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)(unsigned long)h,
 331                                               num_bytes, 1);
 332                 h->block_rsv = &root->fs_info->trans_block_rsv;
 333                 h->bytes_reserved = num_bytes;
 334         }
 335
 336 got_it:
 337         btrfs_record_root_in_trans(h, root);
 338
 339         if (!current->journal_info && type != TRANS_USERSPACE)
 340                 current->journal_info = h;
 341         return h;
 342 }
 343
 344 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 345                                                    int num_items)
 346 {
 347         return start_transaction(root, num_items, TRANS_START);
 348 }
 349 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 350 {
 351         return start_transaction(root, 0, TRANS_JOIN);
 352 }
 353
 354 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 355 {
 356         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 357 }
 358
 359 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 360 {
 361         return start_transaction(root, 0, TRANS_USERSPACE);
 362 }
 363
 364 /* wait for a transaction commit to be fully complete */
 365 static noinline void wait_for_commit(struct btrfs_root *root,
 366                                     struct btrfs_transaction *commit)
 367 {
 368         wait_event(commit->commit_wait, commit->commit_done);
 369 }
 370
 371 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 372 {
 373         struct btrfs_transaction *cur_trans = NULL, *t;
 374         int ret;
 375
 376         ret = 0;
 377         if (transid) {
 378                 if (transid <= root->fs_info->last_trans_committed)
 379                         goto out;
 380
 381                 /* find specified transaction */
 382                 spin_lock(&root->fs_info->trans_lock);
 383                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 384                         if (t->transid == transid) {
 385                                 cur_trans = t;
 386                                 atomic_inc(&cur_trans->use_count);
 387                                 break;
 388                         }
 389                         if (t->transid > transid)
 390                                 break;
 391                 }
 392                 spin_unlock(&root->fs_info->trans_lock);
 393                 ret = -EINVAL;
 394                 if (!cur_trans)
 395                         goto out;  /* bad transid */
 396         } else {
 397                 /* find newest transaction that is committing | committed */
 398                 spin_lock(&root->fs_info->trans_lock);
 399                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 400                                             list) {
 401                         if (t->in_commit) {
 402                                 if (t->commit_done)
 403                                         break;
 404                                 cur_trans = t;
 405                                 atomic_inc(&cur_trans->use_count);
 406                                 break;
 407                         }
 408                 }
 409                 spin_unlock(&root->fs_info->trans_lock);
 410                 if (!cur_trans)
 411                         goto out;  /* nothing committing|committed */
 412         }
 413
 414         wait_for_commit(root, cur_trans);
 415
 416         put_transaction(cur_trans);
 417         ret = 0;
 418 out:
 419         return ret;
 420 }
 421
 422 void btrfs_throttle(struct btrfs_root *root)
 423 {
 424         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 425                 wait_current_trans(root);
 426 }
 427
 428 static int should_end_transaction(struct btrfs_trans_handle *trans,
 429                                   struct btrfs_root *root)
 430 {
 431         int ret;
 432
 433         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 434         return ret ? 1 : 0;
 435 }
 436
 437 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 438                                  struct btrfs_root *root)
 439 {
 440         struct btrfs_transaction *cur_trans = trans->transaction;
 441         struct btrfs_block_rsv *rsv = trans->block_rsv;
 442         int updates;
 443
 444         smp_mb();
 445         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 446                 return 1;
 447
 448         /*
 449          * We need to do this in case we're deleting csums so the global block
 450          * rsv get's used instead of the csum block rsv.
