fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30 #include "inode-map.h"
  31
  32 #define BTRFS_ROOT_TRANS_TAG 0
  33
  34 void put_transaction(struct btrfs_transaction *transaction)
  35 {
  36         WARN_ON(atomic_read(&transaction->use_count) == 0);
  37         if (atomic_dec_and_test(&transaction->use_count)) {
  38                 BUG_ON(!list_empty(&transaction->list));
  39                 WARN_ON(transaction->delayed_refs.root.rb_node);
  40                 WARN_ON(!list_empty(&transaction->delayed_refs.seq_head));
  41                 memset(transaction, 0, sizeof(*transaction));
  42                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  43         }
  44 }
  45
  46 static noinline void switch_commit_root(struct btrfs_root *root)
  47 {
  48         free_extent_buffer(root->commit_root);
  49         root->commit_root = btrfs_root_node(root);
  50 }
  51
  52 /*
  53  * either allocate a new transaction or hop into the existing one
  54  */
  55 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  56 {
  57         struct btrfs_transaction *cur_trans;
  58
  59         spin_lock(&root->fs_info->trans_lock);
  60 loop:
  61         /* The file system has been taken offline. No new transactions. */
  62         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  63                 spin_unlock(&root->fs_info->trans_lock);
  64                 return -EROFS;
  65         }
  66
  67         if (root->fs_info->trans_no_join) {
  68                 if (!nofail) {
  69                         spin_unlock(&root->fs_info->trans_lock);
  70                         return -EBUSY;
  71                 }
  72         }
  73
  74         cur_trans = root->fs_info->running_transaction;
  75         if (cur_trans) {
  76                 if (cur_trans->aborted) {
  77                         spin_unlock(&root->fs_info->trans_lock);
  78                         return cur_trans->aborted;
  79                 }
  80                 atomic_inc(&cur_trans->use_count);
  81                 atomic_inc(&cur_trans->num_writers);
  82                 cur_trans->num_joined++;
  83                 spin_unlock(&root->fs_info->trans_lock);
  84                 return 0;
  85         }
  86         spin_unlock(&root->fs_info->trans_lock);
  87
  88         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  89         if (!cur_trans)
  90                 return -ENOMEM;
  91
  92         spin_lock(&root->fs_info->trans_lock);
  93         if (root->fs_info->running_transaction) {
  94                 /*
  95                  * someone started a transaction after we unlocked.  Make sure
  96                  * to redo the trans_no_join checks above
  97                  */
  98                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  99                 cur_trans = root->fs_info->running_transaction;
 100                 goto loop;
 101         }
 102
 103         atomic_set(&cur_trans->num_writers, 1);
 104         cur_trans->num_joined = 0;
 105         init_waitqueue_head(&cur_trans->writer_wait);
 106         init_waitqueue_head(&cur_trans->commit_wait);
 107         cur_trans->in_commit = 0;
 108         cur_trans->blocked = 0;
 109         /*
 110          * One for this trans handle, one so it will live on until we
 111          * commit the transaction.
 112          */
 113         atomic_set(&cur_trans->use_count, 2);
 114         cur_trans->commit_done = 0;
 115         cur_trans->start_time = get_seconds();
 116
 117         cur_trans->delayed_refs.root = RB_ROOT;
 118         cur_trans->delayed_refs.num_entries = 0;
 119         cur_trans->delayed_refs.num_heads_ready = 0;
 120         cur_trans->delayed_refs.num_heads = 0;
 121         cur_trans->delayed_refs.flushing = 0;
 122         cur_trans->delayed_refs.run_delayed_start = 0;
 123         cur_trans->delayed_refs.seq = 1;
 124         init_waitqueue_head(&cur_trans->delayed_refs.seq_wait);
 125         spin_lock_init(&cur_trans->commit_lock);
 126         spin_lock_init(&cur_trans->delayed_refs.lock);
 127         INIT_LIST_HEAD(&cur_trans->delayed_refs.seq_head);
 128
 129         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 130         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 131         extent_io_tree_init(&cur_trans->dirty_pages,
 132                              root->fs_info->btree_inode->i_mapping);
 133         root->fs_info->generation++;
 134         cur_trans->transid = root->fs_info->generation;
 135         root->fs_info->running_transaction = cur_trans;
 136         cur_trans->aborted = 0;
 137         spin_unlock(&root->fs_info->trans_lock);
 138
 139         return 0;
 140 }
 141
 142 /*
 143  * this does all the record keeping required to make sure that a reference
 144  * counted root is properly recorded in a given transaction.  This is required
 145  * to make sure the old root from before we joined the transaction is deleted
 146  * when the transaction commits
 147  */
 148 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 149                                struct btrfs_root *root)
 150 {
 151         if (root->ref_cows && root->last_trans < trans->transid) {
 152                 WARN_ON(root == root->fs_info->extent_root);
 153                 WARN_ON(root->commit_root != root->node);
 154
 155                 /*
 156                  * see below for in_trans_setup usage rules
 157                  * we have the reloc mutex held now, so there
 158                  * is only one writer in this function
 159                  */
 160                 root->in_trans_setup = 1;
 161
 162                 /* make sure readers find in_trans_setup before
 163                  * they find our root->last_trans update
 164                  */
 165                 smp_wmb();
 166
 167                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 168                 if (root->last_trans == trans->transid) {
 169                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 170                         return 0;
 171                 }
 172                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 173                            (unsigned long)root->root_key.objectid,
 174                            BTRFS_ROOT_TRANS_TAG);
 175                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 176                 root->last_trans = trans->transid;
 177
 178                 /* this is pretty tricky.  We don't want to
 179                  * take the relocation lock in btrfs_record_root_in_trans
 180                  * unless we're really doing the first setup for this root in
 181                  * this transaction.
 182                  *
 183                  * Normally we'd use root->last_trans as a flag to decide
 184                  * if we want to take the expensive mutex.
 185                  *
 186                  * But, we have to set root->last_trans before we
 187                  * init the relocation root, otherwise, we trip over warnings
 188                  * in ctree.c.  The solution used here is to flag ourselves
 189                  * with root->in_trans_setup.  When this is 1, we're still
 190                  * fixing up the reloc trees and everyone must wait.
