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 #include "volumes.h"
  32
  33 #define BTRFS_ROOT_TRANS_TAG 0
  34
  35 void put_transaction(struct btrfs_transaction *transaction)
  36 {
  37         WARN_ON(atomic_read(&transaction->use_count) == 0);
  38         if (atomic_dec_and_test(&transaction->use_count)) {
  39                 BUG_ON(!list_empty(&transaction->list));
  40                 WARN_ON(transaction->delayed_refs.root.rb_node);
  41                 WARN_ON(!list_empty(&transaction->delayed_refs.seq_head));
  42                 memset(transaction, 0, sizeof(*transaction));
  43                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  44         }
  45 }
  46
  47 static noinline void switch_commit_root(struct btrfs_root *root)
  48 {
  49         free_extent_buffer(root->commit_root);
  50         root->commit_root = btrfs_root_node(root);
  51 }
  52
  53 /*
  54  * either allocate a new transaction or hop into the existing one
  55  */
  56 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  57 {
  58         struct btrfs_transaction *cur_trans;
  59         struct btrfs_fs_info *fs_info = root->fs_info;
  60
  61         spin_lock(&fs_info->trans_lock);
  62 loop:
  63         /* The file system has been taken offline. No new transactions. */
  64         if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  65                 spin_unlock(&fs_info->trans_lock);
  66                 return -EROFS;
  67         }
  68
  69         if (fs_info->trans_no_join) {
  70                 if (!nofail) {
  71                         spin_unlock(&fs_info->trans_lock);
  72                         return -EBUSY;
  73                 }
  74         }
  75
  76         cur_trans = fs_info->running_transaction;
  77         if (cur_trans) {
  78                 if (cur_trans->aborted) {
  79                         spin_unlock(&fs_info->trans_lock);
  80                         return cur_trans->aborted;
  81                 }
  82                 atomic_inc(&cur_trans->use_count);
  83                 atomic_inc(&cur_trans->num_writers);
  84                 cur_trans->num_joined++;
  85                 spin_unlock(&fs_info->trans_lock);
  86                 return 0;
  87         }
  88         spin_unlock(&fs_info->trans_lock);
  89
  90         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  91         if (!cur_trans)
  92                 return -ENOMEM;
  93
  94         spin_lock(&fs_info->trans_lock);
  95         if (fs_info->running_transaction) {
  96                 /*
  97                  * someone started a transaction after we unlocked.  Make sure
  98                  * to redo the trans_no_join checks above
  99                  */
 100                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 101                 cur_trans = fs_info->running_transaction;
 102                 goto loop;
 103         }
 104
 105         atomic_set(&cur_trans->num_writers, 1);
 106         cur_trans->num_joined = 0;
 107         init_waitqueue_head(&cur_trans->writer_wait);
 108         init_waitqueue_head(&cur_trans->commit_wait);
 109         cur_trans->in_commit = 0;
 110         cur_trans->blocked = 0;
 111         /*
 112          * One for this trans handle, one so it will live on until we
 113          * commit the transaction.
 114          */
 115         atomic_set(&cur_trans->use_count, 2);
 116         cur_trans->commit_done = 0;
 117         cur_trans->start_time = get_seconds();
 118
 119         cur_trans->delayed_refs.root = RB_ROOT;
 120         cur_trans->delayed_refs.num_entries = 0;
 121         cur_trans->delayed_refs.num_heads_ready = 0;
 122         cur_trans->delayed_refs.num_heads = 0;
 123         cur_trans->delayed_refs.flushing = 0;
 124         cur_trans->delayed_refs.run_delayed_start = 0;
 125         cur_trans->delayed_refs.seq = 1;
 126
 127         /*
 128          * although the tree mod log is per file system and not per transaction,
 129          * the log must never go across transaction boundaries.
 130          */
 131         smp_mb();
 132         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 133                 printk(KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 134                         "creating a fresh transaction\n");
 135                 WARN_ON(1);
 136         }
 137         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) {
 138                 printk(KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 139                         "creating a fresh transaction\n");
 140                 WARN_ON(1);
 141         }
 142         atomic_set(&fs_info->tree_mod_seq, 0);
 143
 144         init_waitqueue_head(&cur_trans->delayed_refs.seq_wait);
 145         spin_lock_init(&cur_trans->commit_lock);
 146         spin_lock_init(&cur_trans->delayed_refs.lock);
 147         INIT_LIST_HEAD(&cur_trans->delayed_refs.seq_head);
 148
 149         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 150         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 151         extent_io_tree_init(&cur_trans->dirty_pages,
 152                              fs_info->btree_inode->i_mapping);
 153         fs_info->generation++;
 154         cur_trans->transid = fs_info->generation;
 155         fs_info->running_transaction = cur_trans;
 156         cur_trans->aborted = 0;
 157         spin_unlock(&fs_info->trans_lock);
 158
 159         return 0;
 160 }
 161
 162 /*
 163  * this does all the record keeping required to make sure that a reference
 164  * counted root is properly recorded in a given transaction.  This is required
 165  * to make sure the old root from before we joined the transaction is deleted
 166  * when the transaction commits
 167  */
 168 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 169                                struct btrfs_root *root)
 170 {
 171         if (root->ref_cows && root->last_trans < trans->transid) {
 172                 WARN_ON(root == root->fs_info->extent_root);
 173                 WARN_ON(root->commit_root != root->node);
 174
 175                 /*
 176                  * see below for in_trans_setup usage rules
 177                  * we have the reloc mutex held now, so there
 178                  * is only one writer in this function
 179                  */
 180                 root->in_trans_setup = 1;
 181
 182                 /* make sure readers find in_trans_setup before
 183                  * they find our root->last_trans update
 184                  */
 185                 smp_wmb();
 186
 187                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 188                 if (root->last_trans == trans->transid) {
 189                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 190                         return 0;
 191                 }
 192                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 193                            (unsigned long)root->root_key.objectid,
 194                            BTRFS_ROOT_TRANS_TAG);
 195                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 196                 root->last_trans = trans->transid;
 197
 198                 /* this is pretty tricky.  We don't want to
 199                  * take the relocation lock in btrfs_record_root_in_trans
 200                  * unless we're really doing the first setup for this root in
 201                  * this transaction.
