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
  31 #define BTRFS_ROOT_TRANS_TAG 0
  32
  33 static noinline void put_transaction(struct btrfs_transaction *transaction)
  34 {
  35         WARN_ON(transaction->use_count == 0);
  36         transaction->use_count--;
  37         if (transaction->use_count == 0) {
  38                 list_del_init(&transaction->list);
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56         cur_trans = root->fs_info->running_transaction;
  57         if (!cur_trans) {
  58                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  59                                              GFP_NOFS);
  60                 BUG_ON(!cur_trans);
  61                 root->fs_info->generation++;
  62                 cur_trans->num_writers = 1;
  63                 cur_trans->num_joined = 0;
  64                 cur_trans->transid = root->fs_info->generation;
  65                 init_waitqueue_head(&cur_trans->writer_wait);
  66                 init_waitqueue_head(&cur_trans->commit_wait);
  67                 cur_trans->in_commit = 0;
  68                 cur_trans->blocked = 0;
  69                 cur_trans->use_count = 1;
  70                 cur_trans->commit_done = 0;
  71                 cur_trans->start_time = get_seconds();
  72
  73                 cur_trans->delayed_refs.root = RB_ROOT;
  74                 cur_trans->delayed_refs.num_entries = 0;
  75                 cur_trans->delayed_refs.num_heads_ready = 0;
  76                 cur_trans->delayed_refs.num_heads = 0;
  77                 cur_trans->delayed_refs.flushing = 0;
  78                 cur_trans->delayed_refs.run_delayed_start = 0;
  79                 spin_lock_init(&cur_trans->delayed_refs.lock);
  80
  81                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  82                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  83                 extent_io_tree_init(&cur_trans->dirty_pages,
  84                                      root->fs_info->btree_inode->i_mapping,
  85                                      GFP_NOFS);
  86                 spin_lock(&root->fs_info->new_trans_lock);
  87                 root->fs_info->running_transaction = cur_trans;
  88                 spin_unlock(&root->fs_info->new_trans_lock);
  89         } else {
  90                 cur_trans->num_writers++;
  91                 cur_trans->num_joined++;
  92         }
  93
  94         return 0;
  95 }
  96
  97 /*
  98  * this does all the record keeping required to make sure that a reference
  99  * counted root is properly recorded in a given transaction.  This is required
 100  * to make sure the old root from before we joined the transaction is deleted
 101  * when the transaction commits
 102  */
 103 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 104                                          struct btrfs_root *root)
 105 {
 106         if (root->ref_cows && root->last_trans < trans->transid) {
 107                 WARN_ON(root == root->fs_info->extent_root);
 108                 WARN_ON(root->commit_root != root->node);
 109
 110                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 111                            (unsigned long)root->root_key.objectid,
 112                            BTRFS_ROOT_TRANS_TAG);
 113                 root->last_trans = trans->transid;
 114                 btrfs_init_reloc_root(trans, root);
 115         }
 116         return 0;
 117 }
 118
 119 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 120                                struct btrfs_root *root)
 121 {
 122         if (!root->ref_cows)
 123                 return 0;
 124
 125         mutex_lock(&root->fs_info->trans_mutex);
 126         if (root->last_trans == trans->transid) {
 127                 mutex_unlock(&root->fs_info->trans_mutex);
 128                 return 0;
 129         }
 130
 131         record_root_in_trans(trans, root);
 132         mutex_unlock(&root->fs_info->trans_mutex);
 133         return 0;
 134 }
 135
 136 /* wait for commit against the current transaction to become unblocked
 137  * when this is done, it is safe to start a new transaction, but the current
 138  * transaction might not be fully on disk.
 139  */
 140 static void wait_current_trans(struct btrfs_root *root)
 141 {
 142         struct btrfs_transaction *cur_trans;
 143
 144         cur_trans = root->fs_info->running_transaction;
 145         if (cur_trans && cur_trans->blocked) {
 146                 DEFINE_WAIT(wait);
 147                 cur_trans->use_count++;
 148                 while (1) {
 149                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 150                                         TASK_UNINTERRUPTIBLE);
 151                         if (!cur_trans->blocked)
 152                                 break;
 153                         mutex_unlock(&root->fs_info->trans_mutex);
 154                         schedule();
 155                         mutex_lock(&root->fs_info->trans_mutex);
 156                 }
 157                 finish_wait(&root->fs_info->transaction_wait, &wait);
 158                 put_transaction(cur_trans);
 159         }
 160 }
 161
 162 enum btrfs_trans_type {
 163         TRANS_START,
 164         TRANS_JOIN,
 165         TRANS_USERSPACE,
 166         TRANS_JOIN_NOLOCK,
 167 };
 168
 169 static int may_wait_transaction(struct btrfs_root *root, int type)
 170 {
 171         if (!root->fs_info->log_root_recovering &&
 172             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 173              type == TRANS_USERSPACE))
 174                 return 1;
 175         return 0;
 176 }
 177
 178 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 179                                                     u64 num_items, int type)
 180 {
 181         struct btrfs_trans_handle *h;
 182         struct btrfs_transaction *cur_trans;
 183         int ret;