 191                  *
 192                  * When this is zero, they can trust root->last_trans and fly
 193                  * through btrfs_record_root_in_trans without having to take the
 194                  * lock.  smp_wmb() makes sure that all the writes above are
 195                  * done before we pop in the zero below
 196                  */
 197                 btrfs_init_reloc_root(trans, root);
 198                 smp_wmb();
 199                 root->in_trans_setup = 0;
 200         }
 201         return 0;
 202 }
 203
 204
 205 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 206                                struct btrfs_root *root)
 207 {
 208         if (!root->ref_cows)
 209                 return 0;
 210
 211         /*
 212          * see record_root_in_trans for comments about in_trans_setup usage
 213          * and barriers
 214          */
 215         smp_rmb();
 216         if (root->last_trans == trans->transid &&
 217             !root->in_trans_setup)
 218                 return 0;
 219
 220         mutex_lock(&root->fs_info->reloc_mutex);
 221         record_root_in_trans(trans, root);
 222         mutex_unlock(&root->fs_info->reloc_mutex);
 223
 224         return 0;
 225 }
 226
 227 /* wait for commit against the current transaction to become unblocked
 228  * when this is done, it is safe to start a new transaction, but the current
 229  * transaction might not be fully on disk.
 230  */
 231 static void wait_current_trans(struct btrfs_root *root)
 232 {
 233         struct btrfs_transaction *cur_trans;
 234
 235         spin_lock(&root->fs_info->trans_lock);
 236         cur_trans = root->fs_info->running_transaction;
 237         if (cur_trans && cur_trans->blocked) {
 238                 atomic_inc(&cur_trans->use_count);
 239                 spin_unlock(&root->fs_info->trans_lock);
 240
 241                 wait_event(root->fs_info->transaction_wait,
 242                            !cur_trans->blocked);
 243                 put_transaction(cur_trans);
 244         } else {
 245                 spin_unlock(&root->fs_info->trans_lock);
 246         }
 247 }
 248
 249 enum btrfs_trans_type {
 250         TRANS_START,
 251         TRANS_JOIN,
 252         TRANS_USERSPACE,
 253         TRANS_JOIN_NOLOCK,
 254 };
 255
 256 static int may_wait_transaction(struct btrfs_root *root, int type)
 257 {
 258         if (root->fs_info->log_root_recovering)
 259                 return 0;
 260
 261         if (type == TRANS_USERSPACE)
 262                 return 1;
 263
 264         if (type == TRANS_START &&
 265             !atomic_read(&root->fs_info->open_ioctl_trans))
 266                 return 1;
 267
 268         return 0;
 269 }
 270
 271 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 272                                                     u64 num_items, int type)
 273 {
 274         struct btrfs_trans_handle *h;
 275         struct btrfs_transaction *cur_trans;
 276         u64 num_bytes = 0;
 277         int ret;
 278
 279         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 280                 return ERR_PTR(-EROFS);
 281
 282         if (current->journal_info) {
 283                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 284                 h = current->journal_info;
 285                 h->use_count++;
 286                 h->orig_rsv = h->block_rsv;
 287                 h->block_rsv = NULL;
 288                 goto got_it;
 289         }
 290
 291         /*
 292          * Do the reservation before we join the transaction so we can do all
 293          * the appropriate flushing if need be.
 294          */
 295         if (num_items > 0 && root != root->fs_info->chunk_root) {
 296                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 297                 ret = btrfs_block_rsv_add(root,
 298                                           &root->fs_info->trans_block_rsv,
 299                                           num_bytes);
 300                 if (ret)
 301                         return ERR_PTR(ret);
 302         }
 303 again:
 304         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 305         if (!h)
 306                 return ERR_PTR(-ENOMEM);
 307
 308         if (may_wait_transaction(root, type))
 309                 wait_current_trans(root);
 310
 311         do {
 312                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 313                 if (ret == -EBUSY)
 314                         wait_current_trans(root);
 315         } while (ret == -EBUSY);
 316
 317         if (ret < 0) {
 318                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 319                 return ERR_PTR(ret);
 320         }
 321
 322         cur_trans = root->fs_info->running_transaction;
 323
 324         h->transid = cur_trans->transid;
 325         h->transaction = cur_trans;
 326         h->blocks_used = 0;
 327         h->bytes_reserved = 0;
 328         h->delayed_ref_updates = 0;
 329         h->use_count = 1;
 330         h->block_rsv = NULL;
 331         h->orig_rsv = NULL;
 332         h->aborted = 0;
 333
 334         smp_mb();
 335         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 336                 btrfs_commit_transaction(h, root);
 337                 goto again;
 338         }
 339
 340         if (num_bytes) {
 341                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 342                                               h->transid, num_bytes, 1);
 343                 h->block_rsv = &root->fs_info->trans_block_rsv;
 344                 h->bytes_reserved = num_bytes;
 345         }
 346
 347 got_it:
 348         btrfs_record_root_in_trans(h, root);
 349
 350         if (!current->journal_info && type != TRANS_USERSPACE)
 351                 current->journal_info = h;
 352         return h;
 353 }
 354
 355 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 356                                                    int num_items)
 357 {
 358         return start_transaction(root, num_items, TRANS_START);
 359 }
 360 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 361 {
 362         return start_transaction(root, 0, TRANS_JOIN);
 363 }
 364
 365 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 366 {
 367         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 368 }
 369
 370 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 371 {
 372         return start_transaction(root, 0, TRANS_USERSPACE);
 373 }
 374
 375 /* wait for a transaction commit to be fully complete */
 376 static noinline void wait_for_commit(struct btrfs_root *root,
 377                                     struct btrfs_transaction *commit)
 378 {
 379         wait_event(commit->commit_wait, commit->commit_done);
 380 }
 381
 382 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 383 {
 384         struct btrfs_transaction *cur_trans = NULL, *t;
 385         int ret;
 386
 387         ret = 0;
 388         if (transid) {
 389                 if (transid <= root->fs_info->last_trans_committed)
