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 #include "dev-replace.h"
  34
  35 #define BTRFS_ROOT_TRANS_TAG 0
  36
  37 void put_transaction(struct btrfs_transaction *transaction)
  38 {
  39         WARN_ON(atomic_read(&transaction->use_count) == 0);
  40         if (atomic_dec_and_test(&transaction->use_count)) {
  41                 BUG_ON(!list_empty(&transaction->list));
  42                 WARN_ON(transaction->delayed_refs.root.rb_node);
  43                 memset(transaction, 0, sizeof(*transaction));
  44                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  45         }
  46 }
  47
  48 static noinline void switch_commit_root(struct btrfs_root *root)
  49 {
  50         free_extent_buffer(root->commit_root);
  51         root->commit_root = btrfs_root_node(root);
  52 }
  53
  54 /*
  55  * either allocate a new transaction or hop into the existing one
  56  */
  57 static noinline int join_transaction(struct btrfs_root *root, int type)
  58 {
  59         struct btrfs_transaction *cur_trans;
  60         struct btrfs_fs_info *fs_info = root->fs_info;
  61
  62         spin_lock(&fs_info->trans_lock);
  63 loop:
  64         /* The file system has been taken offline. No new transactions. */
  65         if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  66                 spin_unlock(&fs_info->trans_lock);
  67                 return -EROFS;
  68         }
  69
  70         if (fs_info->trans_no_join) {
  71                 /*
  72                  * If we are JOIN_NOLOCK we're already committing a current
  73                  * transaction, we just need a handle to deal with something
  74                  * when committing the transaction, such as inode cache and
  75                  * space cache. It is a special case.
  76                  */
  77                 if (type != TRANS_JOIN_NOLOCK) {
  78                         spin_unlock(&fs_info->trans_lock);
  79                         return -EBUSY;
  80                 }
  81         }
  82
  83         cur_trans = fs_info->running_transaction;
  84         if (cur_trans) {
  85                 if (cur_trans->aborted) {
  86                         spin_unlock(&fs_info->trans_lock);
  87                         return cur_trans->aborted;
  88                 }
  89                 atomic_inc(&cur_trans->use_count);
  90                 atomic_inc(&cur_trans->num_writers);
  91                 cur_trans->num_joined++;
  92                 spin_unlock(&fs_info->trans_lock);
  93                 return 0;
  94         }
  95         spin_unlock(&fs_info->trans_lock);
  96
  97         /*
  98          * If we are ATTACH, we just want to catch the current transaction,
  99          * and commit it. If there is no transaction, just return ENOENT.
 100          */
 101         if (type == TRANS_ATTACH)
 102                 return -ENOENT;
 103
 104         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 105         if (!cur_trans)
 106                 return -ENOMEM;
 107
 108         spin_lock(&fs_info->trans_lock);
 109         if (fs_info->running_transaction) {
 110                 /*
 111                  * someone started a transaction after we unlocked.  Make sure
 112                  * to redo the trans_no_join checks above
 113                  */
 114                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 115                 cur_trans = fs_info->running_transaction;
 116                 goto loop;
 117         } else if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 118                 spin_unlock(&fs_info->trans_lock);
 119                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 120                 return -EROFS;
 121         }
 122
 123         atomic_set(&cur_trans->num_writers, 1);
 124         cur_trans->num_joined = 0;
 125         init_waitqueue_head(&cur_trans->writer_wait);
 126         init_waitqueue_head(&cur_trans->commit_wait);
 127         cur_trans->in_commit = 0;
 128         cur_trans->blocked = 0;
 129         /*
 130          * One for this trans handle, one so it will live on until we
 131          * commit the transaction.
 132          */
 133         atomic_set(&cur_trans->use_count, 2);
 134         cur_trans->commit_done = 0;
 135         cur_trans->start_time = get_seconds();
 136
 137         cur_trans->delayed_refs.root = RB_ROOT;
 138         cur_trans->delayed_refs.num_entries = 0;
 139         cur_trans->delayed_refs.num_heads_ready = 0;
 140         cur_trans->delayed_refs.num_heads = 0;
 141         cur_trans->delayed_refs.flushing = 0;
 142         cur_trans->delayed_refs.run_delayed_start = 0;
 143
 144         /*
 145          * although the tree mod log is per file system and not per transaction,
 146          * the log must never go across transaction boundaries.
 147          */
 148         smp_mb();
 149         if (!list_empty(&fs_info->tree_mod_seq_list))
 150                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 151                         "creating a fresh transaction\n");
 152         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 153                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 154                         "creating a fresh transaction\n");
 155         atomic_set(&fs_info->tree_mod_seq, 0);
 156
 157         spin_lock_init(&cur_trans->commit_lock);
 158         spin_lock_init(&cur_trans->delayed_refs.lock);
 159         atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
 160         atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
 161         init_waitqueue_head(&cur_trans->delayed_refs.wait);
 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                 WARN_ON(h->use_count > 2);
 316                 h->orig_rsv = h->block_rsv;
 317                 h->block_rsv = NULL;
 318                 goto got_it;
 319         }
 320
 321         /*
 322          * Do the reservation before we join the transaction so we can do all
 323          * the appropriate flushing if need be.
 324          */
 325         if (num_items > 0 && root != root->fs_info->chunk_root) {
 326                 if (root->fs_info->quota_enabled &&
 327                     is_fstree(root->root_key.objectid)) {
 328                         qgroup_reserved = num_items * root->leafsize;
 329                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 330                         if (ret)
 331                                 return ERR_PTR(ret);
 332                 }
 333
 334                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 335                 ret = btrfs_block_rsv_add(root,
 336                                           &root->fs_info->trans_block_rsv,
 337                                           num_bytes, flush);
 338                 if (ret)
 339                         return ERR_PTR(ret);
 340         }
 341 again:
 342         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 343         if (!h)
 344                 return ERR_PTR(-ENOMEM);
 345
 346         /*
 347          * If we are JOIN_NOLOCK we're already committing a transaction and
 348          * waiting on this guy, so we don't need to do the sb_start_intwrite
 349          * because we're already holding a ref.  We need this because we could
 350          * have raced in and did an fsync() on a file which can kick a commit
 351          * and then we deadlock with somebody doing a freeze.
 352          *
 353          * If we are ATTACH, it means we just want to catch the current
 354          * transaction and commit it, so we needn't do sb_start_intwrite().