 184 again:
 185         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 186         if (!h)
 187                 return ERR_PTR(-ENOMEM);
 188
 189         if (type != TRANS_JOIN_NOLOCK)
 190                 mutex_lock(&root->fs_info->trans_mutex);
 191         if (may_wait_transaction(root, type))
 192                 wait_current_trans(root);
 193
 194         ret = join_transaction(root);
 195         BUG_ON(ret);
 196
 197         cur_trans = root->fs_info->running_transaction;
 198         cur_trans->use_count++;
 199         if (type != TRANS_JOIN_NOLOCK)
 200                 mutex_unlock(&root->fs_info->trans_mutex);
 201
 202         h->transid = cur_trans->transid;
 203         h->transaction = cur_trans;
 204         h->blocks_used = 0;
 205         h->block_group = 0;
 206         h->bytes_reserved = 0;
 207         h->delayed_ref_updates = 0;
 208         h->block_rsv = NULL;
 209
 210         smp_mb();
 211         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 212                 btrfs_commit_transaction(h, root);
 213                 goto again;
 214         }
 215
 216         if (num_items > 0) {
 217                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 218                 if (ret == -EAGAIN) {
 219                         btrfs_commit_transaction(h, root);
 220                         goto again;
 221                 }
 222                 if (ret < 0) {
 223                         btrfs_end_transaction(h, root);
 224                         return ERR_PTR(ret);
 225                 }
 226         }
 227
 228         if (type != TRANS_JOIN_NOLOCK)
 229                 mutex_lock(&root->fs_info->trans_mutex);
 230         record_root_in_trans(h, root);
 231         if (type != TRANS_JOIN_NOLOCK)
 232                 mutex_unlock(&root->fs_info->trans_mutex);
 233
 234         if (!current->journal_info && type != TRANS_USERSPACE)
 235                 current->journal_info = h;
 236         return h;
 237 }
 238
 239 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 240                                                    int num_items)
 241 {
 242         return start_transaction(root, num_items, TRANS_START);
 243 }
 244 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 245                                                    int num_blocks)
 246 {
 247         return start_transaction(root, 0, TRANS_JOIN);
 248 }
 249
 250 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 251                                                           int num_blocks)
 252 {
 253         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 254 }
 255
 256 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 257                                                          int num_blocks)
 258 {
 259         return start_transaction(r, 0, TRANS_USERSPACE);
 260 }
 261
 262 /* wait for a transaction commit to be fully complete */
 263 static noinline int wait_for_commit(struct btrfs_root *root,
 264                                     struct btrfs_transaction *commit)
 265 {
 266         DEFINE_WAIT(wait);
 267         mutex_lock(&root->fs_info->trans_mutex);
 268         while (!commit->commit_done) {
 269                 prepare_to_wait(&commit->commit_wait, &wait,
 270                                 TASK_UNINTERRUPTIBLE);
 271                 if (commit->commit_done)
 272                         break;
 273                 mutex_unlock(&root->fs_info->trans_mutex);
 274                 schedule();
 275                 mutex_lock(&root->fs_info->trans_mutex);
 276         }
 277         mutex_unlock(&root->fs_info->trans_mutex);
 278         finish_wait(&commit->commit_wait, &wait);
 279         return 0;
 280 }
 281
 282 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 283 {
 284         struct btrfs_transaction *cur_trans = NULL, *t;
 285         int ret;
 286
 287         mutex_lock(&root->fs_info->trans_mutex);
 288
 289         ret = 0;
 290         if (transid) {
 291                 if (transid <= root->fs_info->last_trans_committed)
 292                         goto out_unlock;
 293
 294                 /* find specified transaction */
 295                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 296                         if (t->transid == transid) {
 297                                 cur_trans = t;
 298                                 break;
 299                         }
 300                         if (t->transid > transid)
 301                                 break;
 302                 }
 303                 ret = -EINVAL;
 304                 if (!cur_trans)
 305                         goto out_unlock;  /* bad transid */
 306         } else {
 307                 /* find newest transaction that is committing | committed */
 308                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 309                                             list) {
 310                         if (t->in_commit) {
 311                                 if (t->commit_done)
 312                                         goto out_unlock;
 313                                 cur_trans = t;
 314                                 break;
 315                         }
 316                 }
 317                 if (!cur_trans)
 318                         goto out_unlock;  /* nothing committing|committed */
 319         }
 320
 321         cur_trans->use_count++;
 322         mutex_unlock(&root->fs_info->trans_mutex);
 323
 324         wait_for_commit(root, cur_trans);
 325
 326         mutex_lock(&root->fs_info->trans_mutex);
 327         put_transaction(cur_trans);
 328         ret = 0;
 329 out_unlock:
 330         mutex_unlock(&root->fs_info->trans_mutex);
 331         return ret;
 332 }
 333
 334 #if 0
 335 /*
 336  * rate limit against the drop_snapshot code.  This helps to slow down new
 337  * operations if the drop_snapshot code isn't able to keep up.
