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 <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33
  34 #define BTRFS_ROOT_TRANS_TAG 0
  35
  36 void put_transaction(struct btrfs_transaction *transaction)
  37 {
  38         WARN_ON(atomic_read(&transaction->use_count) == 0);
  39         if (atomic_dec_and_test(&transaction->use_count)) {
  40                 BUG_ON(!list_empty(&transaction->list));
  41                 WARN_ON(transaction->delayed_refs.root.rb_node);
  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 type)
  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                 /*
  71                  * If we are JOIN_NOLOCK we're already committing a current
  72                  * transaction, we just need a handle to deal with something
  73                  * when committing the transaction, such as inode cache and
  74                  * space cache. It is a special case.
  75                  */
  76                 if (type != TRANS_JOIN_NOLOCK) {
  77                         spin_unlock(&fs_info->trans_lock);
  78                         return -EBUSY;
  79                 }
  80         }
  81
  82         cur_trans = fs_info->running_transaction;
  83         if (cur_trans) {
  84                 if (cur_trans->aborted) {
  85                         spin_unlock(&fs_info->trans_lock);
  86                         return cur_trans->aborted;
  87                 }
  88                 atomic_inc(&cur_trans->use_count);
  89                 atomic_inc(&cur_trans->num_writers);
  90                 cur_trans->num_joined++;
  91                 spin_unlock(&fs_info->trans_lock);
  92                 return 0;
  93         }
  94         spin_unlock(&fs_info->trans_lock);
  95
  96         /*
  97          * If we are ATTACH, we just want to catch the current transaction,
  98          * and commit it. If there is no transaction, just return ENOENT.
  99          */
 100         if (type == TRANS_ATTACH)
 101                 return -ENOENT;
 102
 103         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 104         if (!cur_trans)
 105                 return -ENOMEM;
 106
 107         spin_lock(&fs_info->trans_lock);
 108         if (fs_info->running_transaction) {
 109                 /*
 110                  * someone started a transaction after we unlocked.  Make sure
 111                  * to redo the trans_no_join checks above
 112                  */
 113                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 114                 cur_trans = fs_info->running_transaction;
 115                 goto loop;
 116         } else if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 117                 spin_unlock(&fs_info->trans_lock);
 118                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 119                 return -EROFS;
 120         }
 121
 122         atomic_set(&cur_trans->num_writers, 1);
 123         cur_trans->num_joined = 0;
 124         init_waitqueue_head(&cur_trans->writer_wait);
 125         init_waitqueue_head(&cur_trans->commit_wait);
 126         cur_trans->in_commit = 0;
 127         cur_trans->blocked = 0;
 128         /*
 129          * One for this trans handle, one so it will live on until we
 130          * commit the transaction.
 131          */
 132         atomic_set(&cur_trans->use_count, 2);
 133         cur_trans->commit_done = 0;
 134         cur_trans->start_time = get_seconds();
 135
 136         cur_trans->delayed_refs.root = RB_ROOT;
 137         cur_trans->delayed_refs.num_entries = 0;
 138         cur_trans->delayed_refs.num_heads_ready = 0;
 139         cur_trans->delayed_refs.num_heads = 0;
 140         cur_trans->delayed_refs.flushing = 0;
 141         cur_trans->delayed_refs.run_delayed_start = 0;
 142
 143         /*
 144          * although the tree mod log is per file system and not per transaction,
 145          * the log must never go across transaction boundaries.
 146          */
 147         smp_mb();
 148         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 149                 printk(KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 150                         "creating a fresh transaction\n");
 151                 WARN_ON(1);
 152         }
 153         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) {
 154                 printk(KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 155                         "creating a fresh transaction\n");
 156                 WARN_ON(1);
 157         }
 158         atomic_set(&fs_info->tree_mod_seq, 0);
 159
 160         spin_lock_init(&cur_trans->commit_lock);
 161         spin_lock_init(&cur_trans->delayed_refs.lock);
 162
 163         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 164         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 165         extent_io_tree_init(&cur_trans->dirty_pages,
 166                              fs_info->btree_inode->i_mapping);
 167         fs_info->generation++;
 168         cur_trans->transid = fs_info->generation;
 169         fs_info->running_transaction = cur_trans;
 170         cur_trans->aborted = 0;
 171         spin_unlock(&fs_info->trans_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * this does all the record keeping required to make sure that a reference
 178  * counted root is properly recorded in a given transaction.  This is required
 179  * to make sure the old root from before we joined the transaction is deleted
 180  * when the transaction commits
 181  */
 182 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 183                                struct btrfs_root *root)
 184 {
 185         if (root->ref_cows && root->last_trans < trans->transid) {
 186                 WARN_ON(root == root->fs_info->extent_root);
 187                 WARN_ON(root->commit_root != root->node);
 188
 189                 /*
 190                  * see below for in_trans_setup usage rules
 191                  * we have the reloc mutex held now, so there
 192                  * is only one writer in this function
 193                  */
 194                 root->in_trans_setup = 1;
 195
 196                 /* make sure readers find in_trans_setup before
 197                  * they find our root->last_trans update
 198                  */
 199                 smp_wmb();
 200
 201                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 202                 if (root->last_trans == trans->transid) {
 203                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 204                         return 0;
 205                 }
 206                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 207                            (unsigned long)root->root_key.objectid,
 208                            BTRFS_ROOT_TRANS_TAG);
 209                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 210                 root->last_trans = trans->transid;
 211
 212                 /* this is pretty tricky.  We don't want to
 213                  * take the relocation lock in btrfs_record_root_in_trans
 214                  * unless we're really doing the first setup for this root in
 215                  * this transaction.
 216                  *
 217                  * Normally we'd use root->last_trans as a flag to decide
 218                  * if we want to take the expensive mutex.
 219                  *
 220                  * But, we have to set root->last_trans before we
 221                  * init the relocation root, otherwise, we trip over warnings
 222                  * in ctree.c.  The solution used here is to flag ourselves
 223                  * with root->in_trans_setup.  When this is 1, we're still
 224                  * fixing up the reloc trees and everyone must wait.
 225                  *
 226                  * When this is zero, they can trust root->last_trans and fly
 227                  * through btrfs_record_root_in_trans without having to take the
 228                  * lock.  smp_wmb() makes sure that all the writes above are
 229                  * done before we pop in the zero below
 230                  */
 231                 btrfs_init_reloc_root(trans, root);
 232                 smp_wmb();
 233                 root->in_trans_setup = 0;
 234         }
 235         return 0;
 236 }
 237
 238
 239 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 240                                struct btrfs_root *root)
 241 {
 242         if (!root->ref_cows)
 243                 return 0;
 244
 245         /*
 246          * see record_root_in_trans for comments about in_trans_setup usage
 247          * and barriers
 248          */
 249         smp_rmb();
 250         if (root->last_trans == trans->transid &&
 251             !root->in_trans_setup)
 252                 return 0;
 253
 254         mutex_lock(&root->fs_info->reloc_mutex);
 255         record_root_in_trans(trans, root);
 256         mutex_unlock(&root->fs_info->reloc_mutex);
 257
 258         return 0;
 259 }
 260
 261 /* wait for commit against the current transaction to become unblocked
 262  * when this is done, it is safe to start a new transaction, but the current
 263  * transaction might not be fully on disk.