 202                  *
 203                  * Normally we'd use root->last_trans as a flag to decide
 204                  * if we want to take the expensive mutex.
 205                  *
 206                  * But, we have to set root->last_trans before we
 207                  * init the relocation root, otherwise, we trip over warnings
 208                  * in ctree.c.  The solution used here is to flag ourselves
 209                  * with root->in_trans_setup.  When this is 1, we're still
 210                  * fixing up the reloc trees and everyone must wait.
 211                  *
 212                  * When this is zero, they can trust root->last_trans and fly
 213                  * through btrfs_record_root_in_trans without having to take the
 214                  * lock.  smp_wmb() makes sure that all the writes above are
 215                  * done before we pop in the zero below
 216                  */
 217                 btrfs_init_reloc_root(trans, root);
 218                 smp_wmb();
 219                 root->in_trans_setup = 0;
 220         }
 221         return 0;
 222 }
 223
 224
 225 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 226                                struct btrfs_root *root)
 227 {
 228         if (!root->ref_cows)
 229                 return 0;
 230
 231         /*
 232          * see record_root_in_trans for comments about in_trans_setup usage
 233          * and barriers
 234          */
 235         smp_rmb();
 236         if (root->last_trans == trans->transid &&
 237             !root->in_trans_setup)
 238                 return 0;
 239
 240         mutex_lock(&root->fs_info->reloc_mutex);
 241         record_root_in_trans(trans, root);
 242         mutex_unlock(&root->fs_info->reloc_mutex);
 243
 244         return 0;
 245 }
 246
 247 /* wait for commit against the current transaction to become unblocked
 248  * when this is done, it is safe to start a new transaction, but the current
 249  * transaction might not be fully on disk.
 250  */
 251 static void wait_current_trans(struct btrfs_root *root)
 252 {
 253         struct btrfs_transaction *cur_trans;
 254
 255         spin_lock(&root->fs_info->trans_lock);
 256         cur_trans = root->fs_info->running_transaction;
 257         if (cur_trans && cur_trans->blocked) {
 258                 atomic_inc(&cur_trans->use_count);
 259                 spin_unlock(&root->fs_info->trans_lock);
 260
 261                 wait_event(root->fs_info->transaction_wait,
 262                            !cur_trans->blocked);
 263                 put_transaction(cur_trans);
 264         } else {
 265                 spin_unlock(&root->fs_info->trans_lock);
 266         }
 267 }
 268
 269 enum btrfs_trans_type {
 270         TRANS_START,
 271         TRANS_JOIN,
 272         TRANS_USERSPACE,
 273         TRANS_JOIN_NOLOCK,
 274 };
 275
 276 static int may_wait_transaction(struct btrfs_root *root, int type)
 277 {
 278         if (root->fs_info->log_root_recovering)
 279                 return 0;
 280
 281         if (type == TRANS_USERSPACE)
 282                 return 1;
 283
 284         if (type == TRANS_START &&
 285             !atomic_read(&root->fs_info->open_ioctl_trans))
 286                 return 1;
 287
 288         return 0;
 289 }
 290
 291 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 292                                                     u64 num_items, int type)
 293 {
 294         struct btrfs_trans_handle *h;
 295         struct btrfs_transaction *cur_trans;
 296         u64 num_bytes = 0;
 297         int ret;
 298
 299         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 300                 return ERR_PTR(-EROFS);
 301
 302         if (current->journal_info) {
 303                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 304                 h = current->journal_info;
 305                 h->use_count++;
 306                 h->orig_rsv = h->block_rsv;
 307                 h->block_rsv = NULL;
 308                 goto got_it;
 309         }
 310
 311         /*
 312          * Do the reservation before we join the transaction so we can do all
 313          * the appropriate flushing if need be.
 314          */
 315         if (num_items > 0 && root != root->fs_info->chunk_root) {
 316                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 317                 ret = btrfs_block_rsv_add(root,
 318                                           &root->fs_info->trans_block_rsv,
 319                                           num_bytes);
 320                 if (ret)
 321                         return ERR_PTR(ret);
 322         }
 323 again:
 324         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 325         if (!h)
 326                 return ERR_PTR(-ENOMEM);
 327
 328         if (may_wait_transaction(root, type))
 329                 wait_current_trans(root);
 330
 331         do {
 332                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 333                 if (ret == -EBUSY)
 334                         wait_current_trans(root);
 335         } while (ret == -EBUSY);
 336
 337         if (ret < 0) {
 338                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 339                 return ERR_PTR(ret);
 340         }
 341
 342         cur_trans = root->fs_info->running_transaction;
 343
 344         h->transid = cur_trans->transid;
 345         h->transaction = cur_trans;
 346         h->blocks_used = 0;
 347         h->bytes_reserved = 0;
 348         h->delayed_ref_updates = 0;
 349         h->use_count = 1;
 350         h->block_rsv = NULL;
 351         h->orig_rsv = NULL;
 352         h->aborted = 0;
 353
 354         smp_mb();
 355         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 356                 btrfs_commit_transaction(h, root);
 357                 goto again;
 358         }
 359
 360         if (num_bytes) {
 361                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 362                                               h->transid, num_bytes, 1);
 363                 h->block_rsv = &root->fs_info->trans_block_rsv;
 364                 h->bytes_reserved = num_bytes;
 365         }
 366
 367 got_it:
 368         btrfs_record_root_in_trans(h, root);
 369
 370         if (!current->journal_info && type != TRANS_USERSPACE)
 371                 current->journal_info = h;
 372         return h;
 373 }
 374
 375 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 376                                                    int num_items)
 377 {
 378         return start_transaction(root, num_items, TRANS_START);
 379 }
 380 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 381 {
 382         return start_transaction(root, 0, TRANS_JOIN);
 383 }
 384
 385 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 386 {
 387         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 388 }