 338  */
 339 static void throttle_on_drops(struct btrfs_root *root)
 340 {
 341         struct btrfs_fs_info *info = root->fs_info;
 342         int harder_count = 0;
 343
 344 harder:
 345         if (atomic_read(&info->throttles)) {
 346                 DEFINE_WAIT(wait);
 347                 int thr;
 348                 thr = atomic_read(&info->throttle_gen);
 349
 350                 do {
 351                         prepare_to_wait(&info->transaction_throttle,
 352                                         &wait, TASK_UNINTERRUPTIBLE);
 353                         if (!atomic_read(&info->throttles)) {
 354                                 finish_wait(&info->transaction_throttle, &wait);
 355                                 break;
 356                         }
 357                         schedule();
 358                         finish_wait(&info->transaction_throttle, &wait);
 359                 } while (thr == atomic_read(&info->throttle_gen));
 360                 harder_count++;
 361
 362                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 363                     harder_count < 2)
 364                         goto harder;
 365
 366                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 367                     harder_count < 10)
 368                         goto harder;
 369
 370                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 371                     harder_count < 20)
 372                         goto harder;
 373         }
 374 }
 375 #endif
 376
 377 void btrfs_throttle(struct btrfs_root *root)
 378 {
 379         mutex_lock(&root->fs_info->trans_mutex);
 380         if (!root->fs_info->open_ioctl_trans)
 381                 wait_current_trans(root);
 382         mutex_unlock(&root->fs_info->trans_mutex);
 383 }
 384
 385 static int should_end_transaction(struct btrfs_trans_handle *trans,
 386                                   struct btrfs_root *root)
 387 {
 388         int ret;
 389         ret = btrfs_block_rsv_check(trans, root,
 390                                     &root->fs_info->global_block_rsv, 0, 5);
 391         return ret ? 1 : 0;
 392 }
 393
 394 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 395                                  struct btrfs_root *root)
 396 {
 397         struct btrfs_transaction *cur_trans = trans->transaction;
 398         int updates;
 399
 400         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 401                 return 1;
 402
 403         updates = trans->delayed_ref_updates;
 404         trans->delayed_ref_updates = 0;
 405         if (updates)
 406                 btrfs_run_delayed_refs(trans, root, updates);
 407
 408         return should_end_transaction(trans, root);
 409 }
 410
 411 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 412                           struct btrfs_root *root, int throttle, int lock)
 413 {
 414         struct btrfs_transaction *cur_trans = trans->transaction;
 415         struct btrfs_fs_info *info = root->fs_info;
 416         int count = 0;
 417
 418         while (count < 4) {
 419                 unsigned long cur = trans->delayed_ref_updates;
 420                 trans->delayed_ref_updates = 0;
 421                 if (cur &&
 422                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 423                         trans->delayed_ref_updates = 0;
 424
 425                         /*
 426                          * do a full flush if the transaction is trying
 427                          * to close
 428                          */
 429                         if (trans->transaction->delayed_refs.flushing)
 430                                 cur = 0;
 431                         btrfs_run_delayed_refs(trans, root, cur);
 432                 } else {
 433                         break;
 434                 }
 435                 count++;
 436         }
 437
 438         btrfs_trans_release_metadata(trans, root);
 439
 440         if (lock && !root->fs_info->open_ioctl_trans &&
 441             should_end_transaction(trans, root))
 442                 trans->transaction->blocked = 1;
 443
 444         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 445                 if (throttle)
 446                         return btrfs_commit_transaction(trans, root);
 447                 else
 448                         wake_up_process(info->transaction_kthread);
 449         }
 450
 451         if (lock)
 452                 mutex_lock(&info->trans_mutex);
 453         WARN_ON(cur_trans != info->running_transaction);
 454         WARN_ON(cur_trans->num_writers < 1);
 455         cur_trans->num_writers--;
 456
 457         smp_mb();
 458         if (waitqueue_active(&cur_trans->writer_wait))
 459                 wake_up(&cur_trans->writer_wait);
 460         put_transaction(cur_trans);
 461         if (lock)
 462                 mutex_unlock(&info->trans_mutex);
 463
 464         if (current->journal_info == trans)
 465                 current->journal_info = NULL;
 466         memset(trans, 0, sizeof(*trans));
 467         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 468
 469         if (throttle)
 470                 btrfs_run_delayed_iputs(root);
 471
 472         return 0;
 473 }
 474
 475 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 476                           struct btrfs_root *root)
 477 {
 478         return __btrfs_end_transaction(trans, root, 0, 1);
 479 }
 480
 481 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 482                                    struct btrfs_root *root)
 483 {
 484         return __btrfs_end_transaction(trans, root, 1, 1);
 485 }
 486
 487 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 488                                  struct btrfs_root *root)
 489 {
 490         return __btrfs_end_transaction(trans, root, 0, 0);
 491 }
 492
 493 /*
 494  * when btree blocks are allocated, they have some corresponding bits set for
 495  * them in one of two extent_io trees.  This is used to make sure all of
 496  * those extents are sent to disk but does not wait on them
 497  */
 498 int btrfs_write_marked_extents(struct btrfs_root *root,
 499                                struct extent_io_tree *dirty_pages, int mark)
 500 {
 501         int ret;
 502         int err = 0;
 503         int werr = 0;
 504         struct page *page;
 505         struct inode *btree_inode = root->fs_info->btree_inode;
 506         u64 start = 0;
 507         u64 end;
 508         unsigned long index;
 509
 510         while (1) {
 511                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 512                                             mark);
 513                 if (ret)
 514                         break;
 515                 while (start <= end) {
 516                         cond_resched();
 517
 518                         index = start >> PAGE_CACHE_SHIFT;
 519                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 520                         page = find_get_page(btree_inode->i_mapping, index);
 521                         if (!page)
 522                                 continue;
 523
 524                         btree_lock_page_hook(page);
 525                         if (!page->mapping) {
 526                                 unlock_page(page);
 527                                 page_cache_release(page);
 528                                 continue;
 529                         }
 530
 531                         if (PageWriteback(page)) {
 532                                 if (PageDirty(page))
 533                                         wait_on_page_writeback(page);
 534                                 else {
 535                                         unlock_page(page);
 536                                         page_cache_release(page);
 537                                         continue;
 538                                 }
 539                         }
 540                         err = write_one_page(page, 0);
 541                         if (err)
 542                                 werr = err;
 543                         page_cache_release(page);
 544                 }
 545         }
 546         if (err)
 547                 werr = err;
 548         return werr;
 549 }
 550
 551 /*
 552  * when btree blocks are allocated, they have some corresponding bits set for
 553  * them in one of two extent_io trees.  This is used to make sure all of
