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 #include "qgroup.h"
  35
  36 #define BTRFS_ROOT_TRANS_TAG 0
  37
  38 static unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  39         [TRANS_STATE_RUNNING]           = 0U,
  40         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  41                                            __TRANS_START),
  42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  43                                            __TRANS_START |
  44                                            __TRANS_ATTACH),
  45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  46                                            __TRANS_START |
  47                                            __TRANS_ATTACH |
  48                                            __TRANS_JOIN),
  49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  50                                            __TRANS_START |
  51                                            __TRANS_ATTACH |
  52                                            __TRANS_JOIN |
  53                                            __TRANS_JOIN_NOLOCK),
  54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  55                                            __TRANS_START |
  56                                            __TRANS_ATTACH |
  57                                            __TRANS_JOIN |
  58                                            __TRANS_JOIN_NOLOCK),
  59 };
  60
  61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  62 {
  63         WARN_ON(atomic_read(&transaction->use_count) == 0);
  64         if (atomic_dec_and_test(&transaction->use_count)) {
  65                 BUG_ON(!list_empty(&transaction->list));
  66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  67                 while (!list_empty(&transaction->pending_chunks)) {
  68                         struct extent_map *em;
  69
  70                         em = list_first_entry(&transaction->pending_chunks,
  71                                               struct extent_map, list);
  72                         list_del_init(&em->list);
  73                         free_extent_map(em);
  74                 }
  75                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  76         }
  77 }
  78
  79 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
  80                                          struct btrfs_fs_info *fs_info)
  81 {
  82         struct btrfs_root *root, *tmp;
  83
  84         down_write(&fs_info->commit_root_sem);
  85         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
  86                                  dirty_list) {
  87                 list_del_init(&root->dirty_list);
  88                 free_extent_buffer(root->commit_root);
  89                 root->commit_root = btrfs_root_node(root);
  90                 if (is_fstree(root->objectid))
  91                         btrfs_unpin_free_ino(root);
  92         }
  93         up_write(&fs_info->commit_root_sem);
  94 }
  95
  96 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
  97                                          unsigned int type)
  98 {
  99         if (type & TRANS_EXTWRITERS)
 100                 atomic_inc(&trans->num_extwriters);
 101 }
 102
 103 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 104                                          unsigned int type)
 105 {
 106         if (type & TRANS_EXTWRITERS)
 107                 atomic_dec(&trans->num_extwriters);
 108 }
 109
 110 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 111                                           unsigned int type)
 112 {
 113         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 114 }
 115
 116 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 117 {
 118         return atomic_read(&trans->num_extwriters);
 119 }
 120
 121 /*
 122  * either allocate a new transaction or hop into the existing one
 123  */
 124 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 125 {
 126         struct btrfs_transaction *cur_trans;
 127         struct btrfs_fs_info *fs_info = root->fs_info;
 128
 129         spin_lock(&fs_info->trans_lock);
 130 loop:
 131         /* The file system has been taken offline. No new transactions. */
 132         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 133                 spin_unlock(&fs_info->trans_lock);
 134                 return -EROFS;
 135         }
 136
 137         cur_trans = fs_info->running_transaction;
 138         if (cur_trans) {
 139                 if (cur_trans->aborted) {
 140                         spin_unlock(&fs_info->trans_lock);
 141                         return cur_trans->aborted;
 142                 }
 143                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 144                         spin_unlock(&fs_info->trans_lock);
 145                         return -EBUSY;
 146                 }
 147                 atomic_inc(&cur_trans->use_count);
 148                 atomic_inc(&cur_trans->num_writers);
 149                 extwriter_counter_inc(cur_trans, type);
 150                 spin_unlock(&fs_info->trans_lock);
 151                 return 0;
 152         }
 153         spin_unlock(&fs_info->trans_lock);
 154
 155         /*
 156          * If we are ATTACH, we just want to catch the current transaction,
 157          * and commit it. If there is no transaction, just return ENOENT.
 158          */
 159         if (type == TRANS_ATTACH)
 160                 return -ENOENT;
 161
 162         /*
 163          * JOIN_NOLOCK only happens during the transaction commit, so
 164          * it is impossible that ->running_transaction is NULL
 165          */
 166         BUG_ON(type == TRANS_JOIN_NOLOCK);
 167
 168         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 169         if (!cur_trans)
 170                 return -ENOMEM;
 171
 172         spin_lock(&fs_info->trans_lock);
 173         if (fs_info->running_transaction) {
 174                 /*
 175                  * someone started a transaction after we unlocked.  Make sure
 176                  * to redo the checks above
 177                  */
 178                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 179                 goto loop;
 180         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 181                 spin_unlock(&fs_info->trans_lock);
 182                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 183                 return -EROFS;
 184         }
 185
 186         atomic_set(&cur_trans->num_writers, 1);
 187         extwriter_counter_init(cur_trans, type);
 188         init_waitqueue_head(&cur_trans->writer_wait);
 189         init_waitqueue_head(&cur_trans->commit_wait);
 190         cur_trans->state = TRANS_STATE_RUNNING;
 191         /*
 192          * One for this trans handle, one so it will live on until we
 193          * commit the transaction.
 194          */
 195         atomic_set(&cur_trans->use_count, 2);
 196         cur_trans->start_time = get_seconds();
 197
 198         cur_trans->delayed_refs.href_root = RB_ROOT;
 199         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 200         cur_trans->delayed_refs.num_heads_ready = 0;
 201         cur_trans->delayed_refs.num_heads = 0;
 202         cur_trans->delayed_refs.flushing = 0;
 203         cur_trans->delayed_refs.run_delayed_start = 0;
 204
 205         /*
 206          * although the tree mod log is per file system and not per transaction,
 207          * the log must never go across transaction boundaries.
 208          */
 209         smp_mb();
 210         if (!list_empty(&fs_info->tree_mod_seq_list))
 211                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 212                         "creating a fresh transaction\n");
 213         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 214                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 215                         "creating a fresh transaction\n");
 216         atomic64_set(&fs_info->tree_mod_seq, 0);
 217
 218         spin_lock_init(&cur_trans->delayed_refs.lock);
 219
 220         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 221         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 222         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 223         INIT_LIST_HEAD(&cur_trans->switch_commits);
 224         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 225         extent_io_tree_init(&cur_trans->dirty_pages,
 226                              fs_info->btree_inode->i_mapping);
 227         fs_info->generation++;
 228         cur_trans->transid = fs_info->generation;
 229         fs_info->running_transaction = cur_trans;
 230         cur_trans->aborted = 0;
 231         spin_unlock(&fs_info->trans_lock);
 232
 233         return 0;
 234 }
 235
 236 /*
 237  * this does all the record keeping required to make sure that a reference
 238  * counted root is properly recorded in a given transaction.  This is required
 239  * to make sure the old root from before we joined the transaction is deleted
 240  * when the transaction commits
 241  */
 242 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 243                                struct btrfs_root *root)
 244 {
 245         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 246             root->last_trans < trans->transid) {
 247                 WARN_ON(root == root->fs_info->extent_root);
 248                 WARN_ON(root->commit_root != root->node);
 249
 250                 /*
 251                  * see below for IN_TRANS_SETUP usage rules
 252                  * we have the reloc mutex held now, so there
 253                  * is only one writer in this function
 254                  */
 255                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 256
 257                 /* make sure readers find IN_TRANS_SETUP before
 258                  * they find our root->last_trans update
 259                  */
 260                 smp_wmb();
 261
 262                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 263                 if (root->last_trans == trans->transid) {
 264                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 265                         return 0;
 266                 }
 267                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 268                            (unsigned long)root->root_key.objectid,
 269                            BTRFS_ROOT_TRANS_TAG);
 270                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 271                 root->last_trans = trans->transid;
 272
 273                 /* this is pretty tricky.  We don't want to
 274                  * take the relocation lock in btrfs_record_root_in_trans
 275                  * unless we're really doing the first setup for this root in
 276                  * this transaction.
 277                  *
 278                  * Normally we'd use root->last_trans as a flag to decide
 279                  * if we want to take the expensive mutex.