 390                         goto out;
 391
 392                 /* find specified transaction */
 393                 spin_lock(&root->fs_info->trans_lock);
 394                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 395                         if (t->transid == transid) {
 396                                 cur_trans = t;
 397                                 atomic_inc(&cur_trans->use_count);
 398                                 break;
 399                         }
 400                         if (t->transid > transid)
 401                                 break;
 402                 }
 403                 spin_unlock(&root->fs_info->trans_lock);
 404                 ret = -EINVAL;
 405                 if (!cur_trans)
 406                         goto out;  /* bad transid */
 407         } else {
 408                 /* find newest transaction that is committing | committed */
 409                 spin_lock(&root->fs_info->trans_lock);
 410                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 411                                             list) {
 412                         if (t->in_commit) {
 413                                 if (t->commit_done)
 414                                         break;
 415                                 cur_trans = t;
 416                                 atomic_inc(&cur_trans->use_count);
 417                                 break;
 418                         }
 419                 }
 420                 spin_unlock(&root->fs_info->trans_lock);
 421                 if (!cur_trans)
 422                         goto out;  /* nothing committing|committed */
 423         }
 424
 425         wait_for_commit(root, cur_trans);
 426
 427         put_transaction(cur_trans);
 428         ret = 0;
 429 out:
 430         return ret;
 431 }
 432
 433 void btrfs_throttle(struct btrfs_root *root)
 434 {
 435         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 436                 wait_current_trans(root);
 437 }
 438
 439 static int should_end_transaction(struct btrfs_trans_handle *trans,
 440                                   struct btrfs_root *root)
 441 {
 442         int ret;
 443
 444         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 445         return ret ? 1 : 0;
 446 }
 447
 448 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 449                                  struct btrfs_root *root)
 450 {
 451         struct btrfs_transaction *cur_trans = trans->transaction;
 452         struct btrfs_block_rsv *rsv = trans->block_rsv;
 453         int updates;
 454         int err;
 455
 456         smp_mb();
 457         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 458                 return 1;
 459
 460         /*
 461          * We need to do this in case we're deleting csums so the global block
 462          * rsv get's used instead of the csum block rsv.
 463          */
 464         trans->block_rsv = NULL;
 465
 466         updates = trans->delayed_ref_updates;
 467         trans->delayed_ref_updates = 0;
 468         if (updates) {
 469                 err = btrfs_run_delayed_refs(trans, root, updates);
 470                 if (err) /* Error code will also eval true */
 471                         return err;
 472         }
 473
 474         trans->block_rsv = rsv;
 475
 476         return should_end_transaction(trans, root);
 477 }
 478
 479 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 480                           struct btrfs_root *root, int throttle, int lock)
 481 {
 482         struct btrfs_transaction *cur_trans = trans->transaction;
 483         struct btrfs_fs_info *info = root->fs_info;
 484         int count = 0;
 485         int err = 0;
 486
 487         if (--trans->use_count) {
 488                 trans->block_rsv = trans->orig_rsv;
 489                 return 0;
 490         }
 491
 492         btrfs_trans_release_metadata(trans, root);
 493         trans->block_rsv = NULL;
 494         while (count < 2) {
 495                 unsigned long cur = trans->delayed_ref_updates;
 496                 trans->delayed_ref_updates = 0;
 497                 if (cur &&
 498                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 499                         trans->delayed_ref_updates = 0;
 500                         btrfs_run_delayed_refs(trans, root, cur);
 501                 } else {
 502                         break;
 503                 }
 504                 count++;
 505         }
 506
 507         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 508             should_end_transaction(trans, root)) {
 509                 trans->transaction->blocked = 1;
 510                 smp_wmb();
 511         }
 512
 513         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 514                 if (throttle) {
 515                         /*
 516                          * We may race with somebody else here so end up having
 517                          * to call end_transaction on ourselves again, so inc
 518                          * our use_count.
 519                          */
 520                         trans->use_count++;
 521                         return btrfs_commit_transaction(trans, root);
 522                 } else {
 523                         wake_up_process(info->transaction_kthread);
 524                 }
 525         }
 526
 527         WARN_ON(cur_trans != info->running_transaction);
 528         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 529         atomic_dec(&cur_trans->num_writers);
 530
 531         smp_mb();
 532         if (waitqueue_active(&cur_trans->writer_wait))
 533                 wake_up(&cur_trans->writer_wait);
 534         put_transaction(cur_trans);
 535
 536         if (current->journal_info == trans)
 537                 current->journal_info = NULL;
 538
 539         if (throttle)
 540                 btrfs_run_delayed_iputs(root);
 541
 542         if (trans->aborted ||
 543             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 544                 err = -EIO;
 545         }
 546
 547         memset(trans, 0, sizeof(*trans));
 548         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 549         return err;
 550 }
 551
 552 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 553                           struct btrfs_root *root)
 554 {
 555         int ret;
 556
 557         ret = __btrfs_end_transaction(trans, root, 0, 1);
 558         if (ret)
 559                 return ret;
 560         return 0;
 561 }
 562
 563 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 564                                    struct btrfs_root *root)
 565 {
 566         int ret;
 567
 568         ret = __btrfs_end_transaction(trans, root, 1, 1);
 569         if (ret)
 570                 return ret;
 571         return 0;
 572 }
 573
 574 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 575                                  struct btrfs_root *root)
 576 {
 577         int ret;
 578
 579         ret = __btrfs_end_transaction(trans, root, 0, 0);