 554  * those extents are on disk for transaction or log commit.  We wait
 555  * on all the pages and clear them from the dirty pages state tree
 556  */
 557 int btrfs_wait_marked_extents(struct btrfs_root *root,
 558                               struct extent_io_tree *dirty_pages, int mark)
 559 {
 560         int ret;
 561         int err = 0;
 562         int werr = 0;
 563         struct page *page;
 564         struct inode *btree_inode = root->fs_info->btree_inode;
 565         u64 start = 0;
 566         u64 end;
 567         unsigned long index;
 568
 569         while (1) {
 570                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 571                                             mark);
 572                 if (ret)
 573                         break;
 574
 575                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 576                 while (start <= end) {
 577                         index = start >> PAGE_CACHE_SHIFT;
 578                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 579                         page = find_get_page(btree_inode->i_mapping, index);
 580                         if (!page)
 581                                 continue;
 582                         if (PageDirty(page)) {
 583                                 btree_lock_page_hook(page);
 584                                 wait_on_page_writeback(page);
 585                                 err = write_one_page(page, 0);
 586                                 if (err)
 587                                         werr = err;
 588                         }
 589                         wait_on_page_writeback(page);
 590                         page_cache_release(page);
 591                         cond_resched();
 592                 }
 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
 603  */
 604 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 605                                 struct extent_io_tree *dirty_pages, int mark)
 606 {
 607         int ret;
 608         int ret2;
 609
 610         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 611         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 612         return ret || ret2;
 613 }
 614
 615 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 616                                      struct btrfs_root *root)
 617 {
 618         if (!trans || !trans->transaction) {
 619                 struct inode *btree_inode;
 620                 btree_inode = root->fs_info->btree_inode;
 621                 return filemap_write_and_wait(btree_inode->i_mapping);
 622         }
 623         return btrfs_write_and_wait_marked_extents(root,
 624                                            &trans->transaction->dirty_pages,
 625                                            EXTENT_DIRTY);
 626 }
 627
 628 /*
 629  * this is used to update the root pointer in the tree of tree roots.
 630  *
 631  * But, in the case of the extent allocation tree, updating the root
 632  * pointer may allocate blocks which may change the root of the extent
 633  * allocation tree.
 634  *
 635  * So, this loops and repeats and makes sure the cowonly root didn't
 636  * change while the root pointer was being updated in the metadata.
 637  */
 638 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 639                                struct btrfs_root *root)
 640 {
 641         int ret;
 642         u64 old_root_bytenr;
 643         u64 old_root_used;
 644         struct btrfs_root *tree_root = root->fs_info->tree_root;
 645
 646         old_root_used = btrfs_root_used(&root->root_item);
 647         btrfs_write_dirty_block_groups(trans, root);
 648
 649         while (1) {
 650                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 651                 if (old_root_bytenr == root->node->start &&
 652                     old_root_used == btrfs_root_used(&root->root_item))
 653                         break;
 654
 655                 btrfs_set_root_node(&root->root_item, root->node);
 656                 ret = btrfs_update_root(trans, tree_root,
 657                                         &root->root_key,
 658                                         &root->root_item);
 659                 BUG_ON(ret);
 660
 661                 old_root_used = btrfs_root_used(&root->root_item);
 662                 ret = btrfs_write_dirty_block_groups(trans, root);
 663                 BUG_ON(ret);
 664         }
 665
 666         if (root != root->fs_info->extent_root)
 667                 switch_commit_root(root);
 668
 669         return 0;
 670 }
 671
 672 /*
 673  * update all the cowonly tree roots on disk
 674  */
 675 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 676                                          struct btrfs_root *root)
 677 {
 678         struct btrfs_fs_info *fs_info = root->fs_info;
 679         struct list_head *next;
 680         struct extent_buffer *eb;
 681         int ret;
 682
 683         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 684         BUG_ON(ret);
 685
 686         eb = btrfs_lock_root_node(fs_info->tree_root);
 687         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 688         btrfs_tree_unlock(eb);
 689         free_extent_buffer(eb);
 690
 691         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 692         BUG_ON(ret);
 693
 694         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 695                 next = fs_info->dirty_cowonly_roots.next;
 696                 list_del_init(next);
 697                 root = list_entry(next, struct btrfs_root, dirty_list);
 698
 699                 update_cowonly_root(trans, root);
 700         }
 701
 702         down_write(&fs_info->extent_commit_sem);
 703         switch_commit_root(fs_info->extent_root);
 704         up_write(&fs_info->extent_commit_sem);
 705
 706         return 0;
 707 }
 708
 709 /*
 710  * dead roots are old snapshots that need to be deleted.  This allocates
 711  * a dirty root struct and adds it into the list of dead roots that need to
 712  * be deleted
 713  */
 714 int btrfs_add_dead_root(struct btrfs_root *root)
 715 {
 716         mutex_lock(&root->fs_info->trans_mutex);
 717         list_add(&root->root_list, &root->fs_info->dead_roots);
 718         mutex_unlock(&root->fs_info->trans_mutex);
 719         return 0;
 720 }
 721
 722 /*
 723  * update all the cowonly tree roots on disk
 724  */
 725 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 726                                     struct btrfs_root *root)
 727 {
 728         struct btrfs_root *gang[8];
 729         struct btrfs_fs_info *fs_info = root->fs_info;
 730         int i;
 731         int ret;
 732         int err = 0;
 733
 734         while (1) {
 735                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 736                                                  (void **)gang, 0,
 737                                                  ARRAY_SIZE(gang),
 738                                                  BTRFS_ROOT_TRANS_TAG);
 739                 if (ret == 0)
 740                         break;
 741                 for (i = 0; i < ret; i++) {
 742                         root = gang[i];
 743                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 744                                         (unsigned long)root->root_key.objectid,
 745                                         BTRFS_ROOT_TRANS_TAG);
 746
 747                         btrfs_free_log(trans, root);
 748                         btrfs_update_reloc_root(trans, root);
 749                         btrfs_orphan_commit_root(trans, root);
 750
 751                         if (root->commit_root != root->node) {
 752                                 switch_commit_root(root);
 753                                 btrfs_set_root_node(&root->root_item,
 754                                                     root->node);
 755                         }
 756
 757                         err = btrfs_update_root(trans, fs_info->tree_root,
 758                                                 &root->root_key,
 759                                                 &root->root_item);
 760                         if (err)
 761                                 break;
 762                 }
 763         }
 764         return err;
 765 }
 766
 767 /*
 768  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 769  * otherwise every leaf in the btree is read and defragged.
