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                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  44         }
  45 }
  46
  47 static noinline void switch_commit_root(struct btrfs_root *root)
  48 {
  49         free_extent_buffer(root->commit_root);
  50         root->commit_root = btrfs_root_node(root);
  51 }
  52
  53 /*
  54  * either allocate a new transaction or hop into the existing one
  55  */
  56 static noinline int join_transaction(struct btrfs_root *root, int type)
  57 {
  58         struct btrfs_transaction *cur_trans;
  59         struct btrfs_fs_info *fs_info = root->fs_info;
  60
  61         spin_lock(&fs_info->trans_lock);
  62 loop:
  63         /* The file system has been taken offline. No new transactions. */
  64         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
  65                 spin_unlock(&fs_info->trans_lock);
  66                 return -EROFS;
  67         }
  68
  69         if (fs_info->trans_no_join) {
  70                 /*
  71                  * If we are JOIN_NOLOCK we're already committing a current
  72                  * transaction, we just need a handle to deal with something
  73                  * when committing the transaction, such as inode cache and
  74                  * space cache. It is a special case.
  75                  */
  76                 if (type != TRANS_JOIN_NOLOCK) {
  77                         spin_unlock(&fs_info->trans_lock);
  78                         return -EBUSY;
  79                 }
  80         }
  81
  82         cur_trans = fs_info->running_transaction;
  83         if (cur_trans) {
  84                 if (cur_trans->aborted) {
  85                         spin_unlock(&fs_info->trans_lock);
  86                         return cur_trans->aborted;
  87                 }
  88                 atomic_inc(&cur_trans->use_count);
  89                 atomic_inc(&cur_trans->num_writers);
  90                 cur_trans->num_joined++;
  91                 spin_unlock(&fs_info->trans_lock);
  92                 return 0;
  93         }
  94         spin_unlock(&fs_info->trans_lock);
  95
  96         /*
  97          * If we are ATTACH, we just want to catch the current transaction,
  98          * and commit it. If there is no transaction, just return ENOENT.
  99          */
 100         if (type == TRANS_ATTACH)
 101                 return -ENOENT;
 102
 103         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 104         if (!cur_trans)
 105                 return -ENOMEM;
 106
 107         spin_lock(&fs_info->trans_lock);
 108         if (fs_info->running_transaction) {
 109                 /*
 110                  * someone started a transaction after we unlocked.  Make sure
 111                  * to redo the trans_no_join checks above
 112                  */
 113                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 114                 cur_trans = fs_info->running_transaction;
 115                 goto loop;
 116         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 117                 spin_unlock(&fs_info->trans_lock);
 118                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 119                 return -EROFS;
 120         }
 121
 122         atomic_set(&cur_trans->num_writers, 1);
 123         cur_trans->num_joined = 0;
 124         init_waitqueue_head(&cur_trans->writer_wait);
 125         init_waitqueue_head(&cur_trans->commit_wait);
 126         cur_trans->in_commit = 0;
 127         cur_trans->blocked = 0;
 128         /*
 129          * One for this trans handle, one so it will live on until we
 130          * commit the transaction.
 131          */
 132         atomic_set(&cur_trans->use_count, 2);
 133         cur_trans->commit_done = 0;
 134         cur_trans->start_time = get_seconds();
 135
 136         cur_trans->delayed_refs.root = RB_ROOT;
 137         cur_trans->delayed_refs.num_entries = 0;
 138         cur_trans->delayed_refs.num_heads_ready = 0;
 139         cur_trans->delayed_refs.num_heads = 0;
 140         cur_trans->delayed_refs.flushing = 0;
 141         cur_trans->delayed_refs.run_delayed_start = 0;
 142
 143         /*
 144          * although the tree mod log is per file system and not per transaction,
 145          * the log must never go across transaction boundaries.