 451          */
 452         trans->block_rsv = NULL;
 453
 454         updates = trans->delayed_ref_updates;
 455         trans->delayed_ref_updates = 0;
 456         if (updates)
 457                 btrfs_run_delayed_refs(trans, root, updates);
 458
 459         trans->block_rsv = rsv;
 460
 461         return should_end_transaction(trans, root);
 462 }
 463
 464 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 465                           struct btrfs_root *root, int throttle, int lock)
 466 {
 467         struct btrfs_transaction *cur_trans = trans->transaction;
 468         struct btrfs_fs_info *info = root->fs_info;
 469         int count = 0;
 470
 471         if (--trans->use_count) {
 472                 trans->block_rsv = trans->orig_rsv;
 473                 return 0;
 474         }
 475
 476         btrfs_trans_release_metadata(trans, root);
 477         trans->block_rsv = NULL;
 478         while (count < 2) {
 479                 unsigned long cur = trans->delayed_ref_updates;
 480                 trans->delayed_ref_updates = 0;
 481                 if (cur &&
 482                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 483                         trans->delayed_ref_updates = 0;
 484                         btrfs_run_delayed_refs(trans, root, cur);
 485                 } else {
 486                         break;
 487                 }
 488                 count++;
 489         }
 490
 491         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 492             should_end_transaction(trans, root)) {
 493                 trans->transaction->blocked = 1;
 494                 smp_wmb();
 495         }
 496
 497         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 498                 if (throttle) {
 499                         /*
 500                          * We may race with somebody else here so end up having
 501                          * to call end_transaction on ourselves again, so inc
 502                          * our use_count.
 503                          */
 504                         trans->use_count++;
 505                         return btrfs_commit_transaction(trans, root);
 506                 } else {
 507                         wake_up_process(info->transaction_kthread);
 508                 }
 509         }
 510
 511         WARN_ON(cur_trans != info->running_transaction);
 512         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 513         atomic_dec(&cur_trans->num_writers);
 514
 515         smp_mb();
 516         if (waitqueue_active(&cur_trans->writer_wait))
 517                 wake_up(&cur_trans->writer_wait);
 518         put_transaction(cur_trans);
 519
 520         if (current->journal_info == trans)
 521                 current->journal_info = NULL;
 522         memset(trans, 0, sizeof(*trans));
 523         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 524
 525         if (throttle)
 526                 btrfs_run_delayed_iputs(root);
 527
 528         return 0;
 529 }
 530
 531 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 532                           struct btrfs_root *root)
 533 {
 534         int ret;
 535
 536         ret = __btrfs_end_transaction(trans, root, 0, 1);
 537         if (ret)
 538                 return ret;
 539         return 0;
 540 }
 541
 542 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 543                                    struct btrfs_root *root)
 544 {
 545         int ret;
 546
 547         ret = __btrfs_end_transaction(trans, root, 1, 1);
 548         if (ret)
 549                 return ret;
 550         return 0;
 551 }
 552
 553 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 554                                  struct btrfs_root *root)
 555 {
 556         int ret;
 557
 558         ret = __btrfs_end_transaction(trans, root, 0, 0);
 559         if (ret)
 560                 return ret;
 561         return 0;
 562 }
 563
 564 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 565                                 struct btrfs_root *root)
 566 {
 567         return __btrfs_end_transaction(trans, root, 1, 1);
 568 }
 569
 570 /*
 571  * when btree blocks are allocated, they have some corresponding bits set for
 572  * them in one of two extent_io trees.  This is used to make sure all of
 573  * those extents are sent to disk but does not wait on them
 574  */
 575 int btrfs_write_marked_extents(struct btrfs_root *root,
 576                                struct extent_io_tree *dirty_pages, int mark)
 577 {
 578         int err = 0;
 579         int werr = 0;
 580         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 581         u64 start = 0;
 582         u64 end;
 583
 584         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 585                                       mark)) {
 586                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 587                                    GFP_NOFS);
 588                 err = filemap_fdatawrite_range(mapping, start, end);
 589                 if (err)
 590                         werr = err;
 591                 cond_resched();
 592                 start = end + 1;
 593         }
 594         if (err)
 595                 werr = err;
 596         return werr;
 597 }
 598
 599 /*
 600  * when btree blocks are allocated, they have some corresponding bits set for
 601  * them in one of two extent_io trees.  This is used to make sure all of
 602  * those extents are on disk for transaction or log commit.  We wait
 603  * on all the pages and clear them from the dirty pages state tree
 604  */
 605 int btrfs_wait_marked_extents(struct btrfs_root *root,
 606                               struct extent_io_tree *dirty_pages, int mark)
 607 {
 608         int err = 0;
 609         int werr = 0;
 610         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 611         u64 start = 0;
 612         u64 end;
 613
 614         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 615                                       EXTENT_NEED_WAIT)) {
 616                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 617                 err = filemap_fdatawait_range(mapping, start, end);
 618                 if (err)
 619                         werr = err;
 620                 cond_resched();
 621                 start = end + 1;
 622         }
 623         if (err)
 624                 werr = err;
 625         return werr;
 626 }
 627
 628 /*
 629  * when btree blocks are allocated, they have some corresponding bits set for
 630  * them in one of two extent_io trees.  This is used to make sure all of
 631  * those extents are on disk for transaction or log commit
 632  */
 633 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 634                                 struct extent_io_tree *dirty_pages, int mark)
 635 {
 636         int ret;
 637         int ret2;
 638
 639         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 640         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 641
 642         if (ret)
 643                 return ret;
 644         if (ret2)
 645                 return ret2;
 646         return 0;
 647 }
 648
 649 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 650                                      struct btrfs_root *root)
 651 {
 652         if (!trans || !trans->transaction) {
 653                 struct inode *btree_inode;
 654                 btree_inode = root->fs_info->btree_inode;
 655                 return filemap_write_and_wait(btree_inode->i_mapping);
 656         }
 657         return btrfs_write_and_wait_marked_extents(root,
 658                                            &trans->transaction->dirty_pages,
 659                                            EXTENT_DIRTY);
 660 }
 661
 662 /*
 663  * this is used to update the root pointer in the tree of tree roots.