 389
 390 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 391 {
 392         return start_transaction(root, 0, TRANS_USERSPACE);
 393 }
 394
 395 /* wait for a transaction commit to be fully complete */
 396 static noinline void wait_for_commit(struct btrfs_root *root,
 397                                     struct btrfs_transaction *commit)
 398 {
 399         wait_event(commit->commit_wait, commit->commit_done);
 400 }
 401
 402 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 403 {
 404         struct btrfs_transaction *cur_trans = NULL, *t;
 405         int ret;
 406
 407         ret = 0;
 408         if (transid) {
 409                 if (transid <= root->fs_info->last_trans_committed)
 410                         goto out;
 411
 412                 /* find specified transaction */
 413                 spin_lock(&root->fs_info->trans_lock);
 414                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 415                         if (t->transid == transid) {
 416                                 cur_trans = t;
 417                                 atomic_inc(&cur_trans->use_count);
 418                                 break;
 419                         }
 420                         if (t->transid > transid)
 421                                 break;
 422                 }
 423                 spin_unlock(&root->fs_info->trans_lock);
 424                 ret = -EINVAL;
 425                 if (!cur_trans)
 426                         goto out;  /* bad transid */
 427         } else {
 428                 /* find newest transaction that is committing | committed */
 429                 spin_lock(&root->fs_info->trans_lock);
 430                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 431                                             list) {
 432                         if (t->in_commit) {
 433                                 if (t->commit_done)
 434                                         break;
 435                                 cur_trans = t;
 436                                 atomic_inc(&cur_trans->use_count);
 437                                 break;
 438                         }
 439                 }
 440                 spin_unlock(&root->fs_info->trans_lock);
 441                 if (!cur_trans)
 442                         goto out;  /* nothing committing|committed */
 443         }
 444
 445         wait_for_commit(root, cur_trans);
 446
 447         put_transaction(cur_trans);
 448         ret = 0;
 449 out:
 450         return ret;
 451 }
 452
 453 void btrfs_throttle(struct btrfs_root *root)
 454 {
 455         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 456                 wait_current_trans(root);
 457 }
 458
 459 static int should_end_transaction(struct btrfs_trans_handle *trans,
 460                                   struct btrfs_root *root)
 461 {
 462         int ret;
 463
 464         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 465         return ret ? 1 : 0;
 466 }
 467
 468 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 469                                  struct btrfs_root *root)
 470 {
 471         struct btrfs_transaction *cur_trans = trans->transaction;
 472         struct btrfs_block_rsv *rsv = trans->block_rsv;
 473         int updates;
 474         int err;
 475
 476         smp_mb();
 477         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 478                 return 1;
 479
 480         /*
 481          * We need to do this in case we're deleting csums so the global block
 482          * rsv get's used instead of the csum block rsv.
 483          */
 484         trans->block_rsv = NULL;
 485
 486         updates = trans->delayed_ref_updates;
 487         trans->delayed_ref_updates = 0;
 488         if (updates) {
 489                 err = btrfs_run_delayed_refs(trans, root, updates);
 490                 if (err) /* Error code will also eval true */
 491                         return err;
 492         }
 493
 494         trans->block_rsv = rsv;
 495
 496         return should_end_transaction(trans, root);
 497 }
 498
 499 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 500                           struct btrfs_root *root, int throttle, int lock)
 501 {
 502         struct btrfs_transaction *cur_trans = trans->transaction;
 503         struct btrfs_fs_info *info = root->fs_info;
 504         int count = 0;
 505         int err = 0;
 506
 507         if (--trans->use_count) {
 508                 trans->block_rsv = trans->orig_rsv;
 509                 return 0;
 510         }
 511
 512         btrfs_trans_release_metadata(trans, root);
 513         trans->block_rsv = NULL;
 514         while (count < 2) {
 515                 unsigned long cur = trans->delayed_ref_updates;
 516                 trans->delayed_ref_updates = 0;
 517                 if (cur &&
 518                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 519                         trans->delayed_ref_updates = 0;
 520                         btrfs_run_delayed_refs(trans, root, cur);
 521                 } else {
 522                         break;
 523                 }
 524                 count++;
 525         }
 526
 527         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 528             should_end_transaction(trans, root)) {
 529                 trans->transaction->blocked = 1;
 530                 smp_wmb();
 531         }
 532
 533         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 534                 if (throttle) {
 535                         /*
 536                          * We may race with somebody else here so end up having
 537                          * to call end_transaction on ourselves again, so inc
 538                          * our use_count.
 539                          */
 540                         trans->use_count++;
 541                         return btrfs_commit_transaction(trans, root);
 542                 } else {
 543                         wake_up_process(info->transaction_kthread);
 544                 }
 545         }
 546
 547         WARN_ON(cur_trans != info->running_transaction);
 548         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 549         atomic_dec(&cur_trans->num_writers);
 550
 551         smp_mb();
 552         if (waitqueue_active(&cur_trans->writer_wait))
 553                 wake_up(&cur_trans->writer_wait);
 554         put_transaction(cur_trans);
 555
 556         if (current->journal_info == trans)
 557                 current->journal_info = NULL;
 558
 559         if (throttle)
 560                 btrfs_run_delayed_iputs(root);
 561
 562         if (trans->aborted ||
 563             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 564                 err = -EIO;
 565         }
 566
 567         memset(trans, 0, sizeof(*trans));