 580         if (ret)
 581                 return ret;
 582         return 0;
 583 }
 584
 585 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 586                                 struct btrfs_root *root)
 587 {
 588         return __btrfs_end_transaction(trans, root, 1, 1);
 589 }
 590
 591 /*
 592  * when btree blocks are allocated, they have some corresponding bits set for
 593  * them in one of two extent_io trees.  This is used to make sure all of
 594  * those extents are sent to disk but does not wait on them
 595  */
 596 int btrfs_write_marked_extents(struct btrfs_root *root,
 597                                struct extent_io_tree *dirty_pages, int mark)
 598 {
 599         int err = 0;
 600         int werr = 0;
 601         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 602         u64 start = 0;
 603         u64 end;
 604
 605         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 606                                       mark)) {
 607                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 608                                    GFP_NOFS);
 609                 err = filemap_fdatawrite_range(mapping, start, end);
 610                 if (err)
 611                         werr = err;
 612                 cond_resched();
 613                 start = end + 1;
 614         }
 615         if (err)
 616                 werr = err;
 617         return werr;
 618 }
 619
 620 /*
 621  * when btree blocks are allocated, they have some corresponding bits set for
 622  * them in one of two extent_io trees.  This is used to make sure all of
 623  * those extents are on disk for transaction or log commit.  We wait
 624  * on all the pages and clear them from the dirty pages state tree
 625  */
 626 int btrfs_wait_marked_extents(struct btrfs_root *root,
 627                               struct extent_io_tree *dirty_pages, int mark)
 628 {
 629         int err = 0;
 630         int werr = 0;
 631         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 632         u64 start = 0;
 633         u64 end;
 634
 635         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 636                                       EXTENT_NEED_WAIT)) {
 637                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 638                 err = filemap_fdatawait_range(mapping, start, end);
 639                 if (err)
 640                         werr = err;
 641                 cond_resched();
 642                 start = end + 1;
 643         }
 644         if (err)
 645                 werr = err;
 646         return werr;
 647 }
 648
 649 /*
 650  * when btree blocks are allocated, they have some corresponding bits set for
 651  * them in one of two extent_io trees.  This is used to make sure all of
 652  * those extents are on disk for transaction or log commit
 653  */
 654 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 655                                 struct extent_io_tree *dirty_pages, int mark)
 656 {
 657         int ret;
 658         int ret2;
 659
 660         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 661         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 662
 663         if (ret)
 664                 return ret;
 665         if (ret2)
 666                 return ret2;
 667         return 0;
 668 }
 669
 670 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 671                                      struct btrfs_root *root)
 672 {
 673         if (!trans || !trans->transaction) {
 674                 struct inode *btree_inode;
 675                 btree_inode = root->fs_info->btree_inode;
 676                 return filemap_write_and_wait(btree_inode->i_mapping);
 677         }
 678         return btrfs_write_and_wait_marked_extents(root,
 679                                            &trans->transaction->dirty_pages,
 680                                            EXTENT_DIRTY);
 681 }
 682
 683 /*
 684  * this is used to update the root pointer in the tree of tree roots.
 685  *
 686  * But, in the case of the extent allocation tree, updating the root
 687  * pointer may allocate blocks which may change the root of the extent
 688  * allocation tree.
 689  *
 690  * So, this loops and repeats and makes sure the cowonly root didn't
 691  * change while the root pointer was being updated in the metadata.
 692  */
 693 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 694                                struct btrfs_root *root)
 695 {
 696         int ret;
 697         u64 old_root_bytenr;
 698         u64 old_root_used;
 699         struct btrfs_root *tree_root = root->fs_info->tree_root;
 700
 701         old_root_used = btrfs_root_used(&root->root_item);
 702         btrfs_write_dirty_block_groups(trans, root);
 703
 704         while (1) {
 705                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 706                 if (old_root_bytenr == root->node->start &&
 707                     old_root_used == btrfs_root_used(&root->root_item))
 708                         break;
 709
 710                 btrfs_set_root_node(&root->root_item, root->node);
 711                 ret = btrfs_update_root(trans, tree_root,
 712                                         &root->root_key,
 713                                         &root->root_item);
 714                 if (ret)
 715                         return ret;
 716
 717                 old_root_used = btrfs_root_used(&root->root_item);
 718                 ret = btrfs_write_dirty_block_groups(trans, root);
 719                 if (ret)
 720                         return ret;
 721         }
 722
 723         if (root != root->fs_info->extent_root)
 724                 switch_commit_root(root);
 725
 726         return 0;
 727 }
 728
 729 /*
 730  * update all the cowonly tree roots on disk
 731  *
 732  * The error handling in this function may not be obvious. Any of the
 733  * failures will cause the file system to go offline. We still need
 734  * to clean up the delayed refs.
 735  */
 736 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 737                                          struct btrfs_root *root)
 738 {
 739         struct btrfs_fs_info *fs_info = root->fs_info;
 740         struct list_head *next;
 741         struct extent_buffer *eb;
 742         int ret;
 743
 744         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 745         if (ret)
 746                 return ret;
 747
 748         eb = btrfs_lock_root_node(fs_info->tree_root);
 749         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 750                               0, &eb);
 751         btrfs_tree_unlock(eb);
 752         free_extent_buffer(eb);
 753
 754         if (ret)
 755                 return ret;
 756
 757         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 758         if (ret)
 759                 return ret;
 760
 761         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 762                 next = fs_info->dirty_cowonly_roots.next;
 763                 list_del_init(next);
