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
 160         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 161         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 162         extent_io_tree_init(&cur_trans->dirty_pages,
 163                              fs_info->btree_inode->i_mapping);
 164         fs_info->generation++;
 165         cur_trans->transid = fs_info->generation;
 166         fs_info->running_transaction = cur_trans;
 167         cur_trans->aborted = 0;
 168         spin_unlock(&fs_info->trans_lock);
 169
 170         return 0;
 171 }
 172
 173 /*
 174  * this does all the record keeping required to make sure that a reference
 175  * counted root is properly recorded in a given transaction.  This is required
 176  * to make sure the old root from before we joined the transaction is deleted
 177  * when the transaction commits
 178  */
 179 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 180                                struct btrfs_root *root)
 181 {
 182         if (root->ref_cows && root->last_trans < trans->transid) {
 183                 WARN_ON(root == root->fs_info->extent_root);
 184                 WARN_ON(root->commit_root != root->node);
 185
 186                 /*
 187                  * see below for in_trans_setup usage rules
 188                  * we have the reloc mutex held now, so there
 189                  * is only one writer in this function
 190                  */
 191                 root->in_trans_setup = 1;
 192
 193                 /* make sure readers find in_trans_setup before
 194                  * they find our root->last_trans update
 195                  */
 196                 smp_wmb();
 197
 198                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 199                 if (root->last_trans == trans->transid) {
 200                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 201                         return 0;
 202                 }
 203                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 204                            (unsigned long)root->root_key.objectid,
 205                            BTRFS_ROOT_TRANS_TAG);
 206                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 207                 root->last_trans = trans->transid;
 208
 209                 /* this is pretty tricky.  We don't want to
 210                  * take the relocation lock in btrfs_record_root_in_trans
 211                  * unless we're really doing the first setup for this root in
 212                  * this transaction.
 213                  *
 214                  * Normally we'd use root->last_trans as a flag to decide
 215                  * if we want to take the expensive mutex.
 216                  *
 217                  * But, we have to set root->last_trans before we
 218                  * init the relocation root, otherwise, we trip over warnings
 219                  * in ctree.c.  The solution used here is to flag ourselves
 220                  * with root->in_trans_setup.  When this is 1, we're still
 221                  * fixing up the reloc trees and everyone must wait.
 222                  *
 223                  * When this is zero, they can trust root->last_trans and fly
 224                  * through btrfs_record_root_in_trans without having to take the
 225                  * lock.  smp_wmb() makes sure that all the writes above are
 226                  * done before we pop in the zero below
 227                  */
 228                 btrfs_init_reloc_root(trans, root);
 229                 smp_wmb();
 230                 root->in_trans_setup = 0;
 231         }
 232         return 0;
 233 }
 234
 235
 236 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 237                                struct btrfs_root *root)
 238 {
 239         if (!root->ref_cows)
 240                 return 0;
 241
 242         /*
 243          * see record_root_in_trans for comments about in_trans_setup usage
 244          * and barriers
 245          */
 246         smp_rmb();
 247         if (root->last_trans == trans->transid &&
 248             !root->in_trans_setup)
 249                 return 0;
 250
 251         mutex_lock(&root->fs_info->reloc_mutex);
 252         record_root_in_trans(trans, root);
 253         mutex_unlock(&root->fs_info->reloc_mutex);
 254
 255         return 0;
 256 }
 257
 258 /* wait for commit against the current transaction to become unblocked
 259  * when this is done, it is safe to start a new transaction, but the current
 260  * transaction might not be fully on disk.
 261  */
 262 static void wait_current_trans(struct btrfs_root *root)
 263 {
 264         struct btrfs_transaction *cur_trans;
 265
 266         spin_lock(&root->fs_info->trans_lock);
 267         cur_trans = root->fs_info->running_transaction;
 268         if (cur_trans && cur_trans->blocked) {
 269                 atomic_inc(&cur_trans->use_count);
 270                 spin_unlock(&root->fs_info->trans_lock);
 271
 272                 wait_event(root->fs_info->transaction_wait,
 273                            !cur_trans->blocked);
 274                 put_transaction(cur_trans);
 275         } else {
 276                 spin_unlock(&root->fs_info->trans_lock);
 277         }
 278 }
 279
 280 static int may_wait_transaction(struct btrfs_root *root, int type)
 281 {
 282         if (root->fs_info->log_root_recovering)
 283                 return 0;
 284
 285         if (type == TRANS_USERSPACE)
 286                 return 1;
 287
 288         if (type == TRANS_START &&
 289             !atomic_read(&root->fs_info->open_ioctl_trans))
 290                 return 1;
 291
 292         return 0;
 293 }
 294
 295 static struct btrfs_trans_handle *
 296 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 297                   enum btrfs_reserve_flush_enum flush)
 298 {
 299         struct btrfs_trans_handle *h;
 300         struct btrfs_transaction *cur_trans;
 301         u64 num_bytes = 0;
 302         int ret;
 303         u64 qgroup_reserved = 0;
 304
 305         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 306                 return ERR_PTR(-EROFS);
 307
 308         if (current->journal_info) {
 309                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 310                 h = current->journal_info;
 311                 h->use_count++;
 312                 WARN_ON(h->use_count > 2);
 313                 h->orig_rsv = h->block_rsv;
 314                 h->block_rsv = NULL;
 315                 goto got_it;
 316         }
 317
 318         /*
 319          * Do the reservation before we join the transaction so we can do all
 320          * the appropriate flushing if need be.
