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 static unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  38         [TRANS_STATE_RUNNING]           = 0U,
  39         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  40                                            __TRANS_START),
  41         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  42                                            __TRANS_START |
  43                                            __TRANS_ATTACH),
  44         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  45                                            __TRANS_START |
  46                                            __TRANS_ATTACH |
  47                                            __TRANS_JOIN),
  48         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  49                                            __TRANS_START |
  50                                            __TRANS_ATTACH |
  51                                            __TRANS_JOIN |
  52                                            __TRANS_JOIN_NOLOCK),
  53         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  54                                            __TRANS_START |
  55                                            __TRANS_ATTACH |
  56                                            __TRANS_JOIN |
  57                                            __TRANS_JOIN_NOLOCK),
  58 };
  59
  60 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  61 {
  62         WARN_ON(atomic_read(&transaction->use_count) == 0);
  63         if (atomic_dec_and_test(&transaction->use_count)) {
  64                 BUG_ON(!list_empty(&transaction->list));
  65                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  66                 while (!list_empty(&transaction->pending_chunks)) {
  67                         struct extent_map *em;
  68
  69                         em = list_first_entry(&transaction->pending_chunks,
  70                                               struct extent_map, list);
  71                         list_del_init(&em->list);
  72                         free_extent_map(em);
  73                 }
  74                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  75         }
  76 }
  77
  78 static noinline void switch_commit_root(struct btrfs_root *root)
  79 {
  80         free_extent_buffer(root->commit_root);
  81         root->commit_root = btrfs_root_node(root);
  82 }
  83
  84 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
  85                                          unsigned int type)
  86 {
  87         if (type & TRANS_EXTWRITERS)
  88                 atomic_inc(&trans->num_extwriters);
  89 }
  90
  91 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
  92                                          unsigned int type)
  93 {
  94         if (type & TRANS_EXTWRITERS)
  95                 atomic_dec(&trans->num_extwriters);
  96 }
  97
  98 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
  99                                           unsigned int type)
 100 {
 101         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 102 }
 103
 104 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 105 {
 106         return atomic_read(&trans->num_extwriters);
 107 }
 108
 109 /*
 110  * either allocate a new transaction or hop into the existing one
 111  */
 112 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 113 {
 114         struct btrfs_transaction *cur_trans;
 115         struct btrfs_fs_info *fs_info = root->fs_info;
 116
 117         spin_lock(&fs_info->trans_lock);
 118 loop:
 119         /* The file system has been taken offline. No new transactions. */
 120         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 121                 spin_unlock(&fs_info->trans_lock);
 122                 return -EROFS;
 123         }
 124
 125         cur_trans = fs_info->running_transaction;
 126         if (cur_trans) {
 127                 if (cur_trans->aborted) {
 128                         spin_unlock(&fs_info->trans_lock);
 129                         return cur_trans->aborted;
 130                 }
 131                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 132                         spin_unlock(&fs_info->trans_lock);
 133                         return -EBUSY;
 134                 }
 135                 atomic_inc(&cur_trans->use_count);
 136                 atomic_inc(&cur_trans->num_writers);
 137                 extwriter_counter_inc(cur_trans, type);
 138                 spin_unlock(&fs_info->trans_lock);
 139                 return 0;
 140         }
 141         spin_unlock(&fs_info->trans_lock);
 142
 143         /*
 144          * If we are ATTACH, we just want to catch the current transaction,
 145          * and commit it. If there is no transaction, just return ENOENT.
 146          */
 147         if (type == TRANS_ATTACH)
 148                 return -ENOENT;
 149
 150         /*
 151          * JOIN_NOLOCK only happens during the transaction commit, so
 152          * it is impossible that ->running_transaction is NULL
 153          */
 154         BUG_ON(type == TRANS_JOIN_NOLOCK);
 155
 156         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 157         if (!cur_trans)
 158                 return -ENOMEM;
 159
 160         spin_lock(&fs_info->trans_lock);
 161         if (fs_info->running_transaction) {
 162                 /*
 163                  * someone started a transaction after we unlocked.  Make sure
 164                  * to redo the checks above
 165                  */
 166                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 167                 goto loop;
 168         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 169                 spin_unlock(&fs_info->trans_lock);
 170                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 171                 return -EROFS;
 172         }
 173
 174         atomic_set(&cur_trans->num_writers, 1);
 175         extwriter_counter_init(cur_trans, type);
 176         init_waitqueue_head(&cur_trans->writer_wait);
 177         init_waitqueue_head(&cur_trans->commit_wait);
 178         cur_trans->state = TRANS_STATE_RUNNING;
 179         /*
 180          * One for this trans handle, one so it will live on until we
 181          * commit the transaction.
 182          */
 183         atomic_set(&cur_trans->use_count, 2);
 184         cur_trans->start_time = get_seconds();
 185
 186         cur_trans->delayed_refs.href_root = RB_ROOT;
 187         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 188         cur_trans->delayed_refs.num_heads_ready = 0;
 189         cur_trans->delayed_refs.num_heads = 0;
 190         cur_trans->delayed_refs.flushing = 0;
 191         cur_trans->delayed_refs.run_delayed_start = 0;
 192
 193         /*
 194          * although the tree mod log is per file system and not per transaction,
 195          * the log must never go across transaction boundaries.
 196          */
 197         smp_mb();
 198         if (!list_empty(&fs_info->tree_mod_seq_list))
 199                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 200                         "creating a fresh transaction\n");
 201         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 202                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 203                         "creating a fresh transaction\n");
 204         atomic64_set(&fs_info->tree_mod_seq, 0);
 205
 206         spin_lock_init(&cur_trans->delayed_refs.lock);
 207
 208         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 209         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 210         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 211         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 212         extent_io_tree_init(&cur_trans->dirty_pages,
 213                              fs_info->btree_inode->i_mapping);
 214         fs_info->generation++;
 215         cur_trans->transid = fs_info->generation;
 216         fs_info->running_transaction = cur_trans;
 217         cur_trans->aborted = 0;
 218         spin_unlock(&fs_info->trans_lock);
 219
 220         return 0;
 221 }
 222
 223 /*
 224  * this does all the record keeping required to make sure that a reference
 225  * counted root is properly recorded in a given transaction.  This is required
 226  * to make sure the old root from before we joined the transaction is deleted
 227  * when the transaction commits
 228  */
 229 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 230                                struct btrfs_root *root)
 231 {
 232         if (root->ref_cows && root->last_trans < trans->transid) {
 233                 WARN_ON(root == root->fs_info->extent_root);
 234                 WARN_ON(root->commit_root != root->node);
 235
 236                 /*
 237                  * see below for in_trans_setup usage rules
 238                  * we have the reloc mutex held now, so there
 239                  * is only one writer in this function
 240                  */
 241                 root->in_trans_setup = 1;
 242
 243                 /* make sure readers find in_trans_setup before
 244                  * they find our root->last_trans update
 245                  */
 246                 smp_wmb();
 247
 248                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 249                 if (root->last_trans == trans->transid) {
 250                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 251                         return 0;
 252                 }
 253                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 254                            (unsigned long)root->root_key.objectid,
 255                            BTRFS_ROOT_TRANS_TAG);
 256                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 257                 root->last_trans = trans->transid;
 258
 259                 /* this is pretty tricky.  We don't want to
 260                  * take the relocation lock in btrfs_record_root_in_trans
 261                  * unless we're really doing the first setup for this root in
 262                  * this transaction.
 263                  *
 264                  * Normally we'd use root->last_trans as a flag to decide
 265                  * if we want to take the expensive mutex.
 266                  *
 267                  * But, we have to set root->last_trans before we
 268                  * init the relocation root, otherwise, we trip over warnings
 269                  * in ctree.c.  The solution used here is to flag ourselves
 270                  * with root->in_trans_setup.  When this is 1, we're still
 271                  * fixing up the reloc trees and everyone must wait.
