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(atomic_read(&transaction->use_count) == 0);
  36         if (atomic_dec_and_test(&transaction->use_count)) {
  37                 memset(transaction, 0, sizeof(*transaction));
  38                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  39         }
  40 }
  41
  42 static noinline void switch_commit_root(struct btrfs_root *root)
  43 {
  44         free_extent_buffer(root->commit_root);
  45         root->commit_root = btrfs_root_node(root);
  46 }
  47
  48 /*
  49  * either allocate a new transaction or hop into the existing one
  50  */
  51 static noinline int join_transaction(struct btrfs_root *root)
  52 {
  53         struct btrfs_transaction *cur_trans;
  54         cur_trans = root->fs_info->running_transaction;
  55         if (!cur_trans) {
  56                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  57                                              GFP_NOFS);
  58                 if (!cur_trans)
  59                         return -ENOMEM;
  60                 root->fs_info->generation++;
  61                 atomic_set(&cur_trans->num_writers, 1);
  62                 cur_trans->num_joined = 0;
  63                 cur_trans->transid = root->fs_info->generation;
  64                 init_waitqueue_head(&cur_trans->writer_wait);
  65                 init_waitqueue_head(&cur_trans->commit_wait);
  66                 cur_trans->in_commit = 0;
  67                 cur_trans->blocked = 0;
  68                 atomic_set(&cur_trans->use_count, 1);
  69                 cur_trans->commit_done = 0;
  70                 cur_trans->start_time = get_seconds();
  71
  72                 cur_trans->delayed_refs.root = RB_ROOT;
  73                 cur_trans->delayed_refs.num_entries = 0;
  74                 cur_trans->delayed_refs.num_heads_ready = 0;
  75                 cur_trans->delayed_refs.num_heads = 0;
  76                 cur_trans->delayed_refs.flushing = 0;
  77                 cur_trans->delayed_refs.run_delayed_start = 0;
  78                 spin_lock_init(&cur_trans->delayed_refs.lock);
  79
  80                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  81                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  82                 extent_io_tree_init(&cur_trans->dirty_pages,
  83                                      root->fs_info->btree_inode->i_mapping);
  84                 spin_lock(&root->fs_info->new_trans_lock);
  85                 root->fs_info->running_transaction = cur_trans;
  86                 spin_unlock(&root->fs_info->new_trans_lock);
  87         } else {
  88                 atomic_inc(&cur_trans->num_writers);
  89                 cur_trans->num_joined++;
  90         }
  91
  92         return 0;
  93 }
  94
  95 /*
  96  * this does all the record keeping required to make sure that a reference
  97  * counted root is properly recorded in a given transaction.  This is required
  98  * to make sure the old root from before we joined the transaction is deleted
  99  * when the transaction commits
 100  */
 101 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 102                                          struct btrfs_root *root)
 103 {
 104         if (root->ref_cows && root->last_trans < trans->transid) {
 105                 WARN_ON(root == root->fs_info->extent_root);
 106                 WARN_ON(root->commit_root != root->node);
 107
 108                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 109                            (unsigned long)root->root_key.objectid,
 110                            BTRFS_ROOT_TRANS_TAG);
 111                 root->last_trans = trans->transid;
 112                 btrfs_init_reloc_root(trans, root);
 113         }
 114         return 0;
 115 }
 116
 117 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 118                                struct btrfs_root *root)
 119 {
 120         if (!root->ref_cows)
 121                 return 0;
 122
 123         mutex_lock(&root->fs_info->trans_mutex);
 124         if (root->last_trans == trans->transid) {
 125                 mutex_unlock(&root->fs_info->trans_mutex);
 126                 return 0;
 127         }
 128
 129         record_root_in_trans(trans, root);
 130         mutex_unlock(&root->fs_info->trans_mutex);
 131         return 0;
 132 }
 133
 134 /* wait for commit against the current transaction to become unblocked
 135  * when this is done, it is safe to start a new transaction, but the current
 136  * transaction might not be fully on disk.
 137  */
 138 static void wait_current_trans(struct btrfs_root *root)
 139 {
 140         struct btrfs_transaction *cur_trans;
 141
 142         cur_trans = root->fs_info->running_transaction;
 143         if (cur_trans && cur_trans->blocked) {
 144                 DEFINE_WAIT(wait);
 145                 atomic_inc(&cur_trans->use_count);
 146                 while (1) {
 147                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 148                                         TASK_UNINTERRUPTIBLE);
 149                         if (!cur_trans->blocked)
 150                                 break;
 151                         mutex_unlock(&root->fs_info->trans_mutex);
 152                         schedule();
 153                         mutex_lock(&root->fs_info->trans_mutex);
 154                 }
 155                 finish_wait(&root->fs_info->transaction_wait, &wait);
 156                 put_transaction(cur_trans);
 157         }
 158 }
 159
 160 enum btrfs_trans_type {
 161         TRANS_START,
 162         TRANS_JOIN,
 163         TRANS_USERSPACE,
 164         TRANS_JOIN_NOLOCK,
 165 };
 166
 167 static int may_wait_transaction(struct btrfs_root *root, int type)
 168 {
 169         if (!root->fs_info->log_root_recovering &&
 170             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 171              type == TRANS_USERSPACE))
 172                 return 1;
 173         return 0;
 174 }
 175
 176 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 177                                                     u64 num_items, int type)
 178 {
 179         struct btrfs_trans_handle *h;
 180         struct btrfs_transaction *cur_trans;
 181         int retries = 0;
 182         int ret;
 183
 184         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 185                 return ERR_PTR(-EROFS);
 186 again:
 187         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 188         if (!h)
 189                 return ERR_PTR(-ENOMEM);
 190
 191         if (type != TRANS_JOIN_NOLOCK)
 192                 mutex_lock(&root->fs_info->trans_mutex);
 193         if (may_wait_transaction(root, type))
 194                 wait_current_trans(root);
 195
 196         ret = join_transaction(root);
 197         if (ret < 0) {
 198                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 199                 if (type != TRANS_JOIN_NOLOCK)
 200                         mutex_unlock(&root->fs_info->trans_mutex);
 201                 return ERR_PTR(ret);
 202         }
 203
 204         cur_trans = root->fs_info->running_transaction;
 205         atomic_inc(&cur_trans->use_count);
 206         if (type != TRANS_JOIN_NOLOCK)
 207                 mutex_unlock(&root->fs_info->trans_mutex);
 208
 209         h->transid = cur_trans->transid;
 210         h->transaction = cur_trans;
 211         h->blocks_used = 0;
 212         h->block_group = 0;
 213         h->bytes_reserved = 0;
 214         h->delayed_ref_updates = 0;
 215         h->block_rsv = NULL;
 216
 217         smp_mb();
 218         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 219                 btrfs_commit_transaction(h, root);
 220                 goto again;
 221         }
 222
 223         if (num_items > 0) {
 224                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 225                 if (ret == -EAGAIN && !retries) {
 226                         retries++;
 227                         btrfs_commit_transaction(h, root);
 228                         goto again;
 229                 } else if (ret == -EAGAIN) {
 230                         /*
 231                          * We have already retried and got EAGAIN, so really we
 232                          * don't have space, so set ret to -ENOSPC.