 568         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 569         return err;
 570 }
 571
 572 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 573                           struct btrfs_root *root)
 574 {
 575         int ret;
 576
 577         ret = __btrfs_end_transaction(trans, root, 0, 1);
 578         if (ret)
 579                 return ret;
 580         return 0;
 581 }
 582
 583 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 584                                    struct btrfs_root *root)
 585 {
 586         int ret;
 587
 588         ret = __btrfs_end_transaction(trans, root, 1, 1);
 589         if (ret)
 590                 return ret;
 591         return 0;
 592 }
 593
 594 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 595                                  struct btrfs_root *root)
 596 {
 597         int ret;
 598
 599         ret = __btrfs_end_transaction(trans, root, 0, 0);
 600         if (ret)
 601                 return ret;
 602         return 0;
 603 }
 604
 605 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 606                                 struct btrfs_root *root)
 607 {
 608         return __btrfs_end_transaction(trans, root, 1, 1);
 609 }
 610
 611 /*
 612  * when btree blocks are allocated, they have some corresponding bits set for
 613  * them in one of two extent_io trees.  This is used to make sure all of
 614  * those extents are sent to disk but does not wait on them
 615  */
 616 int btrfs_write_marked_extents(struct btrfs_root *root,
 617                                struct extent_io_tree *dirty_pages, int mark)
 618 {
 619         int err = 0;
 620         int werr = 0;
 621         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 622         u64 start = 0;
 623         u64 end;
 624
 625         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 626                                       mark)) {
 627                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 628                                    GFP_NOFS);
 629                 err = filemap_fdatawrite_range(mapping, start, end);
 630                 if (err)
 631                         werr = err;
 632                 cond_resched();
 633                 start = end + 1;
 634         }
 635         if (err)
 636                 werr = err;
 637         return werr;
 638 }
 639
 640 /*
 641  * when btree blocks are allocated, they have some corresponding bits set for
 642  * them in one of two extent_io trees.  This is used to make sure all of
 643  * those extents are on disk for transaction or log commit.  We wait
 644  * on all the pages and clear them from the dirty pages state tree
 645  */
 646 int btrfs_wait_marked_extents(struct btrfs_root *root,
 647                               struct extent_io_tree *dirty_pages, int mark)
 648 {
 649         int err = 0;
 650         int werr = 0;
 651         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 652         u64 start = 0;
 653         u64 end;
 654
 655         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 656                                       EXTENT_NEED_WAIT)) {
 657                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 658                 err = filemap_fdatawait_range(mapping, start, end);
 659                 if (err)
 660                         werr = err;
 661                 cond_resched();
 662                 start = end + 1;
 663         }
 664         if (err)
 665                 werr = err;
 666         return werr;
 667 }
 668
 669 /*
 670  * when btree blocks are allocated, they have some corresponding bits set for
 671  * them in one of two extent_io trees.  This is used to make sure all of
 672  * those extents are on disk for transaction or log commit
 673  */
 674 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 675                                 struct extent_io_tree *dirty_pages, int mark)
 676 {
 677         int ret;
 678         int ret2;
 679
 680         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 681         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 682
 683         if (ret)
 684                 return ret;
 685         if (ret2)
 686                 return ret2;
 687         return 0;
 688 }
 689
 690 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 691                                      struct btrfs_root *root)
 692 {
 693         if (!trans || !trans->transaction) {
 694                 struct inode *btree_inode;
 695                 btree_inode = root->fs_info->btree_inode;
 696                 return filemap_write_and_wait(btree_inode->i_mapping);
 697         }
 698         return btrfs_write_and_wait_marked_extents(root,
 699                                            &trans->transaction->dirty_pages,
 700                                            EXTENT_DIRTY);
 701 }
 702
 703 /*
 704  * this is used to update the root pointer in the tree of tree roots.
 705  *
 706  * But, in the case of the extent allocation tree, updating the root
 707  * pointer may allocate blocks which may change the root of the extent
 708  * allocation tree.
 709  *
 710  * So, this loops and repeats and makes sure the cowonly root didn't
 711  * change while the root pointer was being updated in the metadata.
 712  */
 713 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 714                                struct btrfs_root *root)
 715 {
 716         int ret;
 717         u64 old_root_bytenr;
 718         u64 old_root_used;
 719         struct btrfs_root *tree_root = root->fs_info->tree_root;
 720
 721         old_root_used = btrfs_root_used(&root->root_item);
 722         btrfs_write_dirty_block_groups(trans, root);
 723
 724         while (1) {
 725                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 726                 if (old_root_bytenr == root->node->start &&
 727                     old_root_used == btrfs_root_used(&root->root_item))
 728                         break;
 729
 730                 btrfs_set_root_node(&root->root_item, root->node);
 731                 ret = btrfs_update_root(trans, tree_root,
 732                                         &root->root_key,
 733                                         &root->root_item);
 734                 if (ret)
 735                         return ret;
 736
 737                 old_root_used = btrfs_root_used(&root->root_item);
 738                 ret = btrfs_write_dirty_block_groups(trans, root);
 739                 if (ret)
 740                         return ret;
 741         }
 742
 743         if (root != root->fs_info->extent_root)
 744                 switch_commit_root(root);
 745
 746         return 0;
 747 }
 748
 749 /*
 750  * update all the cowonly tree roots on disk
 751  *
 752  * The error handling in this function may not be obvious. Any of the
 753  * failures will cause the file system to go offline. We still need
 754  * to clean up the delayed refs.