 280                  *
 281                  * But, we have to set root->last_trans before we
 282                  * init the relocation root, otherwise, we trip over warnings
 283                  * in ctree.c.  The solution used here is to flag ourselves
 284                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 285                  * fixing up the reloc trees and everyone must wait.
 286                  *
 287                  * When this is zero, they can trust root->last_trans and fly
 288                  * through btrfs_record_root_in_trans without having to take the
 289                  * lock.  smp_wmb() makes sure that all the writes above are
 290                  * done before we pop in the zero below
 291                  */
 292                 btrfs_init_reloc_root(trans, root);
 293                 smp_mb__before_atomic();
 294                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 295         }
 296         return 0;
 297 }
 298
 299
 300 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 301                                struct btrfs_root *root)
 302 {
 303         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 304                 return 0;
 305
 306         /*
 307          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 308          * and barriers
 309          */
 310         smp_rmb();
 311         if (root->last_trans == trans->transid &&
 312             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 313                 return 0;
 314
 315         mutex_lock(&root->fs_info->reloc_mutex);
 316         record_root_in_trans(trans, root);
 317         mutex_unlock(&root->fs_info->reloc_mutex);
 318
 319         return 0;
 320 }
 321
 322 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 323 {
 324         return (trans->state >= TRANS_STATE_BLOCKED &&
 325                 trans->state < TRANS_STATE_UNBLOCKED &&
 326                 !trans->aborted);
 327 }
 328
 329 /* wait for commit against the current transaction to become unblocked
 330  * when this is done, it is safe to start a new transaction, but the current
 331  * transaction might not be fully on disk.
 332  */
 333 static void wait_current_trans(struct btrfs_root *root)
 334 {
 335         struct btrfs_transaction *cur_trans;
 336
 337         spin_lock(&root->fs_info->trans_lock);
 338         cur_trans = root->fs_info->running_transaction;
 339         if (cur_trans && is_transaction_blocked(cur_trans)) {
 340                 atomic_inc(&cur_trans->use_count);
 341                 spin_unlock(&root->fs_info->trans_lock);
 342
 343                 wait_event(root->fs_info->transaction_wait,
 344                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 345                            cur_trans->aborted);
 346                 btrfs_put_transaction(cur_trans);
 347         } else {
 348                 spin_unlock(&root->fs_info->trans_lock);
 349         }
 350 }
 351
 352 static int may_wait_transaction(struct btrfs_root *root, int type)
 353 {
 354         if (root->fs_info->log_root_recovering)
 355                 return 0;
 356
 357         if (type == TRANS_USERSPACE)
 358                 return 1;
 359
 360         if (type == TRANS_START &&
 361             !atomic_read(&root->fs_info->open_ioctl_trans))
 362                 return 1;
 363
 364         return 0;
 365 }
 366
 367 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 368 {
 369         if (!root->fs_info->reloc_ctl ||
 370             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 371             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 372             root->reloc_root)
 373                 return false;
 374
 375         return true;
 376 }
 377
 378 static struct btrfs_trans_handle *
 379 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 380                   enum btrfs_reserve_flush_enum flush)
 381 {
 382         struct btrfs_trans_handle *h;
 383         struct btrfs_transaction *cur_trans;
 384         u64 num_bytes = 0;
 385         u64 qgroup_reserved = 0;
 386         bool reloc_reserved = false;
 387         int ret;
 388
 389         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 390                 return ERR_PTR(-EROFS);
 391
 392         if (current->journal_info &&
 393             current->journal_info != (void *)BTRFS_SEND_TRANS_STUB) {
 394                 WARN_ON(type & TRANS_EXTWRITERS);
 395                 h = current->journal_info;
 396                 h->use_count++;
 397                 WARN_ON(h->use_count > 2);
 398                 h->orig_rsv = h->block_rsv;
 399                 h->block_rsv = NULL;
 400                 goto got_it;
 401         }
 402
 403         /*
 404          * Do the reservation before we join the transaction so we can do all
 405          * the appropriate flushing if need be.
 406          */
 407         if (num_items > 0 && root != root->fs_info->chunk_root) {
 408                 if (root->fs_info->quota_enabled &&
 409                     is_fstree(root->root_key.objectid)) {
 410                         qgroup_reserved = num_items * root->leafsize;
 411                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 412                         if (ret)
 413                                 return ERR_PTR(ret);
 414                 }
 415
 416                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 417                 /*
 418                  * Do the reservation for the relocation root creation
 419                  */
 420                 if (unlikely(need_reserve_reloc_root(root))) {
 421                         num_bytes += root->nodesize;
 422                         reloc_reserved = true;
 423                 }
 424
 425                 ret = btrfs_block_rsv_add(root,
 426                                           &root->fs_info->trans_block_rsv,
 427                                           num_bytes, flush);
 428                 if (ret)
 429                         goto reserve_fail;
 430         }
 431 again:
 432         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 433         if (!h) {
 434                 ret = -ENOMEM;
 435                 goto alloc_fail;
 436         }
 437
 438         /*
 439          * If we are JOIN_NOLOCK we're already committing a transaction and
 440          * waiting on this guy, so we don't need to do the sb_start_intwrite
 441          * because we're already holding a ref.  We need this because we could
 442          * have raced in and did an fsync() on a file which can kick a commit
 443          * and then we deadlock with somebody doing a freeze.
 444          *
 445          * If we are ATTACH, it means we just want to catch the current
 446          * transaction and commit it, so we needn't do sb_start_intwrite().
 447          */
 448         if (type & __TRANS_FREEZABLE)
 449                 sb_start_intwrite(root->fs_info->sb);
 450
 451         if (may_wait_transaction(root, type))
 452                 wait_current_trans(root);
 453
 454         do {
 455                 ret = join_transaction(root, type);
 456                 if (ret == -EBUSY) {
 457                         wait_current_trans(root);
 458                         if (unlikely(type == TRANS_ATTACH))
 459                                 ret = -ENOENT;
 460                 }
 461         } while (ret == -EBUSY);
 462
 463         if (ret < 0) {
 464                 /* We must get the transaction if we are JOIN_NOLOCK. */
 465                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 466                 goto join_fail;
 467         }
 468
 469         cur_trans = root->fs_info->running_transaction;
 470
 471         h->transid = cur_trans->transid;
 472         h->transaction = cur_trans;
 473         h->blocks_used = 0;
 474         h->bytes_reserved = 0;
 475         h->root = root;
 476         h->delayed_ref_updates = 0;
 477         h->use_count = 1;
 478         h->adding_csums = 0;
 479         h->block_rsv = NULL;
 480         h->orig_rsv = NULL;
 481         h->aborted = 0;
 482         h->qgroup_reserved = 0;
 483         h->delayed_ref_elem.seq = 0;
 484         h->type = type;
 485         h->allocating_chunk = false;
 486         h->reloc_reserved = false;
 487         h->sync = false;
 488         INIT_LIST_HEAD(&h->qgroup_ref_list);
 489         INIT_LIST_HEAD(&h->new_bgs);
 490
 491         smp_mb();
 492         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 493             may_wait_transaction(root, type)) {
 494                 btrfs_commit_transaction(h, root);
 495                 goto again;
 496         }
 497
 498         if (num_bytes) {
 499                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 500                                               h->transid, num_bytes, 1);
 501                 h->block_rsv = &root->fs_info->trans_block_rsv;
 502                 h->bytes_reserved = num_bytes;
 503                 h->reloc_reserved = reloc_reserved;
 504         }
 505         h->qgroup_reserved = qgroup_reserved;
 506
 507 got_it:
 508         btrfs_record_root_in_trans(h, root);
 509
 510         if (!current->journal_info && type != TRANS_USERSPACE)
 511                 current->journal_info = h;
 512         return h;
 513
 514 join_fail:
 515         if (type & __TRANS_FREEZABLE)
 516                 sb_end_intwrite(root->fs_info->sb);
 517         kmem_cache_free(btrfs_trans_handle_cachep, h);
 518 alloc_fail:
 519         if (num_bytes)
 520                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 521                                         num_bytes);
 522 reserve_fail:
 523         if (qgroup_reserved)
 524                 btrfs_qgroup_free(root, qgroup_reserved);
 525         return ERR_PTR(ret);
 526 }
 527
 528 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 529                                                    int num_items)
 530 {
 531         return start_transaction(root, num_items, TRANS_START,
 532                                  BTRFS_RESERVE_FLUSH_ALL);
 533 }
 534
 535 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 536                                         struct btrfs_root *root, int num_items)
 537 {
 538         return start_transaction(root, num_items, TRANS_START,
 539                                  BTRFS_RESERVE_FLUSH_LIMIT);
 540 }
 541
 542 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 543 {
 544         return start_transaction(root, 0, TRANS_JOIN, 0);
 545 }
 546
 547 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 548 {
 549         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 550 }
 551
 552 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 553 {
 554         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 555 }
 556
 557 /*
 558  * btrfs_attach_transaction() - catch the running transaction
 559  *
 560  * It is used when we want to commit the current the transaction, but
 561  * don't want to start a new one.