 264  */
 265 static void wait_current_trans(struct btrfs_root *root)
 266 {
 267         struct btrfs_transaction *cur_trans;
 268
 269         spin_lock(&root->fs_info->trans_lock);
 270         cur_trans = root->fs_info->running_transaction;
 271         if (cur_trans && cur_trans->blocked) {
 272                 atomic_inc(&cur_trans->use_count);
 273                 spin_unlock(&root->fs_info->trans_lock);
 274
 275                 wait_event(root->fs_info->transaction_wait,
 276                            !cur_trans->blocked);
 277                 put_transaction(cur_trans);
 278         } else {
 279                 spin_unlock(&root->fs_info->trans_lock);
 280         }
 281 }
 282
 283 static int may_wait_transaction(struct btrfs_root *root, int type)
 284 {
 285         if (root->fs_info->log_root_recovering)
 286                 return 0;
 287
 288         if (type == TRANS_USERSPACE)
 289                 return 1;
 290
 291         if (type == TRANS_START &&
 292             !atomic_read(&root->fs_info->open_ioctl_trans))
 293                 return 1;
 294
 295         return 0;
 296 }
 297
 298 static struct btrfs_trans_handle *
 299 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 300                   enum btrfs_reserve_flush_enum flush)
 301 {
 302         struct btrfs_trans_handle *h;
 303         struct btrfs_transaction *cur_trans;
 304         u64 num_bytes = 0;
 305         int ret;
 306         u64 qgroup_reserved = 0;
 307
 308         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 309                 return ERR_PTR(-EROFS);
 310
 311         if (current->journal_info) {
 312                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 313                 h = current->journal_info;
 314                 h->use_count++;
 315                 h->orig_rsv = h->block_rsv;
 316                 h->block_rsv = NULL;
 317                 goto got_it;
 318         }
 319
 320         /*
 321          * Do the reservation before we join the transaction so we can do all
 322          * the appropriate flushing if need be.
 323          */
 324         if (num_items > 0 && root != root->fs_info->chunk_root) {
 325                 if (root->fs_info->quota_enabled &&
 326                     is_fstree(root->root_key.objectid)) {
 327                         qgroup_reserved = num_items * root->leafsize;
 328                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 329                         if (ret)
 330                                 return ERR_PTR(ret);
 331                 }
 332
 333                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 334                 ret = btrfs_block_rsv_add(root,
 335                                           &root->fs_info->trans_block_rsv,
 336                                           num_bytes, flush);
 337                 if (ret)
 338                         return ERR_PTR(ret);
 339         }
 340 again:
 341         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 342         if (!h)
 343                 return ERR_PTR(-ENOMEM);
 344
 345         /*
 346          * If we are JOIN_NOLOCK we're already committing a transaction and
 347          * waiting on this guy, so we don't need to do the sb_start_intwrite
 348          * because we're already holding a ref.  We need this because we could
 349          * have raced in and did an fsync() on a file which can kick a commit
 350          * and then we deadlock with somebody doing a freeze.
 351          *
 352          * If we are ATTACH, it means we just want to catch the current
 353          * transaction and commit it, so we needn't do sb_start_intwrite().
 354          */
 355         if (type < TRANS_JOIN_NOLOCK)
 356                 sb_start_intwrite(root->fs_info->sb);
 357
 358         if (may_wait_transaction(root, type))
 359                 wait_current_trans(root);
 360
 361         do {
 362                 ret = join_transaction(root, type);
 363                 if (ret == -EBUSY)
 364                         wait_current_trans(root);
 365         } while (ret == -EBUSY);
 366
 367         if (ret < 0) {
 368                 /* We must get the transaction if we are JOIN_NOLOCK. */
 369                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 370
 371                 if (type < TRANS_JOIN_NOLOCK)
 372                         sb_end_intwrite(root->fs_info->sb);
 373                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 374                 return ERR_PTR(ret);
 375         }
 376
 377         cur_trans = root->fs_info->running_transaction;
 378
 379         h->transid = cur_trans->transid;
 380         h->transaction = cur_trans;
 381         h->blocks_used = 0;
 382         h->bytes_reserved = 0;
 383         h->root = root;
 384         h->delayed_ref_updates = 0;
 385         h->use_count = 1;
 386         h->adding_csums = 0;
 387         h->block_rsv = NULL;
 388         h->orig_rsv = NULL;
 389         h->aborted = 0;
 390         h->qgroup_reserved = qgroup_reserved;
 391         h->delayed_ref_elem.seq = 0;
 392         h->type = type;
 393         INIT_LIST_HEAD(&h->qgroup_ref_list);
 394         INIT_LIST_HEAD(&h->new_bgs);
 395
 396         smp_mb();
 397         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 398                 btrfs_commit_transaction(h, root);
 399                 goto again;
 400         }
 401
 402         if (num_bytes) {
 403                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 404                                               h->transid, num_bytes, 1);
 405                 h->block_rsv = &root->fs_info->trans_block_rsv;
 406                 h->bytes_reserved = num_bytes;
 407         }
 408
 409 got_it:
 410         btrfs_record_root_in_trans(h, root);
 411
 412         if (!current->journal_info && type != TRANS_USERSPACE)
 413                 current->journal_info = h;
 414         return h;
 415 }
 416
 417 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 418                                                    int num_items)
 419 {
 420         return start_transaction(root, num_items, TRANS_START,
 421                                  BTRFS_RESERVE_FLUSH_ALL);
 422 }
 423
 424 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 425                                         struct btrfs_root *root, int num_items)
 426 {
 427         return start_transaction(root, num_items, TRANS_START,
 428                                  BTRFS_RESERVE_FLUSH_LIMIT);
 429 }
 430
 431 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 432 {
 433         return start_transaction(root, 0, TRANS_JOIN, 0);
 434 }
 435
 436 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 437 {
 438         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 439 }
 440
 441 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 442 {
 443         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 444 }
 445
 446 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 447 {
 448         return start_transaction(root, 0, TRANS_ATTACH, 0);
 449 }
 450
 451 /* wait for a transaction commit to be fully complete */
 452 static noinline void wait_for_commit(struct btrfs_root *root,
 453                                     struct btrfs_transaction *commit)
 454 {
 455         wait_event(commit->commit_wait, commit->commit_done);
 456 }
 457
 458 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 459 {
 460         struct btrfs_transaction *cur_trans = NULL, *t;
 461         int ret;
 462
 463         ret = 0;
 464         if (transid) {
 465                 if (transid <= root->fs_info->last_trans_committed)
 466                         goto out;
 467
 468                 /* find specified transaction */
 469                 spin_lock(&root->fs_info->trans_lock);
 470                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 471                         if (t->transid == transid) {
 472                                 cur_trans = t;
 473                                 atomic_inc(&cur_trans->use_count);
 474                                 break;
 475                         }
 476                         if (t->transid > transid)
 477                                 break;
 478                 }
 479                 spin_unlock(&root->fs_info->trans_lock);
 480                 ret = -EINVAL;
 481                 if (!cur_trans)
 482                         goto out;  /* bad transid */
 483         } else {
 484                 /* find newest transaction that is committing | committed */
 485                 spin_lock(&root->fs_info->trans_lock);
 486                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 487                                             list) {
 488                         if (t->in_commit) {
 489                                 if (t->commit_done)
 490                                         break;
 491                                 cur_trans = t;
 492                                 atomic_inc(&cur_trans->use_count);
 493                                 break;
 494                         }
 495                 }
 496                 spin_unlock(&root->fs_info->trans_lock);
 497                 if (!cur_trans)
 498                         goto out;  /* nothing committing|committed */
 499         }
 500
 501         wait_for_commit(root, cur_trans);
 502
 503         put_transaction(cur_trans);
 504         ret = 0;
 505 out:
 506         return ret;
 507 }
 508
 509 void btrfs_throttle(struct btrfs_root *root)
 510 {
 511         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 512                 wait_current_trans(root);
 513 }
 514
 515 static int should_end_transaction(struct btrfs_trans_handle *trans,