 770  */
 771 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 772 {
 773         struct btrfs_fs_info *info = root->fs_info;
 774         struct btrfs_trans_handle *trans;
 775         int ret;
 776         unsigned long nr;
 777
 778         if (xchg(&root->defrag_running, 1))
 779                 return 0;
 780
 781         while (1) {
 782                 trans = btrfs_start_transaction(root, 0);
 783                 if (IS_ERR(trans))
 784                         return PTR_ERR(trans);
 785
 786                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 787
 788                 nr = trans->blocks_used;
 789                 btrfs_end_transaction(trans, root);
 790                 btrfs_btree_balance_dirty(info->tree_root, nr);
 791                 cond_resched();
 792
 793                 if (root->fs_info->closing || ret != -EAGAIN)
 794                         break;
 795         }
 796         root->defrag_running = 0;
 797         return ret;
 798 }
 799
 800 #if 0
 801 /*
 802  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 803  * sense not to join the transaction while it is trying to flush the current
 804  * queue of delayed refs out.
 805  *
 806  * This is used by the drop snapshot code only
 807  */
 808 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 809 {
 810         DEFINE_WAIT(wait);
 811
 812         mutex_lock(&info->trans_mutex);
 813         while (info->running_transaction &&
 814                info->running_transaction->delayed_refs.flushing) {
 815                 prepare_to_wait(&info->transaction_wait, &wait,
 816                                 TASK_UNINTERRUPTIBLE);
 817                 mutex_unlock(&info->trans_mutex);
 818
 819                 schedule();
 820
 821                 mutex_lock(&info->trans_mutex);
 822                 finish_wait(&info->transaction_wait, &wait);
 823         }
 824         mutex_unlock(&info->trans_mutex);
 825         return 0;
 826 }
 827
 828 /*
 829  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 830  * all of them
 831  */
 832 int btrfs_drop_dead_root(struct btrfs_root *root)
 833 {
 834         struct btrfs_trans_handle *trans;
 835         struct btrfs_root *tree_root = root->fs_info->tree_root;
 836         unsigned long nr;
 837         int ret;
 838
 839         while (1) {
 840                 /*
 841                  * we don't want to jump in and create a bunch of
 842                  * delayed refs if the transaction is starting to close
 843                  */
 844                 wait_transaction_pre_flush(tree_root->fs_info);
 845                 trans = btrfs_start_transaction(tree_root, 1);
 846
 847                 /*
 848                  * we've joined a transaction, make sure it isn't
 849                  * closing right now
 850                  */
 851                 if (trans->transaction->delayed_refs.flushing) {
 852                         btrfs_end_transaction(trans, tree_root);
 853                         continue;
 854                 }
 855
 856                 ret = btrfs_drop_snapshot(trans, root);
 857                 if (ret != -EAGAIN)
 858                         break;
 859
 860                 ret = btrfs_update_root(trans, tree_root,
 861                                         &root->root_key,
 862                                         &root->root_item);
 863                 if (ret)
 864                         break;
 865
 866                 nr = trans->blocks_used;
 867                 ret = btrfs_end_transaction(trans, tree_root);
 868                 BUG_ON(ret);
 869
 870                 btrfs_btree_balance_dirty(tree_root, nr);
 871                 cond_resched();
 872         }
 873         BUG_ON(ret);
 874
 875         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 876         BUG_ON(ret);
 877
 878         nr = trans->blocks_used;
 879         ret = btrfs_end_transaction(trans, tree_root);
 880         BUG_ON(ret);
 881
 882         free_extent_buffer(root->node);
 883         free_extent_buffer(root->commit_root);
 884         kfree(root);
 885
 886         btrfs_btree_balance_dirty(tree_root, nr);
 887         return ret;
 888 }
 889 #endif
 890
 891 /*
 892  * new snapshots need to be created at a very specific time in the
 893  * transaction commit.  This does the actual creation
 894  */
 895 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 896                                    struct btrfs_fs_info *fs_info,
 897                                    struct btrfs_pending_snapshot *pending)
 898 {
 899         struct btrfs_key key;
 900         struct btrfs_root_item *new_root_item;
 901         struct btrfs_root *tree_root = fs_info->tree_root;
 902         struct btrfs_root *root = pending->root;
 903         struct btrfs_root *parent_root;
 904         struct inode *parent_inode;
 905         struct dentry *parent;
 906         struct dentry *dentry;
 907         struct extent_buffer *tmp;
 908         struct extent_buffer *old;
 909         int ret;
 910         u64 to_reserve = 0;
 911         u64 index = 0;
 912         u64 objectid;
 913
 914         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 915         if (!new_root_item) {
 916                 pending->error = -ENOMEM;
 917                 goto fail;
 918         }
 919
 920         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 921         if (ret) {
 922                 pending->error = ret;
 923                 goto fail;
 924         }
 925
 926         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 927         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 928
 929         if (to_reserve > 0) {
 930                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 931                                           to_reserve);
 932                 if (ret) {
 933                         pending->error = ret;
 934                         goto fail;
 935                 }
 936         }
 937