 146          */
 147         smp_mb();
 148         if (!list_empty(&fs_info->tree_mod_seq_list))
 149                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 150                         "creating a fresh transaction\n");
 151         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 152                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 153                         "creating a fresh transaction\n");
 154         atomic_set(&fs_info->tree_mod_seq, 0);
 155
 156         spin_lock_init(&cur_trans->commit_lock);
 157         spin_lock_init(&cur_trans->delayed_refs.lock);
 158
 159         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 160         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 161         extent_io_tree_init(&cur_trans->dirty_pages,
 162                              fs_info->btree_inode->i_mapping);
 163         fs_info->generation++;
 164         cur_trans->transid = fs_info->generation;
 165         fs_info->running_transaction = cur_trans;
 166         cur_trans->aborted = 0;
 167         spin_unlock(&fs_info->trans_lock);
 168
 169         return 0;
 170 }
 171
 172 /*
 173  * this does all the record keeping required to make sure that a reference
 174  * counted root is properly recorded in a given transaction.  This is required
 175  * to make sure the old root from before we joined the transaction is deleted
 176  * when the transaction commits
 177  */
 178 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 179                                struct btrfs_root *root)
 180 {
 181         if (root->ref_cows && root->last_trans < trans->transid) {
 182                 WARN_ON(root == root->fs_info->extent_root);
 183                 WARN_ON(root->commit_root != root->node);
 184
 185                 /*
 186                  * see below for in_trans_setup usage rules
 187                  * we have the reloc mutex held now, so there
 188                  * is only one writer in this function
 189                  */
 190                 root->in_trans_setup = 1;
 191
 192                 /* make sure readers find in_trans_setup before
 193                  * they find our root->last_trans update
 194                  */
 195                 smp_wmb();
 196
 197                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 198                 if (root->last_trans == trans->transid) {
 199                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 200                         return 0;
 201                 }
 202                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 203                            (unsigned long)root->root_key.objectid,
 204                            BTRFS_ROOT_TRANS_TAG);
 205                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 206                 root->last_trans = trans->transid;
 207
 208                 /* this is pretty tricky.  We don't want to
 209                  * take the relocation lock in btrfs_record_root_in_trans
 210                  * unless we're really doing the first setup for this root in
 211                  * this transaction.
 212                  *
 213                  * Normally we'd use root->last_trans as a flag to decide
 214                  * if we want to take the expensive mutex.
 215                  *
 216                  * But, we have to set root->last_trans before we
 217                  * init the relocation root, otherwise, we trip over warnings
 218                  * in ctree.c.  The solution used here is to flag ourselves
 219                  * with root->in_trans_setup.  When this is 1, we're still
 220                  * fixing up the reloc trees and everyone must wait.
 221                  *
 222                  * When this is zero, they can trust root->last_trans and fly
 223                  * through btrfs_record_root_in_trans without having to take the
 224                  * lock.  smp_wmb() makes sure that all the writes above are
 225                  * done before we pop in the zero below
 226                  */
 227                 btrfs_init_reloc_root(trans, root);
 228                 smp_wmb();
 229                 root->in_trans_setup = 0;
 230         }
 231         return 0;
 232 }
 233
 234
 235 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 236                                struct btrfs_root *root)
 237 {
 238         if (!root->ref_cows)
 239                 return 0;
 240
 241         /*
 242          * see record_root_in_trans for comments about in_trans_setup usage
 243          * and barriers
 244          */
 245         smp_rmb();
 246         if (root->last_trans == trans->transid &&
 247             !root->in_trans_setup)
 248                 return 0;
 249
 250         mutex_lock(&root->fs_info->reloc_mutex);
 251         record_root_in_trans(trans, root);
 252         mutex_unlock(&root->fs_info->reloc_mutex);
 253
 254         return 0;
 255 }
 256
 257 /* wait for commit against the current transaction to become unblocked
 258  * when this is done, it is safe to start a new transaction, but the current
 259  * transaction might not be fully on disk.
 260  */
 261 static void wait_current_trans(struct btrfs_root *root)
 262 {
 263         struct btrfs_transaction *cur_trans;
 264
 265         spin_lock(&root->fs_info->trans_lock);
 266         cur_trans = root->fs_info->running_transaction;
 267         if (cur_trans && cur_trans->blocked) {
 268                 atomic_inc(&cur_trans->use_count);
 269                 spin_unlock(&root->fs_info->trans_lock);
 270
 271                 wait_event(root->fs_info->transaction_wait,
 272                            !cur_trans->blocked);
 273                 put_transaction(cur_trans);
 274         } else {
 275                 spin_unlock(&root->fs_info->trans_lock);
 276         }
 277 }
 278
 279 static int may_wait_transaction(struct btrfs_root *root, int type)
 280 {
 281         if (root->fs_info->log_root_recovering)
 282                 return 0;
 283
 284         if (type == TRANS_USERSPACE)
 285                 return 1;
 286
 287         if (type == TRANS_START &&
 288             !atomic_read(&root->fs_info->open_ioctl_trans))
 289                 return 1;
 290
 291         return 0;
 292 }
 293
 294 static struct btrfs_trans_handle *
 295 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 296                   enum btrfs_reserve_flush_enum flush)
 297 {
 298         struct btrfs_trans_handle *h;
 299         struct btrfs_transaction *cur_trans;
 300         u64 num_bytes = 0;
 301         int ret;
 302         u64 qgroup_reserved = 0;
 303
 304         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 305                 return ERR_PTR(-EROFS);
 306
 307         if (current->journal_info) {
 308                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 309                 h = current->journal_info;
 310                 h->use_count++;
 311                 WARN_ON(h->use_count > 2);
 312                 h->orig_rsv = h->block_rsv;
 313                 h->block_rsv = NULL;
 314                 goto got_it;
 315         }
 316
 317         /*
 318          * Do the reservation before we join the transaction so we can do all
 319          * the appropriate flushing if need be.