 233                          */
 234                         ret = -ENOSPC;
 235                 }
 236
 237                 if (ret < 0) {
 238                         btrfs_end_transaction(h, root);
 239                         return ERR_PTR(ret);
 240                 }
 241         }
 242
 243         if (type != TRANS_JOIN_NOLOCK)
 244                 mutex_lock(&root->fs_info->trans_mutex);
 245         record_root_in_trans(h, root);
 246         if (type != TRANS_JOIN_NOLOCK)
 247                 mutex_unlock(&root->fs_info->trans_mutex);
 248
 249         if (!current->journal_info && type != TRANS_USERSPACE)
 250                 current->journal_info = h;
 251         return h;
 252 }
 253
 254 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 255                                                    int num_items)
 256 {
 257         return start_transaction(root, num_items, TRANS_START);
 258 }
 259 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 260                                                    int num_blocks)
 261 {
 262         return start_transaction(root, 0, TRANS_JOIN);
 263 }
 264
 265 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 266                                                           int num_blocks)
 267 {
 268         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 269 }
 270
 271 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 272                                                          int num_blocks)
 273 {
 274         return start_transaction(r, 0, TRANS_USERSPACE);
 275 }
 276
 277 /* wait for a transaction commit to be fully complete */
 278 static noinline int wait_for_commit(struct btrfs_root *root,
 279                                     struct btrfs_transaction *commit)
 280 {
 281         DEFINE_WAIT(wait);
 282         mutex_lock(&root->fs_info->trans_mutex);
 283         while (!commit->commit_done) {
 284                 prepare_to_wait(&commit->commit_wait, &wait,
 285                                 TASK_UNINTERRUPTIBLE);
 286                 if (commit->commit_done)
 287                         break;
 288                 mutex_unlock(&root->fs_info->trans_mutex);
 289                 schedule();
 290                 mutex_lock(&root->fs_info->trans_mutex);
 291         }
 292         mutex_unlock(&root->fs_info->trans_mutex);
 293         finish_wait(&commit->commit_wait, &wait);
 294         return 0;
 295 }
 296
 297 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 298 {
 299         struct btrfs_transaction *cur_trans = NULL, *t;
 300         int ret;
 301
 302         mutex_lock(&root->fs_info->trans_mutex);
 303
 304         ret = 0;
 305         if (transid) {
 306                 if (transid <= root->fs_info->last_trans_committed)
 307                         goto out_unlock;
 308
 309                 /* find specified transaction */
 310                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 311                         if (t->transid == transid) {
 312                                 cur_trans = t;
 313                                 break;
 314                         }
 315                         if (t->transid > transid)
 316                                 break;
 317                 }
 318                 ret = -EINVAL;
 319                 if (!cur_trans)
 320                         goto out_unlock;  /* bad transid */
 321         } else {
 322                 /* find newest transaction that is committing | committed */
 323                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 324                                             list) {
 325                         if (t->in_commit) {
 326                                 if (t->commit_done)
 327                                         goto out_unlock;
 328                                 cur_trans = t;
 329                                 break;
 330                         }
 331                 }
 332                 if (!cur_trans)
 333                         goto out_unlock;  /* nothing committing|committed */
 334         }
 335
 336         atomic_inc(&cur_trans->use_count);
 337         mutex_unlock(&root->fs_info->trans_mutex);
 338
 339         wait_for_commit(root, cur_trans);
 340
 341         mutex_lock(&root->fs_info->trans_mutex);
 342         put_transaction(cur_trans);
 343         ret = 0;
 344 out_unlock:
 345         mutex_unlock(&root->fs_info->trans_mutex);
 346         return ret;
 347 }
 348
 349 #if 0
 350 /*
 351  * rate limit against the drop_snapshot code.  This helps to slow down new
 352  * operations if the drop_snapshot code isn't able to keep up.