 938         key.objectid = objectid;
 939         key.offset = (u64)-1;
 940         key.type = BTRFS_ROOT_ITEM_KEY;
 941
 942         trans->block_rsv = &pending->block_rsv;
 943
 944         dentry = pending->dentry;
 945         parent = dget_parent(dentry);
 946         parent_inode = parent->d_inode;
 947         parent_root = BTRFS_I(parent_inode)->root;
 948         record_root_in_trans(trans, parent_root);
 949
 950         /*
 951          * insert the directory item
 952          */
 953         ret = btrfs_set_inode_index(parent_inode, &index);
 954         BUG_ON(ret);
 955         ret = btrfs_insert_dir_item(trans, parent_root,
 956                                 dentry->d_name.name, dentry->d_name.len,
 957                                 parent_inode->i_ino, &key,
 958                                 BTRFS_FT_DIR, index);
 959         BUG_ON(ret);
 960
 961         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 962                                          dentry->d_name.len * 2);
 963         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 964         BUG_ON(ret);
 965
 966         record_root_in_trans(trans, root);
 967         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 968         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 969
 970         old = btrfs_lock_root_node(root);
 971         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 972         btrfs_set_lock_blocking(old);
 973
 974         btrfs_copy_root(trans, root, old, &tmp, objectid);
 975         btrfs_tree_unlock(old);
 976         free_extent_buffer(old);
 977
 978         btrfs_set_root_node(new_root_item, tmp);
 979         /* record when the snapshot was created in key.offset */
 980         key.offset = trans->transid;
 981         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 982         btrfs_tree_unlock(tmp);
 983         free_extent_buffer(tmp);
 984         BUG_ON(ret);
 985
 986         /*
 987          * insert root back/forward references
 988          */
 989         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 990                                  parent_root->root_key.objectid,
 991                                  parent_inode->i_ino, index,
 992                                  dentry->d_name.name, dentry->d_name.len);
 993         BUG_ON(ret);
 994         dput(parent);
 995
 996         key.offset = (u64)-1;
 997         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 998         BUG_ON(IS_ERR(pending->snap));
 999
1000         btrfs_reloc_post_snapshot(trans, pending);
1001         btrfs_orphan_post_snapshot(trans, pending);
1002 fail:
1003         kfree(new_root_item);
1004         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1005         return 0;
1006 }
1007
1008 /*
1009  * create all the snapshots we've scheduled for creation
1010  */
1011 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1012                                              struct btrfs_fs_info *fs_info)
1013 {
1014         struct btrfs_pending_snapshot *pending;
1015         struct list_head *head = &trans->transaction->pending_snapshots;
1016         int ret;
1017
1018         list_for_each_entry(pending, head, list) {
1019                 ret = create_pending_snapshot(trans, fs_info, pending);
1020                 BUG_ON(ret);
1021         }
1022         return 0;
1023 }
1024
1025 static void update_super_roots(struct btrfs_root *root)
1026 {
1027         struct btrfs_root_item *root_item;
1028         struct btrfs_super_block *super;
1029
1030         super = &root->fs_info->super_copy;
1031
1032         root_item = &root->fs_info->chunk_root->root_item;
1033         super->chunk_root = root_item->bytenr;
1034         super->chunk_root_generation = root_item->generation;
1035         super->chunk_root_level = root_item->level;
1036
1037         root_item = &root->fs_info->tree_root->root_item;
1038         super->root = root_item->bytenr;
1039         super->generation = root_item->generation;
1040         super->root_level = root_item->level;
1041         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1042                 super->cache_generation = root_item->generation;
1043 }
1044
1045 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1046 {
1047         int ret = 0;
1048         spin_lock(&info->new_trans_lock);
1049         if (info->running_transaction)
1050                 ret = info->running_transaction->in_commit;
1051         spin_unlock(&info->new_trans_lock);
1052         return ret;
1053 }
1054
1055 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1056 {
1057         int ret = 0;
1058         spin_lock(&info->new_trans_lock);
1059         if (info->running_transaction)
1060                 ret = info->running_transaction->blocked;
1061         spin_unlock(&info->new_trans_lock);
1062         return ret;
1063 }
1064
1065 /*
1066  * wait for the current transaction commit to start and block subsequent
1067  * transaction joins
1068  */
1069 static void wait_current_trans_commit_start(struct btrfs_root *root,
1070                                             struct btrfs_transaction *trans)
1071 {
1072         DEFINE_WAIT(wait);
1073
1074         if (trans->in_commit)
1075                 return;
1076
1077         while (1) {
1078                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1079                                 TASK_UNINTERRUPTIBLE);
1080                 if (trans->in_commit) {
1081                         finish_wait(&root->fs_info->transaction_blocked_wait,
1082                                     &wait);
1083                         break;
1084                 }
1085                 mutex_unlock(&root->fs_info->trans_mutex);
1086                 schedule();
1087                 mutex_lock(&root->fs_info->trans_mutex);
1088                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1089         }
1090 }
1091
1092 /*
1093  * wait for the current transaction to start and then become unblocked.
