Pileus Git - ~andy/linux/blob - 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                 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         struct btrfs_fs_info *fs_info = root->fs_info;
  59
  60         spin_lock(&fs_info->trans_lock);
  61 loop:
  62         /* The file system has been taken offline. No new transactions. */
  63         if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  64                 spin_unlock(&fs_info->trans_lock);
  65                 return -EROFS;
  66         }
  67
  68         if (fs_info->trans_no_join) {
  69                 if (!nofail) {
  70                         spin_unlock(&fs_info->trans_lock);
  71                         return -EBUSY;
  72                 }
  73         }
  74
  75         cur_trans = fs_info->running_transaction;
  76         if (cur_trans) {
  77                 if (cur_trans->aborted) {
  78                         spin_unlock(&fs_info->trans_lock);
  79                         return cur_trans->aborted;
  80                 }
  81                 atomic_inc(&cur_trans->use_count);
  82                 atomic_inc(&cur_trans->num_writers);
  83                 cur_trans->num_joined++;
  84                 spin_unlock(&fs_info->trans_lock);
  85                 return 0;
  86         }
  87         spin_unlock(&fs_info->trans_lock);
  88
  89         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  90         if (!cur_trans)
  91                 return -ENOMEM;
  92
  93         spin_lock(&fs_info->trans_lock);
  94         if (fs_info->running_transaction) {
  95                 /*
  96                  * someone started a transaction after we unlocked.  Make sure
  97                  * to redo the trans_no_join checks above
  98                  */
  99                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 100                 cur_trans = fs_info->running_transaction;
 101                 goto loop;
 102         } else if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 103                 spin_unlock(&root->fs_info->trans_lock);
 104                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 105                 return -EROFS;
 106         }
 107
 108         atomic_set(&cur_trans->num_writers, 1);
 109         cur_trans->num_joined = 0;
 110         init_waitqueue_head(&cur_trans->writer_wait);
 111         init_waitqueue_head(&cur_trans->commit_wait);
 112         cur_trans->in_commit = 0;
 113         cur_trans->blocked = 0;
 114         /*
 115          * One for this trans handle, one so it will live on until we
 116          * commit the transaction.
 117          */
 118         atomic_set(&cur_trans->use_count, 2);
 119         cur_trans->commit_done = 0;
 120         cur_trans->start_time = get_seconds();
 121
 122         cur_trans->delayed_refs.root = RB_ROOT;
 123         cur_trans->delayed_refs.num_entries = 0;
 124         cur_trans->delayed_refs.num_heads_ready = 0;
 125         cur_trans->delayed_refs.num_heads = 0;
 126         cur_trans->delayed_refs.flushing = 0;
 127         cur_trans->delayed_refs.run_delayed_start = 0;
 128
 129         /*
 130          * although the tree mod log is per file system and not per transaction,
 131          * the log must never go across transaction boundaries.
 132          */
 133         smp_mb();
 134         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 135                 printk(KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 136                         "creating a fresh transaction\n");
 137                 WARN_ON(1);
 138         }
 139         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) {
 140                 printk(KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 141                         "creating a fresh transaction\n");
 142                 WARN_ON(1);
 143         }
 144         atomic_set(&fs_info->tree_mod_seq, 0);
 145
 146         spin_lock_init(&cur_trans->commit_lock);
 147         spin_lock_init(&cur_trans->delayed_refs.lock);
 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->root = root;
 349         h->delayed_ref_updates = 0;
 350         h->use_count = 1;
 351         h->block_rsv = NULL;
 352         h->orig_rsv = NULL;
 353         h->aborted = 0;
 354         h->delayed_ref_elem.seq = 0;
 355         INIT_LIST_HEAD(&h->qgroup_ref_list);
 356
 357         smp_mb();
 358         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 359                 btrfs_commit_transaction(h, root);
 360                 goto again;
 361         }
 362
 363         if (num_bytes) {
 364                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 365                                               h->transid, num_bytes, 1);
 366                 h->block_rsv = &root->fs_info->trans_block_rsv;
 367                 h->bytes_reserved = num_bytes;
 368         }
 369
 370 got_it:
 371         btrfs_record_root_in_trans(h, root);
 372
 373         if (!current->journal_info && type != TRANS_USERSPACE)
 374                 current->journal_info = h;
 375         return h;
 376 }
 377
 378 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 379                                                    int num_items)
 380 {
 381         return start_transaction(root, num_items, TRANS_START);
 382 }
 383 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 384 {
 385         return start_transaction(root, 0, TRANS_JOIN);
 386 }
 387
 388 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 389 {
 390         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 391 }
 392
 393 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 394 {
 395         return start_transaction(root, 0, TRANS_USERSPACE);
 396 }
 397
 398 /* wait for a transaction commit to be fully complete */
 399 static noinline void wait_for_commit(struct btrfs_root *root,