 664  *
 665  * But, in the case of the extent allocation tree, updating the root
 666  * pointer may allocate blocks which may change the root of the extent
 667  * allocation tree.
 668  *
 669  * So, this loops and repeats and makes sure the cowonly root didn't
 670  * change while the root pointer was being updated in the metadata.
 671  */
 672 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 673                                struct btrfs_root *root)
 674 {
 675         int ret;
 676         u64 old_root_bytenr;
 677         u64 old_root_used;
 678         struct btrfs_root *tree_root = root->fs_info->tree_root;
 679
 680         old_root_used = btrfs_root_used(&root->root_item);
 681         btrfs_write_dirty_block_groups(trans, root);
 682
 683         while (1) {
 684                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 685                 if (old_root_bytenr == root->node->start &&
 686                     old_root_used == btrfs_root_used(&root->root_item))
 687                         break;
 688
 689                 btrfs_set_root_node(&root->root_item, root->node);
 690                 ret = btrfs_update_root(trans, tree_root,
 691                                         &root->root_key,
 692                                         &root->root_item);
 693                 BUG_ON(ret);
 694
 695                 old_root_used = btrfs_root_used(&root->root_item);
 696                 ret = btrfs_write_dirty_block_groups(trans, root);
 697                 BUG_ON(ret);
 698         }
 699
 700         if (root != root->fs_info->extent_root)
 701                 switch_commit_root(root);
 702
 703         return 0;
 704 }
 705
 706 /*
 707  * update all the cowonly tree roots on disk
 708  */
 709 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 710                                          struct btrfs_root *root)
 711 {
 712         struct btrfs_fs_info *fs_info = root->fs_info;
 713         struct list_head *next;
 714         struct extent_buffer *eb;
 715         int ret;
 716
 717         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 718         BUG_ON(ret);
 719
 720         eb = btrfs_lock_root_node(fs_info->tree_root);
 721         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 722         btrfs_tree_unlock(eb);
 723         free_extent_buffer(eb);
 724
 725         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 726         BUG_ON(ret);
 727
 728         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 729                 next = fs_info->dirty_cowonly_roots.next;
 730                 list_del_init(next);
 731                 root = list_entry(next, struct btrfs_root, dirty_list);
 732
 733                 update_cowonly_root(trans, root);
 734         }
 735
 736         down_write(&fs_info->extent_commit_sem);
 737         switch_commit_root(fs_info->extent_root);
 738         up_write(&fs_info->extent_commit_sem);
 739
 740         return 0;
 741 }
 742
 743 /*
 744  * dead roots are old snapshots that need to be deleted.  This allocates
 745  * a dirty root struct and adds it into the list of dead roots that need to
 746  * be deleted
 747  */
 748 int btrfs_add_dead_root(struct btrfs_root *root)
 749 {
 750         spin_lock(&root->fs_info->trans_lock);
 751         list_add(&root->root_list, &root->fs_info->dead_roots);
 752         spin_unlock(&root->fs_info->trans_lock);
 753         return 0;
 754 }
 755
 756 /*
 757  * update all the cowonly tree roots on disk
 758  */
 759 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 760                                     struct btrfs_root *root)
 761 {
 762         struct btrfs_root *gang[8];
 763         struct btrfs_fs_info *fs_info = root->fs_info;
 764         int i;
 765         int ret;
 766         int err = 0;
 767
 768         spin_lock(&fs_info->fs_roots_radix_lock);
 769         while (1) {
 770                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 771                                                  (void **)gang, 0,
 772                                                  ARRAY_SIZE(gang),
 773                                                  BTRFS_ROOT_TRANS_TAG);
 774                 if (ret == 0)
 775                         break;
 776                 for (i = 0; i < ret; i++) {
 777                         root = gang[i];
 778                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 779                                         (unsigned long)root->root_key.objectid,
 780                                         BTRFS_ROOT_TRANS_TAG);
 781                         spin_unlock(&fs_info->fs_roots_radix_lock);
 782
 783                         btrfs_free_log(trans, root);
 784                         btrfs_update_reloc_root(trans, root);
 785                         btrfs_orphan_commit_root(trans, root);
 786
 787                         btrfs_save_ino_cache(root, trans);
 788
 789                         /* see comments in should_cow_block() */
 790                         root->force_cow = 0;
 791                         smp_wmb();