 562  *
 563  * Note: If this function return -ENOENT, it just means there is no
 564  * running transaction. But it is possible that the inactive transaction
 565  * is still in the memory, not fully on disk. If you hope there is no
 566  * inactive transaction in the fs when -ENOENT is returned, you should
 567  * invoke
 568  *     btrfs_attach_transaction_barrier()
 569  */
 570 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 571 {
 572         return start_transaction(root, 0, TRANS_ATTACH, 0);
 573 }
 574
 575 /*
 576  * btrfs_attach_transaction_barrier() - catch the running transaction
 577  *
 578  * It is similar to the above function, the differentia is this one
 579  * will wait for all the inactive transactions until they fully
 580  * complete.
 581  */
 582 struct btrfs_trans_handle *
 583 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 584 {
 585         struct btrfs_trans_handle *trans;
 586
 587         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 588         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 589                 btrfs_wait_for_commit(root, 0);
 590
 591         return trans;
 592 }
 593
 594 /* wait for a transaction commit to be fully complete */
 595 static noinline void wait_for_commit(struct btrfs_root *root,
 596                                     struct btrfs_transaction *commit)
 597 {
 598         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 599 }
 600
 601 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 602 {
 603         struct btrfs_transaction *cur_trans = NULL, *t;
 604         int ret = 0;
 605
 606         if (transid) {
 607                 if (transid <= root->fs_info->last_trans_committed)
 608                         goto out;
 609
 610                 ret = -EINVAL;
 611                 /* find specified transaction */
 612                 spin_lock(&root->fs_info->trans_lock);
 613                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 614                         if (t->transid == transid) {
 615                                 cur_trans = t;
 616                                 atomic_inc(&cur_trans->use_count);
 617                                 ret = 0;
 618                                 break;
 619                         }
 620                         if (t->transid > transid) {
 621                                 ret = 0;
 622                                 break;
 623                         }
 624                 }
 625                 spin_unlock(&root->fs_info->trans_lock);
 626                 /* The specified transaction doesn't exist */
 627                 if (!cur_trans)
 628                         goto out;
 629         } else {
 630                 /* find newest transaction that is committing | committed */
 631                 spin_lock(&root->fs_info->trans_lock);
 632                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 633                                             list) {
 634                         if (t->state >= TRANS_STATE_COMMIT_START) {
 635                                 if (t->state == TRANS_STATE_COMPLETED)
 636                                         break;
 637                                 cur_trans = t;
 638                                 atomic_inc(&cur_trans->use_count);
 639                                 break;
 640                         }
 641                 }
 642                 spin_unlock(&root->fs_info->trans_lock);
 643                 if (!cur_trans)
 644                         goto out;  /* nothing committing|committed */
 645         }
 646
 647         wait_for_commit(root, cur_trans);
 648         btrfs_put_transaction(cur_trans);
 649 out:
 650         return ret;
 651 }
 652
 653 void btrfs_throttle(struct btrfs_root *root)
 654 {
 655         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 656                 wait_current_trans(root);
 657 }
 658
 659 static int should_end_transaction(struct btrfs_trans_handle *trans,
 660                                   struct btrfs_root *root)
 661 {
 662         if (root->fs_info->global_block_rsv.space_info->full &&
 663             btrfs_check_space_for_delayed_refs(trans, root))
 664                 return 1;
 665
 666         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 667 }
 668
 669 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 670                                  struct btrfs_root *root)
 671 {
 672         struct btrfs_transaction *cur_trans = trans->transaction;
 673         int updates;
 674         int err;
 675
 676         smp_mb();
 677         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 678             cur_trans->delayed_refs.flushing)
 679                 return 1;
 680
 681         updates = trans->delayed_ref_updates;
 682         trans->delayed_ref_updates = 0;
 683         if (updates) {
 684                 err = btrfs_run_delayed_refs(trans, root, updates);
 685                 if (err) /* Error code will also eval true */
 686                         return err;
 687         }
 688
 689         return should_end_transaction(trans, root);
 690 }
 691
 692 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 693                           struct btrfs_root *root, int throttle)
 694 {
 695         struct btrfs_transaction *cur_trans = trans->transaction;
 696         struct btrfs_fs_info *info = root->fs_info;
 697         unsigned long cur = trans->delayed_ref_updates;
 698         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 699         int err = 0;
 700         int must_run_delayed_refs = 0;
 701
 702         if (trans->use_count > 1) {
 703                 trans->use_count--;
 704                 trans->block_rsv = trans->orig_rsv;
 705                 return 0;
 706         }
 707
 708         btrfs_trans_release_metadata(trans, root);
 709         trans->block_rsv = NULL;
 710
 711         if (!list_empty(&trans->new_bgs))
 712                 btrfs_create_pending_block_groups(trans, root);
 713
 714         trans->delayed_ref_updates = 0;
 715         if (!trans->sync) {
 716                 must_run_delayed_refs =
 717                         btrfs_should_throttle_delayed_refs(trans, root);
 718                 cur = max_t(unsigned long, cur, 32);
 719
 720                 /*
 721                  * don't make the caller wait if they are from a NOLOCK
 722                  * or ATTACH transaction, it will deadlock with commit
 723                  */
 724                 if (must_run_delayed_refs == 1 &&
 725                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 726                         must_run_delayed_refs = 2;
 727         }
 728
 729         if (trans->qgroup_reserved) {
 730                 /*
 731                  * the same root has to be passed here between start_transaction
 732                  * and end_transaction. Subvolume quota depends on this.