 321          */
 322         if (num_items > 0 && root != root->fs_info->chunk_root) {
 323                 if (root->fs_info->quota_enabled &&
 324                     is_fstree(root->root_key.objectid)) {
 325                         qgroup_reserved = num_items * root->leafsize;
 326                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 327                         if (ret)
 328                                 return ERR_PTR(ret);
 329                 }
 330
 331                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 332                 ret = btrfs_block_rsv_add(root,
 333                                           &root->fs_info->trans_block_rsv,
 334                                           num_bytes, flush);
 335                 if (ret)
 336                         goto reserve_fail;
 337         }
 338 again:
 339         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 340         if (!h) {
 341                 ret = -ENOMEM;
 342                 goto alloc_fail;
 343         }
 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                 goto join_fail;
 371         }
 372
 373         cur_trans = root->fs_info->running_transaction;
 374
 375         h->transid = cur_trans->transid;
 376         h->transaction = cur_trans;
 377         h->blocks_used = 0;
 378         h->bytes_reserved = 0;
 379         h->root = root;
 380         h->delayed_ref_updates = 0;
 381         h->use_count = 1;
 382         h->adding_csums = 0;
 383         h->block_rsv = NULL;
 384         h->orig_rsv = NULL;
 385         h->aborted = 0;
 386         h->qgroup_reserved = qgroup_reserved;
 387         h->delayed_ref_elem.seq = 0;
 388         h->type = type;
 389         INIT_LIST_HEAD(&h->qgroup_ref_list);
 390         INIT_LIST_HEAD(&h->new_bgs);
 391
 392         smp_mb();
 393         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 394                 btrfs_commit_transaction(h, root);
 395                 goto again;
 396         }
 397
 398         if (num_bytes) {
 399                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 400                                               h->transid, num_bytes, 1);
 401                 h->block_rsv = &root->fs_info->trans_block_rsv;
 402                 h->bytes_reserved = num_bytes;
 403         }
 404
 405 got_it:
 406         btrfs_record_root_in_trans(h, root);
 407
 408         if (!current->journal_info && type != TRANS_USERSPACE)
 409                 current->journal_info = h;
 410         return h;
 411
 412 join_fail:
 413         if (type < TRANS_JOIN_NOLOCK)
 414                 sb_end_intwrite(root->fs_info->sb);
 415         kmem_cache_free(btrfs_trans_handle_cachep, h);
 416 alloc_fail:
 417         if (num_bytes)
 418                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 419                                         num_bytes);
 420 reserve_fail:
 421         if (qgroup_reserved)
 422                 btrfs_qgroup_free(root, qgroup_reserved);
 423         return ERR_PTR(ret);
 424 }
 425
 426 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 427                                                    int num_items)
 428 {
 429         return start_transaction(root, num_items, TRANS_START,
 430                                  BTRFS_RESERVE_FLUSH_ALL);
 431 }
 432
 433 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 434                                         struct btrfs_root *root, int num_items)
 435 {
 436         return start_transaction(root, num_items, TRANS_START,
 437                                  BTRFS_RESERVE_FLUSH_LIMIT);
 438 }
 439
 440 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 441 {
 442         return start_transaction(root, 0, TRANS_JOIN, 0);
 443 }
 444
 445 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 446 {
 447         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 448 }
 449
 450 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 451 {
 452         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 453 }
 454
 455 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 456 {
 457         return start_transaction(root, 0, TRANS_ATTACH, 0);
 458 }
 459
 460 /* wait for a transaction commit to be fully complete */
 461 static noinline void wait_for_commit(struct btrfs_root *root,
 462                                     struct btrfs_transaction *commit)
 463 {
 464         wait_event(commit->commit_wait, commit->commit_done);
 465 }
 466
 467 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 468 {
 469         struct btrfs_transaction *cur_trans = NULL, *t;
 470         int ret = 0;
 471
 472         if (transid) {
 473                 if (transid <= root->fs_info->last_trans_committed)
 474                         goto out;
 475
 476                 ret = -EINVAL;
 477                 /* find specified transaction */
 478                 spin_lock(&root->fs_info->trans_lock);
 479                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 480                         if (t->transid == transid) {
 481                                 cur_trans = t;
 482                                 atomic_inc(&cur_trans->use_count);
 483                                 ret = 0;
 484                                 break;
 485                         }
 486                         if (t->transid > transid) {
 487                                 ret = 0;
 488                                 break;
 489                         }
 490                 }
 491                 spin_unlock(&root->fs_info->trans_lock);
 492                 /* The specified transaction doesn't exist */
 493                 if (!cur_trans)
 494                         goto out;
 495         } else {
 496                 /* find newest transaction that is committing | committed */
 497                 spin_lock(&root->fs_info->trans_lock);
 498                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 499                                             list) {
 500                         if (t->in_commit) {
 501                                 if (t->commit_done)
 502                                         break;
 503                                 cur_trans = t;
 504                                 atomic_inc(&cur_trans->use_count);
 505                                 break;
 506                         }
 507                 }
 508                 spin_unlock(&root->fs_info->trans_lock);
 509                 if (!cur_trans)
 510                         goto out;  /* nothing committing|committed */
 511         }
 512
 513         wait_for_commit(root, cur_trans);
 514         put_transaction(cur_trans);
 515 out:
 516         return ret;
 517 }
 518
 519 void btrfs_throttle(struct btrfs_root *root)
 520 {
 521         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 522                 wait_current_trans(root);
 523 }
 524
 525 static int should_end_transaction(struct btrfs_trans_handle *trans,
 526                                   struct btrfs_root *root)
 527 {
 528         int ret;
 529
 530         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 531         return ret ? 1 : 0;
 532 }
 533
 534 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 535                                  struct btrfs_root *root)
 536 {
 537         struct btrfs_transaction *cur_trans = trans->transaction;
 538         int updates;
 539         int err;
 540
 541         smp_mb();
 542         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 543                 return 1;
 544
 545         updates = trans->delayed_ref_updates;
 546         trans->delayed_ref_updates = 0;
 547         if (updates) {
 548                 err = btrfs_run_delayed_refs(trans, root, updates);
 549                 if (err) /* Error code will also eval true */
 550                         return err;
 551         }
 552
 553         return should_end_transaction(trans, root);
 554 }
 555
 556 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 557                           struct btrfs_root *root, int throttle)
 558 {
 559         struct btrfs_transaction *cur_trans = trans->transaction;
 560         struct btrfs_fs_info *info = root->fs_info;
 561         int count = 0;
 562         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 563         int err = 0;
 564
 565         if (--trans->use_count) {
 566                 trans->block_rsv = trans->orig_rsv;
 567                 return 0;
 568         }
 569
 570         /*
 571          * do the qgroup accounting as early as possible
 572          */
 573         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 574
 575         btrfs_trans_release_metadata(trans, root);
 576         trans->block_rsv = NULL;
 577         /*
 578          * the same root has to be passed to start_transaction and
 579          * end_transaction. Subvolume quota depends on this.