 320          */
 321         if (num_items > 0 && root != root->fs_info->chunk_root) {
 322                 if (root->fs_info->quota_enabled &&
 323                     is_fstree(root->root_key.objectid)) {
 324                         qgroup_reserved = num_items * root->leafsize;
 325                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 326                         if (ret)
 327                                 return ERR_PTR(ret);
 328                 }
 329
 330                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 331                 ret = btrfs_block_rsv_add(root,
 332                                           &root->fs_info->trans_block_rsv,
 333                                           num_bytes, flush);
 334                 if (ret)
 335                         goto reserve_fail;
 336         }
 337 again:
 338         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 339         if (!h) {
 340                 ret = -ENOMEM;
 341                 goto alloc_fail;
 342         }
 343
 344         /*
 345          * If we are JOIN_NOLOCK we're already committing a transaction and
 346          * waiting on this guy, so we don't need to do the sb_start_intwrite
 347          * because we're already holding a ref.  We need this because we could
 348          * have raced in and did an fsync() on a file which can kick a commit
 349          * and then we deadlock with somebody doing a freeze.
 350          *
 351          * If we are ATTACH, it means we just want to catch the current
 352          * transaction and commit it, so we needn't do sb_start_intwrite().
 353          */
 354         if (type < TRANS_JOIN_NOLOCK)
 355                 sb_start_intwrite(root->fs_info->sb);
 356
 357         if (may_wait_transaction(root, type))
 358                 wait_current_trans(root);
 359
 360         do {
 361                 ret = join_transaction(root, type);
 362                 if (ret == -EBUSY)
 363                         wait_current_trans(root);
 364         } while (ret == -EBUSY);
 365
 366         if (ret < 0) {
 367                 /* We must get the transaction if we are JOIN_NOLOCK. */
 368                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 369                 goto join_fail;
 370         }
 371
 372         cur_trans = root->fs_info->running_transaction;
 373
 374         h->transid = cur_trans->transid;
 375         h->transaction = cur_trans;
 376         h->blocks_used = 0;
 377         h->bytes_reserved = 0;
 378         h->root = root;
 379         h->delayed_ref_updates = 0;
 380         h->use_count = 1;
 381         h->adding_csums = 0;
 382         h->block_rsv = NULL;
 383         h->orig_rsv = NULL;
 384         h->aborted = 0;
 385         h->qgroup_reserved = qgroup_reserved;
 386         h->delayed_ref_elem.seq = 0;
 387         h->type = type;
 388         h->allocating_chunk = false;
 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             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 649                 err = -EIO;
 650         assert_qgroups_uptodate(trans);
 651
 652         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 653         return err;
 654 }
 655
 656 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 657                           struct btrfs_root *root)
 658 {
 659         int ret;
 660
 661         ret = __btrfs_end_transaction(trans, root, 0);
 662         if (ret)
 663                 return ret;
 664         return 0;
 665 }
 666
 667 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 668                                    struct btrfs_root *root)
 669 {
 670         int ret;
 671
 672         ret = __btrfs_end_transaction(trans, root, 1);
 673         if (ret)
 674                 return ret;
 675         return 0;
 676 }
 677
 678 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 679                                 struct btrfs_root *root)
 680 {
 681         return __btrfs_end_transaction(trans, root, 1);
 682 }
 683
 684 /*
 685  * when btree blocks are allocated, they have some corresponding bits set for
 686  * them in one of two extent_io trees.  This is used to make sure all of
 687  * those extents are sent to disk but does not wait on them
 688  */
 689 int btrfs_write_marked_extents(struct btrfs_root *root,
 690                                struct extent_io_tree *dirty_pages, int mark)
 691 {
 692         int err = 0;
 693         int werr = 0;
 694         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 695         struct extent_state *cached_state = NULL;
 696         u64 start = 0;
 697         u64 end;
 698
 699         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 700                                       mark, &cached_state)) {
 701                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 702                                    mark, &cached_state, GFP_NOFS);
 703                 cached_state = NULL;
 704                 err = filemap_fdatawrite_range(mapping, start, end);