 272                  *
 273                  * When this is zero, they can trust root->last_trans and fly
 274                  * through btrfs_record_root_in_trans without having to take the
 275                  * lock.  smp_wmb() makes sure that all the writes above are
 276                  * done before we pop in the zero below
 277                  */
 278                 btrfs_init_reloc_root(trans, root);
 279                 smp_wmb();
 280                 root->in_trans_setup = 0;
 281         }
 282         return 0;
 283 }
 284
 285
 286 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 287                                struct btrfs_root *root)
 288 {
 289         if (!root->ref_cows)
 290                 return 0;
 291
 292         /*
 293          * see record_root_in_trans for comments about in_trans_setup usage
 294          * and barriers
 295          */
 296         smp_rmb();
 297         if (root->last_trans == trans->transid &&
 298             !root->in_trans_setup)
 299                 return 0;
 300
 301         mutex_lock(&root->fs_info->reloc_mutex);
 302         record_root_in_trans(trans, root);
 303         mutex_unlock(&root->fs_info->reloc_mutex);
 304
 305         return 0;
 306 }
 307
 308 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 309 {
 310         return (trans->state >= TRANS_STATE_BLOCKED &&
 311                 trans->state < TRANS_STATE_UNBLOCKED &&
 312                 !trans->aborted);
 313 }
 314
 315 /* wait for commit against the current transaction to become unblocked
 316  * when this is done, it is safe to start a new transaction, but the current
 317  * transaction might not be fully on disk.
 318  */
 319 static void wait_current_trans(struct btrfs_root *root)
 320 {
 321         struct btrfs_transaction *cur_trans;
 322
 323         spin_lock(&root->fs_info->trans_lock);
 324         cur_trans = root->fs_info->running_transaction;
 325         if (cur_trans && is_transaction_blocked(cur_trans)) {
 326                 atomic_inc(&cur_trans->use_count);
 327                 spin_unlock(&root->fs_info->trans_lock);
 328
 329                 wait_event(root->fs_info->transaction_wait,
 330                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 331                            cur_trans->aborted);
 332                 btrfs_put_transaction(cur_trans);
 333         } else {
 334                 spin_unlock(&root->fs_info->trans_lock);
 335         }
 336 }
 337
 338 static int may_wait_transaction(struct btrfs_root *root, int type)
 339 {
 340         if (root->fs_info->log_root_recovering)
 341                 return 0;
 342
 343         if (type == TRANS_USERSPACE)
 344                 return 1;
 345
 346         if (type == TRANS_START &&
 347             !atomic_read(&root->fs_info->open_ioctl_trans))
 348                 return 1;
 349
 350         return 0;
 351 }
 352
 353 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 354 {
 355         if (!root->fs_info->reloc_ctl ||
 356             !root->ref_cows ||
 357             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 358             root->reloc_root)
 359                 return false;
 360
 361         return true;
 362 }
 363
 364 static struct btrfs_trans_handle *
 365 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 366                   enum btrfs_reserve_flush_enum flush)
 367 {
 368         struct btrfs_trans_handle *h;
 369         struct btrfs_transaction *cur_trans;
 370         u64 num_bytes = 0;
 371         u64 qgroup_reserved = 0;
 372         bool reloc_reserved = false;
 373         int ret;
 374
 375         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 376                 return ERR_PTR(-EROFS);
 377
 378         if (current->journal_info) {
 379                 WARN_ON(type & TRANS_EXTWRITERS);
 380                 h = current->journal_info;
 381                 h->use_count++;
 382                 WARN_ON(h->use_count > 2);
 383                 h->orig_rsv = h->block_rsv;
 384                 h->block_rsv = NULL;
 385                 goto got_it;
 386         }
 387
 388         /*
 389          * Do the reservation before we join the transaction so we can do all
 390          * the appropriate flushing if need be.
 391          */
 392         if (num_items > 0 && root != root->fs_info->chunk_root) {
 393                 if (root->fs_info->quota_enabled &&
 394                     is_fstree(root->root_key.objectid)) {
 395                         qgroup_reserved = num_items * root->leafsize;
 396                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 397                         if (ret)
 398                                 return ERR_PTR(ret);
 399                 }
 400
 401                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 402                 /*
 403                  * Do the reservation for the relocation root creation
 404                  */
 405                 if (unlikely(need_reserve_reloc_root(root))) {
 406                         num_bytes += root->nodesize;
 407                         reloc_reserved = true;
 408                 }
 409
 410                 ret = btrfs_block_rsv_add(root,
 411                                           &root->fs_info->trans_block_rsv,
 412                                           num_bytes, flush);
 413                 if (ret)
 414                         goto reserve_fail;
 415         }
 416 again:
 417         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 418         if (!h) {
 419                 ret = -ENOMEM;
 420                 goto alloc_fail;
 421         }
 422
 423         /*
 424          * If we are JOIN_NOLOCK we're already committing a transaction and
 425          * waiting on this guy, so we don't need to do the sb_start_intwrite
 426          * because we're already holding a ref.  We need this because we could
 427          * have raced in and did an fsync() on a file which can kick a commit
 428          * and then we deadlock with somebody doing a freeze.
 429          *
 430          * If we are ATTACH, it means we just want to catch the current
 431          * transaction and commit it, so we needn't do sb_start_intwrite().
 432          */
 433         if (type & __TRANS_FREEZABLE)
 434                 sb_start_intwrite(root->fs_info->sb);
 435
 436         if (may_wait_transaction(root, type))
 437                 wait_current_trans(root);
 438
 439         do {
 440                 ret = join_transaction(root, type);
 441                 if (ret == -EBUSY) {
 442                         wait_current_trans(root);
 443                         if (unlikely(type == TRANS_ATTACH))
 444                                 ret = -ENOENT;
 445                 }
 446         } while (ret == -EBUSY);
 447
 448         if (ret < 0) {
 449                 /* We must get the transaction if we are JOIN_NOLOCK. */
 450                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 451                 goto join_fail;
 452         }
 453
 454         cur_trans = root->fs_info->running_transaction;
 455
 456         h->transid = cur_trans->transid;
 457         h->transaction = cur_trans;
 458         h->blocks_used = 0;
 459         h->bytes_reserved = 0;
 460         h->root = root;
 461         h->delayed_ref_updates = 0;
 462         h->use_count = 1;
 463         h->adding_csums = 0;
 464         h->block_rsv = NULL;
 465         h->orig_rsv = NULL;
 466         h->aborted = 0;
 467         h->qgroup_reserved = 0;
 468         h->delayed_ref_elem.seq = 0;
 469         h->type = type;
 470         h->allocating_chunk = false;
 471         h->reloc_reserved = false;
 472         h->sync = false;
 473         INIT_LIST_HEAD(&h->qgroup_ref_list);
 474         INIT_LIST_HEAD(&h->new_bgs);
 475
 476         smp_mb();
 477         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 478             may_wait_transaction(root, type)) {
 479                 btrfs_commit_transaction(h, root);
 480                 goto again;
 481         }
 482
 483         if (num_bytes) {
 484                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 485                                               h->transid, num_bytes, 1);
 486                 h->block_rsv = &root->fs_info->trans_block_rsv;
 487                 h->bytes_reserved = num_bytes;
 488                 h->reloc_reserved = reloc_reserved;
 489         }
 490         h->qgroup_reserved = qgroup_reserved;
 491
 492 got_it:
 493         btrfs_record_root_in_trans(h, root);
 494
 495         if (!current->journal_info && type != TRANS_USERSPACE)
 496                 current->journal_info = h;
 497         return h;
 498
 499 join_fail:
 500         if (type & __TRANS_FREEZABLE)
 501                 sb_end_intwrite(root->fs_info->sb);
 502         kmem_cache_free(btrfs_trans_handle_cachep, h);
 503 alloc_fail:
 504         if (num_bytes)
 505                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 506                                         num_bytes);
 507 reserve_fail:
 508         if (qgroup_reserved)
 509                 btrfs_qgroup_free(root, qgroup_reserved);
 510         return ERR_PTR(ret);
 511 }
 512
 513 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 514                                                    int num_items)
 515 {
 516         return start_transaction(root, num_items, TRANS_START,
 517                                  BTRFS_RESERVE_FLUSH_ALL);
 518 }
 519
 520 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 521                                         struct btrfs_root *root, int num_items)
 522 {
 523         return start_transaction(root, num_items, TRANS_START,
 524                                  BTRFS_RESERVE_FLUSH_LIMIT);
 525 }
 526
 527 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 528 {
 529         return start_transaction(root, 0, TRANS_JOIN, 0);
 530 }
 531
 532 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 533 {
 534         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 535 }
 536
 537 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 538 {
 539         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 540 }
 541
 542 /*
 543  * btrfs_attach_transaction() - catch the running transaction
 544  *
 545  * It is used when we want to commit the current the transaction, but
 546  * don't want to start a new one.