 580          */
 581         WARN_ON(trans->root != root);
 582
 583         if (trans->qgroup_reserved) {
 584                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 585                 trans->qgroup_reserved = 0;
 586         }
 587
 588         if (!list_empty(&trans->new_bgs))
 589                 btrfs_create_pending_block_groups(trans, root);
 590
 591         while (count < 2) {
 592                 unsigned long cur = trans->delayed_ref_updates;
 593                 trans->delayed_ref_updates = 0;
 594                 if (cur &&
 595                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 596                         trans->delayed_ref_updates = 0;
 597                         btrfs_run_delayed_refs(trans, root, cur);
 598                 } else {
 599                         break;
 600                 }
 601                 count++;
 602         }
 603         btrfs_trans_release_metadata(trans, root);
 604         trans->block_rsv = NULL;
 605
 606         if (!list_empty(&trans->new_bgs))
 607                 btrfs_create_pending_block_groups(trans, root);
 608
 609         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 610             should_end_transaction(trans, root)) {
 611                 trans->transaction->blocked = 1;
 612                 smp_wmb();
 613         }
 614
 615         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 616                 if (throttle) {
 617                         /*
 618                          * We may race with somebody else here so end up having
 619                          * to call end_transaction on ourselves again, so inc
 620                          * our use_count.
 621                          */
 622                         trans->use_count++;
 623                         return btrfs_commit_transaction(trans, root);
 624                 } else {
 625                         wake_up_process(info->transaction_kthread);
 626                 }
 627         }
 628
 629         if (trans->type < TRANS_JOIN_NOLOCK)
 630                 sb_end_intwrite(root->fs_info->sb);
 631
 632         WARN_ON(cur_trans != info->running_transaction);
 633         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 634         atomic_dec(&cur_trans->num_writers);
 635
 636         smp_mb();
 637         if (waitqueue_active(&cur_trans->writer_wait))
 638                 wake_up(&cur_trans->writer_wait);
 639         put_transaction(cur_trans);
 640
 641         if (current->journal_info == trans)
 642                 current->journal_info = NULL;
 643
 644         if (throttle)
 645                 btrfs_run_delayed_iputs(root);
 646
 647         if (trans->aborted ||
 648             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 649                 err = -EIO;
 650         }
 651         assert_qgroups_uptodate(trans);
 652
 653         memset(trans, 0, sizeof(*trans));
 654         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 655         return err;
 656 }
 657
 658 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 659                           struct btrfs_root *root)
 660 {
 661         int ret;
 662
 663         ret = __btrfs_end_transaction(trans, root, 0);
 664         if (ret)
 665                 return ret;
 666         return 0;
 667 }
 668
 669 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 670                                    struct btrfs_root *root)
 671 {
 672         int ret;
 673
 674         ret = __btrfs_end_transaction(trans, root, 1);
 675         if (ret)
 676                 return ret;
 677         return 0;
 678 }
 679
 680 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 681                                 struct btrfs_root *root)
 682 {
 683         return __btrfs_end_transaction(trans, root, 1);
 684 }
 685
 686 /*
 687  * when btree blocks are allocated, they have some corresponding bits set for
 688  * them in one of two extent_io trees.  This is used to make sure all of
 689  * those extents are sent to disk but does not wait on them
 690  */
 691 int btrfs_write_marked_extents(struct btrfs_root *root,
 692                                struct extent_io_tree *dirty_pages, int mark)
 693 {
 694         int err = 0;
 695         int werr = 0;
 696         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 697         struct extent_state *cached_state = NULL;
 698         u64 start = 0;
 699         u64 end;
 700
 701         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 702                                       mark, &cached_state)) {
 703                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 704                                    mark, &cached_state, GFP_NOFS);
 705                 cached_state = NULL;
 706                 err = filemap_fdatawrite_range(mapping, start, end);
 707                 if (err)
 708                         werr = err;
 709                 cond_resched();
 710                 start = end + 1;
 711         }
 712         if (err)
 713                 werr = err;
 714         return werr;
 715 }
 716
 717 /*
 718  * when btree blocks are allocated, they have some corresponding bits set for
 719  * them in one of two extent_io trees.  This is used to make sure all of
 720  * those extents are on disk for transaction or log commit.  We wait
 721  * on all the pages and clear them from the dirty pages state tree
 722  */
 723 int btrfs_wait_marked_extents(struct btrfs_root *root,
 724                               struct extent_io_tree *dirty_pages, int mark)
 725 {
 726         int err = 0;
 727         int werr = 0;
 728         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 729         struct extent_state *cached_state = NULL;
 730         u64 start = 0;
 731         u64 end;
 732
 733         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 734                                       EXTENT_NEED_WAIT, &cached_state)) {
 735                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 736                                  0, 0, &cached_state, GFP_NOFS);
 737                 err = filemap_fdatawait_range(mapping, start, end);
 738                 if (err)
 739                         werr = err;
 740                 cond_resched();
 741                 start = end + 1;
 742         }
 743         if (err)
 744                 werr = err;
 745         return werr;
 746 }
 747
 748 /*
 749  * when btree blocks are allocated, they have some corresponding bits set for
 750  * them in one of two extent_io trees.  This is used to make sure all of
 751  * those extents are on disk for transaction or log commit
 752  */
 753 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 754                                 struct extent_io_tree *dirty_pages, int mark)
 755 {
 756         int ret;
 757         int ret2;
 758
 759         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 760         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 761
 762         if (ret)
 763                 return ret;
 764         if (ret2)
 765                 return ret2;
 766         return 0;
 767 }
 768
 769 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 770                                      struct btrfs_root *root)
 771 {
 772         if (!trans || !trans->transaction) {
 773                 struct inode *btree_inode;
 774                 btree_inode = root->fs_info->btree_inode;
 775                 return filemap_write_and_wait(btree_inode->i_mapping);
 776         }
 777         return btrfs_write_and_wait_marked_extents(root,
 778                                            &trans->transaction->dirty_pages,
 779                                            EXTENT_DIRTY);
 780 }
 781
 782 /*
 783  * this is used to update the root pointer in the tree of tree roots.
 784  *
 785  * But, in the case of the extent allocation tree, updating the root
 786  * pointer may allocate blocks which may change the root of the extent
 787  * allocation tree.
 788  *
 789  * So, this loops and repeats and makes sure the cowonly root didn't
 790  * change while the root pointer was being updated in the metadata.
 791  */
 792 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 793                                struct btrfs_root *root)
 794 {
 795         int ret;
 796         u64 old_root_bytenr;
 797         u64 old_root_used;
 798         struct btrfs_root *tree_root = root->fs_info->tree_root;
 799
 800         old_root_used = btrfs_root_used(&root->root_item);
 801         btrfs_write_dirty_block_groups(trans, root);
 802
 803         while (1) {
 804                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 805                 if (old_root_bytenr == root->node->start &&
 806                     old_root_used == btrfs_root_used(&root->root_item))
 807                         break;
 808
 809                 btrfs_set_root_node(&root->root_item, root->node);
 810                 ret = btrfs_update_root(trans, tree_root,
 811                                         &root->root_key,
 812                                         &root->root_item);
 813                 if (ret)
 814                         return ret;
 815
 816                 old_root_used = btrfs_root_used(&root->root_item);
 817                 ret = btrfs_write_dirty_block_groups(trans, root);
 818                 if (ret)
 819                         return ret;
 820         }
 821
 822         if (root != root->fs_info->extent_root)
 823                 switch_commit_root(root);
 824
 825         return 0;
 826 }
 827
 828 /*
 829  * update all the cowonly tree roots on disk
 830  *
 831  * The error handling in this function may not be obvious. Any of the
 832  * failures will cause the file system to go offline. We still need
 833  * to clean up the delayed refs.