 733                  */
 734                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 735                 trans->qgroup_reserved = 0;
 736         }
 737
 738         btrfs_trans_release_metadata(trans, root);
 739         trans->block_rsv = NULL;
 740
 741         if (!list_empty(&trans->new_bgs))
 742                 btrfs_create_pending_block_groups(trans, root);
 743
 744         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 745             should_end_transaction(trans, root) &&
 746             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 747                 spin_lock(&info->trans_lock);
 748                 if (cur_trans->state == TRANS_STATE_RUNNING)
 749                         cur_trans->state = TRANS_STATE_BLOCKED;
 750                 spin_unlock(&info->trans_lock);
 751         }
 752
 753         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 754                 if (throttle)
 755                         return btrfs_commit_transaction(trans, root);
 756                 else
 757                         wake_up_process(info->transaction_kthread);
 758         }
 759
 760         if (trans->type & __TRANS_FREEZABLE)
 761                 sb_end_intwrite(root->fs_info->sb);
 762
 763         WARN_ON(cur_trans != info->running_transaction);
 764         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 765         atomic_dec(&cur_trans->num_writers);
 766         extwriter_counter_dec(cur_trans, trans->type);
 767
 768         smp_mb();
 769         if (waitqueue_active(&cur_trans->writer_wait))
 770                 wake_up(&cur_trans->writer_wait);
 771         btrfs_put_transaction(cur_trans);
 772
 773         if (current->journal_info == trans)
 774                 current->journal_info = NULL;
 775
 776         if (throttle)
 777                 btrfs_run_delayed_iputs(root);
 778
 779         if (trans->aborted ||
 780             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 781                 wake_up_process(info->transaction_kthread);
 782                 err = -EIO;
 783         }
 784         assert_qgroups_uptodate(trans);
 785
 786         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 787         if (must_run_delayed_refs) {
 788                 btrfs_async_run_delayed_refs(root, cur,
 789                                              must_run_delayed_refs == 1);
 790         }
 791         return err;
 792 }
 793
 794 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 795                           struct btrfs_root *root)
 796 {
 797         return __btrfs_end_transaction(trans, root, 0);
 798 }
 799
 800 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 801                                    struct btrfs_root *root)
 802 {
 803         return __btrfs_end_transaction(trans, root, 1);
 804 }
 805
 806 /*
 807  * when btree blocks are allocated, they have some corresponding bits set for
 808  * them in one of two extent_io trees.  This is used to make sure all of
 809  * those extents are sent to disk but does not wait on them
 810  */
 811 int btrfs_write_marked_extents(struct btrfs_root *root,
 812                                struct extent_io_tree *dirty_pages, int mark)
 813 {
 814         int err = 0;
 815         int werr = 0;
 816         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 817         struct extent_state *cached_state = NULL;
 818         u64 start = 0;
 819         u64 end;
 820
 821         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 822                                       mark, &cached_state)) {
 823                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 824                                    mark, &cached_state, GFP_NOFS);
 825                 cached_state = NULL;
 826                 err = filemap_fdatawrite_range(mapping, start, end);
 827                 if (err)
 828                         werr = err;
 829                 cond_resched();
 830                 start = end + 1;
 831         }
 832         if (err)
 833                 werr = err;
 834         return werr;
 835 }
 836
 837 /*
 838  * when btree blocks are allocated, they have some corresponding bits set for
 839  * them in one of two extent_io trees.  This is used to make sure all of
 840  * those extents are on disk for transaction or log commit.  We wait
 841  * on all the pages and clear them from the dirty pages state tree
 842  */
 843 int btrfs_wait_marked_extents(struct btrfs_root *root,
 844                               struct extent_io_tree *dirty_pages, int mark)
 845 {
 846         int err = 0;
 847         int werr = 0;
 848         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 849         struct extent_state *cached_state = NULL;
 850         u64 start = 0;
 851         u64 end;
 852
 853         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 854                                       EXTENT_NEED_WAIT, &cached_state)) {
 855                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 856                                  0, 0, &cached_state, GFP_NOFS);
 857                 err = filemap_fdatawait_range(mapping, start, end);
 858                 if (err)
 859                         werr = err;
 860                 cond_resched();
 861                 start = end + 1;
 862         }
 863         if (err)
 864                 werr = err;
 865         return werr;
 866 }
 867
 868 /*
 869  * when btree blocks are allocated, they have some corresponding bits set for
 870  * them in one of two extent_io trees.  This is used to make sure all of
 871  * those extents are on disk for transaction or log commit
 872  */
 873 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 874                                 struct extent_io_tree *dirty_pages, int mark)
 875 {
 876         int ret;
 877         int ret2;
 878         struct blk_plug plug;
 879
 880         blk_start_plug(&plug);
 881         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 882         blk_finish_plug(&plug);
 883         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 884
 885         if (ret)
 886                 return ret;
 887         if (ret2)
 888                 return ret2;
 889         return 0;
 890 }
 891
 892 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 893                                      struct btrfs_root *root)
 894 {
 895         if (!trans || !trans->transaction) {
 896                 struct inode *btree_inode;
 897                 btree_inode = root->fs_info->btree_inode;
 898                 return filemap_write_and_wait(btree_inode->i_mapping);
 899         }
 900         return btrfs_write_and_wait_marked_extents(root,
 901                                            &trans->transaction->dirty_pages,
 902                                            EXTENT_DIRTY);
 903 }
 904
 905 /*
 906  * this is used to update the root pointer in the tree of tree roots.
 907  *
 908  * But, in the case of the extent allocation tree, updating the root
 909  * pointer may allocate blocks which may change the root of the extent
 910  * allocation tree.
 911  *
 912  * So, this loops and repeats and makes sure the cowonly root didn't
 913  * change while the root pointer was being updated in the metadata.
 914  */
 915 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 916                                struct btrfs_root *root)
 917 {
 918         int ret;
 919         u64 old_root_bytenr;
 920         u64 old_root_used;
 921         struct btrfs_root *tree_root = root->fs_info->tree_root;
 922
 923         old_root_used = btrfs_root_used(&root->root_item);
 924         btrfs_write_dirty_block_groups(trans, root);
 925
 926         while (1) {
 927                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 928                 if (old_root_bytenr == root->node->start &&
 929                     old_root_used == btrfs_root_used(&root->root_item))
 930                         break;
 931
 932                 btrfs_set_root_node(&root->root_item, root->node);
 933                 ret = btrfs_update_root(trans, tree_root,
 934                                         &root->root_key,
 935                                         &root->root_item);
 936                 if (ret)
 937                         return ret;
 938
 939                 old_root_used = btrfs_root_used(&root->root_item);
 940                 ret = btrfs_write_dirty_block_groups(trans, root);
 941                 if (ret)
 942                         return ret;
 943         }
 944
 945         return 0;
 946 }
 947
 948 /*
 949  * update all the cowonly tree roots on disk
 950  *
 951  * The error handling in this function may not be obvious. Any of the
 952  * failures will cause the file system to go offline. We still need
 953  * to clean up the delayed refs.
 954  */
 955 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 956                                          struct btrfs_root *root)
 957 {
 958         struct btrfs_fs_info *fs_info = root->fs_info;
 959         struct list_head *next;
 960         struct extent_buffer *eb;
 961         int ret;
 962
 963         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 964         if (ret)
 965                 return ret;
 966
 967         eb = btrfs_lock_root_node(fs_info->tree_root);
 968         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 969                               0, &eb);
 970         btrfs_tree_unlock(eb);
 971         free_extent_buffer(eb);
 972
 973         if (ret)
 974                 return ret;
 975
 976         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 977         if (ret)
 978                 return ret;
 979
 980         ret = btrfs_run_dev_stats(trans, root->fs_info);
 981         if (ret)
 982                 return ret;
 983         ret = btrfs_run_dev_replace(trans, root->fs_info);
 984         if (ret)
 985                 return ret;
 986         ret = btrfs_run_qgroups(trans, root->fs_info);
 987         if (ret)
 988                 return ret;
 989
 990         /* run_qgroups might have added some more refs */
 991         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 992         if (ret)
 993                 return ret;
 994
 995         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 996                 next = fs_info->dirty_cowonly_roots.next;
 997                 list_del_init(next);
 998                 root = list_entry(next, struct btrfs_root, dirty_list);
 999
1000                 if (root != fs_info->extent_root)
1001                         list_add_tail(&root->dirty_list,
1002                                       &trans->transaction->switch_commits);
1003                 ret = update_cowonly_root(trans, root);
1004                 if (ret)
1005                         return ret;
1006         }
1007
1008         list_add_tail(&fs_info->extent_root->dirty_list,
1009                       &trans->transaction->switch_commits);
1010         btrfs_after_dev_replace_commit(fs_info);
1011
1012         return 0;
1013 }
1014
1015 /*
1016  * dead roots are old snapshots that need to be deleted.  This allocates
1017  * a dirty root struct and adds it into the list of dead roots that need to
1018  * be deleted
1019  */
1020 void btrfs_add_dead_root(struct btrfs_root *root)
1021 {
1022         spin_lock(&root->fs_info->trans_lock);
1023         if (list_empty(&root->root_list))
1024                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1025         spin_unlock(&root->fs_info->trans_lock);
1026 }
1027
1028 /*
1029  * update all the cowonly tree roots on disk
1030  */
1031 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1032                                     struct btrfs_root *root)
1033 {
1034         struct btrfs_root *gang[8];
1035         struct btrfs_fs_info *fs_info = root->fs_info;
1036         int i;
1037         int ret;
1038         int err = 0;
1039
1040         spin_lock(&fs_info->fs_roots_radix_lock);
1041         while (1) {
1042                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1043                                                  (void **)gang, 0,
1044                                                  ARRAY_SIZE(gang),
1045                                                  BTRFS_ROOT_TRANS_TAG);
1046                 if (ret == 0)
1047                         break;
1048                 for (i = 0; i < ret; i++) {
1049                         root = gang[i];
1050                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1051                                         (unsigned long)root->root_key.objectid,
1052                                         BTRFS_ROOT_TRANS_TAG);
1053                         spin_unlock(&fs_info->fs_roots_radix_lock);
1054
1055                         btrfs_free_log(trans, root);
1056                         btrfs_update_reloc_root(trans, root);
1057                         btrfs_orphan_commit_root(trans, root);
1058
1059                         btrfs_save_ino_cache(root, trans);
1060
1061                         /* see comments in should_cow_block() */
1062                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1063                         smp_mb__after_atomic();
1064
1065                         if (root->commit_root != root->node) {
1066                                 list_add_tail(&root->dirty_list,
1067                                         &trans->transaction->switch_commits);
1068                                 btrfs_set_root_node(&root->root_item,
1069                                                     root->node);
1070                         }
1071
1072                         err = btrfs_update_root(trans, fs_info->tree_root,
1073                                                 &root->root_key,
1074                                                 &root->root_item);
1075                         spin_lock(&fs_info->fs_roots_radix_lock);
1076                         if (err)
1077                                 break;
1078                 }
1079         }
1080         spin_unlock(&fs_info->fs_roots_radix_lock);
1081         return err;
1082 }
1083
1084 /*
1085  * defrag a given btree.