 516                                   struct btrfs_root *root)
 517 {
 518         int ret;
 519
 520         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 521         return ret ? 1 : 0;
 522 }
 523
 524 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 525                                  struct btrfs_root *root)
 526 {
 527         struct btrfs_transaction *cur_trans = trans->transaction;
 528         int updates;
 529         int err;
 530
 531         smp_mb();
 532         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 533                 return 1;
 534
 535         updates = trans->delayed_ref_updates;
 536         trans->delayed_ref_updates = 0;
 537         if (updates) {
 538                 err = btrfs_run_delayed_refs(trans, root, updates);
 539                 if (err) /* Error code will also eval true */
 540                         return err;
 541         }
 542
 543         return should_end_transaction(trans, root);
 544 }
 545
 546 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 547                           struct btrfs_root *root, int throttle)
 548 {
 549         struct btrfs_transaction *cur_trans = trans->transaction;
 550         struct btrfs_fs_info *info = root->fs_info;
 551         int count = 0;
 552         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 553         int err = 0;
 554
 555         if (--trans->use_count) {
 556                 trans->block_rsv = trans->orig_rsv;
 557                 return 0;
 558         }
 559
 560         /*
 561          * do the qgroup accounting as early as possible
 562          */
 563         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 564
 565         btrfs_trans_release_metadata(trans, root);
 566         trans->block_rsv = NULL;
 567         /*
 568          * the same root has to be passed to start_transaction and
 569          * end_transaction. Subvolume quota depends on this.
 570          */
 571         WARN_ON(trans->root != root);
 572
 573         if (trans->qgroup_reserved) {
 574                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 575                 trans->qgroup_reserved = 0;
 576         }
 577
 578         if (!list_empty(&trans->new_bgs))
 579                 btrfs_create_pending_block_groups(trans, root);
 580
 581         while (count < 2) {
 582                 unsigned long cur = trans->delayed_ref_updates;
 583                 trans->delayed_ref_updates = 0;
 584                 if (cur &&
 585                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 586                         trans->delayed_ref_updates = 0;
 587                         btrfs_run_delayed_refs(trans, root, cur);
 588                 } else {
 589                         break;
 590                 }
 591                 count++;
 592         }
 593         btrfs_trans_release_metadata(trans, root);
 594         trans->block_rsv = NULL;
 595
 596         if (!list_empty(&trans->new_bgs))
 597                 btrfs_create_pending_block_groups(trans, root);
 598
 599         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 600             should_end_transaction(trans, root)) {
 601                 trans->transaction->blocked = 1;
 602                 smp_wmb();
 603         }
 604
 605         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 606                 if (throttle) {
 607                         /*
 608                          * We may race with somebody else here so end up having
 609                          * to call end_transaction on ourselves again, so inc
 610                          * our use_count.
 611                          */
 612                         trans->use_count++;
 613                         return btrfs_commit_transaction(trans, root);
 614                 } else {
 615                         wake_up_process(info->transaction_kthread);
 616                 }
 617         }
 618
 619         if (trans->type < TRANS_JOIN_NOLOCK)
 620                 sb_end_intwrite(root->fs_info->sb);
 621
 622         WARN_ON(cur_trans != info->running_transaction);
 623         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 624         atomic_dec(&cur_trans->num_writers);
 625
 626         smp_mb();
 627         if (waitqueue_active(&cur_trans->writer_wait))
 628                 wake_up(&cur_trans->writer_wait);
 629         put_transaction(cur_trans);
 630
 631         if (current->journal_info == trans)
 632                 current->journal_info = NULL;
 633
 634         if (throttle)
 635                 btrfs_run_delayed_iputs(root);
 636
 637         if (trans->aborted ||
 638             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 639                 err = -EIO;
 640         }
 641         assert_qgroups_uptodate(trans);
 642
 643         memset(trans, 0, sizeof(*trans));
 644         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 645         return err;
 646 }
 647
 648 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 649                           struct btrfs_root *root)
 650 {
 651         int ret;
 652
 653         ret = __btrfs_end_transaction(trans, root, 0);
 654         if (ret)
 655                 return ret;
 656         return 0;
 657 }
 658
 659 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 660                                    struct btrfs_root *root)
 661 {
 662         int ret;
 663
 664         ret = __btrfs_end_transaction(trans, root, 1);
 665         if (ret)
 666                 return ret;
 667         return 0;
 668 }
 669
 670 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 671                                 struct btrfs_root *root)
 672 {
 673         return __btrfs_end_transaction(trans, root, 1);
 674 }
 675
 676 /*
 677  * when btree blocks are allocated, they have some corresponding bits set for
 678  * them in one of two extent_io trees.  This is used to make sure all of
 679  * those extents are sent to disk but does not wait on them
 680  */
 681 int btrfs_write_marked_extents(struct btrfs_root *root,
 682                                struct extent_io_tree *dirty_pages, int mark)
 683 {
 684         int err = 0;
 685         int werr = 0;
 686         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 687         struct extent_state *cached_state = NULL;
 688         u64 start = 0;
 689         u64 end;
 690
 691         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 692                                       mark, &cached_state)) {
 693                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 694                                    mark, &cached_state, GFP_NOFS);
 695                 cached_state = NULL;
 696                 err = filemap_fdatawrite_range(mapping, start, end);
 697                 if (err)
 698                         werr = err;
 699                 cond_resched();
 700                 start = end + 1;
 701         }
 702         if (err)
 703                 werr = err;
 704         return werr;
 705 }
 706
 707 /*
 708  * when btree blocks are allocated, they have some corresponding bits set for
 709  * them in one of two extent_io trees.  This is used to make sure all of
 710  * those extents are on disk for transaction or log commit.  We wait
 711  * on all the pages and clear them from the dirty pages state tree
 712  */
 713 int btrfs_wait_marked_extents(struct btrfs_root *root,
 714                               struct extent_io_tree *dirty_pages, int mark)
 715 {
 716         int err = 0;
 717         int werr = 0;
 718         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 719         struct extent_state *cached_state = NULL;
 720         u64 start = 0;
 721         u64 end;
 722
 723         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 724                                       EXTENT_NEED_WAIT, &cached_state)) {
 725                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 726                                  0, 0, &cached_state, GFP_NOFS);
 727                 err = filemap_fdatawait_range(mapping, start, end);
 728                 if (err)
 729                         werr = err;
 730                 cond_resched();
 731                 start = end + 1;
 732         }
 733         if (err)
 734                 werr = err;
 735         return werr;
 736 }
 737
 738 /*
 739  * when btree blocks are allocated, they have some corresponding bits set for
 740  * them in one of two extent_io trees.  This is used to make sure all of
 741  * those extents are on disk for transaction or log commit
 742  */
 743 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 744                                 struct extent_io_tree *dirty_pages, int mark)
 745 {
 746         int ret;
 747         int ret2;
 748
 749         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 750         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 751
 752         if (ret)
 753                 return ret;
 754         if (ret2)
 755                 return ret2;
 756         return 0;
 757 }
 758
 759 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 760                                      struct btrfs_root *root)
 761 {
 762         if (!trans || !trans->transaction) {
 763                 struct inode *btree_inode;
 764                 btree_inode = root->fs_info->btree_inode;
 765                 return filemap_write_and_wait(btree_inode->i_mapping);
 766         }
 767         return btrfs_write_and_wait_marked_extents(root,
 768                                            &trans->transaction->dirty_pages,
 769                                            EXTENT_DIRTY);
 770 }
 771
 772 /*
 773  * this is used to update the root pointer in the tree of tree roots.