 834  */
 835 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 836                                          struct btrfs_root *root)
 837 {
 838         struct btrfs_fs_info *fs_info = root->fs_info;
 839         struct list_head *next;
 840         struct extent_buffer *eb;
 841         int ret;
 842
 843         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 844         if (ret)
 845                 return ret;
 846
 847         eb = btrfs_lock_root_node(fs_info->tree_root);
 848         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 849                               0, &eb);
 850         btrfs_tree_unlock(eb);
 851         free_extent_buffer(eb);
 852
 853         if (ret)
 854                 return ret;
 855
 856         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 857         if (ret)
 858                 return ret;
 859
 860         ret = btrfs_run_dev_stats(trans, root->fs_info);
 861         WARN_ON(ret);
 862         ret = btrfs_run_dev_replace(trans, root->fs_info);
 863         WARN_ON(ret);
 864
 865         ret = btrfs_run_qgroups(trans, root->fs_info);
 866         BUG_ON(ret);
 867
 868         /* run_qgroups might have added some more refs */
 869         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 870         BUG_ON(ret);
 871
 872         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 873                 next = fs_info->dirty_cowonly_roots.next;
 874                 list_del_init(next);
 875                 root = list_entry(next, struct btrfs_root, dirty_list);
 876
 877                 ret = update_cowonly_root(trans, root);
 878                 if (ret)
 879                         return ret;
 880         }
 881
 882         down_write(&fs_info->extent_commit_sem);
 883         switch_commit_root(fs_info->extent_root);
 884         up_write(&fs_info->extent_commit_sem);
 885
 886         btrfs_after_dev_replace_commit(fs_info);
 887
 888         return 0;
 889 }
 890
 891 /*
 892  * dead roots are old snapshots that need to be deleted.  This allocates
 893  * a dirty root struct and adds it into the list of dead roots that need to
 894  * be deleted
 895  */
 896 int btrfs_add_dead_root(struct btrfs_root *root)
 897 {
 898         spin_lock(&root->fs_info->trans_lock);
 899         list_add(&root->root_list, &root->fs_info->dead_roots);
 900         spin_unlock(&root->fs_info->trans_lock);
 901         return 0;
 902 }
 903
 904 /*
 905  * update all the cowonly tree roots on disk
 906  */
 907 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 908                                     struct btrfs_root *root)
 909 {
 910         struct btrfs_root *gang[8];
 911         struct btrfs_fs_info *fs_info = root->fs_info;
 912         int i;
 913         int ret;
 914         int err = 0;
 915
 916         spin_lock(&fs_info->fs_roots_radix_lock);
 917         while (1) {
 918                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 919                                                  (void **)gang, 0,
 920                                                  ARRAY_SIZE(gang),
 921                                                  BTRFS_ROOT_TRANS_TAG);
 922                 if (ret == 0)
 923                         break;
 924                 for (i = 0; i < ret; i++) {
 925                         root = gang[i];
 926                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 927                                         (unsigned long)root->root_key.objectid,
 928                                         BTRFS_ROOT_TRANS_TAG);
 929                         spin_unlock(&fs_info->fs_roots_radix_lock);
 930
 931                         btrfs_free_log(trans, root);
 932                         btrfs_update_reloc_root(trans, root);
 933                         btrfs_orphan_commit_root(trans, root);
 934
 935                         btrfs_save_ino_cache(root, trans);
 936
 937                         /* see comments in should_cow_block() */
 938                         root->force_cow = 0;
 939                         smp_wmb();
 940
 941                         if (root->commit_root != root->node) {
 942                                 mutex_lock(&root->fs_commit_mutex);
 943                                 switch_commit_root(root);
 944                                 btrfs_unpin_free_ino(root);
 945                                 mutex_unlock(&root->fs_commit_mutex);
 946
 947                                 btrfs_set_root_node(&root->root_item,
 948                                                     root->node);
 949                         }
 950
 951                         err = btrfs_update_root(trans, fs_info->tree_root,
 952                                                 &root->root_key,
 953                                                 &root->root_item);
 954                         spin_lock(&fs_info->fs_roots_radix_lock);
 955                         if (err)
 956                                 break;
 957                 }
 958         }
 959         spin_unlock(&fs_info->fs_roots_radix_lock);
 960         return err;
 961 }
 962
 963 /*
 964  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 965  * otherwise every leaf in the btree is read and defragged.