 764                 root = list_entry(next, struct btrfs_root, dirty_list);
 765
 766                 ret = update_cowonly_root(trans, root);
 767                 if (ret)
 768                         return ret;
 769         }
 770
 771         down_write(&fs_info->extent_commit_sem);
 772         switch_commit_root(fs_info->extent_root);
 773         up_write(&fs_info->extent_commit_sem);
 774
 775         return 0;
 776 }
 777
 778 /*
 779  * dead roots are old snapshots that need to be deleted.  This allocates
 780  * a dirty root struct and adds it into the list of dead roots that need to
 781  * be deleted
 782  */
 783 int btrfs_add_dead_root(struct btrfs_root *root)
 784 {
 785         spin_lock(&root->fs_info->trans_lock);
 786         list_add(&root->root_list, &root->fs_info->dead_roots);
 787         spin_unlock(&root->fs_info->trans_lock);
 788         return 0;
 789 }
 790
 791 /*
 792  * update all the cowonly tree roots on disk
 793  */
 794 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 795                                     struct btrfs_root *root)
 796 {
 797         struct btrfs_root *gang[8];
 798         struct btrfs_fs_info *fs_info = root->fs_info;
 799         int i;
 800         int ret;
 801         int err = 0;
 802
 803         spin_lock(&fs_info->fs_roots_radix_lock);
 804         while (1) {
 805                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 806                                                  (void **)gang, 0,
 807                                                  ARRAY_SIZE(gang),
 808                                                  BTRFS_ROOT_TRANS_TAG);
 809                 if (ret == 0)
 810                         break;
 811                 for (i = 0; i < ret; i++) {
 812                         root = gang[i];
 813                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 814                                         (unsigned long)root->root_key.objectid,
 815                                         BTRFS_ROOT_TRANS_TAG);
 816                         spin_unlock(&fs_info->fs_roots_radix_lock);
 817
 818                         btrfs_free_log(trans, root);
 819                         btrfs_update_reloc_root(trans, root);
 820                         btrfs_orphan_commit_root(trans, root);
 821
 822                         btrfs_save_ino_cache(root, trans);
 823
 824                         /* see comments in should_cow_block() */
 825                         root->force_cow = 0;
 826                         smp_wmb();
 827
 828                         if (root->commit_root != root->node) {
 829                                 mutex_lock(&root->fs_commit_mutex);
 830                                 switch_commit_root(root);
 831                                 btrfs_unpin_free_ino(root);
 832                                 mutex_unlock(&root->fs_commit_mutex);
 833
 834                                 btrfs_set_root_node(&root->root_item,
 835                                                     root->node);
 836                         }
 837
 838                         err = btrfs_update_root(trans, fs_info->tree_root,
 839                                                 &root->root_key,
 840                                                 &root->root_item);
 841                         spin_lock(&fs_info->fs_roots_radix_lock);
 842                         if (err)
 843                                 break;
 844                 }
 845         }
 846         spin_unlock(&fs_info->fs_roots_radix_lock);
 847         return err;
 848 }
 849
 850 /*
 851  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 852  * otherwise every leaf in the btree is read and defragged.
 853  */
 854 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 855 {
 856         struct btrfs_fs_info *info = root->fs_info;
 857         struct btrfs_trans_handle *trans;
 858         int ret;
 859         unsigned long nr;
 860
 861         if (xchg(&root->defrag_running, 1))
 862                 return 0;
 863
 864         while (1) {
 865                 trans = btrfs_start_transaction(root, 0);
 866                 if (IS_ERR(trans))
 867                         return PTR_ERR(trans);
 868
 869                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 870
 871                 nr = trans->blocks_used;
 872                 btrfs_end_transaction(trans, root);
 873                 btrfs_btree_balance_dirty(info->tree_root, nr);
 874                 cond_resched();
 875
 876                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 877                         break;
 878         }
 879         root->defrag_running = 0;
 880         return ret;
 881 }
 882
 883 /*
 884  * new snapshots need to be created at a very specific time in the
 885  * transaction commit.  This does the actual creation
 886  */
 887 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 888                                    struct btrfs_fs_info *fs_info,
 889                                    struct btrfs_pending_snapshot *pending)
 890 {
 891         struct btrfs_key key;
 892         struct btrfs_root_item *new_root_item;
 893         struct btrfs_root *tree_root = fs_info->tree_root;
 894         struct btrfs_root *root = pending->root;
 895         struct btrfs_root *parent_root;
 896         struct btrfs_block_rsv *rsv;
 897         struct inode *parent_inode;
 898         struct dentry *parent;
 899         struct dentry *dentry;
 900         struct extent_buffer *tmp;
 901         struct extent_buffer *old;
 902         int ret;
 903         u64 to_reserve = 0;
 904         u64 index = 0;
 905         u64 objectid;
 906         u64 root_flags;
 907
 908         rsv = trans->block_rsv;
 909
 910         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 911         if (!new_root_item) {
 912                 ret = pending->error = -ENOMEM;
 913                 goto fail;
 914         }
 915
 916         ret = btrfs_find_free_objectid(tree_root, &objectid);
 917         if (ret) {
 918                 pending->error = ret;
 919                 goto fail;
 920         }
 921
 922         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 923
 924         if (to_reserve > 0) {
 925                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 926                                                   to_reserve);
 927                 if (ret) {
 928                         pending->error = ret;
 929                         goto fail;
 930                 }
 931         }
 932
 933         key.objectid = objectid;
 934         key.offset = (u64)-1;
 935         key.type = BTRFS_ROOT_ITEM_KEY;
 936
 937         trans->block_rsv = &pending->block_rsv;
 938
 939         dentry = pending->dentry;
 940         parent = dget_parent(dentry);
 941         parent_inode = parent->d_inode;
 942         parent_root = BTRFS_I(parent_inode)->root;
 943         record_root_in_trans(trans, parent_root);
 944
 945         /*
 946          * insert the directory item
 947          */