 400                                     struct btrfs_transaction *commit)
 401 {
 402         wait_event(commit->commit_wait, commit->commit_done);
 403 }
 404
 405 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 406 {
 407         struct btrfs_transaction *cur_trans = NULL, *t;
 408         int ret;
 409
 410         ret = 0;
 411         if (transid) {
 412                 if (transid <= root->fs_info->last_trans_committed)
 413                         goto out;
 414
 415                 /* find specified transaction */
 416                 spin_lock(&root->fs_info->trans_lock);
 417                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 418                         if (t->transid == transid) {
 419                                 cur_trans = t;
 420                                 atomic_inc(&cur_trans->use_count);
 421                                 break;
 422                         }
 423                         if (t->transid > transid)
 424                                 break;
 425                 }
 426                 spin_unlock(&root->fs_info->trans_lock);
 427                 ret = -EINVAL;
 428                 if (!cur_trans)
 429                         goto out;  /* bad transid */
 430         } else {
 431                 /* find newest transaction that is committing | committed */
 432                 spin_lock(&root->fs_info->trans_lock);
 433                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 434                                             list) {
 435                         if (t->in_commit) {
 436                                 if (t->commit_done)
 437                                         break;
 438                                 cur_trans = t;
 439                                 atomic_inc(&cur_trans->use_count);
 440                                 break;
 441                         }
 442                 }
 443                 spin_unlock(&root->fs_info->trans_lock);
 444                 if (!cur_trans)
 445                         goto out;  /* nothing committing|committed */
 446         }
 447
 448         wait_for_commit(root, cur_trans);
 449
 450         put_transaction(cur_trans);
 451         ret = 0;
 452 out:
 453         return ret;
 454 }
 455
 456 void btrfs_throttle(struct btrfs_root *root)
 457 {
 458         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 459                 wait_current_trans(root);
 460 }
 461
 462 static int should_end_transaction(struct btrfs_trans_handle *trans,
 463                                   struct btrfs_root *root)
 464 {
 465         int ret;
 466
 467         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 468         return ret ? 1 : 0;
 469 }
 470
 471 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 472                                  struct btrfs_root *root)
 473 {
 474         struct btrfs_transaction *cur_trans = trans->transaction;
 475         struct btrfs_block_rsv *rsv = trans->block_rsv;
 476         int updates;
 477         int err;
 478
 479         smp_mb();
 480         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 481                 return 1;
 482
 483         /*
 484          * We need to do this in case we're deleting csums so the global block
 485          * rsv get's used instead of the csum block rsv.
 486          */
 487         trans->block_rsv = NULL;
 488
 489         updates = trans->delayed_ref_updates;
 490         trans->delayed_ref_updates = 0;
 491         if (updates) {
 492                 err = btrfs_run_delayed_refs(trans, root, updates);
 493                 if (err) /* Error code will also eval true */
 494                         return err;
 495         }
 496
 497         trans->block_rsv = rsv;
 498
 499         return should_end_transaction(trans, root);
 500 }
 501
 502 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 503                           struct btrfs_root *root, int throttle, int lock)
 504 {
 505         struct btrfs_transaction *cur_trans = trans->transaction;
 506         struct btrfs_fs_info *info = root->fs_info;
 507         int count = 0;
 508         int err = 0;
 509
 510         if (--trans->use_count) {
 511                 trans->block_rsv = trans->orig_rsv;
 512                 return 0;
 513         }
 514
 515         /*
 516          * do the qgroup accounting as early as possible
 517          */
 518         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 519
 520         btrfs_trans_release_metadata(trans, root);
 521         trans->block_rsv = NULL;
 522         /*
 523          * the same root has to be passed to start_transaction and
 524          * end_transaction. Subvolume quota depends on this.
 525          */
 526         WARN_ON(trans->root != root);
 527         while (count < 2) {
 528                 unsigned long cur = trans->delayed_ref_updates;
 529                 trans->delayed_ref_updates = 0;
 530                 if (cur &&
 531                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 532                         trans->delayed_ref_updates = 0;
 533                         btrfs_run_delayed_refs(trans, root, cur);
 534                 } else {
 535                         break;
 536                 }
 537                 count++;
 538         }
 539
 540         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 541             should_end_transaction(trans, root)) {
 542                 trans->transaction->blocked = 1;
 543                 smp_wmb();
 544         }
 545
 546         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 547                 if (throttle) {
 548                         /*
 549                          * We may race with somebody else here so end up having
 550                          * to call end_transaction on ourselves again, so inc
 551                          * our use_count.