 774  *
 775  * But, in the case of the extent allocation tree, updating the root
 776  * pointer may allocate blocks which may change the root of the extent
 777  * allocation tree.
 778  *
 779  * So, this loops and repeats and makes sure the cowonly root didn't
 780  * change while the root pointer was being updated in the metadata.
 781  */
 782 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 783                                struct btrfs_root *root)
 784 {
 785         int ret;
 786         u64 old_root_bytenr;
 787         u64 old_root_used;
 788         struct btrfs_root *tree_root = root->fs_info->tree_root;
 789
 790         old_root_used = btrfs_root_used(&root->root_item);
 791         btrfs_write_dirty_block_groups(trans, root);
 792
 793         while (1) {
 794                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 795                 if (old_root_bytenr == root->node->start &&
 796                     old_root_used == btrfs_root_used(&root->root_item))
 797                         break;
 798
 799                 btrfs_set_root_node(&root->root_item, root->node);
 800                 ret = btrfs_update_root(trans, tree_root,
 801                                         &root->root_key,
 802                                         &root->root_item);
 803                 if (ret)
 804                         return ret;
 805
 806                 old_root_used = btrfs_root_used(&root->root_item);
 807                 ret = btrfs_write_dirty_block_groups(trans, root);
 808                 if (ret)
 809                         return ret;
 810         }
 811
 812         if (root != root->fs_info->extent_root)
 813                 switch_commit_root(root);
 814
 815         return 0;
 816 }
 817
 818 /*
 819  * update all the cowonly tree roots on disk
 820  *
 821  * The error handling in this function may not be obvious. Any of the
 822  * failures will cause the file system to go offline. We still need
 823  * to clean up the delayed refs.
 824  */
 825 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 826                                          struct btrfs_root *root)
 827 {
 828         struct btrfs_fs_info *fs_info = root->fs_info;
 829         struct list_head *next;
 830         struct extent_buffer *eb;
 831         int ret;
 832
 833         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 834         if (ret)
 835                 return ret;
 836
 837         eb = btrfs_lock_root_node(fs_info->tree_root);
 838         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 839                               0, &eb);
 840         btrfs_tree_unlock(eb);
 841         free_extent_buffer(eb);
 842
 843         if (ret)
 844                 return ret;
 845
 846         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 847         if (ret)
 848                 return ret;
 849
 850         ret = btrfs_run_dev_stats(trans, root->fs_info);
 851         BUG_ON(ret);
 852
 853         ret = btrfs_run_qgroups(trans, root->fs_info);
 854         BUG_ON(ret);
 855
 856         /* run_qgroups might have added some more refs */
 857         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 858         BUG_ON(ret);
 859
 860         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 861                 next = fs_info->dirty_cowonly_roots.next;
 862                 list_del_init(next);
 863                 root = list_entry(next, struct btrfs_root, dirty_list);
 864
 865                 ret = update_cowonly_root(trans, root);
 866                 if (ret)
 867                         return ret;
 868         }
 869
 870         down_write(&fs_info->extent_commit_sem);
 871         switch_commit_root(fs_info->extent_root);
 872         up_write(&fs_info->extent_commit_sem);
 873
 874         return 0;
 875 }
 876
 877 /*
 878  * dead roots are old snapshots that need to be deleted.  This allocates
 879  * a dirty root struct and adds it into the list of dead roots that need to
 880  * be deleted
 881  */
 882 int btrfs_add_dead_root(struct btrfs_root *root)
 883 {
 884         spin_lock(&root->fs_info->trans_lock);
 885         list_add(&root->root_list, &root->fs_info->dead_roots);
 886         spin_unlock(&root->fs_info->trans_lock);
 887         return 0;
 888 }
 889
 890 /*
 891  * update all the cowonly tree roots on disk
 892  */
 893 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 894                                     struct btrfs_root *root)
 895 {
 896         struct btrfs_root *gang[8];
 897         struct btrfs_fs_info *fs_info = root->fs_info;
 898         int i;
 899         int ret;
 900         int err = 0;
 901
 902         spin_lock(&fs_info->fs_roots_radix_lock);
 903         while (1) {
 904                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 905                                                  (void **)gang, 0,
 906                                                  ARRAY_SIZE(gang),
 907                                                  BTRFS_ROOT_TRANS_TAG);
 908                 if (ret == 0)
 909                         break;
 910                 for (i = 0; i < ret; i++) {
 911                         root = gang[i];
 912                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 913                                         (unsigned long)root->root_key.objectid,
 914                                         BTRFS_ROOT_TRANS_TAG);
 915                         spin_unlock(&fs_info->fs_roots_radix_lock);
 916
 917                         btrfs_free_log(trans, root);
 918                         btrfs_update_reloc_root(trans, root);
 919                         btrfs_orphan_commit_root(trans, root);
 920
 921                         btrfs_save_ino_cache(root, trans);
 922
 923                         /* see comments in should_cow_block() */
 924                         root->force_cow = 0;
 925                         smp_wmb();
 926
 927                         if (root->commit_root != root->node) {
 928                                 mutex_lock(&root->fs_commit_mutex);
 929                                 switch_commit_root(root);
 930                                 btrfs_unpin_free_ino(root);
 931                                 mutex_unlock(&root->fs_commit_mutex);
 932
 933                                 btrfs_set_root_node(&root->root_item,
 934                                                     root->node);
 935                         }
 936
 937                         err = btrfs_update_root(trans, fs_info->tree_root,
 938                                                 &root->root_key,
 939                                                 &root->root_item);
 940                         spin_lock(&fs_info->fs_roots_radix_lock);
 941                         if (err)
 942                                 break;
 943                 }
 944         }
 945         spin_unlock(&fs_info->fs_roots_radix_lock);
 946         return err;
 947 }
 948
 949 /*
 950  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 951  * otherwise every leaf in the btree is read and defragged.