1086  * Every leaf in the btree is read and defragged.
1087  */
1088 int btrfs_defrag_root(struct btrfs_root *root)
1089 {
1090         struct btrfs_fs_info *info = root->fs_info;
1091         struct btrfs_trans_handle *trans;
1092         int ret;
1093
1094         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1095                 return 0;
1096
1097         while (1) {
1098                 trans = btrfs_start_transaction(root, 0);
1099                 if (IS_ERR(trans))
1100                         return PTR_ERR(trans);
1101
1102                 ret = btrfs_defrag_leaves(trans, root);
1103
1104                 btrfs_end_transaction(trans, root);
1105                 btrfs_btree_balance_dirty(info->tree_root);
1106                 cond_resched();
1107
1108                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1109                         break;
1110
1111                 if (btrfs_defrag_cancelled(root->fs_info)) {
1112                         pr_debug("BTRFS: defrag_root cancelled\n");
1113                         ret = -EAGAIN;
1114                         break;
1115                 }
1116         }
1117         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1118         return ret;
1119 }
1120
1121 /*
1122  * new snapshots need to be created at a very specific time in the
1123  * transaction commit.  This does the actual creation.
1124  *
1125  * Note:
1126  * If the error which may affect the commitment of the current transaction
1127  * happens, we should return the error number. If the error which just affect
1128  * the creation of the pending snapshots, just return 0.
1129  */
1130 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1131                                    struct btrfs_fs_info *fs_info,
1132                                    struct btrfs_pending_snapshot *pending)
1133 {
1134         struct btrfs_key key;
1135         struct btrfs_root_item *new_root_item;
1136         struct btrfs_root *tree_root = fs_info->tree_root;
1137         struct btrfs_root *root = pending->root;
1138         struct btrfs_root *parent_root;
1139         struct btrfs_block_rsv *rsv;
1140         struct inode *parent_inode;
1141         struct btrfs_path *path;
1142         struct btrfs_dir_item *dir_item;
1143         struct dentry *dentry;
1144         struct extent_buffer *tmp;
1145         struct extent_buffer *old;
1146         struct timespec cur_time = CURRENT_TIME;
1147         int ret = 0;
1148         u64 to_reserve = 0;
1149         u64 index = 0;
1150         u64 objectid;
1151         u64 root_flags;
1152         uuid_le new_uuid;
1153
1154         path = btrfs_alloc_path();
1155         if (!path) {
1156                 pending->error = -ENOMEM;
1157                 return 0;
1158         }
1159
1160         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1161         if (!new_root_item) {
1162                 pending->error = -ENOMEM;
1163                 goto root_item_alloc_fail;
1164         }
1165
1166         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1167         if (pending->error)
1168                 goto no_free_objectid;
1169
1170         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1171
1172         if (to_reserve > 0) {
1173                 pending->error = btrfs_block_rsv_add(root,
1174                                                      &pending->block_rsv,
1175                                                      to_reserve,
1176                                                      BTRFS_RESERVE_NO_FLUSH);
1177                 if (pending->error)
1178                         goto no_free_objectid;
1179         }
1180
1181         key.objectid = objectid;
1182         key.offset = (u64)-1;
1183         key.type = BTRFS_ROOT_ITEM_KEY;
1184
1185         rsv = trans->block_rsv;
1186         trans->block_rsv = &pending->block_rsv;
1187         trans->bytes_reserved = trans->block_rsv->reserved;
1188
1189         dentry = pending->dentry;
1190         parent_inode = pending->dir;
1191         parent_root = BTRFS_I(parent_inode)->root;
1192         record_root_in_trans(trans, parent_root);
1193
1194         /*
1195          * insert the directory item
1196          */
1197         ret = btrfs_set_inode_index(parent_inode, &index);
1198         BUG_ON(ret); /* -ENOMEM */
1199
1200         /* check if there is a file/dir which has the same name. */
1201         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1202                                          btrfs_ino(parent_inode),
1203                                          dentry->d_name.name,
1204                                          dentry->d_name.len, 0);
1205         if (dir_item != NULL && !IS_ERR(dir_item)) {
1206                 pending->error = -EEXIST;
1207                 goto dir_item_existed;
1208         } else if (IS_ERR(dir_item)) {
1209                 ret = PTR_ERR(dir_item);
1210                 btrfs_abort_transaction(trans, root, ret);
1211                 goto fail;
1212         }
1213         btrfs_release_path(path);
1214
1215         /*
1216          * pull in the delayed directory update
1217          * and the delayed inode item
1218          * otherwise we corrupt the FS during
1219          * snapshot
1220          */
1221         ret = btrfs_run_delayed_items(trans, root);
1222         if (ret) {      /* Transaction aborted */
1223                 btrfs_abort_transaction(trans, root, ret);
1224                 goto fail;
1225         }
1226
1227         record_root_in_trans(trans, root);
1228         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1229         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1230         btrfs_check_and_init_root_item(new_root_item);
1231
1232         root_flags = btrfs_root_flags(new_root_item);
1233         if (pending->readonly)
1234                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1235         else
1236                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1237         btrfs_set_root_flags(new_root_item, root_flags);
1238
1239         btrfs_set_root_generation_v2(new_root_item,
1240                         trans->transid);
1241         uuid_le_gen(&new_uuid);
1242         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1243         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1244                         BTRFS_UUID_SIZE);
1245         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1246                 memset(new_root_item->received_uuid, 0,
1247                        sizeof(new_root_item->received_uuid));
1248                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1249                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1250                 btrfs_set_root_stransid(new_root_item, 0);
1251                 btrfs_set_root_rtransid(new_root_item, 0);
1252         }
1253         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1254         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1255         btrfs_set_root_otransid(new_root_item, trans->transid);
1256
1257         old = btrfs_lock_root_node(root);
1258         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1259         if (ret) {
1260                 btrfs_tree_unlock(old);
1261                 free_extent_buffer(old);
1262                 btrfs_abort_transaction(trans, root, ret);
1263                 goto fail;
1264         }
1265
1266         btrfs_set_lock_blocking(old);
1267
1268         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1269         /* clean up in any case */
1270         btrfs_tree_unlock(old);
1271         free_extent_buffer(old);
1272         if (ret) {
1273                 btrfs_abort_transaction(trans, root, ret);
1274                 goto fail;
1275         }
1276
1277         /*
1278          * We need to flush delayed refs in order to make sure all of our quota
1279          * operations have been done before we call btrfs_qgroup_inherit.