 552                          */
 553                         trans->use_count++;
 554                         return btrfs_commit_transaction(trans, root);
 555                 } else {
 556                         wake_up_process(info->transaction_kthread);
 557                 }
 558         }
 559
 560         WARN_ON(cur_trans != info->running_transaction);
 561         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 562         atomic_dec(&cur_trans->num_writers);
 563
 564         smp_mb();
 565         if (waitqueue_active(&cur_trans->writer_wait))
 566                 wake_up(&cur_trans->writer_wait);
 567         put_transaction(cur_trans);
 568
 569         if (current->journal_info == trans)
 570                 current->journal_info = NULL;
 571
 572         if (throttle)
 573                 btrfs_run_delayed_iputs(root);
 574
 575         if (trans->aborted ||
 576             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 577                 err = -EIO;
 578         }
 579         assert_qgroups_uptodate(trans);
 580
 581         memset(trans, 0, sizeof(*trans));
 582         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 583         return err;
 584 }
 585
 586 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 587                           struct btrfs_root *root)
 588 {
 589         int ret;
 590
 591         ret = __btrfs_end_transaction(trans, root, 0, 1);
 592         if (ret)
 593                 return ret;
 594         return 0;
 595 }
 596
 597 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 598                                    struct btrfs_root *root)
 599 {
 600         int ret;
 601
 602         ret = __btrfs_end_transaction(trans, root, 1, 1);
 603         if (ret)
 604                 return ret;
 605         return 0;
 606 }
 607
 608 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 609                                  struct btrfs_root *root)
 610 {
 611         int ret;
 612
 613         ret = __btrfs_end_transaction(trans, root, 0, 0);
 614         if (ret)
 615                 return ret;
 616         return 0;
 617 }
 618
 619 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 620                                 struct btrfs_root *root)
 621 {
 622         return __btrfs_end_transaction(trans, root, 1, 1);
 623 }
 624
 625 /*
 626  * when btree blocks are allocated, they have some corresponding bits set for
 627  * them in one of two extent_io trees.  This is used to make sure all of
 628  * those extents are sent to disk but does not wait on them
 629  */
 630 int btrfs_write_marked_extents(struct btrfs_root *root,
 631                                struct extent_io_tree *dirty_pages, int mark)
 632 {
 633         int err = 0;
 634         int werr = 0;
 635         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 636         u64 start = 0;
 637         u64 end;
 638
 639         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 640                                       mark)) {
 641                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 642                                    GFP_NOFS);
 643                 err = filemap_fdatawrite_range(mapping, start, end);
 644                 if (err)
 645                         werr = err;
 646                 cond_resched();
 647                 start = end + 1;
 648         }
 649         if (err)
 650                 werr = err;
 651         return werr;
 652 }
 653
 654 /*
 655  * when btree blocks are allocated, they have some corresponding bits set for
 656  * them in one of two extent_io trees.  This is used to make sure all of
 657  * those extents are on disk for transaction or log commit.  We wait
 658  * on all the pages and clear them from the dirty pages state tree
 659  */
 660 int btrfs_wait_marked_extents(struct btrfs_root *root,
 661                               struct extent_io_tree *dirty_pages, int mark)
 662 {
 663         int err = 0;
 664         int werr = 0;
 665         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 666         u64 start = 0;
 667         u64 end;
 668
 669         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 670                                       EXTENT_NEED_WAIT)) {
 671                 clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 672                 err = filemap_fdatawait_range(mapping, start, end);
 673                 if (err)
 674                         werr = err;
 675                 cond_resched();
 676                 start = end + 1;
 677         }
 678         if (err)
 679                 werr = err;
 680         return werr;
 681 }
 682
 683 /*
 684  * when btree blocks are allocated, they have some corresponding bits set for
 685  * them in one of two extent_io trees.  This is used to make sure all of
 686  * those extents are on disk for transaction or log commit
 687  */
 688 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 689                                 struct extent_io_tree *dirty_pages, int mark)
 690 {
 691         int ret;
 692         int ret2;
 693
 694         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 695         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 696
 697         if (ret)
 698                 return ret;
 699         if (ret2)
 700                 return ret2;
 701         return 0;
 702 }
 703
 704 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 705                                      struct btrfs_root *root)
 706 {
 707         if (!trans || !trans->transaction) {
 708                 struct inode *btree_inode;
 709                 btree_inode = root->fs_info->btree_inode;
 710                 return filemap_write_and_wait(btree_inode->i_mapping);
 711         }
 712         return btrfs_write_and_wait_marked_extents(root,
 713                                            &trans->transaction->dirty_pages,
 714                                            EXTENT_DIRTY);
 715 }
 716
 717 /*
 718  * this is used to update the root pointer in the tree of tree roots.
 719  *
 720  * But, in the case of the extent allocation tree, updating the root
 721  * pointer may allocate blocks which may change the root of the extent
 722  * allocation tree.
 723  *
 724  * So, this loops and repeats and makes sure the cowonly root didn't
 725  * change while the root pointer was being updated in the metadata.
 726  */
 727 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 728                                struct btrfs_root *root)
 729 {
 730         int ret;
 731         u64 old_root_bytenr;
 732         u64 old_root_used;
 733         struct btrfs_root *tree_root = root->fs_info->tree_root;
 734
 735         old_root_used = btrfs_root_used(&root->root_item);
 736         btrfs_write_dirty_block_groups(trans, root);
 737
 738         while (1) {
 739                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 740                 if (old_root_bytenr == root->node->start &&
 741                     old_root_used == btrfs_root_used(&root->root_item))
 742                         break;
 743
 744                 btrfs_set_root_node(&root->root_item, root->node);
 745                 ret = btrfs_update_root(trans, tree_root,
 746                                         &root->root_key,
 747                                         &root->root_item);
 748                 if (ret)
 749                         return ret;
 750
 751                 old_root_used = btrfs_root_used(&root->root_item);
 752                 ret = btrfs_write_dirty_block_groups(trans, root);
 753                 if (ret)
 754                         return ret;
 755         }
 756
 757         if (root != root->fs_info->extent_root)
 758                 switch_commit_root(root);
 759
 760         return 0;
 761 }
 762
 763 /*
 764  * update all the cowonly tree roots on disk
 765  *
 766  * The error handling in this function may not be obvious. Any of the
 767  * failures will cause the file system to go offline. We still need
 768  * to clean up the delayed refs.