 547  *
 548  * Note: If this function return -ENOENT, it just means there is no
 549  * running transaction. But it is possible that the inactive transaction
 550  * is still in the memory, not fully on disk. If you hope there is no
 551  * inactive transaction in the fs when -ENOENT is returned, you should
 552  * invoke
 553  *     btrfs_attach_transaction_barrier()
 554  */
 555 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 556 {
 557         return start_transaction(root, 0, TRANS_ATTACH, 0);
 558 }
 559
 560 /*
 561  * btrfs_attach_transaction_barrier() - catch the running transaction
 562  *
 563  * It is similar to the above function, the differentia is this one
 564  * will wait for all the inactive transactions until they fully
 565  * complete.
 566  */
 567 struct btrfs_trans_handle *
 568 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 569 {
 570         struct btrfs_trans_handle *trans;
 571
 572         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 573         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 574                 btrfs_wait_for_commit(root, 0);
 575
 576         return trans;
 577 }
 578
 579 /* wait for a transaction commit to be fully complete */
 580 static noinline void wait_for_commit(struct btrfs_root *root,
 581                                     struct btrfs_transaction *commit)
 582 {
 583         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 584 }
 585
 586 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 587 {
 588         struct btrfs_transaction *cur_trans = NULL, *t;
 589         int ret = 0;
 590
 591         if (transid) {
 592                 if (transid <= root->fs_info->last_trans_committed)
 593                         goto out;
 594
 595                 ret = -EINVAL;
 596                 /* find specified transaction */
 597                 spin_lock(&root->fs_info->trans_lock);
 598                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 599                         if (t->transid == transid) {
 600                                 cur_trans = t;
 601                                 atomic_inc(&cur_trans->use_count);
 602                                 ret = 0;
 603                                 break;
 604                         }
 605                         if (t->transid > transid) {
 606                                 ret = 0;
 607                                 break;
 608                         }
 609                 }
 610                 spin_unlock(&root->fs_info->trans_lock);
 611                 /* The specified transaction doesn't exist */
 612                 if (!cur_trans)
 613                         goto out;
 614         } else {
 615                 /* find newest transaction that is committing | committed */
 616                 spin_lock(&root->fs_info->trans_lock);
 617                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 618                                             list) {
 619                         if (t->state >= TRANS_STATE_COMMIT_START) {
 620                                 if (t->state == TRANS_STATE_COMPLETED)
 621                                         break;
 622                                 cur_trans = t;
 623                                 atomic_inc(&cur_trans->use_count);
 624                                 break;
 625                         }
 626                 }
 627                 spin_unlock(&root->fs_info->trans_lock);
 628                 if (!cur_trans)
 629                         goto out;  /* nothing committing|committed */
 630         }
 631
 632         wait_for_commit(root, cur_trans);
 633         btrfs_put_transaction(cur_trans);
 634 out:
 635         return ret;
 636 }
 637
 638 void btrfs_throttle(struct btrfs_root *root)
 639 {
 640         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 641                 wait_current_trans(root);
 642 }
 643
 644 static int should_end_transaction(struct btrfs_trans_handle *trans,
 645                                   struct btrfs_root *root)
 646 {
 647         if (root->fs_info->global_block_rsv.space_info->full &&
 648             btrfs_should_throttle_delayed_refs(trans, root))
 649                 return 1;
 650
 651         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 652 }
 653
 654 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 655                                  struct btrfs_root *root)
 656 {
 657         struct btrfs_transaction *cur_trans = trans->transaction;
 658         int updates;
 659         int err;
 660
 661         smp_mb();
 662         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 663             cur_trans->delayed_refs.flushing)
 664                 return 1;
 665
 666         updates = trans->delayed_ref_updates;
 667         trans->delayed_ref_updates = 0;
 668         if (updates) {
 669                 err = btrfs_run_delayed_refs(trans, root, updates);
 670                 if (err) /* Error code will also eval true */
 671                         return err;
 672         }
 673
 674         return should_end_transaction(trans, root);
 675 }
 676
 677 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 678                           struct btrfs_root *root, int throttle)
 679 {
 680         struct btrfs_transaction *cur_trans = trans->transaction;
 681         struct btrfs_fs_info *info = root->fs_info;
 682         unsigned long cur = trans->delayed_ref_updates;
 683         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 684         int err = 0;
 685
 686         if (--trans->use_count) {
 687                 trans->block_rsv = trans->orig_rsv;
 688                 return 0;
 689         }
 690
 691         /*
 692          * do the qgroup accounting as early as possible
 693          */
 694         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 695
 696         btrfs_trans_release_metadata(trans, root);
 697         trans->block_rsv = NULL;
 698
 699         if (trans->qgroup_reserved) {
 700                 /*
 701                  * the same root has to be passed here between start_transaction
 702                  * and end_transaction. Subvolume quota depends on this.
 703                  */
 704                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 705                 trans->qgroup_reserved = 0;
 706         }
 707
 708         if (!list_empty(&trans->new_bgs))
 709                 btrfs_create_pending_block_groups(trans, root);
 710
 711         trans->delayed_ref_updates = 0;
 712         if (!trans->sync && btrfs_should_throttle_delayed_refs(trans, root)) {
 713                 cur = max_t(unsigned long, cur, 1);
 714                 trans->delayed_ref_updates = 0;
 715                 btrfs_run_delayed_refs(trans, root, cur);
 716         }
 717
 718         btrfs_trans_release_metadata(trans, root);
 719         trans->block_rsv = NULL;
 720
 721         if (!list_empty(&trans->new_bgs))
 722                 btrfs_create_pending_block_groups(trans, root);
 723
 724         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 725             should_end_transaction(trans, root) &&
 726             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 727                 spin_lock(&info->trans_lock);
 728                 if (cur_trans->state == TRANS_STATE_RUNNING)
 729                         cur_trans->state = TRANS_STATE_BLOCKED;
 730                 spin_unlock(&info->trans_lock);
 731         }
 732
 733         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 734                 if (throttle) {
 735                         /*
 736                          * We may race with somebody else here so end up having
 737                          * to call end_transaction on ourselves again, so inc
 738                          * our use_count.