 755  */
 756 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 757                                          struct btrfs_root *root)
 758 {
 759         struct btrfs_fs_info *fs_info = root->fs_info;
 760         struct list_head *next;
 761         struct extent_buffer *eb;
 762         int ret;
 763
 764         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 765         if (ret)
 766                 return ret;
 767
 768         eb = btrfs_lock_root_node(fs_info->tree_root);
 769         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 770                               0, &eb);
 771         btrfs_tree_unlock(eb);
 772         free_extent_buffer(eb);
 773
 774         if (ret)
 775                 return ret;
 776
 777         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 778         if (ret)
 779                 return ret;
 780
 781         ret = btrfs_run_dev_stats(trans, root->fs_info);
 782         BUG_ON(ret);
 783
 784         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 785                 next = fs_info->dirty_cowonly_roots.next;
 786                 list_del_init(next);
 787                 root = list_entry(next, struct btrfs_root, dirty_list);
 788
 789                 ret = update_cowonly_root(trans, root);
 790                 if (ret)
 791                         return ret;
 792         }
 793
 794         down_write(&fs_info->extent_commit_sem);
 795         switch_commit_root(fs_info->extent_root);
 796         up_write(&fs_info->extent_commit_sem);
 797
 798         return 0;
 799 }
 800
 801 /*
 802  * dead roots are old snapshots that need to be deleted.  This allocates
 803  * a dirty root struct and adds it into the list of dead roots that need to
 804  * be deleted
 805  */
 806 int btrfs_add_dead_root(struct btrfs_root *root)
 807 {
 808         spin_lock(&root->fs_info->trans_lock);
 809         list_add(&root->root_list, &root->fs_info->dead_roots);
 810         spin_unlock(&root->fs_info->trans_lock);
 811         return 0;
 812 }
 813
 814 /*
 815  * update all the cowonly tree roots on disk
 816  */
 817 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 818                                     struct btrfs_root *root)
 819 {
 820         struct btrfs_root *gang[8];
 821         struct btrfs_fs_info *fs_info = root->fs_info;
 822         int i;
 823         int ret;
 824         int err = 0;
 825
 826         spin_lock(&fs_info->fs_roots_radix_lock);
 827         while (1) {
 828                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 829                                                  (void **)gang, 0,
 830                                                  ARRAY_SIZE(gang),
 831                                                  BTRFS_ROOT_TRANS_TAG);
 832                 if (ret == 0)
 833                         break;
 834                 for (i = 0; i < ret; i++) {
 835                         root = gang[i];
 836                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 837                                         (unsigned long)root->root_key.objectid,
 838                                         BTRFS_ROOT_TRANS_TAG);
 839                         spin_unlock(&fs_info->fs_roots_radix_lock);
 840
 841                         btrfs_free_log(trans, root);
 842                         btrfs_update_reloc_root(trans, root);
 843                         btrfs_orphan_commit_root(trans, root);
 844
 845                         btrfs_save_ino_cache(root, trans);
 846
 847                         /* see comments in should_cow_block() */
 848                         root->force_cow = 0;
 849                         smp_wmb();
 850
 851                         if (root->commit_root != root->node) {
 852                                 mutex_lock(&root->fs_commit_mutex);
 853                                 switch_commit_root(root);
 854                                 btrfs_unpin_free_ino(root);
 855                                 mutex_unlock(&root->fs_commit_mutex);
 856
 857                                 btrfs_set_root_node(&root->root_item,
 858                                                     root->node);
 859                         }
 860
 861                         err = btrfs_update_root(trans, fs_info->tree_root,
 862                                                 &root->root_key,
 863                                                 &root->root_item);
 864                         spin_lock(&fs_info->fs_roots_radix_lock);
 865                         if (err)
 866                                 break;
 867                 }
 868         }
 869         spin_unlock(&fs_info->fs_roots_radix_lock);
 870         return err;
 871 }
 872
 873 /*
 874  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 875  * otherwise every leaf in the btree is read and defragged.
 876  */
 877 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 878 {
 879         struct btrfs_fs_info *info = root->fs_info;
 880         struct btrfs_trans_handle *trans;
 881         int ret;
 882         unsigned long nr;
 883
 884         if (xchg(&root->defrag_running, 1))
 885                 return 0;
 886
 887         while (1) {
 888                 trans = btrfs_start_transaction(root, 0);
 889                 if (IS_ERR(trans))
 890                         return PTR_ERR(trans);
 891
 892                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 893
 894                 nr = trans->blocks_used;
 895                 btrfs_end_transaction(trans, root);
 896                 btrfs_btree_balance_dirty(info->tree_root, nr);
 897                 cond_resched();
 898
 899                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 900                         break;
 901         }
 902         root->defrag_running = 0;
 903         return ret;
 904 }
 905
 906 /*
 907  * new snapshots need to be created at a very specific time in the
 908  * transaction commit.  This does the actual creation
 909  */
 910 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 911                                    struct btrfs_fs_info *fs_info,
 912                                    struct btrfs_pending_snapshot *pending)
 913 {
 914         struct btrfs_key key;
 915         struct btrfs_root_item *new_root_item;
 916         struct btrfs_root *tree_root = fs_info->tree_root;
 917         struct btrfs_root *root = pending->root;
 918         struct btrfs_root *parent_root;
 919         struct btrfs_block_rsv *rsv;
 920         struct inode *parent_inode;
 921         struct dentry *parent;
 922         struct dentry *dentry;
 923         struct extent_buffer *tmp;
 924         struct extent_buffer *old;
 925         int ret;
 926         u64 to_reserve = 0;
 927         u64 index = 0;
 928         u64 objectid;
 929         u64 root_flags;
 930
 931         rsv = trans->block_rsv;
 932
 933         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 934         if (!new_root_item) {
 935                 ret = pending->error = -ENOMEM;
 936                 goto fail;
 937         }
 938
 939         ret = btrfs_find_free_objectid(tree_root, &objectid);
 940         if (ret) {
 941                 pending->error = ret;
 942                 goto fail;
 943         }
 944
 945         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 946
 947         if (to_reserve > 0) {
 948                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 949                                                   to_reserve);
 950                 if (ret) {
 951                         pending->error = ret;