 948         ret = btrfs_set_inode_index(parent_inode, &index);
 949         BUG_ON(ret); /* -ENOMEM */
 950         ret = btrfs_insert_dir_item(trans, parent_root,
 951                                 dentry->d_name.name, dentry->d_name.len,
 952                                 parent_inode, &key,
 953                                 BTRFS_FT_DIR, index);
 954         if (ret == -EEXIST) {
 955                 pending->error = -EEXIST;
 956                 dput(parent);
 957                 goto fail;
 958         } else if (ret) {
 959                 goto abort_trans_dput;
 960         }
 961
 962         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 963                                          dentry->d_name.len * 2);
 964         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 965         if (ret)
 966                 goto abort_trans_dput;
 967
 968         /*
 969          * pull in the delayed directory update
 970          * and the delayed inode item
 971          * otherwise we corrupt the FS during
 972          * snapshot
 973          */
 974         ret = btrfs_run_delayed_items(trans, root);
 975         if (ret) { /* Transaction aborted */
 976                 dput(parent);
 977                 goto fail;
 978         }
 979
 980         record_root_in_trans(trans, root);
 981         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 982         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 983         btrfs_check_and_init_root_item(new_root_item);
 984
 985         root_flags = btrfs_root_flags(new_root_item);
 986         if (pending->readonly)
 987                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 988         else
 989                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 990         btrfs_set_root_flags(new_root_item, root_flags);
 991
 992         old = btrfs_lock_root_node(root);
 993         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
 994         if (ret) {
 995                 btrfs_tree_unlock(old);
 996                 free_extent_buffer(old);
 997                 goto abort_trans_dput;
 998         }
 999
1000         btrfs_set_lock_blocking(old);
1001
1002         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1003         /* clean up in any case */
1004         btrfs_tree_unlock(old);
1005         free_extent_buffer(old);
1006         if (ret)
1007                 goto abort_trans_dput;
1008
1009         /* see comments in should_cow_block() */
1010         root->force_cow = 1;
1011         smp_wmb();
1012
1013         btrfs_set_root_node(new_root_item, tmp);
1014         /* record when the snapshot was created in key.offset */
1015         key.offset = trans->transid;
1016         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1017         btrfs_tree_unlock(tmp);
1018         free_extent_buffer(tmp);
1019         if (ret)
1020                 goto abort_trans_dput;
1021
1022         /*
1023          * insert root back/forward references
1024          */
1025         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1026                                  parent_root->root_key.objectid,
1027                                  btrfs_ino(parent_inode), index,
1028                                  dentry->d_name.name, dentry->d_name.len);
1029         dput(parent);
1030         if (ret)
1031                 goto fail;
1032
1033         key.offset = (u64)-1;
1034         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1035         if (IS_ERR(pending->snap)) {
1036                 ret = PTR_ERR(pending->snap);
1037                 goto abort_trans;
1038         }
1039
1040         ret = btrfs_reloc_post_snapshot(trans, pending);
1041         if (ret)
1042                 goto abort_trans;
1043         ret = 0;
1044 fail:
1045         kfree(new_root_item);
1046         trans->block_rsv = rsv;
1047         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1048         return ret;
1049
1050 abort_trans_dput:
1051         dput(parent);
1052 abort_trans:
1053         btrfs_abort_transaction(trans, root, ret);
1054         goto fail;
1055 }
1056
1057 /*
1058  * create all the snapshots we've scheduled for creation
1059  */
1060 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1061                                              struct btrfs_fs_info *fs_info)
1062 {
1063         struct btrfs_pending_snapshot *pending;
1064         struct list_head *head = &trans->transaction->pending_snapshots;
1065
1066         list_for_each_entry(pending, head, list)
1067                 create_pending_snapshot(trans, fs_info, pending);
1068         return 0;
1069 }
1070
1071 static void update_super_roots(struct btrfs_root *root)
1072 {
1073         struct btrfs_root_item *root_item;
1074         struct btrfs_super_block *super;
1075
1076         super = root->fs_info->super_copy;
1077
1078         root_item = &root->fs_info->chunk_root->root_item;
1079         super->chunk_root = root_item->bytenr;
1080         super->chunk_root_generation = root_item->generation;
1081         super->chunk_root_level = root_item->level;
1082
1083         root_item = &root->fs_info->tree_root->root_item;
1084         super->root = root_item->bytenr;
1085         super->generation = root_item->generation;
1086         super->root_level = root_item->level;
1087         if (btrfs_test_opt(root, SPACE_CACHE))
1088                 super->cache_generation = root_item->generation;
1089 }
1090
1091 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1092 {
1093         int ret = 0;
1094         spin_lock(&info->trans_lock);
1095         if (info->running_transaction)
1096                 ret = info->running_transaction->in_commit;
1097         spin_unlock(&info->trans_lock);
1098         return ret;
1099 }
1100
1101 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1102 {
1103         int ret = 0;
1104         spin_lock(&info->trans_lock);
1105         if (info->running_transaction)
1106                 ret = info->running_transaction->blocked;
1107         spin_unlock(&info->trans_lock);
1108         return ret;
1109 }
1110
1111 /*
1112  * wait for the current transaction commit to start and block subsequent
1113  * transaction joins
1114  */
1115 static void wait_current_trans_commit_start(struct btrfs_root *root,
1116                                             struct btrfs_transaction *trans)
1117 {
1118         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1119 }
1120
1121 /*
1122  * wait for the current transaction to start and then become unblocked.
1123  * caller holds ref.
1124  */
1125 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1126                                          struct btrfs_transaction *trans)
1127 {
1128         wait_event(root->fs_info->transaction_wait,
1129                    trans->commit_done || (trans->in_commit && !trans->blocked));
1130 }
1131
1132 /*
1133  * commit transactions asynchronously. once btrfs_commit_transaction_async
1134  * returns, any subsequent transaction will not be allowed to join.