 353  */
 354 static void throttle_on_drops(struct btrfs_root *root)
 355 {
 356         struct btrfs_fs_info *info = root->fs_info;
 357         int harder_count = 0;
 358
 359 harder:
 360         if (atomic_read(&info->throttles)) {
 361                 DEFINE_WAIT(wait);
 362                 int thr;
 363                 thr = atomic_read(&info->throttle_gen);
 364
 365                 do {
 366                         prepare_to_wait(&info->transaction_throttle,
 367                                         &wait, TASK_UNINTERRUPTIBLE);
 368                         if (!atomic_read(&info->throttles)) {
 369                                 finish_wait(&info->transaction_throttle, &wait);
 370                                 break;
 371                         }
 372                         schedule();
 373                         finish_wait(&info->transaction_throttle, &wait);
 374                 } while (thr == atomic_read(&info->throttle_gen));
 375                 harder_count++;
 376
 377                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 378                     harder_count < 2)
 379                         goto harder;
 380
 381                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 382                     harder_count < 10)
 383                         goto harder;
 384
 385                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 386                     harder_count < 20)
 387                         goto harder;
 388         }
 389 }
 390 #endif
 391
 392 void btrfs_throttle(struct btrfs_root *root)
 393 {
 394         mutex_lock(&root->fs_info->trans_mutex);
 395         if (!root->fs_info->open_ioctl_trans)
 396                 wait_current_trans(root);
 397         mutex_unlock(&root->fs_info->trans_mutex);
 398 }
 399
 400 static int should_end_transaction(struct btrfs_trans_handle *trans,
 401                                   struct btrfs_root *root)
 402 {
 403         int ret;
 404         ret = btrfs_block_rsv_check(trans, root,
 405                                     &root->fs_info->global_block_rsv, 0, 5);
 406         return ret ? 1 : 0;
 407 }
 408
 409 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 410                                  struct btrfs_root *root)
 411 {
 412         struct btrfs_transaction *cur_trans = trans->transaction;
 413         int updates;
 414
 415         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 416                 return 1;
 417
 418         updates = trans->delayed_ref_updates;
 419         trans->delayed_ref_updates = 0;
 420         if (updates)
 421                 btrfs_run_delayed_refs(trans, root, updates);
 422
 423         return should_end_transaction(trans, root);
 424 }
 425
 426 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 427                           struct btrfs_root *root, int throttle, int lock)
 428 {
 429         struct btrfs_transaction *cur_trans = trans->transaction;
 430         struct btrfs_fs_info *info = root->fs_info;
 431         int count = 0;
 432
 433         while (count < 4) {
 434                 unsigned long cur = trans->delayed_ref_updates;
 435                 trans->delayed_ref_updates = 0;
 436                 if (cur &&
 437                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 438                         trans->delayed_ref_updates = 0;
 439
 440                         /*
 441                          * do a full flush if the transaction is trying
 442                          * to close
 443                          */
 444                         if (trans->transaction->delayed_refs.flushing)
 445                                 cur = 0;
 446                         btrfs_run_delayed_refs(trans, root, cur);
 447                 } else {
 448                         break;
 449                 }
 450                 count++;
 451         }
 452
 453         btrfs_trans_release_metadata(trans, root);
 454
 455         if (lock && !root->fs_info->open_ioctl_trans &&
 456             should_end_transaction(trans, root))
 457                 trans->transaction->blocked = 1;
 458
 459         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 460                 if (throttle)
 461                         return btrfs_commit_transaction(trans, root);
 462                 else
 463                         wake_up_process(info->transaction_kthread);
 464         }
 465
 466         WARN_ON(cur_trans != info->running_transaction);
 467         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 468         atomic_dec(&cur_trans->num_writers);
 469
 470         smp_mb();
 471         if (waitqueue_active(&cur_trans->writer_wait))
 472                 wake_up(&cur_trans->writer_wait);
 473         put_transaction(cur_trans);
 474
 475         if (current->journal_info == trans)
 476                 current->journal_info = NULL;
 477         memset(trans, 0, sizeof(*trans));
 478         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 479
 480         if (throttle)
 481                 btrfs_run_delayed_iputs(root);
 482
 483         return 0;
 484 }
 485
 486 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 487                           struct btrfs_root *root)
 488 {
 489         return __btrfs_end_transaction(trans, root, 0, 1);
 490 }
 491
 492 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 493                                    struct btrfs_root *root)
 494 {
 495         return __btrfs_end_transaction(trans, root, 1, 1);
 496 }
 497
 498 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 499                                  struct btrfs_root *root)
 500 {
 501         return __btrfs_end_transaction(trans, root, 0, 0);
 502 }
 503
 504 /*
 505  * when btree blocks are allocated, they have some corresponding bits set for
 506  * them in one of two extent_io trees.  This is used to make sure all of
 507  * those extents are sent to disk but does not wait on them
 508  */
 509 int btrfs_write_marked_extents(struct btrfs_root *root,
 510                                struct extent_io_tree *dirty_pages, int mark)
 511 {
 512         int ret;
 513         int err = 0;
 514         int werr = 0;
 515         struct page *page;
 516         struct inode *btree_inode = root->fs_info->btree_inode;
 517         u64 start = 0;
 518         u64 end;
 519         unsigned long index;
 520
 521         while (1) {
 522                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 523                                             mark);
 524                 if (ret)
 525                         break;
 526                 while (start <= end) {
 527                         cond_resched();
 528
 529                         index = start >> PAGE_CACHE_SHIFT;
 530                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 531                         page = find_get_page(btree_inode->i_mapping, index);
 532                         if (!page)
 533                                 continue;
 534
 535                         btree_lock_page_hook(page);
 536                         if (!page->mapping) {
 537                                 unlock_page(page);
 538                                 page_cache_release(page);
 539                                 continue;
 540                         }
 541
 542                         if (PageWriteback(page)) {
 543                                 if (PageDirty(page))
 544                                         wait_on_page_writeback(page);
 545                                 else {
 546                                         unlock_page(page);
 547                                         page_cache_release(page);
 548                                         continue;
 549                                 }
 550                         }
 551                         err = write_one_page(page, 0);
 552                         if (err)
 553                                 werr = err;
 554                         page_cache_release(page);
 555                 }
 556         }
 557         if (err)
 558                 werr = err;
 559         return werr;
 560 }
 561
 562 /*
 563  * when btree blocks are allocated, they have some corresponding bits set for
 564  * them in one of two extent_io trees.  This is used to make sure all of
 565  * those extents are on disk for transaction or log commit.  We wait
 566  * on all the pages and clear them from the dirty pages state tree
 567  */
 568 int btrfs_wait_marked_extents(struct btrfs_root *root,