 769  */
 770 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 771                                          struct btrfs_root *root)
 772 {
 773         struct btrfs_fs_info *fs_info = root->fs_info;
 774         struct list_head *next;
 775         struct extent_buffer *eb;
 776         int ret;
 777
 778         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 779         if (ret)
 780                 return ret;
 781
 782         eb = btrfs_lock_root_node(fs_info->tree_root);
 783         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 784                               0, &eb);
 785         btrfs_tree_unlock(eb);
 786         free_extent_buffer(eb);
 787
 788         if (ret)
 789                 return ret;
 790
 791         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 792         if (ret)
 793                 return ret;
 794
 795         ret = btrfs_run_dev_stats(trans, root->fs_info);
 796         BUG_ON(ret);
 797
 798         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 799                 next = fs_info->dirty_cowonly_roots.next;
 800                 list_del_init(next);
 801                 root = list_entry(next, struct btrfs_root, dirty_list);
 802
 803                 ret = update_cowonly_root(trans, root);
 804                 if (ret)
 805                         return ret;
 806         }
 807
 808         down_write(&fs_info->extent_commit_sem);
 809         switch_commit_root(fs_info->extent_root);
 810         up_write(&fs_info->extent_commit_sem);
 811
 812         return 0;
 813 }
 814
 815 /*
 816  * dead roots are old snapshots that need to be deleted.  This allocates
 817  * a dirty root struct and adds it into the list of dead roots that need to
 818  * be deleted
 819  */
 820 int btrfs_add_dead_root(struct btrfs_root *root)
 821 {
 822         spin_lock(&root->fs_info->trans_lock);
 823         list_add(&root->root_list, &root->fs_info->dead_roots);
 824         spin_unlock(&root->fs_info->trans_lock);
 825         return 0;
 826 }
 827
 828 /*
 829  * update all the cowonly tree roots on disk
 830  */
 831 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 832                                     struct btrfs_root *root)
 833 {
 834         struct btrfs_root *gang[8];
 835         struct btrfs_fs_info *fs_info = root->fs_info;
 836         int i;
 837         int ret;
 838         int err = 0;
 839
 840         spin_lock(&fs_info->fs_roots_radix_lock);
 841         while (1) {
 842                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 843                                                  (void **)gang, 0,
 844                                                  ARRAY_SIZE(gang),
 845                                                  BTRFS_ROOT_TRANS_TAG);
 846                 if (ret == 0)
 847                         break;
 848                 for (i = 0; i < ret; i++) {
 849                         root = gang[i];
 850                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 851                                         (unsigned long)root->root_key.objectid,
 852                                         BTRFS_ROOT_TRANS_TAG);
 853                         spin_unlock(&fs_info->fs_roots_radix_lock);
 854
 855                         btrfs_free_log(trans, root);
 856                         btrfs_update_reloc_root(trans, root);
 857                         btrfs_orphan_commit_root(trans, root);
 858
 859                         btrfs_save_ino_cache(root, trans);
 860
 861                         /* see comments in should_cow_block() */
 862                         root->force_cow = 0;
 863                         smp_wmb();
 864
 865                         if (root->commit_root != root->node) {
 866                                 mutex_lock(&root->fs_commit_mutex);
 867                                 switch_commit_root(root);
 868                                 btrfs_unpin_free_ino(root);
 869                                 mutex_unlock(&root->fs_commit_mutex);
 870
 871                                 btrfs_set_root_node(&root->root_item,
 872                                                     root->node);
 873                         }
 874
 875                         err = btrfs_update_root(trans, fs_info->tree_root,
 876                                                 &root->root_key,
 877                                                 &root->root_item);
 878                         spin_lock(&fs_info->fs_roots_radix_lock);
 879                         if (err)
 880                                 break;
 881                 }
 882         }
 883         spin_unlock(&fs_info->fs_roots_radix_lock);
 884         return err;
 885 }
 886
 887 /*
 888  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 889  * otherwise every leaf in the btree is read and defragged.
 890  */
 891 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 892 {
 893         struct btrfs_fs_info *info = root->fs_info;
 894         struct btrfs_trans_handle *trans;
 895         int ret;
 896         unsigned long nr;
 897
 898         if (xchg(&root->defrag_running, 1))
 899                 return 0;
 900
 901         while (1) {
 902                 trans = btrfs_start_transaction(root, 0);
 903                 if (IS_ERR(trans))
 904                         return PTR_ERR(trans);
 905
 906                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 907
 908                 nr = trans->blocks_used;
 909                 btrfs_end_transaction(trans, root);
 910                 btrfs_btree_balance_dirty(info->tree_root, nr);
 911                 cond_resched();
 912
 913                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 914                         break;
 915         }
 916         root->defrag_running = 0;
 917         return ret;
 918 }
 919
 920 /*
 921  * new snapshots need to be created at a very specific time in the
 922  * transaction commit.  This does the actual creation
 923  */
 924 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 925                                    struct btrfs_fs_info *fs_info,
 926                                    struct btrfs_pending_snapshot *pending)
 927 {
 928         struct btrfs_key key;
 929         struct btrfs_root_item *new_root_item;
 930         struct btrfs_root *tree_root = fs_info->tree_root;
 931         struct btrfs_root *root = pending->root;
 932         struct btrfs_root *parent_root;
 933         struct btrfs_block_rsv *rsv;
 934         struct inode *parent_inode;
 935         struct dentry *parent;
 936         struct dentry *dentry;
 937         struct extent_buffer *tmp;
 938         struct extent_buffer *old;
 939         int ret;
 940         u64 to_reserve = 0;
 941         u64 index = 0;
 942         u64 objectid;
 943         u64 root_flags;
 944
 945         rsv = trans->block_rsv;
 946
 947         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 948         if (!new_root_item) {
 949                 ret = pending->error = -ENOMEM;
 950                 goto fail;
 951         }
 952
 953         ret = btrfs_find_free_objectid(tree_root, &objectid);
 954         if (ret) {
 955                 pending->error = ret;
 956                 goto fail;
 957         }
 958
 959         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 960
 961         if (to_reserve > 0) {