 792
 793                         if (root->commit_root != root->node) {
 794                                 mutex_lock(&root->fs_commit_mutex);
 795                                 switch_commit_root(root);
 796                                 btrfs_unpin_free_ino(root);
 797                                 mutex_unlock(&root->fs_commit_mutex);
 798
 799                                 btrfs_set_root_node(&root->root_item,
 800                                                     root->node);
 801                         }
 802
 803                         err = btrfs_update_root(trans, fs_info->tree_root,
 804                                                 &root->root_key,
 805                                                 &root->root_item);
 806                         spin_lock(&fs_info->fs_roots_radix_lock);
 807                         if (err)
 808                                 break;
 809                 }
 810         }
 811         spin_unlock(&fs_info->fs_roots_radix_lock);
 812         return err;
 813 }
 814
 815 /*
 816  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 817  * otherwise every leaf in the btree is read and defragged.
 818  */
 819 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 820 {
 821         struct btrfs_fs_info *info = root->fs_info;
 822         struct btrfs_trans_handle *trans;
 823         int ret;
 824         unsigned long nr;
 825
 826         if (xchg(&root->defrag_running, 1))
 827                 return 0;
 828
 829         while (1) {
 830                 trans = btrfs_start_transaction(root, 0);
 831                 if (IS_ERR(trans))
 832                         return PTR_ERR(trans);
 833
 834                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 835
 836                 nr = trans->blocks_used;
 837                 btrfs_end_transaction(trans, root);
 838                 btrfs_btree_balance_dirty(info->tree_root, nr);
 839                 cond_resched();
 840
 841                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 842                         break;
 843         }
 844         root->defrag_running = 0;
 845         return ret;
 846 }
 847
 848 /*
 849  * new snapshots need to be created at a very specific time in the
 850  * transaction commit.  This does the actual creation
 851  */
 852 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 853                                    struct btrfs_fs_info *fs_info,
 854                                    struct btrfs_pending_snapshot *pending)
 855 {
 856         struct btrfs_key key;
 857         struct btrfs_root_item *new_root_item;
 858         struct btrfs_root *tree_root = fs_info->tree_root;
 859         struct btrfs_root *root = pending->root;
 860         struct btrfs_root *parent_root;
 861         struct btrfs_block_rsv *rsv;
 862         struct inode *parent_inode;
 863         struct dentry *parent;
 864         struct dentry *dentry;
 865         struct extent_buffer *tmp;
 866         struct extent_buffer *old;
 867         int ret;
 868         u64 to_reserve = 0;
 869         u64 index = 0;
 870         u64 objectid;
 871         u64 root_flags;
 872
 873         rsv = trans->block_rsv;
 874
 875         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 876         if (!new_root_item) {
 877                 pending->error = -ENOMEM;
 878                 goto fail;
 879         }
 880
 881         ret = btrfs_find_free_objectid(tree_root, &objectid);
 882         if (ret) {
 883                 pending->error = ret;
 884                 goto fail;
 885         }
 886
 887         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 888
 889         if (to_reserve > 0) {
 890                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 891                                                   to_reserve);
 892                 if (ret) {
 893                         pending->error = ret;
 894                         goto fail;
 895                 }
 896         }
 897
 898         key.objectid = objectid;
 899         key.offset = (u64)-1;
 900         key.type = BTRFS_ROOT_ITEM_KEY;
 901
 902         trans->block_rsv = &pending->block_rsv;
 903
 904         dentry = pending->dentry;
 905         parent = dget_parent(dentry);
 906         parent_inode = parent->d_inode;
 907         parent_root = BTRFS_I(parent_inode)->root;
 908         record_root_in_trans(trans, parent_root);
 909
 910         /*
 911          * insert the directory item
 912          */
 913         ret = btrfs_set_inode_index(parent_inode, &index);
 914         BUG_ON(ret);
 915         ret = btrfs_insert_dir_item(trans, parent_root,
 916                                 dentry->d_name.name, dentry->d_name.len,
 917                                 parent_inode, &key,
 918                                 BTRFS_FT_DIR, index);
 919         if (ret) {
 920                 pending->error = -EEXIST;
 921                 dput(parent);
 922                 goto fail;
 923         }
 924