1094  * caller holds ref.
1095  */
1096 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1097                                          struct btrfs_transaction *trans)
1098 {
1099         DEFINE_WAIT(wait);
1100
1101         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1102                 return;
1103
1104         while (1) {
1105                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1106                                 TASK_UNINTERRUPTIBLE);
1107                 if (trans->commit_done ||
1108                     (trans->in_commit && !trans->blocked)) {
1109                         finish_wait(&root->fs_info->transaction_wait,
1110                                     &wait);
1111                         break;
1112                 }
1113                 mutex_unlock(&root->fs_info->trans_mutex);
1114                 schedule();
1115                 mutex_lock(&root->fs_info->trans_mutex);
1116                 finish_wait(&root->fs_info->transaction_wait,
1117                             &wait);
1118         }
1119 }
1120
1121 /*
1122  * commit transactions asynchronously. once btrfs_commit_transaction_async
1123  * returns, any subsequent transaction will not be allowed to join.
1124  */
1125 struct btrfs_async_commit {
1126         struct btrfs_trans_handle *newtrans;
1127         struct btrfs_root *root;
1128         struct delayed_work work;
1129 };
1130
1131 static void do_async_commit(struct work_struct *work)
1132 {
1133         struct btrfs_async_commit *ac =
1134                 container_of(work, struct btrfs_async_commit, work.work);
1135
1136         btrfs_commit_transaction(ac->newtrans, ac->root);
1137         kfree(ac);
1138 }
1139
1140 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1141                                    struct btrfs_root *root,
1142                                    int wait_for_unblock)
1143 {
1144         struct btrfs_async_commit *ac;
1145         struct btrfs_transaction *cur_trans;
1146
1147         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1148         BUG_ON(!ac);
1149
1150         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1151         ac->root = root;
1152         ac->newtrans = btrfs_join_transaction(root, 0);
1153
1154         /* take transaction reference */
1155         mutex_lock(&root->fs_info->trans_mutex);
1156         cur_trans = trans->transaction;
1157         cur_trans->use_count++;
1158         mutex_unlock(&root->fs_info->trans_mutex);
1159
1160         btrfs_end_transaction(trans, root);
1161         schedule_delayed_work(&ac->work, 0);
1162
1163         /* wait for transaction to start and unblock */
1164         mutex_lock(&root->fs_info->trans_mutex);
1165         if (wait_for_unblock)
1166                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1167         else
1168                 wait_current_trans_commit_start(root, cur_trans);
1169         put_transaction(cur_trans);
1170         mutex_unlock(&root->fs_info->trans_mutex);
1171
1172         return 0;
1173 }
1174
1175 /*
1176  * btrfs_transaction state sequence:
1177  *    in_commit = 0, blocked = 0  (initial)
1178  *    in_commit = 1, blocked = 1
1179  *    blocked = 0
1180  *    commit_done = 1
1181  */
1182 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1183                              struct btrfs_root *root)
1184 {
1185         unsigned long joined = 0;
1186         struct btrfs_transaction *cur_trans;
1187         struct btrfs_transaction *prev_trans = NULL;
1188         DEFINE_WAIT(wait);
1189         int ret;
1190         int should_grow = 0;
1191         unsigned long now = get_seconds();
1192         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1193
1194         btrfs_run_ordered_operations(root, 0);
1195
1196         /* make a pass through all the delayed refs we have so far
1197          * any runnings procs may add more while we are here
1198          */
1199         ret = btrfs_run_delayed_refs(trans, root, 0);
1200         BUG_ON(ret);
1201
1202         btrfs_trans_release_metadata(trans, root);
1203
1204         cur_trans = trans->transaction;
1205         /*
1206          * set the flushing flag so procs in this transaction have to
1207          * start sending their work down.
1208          */
1209         cur_trans->delayed_refs.flushing = 1;
1210
1211         ret = btrfs_run_delayed_refs(trans, root, 0);
1212         BUG_ON(ret);
1213
1214         mutex_lock(&root->fs_info->trans_mutex);
1215         if (cur_trans->in_commit) {
1216                 cur_trans->use_count++;
1217                 mutex_unlock(&root->fs_info->trans_mutex);
1218                 btrfs_end_transaction(trans, root);
1219
1220                 ret = wait_for_commit(root, cur_trans);
1221                 BUG_ON(ret);
1222
1223                 mutex_lock(&root->fs_info->trans_mutex);
1224                 put_transaction(cur_trans);
1225                 mutex_unlock(&root->fs_info->trans_mutex);
1226
1227                 return 0;
1228         }
1229
1230         trans->transaction->in_commit = 1;
1231         trans->transaction->blocked = 1;
1232         wake_up(&root->fs_info->transaction_blocked_wait);
1233
1234         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1235                 prev_trans = list_entry(cur_trans->list.prev,
1236                                         struct btrfs_transaction, list);
1237                 if (!prev_trans->commit_done) {
1238                         prev_trans->use_count++;
1239                         mutex_unlock(&root->fs_info->trans_mutex);
1240
1241                         wait_for_commit(root, prev_trans);
1242
1243                         mutex_lock(&root->fs_info->trans_mutex);
1244                         put_transaction(prev_trans);
1245                 }
1246         }
1247
1248         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1249                 should_grow = 1;
1250
1251         do {
1252                 int snap_pending = 0;
1253                 joined = cur_trans->num_joined;
1254                 if (!list_empty(&trans->transaction->pending_snapshots))
1255                         snap_pending = 1;
1256
1257                 WARN_ON(cur_trans != trans->transaction);