1135  */
1136 struct btrfs_async_commit {
1137         struct btrfs_trans_handle *newtrans;
1138         struct btrfs_root *root;
1139         struct delayed_work work;
1140 };
1141
1142 static void do_async_commit(struct work_struct *work)
1143 {
1144         struct btrfs_async_commit *ac =
1145                 container_of(work, struct btrfs_async_commit, work.work);
1146
1147         btrfs_commit_transaction(ac->newtrans, ac->root);
1148         kfree(ac);
1149 }
1150
1151 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1152                                    struct btrfs_root *root,
1153                                    int wait_for_unblock)
1154 {
1155         struct btrfs_async_commit *ac;
1156         struct btrfs_transaction *cur_trans;
1157
1158         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1159         if (!ac)
1160                 return -ENOMEM;
1161
1162         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1163         ac->root = root;
1164         ac->newtrans = btrfs_join_transaction(root);
1165         if (IS_ERR(ac->newtrans)) {
1166                 int err = PTR_ERR(ac->newtrans);
1167                 kfree(ac);
1168                 return err;
1169         }
1170
1171         /* take transaction reference */
1172         cur_trans = trans->transaction;
1173         atomic_inc(&cur_trans->use_count);
1174
1175         btrfs_end_transaction(trans, root);
1176         schedule_delayed_work(&ac->work, 0);
1177
1178         /* wait for transaction to start and unblock */
1179         if (wait_for_unblock)
1180                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1181         else
1182                 wait_current_trans_commit_start(root, cur_trans);
1183
1184         if (current->journal_info == trans)
1185                 current->journal_info = NULL;
1186
1187         put_transaction(cur_trans);
1188         return 0;
1189 }
1190
1191
1192 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1193                                 struct btrfs_root *root)
1194 {
1195         struct btrfs_transaction *cur_trans = trans->transaction;
1196
1197         WARN_ON(trans->use_count > 1);
1198
1199         spin_lock(&root->fs_info->trans_lock);
1200         list_del_init(&cur_trans->list);
1201         spin_unlock(&root->fs_info->trans_lock);
1202
1203         btrfs_cleanup_one_transaction(trans->transaction, root);
1204
1205         put_transaction(cur_trans);
1206         put_transaction(cur_trans);
1207
1208         trace_btrfs_transaction_commit(root);
1209
1210         btrfs_scrub_continue(root);
1211
1212         if (current->journal_info == trans)
1213                 current->journal_info = NULL;
1214
1215         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1216 }
1217
1218 /*
1219  * btrfs_transaction state sequence:
1220  *    in_commit = 0, blocked = 0  (initial)
1221  *    in_commit = 1, blocked = 1
1222  *    blocked = 0
1223  *    commit_done = 1
1224  */
1225 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1226                              struct btrfs_root *root)
1227 {
1228         unsigned long joined = 0;
1229         struct btrfs_transaction *cur_trans = trans->transaction;
1230         struct btrfs_transaction *prev_trans = NULL;
1231         DEFINE_WAIT(wait);
1232         int ret = -EIO;
1233         int should_grow = 0;
1234         unsigned long now = get_seconds();
1235         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1236
1237         btrfs_run_ordered_operations(root, 0);
1238
1239         btrfs_trans_release_metadata(trans, root);
1240         trans->block_rsv = NULL;
1241
1242         if (cur_trans->aborted)
1243                 goto cleanup_transaction;
1244
1245         /* make a pass through all the delayed refs we have so far
1246          * any runnings procs may add more while we are here
1247          */
1248         ret = btrfs_run_delayed_refs(trans, root, 0);
1249         if (ret)
1250                 goto cleanup_transaction;
1251
1252         cur_trans = trans->transaction;
1253
1254         /*
1255          * set the flushing flag so procs in this transaction have to
1256          * start sending their work down.
1257          */
1258         cur_trans->delayed_refs.flushing = 1;
1259
1260         ret = btrfs_run_delayed_refs(trans, root, 0);
1261         if (ret)
1262                 goto cleanup_transaction;
1263
1264         spin_lock(&cur_trans->commit_lock);
1265         if (cur_trans->in_commit) {
1266                 spin_unlock(&cur_trans->commit_lock);
1267                 atomic_inc(&cur_trans->use_count);
1268                 ret = btrfs_end_transaction(trans, root);
1269
1270                 wait_for_commit(root, cur_trans);
1271
1272                 put_transaction(cur_trans);
1273
1274                 return ret;
1275         }
1276
1277         trans->transaction->in_commit = 1;
1278         trans->transaction->blocked = 1;
1279         spin_unlock(&cur_trans->commit_lock);
1280         wake_up(&root->fs_info->transaction_blocked_wait);
1281
1282         spin_lock(&root->fs_info->trans_lock);
1283         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1284                 prev_trans = list_entry(cur_trans->list.prev,
1285                                         struct btrfs_transaction, list);
1286                 if (!prev_trans->commit_done) {
1287                         atomic_inc(&prev_trans->use_count);
1288                         spin_unlock(&root->fs_info->trans_lock);
1289
1290                         wait_for_commit(root, prev_trans);
1291
1292                         put_transaction(prev_trans);
1293                 } else {
1294                         spin_unlock(&root->fs_info->trans_lock);
1295                 }
1296         } else {
1297                 spin_unlock(&root->fs_info->trans_lock);
1298         }
1299
1300         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1301                 should_grow = 1;
1302
1303         do {
1304                 int snap_pending = 0;
1305
1306                 joined = cur_trans->num_joined;
1307                 if (!list_empty(&trans->transaction->pending_snapshots))
1308                         snap_pending = 1;
1309
1310                 WARN_ON(cur_trans != trans->transaction);
1311
1312                 if (flush_on_commit || snap_pending) {
1313                         btrfs_start_delalloc_inodes(root, 1);
1314                         btrfs_wait_ordered_extents(root, 0, 1);
1315                 }
1316
1317                 ret = btrfs_run_delayed_items(trans, root);
1318                 if (ret)
1319                         goto cleanup_transaction;
1320
1321                 /*
1322                  * rename don't use btrfs_join_transaction, so, once we
1323                  * set the transaction to blocked above, we aren't going
1324                  * to get any new ordered operations.  We can safely run
1325                  * it here and no for sure that nothing new will be added
1326                  * to the list
1327                  */
1328                 btrfs_run_ordered_operations(root, 1);
1329
1330                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1331                                 TASK_UNINTERRUPTIBLE);
1332
1333                 if (atomic_read(&cur_trans->num_writers) > 1)
1334                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1335                 else if (should_grow)
1336                         schedule_timeout(1);
1337
1338                 finish_wait(&cur_trans->writer_wait, &wait);
1339         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1340                  (should_grow && cur_trans->num_joined != joined));
1341
1342         /*
1343          * Ok now we need to make sure to block out any other joins while we
1344          * commit the transaction.  We could have started a join before setting
1345          * no_join so make sure to wait for num_writers to == 1 again.