 952  */
 953 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 954 {
 955         struct btrfs_fs_info *info = root->fs_info;
 956         struct btrfs_trans_handle *trans;
 957         int ret;
 958         unsigned long nr;
 959
 960         if (xchg(&root->defrag_running, 1))
 961                 return 0;
 962
 963         while (1) {
 964                 trans = btrfs_start_transaction(root, 0);
 965                 if (IS_ERR(trans))
 966                         return PTR_ERR(trans);
 967
 968                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 969
 970                 nr = trans->blocks_used;
 971                 btrfs_end_transaction(trans, root);
 972                 btrfs_btree_balance_dirty(info->tree_root, nr);
 973                 cond_resched();
 974
 975                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 976                         break;
 977         }
 978         root->defrag_running = 0;
 979         return ret;
 980 }
 981
 982 /*
 983  * new snapshots need to be created at a very specific time in the
 984  * transaction commit.  This does the actual creation
 985  */
 986 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 987                                    struct btrfs_fs_info *fs_info,
 988                                    struct btrfs_pending_snapshot *pending)
 989 {
 990         struct btrfs_key key;
 991         struct btrfs_root_item *new_root_item;
 992         struct btrfs_root *tree_root = fs_info->tree_root;
 993         struct btrfs_root *root = pending->root;
 994         struct btrfs_root *parent_root;
 995         struct btrfs_block_rsv *rsv;
 996         struct inode *parent_inode;
 997         struct btrfs_path *path;
 998         struct btrfs_dir_item *dir_item;
 999         struct dentry *parent;
1000         struct dentry *dentry;
1001         struct extent_buffer *tmp;
1002         struct extent_buffer *old;
1003         struct timespec cur_time = CURRENT_TIME;
1004         int ret;
1005         u64 to_reserve = 0;
1006         u64 index = 0;
1007         u64 objectid;
1008         u64 root_flags;
1009         uuid_le new_uuid;
1010
1011         path = btrfs_alloc_path();
1012         if (!path) {
1013                 ret = pending->error = -ENOMEM;
1014                 goto path_alloc_fail;
1015         }
1016
1017         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1018         if (!new_root_item) {
1019                 ret = pending->error = -ENOMEM;
1020                 goto root_item_alloc_fail;
1021         }
1022
1023         ret = btrfs_find_free_objectid(tree_root, &objectid);
1024         if (ret) {
1025                 pending->error = ret;
1026                 goto no_free_objectid;
1027         }
1028
1029         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1030
1031         if (to_reserve > 0) {
1032                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1033                                           to_reserve,
1034                                           BTRFS_RESERVE_NO_FLUSH);
1035                 if (ret) {
1036                         pending->error = ret;
1037                         goto no_free_objectid;
1038                 }
1039         }
1040
1041         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1042                                    objectid, pending->inherit);
1043         if (ret) {
1044                 pending->error = ret;
1045                 goto no_free_objectid;
1046         }
1047
1048         key.objectid = objectid;
1049         key.offset = (u64)-1;
1050         key.type = BTRFS_ROOT_ITEM_KEY;
1051
1052         rsv = trans->block_rsv;
1053         trans->block_rsv = &pending->block_rsv;
1054
1055         dentry = pending->dentry;
1056         parent = dget_parent(dentry);
1057         parent_inode = parent->d_inode;
1058         parent_root = BTRFS_I(parent_inode)->root;
1059         record_root_in_trans(trans, parent_root);
1060
1061         /*
1062          * insert the directory item
1063          */
1064         ret = btrfs_set_inode_index(parent_inode, &index);
1065         BUG_ON(ret); /* -ENOMEM */
1066
1067         /* check if there is a file/dir which has the same name. */
1068         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1069                                          btrfs_ino(parent_inode),
1070                                          dentry->d_name.name,
1071                                          dentry->d_name.len, 0);
1072         if (dir_item != NULL && !IS_ERR(dir_item)) {
1073                 pending->error = -EEXIST;
1074                 goto fail;
1075         } else if (IS_ERR(dir_item)) {
1076                 ret = PTR_ERR(dir_item);
1077                 btrfs_abort_transaction(trans, root, ret);
1078                 goto fail;
1079         }
1080         btrfs_release_path(path);
1081
1082         /*
1083          * pull in the delayed directory update
1084          * and the delayed inode item
1085          * otherwise we corrupt the FS during
1086          * snapshot
1087          */
1088         ret = btrfs_run_delayed_items(trans, root);
1089         if (ret) {      /* Transaction aborted */
1090                 btrfs_abort_transaction(trans, root, ret);
1091                 goto fail;
1092         }
1093
1094         record_root_in_trans(trans, root);
1095         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1096         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1097         btrfs_check_and_init_root_item(new_root_item);
1098
1099         root_flags = btrfs_root_flags(new_root_item);
1100         if (pending->readonly)
1101                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1102         else
1103                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1104         btrfs_set_root_flags(new_root_item, root_flags);
1105
1106         btrfs_set_root_generation_v2(new_root_item,
1107                         trans->transid);
1108         uuid_le_gen(&new_uuid);
1109         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1110         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1111                         BTRFS_UUID_SIZE);
1112         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1113         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1114         btrfs_set_root_otransid(new_root_item, trans->transid);
1115         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1116         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1117         btrfs_set_root_stransid(new_root_item, 0);
1118         btrfs_set_root_rtransid(new_root_item, 0);
1119
1120         old = btrfs_lock_root_node(root);
1121         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1122         if (ret) {
1123                 btrfs_tree_unlock(old);
1124                 free_extent_buffer(old);
1125                 btrfs_abort_transaction(trans, root, ret);
1126                 goto fail;
1127         }
1128
1129         btrfs_set_lock_blocking(old);
1130
1131         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1132         /* clean up in any case */
1133         btrfs_tree_unlock(old);
1134         free_extent_buffer(old);
1135         if (ret) {
1136                 btrfs_abort_transaction(trans, root, ret);
1137                 goto fail;
1138         }
1139
1140         /* see comments in should_cow_block() */
1141         root->force_cow = 1;
1142         smp_wmb();
1143
1144         btrfs_set_root_node(new_root_item, tmp);
1145         /* record when the snapshot was created in key.offset */