 705                 if (err)
 706                         werr = err;
 707                 cond_resched();
 708                 start = end + 1;
 709         }
 710         if (err)
 711                 werr = err;
 712         return werr;
 713 }
 714
 715 /*
 716  * when btree blocks are allocated, they have some corresponding bits set for
 717  * them in one of two extent_io trees.  This is used to make sure all of
 718  * those extents are on disk for transaction or log commit.  We wait
 719  * on all the pages and clear them from the dirty pages state tree
 720  */
 721 int btrfs_wait_marked_extents(struct btrfs_root *root,
 722                               struct extent_io_tree *dirty_pages, int mark)
 723 {
 724         int err = 0;
 725         int werr = 0;
 726         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 727         struct extent_state *cached_state = NULL;
 728         u64 start = 0;
 729         u64 end;
 730
 731         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 732                                       EXTENT_NEED_WAIT, &cached_state)) {
 733                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 734                                  0, 0, &cached_state, GFP_NOFS);
 735                 err = filemap_fdatawait_range(mapping, start, end);
 736                 if (err)
 737                         werr = err;
 738                 cond_resched();
 739                 start = end + 1;
 740         }
 741         if (err)
 742                 werr = err;
 743         return werr;
 744 }
 745
 746 /*
 747  * when btree blocks are allocated, they have some corresponding bits set for
 748  * them in one of two extent_io trees.  This is used to make sure all of
 749  * those extents are on disk for transaction or log commit
 750  */
 751 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 752                                 struct extent_io_tree *dirty_pages, int mark)
 753 {
 754         int ret;
 755         int ret2;
 756
 757         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 758         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 759
 760         if (ret)
 761                 return ret;
 762         if (ret2)
 763                 return ret2;
 764         return 0;
 765 }
 766
 767 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 768                                      struct btrfs_root *root)
 769 {
 770         if (!trans || !trans->transaction) {
 771                 struct inode *btree_inode;
 772                 btree_inode = root->fs_info->btree_inode;
 773                 return filemap_write_and_wait(btree_inode->i_mapping);
 774         }
 775         return btrfs_write_and_wait_marked_extents(root,
 776                                            &trans->transaction->dirty_pages,
 777                                            EXTENT_DIRTY);
 778 }
 779
 780 /*
 781  * this is used to update the root pointer in the tree of tree roots.
 782  *
 783  * But, in the case of the extent allocation tree, updating the root
 784  * pointer may allocate blocks which may change the root of the extent
 785  * allocation tree.
 786  *
 787  * So, this loops and repeats and makes sure the cowonly root didn't
 788  * change while the root pointer was being updated in the metadata.
 789  */
 790 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 791                                struct btrfs_root *root)
 792 {
 793         int ret;
 794         u64 old_root_bytenr;
 795         u64 old_root_used;
 796         struct btrfs_root *tree_root = root->fs_info->tree_root;
 797
 798         old_root_used = btrfs_root_used(&root->root_item);
 799         btrfs_write_dirty_block_groups(trans, root);
 800
 801         while (1) {
 802                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 803                 if (old_root_bytenr == root->node->start &&
 804                     old_root_used == btrfs_root_used(&root->root_item))
 805                         break;
 806
 807                 btrfs_set_root_node(&root->root_item, root->node);
 808                 ret = btrfs_update_root(trans, tree_root,
 809                                         &root->root_key,
 810                                         &root->root_item);
 811                 if (ret)
 812                         return ret;
 813
 814                 old_root_used = btrfs_root_used(&root->root_item);
 815                 ret = btrfs_write_dirty_block_groups(trans, root);
 816                 if (ret)
 817                         return ret;
 818         }
 819
 820         if (root != root->fs_info->extent_root)
 821                 switch_commit_root(root);
 822
 823         return 0;
 824 }
 825
 826 /*
 827  * update all the cowonly tree roots on disk
 828  *
 829  * The error handling in this function may not be obvious. Any of the
 830  * failures will cause the file system to go offline. We still need
 831  * to clean up the delayed refs.