 966  */
 967 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 968 {
 969         struct btrfs_fs_info *info = root->fs_info;
 970         struct btrfs_trans_handle *trans;
 971         int ret;
 972
 973         if (xchg(&root->defrag_running, 1))
 974                 return 0;
 975
 976         while (1) {
 977                 trans = btrfs_start_transaction(root, 0);
 978                 if (IS_ERR(trans))
 979                         return PTR_ERR(trans);
 980
 981                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 982
 983                 btrfs_end_transaction(trans, root);
 984                 btrfs_btree_balance_dirty(info->tree_root);
 985                 cond_resched();
 986
 987                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 988                         break;
 989         }
 990         root->defrag_running = 0;
 991         return ret;
 992 }
 993
 994 /*
 995  * new snapshots need to be created at a very specific time in the
 996  * transaction commit.  This does the actual creation
 997  */
 998 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 999                                    struct btrfs_fs_info *fs_info,
1000                                    struct btrfs_pending_snapshot *pending)
1001 {
1002         struct btrfs_key key;
1003         struct btrfs_root_item *new_root_item;
1004         struct btrfs_root *tree_root = fs_info->tree_root;
1005         struct btrfs_root *root = pending->root;
1006         struct btrfs_root *parent_root;
1007         struct btrfs_block_rsv *rsv;
1008         struct inode *parent_inode;
1009         struct btrfs_path *path;
1010         struct btrfs_dir_item *dir_item;
1011         struct dentry *parent;
1012         struct dentry *dentry;
1013         struct extent_buffer *tmp;
1014         struct extent_buffer *old;
1015         struct timespec cur_time = CURRENT_TIME;
1016         int ret;
1017         u64 to_reserve = 0;
1018         u64 index = 0;
1019         u64 objectid;
1020         u64 root_flags;
1021         uuid_le new_uuid;
1022
1023         path = btrfs_alloc_path();
1024         if (!path) {
1025                 ret = pending->error = -ENOMEM;
1026                 goto path_alloc_fail;
1027         }
1028
1029         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1030         if (!new_root_item) {
1031                 ret = pending->error = -ENOMEM;
1032                 goto root_item_alloc_fail;
1033         }
1034
1035         ret = btrfs_find_free_objectid(tree_root, &objectid);
1036         if (ret) {
1037                 pending->error = ret;
1038                 goto no_free_objectid;
1039         }
1040
1041         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1042
1043         if (to_reserve > 0) {
1044                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1045                                           to_reserve,
1046                                           BTRFS_RESERVE_NO_FLUSH);
1047                 if (ret) {
1048                         pending->error = ret;
1049                         goto no_free_objectid;
1050                 }
1051         }
1052
1053         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1054                                    objectid, pending->inherit);
1055         if (ret) {
1056                 pending->error = ret;
1057                 goto no_free_objectid;
1058         }
1059
1060         key.objectid = objectid;
1061         key.offset = (u64)-1;
1062         key.type = BTRFS_ROOT_ITEM_KEY;
1063
1064         rsv = trans->block_rsv;
1065         trans->block_rsv = &pending->block_rsv;
1066
1067         dentry = pending->dentry;
1068         parent = dget_parent(dentry);
1069         parent_inode = parent->d_inode;
1070         parent_root = BTRFS_I(parent_inode)->root;
1071         record_root_in_trans(trans, parent_root);
1072
1073         /*
1074          * insert the directory item
1075          */
1076         ret = btrfs_set_inode_index(parent_inode, &index);
1077         BUG_ON(ret); /* -ENOMEM */
1078
1079         /* check if there is a file/dir which has the same name. */
1080         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1081                                          btrfs_ino(parent_inode),
1082                                          dentry->d_name.name,
1083                                          dentry->d_name.len, 0);
1084         if (dir_item != NULL && !IS_ERR(dir_item)) {
1085                 pending->error = -EEXIST;
1086                 goto fail;
1087         } else if (IS_ERR(dir_item)) {
1088                 ret = PTR_ERR(dir_item);
1089                 btrfs_abort_transaction(trans, root, ret);
1090                 goto fail;
1091         }
1092         btrfs_release_path(path);
1093
1094         /*
1095          * pull in the delayed directory update
1096          * and the delayed inode item
1097          * otherwise we corrupt the FS during
1098          * snapshot
1099          */
1100         ret = btrfs_run_delayed_items(trans, root);
1101         if (ret) {      /* Transaction aborted */
1102                 btrfs_abort_transaction(trans, root, ret);
1103                 goto fail;
1104         }
1105
1106         record_root_in_trans(trans, root);
1107         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1108         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1109         btrfs_check_and_init_root_item(new_root_item);
1110
1111         root_flags = btrfs_root_flags(new_root_item);
1112         if (pending->readonly)
1113                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1114         else
1115                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1116         btrfs_set_root_flags(new_root_item, root_flags);
1117
1118         btrfs_set_root_generation_v2(new_root_item,
1119                         trans->transid);
1120         uuid_le_gen(&new_uuid);
1121         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1122         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1123                         BTRFS_UUID_SIZE);
1124         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1125         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1126         btrfs_set_root_otransid(new_root_item, trans->transid);
1127         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1128         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1129         btrfs_set_root_stransid(new_root_item, 0);
1130         btrfs_set_root_rtransid(new_root_item, 0);
1131
1132         old = btrfs_lock_root_node(root);
1133         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1134         if (ret) {
1135                 btrfs_tree_unlock(old);
1136                 free_extent_buffer(old);
1137                 btrfs_abort_transaction(trans, root, ret);
1138                 goto fail;
1139         }
1140
1141         btrfs_set_lock_blocking(old);
1142
1143         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1144         /* clean up in any case */
1145         btrfs_tree_unlock(old);
1146         free_extent_buffer(old);
1147         if (ret) {
1148                 btrfs_abort_transaction(trans, root, ret);
1149                 goto fail;
1150         }
1151
1152         /* see comments in should_cow_block() */
1153         root->force_cow = 1;
1154         smp_wmb();
1155
1156         btrfs_set_root_node(new_root_item, tmp);
1157         /* record when the snapshot was created in key.offset */
1158         key.offset = trans->transid;
1159         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1160         btrfs_tree_unlock(tmp);
1161         free_extent_buffer(tmp);
1162         if (ret) {
1163                 btrfs_abort_transaction(trans, root, ret);
1164                 goto fail;
1165         }
1166
1167         /*
1168          * insert root back/forward references
1169          */
1170         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1171                                  parent_root->root_key.objectid,
1172                                  btrfs_ino(parent_inode), index,
1173                                  dentry->d_name.name, dentry->d_name.len);
1174         if (ret) {
1175                 btrfs_abort_transaction(trans, root, ret);
1176                 goto fail;
1177         }
1178
1179         key.offset = (u64)-1;
1180         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1181         if (IS_ERR(pending->snap)) {
1182                 ret = PTR_ERR(pending->snap);
1183                 btrfs_abort_transaction(trans, root, ret);
1184                 goto fail;
1185         }
1186
1187         ret = btrfs_reloc_post_snapshot(trans, pending);
1188         if (ret) {
1189                 btrfs_abort_transaction(trans, root, ret);
1190                 goto fail;
1191         }
1192
1193         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1194         if (ret) {
1195                 btrfs_abort_transaction(trans, root, ret);
1196                 goto fail;
1197         }
1198