1146         key.offset = trans->transid;
1147         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1148         btrfs_tree_unlock(tmp);
1149         free_extent_buffer(tmp);
1150         if (ret) {
1151                 btrfs_abort_transaction(trans, root, ret);
1152                 goto fail;
1153         }
1154
1155         /*
1156          * insert root back/forward references
1157          */
1158         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1159                                  parent_root->root_key.objectid,
1160                                  btrfs_ino(parent_inode), index,
1161                                  dentry->d_name.name, dentry->d_name.len);
1162         if (ret) {
1163                 btrfs_abort_transaction(trans, root, ret);
1164                 goto fail;
1165         }
1166
1167         key.offset = (u64)-1;
1168         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1169         if (IS_ERR(pending->snap)) {
1170                 ret = PTR_ERR(pending->snap);
1171                 btrfs_abort_transaction(trans, root, ret);
1172                 goto fail;
1173         }
1174
1175         ret = btrfs_reloc_post_snapshot(trans, pending);
1176         if (ret) {
1177                 btrfs_abort_transaction(trans, root, ret);
1178                 goto fail;
1179         }
1180
1181         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1182         if (ret) {
1183                 btrfs_abort_transaction(trans, root, ret);
1184                 goto fail;
1185         }
1186
1187         ret = btrfs_insert_dir_item(trans, parent_root,
1188                                     dentry->d_name.name, dentry->d_name.len,
1189                                     parent_inode, &key,
1190                                     BTRFS_FT_DIR, index);
1191         /* We have check then name at the beginning, so it is impossible. */
1192         BUG_ON(ret == -EEXIST);
1193         if (ret) {
1194                 btrfs_abort_transaction(trans, root, ret);
1195                 goto fail;
1196         }
1197
1198         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1199                                          dentry->d_name.len * 2);
1200         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1201         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1202         if (ret)
1203                 btrfs_abort_transaction(trans, root, ret);
1204 fail:
1205         dput(parent);
1206         trans->block_rsv = rsv;
1207 no_free_objectid:
1208         kfree(new_root_item);
1209 root_item_alloc_fail:
1210         btrfs_free_path(path);
1211 path_alloc_fail:
1212         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1213         return ret;
1214 }
1215
1216 /*
1217  * create all the snapshots we've scheduled for creation
1218  */
1219 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1220                                              struct btrfs_fs_info *fs_info)
1221 {
1222         struct btrfs_pending_snapshot *pending;
1223         struct list_head *head = &trans->transaction->pending_snapshots;
1224
1225         list_for_each_entry(pending, head, list)
1226                 create_pending_snapshot(trans, fs_info, pending);
1227         return 0;
1228 }
1229
1230 static void update_super_roots(struct btrfs_root *root)
1231 {
1232         struct btrfs_root_item *root_item;
1233         struct btrfs_super_block *super;
1234
1235         super = root->fs_info->super_copy;
1236
1237         root_item = &root->fs_info->chunk_root->root_item;
1238         super->chunk_root = root_item->bytenr;
1239         super->chunk_root_generation = root_item->generation;
1240         super->chunk_root_level = root_item->level;
1241
1242         root_item = &root->fs_info->tree_root->root_item;
1243         super->root = root_item->bytenr;
1244         super->generation = root_item->generation;
1245         super->root_level = root_item->level;
1246         if (btrfs_test_opt(root, SPACE_CACHE))
1247                 super->cache_generation = root_item->generation;
1248 }
1249
1250 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1251 {
1252         int ret = 0;
1253         spin_lock(&info->trans_lock);
1254         if (info->running_transaction)
1255                 ret = info->running_transaction->in_commit;
1256         spin_unlock(&info->trans_lock);
1257         return ret;
1258 }
1259
1260 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1261 {
1262         int ret = 0;
1263         spin_lock(&info->trans_lock);
1264         if (info->running_transaction)
1265                 ret = info->running_transaction->blocked;
1266         spin_unlock(&info->trans_lock);
1267         return ret;
1268 }
1269
1270 /*
1271  * wait for the current transaction commit to start and block subsequent
1272  * transaction joins
1273  */
1274 static void wait_current_trans_commit_start(struct btrfs_root *root,
1275                                             struct btrfs_transaction *trans)
1276 {
1277         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1278 }
1279
1280 /*
1281  * wait for the current transaction to start and then become unblocked.
1282  * caller holds ref.
1283  */
1284 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1285                                          struct btrfs_transaction *trans)
1286 {
1287         wait_event(root->fs_info->transaction_wait,
1288                    trans->commit_done || (trans->in_commit && !trans->blocked));
1289 }
1290
1291 /*
1292  * commit transactions asynchronously. once btrfs_commit_transaction_async
1293  * returns, any subsequent transaction will not be allowed to join.
1294  */
1295 struct btrfs_async_commit {
1296         struct btrfs_trans_handle *newtrans;
1297         struct btrfs_root *root;
1298         struct delayed_work work;
1299 };
1300
1301 static void do_async_commit(struct work_struct *work)
1302 {
1303         struct btrfs_async_commit *ac =
1304                 container_of(work, struct btrfs_async_commit, work.work);
1305
1306         /*
1307          * We've got freeze protection passed with the transaction.
1308          * Tell lockdep about it.
1309          */
1310         rwsem_acquire_read(
1311                 &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1312                 0, 1, _THIS_IP_);
1313
1314         current->journal_info = ac->newtrans;
1315
1316         btrfs_commit_transaction(ac->newtrans, ac->root);
1317         kfree(ac);
1318 }
1319
1320 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1321                                    struct btrfs_root *root,
1322                                    int wait_for_unblock)
1323 {
1324         struct btrfs_async_commit *ac;
1325         struct btrfs_transaction *cur_trans;
1326
1327         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1328         if (!ac)
1329                 return -ENOMEM;
1330
1331         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1332         ac->root = root;
1333         ac->newtrans = btrfs_join_transaction(root);
1334         if (IS_ERR(ac->newtrans)) {
1335                 int err = PTR_ERR(ac->newtrans);
1336                 kfree(ac);
1337                 return err;
1338         }
1339
1340         /* take transaction reference */
1341         cur_trans = trans->transaction;
1342         atomic_inc(&cur_trans->use_count);
1343
1344         btrfs_end_transaction(trans, root);
1345
1346         /*
1347          * Tell lockdep we've released the freeze rwsem, since the
1348          * async commit thread will be the one to unlock it.