 925         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 926                                          dentry->d_name.len * 2);
 927         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 928         BUG_ON(ret);
 929
 930         /*
 931          * pull in the delayed directory update
 932          * and the delayed inode item
 933          * otherwise we corrupt the FS during
 934          * snapshot
 935          */
 936         ret = btrfs_run_delayed_items(trans, root);
 937         BUG_ON(ret);
 938
 939         record_root_in_trans(trans, root);
 940         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 941         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 942         btrfs_check_and_init_root_item(new_root_item);
 943
 944         root_flags = btrfs_root_flags(new_root_item);
 945         if (pending->readonly)
 946                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 947         else
 948                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 949         btrfs_set_root_flags(new_root_item, root_flags);
 950
 951         old = btrfs_lock_root_node(root);
 952         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 953         btrfs_set_lock_blocking(old);
 954
 955         btrfs_copy_root(trans, root, old, &tmp, objectid);
 956         btrfs_tree_unlock(old);
 957         free_extent_buffer(old);
 958
 959         /* see comments in should_cow_block() */
 960         root->force_cow = 1;
 961         smp_wmb();
 962
 963         btrfs_set_root_node(new_root_item, tmp);
 964         /* record when the snapshot was created in key.offset */
 965         key.offset = trans->transid;
 966         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 967         btrfs_tree_unlock(tmp);
 968         free_extent_buffer(tmp);
 969         BUG_ON(ret);
 970
 971         /*
 972          * insert root back/forward references
 973          */
 974         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 975                                  parent_root->root_key.objectid,
 976                                  btrfs_ino(parent_inode), index,
 977                                  dentry->d_name.name, dentry->d_name.len);
 978         BUG_ON(ret);
 979         dput(parent);
 980
 981         key.offset = (u64)-1;
 982         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 983         BUG_ON(IS_ERR(pending->snap));
 984
 985         btrfs_reloc_post_snapshot(trans, pending);
 986 fail:
 987         kfree(new_root_item);
 988         trans->block_rsv = rsv;
 989         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 990         return 0;
 991 }
 992
 993 /*
 994  * create all the snapshots we've scheduled for creation
 995  */
 996 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 997                                              struct btrfs_fs_info *fs_info)
 998 {
 999         struct btrfs_pending_snapshot *pending;
1000         struct list_head *head = &trans->transaction->pending_snapshots;
1001
1002         list_for_each_entry(pending, head, list)
1003                 create_pending_snapshot(trans, fs_info, pending);
1004         return 0;
1005 }
1006
1007 static void update_super_roots(struct btrfs_root *root)
1008 {
1009         struct btrfs_root_item *root_item;
1010         struct btrfs_super_block *super;
1011
1012         super = root->fs_info->super_copy;
1013
1014         root_item = &root->fs_info->chunk_root->root_item;
1015         super->chunk_root = root_item->bytenr;
1016         super->chunk_root_generation = root_item->generation;
1017         super->chunk_root_level = root_item->level;
1018
1019         root_item = &root->fs_info->tree_root->root_item;
1020         super->root = root_item->bytenr;
1021         super->generation = root_item->generation;
1022         super->root_level = root_item->level;
1023         if (btrfs_test_opt(root, SPACE_CACHE))
1024                 super->cache_generation = root_item->generation;
1025 }
1026
1027 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1028 {
1029         int ret = 0;
1030         spin_lock(&info->trans_lock);
1031         if (info->running_transaction)
1032                 ret = info->running_transaction->in_commit;
1033         spin_unlock(&info->trans_lock);
1034         return ret;
1035 }
1036
1037 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1038 {
1039         int ret = 0;
1040         spin_lock(&info->trans_lock);
1041         if (info->running_transaction)
1042                 ret = info->running_transaction->blocked;
1043         spin_unlock(&info->trans_lock);
1044         return ret;
1045 }
1046
1047 /*
1048  * wait for the current transaction commit to start and block subsequent
1049  * transaction joins
1050  */
1051 static void wait_current_trans_commit_start(struct btrfs_root *root,
1052                                             struct btrfs_transaction *trans)
1053 {
1054         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1055 }
1056
1057 /*
1058  * wait for the current transaction to start and then become unblocked.