 355          */
 356         if (type < TRANS_JOIN_NOLOCK)
 357                 sb_start_intwrite(root->fs_info->sb);
 358
 359         if (may_wait_transaction(root, type))
 360                 wait_current_trans(root);
 361
 362         do {
 363                 ret = join_transaction(root, type);
 364                 if (ret == -EBUSY)
 365                         wait_current_trans(root);
 366         } while (ret == -EBUSY);
 367
 368         if (ret < 0) {
 369                 /* We must get the transaction if we are JOIN_NOLOCK. */
 370                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 371
 372                 if (type < TRANS_JOIN_NOLOCK)
 373                         sb_end_intwrite(root->fs_info->sb);
 374                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 375                 return ERR_PTR(ret);
 376         }
 377
 378         cur_trans = root->fs_info->running_transaction;
 379
 380         h->transid = cur_trans->transid;
 381         h->transaction = cur_trans;
 382         h->blocks_used = 0;
 383         h->bytes_reserved = 0;
 384         h->root = root;
 385         h->delayed_ref_updates = 0;
 386         h->use_count = 1;
 387         h->adding_csums = 0;
 388         h->block_rsv = NULL;
 389         h->orig_rsv = NULL;
 390         h->aborted = 0;
 391         h->qgroup_reserved = qgroup_reserved;
 392         h->delayed_ref_elem.seq = 0;
 393         h->type = type;
 394         INIT_LIST_HEAD(&h->qgroup_ref_list);
 395         INIT_LIST_HEAD(&h->new_bgs);
 396
 397         smp_mb();
 398         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 399                 btrfs_commit_transaction(h, root);
 400                 goto again;
 401         }
 402
 403         if (num_bytes) {
 404                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 405                                               h->transid, num_bytes, 1);
 406                 h->block_rsv = &root->fs_info->trans_block_rsv;
 407                 h->bytes_reserved = num_bytes;
 408         }
 409
 410 got_it:
 411         btrfs_record_root_in_trans(h, root);
 412
 413         if (!current->journal_info && type != TRANS_USERSPACE)
 414                 current->journal_info = h;
 415         return h;
 416 }
 417
 418 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 419                                                    int num_items)
 420 {
 421         return start_transaction(root, num_items, TRANS_START,
 422                                  BTRFS_RESERVE_FLUSH_ALL);
 423 }
 424
 425 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 426                                         struct btrfs_root *root, int num_items)
 427 {
 428         return start_transaction(root, num_items, TRANS_START,
 429                                  BTRFS_RESERVE_FLUSH_LIMIT);
 430 }
 431
 432 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 433 {
 434         return start_transaction(root, 0, TRANS_JOIN, 0);
 435 }
 436
 437 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 438 {
 439         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 440 }
 441
 442 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 443 {
 444         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 445 }
 446
 447 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 448 {
 449         return start_transaction(root, 0, TRANS_ATTACH, 0);
 450 }
 451
 452 /* wait for a transaction commit to be fully complete */
 453 static noinline void wait_for_commit(struct btrfs_root *root,
 454                                     struct btrfs_transaction *commit)
 455 {
 456         wait_event(commit->commit_wait, commit->commit_done);
 457 }
 458
 459 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 460 {
 461         struct btrfs_transaction *cur_trans = NULL, *t;
 462         int ret = 0;
 463
 464         if (transid) {
 465                 if (transid <= root->fs_info->last_trans_committed)
 466                         goto out;
 467
 468                 ret = -EINVAL;
 469                 /* find specified transaction */
 470                 spin_lock(&root->fs_info->trans_lock);
 471                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 472                         if (t->transid == transid) {
 473                                 cur_trans = t;
 474                                 atomic_inc(&cur_trans->use_count);
 475                                 ret = 0;
 476                                 break;
 477                         }
 478                         if (t->transid > transid) {
 479                                 ret = 0;
 480                                 break;
 481                         }
 482                 }
 483                 spin_unlock(&root->fs_info->trans_lock);
 484                 /* The specified transaction doesn't exist */
 485                 if (!cur_trans)
 486                         goto out;
 487         } else {
 488                 /* find newest transaction that is committing | committed */
 489                 spin_lock(&root->fs_info->trans_lock);
 490                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 491                                             list) {
 492                         if (t->in_commit) {
 493                                 if (t->commit_done)
 494                                         break;
 495                                 cur_trans = t;
 496                                 atomic_inc(&cur_trans->use_count);
 497                                 break;
 498                         }
 499                 }
 500                 spin_unlock(&root->fs_info->trans_lock);
 501                 if (!cur_trans)
 502                         goto out;  /* nothing committing|committed */
 503         }
 504
 505         wait_for_commit(root, cur_trans);
 506         put_transaction(cur_trans);
 507 out:
 508         return ret;
 509 }
 510
 511 void btrfs_throttle(struct btrfs_root *root)
 512 {
 513         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 514                 wait_current_trans(root);
 515 }
 516
 517 static int should_end_transaction(struct btrfs_trans_handle *trans,
 518                                   struct btrfs_root *root)
 519 {
 520         int ret;
 521
 522         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 523         return ret ? 1 : 0;
 524 }
 525
 526 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 527                                  struct btrfs_root *root)
 528 {
 529         struct btrfs_transaction *cur_trans = trans->transaction;
 530         int updates;
 531         int err;
 532
 533         smp_mb();
 534         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 535                 return 1;
 536
 537         updates = trans->delayed_ref_updates;
 538         trans->delayed_ref_updates = 0;
 539         if (updates) {
 540                 err = btrfs_run_delayed_refs(trans, root, updates);
 541                 if (err) /* Error code will also eval true */
 542                         return err;
 543         }
 544
 545         return should_end_transaction(trans, root);
 546 }
 547
 548 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 549                           struct btrfs_root *root, int throttle)
 550 {
 551         struct btrfs_transaction *cur_trans = trans->transaction;
 552         struct btrfs_fs_info *info = root->fs_info;
 553         int count = 0;
 554         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 555         int err = 0;
 556
 557         if (--trans->use_count) {
 558                 trans->block_rsv = trans->orig_rsv;
 559                 return 0;
 560         }
 561
 562         /*
 563          * do the qgroup accounting as early as possible
 564          */
 565         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 566
 567         btrfs_trans_release_metadata(trans, root);
 568         trans->block_rsv = NULL;
 569         /*
 570          * the same root has to be passed to start_transaction and
 571          * end_transaction. Subvolume quota depends on this.