1280          */
1281         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1282         if (ret) {
1283                 btrfs_abort_transaction(trans, root, ret);
1284                 goto fail;
1285         }
1286
1287         ret = btrfs_qgroup_inherit(trans, fs_info,
1288                                    root->root_key.objectid,
1289                                    objectid, pending->inherit);
1290         if (ret) {
1291                 btrfs_abort_transaction(trans, root, ret);
1292                 goto fail;
1293         }
1294
1295         /* see comments in should_cow_block() */
1296         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1297         smp_wmb();
1298
1299         btrfs_set_root_node(new_root_item, tmp);
1300         /* record when the snapshot was created in key.offset */
1301         key.offset = trans->transid;
1302         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1303         btrfs_tree_unlock(tmp);
1304         free_extent_buffer(tmp);
1305         if (ret) {
1306                 btrfs_abort_transaction(trans, root, ret);
1307                 goto fail;
1308         }
1309
1310         /*
1311          * insert root back/forward references
1312          */
1313         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1314                                  parent_root->root_key.objectid,
1315                                  btrfs_ino(parent_inode), index,
1316                                  dentry->d_name.name, dentry->d_name.len);
1317         if (ret) {
1318                 btrfs_abort_transaction(trans, root, ret);
1319                 goto fail;
1320         }
1321
1322         key.offset = (u64)-1;
1323         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1324         if (IS_ERR(pending->snap)) {
1325                 ret = PTR_ERR(pending->snap);
1326                 btrfs_abort_transaction(trans, root, ret);
1327                 goto fail;
1328         }
1329
1330         ret = btrfs_reloc_post_snapshot(trans, pending);
1331         if (ret) {
1332                 btrfs_abort_transaction(trans, root, ret);
1333                 goto fail;
1334         }
1335
1336         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1337         if (ret) {
1338                 btrfs_abort_transaction(trans, root, ret);
1339                 goto fail;
1340         }
1341
1342         ret = btrfs_insert_dir_item(trans, parent_root,
1343                                     dentry->d_name.name, dentry->d_name.len,
1344                                     parent_inode, &key,
1345                                     BTRFS_FT_DIR, index);
1346         /* We have check then name at the beginning, so it is impossible. */
1347         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1348         if (ret) {
1349                 btrfs_abort_transaction(trans, root, ret);
1350                 goto fail;
1351         }
1352
1353         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1354                                          dentry->d_name.len * 2);
1355         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1356         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1357         if (ret) {
1358                 btrfs_abort_transaction(trans, root, ret);
1359                 goto fail;
1360         }
1361         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1362                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1363         if (ret) {
1364                 btrfs_abort_transaction(trans, root, ret);
1365                 goto fail;
1366         }
1367         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1368                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1369                                           new_root_item->received_uuid,
1370                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1371                                           objectid);
1372                 if (ret && ret != -EEXIST) {
1373                         btrfs_abort_transaction(trans, root, ret);
1374                         goto fail;
1375                 }
1376         }
1377 fail:
1378         pending->error = ret;
1379 dir_item_existed:
1380         trans->block_rsv = rsv;
1381         trans->bytes_reserved = 0;
1382 no_free_objectid:
1383         kfree(new_root_item);
1384 root_item_alloc_fail:
1385         btrfs_free_path(path);
1386         return ret;
1387 }
1388
1389 /*
1390  * create all the snapshots we've scheduled for creation
1391  */
1392 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1393                                              struct btrfs_fs_info *fs_info)
1394 {
1395         struct btrfs_pending_snapshot *pending, *next;
1396         struct list_head *head = &trans->transaction->pending_snapshots;
1397         int ret = 0;
1398
1399         list_for_each_entry_safe(pending, next, head, list) {
1400                 list_del(&pending->list);
1401                 ret = create_pending_snapshot(trans, fs_info, pending);
1402                 if (ret)
1403                         break;
1404         }
1405         return ret;
1406 }
1407
1408 static void update_super_roots(struct btrfs_root *root)
1409 {
1410         struct btrfs_root_item *root_item;
1411         struct btrfs_super_block *super;
1412
1413         super = root->fs_info->super_copy;
1414
1415         root_item = &root->fs_info->chunk_root->root_item;
1416         super->chunk_root = root_item->bytenr;
1417         super->chunk_root_generation = root_item->generation;
1418         super->chunk_root_level = root_item->level;
1419
1420         root_item = &root->fs_info->tree_root->root_item;
1421         super->root = root_item->bytenr;
1422         super->generation = root_item->generation;
1423         super->root_level = root_item->level;
1424         if (btrfs_test_opt(root, SPACE_CACHE))
1425                 super->cache_generation = root_item->generation;
1426         if (root->fs_info->update_uuid_tree_gen)
1427                 super->uuid_tree_generation = root_item->generation;
1428 }
1429
1430 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1431 {
1432         struct btrfs_transaction *trans;
1433         int ret = 0;
1434
1435         spin_lock(&info->trans_lock);
1436         trans = info->running_transaction;
1437         if (trans)
1438                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1439         spin_unlock(&info->trans_lock);
1440         return ret;
1441 }
1442
1443 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1444 {
1445         struct btrfs_transaction *trans;
1446         int ret = 0;
1447
1448         spin_lock(&info->trans_lock);
1449         trans = info->running_transaction;
1450         if (trans)
1451                 ret = is_transaction_blocked(trans);
1452         spin_unlock(&info->trans_lock);
1453         return ret;
1454 }
1455
1456 /*
1457  * wait for the current transaction commit to start and block subsequent
1458  * transaction joins
1459  */
1460 static void wait_current_trans_commit_start(struct btrfs_root *root,
1461                                             struct btrfs_transaction *trans)
1462 {
1463         wait_event(root->fs_info->transaction_blocked_wait,
1464                    trans->state >= TRANS_STATE_COMMIT_START ||
1465                    trans->aborted);
1466 }
1467
1468 /*
1469  * wait for the current transaction to start and then become unblocked.
1470  * caller holds ref.
1471  */
1472 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1473                                          struct btrfs_transaction *trans)
1474 {
1475         wait_event(root->fs_info->transaction_wait,
1476                    trans->state >= TRANS_STATE_UNBLOCKED ||
1477                    trans->aborted);
1478 }
1479
1480 /*
1481  * commit transactions asynchronously. once btrfs_commit_transaction_async
1482  * returns, any subsequent transaction will not be allowed to join.
1483  */
1484 struct btrfs_async_commit {
1485         struct btrfs_trans_handle *newtrans;
1486         struct btrfs_root *root;
1487         struct work_struct work;
1488 };
1489
1490 static void do_async_commit(struct work_struct *work)
1491 {
1492         struct btrfs_async_commit *ac =
1493                 container_of(work, struct btrfs_async_commit, work);
1494
1495         /*
1496          * We've got freeze protection passed with the transaction.
1497          * Tell lockdep about it.
1498          */
1499         if (ac->newtrans->type & __TRANS_FREEZABLE)
1500                 rwsem_acquire_read(
1501                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1502                      0, 1, _THIS_IP_);
1503
1504         current->journal_info = ac->newtrans;
1505
1506         btrfs_commit_transaction(ac->newtrans, ac->root);
1507         kfree(ac);
1508 }
1509
1510 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1511                                    struct btrfs_root *root,
1512                                    int wait_for_unblock)
1513 {
1514         struct btrfs_async_commit *ac;
1515         struct btrfs_transaction *cur_trans;
1516
1517         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1518         if (!ac)
1519                 return -ENOMEM;
1520
1521         INIT_WORK(&ac->work, do_async_commit);
1522         ac->root = root;
1523         ac->newtrans = btrfs_join_transaction(root);
1524         if (IS_ERR(ac->newtrans)) {
1525                 int err = PTR_ERR(ac->newtrans);
1526                 kfree(ac);
1527                 return err;
1528         }
1529
1530         /* take transaction reference */
1531         cur_trans = trans->transaction;
1532         atomic_inc(&cur_trans->use_count);
1533
1534         btrfs_end_transaction(trans, root);
1535
1536         /*
1537          * Tell lockdep we've released the freeze rwsem, since the
1538          * async commit thread will be the one to unlock it.