1199         ret = btrfs_insert_dir_item(trans, parent_root,
1200                                     dentry->d_name.name, dentry->d_name.len,
1201                                     parent_inode, &key,
1202                                     BTRFS_FT_DIR, index);
1203         /* We have check then name at the beginning, so it is impossible. */
1204         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1205         if (ret) {
1206                 btrfs_abort_transaction(trans, root, ret);
1207                 goto fail;
1208         }
1209
1210         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1211                                          dentry->d_name.len * 2);
1212         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1213         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1214         if (ret)
1215                 btrfs_abort_transaction(trans, root, ret);
1216 fail:
1217         dput(parent);
1218         trans->block_rsv = rsv;
1219 no_free_objectid:
1220         kfree(new_root_item);
1221 root_item_alloc_fail:
1222         btrfs_free_path(path);
1223 path_alloc_fail:
1224         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1225         return ret;
1226 }
1227
1228 /*
1229  * create all the snapshots we've scheduled for creation
1230  */
1231 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1232                                              struct btrfs_fs_info *fs_info)
1233 {
1234         struct btrfs_pending_snapshot *pending;
1235         struct list_head *head = &trans->transaction->pending_snapshots;
1236
1237         list_for_each_entry(pending, head, list)
1238                 create_pending_snapshot(trans, fs_info, pending);
1239         return 0;
1240 }
1241
1242 static void update_super_roots(struct btrfs_root *root)
1243 {
1244         struct btrfs_root_item *root_item;
1245         struct btrfs_super_block *super;
1246
1247         super = root->fs_info->super_copy;
1248
1249         root_item = &root->fs_info->chunk_root->root_item;
1250         super->chunk_root = root_item->bytenr;
1251         super->chunk_root_generation = root_item->generation;
1252         super->chunk_root_level = root_item->level;
1253
1254         root_item = &root->fs_info->tree_root->root_item;
1255         super->root = root_item->bytenr;
1256         super->generation = root_item->generation;
1257         super->root_level = root_item->level;
1258         if (btrfs_test_opt(root, SPACE_CACHE))
1259                 super->cache_generation = root_item->generation;
1260 }
1261
1262 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1263 {
1264         int ret = 0;
1265         spin_lock(&info->trans_lock);
1266         if (info->running_transaction)
1267                 ret = info->running_transaction->in_commit;
1268         spin_unlock(&info->trans_lock);
1269         return ret;
1270 }
1271
1272 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1273 {
1274         int ret = 0;
1275         spin_lock(&info->trans_lock);
1276         if (info->running_transaction)
1277                 ret = info->running_transaction->blocked;
1278         spin_unlock(&info->trans_lock);
1279         return ret;
1280 }
1281
1282 /*
1283  * wait for the current transaction commit to start and block subsequent
1284  * transaction joins
1285  */
1286 static void wait_current_trans_commit_start(struct btrfs_root *root,
1287                                             struct btrfs_transaction *trans)
1288 {
1289         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1290 }
1291
1292 /*
1293  * wait for the current transaction to start and then become unblocked.
1294  * caller holds ref.
1295  */
1296 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1297                                          struct btrfs_transaction *trans)
1298 {
1299         wait_event(root->fs_info->transaction_wait,
1300                    trans->commit_done || (trans->in_commit && !trans->blocked));
1301 }
1302
1303 /*
1304  * commit transactions asynchronously. once btrfs_commit_transaction_async
1305  * returns, any subsequent transaction will not be allowed to join.
1306  */
1307 struct btrfs_async_commit {
1308         struct btrfs_trans_handle *newtrans;
1309         struct btrfs_root *root;
1310         struct delayed_work work;
1311 };
1312
1313 static void do_async_commit(struct work_struct *work)
1314 {
1315         struct btrfs_async_commit *ac =
1316                 container_of(work, struct btrfs_async_commit, work.work);
1317
1318         /*
1319          * We've got freeze protection passed with the transaction.
1320          * Tell lockdep about it.
1321          */
1322         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1323                 rwsem_acquire_read(
1324                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1325                      0, 1, _THIS_IP_);
1326
1327         current->journal_info = ac->newtrans;
1328
1329         btrfs_commit_transaction(ac->newtrans, ac->root);
1330         kfree(ac);
1331 }
1332
1333 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1334                                    struct btrfs_root *root,
1335                                    int wait_for_unblock)
1336 {
1337         struct btrfs_async_commit *ac;
1338         struct btrfs_transaction *cur_trans;
1339
1340         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1341         if (!ac)
1342                 return -ENOMEM;
1343
1344         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1345         ac->root = root;
1346         ac->newtrans = btrfs_join_transaction(root);
1347         if (IS_ERR(ac->newtrans)) {
1348                 int err = PTR_ERR(ac->newtrans);
1349                 kfree(ac);
1350                 return err;
1351         }
1352
1353         /* take transaction reference */
1354         cur_trans = trans->transaction;
1355         atomic_inc(&cur_trans->use_count);
1356
1357         btrfs_end_transaction(trans, root);
1358
1359         /*
1360          * Tell lockdep we've released the freeze rwsem, since the
1361          * async commit thread will be the one to unlock it.
1362          */
1363         if (trans->type < TRANS_JOIN_NOLOCK)
1364                 rwsem_release(
1365                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1366                         1, _THIS_IP_);
1367
1368         schedule_delayed_work(&ac->work, 0);
1369
1370         /* wait for transaction to start and unblock */
1371         if (wait_for_unblock)
1372                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1373         else
1374                 wait_current_trans_commit_start(root, cur_trans);
1375
1376         if (current->journal_info == trans)
1377                 current->journal_info = NULL;
1378
1379         put_transaction(cur_trans);
1380         return 0;
1381 }
1382
1383
1384 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1385                                 struct btrfs_root *root, int err)
1386 {
1387         struct btrfs_transaction *cur_trans = trans->transaction;
1388
1389         WARN_ON(trans->use_count > 1);
1390
1391         btrfs_abort_transaction(trans, root, err);
1392
1393         spin_lock(&root->fs_info->trans_lock);
1394         list_del_init(&cur_trans->list);
1395         if (cur_trans == root->fs_info->running_transaction) {
1396                 root->fs_info->running_transaction = NULL;
1397                 root->fs_info->trans_no_join = 0;
1398         }
1399         spin_unlock(&root->fs_info->trans_lock);
1400
1401         btrfs_cleanup_one_transaction(trans->transaction, root);
1402
1403         put_transaction(cur_trans);
1404         put_transaction(cur_trans);
1405
1406         trace_btrfs_transaction_commit(root);
1407
1408         btrfs_scrub_continue(root);
1409
1410         if (current->journal_info == trans)
1411                 current->journal_info = NULL;
1412
1413         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1414 }
1415
1416 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1417                                           struct btrfs_root *root)
1418 {
1419         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1420         int snap_pending = 0;
1421         int ret;
1422
1423         if (!flush_on_commit) {
1424                 spin_lock(&root->fs_info->trans_lock);
1425                 if (!list_empty(&trans->transaction->pending_snapshots))
1426                         snap_pending = 1;
1427                 spin_unlock(&root->fs_info->trans_lock);
1428         }
1429
1430         if (flush_on_commit || snap_pending) {
1431                 btrfs_start_delalloc_inodes(root, 1);
1432                 btrfs_wait_ordered_extents(root, 1);
1433         }
1434
1435         ret = btrfs_run_delayed_items(trans, root);
1436         if (ret)
1437                 return ret;
1438
1439         /*
1440          * running the delayed items may have added new refs. account
1441          * them now so that they hinder processing of more delayed refs
1442          * as little as possible.