 962                 ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 963                                                   to_reserve);
 964                 if (ret) {
 965                         pending->error = ret;
 966                         goto fail;
 967                 }
 968         }
 969
 970         key.objectid = objectid;
 971         key.offset = (u64)-1;
 972         key.type = BTRFS_ROOT_ITEM_KEY;
 973
 974         trans->block_rsv = &pending->block_rsv;
 975
 976         dentry = pending->dentry;
 977         parent = dget_parent(dentry);
 978         parent_inode = parent->d_inode;
 979         parent_root = BTRFS_I(parent_inode)->root;
 980         record_root_in_trans(trans, parent_root);
 981
 982         /*
 983          * insert the directory item
 984          */
 985         ret = btrfs_set_inode_index(parent_inode, &index);
 986         BUG_ON(ret); /* -ENOMEM */
 987         ret = btrfs_insert_dir_item(trans, parent_root,
 988                                 dentry->d_name.name, dentry->d_name.len,
 989                                 parent_inode, &key,
 990                                 BTRFS_FT_DIR, index);
 991         if (ret == -EEXIST) {
 992                 pending->error = -EEXIST;
 993                 dput(parent);
 994                 goto fail;
 995         } else if (ret) {
 996                 goto abort_trans_dput;
 997         }
 998
 999         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1000                                          dentry->d_name.len * 2);
1001         ret = btrfs_update_inode(trans, parent_root, parent_inode);
1002         if (ret)
1003                 goto abort_trans_dput;
1004
1005         /*
1006          * pull in the delayed directory update
1007          * and the delayed inode item
1008          * otherwise we corrupt the FS during
1009          * snapshot
1010          */
1011         ret = btrfs_run_delayed_items(trans, root);
1012         if (ret) { /* Transaction aborted */
1013                 dput(parent);
1014                 goto fail;
1015         }
1016
1017         record_root_in_trans(trans, root);
1018         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1019         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1020         btrfs_check_and_init_root_item(new_root_item);
1021
1022         root_flags = btrfs_root_flags(new_root_item);
1023         if (pending->readonly)
1024                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1025         else
1026                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1027         btrfs_set_root_flags(new_root_item, root_flags);
1028
1029         old = btrfs_lock_root_node(root);
1030         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1031         if (ret) {
1032                 btrfs_tree_unlock(old);
1033                 free_extent_buffer(old);
1034                 goto abort_trans_dput;
1035         }
1036
1037         btrfs_set_lock_blocking(old);
1038
1039         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1040         /* clean up in any case */
1041         btrfs_tree_unlock(old);
1042         free_extent_buffer(old);
1043         if (ret)
1044                 goto abort_trans_dput;
1045
1046         /* see comments in should_cow_block() */
1047         root->force_cow = 1;
1048         smp_wmb();
1049
1050         btrfs_set_root_node(new_root_item, tmp);
1051         /* record when the snapshot was created in key.offset */
1052         key.offset = trans->transid;
1053         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1054         btrfs_tree_unlock(tmp);
1055         free_extent_buffer(tmp);
1056         if (ret)
1057                 goto abort_trans_dput;
1058
1059         /*
1060          * insert root back/forward references
1061          */
1062         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1063                                  parent_root->root_key.objectid,
1064                                  btrfs_ino(parent_inode), index,
1065                                  dentry->d_name.name, dentry->d_name.len);
1066         dput(parent);
1067         if (ret)
1068                 goto fail;
1069
1070         key.offset = (u64)-1;
1071         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1072         if (IS_ERR(pending->snap)) {
1073                 ret = PTR_ERR(pending->snap);
1074                 goto abort_trans;
1075         }
1076
1077         ret = btrfs_reloc_post_snapshot(trans, pending);
1078         if (ret)
1079                 goto abort_trans;
1080         ret = 0;
1081 fail:
1082         kfree(new_root_item);
1083         trans->block_rsv = rsv;
1084         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1085         return ret;
1086
1087 abort_trans_dput:
1088         dput(parent);
1089 abort_trans:
1090         btrfs_abort_transaction(trans, root, ret);
1091         goto fail;
1092 }
1093
1094 /*
1095  * create all the snapshots we've scheduled for creation
1096  */
1097 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1098                                              struct btrfs_fs_info *fs_info)
1099 {
1100         struct btrfs_pending_snapshot *pending;
1101         struct list_head *head = &trans->transaction->pending_snapshots;
1102
1103         list_for_each_entry(pending, head, list)
1104                 create_pending_snapshot(trans, fs_info, pending);
1105         return 0;
1106 }
1107
1108 static void update_super_roots(struct btrfs_root *root)
1109 {
1110         struct btrfs_root_item *root_item;
1111         struct btrfs_super_block *super;
1112
1113         super = root->fs_info->super_copy;
1114
1115         root_item = &root->fs_info->chunk_root->root_item;
1116         super->chunk_root = root_item->bytenr;
1117         super->chunk_root_generation = root_item->generation;
1118         super->chunk_root_level = root_item->level;
1119
1120         root_item = &root->fs_info->tree_root->root_item;
1121         super->root = root_item->bytenr;
1122         super->generation = root_item->generation;
1123         super->root_level = root_item->level;
1124         if (btrfs_test_opt(root, SPACE_CACHE))
1125                 super->cache_generation = root_item->generation;
1126 }
1127
1128 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1129 {
1130         int ret = 0;
1131         spin_lock(&info->trans_lock);
1132         if (info->running_transaction)
1133                 ret = info->running_transaction->in_commit;
1134         spin_unlock(&info->trans_lock);
1135         return ret;
1136 }
1137
1138 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1139 {
1140         int ret = 0;
1141         spin_lock(&info->trans_lock);
1142         if (info->running_transaction)
1143                 ret = info->running_transaction->blocked;
1144         spin_unlock(&info->trans_lock);
1145         return ret;
1146 }
1147
1148 /*
1149  * wait for the current transaction commit to start and block subsequent
1150  * transaction joins
1151  */
1152 static void wait_current_trans_commit_start(struct btrfs_root *root,
1153                                             struct btrfs_transaction *trans)
1154 {
1155         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1156 }
1157
1158 /*
1159  * wait for the current transaction to start and then become unblocked.