 572          */
 573         WARN_ON(trans->root != root);
 574
 575         if (trans->qgroup_reserved) {
 576                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 577                 trans->qgroup_reserved = 0;
 578         }
 579
 580         if (!list_empty(&trans->new_bgs))
 581                 btrfs_create_pending_block_groups(trans, root);
 582
 583         while (count < 1) {
 584                 unsigned long cur = trans->delayed_ref_updates;
 585                 trans->delayed_ref_updates = 0;
 586                 if (cur &&
 587                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 588                         trans->delayed_ref_updates = 0;
 589                         btrfs_run_delayed_refs(trans, root, cur);
 590                 } else {
 591                         break;
 592                 }
 593                 count++;
 594         }
 595
 596         btrfs_trans_release_metadata(trans, root);
 597         trans->block_rsv = NULL;
 598
 599         if (!list_empty(&trans->new_bgs))
 600                 btrfs_create_pending_block_groups(trans, root);
 601
 602         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 603             should_end_transaction(trans, root)) {
 604                 trans->transaction->blocked = 1;
 605                 smp_wmb();
 606         }
 607
 608         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 609                 if (throttle) {
 610                         /*
 611                          * We may race with somebody else here so end up having
 612                          * to call end_transaction on ourselves again, so inc
 613                          * our use_count.
 614                          */
 615                         trans->use_count++;
 616                         return btrfs_commit_transaction(trans, root);
 617                 } else {
 618                         wake_up_process(info->transaction_kthread);
 619                 }
 620         }
 621
 622         if (trans->type < TRANS_JOIN_NOLOCK)
 623                 sb_end_intwrite(root->fs_info->sb);
 624
 625         WARN_ON(cur_trans != info->running_transaction);
 626         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 627         atomic_dec(&cur_trans->num_writers);
 628
 629         smp_mb();
 630         if (waitqueue_active(&cur_trans->writer_wait))
 631                 wake_up(&cur_trans->writer_wait);
 632         put_transaction(cur_trans);
 633
 634         if (current->journal_info == trans)
 635                 current->journal_info = NULL;
 636
 637         if (throttle)
 638                 btrfs_run_delayed_iputs(root);
 639
 640         if (trans->aborted ||
 641             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 642                 err = -EIO;
 643         }
 644         assert_qgroups_uptodate(trans);
 645
 646         memset(trans, 0, sizeof(*trans));
 647         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 648         return err;
 649 }
 650
 651 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 652                           struct btrfs_root *root)
 653 {
 654         int ret;
 655
 656         ret = __btrfs_end_transaction(trans, root, 0);
 657         if (ret)
 658                 return ret;
 659         return 0;
 660 }
 661
 662 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 663                                    struct btrfs_root *root)
 664 {
 665         int ret;
 666
 667         ret = __btrfs_end_transaction(trans, root, 1);
 668         if (ret)
 669                 return ret;
 670         return 0;
 671 }
 672
 673 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 674                                 struct btrfs_root *root)
 675 {
 676         return __btrfs_end_transaction(trans, root, 1);
 677 }
 678
 679 /*
 680  * when btree blocks are allocated, they have some corresponding bits set for
 681  * them in one of two extent_io trees.  This is used to make sure all of
 682  * those extents are sent to disk but does not wait on them
 683  */
 684 int btrfs_write_marked_extents(struct btrfs_root *root,
 685                                struct extent_io_tree *dirty_pages, int mark)
 686 {
 687         int err = 0;
 688         int werr = 0;
 689         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 690         struct extent_state *cached_state = NULL;
 691         u64 start = 0;
 692         u64 end;
 693         struct blk_plug plug;
 694
 695         blk_start_plug(&plug);
 696         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 697                                       mark, &cached_state)) {
 698                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 699                                    mark, &cached_state, GFP_NOFS);
 700                 cached_state = NULL;
 701                 err = filemap_fdatawrite_range(mapping, start, end);
 702                 if (err)
 703                         werr = err;
 704                 cond_resched();
 705                 start = end + 1;
 706         }
 707         if (err)
 708                 werr = err;
 709         blk_finish_plug(&plug);
 710         return werr;
 711 }
 712
 713 /*
 714  * when btree blocks are allocated, they have some corresponding bits set for
 715  * them in one of two extent_io trees.  This is used to make sure all of
 716  * those extents are on disk for transaction or log commit.  We wait
 717  * on all the pages and clear them from the dirty pages state tree
 718  */
 719 int btrfs_wait_marked_extents(struct btrfs_root *root,
 720                               struct extent_io_tree *dirty_pages, int mark)
 721 {
 722         int err = 0;
 723         int werr = 0;
 724         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 725         struct extent_state *cached_state = NULL;
 726         u64 start = 0;
 727         u64 end;
 728
 729         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 730                                       EXTENT_NEED_WAIT, &cached_state)) {
 731                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 732                                  0, 0, &cached_state, GFP_NOFS);
 733                 err = filemap_fdatawait_range(mapping, start, end);
 734                 if (err)
 735                         werr = err;
 736                 cond_resched();
 737                 start = end + 1;
 738         }
 739         if (err)
 740                 werr = err;
 741         return werr;
 742 }
 743
 744 /*
 745  * when btree blocks are allocated, they have some corresponding bits set for
 746  * them in one of two extent_io trees.  This is used to make sure all of
 747  * those extents are on disk for transaction or log commit
 748  */
 749 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 750                                 struct extent_io_tree *dirty_pages, int mark)
 751 {
 752         int ret;
 753         int ret2;
 754
 755         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 756         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 757
 758         if (ret)
 759                 return ret;
 760         if (ret2)
 761                 return ret2;
 762         return 0;
 763 }
 764
 765 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 766                                      struct btrfs_root *root)
 767 {
 768         if (!trans || !trans->transaction) {
 769                 struct inode *btree_inode;
 770                 btree_inode = root->fs_info->btree_inode;
 771                 return filemap_write_and_wait(btree_inode->i_mapping);
 772         }
 773         return btrfs_write_and_wait_marked_extents(root,
 774                                            &trans->transaction->dirty_pages,
 775                                            EXTENT_DIRTY);
 776 }
 777
 778 /*
 779  * this is used to update the root pointer in the tree of tree roots.
 780  *
 781  * But, in the case of the extent allocation tree, updating the root
 782  * pointer may allocate blocks which may change the root of the extent
 783  * allocation tree.
 784  *
 785  * So, this loops and repeats and makes sure the cowonly root didn't
 786  * change while the root pointer was being updated in the metadata.