1059  * caller holds ref.
1060  */
1061 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1062                                          struct btrfs_transaction *trans)
1063 {
1064         wait_event(root->fs_info->transaction_wait,
1065                    trans->commit_done || (trans->in_commit && !trans->blocked));
1066 }
1067
1068 /*
1069  * commit transactions asynchronously. once btrfs_commit_transaction_async
1070  * returns, any subsequent transaction will not be allowed to join.
1071  */
1072 struct btrfs_async_commit {
1073         struct btrfs_trans_handle *newtrans;
1074         struct btrfs_root *root;
1075         struct delayed_work work;
1076 };
1077
1078 static void do_async_commit(struct work_struct *work)
1079 {
1080         struct btrfs_async_commit *ac =
1081                 container_of(work, struct btrfs_async_commit, work.work);
1082
1083         btrfs_commit_transaction(ac->newtrans, ac->root);
1084         kfree(ac);
1085 }
1086
1087 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1088                                    struct btrfs_root *root,
1089                                    int wait_for_unblock)
1090 {
1091         struct btrfs_async_commit *ac;
1092         struct btrfs_transaction *cur_trans;
1093
1094         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1095         if (!ac)
1096                 return -ENOMEM;
1097
1098         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1099         ac->root = root;
1100         ac->newtrans = btrfs_join_transaction(root);
1101         if (IS_ERR(ac->newtrans)) {
1102                 int err = PTR_ERR(ac->newtrans);
1103                 kfree(ac);
1104                 return err;
1105         }
1106
1107         /* take transaction reference */
1108         cur_trans = trans->transaction;
1109         atomic_inc(&cur_trans->use_count);
1110
1111         btrfs_end_transaction(trans, root);
1112         schedule_delayed_work(&ac->work, 0);
1113
1114         /* wait for transaction to start and unblock */
1115         if (wait_for_unblock)
1116                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1117         else
1118                 wait_current_trans_commit_start(root, cur_trans);
1119
1120         if (current->journal_info == trans)
1121                 current->journal_info = NULL;
1122
1123         put_transaction(cur_trans);
1124         return 0;
1125 }
1126
1127 /*
1128  * btrfs_transaction state sequence:
1129  *    in_commit = 0, blocked = 0  (initial)
1130  *    in_commit = 1, blocked = 1
1131  *    blocked = 0
1132  *    commit_done = 1
1133  */
1134 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1135                              struct btrfs_root *root)
1136 {
1137         unsigned long joined = 0;
1138         struct btrfs_transaction *cur_trans;
1139         struct btrfs_transaction *prev_trans = NULL;
1140         DEFINE_WAIT(wait);
1141         int ret;
1142         int should_grow = 0;
1143         unsigned long now = get_seconds();
1144         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1145
1146         btrfs_run_ordered_operations(root, 0);
1147
1148         btrfs_trans_release_metadata(trans, root);
1149         trans->block_rsv = NULL;
1150
1151         /* make a pass through all the delayed refs we have so far
1152          * any runnings procs may add more while we are here
1153          */
1154         ret = btrfs_run_delayed_refs(trans, root, 0);
1155         BUG_ON(ret);
1156
1157         cur_trans = trans->transaction;
1158         /*
1159          * set the flushing flag so procs in this transaction have to
1160          * start sending their work down.
