spin_unlock(&cur_trans->commit_lock);
1168                 atomic_inc(&cur_trans->use_count);
1169                 btrfs_end_transaction(trans, root);
1170
1171                 wait_for_commit(root, cur_trans);
1172
1173                 put_transaction(cur_trans);
1174
1175                 return 0;
1176         }
1177
1178         trans->transaction->in_commit = 1;
1179         trans->transaction->blocked = 1;
1180         spin_unlock(&cur_trans->commit_lock);
1181         wake_up(&root->fs_info->transaction_blocked_wait);
1182
1183         spin_lock(&root->fs_info->trans_lock);
1184         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1185                 prev_trans = list_entry(cur_trans->list.prev,
1186                                         struct btrfs_transaction, list);
1187                 if (!prev_trans->commit_done) {
1188                         atomic_inc(&prev_trans->use_count);
1189                         spin_unlock(&root->fs_info->trans_lock);
1190
1191                         wait_for_commit(root, prev_trans);
1192
1193                         put_transaction(prev_trans);
1194                 } else {
1195                         spin_unlock(&root->fs_info->trans_lock);
1196                 }
1197         } else {
1198                 spin_unlock(&root->fs_info->trans_lock);
1199         }
1200
1201         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1202                 should_grow = 1;
1203
1204         do {
1205                 int snap_pending = 0;
1206
1207                 joined = cur_trans->num_joined;
1208                 if (!list_empty(&trans->transaction->pending_snapshots))
1209                         snap_pending = 1;
1210
1211                 WARN_ON(cur_trans != trans->transaction);
1212
1213                 if (flush_on_commit || snap_pending) {
1214                         btrfs_start_delalloc_inodes(root, 1);
1215                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1216                         BUG_ON(ret);
1217                 }
1218
1219                 ret = btrfs_run_delayed_items(trans, root);
1220                 BUG_ON(ret);
1221
1222                 /*
1223                  * rename don't use btrfs_join_transaction, so, once we
1224                  * set the transaction to blocked above, we aren't going
1225                  * to get any new ordered operations.  We can safely run
1226                  * it here and no for sure that nothing new will be added
1227                  * to the list
1228                  */
1229                 btrfs_run_ordered_operations(root, 1);
1230
1231                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1232                                 TASK_UNINTERRUPTIBLE);
1233
1234                 if (atomic_read(&cur_trans->num_writers) > 1)
1235                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1236                 else if (should_grow)
1237                         schedule_timeout(1);
1238
1239                 finish_wait(&cur_trans->writer_wait, &wait);
1240         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1241                  (should_grow && cur_trans->num_joined != joined));
1242
1243         /*
1244          * Ok now we need to make sure to block out any other joins while we
1245          * commit the transaction.  We could have started a join before setting
1246          * no_join so make sure to wait for num_writers to == 1 again.
1247          */
1248         spin_lock(&root->fs_info->trans_lock);
1249         root->fs_info->trans_no_join = 1;
1250         spin_unlock(&root->fs_info->trans_lock);
1251         wait_event(cur_trans->writer_wait,
1252                    atomic_read(&cur_trans->num_writers) == 1);
1253
1254         /*
1255          * the reloc mutex makes sure that we stop
1256          * the balancing code from coming in and moving
1257          * extents around in the middle of the commit
1258          */
1259         mutex_lock(&root->fs_info->reloc_mutex);
1260
1261         ret = btrfs_run_delayed_items(trans, root);
1262         BUG_ON(ret);
1263
1264         ret = create_pending_snapshots(trans, root->fs_info);
1265         BUG_ON(ret);
1266
1267         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1268         BUG_ON(ret);
1269
1270         /*
1271          * make sure none of the code above managed to slip in a
1272          * delayed item
1273          */
1274         btrfs_assert_delayed_root_empty(root);
1275
1276         WARN_ON(cur_trans != trans->transaction);
1277
1278         btrfs_scrub_pause(root);
1279         /* btrfs_commit_tree_roots is responsible for getting the
1280          * various roots consistent with each other.  Every pointer
1281          * in the tree of tree roots has to point to the most up to date
1282          * root for every subvolume and other tree.  So, we have to keep
1283          * the tree logging code from jumping in and changing any
1284          * of the trees.