1443          */
1444         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1445
1446         /*
1447          * rename don't use btrfs_join_transaction, so, once we
1448          * set the transaction to blocked above, we aren't going
1449          * to get any new ordered operations.  We can safely run
1450          * it here and no for sure that nothing new will be added
1451          * to the list
1452          */
1453         btrfs_run_ordered_operations(root, 1);
1454
1455         return 0;
1456 }
1457
1458 /*
1459  * btrfs_transaction state sequence:
1460  *    in_commit = 0, blocked = 0  (initial)
1461  *    in_commit = 1, blocked = 1
1462  *    blocked = 0
1463  *    commit_done = 1
1464  */
1465 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1466                              struct btrfs_root *root)
1467 {
1468         unsigned long joined = 0;
1469         struct btrfs_transaction *cur_trans = trans->transaction;
1470         struct btrfs_transaction *prev_trans = NULL;
1471         DEFINE_WAIT(wait);
1472         int ret;
1473         int should_grow = 0;
1474         unsigned long now = get_seconds();
1475
1476         ret = btrfs_run_ordered_operations(root, 0);
1477         if (ret) {
1478                 btrfs_abort_transaction(trans, root, ret);
1479                 goto cleanup_transaction;
1480         }
1481
1482         /* Stop the commit early if ->aborted is set */
1483         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1484                 ret = cur_trans->aborted;
1485                 goto cleanup_transaction;
1486         }
1487
1488         /* make a pass through all the delayed refs we have so far
1489          * any runnings procs may add more while we are here
1490          */
1491         ret = btrfs_run_delayed_refs(trans, root, 0);
1492         if (ret)
1493                 goto cleanup_transaction;
1494
1495         btrfs_trans_release_metadata(trans, root);
1496         trans->block_rsv = NULL;
1497
1498         cur_trans = trans->transaction;
1499
1500         /*
1501          * set the flushing flag so procs in this transaction have to
1502          * start sending their work down.
1503          */
1504         cur_trans->delayed_refs.flushing = 1;
1505
1506         if (!list_empty(&trans->new_bgs))
1507                 btrfs_create_pending_block_groups(trans, root);
1508
1509         ret = btrfs_run_delayed_refs(trans, root, 0);
1510         if (ret)
1511                 goto cleanup_transaction;
1512
1513         spin_lock(&cur_trans->commit_lock);
1514         if (cur_trans->in_commit) {
1515                 spin_unlock(&cur_trans->commit_lock);
1516                 atomic_inc(&cur_trans->use_count);
1517                 ret = btrfs_end_transaction(trans, root);
1518
1519                 wait_for_commit(root, cur_trans);
1520
1521                 put_transaction(cur_trans);
1522
1523                 return ret;
1524         }
1525
1526         trans->transaction->in_commit = 1;
1527         trans->transaction->blocked = 1;
1528         spin_unlock(&cur_trans->commit_lock);
1529         wake_up(&root->fs_info->transaction_blocked_wait);
1530
1531         spin_lock(&root->fs_info->trans_lock);
1532         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1533                 prev_trans = list_entry(cur_trans->list.prev,
1534                                         struct btrfs_transaction, list);
1535                 if (!prev_trans->commit_done) {
1536                         atomic_inc(&prev_trans->use_count);
1537                         spin_unlock(&root->fs_info->trans_lock);
1538
1539                         wait_for_commit(root, prev_trans);
1540
1541                         put_transaction(prev_trans);
1542                 } else {
1543                         spin_unlock(&root->fs_info->trans_lock);
1544                 }
1545         } else {
1546                 spin_unlock(&root->fs_info->trans_lock);
1547         }
1548
1549         if (!btrfs_test_opt(root, SSD) &&
1550             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1551                 should_grow = 1;
1552
1553         do {
1554                 joined = cur_trans->num_joined;
1555
1556                 WARN_ON(cur_trans != trans->transaction);
1557
1558                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1559                 if (ret)
1560                         goto cleanup_transaction;
1561
1562                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1563                                 TASK_UNINTERRUPTIBLE);
1564
1565                 if (atomic_read(&cur_trans->num_writers) > 1)
1566                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1567                 else if (should_grow)
1568                         schedule_timeout(1);
1569
1570                 finish_wait(&cur_trans->writer_wait, &wait);
1571         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1572                  (should_grow && cur_trans->num_joined != joined));
1573
1574         ret = btrfs_flush_all_pending_stuffs(trans, root);
1575         if (ret)
1576                 goto cleanup_transaction;
1577
1578         /*
1579          * Ok now we need to make sure to block out any other joins while we
1580          * commit the transaction.  We could have started a join before setting
1581          * no_join so make sure to wait for num_writers to == 1 again.
1582          */
1583         spin_lock(&root->fs_info->trans_lock);
1584         root->fs_info->trans_no_join = 1;
1585         spin_unlock(&root->fs_info->trans_lock);
1586         wait_event(cur_trans->writer_wait,
1587                    atomic_read(&cur_trans->num_writers) == 1);
1588
1589         /* ->aborted might be set after the previous check, so check it */
1590         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1591                 ret = cur_trans->aborted;
1592                 goto cleanup_transaction;
1593         }
1594         /*
1595          * the reloc mutex makes sure that we stop
1596          * the balancing code from coming in and moving
1597          * extents around in the middle of the commit
1598          */
1599         mutex_lock(&root->fs_info->reloc_mutex);
1600
1601         /*
1602          * We needn't worry about the delayed items because we will
1603          * deal with them in create_pending_snapshot(), which is the
1604          * core function of the snapshot creation.
1605          */
1606         ret = create_pending_snapshots(trans, root->fs_info);
1607         if (ret) {
1608                 mutex_unlock(&root->fs_info->reloc_mutex);
1609                 goto cleanup_transaction;
1610         }
1611
1612         /*
1613          * We insert the dir indexes of the snapshots and update the inode
1614          * of the snapshots' parents after the snapshot creation, so there
1615          * are some delayed items which are not dealt with. Now deal with
1616          * them.
1617          *
1618          * We needn't worry that this operation will corrupt the snapshots,
1619          * because all the tree which are snapshoted will be forced to COW
1620          * the nodes and leaves.