 569                               struct extent_io_tree *dirty_pages, int mark)
 570 {
 571         int ret;
 572         int err = 0;
 573         int werr = 0;
 574         struct page *page;
 575         struct inode *btree_inode = root->fs_info->btree_inode;
 576         u64 start = 0;
 577         u64 end;
 578         unsigned long index;
 579
 580         while (1) {
 581                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 582                                             mark);
 583                 if (ret)
 584                         break;
 585
 586                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 587                 while (start <= end) {
 588                         index = start >> PAGE_CACHE_SHIFT;
 589                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 590                         page = find_get_page(btree_inode->i_mapping, index);
 591                         if (!page)
 592                                 continue;
 593                         if (PageDirty(page)) {
 594                                 btree_lock_page_hook(page);
 595                                 wait_on_page_writeback(page);
 596                                 err = write_one_page(page, 0);
 597                                 if (err)
 598                                         werr = err;
 599                         }
 600                         wait_on_page_writeback(page);
 601                         page_cache_release(page);
 602                         cond_resched();
 603                 }
 604         }
 605         if (err)
 606                 werr = err;
 607         return werr;
 608 }
 609
 610 /*
 611  * when btree blocks are allocated, they have some corresponding bits set for
 612  * them in one of two extent_io trees.  This is used to make sure all of
 613  * those extents are on disk for transaction or log commit
 614  */
 615 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 616                                 struct extent_io_tree *dirty_pages, int mark)
 617 {
 618         int ret;
 619         int ret2;
 620
 621         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 622         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 623         return ret || ret2;
 624 }
 625
 626 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 627                                      struct btrfs_root *root)
 628 {
 629         if (!trans || !trans->transaction) {
 630                 struct inode *btree_inode;
 631                 btree_inode = root->fs_info->btree_inode;
 632                 return filemap_write_and_wait(btree_inode->i_mapping);
 633         }
 634         return btrfs_write_and_wait_marked_extents(root,
 635                                            &trans->transaction->dirty_pages,
 636                                            EXTENT_DIRTY);
 637 }
 638
 639 /*
 640  * this is used to update the root pointer in the tree of tree roots.
 641  *
 642  * But, in the case of the extent allocation tree, updating the root
 643  * pointer may allocate blocks which may change the root of the extent
 644  * allocation tree.
 645  *
 646  * So, this loops and repeats and makes sure the cowonly root didn't
 647  * change while the root pointer was being updated in the metadata.
 648  */
 649 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 650                                struct btrfs_root *root)
 651 {
 652         int ret;
 653         u64 old_root_bytenr;
 654         u64 old_root_used;
 655         struct btrfs_root *tree_root = root->fs_info->tree_root;
 656
 657         old_root_used = btrfs_root_used(&root->root_item);
 658         btrfs_write_dirty_block_groups(trans, root);
 659
 660         while (1) {
 661                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 662                 if (old_root_bytenr == root->node->start &&
 663                     old_root_used == btrfs_root_used(&root->root_item))
 664                         break;
 665
 666                 btrfs_set_root_node(&root->root_item, root->node);
 667                 ret = btrfs_update_root(trans, tree_root,
 668                                         &root->root_key,
 669                                         &root->root_item);
 670                 BUG_ON(ret);
 671
 672                 old_root_used = btrfs_root_used(&root->root_item);
 673                 ret = btrfs_write_dirty_block_groups(trans, root);
 674                 BUG_ON(ret);
 675         }
 676
 677         if (root != root->fs_info->extent_root)
 678                 switch_commit_root(root);
 679
 680         return 0;
 681 }
 682
 683 /*
 684  * update all the cowonly tree roots on disk
 685  */
 686 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 687                                          struct btrfs_root *root)
 688 {
 689         struct btrfs_fs_info *fs_info = root->fs_info;
 690         struct list_head *next;
 691         struct extent_buffer *eb;
 692         int ret;
 693
 694         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 695         BUG_ON(ret);
 696
 697         eb = btrfs_lock_root_node(fs_info->tree_root);
 698         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 699         btrfs_tree_unlock(eb);
 700         free_extent_buffer(eb);
 701
 702         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 703         BUG_ON(ret);
 704
 705         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 706                 next = fs_info->dirty_cowonly_roots.next;
 707                 list_del_init(next);
 708                 root = list_entry(next, struct btrfs_root, dirty_list);
 709
 710                 update_cowonly_root(trans, root);
 711         }
 712
 713         down_write(&fs_info->extent_commit_sem);
 714         switch_commit_root(fs_info->extent_root);
 715         up_write(&fs_info->extent_commit_sem);
 716
 717         return 0;
 718 }
 719
 720 /*
 721  * dead roots are old snapshots that need to be deleted.  This allocates
 722  * a dirty root struct and adds it into the list of dead roots that need to
 723  * be deleted
 724  */
 725 int btrfs_add_dead_root(struct btrfs_root *root)
 726 {
 727         mutex_lock(&root->fs_info->trans_mutex);
 728         list_add(&root->root_list, &root->fs_info->dead_roots);
 729         mutex_unlock(&root->fs_info->trans_mutex);
 730         return 0;
 731 }
 732
 733 /*
 734  * update all the cowonly tree roots on disk
 735  */
 736 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 737                                     struct btrfs_root *root)
 738 {
 739         struct btrfs_root *gang[8];
 740         struct btrfs_fs_info *fs_info = root->fs_info;
 741         int i;
 742         int ret;
 743         int err = 0;
 744
 745         while (1) {
 746                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 747                                                  (void **)gang, 0,
 748                                                  ARRAY_SIZE(gang),
 749                                                  BTRFS_ROOT_TRANS_TAG);
 750                 if (ret == 0)
 751                         break;
 752                 for (i = 0; i < ret; i++) {
 753                         root = gang[i];
 754                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 755                                         (unsigned long)root->root_key.objectid,
 756                                         BTRFS_ROOT_TRANS_TAG);
 757
 758                         btrfs_free_log(trans, root);
 759                         btrfs_update_reloc_root(trans, root);
 760                         btrfs_orphan_commit_root(trans, root);
 761
 762                         if (root->commit_root != root->node) {
 763                                 switch_commit_root(root);
 764                                 btrfs_set_root_node(&root->root_item,
 765                                                     root->node);
 766                         }
 767
 768                         err = btrfs_update_root(trans, fs_info->tree_root,
 769                                                 &root->root_key,
 770                                                 &root->root_item);
 771                         if (err)
 772                                 break;
 773                 }
 774         }
 775         return err;
 776 }
 777
 778 /*
 779  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 780  * otherwise every leaf in the btree is read and defragged.