 952                         goto fail;
 953                 }
 954         }
 955
 956         key.objectid = objectid;
 957         key.offset = (u64)-1;
 958         key.type = BTRFS_ROOT_ITEM_KEY;
 959
 960         trans->block_rsv = &pending->block_rsv;
 961
 962         dentry = pending->dentry;
 963         parent = dget_parent(dentry);
 964         parent_inode = parent->d_inode;
 965         parent_root = BTRFS_I(parent_inode)->root;
 966         record_root_in_trans(trans, parent_root);
 967
 968         /*
 969          * insert the directory item
 970          */
 971         ret = btrfs_set_inode_index(parent_inode, &index);
 972         BUG_ON(ret); /* -ENOMEM */
 973         ret = btrfs_insert_dir_item(trans, parent_root,
 974                                 dentry->d_name.name, dentry->d_name.len,
 975                                 parent_inode, &key,
 976                                 BTRFS_FT_DIR, index);
 977         if (ret == -EEXIST) {
 978                 pending->error = -EEXIST;
 979                 dput(parent);
 980                 goto fail;
 981         } else if (ret) {
 982                 goto abort_trans_dput;
 983         }
 984
 985         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 986                                          dentry->d_name.len * 2);
 987         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 988         if (ret)
 989                 goto abort_trans_dput;
 990
 991         /*
 992          * pull in the delayed directory update
 993          * and the delayed inode item
 994          * otherwise we corrupt the FS during
 995          * snapshot
 996          */
 997         ret = btrfs_run_delayed_items(trans, root);
 998         if (ret) { /* Transaction aborted */
 999                 dput(parent);
1000                 goto fail;
1001         }
1002
1003         record_root_in_trans(trans, root);
1004         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1005         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1006         btrfs_check_and_init_root_item(new_root_item);
1007
1008         root_flags = btrfs_root_flags(new_root_item);
1009         if (pending->readonly)
1010                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1011         else
1012                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1013         btrfs_set_root_flags(new_root_item, root_flags);
1014
1015         old = btrfs_lock_root_node(root);
1016         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1017         if (ret) {
1018                 btrfs_tree_unlock(old);
1019                 free_extent_buffer(old);
1020                 goto abort_trans_dput;
1021         }
1022
1023         btrfs_set_lock_blocking(old);
1024
1025         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1026         /* clean up in any case */
1027         btrfs_tree_unlock(old);
1028         free_extent_buffer(old);
1029         if (ret)
1030                 goto abort_trans_dput;
1031
1032         /* see comments in should_cow_block() */
1033         root->force_cow = 1;
1034         smp_wmb();
1035
1036         btrfs_set_root_node(new_root_item, tmp);
1037         /* record when the snapshot was created in key.offset */
1038         key.offset = trans->transid;
1039         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1040         btrfs_tree_unlock(tmp);
1041         free_extent_buffer(tmp);
1042         if (ret)
1043                 goto abort_trans_dput;
1044
1045         /*
1046          * insert root back/forward references
1047          */
1048         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1049                                  parent_root->root_key.objectid,
1050                                  btrfs_ino(parent_inode), index,
1051                                  dentry->d_name.name, dentry->d_name.len);
1052         dput(parent);
1053         if (ret)
1054                 goto fail;
1055
1056         key.offset = (u64)-1;
1057         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1058         if (IS_ERR(pending->snap)) {
1059                 ret = PTR_ERR(pending->snap);
1060                 goto abort_trans;
1061         }
1062
1063         ret = btrfs_reloc_post_snapshot(trans, pending);
1064         if (ret)
1065                 goto abort_trans;
1066         ret = 0;
1067 fail:
1068         kfree(new_root_item);
1069         trans->block_rsv = rsv;
1070         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1071         return ret;
1072
1073 abort_trans_dput:
1074         dput(parent);
1075 abort_trans:
1076         btrfs_abort_transaction(trans, root, ret);
1077         goto fail;
1078 }
1079
1080 /*
1081  * create all the snapshots we've scheduled for creation
1082  */
1083 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1084                                              struct btrfs_fs_info *fs_info)
1085 {
1086         struct btrfs_pending_snapshot *pending;
1087         struct list_head *head = &trans->transaction->pending_snapshots;
1088
1089         list_for_each_entry(pending, head, list)
1090                 create_pending_snapshot(trans, fs_info, pending);
1091         return 0;
1092 }
1093
1094 static void update_super_roots(struct btrfs_root *root)
1095 {
1096         struct btrfs_root_item *root_item;
1097         struct btrfs_super_block *super;
1098
1099         super = root->fs_info->super_copy;
1100
1101         root_item = &root->fs_info->chunk_root->root_item;
1102         super->chunk_root = root_item->bytenr;
1103         super->chunk_root_generation = root_item->generation;
1104         super->chunk_root_level = root_item->level;
1105
1106         root_item = &root->fs_info->tree_root->root_item;
1107         super->root = root_item->bytenr;
1108         super->generation = root_item->generation;
1109         super->root_level = root_item->level;
1110         if (btrfs_test_opt(root, SPACE_CACHE))
1111                 super->cache_generation = root_item->generation;
1112 }
1113
1114 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1115 {
1116         int ret = 0;
1117         spin_lock(&info->trans_lock);
1118         if (info->running_transaction)
1119                 ret = info->running_transaction->in_commit;
1120         spin_unlock(&info->trans_lock);
1121         return ret;
1122 }
1123
1124 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1125 {
1126         int ret = 0;
1127         spin_lock(&info->trans_lock);
1128         if (info->running_transaction)
1129                 ret = info->running_transaction->blocked;
1130         spin_unlock(&info->trans_lock);
1131         return ret;
1132 }
1133
1134 /*
1135  * wait for the current transaction commit to start and block subsequent
1136  * transaction joins
1137  */
1138 static void wait_current_trans_commit_start(struct btrfs_root *root,
1139                                             struct btrfs_transaction *trans)
1140 {
1141         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1142 }
1143
1144 /*
1145  * wait for the current transaction to start and then become unblocked.
1146  * caller holds ref.
1147  */
1148 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1149                                          struct btrfs_transaction *trans)
1150 {
1151         wait_event(root->fs_info->transaction_wait,
1152                    trans->commit_done || (trans->in_commit && !trans->blocked));
1153 }
1154
1155 /*
1156  * commit transactions asynchronously. once btrfs_commit_transaction_async
1157  * returns, any subsequent transaction will not be allowed to join.