1621          */
1622         ret = btrfs_run_delayed_items(trans, root);
1623         if (ret) {
1624                 mutex_unlock(&root->fs_info->reloc_mutex);
1625                 goto cleanup_transaction;
1626         }
1627
1628         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1629         if (ret) {
1630                 mutex_unlock(&root->fs_info->reloc_mutex);
1631                 goto cleanup_transaction;
1632         }
1633
1634         /*
1635          * make sure none of the code above managed to slip in a
1636          * delayed item
1637          */
1638         btrfs_assert_delayed_root_empty(root);
1639
1640         WARN_ON(cur_trans != trans->transaction);
1641
1642         btrfs_scrub_pause(root);
1643         /* btrfs_commit_tree_roots is responsible for getting the
1644          * various roots consistent with each other.  Every pointer
1645          * in the tree of tree roots has to point to the most up to date
1646          * root for every subvolume and other tree.  So, we have to keep
1647          * the tree logging code from jumping in and changing any
1648          * of the trees.
1649          *
1650          * At this point in the commit, there can't be any tree-log
1651          * writers, but a little lower down we drop the trans mutex
1652          * and let new people in.  By holding the tree_log_mutex
1653          * from now until after the super is written, we avoid races
1654          * with the tree-log code.
1655          */
1656         mutex_lock(&root->fs_info->tree_log_mutex);
1657
1658         ret = commit_fs_roots(trans, root);
1659         if (ret) {
1660                 mutex_unlock(&root->fs_info->tree_log_mutex);
1661                 mutex_unlock(&root->fs_info->reloc_mutex);
1662                 goto cleanup_transaction;
1663         }
1664
1665         /* commit_fs_roots gets rid of all the tree log roots, it is now
1666          * safe to free the root of tree log roots
1667          */
1668         btrfs_free_log_root_tree(trans, root->fs_info);
1669
1670         ret = commit_cowonly_roots(trans, root);
1671         if (ret) {
1672                 mutex_unlock(&root->fs_info->tree_log_mutex);
1673                 mutex_unlock(&root->fs_info->reloc_mutex);
1674                 goto cleanup_transaction;
1675         }
1676
1677         /*
1678          * The tasks which save the space cache and inode cache may also
1679          * update ->aborted, check it.
1680          */
1681         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1682                 ret = cur_trans->aborted;
1683                 mutex_unlock(&root->fs_info->tree_log_mutex);
1684                 mutex_unlock(&root->fs_info->reloc_mutex);
1685                 goto cleanup_transaction;
1686         }
1687
1688         btrfs_prepare_extent_commit(trans, root);
1689
1690         cur_trans = root->fs_info->running_transaction;
1691
1692         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1693                             root->fs_info->tree_root->node);
1694         switch_commit_root(root->fs_info->tree_root);
1695
1696         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1697                             root->fs_info->chunk_root->node);
1698         switch_commit_root(root->fs_info->chunk_root);
1699
1700         assert_qgroups_uptodate(trans);
1701         update_super_roots(root);
1702
1703         if (!root->fs_info->log_root_recovering) {
1704                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1705                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1706         }
1707
1708         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1709                sizeof(*root->fs_info->super_copy));
1710
1711         trans->transaction->blocked = 0;
1712         spin_lock(&root->fs_info->trans_lock);
1713         root->fs_info->running_transaction = NULL;
1714         root->fs_info->trans_no_join = 0;
1715         spin_unlock(&root->fs_info->trans_lock);
1716         mutex_unlock(&root->fs_info->reloc_mutex);
1717
1718         wake_up(&root->fs_info->transaction_wait);
1719
1720         ret = btrfs_write_and_wait_transaction(trans, root);
1721         if (ret) {
1722                 btrfs_error(root->fs_info, ret,
1723                             "Error while writing out transaction.");
1724                 mutex_unlock(&root->fs_info->tree_log_mutex);
1725                 goto cleanup_transaction;
1726         }
1727
1728         ret = write_ctree_super(trans, root, 0);
1729         if (ret) {
1730                 mutex_unlock(&root->fs_info->tree_log_mutex);
1731                 goto cleanup_transaction;
1732         }
1733
1734         /*
1735          * the super is written, we can safely allow the tree-loggers
1736          * to go about their business
1737          */
1738         mutex_unlock(&root->fs_info->tree_log_mutex);
1739
1740         btrfs_finish_extent_commit(trans, root);
1741
1742         cur_trans->commit_done = 1;
1743
1744         root->fs_info->last_trans_committed = cur_trans->transid;
1745
1746         wake_up(&cur_trans->commit_wait);
1747
1748         spin_lock(&root->fs_info->trans_lock);
1749         list_del_init(&cur_trans->list);
1750         spin_unlock(&root->fs_info->trans_lock);
1751
1752         put_transaction(cur_trans);
1753         put_transaction(cur_trans);
1754
1755         if (trans->type < TRANS_JOIN_NOLOCK)
1756                 sb_end_intwrite(root->fs_info->sb);
1757
1758         trace_btrfs_transaction_commit(root);
1759
1760         btrfs_scrub_continue(root);
1761
1762         if (current->journal_info == trans)
1763                 current->journal_info = NULL;
1764
1765         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1766
1767         if (current != root->fs_info->transaction_kthread)
1768                 btrfs_run_delayed_iputs(root);
1769
1770         return ret;
1771
1772 cleanup_transaction:
1773         btrfs_trans_release_metadata(trans, root);
1774         trans->block_rsv = NULL;
1775         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1776 //      WARN_ON(1);
1777         if (current->journal_info == trans)
1778                 current->journal_info = NULL;
1779         cleanup_transaction(trans, root, ret);
1780
1781         return ret;
1782 }
1783
1784 /*
1785  * interface function to delete all the snapshots we have scheduled for deletion
1786  */
1787 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1788 {
1789         LIST_HEAD(list);
1790         struct btrfs_fs_info *fs_info = root->fs_info;
1791
1792         spin_lock(&fs_info->trans_lock);
1793         list_splice_init(&fs_info->dead_roots, &list);
1794         spin_unlock(&fs_info->trans_lock);
1795
1796         while (!list_empty(&list)) {
1797                 int ret;
1798
1799                 root = list_entry(list.next, struct btrfs_root, root_list);
1800                 list_del(&root->root_list);
1801
1802                 btrfs_kill_all_delayed_nodes(root);
1803
1804                 if (btrfs_header_backref_rev(root->node) <
1805                     BTRFS_MIXED_BACKREF_REV)
1806                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1807                 else
1808                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1809                 BUG_ON(ret < 0);
1810         }
1811         return 0;
1812 }