 739                          */
 740                         trans->use_count++;
 741                         return btrfs_commit_transaction(trans, root);
 742                 } else {
 743                         wake_up_process(info->transaction_kthread);
 744                 }
 745         }
 746
 747         if (trans->type & __TRANS_FREEZABLE)
 748                 sb_end_intwrite(root->fs_info->sb);
 749
 750         WARN_ON(cur_trans != info->running_transaction);
 751         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 752         atomic_dec(&cur_trans->num_writers);
 753         extwriter_counter_dec(cur_trans, trans->type);
 754
 755         smp_mb();
 756         if (waitqueue_active(&cur_trans->writer_wait))
 757                 wake_up(&cur_trans->writer_wait);
 758         btrfs_put_transaction(cur_trans);
 759
 760         if (current->journal_info == trans)
 761                 current->journal_info = NULL;
 762
 763         if (throttle)
 764                 btrfs_run_delayed_iputs(root);
 765
 766         if (trans->aborted ||
 767             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 768                 wake_up_process(info->transaction_kthread);
 769                 err = -EIO;
 770         }
 771         assert_qgroups_uptodate(trans);
 772
 773         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 774         return err;
 775 }
 776
 777 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 778                           struct btrfs_root *root)
 779 {
 780         return __btrfs_end_transaction(trans, root, 0);
 781 }
 782
 783 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 784                                    struct btrfs_root *root)
 785 {
 786         return __btrfs_end_transaction(trans, root, 1);
 787 }
 788
 789 /*
 790  * when btree blocks are allocated, they have some corresponding bits set for
 791  * them in one of two extent_io trees.  This is used to make sure all of
 792  * those extents are sent to disk but does not wait on them
 793  */
 794 int btrfs_write_marked_extents(struct btrfs_root *root,
 795                                struct extent_io_tree *dirty_pages, int mark)
 796 {
 797         int err = 0;
 798         int werr = 0;
 799         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 800         struct extent_state *cached_state = NULL;
 801         u64 start = 0;
 802         u64 end;
 803
 804         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 805                                       mark, &cached_state)) {
 806                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 807                                    mark, &cached_state, GFP_NOFS);
 808                 cached_state = NULL;
 809                 err = filemap_fdatawrite_range(mapping, start, end);
 810                 if (err)
 811                         werr = err;
 812                 cond_resched();
 813                 start = end + 1;
 814         }
 815         if (err)
 816                 werr = err;
 817         return werr;
 818 }
 819
 820 /*
 821  * when btree blocks are allocated, they have some corresponding bits set for
 822  * them in one of two extent_io trees.  This is used to make sure all of
 823  * those extents are on disk for transaction or log commit.  We wait
 824  * on all the pages and clear them from the dirty pages state tree
 825  */
 826 int btrfs_wait_marked_extents(struct btrfs_root *root,
 827                               struct extent_io_tree *dirty_pages, int mark)
 828 {
 829         int err = 0;
 830         int werr = 0;
 831         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 832         struct extent_state *cached_state = NULL;
 833         u64 start = 0;
 834         u64 end;
 835
 836         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 837                                       EXTENT_NEED_WAIT, &cached_state)) {
 838                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 839                                  0, 0, &cached_state, GFP_NOFS);
 840                 err = filemap_fdatawait_range(mapping, start, end);
 841                 if (err)
 842                         werr = err;
 843                 cond_resched();
 844                 start = end + 1;
 845         }
 846         if (err)
 847                 werr = err;
 848         return werr;
 849 }
 850
 851 /*
 852  * when btree blocks are allocated, they have some corresponding bits set for
 853  * them in one of two extent_io trees.  This is used to make sure all of
 854  * those extents are on disk for transaction or log commit
 855  */
 856 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 857                                 struct extent_io_tree *dirty_pages, int mark)
 858 {
 859         int ret;
 860         int ret2;
 861         struct blk_plug plug;
 862
 863         blk_start_plug(&plug);
 864         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 865         blk_finish_plug(&plug);
 866         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 867
 868         if (ret)
 869                 return ret;
 870         if (ret2)
 871                 return ret2;
 872         return 0;
 873 }
 874
 875 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 876                                      struct btrfs_root *root)
 877 {
 878         if (!trans || !trans->transaction) {
 879                 struct inode *btree_inode;
 880                 btree_inode = root->fs_info->btree_inode;
 881                 return filemap_write_and_wait(btree_inode->i_mapping);
 882         }
 883         return btrfs_write_and_wait_marked_extents(root,
 884                                            &trans->transaction->dirty_pages,
 885                                            EXTENT_DIRTY);
 886 }
 887
 888 /*
 889  * this is used to update the root pointer in the tree of tree roots.
 890  *
 891  * But, in the case of the extent allocation tree, updating the root
 892  * pointer may allocate blocks which may change the root of the extent
 893  * allocation tree.
 894  *
 895  * So, this loops and repeats and makes sure the cowonly root didn't
 896  * change while the root pointer was being updated in the metadata.
 897  */
 898 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 899                                struct btrfs_root *root)
 900 {
 901         int ret;
 902         u64 old_root_bytenr;
 903         u64 old_root_used;
 904         struct btrfs_root *tree_root = root->fs_info->tree_root;
 905
 906         old_root_used = btrfs_root_used(&root->root_item);
 907         btrfs_write_dirty_block_groups(trans, root);
 908
 909         while (1) {
 910                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 911                 if (old_root_bytenr == root->node->start &&
 912                     old_root_used == btrfs_root_used(&root->root_item))
 913                         break;
 914
 915                 btrfs_set_root_node(&root->root_item, root->node);
 916                 ret = btrfs_update_root(trans, tree_root,
 917                                         &root->root_key,
 918                                         &root->root_item);
 919                 if (ret)
 920                         return ret;
 921
 922                 old_root_used = btrfs_root_used(&root->root_item);
 923                 ret = btrfs_write_dirty_block_groups(trans, root);
 924                 if (ret)
 925                         return ret;
 926         }
 927
 928         if (root != root->fs_info->extent_root)
 929                 switch_commit_root(root);
 930
 931         return 0;
 932 }
 933
 934 /*
 935  * update all the cowonly tree roots on disk
 936  *
 937  * The error handling in this function may not be obvious. Any of the
 938  * failures will cause the file system to go offline. We still need
 939  * to clean up the delayed refs.
 940  */
 941 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 942                                          struct btrfs_root *root)
 943 {
 944         struct btrfs_fs_info *fs_info = root->fs_info;
 945         struct list_head *next;
 946         struct extent_buffer *eb;
 947         int ret;
 948
 949         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 950         if (ret)
 951                 return ret;
 952
 953         eb = btrfs_lock_root_node(fs_info->tree_root);
 954         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 955                               0, &eb);
 956         btrfs_tree_unlock(eb);
 957         free_extent_buffer(eb);
 958
 959         if (ret)
 960                 return ret;
 961
 962         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 963         if (ret)
 964                 return ret;
 965
 966         ret = btrfs_run_dev_stats(trans, root->fs_info);
 967         if (ret)
 968                 return ret;
 969         ret = btrfs_run_dev_replace(trans, root->fs_info);
 970         if (ret)
 971                 return ret;
 972         ret = btrfs_run_qgroups(trans, root->fs_info);
 973         if (ret)
 974                 return ret;
 975
 976         /* run_qgroups might have added some more refs */
 977         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 978         if (ret)
 979                 return ret;
 980
 981         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 982                 next = fs_info->dirty_cowonly_roots.next;
 983                 list_del_init(next);
 984                 root = list_entry(next, struct btrfs_root, dirty_list);
 985
 986                 ret = update_cowonly_root(trans, root);
 987                 if (ret)
 988                         return ret;
 989         }
 990
 991         down_write(&fs_info->extent_commit_sem);
 992         switch_commit_root(fs_info->extent_root);
 993         up_write(&fs_info->extent_commit_sem);
 994
 995         btrfs_after_dev_replace_commit(fs_info);
 996
 997         return 0;
 998 }
 999
1000 /*
1001  * dead roots are old snapshots that need to be deleted.  This allocates
1002  * a dirty root struct and adds it into the list of dead roots that need to
1003  * be deleted
1004  */
1005 void btrfs_add_dead_root(struct btrfs_root *root)
1006 {
1007         spin_lock(&root->fs_info->trans_lock);
1008         if (list_empty(&root->root_list))
1009                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1010         spin_unlock(&root->fs_info->trans_lock);
1011 }
1012
1013 /*
1014  * update all the cowonly tree roots on disk
1015  */
1016 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1017                                     struct btrfs_root *root)
1018 {
1019         struct btrfs_root *gang[8];
1020         struct btrfs_fs_info *fs_info = root->fs_info;
1021         int i;
1022         int ret;
1023         int err = 0;
1024
1025         spin_lock(&fs_info->fs_roots_radix_lock);
1026         while (1) {
1027                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1028                                                  (void **)gang, 0,
1029                                                  ARRAY_SIZE(gang),
1030                                                  BTRFS_ROOT_TRANS_TAG);
1031                 if (ret == 0)
1032                         break;
1033                 for (i = 0; i < ret; i++) {
1034                         root = gang[i];
1035                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1036                                         (unsigned long)root->root_key.objectid,
1037                                         BTRFS_ROOT_TRANS_TAG);
1038                         spin_unlock(&fs_info->fs_roots_radix_lock);
1039
1040                         btrfs_free_log(trans, root);
1041                         btrfs_update_reloc_root(trans, root);
1042                         btrfs_orphan_commit_root(trans, root);
1043
1044                         btrfs_save_ino_cache(root, trans);
1045
1046                         /* see comments in should_cow_block() */
1047                         root->force_cow = 0;
1048                         smp_wmb();
1049
1050                         if (root->commit_root != root->node) {
1051                                 mutex_lock(&root->fs_commit_mutex);
1052                                 switch_commit_root(root);
1053                                 btrfs_unpin_free_ino(root);
1054                                 mutex_unlock(&root->fs_commit_mutex);
1055
1056                                 btrfs_set_root_node(&root->root_item,
1057                                                     root->node);
1058                         }
1059
1060                         err = btrfs_update_root(trans, fs_info->tree_root,
1061                                                 &root->root_key,
1062                                                 &root->root_item);
1063                         spin_lock(&fs_info->fs_roots_radix_lock);
1064                         if (err)
1065                                 break;
1066                 }
1067         }
1068         spin_unlock(&fs_info->fs_roots_radix_lock);
1069         return err;
1070 }
1071
1072 /*
1073  * defrag a given btree.