1158  */
1159 struct btrfs_async_commit {
1160         struct btrfs_trans_handle *newtrans;
1161         struct btrfs_root *root;
1162         struct delayed_work work;
1163 };
1164
1165 static void do_async_commit(struct work_struct *work)
1166 {
1167         struct btrfs_async_commit *ac =
1168                 container_of(work, struct btrfs_async_commit, work.work);
1169
1170         btrfs_commit_transaction(ac->newtrans, ac->root);
1171         kfree(ac);
1172 }
1173
1174 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1175                                    struct btrfs_root *root,
1176                                    int wait_for_unblock)
1177 {
1178         struct btrfs_async_commit *ac;
1179         struct btrfs_transaction *cur_trans;
1180
1181         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1182         if (!ac)
1183                 return -ENOMEM;
1184
1185         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1186         ac->root = root;
1187         ac->newtrans = btrfs_join_transaction(root);
1188         if (IS_ERR(ac->newtrans)) {
1189                 int err = PTR_ERR(ac->newtrans);
1190                 kfree(ac);
1191                 return err;
1192         }
1193
1194         /* take transaction reference */
1195         cur_trans = trans->transaction;
1196         atomic_inc(&cur_trans->use_count);
1197
1198         btrfs_end_transaction(trans, root);
1199         schedule_delayed_work(&ac->work, 0);
1200
1201         /* wait for transaction to start and unblock */
1202         if (wait_for_unblock)
1203                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1204         else
1205                 wait_current_trans_commit_start(root, cur_trans);
1206
1207         if (current->journal_info == trans)
1208                 current->journal_info = NULL;
1209
1210         put_transaction(cur_trans);
1211         return 0;
1212 }
1213
1214
1215 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1216                                 struct btrfs_root *root)
1217 {
1218         struct btrfs_transaction *cur_trans = trans->transaction;
1219
1220         WARN_ON(trans->use_count > 1);
1221
1222         spin_lock(&root->fs_info->trans_lock);
1223         list_del_init(&cur_trans->list);
1224         spin_unlock(&root->fs_info->trans_lock);
1225
1226         btrfs_cleanup_one_transaction(trans->transaction, root);
1227
1228         put_transaction(cur_trans);
1229         put_transaction(cur_trans);
1230
1231         trace_btrfs_transaction_commit(root);
1232
1233         btrfs_scrub_continue(root);
1234
1235         if (current->journal_info == trans)
1236                 current->journal_info = NULL;
1237
1238         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1239 }
1240
1241 /*
1242  * btrfs_transaction state sequence:
1243  *    in_commit = 0, blocked = 0  (initial)
1244  *    in_commit = 1, blocked = 1
1245  *    blocked = 0
1246  *    commit_done = 1
1247  */
1248 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1249                              struct btrfs_root *root)
1250 {
1251         unsigned long joined = 0;
1252         struct btrfs_transaction *cur_trans = trans->transaction;
1253         struct btrfs_transaction *prev_trans = NULL;
1254         DEFINE_WAIT(wait);
1255         int ret = -EIO;
1256         int should_grow = 0;
1257         unsigned long now = get_seconds();
1258         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1259
1260         btrfs_run_ordered_operations(root, 0);
1261
1262         btrfs_trans_release_metadata(trans, root);
1263         trans->block_rsv = NULL;
1264
1265         if (cur_trans->aborted)
1266                 goto cleanup_transaction;
1267
1268         /* make a pass through all the delayed refs we have so far
1269          * any runnings procs may add more while we are here
1270          */
1271         ret = btrfs_run_delayed_refs(trans, root, 0);
1272         if (ret)
1273                 goto cleanup_transaction;
1274
1275         cur_trans = trans->transaction;
1276
1277         /*
1278          * set the flushing flag so procs in this transaction have to
1279          * start sending their work down.
1280          */
1281         cur_trans->delayed_refs.flushing = 1;
1282
1283         ret = btrfs_run_delayed_refs(trans, root, 0);
1284         if (ret)
1285                 goto cleanup_transaction;
1286
1287         spin_lock(&cur_trans->commit_lock);
1288         if (cur_trans->in_commit) {
1289                 spin_unlock(&cur_trans->commit_lock);
1290                 atomic_inc(&cur_trans->use_count);
1291                 ret = btrfs_end_transaction(trans, root);
1292
1293                 wait_for_commit(root, cur_trans);
1294
1295                 put_transaction(cur_trans);
1296
1297                 return ret;
1298         }
1299
1300         trans->transaction->in_commit = 1;
1301         trans->transaction->blocked = 1;
1302         spin_unlock(&cur_trans->commit_lock);
1303         wake_up(&root->fs_info->transaction_blocked_wait);
1304
1305         spin_lock(&root->fs_info->trans_lock);
1306         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1307                 prev_trans = list_entry(cur_trans->list.prev,
1308                                         struct btrfs_transaction, list);
1309                 if (!prev_trans->commit_done) {
1310                         atomic_inc(&prev_trans->use_count);
1311                         spin_unlock(&root->fs_info->trans_lock);
1312
1313                         wait_for_commit(root, prev_trans);
1314
1315                         put_transaction(prev_trans);
1316                 } else {
1317                         spin_unlock(&root->fs_info->trans_lock);
1318                 }
1319         } else {
1320                 spin_unlock(&root->fs_info->trans_lock);
1321         }
1322
1323         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1324                 should_grow = 1;
1325
1326         do {
1327                 int snap_pending = 0;
1328
1329                 joined = cur_trans->num_joined;
1330                 if (!list_empty(&trans->transaction->pending_snapshots))
1331                         snap_pending = 1;
1332
1333                 WARN_ON(cur_trans != trans->transaction);
1334
1335                 if (flush_on_commit || snap_pending) {
1336                         btrfs_start_delalloc_inodes(root, 1);
1337                         btrfs_wait_ordered_extents(root, 0, 1);
1338                 }
1339
1340                 ret = btrfs_run_delayed_items(trans, root);
1341                 if (ret)
1342                         goto cleanup_transaction;
1343
1344                 /*
1345                  * rename don't use btrfs_join_transaction, so, once we
1346                  * set the transaction to blocked above, we aren't going
1347                  * to get any new ordered operations.  We can safely run
1348                  * it here and no for sure that nothing new will be added
1349                  * to the list
1350                  */
1351                 btrfs_run_ordered_operations(root, 1);
1352
1353                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1354                                 TASK_UNINTERRUPTIBLE);
1355
1356                 if (atomic_read(&cur_trans->num_writers) > 1)
1357                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1358                 else if (should_grow)
1359                         schedule_timeout(1);
1360
1361                 finish_wait(&cur_trans->writer_wait, &wait);
1362         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1363                  (should_grow && cur_trans->num_joined != joined));
1364
1365         /*
1366          * Ok now we need to make sure to block out any other joins while we
1367          * commit the transaction.  We could have started a join before setting
1368          * no_join so make sure to wait for num_writers to == 1 again.