 832  */
 833 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 834                                          struct btrfs_root *root)
 835 {
 836         struct btrfs_fs_info *fs_info = root->fs_info;
 837         struct list_head *next;
 838         struct extent_buffer *eb;
 839         int ret;
 840
 841         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 842         if (ret)
 843                 return ret;
 844
 845         eb = btrfs_lock_root_node(fs_info->tree_root);
 846         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 847                               0, &eb);
 848         btrfs_tree_unlock(eb);
 849         free_extent_buffer(eb);
 850
 851         if (ret)
 852                 return ret;
 853
 854         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 855         if (ret)
 856                 return ret;
 857
 858         ret = btrfs_run_dev_stats(trans, root->fs_info);
 859         WARN_ON(ret);
 860         ret = btrfs_run_dev_replace(trans, root->fs_info);
 861         WARN_ON(ret);
 862
 863         ret = btrfs_run_qgroups(trans, root->fs_info);
 864         BUG_ON(ret);
 865
 866         /* run_qgroups might have added some more refs */
 867         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 868         BUG_ON(ret);
 869
 870         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 871                 next = fs_info->dirty_cowonly_roots.next;
 872                 list_del_init(next);
 873                 root = list_entry(next, struct btrfs_root, dirty_list);
 874
 875                 ret = update_cowonly_root(trans, root);
 876                 if (ret)
 877                         return ret;
 878         }
 879
 880         down_write(&fs_info->extent_commit_sem);
 881         switch_commit_root(fs_info->extent_root);
 882         up_write(&fs_info->extent_commit_sem);
 883
 884         btrfs_after_dev_replace_commit(fs_info);
 885
 886         return 0;
 887 }
 888
 889 /*
 890  * dead roots are old snapshots that need to be deleted.  This allocates
 891  * a dirty root struct and adds it into the list of dead roots that need to
 892  * be deleted
 893  */
 894 int btrfs_add_dead_root(struct btrfs_root *root)
 895 {
 896         spin_lock(&root->fs_info->trans_lock);
 897         list_add(&root->root_list, &root->fs_info->dead_roots);
 898         spin_unlock(&root->fs_info->trans_lock);
 899         return 0;
 900 }
 901
 902 /*
 903  * update all the cowonly tree roots on disk
 904  */
 905 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 906                                     struct btrfs_root *root)
 907 {
 908         struct btrfs_root *gang[8];
 909         struct btrfs_fs_info *fs_info = root->fs_info;
 910         int i;
 911         int ret;
 912         int err = 0;
 913
 914         spin_lock(&fs_info->fs_roots_radix_lock);
 915         while (1) {
 916                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 917                                                  (void **)gang, 0,
 918                                                  ARRAY_SIZE(gang),
 919                                                  BTRFS_ROOT_TRANS_TAG);
 920                 if (ret == 0)
 921                         break;
 922                 for (i = 0; i < ret; i++) {
 923                         root = gang[i];
 924                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 925                                         (unsigned long)root->root_key.objectid,
 926                                         BTRFS_ROOT_TRANS_TAG);
 927                         spin_unlock(&fs_info->fs_roots_radix_lock);
 928
 929                         btrfs_free_log(trans, root);
 930                         btrfs_update_reloc_root(trans, root);
 931                         btrfs_orphan_commit_root(trans, root);
 932
 933                         btrfs_save_ino_cache(root, trans);
 934
 935                         /* see comments in should_cow_block() */
 936                         root->force_cow = 0;
 937                         smp_wmb();
 938
 939                         if (root->commit_root != root->node) {
 940                                 mutex_lock(&root->fs_commit_mutex);
 941                                 switch_commit_root(root);
 942                                 btrfs_unpin_free_ino(root);
 943                                 mutex_unlock(&root->fs_commit_mutex);
 944
 945                                 btrfs_set_root_node(&root->root_item,
 946                                                     root->node);
 947                         }
 948
 949                         err = btrfs_update_root(trans, fs_info->tree_root,
 950                                                 &root->root_key,
 951                                                 &root->root_item);
 952                         spin_lock(&fs_info->fs_roots_radix_lock);
 953                         if (err)
 954                                 break;
 955                 }
 956         }
 957         spin_unlock(&fs_info->fs_roots_radix_lock);
 958         return err;
 959 }
 960
 961 /*
 962  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 963  * otherwise every leaf in the btree is read and defragged.