1161          */
1162         cur_trans->delayed_refs.flushing = 1;
1163
1164         ret = btrfs_run_delayed_refs(trans, root, 0);
1165         BUG_ON(ret);
1166
1167         spin_lock(&cur_trans->commit_lock);
1168         if (cur_trans->in_commit) {
1169                 spin_unlock(&cur_trans->commit_lock);
1170                 atomic_inc(&cur_trans->use_count);
1171                 btrfs_end_transaction(trans, root);
1172
1173                 wait_for_commit(root, cur_trans);
1174
1175                 put_transaction(cur_trans);
1176
1177                 return 0;
1178         }
1179
1180         trans->transaction->in_commit = 1;
1181         trans->transaction->blocked = 1;
1182         spin_unlock(&cur_trans->commit_lock);
1183         wake_up(&root->fs_info->transaction_blocked_wait);
1184
1185         spin_lock(&root->fs_info->trans_lock);
1186         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1187                 prev_trans = list_entry(cur_trans->list.prev,
1188                                         struct btrfs_transaction, list);
1189                 if (!prev_trans->commit_done) {
1190                         atomic_inc(&prev_trans->use_count);
1191                         spin_unlock(&root->fs_info->trans_lock);
1192
1193                         wait_for_commit(root, prev_trans);
1194
1195                         put_transaction(prev_trans);
1196                 } else {
1197                         spin_unlock(&root->fs_info->trans_lock);
1198                 }
1199         } else {
1200                 spin_unlock(&root->fs_info->trans_lock);
1201         }
1202
1203         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1204                 should_grow = 1;
1205
1206         do {
1207                 int snap_pending = 0;
1208
1209                 joined = cur_trans->num_joined;
1210                 if (!list_empty(&trans->transaction->pending_snapshots))
1211                         snap_pending = 1;
1212
1213                 WARN_ON(cur_trans != trans->transaction);
1214
1215                 if (flush_on_commit || snap_pending) {
1216                         btrfs_start_delalloc_inodes(root, 1);
1217                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1218                         BUG_ON(ret);
1219                 }
1220
1221                 ret = btrfs_run_delayed_items(trans, root);
1222                 BUG_ON(ret);
1223
1224                 /*
1225                  * rename don't use btrfs_join_transaction, so, once we
1226                  * set the transaction to blocked above, we aren't going
1227                  * to get any new ordered operations.  We can safely run
1228                  * it here and no for sure that nothing new will be added
1229                  * to the list
1230                  */
1231                 btrfs_run_ordered_operations(root, 1);
1232
1233                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1234                                 TASK_UNINTERRUPTIBLE);
1235
1236                 if (atomic_read(&cur_trans->num_writers) > 1)
1237                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1238                 else if (should_grow)
1239                         schedule_timeout(1);
1240
1241                 finish_wait(&cur_trans->writer_wait, &wait);
1242         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1243                  (should_grow && cur_trans->num_joined != joined));
1244
1245         /*
1246          * Ok now we need to make sure to block out any other joins while we
1247          * commit the transaction.  We could have started a join before setting
1248          * no_join so make sure to wait for num_writers to == 1 again.
1249          */
1250         spin_lock(&root->fs_info->trans_lock);
1251         root->fs_info->trans_no_join = 1;
1252         spin_unlock(&root->fs_info->trans_lock);
1253         wait_event(cur_trans->writer_wait,
1254                    atomic_read(&cur_trans->num_writers) == 1);
1255
1256         /*
1257          * the reloc mutex makes sure that we stop
1258          * the balancing code from coming in and moving
1259          * extents around in the middle of the commit
1260          */
1261         mutex_lock(&root->fs_info->reloc_mutex);
1262
1263         ret = btrfs_run_delayed_items(trans, root);
1264         BUG_ON(ret);
1265
1266         ret = create_pending_snapshots(trans, root->fs_info);
1267         BUG_ON(ret);
1268
1269         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1270         BUG_ON(ret);
1271
1272         /*
1273          * make sure none of the code above managed to slip in a
1274          * delayed item
1275          */
1276         btrfs_assert_delayed_root_empty(root);
1277
1278         WARN_ON(cur_trans != trans->transaction);
1279
1280         btrfs_scrub_pause(root);
1281         /* btrfs_commit_tree_roots is responsible for getting the
1282          * various roots consistent with each other.  Every pointer
1283          * in the tree of tree roots has to point to the most up to date
1284          * root for every subvolume and other tree.  So, we have to keep
1285          * the tree logging code from jumping in and changing any
1286          * of the trees.