1346          */
1347         spin_lock(&root->fs_info->trans_lock);
1348         root->fs_info->trans_no_join = 1;
1349         spin_unlock(&root->fs_info->trans_lock);
1350         wait_event(cur_trans->writer_wait,
1351                    atomic_read(&cur_trans->num_writers) == 1);
1352
1353         /*
1354          * the reloc mutex makes sure that we stop
1355          * the balancing code from coming in and moving
1356          * extents around in the middle of the commit
1357          */
1358         mutex_lock(&root->fs_info->reloc_mutex);
1359
1360         ret = btrfs_run_delayed_items(trans, root);
1361         if (ret) {
1362                 mutex_unlock(&root->fs_info->reloc_mutex);
1363                 goto cleanup_transaction;
1364         }
1365
1366         ret = create_pending_snapshots(trans, root->fs_info);
1367         if (ret) {
1368                 mutex_unlock(&root->fs_info->reloc_mutex);
1369                 goto cleanup_transaction;
1370         }
1371
1372         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1373         if (ret) {
1374                 mutex_unlock(&root->fs_info->reloc_mutex);
1375                 goto cleanup_transaction;
1376         }
1377
1378         /*
1379          * make sure none of the code above managed to slip in a
1380          * delayed item
1381          */
1382         btrfs_assert_delayed_root_empty(root);
1383
1384         WARN_ON(cur_trans != trans->transaction);
1385
1386         btrfs_scrub_pause(root);
1387         /* btrfs_commit_tree_roots is responsible for getting the
1388          * various roots consistent with each other.  Every pointer
1389          * in the tree of tree roots has to point to the most up to date
1390          * root for every subvolume and other tree.  So, we have to keep
1391          * the tree logging code from jumping in and changing any
1392          * of the trees.
1393          *
1394          * At this point in the commit, there can't be any tree-log
1395          * writers, but a little lower down we drop the trans mutex
1396          * and let new people in.  By holding the tree_log_mutex
1397          * from now until after the super is written, we avoid races
1398          * with the tree-log code.
1399          */
1400         mutex_lock(&root->fs_info->tree_log_mutex);
1401
1402         ret = commit_fs_roots(trans, root);
1403         if (ret) {
1404                 mutex_unlock(&root->fs_info->tree_log_mutex);
1405                 mutex_unlock(&root->fs_info->reloc_mutex);
1406                 goto cleanup_transaction;
1407         }
1408
1409         /* commit_fs_roots gets rid of all the tree log roots, it is now
1410          * safe to free the root of tree log roots
1411          */
1412         btrfs_free_log_root_tree(trans, root->fs_info);
1413
1414         ret = commit_cowonly_roots(trans, root);
1415         if (ret) {
1416                 mutex_unlock(&root->fs_info->tree_log_mutex);
1417                 mutex_unlock(&root->fs_info->reloc_mutex);
1418                 goto cleanup_transaction;
1419         }
1420
1421         btrfs_prepare_extent_commit(trans, root);
1422
1423         cur_trans = root->fs_info->running_transaction;
1424
1425         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1426                             root->fs_info->tree_root->node);
1427         switch_commit_root(root->fs_info->tree_root);
1428
1429         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1430                             root->fs_info->chunk_root->node);
1431         switch_commit_root(root->fs_info->chunk_root);
1432
1433         update_super_roots(root);
1434
1435         if (!root->fs_info->log_root_recovering) {
1436                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1437                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1438         }
1439
1440         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1441                sizeof(*root->fs_info->super_copy));
1442
1443         trans->transaction->blocked = 0;
1444         spin_lock(&root->fs_info->trans_lock);
1445         root->fs_info->running_transaction = NULL;
1446         root->fs_info->trans_no_join = 0;
1447         spin_unlock(&root->fs_info->trans_lock);
1448         mutex_unlock(&root->fs_info->reloc_mutex);
1449
1450         wake_up(&root->fs_info->transaction_wait);
1451
1452         ret = btrfs_write_and_wait_transaction(trans, root);
1453         if (ret) {
1454                 btrfs_error(root->fs_info, ret,
1455                             "Error while writing out transaction.");
1456                 mutex_unlock(&root->fs_info->tree_log_mutex);
1457                 goto cleanup_transaction;
1458         }
1459
1460         ret = write_ctree_super(trans, root, 0);
1461         if (ret) {
1462                 mutex_unlock(&root->fs_info->tree_log_mutex);
1463                 goto cleanup_transaction;
1464         }
1465
1466         /*
1467          * the super is written, we can safely allow the tree-loggers
1468          * to go about their business
1469          */
1470         mutex_unlock(&root->fs_info->tree_log_mutex);
1471
1472         btrfs_finish_extent_commit(trans, root);
1473
1474         cur_trans->commit_done = 1;
1475
1476         root->fs_info->last_trans_committed = cur_trans->transid;
1477
1478         wake_up(&cur_trans->commit_wait);
1479
1480         spin_lock(&root->fs_info->trans_lock);
1481         list_del_init(&cur_trans->list);
1482         spin_unlock(&root->fs_info->trans_lock);
1483
1484         put_transaction(cur_trans);
1485         put_transaction(cur_trans);
1486
1487         trace_btrfs_transaction_commit(root);
1488
1489         btrfs_scrub_continue(root);
1490
1491         if (current->journal_info == trans)
1492                 current->journal_info = NULL;
1493
1494         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1495
1496         if (current != root->fs_info->transaction_kthread)
1497                 btrfs_run_delayed_iputs(root);
1498
1499         return ret;
1500
1501 cleanup_transaction:
1502         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1503 //      WARN_ON(1);
1504         if (current->journal_info == trans)
1505                 current->journal_info = NULL;
1506         cleanup_transaction(trans, root);
1507
1508         return ret;
1509 }
1510
1511 /*
1512  * interface function to delete all the snapshots we have scheduled for deletion
1513  */
1514 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1515 {
1516         LIST_HEAD(list);
1517         struct btrfs_fs_info *fs_info = root->fs_info;
1518
1519         spin_lock(&fs_info->trans_lock);
1520         list_splice_init(&fs_info->dead_roots, &list);
1521         spin_unlock(&fs_info->trans_lock);
1522
1523         while (!list_empty(&list)) {
1524                 int ret;
1525
1526                 root = list_entry(list.next, struct btrfs_root, root_list);
1527                 list_del(&root->root_list);
1528
1529                 btrfs_kill_all_delayed_nodes(root);
1530
1531                 if (btrfs_header_backref_rev(root->node) <
1532                     BTRFS_MIXED_BACKREF_REV)
1533                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1534                 else
1535                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1536                 BUG_ON(ret < 0);
1537         }
1538         return 0;
1539 }