1258                 mutex_unlock(&root->fs_info->trans_mutex);
1259
1260                 if (flush_on_commit || snap_pending) {
1261                         btrfs_start_delalloc_inodes(root, 1);
1262                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1263                         BUG_ON(ret);
1264                 }
1265
1266                 /*
1267                  * rename don't use btrfs_join_transaction, so, once we
1268                  * set the transaction to blocked above, we aren't going
1269                  * to get any new ordered operations.  We can safely run
1270                  * it here and no for sure that nothing new will be added
1271                  * to the list
1272                  */
1273                 btrfs_run_ordered_operations(root, 1);
1274
1275                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1276                                 TASK_UNINTERRUPTIBLE);
1277
1278                 smp_mb();
1279                 if (cur_trans->num_writers > 1)
1280                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1281                 else if (should_grow)
1282                         schedule_timeout(1);
1283
1284                 mutex_lock(&root->fs_info->trans_mutex);
1285                 finish_wait(&cur_trans->writer_wait, &wait);
1286         } while (cur_trans->num_writers > 1 ||
1287                  (should_grow && cur_trans->num_joined != joined));
1288
1289         ret = create_pending_snapshots(trans, root->fs_info);
1290         BUG_ON(ret);
1291
1292         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1293         BUG_ON(ret);
1294
1295         WARN_ON(cur_trans != trans->transaction);
1296
1297         /* btrfs_commit_tree_roots is responsible for getting the
1298          * various roots consistent with each other.  Every pointer
1299          * in the tree of tree roots has to point to the most up to date
1300          * root for every subvolume and other tree.  So, we have to keep
1301          * the tree logging code from jumping in and changing any
1302          * of the trees.
1303          *
1304          * At this point in the commit, there can't be any tree-log
1305          * writers, but a little lower down we drop the trans mutex
1306          * and let new people in.  By holding the tree_log_mutex
1307          * from now until after the super is written, we avoid races
1308          * with the tree-log code.
1309          */
1310         mutex_lock(&root->fs_info->tree_log_mutex);
1311
1312         ret = commit_fs_roots(trans, root);
1313         BUG_ON(ret);
1314
1315         /* commit_fs_roots gets rid of all the tree log roots, it is now
1316          * safe to free the root of tree log roots
1317          */
1318         btrfs_free_log_root_tree(trans, root->fs_info);
1319
1320         ret = commit_cowonly_roots(trans, root);
1321         BUG_ON(ret);
1322
1323         btrfs_prepare_extent_commit(trans, root);
1324
1325         cur_trans = root->fs_info->running_transaction;
1326         spin_lock(&root->fs_info->new_trans_lock);
1327         root->fs_info->running_transaction = NULL;
1328         spin_unlock(&root->fs_info->new_trans_lock);
1329
1330         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1331                             root->fs_info->tree_root->node);
1332         switch_commit_root(root->fs_info->tree_root);
1333
1334         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1335                             root->fs_info->chunk_root->node);
1336         switch_commit_root(root->fs_info->chunk_root);
1337
1338         update_super_roots(root);
1339
1340         if (!root->fs_info->log_root_recovering) {
1341                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1342                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1343         }
1344
1345         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1346                sizeof(root->fs_info->super_copy));
1347
1348         trans->transaction->blocked = 0;
1349
1350         wake_up(&root->fs_info->transaction_wait);
1351
1352         mutex_unlock(&root->fs_info->trans_mutex);
1353         ret = btrfs_write_and_wait_transaction(trans, root);
1354         BUG_ON(ret);
1355         write_ctree_super(trans, root, 0);
1356
1357         /*
1358          * the super is written, we can safely allow the tree-loggers
1359          * to go about their business
1360          */
1361         mutex_unlock(&root->fs_info->tree_log_mutex);
1362
1363         btrfs_finish_extent_commit(trans, root);
1364
1365         mutex_lock(&root->fs_info->trans_mutex);
1366
1367         cur_trans->commit_done = 1;
1368
1369         root->fs_info->last_trans_committed = cur_trans->transid;
1370
1371         wake_up(&cur_trans->commit_wait);
1372
1373         put_transaction(cur_trans);
1374         put_transaction(cur_trans);
1375
1376         mutex_unlock(&root->fs_info->trans_mutex);
1377
1378         if (current->journal_info == trans)
1379                 current->journal_info = NULL;
1380
1381         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1382
1383         if (current != root->fs_info->transaction_kthread)
1384                 btrfs_run_delayed_iputs(root);
1385
1386         return ret;
1387 }
1388
1389 /*
1390  * interface function to delete all the snapshots we have scheduled for deletion
1391  */
1392 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1393 {
1394         LIST_HEAD(list);
1395         struct btrfs_fs_info *fs_info = root->fs_info;
1396
1397         mutex_lock(&fs_info->trans_mutex);
1398         list_splice_init(&fs_info->dead_roots, &list);
1399         mutex_unlock(&fs_info->trans_mutex);
1400
1401         while (!list_empty(&list)) {
1402                 root = list_entry(list.next, struct btrfs_root, root_list);
1403                 list_del(&root->root_list);
1404
1405                 if (btrfs_header_backref_rev(root->node) <
1406                     BTRFS_MIXED_BACKREF_REV)
1407                         btrfs_drop_snapshot(root, NULL, 0);
1408                 else
1409                         btrfs_drop_snapshot(root, NULL, 1);
1410         }
1411         return 0;
1412 }