 787  */
 788 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 789                                struct btrfs_root *root)
 790 {
 791         int ret;
 792         u64 old_root_bytenr;
 793         u64 old_root_used;
 794         struct btrfs_root *tree_root = root->fs_info->tree_root;
 795
 796         old_root_used = btrfs_root_used(&root->root_item);
 797         btrfs_write_dirty_block_groups(trans, root);
 798
 799         while (1) {
 800                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 801                 if (old_root_bytenr == root->node->start &&
 802                     old_root_used == btrfs_root_used(&root->root_item))
 803                         break;
 804
 805                 btrfs_set_root_node(&root->root_item, root->node);
 806                 ret = btrfs_update_root(trans, tree_root,
 807                                         &root->root_key,
 808                                         &root->root_item);
 809                 if (ret)
 810                         return ret;
 811
 812                 old_root_used = btrfs_root_used(&root->root_item);
 813                 ret = btrfs_write_dirty_block_groups(trans, root);
 814                 if (ret)
 815                         return ret;
 816         }
 817
 818         if (root != root->fs_info->extent_root)
 819                 switch_commit_root(root);
 820
 821         return 0;
 822 }
 823
 824 /*
 825  * update all the cowonly tree roots on disk
 826  *
 827  * The error handling in this function may not be obvious. Any of the
 828  * failures will cause the file system to go offline. We still need
 829  * to clean up the delayed refs.
 830  */
 831 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 832                                          struct btrfs_root *root)
 833 {
 834         struct btrfs_fs_info *fs_info = root->fs_info;
 835         struct list_head *next;
 836         struct extent_buffer *eb;
 837         int ret;
 838
 839         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 840         if (ret)
 841                 return ret;
 842
 843         eb = btrfs_lock_root_node(fs_info->tree_root);
 844         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 845                               0, &eb);
 846         btrfs_tree_unlock(eb);
 847         free_extent_buffer(eb);
 848
 849         if (ret)
 850                 return ret;
 851
 852         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 853         if (ret)
 854                 return ret;
 855
 856         ret = btrfs_run_dev_stats(trans, root->fs_info);
 857         WARN_ON(ret);
 858         ret = btrfs_run_dev_replace(trans, root->fs_info);
 859         WARN_ON(ret);
 860
 861         ret = btrfs_run_qgroups(trans, root->fs_info);
 862         BUG_ON(ret);
 863
 864         /* run_qgroups might have added some more refs */
 865         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 866         BUG_ON(ret);
 867
 868         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 869                 next = fs_info->dirty_cowonly_roots.next;
 870                 list_del_init(next);
 871                 root = list_entry(next, struct btrfs_root, dirty_list);
 872
 873                 ret = update_cowonly_root(trans, root);
 874                 if (ret)
 875                         return ret;
 876         }
 877
 878         down_write(&fs_info->extent_commit_sem);
 879         switch_commit_root(fs_info->extent_root);
 880         up_write(&fs_info->extent_commit_sem);
 881
 882         btrfs_after_dev_replace_commit(fs_info);
 883
 884         return 0;
 885 }
 886
 887 /*
 888  * dead roots are old snapshots that need to be deleted.  This allocates
 889  * a dirty root struct and adds it into the list of dead roots that need to
 890  * be deleted
 891  */
 892 int btrfs_add_dead_root(struct btrfs_root *root)
 893 {
 894         spin_lock(&root->fs_info->trans_lock);
 895         list_add(&root->root_list, &root->fs_info->dead_roots);
 896         spin_unlock(&root->fs_info->trans_lock);
 897         return 0;
 898 }
 899
 900 /*
 901  * update all the cowonly tree roots on disk
 902  */
 903 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 904                                     struct btrfs_root *root)
 905 {
 906         struct btrfs_root *gang[8];
 907         struct btrfs_fs_info *fs_info = root->fs_info;
 908         int i;
 909         int ret;
 910         int err = 0;
 911
 912         spin_lock(&fs_info->fs_roots_radix_lock);
 913         while (1) {
 914                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 915                                                  (void **)gang, 0,
 916                                                  ARRAY_SIZE(gang),
 917                                                  BTRFS_ROOT_TRANS_TAG);
 918                 if (ret == 0)
 919                         break;
 920                 for (i = 0; i < ret; i++) {
 921                         root = gang[i];
 922                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 923                                         (unsigned long)root->root_key.objectid,
 924                                         BTRFS_ROOT_TRANS_TAG);
 925                         spin_unlock(&fs_info->fs_roots_radix_lock);
 926
 927                         btrfs_free_log(trans, root);
 928                         btrfs_update_reloc_root(trans, root);
 929                         btrfs_orphan_commit_root(trans, root);
 930
 931                         btrfs_save_ino_cache(root, trans);
 932
 933                         /* see comments in should_cow_block() */
 934                         root->force_cow = 0;
 935                         smp_wmb();
 936
 937                         if (root->commit_root != root->node) {
 938                                 mutex_lock(&root->fs_commit_mutex);
 939                                 switch_commit_root(root);
 940                                 btrfs_unpin_free_ino(root);
 941                                 mutex_unlock(&root->fs_commit_mutex);
 942
 943                                 btrfs_set_root_node(&root->root_item,
 944                                                     root->node);
 945                         }
 946
 947                         err = btrfs_update_root(trans, fs_info->tree_root,
 948                                                 &root->root_key,
 949                                                 &root->root_item);
 950                         spin_lock(&fs_info->fs_roots_radix_lock);
 951                         if (err)
 952                                 break;
 953                 }
 954         }
 955         spin_unlock(&fs_info->fs_roots_radix_lock);
 956         return err;
 957 }
 958
 959 /*
 960  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 961  * otherwise every leaf in the btree is read and defragged.