 964  */
 965 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 966 {
 967         struct btrfs_fs_info *info = root->fs_info;
 968         struct btrfs_trans_handle *trans;
 969         int ret;
 970
 971         if (xchg(&root->defrag_running, 1))
 972                 return 0;
 973
 974         while (1) {
 975                 trans = btrfs_start_transaction(root, 0);
 976                 if (IS_ERR(trans))
 977                         return PTR_ERR(trans);
 978
 979                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 980
 981                 btrfs_end_transaction(trans, root);
 982                 btrfs_btree_balance_dirty(info->tree_root);
 983                 cond_resched();
 984
 985                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 986                         break;
 987         }
 988         root->defrag_running = 0;
 989         return ret;
 990 }
 991
 992 /*
 993  * new snapshots need to be created at a very specific time in the
 994  * transaction commit.  This does the actual creation
 995  */
 996 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 997                                    struct btrfs_fs_info *fs_info,
 998                                    struct btrfs_pending_snapshot *pending)
 999 {
1000         struct btrfs_key key;
1001         struct btrfs_root_item *new_root_item;
1002         struct btrfs_root *tree_root = fs_info->tree_root;
1003         struct btrfs_root *root = pending->root;
1004         struct btrfs_root *parent_root;
1005         struct btrfs_block_rsv *rsv;
1006         struct inode *parent_inode;
1007         struct btrfs_path *path;
1008         struct btrfs_dir_item *dir_item;
1009         struct dentry *parent;
1010         struct dentry *dentry;
1011         struct extent_buffer *tmp;
1012         struct extent_buffer *old;
1013         struct timespec cur_time = CURRENT_TIME;
1014         int ret;
1015         u64 to_reserve = 0;
1016         u64 index = 0;
1017         u64 objectid;
1018         u64 root_flags;
1019         uuid_le new_uuid;
1020
1021         path = btrfs_alloc_path();
1022         if (!path) {
1023                 ret = pending->error = -ENOMEM;
1024                 goto path_alloc_fail;
1025         }
1026
1027         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1028         if (!new_root_item) {
1029                 ret = pending->error = -ENOMEM;
1030                 goto root_item_alloc_fail;
1031         }
1032
1033         ret = btrfs_find_free_objectid(tree_root, &objectid);
1034         if (ret) {
1035                 pending->error = ret;
1036                 goto no_free_objectid;
1037         }
1038
1039         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1040
1041         if (to_reserve > 0) {
1042                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1043                                           to_reserve,
1044                                           BTRFS_RESERVE_NO_FLUSH);
1045                 if (ret) {
1046                         pending->error = ret;
1047                         goto no_free_objectid;
1048                 }
1049         }
1050
1051         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1052                                    objectid, pending->inherit);
1053         if (ret) {
1054                 pending->error = ret;
1055                 goto no_free_objectid;
1056         }
1057
1058         key.objectid = objectid;
1059         key.offset = (u64)-1;
1060         key.type = BTRFS_ROOT_ITEM_KEY;
1061
1062         rsv = trans->block_rsv;
1063         trans->block_rsv = &pending->block_rsv;
1064
1065         dentry = pending->dentry;
1066         parent = dget_parent(dentry);
1067         parent_inode = parent->d_inode;
1068         parent_root = BTRFS_I(parent_inode)->root;
1069         record_root_in_trans(trans, parent_root);
1070
1071         /*
1072          * insert the directory item
1073          */
1074         ret = btrfs_set_inode_index(parent_inode, &index);
1075         BUG_ON(ret); /* -ENOMEM */
1076
1077         /* check if there is a file/dir which has the same name. */
1078         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1079                                          btrfs_ino(parent_inode),
1080                                          dentry->d_name.name,
1081                                          dentry->d_name.len, 0);
1082         if (dir_item != NULL && !IS_ERR(dir_item)) {
1083                 pending->error = -EEXIST;
1084                 goto fail;
1085         } else if (IS_ERR(dir_item)) {
1086                 ret = PTR_ERR(dir_item);
1087                 btrfs_abort_transaction(trans, root, ret);
1088                 goto fail;
1089         }
1090         btrfs_release_path(path);
1091
1092         /*
1093          * pull in the delayed directory update
1094          * and the delayed inode item
1095          * otherwise we corrupt the FS during
1096          * snapshot
1097          */
1098         ret = btrfs_run_delayed_items(trans, root);
1099         if (ret) {      /* Transaction aborted */
1100                 btrfs_abort_transaction(trans, root, ret);
1101                 goto fail;
1102         }
1103
1104         record_root_in_trans(trans, root);
1105         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1106         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1107         btrfs_check_and_init_root_item(new_root_item);
1108
1109         root_flags = btrfs_root_flags(new_root_item);
1110         if (pending->readonly)
1111                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1112         else
1113                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1114         btrfs_set_root_flags(new_root_item, root_flags);
1115
1116         btrfs_set_root_generation_v2(new_root_item,
1117                         trans->transid);
1118         uuid_le_gen(&new_uuid);
1119         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1120         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1121                         BTRFS_UUID_SIZE);
1122         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1123         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1124         btrfs_set_root_otransid(new_root_item, trans->transid);