 781  */
 782 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 783 {
 784         struct btrfs_fs_info *info = root->fs_info;
 785         struct btrfs_trans_handle *trans;
 786         int ret;
 787         unsigned long nr;
 788
 789         if (xchg(&root->defrag_running, 1))
 790                 return 0;
 791
 792         while (1) {
 793                 trans = btrfs_start_transaction(root, 0);
 794                 if (IS_ERR(trans))
 795                         return PTR_ERR(trans);
 796
 797                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 798
 799                 nr = trans->blocks_used;
 800                 btrfs_end_transaction(trans, root);
 801                 btrfs_btree_balance_dirty(info->tree_root, nr);
 802                 cond_resched();
 803
 804                 if (root->fs_info->closing || ret != -EAGAIN)
 805                         break;
 806         }
 807         root->defrag_running = 0;
 808         return ret;
 809 }
 810
 811 #if 0
 812 /*
 813  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 814  * sense not to join the transaction while it is trying to flush the current
 815  * queue of delayed refs out.
 816  *
 817  * This is used by the drop snapshot code only
 818  */
 819 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 820 {
 821         DEFINE_WAIT(wait);
 822
 823         mutex_lock(&info->trans_mutex);
 824         while (info->running_transaction &&
 825                info->running_transaction->delayed_refs.flushing) {
 826                 prepare_to_wait(&info->transaction_wait, &wait,
 827                                 TASK_UNINTERRUPTIBLE);
 828                 mutex_unlock(&info->trans_mutex);
 829
 830                 schedule();
 831
 832                 mutex_lock(&info->trans_mutex);
 833                 finish_wait(&info->transaction_wait, &wait);
 834         }
 835         mutex_unlock(&info->trans_mutex);
 836         return 0;
 837 }
 838
 839 /*
 840  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 841  * all of them
 842  */
 843 int btrfs_drop_dead_root(struct btrfs_root *root)
 844 {
 845         struct btrfs_trans_handle *trans;
 846         struct btrfs_root *tree_root = root->fs_info->tree_root;
 847         unsigned long nr;
 848         int ret;
 849
 850         while (1) {
 851                 /*
 852                  * we don't want to jump in and create a bunch of
 853                  * delayed refs if the transaction is starting to close
 854                  */
 855                 wait_transaction_pre_flush(tree_root->fs_info);
 856                 trans = btrfs_start_transaction(tree_root, 1);
 857
 858                 /*
 859                  * we've joined a transaction, make sure it isn't
 860                  * closing right now
 861                  */
 862                 if (trans->transaction->delayed_refs.flushing) {
 863                         btrfs_end_transaction(trans, tree_root);
 864                         continue;
 865                 }
 866
 867                 ret = btrfs_drop_snapshot(trans, root);
 868                 if (ret != -EAGAIN)
 869                         break;
 870
 871                 ret = btrfs_update_root(trans, tree_root,
 872                                         &root->root_key,
 873                                         &root->root_item);
 874                 if (ret)
 875                         break;
 876
 877                 nr = trans->blocks_used;
 878                 ret = btrfs_end_transaction(trans, tree_root);
 879                 BUG_ON(ret);
 880
 881                 btrfs_btree_balance_dirty(tree_root, nr);
 882                 cond_resched();
 883         }
 884         BUG_ON(ret);
 885
 886         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 887         BUG_ON(ret);
 888
 889         nr = trans->blocks_used;
 890         ret = btrfs_end_transaction(trans, tree_root);
 891         BUG_ON(ret);
 892
 893         free_extent_buffer(root->node);
 894         free_extent_buffer(root->commit_root);
 895         kfree(root);
 896
 897         btrfs_btree_balance_dirty(tree_root, nr);
 898         return ret;
 899 }
 900 #endif
 901
 902 /*
 903  * new snapshots need to be created at a very specific time in the
 904  * transaction commit.  This does the actual creation
 905  */
 906 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 907                                    struct btrfs_fs_info *fs_info,
 908                                    struct btrfs_pending_snapshot *pending)
 909 {
 910         struct btrfs_key key;
 911         struct btrfs_root_item *new_root_item;
 912         struct btrfs_root *tree_root = fs_info->tree_root;
 913         struct btrfs_root *root = pending->root;
 914         struct btrfs_root *parent_root;
 915         struct inode *parent_inode;
 916         struct dentry *parent;
 917         struct dentry *dentry;
 918         struct extent_buffer *tmp;
 919         struct extent_buffer *old;
 920         int ret;
 921         u64 to_reserve = 0;
 922         u64 index = 0;
 923         u64 objectid;
 924         u64 root_flags;
 925
 926         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 927         if (!new_root_item) {
 928                 pending->error = -ENOMEM;
 929                 goto fail;
 930         }
 931
 932         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 933         if (ret) {
 934                 pending->error = ret;
 935                 goto fail;
 936         }
 937
 938         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 939         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 940
 941         if (to_reserve > 0) {
 942                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 943                                           to_reserve);
 944                 if (ret) {
 945                         pending->error = ret;
 946                         goto fail;
 947                 }
 948         }
 949
 950         key.objectid = objectid;
 951         key.offset = (u64)-1;
 952         key.type = BTRFS_ROOT_ITEM_KEY;
 953
 954         trans->block_rsv = &pending->block_rsv;
 955