 962  */
 963 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 964 {
 965         struct btrfs_fs_info *info = root->fs_info;
 966         struct btrfs_trans_handle *trans;
 967         int ret;
 968
 969         if (xchg(&root->defrag_running, 1))
 970                 return 0;
 971
 972         while (1) {
 973                 trans = btrfs_start_transaction(root, 0);
 974                 if (IS_ERR(trans))
 975                         return PTR_ERR(trans);
 976
 977                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 978
 979                 btrfs_end_transaction(trans, root);
 980                 btrfs_btree_balance_dirty(info->tree_root);
 981                 cond_resched();
 982
 983                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 984                         break;
 985         }
 986         root->defrag_running = 0;
 987         return ret;
 988 }
 989
 990 /*
 991  * new snapshots need to be created at a very specific time in the
 992  * transaction commit.  This does the actual creation
 993  */
 994 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 995                                    struct btrfs_fs_info *fs_info,
 996                                    struct btrfs_pending_snapshot *pending)
 997 {
 998         struct btrfs_key key;
 999         struct btrfs_root_item *new_root_item;
1000         struct btrfs_root *tree_root = fs_info->tree_root;
1001         struct btrfs_root *root = pending->root;
1002         struct btrfs_root *parent_root;
1003         struct btrfs_block_rsv *rsv;
1004         struct inode *parent_inode;
1005         struct btrfs_path *path;
1006         struct btrfs_dir_item *dir_item;
1007         struct dentry *parent;
1008         struct dentry *dentry;
1009         struct extent_buffer *tmp;
1010         struct extent_buffer *old;
1011         struct timespec cur_time = CURRENT_TIME;
1012         int ret;
1013         u64 to_reserve = 0;
1014         u64 index = 0;
1015         u64 objectid;
1016         u64 root_flags;
1017         uuid_le new_uuid;
1018
1019         path = btrfs_alloc_path();
1020         if (!path) {
1021                 ret = pending->error = -ENOMEM;
1022                 goto path_alloc_fail;
1023         }
1024
1025         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1026         if (!new_root_item) {
1027                 ret = pending->error = -ENOMEM;
1028                 goto root_item_alloc_fail;
1029         }
1030
1031         ret = btrfs_find_free_objectid(tree_root, &objectid);
1032         if (ret) {
1033                 pending->error = ret;
1034                 goto no_free_objectid;
1035         }
1036
1037         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1038
1039         if (to_reserve > 0) {
1040                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1041                                           to_reserve,
1042                                           BTRFS_RESERVE_NO_FLUSH);
1043                 if (ret) {
1044                         pending->error = ret;
1045                         goto no_free_objectid;
1046                 }
1047         }
1048
1049         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1050                                    objectid, pending->inherit);
1051         if (ret) {
1052                 pending->error = ret;
1053                 goto no_free_objectid;
1054         }
1055
1056         key.objectid = objectid;
1057         key.offset = (u64)-1;
1058         key.type = BTRFS_ROOT_ITEM_KEY;
1059
1060         rsv = trans->block_rsv;
1061         trans->block_rsv = &pending->block_rsv;
1062
1063         dentry = pending->dentry;
1064         parent = dget_parent(dentry);
1065         parent_inode = parent->d_inode;
1066         parent_root = BTRFS_I(parent_inode)->root;
1067         record_root_in_trans(trans, parent_root);
1068
1069         /*
1070          * insert the directory item
1071          */
1072         ret = btrfs_set_inode_index(parent_inode, &index);
1073         BUG_ON(ret); /* -ENOMEM */
1074
1075         /* check if there is a file/dir which has the same name. */
1076         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1077                                          btrfs_ino(parent_inode),
1078                                          dentry->d_name.name,
1079                                          dentry->d_name.len, 0);
1080         if (dir_item != NULL && !IS_ERR(dir_item)) {
1081                 pending->error = -EEXIST;
1082                 goto fail;
1083         } else if (IS_ERR(dir_item)) {
1084                 ret = PTR_ERR(dir_item);
1085                 btrfs_abort_transaction(trans, root, ret);
1086                 goto fail;
1087         }
1088         btrfs_release_path(path);
1089
1090         /*
1091          * pull in the delayed directory update
1092          * and the delayed inode item
1093          * otherwise we corrupt the FS during
1094          * snapshot
1095          */
1096         ret = btrfs_run_delayed_items(trans, root);
1097         if (ret) {      /* Transaction aborted */
1098                 btrfs_abort_transaction(trans, root, ret);
1099                 goto fail;
1100         }
1101
1102         record_root_in_trans(trans, root);
1103         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1104         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1105         btrfs_check_and_init_root_item(new_root_item);
1106
1107         root_flags = btrfs_root_flags(new_root_item);
1108         if (pending->readonly)
1109                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1110         else
1111                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1112         btrfs_set_root_flags(new_root_item, root_flags);
1113
1114         btrfs_set_root_generation_v2(new_root_item,
1115                         trans->transid);
1116         uuid_le_gen(&new_uuid);
1117         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1118         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1119                         BTRFS_UUID_SIZE);
1120         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1121         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1122         btrfs_set_root_otransid(new_root_item, trans->transid);
1123         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1124         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1125         btrfs_set_root_stransid(new_root_item, 0);
1126         btrfs_set_root_rtransid(new_root_item, 0);
1127
1128         old = btrfs_lock_root_node(root);
1129         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1130         if (ret) {
1131                 btrfs_tree_unlock(old);
1132                 free_extent_buffer(old);
1133                 btrfs_abort_transaction(trans, root, ret);
1134                 goto fail;
1135         }
1136
1137         btrfs_set_lock_blocking(old);
1138
1139         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1140         /* clean up in any case */
1141         btrfs_tree_unlock(old);
1142         free_extent_buffer(old);
1143         if (ret) {
1144                 btrfs_abort_transaction(trans, root, ret);
1145                 goto fail;
1146         }
1147
1148         /* see comments in should_cow_block() */
1149         root->force_cow = 1;
1150         smp_wmb();
1151
1152         btrfs_set_root_node(new_root_item, tmp);
1153         /* record when the snapshot was created in key.offset */
1154         key.offset = trans->transid;
1155         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1156         btrfs_tree_unlock(tmp);
1157         free_extent_buffer(tmp);
1158         if (ret) {
1159                 btrfs_abort_transaction(trans, root, ret);
1160                 goto fail;
1161         }
1162
1163         /*
1164          * insert root back/forward references
1165          */
1166         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1167                                  parent_root->root_key.objectid,
1168                                  btrfs_ino(parent_inode), index,
1169                                  dentry->d_name.name, dentry->d_name.len);
1170         if (ret) {
1171                 btrfs_abort_transaction(trans, root, ret);
1172                 goto fail;
1173         }
1174
1175         key.offset = (u64)-1;
1176         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1177         if (IS_ERR(pending->snap)) {
1178                 ret = PTR_ERR(pending->snap);
1179                 btrfs_abort_transaction(trans, root, ret);
1180                 goto fail;
1181         }
1182
1183         ret = btrfs_reloc_post_snapshot(trans, pending);
1184         if (ret) {
1185                 btrfs_abort_transaction(trans, root, ret);
1186                 goto fail;
1187         }
1188
1189         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1190         if (ret) {
1191                 btrfs_abort_transaction(trans, root, ret);
1192                 goto fail;
1193         }
1194