1125         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1126         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1127         btrfs_set_root_stransid(new_root_item, 0);
1128         btrfs_set_root_rtransid(new_root_item, 0);
1129
1130         old = btrfs_lock_root_node(root);
1131         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1132         if (ret) {
1133                 btrfs_tree_unlock(old);
1134                 free_extent_buffer(old);
1135                 btrfs_abort_transaction(trans, root, ret);
1136                 goto fail;
1137         }
1138
1139         btrfs_set_lock_blocking(old);
1140
1141         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1142         /* clean up in any case */
1143         btrfs_tree_unlock(old);
1144         free_extent_buffer(old);
1145         if (ret) {
1146                 btrfs_abort_transaction(trans, root, ret);
1147                 goto fail;
1148         }
1149
1150         /* see comments in should_cow_block() */
1151         root->force_cow = 1;
1152         smp_wmb();
1153
1154         btrfs_set_root_node(new_root_item, tmp);
1155         /* record when the snapshot was created in key.offset */
1156         key.offset = trans->transid;
1157         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1158         btrfs_tree_unlock(tmp);
1159         free_extent_buffer(tmp);
1160         if (ret) {
1161                 btrfs_abort_transaction(trans, root, ret);
1162                 goto fail;
1163         }
1164
1165         /*
1166          * insert root back/forward references
1167          */
1168         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1169                                  parent_root->root_key.objectid,
1170                                  btrfs_ino(parent_inode), index,
1171                                  dentry->d_name.name, dentry->d_name.len);
1172         if (ret) {
1173                 btrfs_abort_transaction(trans, root, ret);
1174                 goto fail;
1175         }
1176
1177         key.offset = (u64)-1;
1178         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1179         if (IS_ERR(pending->snap)) {
1180                 ret = PTR_ERR(pending->snap);
1181                 btrfs_abort_transaction(trans, root, ret);
1182                 goto fail;
1183         }
1184
1185         ret = btrfs_reloc_post_snapshot(trans, pending);
1186         if (ret) {
1187                 btrfs_abort_transaction(trans, root, ret);
1188                 goto fail;
1189         }
1190
1191         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1192         if (ret) {
1193                 btrfs_abort_transaction(trans, root, ret);
1194                 goto fail;
1195         }
1196
1197         ret = btrfs_insert_dir_item(trans, parent_root,
1198                                     dentry->d_name.name, dentry->d_name.len,
1199                                     parent_inode, &key,
1200                                     BTRFS_FT_DIR, index);
1201         /* We have check then name at the beginning, so it is impossible. */
1202         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1203         if (ret) {
1204                 btrfs_abort_transaction(trans, root, ret);
1205                 goto fail;
1206         }
1207
1208         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1209                                          dentry->d_name.len * 2);
1210         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1211         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1212         if (ret)
1213                 btrfs_abort_transaction(trans, root, ret);
1214 fail:
1215         dput(parent);
1216         trans->block_rsv = rsv;
1217 no_free_objectid:
1218         kfree(new_root_item);
1219 root_item_alloc_fail:
1220         btrfs_free_path(path);
1221 path_alloc_fail:
1222         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1223         return ret;
1224 }
1225
1226 /*
1227  * create all the snapshots we've scheduled for creation
1228  */
1229 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1230                                              struct btrfs_fs_info *fs_info)
1231 {
1232         struct btrfs_pending_snapshot *pending;
1233         struct list_head *head = &trans->transaction->pending_snapshots;
1234
1235         list_for_each_entry(pending, head, list)
1236                 create_pending_snapshot(trans, fs_info, pending);
1237         return 0;
1238 }
1239
1240 static void update_super_roots(struct btrfs_root *root)
1241 {
1242         struct btrfs_root_item *root_item;
1243         struct btrfs_super_block *super;
1244
1245         super = root->fs_info->super_copy;
1246
1247         root_item = &root->fs_info->chunk_root->root_item;
1248         super->chunk_root = root_item->bytenr;
1249         super->chunk_root_generation = root_item->generation;
1250         super->chunk_root_level = root_item->level;
1251
1252         root_item = &root->fs_info->tree_root->root_item;
1253         super->root = root_item->bytenr;
1254         super->generation = root_item->generation;
1255         super->root_level = root_item->level;
1256         if (btrfs_test_opt(root, SPACE_CACHE))
1257                 super->cache_generation = root_item->generation;
1258 }
1259
1260 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1261 {
1262         int ret = 0;
1263         spin_lock(&info->trans_lock);
1264         if (info->running_transaction)
1265                 ret = info->running_transaction->in_commit;
1266         spin_unlock(&info->trans_lock);
1267         return ret;
1268 }
1269
1270 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1271 {
1272         int ret = 0;
1273         spin_lock(&info->trans_lock);
1274         if (info->running_transaction)
1275                 ret = info->running_transaction->blocked;
1276         spin_unlock(&info->trans_lock);
1277         return ret;
1278 }
1279
1280 /*
1281  * wait for the current transaction commit to start and block subsequent
1282  * transaction joins
1283  */
1284 static void wait_current_trans_commit_start(struct btrfs_root *root,
1285                                             struct btrfs_transaction *trans)
1286 {
1287         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1288 }
1289
1290 /*
1291  * wait for the current transaction to start and then become unblocked.