1285          *
1286          * At this point in the commit, there can't be any tree-log
1287          * writers, but a little lower down we drop the trans mutex
1288          * and let new people in.  By holding the tree_log_mutex
1289          * from now until after the super is written, we avoid races
1290          * with the tree-log code.
1291          */
1292         mutex_lock(&root->fs_info->tree_log_mutex);
1293
1294         ret = commit_fs_roots(trans, root);
1295         BUG_ON(ret);
1296
1297         /* commit_fs_roots gets rid of all the tree log roots, it is now
1298          * safe to free the root of tree log roots
1299          */
1300         btrfs_free_log_root_tree(trans, root->fs_info);
1301
1302         ret = commit_cowonly_roots(trans, root);
1303         BUG_ON(ret);
1304
1305         btrfs_prepare_extent_commit(trans, root);
1306
1307         cur_trans = root->fs_info->running_transaction;
1308
1309         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1310                             root->fs_info->tree_root->node);
1311         switch_commit_root(root->fs_info->tree_root);
1312
1313         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1314                             root->fs_info->chunk_root->node);
1315         switch_commit_root(root->fs_info->chunk_root);
1316
1317         update_super_roots(root);
1318
1319         if (!root->fs_info->log_root_recovering) {
1320                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1321                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1322         }
1323
1324         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1325                sizeof(*root->fs_info->super_copy));
1326
1327         trans->transaction->blocked = 0;
1328         spin_lock(&root->fs_info->trans_lock);
1329         root->fs_info->running_transaction = NULL;
1330         root->fs_info->trans_no_join = 0;
1331         spin_unlock(&root->fs_info->trans_lock);
1332         mutex_unlock(&root->fs_info->reloc_mutex);
1333
1334         wake_up(&root->fs_info->transaction_wait);
1335
1336         ret = btrfs_write_and_wait_transaction(trans, root);
1337         BUG_ON(ret);
1338         write_ctree_super(trans, root, 0);
1339
1340         /*
1341          * the super is written, we can safely allow the tree-loggers
1342          * to go about their business
1343          */
1344         mutex_unlock(&root->fs_info->tree_log_mutex);
1345
1346         btrfs_finish_extent_commit(trans, root);
1347
1348         cur_trans->commit_done = 1;
1349
1350         root->fs_info->last_trans_committed = cur_trans->transid;
1351
1352         wake_up(&cur_trans->commit_wait);
1353
1354         spin_lock(&root->fs_info->trans_lock);
1355         list_del_init(&cur_trans->list);
1356         spin_unlock(&root->fs_info->trans_lock);
1357
1358         put_transaction(cur_trans);
1359         put_transaction(cur_trans);
1360
1361         trace_btrfs_transaction_commit(root);
1362
1363         btrfs_scrub_continue(root);
1364
1365         if (current->journal_info == trans)
1366                 current->journal_info = NULL;
1367
1368         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1369
1370         if (current != root->fs_info->transaction_kthread)
1371                 btrfs_run_delayed_iputs(root);
1372
1373         return ret;
1374 }
1375
1376 /*
1377  * interface function to delete all the snapshots we have scheduled for deletion
1378  */
1379 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1380 {
1381         LIST_HEAD(list);
1382         struct btrfs_fs_info *fs_info = root->fs_info;
1383
1384         spin_lock(&fs_info->trans_lock);
1385         list_splice_init(&fs_info->dead_roots, &list);
1386         spin_unlock(&fs_info->trans_lock);
1387
1388         while (!list_empty(&list)) {
1389                 root = list_entry(list.next, struct btrfs_root, root_list);
1390                 list_del(&root->root_list);
1391
1392                 btrfs_kill_all_delayed_nodes(root);
1393
1394                 if (btrfs_header_backref_rev(root->node) <
1395                     BTRFS_MIXED_BACKREF_REV)
1396                         btrfs_drop_snapshot(root, NULL, 0, 0);
1397                 else
1398                         btrfs_drop_snapshot(root, NULL, 1, 0);
1399         }
1400         return 0;
1401 }