1195         ret = btrfs_insert_dir_item(trans, parent_root,
1196                                     dentry->d_name.name, dentry->d_name.len,
1197                                     parent_inode, &key,
1198                                     BTRFS_FT_DIR, index);
1199         /* We have check then name at the beginning, so it is impossible. */
1200         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1201         if (ret) {
1202                 btrfs_abort_transaction(trans, root, ret);
1203                 goto fail;
1204         }
1205
1206         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1207                                          dentry->d_name.len * 2);
1208         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1209         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1210         if (ret)
1211                 btrfs_abort_transaction(trans, root, ret);
1212 fail:
1213         dput(parent);
1214         trans->block_rsv = rsv;
1215 no_free_objectid:
1216         kfree(new_root_item);
1217 root_item_alloc_fail:
1218         btrfs_free_path(path);
1219 path_alloc_fail:
1220         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1221         return ret;
1222 }
1223
1224 /*
1225  * create all the snapshots we've scheduled for creation
1226  */
1227 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1228                                              struct btrfs_fs_info *fs_info)
1229 {
1230         struct btrfs_pending_snapshot *pending;
1231         struct list_head *head = &trans->transaction->pending_snapshots;
1232
1233         list_for_each_entry(pending, head, list)
1234                 create_pending_snapshot(trans, fs_info, pending);
1235         return 0;
1236 }
1237
1238 static void update_super_roots(struct btrfs_root *root)
1239 {
1240         struct btrfs_root_item *root_item;
1241         struct btrfs_super_block *super;
1242
1243         super = root->fs_info->super_copy;
1244
1245         root_item = &root->fs_info->chunk_root->root_item;
1246         super->chunk_root = root_item->bytenr;
1247         super->chunk_root_generation = root_item->generation;
1248         super->chunk_root_level = root_item->level;
1249
1250         root_item = &root->fs_info->tree_root->root_item;
1251         super->root = root_item->bytenr;
1252         super->generation = root_item->generation;
1253         super->root_level = root_item->level;
1254         if (btrfs_test_opt(root, SPACE_CACHE))
1255                 super->cache_generation = root_item->generation;
1256 }
1257
1258 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1259 {
1260         int ret = 0;
1261         spin_lock(&info->trans_lock);
1262         if (info->running_transaction)
1263                 ret = info->running_transaction->in_commit;
1264         spin_unlock(&info->trans_lock);
1265         return ret;
1266 }
1267
1268 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1269 {
1270         int ret = 0;
1271         spin_lock(&info->trans_lock);
1272         if (info->running_transaction)
1273                 ret = info->running_transaction->blocked;
1274         spin_unlock(&info->trans_lock);
1275         return ret;
1276 }
1277
1278 /*
1279  * wait for the current transaction commit to start and block subsequent
1280  * transaction joins
1281  */
1282 static void wait_current_trans_commit_start(struct btrfs_root *root,
1283                                             struct btrfs_transaction *trans)
1284 {
1285         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1286 }
1287
1288 /*
1289  * wait for the current transaction to start and then become unblocked.
1290  * caller holds ref.
1291  */
1292 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1293                                          struct btrfs_transaction *trans)
1294 {
1295         wait_event(root->fs_info->transaction_wait,
1296                    trans->commit_done || (trans->in_commit && !trans->blocked));
1297 }
1298
1299 /*
1300  * commit transactions asynchronously. once btrfs_commit_transaction_async
1301  * returns, any subsequent transaction will not be allowed to join.
1302  */
1303 struct btrfs_async_commit {
1304         struct btrfs_trans_handle *newtrans;
1305         struct btrfs_root *root;
1306         struct delayed_work work;
1307 };
1308
1309 static void do_async_commit(struct work_struct *work)
1310 {
1311         struct btrfs_async_commit *ac =
1312                 container_of(work, struct btrfs_async_commit, work.work);
1313
1314         /*
1315          * We've got freeze protection passed with the transaction.
1316          * Tell lockdep about it.
1317          */
1318         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1319                 rwsem_acquire_read(
1320                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1321                      0, 1, _THIS_IP_);
1322
1323         current->journal_info = ac->newtrans;
1324
1325         btrfs_commit_transaction(ac->newtrans, ac->root);
1326         kfree(ac);
1327 }
1328
1329 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1330                                    struct btrfs_root *root,
1331                                    int wait_for_unblock)
1332 {
1333         struct btrfs_async_commit *ac;
1334         struct btrfs_transaction *cur_trans;
1335
1336         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1337         if (!ac)
1338                 return -ENOMEM;
1339
1340         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1341         ac->root = root;
1342         ac->newtrans = btrfs_join_transaction(root);
1343         if (IS_ERR(ac->newtrans)) {
1344                 int err = PTR_ERR(ac->newtrans);
1345                 kfree(ac);
1346                 return err;
1347         }
1348
1349         /* take transaction reference */
1350         cur_trans = trans->transaction;
1351         atomic_inc(&cur_trans->use_count);
1352
1353         btrfs_end_transaction(trans, root);
1354
1355         /*
1356          * Tell lockdep we've released the freeze rwsem, since the
1357          * async commit thread will be the one to unlock it.
1358          */
1359         if (trans->type < TRANS_JOIN_NOLOCK)
1360                 rwsem_release(
1361                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1362                         1, _THIS_IP_);
1363
1364         schedule_delayed_work(&ac->work, 0);
1365
1366         /* wait for transaction to start and unblock */
1367         if (wait_for_unblock)
1368                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1369         else
1370                 wait_current_trans_commit_start(root, cur_trans);
1371
1372         if (current->journal_info == trans)
1373                 current->journal_info = NULL;
1374
1375         put_transaction(cur_trans);
1376         return 0;
1377 }
1378
1379
1380 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1381                                 struct btrfs_root *root, int err)
1382 {
1383         struct btrfs_transaction *cur_trans = trans->transaction;
1384
1385         WARN_ON(trans->use_count > 1);
1386
1387         btrfs_abort_transaction(trans, root, err);
1388
1389         spin_lock(&root->fs_info->trans_lock);
1390         list_del_init(&cur_trans->list);
1391         if (cur_trans == root->fs_info->running_transaction) {
1392                 root->fs_info->running_transaction = NULL;
1393                 root->fs_info->trans_no_join = 0;
1394         }
1395         spin_unlock(&root->fs_info->trans_lock);
1396
1397         btrfs_cleanup_one_transaction(trans->transaction, root);
1398
1399         put_transaction(cur_trans);
1400         put_transaction(cur_trans);
1401
1402         trace_btrfs_transaction_commit(root);
1403
1404         btrfs_scrub_continue(root);
1405
1406         if (current->journal_info == trans)
1407                 current->journal_info = NULL;
1408
1409         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1410 }
1411
1412 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1413                                           struct btrfs_root *root)
1414 {
1415         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1416         int snap_pending = 0;
1417         int ret;
1418
1419         if (!flush_on_commit) {
1420                 spin_lock(&root->fs_info->trans_lock);
1421                 if (!list_empty(&trans->transaction->pending_snapshots))
1422                         snap_pending = 1;
1423                 spin_unlock(&root->fs_info->trans_lock);
1424         }
1425
1426         if (flush_on_commit || snap_pending) {
1427                 btrfs_start_delalloc_inodes(root, 1);
1428                 btrfs_wait_ordered_extents(root, 1);
1429         }
1430
1431         ret = btrfs_run_delayed_items(trans, root);
1432         if (ret)
1433                 return ret;
1434
1435         /*
1436          * running the delayed items may have added new refs. account
1437          * them now so that they hinder processing of more delayed refs
1438          * as little as possible.