1292  * caller holds ref.
1293  */
1294 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1295                                          struct btrfs_transaction *trans)
1296 {
1297         wait_event(root->fs_info->transaction_wait,
1298                    trans->commit_done || (trans->in_commit && !trans->blocked));
1299 }
1300
1301 /*
1302  * commit transactions asynchronously. once btrfs_commit_transaction_async
1303  * returns, any subsequent transaction will not be allowed to join.
1304  */
1305 struct btrfs_async_commit {
1306         struct btrfs_trans_handle *newtrans;
1307         struct btrfs_root *root;
1308         struct work_struct work;
1309 };
1310
1311 static void do_async_commit(struct work_struct *work)
1312 {
1313         struct btrfs_async_commit *ac =
1314                 container_of(work, struct btrfs_async_commit, work);
1315
1316         /*
1317          * We've got freeze protection passed with the transaction.
1318          * Tell lockdep about it.
1319          */
1320         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1321                 rwsem_acquire_read(
1322                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1323                      0, 1, _THIS_IP_);
1324
1325         current->journal_info = ac->newtrans;
1326
1327         btrfs_commit_transaction(ac->newtrans, ac->root);
1328         kfree(ac);
1329 }
1330
1331 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1332                                    struct btrfs_root *root,
1333                                    int wait_for_unblock)
1334 {
1335         struct btrfs_async_commit *ac;
1336         struct btrfs_transaction *cur_trans;
1337
1338         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1339         if (!ac)
1340                 return -ENOMEM;
1341
1342         INIT_WORK(&ac->work, do_async_commit);
1343         ac->root = root;
1344         ac->newtrans = btrfs_join_transaction(root);
1345         if (IS_ERR(ac->newtrans)) {
1346                 int err = PTR_ERR(ac->newtrans);
1347                 kfree(ac);
1348                 return err;
1349         }
1350
1351         /* take transaction reference */
1352         cur_trans = trans->transaction;
1353         atomic_inc(&cur_trans->use_count);
1354
1355         btrfs_end_transaction(trans, root);
1356
1357         /*
1358          * Tell lockdep we've released the freeze rwsem, since the
1359          * async commit thread will be the one to unlock it.
1360          */
1361         if (trans->type < TRANS_JOIN_NOLOCK)
1362                 rwsem_release(
1363                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1364                         1, _THIS_IP_);
1365
1366         schedule_work(&ac->work);
1367
1368         /* wait for transaction to start and unblock */
1369         if (wait_for_unblock)
1370                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1371         else
1372                 wait_current_trans_commit_start(root, cur_trans);
1373
1374         if (current->journal_info == trans)
1375                 current->journal_info = NULL;
1376
1377         put_transaction(cur_trans);
1378         return 0;
1379 }
1380
1381
1382 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1383                                 struct btrfs_root *root, int err)
1384 {
1385         struct btrfs_transaction *cur_trans = trans->transaction;
1386
1387         WARN_ON(trans->use_count > 1);
1388
1389         btrfs_abort_transaction(trans, root, err);
1390
1391         spin_lock(&root->fs_info->trans_lock);
1392         list_del_init(&cur_trans->list);
1393         if (cur_trans == root->fs_info->running_transaction) {
1394                 root->fs_info->running_transaction = NULL;
1395                 root->fs_info->trans_no_join = 0;
1396         }
1397         spin_unlock(&root->fs_info->trans_lock);
1398
1399         btrfs_cleanup_one_transaction(trans->transaction, root);
1400
1401         put_transaction(cur_trans);
1402         put_transaction(cur_trans);
1403
1404         trace_btrfs_transaction_commit(root);
1405
1406         btrfs_scrub_continue(root);
1407
1408         if (current->journal_info == trans)
1409                 current->journal_info = NULL;
1410
1411         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1412 }
1413
1414 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1415                                           struct btrfs_root *root)
1416 {
1417         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1418         int snap_pending = 0;
1419         int ret;
1420
1421         if (!flush_on_commit) {
1422                 spin_lock(&root->fs_info->trans_lock);
1423                 if (!list_empty(&trans->transaction->pending_snapshots))
1424                         snap_pending = 1;
1425                 spin_unlock(&root->fs_info->trans_lock);
1426         }
1427
1428         if (flush_on_commit || snap_pending) {
1429                 ret = btrfs_start_delalloc_inodes(root, 1);
1430                 if (ret)
1431                         return ret;
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         ret = btrfs_run_ordered_operations(root, 1);
1454
1455         return ret;
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 }