1539          */
1540         if (ac->newtrans->type & __TRANS_FREEZABLE)
1541                 rwsem_release(
1542                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1543                         1, _THIS_IP_);
1544
1545         schedule_work(&ac->work);
1546
1547         /* wait for transaction to start and unblock */
1548         if (wait_for_unblock)
1549                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1550         else
1551                 wait_current_trans_commit_start(root, cur_trans);
1552
1553         if (current->journal_info == trans)
1554                 current->journal_info = NULL;
1555
1556         btrfs_put_transaction(cur_trans);
1557         return 0;
1558 }
1559
1560
1561 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1562                                 struct btrfs_root *root, int err)
1563 {
1564         struct btrfs_transaction *cur_trans = trans->transaction;
1565         DEFINE_WAIT(wait);
1566
1567         WARN_ON(trans->use_count > 1);
1568
1569         btrfs_abort_transaction(trans, root, err);
1570
1571         spin_lock(&root->fs_info->trans_lock);
1572
1573         /*
1574          * If the transaction is removed from the list, it means this
1575          * transaction has been committed successfully, so it is impossible
1576          * to call the cleanup function.
1577          */
1578         BUG_ON(list_empty(&cur_trans->list));
1579
1580         list_del_init(&cur_trans->list);
1581         if (cur_trans == root->fs_info->running_transaction) {
1582                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1583                 spin_unlock(&root->fs_info->trans_lock);
1584                 wait_event(cur_trans->writer_wait,
1585                            atomic_read(&cur_trans->num_writers) == 1);
1586
1587                 spin_lock(&root->fs_info->trans_lock);
1588         }
1589         spin_unlock(&root->fs_info->trans_lock);
1590
1591         btrfs_cleanup_one_transaction(trans->transaction, root);
1592
1593         spin_lock(&root->fs_info->trans_lock);
1594         if (cur_trans == root->fs_info->running_transaction)
1595                 root->fs_info->running_transaction = NULL;
1596         spin_unlock(&root->fs_info->trans_lock);
1597
1598         if (trans->type & __TRANS_FREEZABLE)
1599                 sb_end_intwrite(root->fs_info->sb);
1600         btrfs_put_transaction(cur_trans);
1601         btrfs_put_transaction(cur_trans);
1602
1603         trace_btrfs_transaction_commit(root);
1604
1605         if (current->journal_info == trans)
1606                 current->journal_info = NULL;
1607         btrfs_scrub_cancel(root->fs_info);
1608
1609         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1610 }
1611
1612 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1613                                           struct btrfs_root *root)
1614 {
1615         int ret;
1616
1617         ret = btrfs_run_delayed_items(trans, root);
1618         /*
1619          * running the delayed items may have added new refs. account
1620          * them now so that they hinder processing of more delayed refs
1621          * as little as possible.
1622          */
1623         if (ret)
1624                 return ret;
1625
1626         /*
1627          * rename don't use btrfs_join_transaction, so, once we
1628          * set the transaction to blocked above, we aren't going
1629          * to get any new ordered operations.  We can safely run
1630          * it here and no for sure that nothing new will be added
1631          * to the list
1632          */
1633         ret = btrfs_run_ordered_operations(trans, root, 1);
1634
1635         return ret;
1636 }
1637
1638 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1639 {
1640         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1641                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1642         return 0;
1643 }
1644
1645 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1646 {
1647         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1648                 btrfs_wait_ordered_roots(fs_info, -1);
1649 }
1650
1651 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1652                              struct btrfs_root *root)
1653 {
1654         struct btrfs_transaction *cur_trans = trans->transaction;
1655         struct btrfs_transaction *prev_trans = NULL;
1656         int ret;
1657
1658         ret = btrfs_run_ordered_operations(trans, root, 0);
1659         if (ret) {
1660                 btrfs_abort_transaction(trans, root, ret);
1661                 btrfs_end_transaction(trans, root);
1662                 return ret;
1663         }
1664
1665         /* Stop the commit early if ->aborted is set */
1666         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1667                 ret = cur_trans->aborted;
1668                 btrfs_end_transaction(trans, root);
1669                 return ret;
1670         }
1671
1672         /* make a pass through all the delayed refs we have so far
1673          * any runnings procs may add more while we are here
1674          */
1675         ret = btrfs_run_delayed_refs(trans, root, 0);
1676         if (ret) {
1677                 btrfs_end_transaction(trans, root);
1678                 return ret;
1679         }
1680
1681         btrfs_trans_release_metadata(trans, root);
1682         trans->block_rsv = NULL;
1683         if (trans->qgroup_reserved) {
1684                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1685                 trans->qgroup_reserved = 0;
1686         }
1687
1688         cur_trans = trans->transaction;
1689
1690         /*
1691          * set the flushing flag so procs in this transaction have to
1692          * start sending their work down.
1693          */
1694         cur_trans->delayed_refs.flushing = 1;
1695         smp_wmb();
1696
1697         if (!list_empty(&trans->new_bgs))
1698                 btrfs_create_pending_block_groups(trans, root);
1699
1700         ret = btrfs_run_delayed_refs(trans, root, 0);
1701         if (ret) {
1702                 btrfs_end_transaction(trans, root);
1703                 return ret;
1704         }
1705
1706         spin_lock(&root->fs_info->trans_lock);
1707         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1708                 spin_unlock(&root->fs_info->trans_lock);
1709                 atomic_inc(&cur_trans->use_count);
1710                 ret = btrfs_end_transaction(trans, root);
1711
1712                 wait_for_commit(root, cur_trans);
1713
1714                 btrfs_put_transaction(cur_trans);
1715
1716                 return ret;
1717         }
1718
1719         cur_trans->state = TRANS_STATE_COMMIT_START;
1720         wake_up(&root->fs_info->transaction_blocked_wait);
1721
1722         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1723                 prev_trans = list_entry(cur_trans->list.prev,
1724                                         struct btrfs_transaction, list);
1725                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1726                         atomic_inc(&prev_trans->use_count);
1727                         spin_unlock(&root->fs_info->trans_lock);
1728
1729                         wait_for_commit(root, prev_trans);
1730
1731                         btrfs_put_transaction(prev_trans);
1732                 } else {
1733                         spin_unlock(&root->fs_info->trans_lock);
1734                 }
1735         } else {
1736                 spin_unlock(&root->fs_info->trans_lock);
1737         }
1738
1739         extwriter_counter_dec(cur_trans, trans->type);
1740
1741         ret = btrfs_start_delalloc_flush(root->fs_info);
1742         if (ret)
1743                 goto cleanup_transaction;
1744
1745         ret = btrfs_flush_all_pending_stuffs(trans, root);
1746         if (ret)
1747                 goto cleanup_transaction;
1748
1749         wait_event(cur_trans->writer_wait,
1750                    extwriter_counter_read(cur_trans) == 0);
1751
1752         /* some pending stuffs might be added after the previous flush. */
1753         ret = btrfs_flush_all_pending_stuffs(trans, root);
1754         if (ret)
1755                 goto cleanup_transaction;
1756
1757         btrfs_wait_delalloc_flush(root->fs_info);
1758
1759         btrfs_scrub_pause(root);
1760         /*
1761          * Ok now we need to make sure to block out any other joins while we
1762          * commit the transaction.  We could have started a join before setting
1763          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1764          */
1765         spin_lock(&root->fs_info->trans_lock);
1766         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1767         spin_unlock(&root->fs_info->trans_lock);
1768         wait_event(cur_trans->writer_wait,
1769                    atomic_read(&cur_trans->num_writers) == 1);
1770
1771         /* ->aborted might be set after the previous check, so check it */
1772         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1773                 ret = cur_trans->aborted;
1774                 goto scrub_continue;
1775         }
1776         /*
1777          * the reloc mutex makes sure that we stop
1778          * the balancing code from coming in and moving
1779          * extents around in the middle of the commit
1780          */
1781         mutex_lock(&root->fs_info->reloc_mutex);
1782
1783         /*
1784          * We needn't worry about the delayed items because we will
1785          * deal with them in create_pending_snapshot(), which is the
1786          * core function of the snapshot creation.
1787          */
1788         ret = create_pending_snapshots(trans, root->fs_info);
1789         if (ret) {
1790                 mutex_unlock(&root->fs_info->reloc_mutex);
1791                 goto scrub_continue;
1792         }
1793
1794         /*
1795          * We insert the dir indexes of the snapshots and update the inode
1796          * of the snapshots' parents after the snapshot creation, so there
1797          * are some delayed items which are not dealt with. Now deal with
1798          * them.