1349          */
1350         rwsem_release(&root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1351                       1, _THIS_IP_);
1352
1353         schedule_delayed_work(&ac->work, 0);
1354
1355         /* wait for transaction to start and unblock */
1356         if (wait_for_unblock)
1357                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1358         else
1359                 wait_current_trans_commit_start(root, cur_trans);
1360
1361         if (current->journal_info == trans)
1362                 current->journal_info = NULL;
1363
1364         put_transaction(cur_trans);
1365         return 0;
1366 }
1367
1368
1369 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1370                                 struct btrfs_root *root, int err)
1371 {
1372         struct btrfs_transaction *cur_trans = trans->transaction;
1373
1374         WARN_ON(trans->use_count > 1);
1375
1376         btrfs_abort_transaction(trans, root, err);
1377
1378         spin_lock(&root->fs_info->trans_lock);
1379         list_del_init(&cur_trans->list);
1380         if (cur_trans == root->fs_info->running_transaction) {
1381                 root->fs_info->running_transaction = NULL;
1382                 root->fs_info->trans_no_join = 0;
1383         }
1384         spin_unlock(&root->fs_info->trans_lock);
1385
1386         btrfs_cleanup_one_transaction(trans->transaction, root);
1387
1388         put_transaction(cur_trans);
1389         put_transaction(cur_trans);
1390
1391         trace_btrfs_transaction_commit(root);
1392
1393         btrfs_scrub_continue(root);
1394
1395         if (current->journal_info == trans)
1396                 current->journal_info = NULL;
1397
1398         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1399 }
1400
1401 /*
1402  * btrfs_transaction state sequence:
1403  *    in_commit = 0, blocked = 0  (initial)
1404  *    in_commit = 1, blocked = 1
1405  *    blocked = 0
1406  *    commit_done = 1
1407  */
1408 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1409                              struct btrfs_root *root)
1410 {
1411         unsigned long joined = 0;
1412         struct btrfs_transaction *cur_trans = trans->transaction;
1413         struct btrfs_transaction *prev_trans = NULL;
1414         DEFINE_WAIT(wait);
1415         int ret;
1416         int should_grow = 0;
1417         unsigned long now = get_seconds();
1418         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1419
1420         ret = btrfs_run_ordered_operations(root, 0);
1421         if (ret) {
1422                 btrfs_abort_transaction(trans, root, ret);
1423                 goto cleanup_transaction;
1424         }
1425
1426         if (cur_trans->aborted) {
1427                 ret = cur_trans->aborted;
1428                 goto cleanup_transaction;
1429         }
1430
1431         /* make a pass through all the delayed refs we have so far
1432          * any runnings procs may add more while we are here
1433          */
1434         ret = btrfs_run_delayed_refs(trans, root, 0);
1435         if (ret)
1436                 goto cleanup_transaction;
1437
1438         btrfs_trans_release_metadata(trans, root);
1439         trans->block_rsv = NULL;
1440
1441         cur_trans = trans->transaction;
1442
1443         /*
1444          * set the flushing flag so procs in this transaction have to
1445          * start sending their work down.
1446          */
1447         cur_trans->delayed_refs.flushing = 1;
1448
1449         if (!list_empty(&trans->new_bgs))
1450                 btrfs_create_pending_block_groups(trans, root);
1451
1452         ret = btrfs_run_delayed_refs(trans, root, 0);
1453         if (ret)
1454                 goto cleanup_transaction;
1455
1456         spin_lock(&cur_trans->commit_lock);
1457         if (cur_trans->in_commit) {
1458                 spin_unlock(&cur_trans->commit_lock);
1459                 atomic_inc(&cur_trans->use_count);
1460                 ret = btrfs_end_transaction(trans, root);
1461
1462                 wait_for_commit(root, cur_trans);
1463
1464                 put_transaction(cur_trans);
1465
1466                 return ret;
1467         }
1468
1469         trans->transaction->in_commit = 1;
1470         trans->transaction->blocked = 1;
1471         spin_unlock(&cur_trans->commit_lock);
1472         wake_up(&root->fs_info->transaction_blocked_wait);
1473
1474         spin_lock(&root->fs_info->trans_lock);
1475         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1476                 prev_trans = list_entry(cur_trans->list.prev,
1477                                         struct btrfs_transaction, list);
1478                 if (!prev_trans->commit_done) {
1479                         atomic_inc(&prev_trans->use_count);
1480                         spin_unlock(&root->fs_info->trans_lock);
1481
1482                         wait_for_commit(root, prev_trans);
1483
1484                         put_transaction(prev_trans);
1485                 } else {
1486                         spin_unlock(&root->fs_info->trans_lock);
1487                 }
1488         } else {
1489                 spin_unlock(&root->fs_info->trans_lock);
1490         }
1491
1492         if (!btrfs_test_opt(root, SSD) &&
1493             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1494                 should_grow = 1;
1495
1496         do {
1497                 int snap_pending = 0;
1498
1499                 joined = cur_trans->num_joined;
1500                 if (!list_empty(&trans->transaction->pending_snapshots))
1501                         snap_pending = 1;
1502
1503                 WARN_ON(cur_trans != trans->transaction);
1504
1505                 if (flush_on_commit || snap_pending) {
1506                         ret = btrfs_start_delalloc_inodes(root, 1);
1507                         if (ret) {
1508                                 btrfs_abort_transaction(trans, root, ret);
1509                                 goto cleanup_transaction;
1510                         }
1511                         btrfs_wait_ordered_extents(root, 1);
1512                 }
1513
1514                 ret = btrfs_run_delayed_items(trans, root);
1515                 if (ret)
1516                         goto cleanup_transaction;
1517
1518                 /*
1519                  * running the delayed items may have added new refs. account
1520                  * them now so that they hinder processing of more delayed refs
1521                  * as little as possible.
1522                  */
1523                 btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1524
1525                 /*
1526                  * rename don't use btrfs_join_transaction, so, once we
1527                  * set the transaction to blocked above, we aren't going
1528                  * to get any new ordered operations.  We can safely run
1529                  * it here and no for sure that nothing new will be added
1530                  * to the list
1531                  */
1532                 ret = btrfs_run_ordered_operations(root, 1);
1533                 if (ret) {
1534                         btrfs_abort_transaction(trans, root, ret);
1535                         goto cleanup_transaction;
1536                 }
1537
1538                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1539                                 TASK_UNINTERRUPTIBLE);
1540
1541                 if (atomic_read(&cur_trans->num_writers) > 1)
1542                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1543                 else if (should_grow)
1544                         schedule_timeout(1);
1545
1546                 finish_wait(&cur_trans->writer_wait, &wait);
1547         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1548                  (should_grow && cur_trans->num_joined != joined));
1549
1550         /*
1551          * Ok now we need to make sure to block out any other joins while we
1552          * commit the transaction.  We could have started a join before setting
1553          * no_join so make sure to wait for num_writers to == 1 again.
1554          */
1555         spin_lock(&root->fs_info->trans_lock);
1556         root->fs_info->trans_no_join = 1;
1557         spin_unlock(&root->fs_info->trans_lock);
1558         wait_event(cur_trans->writer_wait,
1559                    atomic_read(&cur_trans->num_writers) == 1);
1560
1561         /*
1562          * the reloc mutex makes sure that we stop
1563          * the balancing code from coming in and moving
1564          * extents around in the middle of the commit
1565          */
1566         mutex_lock(&root->fs_info->reloc_mutex);
1567
1568         /*
1569          * We needn't worry about the delayed items because we will
1570          * deal with them in create_pending_snapshot(), which is the
1571          * core function of the snapshot creation.