1074  * Every leaf in the btree is read and defragged.
1075  */
1076 int btrfs_defrag_root(struct btrfs_root *root)
1077 {
1078         struct btrfs_fs_info *info = root->fs_info;
1079         struct btrfs_trans_handle *trans;
1080         int ret;
1081
1082         if (xchg(&root->defrag_running, 1))
1083                 return 0;
1084
1085         while (1) {
1086                 trans = btrfs_start_transaction(root, 0);
1087                 if (IS_ERR(trans))
1088                         return PTR_ERR(trans);
1089
1090                 ret = btrfs_defrag_leaves(trans, root);
1091
1092                 btrfs_end_transaction(trans, root);
1093                 btrfs_btree_balance_dirty(info->tree_root);
1094                 cond_resched();
1095
1096                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1097                         break;
1098
1099                 if (btrfs_defrag_cancelled(root->fs_info)) {
1100                         pr_debug("BTRFS: defrag_root cancelled\n");
1101                         ret = -EAGAIN;
1102                         break;
1103                 }
1104         }
1105         root->defrag_running = 0;
1106         return ret;
1107 }
1108
1109 /*
1110  * new snapshots need to be created at a very specific time in the
1111  * transaction commit.  This does the actual creation.
1112  *
1113  * Note:
1114  * If the error which may affect the commitment of the current transaction
1115  * happens, we should return the error number. If the error which just affect
1116  * the creation of the pending snapshots, just return 0.
1117  */
1118 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1119                                    struct btrfs_fs_info *fs_info,
1120                                    struct btrfs_pending_snapshot *pending)
1121 {
1122         struct btrfs_key key;
1123         struct btrfs_root_item *new_root_item;
1124         struct btrfs_root *tree_root = fs_info->tree_root;
1125         struct btrfs_root *root = pending->root;
1126         struct btrfs_root *parent_root;
1127         struct btrfs_block_rsv *rsv;
1128         struct inode *parent_inode;
1129         struct btrfs_path *path;
1130         struct btrfs_dir_item *dir_item;
1131         struct dentry *dentry;
1132         struct extent_buffer *tmp;
1133         struct extent_buffer *old;
1134         struct timespec cur_time = CURRENT_TIME;
1135         int ret = 0;
1136         u64 to_reserve = 0;
1137         u64 index = 0;
1138         u64 objectid;
1139         u64 root_flags;
1140         uuid_le new_uuid;
1141
1142         path = btrfs_alloc_path();
1143         if (!path) {
1144                 pending->error = -ENOMEM;
1145                 return 0;
1146         }
1147
1148         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1149         if (!new_root_item) {
1150                 pending->error = -ENOMEM;
1151                 goto root_item_alloc_fail;
1152         }
1153
1154         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1155         if (pending->error)
1156                 goto no_free_objectid;
1157
1158         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1159
1160         if (to_reserve > 0) {
1161                 pending->error = btrfs_block_rsv_add(root,
1162                                                      &pending->block_rsv,
1163                                                      to_reserve,
1164                                                      BTRFS_RESERVE_NO_FLUSH);
1165                 if (pending->error)
1166                         goto no_free_objectid;
1167         }
1168
1169         pending->error = btrfs_qgroup_inherit(trans, fs_info,
1170                                               root->root_key.objectid,
1171                                               objectid, pending->inherit);
1172         if (pending->error)
1173                 goto no_free_objectid;
1174
1175         key.objectid = objectid;
1176         key.offset = (u64)-1;
1177         key.type = BTRFS_ROOT_ITEM_KEY;
1178
1179         rsv = trans->block_rsv;
1180         trans->block_rsv = &pending->block_rsv;
1181         trans->bytes_reserved = trans->block_rsv->reserved;
1182
1183         dentry = pending->dentry;
1184         parent_inode = pending->dir;
1185         parent_root = BTRFS_I(parent_inode)->root;
1186         record_root_in_trans(trans, parent_root);
1187
1188         /*
1189          * insert the directory item
1190          */
1191         ret = btrfs_set_inode_index(parent_inode, &index);
1192         BUG_ON(ret); /* -ENOMEM */
1193
1194         /* check if there is a file/dir which has the same name. */
1195         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1196                                          btrfs_ino(parent_inode),
1197                                          dentry->d_name.name,
1198                                          dentry->d_name.len, 0);
1199         if (dir_item != NULL && !IS_ERR(dir_item)) {
1200                 pending->error = -EEXIST;
1201                 goto dir_item_existed;
1202         } else if (IS_ERR(dir_item)) {
1203                 ret = PTR_ERR(dir_item);
1204                 btrfs_abort_transaction(trans, root, ret);
1205                 goto fail;
1206         }
1207         btrfs_release_path(path);
1208
1209         /*
1210          * pull in the delayed directory update
1211          * and the delayed inode item
1212          * otherwise we corrupt the FS during
1213          * snapshot
1214          */
1215         ret = btrfs_run_delayed_items(trans, root);
1216         if (ret) {      /* Transaction aborted */
1217                 btrfs_abort_transaction(trans, root, ret);
1218                 goto fail;
1219         }
1220
1221         record_root_in_trans(trans, root);
1222         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1223         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1224         btrfs_check_and_init_root_item(new_root_item);
1225
1226         root_flags = btrfs_root_flags(new_root_item);
1227         if (pending->readonly)
1228                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1229         else
1230                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1231         btrfs_set_root_flags(new_root_item, root_flags);
1232
1233         btrfs_set_root_generation_v2(new_root_item,
1234                         trans->transid);
1235         uuid_le_gen(&new_uuid);
1236         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1237         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1238                         BTRFS_UUID_SIZE);
1239         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1240                 memset(new_root_item->received_uuid, 0,
1241                        sizeof(new_root_item->received_uuid));
1242                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1243                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1244                 btrfs_set_root_stransid(new_root_item, 0);
1245                 btrfs_set_root_rtransid(new_root_item, 0);
1246         }
1247         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1248         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1249         btrfs_set_root_otransid(new_root_item, trans->transid);
1250
1251         old = btrfs_lock_root_node(root);
1252         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1253         if (ret) {
1254                 btrfs_tree_unlock(old);
1255                 free_extent_buffer(old);
1256                 btrfs_abort_transaction(trans, root, ret);
1257                 goto fail;
1258         }
1259
1260         btrfs_set_lock_blocking(old);
1261
1262         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1263         /* clean up in any case */
1264         btrfs_tree_unlock(old);
1265         free_extent_buffer(old);
1266         if (ret) {
1267                 btrfs_abort_transaction(trans, root, ret);
1268                 goto fail;
1269         }
1270
1271         /* see comments in should_cow_block() */
1272         root->force_cow = 1;
1273         smp_wmb();
1274
1275         btrfs_set_root_node(new_root_item, tmp);
1276         /* record when the snapshot was created in key.offset */
1277         key.offset = trans->transid;
1278         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1279         btrfs_tree_unlock(tmp);
1280         free_extent_buffer(tmp);
1281         if (ret) {
1282                 btrfs_abort_transaction(trans, root, ret);
1283                 goto fail;
1284         }
1285
1286         /*
1287          * insert root back/forward references
1288          */