1287          *
1288          * At this point in the commit, there can't be any tree-log
1289          * writers, but a little lower down we drop the trans mutex
1290          * and let new people in.  By holding the tree_log_mutex
1291          * from now until after the super is written, we avoid races
1292          * with the tree-log code.
1293          */
1294         mutex_lock(&root->fs_info->tree_log_mutex);
1295
1296         ret = commit_fs_roots(trans, root);
1297         BUG_ON(ret);
1298
1299         /* commit_fs_roots gets rid of all the tree log roots, it is now
1300          * safe to free the root of tree log roots
1301          */
1302         btrfs_free_log_root_tree(trans, root->fs_info);
1303
1304         ret = commit_cowonly_roots(trans, root);
1305         BUG_ON(ret);
1306
1307         btrfs_prepare_extent_commit(trans, root);
1308
1309         cur_trans = root->fs_info->running_transaction;
1310
1311         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1312                             root->fs_info->tree_root->node);
1313         switch_commit_root(root->fs_info->tree_root);
1314
1315         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1316                             root->fs_info->chunk_root->node);
1317         switch_commit_root(root->fs_info->chunk_root);
1318
1319         update_super_roots(root);
1320
1321         if (!root->fs_info->log_root_recovering) {
1322                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1323                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1324         }
1325
1326         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1327                sizeof(*root->fs_info->super_copy));
1328
1329         trans->transaction->blocked = 0;
1330         spin_lock(&root->fs_info->trans_lock);
1331         root->fs_info->running_transaction = NULL;
1332         root->fs_info->trans_no_join = 0;
1333         spin_unlock(&root->fs_info->trans_lock);
1334         mutex_unlock(&root->fs_info->reloc_mutex);
1335
1336         wake_up(&root->fs_info->transaction_wait);
1337
1338         ret = btrfs_write_and_wait_transaction(trans, root);
1339         BUG_ON(ret);
1340         write_ctree_super(trans, root, 0);
1341
1342         /*
1343          * the super is written, we can safely allow the tree-loggers
1344          * to go about their business
1345          */
1346         mutex_unlock(&root->fs_info->tree_log_mutex);
1347
1348         btrfs_finish_extent_commit(trans, root);
1349
1350         cur_trans->commit_done = 1;
1351
1352         root->fs_info->last_trans_committed = cur_trans->transid;
1353
1354         wake_up(&cur_trans->commit_wait);
1355
1356         spin_lock(&root->fs_info->trans_lock);
1357         list_del_init(&cur_trans->list);
1358         spin_unlock(&root->fs_info->trans_lock);
1359
1360         put_transaction(cur_trans);
1361         put_transaction(cur_trans);
1362
1363         trace_btrfs_transaction_commit(root);
1364
1365         btrfs_scrub_continue(root);
1366
1367         if (current->journal_info == trans)
1368                 current->journal_info = NULL;
1369
1370         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1371
1372         if (current != root->fs_info->transaction_kthread)
1373                 btrfs_run_delayed_iputs(root);
1374
1375         return ret;
1376 }
1377
1378 /*
1379  * interface function to delete all the snapshots we have scheduled for deletion
1380  */
1381 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1382 {
1383         LIST_HEAD(list);
1384         struct btrfs_fs_info *fs_info = root->fs_info;
1385
1386         spin_lock(&fs_info->trans_lock);
1387         list_splice_init(&fs_info->dead_roots, &list);
1388         spin_unlock(&fs_info->trans_lock);
1389
1390         while (!list_empty(&list)) {
1391                 root = list_entry(list.next, struct btrfs_root, root_list);
1392                 list_del(&root->root_list);
1393
1394                 btrfs_kill_all_delayed_nodes(root);
1395
1396                 if (btrfs_header_backref_rev(root->node) <
1397                     BTRFS_MIXED_BACKREF_REV)
1398                         btrfs_drop_snapshot(root, NULL, 0, 0);
1399                 else
1400                         btrfs_drop_snapshot(root, NULL, 1, 0);
1401         }
1402         return 0;
1403 }