1572          */
1573         ret = create_pending_snapshots(trans, root->fs_info);
1574         if (ret) {
1575                 mutex_unlock(&root->fs_info->reloc_mutex);
1576                 goto cleanup_transaction;
1577         }
1578
1579         /*
1580          * We insert the dir indexes of the snapshots and update the inode
1581          * of the snapshots' parents after the snapshot creation, so there
1582          * are some delayed items which are not dealt with. Now deal with
1583          * them.
1584          *
1585          * We needn't worry that this operation will corrupt the snapshots,
1586          * because all the tree which are snapshoted will be forced to COW
1587          * the nodes and leaves.
1588          */
1589         ret = btrfs_run_delayed_items(trans, root);
1590         if (ret) {
1591                 mutex_unlock(&root->fs_info->reloc_mutex);
1592                 goto cleanup_transaction;
1593         }
1594
1595         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1596         if (ret) {
1597                 mutex_unlock(&root->fs_info->reloc_mutex);
1598                 goto cleanup_transaction;
1599         }
1600
1601         /*
1602          * make sure none of the code above managed to slip in a
1603          * delayed item
1604          */
1605         btrfs_assert_delayed_root_empty(root);
1606
1607         WARN_ON(cur_trans != trans->transaction);
1608
1609         btrfs_scrub_pause(root);
1610         /* btrfs_commit_tree_roots is responsible for getting the
1611          * various roots consistent with each other.  Every pointer
1612          * in the tree of tree roots has to point to the most up to date
1613          * root for every subvolume and other tree.  So, we have to keep
1614          * the tree logging code from jumping in and changing any
1615          * of the trees.
1616          *
1617          * At this point in the commit, there can't be any tree-log
1618          * writers, but a little lower down we drop the trans mutex
1619          * and let new people in.  By holding the tree_log_mutex
1620          * from now until after the super is written, we avoid races
1621          * with the tree-log code.
1622          */
1623         mutex_lock(&root->fs_info->tree_log_mutex);
1624
1625         ret = commit_fs_roots(trans, root);
1626         if (ret) {
1627                 mutex_unlock(&root->fs_info->tree_log_mutex);
1628                 mutex_unlock(&root->fs_info->reloc_mutex);
1629                 goto cleanup_transaction;
1630         }
1631
1632         /* commit_fs_roots gets rid of all the tree log roots, it is now
1633          * safe to free the root of tree log roots
1634          */
1635         btrfs_free_log_root_tree(trans, root->fs_info);
1636
1637         ret = commit_cowonly_roots(trans, root);
1638         if (ret) {
1639                 mutex_unlock(&root->fs_info->tree_log_mutex);
1640                 mutex_unlock(&root->fs_info->reloc_mutex);
1641                 goto cleanup_transaction;
1642         }
1643
1644         btrfs_prepare_extent_commit(trans, root);
1645
1646         cur_trans = root->fs_info->running_transaction;
1647
1648         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1649                             root->fs_info->tree_root->node);
1650         switch_commit_root(root->fs_info->tree_root);
1651
1652         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1653                             root->fs_info->chunk_root->node);
1654         switch_commit_root(root->fs_info->chunk_root);
1655
1656         assert_qgroups_uptodate(trans);
1657         update_super_roots(root);
1658
1659         if (!root->fs_info->log_root_recovering) {
1660                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1661                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1662         }
1663
1664         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1665                sizeof(*root->fs_info->super_copy));
1666
1667         trans->transaction->blocked = 0;
1668         spin_lock(&root->fs_info->trans_lock);
1669         root->fs_info->running_transaction = NULL;
1670         root->fs_info->trans_no_join = 0;
1671         spin_unlock(&root->fs_info->trans_lock);
1672         mutex_unlock(&root->fs_info->reloc_mutex);
1673
1674         wake_up(&root->fs_info->transaction_wait);
1675
1676         ret = btrfs_write_and_wait_transaction(trans, root);
1677         if (ret) {
1678                 btrfs_error(root->fs_info, ret,
1679                             "Error while writing out transaction.");
1680                 mutex_unlock(&root->fs_info->tree_log_mutex);
1681                 goto cleanup_transaction;
1682         }
1683
1684         ret = write_ctree_super(trans, root, 0);
1685         if (ret) {
1686                 mutex_unlock(&root->fs_info->tree_log_mutex);
1687                 goto cleanup_transaction;
1688         }
1689
1690         /*
1691          * the super is written, we can safely allow the tree-loggers
1692          * to go about their business
1693          */
1694         mutex_unlock(&root->fs_info->tree_log_mutex);
1695
1696         btrfs_finish_extent_commit(trans, root);
1697
1698         cur_trans->commit_done = 1;
1699
1700         root->fs_info->last_trans_committed = cur_trans->transid;
1701
1702         wake_up(&cur_trans->commit_wait);
1703
1704         spin_lock(&root->fs_info->trans_lock);
1705         list_del_init(&cur_trans->list);
1706         spin_unlock(&root->fs_info->trans_lock);
1707
1708         put_transaction(cur_trans);
1709         put_transaction(cur_trans);
1710
1711         if (trans->type < TRANS_JOIN_NOLOCK)
1712                 sb_end_intwrite(root->fs_info->sb);
1713
1714         trace_btrfs_transaction_commit(root);
1715
1716         btrfs_scrub_continue(root);
1717
1718         if (current->journal_info == trans)
1719                 current->journal_info = NULL;
1720
1721         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1722
1723         if (current != root->fs_info->transaction_kthread)
1724                 btrfs_run_delayed_iputs(root);
1725
1726         return ret;
1727
1728 cleanup_transaction:
1729         btrfs_trans_release_metadata(trans, root);
1730         trans->block_rsv = NULL;
1731         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1732 //      WARN_ON(1);
1733         if (current->journal_info == trans)
1734                 current->journal_info = NULL;
1735         cleanup_transaction(trans, root, ret);
1736
1737         return ret;
1738 }
1739
1740 /*
1741  * interface function to delete all the snapshots we have scheduled for deletion
1742  */
1743 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1744 {
1745         LIST_HEAD(list);
1746         struct btrfs_fs_info *fs_info = root->fs_info;
1747
1748         spin_lock(&fs_info->trans_lock);
1749         list_splice_init(&fs_info->dead_roots, &list);
1750         spin_unlock(&fs_info->trans_lock);
1751
1752         while (!list_empty(&list)) {
1753                 int ret;
1754
1755                 root = list_entry(list.next, struct btrfs_root, root_list);
1756                 list_del(&root->root_list);
1757
1758                 btrfs_kill_all_delayed_nodes(root);
1759
1760                 if (btrfs_header_backref_rev(root->node) <
1761                     BTRFS_MIXED_BACKREF_REV)
1762                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1763                 else
1764                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1765                 BUG_ON(ret < 0);
1766         }
1767         return 0;
1768 }