1289         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1290                                  parent_root->root_key.objectid,
1291                                  btrfs_ino(parent_inode), index,
1292                                  dentry->d_name.name, dentry->d_name.len);
1293         if (ret) {
1294                 btrfs_abort_transaction(trans, root, ret);
1295                 goto fail;
1296         }
1297
1298         key.offset = (u64)-1;
1299         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1300         if (IS_ERR(pending->snap)) {
1301                 ret = PTR_ERR(pending->snap);
1302                 btrfs_abort_transaction(trans, root, ret);
1303                 goto fail;
1304         }
1305
1306         ret = btrfs_reloc_post_snapshot(trans, pending);
1307         if (ret) {
1308                 btrfs_abort_transaction(trans, root, ret);
1309                 goto fail;
1310         }
1311
1312         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1313         if (ret) {
1314                 btrfs_abort_transaction(trans, root, ret);
1315                 goto fail;
1316         }
1317
1318         ret = btrfs_insert_dir_item(trans, parent_root,
1319                                     dentry->d_name.name, dentry->d_name.len,
1320                                     parent_inode, &key,
1321                                     BTRFS_FT_DIR, index);
1322         /* We have check then name at the beginning, so it is impossible. */
1323         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1324         if (ret) {
1325                 btrfs_abort_transaction(trans, root, ret);
1326                 goto fail;
1327         }
1328
1329         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1330                                          dentry->d_name.len * 2);
1331         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1332         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1333         if (ret) {
1334                 btrfs_abort_transaction(trans, root, ret);
1335                 goto fail;
1336         }
1337         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1338                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1339         if (ret) {
1340                 btrfs_abort_transaction(trans, root, ret);
1341                 goto fail;
1342         }
1343         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1344                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1345                                           new_root_item->received_uuid,
1346                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1347                                           objectid);
1348                 if (ret && ret != -EEXIST) {
1349                         btrfs_abort_transaction(trans, root, ret);
1350                         goto fail;
1351                 }
1352         }
1353 fail:
1354         pending->error = ret;
1355 dir_item_existed:
1356         trans->block_rsv = rsv;
1357         trans->bytes_reserved = 0;
1358 no_free_objectid:
1359         kfree(new_root_item);
1360 root_item_alloc_fail:
1361         btrfs_free_path(path);
1362         return ret;
1363 }
1364
1365 /*
1366  * create all the snapshots we've scheduled for creation
1367  */
1368 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1369                                              struct btrfs_fs_info *fs_info)
1370 {
1371         struct btrfs_pending_snapshot *pending, *next;
1372         struct list_head *head = &trans->transaction->pending_snapshots;
1373         int ret = 0;
1374
1375         list_for_each_entry_safe(pending, next, head, list) {
1376                 list_del(&pending->list);
1377                 ret = create_pending_snapshot(trans, fs_info, pending);
1378                 if (ret)
1379                         break;
1380         }
1381         return ret;
1382 }
1383
1384 static void update_super_roots(struct btrfs_root *root)
1385 {
1386         struct btrfs_root_item *root_item;
1387         struct btrfs_super_block *super;
1388
1389         super = root->fs_info->super_copy;
1390
1391         root_item = &root->fs_info->chunk_root->root_item;
1392         super->chunk_root = root_item->bytenr;
1393         super->chunk_root_generation = root_item->generation;
1394         super->chunk_root_level = root_item->level;
1395
1396         root_item = &root->fs_info->tree_root->root_item;
1397         super->root = root_item->bytenr;
1398         super->generation = root_item->generation;
1399         super->root_level = root_item->level;
1400         if (btrfs_test_opt(root, SPACE_CACHE))
1401                 super->cache_generation = root_item->generation;
1402         if (root->fs_info->update_uuid_tree_gen)
1403                 super->uuid_tree_generation = root_item->generation;
1404 }
1405
1406 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1407 {
1408         struct btrfs_transaction *trans;
1409         int ret = 0;
1410
1411         spin_lock(&info->trans_lock);
1412         trans = info->running_transaction;
1413         if (trans)
1414                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1415         spin_unlock(&info->trans_lock);
1416         return ret;
1417 }
1418
1419 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1420 {
1421         struct btrfs_transaction *trans;
1422         int ret = 0;
1423
1424         spin_lock(&info->trans_lock);
1425         trans = info->running_transaction;
1426         if (trans)
1427                 ret = is_transaction_blocked(trans);
1428         spin_unlock(&info->trans_lock);
1429         return ret;
1430 }
1431
1432 /*
1433  * wait for the current transaction commit to start and block subsequent
1434  * transaction joins
1435  */
1436 static void wait_current_trans_commit_start(struct btrfs_root *root,
1437                                             struct btrfs_transaction *trans)
1438 {
1439         wait_event(root->fs_info->transaction_blocked_wait,
1440                    trans->state >= TRANS_STATE_COMMIT_START ||
1441                    trans->aborted);
1442 }
1443
1444 /*
1445  * wait for the current transaction to start and then become unblocked.
1446  * caller holds ref.
1447  */
1448 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1449                                          struct btrfs_transaction *trans)
1450 {
1451         wait_event(root->fs_info->transaction_wait,
1452                    trans->state >= TRANS_STATE_UNBLOCKED ||
1453                    trans->aborted);
1454 }
1455
1456 /*
1457  * commit transactions asynchronously. once btrfs_commit_transaction_async
1458  * returns, any subsequent transaction will not be allowed to join.
1459  */
1460 struct btrfs_async_commit {
1461         struct btrfs_trans_handle *newtrans;
1462         struct btrfs_root *root;
1463         struct work_struct work;
1464 };
1465
1466 static void do_async_commit(struct work_struct *work)
1467 {
1468         struct btrfs_async_commit *ac =
1469                 container_of(work, struct btrfs_async_commit, work);
1470
1471         /*
1472          * We've got freeze protection passed with the transaction.
1473          * Tell lockdep about it.
1474          */
1475         if (ac->newtrans->type & __TRANS_FREEZABLE)
1476                 rwsem_acquire_read(
1477                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1478                      0, 1, _THIS_IP_);
1479
1480         current->journal_info = ac->newtrans;
1481
1482         btrfs_commit_transaction(ac->newtrans, ac->root);
1483         kfree(ac);
1484 }
1485
1486 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1487                                    struct btrfs_root *root,
1488                                    int wait_for_unblock)
1489 {
1490         struct btrfs_async_commit *ac;
1491         struct btrfs_transaction *cur_trans;
1492
1493         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1494         if (!ac)
1495                 return -ENOMEM;
1496
1497         INIT_WORK(&ac->work, do_async_commit);
1498         ac->root = root;
1499         ac->newtrans = btrfs_join_transaction(root);
1500         if (IS_ERR(ac->newtrans)) {
1501                 int err = PTR_ERR(ac->newtrans);
1502                 kfree(ac);
1503                 return err;
1504         }
1505
1506         /* take transaction reference */
1507         cur_trans = trans->transaction;
1508         atomic_inc(&cur_trans->use_count);
1509
1510         btrfs_end_transaction(trans, root);
1511
1512         /*
1513          * Tell lockdep we've released the freeze rwsem, since the
1514          * async commit thread will be the one to unlock it.