1439          */
1440         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1441
1442         /*
1443          * rename don't use btrfs_join_transaction, so, once we
1444          * set the transaction to blocked above, we aren't going
1445          * to get any new ordered operations.  We can safely run
1446          * it here and no for sure that nothing new will be added
1447          * to the list
1448          */
1449         btrfs_run_ordered_operations(root, 1);
1450
1451         return 0;
1452 }
1453
1454 /*
1455  * btrfs_transaction state sequence:
1456  *    in_commit = 0, blocked = 0  (initial)
1457  *    in_commit = 1, blocked = 1
1458  *    blocked = 0
1459  *    commit_done = 1
1460  */
1461 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1462                              struct btrfs_root *root)
1463 {
1464         unsigned long joined = 0;
1465         struct btrfs_transaction *cur_trans = trans->transaction;
1466         struct btrfs_transaction *prev_trans = NULL;
1467         DEFINE_WAIT(wait);
1468         int ret;
1469         int should_grow = 0;
1470         unsigned long now = get_seconds();
1471
1472         ret = btrfs_run_ordered_operations(root, 0);
1473         if (ret) {
1474                 btrfs_abort_transaction(trans, root, ret);
1475                 goto cleanup_transaction;
1476         }
1477
1478         if (cur_trans->aborted) {
1479                 ret = cur_trans->aborted;
1480                 goto cleanup_transaction;
1481         }
1482
1483         /* make a pass through all the delayed refs we have so far
1484          * any runnings procs may add more while we are here
1485          */
1486         ret = btrfs_run_delayed_refs(trans, root, 0);
1487         if (ret)
1488                 goto cleanup_transaction;
1489
1490         btrfs_trans_release_metadata(trans, root);
1491         trans->block_rsv = NULL;
1492
1493         cur_trans = trans->transaction;
1494
1495         /*
1496          * set the flushing flag so procs in this transaction have to
1497          * start sending their work down.
1498          */
1499         cur_trans->delayed_refs.flushing = 1;
1500
1501         if (!list_empty(&trans->new_bgs))
1502                 btrfs_create_pending_block_groups(trans, root);
1503
1504         ret = btrfs_run_delayed_refs(trans, root, 0);
1505         if (ret)
1506                 goto cleanup_transaction;
1507
1508         spin_lock(&cur_trans->commit_lock);
1509         if (cur_trans->in_commit) {
1510                 spin_unlock(&cur_trans->commit_lock);
1511                 atomic_inc(&cur_trans->use_count);
1512                 ret = btrfs_end_transaction(trans, root);
1513
1514                 wait_for_commit(root, cur_trans);
1515
1516                 put_transaction(cur_trans);
1517
1518                 return ret;
1519         }
1520
1521         trans->transaction->in_commit = 1;
1522         trans->transaction->blocked = 1;
1523         spin_unlock(&cur_trans->commit_lock);
1524         wake_up(&root->fs_info->transaction_blocked_wait);
1525
1526         spin_lock(&root->fs_info->trans_lock);
1527         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1528                 prev_trans = list_entry(cur_trans->list.prev,
1529                                         struct btrfs_transaction, list);
1530                 if (!prev_trans->commit_done) {
1531                         atomic_inc(&prev_trans->use_count);
1532                         spin_unlock(&root->fs_info->trans_lock);
1533
1534                         wait_for_commit(root, prev_trans);
1535
1536                         put_transaction(prev_trans);
1537                 } else {
1538                         spin_unlock(&root->fs_info->trans_lock);
1539                 }
1540         } else {
1541                 spin_unlock(&root->fs_info->trans_lock);
1542         }
1543
1544         if (!btrfs_test_opt(root, SSD) &&
1545             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1546                 should_grow = 1;
1547
1548         do {
1549                 joined = cur_trans->num_joined;
1550
1551                 WARN_ON(cur_trans != trans->transaction);
1552
1553                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1554                 if (ret)
1555                         goto cleanup_transaction;
1556
1557                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1558                                 TASK_UNINTERRUPTIBLE);
1559
1560                 if (atomic_read(&cur_trans->num_writers) > 1)
1561                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1562                 else if (should_grow)
1563                         schedule_timeout(1);
1564
1565                 finish_wait(&cur_trans->writer_wait, &wait);
1566         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1567                  (should_grow && cur_trans->num_joined != joined));
1568
1569         ret = btrfs_flush_all_pending_stuffs(trans, root);
1570         if (ret)
1571                 goto cleanup_transaction;
1572
1573         /*
1574          * Ok now we need to make sure to block out any other joins while we
1575          * commit the transaction.  We could have started a join before setting
1576          * no_join so make sure to wait for num_writers to == 1 again.
1577          */
1578         spin_lock(&root->fs_info->trans_lock);
1579         root->fs_info->trans_no_join = 1;
1580         spin_unlock(&root->fs_info->trans_lock);
1581         wait_event(cur_trans->writer_wait,
1582                    atomic_read(&cur_trans->num_writers) == 1);
1583
1584         /*
1585          * the reloc mutex makes sure that we stop
1586          * the balancing code from coming in and moving
1587          * extents around in the middle of the commit
1588          */
1589         mutex_lock(&root->fs_info->reloc_mutex);
1590
1591         /*
1592          * We needn't worry about the delayed items because we will
1593          * deal with them in create_pending_snapshot(), which is the
1594          * core function of the snapshot creation.
1595          */
1596         ret = create_pending_snapshots(trans, root->fs_info);
1597         if (ret) {
1598                 mutex_unlock(&root->fs_info->reloc_mutex);
1599                 goto cleanup_transaction;
1600         }
1601
1602         /*
1603          * We insert the dir indexes of the snapshots and update the inode
1604          * of the snapshots' parents after the snapshot creation, so there
1605          * are some delayed items which are not dealt with. Now deal with
1606          * them.