1369          */
1370         spin_lock(&root->fs_info->trans_lock);
1371         root->fs_info->trans_no_join = 1;
1372         spin_unlock(&root->fs_info->trans_lock);
1373         wait_event(cur_trans->writer_wait,
1374                    atomic_read(&cur_trans->num_writers) == 1);
1375
1376         /*
1377          * the reloc mutex makes sure that we stop
1378          * the balancing code from coming in and moving
1379          * extents around in the middle of the commit
1380          */
1381         mutex_lock(&root->fs_info->reloc_mutex);
1382
1383         ret = btrfs_run_delayed_items(trans, root);
1384         if (ret) {
1385                 mutex_unlock(&root->fs_info->reloc_mutex);
1386                 goto cleanup_transaction;
1387         }
1388
1389         ret = create_pending_snapshots(trans, root->fs_info);
1390         if (ret) {
1391                 mutex_unlock(&root->fs_info->reloc_mutex);
1392                 goto cleanup_transaction;
1393         }
1394
1395         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1396         if (ret) {
1397                 mutex_unlock(&root->fs_info->reloc_mutex);
1398                 goto cleanup_transaction;
1399         }
1400
1401         /*
1402          * make sure none of the code above managed to slip in a
1403          * delayed item
1404          */
1405         btrfs_assert_delayed_root_empty(root);
1406
1407         WARN_ON(cur_trans != trans->transaction);
1408
1409         btrfs_scrub_pause(root);
1410         /* btrfs_commit_tree_roots is responsible for getting the
1411          * various roots consistent with each other.  Every pointer
1412          * in the tree of tree roots has to point to the most up to date
1413          * root for every subvolume and other tree.  So, we have to keep
1414          * the tree logging code from jumping in and changing any
1415          * of the trees.
1416          *
1417          * At this point in the commit, there can't be any tree-log
1418          * writers, but a little lower down we drop the trans mutex
1419          * and let new people in.  By holding the tree_log_mutex
1420          * from now until after the super is written, we avoid races
1421          * with the tree-log code.
1422          */
1423         mutex_lock(&root->fs_info->tree_log_mutex);
1424
1425         ret = commit_fs_roots(trans, root);
1426         if (ret) {
1427                 mutex_unlock(&root->fs_info->tree_log_mutex);
1428                 mutex_unlock(&root->fs_info->reloc_mutex);
1429                 goto cleanup_transaction;
1430         }
1431
1432         /* commit_fs_roots gets rid of all the tree log roots, it is now
1433          * safe to free the root of tree log roots
1434          */
1435         btrfs_free_log_root_tree(trans, root->fs_info);
1436
1437         ret = commit_cowonly_roots(trans, root);
1438         if (ret) {
1439                 mutex_unlock(&root->fs_info->tree_log_mutex);
1440                 mutex_unlock(&root->fs_info->reloc_mutex);
1441                 goto cleanup_transaction;
1442         }
1443
1444         btrfs_prepare_extent_commit(trans, root);
1445
1446         cur_trans = root->fs_info->running_transaction;
1447
1448         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1449                             root->fs_info->tree_root->node);
1450         switch_commit_root(root->fs_info->tree_root);
1451
1452         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1453                             root->fs_info->chunk_root->node);
1454         switch_commit_root(root->fs_info->chunk_root);
1455
1456         update_super_roots(root);
1457
1458         if (!root->fs_info->log_root_recovering) {
1459                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1460                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1461         }
1462
1463         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1464                sizeof(*root->fs_info->super_copy));
1465
1466         trans->transaction->blocked = 0;
1467         spin_lock(&root->fs_info->trans_lock);
1468         root->fs_info->running_transaction = NULL;
1469         root->fs_info->trans_no_join = 0;
1470         spin_unlock(&root->fs_info->trans_lock);
1471         mutex_unlock(&root->fs_info->reloc_mutex);
1472
1473         wake_up(&root->fs_info->transaction_wait);
1474
1475         ret = btrfs_write_and_wait_transaction(trans, root);
1476         if (ret) {
1477                 btrfs_error(root->fs_info, ret,
1478                             "Error while writing out transaction.");
1479                 mutex_unlock(&root->fs_info->tree_log_mutex);
1480                 goto cleanup_transaction;
1481         }
1482
1483         ret = write_ctree_super(trans, root, 0);
1484         if (ret) {
1485                 mutex_unlock(&root->fs_info->tree_log_mutex);
1486                 goto cleanup_transaction;
1487         }
1488
1489         /*
1490          * the super is written, we can safely allow the tree-loggers
1491          * to go about their business
1492          */
1493         mutex_unlock(&root->fs_info->tree_log_mutex);
1494
1495         btrfs_finish_extent_commit(trans, root);
1496
1497         cur_trans->commit_done = 1;
1498
1499         root->fs_info->last_trans_committed = cur_trans->transid;
1500
1501         wake_up(&cur_trans->commit_wait);
1502
1503         spin_lock(&root->fs_info->trans_lock);
1504         list_del_init(&cur_trans->list);
1505         spin_unlock(&root->fs_info->trans_lock);
1506
1507         put_transaction(cur_trans);
1508         put_transaction(cur_trans);
1509
1510         trace_btrfs_transaction_commit(root);
1511
1512         btrfs_scrub_continue(root);
1513
1514         if (current->journal_info == trans)
1515                 current->journal_info = NULL;
1516
1517         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1518
1519         if (current != root->fs_info->transaction_kthread)
1520                 btrfs_run_delayed_iputs(root);
1521
1522         return ret;
1523
1524 cleanup_transaction:
1525         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1526 //      WARN_ON(1);
1527         if (current->journal_info == trans)
1528                 current->journal_info = NULL;
1529         cleanup_transaction(trans, root);
1530
1531         return ret;
1532 }
1533
1534 /*
1535  * interface function to delete all the snapshots we have scheduled for deletion
1536  */
1537 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1538 {
1539         LIST_HEAD(list);
1540         struct btrfs_fs_info *fs_info = root->fs_info;
1541
1542         spin_lock(&fs_info->trans_lock);
1543         list_splice_init(&fs_info->dead_roots, &list);
1544         spin_unlock(&fs_info->trans_lock);
1545
1546         while (!list_empty(&list)) {
1547                 int ret;
1548
1549                 root = list_entry(list.next, struct btrfs_root, root_list);
1550                 list_del(&root->root_list);
1551
1552                 btrfs_kill_all_delayed_nodes(root);
1553
1554                 if (btrfs_header_backref_rev(root->node) <
1555                     BTRFS_MIXED_BACKREF_REV)
1556                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1557                 else
1558                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1559                 BUG_ON(ret < 0);
1560         }
1561         return 0;
1562 }