1515          */
1516         if (ac->newtrans->type & __TRANS_FREEZABLE)
1517                 rwsem_release(
1518                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1519                         1, _THIS_IP_);
1520
1521         schedule_work(&ac->work);
1522
1523         /* wait for transaction to start and unblock */
1524         if (wait_for_unblock)
1525                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1526         else
1527                 wait_current_trans_commit_start(root, cur_trans);
1528
1529         if (current->journal_info == trans)
1530                 current->journal_info = NULL;
1531
1532         btrfs_put_transaction(cur_trans);
1533         return 0;
1534 }
1535
1536
1537 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1538                                 struct btrfs_root *root, int err)
1539 {
1540         struct btrfs_transaction *cur_trans = trans->transaction;
1541         DEFINE_WAIT(wait);
1542
1543         WARN_ON(trans->use_count > 1);
1544
1545         btrfs_abort_transaction(trans, root, err);
1546
1547         spin_lock(&root->fs_info->trans_lock);
1548
1549         /*
1550          * If the transaction is removed from the list, it means this
1551          * transaction has been committed successfully, so it is impossible
1552          * to call the cleanup function.
1553          */
1554         BUG_ON(list_empty(&cur_trans->list));
1555
1556         list_del_init(&cur_trans->list);
1557         if (cur_trans == root->fs_info->running_transaction) {
1558                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1559                 spin_unlock(&root->fs_info->trans_lock);
1560                 wait_event(cur_trans->writer_wait,
1561                            atomic_read(&cur_trans->num_writers) == 1);
1562
1563                 spin_lock(&root->fs_info->trans_lock);
1564         }
1565         spin_unlock(&root->fs_info->trans_lock);
1566
1567         btrfs_cleanup_one_transaction(trans->transaction, root);
1568
1569         spin_lock(&root->fs_info->trans_lock);
1570         if (cur_trans == root->fs_info->running_transaction)
1571                 root->fs_info->running_transaction = NULL;
1572         spin_unlock(&root->fs_info->trans_lock);
1573
1574         if (trans->type & __TRANS_FREEZABLE)
1575                 sb_end_intwrite(root->fs_info->sb);
1576         btrfs_put_transaction(cur_trans);
1577         btrfs_put_transaction(cur_trans);
1578
1579         trace_btrfs_transaction_commit(root);
1580
1581         btrfs_scrub_continue(root);
1582
1583         if (current->journal_info == trans)
1584                 current->journal_info = NULL;
1585
1586         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1587 }
1588
1589 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1590                                           struct btrfs_root *root)
1591 {
1592         int ret;
1593
1594         ret = btrfs_run_delayed_items(trans, root);
1595         /*
1596          * running the delayed items may have added new refs. account
1597          * them now so that they hinder processing of more delayed refs
1598          * as little as possible.
1599          */
1600         if (ret) {
1601                 btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1602                 return ret;
1603         }
1604
1605         ret = btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1606         if (ret)
1607                 return ret;
1608
1609         /*
1610          * rename don't use btrfs_join_transaction, so, once we
1611          * set the transaction to blocked above, we aren't going
1612          * to get any new ordered operations.  We can safely run
1613          * it here and no for sure that nothing new will be added
1614          * to the list
1615          */
1616         ret = btrfs_run_ordered_operations(trans, root, 1);
1617
1618         return ret;
1619 }
1620
1621 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1622 {
1623         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1624                 return btrfs_start_delalloc_roots(fs_info, 1);
1625         return 0;
1626 }
1627
1628 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1629 {
1630         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1631                 btrfs_wait_ordered_roots(fs_info, -1);
1632 }
1633
1634 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1635                              struct btrfs_root *root)
1636 {
1637         struct btrfs_transaction *cur_trans = trans->transaction;
1638         struct btrfs_transaction *prev_trans = NULL;
1639         int ret;
1640
1641         ret = btrfs_run_ordered_operations(trans, root, 0);
1642         if (ret) {
1643                 btrfs_abort_transaction(trans, root, ret);
1644                 btrfs_end_transaction(trans, root);
1645                 return ret;
1646         }
1647
1648         /* Stop the commit early if ->aborted is set */
1649         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1650                 ret = cur_trans->aborted;
1651                 btrfs_end_transaction(trans, root);
1652                 return ret;
1653         }
1654
1655         /* make a pass through all the delayed refs we have so far
1656          * any runnings procs may add more while we are here
1657          */
1658         ret = btrfs_run_delayed_refs(trans, root, 0);
1659         if (ret) {
1660                 btrfs_end_transaction(trans, root);
1661                 return ret;
1662         }
1663
1664         btrfs_trans_release_metadata(trans, root);
1665         trans->block_rsv = NULL;
1666         if (trans->qgroup_reserved) {
1667                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1668                 trans->qgroup_reserved = 0;
1669         }
1670
1671         cur_trans = trans->transaction;
1672
1673         /*
1674          * set the flushing flag so procs in this transaction have to
1675          * start sending their work down.
1676          */
1677         cur_trans->delayed_refs.flushing = 1;
1678         smp_wmb();
1679
1680         if (!list_empty(&trans->new_bgs))
1681                 btrfs_create_pending_block_groups(trans, root);
1682
1683         ret = btrfs_run_delayed_refs(trans, root, 0);
1684         if (ret) {
1685                 btrfs_end_transaction(trans, root);
1686                 return ret;
1687         }
1688
1689         spin_lock(&root->fs_info->trans_lock);
1690         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1691                 spin_unlock(&root->fs_info->trans_lock);
1692                 atomic_inc(&cur_trans->use_count);
1693                 ret = btrfs_end_transaction(trans, root);
1694
1695                 wait_for_commit(root, cur_trans);
1696
1697                 btrfs_put_transaction(cur_trans);
1698
1699                 return ret;
1700         }
1701
1702         cur_trans->state = TRANS_STATE_COMMIT_START;
1703         wake_up(&root->fs_info->transaction_blocked_wait);
1704
1705         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1706                 prev_trans = list_entry(cur_trans->list.prev,
1707                                         struct btrfs_transaction, list);
1708                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1709                         atomic_inc(&prev_trans->use_count);
1710                         spin_unlock(&root->fs_info->trans_lock);
1711
1712                         wait_for_commit(root, prev_trans);
1713
1714                         btrfs_put_transaction(prev_trans);
1715                 } else {
1716                         spin_unlock(&root->fs_info->trans_lock);
1717                 }
1718         } else {
1719                 spin_unlock(&root->fs_info->trans_lock);
1720         }
1721
1722         extwriter_counter_dec(cur_trans, trans->type);
1723
1724         ret = btrfs_start_delalloc_flush(root->fs_info);
1725         if (ret)
1726                 goto cleanup_transaction;
1727
1728         ret = btrfs_flush_all_pending_stuffs(trans, root);
1729         if (ret)
1730                 goto cleanup_transaction;
1731
1732         wait_event(cur_trans->writer_wait,
1733                    extwriter_counter_read(cur_trans) == 0);
1734
1735         /* some pending stuffs might be added after the previous flush. */
1736         ret = btrfs_flush_all_pending_stuffs(trans, root);
1737         if (ret)
1738                 goto cleanup_transaction;
1739
1740         btrfs_wait_delalloc_flush(root->fs_info);
1741
1742         btrfs_scrub_pause(root);
1743         /*
1744          * Ok now we need to make sure to block out any other joins while we
1745          * commit the transaction.  We could have started a join before setting
1746          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1747          */
1748         spin_lock(&root->fs_info->trans_lock);
1749         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1750         spin_unlock(&root->fs_info->trans_lock);
1751         wait_event(cur_trans->writer_wait,
1752                    atomic_read(&cur_trans->num_writers) == 1);
1753
1754         /* ->aborted might be set after the previous check, so check it */
1755         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1756                 ret = cur_trans->aborted;
1757                 goto cleanup_transaction;
1758         }
1759         /*
1760          * the reloc mutex makes sure that we stop
1761          * the balancing code from coming in and moving
1762          * extents around in the middle of the commit
1763          */
1764         mutex_lock(&root->fs_info->reloc_mutex);
1765
1766         /*
1767          * We needn't worry about the delayed items because we will
1768          * deal with them in create_pending_snapshot(), which is the
1769          * core function of the snapshot creation.