1607          *
1608          * We needn't worry that this operation will corrupt the snapshots,
1609          * because all the tree which are snapshoted will be forced to COW
1610          * the nodes and leaves.
1611          */
1612         ret = btrfs_run_delayed_items(trans, root);
1613         if (ret) {
1614                 mutex_unlock(&root->fs_info->reloc_mutex);
1615                 goto cleanup_transaction;
1616         }
1617
1618         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1619         if (ret) {
1620                 mutex_unlock(&root->fs_info->reloc_mutex);
1621                 goto cleanup_transaction;
1622         }
1623
1624         /*
1625          * make sure none of the code above managed to slip in a
1626          * delayed item
1627          */
1628         btrfs_assert_delayed_root_empty(root);
1629
1630         WARN_ON(cur_trans != trans->transaction);
1631
1632         btrfs_scrub_pause(root);
1633         /* btrfs_commit_tree_roots is responsible for getting the
1634          * various roots consistent with each other.  Every pointer
1635          * in the tree of tree roots has to point to the most up to date
1636          * root for every subvolume and other tree.  So, we have to keep
1637          * the tree logging code from jumping in and changing any
1638          * of the trees.
1639          *
1640          * At this point in the commit, there can't be any tree-log
1641          * writers, but a little lower down we drop the trans mutex
1642          * and let new people in.  By holding the tree_log_mutex
1643          * from now until after the super is written, we avoid races
1644          * with the tree-log code.
1645          */
1646         mutex_lock(&root->fs_info->tree_log_mutex);
1647
1648         ret = commit_fs_roots(trans, root);
1649         if (ret) {
1650                 mutex_unlock(&root->fs_info->tree_log_mutex);
1651                 mutex_unlock(&root->fs_info->reloc_mutex);
1652                 goto cleanup_transaction;
1653         }
1654
1655         /* commit_fs_roots gets rid of all the tree log roots, it is now
1656          * safe to free the root of tree log roots
1657          */
1658         btrfs_free_log_root_tree(trans, root->fs_info);
1659
1660         ret = commit_cowonly_roots(trans, root);
1661         if (ret) {
1662                 mutex_unlock(&root->fs_info->tree_log_mutex);
1663                 mutex_unlock(&root->fs_info->reloc_mutex);
1664                 goto cleanup_transaction;
1665         }
1666
1667         btrfs_prepare_extent_commit(trans, root);
1668
1669         cur_trans = root->fs_info->running_transaction;
1670
1671         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1672                             root->fs_info->tree_root->node);
1673         switch_commit_root(root->fs_info->tree_root);
1674
1675         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1676                             root->fs_info->chunk_root->node);
1677         switch_commit_root(root->fs_info->chunk_root);
1678
1679         assert_qgroups_uptodate(trans);
1680         update_super_roots(root);
1681
1682         if (!root->fs_info->log_root_recovering) {
1683                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1684                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1685         }
1686
1687         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1688                sizeof(*root->fs_info->super_copy));
1689
1690         trans->transaction->blocked = 0;
1691         spin_lock(&root->fs_info->trans_lock);
1692         root->fs_info->running_transaction = NULL;
1693         root->fs_info->trans_no_join = 0;
1694         spin_unlock(&root->fs_info->trans_lock);
1695         mutex_unlock(&root->fs_info->reloc_mutex);
1696
1697         wake_up(&root->fs_info->transaction_wait);
1698
1699         ret = btrfs_write_and_wait_transaction(trans, root);
1700         if (ret) {
1701                 btrfs_error(root->fs_info, ret,
1702                             "Error while writing out transaction.");
1703                 mutex_unlock(&root->fs_info->tree_log_mutex);
1704                 goto cleanup_transaction;
1705         }
1706
1707         ret = write_ctree_super(trans, root, 0);
1708         if (ret) {
1709                 mutex_unlock(&root->fs_info->tree_log_mutex);
1710                 goto cleanup_transaction;
1711         }
1712
1713         /*
1714          * the super is written, we can safely allow the tree-loggers
1715          * to go about their business
1716          */
1717         mutex_unlock(&root->fs_info->tree_log_mutex);
1718
1719         btrfs_finish_extent_commit(trans, root);
1720
1721         cur_trans->commit_done = 1;
1722
1723         root->fs_info->last_trans_committed = cur_trans->transid;
1724
1725         wake_up(&cur_trans->commit_wait);
1726
1727         spin_lock(&root->fs_info->trans_lock);
1728         list_del_init(&cur_trans->list);
1729         spin_unlock(&root->fs_info->trans_lock);
1730
1731         put_transaction(cur_trans);
1732         put_transaction(cur_trans);
1733
1734         if (trans->type < TRANS_JOIN_NOLOCK)
1735                 sb_end_intwrite(root->fs_info->sb);
1736
1737         trace_btrfs_transaction_commit(root);
1738
1739         btrfs_scrub_continue(root);
1740
1741         if (current->journal_info == trans)
1742                 current->journal_info = NULL;
1743
1744         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1745
1746         if (current != root->fs_info->transaction_kthread)
1747                 btrfs_run_delayed_iputs(root);
1748
1749         return ret;
1750
1751 cleanup_transaction:
1752         btrfs_trans_release_metadata(trans, root);
1753         trans->block_rsv = NULL;
1754         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1755 //      WARN_ON(1);
1756         if (current->journal_info == trans)
1757                 current->journal_info = NULL;
1758         cleanup_transaction(trans, root, ret);
1759
1760         return ret;
1761 }
1762
1763 /*
1764  * interface function to delete all the snapshots we have scheduled for deletion
1765  */
1766 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1767 {
1768         LIST_HEAD(list);
1769         struct btrfs_fs_info *fs_info = root->fs_info;
1770
1771         spin_lock(&fs_info->trans_lock);
1772         list_splice_init(&fs_info->dead_roots, &list);
1773         spin_unlock(&fs_info->trans_lock);
1774
1775         while (!list_empty(&list)) {
1776                 int ret;
1777
1778                 root = list_entry(list.next, struct btrfs_root, root_list);
1779                 list_del(&root->root_list);
1780
1781                 btrfs_kill_all_delayed_nodes(root);
1782
1783                 if (btrfs_header_backref_rev(root->node) <
1784                     BTRFS_MIXED_BACKREF_REV)
1785                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1786                 else
1787                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1788                 BUG_ON(ret < 0);
1789         }
1790         return 0;
1791 }