1160  * caller holds ref.
1161  */
1162 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1163                                          struct btrfs_transaction *trans)
1164 {
1165         wait_event(root->fs_info->transaction_wait,
1166                    trans->commit_done || (trans->in_commit && !trans->blocked));
1167 }
1168
1169 /*
1170  * commit transactions asynchronously. once btrfs_commit_transaction_async
1171  * returns, any subsequent transaction will not be allowed to join.
1172  */
1173 struct btrfs_async_commit {
1174         struct btrfs_trans_handle *newtrans;
1175         struct btrfs_root *root;
1176         struct delayed_work work;
1177 };
1178
1179 static void do_async_commit(struct work_struct *work)
1180 {
1181         struct btrfs_async_commit *ac =
1182                 container_of(work, struct btrfs_async_commit, work.work);
1183
1184         btrfs_commit_transaction(ac->newtrans, ac->root);
1185         kfree(ac);
1186 }
1187
1188 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1189                                    struct btrfs_root *root,
1190                                    int wait_for_unblock)
1191 {
1192         struct btrfs_async_commit *ac;
1193         struct btrfs_transaction *cur_trans;
1194
1195         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1196         if (!ac)
1197                 return -ENOMEM;
1198
1199         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1200         ac->root = root;
1201         ac->newtrans = btrfs_join_transaction(root);
1202         if (IS_ERR(ac->newtrans)) {
1203                 int err = PTR_ERR(ac->newtrans);
1204                 kfree(ac);
1205                 return err;
1206         }
1207
1208         /* take transaction reference */
1209         cur_trans = trans->transaction;
1210         atomic_inc(&cur_trans->use_count);
1211
1212         btrfs_end_transaction(trans, root);
1213         schedule_delayed_work(&ac->work, 0);
1214
1215         /* wait for transaction to start and unblock */
1216         if (wait_for_unblock)
1217                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1218         else
1219                 wait_current_trans_commit_start(root, cur_trans);
1220
1221         if (current->journal_info == trans)
1222                 current->journal_info = NULL;
1223
1224         put_transaction(cur_trans);
1225         return 0;
1226 }
1227
1228
1229 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1230                                 struct btrfs_root *root, int err)
1231 {
1232         struct btrfs_transaction *cur_trans = trans->transaction;
1233
1234         WARN_ON(trans->use_count > 1);
1235
1236         btrfs_abort_transaction(trans, root, err);
1237
1238         spin_lock(&root->fs_info->trans_lock);
1239         list_del_init(&cur_trans->list);
1240         if (cur_trans == root->fs_info->running_transaction) {
1241                 root->fs_info->running_transaction = NULL;
1242                 root->fs_info->trans_no_join = 0;
1243         }
1244         spin_unlock(&root->fs_info->trans_lock);
1245
1246         btrfs_cleanup_one_transaction(trans->transaction, root);
1247
1248         put_transaction(cur_trans);
1249         put_transaction(cur_trans);
1250
1251         trace_btrfs_transaction_commit(root);
1252
1253         btrfs_scrub_continue(root);
1254
1255         if (current->journal_info == trans)
1256                 current->journal_info = NULL;
1257
1258         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1259 }
1260
1261 /*
1262  * btrfs_transaction state sequence:
1263  *    in_commit = 0, blocked = 0  (initial)
1264  *    in_commit = 1, blocked = 1
1265  *    blocked = 0
1266  *    commit_done = 1
1267  */
1268 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1269                              struct btrfs_root *root)
1270 {
1271         unsigned long joined = 0;
1272         struct btrfs_transaction *cur_trans = trans->transaction;
1273         struct btrfs_transaction *prev_trans = NULL;
1274         DEFINE_WAIT(wait);
1275         int ret = -EIO;
1276         int should_grow = 0;
1277         unsigned long now = get_seconds();
1278         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1279
1280         btrfs_run_ordered_operations(root, 0);
1281
1282         btrfs_trans_release_metadata(trans, root);
1283         trans->block_rsv = NULL;
1284
1285         if (cur_trans->aborted)
1286                 goto cleanup_transaction;
1287
1288         /* make a pass through all the delayed refs we have so far
1289          * any runnings procs may add more while we are here
1290          */
1291         ret = btrfs_run_delayed_refs(trans, root, 0);
1292         if (ret)
1293                 goto cleanup_transaction;
1294
1295         cur_trans = trans->transaction;
1296
1297         /*
1298          * set the flushing flag so procs in this transaction have to
1299          * start sending their work down.