1770          */
1771         ret = create_pending_snapshots(trans, root->fs_info);
1772         if (ret) {
1773                 mutex_unlock(&root->fs_info->reloc_mutex);
1774                 goto cleanup_transaction;
1775         }
1776
1777         /*
1778          * We insert the dir indexes of the snapshots and update the inode
1779          * of the snapshots' parents after the snapshot creation, so there
1780          * are some delayed items which are not dealt with. Now deal with
1781          * them.
1782          *
1783          * We needn't worry that this operation will corrupt the snapshots,
1784          * because all the tree which are snapshoted will be forced to COW
1785          * the nodes and leaves.
1786          */
1787         ret = btrfs_run_delayed_items(trans, root);
1788         if (ret) {
1789                 mutex_unlock(&root->fs_info->reloc_mutex);
1790                 goto cleanup_transaction;
1791         }
1792
1793         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1794         if (ret) {
1795                 mutex_unlock(&root->fs_info->reloc_mutex);
1796                 goto cleanup_transaction;
1797         }
1798
1799         /*
1800          * make sure none of the code above managed to slip in a
1801          * delayed item
1802          */
1803         btrfs_assert_delayed_root_empty(root);
1804
1805         WARN_ON(cur_trans != trans->transaction);
1806
1807         /* btrfs_commit_tree_roots is responsible for getting the
1808          * various roots consistent with each other.  Every pointer
1809          * in the tree of tree roots has to point to the most up to date
1810          * root for every subvolume and other tree.  So, we have to keep
1811          * the tree logging code from jumping in and changing any
1812          * of the trees.
1813          *
1814          * At this point in the commit, there can't be any tree-log
1815          * writers, but a little lower down we drop the trans mutex
1816          * and let new people in.  By holding the tree_log_mutex
1817          * from now until after the super is written, we avoid races
1818          * with the tree-log code.
1819          */
1820         mutex_lock(&root->fs_info->tree_log_mutex);
1821
1822         ret = commit_fs_roots(trans, root);
1823         if (ret) {
1824                 mutex_unlock(&root->fs_info->tree_log_mutex);
1825                 mutex_unlock(&root->fs_info->reloc_mutex);
1826                 goto cleanup_transaction;
1827         }
1828
1829         /* commit_fs_roots gets rid of all the tree log roots, it is now
1830          * safe to free the root of tree log roots
1831          */
1832         btrfs_free_log_root_tree(trans, root->fs_info);
1833
1834         ret = commit_cowonly_roots(trans, root);
1835         if (ret) {
1836                 mutex_unlock(&root->fs_info->tree_log_mutex);
1837                 mutex_unlock(&root->fs_info->reloc_mutex);
1838                 goto cleanup_transaction;
1839         }
1840
1841         /*
1842          * The tasks which save the space cache and inode cache may also
1843          * update ->aborted, check it.
1844          */
1845         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1846                 ret = cur_trans->aborted;
1847                 mutex_unlock(&root->fs_info->tree_log_mutex);
1848                 mutex_unlock(&root->fs_info->reloc_mutex);
1849                 goto cleanup_transaction;
1850         }
1851
1852         btrfs_prepare_extent_commit(trans, root);
1853
1854         cur_trans = root->fs_info->running_transaction;
1855
1856         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1857                             root->fs_info->tree_root->node);
1858         switch_commit_root(root->fs_info->tree_root);
1859
1860         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1861                             root->fs_info->chunk_root->node);
1862         switch_commit_root(root->fs_info->chunk_root);
1863
1864         assert_qgroups_uptodate(trans);
1865         update_super_roots(root);
1866
1867         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1868         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1869         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1870                sizeof(*root->fs_info->super_copy));
1871
1872         spin_lock(&root->fs_info->trans_lock);
1873         cur_trans->state = TRANS_STATE_UNBLOCKED;
1874         root->fs_info->running_transaction = NULL;
1875         spin_unlock(&root->fs_info->trans_lock);
1876         mutex_unlock(&root->fs_info->reloc_mutex);
1877
1878         wake_up(&root->fs_info->transaction_wait);
1879
1880         ret = btrfs_write_and_wait_transaction(trans, root);
1881         if (ret) {
1882                 btrfs_error(root->fs_info, ret,
1883                             "Error while writing out transaction");
1884                 mutex_unlock(&root->fs_info->tree_log_mutex);
1885                 goto cleanup_transaction;
1886         }
1887
1888         ret = write_ctree_super(trans, root, 0);
1889         if (ret) {
1890                 mutex_unlock(&root->fs_info->tree_log_mutex);
1891                 goto cleanup_transaction;
1892         }
1893
1894         /*
1895          * the super is written, we can safely allow the tree-loggers
1896          * to go about their business
1897          */
1898         mutex_unlock(&root->fs_info->tree_log_mutex);
1899
1900         btrfs_finish_extent_commit(trans, root);
1901
1902         root->fs_info->last_trans_committed = cur_trans->transid;
1903         /*
1904          * We needn't acquire the lock here because there is no other task
1905          * which can change it.
1906          */
1907         cur_trans->state = TRANS_STATE_COMPLETED;
1908         wake_up(&cur_trans->commit_wait);
1909
1910         spin_lock(&root->fs_info->trans_lock);
1911         list_del_init(&cur_trans->list);
1912         spin_unlock(&root->fs_info->trans_lock);
1913
1914         btrfs_put_transaction(cur_trans);
1915         btrfs_put_transaction(cur_trans);
1916
1917         if (trans->type & __TRANS_FREEZABLE)
1918                 sb_end_intwrite(root->fs_info->sb);
1919
1920         trace_btrfs_transaction_commit(root);
1921
1922         btrfs_scrub_continue(root);
1923
1924         if (current->journal_info == trans)
1925                 current->journal_info = NULL;
1926
1927         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1928
1929         if (current != root->fs_info->transaction_kthread)
1930                 btrfs_run_delayed_iputs(root);
1931
1932         return ret;
1933
1934 cleanup_transaction:
1935         btrfs_trans_release_metadata(trans, root);
1936         trans->block_rsv = NULL;
1937         if (trans->qgroup_reserved) {
1938                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1939                 trans->qgroup_reserved = 0;
1940         }
1941         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
1942         if (current->journal_info == trans)
1943                 current->journal_info = NULL;
1944         cleanup_transaction(trans, root, ret);
1945
1946         return ret;
1947 }
1948
1949 /*
1950  * return < 0 if error
1951  * 0 if there are no more dead_roots at the time of call
1952  * 1 there are more to be processed, call me again
1953  *
1954  * The return value indicates there are certainly more snapshots to delete, but
1955  * if there comes a new one during processing, it may return 0. We don't mind,
1956  * because btrfs_commit_super will poke cleaner thread and it will process it a
1957  * few seconds later.
1958  */
1959 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
1960 {
1961         int ret;
1962         struct btrfs_fs_info *fs_info = root->fs_info;
1963
1964         spin_lock(&fs_info->trans_lock);
1965         if (list_empty(&fs_info->dead_roots)) {
1966                 spin_unlock(&fs_info->trans_lock);
1967                 return 0;
1968         }
1969         root = list_first_entry(&fs_info->dead_roots,
1970                         struct btrfs_root, root_list);
1971         /*
1972          * Make sure root is not involved in send,
1973          * if we fail with first root, we return
1974          * directly rather than continue.
1975          */
1976         spin_lock(&root->root_item_lock);
1977         if (root->send_in_progress) {
1978                 spin_unlock(&fs_info->trans_lock);
1979                 spin_unlock(&root->root_item_lock);
1980                 return 0;
1981         }
1982         spin_unlock(&root->root_item_lock);
1983
1984         list_del_init(&root->root_list);
1985         spin_unlock(&fs_info->trans_lock);
1986
1987         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
1988
1989         btrfs_kill_all_delayed_nodes(root);
1990
1991         if (btrfs_header_backref_rev(root->node) <
1992                         BTRFS_MIXED_BACKREF_REV)
1993                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1994         else
1995                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
1996         /*
1997          * If we encounter a transaction abort during snapshot cleaning, we
1998          * don't want to crash here
1999          */
2000         return (ret < 0) ? 0 : 1;
2001 }