1799          *
1800          * We needn't worry that this operation will corrupt the snapshots,
1801          * because all the tree which are snapshoted will be forced to COW
1802          * the nodes and leaves.
1803          */
1804         ret = btrfs_run_delayed_items(trans, root);
1805         if (ret) {
1806                 mutex_unlock(&root->fs_info->reloc_mutex);
1807                 goto scrub_continue;
1808         }
1809
1810         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1811         if (ret) {
1812                 mutex_unlock(&root->fs_info->reloc_mutex);
1813                 goto scrub_continue;
1814         }
1815
1816         /*
1817          * make sure none of the code above managed to slip in a
1818          * delayed item
1819          */
1820         btrfs_assert_delayed_root_empty(root);
1821
1822         WARN_ON(cur_trans != trans->transaction);
1823
1824         /* btrfs_commit_tree_roots is responsible for getting the
1825          * various roots consistent with each other.  Every pointer
1826          * in the tree of tree roots has to point to the most up to date
1827          * root for every subvolume and other tree.  So, we have to keep
1828          * the tree logging code from jumping in and changing any
1829          * of the trees.
1830          *
1831          * At this point in the commit, there can't be any tree-log
1832          * writers, but a little lower down we drop the trans mutex
1833          * and let new people in.  By holding the tree_log_mutex
1834          * from now until after the super is written, we avoid races
1835          * with the tree-log code.
1836          */
1837         mutex_lock(&root->fs_info->tree_log_mutex);
1838
1839         ret = commit_fs_roots(trans, root);
1840         if (ret) {
1841                 mutex_unlock(&root->fs_info->tree_log_mutex);
1842                 mutex_unlock(&root->fs_info->reloc_mutex);
1843                 goto scrub_continue;
1844         }
1845
1846         /*
1847          * Since the transaction is done, we should set the inode map cache flag
1848          * before any other comming transaction.
1849          */
1850         if (btrfs_test_opt(root, CHANGE_INODE_CACHE))
1851                 btrfs_set_opt(root->fs_info->mount_opt, INODE_MAP_CACHE);
1852         else
1853                 btrfs_clear_opt(root->fs_info->mount_opt, INODE_MAP_CACHE);
1854
1855         /* commit_fs_roots gets rid of all the tree log roots, it is now
1856          * safe to free the root of tree log roots
1857          */
1858         btrfs_free_log_root_tree(trans, root->fs_info);
1859
1860         ret = commit_cowonly_roots(trans, root);
1861         if (ret) {
1862                 mutex_unlock(&root->fs_info->tree_log_mutex);
1863                 mutex_unlock(&root->fs_info->reloc_mutex);
1864                 goto scrub_continue;
1865         }
1866
1867         /*
1868          * The tasks which save the space cache and inode cache may also
1869          * update ->aborted, check it.
1870          */
1871         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1872                 ret = cur_trans->aborted;
1873                 mutex_unlock(&root->fs_info->tree_log_mutex);
1874                 mutex_unlock(&root->fs_info->reloc_mutex);
1875                 goto scrub_continue;
1876         }
1877
1878         btrfs_prepare_extent_commit(trans, root);
1879
1880         cur_trans = root->fs_info->running_transaction;
1881
1882         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1883                             root->fs_info->tree_root->node);
1884         list_add_tail(&root->fs_info->tree_root->dirty_list,
1885                       &cur_trans->switch_commits);
1886
1887         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1888                             root->fs_info->chunk_root->node);
1889         list_add_tail(&root->fs_info->chunk_root->dirty_list,
1890                       &cur_trans->switch_commits);
1891
1892         switch_commit_roots(cur_trans, root->fs_info);
1893
1894         assert_qgroups_uptodate(trans);
1895         update_super_roots(root);
1896
1897         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1898         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1899         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1900                sizeof(*root->fs_info->super_copy));
1901
1902         spin_lock(&root->fs_info->trans_lock);
1903         cur_trans->state = TRANS_STATE_UNBLOCKED;
1904         root->fs_info->running_transaction = NULL;
1905         spin_unlock(&root->fs_info->trans_lock);
1906         mutex_unlock(&root->fs_info->reloc_mutex);
1907
1908         wake_up(&root->fs_info->transaction_wait);
1909
1910         ret = btrfs_write_and_wait_transaction(trans, root);
1911         if (ret) {
1912                 btrfs_error(root->fs_info, ret,
1913                             "Error while writing out transaction");
1914                 mutex_unlock(&root->fs_info->tree_log_mutex);
1915                 goto scrub_continue;
1916         }
1917
1918         ret = write_ctree_super(trans, root, 0);
1919         if (ret) {
1920                 mutex_unlock(&root->fs_info->tree_log_mutex);
1921                 goto scrub_continue;
1922         }
1923
1924         /*
1925          * the super is written, we can safely allow the tree-loggers
1926          * to go about their business
1927          */
1928         mutex_unlock(&root->fs_info->tree_log_mutex);
1929
1930         btrfs_finish_extent_commit(trans, root);
1931
1932         root->fs_info->last_trans_committed = cur_trans->transid;
1933         /*
1934          * We needn't acquire the lock here because there is no other task
1935          * which can change it.
1936          */
1937         cur_trans->state = TRANS_STATE_COMPLETED;
1938         wake_up(&cur_trans->commit_wait);
1939
1940         spin_lock(&root->fs_info->trans_lock);
1941         list_del_init(&cur_trans->list);
1942         spin_unlock(&root->fs_info->trans_lock);
1943
1944         btrfs_put_transaction(cur_trans);
1945         btrfs_put_transaction(cur_trans);
1946
1947         if (trans->type & __TRANS_FREEZABLE)
1948                 sb_end_intwrite(root->fs_info->sb);
1949
1950         trace_btrfs_transaction_commit(root);
1951
1952         btrfs_scrub_continue(root);
1953
1954         if (current->journal_info == trans)
1955                 current->journal_info = NULL;
1956
1957         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1958
1959         if (current != root->fs_info->transaction_kthread)
1960                 btrfs_run_delayed_iputs(root);
1961
1962         return ret;
1963
1964 scrub_continue:
1965         btrfs_scrub_continue(root);
1966 cleanup_transaction:
1967         btrfs_trans_release_metadata(trans, root);
1968         trans->block_rsv = NULL;
1969         if (trans->qgroup_reserved) {
1970                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1971                 trans->qgroup_reserved = 0;
1972         }
1973         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
1974         if (current->journal_info == trans)
1975                 current->journal_info = NULL;
1976         cleanup_transaction(trans, root, ret);
1977
1978         return ret;
1979 }
1980
1981 /*
1982  * return < 0 if error
1983  * 0 if there are no more dead_roots at the time of call
1984  * 1 there are more to be processed, call me again
1985  *
1986  * The return value indicates there are certainly more snapshots to delete, but
1987  * if there comes a new one during processing, it may return 0. We don't mind,
1988  * because btrfs_commit_super will poke cleaner thread and it will process it a
1989  * few seconds later.
1990  */
1991 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
1992 {
1993         int ret;
1994         struct btrfs_fs_info *fs_info = root->fs_info;
1995
1996         spin_lock(&fs_info->trans_lock);
1997         if (list_empty(&fs_info->dead_roots)) {
1998                 spin_unlock(&fs_info->trans_lock);
1999                 return 0;
2000         }
2001         root = list_first_entry(&fs_info->dead_roots,
2002                         struct btrfs_root, root_list);
2003         list_del_init(&root->root_list);
2004         spin_unlock(&fs_info->trans_lock);
2005
2006         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2007
2008         btrfs_kill_all_delayed_nodes(root);
2009
2010         if (btrfs_header_backref_rev(root->node) <
2011                         BTRFS_MIXED_BACKREF_REV)
2012                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2013         else
2014                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2015         /*
2016          * If we encounter a transaction abort during snapshot cleaning, we
2017          * don't want to crash here
2018          */
2019         return (ret < 0) ? 0 : 1;
2020 }