1300          */
1301         cur_trans->delayed_refs.flushing = 1;
1302
1303         ret = btrfs_run_delayed_refs(trans, root, 0);
1304         if (ret)
1305                 goto cleanup_transaction;
1306
1307         spin_lock(&cur_trans->commit_lock);
1308         if (cur_trans->in_commit) {
1309                 spin_unlock(&cur_trans->commit_lock);
1310                 atomic_inc(&cur_trans->use_count);
1311                 ret = btrfs_end_transaction(trans, root);
1312
1313                 wait_for_commit(root, cur_trans);
1314
1315                 put_transaction(cur_trans);
1316
1317                 return ret;
1318         }
1319
1320         trans->transaction->in_commit = 1;
1321         trans->transaction->blocked = 1;
1322         spin_unlock(&cur_trans->commit_lock);
1323         wake_up(&root->fs_info->transaction_blocked_wait);
1324
1325         spin_lock(&root->fs_info->trans_lock);
1326         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1327                 prev_trans = list_entry(cur_trans->list.prev,
1328                                         struct btrfs_transaction, list);
1329                 if (!prev_trans->commit_done) {
1330                         atomic_inc(&prev_trans->use_count);
1331                         spin_unlock(&root->fs_info->trans_lock);
1332
1333                         wait_for_commit(root, prev_trans);
1334
1335                         put_transaction(prev_trans);
1336                 } else {
1337                         spin_unlock(&root->fs_info->trans_lock);
1338                 }
1339         } else {
1340                 spin_unlock(&root->fs_info->trans_lock);
1341         }
1342
1343         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1344                 should_grow = 1;
1345
1346         do {
1347                 int snap_pending = 0;
1348
1349                 joined = cur_trans->num_joined;
1350                 if (!list_empty(&trans->transaction->pending_snapshots))
1351                         snap_pending = 1;
1352
1353                 WARN_ON(cur_trans != trans->transaction);
1354
1355                 if (flush_on_commit || snap_pending) {
1356                         btrfs_start_delalloc_inodes(root, 1);
1357                         btrfs_wait_ordered_extents(root, 0, 1);
1358                 }
1359
1360                 ret = btrfs_run_delayed_items(trans, root);
1361                 if (ret)
1362                         goto cleanup_transaction;
1363
1364                 /*
1365                  * running the delayed items may have added new refs. account
1366                  * them now so that they hinder processing of more delayed refs
1367                  * as little as possible.
1368                  */
1369                 btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1370
1371                 /*
1372                  * rename don't use btrfs_join_transaction, so, once we
1373                  * set the transaction to blocked above, we aren't going
1374                  * to get any new ordered operations.  We can safely run
1375                  * it here and no for sure that nothing new will be added
1376                  * to the list
1377                  */
1378                 btrfs_run_ordered_operations(root, 1);
1379
1380                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1381                                 TASK_UNINTERRUPTIBLE);
1382
1383                 if (atomic_read(&cur_trans->num_writers) > 1)
1384                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1385                 else if (should_grow)
1386                         schedule_timeout(1);
1387
1388                 finish_wait(&cur_trans->writer_wait, &wait);
1389         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1390                  (should_grow && cur_trans->num_joined != joined));
1391
1392         /*
1393          * Ok now we need to make sure to block out any other joins while we
1394          * commit the transaction.  We could have started a join before setting
1395          * no_join so make sure to wait for num_writers to == 1 again.
1396          */
1397         spin_lock(&root->fs_info->trans_lock);
1398         root->fs_info->trans_no_join = 1;
1399         spin_unlock(&root->fs_info->trans_lock);
1400         wait_event(cur_trans->writer_wait,
1401                    atomic_read(&cur_trans->num_writers) == 1);
1402
1403         /*
1404          * the reloc mutex makes sure that we stop
1405          * the balancing code from coming in and moving
1406          * extents around in the middle of the commit
1407          */
1408         mutex_lock(&root->fs_info->reloc_mutex);
1409
1410         ret = btrfs_run_delayed_items(trans, root);
1411         if (ret) {
1412                 mutex_unlock(&root->fs_info->reloc_mutex);
1413                 goto cleanup_transaction;
1414         }
1415
1416         ret = create_pending_snapshots(trans, root->fs_info);
1417         if (ret) {
1418                 mutex_unlock(&root->fs_info->reloc_mutex);
1419                 goto cleanup_transaction;
1420         }
1421
1422         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1423         if (ret) {
1424                 mutex_unlock(&root->fs_info->reloc_mutex);
1425                 goto cleanup_transaction;
1426         }
1427
1428         /*
1429          * make sure none of the code above managed to slip in a
1430          * delayed item
1431          */
1432         btrfs_assert_delayed_root_empty(root);
1433
1434         WARN_ON(cur_trans != trans->transaction);
1435
1436         btrfs_scrub_pause(root);
1437         /* btrfs_commit_tree_roots is responsible for getting the
1438          * various roots consistent with each other.  Every pointer
1439          * in the tree of tree roots has to point to the most up to date
1440          * root for every subvolume and other tree.  So, we have to keep
1441          * the tree logging code from jumping in and changing any
1442          * of the trees.