 956         dentry = pending->dentry;
 957         parent = dget_parent(dentry);
 958         parent_inode = parent->d_inode;
 959         parent_root = BTRFS_I(parent_inode)->root;
 960         record_root_in_trans(trans, parent_root);
 961
 962         /*
 963          * insert the directory item
 964          */
 965         ret = btrfs_set_inode_index(parent_inode, &index);
 966         BUG_ON(ret);
 967         ret = btrfs_insert_dir_item(trans, parent_root,
 968                                 dentry->d_name.name, dentry->d_name.len,
 969                                 parent_inode->i_ino, &key,
 970                                 BTRFS_FT_DIR, index);
 971         BUG_ON(ret);
 972
 973         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 974                                          dentry->d_name.len * 2);
 975         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 976         BUG_ON(ret);
 977
 978         record_root_in_trans(trans, root);
 979         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 980         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 981         btrfs_check_and_init_root_item(new_root_item);
 982
 983         root_flags = btrfs_root_flags(new_root_item);
 984         if (pending->readonly)
 985                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 986         else
 987                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 988         btrfs_set_root_flags(new_root_item, root_flags);
 989
 990         old = btrfs_lock_root_node(root);
 991         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 992         btrfs_set_lock_blocking(old);
 993
 994         btrfs_copy_root(trans, root, old, &tmp, objectid);
 995         btrfs_tree_unlock(old);
 996         free_extent_buffer(old);
 997
 998         btrfs_set_root_node(new_root_item, tmp);
 999         /* record when the snapshot was created in key.offset */
1000         key.offset = trans->transid;
1001         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1002         btrfs_tree_unlock(tmp);
1003         free_extent_buffer(tmp);
1004         BUG_ON(ret);
1005
1006         /*
1007          * insert root back/forward references
1008          */
1009         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1010                                  parent_root->root_key.objectid,
1011                                  parent_inode->i_ino, index,
1012                                  dentry->d_name.name, dentry->d_name.len);
1013         BUG_ON(ret);
1014         dput(parent);
1015
1016         key.offset = (u64)-1;
1017         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1018         BUG_ON(IS_ERR(pending->snap));
1019
1020         btrfs_reloc_post_snapshot(trans, pending);
1021         btrfs_orphan_post_snapshot(trans, pending);
1022 fail:
1023         kfree(new_root_item);
1024         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1025         return 0;
1026 }
1027
1028 /*
1029  * create all the snapshots we've scheduled for creation
1030  */
1031 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1032                                              struct btrfs_fs_info *fs_info)
1033 {
1034         struct btrfs_pending_snapshot *pending;
1035         struct list_head *head = &trans->transaction->pending_snapshots;
1036         int ret;
1037
1038         list_for_each_entry(pending, head, list) {
1039                 ret = create_pending_snapshot(trans, fs_info, pending);
1040                 BUG_ON(ret);
1041         }
1042         return 0;
1043 }
1044
1045 static void update_super_roots(struct btrfs_root *root)
1046 {
1047         struct btrfs_root_item *root_item;
1048         struct btrfs_super_block *super;
1049
1050         super = &root->fs_info->super_copy;
1051
1052         root_item = &root->fs_info->chunk_root->root_item;
1053         super->chunk_root = root_item->bytenr;
1054         super->chunk_root_generation = root_item->generation;
1055         super->chunk_root_level = root_item->level;
1056
1057         root_item = &root->fs_info->tree_root->root_item;
1058         super->root = root_item->bytenr;
1059         super->generation = root_item->generation;
1060         super->root_level = root_item->level;
1061         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1062                 super->cache_generation = root_item->generation;
1063 }
1064
1065 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1066 {
1067         int ret = 0;
1068         spin_lock(&info->new_trans_lock);
1069         if (info->running_transaction)
1070                 ret = info->running_transaction->in_commit;
1071         spin_unlock(&info->new_trans_lock);
1072         return ret;
1073 }
1074
1075 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1076 {
1077         int ret = 0;
1078         spin_lock(&info->new_trans_lock);
1079         if (info->running_transaction)
1080                 ret = info->running_transaction->blocked;
1081         spin_unlock(&info->new_trans_lock);
1082         return ret;
1083 }
1084
1085 /*
1086  * wait for the current transaction commit to start and block subsequent
1087  * transaction joins
1088  */
1089 static void wait_current_trans_commit_start(struct btrfs_root *root,
1090                                             struct btrfs_transaction *trans)
1091 {
1092         DEFINE_WAIT(wait);
1093
1094         if (trans->in_commit)
1095                 return;
1096
1097         while (1) {
1098                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1099                                 TASK_UNINTERRUPTIBLE);
1100                 if (trans->in_commit) {
1101                         finish_wait(&root->fs_info->transaction_blocked_wait,
1102                                     &wait);
1103                         break;
1104                 }
1105                 mutex_unlock(&root->fs_info->trans_mutex);
1106                 schedule();
1107                 mutex_lock(&root->fs_info->trans_mutex);
1108                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1109         }
1110 }
1111
1112 /*
1113  * wait for the current transaction to start and then become unblocked.
1114  * caller holds ref.
1115  */
1116 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1117                                          struct btrfs_transaction *trans)
1118 {
1119         DEFINE_WAIT(wait);
1120
1121         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1122                 return;
1123
1124         while (1) {
1125                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1126                                 TASK_UNINTERRUPTIBLE);
1127                 if (trans->commit_done ||
1128                     (trans->in_commit && !trans->blocked)) {
1129                         finish_wait(&root->fs_info->transaction_wait,
1130                                     &wait);
1131                         break;
1132                 }
1133                 mutex_unlock(&root->fs_info->trans_mutex);
1134                 schedule();
1135                 mutex_lock(&root->fs_info->trans_mutex);
1136                 finish_wait(&root->fs_info->transaction_wait,
1137                             &wait);
1138         }
1139 }
1140
1141 /*
1142  * commit transactions asynchronously. once btrfs_commit_transaction_async
1143  * returns, any subsequent transaction will not be allowed to join.
1144  */
1145 struct btrfs_async_commit {
1146         struct btrfs_trans_handle *newtrans;
1147         struct btrfs_root *root;
1148         struct delayed_work work;
1149 };
1150
1151 static void do_async_commit(struct work_struct *work)
1152 {
1153         struct btrfs_async_commit *ac =
1154                 container_of(work, struct btrfs_async_commit, work.work);
1155
1156         btrfs_commit_transaction(ac->newtrans, ac->root);
1157         kfree(ac);
1158 }
1159
1160 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1161                                    struct btrfs_root *root,
1162                                    int wait_for_unblock)
1163 {
1164         struct btrfs_async_commit *ac;
1165         struct btrfs_transaction *cur_trans;
1166
1167         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1168         if (!ac)
1169                 return -ENOMEM;
1170
1171         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1172         ac->root = root;
1173         ac->newtrans = btrfs_join_transaction(root, 0);
1174         if (IS_ERR(ac->newtrans)) {
1175                 int err = PTR_ERR(ac->newtrans);
1176                 kfree(ac);
1177                 return err;
1178         }
1179
1180         /* take transaction reference */
1181         mutex_lock(&root->fs_info->trans_mutex);
1182         cur_trans = trans->transaction;
1183         atomic_inc(&cur_trans->use_count);
1184         mutex_unlock(&root->fs_info->trans_mutex);
1185
1186         btrfs_end_transaction(trans, root);
1187         schedule_delayed_work(&ac->work, 0);
1188
1189         /* wait for transaction to start and unblock */
1190         mutex_lock(&root->fs_info->trans_mutex);
1191         if (wait_for_unblock)
1192                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1193         else
1194                 wait_current_trans_commit_start(root, cur_trans);
1195         put_transaction(cur_trans);
1196         mutex_unlock(&root->fs_info->trans_mutex);
1197
1198         return 0;
1199 }
1200
1201 /*
1202  * btrfs_transaction state sequence:
1203  *    in_commit = 0, blocked = 0  (initial)
1204  *    in_commit = 1, blocked = 1
1205  *    blocked = 0
1206  *    commit_done = 1
1207  */
1208 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1209                              struct btrfs_root *root)
1210 {
1211         unsigned long joined = 0;
1212         struct btrfs_transaction *cur_trans;
1213         struct btrfs_transaction *prev_trans = NULL;
1214         DEFINE_WAIT(wait);
1215         int ret;
1216         int should_grow = 0;
1217         unsigned long now = get_seconds();
1218         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1219
1220         btrfs_run_ordered_operations(root, 0);
1221
1222         /* make a pass through all the delayed refs we have so far
1223          * any runnings procs may add more while we are here
1224          */
1225         ret = btrfs_run_delayed_refs(trans, root, 0);
1226         BUG_ON(ret);
1227
1228         btrfs_trans_release_metadata(trans, root);
1229
1230         cur_trans = trans->transaction;
1231         /*
1232          * set the flushing flag so procs in this transaction have to
1233          * start sending their work down.