1443          *
1444          * At this point in the commit, there can't be any tree-log
1445          * writers, but a little lower down we drop the trans mutex
1446          * and let new people in.  By holding the tree_log_mutex
1447          * from now until after the super is written, we avoid races
1448          * with the tree-log code.
1449          */
1450         mutex_lock(&root->fs_info->tree_log_mutex);
1451
1452         ret = commit_fs_roots(trans, root);
1453         if (ret) {
1454                 mutex_unlock(&root->fs_info->tree_log_mutex);
1455                 mutex_unlock(&root->fs_info->reloc_mutex);
1456                 goto cleanup_transaction;
1457         }
1458
1459         /* commit_fs_roots gets rid of all the tree log roots, it is now
1460          * safe to free the root of tree log roots
1461          */
1462         btrfs_free_log_root_tree(trans, root->fs_info);
1463
1464         ret = commit_cowonly_roots(trans, root);
1465         if (ret) {
1466                 mutex_unlock(&root->fs_info->tree_log_mutex);
1467                 mutex_unlock(&root->fs_info->reloc_mutex);
1468                 goto cleanup_transaction;
1469         }
1470
1471         btrfs_prepare_extent_commit(trans, root);
1472
1473         cur_trans = root->fs_info->running_transaction;
1474
1475         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1476                             root->fs_info->tree_root->node);
1477         switch_commit_root(root->fs_info->tree_root);
1478
1479         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1480                             root->fs_info->chunk_root->node);
1481         switch_commit_root(root->fs_info->chunk_root);
1482
1483         assert_qgroups_uptodate(trans);
1484         update_super_roots(root);
1485
1486         if (!root->fs_info->log_root_recovering) {
1487                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1488                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1489         }
1490
1491         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1492                sizeof(*root->fs_info->super_copy));
1493
1494         trans->transaction->blocked = 0;
1495         spin_lock(&root->fs_info->trans_lock);
1496         root->fs_info->running_transaction = NULL;
1497         root->fs_info->trans_no_join = 0;
1498         spin_unlock(&root->fs_info->trans_lock);
1499         mutex_unlock(&root->fs_info->reloc_mutex);
1500
1501         wake_up(&root->fs_info->transaction_wait);
1502
1503         ret = btrfs_write_and_wait_transaction(trans, root);
1504         if (ret) {
1505                 btrfs_error(root->fs_info, ret,
1506                             "Error while writing out transaction.");
1507                 mutex_unlock(&root->fs_info->tree_log_mutex);
1508                 goto cleanup_transaction;
1509         }
1510
1511         ret = write_ctree_super(trans, root, 0);
1512         if (ret) {
1513                 mutex_unlock(&root->fs_info->tree_log_mutex);
1514                 goto cleanup_transaction;
1515         }
1516
1517         /*
1518          * the super is written, we can safely allow the tree-loggers
1519          * to go about their business
1520          */
1521         mutex_unlock(&root->fs_info->tree_log_mutex);
1522
1523         btrfs_finish_extent_commit(trans, root);
1524
1525         cur_trans->commit_done = 1;
1526
1527         root->fs_info->last_trans_committed = cur_trans->transid;
1528
1529         wake_up(&cur_trans->commit_wait);
1530
1531         spin_lock(&root->fs_info->trans_lock);
1532         list_del_init(&cur_trans->list);
1533         spin_unlock(&root->fs_info->trans_lock);
1534
1535         put_transaction(cur_trans);
1536         put_transaction(cur_trans);
1537
1538         trace_btrfs_transaction_commit(root);
1539
1540         btrfs_scrub_continue(root);
1541
1542         if (current->journal_info == trans)
1543                 current->journal_info = NULL;
1544
1545         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1546
1547         if (current != root->fs_info->transaction_kthread)
1548                 btrfs_run_delayed_iputs(root);
1549
1550         return ret;
1551
1552 cleanup_transaction:
1553         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1554 //      WARN_ON(1);
1555         if (current->journal_info == trans)
1556                 current->journal_info = NULL;
1557         cleanup_transaction(trans, root, ret);
1558
1559         return ret;
1560 }
1561
1562 /*
1563  * interface function to delete all the snapshots we have scheduled for deletion
1564  */
1565 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1566 {
1567         LIST_HEAD(list);
1568         struct btrfs_fs_info *fs_info = root->fs_info;
1569
1570         spin_lock(&fs_info->trans_lock);
1571         list_splice_init(&fs_info->dead_roots, &list);
1572         spin_unlock(&fs_info->trans_lock);
1573
1574         while (!list_empty(&list)) {
1575                 int ret;
1576
1577                 root = list_entry(list.next, struct btrfs_root, root_list);
1578                 list_del(&root->root_list);
1579
1580                 btrfs_kill_all_delayed_nodes(root);
1581
1582                 if (btrfs_header_backref_rev(root->node) <
1583                     BTRFS_MIXED_BACKREF_REV)
1584                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1585                 else
1586                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1587                 BUG_ON(ret < 0);
1588         }
1589         return 0;
1590 }