1234          */
1235         cur_trans->delayed_refs.flushing = 1;
1236
1237         ret = btrfs_run_delayed_refs(trans, root, 0);
1238         BUG_ON(ret);
1239
1240         mutex_lock(&root->fs_info->trans_mutex);
1241         if (cur_trans->in_commit) {
1242                 atomic_inc(&cur_trans->use_count);
1243                 mutex_unlock(&root->fs_info->trans_mutex);
1244                 btrfs_end_transaction(trans, root);
1245
1246                 ret = wait_for_commit(root, cur_trans);
1247                 BUG_ON(ret);
1248
1249                 mutex_lock(&root->fs_info->trans_mutex);
1250                 put_transaction(cur_trans);
1251                 mutex_unlock(&root->fs_info->trans_mutex);
1252
1253                 return 0;
1254         }
1255
1256         trans->transaction->in_commit = 1;
1257         trans->transaction->blocked = 1;
1258         wake_up(&root->fs_info->transaction_blocked_wait);
1259
1260         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1261                 prev_trans = list_entry(cur_trans->list.prev,
1262                                         struct btrfs_transaction, list);
1263                 if (!prev_trans->commit_done) {
1264                         atomic_inc(&prev_trans->use_count);
1265                         mutex_unlock(&root->fs_info->trans_mutex);
1266
1267                         wait_for_commit(root, prev_trans);
1268
1269                         mutex_lock(&root->fs_info->trans_mutex);
1270                         put_transaction(prev_trans);
1271                 }
1272         }
1273
1274         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1275                 should_grow = 1;
1276
1277         do {
1278                 int snap_pending = 0;
1279                 joined = cur_trans->num_joined;
1280                 if (!list_empty(&trans->transaction->pending_snapshots))
1281                         snap_pending = 1;
1282
1283                 WARN_ON(cur_trans != trans->transaction);
1284                 mutex_unlock(&root->fs_info->trans_mutex);
1285
1286                 if (flush_on_commit || snap_pending) {
1287                         btrfs_start_delalloc_inodes(root, 1);
1288                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1289                         BUG_ON(ret);
1290                 }
1291
1292                 /*
1293                  * rename don't use btrfs_join_transaction, so, once we
1294                  * set the transaction to blocked above, we aren't going
1295                  * to get any new ordered operations.  We can safely run
1296                  * it here and no for sure that nothing new will be added
1297                  * to the list
1298                  */
1299                 btrfs_run_ordered_operations(root, 1);
1300
1301                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1302                                 TASK_UNINTERRUPTIBLE);
1303
1304                 smp_mb();
1305                 if (atomic_read(&cur_trans->num_writers) > 1)
1306                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1307                 else if (should_grow)
1308                         schedule_timeout(1);
1309
1310                 mutex_lock(&root->fs_info->trans_mutex);
1311                 finish_wait(&cur_trans->writer_wait, &wait);
1312         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1313                  (should_grow && cur_trans->num_joined != joined));
1314
1315         ret = create_pending_snapshots(trans, root->fs_info);
1316         BUG_ON(ret);
1317
1318         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1319         BUG_ON(ret);
1320
1321         WARN_ON(cur_trans != trans->transaction);
1322
1323         /* btrfs_commit_tree_roots is responsible for getting the
1324          * various roots consistent with each other.  Every pointer
1325          * in the tree of tree roots has to point to the most up to date
1326          * root for every subvolume and other tree.  So, we have to keep
1327          * the tree logging code from jumping in and changing any
1328          * of the trees.
1329          *
1330          * At this point in the commit, there can't be any tree-log
1331          * writers, but a little lower down we drop the trans mutex
1332          * and let new people in.  By holding the tree_log_mutex
1333          * from now until after the super is written, we avoid races
1334          * with the tree-log code.
1335          */
1336         mutex_lock(&root->fs_info->tree_log_mutex);
1337
1338         ret = commit_fs_roots(trans, root);
1339         BUG_ON(ret);
1340
1341         /* commit_fs_roots gets rid of all the tree log roots, it is now
1342          * safe to free the root of tree log roots
1343          */
1344         btrfs_free_log_root_tree(trans, root->fs_info);
1345
1346         ret = commit_cowonly_roots(trans, root);
1347         BUG_ON(ret);
1348
1349         btrfs_prepare_extent_commit(trans, root);
1350
1351         cur_trans = root->fs_info->running_transaction;
1352         spin_lock(&root->fs_info->new_trans_lock);
1353         root->fs_info->running_transaction = NULL;
1354         spin_unlock(&root->fs_info->new_trans_lock);
1355
1356         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1357                             root->fs_info->tree_root->node);
1358         switch_commit_root(root->fs_info->tree_root);
1359
1360         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1361                             root->fs_info->chunk_root->node);
1362         switch_commit_root(root->fs_info->chunk_root);
1363
1364         update_super_roots(root);
1365
1366         if (!root->fs_info->log_root_recovering) {
1367                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1368                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1369         }
1370
1371         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1372                sizeof(root->fs_info->super_copy));
1373
1374         trans->transaction->blocked = 0;
1375
1376         wake_up(&root->fs_info->transaction_wait);
1377
1378         mutex_unlock(&root->fs_info->trans_mutex);
1379         ret = btrfs_write_and_wait_transaction(trans, root);
1380         BUG_ON(ret);
1381         write_ctree_super(trans, root, 0);
1382
1383         /*
1384          * the super is written, we can safely allow the tree-loggers
1385          * to go about their business
1386          */
1387         mutex_unlock(&root->fs_info->tree_log_mutex);
1388
1389         btrfs_finish_extent_commit(trans, root);
1390
1391         mutex_lock(&root->fs_info->trans_mutex);
1392
1393         cur_trans->commit_done = 1;
1394
1395         root->fs_info->last_trans_committed = cur_trans->transid;
1396
1397         wake_up(&cur_trans->commit_wait);
1398
1399         list_del_init(&cur_trans->list);
1400         put_transaction(cur_trans);
1401         put_transaction(cur_trans);
1402
1403         trace_btrfs_transaction_commit(root);
1404
1405         mutex_unlock(&root->fs_info->trans_mutex);
1406
1407         if (current->journal_info == trans)
1408                 current->journal_info = NULL;
1409
1410         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1411
1412         if (current != root->fs_info->transaction_kthread)
1413                 btrfs_run_delayed_iputs(root);
1414
1415         return ret;
1416 }
1417
1418 /*
1419  * interface function to delete all the snapshots we have scheduled for deletion
1420  */
1421 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1422 {
1423         LIST_HEAD(list);
1424         struct btrfs_fs_info *fs_info = root->fs_info;
1425
1426         mutex_lock(&fs_info->trans_mutex);
1427         list_splice_init(&fs_info->dead_roots, &list);
1428         mutex_unlock(&fs_info->trans_mutex);
1429
1430         while (!list_empty(&list)) {
1431                 root = list_entry(list.next, struct btrfs_root, root_list);
1432                 list_del(&root->root_list);
1433
1434                 if (btrfs_header_backref_rev(root->node) <
1435                     BTRFS_MIXED_BACKREF_REV)
1436                         btrfs_drop_snapshot(root, NULL, 0);
1437                 else
1438                         btrfs_drop_snapshot(root, NULL, 1);
1439         }
1440         return 0;
1441 }